From fcc1e5a7656186a8c8de64f15af8eba370a7f4db Mon Sep 17 00:00:00 2001
From: yao <yao.zhang@greenwaves-technologies.com>
Date: Thu, 16 Dec 2021 18:53:38 +0100
Subject: [PATCH 1/3] [SDK] Update SDK
- GAP8: fix the i2s1 issue in freeRTOS
- NNTool: Fixes in ProcessArgPadding, Quantizer generator
---
configs/common.sh | 4 +
examples/autotiler/BilinearResize/Makefile | 2 +-
examples/nntool/mnist_gru/Makefile | 5 +-
examples/nntool/mnist_rnn/Makefile | 3 +
.../cpu/iss/vp/include/platform_wrapper.hpp | 2 +-
.../models/pulp/udma/i2s/udma_i2s_v3.cpp | 15 +-
.../models/pulp/udma/udma_v4_addrgens.cpp | 22 +-
.../models/pulp/udma/udma_v4_rx_channels.cpp | 3 +-
.../models/pulp/udma/udma_v4_tx_channels.cpp | 5 +-
libs/gap_lib/testbench/testlib.c | 376 ++++-
libs/gap_lib/testbench/testlib.h | 19 +-
.../pmsis/drivers/udma/i2s/i2s_internal.c | 6 +-
.../vendors/gwt/libs/include/string.h | 8 +
.../vendors/gwt/pmsis/include/pmsis.h | 5 +
.../gwt/pmsis/rtos/include/pmsis/rtos/os/os.h | 8 +
.../vendors/gwt/rules/freeRTOS_rules.mk | 9 +-
.../pmsis_api/include/pmsis/drivers/i2s.h | 187 ++-
rtos/pulp/pulpos-2/include/pmsis.h | 2 +
rtos/pulp/pulpos-2/include/pos/data/data.h | 9 +
.../pulp/pulpos-2/include/pos/implem/implem.h | 1 +
rtos/pulp/pulpos-2/rules/pulpos/src.mk | 2 +-
testplan.cfg | 2 +
.../CNN_Generators/CNN_Copy_Generators.c | 90 +-
.../CNN_Generators_NE16/CNN_Generators_NE16.c | 90 +-
.../CNN_Generators_NE16/CNN_Generators_NE16.h | 2 +-
.../CNN_Generators_SQ8/CNN_Generators_SQ8.c | 390 ++++-
.../CNN_Generators_SQ8/CNN_Generators_SQ8.h | 128 ++
.../CNN_Generators_fp16/CNN_Generators_fp16.c | 482 ++++--
.../CNN_Generators_fp16/CNN_Generators_fp16.h | 184 ++-
tools/autotiler_v3/CNN_Libraries/CNN_Copy.c | 246 ++-
tools/autotiler_v3/CNN_Libraries/CNN_Copy.h | 82 +
.../CNN_BasicKernels_NE16.c | 104 +-
.../CNN_BasicKernels_NE16.h | 4 +-
.../CNN_Libraries_SQ8/CNN_BasicKernels_SQ8.h | 3 +
.../CNN_Libraries_SQ8/CNN_MatAlgebra_SQ8.c | 34 +-
.../CNN_Libraries_SQ8/CNN_MatMul_Conv_SQ8.c | 1313 ++++++++++++-----
.../CNN_BasicKernels_fp16.h | 12 +-
.../CNN_Libraries_fp16/CNN_MatAlgebra_fp16.c | 14 +-
.../CNN_Libraries_fp16/CNN_MatMul_Conv_fp16.c | 246 ++-
.../autotiler_v3/DSP_Libraries/FloatDefines.h | 1 +
.../LUT_Tables/gen_scripts/GenMFCCLUT.py | 2 +-
tools/autotiler_v3/Emulation/GapBuiltins.h | 21 +
.../BilinearResizes/ResizeBasicKernels.c | 85 ++
.../BilinearResizes/ResizeBasicKernels.h | 21 +
.../BilinearResizes/ResizeGenerator.c | 71 +
.../BilinearResizes/ResizeGenerator.h | 1 +
tools/autotiler_v3/Makefile | 2 +-
tools/autotiler_v3/version.cfg | 2 +-
tools/jenkins/gap_sdk_version.txt | 2 +-
tools/nntool/.gitignore | 3 +
tools/nntool/generation/code_block.py | 6 +-
.../general/resizer_kernel_generator.py | 13 +-
.../pow2/conv_pool_relu_kernels_generator.py | 2 +-
.../new_generators/general/dsp_generators.py | 2 +-
.../general/quantize_parameters.py | 24 +-
.../new_generators/mult8/matmul_mult8.py | 18 +-
tools/nntool/graph/dim.py | 6 +
.../graph/manipulations/adjust_order.py | 4 +-
.../nntool/graph/manipulations/dimensions.py | 2 +-
.../eliminate_transposes.py | 406 +++--
.../eliminate_transposes_actions.py | 73 +-
.../graph/matches/matchers/move_node_up.py | 33 +-
.../graph/matches/matchers/slice_to_split.py | 2 +-
tools/nntool/graph/nngraph.py | 11 +-
tools/nntool/graph/types/conv2d.py | 9 +-
tools/nntool/graph/types/dsp_preprocessing.py | 8 +-
tools/nntool/graph/types/input_output.py | 8 +-
tools/nntool/graph/types/pooling.py | 14 +-
tools/nntool/graph/types/tensor_arithmetic.py | 8 +-
.../onnx/handlers/backend/conv_mixin.py | 159 +-
.../importer/onnx/handlers/backend/gather.py | 12 +-
.../onnx/handlers/backend/mat_mul_mixin.py | 8 +-
.../onnx/handlers/backend/pad_mixin.py | 1 +
.../float/kernels/matrix_operations.py | 5 +-
.../float/quantizers/conv_fusion_float.py | 13 +-
.../multiplicative/quantizers/filter_mult.py | 13 +-
.../symmetric/kernels/dsp_preprocessing.py | 8 +-
.../symmetric/kernels/matrix_operations.py | 6 +-
tools/nntool/utils/gen_twiddles.py | 15 +-
utils/gapy/run.py | 5 +
utils/gapy/runner/board/board_runner.py | 13 +-
utils/rules/pulp_rules.mk | 10 +-
82 files changed, 4110 insertions(+), 1137 deletions(-)
diff --git a/configs/common.sh b/configs/common.sh
index 981aff43a..e3d675d5d 100644
--- a/configs/common.sh
+++ b/configs/common.sh
@@ -139,3 +139,7 @@ else
export XCSIM_PLATFORM=$XCSIM_PATH
fi
fi
+
+if [ -f "$GAP_SDK_HOME/configs/wsl.sh" ]; then
+ source $GAP_SDK_HOME/configs/wsl.sh
+fi
diff --git a/examples/autotiler/BilinearResize/Makefile b/examples/autotiler/BilinearResize/Makefile
index f7259a0d4..91e74b484 100644
--- a/examples/autotiler/BilinearResize/Makefile
+++ b/examples/autotiler/BilinearResize/Makefile
@@ -12,7 +12,7 @@ RESIZE_KER_PATH = $(TILER_BILINEAR_RESIZE_KERNEL_PATH)
APP_SRCS += Bilinear_Resize.c ImgIO.c
APP_SRCS += ResizeKernels.c $(RESIZE_KER_PATH)/ResizeBasicKernels.c
-APP_INC += . $(TILER_EMU_INC) $(TILER_INC) $(RESIZE_KER_PATH)
+APP_INC += . $(TILER_EMU_INC) $(TILER_INC) $(RESIZE_KER_PATH) $(TILER_DSP_KERNEL_PATH) $(TILER_CNN_KERNEL_PATH_FP16)
APP_CFLAGS += -O3 -mno-memcpy -fno-tree-loop-distribute-patterns $(USER_FLAGS)
APP_CFLAGS += -Wno-maybe-uninitialized -Wno-unused-but-set-variable -Wno-unused-variable
diff --git a/examples/nntool/mnist_gru/Makefile b/examples/nntool/mnist_gru/Makefile
index ca053caf2..576063fae 100644
--- a/examples/nntool/mnist_gru/Makefile
+++ b/examples/nntool/mnist_gru/Makefile
@@ -64,7 +64,10 @@ PULP_APP = mnist
APP = mnist
APP_SRCS += $(MODEL_PREFIX).c $(MODEL_GEN_C) $(MODEL_COMMON_SRCS) $(CNN_LIB)
-APP_CFLAGS += -g -O3 -mno-memcpy -fno-tree-loop-distribute-patterns
+ifeq '$(TARGET_CHIP_FAMILY)' 'GAP8'
+ APP_CFLAGS += -march=rv32imcxgap8
+endif
+APP_CFLAGS += -g -O3 -mno-memcpy -fno-tree-loop-distribute-patterns
APP_CFLAGS += -I. -I$(MODEL_COMMON_INC) -I$(TILER_EMU_INC) -I$(TILER_INC) $(CNN_LIB_INCLUDE) -I$(MODEL_BUILD)
APP_CFLAGS += -DPERF -DAT_MODEL_PREFIX=$(MODEL_PREFIX) $(MODEL_SIZE_CFLAGS)
APP_CFLAGS += -DSTACK_SIZE=$(CLUSTER_STACK_SIZE) -DSLAVE_STACK_SIZE=$(CLUSTER_SLAVE_STACK_SIZE)
diff --git a/examples/nntool/mnist_rnn/Makefile b/examples/nntool/mnist_rnn/Makefile
index 16ee1c51b..04db45099 100644
--- a/examples/nntool/mnist_rnn/Makefile
+++ b/examples/nntool/mnist_rnn/Makefile
@@ -62,6 +62,9 @@ PULP_APP = mnist
APP = mnist
APP_SRCS += $(MODEL_PREFIX).c $(MODEL_GEN_C) $(MODEL_COMMON_SRCS) $(CNN_LIB)
+ifeq '$(TARGET_CHIP_FAMILY)' 'GAP8'
+ APP_CFLAGS += -march=rv32imcxgap8
+endif
APP_CFLAGS += -g -O3 -mno-memcpy -fno-tree-loop-distribute-patterns
APP_CFLAGS += -I. -I$(MODEL_COMMON_INC) -I$(TILER_EMU_INC) -I$(TILER_INC) $(CNN_LIB_INCLUDE) -I$(MODEL_BUILD)
APP_CFLAGS += -DPERF -DAT_MODEL_PREFIX=$(MODEL_PREFIX) $(MODEL_SIZE_CFLAGS)
diff --git a/gvsoc/gvsoc/models/cpu/iss/vp/include/platform_wrapper.hpp b/gvsoc/gvsoc/models/cpu/iss/vp/include/platform_wrapper.hpp
index a386c3768..f1358d239 100644
--- a/gvsoc/gvsoc/models/cpu/iss/vp/include/platform_wrapper.hpp
+++ b/gvsoc/gvsoc/models/cpu/iss/vp/include/platform_wrapper.hpp
@@ -170,7 +170,7 @@ static inline int iss_fetch_req_common(iss_t *_this, uint64_t addr, uint8_t *dat
if (err != vp::IO_REQ_OK)
{
if (err == vp::IO_REQ_INVALID)
- _this->trace.fatal("Invalid fetch request (addr: 0x%x, size: 0x%x)\n", addr, size);
+ _this->trace.force_warning("Invalid fetch request (addr: 0x%x, size: 0x%x)\n", addr, size);
else
{
iss_exec_insn_stall(_this);
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/i2s/udma_i2s_v3.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/i2s/udma_i2s_v3.cpp
index 3296d1307..8eecde9f1 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/i2s/udma_i2s_v3.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/i2s/udma_i2s_v3.cpp
@@ -239,29 +239,32 @@ void I2s_periph::handle_sdo(bool no_restart)
{
this->tx_wait_data_init &= ~(1 << this->active_channel);
data = this->tx_fifo[this->active_channel].front();
-
this->tx_fifo[this->active_channel].pop();
}
else
{
if (((this->tx_wait_data_init >> this->active_channel) & 1) == 0)
{
+ this->trace.msg(vp::trace::LEVEL_DEBUG, "Generating TX error (slot: %d)\n", this->active_channel);
this->regmap.err_status.set(this->regmap.err_status.get() | (1 << (this->active_channel + 16)));
}
}
this->tx_pending_value = this->handle_tx_format(channel, data);
- this->trace.msg(vp::trace::LEVEL_DEBUG, "Got new TX sample (value: 0x%x, width: %d)\n", this->tx_pending_value, this->tx_pending_bits);
-
+ this->trace.msg(vp::trace::LEVEL_DEBUG, "Got new TX sample (slot: %d, value: 0x%x, width: %d)\n", this->active_channel, this->tx_pending_value, this->tx_pending_bits);
// Now ask the next sample to the channel so that it is ready for the
// next frame.
// The channel id is pushed now since we the samples are received in order
// and we don't know what is the slot when we receive the sample from the channel
- this->tx_fifo_slot_id.push(this->active_channel);
- channel->get_data(channel->slot_cfg->tx_dsize_get() + 1, channel->slot_cfg->tx_id_get());
+ if (channel->is_active() || channel->slot_cfg->tx_id_get() >= 0xe0)
+ {
+ this->tx_fifo_slot_id.push(this->active_channel);
+
+ channel->get_data(channel->slot_cfg->tx_dsize_get() + 1, channel->slot_cfg->tx_id_get());
+ }
}
if (this->tx_pending_bits > 0)
@@ -1065,7 +1068,7 @@ void I2s_tx_channel::push_data(uint8_t *data, int size)
uint32_t value = 0;
memcpy((void *)&value, (void *)data, size);
- this->periph->trace.msg(vp::trace::LEVEL_INFO, "Received TX sample from memory (value: 0x%x)\n", value);
+ this->periph->trace.msg(vp::trace::LEVEL_INFO, "Received TX sample from memory (slot: %d, value: 0x%x)\n", this->periph->tx_fifo_slot_id.front(), value);
this->periph->tx_fifo[this->periph->tx_fifo_slot_id.front()].push(value);
this->periph->tx_fifo_slot_id.pop();
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_addrgens.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_addrgens.cpp
index a427b1787..43d48b9cb 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_addrgens.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_addrgens.cpp
@@ -75,7 +75,7 @@ void Udma_addrgen_linear::check_pending_transfer()
{
if (!this->active_transfer && this->nb_pending_transfers)
{
- trace.msg(vp::trace::LEVEL_TRACE, "Starting new buffer (addr: 0x%x, size: 0x%x)\n", this->pending_addr, this->pending_size);
+ trace.msg(vp::trace::LEVEL_TRACE, "Starting new buffer (addr: 0x%x, size: 0x%x, pending: %d)\n", this->pending_addr, this->pending_size, this->nb_pending_transfers);
this->nb_pending_transfers--;
this->set_active_transfer(true);
@@ -101,17 +101,25 @@ void Udma_addrgen_linear::cfg_ctrl_req(uint64_t reg_offset, int size, uint8_t *v
}
else if (this->regmap.cfg_ctrl.en_get())
{
+ this->trace.msg(vp::trace::LEVEL_TRACE, "Enqeueing transfer (pending: %d)\n", this->nb_pending_transfers);
- if (this->nb_pending_transfers > 0)
+ if (this->nb_pending_transfers == 1)
{
- this->remaining_size -= this->pending_size;
+ this->trace.force_warning("Trying to enqueue while alreay 2 transfers are enqueued\n");
}
+ else
+ {
+ if (this->nb_pending_transfers > 0)
+ {
+ this->remaining_size -= this->pending_size;
+ }
- this->pending_addr = this->regmap.cfg_sa_buf0.get();
- this->pending_size = this->regmap.cfg_size.get();
- this->remaining_size += this->pending_size;
+ this->pending_addr = this->regmap.cfg_sa_buf0.get();
+ this->pending_size = this->regmap.cfg_size.get();
+ this->remaining_size += this->pending_size;
- this->nb_pending_transfers++;
+ this->nb_pending_transfers++;
+ }
this->check_pending_transfer();
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_rx_channels.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_rx_channels.cpp
index fc2fb58b0..9553e0ccc 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_rx_channels.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_rx_channels.cpp
@@ -186,8 +186,7 @@ void Udma_rx_channels::handle_pending(void *__this, vp::clock_event *event)
if (addr == 0)
{
- fflush(NULL);
- abort();
+ _this->top->trace.fatal("UDMA trying to access NULL\n");
}
if (err == vp::IO_REQ_OK)
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_tx_channels.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_tx_channels.cpp
index 90e97e6d4..202cd2f10 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_tx_channels.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_tx_channels.cpp
@@ -83,7 +83,7 @@ void Udma_tx_channel::get_data(int size, int channel)
{
this->requested_size_queue.push_back(size);
this->requested_size += size;
- if (this->requested_size_queue.size() == 1 && this->is_active() && !this->enqueued)
+ if (this->requested_size_queue.size() > 0 && this->is_active() && !this->enqueued)
{
this->top->tx_channels->push_ready_channel(this);
this->check_state();
@@ -243,7 +243,7 @@ void Udma_tx_channels::handle_pending(void *__this, vp::clock_event *event)
channel->requested_size_queue.push_front(requested_size_queue - size);
}
- if (channel->requested_size_queue.size() && !channel->enqueued)
+ if (channel->requested_size_queue.size() && !channel->enqueued && channel->is_active())
{
_this->push_ready_channel(channel);
}
@@ -307,7 +307,6 @@ void Udma_tx_channels::check_state()
{
if (!this->send_reqs_event->is_enqueued())
{
- if (!this->pending_channels.empty() && !this->l2_free_reqs->is_empty())
if (!this->pending_channels.empty() && !this->l2_free_reqs->is_empty() && this->pending_channels.front()->is_active())
{
this->top->event_enqueue(this->send_reqs_event, 1);
diff --git a/libs/gap_lib/testbench/testlib.c b/libs/gap_lib/testbench/testlib.c
index 5cfd79862..1619bdc2a 100644
--- a/libs/gap_lib/testbench/testlib.c
+++ b/libs/gap_lib/testbench/testlib.c
@@ -185,7 +185,14 @@ void i2s_slot_deinit(i2s_slot_test_t *i2s_slot)
struct pi_i2s_channel_conf i2s_conf;
pi_i2s_channel_conf_init(&i2s_conf);
i2s_conf.options = PI_I2S_OPT_DISABLED | (i2s_slot->is_rx ? PI_I2S_OPT_IS_RX: PI_I2S_OPT_IS_TX);
- pi_i2s_channel_conf_set(i2s_slot->i2s, i2s_slot->slot, &i2s_conf);
+ if (i2s_slot->frame)
+ {
+ pi_i2s_frame_channel_conf_set(i2s_slot->i2s, i2s_slot->frame, i2s_slot->slot, &i2s_conf);
+ }
+ else
+ {
+ pi_i2s_channel_conf_set(i2s_slot->i2s, i2s_slot->slot, &i2s_conf);
+ }
pi_l2_free(i2s_slot->buffers[0], i2s_slot->buffer_size);
pi_l2_free(i2s_slot->buffers[1], i2s_slot->buffer_size);
@@ -193,7 +200,12 @@ void i2s_slot_deinit(i2s_slot_test_t *i2s_slot)
}
-int i2s_slot_init(i2s_slot_test_t *i2s_slot, struct pi_device *i2s, i2s_slot_config_t *slot_config)
+void i2s_slot_new(i2s_slot_test_t *i2s_slot)
+{
+ pi_task_block(&i2s_slot->end_task);
+}
+
+int i2s_slot_init(i2s_test_t *test, i2s_slot_test_t *i2s_slot, struct pi_device *i2s, i2s_slot_config_t *slot_config)
{
i2s_slot->i2s = i2s;
i2s_slot->itf = slot_config->itf;
@@ -209,10 +221,12 @@ int i2s_slot_init(i2s_slot_test_t *i2s_slot, struct pi_device *i2s, i2s_slot_con
i2s_slot->random_mute = slot_config->random_mute;
i2s_slot->format = slot_config->format;
i2s_slot->bypass = slot_config->bypass;
+ i2s_slot->frame = slot_config->frame;
+ i2s_slot->test = test;
+ i2s_slot->flags.use_slab = slot_config->slab != 0;
if (!i2s_slot->bypass || !i2s_slot->is_rx)
{
- pi_task_block(&i2s_slot->end_task);
struct pi_i2s_channel_conf i2s_conf;
pi_i2s_channel_conf_init(&i2s_conf);
@@ -229,21 +243,85 @@ int i2s_slot_init(i2s_slot_test_t *i2s_slot, struct pi_device *i2s, i2s_slot_con
{
i2s_conf.options = PI_I2S_OPT_PINGPONG | PI_I2S_OPT_IS_TX | PI_I2S_OPT_ENABLED;
}
- i2s_slot->buffers[0] = pi_l2_malloc(buffer_size);
- if (i2s_slot->buffers[0] == NULL)
+
+ if (slot_config->slab)
{
- printf("Failed to allocate\n");
- return -1;
+ if (slot_config->frame)
+ {
+ if (__FF1(slot_config->frame) == i2s_slot->slot)
+ {
+ int nb_slots = __builtin_popcount(slot_config->frame);
+
+ void *buffers = pi_l2_malloc(buffer_size * nb_slots * slot_config->slab);
+ if (buffers == NULL)
+ {
+ printf("Failed to allocate\n");
+ return -1;
+ }
+
+ pi_mem_slab_init(&i2s_slot->slab, buffers, buffer_size * nb_slots, slot_config->slab);
+ }
+ }
+ else
+ {
+ void *buffers = pi_l2_malloc(buffer_size * slot_config->slab);
+ if (buffers == NULL)
+ {
+ printf("Failed to allocate\n");
+ return -1;
+ }
+
+ pi_mem_slab_init(&i2s_slot->slab, buffers, buffer_size, slot_config->slab);
+ }
}
- i2s_slot->buffers[1] = pi_l2_malloc(buffer_size);
- if (i2s_slot->buffers[1] == NULL)
+ else
{
- printf("Failed to allocate\n");
- return -1;
+ if (slot_config->frame)
+ {
+ if (__FF1(slot_config->frame) == i2s_slot->slot)
+ {
+ int nb_slots = __builtin_popcount(slot_config->frame);
+
+ i2s_slot->buffers[0] = pi_l2_malloc(buffer_size*nb_slots);
+ if (i2s_slot->buffers[0] == NULL)
+ {
+ printf("Failed to allocate\n");
+ return -1;
+ }
+ i2s_slot->buffers[1] = pi_l2_malloc(buffer_size*nb_slots);
+ if (i2s_slot->buffers[1] == NULL)
+ {
+ printf("Failed to allocate\n");
+ return -1;
+ }
+ }
+ }
+ else
+ {
+ i2s_slot->buffers[0] = pi_l2_malloc(buffer_size);
+ if (i2s_slot->buffers[0] == NULL)
+ {
+ printf("Failed to allocate\n");
+ return -1;
+ }
+ i2s_slot->buffers[1] = pi_l2_malloc(buffer_size);
+ if (i2s_slot->buffers[1] == NULL)
+ {
+ printf("Failed to allocate\n");
+ return -1;
+ }
+ }
}
- i2s_conf.pingpong_buffers[0] = i2s_slot->buffers[0];
- i2s_conf.pingpong_buffers[1] = i2s_slot->buffers[1];
+ if (slot_config->slab)
+ {
+ i2s_conf.mem_slab = &i2s_slot->slab;
+ }
+ else
+ {
+ i2s_conf.pingpong_buffers[0] = i2s_slot->buffers[0];
+ i2s_conf.pingpong_buffers[1] = i2s_slot->buffers[1];
+ }
i2s_conf.block_size = buffer_size;
i2s_conf.word_size = slot_config->word_size;
@@ -287,8 +365,16 @@ int i2s_slot_init(i2s_slot_test_t *i2s_slot, struct pi_device *i2s, i2s_slot_con
i2s_conf.slot_enable = 0;
}
- if (pi_i2s_channel_conf_set(i2s, slot_config->slot, &i2s_conf))
- return -1;
+ if (slot_config->frame)
+ {
+ if (pi_i2s_frame_channel_conf_set(i2s, slot_config->frame, slot_config->slot, &i2s_conf))
+ return -1;
+ }
+ else
+ {
+ if (pi_i2s_channel_conf_set(i2s, slot_config->slot, &i2s_conf))
+ return -1;
+ }
if (slot_config -> ts_evt_en)
{
@@ -343,14 +429,53 @@ void i2s_slot_callback_tx_file_dumper(void *arg)
if (i2s_slot->nb_sample > 0 || i2s_slot->nb_sample == -1)
{
- void *buffer = i2s_slot->tx_buffers[i2s_slot->tx_buffer];
- i2s_slot->tx_buffer ^= 1;
- for (int i=0; i<i2s_slot->nb_elem; i++)
+ void *buffer;
+
+ if (i2s_slot->flags.use_slab)
+ {
+ pi_mem_slab_alloc(&i2s_slot->slab, (void **)&buffer, 0);
+ }
+ else
{
- set_buffer_elem_iter(i2s_slot, buffer, i);
+ buffer = i2s_slot->tx_buffers[i2s_slot->tx_buffer];
+ i2s_slot->tx_buffer ^= 1;
}
- int err = pi_i2s_channel_write_async(i2s_slot->i2s, i2s_slot->slot, NULL, i2s_slot->buffer_size, pi_task_callback(&i2s_slot->task, i2s_slot_callback_tx_file_dumper, (void *)i2s_slot));
+ if (i2s_slot->frame)
+ {
+ uint32_t frame = i2s_slot->frame;
+ uint32_t addr = (uint32_t)buffer;
+ while(frame)
+ {
+ int slot_id = __FF1(frame);
+ frame = __BITCLR_R(frame, 1, slot_id);
+
+ for (int i=0; i<i2s_slot->nb_elem; i++)
+ {
+ set_buffer_elem_iter(&i2s_slot->test->slot_test_tx[slot_id], (void *)addr, i);
+ }
+
+ addr = addr + i2s_slot->buffer_size;
+ }
+ }
+ else
+ {
+ for (int i=0; i<i2s_slot->nb_elem; i++)
+ {
+ set_buffer_elem_iter(i2s_slot, buffer, i);
+ }
+ }
+
+
+
+ if (i2s_slot->frame)
+ {
+ int err = pi_i2s_frame_write_async(i2s_slot->i2s, i2s_slot->frame, buffer, i2s_slot->buffer_size, pi_task_callback(&i2s_slot->task, i2s_slot_callback_tx_file_dumper, (void *)i2s_slot));
+ }
+ else
+ {
+ int err = pi_i2s_channel_write_async(i2s_slot->i2s, i2s_slot->slot, buffer, i2s_slot->buffer_size, pi_task_callback(&i2s_slot->task, i2s_slot_callback_tx_file_dumper, (void *)i2s_slot));
+ }
}
else
{
@@ -389,18 +514,8 @@ static uint32_t buffer_get_elem(void *buffer, int index, int word_size, void **a
}
}
-void i2s_slot_callback_rx_iter(void *arg)
+int i2s_slot_callback_rx_iter_check(i2s_slot_test_t *i2s_slot, void *chunk, int size)
{
- i2s_slot_test_t *i2s_slot = (i2s_slot_test_t *)arg;
- void *chunk;
- int size;
-
- if (pi_i2s_read_status(&i2s_slot->task, &chunk, (size_t *)&size))
- {
- i2s_slot->retval++;
- goto end;
- }
-
int nb_elem = size / i2s_slot->elem_size;
for (int i=0; i<nb_elem && i2s_slot->nb_sample > 0; i++)
@@ -428,15 +543,65 @@ void i2s_slot_callback_rx_iter(void *arg)
{
printf("Detected error (itf: %d, slot: %d, index: %d, nb_elem: %d, expected: 0x%x, got: 0x%x, address: %p)\n", i2s_slot->itf, i2s_slot->slot, i, nb_elem, current_value, value, address);
i2s_slot->retval++;
- goto end;
+ return 1;
}
i2s_slot->nb_sample--;
if (i2s_slot->nb_sample == 0)
+ return 1;
+ }
+
+ return 0;
+}
+
+
+void i2s_slot_callback_rx_iter(void *arg)
+{
+ i2s_slot_test_t *i2s_slot = (i2s_slot_test_t *)arg;
+ void *chunk;
+ int size;
+
+ if (pi_i2s_read_status(&i2s_slot->task, &chunk, (size_t *)&size))
+ {
+ i2s_slot->retval++;
+ goto end;
+ }
+
+ if (i2s_slot->frame)
+ {
+ uint32_t frame = i2s_slot->frame;
+ uint32_t buffer = (uint32_t)chunk;
+ while(frame)
+ {
+ int slot_id = __FF1(frame);
+ frame = __BITCLR_R(frame, 1, slot_id);
+
+ if (i2s_slot_callback_rx_iter_check(&i2s_slot->test->slot_test_rx[slot_id], (void *)buffer, size))
+ goto end;
+
+ buffer += size;
+ }
+
+ pi_i2s_frame_read_async(i2s_slot->i2s, i2s_slot->frame, pi_task_callback(&i2s_slot->task, i2s_slot_callback_rx_iter, (void *)i2s_slot));
+
+ if (i2s_slot->flags.use_slab)
+ {
+ pi_mem_slab_free(&i2s_slot->slab, &chunk);
+ }
+ }
+ else
+ {
+ if (i2s_slot_callback_rx_iter_check(i2s_slot, chunk, size))
goto end;
+
+ pi_i2s_channel_read_async(i2s_slot->i2s, i2s_slot->slot, pi_task_callback(&i2s_slot->task, i2s_slot_callback_rx_iter, (void *)i2s_slot));
+
+ if (i2s_slot->flags.use_slab)
+ {
+ pi_mem_slab_free(&i2s_slot->slab, &chunk);
+ }
}
- pi_i2s_channel_read_async(i2s_slot->i2s, i2s_slot->slot, pi_task_callback(&i2s_slot->task, i2s_slot_callback_rx_iter, (void *)i2s_slot));
return;
end:
@@ -471,7 +636,17 @@ int i2s_slot_start(i2s_slot_test_t *i2s_slot, i2s_slot_start_config_t *config)
if (i2s_slot->incr_value >= i2s_slot->incr_end)
i2s_slot->incr_value = 0;
- pi_i2s_channel_read_async(i2s_slot->i2s, i2s_slot->slot, pi_task_callback(&i2s_slot->task, i2s_slot_callback_rx_iter, (void *)i2s_slot));
+ if (i2s_slot->frame == 0 || __FF1(i2s_slot->frame) == i2s_slot->slot)
+ {
+ if (i2s_slot->frame)
+ {
+ pi_i2s_frame_read_async(i2s_slot->i2s, i2s_slot->frame, pi_task_callback(&i2s_slot->task, i2s_slot_callback_rx_iter, (void *)i2s_slot));
+ }
+ else
+ {
+ pi_i2s_channel_read_async(i2s_slot->i2s, i2s_slot->slot, pi_task_callback(&i2s_slot->task, i2s_slot_callback_rx_iter, (void *)i2s_slot));
+ }
+ }
}
else if (config->type == I2S_VERIF_TX_ITER)
{
@@ -484,16 +659,92 @@ int i2s_slot_start(i2s_slot_test_t *i2s_slot, i2s_slot_start_config_t *config)
if (i2s_slot->incr_value >= i2s_slot->incr_end)
i2s_slot->incr_value = 0;
- for (int i=0; i<i2s_slot->nb_elem; i++)
+ void *buffers[2] = {0};
+
+ if (i2s_slot->frame)
{
- set_buffer_elem_iter(i2s_slot, i2s_slot->tx_buffers[0], i);
+ if (__FL1(i2s_slot->frame) == i2s_slot->slot)
+ {
+ i2s_slot_test_t *first_slot = &i2s_slot->test->slot_test_tx[__FF1(i2s_slot->frame)];
+ uint32_t frame = i2s_slot->frame;
+ if (i2s_slot->flags.use_slab)
+ {
+ pi_mem_slab_alloc(&first_slot->slab, (void **)&buffers[0], 0);
+ pi_mem_slab_alloc(&first_slot->slab, (void **)&buffers[1], 0);
+ }
+ else
+ {
+ buffers[0] = first_slot->tx_buffers[0];
+ buffers[1] = first_slot->tx_buffers[1];
+ }
+
+ uint32_t buffer0 = (uint32_t)buffers[0];
+ uint32_t buffer1 = (uint32_t)buffers[1];
+
+ while(frame)
+ {
+ int slot_id = __FF1(frame);
+ frame = __BITCLR_R(frame, 1, slot_id);
+
+ i2s_slot_test_t *frame_slot = &i2s_slot->test->slot_test_tx[slot_id];
+
+
+ for (int i=0; i<i2s_slot->nb_elem; i++)
+ {
+ set_buffer_elem_iter(frame_slot, (void *)buffer0, i);
+ }
+
+ for (int i=0; i<i2s_slot->nb_elem; i++)
+ {
+ set_buffer_elem_iter(frame_slot, (void *)buffer1, i);
+ }
+
+ buffer0 += i2s_slot->buffer_size;
+ buffer1 += i2s_slot->buffer_size;
+ }
+ }
}
- for (int i=0; i<i2s_slot->nb_elem; i++)
+ else
{
- set_buffer_elem_iter(i2s_slot, i2s_slot->tx_buffers[1], i);
+ if (i2s_slot->flags.use_slab)
+ {
+ pi_mem_slab_alloc(&i2s_slot->slab, (void **)&buffers[0], 0);
+ pi_mem_slab_alloc(&i2s_slot->slab, (void **)&buffers[1], 0);
+ }
+ else
+ {
+ buffers[0] = i2s_slot->tx_buffers[0];
+ buffers[1] = i2s_slot->tx_buffers[1];
+ }
+
+ if (i2s_slot->incr_value >= i2s_slot->incr_end)
+ i2s_slot->incr_value = 0;
+
+ for (int i=0; i<i2s_slot->nb_elem; i++)
+ {
+ set_buffer_elem_iter(i2s_slot, buffers[0], i);
+ }
+ for (int i=0; i<i2s_slot->nb_elem; i++)
+ {
+ set_buffer_elem_iter(i2s_slot, buffers[1], i);
+ }
}
- pi_i2s_channel_write_async(i2s_slot->i2s, i2s_slot->slot, NULL, i2s_slot->buffer_size, pi_task_callback(&i2s_slot->task, i2s_slot_callback_tx_file_dumper, (void *)i2s_slot));
+ if (i2s_slot->frame == 0 || __FL1(i2s_slot->frame) == i2s_slot->slot)
+ {
+ if (i2s_slot->frame)
+ {
+ i2s_slot_test_t *first_slot = &i2s_slot->test->slot_test_tx[__FF1(i2s_slot->frame)];
+ pi_i2s_frame_write_async(first_slot->i2s, first_slot->frame, buffers[0], first_slot->buffer_size, pi_task_callback(&first_slot->task, i2s_slot_callback_tx_file_dumper, (void *)first_slot));
+ pi_i2s_frame_write_async(first_slot->i2s, first_slot->frame, buffers[1], first_slot->buffer_size, pi_task_callback(&first_slot->task, i2s_slot_callback_tx_file_dumper, (void *)first_slot));
+
+ }
+ else
+ {
+ pi_i2s_channel_write_async(i2s_slot->i2s, i2s_slot->slot, buffers[0], i2s_slot->buffer_size, pi_task_callback(&i2s_slot->task, i2s_slot_callback_tx_file_dumper, (void *)i2s_slot));
+ pi_i2s_channel_write_async(i2s_slot->i2s, i2s_slot->slot, buffers[1], i2s_slot->buffer_size, pi_task_callback(&i2s_slot->task, i2s_slot_callback_tx_file_dumper, (void *)i2s_slot));
+ }
+ }
}
}
@@ -506,7 +757,7 @@ int i2s_slot_start(i2s_slot_test_t *i2s_slot, i2s_slot_start_config_t *config)
int i2s_slot_wait(i2s_slot_test_t *i2s_slot)
{
- if (i2s_slot->bypass)
+ if (i2s_slot->bypass || (i2s_slot->frame && i2s_slot->slot != __FF1(i2s_slot->frame)))
{
return 0;
}
@@ -557,8 +808,13 @@ int i2s_test_init(i2s_test_t *test, i2s_test_config_t *config)
fifo_id = config->fifo_id;
+ uint32_t rx_frames[4] = { (config->rx_frames >> 0) & 0xffff, (config->rx_frames >> 16) & 0xffff, (config->rx_frames >> 32) & 0xffff, (config->rx_frames >> 48) & 0xffff};
+ uint32_t tx_frames[4] = { (config->tx_frames >> 0) & 0xffff, (config->tx_frames >> 16) & 0xffff, (config->tx_frames >> 32) & 0xffff, (config->tx_frames >> 48) & 0xffff};
+
for (int i=0; i<16; i++)
{
+ uint32_t slab = (config->rx_slabs >> (i*4)) & 0xf;
+
test->slot_test_rx[i].i2s = NULL;
if ((config->rx_slots >> i) & 1)
@@ -567,14 +823,25 @@ int i2s_test_init(i2s_test_t *test, i2s_test_config_t *config)
int slot_format = (config->rx_slots_format >> (i*4)) & 0xF;
int random_mute = (config->random_mute >> i) & 1;
int ts_evt_en = (config->ts_evt >> i) & 1;
+ uint32_t frame = 0;
+
+ for (int j=0; j<4; j++)
+ {
+ if (((rx_frames[j] >> i) & 1) != 0)
+ {
+ frame = rx_frames[j];
+ }
+ }
i2s_slot_config_t i2s_slot_config = {
.itf=config->itf, .slot=i, .is_rx=1, .word_size=slot_width, .nb_elem=config->buffer_nb_elem, .elem_size=config->elem_size, .format=slot_format,
- .mute_delay_start=30, .mute_delay_incr=20, .mute_delay_end=150, .random_mute=random_mute, .ts_evt_en=ts_evt_en
+ .mute_delay_start=30, .mute_delay_incr=20, .mute_delay_end=150, .random_mute=random_mute, .ts_evt_en=ts_evt_en, .frame=frame, .slab=slab
};
- if (i2s_slot_init(&test->slot_test_rx[i], &test->i2s, &i2s_slot_config))
+ i2s_slot_new(&test->slot_test_rx[i]);
+
+ if (i2s_slot_init(test, &test->slot_test_rx[i], &test->i2s, &i2s_slot_config))
return -1;
int iter = config->nb_slots;
@@ -600,6 +867,8 @@ int i2s_test_init(i2s_test_t *test, i2s_test_config_t *config)
for (int i=0; i<16; i++)
{
+ uint32_t slab = (config->tx_slabs >> (i*4)) & 0xf;
+
test->slot_test_tx[i].i2s = NULL;
if ((config->tx_slots >> i) & 1)
@@ -609,15 +878,32 @@ int i2s_test_init(i2s_test_t *test, i2s_test_config_t *config)
int random_mute = (config->random_mute >> i) & 1;
int ts_evt_en = (config->ts_evt >> (16+i)) & 1;
int bypass = (config->tx_slots_bypass >> i) & 1;
+ uint32_t frame = 0;
+
+ for (int j=0; j<4; j++)
+ {
+ if (((tx_frames[j] >> i) & 1) != 0)
+ {
+ frame = tx_frames[j];
+ }
+ }
i2s_slot_config_t i2s_slot_config = {
.itf=config->itf, .slot=i, .is_rx=0, .word_size=slot_width, .nb_elem=config->buffer_nb_elem, .elem_size=config->elem_size, .format=slot_format,
- .mute_delay_start=30, .mute_delay_incr=20, .mute_delay_end=150, .random_mute=random_mute, .ts_evt_en=ts_evt_en, .bypass=bypass
+ .mute_delay_start=30, .mute_delay_incr=20, .mute_delay_end=150, .random_mute=random_mute, .ts_evt_en=ts_evt_en, .bypass=bypass, .frame=frame, .slab=slab
};
- if (i2s_slot_init(&test->slot_test_tx[i], &test->i2s, &i2s_slot_config))
+ i2s_slot_new(&test->slot_test_tx[i]);
+
+ if (i2s_slot_init(test, &test->slot_test_tx[i], &test->i2s, &i2s_slot_config))
return -1;
-
+ }
+ }
+
+ for (int i=0; i<16; i++)
+ {
+ if ((config->tx_slots >> i) & 1)
+ {
int iter = config->nb_slots;
if (config->full_duplex)
{
diff --git a/libs/gap_lib/testbench/testlib.h b/libs/gap_lib/testbench/testlib.h
index 982e59785..1526893cc 100644
--- a/libs/gap_lib/testbench/testlib.h
+++ b/libs/gap_lib/testbench/testlib.h
@@ -16,6 +16,8 @@
#define I2S_SLOT_STATIC_INIT {0}
+typedef struct i2s_test_s i2s_test_t;
+
typedef struct
{
int itf;
@@ -60,6 +62,8 @@ typedef struct
int ts_evt_en;
int slot_disable;
int bypass;
+ uint32_t frame;
+ uint32_t slab;
union
{
struct
@@ -132,6 +136,12 @@ typedef struct
int bypass;
int mute_ack;
void *buffers[2];
+ uint32_t frame;
+ i2s_test_t *test;
+ pi_mem_slab_t slab;
+ struct {
+ int use_slab:1;
+ } flags;
}
i2s_slot_test_t;
@@ -171,11 +181,15 @@ typedef struct
uint32_t fifo_id;
int8_t ws_delay;
uint32_t incr;
+ uint64_t rx_frames;
+ uint64_t tx_frames;
+ uint64_t rx_slabs;
+ uint64_t tx_slabs;
}
i2s_test_config_t;
-typedef struct
+typedef struct i2s_test_s
{
struct pi_device i2s;
i2s_slot_test_t slot_test_rx[16];
@@ -194,7 +208,8 @@ pi_device_t *get_testbench();
pi_device_t *i2s_init(struct pi_device *i2s, i2s_config_t *config);
int i2s_deinit(struct pi_device *i2s, int itf);
-int i2s_slot_init(i2s_slot_test_t *i2s_slot, struct pi_device *i2s, i2s_slot_config_t *config);
+void i2s_slot_new(i2s_slot_test_t *i2s_slot);
+int i2s_slot_init(i2s_test_t *test, i2s_slot_test_t *i2s_slot, struct pi_device *i2s, i2s_slot_config_t *config);
void i2s_slot_deinit(i2s_slot_test_t *i2s_slot);
int i2s_slot_start(i2s_slot_test_t *i2s_slot, i2s_slot_start_config_t *config);
int i2s_slot_stop(i2s_slot_test_t *i2s_slot);
diff --git a/rtos/freeRTOS/vendors/gwt/gap8/pmsis/drivers/udma/i2s/i2s_internal.c b/rtos/freeRTOS/vendors/gwt/gap8/pmsis/drivers/udma/i2s/i2s_internal.c
index cba50c71b..3c1b9bd9a 100644
--- a/rtos/freeRTOS/vendors/gwt/gap8/pmsis/drivers/udma/i2s/i2s_internal.c
+++ b/rtos/freeRTOS/vendors/gwt/gap8/pmsis/drivers/udma/i2s/i2s_internal.c
@@ -162,7 +162,9 @@ static uint8_t __pi_i2s_enqueue(struct i2s_itf_data_s *itf_data)
itf_data->pending_size = 0;
}
/* Enqueue first in HW fifo. */
- hal_i2s_enqueue(device_id, itf_data->i2s_id, buffer, size_enqueue, itf_data->udma_cfg);
+ I2S_TRACE("i2s(%d): enqueue l2_buf=%lx, size=%ld, cfg=%lx\n",
+ itf_data->i2s_id, buffer, size_enqueue, itf_data->udma_cfg);
+ hal_i2s_enqueue(device_id, itf_data->i2s_id << 4, buffer, size_enqueue, itf_data->udma_cfg);
}
return done;
}
@@ -238,7 +240,7 @@ static void __pi_i2s_clock_enable(struct i2s_itf_data_s *itf_data)
if ((itf_data->clk & PI_I2S_OPT_EXT_CLK) == 0)
{
uint16_t clk_div = __pi_i2s_clk_div(itf_data->frequency);
- I2S_TRACE("I2S(%d) clk_div = %d\n", device_id, clk_div);
+ I2S_TRACE("I2S(%ld) clk_div = %d\n", device_id, clk_div);
switch (itf_data->clk & 0x1)
{
case 0 :
diff --git a/rtos/freeRTOS/vendors/gwt/libs/include/string.h b/rtos/freeRTOS/vendors/gwt/libs/include/string.h
index 025caa471..51e9afeec 100644
--- a/rtos/freeRTOS/vendors/gwt/libs/include/string.h
+++ b/rtos/freeRTOS/vendors/gwt/libs/include/string.h
@@ -34,6 +34,10 @@
#include <stddef.h>
#include <stdint.h>
+#ifdef __cplusplus
+extern "C" {
+#endif
+
int32_t memcmp(const void *str1, const void *str2, size_t n);
void *memcpy(void *str1, const void *str2, size_t n);
@@ -60,4 +64,8 @@ size_t strcspn(const char *s, const char *reject);
char *strchr(const char *s, int c);
+#ifdef __cplusplus
+}
+#endif
+
#endif /* __STRING_H__ */
diff --git a/rtos/freeRTOS/vendors/gwt/pmsis/include/pmsis.h b/rtos/freeRTOS/vendors/gwt/pmsis/include/pmsis.h
index 1ec6ca35b..e554d5e2d 100644
--- a/rtos/freeRTOS/vendors/gwt/pmsis/include/pmsis.h
+++ b/rtos/freeRTOS/vendors/gwt/pmsis/include/pmsis.h
@@ -72,4 +72,9 @@
/* Hal includes. */
#include "pmsis/implem/hal/hal.h"
+/* GVSOC proxy includes. */
+#if defined(__PLATFORM_GVSOC__)
+#include "pmsis/platforms/gvsoc.h"
+#endif /* __PLATFORM_GVSOC__ */
+
#endif /* __PMSIS_H__ */
diff --git a/rtos/freeRTOS/vendors/gwt/pmsis/rtos/include/pmsis/rtos/os/os.h b/rtos/freeRTOS/vendors/gwt/pmsis/rtos/include/pmsis/rtos/os/os.h
index b7ae00392..7a0fdae6a 100644
--- a/rtos/freeRTOS/vendors/gwt/pmsis/rtos/include/pmsis/rtos/os/os.h
+++ b/rtos/freeRTOS/vendors/gwt/pmsis/rtos/include/pmsis/rtos/os/os.h
@@ -35,6 +35,10 @@
#include "pmsis/backend/implementation_specific_defines.h"
#include "pmsis/backend/pmsis_backend_native_task_api.h"
+#ifdef __cplusplus
+extern "C" {
+#endif
+
/*******************************************************************************
* Definitions
******************************************************************************/
@@ -290,4 +294,8 @@ static inline void pi_os_reboot(void)
#endif /* __GAP8__ */
+#ifdef __cplusplus
+}
+#endif
+
#endif /* __PI_RTOS_IMPLEM_OS_H__ */
diff --git a/rtos/freeRTOS/vendors/gwt/rules/freeRTOS_rules.mk b/rtos/freeRTOS/vendors/gwt/rules/freeRTOS_rules.mk
index 96a66c543..9a2dfe3a0 100644
--- a/rtos/freeRTOS/vendors/gwt/rules/freeRTOS_rules.mk
+++ b/rtos/freeRTOS/vendors/gwt/rules/freeRTOS_rules.mk
@@ -174,7 +174,7 @@ FREERTOS_FLAGS += -DMAIN_APP_STACK_SIZE=$(MAIN_APP_STACK_SIZE)
GCC_OPTIM_LEVEL = -Os # Optimize for size.
COMMON = $(RISCV_FLAGS) \
- -c -g -ffunction-sections -fdata-sections \
+ -c -gdwarf-2 -gstrict-dwarf -ffunction-sections -fdata-sections \
-fno-delete-null-pointer-checks -fomit-frame-pointer \
-fno-tree-loop-distribute-patterns -fno-jump-tables \
$(FEATURE_FLAGS) $(FREERTOS_FLAGS)
@@ -345,6 +345,11 @@ OBJS_DEP = $(patsubst %.o, %.d, $(OBJS))
APP ?= test
BIN = $(BUILDDIR)/$(APP)
+
+ifneq ($(wsl),)
+WSL_ENV=--wsl=$(wsl)
+endif
+
# Makefile targets :
# Build objects (*.o) amd associated dependecies (*.d) with disassembly (*.dump).
#------------------------------------------
@@ -410,7 +415,7 @@ image: $(BIN)
gapy $(GAPY_TARGET_OPT) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --image --binary=$(BIN) $(runner_args)
run: $(BIN)
- gapy $(GAPY_TARGET_OPT) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec-prepare --exec --binary=$(BIN) $(runner_args)
+ gapy $(GAPY_TARGET_OPT) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec-prepare --exec --binary=$(BIN) $(runner_args) $(WSL_ENV)
traces:
gapy $(GAPY_TARGET_OPT) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec --binary=$(BIN) --no-run --extend-traces $(runner_args)
diff --git a/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h b/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h
index d043479f6..62c560de7 100644
--- a/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h
+++ b/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h
@@ -19,6 +19,10 @@
#include <stdint.h>
+#ifndef PI_INLINE_I2S_LVL_0
+#define PI_INLINE_I2S_LVL_0
+#endif
+
#ifdef __cplusplus
extern "C" {
#endif
@@ -756,9 +760,30 @@ void pi_i2s_channel_conf_init(struct pi_i2s_channel_conf *conf);
* @retval 0 If successful.
* @retval -1 An error occured.
*/
-int pi_i2s_channel_conf_set(struct pi_device *dev, int channel,
+PI_INLINE_I2S_LVL_0 int pi_i2s_channel_conf_set(struct pi_device *dev, int channel,
struct pi_i2s_channel_conf *conf);
+/**
+ * \brief Configure a channel of a frame in TDM mode.
+ *
+ * A frame is a set of channels gathered together so that they can be controlled
+ * all together in terms of data transfer.
+ * This function can be used to configure a channel which is part of a frame.
+ * In TDM mode, the same interface is time-multiplexed to transmit data
+ * for multiple channels, and each channel can have a specific
+ * configuration. This function can be used to give the configuration
+ * of one channel.
+ * @param dev Pointer to the device structure for the driver instance.
+ * @param frame A bitfield containing the channels of the part (one bit per channel).
+ * @param channel ID of the slot, from 0 to the number of channels minus 1.
+ * @param conf A pointer to the I2S channel configuration.
+ *
+ * @retval 0 If successful.
+ * @retval -1 An error occured.
+ */
+PI_INLINE_I2S_LVL_0 int pi_i2s_frame_channel_conf_set(struct pi_device *dev,
+ uint32_t frame, int channel, struct pi_i2s_channel_conf *conf);
+
/**
* \brief Get the current configuration of a channel in TDM mode.
@@ -899,8 +924,8 @@ int pi_i2s_slots_stop(struct pi_device *dev, uint32_t slots);
*
* @retval 0 If successful, -1 if not.
*/
-int pi_i2s_channel_read(struct pi_device *dev, int channel, void **mem_block,
- size_t *size);
+int pi_i2s_channel_read(struct pi_device *dev, int channel,
+ void **mem_block, size_t *size);
/**
* @brief Read data asynchronously from the RX queue of a channel in TDM mode.
@@ -929,8 +954,77 @@ int pi_i2s_channel_read(struct pi_device *dev, int channel, void **mem_block,
*
* @retval 0 If successful, -1 if not.
*/
-int pi_i2s_channel_read_async(struct pi_device *dev, int channel,
- pi_task_t *task);
+int pi_i2s_channel_read_async(struct pi_device *dev,
+ int channel, pi_task_t *task);
+
+/**
+ * @brief Read data from the RX queue of a frame channel in TDM mode.
+ *
+ * Data received by the I2S interface is stored in the RX queue consisting
+ * of two memory blocks preallocated by the user and given to the driver in
+ * the configuration. Calling this function will return the next available
+ * buffer to the caller, which has to use it before the sampling for this buffer
+ * starts again.
+ *
+ * The data is read in chunks equal to the size of the memory block.
+ *
+ * This will return data for the specified channel and must only be used in
+ * TDM mode.
+ *
+ * If there is no data in the RX queue the function will block waiting for
+ * the next RX memory block to fill in.
+ *
+ * This function is reading for the whole frame. The returned buffer contains
+ * the samples for the whole frame, but the size is the size of one channel.
+ *
+ * Due to hardware constraints, the address of the buffer must be aligned on
+ * 4 bytes.
+ *
+ * @param dev Pointer to the device structure for the driver instance.
+ * @param frame A bitfield containing the channels of the part (one bit per channel).
+ * @param channel ID of the slot, from 0 to the number of channels minus 1.
+ * @param mem_block Pointer to the variable storing the address of the RX memory
+ * block containing received data.
+ * @param size Pointer to the variable storing the number of bytes read.
+ *
+ * @retval 0 If successful, -1 if not.
+ */
+PI_INLINE_I2S_LVL_0 int pi_i2s_frame_read(struct pi_device *dev, uint32_t frame,
+ void **mem_block, size_t *size);
+
+/**
+ * @brief Read data asynchronously from the RX queue of a frame channel in TDM mode.
+ *
+ * Data received by the I2S interface is stored in the RX queue consisting
+ * of two memory blocks preallocated by the user and given to the driver in
+ * the configuration. Calling this function will return the next available
+ * buffer to the caller, which has to use it before the sampling for this buffer
+ * starts again.
+ *
+ * The data is read in chunks equal to the size of the memory block.
+ *
+ * This will return data for the specified channel and must only be used in
+ * TDM mode.
+ *
+ * The specified task will be pushed as soon as data is ready in the RX queue,
+ * and the information about the memory block and the size will be available
+ * in the task.
+ *
+ * This function is reading for the whole frame. The returned buffer contains
+ * the samples for the whole frame, but the size is the size of one channel.
+ *
+ * Due to hardware constraints, the address of the buffer must be aligned on
+ * 4 bytes.
+ *
+ * @param dev Pointer to the device structure for the driver instance.
+ * @param frame A bitfield containing the channels of the part (one bit per channel).
+ * @param channel ID of the slot, from 0 to the number of channels minus 1.
+ * @param task The task used to notify the end of transfer.
+ *
+ * @retval 0 If successful, -1 if not.
+ */
+PI_INLINE_I2S_LVL_0 int pi_i2s_frame_read_async(struct pi_device *dev,
+ uint32_t frame, pi_task_t *task);
/**
* @brief Read the status of an asynchronous read.
@@ -979,8 +1073,8 @@ int pi_i2s_read_status(pi_task_t *task, void **mem_block, size_t *size);
* @retval 0 If successful.
* @retval -1 An error occured.
*/
-int pi_i2s_channel_write(struct pi_device *dev, int channel, void *mem_block,
- size_t size);
+int pi_i2s_channel_write(struct pi_device *dev, int channel,
+ void *mem_block, size_t size);
/**
* @brief Write data asynchronously to the TX queue of a channel in TDM mode.
@@ -1013,8 +1107,83 @@ int pi_i2s_channel_write(struct pi_device *dev, int channel, void *mem_block,
* @retval 0 If successful.
* @retval -1 An error occured.
*/
-int pi_i2s_channel_write_async(struct pi_device *dev, int channel,
- void *mem_block, size_t size, pi_task_t *task);
+int pi_i2s_channel_write_async(struct pi_device *dev,
+ int channel, void *mem_block, size_t size, pi_task_t *task);
+
+/**
+ * @brief Write data to the TX queue of a frame channel in TDM mode.
+ *
+ * Data to be sent by the I2S interface is stored first in the TX queue
+ * consisting of memory blocks preallocated by the user with either pingpong
+ * buffers or a memory slab allocator.
+ *
+ * In pingpong mode, the driver will automatically alternate between 2 buffers
+ * and the user code is supposed to call this function to notify the driver
+ * that the specified buffer is ready to be sent. This is used by the driver
+ * to report when an underrun or an overrun occurs.
+ *
+ * In memory slab allocator mode, the user has to allocate buffers from the
+ * memory slab allocator and pass them to the driver by calling this function
+ * when they are ready to be sent.
+ *
+ * This fonction will block until the specified buffer has been transfered.
+ *
+ * This function is writing for the whole frame. The buffer must contain
+ * the samples for the whole frame, but the size is the size of one channel.
+ *
+ * This will sent data to the specified channel and must only be used in
+ * TDM mode.
+ *
+ * @param dev Pointer to the device structure for the driver instance.
+ * @param frame A bitfield containing the channels of the part (one bit per channel).
+ * @param channel ID of the slot, from 0 to the number of channels minus 1.
+ * @param mem_block Pointer to the TX memory block containing data to be sent.
+ * @param size Number of bytes to write. This value has to be equal or smaller
+ * than the size of the memory block.
+ *
+ * @retval 0 If successful.
+ * @retval -1 An error occured.
+ */
+PI_INLINE_I2S_LVL_0 int pi_i2s_frame_write(struct pi_device *dev, uint32_t frame,
+ void *mem_block, size_t size);
+
+/**
+ * @brief Write data asynchronously to the TX queue of a frame channel in TDM mode.
+ *
+ * Data to be sent by the I2S interface is stored first in the TX queue
+ * consisting of memory blocks preallocated by the user with either pingpong
+ * buffers or a memory slab allocator.
+ *
+ * In pingpong mode, the driver will automatically alternate between 2 buffers
+ * and the user code is supposed to call this function to notify the driver
+ * that the specified buffer is ready to be sent. This is used by the driver
+ * to report when an underrun or an overrun occurs.
+ *
+ * In memory slab allocator mode, the user has to allocate buffers from the
+ * memory slab allocator and pass them to the driver by calling this function
+ * when they are ready to be sent.
+ *
+ * This will sent data to the specified channel and must only be used in
+ * TDM mode.
+ *
+ * This function is writing for the whole frame. The buffer must contain
+ * the samples for the whole frame, but the size is the size of one channel.
+ *
+ * The specified task will be pushed as soon as data is has been transfered.
+ *
+ * @param dev Pointer to the device structure for the driver instance.
+ * @param frame A bitfield containing the channels of the part (one bit per channel).
+ * @param channel ID of the slot, from 0 to the number of channels minus 1.
+ * @param mem_block Pointer to the TX memory block containing data to be sent.
+ * @param size Number of bytes to write. This value has to be equal or smaller
+ * than the size of the memory block.
+ * @param task The task used to notify the end of transfer.
+ *
+ * @retval 0 If successful.
+ * @retval -1 An error occured.
+ */
+PI_INLINE_I2S_LVL_0 int pi_i2s_frame_write_async(struct pi_device *dev,
+ uint32_t frame, void *mem_block, size_t size, pi_task_t *task);
/**
* @brief Read the status of an asynchronous write.
diff --git a/rtos/pulp/pulpos-2/include/pmsis.h b/rtos/pulp/pulpos-2/include/pmsis.h
index 830a86669..d8441d061 100644
--- a/rtos/pulp/pulpos-2/include/pmsis.h
+++ b/rtos/pulp/pulpos-2/include/pmsis.h
@@ -35,6 +35,7 @@
#define PI_INLINE_CL_TEAM_1 static inline
#define PI_INLINE_HYPER_LVL_0 static inline
#define PI_INLINE_OCTOSPI_LVL_0 static inline
+#define PI_INLINE_I2S_LVL_0 static inline
#if defined(__GAP9__)
#include "pmsis/chips/gap8/perf.h"
@@ -57,6 +58,7 @@
#include "pmsis/rtos/malloc/cl_l1_malloc.h"
#include "pmsis/rtos/malloc/l2_malloc.h"
#include "pmsis/rtos/malloc/fc_l1_malloc.h"
+#include "pmsis/mem_slab.h"
#include "pmsis/drivers/perf.h"
#include "pmsis/drivers/hyperbus.h"
#include "pmsis/drivers/octospi.h"
diff --git a/rtos/pulp/pulpos-2/include/pos/data/data.h b/rtos/pulp/pulpos-2/include/pos/data/data.h
index 9dc3acf5a..0250d1701 100644
--- a/rtos/pulp/pulpos-2/include/pos/data/data.h
+++ b/rtos/pulp/pulpos-2/include/pos/data/data.h
@@ -132,6 +132,15 @@ typedef struct pi_cluster_pe_task_s
uint8_t nb_done_cores;
} pi_cluster_pe_task_t;
+
+struct pi_mem_slab {
+ uint32_t num_blocks;
+ size_t block_size;
+ char *buffer;
+ char *free_list;
+ uint32_t num_used;
+};
+
#endif
diff --git a/rtos/pulp/pulpos-2/include/pos/implem/implem.h b/rtos/pulp/pulpos-2/include/pos/implem/implem.h
index 5484ca05d..ab0701ede 100644
--- a/rtos/pulp/pulpos-2/include/pos/implem/implem.h
+++ b/rtos/pulp/pulpos-2/include/pos/implem/implem.h
@@ -141,6 +141,7 @@ static inline void pmsis_exit(int err)
#if defined(__GAP9__)
#include "pos/implem/hyperbus-v2.h"
#include "pos/implem/octospi-v2.h"
+#include "pos/implem/i2s-v3.h"
#endif
#endif
diff --git a/rtos/pulp/pulpos-2/rules/pulpos/src.mk b/rtos/pulp/pulpos-2/rules/pulpos/src.mk
index 09d748f17..3726fe2f0 100644
--- a/rtos/pulp/pulpos-2/rules/pulpos/src.mk
+++ b/rtos/pulp/pulpos-2/rules/pulpos/src.mk
@@ -20,7 +20,7 @@ endif
ifdef CONFIG_KERNEL
PULP_SRCS += kernel/init.c kernel/kernel.c kernel/device.c kernel/task.c kernel/alloc.c \
- kernel/alloc_pool.c kernel/irq.c kernel/soc_event.c kernel/log.c kernel/time.c
+ kernel/alloc_pool.c kernel/irq.c kernel/soc_event.c kernel/log.c kernel/time.c kernel/mem_slab.c
PULP_ASM_SRCS += kernel/irq_asm.S kernel/time_asm.S
diff --git a/testplan.cfg b/testplan.cfg
index e65b66319..181e6384c 100644
--- a/testplan.cfg
+++ b/testplan.cfg
@@ -24,6 +24,8 @@ i2s.add_feature('Slot dynamic reconfiguration')
i2s.add_feature('Cross-interface synchronization')
i2s.add_feature('PDM SFU connection')
i2s.add_feature('UDMA core PCM connection')
+i2s.add_feature('Memory slab allocators')
+i2s.add_feature('Channel frames')
sfu = testplan.add_category('udma:sfu')
sfu.add_feature('MEM_IN - Memory input')
diff --git a/tools/autotiler_v3/CNN_Generators/CNN_Copy_Generators.c b/tools/autotiler_v3/CNN_Generators/CNN_Copy_Generators.c
index 1af0778c6..a6030494d 100644
--- a/tools/autotiler_v3/CNN_Generators/CNN_Copy_Generators.c
+++ b/tools/autotiler_v3/CNN_Generators/CNN_Copy_Generators.c
@@ -336,6 +336,78 @@ void LoadCNN_Copy_Library()
TCArg("signed char *__restrict__", "Infos")
)
);
+ LibKernelTemplate("CNN_Float32Fp_T",
+ CArgs(5,
+ TCArg("float *__restrict__", "In"),
+ TCArg("signed short *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
+ LibKernelTemplate("CNN_Float32UFp_T",
+ CArgs(5,
+ TCArg("float *__restrict__", "In"),
+ TCArg("unsigned short *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
+ LibKernelTemplate("CNN_Float32Fps_T",
+ CArgs(5,
+ TCArg("float *__restrict__", "In"),
+ TCArg("signed char *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
+ LibKernelTemplate("CNN_Float32UFps_T",
+ CArgs(5,
+ TCArg("float *__restrict__", "In"),
+ TCArg("unsigned char *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
+ LibKernelTemplate("CNN_FpFloat32_T",
+ CArgs(5,
+ TCArg("signed short *__restrict__", "In"),
+ TCArg("float *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
+ LibKernelTemplate("CNN_UFpFloat32_T",
+ CArgs(5,
+ TCArg("unsigned short *__restrict__", "In"),
+ TCArg("float *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
+ LibKernelTemplate("CNN_FpsFloat32_T",
+ CArgs(5,
+ TCArg("signed char *__restrict__", "In"),
+ TCArg("float *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
+ LibKernelTemplate("CNN_UFpsFloat32_T",
+ CArgs(5,
+ TCArg("unsigned char *__restrict__", "In"),
+ TCArg("float *__restrict__", "Out"),
+ TCArg("unsigned short int", "W"),
+ TCArg("unsigned short int", "H"),
+ TCArg("signed char *__restrict__", "Infos")
+ )
+ );
LibKernelTemplate("KerMatTranspose_fps_T",
CArgs(7,
TCArg("signed char *__restrict__", "In"),
@@ -506,6 +578,16 @@ void LoadCNN_Copy_Library()
LibKernel("CNN_FpsFloat16", CALL_PARALLEL, 0, "CNN_FpsFloat16_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FP_FL), 0, 1, CNN_Type(1,0,0,0,2), 0,0,0,0,1,1));
LibKernel("CNN_UFpsFloat16", CALL_PARALLEL, 0, "CNN_UFpsFloat16_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FP_FL), 0, 1, CNN_Type(-1,0,0,0,2), 0,0,0,0,1,1));
+ LibKernel("CNN_Float32Fp", CALL_PARALLEL, 0, "CNN_Float32Fp_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FL_FP), 0, 1, CNN_Type(4,0,0,0,2), 0,0,0,0,1,1));
+ LibKernel("CNN_Float32UFp", CALL_PARALLEL, 0, "CNN_Float32UFp_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FL_FP), 0, 1, CNN_Type(4,0,0,0,-2), 0,0,0,0,1,1));
+ LibKernel("CNN_Float32Fps", CALL_PARALLEL, 0, "CNN_Float32Fps_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FL_FP), 0, 1, CNN_Type(4,0,0,0,1), 0,0,0,0,1,1));
+ LibKernel("CNN_Float32UFps", CALL_PARALLEL, 0, "CNN_Float32UFps_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FL_FP), 0, 1, CNN_Type(4,0,0,0,-1), 0,0,0,0,1,1));
+
+ LibKernel("CNN_FpFloat32", CALL_PARALLEL, 0, "CNN_FpFloat32_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FP_FL), 0, 1, CNN_Type(2,0,0,0,4), 0,0,0,0,1,1));
+ LibKernel("CNN_UFpFloat32", CALL_PARALLEL, 0, "CNN_UFpFloat32_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FP_FL), 0, 1, CNN_Type(-2,0,0,0,4), 0,0,0,0,1,1));
+ LibKernel("CNN_FpsFloat32", CALL_PARALLEL, 0, "CNN_FpsFloat32_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FP_FL), 0, 1, CNN_Type(1,0,0,0,4), 0,0,0,0,1,1));
+ LibKernel("CNN_UFpsFloat32", CALL_PARALLEL, 0, "CNN_UFpsFloat32_T", CNN_Match(CNN_OperList(1, KOP_CONVERT_FP_FL), 0, 1, CNN_Type(-1,0,0,0,4), 0,0,0,0,1,1));
+
/****************************************************************************************************************/
/* Kernels tensor/matrix transposes and permutations */
/****************************************************************************************************************/
@@ -1200,7 +1282,7 @@ int CNN_Convert(
PKerArgs[0] = KerArg("In", KerArgSpace(1,D0), O_IN|O_DB, 1, 1, In_DataSize, 0, 0, 0, "In");
PKerArgs[1] = KerArg("Out", KerArgSpace(1,D0), O_OUT|O_DB, 1, 1, Out_DataSize, 0, 0, 0, "Out");
if (kop!=KOP_CONVERT_FP_FP)
- PKerArgs[2] = KerArg("Infos",KerArgSpace(1,T0), O_IN|O_BUFF|O_NTILED, 1, 1, 7, 0, 0, 0, "Infos");
+ PKerArgs[2] = KerArg("Infos",KerArgSpace(1,T0), O_IN|O_BUFF|O_NTILED, 1, 1, 8, 0, 0, 0, "Infos");
Kernel_T *Kernel = UserKernel(Name,
KernelIterSpace(2, IterTiledSpace(T0), IterParSpace(D0, Sz, 8)),
TILE_HOR,
@@ -1225,10 +1307,10 @@ int CNN_Convert(
if (Kernel) {
AddKernelInfos(Name, AT_KERINFO_OPER, LayerOp, 0);
AddKernelInfos(Name, AT_KERINFO_BANDWIDTH, LayerBandwidth, 0);
- AddKernelArgDim(Name, "In", 3, 1, Sz, In_DataSize);
- AddKernelArgDim(Name, "Out", 3, 1, Sz, Out_DataSize);
+ In_Float? AddKernelFloatArgDim(Name, "In", 3, 1, Sz, In_DataSize) :AddKernelArgDim(Name, "In", 3, 1, Sz, In_DataSize);
+ Out_Float?AddKernelFloatArgDim(Name, "Out", 3, 1, Sz, Out_DataSize):AddKernelArgDim(Name, "Out", 3, 1, Sz, Out_DataSize);
if (kop!=KOP_CONVERT_FP_FP)
- AddKernelArgDim(Name, "Infos", 2, 1, 1);
+ AddKernelArgDim(Name, "Infos", 2, 8, 1);
}
return (Kernel!=0);
}
diff --git a/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.c b/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.c
index bc010a175..f4a50abbf 100644
--- a/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.c
+++ b/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.c
@@ -134,7 +134,7 @@ void LoadCNN_NE16_SQ8_Library()
TCArg("unsigned char", "Dy")
)
);
- LibKernelTemplate("Ker_MM_Conv_NE16_T",
+ LibKernelTemplate("KerConv_MM_NE16_T",
CArgs(28,
TCArg("unsigned char * __restrict__", "In"),
TCArg("unsigned char * __restrict__", "ColBuff"),
@@ -239,13 +239,13 @@ void LoadCNN_NE16_SQ8_Library()
CNN_Match(CNN_OperList(1, KOP_CONV_DW), 0, -1, CNN_Type(0,0,0,0,4), 3,3,1,1,1,1));
LibKernel("KerConvDW3x3Stride2_NE16", CALL_SEQUENTIAL_STRUCT|CALL_NE16_KER, 0, "KerConv_NE16_T",
CNN_Match(CNN_OperList(1, KOP_CONV_DW), 0, -1, CNN_Type(0,0,0,0,4), 3,3,1,1,2,2));
+ LibKernel("Ker_MM_Conv2D_NE16", CALL_PARALLEL_CC|CALL_NE16_KER, 0, "KerConv_MM_NE16_T",
+ CNN_Match(CNN_OperList(1, KOP_MM_CONV), 0, -1, CNN_Type(0,0,0,0,4), -1,-1, 1, 1,-1,-1));
LibKernel("KerLinear_8a_NE16", CALL_SEQUENTIAL_STRUCT|CALL_NE16_KER, 0, "KerLinear_NE16_T",
CNN_Match(CNN_OperList(1, KOP_LINEAR), CNN_OperList(1, KOP_NONE), -1, CNN_Type(1,0,4,0,0), 0,0,0,0,0,0));
LibKernel("KerLinear_16a_NE16", CALL_SEQUENTIAL_STRUCT|CALL_NE16_KER, 0, "KerLinear_NE16_T",
CNN_Match(CNN_OperList(1, KOP_LINEAR), CNN_OperList(1, KOP_NONE), -1, CNN_Type(2,0,4,0,0), 0,0,0,0,0,0));
- LibKernel("Ker_MM_Conv2D_NE16", CALL_PARALLEL_CC|CALL_NE16_KER, 0, "Ker_MM_Conv_NE16_T",
- CNN_Match(CNN_OperList(1, KOP_MM_CONV), 0, -1, CNN_Type(1,1,0,0,4), -1,-1,1,1,-1,-1));
LibKernel("KerMatMul_8a_NE16", CALL_PARALLEL_CC|CALL_NE16_KER, 0, "KerMatMul_NE16_T",
CNN_Match(CNN_OperList(1, KOP_MATMUL_TRANSPOSED), CNN_OperList(1, KOP_NONE), 0, CNN_Type(1,0,4,0,0), 0,0,0,0,0,0));
@@ -259,7 +259,7 @@ void LoadCNN_NE16_SQ8_Library()
}
-int CNN_MM_ConvolutionNE16(
+Kernel_T *CNN_MM_ConvolutionNE16(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -335,7 +335,7 @@ int CNN_MM_ConvolutionNE16(
if (!(ActOper == KOP_NONE || ActOper == KOP_RELU || ActOper == KOP_RELUN || ActOper == KOP_RELUM || ActOper == KOP_HSIGMOID || ActOper == KOP_HSWISH || ActOper == KOP_LEAKYRELU || ActOper == KOP_SIGMOID || ActOper == KOP_TANH))
GenTilingError("CNN_MM_ConvolutionNE16 Kernel: %s, ActOper, expecting KOP_NONE, KOP_RELU, KOP_RELUN, KOP_RELUM, KOP_HSIGMOID, KOP_HSWISH, KOP_LEAKYRELU, KOP_SIGMOID or KOP_TANH", Name);
- Wa |= O_NE16_RNN;
+ Wa |= O_NE16_LIN | O_LINEAR;
/* When there is a special activation (not supported by the accelerator itself), you need to streamout 32bits and do the act in the cluster but the ((*S) >> N) is done in the accelerator (KOP_DP_REDUCT_NOSCALE) */
int NeedReductNoScale = !(ActOper == KOP_RELU || ActOper == KOP_NONE);
/* Also when in/out are 16bits you need to streamout 32bits but here the reduction step will be done in the cluster (KOP_DP_REDUCT) */
@@ -365,7 +365,7 @@ int CNN_MM_ConvolutionNE16(
/* Im2Col Size is aligned to 16bits for linear padding */
int Im2ColSize = InFeat*Fcx*Fcy;
- int WBuffSize = ALIGN(Im2ColSize, 4);
+ int WBuffSize = Im2ColSize;
int BuffS = 2*ALIGN(Im2ColSize, 4);
/* Layer number of operations and memory bandwidth requirements */
@@ -432,19 +432,18 @@ int CNN_MM_ConvolutionNE16(
unsigned int DEFAULT_NE16_JOB_CFG = NE16_DefaultConfig(Filter_DataSizeBits, Mode16, StreamoutMode, FilterMode, LinearMode, StridedMode, NormBits, Streamin, \
WOffsetCfg, QuantRightShift, QuantBits, QuantNoRect, NormShift, NormBias);
- int Ca=0;
- KCArgs[Ca++] = TCArg(In_DataSize>0?CNN_ArgDataType(In_DataSize,1,1):CNN_ArgDataTypeUns(-In_DataSize,1,1), "In");
- KCArgs[Ca++] = TCArg(CNN_ArgDataTypeUns(1, 1,1), "Filter"); // int16_t for 16 chin contributions
- KCArgs[Ca++] = TCArg(CNN_ArgDataType(Bias_DataSize,1,1), "Bias");
- KCArgs[Ca++] = TCArg(Out_DataSize>0?CNN_ArgDataType(Out_DataSize,1,1):CNN_ArgDataTypeUns(-Out_DataSize,1,1), "Out");
- KCArgs[Ca++] = TCArg(CNN_ArgDataTypeUns(1, 1,1), "Scale");
- KCArgs[Ca++] = TCArg(CNN_ArgDataType(1, 1,1), "ScaleN");
- KCArgs[Ca++] = TCArg(CNN_ArgDataType(1, 1,1), "Infos");
-
Kernel = UserKernel(Name,
- KernelIterSpace(3, IterParSpace(D1, OutFeat, OutTileCons), IterTiledSpace(T0), IterParSpace(D0, InFeat, InTileCons)),
+ KernelIterSpace(3, IterParSpace(D1, OutFeat, OutTileCons), IterTiledSpace(T0), IterParSpace(D0|SPACE_PROP_ONE_TILE, InFeat, InTileCons)),
TILE_HOR|TILE_HWC,
- KCArgs,
+ CArgs(7,
+ TCArg(In_DataSize>0?CNN_ArgDataType(In_DataSize,1,1):CNN_ArgDataTypeUns(-In_DataSize,1,1), "In"),
+ TCArg(CNN_ArgDataTypeUns(1, 1,1), "Filter"), // int16_t for 16 chin contribution
+ TCArg(CNN_ArgDataType(Bias_DataSize,1,1), "Bias"),
+ TCArg(Out_DataSize>0?CNN_ArgDataType(Out_DataSize,1,1):CNN_ArgDataTypeUns(-Out_DataSize,1,1), "Out"),
+ TCArg(CNN_ArgDataTypeUns(1, 1,1), "Scale"),
+ TCArg(CNN_ArgDataType(1, 1,1), "ScaleN"),
+ TCArg(CNN_ArgDataType(1, 1,1), "Infos")
+ ),
Calls(6,
Call("NE16_Enable", LOC_D1_PROLOG, Bindings(0)),
Call("NE16_SoftReset", LOC_D0, Bindings(0)),
@@ -544,7 +543,7 @@ int CNN_MM_ConvolutionNE16(
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PadInp,
ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
static Kernel_T *CNN_ConvolutionNE16_Internal(
@@ -604,6 +603,7 @@ static Kernel_T *CNN_ConvolutionNE16_Internal(
Fpx=1; Dpx=1; Spx=1; Fpy=1; Dpy=1; Spy=1;
}
if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) TileOrientation = (Ctrl->TileOrientation==0)?TILE_HOR:TILE_VER;
if (Ctrl->PadType != -1) PadType = Ctrl->PadType;
}
int OverlapC, OverlapP;
@@ -680,7 +680,7 @@ static Kernel_T *CNN_ConvolutionNE16_Internal(
LayerBandwidth += (Fcx*Fcy*Filter_DataSizeBits*InFeat*(DWConv?1:OutFeat)+7)/8;
LayerBandwidth += Bias_DataSize*OutFeat;
- if (ConvOper == KOP_CONV && Height == 1 && Fcx != 1 && Fcy == 1) ConvOper = KOP_CONV1D;
+ if (ConvOper == KOP_CONV && Height == 1 && Fcy == 1) ConvOper = KOP_CONV1D;
ConvKerName = CNN_FindMatchingKernelAttr(ConvOper, KOP_NONE, ParFeat, CALL_NE16_KER, Abs(In_DataSize), Abs(Out_DataSize), Bias_DataSize, 0, 4, Fcx, Fcy, Dcx, Dcy, Scx, Scy,
&NeedFcx, &NeedFcy, &NeedDcx, &NeedDcy, &NeedScx, &NeedScy, 0);
if (ConvKerName==0) GenTilingError("CNN_ConvolutionPoolAct_NE16 Kernel: %s, Can't find a matching Convolution basic kernel", Name);
@@ -778,7 +778,7 @@ static Kernel_T *CNN_ConvolutionNE16_Internal(
(DWConv?
KernelIterSpace(2, IterParSpace(D0|InFeatProp, InFeat, InTileCons), IterTiledSpace(T0)):
KernelIterSpace(3, IterParSpace(D1, OutFeat, OutTileCons), IterTiledSpace(T0), IterParSpace(D0|InFeatProp, InFeat, InTileCons))),
- TILE_HOR|TILE_HWC,
+ TileOrientation|TILE_HWC,
KCArgs,
Calls(10,
Call("NE16_Enable", DWConv?LOC_D0_PROLOG:LOC_D1_PROLOG, Bindings(0)),
@@ -1122,13 +1122,13 @@ static Kernel_T *CNN_LinearAct_NE16_Internal(
LinearKerName = CNN_FindMatchingKernelAttr(LinearOper, KOP_NONE, 0, CALL_NE16_KER, Abs(In_DataSize), 0, Bias_DataSize, 0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0);
if (LinearKerName==0) GenTilingError("CNN_LinearAct_NE16 Kernel: %s, Can't find a matching Linear basic kernel: %d %d", Name, Abs(In_DataSize), Bias_DataSize);
- if (!(ActOper == KOP_NONE || ActOper == KOP_RELU || ActOper == KOP_RELUN || ActOper == KOP_RELUM || ActOper == KOP_HSIGMOID || ActOper == KOP_HSWISH || ActOper == KOP_LEAKYRELU))
+ if (!(ActOper == KOP_NONE || ActOper == KOP_RELU || ActOper == KOP_RELUN || ActOper == KOP_RELUM || ActOper == KOP_HSIGMOID || ActOper == KOP_SIGMOID || ActOper == KOP_HSWISH || ActOper == KOP_LEAKYRELU || ActOper == KOP_TANH))
GenTilingError("CNN_ConvolutionPoolAct_NE16 Kernel: %s, ActOper, expecting KOP_NONE, KOP_RELU, KOP_RELUN, KOP_RELUM, KOP_HSIGMOID, KOP_HSWISH or KOP_LEAKYRELU", Name);
/* Also when in/out are 16bits you need to streamout 32bits but here the reduction step will be done in the cluster (KOP_DP_REDUCT) */
int NeedReductScale = Mode16; //Abs(Out_DataSize) == 2;
int NeedSetBias = Mode16;
- int NeedReductNoScale = ((InTileCons < InDim) && !InFeatOneTile) || NeedSetBias;
+ int NeedReductNoScale = ((InTileCons < InDim) && !InFeatOneTile) || NeedSetBias || (!(ActOper == KOP_RELU || ActOper == KOP_NONE));
int NeedReduct = NeedReductNoScale || NeedReductScale;
int NeedLinOut = NeedReduct || NeedSetBias;
@@ -1173,32 +1173,18 @@ static Kernel_T *CNN_LinearAct_NE16_Internal(
}
Kernel_T *Kernel;
- CKernel_Arg_T **KCArgs = AllocateCArgs(7);
- int Ca=0;
- KCArgs[Ca++] = TCArg(In_DataSize>0?CNN_ArgDataType(In_DataSize,1,1):CNN_ArgDataTypeUns(-In_DataSize,1,1), "In");
- KCArgs[Ca++] = TCArg(CNN_ArgDataTypeUns(1,1,1), "Filter");
- KCArgs[Ca++] = TCArg(CNN_ArgDataType(Bs,1,1), "Bias");
- KCArgs[Ca++] = TCArg(Out_DataSize>0?CNN_ArgDataType(Out_DataSize,1,1):CNN_ArgDataTypeUns(-Out_DataSize,1,1), "Out");
- KCArgs[Ca++] = TCArg(CNN_ArgDataTypeUns(1,1,1), "Scale");
- KCArgs[Ca++] = TCArg(CNN_ArgDataType(1,1,1), "ScaleN");
- KCArgs[Ca++] = TCArg(CNN_ArgDataType(1,1,1), "Infos");
-
- Object_T **KArgs = AllocateKerArgs(7+NeedLinOut);
- int Ka=0;
- KArgs[Ka++] = KerArg("In", KerArgSpace(1,D0|SPACE_PROP_PAD2PREF), OBJ_IN_DB, 1, 1, Abs(In_DataSize), 0, 0, 0, "In");
- KArgs[Ka++] = KerArg("Filter", KerArgSpace(2,D1,D0|Wp), OBJ_IN_DB|O_CONST|Wa, 1, 1, Ws, 0, 0, 0, "Filter");
- KArgs[Ka++] = KerArg("Bias", KerArgSpace(1,D1), OBJ_IN_DB|O_CONST, 1, 1, Bs, 0, 0, 0, "Bias");
- if (NeedLinOut)
- KArgs[Ka++] = KerArg("LinOut", KerArgSpace(1,D1), O_BUFF|O_ONETILE, 1, 1, Ls, 0, 0, 0, "");
- KArgs[Ka++] = KerArg("Out", KerArgSpace(1,D1), OBJ_OUT_DB, 1, 1, Abs(Out_DataSize), 0, 0, 0, "Out");
- KArgs[Ka++] = KerArg("Scale", KerArgSpace(1,D1), OBJ_IN_DB|O_CONST, 1, 1, 1, 0, 0, 0, "Scale");
- KArgs[Ka++] = KerArg("ScaleN", KerArgSpace(1,D1), OBJ_IN_DB|O_CONST, 1, 1, 1, 0, 0, 0, "ScaleN");
- KArgs[Ka++] = KerArg("Infos", KerArgSpace(1,T0), O_IN|O_BUFF|O_NTILED, 1, 1, AT_INF_NE16_DIM*1, 0, 0, 0, "Infos");
-
Kernel = UserKernel(Name,
KernelIterSpace(3, IterTiledSpace(T0), IterParSpace(D1, OutDim, OutTileCons), IterParSpace(D0|InFeatProp, InDim, InTileCons)),
TileOrientation,
- KCArgs,
+ CArgs(7,
+ TCArg(In_DataSize>0?CNN_ArgDataType(In_DataSize,1,1):CNN_ArgDataTypeUns(-In_DataSize,1,1), "In"),
+ TCArg(CNN_ArgDataTypeUns(1,1,1), "Filter"),
+ TCArg(CNN_ArgDataType(Bs,1,1), "Bias"),
+ TCArg(Out_DataSize>0?CNN_ArgDataType(Out_DataSize,1,1):CNN_ArgDataTypeUns(-Out_DataSize,1,1), "Out"),
+ TCArg(CNN_ArgDataTypeUns(1,1,1), "Scale"),
+ TCArg(CNN_ArgDataType(1,1,1), "ScaleN"),
+ TCArg(CNN_ArgDataType(1,1,1), "Infos")
+ ),
Calls(6,
Call("NE16_Enable", LOC_PROLOG, Bindings(0)),
SetBiasKerName?Call(SetBiasKerName, LOC_D0_PROLOG,
@@ -1244,7 +1230,16 @@ static Kernel_T *CNN_LinearAct_NE16_Internal(
),
Call("NE16_Disable", LOC_EPILOG, Bindings(0))
),
- KArgs
+ KerArgs(8,
+ KerArg("In", KerArgSpace(1,D0|SPACE_PROP_PAD2PREF),OBJ_IN_DB, 1, 1, Abs(In_DataSize), 0, 0, 0, "In"),
+ KerArg("Filter", KerArgSpace(2,D1,D0|Wp), OBJ_IN_DB|O_CONST|Wa, 1, 1, Ws, 0, 0, 0, "Filter"),
+ KerArg("Bias", KerArgSpace(1,D1), OBJ_IN_DB|O_CONST, 1, 1, Bs, 0, 0, 0, "Bias"),
+ NeedLinOut?KerArg("LinOut", KerArgSpace(1,D1), O_BUFF|O_ONETILE, 1, 1, Ls, 0, 0, 0, ""):AT_NO_KER_ARG,
+ KerArg("Out", KerArgSpace(1,D1), OBJ_OUT_DB, 1, 1, Abs(Out_DataSize), 0, 0, 0, "Out"),
+ KerArg("Scale", KerArgSpace(1,D1), OBJ_IN_DB|O_CONST, 1, 1, 1, 0, 0, 0, "Scale"),
+ KerArg("ScaleN", KerArgSpace(1,D1), OBJ_IN_DB|O_CONST, 1, 1, 1, 0, 0, 0, "ScaleN"),
+ KerArg("Infos", KerArgSpace(1,T0), O_IN|O_BUFF|O_NTILED, 1, 1, AT_INF_NE16_DIM*1, 0, 0, 0, "Infos")
+ )
);
if (Kernel) {
AddKernelInfos(Name, AT_KERINFO_OPER, LayerOp, 0);
@@ -1385,6 +1380,8 @@ int CNN_MatMulAct_NE16(
)
{
+ return CNN_ConvolutionNE16(Name, Ctrl, In_DataSize, Out_DataSize, Bias_DataSize, 1, Filter_DataSizeBits, ColM1, ColM2, LineM1, 1, KOP_CONV,1,1,1,1,1,1,0,0, KOP_NONE,0,0,0,0,0,0,0, ActOper);
+#if 0
int Log=1;
if (Abs(In_DataSize)!=1 && Abs(In_DataSize)!=2) GenTilingError("Node: %s Input DataSize %d not supported in NE16", Name, In_DataSize);
@@ -1555,4 +1552,5 @@ int CNN_MatMulAct_NE16(
AddKernelArgDim(Name, "Infos", 2, AT_INF_DIM, 1);
}
return Kernel!=0;
+#endif
}
\ No newline at end of file
diff --git a/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.h b/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.h
index ac6285a00..b67494872 100644
--- a/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.h
+++ b/tools/autotiler_v3/CNN_Generators_NE16/CNN_Generators_NE16.h
@@ -59,7 +59,7 @@ int CNN_ConvolutionNE16(
KernelOper_T ActOper
);
-int CNN_MM_ConvolutionNE16(
+Kernel_T *CNN_MM_ConvolutionNE16(
char *Name,
CNN_GenControl_T *Ctrl,
diff --git a/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.c b/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.c
index edc19bab0..99d97952d 100644
--- a/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.c
+++ b/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.c
@@ -730,6 +730,10 @@ void LoadCNN_SQ8_Library()
CNN_Match(CNN_OperList(1, KOP_MM_CONV), CNN_OperList(1, KOP_NONE), 1, CNN_Type(1,1,4,0,1), 1,1,1,1,-1,-1));
LibKernel("KerPar_MM_Conv1x1_ReLU_HWC_SQ8", CALL_PARALLEL|CALL_HWC_KER, 0, "Ker_MM_Conv_SQ8_T",
CNN_Match(CNN_OperList(1, KOP_MM_CONV), CNN_OperList(1, KOP_RELU), 1, CNN_Type(1,1,4,0,1), 1,1,1,1,-1,-1));
+ LibKernel("Ker_MM_Conv1x1_HWC_SQ8", CALL_PARALLEL|CALL_HWC_KER, 0, "Ker_MM_Conv_SQ8_T",
+ CNN_Match(CNN_OperList(1, KOP_MM_CONV), CNN_OperList(1, KOP_NONE), 0, CNN_Type(1,1,4,0,1), 1,1,1,1,-1,-1));
+ LibKernel("Ker_MM_Conv1x1_ReLU_HWC_SQ8", CALL_PARALLEL|CALL_HWC_KER, 0, "Ker_MM_Conv_SQ8_T",
+ CNN_Match(CNN_OperList(1, KOP_MM_CONV), CNN_OperList(1, KOP_RELU), 0, CNN_Type(1,1,4,0,1), 1,1,1,1,-1,-1));
LibKernel("KerPar_MM_Conv1D_HWC_SQ8", CALL_PARALLEL|CALL_HWC_KER, 0, "Ker_MM_Conv_SQ8_T",
CNN_Match(CNN_OperList(1, KOP_MM_CONV), CNN_OperList(1, KOP_NONE), 1, CNN_Type(1,1,4,0,1), -1,1,1,1,-1,-1));
LibKernel("KerPar_MM_Conv1D_DxDy_HWC_SQ8", CALL_PARALLEL|CALL_HWC_KER, 0, "Ker_MM_Conv_SQ8_T",
@@ -1140,7 +1144,7 @@ void LoadCNN_SQ8_Library()
*********************************************************************************************************************************************************************/
-int CNN_MM_ConvolutionPoolAct_SQ8(
+static Kernel_T *CNN_MM_ConvolutionPoolAct_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -1177,9 +1181,9 @@ int CNN_MM_ConvolutionPoolAct_SQ8(
{
if (ConvOper==KOP_NONE) {
if (PoolOper!=KOP_NONE)
- return CNN_PoolAct_SQ8(Name, Ctrl, InFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ return CNN_PoolAct_SQ8_Internal(Name, Ctrl, InFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else if (ActOper!=KOP_NONE)
- return CNN_Act_SQ8(Name, Ctrl, InFeat, Width, Height, ActOper);
+ return CNN_Act_SQ8_Internal(Name, Ctrl, InFeat, Width, Height, ActOper);
else GenTilingError("CNN_MM_ConvolutionPoolAct_SQ8 Kernel: %s, All requested operations are KOP_NONE", Name);
}
@@ -1197,6 +1201,10 @@ int CNN_MM_ConvolutionPoolAct_SQ8(
if (Ctrl->HWC) HWC = 1;
if (Ctrl->ParallelFeatures != -1) ParFeatConv = Ctrl->ParallelFeatures;
}
+
+ if (HWC && Fcy==1 && Fcx==1 && Scy==1 && Scx==1 && Dcy==1 && Dcx==1)
+ return CNN_MatMulAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, Height*Width, OutFeat, InFeat, 0,0,0,0, KOP_MATMUL_TRANSPOSED, ActOper, 0);
+
if (ParFeatConv == 2 && HWC && Fcy>1 && (InFeat < 8))
ParFeatConv = 0;
else
@@ -1253,7 +1261,7 @@ int CNN_MM_ConvolutionPoolAct_SQ8(
int BuffS = ALIGN(InFeat*Fcx*Fcy, 3);
if (HWC) {
if (Fcx==1&&Fcy==1) BuffS = 1;
- else if (ParFeatConv) BuffS = ALIGN(InFeat*Fcx*Fcy, 3);
+ else if (ParFeatConv) BuffS = 2*ALIGN(InFeat*Fcx*Fcy, 3);
else BuffS = 2 * InFeat*Fcx*Fcy*8;
}
@@ -1453,11 +1461,11 @@ int CNN_MM_ConvolutionPoolAct_SQ8(
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PadInp,
ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
-int CNN_HWC_DWConvolutionPoolAct_SQ8(
+static Kernel_T *CNN_HWC_DWConvolutionPoolAct_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -1494,9 +1502,9 @@ int CNN_HWC_DWConvolutionPoolAct_SQ8(
{
if (ConvOper==KOP_NONE) {
if (PoolOper!=KOP_NONE)
- return CNN_PoolAct_SQ8(Name, Ctrl, InFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ return CNN_PoolAct_SQ8_Internal(Name, Ctrl, InFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else if (ActOper!=KOP_NONE)
- return CNN_Act_SQ8(Name, Ctrl, InFeat, Width, Height, ActOper);
+ return CNN_Act_SQ8_Internal(Name, Ctrl, InFeat, Width, Height, ActOper);
else GenTilingError("CNN_HWC_DWConvolutionPoolAct_SQ8: %s, All requested operations are KOP_NONE", Name);
}
@@ -1606,7 +1614,7 @@ int CNN_HWC_DWConvolutionPoolAct_SQ8(
}
if (Log) {
- printf("InFeat: %d, OutFeat: %d%s\n", InFeat, OutFeat, HWC?", HWC":", CHW");
+ printf("InFeat: %d, OutFeat: %d%s - TileOrientation: %s\n", InFeat, OutFeat, HWC?", HWC":", CHW", TileOrientation==TILE_HOR?"TILE_HOR":"TILE_VER");
printf("Conv => W: %d, Pad:[%d,%d] PadT:[%d,%d] => Wc: %d, Filter:[%d,%d]\n", Width, PadInc[0], PadInc[1], PadIncT[0], PadIncT[1], Wc, Fcx, Fcy);
printf(" => H: %d, Pad:[%d,%d] PadT:[%d,%d] => Hc: %d\n", Height, PadInc[2], PadInc[3], PadIncT[2], PadIncT[3], Hc);
printf("Pool => Wc: %d, Pad:[%d,%d] => Wo: %d, Filter:[%d,%d]\n", Wc, PadInp[0], PadInp[1], Wo, Fpx, Fpy);
@@ -1744,7 +1752,7 @@ int CNN_HWC_DWConvolutionPoolAct_SQ8(
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PadInp,
ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -1790,7 +1798,7 @@ int CNN_HWC_DWConvolutionPoolAct_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_ConvolutionPoolAct_SQ8(
+Kernel_T *CNN_ConvolutionPoolAct_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -1850,31 +1858,31 @@ int CNN_ConvolutionPoolAct_SQ8(
}
if (ConvOper==KOP_NONE) {
if (PoolOper!=KOP_NONE)
- return CNN_PoolAct_SQ8(Name, Ctrl, InFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ return CNN_PoolAct_SQ8_Internal(Name, Ctrl, InFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else if (ActOper!=KOP_NONE)
- return CNN_Act_SQ8(Name, Ctrl, InFeat, Width, Height, ActOper);
+ return CNN_Act_SQ8_Internal(Name, Ctrl, InFeat, Width, Height, ActOper);
else GenTilingError("CNN_ConvolutionPoolAct_SQ8 Kernel: %s, All requested operations are KOP_NONE", Name);
} else if (HWC) {
if (ConvOper == KOP_CONV_DW)
- return CNN_HWC_DWConvolutionPoolAct_SQ8(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height,
+ return CNN_HWC_DWConvolutionPoolAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height,
ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad,
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else
- return CNN_MM_ConvolutionPoolAct_SQ8(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height,
+ return CNN_MM_ConvolutionPoolAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height,
ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad,
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
} else if (ConvOper==KOP_CONV && ((Fcx > 1 && Fcy == 1))) {
- AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
- int Ok = CNN_MM_ConvolutionPoolAct_SQ8(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height,
+ // AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+ Kernel_T *Ok = CNN_MM_ConvolutionPoolAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height,
ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad,
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
- AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
- if (Ok) return Ok;
+ // AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ if (Ok!=0) return Ok;
if (Log) printf("No solution found for im2col scheme, reverting to standard implementation\n");
}
if (Fcx==1 && Fcy==1 && Scx==1 && Scy==1 && Dcx==1 && Dcy==1 && Height==1 && Width==1) {
printf("This is a pointwise on 1x1 input --> Mapping to CNN_Linear_NE16\n");
- return CNN_LinearAct_SQ8(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, KOP_LINEAR, ActOper);
+ return CNN_LinearAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, KOP_LINEAR, ActOper);
}
if (PoolOper==KOP_NONE) {
@@ -1997,12 +2005,12 @@ int CNN_ConvolutionPoolAct_SQ8(
// if ((InFeat+OutFeat)<80) {
if ((InFeat+OutFeat)<100 && (Scx==1) && (Scy==1)) {
if (Log) printf("Mapping this convolution to matrix multiplication with small first operand\n");
- int Ok = CNN_MatMulSmallM1Act_SQ8(Name, 0, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL_SM1, ActOper);
+ Kernel_T *Ok = CNN_MatMulSmallM1Act_SQ8_Internal(Name, 0, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL_SM1, ActOper);
if (!Ok&&Log) printf("Mapping this convolution to matrix multiplication with small first operand FAILED, trying with standard mult implementation\n");
if (Ok) return Ok;
}
if (Log) printf("Mapping this convolution to matrix multiplication\n");
- int Ok = CNN_MatMulAct_SQ8(Name, 0, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL, ActOper);
+ Kernel_T *Ok = CNN_MatMulAct_SQ8_Internal(Name, 0, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL, ActOper, 1);
AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
if (Ok) return Ok;
if (Log) printf("Mapping this convolution to matrix multiplication FAILED, reverting to standard implementation\n");
@@ -2159,7 +2167,7 @@ int CNN_ConvolutionPoolAct_SQ8(
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PadInp,
ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -2345,7 +2353,7 @@ int CNN_GroupedConvolutionPoolAct_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_PoolAct_SQ8(
+Kernel_T * CNN_PoolAct_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2367,7 +2375,7 @@ int CNN_PoolAct_SQ8(
)
{
- if (PoolOper==KOP_NONE && ActOper!=KOP_NONE) return CNN_Act_SQ8(Name, Ctrl, Feat, Width, Height, ActOper);
+ if (PoolOper==KOP_NONE && ActOper!=KOP_NONE) return CNN_Act_SQ8_Internal(Name, Ctrl, Feat, Width, Height, ActOper);
Tile_Orientation_T TileOrientation = TILE_HOR;
int ParFeat = 1, HWC = 0;
@@ -2524,10 +2532,11 @@ int CNN_PoolAct_SQ8(
AT_PrepareForTest_SQ8(Name, Feat, Feat, Width, Height, 1, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PadInp, 0, 0,0,0,0,0,0,(v4s) 0, ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
+
/*********************************************************************************************************************************************************************
Generator for Activation with tensor centric scaling
@@ -2549,7 +2558,7 @@ int CNN_PoolAct_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_Act_SQ8(
+Kernel_T * CNN_Act_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2626,9 +2635,10 @@ int CNN_Act_SQ8(
AT_PrepareForTest_SQ8(Name, Feat,Feat,Width,Height, 1, ActOper, 0,0,0,0,0,0,(v4s) 0, 0, 0,0,0,0,0,0,(v4s) 0, KOP_NONE);
}
- return (Kernel!=0);
+ return Kernel;
}
+
/*********************************************************************************************************************************************************************
Generator for Global Pooling (Max or Average) with tensor centric scaling and optional activation
@@ -2652,7 +2662,7 @@ int CNN_Act_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_GlobalPoolAct_SQ8(
+static Kernel_T *CNN_GlobalPoolAct_SQ8_Interal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2843,10 +2853,9 @@ int CNN_GlobalPoolAct_SQ8(
AT_PrepareForTest_SQ8(Name, Feat,Feat,Width,Height, 1, PoolOper, 0,0,0,0,0,0, (v4s)0, 0, 0,0,0,0,0,0,(v4s) 0, ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
-
/*********************************************************************************************************************************************************************
Generator for Linear layers followed wth channel centric scaling followed by an optional activation
@@ -2870,7 +2879,7 @@ int CNN_GlobalPoolAct_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_LinearAct_SQ8(
+Kernel_T * CNN_LinearAct_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -3074,7 +3083,7 @@ int CNN_LinearAct_SQ8(
AT_PrepareForTest_SQ8(Name, InDim,OutDim,1,1, Bias_DataSize, LinearOper, 0,0,0,0,0,0,(v4s)0, 0, 0,0,0,0,0,0,(v4s)0, ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -3096,7 +3105,7 @@ int CNN_LinearAct_SQ8(
a different code. By definition Output contains value is the [0.0 .. 1.0] range with sum(Output)=1.0. Results are always represented in Q15
*********************************************************************************************************************************************************************/
-int CNN_SoftMax_SQ8(
+static Kernel_T * CNN_SoftMax_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -3158,10 +3167,10 @@ int CNN_SoftMax_SQ8(
AT_PrepareForTest_SQ8(Name, Dim,Dim,1,1, 1, SoftMaxOper, 0,0,0,0,0,0,(v4s)0, 0, 0,0,0,0,0,0,(v4s)0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
-int CNN_SoftMax2D_SQ8(
+static Kernel_T * CNN_SoftMax2D_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -3231,7 +3240,7 @@ int CNN_SoftMax2D_SQ8(
AT_PrepareForTest_SQ8(Name, Dim,Dim,1,1, 1, SoftMaxOper, 0,0,0,0,0,0,(v4s)0, 0, 0,0,0,0,0,0,(v4s)0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -3255,7 +3264,7 @@ int CNN_SoftMax2D_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_MatAddAct_SQ8(
+static Kernel_T * CNN_MatAddAct_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -3350,7 +3359,7 @@ int CNN_MatAddAct_SQ8(
AT_PrepareForTest_SQ8(Name, Feat,Feat,Width,Height, 1, AddMatOper, 0,0,0,0,0,0,(v4s)0, 0, 0,0,0,0,0,0,(v4s)0, ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -3711,7 +3720,7 @@ int CNN_TensorVectMultAct_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_MatMulAct_SQ8(
+Kernel_T *CNN_MatMulAct_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -3730,7 +3739,8 @@ int CNN_MatMulAct_SQ8(
int Scy,
KernelOper_T MatMulOper,
- KernelOper_T ActOper
+ KernelOper_T ActOper,
+ int InvertInputs
)
{
@@ -3804,8 +3814,8 @@ int CNN_MatMulAct_SQ8(
KernelIterSpace(2, IterTiledSpace(T1), IterTiledSpace(T0)),
TILE_HOR,
CArgs(7,
- TCArg(CNN_ArgDataType(1,1,1), "In2"),
- TCArg(CNN_ArgDataType(1,1,1), "In1"),
+ TCArg(CNN_ArgDataType(1,1,1), InvertInputs?"In2":"In1"),
+ TCArg(CNN_ArgDataType(1,1,1), InvertInputs?"In1":"In2"),
Bias_DataSize?TCArg(CNN_ArgDataType(Bias_DataSize,1,1), "Bias"):AT_NO_C_ARG,
TCArg(CNN_ArgDataType(1,1,1), "Out"),
!ScaleScalar?TCArg(CNN_ArgDataTypeUns(1,1,1),"Scale"):AT_NO_C_ARG,
@@ -3879,7 +3889,7 @@ int CNN_MatMulAct_SQ8(
AT_PrepareForTest_SQ8(Name, ColM1,LineM1,Width,Height, Bias_DataSize, MatMulOper, 1,1,1,1,Scx,Scy,(v4s)0, 0, 0,0,0,0,0,0,(v4s)0, ActOper);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -3920,7 +3930,7 @@ int CNN_MatMulAct_SQ8(
*********************************************************************************************************************************************************************/
-int CNN_MatMulSmallM1Act_SQ8(
+Kernel_T * CNN_MatMulSmallM1Act_SQ8_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -4077,7 +4087,295 @@ int CNN_MatMulSmallM1Act_SQ8(
AT_PrepareForTest_SQ8(Name, ColM1,LineM1,Width,Height, Bias_DataSize, MatMulOper, 1,1,1,1,Scx,Scy,(v4s)0, 0, 0,0,0,0,0,0,(v4s)0, ActOper);
}
- return (Kernel!=0);
+ return Kernel;
+
+}
+
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+
+int CNN_MM_ConvolutionPoolAct_SQ8(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Bias_DataSize,
+ int Scale_DataSize,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ )
+{
+ Kernel_T *Ker = 0, *Sol1 = 0, *Sol2 = 0;
+ float K = 0.9;
+ Tile_Orientation_T TileOrientation = TILE_HOR;
+ if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) {
+ printf("TileOrientation set by user\n");
+ Ker = CNN_MM_ConvolutionPoolAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker!=0) return 1;
+ else GenTilingError("CNN_MM_ConvolutionPoolAct_SQ8: %s, Failed to gen with set tiling orientation, try to let the Autotiler set it for you", Name);
+ }
+ }
+
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+
+ CNN_GenControl_T InternalCtrl;
+ if (!Ctrl) CNN_InitGenCtrl(&InternalCtrl);
+ else InternalCtrl = *Ctrl;
+
+ printf("\n\n=============================== Trying Tile Orientation: TILE_HOR ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(0));
+ Ker = CNN_MM_ConvolutionPoolAct_SQ8_Internal(Name, &InternalCtrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol1 = CopyAndPopUserKernel(Ker);
+
+ printf("\n=============================== Trying Tile Orientation: TILE_VER ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(1));
+ Ker = CNN_MM_ConvolutionPoolAct_SQ8_Internal(Name, &InternalCtrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol2 = CopyAndPopUserKernel(Ker);
+
+ if (Sol1 && Sol2) {
+ int TakeSol1 = ((K*Sol1->Cost->TileOverhead) < Sol2->Cost->TileOverhead); // K close to 1.0if (TakeSol1) {
+ printf(">>>>>>>>>>>>>>>>>> %s is better: %.3f vs %.3f \n\n\n", TakeSol1?"TILE_HOR":"TILE_VER", Sol1->Cost->TileOverhead, Sol2->Cost->TileOverhead);
+ if (TakeSol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ }
+ } else if (Sol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else if (Sol2) {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ } else {
+ GenTilingError("Failed to Generate code for Kernel: %s", Name);
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ return 1;
+}
+
+int CNN_HWC_DWConvolutionPoolAct_SQ8(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Bias_DataSize,
+ int Scale_DataSize,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ )
+{
+ Kernel_T *Ker = 0, *Sol1 = 0, *Sol2 = 0;
+ float K = 0.9;
+ Tile_Orientation_T TileOrientation = TILE_HOR;
+ if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) {
+ printf("TileOrientation set by user\n");
+ Ker = CNN_HWC_DWConvolutionPoolAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker!=0) return 1;
+ else GenTilingError("CNN_MM_ConvolutionPoolAct_SQ8: %s, Failed to gen with set tiling orientation, try to let the Autotiler set it for you", Name);
+ }
+ }
+
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+
+ CNN_GenControl_T InternalCtrl;
+ if (!Ctrl) CNN_InitGenCtrl(&InternalCtrl);
+ else InternalCtrl = *Ctrl;
+
+ printf("\n\n=============================== Trying Tile Orientation: TILE_HOR ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(0));
+ Ker = CNN_HWC_DWConvolutionPoolAct_SQ8_Internal(Name, &InternalCtrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol1 = CopyAndPopUserKernel(Ker);
+
+ printf("\n=============================== Trying Tile Orientation: TILE_VER ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(1));
+ Ker = CNN_HWC_DWConvolutionPoolAct_SQ8_Internal(Name, &InternalCtrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol2 = CopyAndPopUserKernel(Ker);
+
+ if (Sol1 && Sol2) {
+ int TakeSol1 = ((K*Sol1->Cost->TileOverhead) < Sol2->Cost->TileOverhead); // K close to 1.0if (TakeSol1) {
+ printf(">>>>>>>>>>>>>>>>>> %s is better: %.3f vs %.3f \n\n\n", TakeSol1?"TILE_HOR":"TILE_VER", Sol1->Cost->TileOverhead, Sol2->Cost->TileOverhead);
+ if (TakeSol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ }
+ } else if (Sol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else if (Sol2) {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ } else {
+ GenTilingError("Failed to Generate code for Kernel: %s", Name);
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ return 1;
+}
+
+int CNN_ConvolutionPoolAct_SQ8(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Bias_DataSize,
+ int Scale_DataSize,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ )
+{
+ Kernel_T *Ker = 0, *Sol1 = 0, *Sol2 = 0;
+ float K = 0.9;
+ Tile_Orientation_T TileOrientation = TILE_HOR;
+ if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) {
+ printf("TileOrientation set by user\n");
+ Ker = CNN_ConvolutionPoolAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker!=0) return 1;
+ else GenTilingError("CNN_ConvolutionPoolAct_SQ8: %s, Failed to gen with set tiling orientation, try to let the Autotiler set it for you", Name);
+ }
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+
+ CNN_GenControl_T InternalCtrl;
+ if (!Ctrl) CNN_InitGenCtrl(&InternalCtrl);
+ else InternalCtrl = *Ctrl;
+
+ printf("\n\n=============================== Trying Tile Orientation: TILE_HOR ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(0));
+ Ker = CNN_ConvolutionPoolAct_SQ8_Internal(Name, &InternalCtrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol1 = CopyAndPopUserKernel(Ker);
+
+ printf("\n=============================== Trying Tile Orientation: TILE_VER ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(1));
+ Ker = CNN_ConvolutionPoolAct_SQ8_Internal(Name, &InternalCtrl, Bias_DataSize, Scale_DataSize, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol2 = CopyAndPopUserKernel(Ker);
+
+ if (Sol1 && Sol2) {
+ int TakeSol1 = ((K*Sol1->Cost->TileOverhead) < Sol2->Cost->TileOverhead); // K close to 1.0if (TakeSol1) {
+ printf(">>>>>>>>>>>>>>>>>> %s is better: %.3f vs %.3f \n\n\n", TakeSol1?"TILE_HOR":"TILE_VER", Sol1->Cost->TileOverhead, Sol2->Cost->TileOverhead);
+ if (TakeSol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ }
+ } else if (Sol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else if (Sol2) {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ } else {
+ GenTilingError("Failed to Generate code for Kernel: %s", Name);
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ return 1;
+}
+
+int CNN_PoolAct_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Feat, int Width, int Height, KernelOper_T PoolOper, int Fpx, int Fpy, int Dpx, int Dpy, int Spx, int Spy, int PoolPad, KernelOper_T ActOper)
+{
+ return (CNN_PoolAct_SQ8_Internal(Name, Ctrl, Feat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper)!=0);
+}
+
+int CNN_Act_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Feat, int Width, int Height,KernelOper_T ActOper) {
+ return (CNN_Act_SQ8_Internal(Name, Ctrl, Feat, Width, Height, ActOper)!=0);
+}
+
+int CNN_GlobalPoolAct_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Feat, int Width, int Height, KernelOper_T PoolOper, KernelOper_T ActOper) {
+ return (CNN_GlobalPoolAct_SQ8_Interal(Name, Ctrl, Feat, Width, Height, PoolOper, ActOper)!=0);
+}
+
+int CNN_LinearAct_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Bias_DataSize, int Scale_DataSize, int InDim, int OutDim, KernelOper_T LinearOper, KernelOper_T ActOper)
+{
+ return (CNN_LinearAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, InDim, OutDim, LinearOper, ActOper)!=0);
+}
+
+int CNN_SoftMax_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Dim, KernelOper_T SoftMaxOper) {
+ return (CNN_SoftMax_SQ8_Internal(Name, Ctrl, Dim, SoftMaxOper)!=0);
+}
+
+int CNN_SoftMax2D_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Dim, int N, KernelOper_T SoftMaxOper) {
+ return (CNN_SoftMax2D_SQ8_Internal(Name, Ctrl, Dim, N, SoftMaxOper)!=0);
+}
+
+int CNN_MatAddAct_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Feat, int Width, int Height, KernelOper_T AddMatOper, KernelOper_T ActOper) {
+ return (CNN_MatAddAct_SQ8_Internal(Name, Ctrl, Feat, Width, Height, AddMatOper, ActOper)!=0);
+}
+
+int CNN_MatMulAct_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Bias_DataSize, int Scale_DataSize, int ColM1, int LineM1, int ColM2, int LineM2, int Width, int Height, int Scx, int Scy, KernelOper_T MatMulOper, KernelOper_T ActOper) {
+ return (CNN_MatMulAct_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, ColM1, LineM1, ColM2, LineM2, Width, Height, Scx, Scy, MatMulOper, ActOper, 1)!=0);
+}
+int CNN_MatMulSmallM1Act_SQ8(char *Name, CNN_GenControl_T *Ctrl, int Bias_DataSize, int Scale_DataSize, int ColM1, int LineM1, int ColM2, int LineM2, int Width, int Height, int Scx, int Scy, KernelOper_T MatMulOper, KernelOper_T ActOper) {
+ return (CNN_MatMulSmallM1Act_SQ8_Internal(Name, Ctrl, Bias_DataSize, Scale_DataSize, ColM1, LineM1, ColM2, LineM2, Width, Height, Scx, Scy, MatMulOper, ActOper)!=0);
}
diff --git a/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.h b/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.h
index 6b0ca7d15..ba223b55d 100644
--- a/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.h
+++ b/tools/autotiler_v3/CNN_Generators_SQ8/CNN_Generators_SQ8.h
@@ -67,6 +67,40 @@ void LoadCNN_SQ8_Library();
*********************************************************************************************************************************************************************/
+Kernel_T *CNN_ConvolutionPoolAct_SQ8_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Bias_DataSize,
+ int Scale_DataSize,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ );
+
int CNN_ConvolutionPoolAct_SQ8(
char *Name,
@@ -302,6 +336,27 @@ int CNN_PoolAct_SQ8(
KernelOper_T ActOper
);
+Kernel_T *CNN_PoolAct_SQ8_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Feat,
+ int Width,
+ int Height,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ );
+
/*********************************************************************************************************************************************************************
Generator for Activation with tensor centric scaling
@@ -335,6 +390,18 @@ int CNN_Act_SQ8(
KernelOper_T ActOper
);
+Kernel_T *CNN_Act_SQ8_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Feat,
+ int Width,
+ int Height,
+
+ KernelOper_T ActOper
+ );
+
/*********************************************************************************************************************************************************************
Generator for Global Pooling (Max or Average) with tensor centric scaling and optional activation
@@ -411,6 +478,21 @@ int CNN_LinearAct_SQ8(
KernelOper_T ActOper
);
+Kernel_T *CNN_LinearAct_SQ8_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Bias_DataSize,
+ int Scale_DataSize,
+
+ int InDim,
+ int OutDim,
+
+ KernelOper_T LinearOper,
+ KernelOper_T ActOper
+ );
+
/*********************************************************************************************************************************************************************
Generator for SoftMax layers, no scaling
@@ -616,6 +698,30 @@ int CNN_MatMulAct_SQ8(
KernelOper_T ActOper
);
+Kernel_T *CNN_MatMulAct_SQ8_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Bias_DataSize,
+ int Scale_DataSize,
+
+ int ColM1,
+ int LineM1,
+ int ColM2,
+ int LineM2,
+
+ int Width,
+ int Height,
+ int Scx,
+ int Scy,
+
+ KernelOper_T MatMulOper,
+ KernelOper_T ActOper,
+ int InvertInputs
+ );
+
+
/*********************************************************************************************************************************************************************
Generator for Matrix Multiplication layers with channel centric scaling followed by an optional Activation.
Special form to handle small form factor In1 (InFeat x OutFeat)
@@ -676,4 +782,26 @@ int CNN_MatMulSmallM1Act_SQ8(
KernelOper_T ActOper
);
+Kernel_T *CNN_MatMulSmallM1Act_SQ8_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Bias_DataSize,
+ int Scale_DataSize,
+
+ int ColM1,
+ int LineM1,
+ int ColM2,
+ int LineM2,
+
+ int Width,
+ int Height,
+ int Scx,
+ int Scy,
+
+ KernelOper_T MatMulOper,
+ KernelOper_T ActOper
+ );
+
#endif
diff --git a/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.c b/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.c
index 40ca2284c..659e2e34e 100644
--- a/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.c
+++ b/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.c
@@ -595,7 +595,7 @@ void LoadCNNLibrary_fp16()
*********************************************************************************************************************************************************************/
-int CNN_MM_ConvolutionPoolAct_fp16(
+Kernel_T *CNN_MM_ConvolutionPoolAct_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -629,20 +629,21 @@ int CNN_MM_ConvolutionPoolAct_fp16(
{
if (ConvOper==KOP_NONE) {
if (PoolOper!=KOP_NONE)
- return CNN_PoolAct_fp16(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ return CNN_PoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else if (ActOper!=KOP_NONE)
- return CNN_Act_fp16(Name, Ctrl, InFeat, Width, Height, ActOper);
+ return CNN_Act_fp16_Internal(Name, Ctrl, InFeat, Width, Height, ActOper);
else GenTilingError("CNN_MM_ConvolutionPoolAct_fp16 Kernel: %s, All requested operations are KOP_NONE", Name);
}
int ParFeat = 1, HWC = 0, ParFeatConv = 2;
float UB = (ActOper==KOP_HSIGMOID)?3.0:6.0; // In Case of HSIGMOID, UB is the Offset (default: 3.0)
- Tile_Orientation_T TileOrientation = Height>1?TILE_HOR:TILE_VER;
+ Tile_Orientation_T TileOrientation = TILE_HOR;
AT_PadType PadType = PAD_BALANCED_LEFT;
if (PoolOper==KOP_NONE) {
Fpx=1; Dpx=1; Spx=1; Fpy=1; Dpy=1; Spy=1;
}
if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) TileOrientation = (Ctrl->TileOrientation==0)?TILE_HOR:TILE_VER;
if (Ctrl->PadType != -1) PadType = Ctrl->PadType;
if (Ctrl->HWC) HWC = 1;
if (Ctrl->ParallelFeatures != -1) ParFeatConv = Ctrl->ParallelFeatures;
@@ -742,7 +743,7 @@ int CNN_MM_ConvolutionPoolAct_fp16(
}
if (Log) {
- printf("InFeat: %d, OutFeat: %d%s%s\n", InFeat, OutFeat, HWC?", HWC":", CHW", ParFeatConv?", Out Chan Parallel":", H Parallel");
+ printf("InFeat: %d, OutFeat: %d%s - TileOrientation: %s\n", InFeat, OutFeat, HWC?", HWC":", CHW", TileOrientation==TILE_HOR?"TILE_HOR":"TILE_VER");
printf("Conv => W: %d, Pad:[%d,%d] PadT:[%d,%d] => Wc: %d, Filter:[%d,%d]\n", Width, PadInc[0], PadInc[1], PadIncT[0], PadIncT[1], Wc, Fcx, Fcy);
printf(" => H: %d, Pad:[%d,%d] PadT:[%d,%d] => Hc: %d\n", Height, PadInc[2], PadInc[3], PadIncT[2], PadIncT[3], Hc);
printf("Pool => Wc: %d, Pad:[%d,%d] => Wo: %d, Filter:[%d,%d]\n", Wc, PadInp[0], PadInp[1], Wo, Fpx, Fpy);
@@ -881,11 +882,11 @@ int CNN_MM_ConvolutionPoolAct_fp16(
Fcx, Fcy, Scx, Scy, Dcx, Dcy, PadInc, Fpx, Fpy, Spx, Spy, Dpx, Dpy, PadInp, KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
-int CNN_HWC_DWConvolutionPoolAct_fp16(
+Kernel_T *CNN_HWC_DWConvolutionPoolAct_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -919,9 +920,9 @@ int CNN_HWC_DWConvolutionPoolAct_fp16(
{
if (ConvOper==KOP_NONE) {
if (PoolOper!=KOP_NONE)
- return CNN_PoolAct_fp16(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ return CNN_PoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else if (ActOper!=KOP_NONE)
- return CNN_Act_fp16(Name, Ctrl, InFeat, Width, Height, ActOper);
+ return CNN_Act_fp16_Internal(Name, Ctrl, InFeat, Width, Height, ActOper);
else GenTilingError("CNN_HWC_DWConvolutionPoolAct_fp16: %s, All requested operations are KOP_NONE", Name);
}
@@ -933,6 +934,7 @@ int CNN_HWC_DWConvolutionPoolAct_fp16(
Fpx=1; Dpx=1; Spx=1; Fpy=1; Dpy=1; Spy=1;
}
if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) TileOrientation = (Ctrl->TileOrientation==0)?TILE_HOR:TILE_VER;
if (Ctrl->PadType != -1) PadType = Ctrl->PadType;
if (Ctrl->ReluN != -1) UB = Ctrl->ReluN;
}
@@ -1024,7 +1026,7 @@ int CNN_HWC_DWConvolutionPoolAct_fp16(
}
if (Log) {
- printf("InFeat: %d, OutFeat: %d%s\n", InFeat, OutFeat, HWC?", HWC":", CHW");
+ printf("InFeat: %d, OutFeat: %d%s - TileOrientation: %s\n", InFeat, OutFeat, HWC?", HWC":", CHW", TileOrientation==TILE_HOR?"TILE_HOR":"TILE_VER");
printf("Conv => W: %d, Pad:[%d,%d] PadT:[%d,%d] => Wc: %d, Filter:[%d,%d]\n", Width, PadInc[0], PadInc[1], PadIncT[0], PadIncT[1], Wc, Fcx, Fcy);
printf(" => H: %d, Pad:[%d,%d] PadT:[%d,%d] => Hc: %d\n", Height, PadInc[2], PadInc[3], PadIncT[2], PadIncT[3], Hc);
printf("Pool => Wc: %d, Pad:[%d,%d] => Wo: %d, Filter:[%d,%d]\n", Wc, PadInp[0], PadInp[1], Wo, Fpx, Fpy);
@@ -1046,7 +1048,7 @@ int CNN_HWC_DWConvolutionPoolAct_fp16(
UserSymbols(1, US_Float("UB", UB));
Kernel_T *Kernel = UserKernel(Name,
KernelIterSpace(2, IterTiledSpace(T0), IterParSpace(D0, InFeat, 8)),
- TILE_HOR,
+ TileOrientation,
CArgs(4,
TCArg(CNN_ArgDataTypeF(2,1,1), "In"),
TCArg(CNN_ArgDataTypeF(2,1,1), "Filter"),
@@ -1154,7 +1156,7 @@ int CNN_HWC_DWConvolutionPoolAct_fp16(
Fcx, Fcy, Scx, Scy, Dcx, Dcy, PadInc, Fpx, Fpy, Spx, Spy, Dpx, Dpy, PadInp, KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -1196,36 +1198,36 @@ int CNN_HWC_DWConvolutionPoolAct_fp16(
*********************************************************************************************************************************************************************/
-int CNN_ConvolutionPoolAct_fp16(
- char *Name,
+Kernel_T *CNN_ConvolutionPoolAct_fp16_Internal(
+ char *Name,
- CNN_GenControl_T *Ctrl,
+ CNN_GenControl_T *Ctrl,
- int InFeat,
- int OutFeat,
- int Width,
- int Height,
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
- KernelOper_T ConvOper,
- int Fcx,
- int Fcy,
- int Dcx,
- int Dcy,
- int Scx,
- int Scy,
- int ConvPad,
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
- KernelOper_T PoolOper,
- int Fpx,
- int Fpy,
- int Dpx,
- int Dpy,
- int Spx,
- int Spy,
- int PoolPad,
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
- KernelOper_T ActOper
- )
+ KernelOper_T ActOper
+ )
{
int Log=1;
@@ -1255,29 +1257,32 @@ int CNN_ConvolutionPoolAct_fp16(
if (ConvOper==KOP_NONE) {
if (PoolOper!=KOP_NONE)
- return CNN_PoolAct_fp16(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ return CNN_PoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else if (ActOper!=KOP_NONE)
- return CNN_Act_fp16(Name, Ctrl, InFeat, Width, Height, ActOper);
+ return CNN_Act_fp16_Internal(Name, Ctrl, InFeat, Width, Height, ActOper);
else GenTilingError("CNN_ConvolutionPoolAct_fp16 Kernel: %s, All requested operations are KOP_NONE", Name);
} else if (HWC) {
if (ConvOper == KOP_CONV_DW)
- return CNN_HWC_DWConvolutionPoolAct_fp16(Name, Ctrl, InFeat, OutFeat, Width, Height,
+ return CNN_HWC_DWConvolutionPoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height,
ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad,
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
else
- return CNN_MM_ConvolutionPoolAct_fp16(Name, Ctrl, InFeat, OutFeat, Width, Height,
+ return CNN_MM_ConvolutionPoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height,
ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad,
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
} else if (ConvOper==KOP_CONV && ((Fcx > 1 && Fcy == 1))) {
- AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
- int Ok = CNN_MM_ConvolutionPoolAct_fp16(Name, Ctrl, InFeat, OutFeat, Width, Height,
+ //AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+ Kernel_T *Ok = CNN_MM_ConvolutionPoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height,
ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad,
PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
- AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
- if (Ok) return Ok;
+ //AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ if (Ok!=0) return Ok;
if (Log) printf("Mapping this convolution to im2col scheme failed, reverting to standard implementation\n");
}
-
+ if (Fcx==1 && Fcy==1 && Scx==1 && Scy==1 && Dcx==1 && Dcy==1 && Height==1 && Width==1) {
+ printf("This is a pointwise on 1x1 input --> Mapping to CNN_Linear_NE16\n");
+ return CNN_LinearAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, KOP_LINEAR, ActOper);
+ }
if (PoolOper==KOP_NONE) {
Fpx=1; Dpx=1; Spx=1; Fpy=1; Dpy=1; Spy=1;
@@ -1363,7 +1368,7 @@ int CNN_ConvolutionPoolAct_fp16(
if (ActKerName==0) GenTilingError("CNN_ConvolutionPoolAct_fp16 Kernel: %s, Can't find a matching Activation basic kernel", Name);
}
if (Log) {
- printf("InFeat: %d, OutFeat: %d\n", InFeat, OutFeat);
+ printf("InFeat: %d, OutFeat: %d%s - TileOrientation: %s\n", InFeat, OutFeat, HWC?", HWC":", CHW", TileOrientation==TILE_HOR?"TILE_HOR":"TILE_VER");
printf("Conv => W: %d, Pad:[%d,%d] PadT:[%d,%d] => Wc: %d, Filter:[%d,%d]\n", Width, PadInc[0], PadInc[1], PadIncT[0], PadIncT[1], Wc, Fcx, Fcy);
printf(" => H: %d, Pad:[%d,%d] PadT:[%d,%d] => Hc: %d\n", Height, PadInc[2], PadInc[3], PadIncT[2], PadIncT[3], Hc);
printf("Pool => Wc: %d, Pad:[%d,%d] => Wo: %d, Filter:[%d,%d]\n", Wc, PadInp[0], PadInp[1], Wo, Fpx, Fpy);
@@ -1384,12 +1389,12 @@ int CNN_ConvolutionPoolAct_fp16(
AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
if ((InFeat+OutFeat)<100) {
if (Log) printf("Mapping this convolution to matrix multiplication with small first operand\n");
- int Ok = CNN_MatMulSmallM1Act_fp16(Name, 0, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL_SM1, ActOper);
+ Kernel_T *Ok = CNN_MatMulSmallM1Act_fp16_Internal(Name, 0, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL_SM1, ActOper);
if (!Ok&&Log) printf("Mapping this convolution to matrix multiplication with small first operand FAILED, trying with standard mult implementation\n");
if (Ok) return Ok;
}
if (Log) printf("Mapping this convolution to matrix multiplication\n");
- int Ok = CNN_MatMulAct_fp16(Name, 0, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL, ActOper);
+ Kernel_T *Ok = CNN_MatMulAct_fp16_Internal(Name, 0, InFeat, OutFeat, Width*Height, InFeat, Width, Height, Scx, Scy, KOP_MATMUL, ActOper, 1);
AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
if (Ok) return Ok;
if (Log) printf("Mapping this convolution to matrix multiplication FAILED, reverting to standard implementation\n");
@@ -1519,7 +1524,7 @@ int CNN_ConvolutionPoolAct_fp16(
Fcx, Fcy, Scx, Scy, Dcx, Dcy, PadInc, Fpx, Fpy, Spx, Spy, Dpx, Dpy, PadInp, KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -1692,30 +1697,30 @@ int CNN_GroupedConvolutionPoolAct_fp16(
*********************************************************************************************************************************************************************/
-int CNN_PoolAct_fp16(
- char *Name,
+Kernel_T *CNN_PoolAct_fp16_Internal(
+ char *Name,
- CNN_GenControl_T *Ctrl,
+ CNN_GenControl_T *Ctrl,
- int InFeat,
- int OutFeat,
- int Width,
- int Height,
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
- KernelOper_T PoolOper,
- int Fpx,
- int Fpy,
- int Dpx,
- int Dpy,
- int Spx,
- int Spy,
- int PoolPad,
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
- KernelOper_T ActOper
- )
+ KernelOper_T ActOper
+ )
{
- if (PoolOper==KOP_NONE && ActOper!=KOP_NONE) return CNN_Act_fp16(Name, Ctrl, InFeat, Width, Height, ActOper);
+ if (PoolOper==KOP_NONE && ActOper!=KOP_NONE) return CNN_Act_fp16_Internal(Name, Ctrl, InFeat, Width, Height, ActOper);
Tile_Orientation_T TileOrientation = TILE_HOR;
int ParFeat = 1, HWC = 0;
@@ -1875,7 +1880,7 @@ int CNN_PoolAct_fp16(
KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
@@ -1901,19 +1906,19 @@ int CNN_PoolAct_fp16(
*********************************************************************************************************************************************************************/
-int CNN_GlobalPoolAct_fp16(
- char *Name,
+static Kernel_T *CNN_GlobalPoolAct_fp16_Internal(
+ char *Name,
- CNN_GenControl_T *Ctrl,
+ CNN_GenControl_T *Ctrl,
- int InFeat,
- int OutFeat,
- int Width,
- int Height,
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
- KernelOper_T PoolOper,
- KernelOper_T ActOper
- )
+ KernelOper_T PoolOper,
+ KernelOper_T ActOper
+ )
{
Tile_Orientation_T TileOrientation = TILE_HOR;
@@ -2017,7 +2022,7 @@ int CNN_GlobalPoolAct_fp16(
KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -2040,7 +2045,7 @@ int CNN_GlobalPoolAct_fp16(
*********************************************************************************************************************************************************************/
-int CNN_Act_fp16(
+Kernel_T *CNN_Act_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2130,7 +2135,7 @@ int CNN_Act_fp16(
ActOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -2154,7 +2159,7 @@ int CNN_Act_fp16(
*********************************************************************************************************************************************************************/
-int CNN_LinearAct_fp16(
+Kernel_T *CNN_LinearAct_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2360,7 +2365,7 @@ int CNN_LinearAct_fp16(
0,
(2==1)?7:15);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -2384,7 +2389,7 @@ int CNN_LinearAct_fp16(
*********************************************************************************************************************************************************************/
-int CNN_SoftMax_fp16(
+static Kernel_T *CNN_SoftMax_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2448,10 +2453,10 @@ int CNN_SoftMax_fp16(
KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
-int CNN_SoftMax2D_fp16(
+static Kernel_T *CNN_SoftMax2D_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2517,7 +2522,7 @@ int CNN_SoftMax2D_fp16(
KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
@@ -2544,7 +2549,7 @@ int CNN_SoftMax2D_fp16(
*********************************************************************************************************************************************************************/
-int CNN_MatAddAct_fp16(
+static Kernel_T *CNN_MatAddAct_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2642,7 +2647,7 @@ int CNN_MatAddAct_fp16(
KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -2684,7 +2689,7 @@ int CNN_MatAddPaddedAct_fp16(
{
if (PadBot == 0 && PadTop == 0) return CNN_MatAddAct_fp16(Name, Ctrl, Feat, Feat, Width, Height, AddMatOper, ActOper);
- if (PadTop + PadBot > Feat) GenTilingError("int CNN_MatAddPaddedAct_SQ8 Kernel: %s, Padding exceeds channel size", Name);
+ if (PadTop + PadBot > Feat) GenTilingError("int CNN_MatAddPaddedAct_fp16 Kernel: %s, Padding exceeds channel size", Name);
int FeatBody = Feat - PadTop - PadBot;
int Ok = 1;
@@ -2799,7 +2804,7 @@ int CNN_MatAddPaddedAct_fp16(
*********************************************************************************************************************************************************************/
-int CNN_MatMulAct_fp16(
+Kernel_T *CNN_MatMulAct_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -2815,7 +2820,8 @@ int CNN_MatMulAct_fp16(
int Scy,
KernelOper_T MatMulOper,
- KernelOper_T ActOper
+ KernelOper_T ActOper,
+ int InvertInputs
)
{
@@ -2891,8 +2897,8 @@ int CNN_MatMulAct_fp16(
KernelIterSpace(2, IterTiledSpace(T1), IterTiledSpace(T0)),
TILE_HOR,
CArgs(4,
- TCArg(CNN_ArgDataTypeF(2,1,1), "In2"),
- TCArg(CNN_ArgDataTypeF(2,1,1), "In1"),
+ TCArg(CNN_ArgDataTypeF(2,1,1), InvertInputs?"In2":"In1"),
+ TCArg(CNN_ArgDataTypeF(2,1,1), InvertInputs?"In1":"In2"),
!NoBias?TCArg(CNN_ArgDataTypeF(2,1,1), "Bias"):AT_NO_C_ARG,
TCArg(CNN_ArgDataTypeF(2,1,1), "Out")
),
@@ -2959,7 +2965,7 @@ int CNN_MatMulAct_fp16(
KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
}
/*********************************************************************************************************************************************************************
@@ -2995,7 +3001,7 @@ int CNN_MatMulAct_fp16(
*********************************************************************************************************************************************************************/
-int CNN_MatMulSmallM1Act_fp16(
+Kernel_T *CNN_MatMulSmallM1Act_fp16_Internal(
char *Name,
CNN_GenControl_T *Ctrl,
@@ -3146,7 +3152,287 @@ int CNN_MatMulSmallM1Act_fp16(
KernelOper,
0, 0);
}
- return (Kernel!=0);
+ return Kernel;
+
+}
+
+
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+/* ============================================================================================================================================================== */
+
+int CNN_MM_ConvolutionPoolAct_fp16(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ )
+{
+ Kernel_T *Ker = 0, *Sol1 = 0, *Sol2 = 0;
+ float K = 0.9;
+ Tile_Orientation_T TileOrientation = TILE_HOR;
+ if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) {
+ printf("TileOrientation set by user\n");
+ Ker = CNN_MM_ConvolutionPoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker!=0) return 1;
+ else GenTilingError("CNN_MM_ConvolutionPoolAct_fp16: %s, Failed to gen with set tiling orientation, try to let the Autotiler set it for you", Name);
+ }
+ }
+
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+
+ CNN_GenControl_T InternalCtrl;
+ if (!Ctrl) CNN_InitGenCtrl(&InternalCtrl);
+ else InternalCtrl = *Ctrl;
+
+ printf("\n\n=============================== Trying Tile Orientation: TILE_HOR ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(0));
+ Ker = CNN_MM_ConvolutionPoolAct_fp16_Internal(Name, &InternalCtrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol1 = CopyAndPopUserKernel(Ker);
+
+ printf("\n=============================== Trying Tile Orientation: TILE_VER ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(1));
+ Ker = CNN_MM_ConvolutionPoolAct_fp16_Internal(Name, &InternalCtrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol2 = CopyAndPopUserKernel(Ker);
+
+ if (Sol1 && Sol2) {
+ int TakeSol1 = ((K*Sol1->Cost->TileOverhead) < Sol2->Cost->TileOverhead); // K close to 1.0if (TakeSol1) {
+ printf(">>>>>>>>>>>>>>>>>> %s is better: %.3f vs %.3f \n\n\n", TakeSol1?"TILE_HOR":"TILE_VER", Sol1->Cost->TileOverhead, Sol2->Cost->TileOverhead);
+ if (TakeSol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ }
+ } else if (Sol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else if (Sol2) {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ } else {
+ GenTilingError("Failed to Generate code for Kernel: %s", Name);
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ return 1;
+}
+
+int CNN_HWC_DWConvolutionPoolAct_fp16(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ )
+{
+ Kernel_T *Ker = 0, *Sol1 = 0, *Sol2 = 0;
+ float K = 0.9;
+ Tile_Orientation_T TileOrientation = TILE_HOR;
+ if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) {
+ printf("TileOrientation set by user\n");
+ Ker = CNN_HWC_DWConvolutionPoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker!=0) return 1;
+ else GenTilingError("CNN_MM_ConvolutionPoolAct_fp16: %s, Failed to gen with set tiling orientation, try to let the Autotiler set it for you", Name);
+ }
+ }
+
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+
+ CNN_GenControl_T InternalCtrl;
+ if (!Ctrl) CNN_InitGenCtrl(&InternalCtrl);
+ else InternalCtrl = *Ctrl;
+
+ printf("\n\n=============================== Trying Tile Orientation: TILE_HOR ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(0));
+ Ker = CNN_HWC_DWConvolutionPoolAct_fp16_Internal(Name, &InternalCtrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol1 = CopyAndPopUserKernel(Ker);
+
+ printf("\n=============================== Trying Tile Orientation: TILE_VER ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(1));
+ Ker = CNN_HWC_DWConvolutionPoolAct_fp16_Internal(Name, &InternalCtrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol2 = CopyAndPopUserKernel(Ker);
+
+ if (Sol1 && Sol2) {
+ int TakeSol1 = ((K*Sol1->Cost->TileOverhead) < Sol2->Cost->TileOverhead); // K close to 1.0if (TakeSol1) {
+ printf(">>>>>>>>>>>>>>>>>> %s is better: %.3f vs %.3f \n\n\n", TakeSol1?"TILE_HOR":"TILE_VER", Sol1->Cost->TileOverhead, Sol2->Cost->TileOverhead);
+ if (TakeSol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ }
+ } else if (Sol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else if (Sol2) {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ } else {
+ GenTilingError("Failed to Generate code for Kernel: %s", Name);
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ return 1;
+}
+
+int CNN_ConvolutionPoolAct_fp16(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ )
+{
+ Kernel_T *Ker = 0, *Sol1 = 0, *Sol2 = 0;
+ float K = 0.9;
+ Tile_Orientation_T TileOrientation = TILE_HOR;
+ if (Ctrl) {
+ if (Ctrl->TileOrientation != -1) {
+ printf("TileOrientation set by user\n");
+ Ker = CNN_ConvolutionPoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker!=0) return 1;
+ else GenTilingError("CNN_ConvolutionPoolAct_fp16: %s, Failed to gen with set tiling orientation, try to let the Autotiler set it for you", Name);
+ }
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_OFF);
+
+ CNN_GenControl_T InternalCtrl;
+ if (!Ctrl) CNN_InitGenCtrl(&InternalCtrl);
+ else InternalCtrl = *Ctrl;
+
+ printf("\n\n=============================== Trying Tile Orientation: TILE_HOR ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(0));
+ Ker = CNN_ConvolutionPoolAct_fp16_Internal(Name, &InternalCtrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol1 = CopyAndPopUserKernel(Ker);
+
+ printf("\n=============================== Trying Tile Orientation: TILE_VER ===============================\n\n");
+ CNN_SetGenCtrl(&InternalCtrl, "TILEORIENTATION", AT_OPT_VAL(1));
+ Ker = CNN_ConvolutionPoolAct_fp16_Internal(Name, &InternalCtrl, InFeat, OutFeat, Width, Height, ConvOper, Fcx, Fcy, Dcx, Dcy, Scx, Scy, ConvPad, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper);
+ if (Ker) Sol2 = CopyAndPopUserKernel(Ker);
+
+ if (Sol1 && Sol2) {
+ int TakeSol1 = ((K*Sol1->Cost->TileOverhead) < Sol2->Cost->TileOverhead); // K close to 1.0if (TakeSol1) {
+ printf(">>>>>>>>>>>>>>>>>> %s is better: %.3f vs %.3f \n\n\n", TakeSol1?"TILE_HOR":"TILE_VER", Sol1->Cost->TileOverhead, Sol2->Cost->TileOverhead);
+ if (TakeSol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ }
+ } else if (Sol1) {
+ PushBackUserKernel(Sol1); ReleaseUserKerne(Sol2);
+ } else if (Sol2) {
+ PushBackUserKernel(Sol2); ReleaseUserKerne(Sol1);
+ } else {
+ GenTilingError("Failed to Generate code for Kernel: %s", Name);
+ }
+ AT_SetKernelCtrl(AT_KERNEL_NOSOLUTION_ERROR, AT_OPT_ON);
+ return 1;
+}
+
+int CNN_PoolAct_fp16(char *Name, CNN_GenControl_T *Ctrl, int InFeat, int OutFeat, int Width, int Height, KernelOper_T PoolOper, int Fpx, int Fpy, int Dpx, int Dpy, int Spx, int Spy, int PoolPad, KernelOper_T ActOper)
+{
+ return (CNN_PoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, Fpx, Fpy, Dpx, Dpy, Spx, Spy, PoolPad, ActOper)!=0);
+}
+
+int CNN_Act_fp16(char *Name, CNN_GenControl_T *Ctrl, int Feat, int Width, int Height,KernelOper_T ActOper) {
+ return (CNN_Act_fp16_Internal(Name, Ctrl, Feat, Width, Height, ActOper)!=0);
+}
+
+int CNN_GlobalPoolAct_fp16(char *Name, CNN_GenControl_T *Ctrl, int InFeat, int OutFeat, int Width, int Height, KernelOper_T PoolOper, KernelOper_T ActOper) {
+ return (CNN_GlobalPoolAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, PoolOper, ActOper)!=0);
+}
+
+int CNN_LinearAct_fp16(char *Name, CNN_GenControl_T *Ctrl, int InDim, int OutDim, KernelOper_T LinearOper, KernelOper_T ActOper)
+{
+ return (CNN_LinearAct_fp16_Internal(Name, Ctrl, InDim, OutDim, LinearOper, ActOper)!=0);
+}
+
+int CNN_SoftMax_fp16(char *Name, CNN_GenControl_T *Ctrl, int Dim, KernelOper_T SoftMaxOper) {
+ return (CNN_SoftMax_fp16_Internal(Name, Ctrl, Dim, SoftMaxOper)!=0);
+}
+
+int CNN_SoftMax2D_fp16(char *Name, CNN_GenControl_T *Ctrl, int Dim, int N, KernelOper_T SoftMaxOper) {
+ return (CNN_SoftMax2D_fp16_Internal(Name, Ctrl, Dim, N, SoftMaxOper)!=0);
+}
+
+int CNN_MatAddAct_fp16(char *Name, CNN_GenControl_T *Ctrl, int InFeat, int OutFeat, int Width, int Height, KernelOper_T AddMatOper, KernelOper_T ActOper) {
+ return (CNN_MatAddAct_fp16_Internal(Name, Ctrl, InFeat, OutFeat, Width, Height, AddMatOper, ActOper)!=0);
+}
+
+int CNN_MatMulAct_fp16(char *Name, CNN_GenControl_T *Ctrl, int ColM1, int LineM1, int ColM2, int LineM2, int Width, int Height, int Scx, int Scy, KernelOper_T MatMulOper, KernelOper_T ActOper) {
+ return (CNN_MatMulAct_fp16_Internal(Name, Ctrl, ColM1, LineM1, ColM2, LineM2, Width, Height, Scx, Scy, MatMulOper, ActOper, 1)!=0);
+}
+int CNN_MatMulSmallM1Act_fp16(char *Name, CNN_GenControl_T *Ctrl, int ColM1, int LineM1, int ColM2, int LineM2, int Width, int Height, int Scx, int Scy, KernelOper_T MatMulOper, KernelOper_T ActOper) {
+ return (CNN_MatMulSmallM1Act_fp16_Internal(Name, Ctrl, ColM1, LineM1, ColM2, LineM2, Width, Height, Scx, Scy, MatMulOper, ActOper)!=0);
}
diff --git a/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.h b/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.h
index 5e942f2b6..7096cbc26 100644
--- a/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.h
+++ b/tools/autotiler_v3/CNN_Generators_fp16/CNN_Generators_fp16.h
@@ -67,6 +67,99 @@ extern void LoadCNNLibrary_fp16();
*/
+Kernel_T *CNN_MM_ConvolutionPoolAct_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ );
+
+Kernel_T *CNN_HWC_DWConvolutionPoolAct_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ );
+
+Kernel_T *CNN_ConvolutionPoolAct_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T ConvOper,
+ int Fcx,
+ int Fcy,
+ int Dcx,
+ int Dcy,
+ int Scx,
+ int Scy,
+ int ConvPad,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ );
+
extern int CNN_ConvolutionPoolAct_fp16(
char *Name,
@@ -84,7 +177,7 @@ extern int CNN_ConvolutionPoolAct_fp16(
int Dcy,
int Scx,
int Scy,
- int ConvPad,
+ int ConvPad,
KernelOper_T PoolOper,
int Fpx,
@@ -93,7 +186,7 @@ extern int CNN_ConvolutionPoolAct_fp16(
int Dpy,
int Spx,
int Spy,
- int PoolPad,
+ int PoolPad,
KernelOper_T ActOper
);
@@ -200,6 +293,28 @@ extern int CNN_GroupedConvolutionPoolAct_fp16(
*/
+Kernel_T *CNN_PoolAct_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InFeat,
+ int OutFeat,
+ int Width,
+ int Height,
+
+ KernelOper_T PoolOper,
+ int Fpx,
+ int Fpy,
+ int Dpx,
+ int Dpy,
+ int Spx,
+ int Spy,
+ int PoolPad,
+
+ KernelOper_T ActOper
+ );
+
extern int CNN_PoolAct_fp16(
char *Name,
@@ -276,6 +391,18 @@ extern int CNN_GlobalPoolAct_fp16(
*********************************************************************************************************************************************************************/
+Kernel_T *CNN_Act_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int Feat,
+ int Width,
+ int Height,
+
+ KernelOper_T ActOper
+ );
+
extern int CNN_Act_fp16(
char *Name,
@@ -309,6 +436,18 @@ extern int CNN_Act_fp16(
*/
+Kernel_T *CNN_LinearAct_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int InDim,
+ int OutDim,
+
+ KernelOper_T LinearOper,
+ KernelOper_T ActOper
+ );
+
extern int CNN_LinearAct_fp16(
char *Name,
@@ -456,7 +595,27 @@ extern int CNN_MatAddPaddedAct_fp16(
\param Signature: Name(In2, In1, Bias, Out)
*/
-
+
+Kernel_T *CNN_MatMulAct_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int ColM1,
+ int LineM1,
+ int ColM2,
+ int LineM2,
+
+ int Width,
+ int Height,
+ int Scx,
+ int Scy,
+
+ KernelOper_T MatMulOper,
+ KernelOper_T ActOper,
+ int InvertInputs
+ );
+
extern int CNN_MatMulAct_fp16(
char *Name,
@@ -509,6 +668,25 @@ extern int CNN_MatMulAct_fp16(
*********************************************************************************************************************************************************************/
+Kernel_T *CNN_MatMulSmallM1Act_fp16_Internal(
+ char *Name,
+
+ CNN_GenControl_T *Ctrl,
+
+ int ColM1,
+ int LineM1,
+ int ColM2,
+ int LineM2,
+
+ int Width,
+ int Height,
+ int Scx,
+ int Scy,
+
+ KernelOper_T MatMulOper,
+ KernelOper_T ActOper
+);
+
extern int CNN_MatMulSmallM1Act_fp16(
char *Name,
diff --git a/tools/autotiler_v3/CNN_Libraries/CNN_Copy.c b/tools/autotiler_v3/CNN_Libraries/CNN_Copy.c
index 3ccdeccf2..8b7b62ecd 100644
--- a/tools/autotiler_v3/CNN_Libraries/CNN_Copy.c
+++ b/tools/autotiler_v3/CNN_Libraries/CNN_Copy.c
@@ -39,10 +39,15 @@ float to_float(unsigned short v) {
#endif
#else
#ifndef TO_FLOAT
-#define TO_FLOAT(x) *((F16 *)&x)
+#define TO_FLOAT(x) *((F16 *) (&x))
#endif
#endif
+typedef union {
+ float F;
+ int I;
+} f32_cast_t;
+
static inline unsigned int __attribute__((always_inline)) ChunkSize(unsigned int X)
{
@@ -1706,8 +1711,13 @@ void CNN_FpFloat16(CNN_FpFloat16_T * Arg)
unsigned int Iter = Max(0, Last-First);
signed short *pIn = (signed short *) (Arg->In + First);
F16V *pOut = (F16V *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
@@ -1731,8 +1741,13 @@ void CNN_UFpFloat16(CNN_UFpFloat16_T * Arg)
unsigned int Iter = Max(0, Last-First);
unsigned short *pIn = (unsigned short *) (Arg->In + First);
F16V *pOut = (F16V *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
@@ -1756,8 +1771,13 @@ void CNN_FpsFloat16(CNN_FpsFloat16_T * Arg)
unsigned int Iter = Max(0, Last-First);
signed char *pIn = (signed char *) (Arg->In + First);
F16V *pOut = (F16V *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
@@ -1781,8 +1801,13 @@ void CNN_UFpsFloat16(CNN_UFpsFloat16_T * Arg)
unsigned int Iter = Max(0, Last-First);
unsigned char *pIn = (signed char *) (Arg->In + First);
F16V *pOut = (F16V *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
@@ -1806,8 +1831,13 @@ void CNN_Float16Fp(CNN_Float16Fp_T * Arg)
unsigned int Iter = Max(0, Last-First);
F16V *pIn = (F16V *) (Arg->In + First);
v2s *pOut = (v2s *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
F16V max_v = gap_pack2f16(32767.0F, 32767.0F);
@@ -1836,8 +1866,13 @@ void CNN_Float16UFp(CNN_Float16UFp_T * Arg)
unsigned int Iter = Max(0, Last-First);
F16V *pIn = (F16V *) (Arg->In + First);
v2u *pOut = (v2u *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
F16V max_v = gap_pack2f16(65535.0F, 65535.0F);
@@ -1866,8 +1901,13 @@ void CNN_Float16Fps(CNN_Float16Fps_T * Arg)
unsigned int Iter = Max(0, Last-First);
F16V *pIn = (F16V *) (Arg->In + First);
v4s *pOut = (v4s *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
F16V max_v = gap_pack2f16(127.0F, 127.0F);
@@ -1903,8 +1943,13 @@ void CNN_Float16UFps(CNN_Float16UFps_T * Arg)
unsigned int Iter = Max(0, Last-First);
F16V *pIn = (F16V *) (Arg->In + First);
v4u *pOut = (v4u *) (Arg->Out + First);
- F16 zero_diff = TO_FLOAT(Arg->Infos[AT_INF_QUANT_ZERO_DIFF]);
- F16 scale = TO_FLOAT(Arg->Infos[AT_INF_QUANT_SCALE]);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ F16 zero_diff = (F16) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ F16 scale = (F16) caster.F;
+
F16V scale_v = gap_pack2f16(scale, scale);
F16V zero_v = gap_pack2f16(zero_diff, zero_diff);
F16V max_v = gap_pack2f16(255.0F, 255.0F);
@@ -1932,4 +1977,171 @@ void CNN_Float16UFps(CNN_Float16UFps_T * Arg)
gap_waitbarrier(0);
}
+void CNN_FpFloat32(CNN_FpFloat32_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ signed short *pIn = (signed short *) (Arg->In);
+ float *pOut = (float *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (pIn[i] + zero_diff) * scale;
+ }
+ gap_waitbarrier(0);
+}
+
+void CNN_UFpFloat32(CNN_UFpFloat32_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ unsigned short *pIn = (unsigned short *) (Arg->In);
+ float *pOut = (float *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (pIn[i] + zero_diff) * scale;
+ }
+ gap_waitbarrier(0);
+}
+
+void CNN_FpsFloat32(CNN_FpsFloat32_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ signed char *pIn = (signed char *) (Arg->In);
+ float *pOut = (float *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (pIn[i] + zero_diff) * scale;
+ }
+ gap_waitbarrier(0);
+}
+
+void CNN_UFpsFloat32(CNN_UFpsFloat32_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ unsigned char *pIn = (signed char *) (Arg->In);
+ float *pOut = (float *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (pIn[i] + zero_diff) * scale;
+ }
+ gap_waitbarrier(0);
+}
+
+void CNN_Float32Fp(CNN_Float32Fp_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ float *pIn = (float *) (Arg->In);
+ signed short int *pOut = (short int *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (signed short int) Clipf32((pIn[i] + zero_diff) * scale, 32767.0F, -32768.0F);
+ }
+ gap_waitbarrier(0);
+}
+
+void CNN_Float32UFp(CNN_Float32UFp_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ float *pIn = (float *) (Arg->In);
+ unsigned short int *pOut = (unsigned short int *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (unsigned short int) Clipf32((pIn[i] + zero_diff) * scale, 65535.0F, 0.0F);
+ }
+ gap_waitbarrier(0);
+}
+
+void CNN_Float32Fps(CNN_Float32Fps_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ float *pIn = (float *) (Arg->In);
+ signed char *pOut = (signed char *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (signed char) Clipf32((pIn[i] + zero_diff) * scale, 127.0F, -128.0F);
+ }
+ gap_waitbarrier(0);
+}
+
+void CNN_Float32UFps(CNN_Float32UFps_T * Arg)
+{
+ unsigned int Size = Arg->W * Arg->H;
+ unsigned int CoreId = gap_coreid();
+ unsigned int Chunk = ChunkSize(Size), First = Chunk*CoreId, Last = Min(First+Chunk, Size);
+ unsigned int Iter = Max(0, Last-First);
+ float *pIn = (float *) (Arg->In);
+ unsigned char *pOut = (unsigned char *) (Arg->Out);
+
+ f32_cast_t caster;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_ZERO_DIFF));
+ float zero_diff = (float) caster.F;
+ caster.I = *((int *) (Arg->Infos + AT_INF_QUANT_SCALE));
+ float scale = (float) caster.F;
+
+ for (int i=First; i<Last; i++) {
+ pOut[i] = (unsigned char) Clipf32((pIn[i] + zero_diff) * scale, 255.0F, 0.0F);
+ }
+}
+
#pragma GCC diagnostic pop
diff --git a/tools/autotiler_v3/CNN_Libraries/CNN_Copy.h b/tools/autotiler_v3/CNN_Libraries/CNN_Copy.h
index 7ce10a20a..a53a8e0ce 100644
--- a/tools/autotiler_v3/CNN_Libraries/CNN_Copy.h
+++ b/tools/autotiler_v3/CNN_Libraries/CNN_Copy.h
@@ -310,6 +310,78 @@ typedef struct {
signed char *__restrict__ Infos;
} CNN_UFpsFloat16_T;
+// float32 -> signed short
+typedef struct {
+ float *__restrict__ In;
+ short int *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_Float32Fp_T;
+
+// float32 -> unsigned short
+typedef struct {
+ float *__restrict__ In;
+ unsigned short int *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_Float32UFp_T;
+
+// float32 -> signed char
+typedef struct {
+ float *__restrict__ In;
+ signed char *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_Float32Fps_T;
+
+// float32 -> unsigned char
+typedef struct {
+ float *__restrict__ In;
+ unsigned char *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_Float32UFps_T;
+
+// signed short -> float32
+typedef struct {
+ short int *__restrict__ In;
+ float *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_FpFloat32_T;
+
+// unsigned short -> float32
+typedef struct {
+ unsigned short int *__restrict__ In;
+ float *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_UFpFloat32_T;
+
+// signed char -> float32
+typedef struct {
+ signed char *__restrict__ In;
+ float *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_FpsFloat32_T;
+
+// unsigned char -> float32
+typedef struct {
+ unsigned char *__restrict__ In;
+ float *__restrict__ Out;
+ unsigned short int W;
+ unsigned short int H;
+ signed char *__restrict__ Infos;
+} CNN_UFpsFloat32_T;
+
#define AT_INF_QUANT_ZERO_DIFF 0
#define AT_INF_QUANT_SCALE 4
#define AT_INF_QUANT_NORM 6
@@ -405,5 +477,15 @@ extern void CNN_UFpFloat16(CNN_UFpFloat16_T * Arg);
extern void CNN_FpsFloat16(CNN_FpsFloat16_T * Arg);
extern void CNN_UFpsFloat16(CNN_UFpsFloat16_T * Arg);
+extern void CNN_FpFloat32(CNN_FpFloat32_T * Arg);
+extern void CNN_UFpFloat32(CNN_UFpFloat32_T * Arg);
+extern void CNN_FpsFloat32(CNN_FpsFloat32_T * Arg);
+extern void CNN_UFpsFloat32(CNN_UFpsFloat32_T * Arg);
+
+extern void CNN_Float32Fp(CNN_Float32Fp_T * Arg);
+extern void CNN_Float32UFp(CNN_Float32UFp_T * Arg);
+extern void CNN_Float32Fps(CNN_Float32Fps_T * Arg);
+extern void CNN_Float32UFps(CNN_Float32UFps_T * Arg);
+
#endif
diff --git a/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.c b/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.c
index e8d39e821..1c302417a 100644
--- a/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.c
+++ b/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.c
@@ -244,7 +244,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
int PadL = Arg->Pad[0], PadR = Arg->Pad[1], PadT = Arg->Pad[2], PadB = Arg->Pad[3];
int Nb_KI = Tile_InFeat/16 + (Tile_InFeat%16?1:0);
- int Nb_LoadedKI = Arg->TotalInFeatures/16 + (Arg->TotalInFeatures%16?1:0);
+ int Nb_LoadedKI = Arg->TotalInFeatures/(Mode16?8:16) + (Arg->TotalInFeatures%(Mode16?8:16)?1:0);
int Rem_KI = Tile_InFeat%16?Tile_InFeat%16:16;
int Nb_KO = Tile_OutFeat/32 + (Tile_OutFeat%32?1:0);
int Rem_KO = Tile_OutFeat%32?Tile_OutFeat%32:32;
@@ -289,7 +289,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer ((void *) Max((unsigned int) In, (unsigned int) InPointer + Tile_InFeat*(subfilter_i*Dx + w*Sx-PadL + Tile_InW*(Ho_F*Sy-PadT + subfilter_j*Dy))));
SetNE16_OutPointer (Out + Tile_OutFeat*(w + Ho_F*Tile_OutW));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (1, Rem_H_Subtiles, Rem_KI, Rem_KO, 1, Rem_H_Subtiles);
SetNE16_Dim (Nb_KI, Nb_KO, 1, H_SubTiles);
SetNE16_GenConfig (Gen_Cfg);
@@ -332,7 +332,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer (InPointer + Tile_InFeat*(w*Sx-PadL + subfilter_i*Dx + Tile_InW*(Ho_F*Sy-PadT + subfilter_j*Dy)));
SetNE16_OutPointer (Out + Tile_OutFeat*(w + Ho_F*Tile_OutW));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (1, Rem_H_Subtiles, Rem_KI, Rem_KO, 1, Rem_H_Subtiles);
SetNE16_Dim (Nb_KI, Nb_KO, 1, H_SubTiles);
SetNE16_GenConfig (Gen_Cfg);
@@ -375,7 +375,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer ((void *) Max((unsigned int) InPointer, (unsigned int) InPointer + Tile_InFeat*(Wo_F*Sx-PadL + subfilter_i*Dx + (subfilter_j*Dy+h*Sy-PadT)*Tile_InW)));
SetNE16_OutPointer (Out + Tile_OutFeat*(Wo_F + h*Tile_OutW));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (Rem_W_Subtiles, 1, Rem_KI, Rem_KO, Rem_W_Subtiles, 1);
SetNE16_Dim (Nb_KI, Nb_KO, W_SubTiles, 1);
SetNE16_GenConfig (Gen_Cfg);
@@ -418,7 +418,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer (InPointer + Tile_InFeat*(Wo_F*Sx-PadL + subfilter_i*Dx + (subfilter_j*Dy+h*Sy-PadT)*Tile_InW));
SetNE16_OutPointer (Out + Tile_OutFeat*(Wo_F + h*Tile_OutW));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (Rem_W_Subtiles, 1, Rem_KI, Rem_KO, Rem_W_Subtiles, 1);
SetNE16_Dim (Nb_KI, Nb_KO, W_SubTiles, 1);
SetNE16_GenConfig (Gen_Cfg);
@@ -463,7 +463,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer ((void *) Max((unsigned int) In, (unsigned int) InPointer + Tile_InFeat*(subfilter_i*Dx + w*Sx - PadL + (subfilter_j*Dy+h*Sy-PadT)*Tile_InW)));
SetNE16_OutPointer (Out + Tile_OutFeat*(w + h*Tile_OutW));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (1, 1, Rem_KI, Rem_KO, 1, 1);
SetNE16_Dim (Nb_KI, Nb_KO, 1, 1);
SetNE16_GenConfig (Gen_Cfg);
@@ -509,7 +509,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer ((void *) Max((unsigned int) In, (unsigned int) InPointer + Tile_InFeat*(subfilter_i*Dx + (w*Sx-PadL) + (subfilter_j*Dy+h*Sy-PadT)*Tile_InW)));
SetNE16_OutPointer (Out + Tile_OutFeat*(w + h*Tile_OutW));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (1, 1, Rem_KI, Rem_KO, 1, 1);
SetNE16_Dim (Nb_KI, Nb_KO, 1, 1);
SetNE16_GenConfig (Gen_Cfg);
@@ -555,7 +555,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer ((void *) Max((unsigned int) In, (unsigned int) InPointer + Tile_InFeat*(subfilter_i*Dx + w*Sx - PadL + (subfilter_j*Dy+h*Sy-PadT)*Tile_InW)));
SetNE16_OutPointer (Out + Tile_OutFeat*(h*Tile_OutW + w));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (1, 1, Rem_KI, Rem_KO, 1, 1);
SetNE16_Dim (Nb_KI, Nb_KO, 1, 1);
SetNE16_GenConfig (Gen_Cfg);
@@ -601,7 +601,7 @@ void NE16_ComputeBorders(KerConv_NE16_T *Arg, int Wo_F, int Wo_L, int Wo, int Ho
SetNE16_InPointer (InPointer + Tile_InFeat*(subfilter_i*Dx + w*Sx-PadL + (subfilter_j*Dy+h*Sy-PadT)*Tile_InW));
SetNE16_OutPointer (Out + Tile_OutFeat*(h*Tile_OutW + w));
// TODO - checkme I think here you need the total number of loaded chin
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*Fx*subfilter_j);
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*Fx*subfilter_j);
SetNE16_Reminders (1, 1, Rem_KI, Rem_KO, 1, 1);
SetNE16_Dim (Nb_KI, Nb_KO, 1, 1);
SetNE16_GenConfig (Gen_Cfg);
@@ -1034,14 +1034,6 @@ void KerConv3x3Stride2_NE16(KerConv_NE16_T *Arg)
SetNE16_WOffset (Arg->W_Offset);
SetNE16_GenConfig (Gen_Cfg);
- subtile_j_major ++;
- if(subtile_j_major==W_subtiles) {
- subtile_j_major = 0;
- subtile_i_major ++;
- IsLastSubtileH = subtile_i_major>=(H_subtiles-1);
- }
- IsLastSubtileW = subtile_j_major==(W_subtiles-1);
-
// already commit and trigger NE16 computation, while programming the next one
NE16_WRITE_CMD(NE16_COMMIT_AND_TRIGGER, NE16_TRIGGER_CMD);
if (IsLastSubtileH && IsLastSubtileW) {
@@ -1052,6 +1044,14 @@ void KerConv3x3Stride2_NE16(KerConv_NE16_T *Arg)
return;
}
+ subtile_j_major ++;
+ if(subtile_j_major==W_subtiles) {
+ subtile_j_major = 0;
+ subtile_i_major ++;
+ IsLastSubtileH = subtile_i_major>=(H_subtiles-1);
+ }
+ IsLastSubtileW = subtile_j_major==(W_subtiles-1);
+
// acquire second job
NE16_BARRIER_ACQUIRE(job_id);
@@ -1104,14 +1104,6 @@ void KerConv3x3Stride2_NE16(KerConv_NE16_T *Arg)
SetNE16_WOffset (Arg->W_Offset);
SetNE16_GenConfig (Gen_Cfg);
- subtile_j_major ++;
- if(subtile_j_major==W_subtiles) {
- subtile_j_major = 0;
- subtile_i_major ++;
- IsLastSubtileH = subtile_i_major>=(H_subtiles-1);
- }
- IsLastSubtileW = subtile_j_major==(W_subtiles-1);
-
// already commit and trigger NE16 computation, while programming the next one
NE16_WRITE_CMD(NE16_COMMIT_AND_TRIGGER, NE16_TRIGGER_CMD);
if (IsLastSubtileH && IsLastSubtileW) {
@@ -1122,6 +1114,14 @@ void KerConv3x3Stride2_NE16(KerConv_NE16_T *Arg)
return;
}
+ subtile_j_major ++;
+ if(subtile_j_major==W_subtiles) {
+ subtile_j_major = 0;
+ subtile_i_major ++;
+ IsLastSubtileH = subtile_i_major>=(H_subtiles-1);
+ }
+ IsLastSubtileW = subtile_j_major==(W_subtiles-1);
+
// main iteration strided conv iteration loop (does not need ne16_c0_config any more)
do {
// acquire job
@@ -1283,14 +1283,6 @@ void KerConvDW3x3Stride2_NE16(KerConv_NE16_T *Arg)
SetNE16_WOffset (Arg->W_Offset);
SetNE16_GenConfig (Gen_Cfg);
- subtile_j_major ++;
- if(subtile_j_major==W_subtiles) {
- subtile_j_major = 0;
- subtile_i_major ++;
- IsLastSubtileH = subtile_i_major==(H_subtiles-1);
- }
- IsLastSubtileW = subtile_j_major==(W_subtiles-1);
-
// already commit and trigger NE16 computation, while programming the next one
NE16_WRITE_CMD(NE16_COMMIT_AND_TRIGGER, NE16_TRIGGER_CMD);
if (IsLastSubtileH && IsLastSubtileW) {
@@ -1301,6 +1293,14 @@ void KerConvDW3x3Stride2_NE16(KerConv_NE16_T *Arg)
return;
}
+ subtile_j_major ++;
+ if(subtile_j_major==W_subtiles) {
+ subtile_j_major = 0;
+ subtile_i_major ++;
+ IsLastSubtileH = subtile_i_major>=(H_subtiles-1);
+ }
+ IsLastSubtileW = subtile_j_major==(W_subtiles-1);
+
// acquire second job
NE16_BARRIER_ACQUIRE(job_id);
@@ -1353,14 +1353,6 @@ void KerConvDW3x3Stride2_NE16(KerConv_NE16_T *Arg)
SetNE16_WOffset (Arg->W_Offset);
SetNE16_GenConfig (Gen_Cfg);
- subtile_j_major ++;
- if(subtile_j_major==W_subtiles) {
- subtile_j_major = 0;
- subtile_i_major ++;
- IsLastSubtileH = subtile_i_major==(H_subtiles-1);
- }
- IsLastSubtileW = subtile_j_major==(W_subtiles-1);
-
// already commit and trigger NE16 computation, while programming the next one
NE16_WRITE_CMD(NE16_COMMIT_AND_TRIGGER, NE16_TRIGGER_CMD);
if (IsLastSubtileH && IsLastSubtileW) {
@@ -1371,13 +1363,21 @@ void KerConvDW3x3Stride2_NE16(KerConv_NE16_T *Arg)
return;
}
+ subtile_j_major ++;
+ if(subtile_j_major==W_subtiles) {
+ subtile_j_major = 0;
+ subtile_i_major ++;
+ IsLastSubtileH = subtile_i_major>=(H_subtiles-1);
+ }
+ IsLastSubtileW = subtile_j_major==(W_subtiles-1);
+
// main iteration strided conv iteration loop (does not need ne16_c0_config any more)
do {
// acquire second job
NE16_BARRIER_ACQUIRE(job_id);
- IsLastSubtileH = subtile_i_major==(H_subtiles-1);
- IsLastSubtileW = subtile_j_major==(W_subtiles-1);
+ IsLastSubtileH = subtile_i_major>=(H_subtiles-1);
+ IsLastSubtileW = subtile_j_major>=(W_subtiles-1);
// update input / output pointers
if(IsLastSubtileH && IsLastSubtileW) {
Rem_HO = Last_Rem_HO; Rem_WO = Last_Rem_WO;
@@ -1508,7 +1508,7 @@ void KerConv1D_StrideS_NE16(KerConv_NE16_T *Arg)
unsigned char *pIn = In + Tile_InFeat*(w*Sx - Arg->Pad[0] + subfilter_i*Dx);
unsigned char *pOut = Out + OutBytes*Tile_OutFeat*w;
- unsigned char *pFilt = Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*subfilter_i;
+ unsigned char *pFilt = Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*subfilter_i;
if (LastSubtileW) {
Rem_HI = LastSubtileHeight;
w += LastSubtileSize;
@@ -1628,7 +1628,7 @@ void KerConvNxMDxDy_StrideSxSy_NE16(KerConv_NE16_T *Arg)
// update input / output pointers
SetNE16_InPointer (In + Tile_InFeat*(Wo_F*Sx - Arg->Pad[0] + subfilter_i*Dx + Tile_InW*(Ho_F*Sy - Arg->Pad[2] + subfilter_j*Dy)));
SetNE16_OutPointer (Out + OutBytes*Tile_OutFeat*(Wo_F + Tile_OutW*Ho_F));
- SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?8:16)*Nb_LoadedKI*(subfilter_i + Fx*subfilter_j));
+ SetNE16_WeightsPointer(Filter + Tile_OutFeat*(Mode16?1:2)*Arg->Qw*Nb_LoadedKI*(subfilter_i + Fx*subfilter_j));
SetNE16_Reminders (Rem_WO, Rem_HO, Rem_KI, Rem_KO, Rem_WO, Rem_HO);
SetNE16_Dim (Nb_KI, Nb_KO, Nb_WO, Nb_HO);
SetNE16_GenConfig (Gen_Cfg);
@@ -2176,7 +2176,7 @@ void KerLinear_16a_NE16(KerLinear_NE16_T *Arg)
NE16_BARRIER_ACQUIRE(job_id);
// load configuration for the layer
SetNE16_OutPointer (Out);
- SetNE16_WeightsPointer(Filter+subtile_ki*256*Tile_OutFeat);
+ SetNE16_WeightsPointer(Filter+subtile_ki*256*Tile_OutFeat*Arg->Qw/8);
SetNE16_BiasPointer (Bias);
SetNE16_ScalePointer (Scale);
SetNE16_ScaleNPointer (ScaleN);
@@ -2433,7 +2433,7 @@ void KerMatMul_8a_NE16(KerMatMul_NE16_T *Arg)
}
void Ker_MM_Conv2D_NE16(
- Ker_MM_Conv_NE16_T *Arg
+ KerConv_MM_NE16_T *Arg
)
{
@@ -2443,14 +2443,13 @@ void Ker_MM_Conv2D_NE16(
int Fx = Arg->Fx, Sx = Arg->Sx;
int Fy = Arg->Fy, Sy = Arg->Sy;
int FS = Fx*Fy;
- int PadL = Arg->Pad[0], PadT = Arg->Pad[2];
+ int PadL = Arg->Pad[0], PadT = Arg->Pad[2], Pad_Val = Arg->Pad_Val;
int InFeat = Arg->Tile_InFeat, OutFeat = Arg->Tile_OutFeat;
int * __restrict__ Bias = Arg->Bias;
signed char * __restrict__ Out = Arg->Out;
unsigned char * __restrict__ Scale = Arg->Scale;
unsigned char * __restrict__ ScaleN = Arg->ScaleN;
unsigned char * __restrict__ ColBuff1 = Arg->ColBuff;
- unsigned char * __restrict__ ColBuff2 = Arg->ColBuff + InFeat*FS;
int Wo = Arg->Tile_OutW, Ho = Arg->Tile_OutH;
unsigned int * Semaphores = Arg->Semaphores;
@@ -2469,6 +2468,7 @@ void Ker_MM_Conv2D_NE16(
int ColBuffSize = ((W_In1+15)/16)*16;
int Tail = ColBuffSize / 4;
+ unsigned char * __restrict__ ColBuff2 = Arg->ColBuff + ColBuffSize;
((int *)ColBuff1)[Tail-1] = 0; ((int *)ColBuff1)[Tail-2] = 0; ((int *)ColBuff1)[Tail-3] = 0; ((int *)ColBuff1)[Tail-4] = 0;
((int *)ColBuff2)[Tail-1] = 0; ((int *)ColBuff2)[Tail-2] = 0; ((int *)ColBuff2)[Tail-3] = 0; ((int *)ColBuff2)[Tail-4] = 0;
int PosL = Arg->FirstTile?(-PadT):0;
@@ -2518,9 +2518,9 @@ void Ker_MM_Conv2D_NE16(
gap_waitbarrier(0);
/* Im2Col */
- for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*FS))[i]=0;
- if (Iter1&0x2) ((short int *)(ColBuff+F*FS))[Iter1/2-1]=0;
- if (Iter1&0x1) ((signed char *)(ColBuff+F*FS))[Iter1-1]=0;
+ for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*FS))[i]=Pad_Val;
+ if (Iter1&0x2) ((short int *)(ColBuff+F*FS))[Iter1/2-1]=Pad_Val;
+ if (Iter1&0x1) ((signed char *)(ColBuff+F*FS))[Iter1-1]=Pad_Val;
int Lb = Max(PosC, 0), Rb = Min(PosC+Fx, W);
int OffC = -Lb - Min(PosC, 0);
int Size = Rb-Lb;
diff --git a/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.h b/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.h
index 38311e356..17d64ec8f 100644
--- a/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.h
+++ b/tools/autotiler_v3/CNN_Libraries_NE16/CNN_BasicKernels_NE16.h
@@ -83,7 +83,7 @@ typedef struct {
unsigned char Dx;
unsigned char Dy;
unsigned int Semaphores[2];
-} Ker_MM_Conv_NE16_T;
+} KerConv_MM_NE16_T;
typedef struct {
void * __restrict__ In;
@@ -243,7 +243,7 @@ void KerConvNxMDxDy_StrideSxSy_NE16(KerConv_NE16_T *Arg);
void KerConv3x3Stride1_DxDy_NE16(KerConv_NE16_T *Arg);
void KerLinear_8a_NE16(KerLinear_NE16_T *Arg);
void KerLinear_16a_NE16(KerLinear_NE16_T *Arg);
-void Ker_MM_Conv2D_NE16(Ker_MM_Conv_NE16_T *Arg);
+void Ker_MM_Conv2D_NE16(KerConv_MM_NE16_T *Arg);
void KerMatMul_8a_NE16(KerMatMul_NE16_T *Arg);
void KerMatMul_8aFast_NE16(KerMatMul_NE16_T *Arg);
void KerMatMul_16a_NE16(KerMatMul_NE16_T *Arg);
diff --git a/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_BasicKernels_SQ8.h b/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_BasicKernels_SQ8.h
index eb5fdb638..0475c801a 100644
--- a/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_BasicKernels_SQ8.h
+++ b/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_BasicKernels_SQ8.h
@@ -1206,7 +1206,10 @@ extern void KerParMatMulB32_ReLUN_SF_SQ8(KerMatMul_SQ8_T *Arg);
extern void KerPar_MM_Conv1x1_HWC_SQ8(Ker_MM_Conv_SQ8_T *Arg);
extern void KerPar_MM_Conv1x1_ReLU_HWC_SQ8(Ker_MM_Conv_SQ8_T *Arg);
+
+extern void Ker_MM_Conv1x1_HWC_SQ8(Ker_MM_Conv_SQ8_T *Arg);
extern void Ker_MM_Conv1x1_ReLU_HWC_SQ8(Ker_MM_Conv_SQ8_T *Arg);
+
extern void KerPar_MM_Conv1D_SQ8(Ker_MM_Conv_SQ8_T *Arg);
extern void KerPar_MM_Conv1D_HWC_SQ8(Ker_MM_Conv_SQ8_T *Arg);
extern void KerPar_MM_Conv1D_ReLU_SQ8(Ker_MM_Conv_SQ8_T *Arg);
diff --git a/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatAlgebra_SQ8.c b/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatAlgebra_SQ8.c
index 6f713b20b..1d9b4c173 100644
--- a/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatAlgebra_SQ8.c
+++ b/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatAlgebra_SQ8.c
@@ -3205,11 +3205,39 @@ void KerParMatMulB32_SF_SQ8(KerMatMul_SQ8_T *Arg)
for (int i=0; i<Iter/4; i++) {
int l2 = 4*i+First;
v4s *pIn2_0 = (v4s *) (In2 + (l2+0)*W_In2), *pIn2_1 = (v4s *) (In2 + (l2+1)*W_In2), *pIn2_2 = (v4s *) (In2 + (l2+2)*W_In2), *pIn2_3 = (v4s *) (In2 + (l2+3)*W_In2);
- for (int l1=0; l1<H_In1; l1++) {
- v4s *pIn1 = (v4s *) (In1 + l1*W_In1);
+ for (int j=0; j<H_In1/2; j++) {
+ int l1 = 2*j;
+ v4s *pIn1_0 = (v4s *) (In1 + l1*W_In1);
+ v4s *pIn1_1 = (v4s *) (In1 + (l1+1)*W_In1);
+ int S0 = Bias[l1] <<NormBias, S1=S0, S2=S0, S3=S0;
+ int S4 = Bias[l1+1]<<NormBias, S5=S4, S6=S4, S7=S4;
+ for (int c=0; c<W_In1/4; c++) {
+ v4s C0 = pIn1_0[c], V0 = pIn2_0[c], V1 = pIn2_1[c], V2 = pIn2_2[c], V3 = pIn2_3[c];
+ S0 = gap_sumdotp4(C0, V0, S0); S1 = gap_sumdotp4(C0, V1, S1); S2 = gap_sumdotp4(C0, V2, S2); S3 = gap_sumdotp4(C0, V3, S3);
+ v4s C1 = pIn1_1[c];
+ S4 = gap_sumdotp4(C1, V0, S4); S5 = gap_sumdotp4(C1, V1, S5); S6 = gap_sumdotp4(C1, V2, S6); S7 = gap_sumdotp4(C1, V3, S7);
+ }
+ for (int c=(W_In1/4)*4; c<W_In1; c++) {
+ int C0 = In1[l1*W_In1+c];
+ S0 += C0 * In2[(l2+0)*W_In2+c]; S1 += C0 * In2[(l2+1)*W_In2+c]; S2 += C0 * In2[(l2+2)*W_In2+c]; S3 += C0 * In2[(l2+3)*W_In2+c];
+ int C1 = In1[(l1+1)*W_In1+c];
+ S4 += C1 * In2[(l2+0)*W_In2+c]; S5 += C1 * In2[(l2+1)*W_In2+c]; S6 += C1 * In2[(l2+2)*W_In2+c]; S7 += C1 * In2[(l2+3)*W_In2+c];
+ }
+ unsigned int Sc = Scale[l1], ScN = ScaleN[l1];
+ v4s R = gap_pack4(gap_clip(AT_SCALE(S0, Sc, ScN), 7), gap_clip(AT_SCALE(S1, Sc, ScN), 7),
+ gap_clip(AT_SCALE(S2, Sc, ScN), 7), gap_clip(AT_SCALE(S3, Sc, ScN), 7));
+ *((v4s *) (Out+l1*H_In2 + l2)) = R;
+ unsigned int Sc1 = Scale[l1+1], ScN1 = ScaleN[l1+1];
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S4, Sc1, ScN1), 7), gap_clip(AT_SCALE(S5, Sc1, ScN1), 7),
+ gap_clip(AT_SCALE(S6, Sc1, ScN1), 7), gap_clip(AT_SCALE(S7, Sc1, ScN1), 7));
+ *((v4s *) (Out+(l1+1)*H_In2 + l2)) = R1;
+ }
+ if (H_In1&0x1) {
+ int l1 = H_In1 - 1;
+ v4s *pIn1_0 = (v4s *) (In1 + l1*W_In1);
int S0 = Bias[l1]<<NormBias, S1=S0, S2=S0, S3=S0;
for (int c=0; c<W_In1/4; c++) {
- v4s C0 = pIn1[c], V0 = pIn2_0[c], V1 = pIn2_1[c], V2 = pIn2_2[c], V3 = pIn2_3[c];
+ v4s C0 = pIn1_0[c], V0 = pIn2_0[c], V1 = pIn2_1[c], V2 = pIn2_2[c], V3 = pIn2_3[c];
S0 = gap_sumdotp4(C0, V0, S0); S1 = gap_sumdotp4(C0, V1, S1); S2 = gap_sumdotp4(C0, V2, S2); S3 = gap_sumdotp4(C0, V3, S3);
}
for (int c=(W_In1/4)*4; c<W_In1; c++) {
diff --git a/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatMul_Conv_SQ8.c b/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatMul_Conv_SQ8.c
index 123186654..81e4e8db1 100644
--- a/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatMul_Conv_SQ8.c
+++ b/tools/autotiler_v3/CNN_Libraries_SQ8/CNN_MatMul_Conv_SQ8.c
@@ -84,6 +84,7 @@ void KerPar_MM_Conv1D_SQ8(
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
int Iter = (L-F)*Fx;
+ int IterOut = Last-First;
int PosL = 0;
for (int l=0; l<Ho; l++) {
int PosC = -PadL;
@@ -156,154 +157,6 @@ void KerPar_MM_Conv1D_SQ8(
}
}
-void _KerPar_MM_Conv1x1_HWC_SQ8(
- Ker_MM_Conv_SQ8_T *Arg
- )
-
-{
- /*
- For HWC weights (4D Tensor) are expected to be organized as [OutFeat x Fy x Fx x InFeat]
- ColBuff must be 4 byte aligned large enough to accomodate 2*Align(Feat, 4)
- */
- signed char *__restrict__ In = Arg->In;
- int W = Arg->W, H = Arg->H;
- signed char *__restrict__ Filter = Arg->Filter;
- int Sx = Arg->Sx, Sy = Arg->Sy;
- int PadL = Arg->Pad[0];
- int InFeat = Arg->InFeat, OutFeat = Arg->OutFeat;
-
- int * __restrict__ Bias = Arg->Bias;
- int NormBias = Arg->Infos[AT_INF_BIASN];
- signed char * __restrict__ Out = Arg->Out;
- unsigned char * __restrict__ Scale = Arg->Scale;
- unsigned char * __restrict__ ScaleN = Arg->ScaleN;
- signed char * __restrict__ ColBuff0 = Arg->ColBuff;
- signed char * __restrict__ ColBuff1 = Arg->ColBuff+ALIGN(InFeat, 2);
-
- int Wo = Arg->Wo, Ho = Arg->Ho;
-
- /* ColBuff must be large enough to accomodate Align(InFeat, 8) elements */
- v4s * __restrict__ VBuff0 = (v4s *) ColBuff0;
- v4s * __restrict__ VBuff1 = (v4s *) ColBuff1;
- unsigned int W_In1 = InFeat;
- unsigned int CoreId = gap_coreid(), C = ChunkSize(InFeat), F = Min(CoreId*C, InFeat), L = Min(InFeat, F+C);
- unsigned int ChunkCell = ChunkSize(OutFeat), First = CoreId*ChunkCell, Last = Min(OutFeat, First+ChunkCell);
-
- int Tail = ((W_In1+3)/4);
- ((int *)ColBuff0)[Tail-1] = 0;
- ((int *)ColBuff1)[Tail-1] = 0;
-
- int PosL = 0;
- int Iter = L-F;
- for (int l=0; l<Ho; l++) {
- int PosC = 0;
- for (int c=0; c<Wo/2; c++) {
- if (Iter>=4) {
- for (int f=0; f<(Iter/4); f++) {
- int X0 = ((int *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[f];
- int X1 = ((int *)(In+PosL*W*InFeat + (PosC+1)*InFeat+F))[f];
- ((int *)(ColBuff0+F))[f] = X0;
- ((int *)(ColBuff1+F))[f] = X1;
- }
- if (Iter&0x2) {
- int X0 = ((short int *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter/2-1];
- int X1 = ((short int *)(In+PosL*W*InFeat + (PosC+1)*InFeat+F))[Iter/2-1];
- ((short int *)(ColBuff0+F))[Iter/2-1] = X0;
- ((short int *)(ColBuff1+F))[Iter/2-1] = X1;
- }
- if (Iter&0x1) {
- int X0 = ((signed char *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter-1];
- int X1 = ((signed char *)(In+PosL*W*InFeat + (PosC+1)*InFeat+F))[Iter-1];
- ((signed char *)(ColBuff0+F))[Iter-1] = X0;
- ((signed char *)(ColBuff1+F))[Iter-1] = X1;
- }
- } else if (Iter>=2) {
- if (Iter&0x2) {
- int X0 = ((short int *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter/2-1];
- int X1 = ((short int *)(In+PosL*W*InFeat + (PosC+1)*InFeat+F))[Iter/2-1];
- ((short int *)(ColBuff0+F))[Iter/2-1] = X0;
- ((short int *)(ColBuff1+F))[Iter/2-1] = X1;
- }
- if (Iter&0x1) {
- int X0 = ((signed char *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter-1];
- int X1 = ((signed char *)(In+PosL*W*InFeat + (PosC+1)*InFeat+F))[Iter-1];
- ((signed char *)(ColBuff0+F))[Iter-1] = X0;
- ((signed char *)(ColBuff1+F))[Iter-1] = X1;
- }
- } else {
- int X0 = In[PosL*W*InFeat + (PosC+0)*InFeat + F];
- int X1 = In[PosL*W*InFeat + (PosC+1)*InFeat + F];
- ColBuff0[F] = X0;
- ColBuff1[F] = X1;
- }
- PosC += 2*Sx;
- gap_waitbarrier(0);
- for (int Line=First; Line<Last; Line++) {
- v4s *VIn1 = (v4s *) (&Filter[Line*W_In1 + 0]);
- int S0 = (Bias[Line]<<NormBias);
- int S1 = S0;
- for (int i=0; i<((W_In1+3)/4); i++) {
- v4s C0 = VIn1[i];
- v4s V0 = VBuff0[i], V1 = VBuff1[i];
- S0 = gap_sumdotp4(V0, C0, S0);
- S1 = gap_sumdotp4(V1, C0, S1);
- }
- unsigned int Sc = Scale[Line], ScN = ScaleN[Line];
- Out[l*Wo*OutFeat + (2*c+0)*OutFeat + Line] = gap_clip(AT_SCALE(S0, Sc, ScN), 7);
- Out[l*Wo*OutFeat + (2*c+1)*OutFeat + Line] = gap_clip(AT_SCALE(S1, Sc, ScN), 7);
- }
- gap_waitbarrier(0);
- }
- if (Wo&0X1) {
- int c = Wo-1;
- ((int *)ColBuff0)[Tail] = 0;
- if (Iter>=4) {
- for (int f=0; f<(Iter/4); f++) {
- int X0 = ((int *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[f];
- ((int *)(ColBuff0+F))[f] = X0;
- }
- if (Iter&0x2) {
- int X0 = ((short int *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter/2-1];
- ((short int *)(ColBuff0+F))[Iter/2-1] = X0;
- }
- if (Iter&0x1) {
- int X0 = ((signed char *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter-1];
- ((signed char *)(ColBuff0+F))[Iter-1] = X0;
- }
- } else if (Iter>=2) {
- if (Iter&0x2) {
- int X0 = ((short int *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter/2-1];
- ((short int *)(ColBuff0+F))[Iter/2-1] = X0;
- }
- if (Iter&0x1) {
- int X0 = ((signed char *)(In+PosL*W*InFeat + (PosC+0)*InFeat+F))[Iter-1];
- ((signed char *)(ColBuff0+F))[Iter-1] = X0;
- }
- } else {
- int X0 = In[PosL*W*InFeat + (PosC+0)*InFeat + F];
- ColBuff0[F] = X0;
- }
- PosC += Sx;
- gap_waitbarrier(0);
- for (int Line=First; Line<Last; Line++) {
- v4s *VIn1 = (v4s *) (&Filter[Line*W_In1 + 0]);
- int S0 = (Bias[Line]<<NormBias);
- int S1 = S0;
- for (int i=0; i<((W_In1+7)/8); i++) {
- v4s V0 = VIn1[2*i], V1 = VIn1[2*i+1];
- v4s C0 = VBuff0[2*i], C1 = VBuff0[2*i+1];
- S0 = gap_sumdotp4(V0, C0, S0);
- S0 = gap_sumdotp4(V1, C1, S0);
- }
- unsigned int Sc = Scale[Line], ScN = ScaleN[Line];
- Out[l*Wo*OutFeat + (c)*OutFeat + Line] = gap_clip(AT_SCALE(S0, Sc, ScN), 7);
- }
- gap_waitbarrier(0);
- }
- PosL += Sy;
- }
-}
-
static void __attribute__ ((noinline)) MatMul2Out(
signed char *__restrict__ pI,
signed char *__restrict__ pC,
@@ -702,6 +555,166 @@ void KerPar_MM_Conv1x1_ReLU_HWC_SQ8(
gap_waitbarrier(0);
}
+void Ker_MM_Conv1x1_HWC_SQ8(
+ Ker_MM_Conv_SQ8_T *Arg
+ )
+
+{
+ /*
+ For HWC weights (4D Tensor) are expected to be organized as [OutFeat x Fy x Fx x InFeat]
+ */
+ signed char *__restrict__ In = Arg->In;
+ int W = Arg->W, H = Arg->H;
+ signed char *__restrict__ Filter = Arg->Filter;
+ int Sx = Arg->Sx, Sy = Arg->Sy;
+ unsigned int InFeat = Arg->InFeat, OutFeat = Arg->OutFeat;
+
+ int * __restrict__ Bias = Arg->Bias;
+ int NormBias = Arg->Infos[AT_INF_BIASN];
+ signed char * __restrict__ Out = Arg->Out;
+ unsigned char * __restrict__ Scale = Arg->Scale;
+ unsigned char * __restrict__ ScaleN = Arg->ScaleN;
+
+ int Wo = Arg->Wo, Ho = Arg->Ho;
+
+ unsigned int CoreId = gap_coreid();
+ unsigned int ChunkCell = ChunkSize(Ho), First = Min(Ho, CoreId*ChunkCell), Last = Min(Ho, First+ChunkCell);
+ int IterOut = Last - First;
+
+ for (int l=First; l<Last; l++) {
+ int PosC = 0;
+ for (int c=0; c<Wo/2; c++) {
+ int *pBias = Bias;
+ signed char *pOut0 = Out+l*Wo*OutFeat + (2*c+0)*OutFeat;
+ signed char *pOut1 = Out+l*Wo*OutFeat + (2*c+1)*OutFeat;
+ signed char *pC = Filter;
+ signed char *pI = (In+l*Sy*W*InFeat + (PosC+0)*InFeat);
+ unsigned char *pSc = Scale;
+ unsigned char *pScN = ScaleN;
+
+ for (int i=0; i<(OutFeat/4); i++) {
+ signed char *pIn0 = pI, *pIn1 = pIn0 + Sx*InFeat,
+ *pC0 = pC, *pC1 = pC0+InFeat, *pC2 = pC1+InFeat, *pC3 = pC2+InFeat;
+ pC=pC3+InFeat;
+ int S00 = (*pBias)<<NormBias, S01 = S00; pBias++;
+ int S10 = (*pBias)<<NormBias, S11 = S10; pBias++;
+ int S20 = (*pBias)<<NormBias, S21 = S20; pBias++;
+ int S30 = (*pBias)<<NormBias, S31 = S30; pBias++;
+ for (int f=0; f<(InFeat/4); f++) {
+ v4s V0 = *((v4s *)pIn0), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S00 = gap_sumdotp4(V0, C0, S00); S01 = gap_sumdotp4(V1, C0, S01);
+ S10 = gap_sumdotp4(V0, C1, S10); S11 = gap_sumdotp4(V1, C1, S11);
+ S20 = gap_sumdotp4(V0, C2, S20); S21 = gap_sumdotp4(V1, C2, S21);
+ S30 = gap_sumdotp4(V0, C3, S30); S31 = gap_sumdotp4(V1, C3, S31);
+ pIn0+=4; pIn1+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(InFeat/4); f<InFeat; f++) {
+ int V0 = *pIn0, V1 = *pIn1, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S00 += V0*C0; S01 += V1*C0;
+ S10 += V0*C1; S11 += V1*C1;
+ S20 += V0*C2; S21 += V1*C2;
+ S30 += V0*C3; S31 += V1*C3;
+ pIn0++; pIn1++; pC0++; pC1++; pC2++; pC3++;
+ }
+ unsigned int Sc, ScN;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S00, Sc, ScN), 7); pOut0++;
+ *pOut1 = gap_clip(AT_SCALE(S01, Sc, ScN), 7); pOut1++;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S10, Sc, ScN), 7); pOut0++;
+ *pOut1 = gap_clip(AT_SCALE(S11, Sc, ScN), 7); pOut1++;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S20, Sc, ScN), 7); pOut0++;
+ *pOut1 = gap_clip(AT_SCALE(S21, Sc, ScN), 7); pOut1++;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S30, Sc, ScN), 7); pOut0++;
+ *pOut1 = gap_clip(AT_SCALE(S31, Sc, ScN), 7); pOut1++;
+ }
+ for (int i=4*(OutFeat/4); i<OutFeat; i++) {
+ signed char *pIn0 = pI, *pIn1 = pIn0 + InFeat;
+ int S00 = (*pBias)<<NormBias, S01 = S00; pBias++;
+ for (int f=0; f<(InFeat/4); f++) {
+ v4s V0 = *((v4s *)pIn0), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC);
+ S00 = gap_sumdotp4(V0, C0, S00); S01 = gap_sumdotp4(V1, C0, S01);
+ pIn0+=4; pIn1+=4; pC+=4;
+ }
+ for (int f=4*(InFeat/4); f<InFeat; f++) {
+ int V0 = *pIn0, V1 = *pIn1, C0 = *pC;
+ S00 += V0*C0; S01 += V1*C0;
+ pIn0++; pIn1++; pC++;
+ }
+ unsigned int Sc, ScN;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S00, Sc, ScN), 7); pOut0++;
+ *pOut1 = gap_clip(AT_SCALE(S01, Sc, ScN), 7); pOut1++;
+ }
+ PosC += 2*Sx;
+ }
+ if (Wo&0X1) {
+ PosC = (Wo/2)*2*Sx;
+ int *pBias = Bias;
+ signed char *pOut0 = Out+l*Wo*OutFeat + (Wo-1)*OutFeat;
+ signed char *pC = Filter;
+ signed char *pI = (In+Sy*l*W*InFeat + (PosC+0)*InFeat);
+ unsigned char *pSc = Scale;
+ unsigned char *pScN = ScaleN;
+
+ for (int i=0; i<(OutFeat/4); i++) {
+ signed char *pIn0 = pI,
+ *pC0 = pC, *pC1 = pC0+InFeat, *pC2 = pC1+InFeat, *pC3 = pC2+InFeat;
+ pC=pC3+InFeat;
+ int S00 = (*pBias)<<NormBias; pBias++;
+ int S10 = (*pBias)<<NormBias; pBias++;
+ int S20 = (*pBias)<<NormBias; pBias++;
+ int S30 = (*pBias)<<NormBias; pBias++;
+ for (int f=0; f<(InFeat/4); f++) {
+ v4s V0 = *((v4s *)pIn0), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S00 = gap_sumdotp4(V0, C0, S00);
+ S10 = gap_sumdotp4(V0, C1, S10);
+ S20 = gap_sumdotp4(V0, C2, S20);
+ S30 = gap_sumdotp4(V0, C3, S30);
+ pIn0+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(InFeat/4); f<InFeat; f++) {
+ int V0 = *pIn0, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S00 += V0*C0;
+ S10 += V0*C1;
+ S20 += V0*C2;
+ S30 += V0*C3;
+ pIn0++; pC0++; pC1++; pC2++; pC3++;
+ }
+ unsigned int Sc, ScN;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S00, Sc, ScN), 7); pOut0++;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S10, Sc, ScN), 7); pOut0++;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S20, Sc, ScN), 7); pOut0++;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S30, Sc, ScN), 7); pOut0++;
+ }
+ for (int i=4*(OutFeat/4); i<OutFeat; i++) {
+ signed char *pIn0 = pI;
+ int S00 = (*pBias)<<NormBias; pBias++;
+ for (int f=0; f<(InFeat/4); f++) {
+ v4s V0 = *((v4s *)pIn0), C0 = *((v4s *)pC);
+ S00 = gap_sumdotp4(V0, C0, S00);
+ pIn0+=4; pC+=4;
+ }
+ for (int f=4*(InFeat/4); f<InFeat; f++) {
+ int V0 = *pIn0, C0 = *pC;
+ S00 += V0*C0;
+ pIn0++; pC++;
+ }
+ unsigned int Sc, ScN;
+ Sc = *pSc; ScN = *pScN; pSc++; pScN++;
+ *pOut0 = gap_clip(AT_SCALE(S00, Sc, ScN), 7); pOut0++;
+ }
+ }
+ }
+ gap_waitbarrier(0);
+}
+
void Ker_MM_Conv1x1_ReLU_HWC_SQ8(
Ker_MM_Conv_SQ8_T *Arg
)
@@ -893,14 +906,120 @@ void KerPar_MM_Conv1D_HWC_SQ8(
unsigned int ChunkCell = ChunkSize(OutFeat), First = CoreId*ChunkCell, Last = Min(OutFeat, First+ChunkCell);
int Tail = 2*((W_In1+7)/8);
+ signed char * __restrict__ ColBuff1 = ColBuff + 4*Tail;
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
-
+ ((int *)ColBuff1)[Tail-1] = 0; ((int *)ColBuff1)[Tail-2] = 0;
int PosL = 0;
int Iter = L-F;
int Iter1 = Iter*Fx;
- for (int l=0; l<Ho; l++) {
+ int IterOut = Max(0, Last - First);
+ // for (int l=0; l<Ho; l++) {
int PosC = -PadL;
- for (int c=0; c<Wo; c++) {
+ for (int c=0; c<(Wo/2); c++) {
+ for (int i=0; i<(Iter1/4); i++) {
+ ((int *)(ColBuff+F*Fx))[i]=0;
+ ((int *)(ColBuff1+F*Fx))[i]=0;
+ }
+ if (Iter1&0x2) {
+ ((short int *)(ColBuff+F*Fx))[Iter1/2-1]=0;
+ ((short int *)(ColBuff1+F*Fx))[Iter1/2-1]=0;
+ }
+ if (Iter1&0x1) {
+ ((signed char *)(ColBuff+F*Fx))[Iter1-1]=0;
+ ((signed char *)(ColBuff1+F*Fx))[Iter1-1]=0;
+ }
+ int Lb = Max(PosC, 0), Rb = Min(PosC+Fx, W), Off = -Lb - Min(PosC, 0);
+ int Lb1 = Max(PosC+Sx, 0), Rb1 = Min(PosC+Fx+Sx, W), Off1 = -Lb1 - Min(PosC+Sx, 0);
+ if (Iter>0) {
+ if (Iter>=4) {
+ for (int f=0; f<(Iter/4); f++) {
+ for (int i=Lb; i<Rb; i++) ((int *)(ColBuff+(i+Off)*InFeat+F))[f] = ((int *)(In+ i*InFeat+F))[f];
+ for (int i=Lb1; i<Rb1; i++) ((int *)(ColBuff1+(i+Off1)*InFeat+F))[f] = ((int *)(In+ i*InFeat+F))[f];
+ }
+ if (Iter&0x2) {
+ for (int i=Lb; i<Rb; i++) ((short int *)(ColBuff+(i+Off)*InFeat+F))[Iter/2-1] = ((short int *)(In+ i*InFeat+F))[Iter/2-1];
+ for (int i=Lb1; i<Rb1; i++) ((short int *)(ColBuff1+(i+Off1)*InFeat+F))[Iter/2-1] = ((short int *)(In+ i*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1) {
+ for (int i=Lb; i<Rb; i++) ((signed char *)(ColBuff+(i+Off)*InFeat+F))[Iter-1] = ((signed char *)(In+ i*InFeat+F))[Iter-1];
+ for (int i=Lb1; i<Rb1; i++) ((signed char *)(ColBuff1+(i+Off1)*InFeat+F))[Iter-1] = ((signed char *)(In+ i*InFeat+F))[Iter-1];
+ }
+ } else if (Iter>=2) {
+ if (Iter&0x2) {
+ for (int i=Lb; i<Rb; i++) ((short int *)(ColBuff+(i+Off)*InFeat+F))[Iter/2-1] = ((short int *)(In+ i*InFeat+F))[Iter/2-1];
+ for (int i=Lb1; i<Rb1; i++) ((short int *)(ColBuff1+(i+Off1)*InFeat+F))[Iter/2-1] = ((short int *)(In+ i*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1) {
+ for (int i=Lb; i<Rb; i++) ((signed char *)(ColBuff+(i+Off)*InFeat+F))[Iter-1] = ((signed char *)(In+ i*InFeat+F))[Iter-1];
+ for (int i=Lb1; i<Rb1; i++) ((signed char *)(ColBuff1+(i+Off1)*InFeat+F))[Iter-1] = ((signed char *)(In+ i*InFeat+F))[Iter-1];
+ }
+ } else {
+ for (int i=Lb; i<Rb; i++) ColBuff[(i+Off)*InFeat + F] = In[ i*InFeat + F];
+ for (int i=Lb1; i<Rb1; i++) ColBuff1[(i+Off1)*InFeat + F] = In[ i*InFeat + F];
+ }
+ }
+ PosC += 2*Sx;
+ gap_waitbarrier(0);
+
+ int *pBias = Bias + First;
+ signed char *pC = Filter + W_In1*First;
+ signed char *pOut0 = Out + (2*c+0)*OutFeat+First;
+ signed char *pOut1 = Out + (2*c+1)*OutFeat+First;
+ unsigned char *pSc = Scale + First;
+ unsigned char *pScN = ScaleN + First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ signed char *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ int S0 = (*pBias)<<NormBias, S4=S0; pBias++;
+ int S1 = (*pBias)<<NormBias, S5=S1; pBias++;
+ int S2 = (*pBias)<<NormBias, S6=S2; pBias++;
+ int S3 = (*pBias)<<NormBias, S7=S3; pBias++;
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ S1 = gap_sumdotp4(V0, C1, S1); S5 = gap_sumdotp4(V1, C1, S5);
+ S2 = gap_sumdotp4(V0, C2, S2); S6 = gap_sumdotp4(V1, C2, S6);
+ S3 = gap_sumdotp4(V0, C3, S3); S7 = gap_sumdotp4(V1, C3, S7);
+ pIn+=4; pIn1+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S0 += V0*C0; S4 += V1*C0;
+ S1 += V0*C1; S5 += V1*C1;
+ S2 += V0*C2; S6 += V1*C2;
+ S3 += V0*C3; S7 += V1*C3;
+ pIn++; pIn1++; pC0++; pC1++; pC2++; pC3++;
+ }
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ v4s R2 = gap_pack4(gap_clip(AT_SCALE(S4, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S5, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S6, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S7, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
+ *((v4s *) (pOut1+4*Line)) = R2;
+ }
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ int S0 = (*pBias)<<NormBias, S4=S0; pBias++;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ pIn+=4; pIn1+=4; pC+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC;
+ S0 += V0*C0; S4 += V1*C0;
+ pIn++; pIn1++; pC++;
+ }
+ *(pOut0+i) = gap_clip(AT_SCALE(S0, pSc[i], pScN[i]), 7);
+ *(pOut1+i) = gap_clip(AT_SCALE(S4, pSc[i], pScN[i]), 7);
+ }
+ gap_waitbarrier(0);
+ }
+ if (Wo&0x1) {
+ int c = Wo-1;
for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*Fx))[i]=0;
if (Iter1&0x2) ((short int *)(ColBuff+F*Fx))[Iter1/2-1]=0;
if (Iter1&0x1) ((signed char *)(ColBuff+F*Fx))[Iter1-1]=0;
@@ -919,134 +1038,61 @@ void KerPar_MM_Conv1D_HWC_SQ8(
}
PosC += Sx;
gap_waitbarrier(0);
- for (int Line=First; Line<Last; Line++) {
- v4s *VIn1 = (v4s *) (&Filter[Line*W_In1 + 0]);
- int S0 = (Bias[Line]<<NormBias);
- for (int i=0; i<((W_In1+7)/8); i++) {
- v4s V0 = VIn1[2*i], V1 = VIn1[2*i+1];
- v4s C0 = VBuff[2*i], C1 = VBuff[2*i+1];
+
+ int *pBias = Bias + First;
+ signed char *pC = Filter + W_In1*First;
+ signed char *pOut0 = Out + (c)*OutFeat+First;
+ unsigned char *pSc = Scale + First;
+ unsigned char *pScN = ScaleN + First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ signed char *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ int S0 = (*pBias)<<NormBias; pBias++;
+ int S1 = (*pBias)<<NormBias; pBias++;
+ int S2 = (*pBias)<<NormBias; pBias++;
+ int S3 = (*pBias)<<NormBias; pBias++;
+ signed char *pIn = ColBuff;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
S0 = gap_sumdotp4(V0, C0, S0);
- S0 = gap_sumdotp4(V1, C1, S0);
+ S1 = gap_sumdotp4(V0, C1, S1);
+ S2 = gap_sumdotp4(V0, C2, S2);
+ S3 = gap_sumdotp4(V0, C3, S3);
+ pIn+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
}
- unsigned int Sc = Scale[Line], ScN = ScaleN[Line];
- Out[l*Wo*OutFeat + c*OutFeat + Line] = gap_clip(AT_SCALE(S0, Sc, ScN), 7);
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ pIn++; pC0++; pC1++; pC2++; pC3++;
+ }
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
}
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ signed char *pIn = ColBuff;
+ int S0 = (*pBias)<<NormBias; pBias++;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), C0 = *((v4s *)pC);
+ S0 = gap_sumdotp4(V0, C0, S0);
+ pIn+=4; pC+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, C0 = *pC;
+ S0 += V0*C0;
+ pIn++; pC++;
+ }
+ *(pOut0+i) = gap_clip(AT_SCALE(S0, pSc[i], pScN[i]), 7);
+ }
gap_waitbarrier(0);
}
- PosL += Sy;
- }
+ // PosL += Sy;
+ // }
}
-// void KerPar_MM_Conv1D_HWC_SQ8(
-// Ker_MM_Conv_SQ8_T *Arg
-// )
-
-// {
-// /*
-// For HWC weights (4D Tensor) are expected to be organized as [OutFeat x Fy x Fx x InFeat]
-// */
-// signed char *__restrict__ In = Arg->In;
-// int W = Arg->W, H = Arg->H;
-// signed char *__restrict__ Filter = Arg->Filter;
-// int Fx = Arg->Fx, Sx = Arg->Sx, Sy = Arg->Sy;
-// int PadL = Arg->Pad[0];
-// int InFeat = Arg->InFeat, OutFeat = Arg->OutFeat;
-
-// int * __restrict__ Bias = Arg->Bias;
-// int NormBias = Arg->Infos[AT_INF_BIASN];
-// signed char * __restrict__ Out = Arg->Out;
-// unsigned char * __restrict__ Scale = Arg->Scale;
-// unsigned char * __restrict__ ScaleN = Arg->ScaleN;
-
-// int Wo = Arg->Wo, Ho = Arg->Ho;
-
-// /* ColBuff must be large enough to accomodate Align(Fx*InFeat, 8) elements */
-// unsigned int W_In1 = InFeat*Fx;
-// unsigned int CoreId = gap_coreid(), C = ChunkSize(InFeat), F = Min(CoreId*C, InFeat), L = Min(InFeat, F+C);
-// unsigned int ChunkCell = ChunkSize(OutFeat), First = CoreId*ChunkCell, Last = Min(OutFeat, First+ChunkCell);
-// unsigned int Iter = (Last>First)?(Last-First):0;
-
-// int PosL = 0;
-// int PosC = -PadL;
-// for (int c=0; c<Wo/4; c++) {
-// v4s * __restrict__ VIn1 = (v4s *) (In + (4*c )*Sx*InFeat);
-// v4s * __restrict__ VIn2 = (v4s *) (In + (4*c+1)*Sx*InFeat);
-// v4s * __restrict__ VIn3 = (v4s *) (In + (4*c+2)*Sx*InFeat);
-// v4s * __restrict__ VIn4 = (v4s *) (In + (4*c+3)*Sx*InFeat);
-// for (int c_out=0; c_out<Iter/2; c_out++) {
-// int col = First + 2*c_out;
-// v4s *VFilter1 = (v4s *) (&Filter[(col )*W_In1]);
-// v4s *VFilter2 = (v4s *) (&Filter[(col+1)*W_In1]);
-// int S1 = (Bias[col ]<<NormBias), S3=S1, S5=S1, S7=S1;
-// int S2 = (Bias[col+1]<<NormBias), S4=S2, S6=S2, S8=S2;
-// for (int i=0; i<(W_In1/4); i++) {
-// S1 = gap_sumdotp4(VIn1[i], VFilter1[i], S1); S2 = gap_sumdotp4(VIn1[i], VFilter2[i], S2);
-// S3 = gap_sumdotp4(VIn2[i], VFilter1[i], S3); S4 = gap_sumdotp4(VIn2[i], VFilter2[i], S4);
-// S5 = gap_sumdotp4(VIn3[i], VFilter1[i], S5); S6 = gap_sumdotp4(VIn3[i], VFilter2[i], S6);
-// S7 = gap_sumdotp4(VIn4[i], VFilter1[i], S7); S8 = gap_sumdotp4(VIn4[i], VFilter2[i], S8);
-// }
-// for (int i=(W_In1/4)*4; i<W_In1; i++) {
-// S1 += In[(4*c )*Sx*InFeat+i]*Filter[(col)*W_In1+i]; S2 += In[(4*c )*Sx*InFeat+i]*Filter[(col+1)*W_In1+i];
-// S3 += In[(4*c+1)*Sx*InFeat+i]*Filter[(col)*W_In1+i]; S4 += In[(4*c+1)*Sx*InFeat+i]*Filter[(col+1)*W_In1+i];
-// S5 += In[(4*c+2)*Sx*InFeat+i]*Filter[(col)*W_In1+i]; S6 += In[(4*c+2)*Sx*InFeat+i]*Filter[(col+1)*W_In1+i];
-// S7 += In[(4*c+3)*Sx*InFeat+i]*Filter[(col)*W_In1+i]; S8 += In[(4*c+3)*Sx*InFeat+i]*Filter[(col+1)*W_In1+i];
-// }
-// unsigned int Sc1 = Scale[col ], ScN1 = ScaleN[col ];
-// unsigned int Sc2 = Scale[col+1], ScN2 = ScaleN[col+1];
-// v2s R1 = gap_pack2(gap_clip(AT_SCALE(S1, Sc1, ScN1), 7), gap_clip(AT_SCALE(S2, Sc2, ScN2), 7));
-// v2s R2 = gap_pack2(gap_clip(AT_SCALE(S3, Sc1, ScN1), 7), gap_clip(AT_SCALE(S4, Sc2, ScN2), 7));
-// v2s R3 = gap_pack2(gap_clip(AT_SCALE(S5, Sc1, ScN1), 7), gap_clip(AT_SCALE(S6, Sc2, ScN2), 7));
-// v2s R4 = gap_pack2(gap_clip(AT_SCALE(S7, Sc1, ScN1), 7), gap_clip(AT_SCALE(S8, Sc2, ScN2), 7));
-// *((v2s *) (&Out[(4*c )*OutFeat + (col)])) = R1;
-// *((v2s *) (&Out[(4*c+1)*OutFeat + (col)])) = R2;
-// *((v2s *) (&Out[(4*c+2)*OutFeat + (col)])) = R3;
-// *((v2s *) (&Out[(4*c+3)*OutFeat + (col)])) = R4;
-// }
-// if (Iter&0x1) {
-// int col = Last-1;
-// v4s *VFilter1 = (v4s *) (&Filter[(col)*W_In1]);
-// int S1 = (Bias[col ]<<NormBias), S3=S1, S5=S1, S7=S1;
-// for (int i=0; i<(W_In1/4); i++) {
-// S1 = gap_sumdotp4(VIn1[i], VFilter1[i], S1);
-// S3 = gap_sumdotp4(VIn2[i], VFilter1[i], S3);
-// S5 = gap_sumdotp4(VIn3[i], VFilter1[i], S5);
-// S7 = gap_sumdotp4(VIn4[i], VFilter1[i], S7);
-// }
-// for (int i=(W_In1/4)*4; i<W_In1; i++) {
-// S1 += In[(4*c )*Sx*InFeat+i]*Filter[(col)*W_In1+i];
-// S3 += In[(4*c+1)*Sx*InFeat+i]*Filter[(col)*W_In1+i];
-// S5 += In[(4*c+2)*Sx*InFeat+i]*Filter[(col)*W_In1+i];
-// S7 += In[(4*c+3)*Sx*InFeat+i]*Filter[(col)*W_In1+i];
-// }
-// unsigned int Sc1 = Scale[col ], ScN1 = ScaleN[col ];
-// Out[(4*c )*OutFeat + (col)] = gap_clip(AT_SCALE(S1, Sc1, ScN1), 7);
-// Out[(4*c+1)*OutFeat + (col)] = gap_clip(AT_SCALE(S3, Sc1, ScN1), 7);
-// Out[(4*c+2)*OutFeat + (col)] = gap_clip(AT_SCALE(S5, Sc1, ScN1), 7);
-// Out[(4*c+3)*OutFeat + (col)] = gap_clip(AT_SCALE(S7, Sc1, ScN1), 7);
-
-// }
-// gap_waitbarrier(0);
-// }
-// for (int c=(Wo/4)*4; c<Wo; c++) {
-// v4s * __restrict__ VBuff1 = (v4s *) (In + (c)*Sx*InFeat);
-// for (int c_out=First; c_out<Last; c_out++) {
-// v4s *VIn1 = (v4s *) (&Filter[c_out*W_In1]);
-// int S1 = (Bias[c_out]<<NormBias);
-// for (int i=0; i<(W_In1/4); i++) {
-// v4s V1 = VIn1[i];
-// v4s C11 = VBuff1[i];
-// S1 = gap_sumdotp4(V1, C11, S1);
-// }
-// for (int i=(W_In1/4)*4; i<W_In1; i++) {
-// S1 += In[(c)*Sx*InFeat+i]*Filter[c_out*W_In1+i];
-// }
-// unsigned int Sc = Scale[c_out], ScN = ScaleN[c_out];
-// Out[(c)*OutFeat + c_out] = gap_clip(AT_SCALE(S1, Sc, ScN), 7);
-// }
-// gap_waitbarrier(0);
-// }
-// }
-
void KerPar_MM_Conv1D_DxDy_SQ8(
Ker_MM_Conv_SQ8_T *Arg
)
@@ -1150,7 +1196,11 @@ void KerPar_MM_Conv1D_DxDy_HWC_SQ8(
unsigned int ChunkCell = ChunkSize(OutFeat), First = CoreId*ChunkCell, Last = Min(OutFeat, First+ChunkCell);
int Tail = 2*((W_In1+7)/8);
+ signed char * __restrict__ ColBuff1 = ColBuff + 4*Tail;
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
+ ((int *)ColBuff1)[Tail-1] = 0; ((int *)ColBuff1)[Tail-2] = 0;
+
+ int IterOut = Max(0, Last - First);
int DFx = Dx*(Fx-1)+1;
// int Prec=10;
@@ -1159,53 +1209,210 @@ void KerPar_MM_Conv1D_DxDy_HWC_SQ8(
int PosL = 0;
int Iter = L-F;
int Iter1 = Iter*Fx;
- for (int l=0; l<Ho; l++) {
- int PosC = -PadL;
- for (int c=0; c<Wo; c++) {
- for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*Fx))[i]=0;
- if (Iter1&0x2) ((short int *)(ColBuff+F*Fx))[Iter1/2-1]=0;
- if (Iter1&0x1) ((signed char *)(ColBuff+F*Fx))[Iter1-1]=0;
- int Lb = Max(PosC, 0), Rb = Min(PosC+DFx, W);
- int OffBuffX = Max(0, gap_mulsN(-PosC+Dx-1, InvDx, Prec));
- int OffInX = OffBuffX?(Dx*OffBuffX+PosC):0;
- int IterX = gap_mulsN(Rb-Lb-1, InvDx, Prec) + 1;
- if (Iter>=4) {
- for (int f=0; f<(Iter/4); f++)
- for (int i=0; i<IterX; i++)
- ((int *)(ColBuff+(i+OffBuffX)*InFeat+F))[f] = ((int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[f];
- if (Iter&0x2)
- for (int i=0; i<IterX; i++)
- ((short int *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter/2-1];
- if (Iter&0x1)
- for (int i=0; i<IterX; i++)
- ((signed char *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter-1];
- } else if (Iter>=2) {
- if (Iter&0x2)
- for (int i=0; i<IterX; i++)
- ((short int *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter/2-1];
- if (Iter&0x1)
- for (int i=0; i<IterX; i++)
- ((signed char *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter-1];
- } else if (Iter>0)
+ int PosC = -PadL;
+ for (int c=0; c<(Wo/2); c++) {
+ for (int i=0; i<(Iter1/4); i++) {
+ ((int *)(ColBuff+F*Fx))[i]=0;
+ ((int *)(ColBuff1+F*Fx))[i]=0;
+ }
+ if (Iter1&0x2) {
+ ((short int *)(ColBuff+F*Fx))[Iter1/2-1]=0;
+ ((short int *)(ColBuff1+F*Fx))[Iter1/2-1]=0;
+ }
+ if (Iter1&0x1) {
+ ((signed char *)(ColBuff+F*Fx))[Iter1-1]=0;
+ ((signed char *)(ColBuff1+F*Fx))[Iter1-1]=0;
+ }
+ int Lb = Max(PosC, 0), Rb = Min(PosC+DFx, W);
+ int OffBuffX = Max(0, gap_mulsN(-PosC+Dx-1, InvDx, Prec));
+ int OffInX = OffBuffX?(Dx*OffBuffX+PosC):0;
+ int IterX = gap_mulsN(Rb-Lb-1, InvDx, Prec) + 1;
+
+ int Lb1 = Max(PosC+Sx, 0), Rb1 = Min(PosC+Sx+DFx, W);
+ int OffBuffX1 = Max(0, gap_mulsN(-PosC-Sx+Dx-1, InvDx, Prec));
+ int OffInX1 = OffBuffX1?(Dx*OffBuffX1+PosC+Sx):0;
+ int IterX1 = gap_mulsN(Rb1-Lb1-1, InvDx, Prec) + 1;
+
+ if (Iter>=4) {
+ for (int f=0; f<(Iter/4); f++) {
+ for (int i=0; i<IterX; i++)
+ ((int *)(ColBuff+(i+OffBuffX)*InFeat+F))[f] = ((int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[f];
+ for (int i=0; i<IterX1; i++)
+ ((int *)(ColBuff1+(i+OffBuffX1)*InFeat+F))[f] = ((int *)(In+PosL*W*InFeat + (i*Dx+OffInX1+Lb1)*InFeat+F))[f];
+ }
+ if (Iter&0x2) {
+ for (int i=0; i<IterX; i++)
+ ((short int *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter/2-1];
+ for (int i=0; i<IterX1; i++)
+ ((short int *)(ColBuff1+(i+OffBuffX1)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX1+Lb1)*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1) {
+ for (int i=0; i<IterX; i++)
+ ((signed char *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter-1];
+ for (int i=0; i<IterX1; i++)
+ ((signed char *)(ColBuff1+(i+OffBuffX1)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX1+Lb1)*InFeat+F))[Iter-1];
+ }
+ } else if (Iter>=2) {
+ if (Iter&0x2) {
+ for (int i=0; i<IterX; i++)
+ ((short int *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter/2-1];
+ for (int i=0; i<IterX1; i++)
+ ((short int *)(ColBuff1+(i+OffBuffX1)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX1+Lb1)*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1){
+ for (int i=0; i<IterX; i++)
+ ((signed char *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter-1];
+ for (int i=0; i<IterX1; i++)
+ ((signed char *)(ColBuff1+(i+OffBuffX1)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX1+Lb1)*InFeat+F))[Iter-1];
+ }
+ } else if (Iter>0) {
+ for (int i=0; i<IterX; i++)
+ ColBuff[(i+OffBuffX)*InFeat + F] = In[PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat + F];
+ for (int i=0; i<IterX1; i++)
+ ColBuff1[(i+OffBuffX1)*InFeat + F] = In[PosL*W*InFeat + (i*Dx+OffInX1+Lb1)*InFeat + F];
+ }
+ PosC += 2*Sx;
+ gap_waitbarrier(0);
+
+
+ int *pBias = Bias + First;
+ signed char *pC = Filter + W_In1*First;
+ signed char *pOut0 = Out + (2*c+0)*OutFeat+First;
+ signed char *pOut1 = Out + (2*c+1)*OutFeat+First;
+ unsigned char *pSc = Scale + First;
+ unsigned char *pScN = ScaleN + First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ signed char *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ int S0 = (*pBias)<<NormBias, S4=S0; pBias++;
+ int S1 = (*pBias)<<NormBias, S5=S1; pBias++;
+ int S2 = (*pBias)<<NormBias, S6=S2; pBias++;
+ int S3 = (*pBias)<<NormBias, S7=S3; pBias++;
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ S1 = gap_sumdotp4(V0, C1, S1); S5 = gap_sumdotp4(V1, C1, S5);
+ S2 = gap_sumdotp4(V0, C2, S2); S6 = gap_sumdotp4(V1, C2, S6);
+ S3 = gap_sumdotp4(V0, C3, S3); S7 = gap_sumdotp4(V1, C3, S7);
+ pIn+=4; pIn1+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S0 += V0*C0; S4 += V1*C0;
+ S1 += V0*C1; S5 += V1*C1;
+ S2 += V0*C2; S6 += V1*C2;
+ S3 += V0*C3; S7 += V1*C3;
+ pIn++; pIn1++; pC0++; pC1++; pC2++; pC3++;
+ }
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ v4s R2 = gap_pack4(gap_clip(AT_SCALE(S4, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S5, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S6, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S7, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
+ *((v4s *) (pOut1+4*Line)) = R2;
+ }
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ int S0 = (*pBias)<<NormBias, S4=S0; pBias++;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ pIn+=4; pIn1+=4; pC+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC;
+ S0 += V0*C0; S4 += V1*C0;
+ pIn++; pIn1++; pC++;
+ }
+ *(pOut0+i) = gap_clip(AT_SCALE(S0, pSc[i], pScN[i]), 7);
+ *(pOut1+i) = gap_clip(AT_SCALE(S4, pSc[i], pScN[i]), 7);
+ }
+ gap_waitbarrier(0);
+ }
+ if (Wo&0x1) {
+ int c = Wo-1;
+ for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*Fx))[i]=0;
+ if (Iter1&0x2) ((short int *)(ColBuff+F*Fx))[Iter1/2-1]=0;
+ if (Iter1&0x1) ((signed char *)(ColBuff+F*Fx))[Iter1-1]=0;
+ int Lb = Max(PosC, 0), Rb = Min(PosC+DFx, W);
+ int OffBuffX = Max(0, gap_mulsN(-PosC+Dx-1, InvDx, Prec));
+ int OffInX = OffBuffX?(Dx*OffBuffX+PosC):0;
+ int IterX = gap_mulsN(Rb-Lb-1, InvDx, Prec) + 1;
+ if (Iter>=4) {
+ for (int f=0; f<(Iter/4); f++)
+ for (int i=0; i<IterX; i++)
+ ((int *)(ColBuff+(i+OffBuffX)*InFeat+F))[f] = ((int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[f];
+ if (Iter&0x2)
for (int i=0; i<IterX; i++)
- ColBuff[(i+OffBuffX)*InFeat + F] = In[PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat + F];
- PosC += Sx;
- gap_waitbarrier(0);
- for (int Line=First; Line<Last; Line++) {
- v4s *VIn1 = (v4s *) (&Filter[Line*W_In1 + 0]);
- int S0 = (Bias[Line]<<NormBias);
- for (int i=0; i<((W_In1+7)/8); i++) {
- v4s V0 = VIn1[2*i], V1 = VIn1[2*i+1];
- v4s C0 = VBuff[2*i], C1 = VBuff[2*i+1];
- S0 = gap_sumdotp4(V0, C0, S0);
- S0 = gap_sumdotp4(V1, C1, S0);
- }
- unsigned int Sc = Scale[Line], ScN = ScaleN[Line];
- Out[l*Wo*OutFeat + c*OutFeat + Line] = gap_clip(AT_SCALE(S0, Sc, ScN), 7);
- }
- gap_waitbarrier(0);
+ ((short int *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter/2-1];
+ if (Iter&0x1)
+ for (int i=0; i<IterX; i++)
+ ((signed char *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter-1];
+ } else if (Iter>=2) {
+ if (Iter&0x2)
+ for (int i=0; i<IterX; i++)
+ ((short int *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter/2-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter/2-1];
+ if (Iter&0x1)
+ for (int i=0; i<IterX; i++)
+ ((signed char *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter-1] = ((signed char *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter-1];
+ } else if (Iter>0)
+ for (int i=0; i<IterX; i++)
+ ColBuff[(i+OffBuffX)*InFeat + F] = In[PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat + F];
+ PosC += Sx;
+ gap_waitbarrier(0);
+
+ int *pBias = Bias + First;
+ signed char *pC = Filter + W_In1*First;
+ signed char *pOut0 = Out + (c)*OutFeat+First;
+ unsigned char *pSc = Scale + First;
+ unsigned char *pScN = ScaleN + First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ signed char *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ int S0 = (*pBias)<<NormBias; pBias++;
+ int S1 = (*pBias)<<NormBias; pBias++;
+ int S2 = (*pBias)<<NormBias; pBias++;
+ int S3 = (*pBias)<<NormBias; pBias++;
+ signed char *pIn = ColBuff;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S0 = gap_sumdotp4(V0, C0, S0);
+ S1 = gap_sumdotp4(V0, C1, S1);
+ S2 = gap_sumdotp4(V0, C2, S2);
+ S3 = gap_sumdotp4(V0, C3, S3);
+ pIn+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ pIn++; pC0++; pC1++; pC2++; pC3++;
+ }
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
+ }
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ signed char *pIn = ColBuff;
+ int S0 = (*pBias)<<NormBias; pBias++;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), C0 = *((v4s *)pC);
+ S0 = gap_sumdotp4(V0, C0, S0);
+ pIn+=4; pC+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, C0 = *pC;
+ S0 += V0*C0;
+ pIn++; pC++;
+ }
+ *(pOut0+i) = gap_clip(AT_SCALE(S0, pSc[i], pScN[i]), 7);
}
- PosL += Sy;
+ gap_waitbarrier(0);
}
}
@@ -1733,7 +1940,10 @@ void KerPar_MM_Conv2D_HWC_SQ8(
int FS = Fx*Fy;
int Tail = 2*((W_In1+7)/8);
+
+ signed char * __restrict__ ColBuff1 = ColBuff + 4*Tail;
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
+ ((int *)ColBuff1)[Tail-1] = 0; ((int *)ColBuff1)[Tail-2] = 0;
int PosL = Arg->FirstTile?(-PadT):0;
int Iter = L-F;
@@ -1743,7 +1953,133 @@ void KerPar_MM_Conv2D_HWC_SQ8(
int PosC = -PadL;
int Tb = Max(PosL, 0), Db = Min(PosL+Fy, H);
int OffL = -Tb - Min(PosL, 0);
- for (int c=0; c<Wo; c++) {
+ for (int c=0; c<(Wo/2); c++) {
+ for (int i=0; i<(Iter1/4); i++) {
+ ((int *)(ColBuff+F*FS))[i]=0;
+ ((int *)(ColBuff1+F*FS))[i]=0;
+ }
+ if (Iter1&0x2) {
+ ((short int *)(ColBuff+F*FS))[Iter1/2-1]=0;
+ ((short int *)(ColBuff1+F*FS))[Iter1/2-1]=0;
+ }
+ if (Iter1&0x1) {
+ ((signed char *)(ColBuff+F*FS))[Iter1-1]=0;
+ ((signed char *)(ColBuff1+F*FS))[Iter1-1]=0;
+ }
+ int Lb = Max(PosC, 0), Rb = Min(PosC+Fx, W);
+ int Lb1 = Max(PosC+Sx, 0), Rb1 = Min(PosC+Sx+Fx, W);
+ int OffC = -Lb - Min(PosC, 0);
+ int OffC1 = -Lb1 - Min(PosC+Sx, 0);
+ if (Iter>=4) {
+ for (int f=0; f<(Iter/4); f++)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++) {
+ ((int *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[f] = ((int *)(In+j*W*InFeat + i*InFeat+F))[f];
+ }
+ for (int i=Lb1; i<Rb1; i++) {
+ ((int *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[f] = ((int *)(In+j*W*InFeat + i*InFeat+F))[f];
+ }
+ }
+ if (Iter&0x2)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((short int *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((short int *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((signed char *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((signed char *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ }
+ } else if (Iter>=2) {
+ if (Iter&0x2)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((short int *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((short int *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((signed char *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((signed char *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ }
+ } else if (Iter>0) {
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ColBuff[(j+OffL)*InFeat*Fx+(i+OffC)*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+ for (int i=Lb1; i<Rb1; i++)
+ ColBuff1[(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+ }
+ }
+ PosC += 2*Sx;
+ gap_waitbarrier(0);
+
+ int *pBias = Bias + First;
+ signed char *pC = Filter + W_In1*First;
+ signed char *pOut0 = Out+l*Wo*OutFeat + (2*c+0)*OutFeat+First;
+ signed char *pOut1 = Out+l*Wo*OutFeat + (2*c+1)*OutFeat+First;
+ unsigned char *pSc = Scale + First;
+ unsigned char *pScN = ScaleN + First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ signed char *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ int S0 = (pBias[4*Line ])<<NormBias, S4=S0;
+ int S1 = (pBias[4*Line+1])<<NormBias, S5=S1;
+ int S2 = (pBias[4*Line+2])<<NormBias, S6=S2;
+ int S3 = (pBias[4*Line+3])<<NormBias, S7=S3;
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ S1 = gap_sumdotp4(V0, C1, S1); S5 = gap_sumdotp4(V1, C1, S5);
+ S2 = gap_sumdotp4(V0, C2, S2); S6 = gap_sumdotp4(V1, C2, S6);
+ S3 = gap_sumdotp4(V0, C3, S3); S7 = gap_sumdotp4(V1, C3, S7);
+ pIn+=4; pIn1+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S0 += V0*C0; S4 += V1*C0;
+ S1 += V0*C1; S5 += V1*C1;
+ S2 += V0*C2; S6 += V1*C2;
+ S3 += V0*C3; S7 += V1*C3;
+ pIn++; pIn1++; pC0++; pC1++; pC2++; pC3++;
+ }
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ v4s R2 = gap_pack4(gap_clip(AT_SCALE(S4, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S5, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S6, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S7, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
+ *((v4s *) (pOut1+4*Line)) = R2;
+ }
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ int S0 = (pBias[i])<<NormBias, S4=S0;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ pIn+=4; pIn1+=4; pC+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC;
+ S0 += V0*C0; S4 += V1*C0;
+ pIn++; pIn1++; pC++;
+ }
+ *(pOut0+i) = gap_clip(AT_SCALE(S0, pSc[i], pScN[i]), 7);
+ *(pOut1+i) = gap_clip(AT_SCALE(S4, pSc[i], pScN[i]), 7);
+ }
+ gap_waitbarrier(0);
+ }
+ if (Wo&0x1) {
+ int c = Wo-1;
+ // for (int c=0; c<Wo; c++) {
for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*FS))[i]=0;
if (Iter1&0x2) ((short int *)(ColBuff+F*FS))[Iter1/2-1]=0;
if (Iter1&0x1) ((signed char *)(ColBuff+F*FS))[Iter1-1]=0;
@@ -1772,16 +2108,16 @@ void KerPar_MM_Conv2D_HWC_SQ8(
for (int i=Lb; i<Rb; i++)
((signed char *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
} else if (Iter>0)
- for (int j=Tb; j<Db; j++)
+ for (int j=Tb; j<Db; j++)
for (int i=Lb; i<Rb; i++)
ColBuff[(j+OffL)*InFeat*Fx+(i+OffC)*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
PosC += Sx;
gap_waitbarrier(0);
- // for (int Line=First; Line<Last; Line++) {
+
int *pBias = Bias + First;
signed char *pC = Filter + W_In1*First;
- signed char *pOut0 = Out+l*Wo*OutFeat + (c+0)*OutFeat+First;
+ signed char *pOut0 = Out+l*Wo*OutFeat + (c)*OutFeat+First;
unsigned char *pSc = Scale + First;
unsigned char *pScN = ScaleN + First;
for (int Line=0; Line<IterOut/4; Line++) {
@@ -1808,15 +2144,9 @@ void KerPar_MM_Conv2D_HWC_SQ8(
S3 += V0*C3;
pIn++; pC0++; pC1++; pC2++; pC3++;
}
- unsigned int Sc, ScN;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = gap_clip(AT_SCALE(S0, Sc, ScN), 7); pOut0++;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = gap_clip(AT_SCALE(S1, Sc, ScN), 7); pOut0++;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = gap_clip(AT_SCALE(S2, Sc, ScN), 7); pOut0++;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = gap_clip(AT_SCALE(S3, Sc, ScN), 7); pOut0++;
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
}
for (int i=4*(IterOut/4); i<IterOut; i++) {
signed char *pIn = ColBuff;
@@ -1831,18 +2161,14 @@ void KerPar_MM_Conv2D_HWC_SQ8(
S0 += V0*C0;
pIn++; pC++;
}
- unsigned int Sc, ScN;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = gap_clip(AT_SCALE(S0, Sc, ScN), 7); pOut0++;
+ *(pOut0+i) = gap_clip(AT_SCALE(S0, pSc[i], pScN[i]), 7);
}
gap_waitbarrier(0);
}
PosL += Sy;
}
- gap_waitbarrier(0);
}
-
void KerPar_MM_Conv2D_ReLU_HWC_SQ8(
Ker_MM_Conv_SQ8_T *Arg
)
@@ -1875,7 +2201,10 @@ void KerPar_MM_Conv2D_ReLU_HWC_SQ8(
int FS = Fx*Fy;
int Tail = 2*((W_In1+7)/8);
+
+ signed char * __restrict__ ColBuff1 = ColBuff + 4*Tail;
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
+ ((int *)ColBuff1)[Tail-1] = 0; ((int *)ColBuff1)[Tail-2] = 0;
int PosL = Arg->FirstTile?(-PadT):0;
int Iter = L-F;
@@ -1885,7 +2214,133 @@ void KerPar_MM_Conv2D_ReLU_HWC_SQ8(
int PosC = -PadL;
int Tb = Max(PosL, 0), Db = Min(PosL+Fy, H);
int OffL = -Tb - Min(PosL, 0);
- for (int c=0; c<Wo; c++) {
+ for (int c=0; c<(Wo/2); c++) {
+ for (int i=0; i<(Iter1/4); i++) {
+ ((int *)(ColBuff+F*FS))[i]=0;
+ ((int *)(ColBuff1+F*FS))[i]=0;
+ }
+ if (Iter1&0x2) {
+ ((short int *)(ColBuff+F*FS))[Iter1/2-1]=0;
+ ((short int *)(ColBuff1+F*FS))[Iter1/2-1]=0;
+ }
+ if (Iter1&0x1) {
+ ((signed char *)(ColBuff+F*FS))[Iter1-1]=0;
+ ((signed char *)(ColBuff1+F*FS))[Iter1-1]=0;
+ }
+ int Lb = Max(PosC, 0), Rb = Min(PosC+Fx, W);
+ int Lb1 = Max(PosC+Sx, 0), Rb1 = Min(PosC+Sx+Fx, W);
+ int OffC = -Lb - Min(PosC, 0);
+ int OffC1 = -Lb1 - Min(PosC+Sx, 0);
+ if (Iter>=4) {
+ for (int f=0; f<(Iter/4); f++)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++) {
+ ((int *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[f] = ((int *)(In+j*W*InFeat + i*InFeat+F))[f];
+ }
+ for (int i=Lb1; i<Rb1; i++) {
+ ((int *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[f] = ((int *)(In+j*W*InFeat + i*InFeat+F))[f];
+ }
+ }
+ if (Iter&0x2)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((short int *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((short int *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((signed char *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((signed char *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ }
+ } else if (Iter>=2) {
+ if (Iter&0x2)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((short int *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((short int *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter/2-1] = ((short int *)(In+j*W*InFeat + i*InFeat+F))[Iter/2-1];
+ }
+ if (Iter&0x1)
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ((signed char *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ for (int i=Lb1; i<Rb1; i++)
+ ((signed char *)(ColBuff1+(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
+ }
+ } else if (Iter>0) {
+ for (int j=Tb; j<Db; j++) {
+ for (int i=Lb; i<Rb; i++)
+ ColBuff[(j+OffL)*InFeat*Fx+(i+OffC)*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+ for (int i=Lb1; i<Rb1; i++)
+ ColBuff1[(j+OffL)*InFeat*Fx+(i+OffC1)*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+ }
+ }
+ PosC += 2*Sx;
+ gap_waitbarrier(0);
+
+ int *pBias = Bias + First;
+ signed char *pC = Filter + W_In1*First;
+ signed char *pOut0 = Out+l*Wo*OutFeat + (2*c+0)*OutFeat+First;
+ signed char *pOut1 = Out+l*Wo*OutFeat + (2*c+1)*OutFeat+First;
+ unsigned char *pSc = Scale + First;
+ unsigned char *pScN = ScaleN + First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ signed char *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ int S0 = (pBias[4*Line ])<<NormBias, S4=S0;
+ int S1 = (pBias[4*Line+1])<<NormBias, S5=S1;
+ int S2 = (pBias[4*Line+2])<<NormBias, S6=S2;
+ int S3 = (pBias[4*Line+3])<<NormBias, S7=S3;
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ S1 = gap_sumdotp4(V0, C1, S1); S5 = gap_sumdotp4(V1, C1, S5);
+ S2 = gap_sumdotp4(V0, C2, S2); S6 = gap_sumdotp4(V1, C2, S6);
+ S3 = gap_sumdotp4(V0, C3, S3); S7 = gap_sumdotp4(V1, C3, S7);
+ pIn+=4; pIn1+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S0 += V0*C0; S4 += V1*C0;
+ S1 += V0*C1; S5 += V1*C1;
+ S2 += V0*C2; S6 += V1*C2;
+ S3 += V0*C3; S7 += V1*C3;
+ pIn++; pIn1++; pC0++; pC1++; pC2++; pC3++;
+ }
+ v4s R1 = gap_pack4(AT_CLIP_POS_IMM(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), AT_CLIP_POS_IMM(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ AT_CLIP_POS_IMM(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), AT_CLIP_POS_IMM(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ v4s R2 = gap_pack4(AT_CLIP_POS_IMM(AT_SCALE(S4, pSc[4*Line ], pScN[4*Line ]), 7), AT_CLIP_POS_IMM(AT_SCALE(S5, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ AT_CLIP_POS_IMM(AT_SCALE(S6, pSc[4*Line+2], pScN[4*Line+2]), 7), AT_CLIP_POS_IMM(AT_SCALE(S7, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
+ *((v4s *) (pOut1+4*Line)) = R2;
+ }
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ int S0 = (pBias[i])<<NormBias, S4=S0;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ pIn+=4; pIn1+=4; pC+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC;
+ S0 += V0*C0; S4 += V1*C0;
+ pIn++; pIn1++; pC++;
+ }
+ *(pOut0+i) = AT_CLIP_POS_IMM(AT_SCALE(S0, pSc[i], pScN[i]), 7);
+ *(pOut1+i) = AT_CLIP_POS_IMM(AT_SCALE(S4, pSc[i], pScN[i]), 7);
+ }
+ gap_waitbarrier(0);
+ }
+ if (Wo&0x1) {
+ int c = Wo-1;
+ // for (int c=0; c<Wo; c++) {
for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*FS))[i]=0;
if (Iter1&0x2) ((short int *)(ColBuff+F*FS))[Iter1/2-1]=0;
if (Iter1&0x1) ((signed char *)(ColBuff+F*FS))[Iter1-1]=0;
@@ -1914,15 +2369,16 @@ void KerPar_MM_Conv2D_ReLU_HWC_SQ8(
for (int i=Lb; i<Rb; i++)
((signed char *)(ColBuff+(j+OffL)*InFeat*Fx+(i+OffC)*InFeat+F))[Iter-1] = ((signed char *)(In+j*W*InFeat + i*InFeat+F))[Iter-1];
} else if (Iter>0)
- for (int j=Tb; j<Db; j++)
+ for (int j=Tb; j<Db; j++)
for (int i=Lb; i<Rb; i++)
ColBuff[(j+OffL)*InFeat*Fx+(i+OffC)*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+
PosC += Sx;
gap_waitbarrier(0);
- // for (int Line=First; Line<Last; Line++) {
+
int *pBias = Bias + First;
signed char *pC = Filter + W_In1*First;
- signed char *pOut0 = Out+l*Wo*OutFeat + (c+0)*OutFeat+First;
+ signed char *pOut0 = Out+l*Wo*OutFeat + (c)*OutFeat+First;
unsigned char *pSc = Scale + First;
unsigned char *pScN = ScaleN + First;
for (int Line=0; Line<IterOut/4; Line++) {
@@ -1949,15 +2405,9 @@ void KerPar_MM_Conv2D_ReLU_HWC_SQ8(
S3 += V0*C3;
pIn++; pC0++; pC1++; pC2++; pC3++;
}
- unsigned int Sc, ScN;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = AT_CLIP_POS_IMM(AT_SCALE(S0, Sc, ScN), 7); pOut0++;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = AT_CLIP_POS_IMM(AT_SCALE(S1, Sc, ScN), 7); pOut0++;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = AT_CLIP_POS_IMM(AT_SCALE(S2, Sc, ScN), 7); pOut0++;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = AT_CLIP_POS_IMM(AT_SCALE(S3, Sc, ScN), 7); pOut0++;
+ v4s R1 = gap_pack4(AT_CLIP_POS_IMM(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), AT_CLIP_POS_IMM(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ AT_CLIP_POS_IMM(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), AT_CLIP_POS_IMM(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
}
for (int i=4*(IterOut/4); i<IterOut; i++) {
signed char *pIn = ColBuff;
@@ -1972,15 +2422,12 @@ void KerPar_MM_Conv2D_ReLU_HWC_SQ8(
S0 += V0*C0;
pIn++; pC++;
}
- unsigned int Sc, ScN;
- Sc = *pSc; ScN = *pScN; pSc++; pScN++;
- *pOut0 = AT_CLIP_POS_IMM(AT_SCALE(S0, Sc, ScN), 7); pOut0++;
+ *(pOut0+i) = AT_CLIP_POS_IMM(AT_SCALE(S0, pSc[i], pScN[i]), 7);
}
gap_waitbarrier(0);
}
PosL += Sy;
}
- gap_waitbarrier(0);
}
void Ker_MM_Conv2D_HWC_SQ8(
@@ -2654,7 +3101,9 @@ void KerPar_MM_Conv2D_DxDy_HWC_SQ8(
int FS = Fx*Fy;
int Tail = 2*((W_In1+7)/8);
+ signed char * __restrict__ ColBuff1 = ColBuff + 4*Tail;
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
+ ((int *)ColBuff1)[Tail-1] = 0; ((int *)ColBuff1)[Tail-2] = 0;
int PosL = Arg->FirstTile?(-PadT):0;
int DFx = Dx*Fx, DFy = Dy*Fy;
@@ -2664,44 +3113,132 @@ void KerPar_MM_Conv2D_DxDy_HWC_SQ8(
for (int l=0; l<Ho; l++) {
int PosC = -PadL;
int Tb = PosL, Db = PosL+DFy;
- for (int c=0; c<Wo; c++) {
+ for (int c=0; c<(Wo/2); c++) {
+ for (int i=0; i<(Iter1/4); i++) {
+ ((int *)(ColBuff+F*FS))[i]=0;
+ ((int *)(ColBuff1+F*FS))[i]=0;
+ }
+ if (Iter1&0x2) {
+ ((short int *)(ColBuff+F*FS))[Iter1/2-1]=0;
+ ((short int *)(ColBuff1+F*FS))[Iter1/2-1]=0;
+ }
+ if (Iter1&0x1) {
+ ((signed char *)(ColBuff+F*FS))[Iter1-1]=0;
+ ((signed char *)(ColBuff1+F*FS))[Iter1-1]=0;
+ }
+ int Lb = PosC, Rb = PosC+DFx;
+ int Lb1 = PosC+Sx, Rb1 = PosC+Sx+DFx;
+/*
+This part is more efficient but NOT WORKING ???? TOCHECK
+ if (Iter>=4) {
+ for (int f=0; f<(Iter/4); f++)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ((int *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[f] = ((int *)In+j*W*InFeat + i*InFeat + F)[f];
+ if (Iter&0x2)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ((short int *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter/2-1] = ((short int *)In+j*W*InFeat + i*InFeat + F)[Iter/2-1];
+ if (Iter&0x1)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ((signed char *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter-1] = ((signed char *)In+j*W*InFeat + i*InFeat + F)[Iter-1];
+ } else if (Iter>=2) {
+ if (Iter&0x2)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ((short int *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter/2-1] = ((short int *)In+j*W*InFeat + i*InFeat + F)[Iter/2-1];
+ if (Iter&0x1)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ((signed char *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter-1] = ((signed char *)In+j*W*InFeat + i*InFeat + F)[Iter-1];
+ } else if (Iter>0)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ColBuff[jj*InFeat*Fx+ii*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+*/
+ for (int f=0; f<Iter; f++)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++) {
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ColBuff[jj*InFeat*Fx+ii*InFeat + F+f] = In[j*W*InFeat + i*InFeat + F+f];
+ for (int i=Lb1, ii=0; i<Rb1; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ColBuff1[jj*InFeat*Fx+ii*InFeat + F+f] = In[j*W*InFeat + i*InFeat + F+f];
+ }
+ PosC += 2*Sx;
+ gap_waitbarrier(0);
+
+ int *pBias = Bias + First;
+ signed char *pC = Filter + W_In1*First;
+ signed char *pOut0 = Out+l*Wo*OutFeat + (2*c+0)*OutFeat+First;
+ signed char *pOut1 = Out+l*Wo*OutFeat + (2*c+1)*OutFeat+First;
+ unsigned char *pSc = Scale + First;
+ unsigned char *pScN = ScaleN + First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ signed char *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ int S0 = (pBias[4*Line ])<<NormBias, S4=S0;
+ int S1 = (pBias[4*Line+1])<<NormBias, S5=S1;
+ int S2 = (pBias[4*Line+2])<<NormBias, S6=S2;
+ int S3 = (pBias[4*Line+3])<<NormBias, S7=S3;
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC0), C1 = *((v4s *)pC1), C2 = *((v4s *)pC2), C3 = *((v4s *)pC3);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ S1 = gap_sumdotp4(V0, C1, S1); S5 = gap_sumdotp4(V1, C1, S5);
+ S2 = gap_sumdotp4(V0, C2, S2); S6 = gap_sumdotp4(V1, C2, S6);
+ S3 = gap_sumdotp4(V0, C3, S3); S7 = gap_sumdotp4(V1, C3, S7);
+ pIn+=4; pIn1+=4; pC0+=4; pC1+=4; pC2+=4; pC3+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC0, C1 = *pC1, C2 = *pC2, C3 = *pC3;
+ S0 += V0*C0; S4 += V1*C0;
+ S1 += V0*C1; S5 += V1*C1;
+ S2 += V0*C2; S6 += V1*C2;
+ S3 += V0*C3; S7 += V1*C3;
+ pIn++; pIn1++; pC0++; pC1++; pC2++; pC3++;
+ }
+ v4s R1 = gap_pack4(gap_clip(AT_SCALE(S0, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S1, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S2, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S3, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ v4s R2 = gap_pack4(gap_clip(AT_SCALE(S4, pSc[4*Line ], pScN[4*Line ]), 7), gap_clip(AT_SCALE(S5, pSc[4*Line+1], pScN[4*Line+1]), 7),
+ gap_clip(AT_SCALE(S6, pSc[4*Line+2], pScN[4*Line+2]), 7), gap_clip(AT_SCALE(S7, pSc[4*Line+3], pScN[4*Line+3]), 7));
+ *((v4s *) (pOut0+4*Line)) = R1;
+ *((v4s *) (pOut1+4*Line)) = R2;
+ }
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ signed char *pIn = ColBuff;
+ signed char *pIn1 = ColBuff1;
+ int S0 = (pBias[i])<<NormBias, S4=S0;
+ for (int i=0; i<(W_In1/4); i++) {
+ v4s V0 = *((v4s *)pIn), V1 = *((v4s *)pIn1), C0 = *((v4s *)pC);
+ S0 = gap_sumdotp4(V0, C0, S0); S4 = gap_sumdotp4(V1, C0, S4);
+ pIn+=4; pIn1+=4; pC+=4;
+ }
+ for (int f=4*(W_In1/4); f<W_In1; f++) {
+ int V0 = *pIn, V1 = *pIn1, C0 = *pC;
+ S0 += V0*C0; S4 += V1*C0;
+ pIn++; pIn1++; pC++;
+ }
+ *(pOut0+i) = gap_clip(AT_SCALE(S0, pSc[i], pScN[i]), 7);
+ *(pOut1+i) = gap_clip(AT_SCALE(S4, pSc[i], pScN[i]), 7);
+ }
+ gap_waitbarrier(0);
+ }
+ if (Wo&0x1) {
+ int c = Wo-1;
for (int i=0; i<(Iter1/4); i++) ((int *)(ColBuff+F*FS))[i]=0;
if (Iter1&0x2) ((short int *)(ColBuff+F*FS))[Iter1/2-1]=0;
if (Iter1&0x1) ((signed char *)(ColBuff+F*FS))[Iter1-1]=0;
int Lb = PosC, Rb = PosC+DFx;
- // This part is more efficient but NOT WORKING ???? TOCHECK
- // if (Iter>=4) {
- // for (int f=0; f<(Iter/4); f++)
- // for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
- // for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
- // if (j>=0 && j<H && i>=0 && i<W)
- // ((int *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[f] = ((int *)In+j*W*InFeat + i*InFeat + F)[f];
- // if (Iter&0x2)
- // for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
- // for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
- // if (j>=0 && j<H && i>=0 && i<W)
- // ((short int *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter/2-1] = ((short int *)In+j*W*InFeat + i*InFeat + F)[Iter/2-1];
- // if (Iter&0x1)
- // for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
- // for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
- // if (j>=0 && j<H && i>=0 && i<W)
- // ((signed char *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter-1] = ((signed char *)In+j*W*InFeat + i*InFeat + F)[Iter-1];
- // } else if (Iter>=2) {
- // if (Iter&0x2)
- // for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
- // for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
- // if (j>=0 && j<H && i>=0 && i<W)
- // ((short int *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter/2-1] = ((short int *)In+j*W*InFeat + i*InFeat + F)[Iter/2-1];
- // if (Iter&0x1)
- // for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
- // for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
- // if (j>=0 && j<H && i>=0 && i<W)
- // ((signed char *)ColBuff+jj*InFeat*Fx+ii*InFeat + F)[Iter-1] = ((signed char *)In+j*W*InFeat + i*InFeat + F)[Iter-1];
- // } else if (Iter>0)
- // for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
- // for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
- // if (j>=0 && j<H && i>=0 && i<W)
- // ColBuff[jj*InFeat*Fx+ii*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+
for (int f=0; f<Iter; f++)
for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
@@ -2709,7 +3246,7 @@ void KerPar_MM_Conv2D_DxDy_HWC_SQ8(
ColBuff[jj*InFeat*Fx+ii*InFeat + F+f] = In[j*W*InFeat + i*InFeat + F+f];
PosC += Sx;
gap_waitbarrier(0);
- // for (int Line=First; Line<Last; Line++) {
+
int *pBias = Bias + First;
signed char *pC = Filter + W_In1*First;
signed char *pOut0 = Out+l*Wo*OutFeat + (c+0)*OutFeat+First;
@@ -2851,6 +3388,7 @@ void KerPar_MM_Conv2D_DxDy_ReLU_SQ8(
gap_waitbarrier(0);
}
+#if 0
void KerPar_MM_ConvDW1D_HWC_SQ8(
Ker_MM_Conv_SQ8_T *Arg
)
@@ -2998,4 +3536,5 @@ void KerPar_MM_ConvDW2D_HWC_SQ8(
}
gap_waitbarrier(0);
}
+#endif
#pragma GCC diagnostic pop
diff --git a/tools/autotiler_v3/CNN_Libraries_fp16/CNN_BasicKernels_fp16.h b/tools/autotiler_v3/CNN_Libraries_fp16/CNN_BasicKernels_fp16.h
index e6189463f..bca36c985 100644
--- a/tools/autotiler_v3/CNN_Libraries_fp16/CNN_BasicKernels_fp16.h
+++ b/tools/autotiler_v3/CNN_Libraries_fp16/CNN_BasicKernels_fp16.h
@@ -583,13 +583,13 @@ extern void KerParMatMulLeakyreluSxSy_fp16(KerMatMul_fp16_T *Arg);
extern void KerParMatMulSmallFeat_fp16(KerMatMul_fp16_T *Arg);
extern void KerParMatMulSmallFeatReLU_fp16(KerMatMul_fp16_T *Arg);
extern void KerParMatMulSmallFeatReLUN_fp16(KerMatMul_fp16_T *Arg);
-extern void KerParMatMulSwishSmallFeat_fp16(KerMatMul_fp16_T *Arg);
-extern void KerParMatMulHSwishSmallFeat_fp16(KerMatMul_fp16_T *Arg);
+extern void KerParMatMulSmallFeatSwish_fp16(KerMatMul_fp16_T *Arg);
+extern void KerParMatMulSmallFeatHSwish_fp16(KerMatMul_fp16_T *Arg);
extern void KerParMatMulSigmoidSmallFeat_fp16(KerMatMul_fp16_T *Arg);
-extern void KerParMatMulHSigmoidSmallFeat_fp16(KerMatMul_fp16_T *Arg);
-extern void KerParMatMulTanhSmallFeat_fp16(KerMatMul_fp16_T *Arg);
-extern void KerParMatMulHTanhSmallFeat_fp16(KerMatMul_fp16_T *Arg);
-extern void KerParMatMulLeakyreluSmallFeat_fp16(KerMatMul_fp16_T *Arg);
+extern void KerParMatMulSmallFeatHSigmoid_fp16(KerMatMul_fp16_T *Arg);
+extern void KerParMatMulSmallFeatTanh_fp16(KerMatMul_fp16_T *Arg);
+extern void KerParMatMulSmallFeatHTanh_fp16(KerMatMul_fp16_T *Arg);
+extern void KerParMatMulSmallFeatLeakyrelu_fp16(KerMatMul_fp16_T *Arg);
/******************************************************************************************************************************/
/******************* SOFT MAX *************************************************************************************************/
diff --git a/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatAlgebra_fp16.c b/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatAlgebra_fp16.c
index fa2c466d1..ece4f10f6 100644
--- a/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatAlgebra_fp16.c
+++ b/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatAlgebra_fp16.c
@@ -6799,7 +6799,7 @@ void KerParMatMulSmallFeatReLUN_fp16(KerMatMul_fp16_T *Arg)
gap_waitbarrier(0);
}
-void KerParMatMulSwishSmallFeat_fp16(KerMatMul_fp16_T *Arg)
+void KerParMatMulSmallFeatSwish_fp16(KerMatMul_fp16_T *Arg)
{
F16 * __restrict__ In1 = Arg->In1;
@@ -6833,7 +6833,7 @@ void KerParMatMulSwishSmallFeat_fp16(KerMatMul_fp16_T *Arg)
gap_waitbarrier(0);
}
-void KerParMatMulHSwishSmallFeat_fp16(KerMatMul_fp16_T *Arg)
+void KerParMatMulSmallFeatHSwish_fp16(KerMatMul_fp16_T *Arg)
{
F16 * __restrict__ In1 = Arg->In1;
@@ -6867,7 +6867,7 @@ void KerParMatMulHSwishSmallFeat_fp16(KerMatMul_fp16_T *Arg)
gap_waitbarrier(0);
}
-void KerParMatMulSigmoidSmallFeat_fp16(KerMatMul_fp16_T *Arg)
+void KerParMatMulSmallFeatSigmoid_fp16(KerMatMul_fp16_T *Arg)
{
F16 * __restrict__ In1 = Arg->In1;
@@ -6901,7 +6901,7 @@ void KerParMatMulSigmoidSmallFeat_fp16(KerMatMul_fp16_T *Arg)
gap_waitbarrier(0);
}
-void KerParMatMulHSigmoidSmallFeat_fp16(KerMatMul_fp16_T *Arg)
+void KerParMatMulSmallFeatHSigmoid_fp16(KerMatMul_fp16_T *Arg)
{
F16 * __restrict__ In1 = Arg->In1;
@@ -6936,7 +6936,7 @@ void KerParMatMulHSigmoidSmallFeat_fp16(KerMatMul_fp16_T *Arg)
gap_waitbarrier(0);
}
-void KerParMatMulTanhSmallFeat_fp16(KerMatMul_fp16_T *Arg)
+void KerParMatMulSmallFeatTanh_fp16(KerMatMul_fp16_T *Arg)
{
F16 * __restrict__ In1 = Arg->In1;
@@ -6970,7 +6970,7 @@ void KerParMatMulTanhSmallFeat_fp16(KerMatMul_fp16_T *Arg)
gap_waitbarrier(0);
}
-void KerParMatMulHTanhSmallFeat_fp16(KerMatMul_fp16_T *Arg)
+void KerParMatMulSmallFeatHTanh_fp16(KerMatMul_fp16_T *Arg)
{
F16 * __restrict__ In1 = Arg->In1;
@@ -7004,7 +7004,7 @@ void KerParMatMulHTanhSmallFeat_fp16(KerMatMul_fp16_T *Arg)
gap_waitbarrier(0);
}
-void KerParMatMulLeakyreluSmallFeat_fp16(KerMatMul_fp16_T *Arg)
+void KerParMatMulSmallFeatLeakyrelu_fp16(KerMatMul_fp16_T *Arg)
{
F16 * __restrict__ In1 = Arg->In1;
diff --git a/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatMul_Conv_fp16.c b/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatMul_Conv_fp16.c
index f5137a37f..5af1deb3a 100644
--- a/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatMul_Conv_fp16.c
+++ b/tools/autotiler_v3/CNN_Libraries_fp16/CNN_MatMul_Conv_fp16.c
@@ -17,7 +17,7 @@ static inline unsigned int __attribute__((always_inline)) ChunkSize(unsigned int
return Chunk;
}
-static int FirstDefinedOutput(unsigned int F, unsigned int Pad, unsigned int Stride)
+static int FirstDefinedOutput(int F, int Pad, int Stride)
{
// k*S - (F-1)/2 >=0 => k >= (((F-1)/2) + S-1)/S
@@ -25,12 +25,12 @@ static int FirstDefinedOutput(unsigned int F, unsigned int Pad, unsigned int Str
return ((Pad+Stride-1)/Stride);
}
-static int LastDefinedOutput(unsigned int DimIn, unsigned int F, unsigned int PadL, unsigned int Stride)
+static int LastDefinedOutput(int DimIn, int F, int PadL, int Stride, int D)
{
// k*S + ((F-1)/2 - PadL + F/2) < Dim => k < (Dim-((F-1)/2 - PadL + (F/2)) + S-1)/S
- return ((DimIn - ((F-1)/2 - PadL + (F/2)) + Stride-1)/Stride);
+ return ((DimIn - ((F-1)/2*D - PadL + (F/2)*D) + Stride-1)/Stride);
}
void KerPar_MM_Conv1D_fp16(
@@ -266,6 +266,7 @@ void KerPar_MM_Conv1D_HWC_fp16(
int PosL = 0;
int Iter = L-F;
int Iter1 = Iter*Fx;
+ int IterOut = Last - First;
for (int l=0; l<Ho; l++) {
int PosC = -PadL;
for (int c=0; c<Wo; c++) {
@@ -280,17 +281,53 @@ void KerPar_MM_Conv1D_HWC_fp16(
} else for (int i=Lb; i<Rb; i++) ColBuff[(i+Off)*InFeat + F] = In[PosL*W*InFeat + i*InFeat + F];
PosC += Sx;
gap_waitbarrier(0);
- for (int Line=First; Line<Last; Line++) {
- F16V *VIn1 = (F16V *) (&Filter[Line*W_In1 + 0]);
- F16V S0 = (F16V){Bias[Line],0.0}, S1 = (F16V) 0;
- for (int i=0; i<((W_In1+3)/4); i++) {
- F16V V0 = VIn1[2*i], V1 = VIn1[2*i+1];
- F16V C0 = VBuff[2*i], C1 = VBuff[2*i+1];
- S0 += V0 * C0;
- S1 += V1 * C1;
+
+ // for (int Line=First; Line<Last; Line++) {
+ F16 *pBias = Bias + First;
+ F16 *pC = Filter + W_In1*First;
+ F16 *pOut0 = Out+l*Wo*OutFeat + (c+0)*OutFeat+First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ F16 *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ F16V S0 = (F16V){*pBias,0.0}; pBias++;
+ F16V S1 = (F16V){*pBias,0.0}; pBias++;
+ F16V S2 = (F16V){*pBias,0.0}; pBias++;
+ F16V S3 = (F16V){*pBias,0.0}; pBias++;
+ F16 *pIn = ColBuff;
+ for (int i=0; i<(W_In1/2); i++) {
+ F16V V0 = *((F16V *)pIn), C0 = *((F16V *)pC0), C1 = *((F16V *)pC1), C2 = *((F16V *)pC2), C3 = *((F16V *)pC3);
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ pIn+=2; pC0+=2; pC1+=2; pC2+=2; pC3+=2;
}
- Out[l*Wo*OutFeat + c*OutFeat + Line] = S0[0]+S0[1]+S1[0]+S1[1];
+ if (W_In1&0x1) {
+ F16V V0 = (F16V){*pIn,0.0}, C0 = *((F16V *)pC0), C1 = *((F16V *)pC1), C2 = *((F16V *)pC2), C3 = *((F16V *)pC3);
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ }
+ *pOut0 = S0[0]+S0[1]; pOut0++;
+ *pOut0 = S1[0]+S1[1]; pOut0++;
+ *pOut0 = S2[0]+S2[1]; pOut0++;
+ *pOut0 = S3[0]+S3[1]; pOut0++;
}
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ F16 *pIn = ColBuff;
+ F16V S0 = (F16V){*pBias,0.0}; pBias++;
+ for (int i=0; i<(W_In1/2); i++) {
+ F16V V0 = *((F16V *)pIn), C0 = *((F16V *)pC);
+ S0 += V0*C0;
+ pIn+=2; pC+=2;
+ }
+ if (W_In1&0x1) {
+ F16V V0 = (F16V){*pIn,0.0}, C0 = *((F16V *)pC);
+ S0 += V0*C0;
+ }
+ *pOut0 = S0[0]+S0[1]; pOut0++;
+ }
gap_waitbarrier(0);
}
PosL += Sy;
@@ -395,44 +432,80 @@ void KerPar_MM_Conv1D_DxDy_HWC_fp16(
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
int DFx = Dx*(Fx-1)+1;
- int Prec=10;
- int InvDx = ((1<<Prec)+Dx-1)/Dx;
int PosL = 0;
int Iter = L-F;
int Iter1 = Iter*Fx;
+ int IterOut = Last - First;
for (int l=0; l<Ho; l++) {
int PosC = -PadL;
for (int c=0; c<Wo; c++) {
for (int i=0; i<(Iter1/2); i++) ((int *)(ColBuff+F*Fx))[i]=0;
if (Iter1&0x1) ((short int *)(ColBuff+F*Fx))[Iter1-1]=0;
- int Lb = Max(PosC, 0), Rb = Min(PosC+DFx, W);
- int OffBuffX = Max(0, gap_mulsN(-PosC+Dx-1, InvDx, Prec));
- int OffInX = OffBuffX?(Dx*OffBuffX+PosC):0;
- int IterX = gap_mulsN(Rb-Lb-1, InvDx, Prec) + 1;
+ int Lb = PosC, Rb = PosC+DFx;
if (Iter>=2) {
for (int f=0; f<(Iter/2); f++)
- for (int i=0; i<IterX; i++)
- ((int *)(ColBuff+(i+OffBuffX)*InFeat+F))[f] = ((int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[f];
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (i>=0 && i<W)
+ ((int *) (ColBuff + ii*InFeat + F))[f] = ((int *)(In+i*InFeat + F))[f];
if (Iter&0x1)
- for (int i=0; i<IterX; i++)
- ((short int *)(ColBuff+(i+OffBuffX)*InFeat+F))[Iter-1] = ((short int *)(In+PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat+F))[Iter-1];
- } else
- for (int i=0; i<IterX; i++)
- ColBuff[(i+OffBuffX)*InFeat + F] = In[PosL*W*InFeat + (i*Dx+OffInX+Lb)*InFeat + F];
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (i>=0 && i<W)
+ ((short int *) (ColBuff + ii*InFeat + F))[Iter-1] = ((short int *)(In+i*InFeat + F))[Iter-1];
+ } else {
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (i>=0 && i<W)
+ ColBuff[ii*InFeat + F] = In[i*InFeat + F];
+ }
PosC += Sx;
gap_waitbarrier(0);
- for (int Line=First; Line<Last; Line++) {
- F16V *VIn1 = (F16V *) (&Filter[Line*W_In1 + 0]);
- F16V S0 = (F16V){Bias[Line],0.0}, S1 = (F16V) 0;
- for (int i=0; i<((W_In1+3)/4); i++) {
- F16V V0 = VIn1[2*i], V1 = VIn1[2*i+1];
- F16V C0 = VBuff[2*i], C1 = VBuff[2*i+1];
- S0 += V0 * C0;
- S1 += V1 * C1;
+
+ // for (int Line=First; Line<Last; Line++) {
+ F16 *pBias = Bias + First;
+ F16 *pC = Filter + W_In1*First;
+ F16 *pOut0 = Out+l*Wo*OutFeat + (c+0)*OutFeat+First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ F16 *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ F16V S0 = (F16V){*pBias,0.0}; pBias++;
+ F16V S1 = (F16V){*pBias,0.0}; pBias++;
+ F16V S2 = (F16V){*pBias,0.0}; pBias++;
+ F16V S3 = (F16V){*pBias,0.0}; pBias++;
+ F16 *pIn = ColBuff;
+ for (int i=0; i<(W_In1/2); i++) {
+ F16V V0 = *((F16V *)pIn), C0 = *((F16V *)pC0), C1 = *((F16V *)pC1), C2 = *((F16V *)pC2), C3 = *((F16V *)pC3);
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ pIn+=2; pC0+=2; pC1+=2; pC2+=2; pC3+=2;
}
- Out[l*Wo*OutFeat + c*OutFeat + Line] = S0[0]+S0[1]+S1[0]+S1[1];
+ if (W_In1&0x1) {
+ F16V V0 = (F16V){*pIn,0.0}, C0 = *((F16V *)pC0), C1 = *((F16V *)pC1), C2 = *((F16V *)pC2), C3 = *((F16V *)pC3);
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ }
+ *pOut0 = S0[0]+S0[1]; pOut0++;
+ *pOut0 = S1[0]+S1[1]; pOut0++;
+ *pOut0 = S2[0]+S2[1]; pOut0++;
+ *pOut0 = S3[0]+S3[1]; pOut0++;
}
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ F16 *pIn = ColBuff;
+ F16V S0 = (F16V){*pBias,0.0}; pBias++;
+ for (int i=0; i<(W_In1/2); i++) {
+ F16V V0 = *((F16V *)pIn), C0 = *((F16V *)pC);
+ S0 += V0*C0;
+ pIn+=2; pC+=2;
+ }
+ if (W_In1&0x1) {
+ F16V V0 = (F16V){*pIn,0.0}, C0 = *((F16V *)pC);
+ S0 += V0*C0;
+ }
+ *pOut0 = S0[0]+S0[1]; pOut0++;
+ }
gap_waitbarrier(0);
}
PosL += Sy;
@@ -1156,55 +1229,86 @@ void KerPar_MM_Conv2D_DxDy_HWC_fp16(
((int *)ColBuff)[Tail-1] = 0; ((int *)ColBuff)[Tail-2] = 0;
int PosL = Arg->FirstTile?(-PadT):0;
int DFx = Dx*(Fx-1)+1, DFy = Dy*(Fy-1)+1;
- int Prec=10;
- int InvDx = ((1<<Prec)+Dx-1)/Dx;
- int InvDy = ((1<<Prec)+Dy-1)/Dy;
+ // int Prec=10;
+ // int InvDx = ((1<<Prec)+Dx-1)/Dx;
+ // int InvDy = ((1<<Prec)+Dy-1)/Dy;
int Iter = L-F;
int Iter1 = Iter*FS;
+ int IterOut = Last - First;
for (int l=0; l<Ho; l++) {
int PosC = -PadL;
- int Tb = Max(PosL, 0), Db = Min(PosL+DFy, H);
- int OffLBuffY = Max(0, gap_mulsN(-PosL+Dy-1, InvDy, Prec));
- int OffLInY = OffLBuffY?(Dy*OffLBuffY+PosL):0;
+ int Tb = PosL, Db = PosL+DFy;
for (int c=0; c<Wo; c++) {
for (int i=0; i<(Iter1/2); i++) ((int *)(ColBuff+F*FS))[i]=0;
if (Iter1&0x1) ((short int *)(ColBuff+F*FS))[Iter1-1]=0;
- int Lb = Max(PosC, 0), Rb = Min(PosC+DFx, W);
- int OffCBuffX = (Lb==0)?Max(0, gap_mulsN(-PosC+Dx-1, InvDx, Prec)):0;
- int OffCInX = OffCBuffX?(Dx*OffCBuffX+PosC):0;
- int IterY = gap_mulsN(Db-Tb-1, InvDy, Prec) + 1;
- int IterX = gap_mulsN(Rb-Lb-1, InvDx, Prec) + 1;
- if (Iter>=2) {
- for (int f=0; f<(Iter/2); f++)
- for (int j=Tb; j<Db; j++)
- for (int i=Lb; i<Rb; i++)
- ((int *)(ColBuff+(j+OffLBuffY)*InFeat*Fx+(i+OffCBuffX)*InFeat+F))[f] =
- ((int *)(In+(Tb+j*Dy+OffLInY)*W*InFeat + (Lb+i*Dx+OffCInX)*InFeat+F))[f];
- if (Iter&0x1)
- for (int j=Tb; j<Db; j++)
- for (int i=Lb; i<Rb; i++)
- ((short int *)(ColBuff+(j+OffLBuffY)*InFeat*Fx+(i+OffCBuffX)*InFeat+F))[Iter-1] =
- ((short int *)(In+(Tb+j*Dy+OffLInY)*W*InFeat + (Lb+i*Dx+OffCInX)*InFeat+F))[Iter-1];
- } else
- for (int j=Tb; j<Db; j++)
- for (int i=Lb; i<Rb; i++)
- ColBuff[(j+OffLBuffY)*InFeat*Fx+(i+OffCBuffX)*InFeat + F] =
- In[(Tb+j*Dy+OffLInY)*W*InFeat + (Lb+i*Dx+OffCInX)*InFeat + F];
-
+ int Lb = PosC, Rb = PosC+DFx;
+ if (Iter>=2) {
+ for (int f=0; f<(Iter/2); f++)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ((int *) (ColBuff + jj*InFeat*Fx+ii*InFeat + F))[f] = ((int *)(In+j*W*InFeat + i*InFeat + F))[f];
+ if (Iter&0x1)
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ((short int *) (ColBuff + jj*InFeat*Fx+ii*InFeat + F))[Iter-1] = ((short int *)(In+j*W*InFeat + i*InFeat + F))[Iter-1];
+ } else {
+ for (int j=Tb, jj=0; j<Db; j+=Dy, jj++)
+ for (int i=Lb, ii=0; i<Rb; i+=Dx, ii++)
+ if (j>=0 && j<H && i>=0 && i<W)
+ ColBuff[jj*InFeat*Fx+ii*InFeat + F] = In[j*W*InFeat + i*InFeat + F];
+ }
PosC += Sx;
gap_waitbarrier(0);
- for (int Line=First; Line<Last; Line++) {
- F16V *VIn1 = (F16V *) (&Filter[Line*W_In1 + 0]);
- F16V S0 = (F16V){Bias[Line],0.0};
- for (int i=0; i<((W_In1+3)/4); i++) {
- F16V V0 = VIn1[2*i], V1 = VIn1[2*i+1];
- F16V C0 = VBuff[2*i], C1 = VBuff[2*i+1];
- S0 += V0 * C0;
- S0 += V1 * C1;
+
+ // for (int Line=First; Line<Last; Line++) {
+ F16 *pBias = Bias + First;
+ F16 *pC = Filter + W_In1*First;
+ F16 *pOut0 = Out+l*Wo*OutFeat + (c+0)*OutFeat+First;
+ for (int Line=0; Line<IterOut/4; Line++) {
+ F16 *pC0 = pC, *pC1 = pC0+W_In1, *pC2 = pC1+W_In1, *pC3 = pC2+W_In1;
+ pC=pC3+W_In1;
+ F16V S0 = (F16V){*pBias,0.0}; pBias++;
+ F16V S1 = (F16V){*pBias,0.0}; pBias++;
+ F16V S2 = (F16V){*pBias,0.0}; pBias++;
+ F16V S3 = (F16V){*pBias,0.0}; pBias++;
+ F16 *pIn = ColBuff;
+ for (int i=0; i<(W_In1/2); i++) {
+ F16V V0 = *((F16V *)pIn), C0 = *((F16V *)pC0), C1 = *((F16V *)pC1), C2 = *((F16V *)pC2), C3 = *((F16V *)pC3);
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ pIn+=2; pC0+=2; pC1+=2; pC2+=2; pC3+=2;
}
- Out[l*Wo*OutFeat + c*OutFeat + Line] = S0[0]+S0[1];
+ if (W_In1&0x1) {
+ F16V V0 = (F16V){*pIn,0.0}, C0 = *((F16V *)pC0), C1 = *((F16V *)pC1), C2 = *((F16V *)pC2), C3 = *((F16V *)pC3);
+ S0 += V0*C0;
+ S1 += V0*C1;
+ S2 += V0*C2;
+ S3 += V0*C3;
+ }
+ *pOut0 = S0[0]+S0[1]; pOut0++;
+ *pOut0 = S1[0]+S1[1]; pOut0++;
+ *pOut0 = S2[0]+S2[1]; pOut0++;
+ *pOut0 = S3[0]+S3[1]; pOut0++;
}
+ for (int i=4*(IterOut/4); i<IterOut; i++) {
+ F16 *pIn = ColBuff;
+ F16V S0 = (F16V){*pBias,0.0}; pBias++;
+ for (int i=0; i<(W_In1/2); i++) {
+ F16V V0 = *((F16V *)pIn), C0 = *((F16V *)pC);
+ S0 += V0*C0;
+ pIn+=2; pC+=2;
+ }
+ if (W_In1&0x1) {
+ F16V V0 = (F16V){*pIn,0.0}, C0 = *((F16V *)pC);
+ S0 += V0*C0;
+ }
+ *pOut0 = S0[0]+S0[1]; pOut0++;
+ }
gap_waitbarrier(0);
}
PosL += Sy;
diff --git a/tools/autotiler_v3/DSP_Libraries/FloatDefines.h b/tools/autotiler_v3/DSP_Libraries/FloatDefines.h
index 46b19828f..27317e9ad 100644
--- a/tools/autotiler_v3/DSP_Libraries/FloatDefines.h
+++ b/tools/autotiler_v3/DSP_Libraries/FloatDefines.h
@@ -110,5 +110,6 @@
#define Cvt_v2h_v2ah(a) (v2ah) (a)
#define Cvt_v2ah_v2h(a) (v2h) (a)
#endif //__gap9__
+#define Clipf32(a, upper, lower) ((float) Maxf32(Minf32((a), (upper)), (lower)))
#endif
diff --git a/tools/autotiler_v3/DSP_Libraries/LUT_Tables/gen_scripts/GenMFCCLUT.py b/tools/autotiler_v3/DSP_Libraries/LUT_Tables/gen_scripts/GenMFCCLUT.py
index 6581ab8ef..947e2d625 100644
--- a/tools/autotiler_v3/DSP_Libraries/LUT_Tables/gen_scripts/GenMFCCLUT.py
+++ b/tools/autotiler_v3/DSP_Libraries/LUT_Tables/gen_scripts/GenMFCCLUT.py
@@ -31,7 +31,7 @@ def create_parser():
parser.add_argument('--name_suffix', default="", type=str)
parser.add_argument('--frame_size', required="--params_json" not in sys.argv, type=int,
help="size in number of samples of one frame")
- parser.add_argument('--frame_step', required="--params_json" not in sys.argv, type=int,
+ parser.add_argument('--frame_step', type=int, default=0,
help="step in number of samples between two consecutive frames")
parser.add_argument('--win_func', default="hanning", type=str,
help="numpy window function (e.g. hanning)")
diff --git a/tools/autotiler_v3/Emulation/GapBuiltins.h b/tools/autotiler_v3/Emulation/GapBuiltins.h
index 0df6a71cf..03f22c167 100644
--- a/tools/autotiler_v3/Emulation/GapBuiltins.h
+++ b/tools/autotiler_v3/Emulation/GapBuiltins.h
@@ -230,6 +230,16 @@ static inline unsigned int ExtInsMaskSafe(unsigned int Size, unsigned int Offset
#define gap_norm(x, scale) __builtin_pulp_addN((x), 0, (scale))
#define gap_norm_reg(x, scale) __builtin_pulp_addN_r((x), 0, (scale))
+/* Floating point */
+#define gap_f32min(a,b) __builtin_pulp_f32min((a), (b))
+#define gap_f32max(a,b) __builtin_pulp_f32max((a), (b))
+#define gap_f32sqrt(a) __builtin_pulp_f32sqrt((a))
+#define gap_f32abs(a) __builtin_pulp_f32abs((a))
+#define gap_f32rmm(a) __builtin_pulp_rintsf2((a))
+#define gap_f32rdn(a) __builtin_pulp_rdownsf2((a))
+#define gap_f32rup(a) __builtin_pulp_rupsf2((a))
+
+
#else
/* Emulation */
@@ -452,6 +462,17 @@ static int _VitT0_Flag, _VitT1_Flag;
#define gap_norm(x, scale) ((int)(x)>>(scale))
#define gap_norm_reg(x, scale) ((int)(x)>>(scale))
+/* Floating point */
+#include <math.h>
+#define gap_f32min(a,b) fminf((a), (b))
+#define gap_f32max(a,b) fmaxf((a), (b))
+#define gap_f32sqrt(a) sqrtf((a))
+#define gap_f32abs(a) fabsf((a))
+#define gap_f32rmm(a) (((a)>=0)?((int)((a)+0.5f)):(-((int)(-(a)+0.5f))))
+#define gap_f32rdn(a) floorf((a))
+#define gap_f32rup(a) ceilf((a))
+
+
#endif
#define FIX2FP(Val, Precision) ((float) (Val) / (float) (1<<(Precision)))
diff --git a/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.c b/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.c
index ccbbbd07f..a78820daa 100644
--- a/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.c
+++ b/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.c
@@ -281,3 +281,88 @@ void KerResizeNearestNeighborSigned_Q16(KerResizeSigned16_ArgT *Arg)
}
gap_waitbarrier(0);
}
+
+#ifdef __gap9__
+void KerResizeBilinear_fp16(KerResize_fp16_ArgT *Arg)
+
+{
+ F16 * __restrict__ In = Arg->In;
+ unsigned int Win = Arg->Win;
+ unsigned int Hin = Arg->Hin;
+ F16 * __restrict__ Out = Arg->Out;
+ unsigned int Wout = Arg->Wout;
+ unsigned int Hout = Arg->Hout;
+ unsigned int HTileOut = Arg->HTileOut;
+ unsigned int FirstLineIndex = Arg->FirstLineIndex;
+
+ unsigned int CoreId = gap_coreid();
+ unsigned int ChunkCell = ChunkSize(Wout);
+ unsigned int First = CoreId*ChunkCell, Last = Min(Wout, First+ChunkCell);
+
+ F16 WStep = ((F16) (Win-1))/Wout;
+ F16 HStep = ((F16) (Hin-1))/Hout;
+
+ unsigned int x, y;
+ F16 hCoeff = ((F16) HStep)*FirstLineIndex;
+ F16 BaseY = hCoeff;
+ for (y = 0 ; y < HTileOut ; y++) {
+ int offsetY = (int) (hCoeff - BaseY);
+ F16 hc2 = hCoeff;
+ F16 hc1 = 1.0 - hc2;
+ // unsigned int wCoeff = 0;
+ F16 wCoeff = First*WStep;
+
+ // for (x = 0 ; x < Wout ; x++) {
+ for (x = First ; x < Last ; x++) {
+ int offsetX = (int) wCoeff;
+ F16 wc2 = wCoeff;
+ F16 wc1 = 1.0 - wc2;
+ F16 P1 = In[offsetY*Win + offsetX ];
+ F16 P2 = In[(offsetY + 1)*Win + offsetX ];
+ F16 P3 = In[offsetY*Win + offsetX + 1];
+ F16 P4 = In[(offsetY + 1)*Win + offsetX + 1];
+
+ Out[y*Wout + x] = ((P1*hc1 + P2*hc2)*wc1 + (P3*hc1 + P4*hc2)*wc2);
+ wCoeff += WStep;
+ }
+ hCoeff += HStep;
+ }
+ gap_waitbarrier(0);
+}
+
+void KerResizeNearestNeighbor_fp16(KerResize_fp16_ArgT *Arg)
+
+{
+ F16 * __restrict__ In = Arg->In;
+ unsigned int Win = Arg->Win;
+ unsigned int Hin = Arg->Hin;
+ F16 * __restrict__ Out = Arg->Out;
+ unsigned int Wout = Arg->Wout;
+ unsigned int Hout = Arg->Hout;
+ unsigned int HTileOut = Arg->HTileOut;
+ unsigned int FirstLineIndex = Arg->FirstLineIndex;
+
+ unsigned int CoreId = gap_coreid();
+ unsigned int ChunkCell = ChunkSize(Wout);
+ unsigned int First = CoreId*ChunkCell, Last = Min(Wout, First+ChunkCell);
+
+ F16 WStep = ((F16) (Win-1))/(Wout-1);
+ F16 HStep = ((F16) (Hin-1))/(Hout-1);
+
+ unsigned int x, y;
+ F16 hCoeff = HStep*FirstLineIndex;
+ F16 BaseY = hCoeff;
+ for (y = 0 ; y < HTileOut ; y++) {
+ int h_rounded = (int) (hCoeff - BaseY + 0.5);
+ F16 wCoeff = First*WStep;
+ for (x = First ; x < Last ; x++) {
+ int w_rounded = (int) (wCoeff + 0.5);
+
+ Out[y*Wout + x] = In[h_rounded*Win + w_rounded];
+ wCoeff += WStep;
+ }
+ hCoeff += HStep;
+ }
+ gap_waitbarrier(0);
+}
+#endif
diff --git a/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.h b/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.h
index fd3216f29..6d1406d4f 100644
--- a/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.h
+++ b/tools/autotiler_v3/Generators/BilinearResizes/ResizeBasicKernels.h
@@ -12,6 +12,11 @@
#include "Gap.h"
+#ifdef __gap9__
+#include "CNN_Defines_fp16.h"
+#include "CNN_FloatType.h"
+#endif
+
#ifndef Max
#define Max(a, b) (((a)>(b))?(a):(b))
#endif
@@ -74,6 +79,22 @@ typedef struct {
unsigned int FirstLineIndex;
} KerResizeSigned16_ArgT;
+#ifdef __gap9__
+typedef struct {
+ F16 * __restrict__ In;
+ unsigned int Win;
+ unsigned int Hin;
+ F16 * __restrict__ Out;
+ unsigned int Wout;
+ unsigned int Hout;
+ unsigned int HTileOut;
+ unsigned int FirstLineIndex;
+} KerResize_fp16_ArgT;
+
+void KerResizeBilinear_fp16(KerResize_fp16_ArgT *Arg);
+void KerResizeNearestNeighbor_fp16(KerResize_fp16_ArgT *Arg);
+#endif
+
void KerResizeBilinear(KerResizeBilinear_ArgT *Arg);
void KerResizeNearestNeighbor(KerResizeNearestNeighbor_ArgT *Arg);
void KerResizeBilinearSigned(KerResizeBilinearSigned_ArgT *Arg);
diff --git a/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.c b/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.c
index f74baea96..5133a2148 100644
--- a/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.c
+++ b/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.c
@@ -92,6 +92,32 @@ void LoadResizeLibrary()
"KerResizeSigned16_ArgT",
NULL
);
+ LibKernel("KerResizeNearestNeighbor_fp16", CALL_PARALLEL,
+ CArgs(8,
+ TCArg("F16 * __restrict__", "In"),
+ TCArg("unsigned int", "Win"),
+ TCArg("unsigned int", "Hin"),
+ TCArg("F16 * __restrict__", "Out"),
+ TCArg("unsigned int", "Wout"),
+ TCArg("unsigned int", "Hout"),
+ TCArg("unsigned int", "HTileOut"),
+ TCArg("unsigned int", "FirstLineIndex")),
+ "KerResize_fp16_ArgT",
+ NULL
+ );
+ LibKernel("KerResizeBilinear_fp16", CALL_PARALLEL,
+ CArgs(8,
+ TCArg("F16 * __restrict__", "In"),
+ TCArg("unsigned int", "Win"),
+ TCArg("unsigned int", "Hin"),
+ TCArg("F16 * __restrict__", "Out"),
+ TCArg("unsigned int", "Wout"),
+ TCArg("unsigned int", "Hout"),
+ TCArg("unsigned int", "HTileOut"),
+ TCArg("unsigned int", "FirstLineIndex")),
+ "KerResize_fp16_ArgT",
+ NULL
+ );
}
int GenerateResizeMultiChannel(char *Name, unsigned int Win, unsigned int Hin, unsigned int Wout, unsigned int Hout, unsigned int Channels, InOut_t InOut_Type, resize_kop_t Type)
@@ -186,6 +212,51 @@ int GenerateResizeMultiChannelQ16(char *Name, unsigned int Win, unsigned int Hin
}
+int GenerateResizeMultiChannel_fp16(char *Name, unsigned int Win, unsigned int Hin, unsigned int Wout, unsigned int Hout, unsigned int Channels, InOut_t InOut_Type, resize_kop_t Type)
+
+{
+ char *ResizeKerName;
+ switch (Type){
+ case KOP_BILINEAR_RESIZE:
+ ResizeKerName = "KerResizeBilinear_fp16";
+ break;
+ case KOP_NEAREST_NEIGHBOR_RESIZE:
+ ResizeKerName = "KerResizeNearestNeighbor_fp16";
+ break;
+ default:
+ ResizeKerName = "KerResizeBilinear_fp16";
+ }
+ printf("%s\n", ResizeKerName);
+ int LayerOp = Channels * Wout * Hout * (3 + 6 + 3);
+ int LayerBandwidth = Channels * Win * Hin + Channels * Hout * Wout;
+ Kernel_T *Kernel = UserKernel(Name,
+ KernelIterSpace(2, IterFixedSpace(KER_ITER_D0, Channels), IterTiledSpace(KER_ITER_TILE0)),
+ (Hin==1)?TILE_VER:TILE_HOR,
+ CArgs(2, TCArg("F16 *", "In"), TCArg("F16 *", "Out")),
+ Calls(1, Call(ResizeKerName, LOC_LOOP,
+ Bindings(8, K_Arg("In", KER_ARG_TILE),
+ K_Arg("In", KER_ARG_W),
+ K_Arg("In", KER_ARG_H),
+ K_Arg("Out", KER_ARG_TILE),
+ K_Arg("Out", KER_ARG_W),
+ K_Arg("Out", KER_ARG_H),
+ K_Arg("Out", KER_ARG_TILE_H),
+ K_Arg("In", KER_ARG_TILE_BASE)))),
+ KerArgs(2,
+ KerArg("In" , KerArgSpace(2,KER_ITER_D0,KER_ITER_TILE0), OBJ_IN_DB, Win, Hin, sizeof(short), 1, OBJ_CONSTRAINTS_DYNAMIC, 0, "In"),
+ KerArg("Out", KerArgSpace(2,KER_ITER_D0,KER_ITER_TILE0), OBJ_OUT_DB, Wout, Hout, sizeof(short), 0, OBJ_CONSTRAINTS_DYNAMIC, 0, "Out")
+ )
+ );
+ if (Kernel) {
+ AddKernelInfos(Name, AT_KERINFO_OPER, LayerOp, 0);
+ AddKernelInfos(Name, AT_KERINFO_BANDWIDTH, LayerBandwidth, 0);
+
+ AddKernelFloatArgDim(Name, "In", 4, Channels, Hin, Win, 2);
+ AddKernelFloatArgDim(Name, "Out", 4, Channels, Hout, Wout, 2);
+ }
+ return (Kernel!=0);
+}
+
void ResizeConfiguration(unsigned int L1Memory)
{
diff --git a/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.h b/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.h
index 83e8b7539..901623ab0 100644
--- a/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.h
+++ b/tools/autotiler_v3/Generators/BilinearResizes/ResizeGenerator.h
@@ -57,6 +57,7 @@ Generate Resizing for a given set of parameters
*/
int GenerateResizeMultiChannel(char *Name, unsigned int Win, unsigned int Hin, unsigned int Wout, unsigned int Hout, unsigned int Channels, InOut_t InOut_Type, resize_kop_t Type);
int GenerateResizeMultiChannelQ16(char *Name, unsigned int Win, unsigned int Hin, unsigned int Wout, unsigned int Hout, unsigned int Channels, InOut_t InOut_Type, resize_kop_t Type);
+int GenerateResizeMultiChannel_fp16(char *Name, unsigned int Win, unsigned int Hin, unsigned int Wout, unsigned int Hout, unsigned int Channels, InOut_t InOut_Type, resize_kop_t Type);
#define GenerateResizeNew(Name, Win, Hin, Wout, Hout, Type) GenerateResizeMultiChannel(Name, Win, Hin, Wout, Hout, 1, UNSIGNED_INOUT, Type)
#define GenerateResize(Name, Win, Hin, Wout, Hout) GenerateResizeNew(Name, Win, Hin, Wout, Hout, KOP_BILINEAR_RESIZE)
diff --git a/tools/autotiler_v3/Makefile b/tools/autotiler_v3/Makefile
index 25a7ce5dd..c6f308357 100644
--- a/tools/autotiler_v3/Makefile
+++ b/tools/autotiler_v3/Makefile
@@ -1,4 +1,4 @@
-TILER_VER=4.2.0
+TILER_VER=4.1.0
export TILER_LIB=libtile.${TILER_VER}.a
ifdef GAP_SDK_HOME
export TILER_URL=$(GAP_SDK_HOME)/.tiler_url
diff --git a/tools/autotiler_v3/version.cfg b/tools/autotiler_v3/version.cfg
index 22bf7a2cf..dbce13460 100644
--- a/tools/autotiler_v3/version.cfg
+++ b/tools/autotiler_v3/version.cfg
@@ -3,7 +3,7 @@
{
"version": "autotiler-v3",
"magicNum": 718930176,
- "git-hash": "29855c78d247eb5b3b3b0758b10dfc08306d06c9"
+ "git-hash": "bd6fd2447d2ccd4dcdd8c774a3a130b3141fac22"
}
]
}
\ No newline at end of file
diff --git a/tools/jenkins/gap_sdk_version.txt b/tools/jenkins/gap_sdk_version.txt
index 642d57e4b..1f8776c61 100644
--- a/tools/jenkins/gap_sdk_version.txt
+++ b/tools/jenkins/gap_sdk_version.txt
@@ -1 +1 @@
-c6603c58b94f3ac80ade4370c2b3c4e558f246f2
+0addfcd309a48d5a5dbd3e058de0cf1ffc660a2b
diff --git a/tools/nntool/.gitignore b/tools/nntool/.gitignore
index bd00d96a3..45c264225 100644
--- a/tools/nntool/.gitignore
+++ b/tools/nntool/.gitignore
@@ -16,3 +16,6 @@ importer/tflite
.env
gsync
tempdir/
+notebooks/
+utils_nocheckin/
+graphs_nocheckin/
diff --git a/tools/nntool/generation/code_block.py b/tools/nntool/generation/code_block.py
index b072c79bd..e615c6844 100644
--- a/tools/nntool/generation/code_block.py
+++ b/tools/nntool/generation/code_block.py
@@ -62,7 +62,11 @@ def write_start(self, fmt, *args):
def comment(self, fmt, *args):
fmt = self.get_indent() + '// ' + fmt
- self._lines.append(fmt.format(*args))
+ if args:
+ self._lines.append(fmt.format(*args))
+ else:
+ self._lines.append(fmt)
+
return self
def __str__(self):
diff --git a/tools/nntool/generation/generators/kernels/general/resizer_kernel_generator.py b/tools/nntool/generation/generators/kernels/general/resizer_kernel_generator.py
index c22bb788f..439282e95 100644
--- a/tools/nntool/generation/generators/kernels/general/resizer_kernel_generator.py
+++ b/tools/nntool/generation/generators/kernels/general/resizer_kernel_generator.py
@@ -26,7 +26,7 @@
"nearest_neighbor": "KOP_NEAREST_NEIGHBOR_RESIZE"}
-def gen_at_resizer(code_block, name, in_dim, new_shape, inout_t, resize_kop, q16mode):
+def gen_at_resizer(code_block, name, in_dim, new_shape, inout_t, resize_kop, q16mode, fp16):
if in_dim.has_key('w') and in_dim.has_key('h'):
in_dim_w, in_dim_h, in_dim_c = in_dim.w, in_dim.h, in_dim.c
else:
@@ -47,7 +47,10 @@ def gen_at_resizer(code_block, name, in_dim, new_shape, inout_t, resize_kop, q16
# If both HW change from 1x1 to HxW this is not going to work
LOG.warning(f"Resize Node {name} has 1x1xc input but resizes both HW dimension, could not work in autotiler")
- GenKernel = "GenerateResizeMultiChannelQ16" if q16mode else "GenerateResizeMultiChannel"
+ if fp16:
+ GenKernel = "GenerateResizeMultiChannel_fp16"
+ else:
+ GenKernel = "GenerateResizeMultiChannelQ16" if q16mode else "GenerateResizeMultiChannel"
code_block.write('{}("{}", {}, {}, {}, {}, {}, {}, {});', GenKernel,
name, win, hin, wout, hout, chin, inout_t, resize_kop)
@@ -67,7 +70,8 @@ def __init__(self, cname, params, qrec):
self.inout_type = "SIGNED_INOUT" if qrec.in_qs[0].signed else "UNSIGNED_INOUT"
self.type = params.op_name
self.new_shape = params.new_shape
- self.q16 = True if qrec.in_qs[0].dtype_bits == 16 else False
+ self.q16 = qrec.in_qs[0].dtype_bits == 16
+ self.fp16 = qrec.in_qs[0].is_floating
def code(self, code_block=None):
if code_block is None:
@@ -76,5 +80,6 @@ def code(self, code_block=None):
code_block.comment("generator for {}", self.node_name)
gen_at_resizer(code_block, self.cname, self.in_dim,
- self.new_shape, self.inout_type, RESIZE_KOP[self.type], self.q16)
+ self.new_shape, self.inout_type, RESIZE_KOP[self.type],
+ self.q16, self.fp16)
return code_block
diff --git a/tools/nntool/generation/generators/kernels/pow2/conv_pool_relu_kernels_generator.py b/tools/nntool/generation/generators/kernels/pow2/conv_pool_relu_kernels_generator.py
index e52a9064d..6ef573390 100644
--- a/tools/nntool/generation/generators/kernels/pow2/conv_pool_relu_kernels_generator.py
+++ b/tools/nntool/generation/generators/kernels/pow2/conv_pool_relu_kernels_generator.py
@@ -236,7 +236,7 @@ def __init__(self, node_name, cname, conv_params, conv_q,
"only homogenious operations are supported at present")
LOG.debug("%s: pool relu inq %s outq %s control block",
node_name, in_q, out_q)
- if at_pool_params.PoolOper == 'KOP_NONE' and not in_dim.is_named and in_dim.has_keys(['c', 'w', 'h']):
+ if at_pool_params.PoolOper == 'KOP_NONE' and (not in_dim.is_named or not in_dim.has_keys(['c', 'w', 'h'])):
in_shape = in_dim.shape + ([1] * (3 - len(in_dim.shape)))
in_c, in_h, in_w = in_shape[0], in_shape[1], in_shape[2]
else:
diff --git a/tools/nntool/generation/new_generators/general/dsp_generators.py b/tools/nntool/generation/new_generators/general/dsp_generators.py
index f500cb61a..3186e8073 100644
--- a/tools/nntool/generation/new_generators/general/dsp_generators.py
+++ b/tools/nntool/generation/new_generators/general/dsp_generators.py
@@ -169,7 +169,7 @@ def __init__(self, cname, params, qrec, gen_ctrl=None):
'frame_stride': params.frame_step,
'n_fft': params.n_fft,
'n_melbanks': params.n_fbanks,
- 'size_mel_coeff': params.get_melfilter_size()[1],
+ 'size_mel_coeff': params.get_melfilter_size()[0],
'n_dct': params.n_dct,
'preemp_factor': params.preemp_factor,
'no_window': int(params.win_fn is None),
diff --git a/tools/nntool/generation/new_generators/general/quantize_parameters.py b/tools/nntool/generation/new_generators/general/quantize_parameters.py
index 8ed6b6134..d24c7c37d 100644
--- a/tools/nntool/generation/new_generators/general/quantize_parameters.py
+++ b/tools/nntool/generation/new_generators/general/quantize_parameters.py
@@ -52,14 +52,12 @@ def cache_values(cls, node, qrec):
qrec.cache['float_conversion'] = float_conversion = in_q.is_floating or out_q.is_floating
qrec.cache['bit_conversion'] = bit_conversion = in_q.bits != out_q.bits
if float_conversion:
- if in_q.dtype in [bfloat16, np.float16]:
+ if in_q.is_floating:
qrec.cache['kernel_type'] = 'KOP_CONVERT_FL_FP'
- qrec.cache['in_at_size'] = 2
- qrec.cache['out_at_size'] = at_bits(out_q)
else:
qrec.cache['kernel_type'] = 'KOP_CONVERT_FP_FL'
- qrec.cache['in_at_size'] = at_bits(in_q)
- qrec.cache['out_at_size'] = 2
+ qrec.cache['in_at_size'] = at_bits(in_q)
+ qrec.cache['out_at_size'] = at_bits(out_q)
return True
else:
qrec.cache['in_at_size'] = at_bits(in_q)
@@ -107,7 +105,7 @@ def globals_generator(cls, gen, node, qrec, pnode, fnode) -> bool:
else:
offset = (int(math.pow(2, bits)) - in_q.zero_point[0] +
out_q.zero_point[0]).astype(out_q.dtype)
- contents = np.array(list(offset.tobytes()) + ([0] * 6), dtype=np.uint8)
+ contents = np.array(list(offset.tobytes()) + ([0] * 7), dtype=np.uint8)
elif qrec.cache['kernel_type'] == 'KOP_CONVERT_FP_FP':
# no infos needed
return True
@@ -133,15 +131,15 @@ def globals_generator(cls, gen, node, qrec, pnode, fnode) -> bool:
qbias = list(scale_adjust.qbiases.tobytes())
qbias = qbias + [0] * (2 - len(qbias))
qnorm = list(scale_adjust.qnorms.tobytes())
- contents = np.array(zero_adjust + qbias + qnorm, dtype=np.int8)
+ contents = np.array(zero_adjust + qbias + qnorm + [0], dtype=np.int8)
elif qrec.cache['kernel_type'] == 'KOP_CONVERT_FL_FP':
- qbias = list((1/out_q.scale).astype(in_q.dtype).tobytes())
- zero_adjust = list((out_q.zero_point.astype(np.int32) * out_q.scale).astype(in_q.dtype).tobytes())
- contents = np.array(zero_adjust + [0, 0] + qbias + [0], dtype=np.int8)
+ qbias = list((1/out_q.scale).astype(np.float32).tobytes())
+ zero_adjust = list((out_q.zero_point.astype(np.int32) * out_q.scale).astype(np.float32).tobytes())
+ contents = np.array(zero_adjust + qbias, dtype=np.int8)
elif qrec.cache['kernel_type'] == 'KOP_CONVERT_FP_FL':
- qbias = list((in_q.scale).astype(out_q.dtype).tobytes())
- zero_adjust = list((-in_q.zero_point.astype(np.int32)).astype(out_q.dtype).tobytes())
- contents = np.array(zero_adjust + [0, 0] + qbias + [0], dtype=np.int8)
+ qbias = list((in_q.scale).astype(np.float32).tobytes())
+ zero_adjust = list((-in_q.zero_point.astype(np.int32)).astype(np.float32).tobytes())
+ contents = np.array(zero_adjust + qbias, dtype=np.int8)
else:
raise ValueError(f"strange dtype change in {pnode.name}")
cname, file_name = gen_constant(gen, pnode, pnode, INFOS)
diff --git a/tools/nntool/generation/new_generators/mult8/matmul_mult8.py b/tools/nntool/generation/new_generators/mult8/matmul_mult8.py
index 02bbaae2f..c09c847db 100644
--- a/tools/nntool/generation/new_generators/mult8/matmul_mult8.py
+++ b/tools/nntool/generation/new_generators/mult8/matmul_mult8.py
@@ -136,13 +136,15 @@ def set_matmul_bindings(gen, in_eparams, out_eparams, cname, node, node_q, out_q
"Node {} inq {} outq {}", node.name,
str(node_q.in_qs[0]), str(out_q.out_qs[0]))
)
+ idx_0 = 0 if node_q.cache.get("ne16") else 1
+ idx_1 = 1 if node_q.cache.get("ne16") else 0
if len(node.in_dims) == 3:
if len(node_q.in_qs[1].scale) > 1:
gen.bindings.append(
NodeBindingList(
cname,
- GNodeArgEdge(in_eparams[1]),
- GNodeArgEdge(in_eparams[0]),
+ GNodeArgEdge(in_eparams[idx_0]),
+ GNodeArgEdge(in_eparams[idx_1]),
GNodeArgEdge(in_eparams[2]),
GNodeArgEdge(out_eparams[0], "GNA_OUT"),
GNodeArgNode(node, MULSCALE),
@@ -152,8 +154,8 @@ def set_matmul_bindings(gen, in_eparams, out_eparams, cname, node, node_q, out_q
gen.bindings.append(
NodeBindingList(
cname,
- GNodeArgEdge(in_eparams[1]),
- GNodeArgEdge(in_eparams[0]),
+ GNodeArgEdge(in_eparams[idx_0]),
+ GNodeArgEdge(in_eparams[idx_1]),
GNodeArgEdge(in_eparams[2]),
GNodeArgEdge(out_eparams[0], "GNA_OUT"),
GNodeArgNode(node, INFOS)))
@@ -162,8 +164,8 @@ def set_matmul_bindings(gen, in_eparams, out_eparams, cname, node, node_q, out_q
gen.bindings.append(
NodeBindingList(
cname,
- GNodeArgEdge(in_eparams[1]),
- GNodeArgEdge(in_eparams[0]),
+ GNodeArgEdge(in_eparams[idx_0]),
+ GNodeArgEdge(in_eparams[idx_1]),
GNodeArgEdge(out_eparams[0], "GNA_OUT"),
GNodeArgNode(node, MULSCALE),
GNodeArgNode(node, MULSHIFT),
@@ -172,8 +174,8 @@ def set_matmul_bindings(gen, in_eparams, out_eparams, cname, node, node_q, out_q
gen.bindings.append(
NodeBindingList(
cname,
- GNodeArgEdge(in_eparams[1]),
- GNodeArgEdge(in_eparams[0]),
+ GNodeArgEdge(in_eparams[idx_0]),
+ GNodeArgEdge(in_eparams[idx_1]),
GNodeArgEdge(out_eparams[0], "GNA_OUT"),
GNodeArgNode(node, INFOS)))
diff --git a/tools/nntool/graph/dim.py b/tools/nntool/graph/dim.py
index d2962636f..c1a2fd5e4 100644
--- a/tools/nntool/graph/dim.py
+++ b/tools/nntool/graph/dim.py
@@ -64,6 +64,8 @@ class MoreThanOneInputError(DimError):
class Dim():
def __init__(self, shape=None, names=None, is_ordered=False, is_unknown=False):
set_shape = shape if shape is not None else [] if names is None else [None] * len(names)
+ if any(dim is not None and dim < 0 for dim in set_shape):
+ raise ValueError('invalid dim')
super().__setattr__('_shape', set_shape)
super().__setattr__('_names', names)
super().__setattr__('_is_ordered', is_ordered)
@@ -486,6 +488,8 @@ def _do_operation(self, other, op) -> 'Dim':
res = self.clone()
for k in self.keys:
setattr(res, k, op(getattr(res, k), other))
+ if any(dim is not None and dim < 0 for dim in self.shape):
+ raise ValueError('invalid dim')
return res
if isinstance(other, Dim):
@@ -500,6 +504,8 @@ def _do_operation(self, other, op) -> 'Dim':
setattr(res, k, getattr(other, k))
elif getattr(other, k) is not None:
setattr(res, k, op(getattr(self, k), getattr(other, k)))
+ if any(dim is not None and dim < 0 for dim in self.shape):
+ raise ValueError('invalid dim')
return res
raise TypeError("Inapropriate types for operation")
diff --git a/tools/nntool/graph/manipulations/adjust_order.py b/tools/nntool/graph/manipulations/adjust_order.py
index eba46fffa..4067c3397 100644
--- a/tools/nntool/graph/manipulations/adjust_order.py
+++ b/tools/nntool/graph/manipulations/adjust_order.py
@@ -26,7 +26,7 @@
LOG = logging.getLogger("nntool." + __name__)
-def adjust_order(G, reshape_weights=True, postprocess=True, debug_function=None, one_cycle=False):
+def adjust_order(G, reshape_weights=True, postprocess=True, debug_function=None, one_cycle=False, single_step=False):
opts = {'reshape_weights': reshape_weights}
selector = AdjusterBase.get_all_handlers(opts)
LOG.info("adding transposes to correct tensor order for AT kernels")
@@ -55,5 +55,5 @@ def adjust_order(G, reshape_weights=True, postprocess=True, debug_function=None,
if debug_function:
debug_function(G)
if postprocess:
- eliminate_transposes(G, debug_function=debug_function, one_cycle=one_cycle)
+ eliminate_transposes(G, debug_function=debug_function, one_cycle=one_cycle, single_step=single_step)
add_dimensions(G)
diff --git a/tools/nntool/graph/manipulations/dimensions.py b/tools/nntool/graph/manipulations/dimensions.py
index 6ebaf0b71..f883d4d02 100644
--- a/tools/nntool/graph/manipulations/dimensions.py
+++ b/tools/nntool/graph/manipulations/dimensions.py
@@ -33,7 +33,7 @@ def set_out_edges_multi(G, node: Parameters, dims: Sequence[Dim], step_idx: int,
naming_convension: NamingConvension, edge_type: str = "in_out"):
# clone the dims first so that the edge dims are the same objects as the node output dims
dims = node.set_output_size(dims)
- out_edges = G.indexed_out_edges(node.name)
+ out_edges = G.indexed_out_edges(node)
is_multi_out = len(out_edges) > 1
for edge_idx, edge_group in enumerate(out_edges):
if not edge_group:
diff --git a/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py b/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py
index f77cfcf64..3be578e8d 100644
--- a/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py
+++ b/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py
@@ -13,6 +13,7 @@
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
+from functools import reduce
import logging
from collections.abc import MutableSet
from copy import deepcopy
@@ -24,8 +25,9 @@
InputParameters, LinearFusionParameters,
OutputParameters, PadParameters, ReshapeParameters,
ReverseParameters, StridedSliceParameters,
- TransposeParameters)
+ TransposeParameters, ActivationParameters)
from graph.types.base import NNEdge, SensitiveToOrder
+from graph.types.others import CopyParameters, UnaryOpParameters
from graph.types.tensor_arithmetic import Broadcastable
from utils.compatible_transposes import (find_all_compatible_transposes,
find_combination)
@@ -58,6 +60,10 @@ def info(msg):
LOG.info(msg)
+def debug(msg):
+ LOG.debug(msg)
+
+
TRANSIENT_ACTIONS = {
PadParameters: TransposePad,
ReverseParameters: TransposeReverse,
@@ -76,31 +82,23 @@ class CantContinueError(Exception):
class TransposeHistory():
- def __init__(self, node, from_shape, from_transpose, to_shape=None, to_transpose=None) -> None:
- # can be initialized with to_shape set to none in which case from == to
- # to_transpose == None means that the transpose has been eliminated by a reshape
+ def __init__(self, node, from_shape=None, transpose=None, to_shape=None) -> None:
self.node = node
- self._from = (from_shape, from_transpose)
- if to_shape is None:
- self._to = (from_shape, from_transpose)
- else:
- self._to = (to_shape, to_transpose)
+ self._from = from_shape
+ self._transpose = transpose
+ self._to = to_shape
@property
def from_shape(self):
- return self._from[0]
+ return self._from
@property
def to_shape(self):
- return self._to[0]
+ return self._to
@property
- def from_transpose(self):
- return self._from[1]
-
- @property
- def to_transpose(self):
- return self._to[1]
+ def transpose(self):
+ return self._transpose
class VisitedNodes(MutableSet):
@@ -215,9 +213,26 @@ def none_or_idx(trans, idx):
return None if trans[idx] is None else idx
+def reverse_broadcast(old_shape, new_shape, transpose):
+ old_shape_idx = new_shape_idx = 0
+ res_pos = {}
+ while old_shape_idx < len(old_shape) or new_shape_idx < len(new_shape):
+ if old_shape_idx < len(old_shape) and old_shape[old_shape_idx] == new_shape[new_shape_idx]:
+ res_pos[old_shape_idx] = new_shape_idx
+ old_shape_idx += 1
+ new_shape_idx += 1
+ elif new_shape_idx < len(new_shape) and new_shape[new_shape_idx] == 1:
+ new_shape_idx += 1
+ else:
+ raise ValueError(
+ f'reverse broadcast not possible between {old_shape} and {new_shape}')
+ return tuple([res_pos[idx] for idx in transpose] + [idx for idx, _ in enumerate(new_shape) if idx not in res_pos.values()])
+
+
def requires_reshape(trans1, trans2, dim):
"""Checks if layout shape doesn't change but a reshape is necessary due to 1 position"""
- if tuple(dim.shape) != tuple(dim.layout_shape) and dim.layout_shape == dim.calc_transpose(trans1).calc_transpose(trans2).layout_shape:
+ if (tuple(dim.shape) != tuple(dim.layout_shape) and
+ dim.layout_shape == dim.calc_transpose(trans1).calc_transpose(trans2).layout_shape):
from_shape = dim.calc_transpose(trans1)
to_shape = dim.calc_transpose(trans2)
if from_shape.shape != to_shape.shape:
@@ -302,7 +317,26 @@ def compute_max_shape(dims):
return [max(dims) for dims in zip(*shapes)]
-def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, transpose_history: Sequence[TransposeHistory]):
+def broadcasted_axes(shape, full_shape):
+ return tuple(range(len(full_shape) - len(shape)))
+
+
+def reduce_dimension(axis, stripped_axes):
+ """reduces axis by the number of stripped axes that are less than it"""
+ return axis - sum(1 if stripped_axis < axis else 0 for stripped_axis in stripped_axes)
+
+
+def expand_axes_in_transpose(transpose, num_new_axes):
+ """increases axis by the number of new axes"""
+ return tuple(list(range(num_new_axes)) + [axis + num_new_axes for axis in transpose])
+
+
+def strip_axes_from_transpose(transpose, stripped_axes):
+ return tuple(reduce_dimension(axis, stripped_axes) for axis in transpose if axis not in stripped_axes)
+
+
+def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge,
+ transpose_history: Sequence[TransposeHistory]):
"""Searches down the graph for something that eliminates transpose
Args:
@@ -311,7 +345,8 @@ def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, tr
visited_nodes : Nodes already traversed
in_edge : The edge we are arriving on at this node
transpose_history : A history of the reshapes passed that did not allow us to determine the transpose
- transpose : The current transpose being propagated. Can be None to indicate that we cannot translate the transpose via that reshape
+ transpose : The current transpose being propagated. Can be None to indicate that we cannot translate
+ the transpose via that reshape
Returns:
A tuple of a list of actions and a list of nodes traversed
@@ -319,15 +354,16 @@ def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, tr
cur_visited_nodes = VisitedNodes()
cur_visited_nodes.visit_down(node, in_edge.to_idx)
- transpose = transpose_history[-1].to_transpose
+ transpose = transpose_history[-1].transpose
in_shape = node.in_dims[in_edge.to_idx].shape
+ debug(f'down at {node.name} trans {transpose} shape {in_shape}')
if transpose is not None and len(transpose) == 1:
return [EndActionDown(node)], []
- if isinstance(node, SensitiveToOrder) and transpose_does_nothing(transpose, in_shape):
+ if isinstance(node, SensitiveToOrder) and transpose_does_nothing(reverse_transpose(transpose), in_shape):
check_for_null_transpose(node, transpose)
new_shape = apply_transpose(
- in_shape, transpose)
+ in_shape, reverse_transpose(transpose))
# could be that the transpose does nothing to the data layout but still changes the positions of
# the ones in the shape
if new_shape == in_shape:
@@ -343,30 +379,30 @@ def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, tr
check_for_null_transpose(node, transpose)
info(
f'rejected {node.name} - sensitive to order - inserting transpose {transpose}')
- return [InsertTransposeAction(node, direction='in', idx=in_edge.to_idx, transpose=transpose), EndActionDown(node)], cur_visited_nodes
+ return ([InsertTransposeAction(node, direction='in', idx=in_edge.to_idx, transpose=transpose),
+ EndActionDown(node)], cur_visited_nodes)
cur_actions = []
- # if arriiving on a broadcasted input the transpose needs to be expanded and a reshape inserted
- # TODO - Expand to handle multiple outputs on expressions
- if isinstance(node, Broadcastable) and node.in_dims[in_edge.to_idx].rank != node.out_dims[0].rank:
+ # if arriving on a broadcasted input the transpose needs to be expanded
+ # since the transpose is only acting on the broadcasted dimensions no reshape is necessary
+ if isinstance(node, Broadcastable) and len(in_shape) != node.out_dims[0].rank:
check_for_null_transpose(node, transpose)
+ # This could be an expression so need to broadcaset the output
max_shape = compute_max_shape(node.out_dims)
- in_shape, new_shape, new_transpose = broadcast_expand(
- max_shape, in_shape, transpose)
- from_shape = apply_transpose(in_shape, reverse_transpose(transpose))
- to_shape = apply_transpose(new_shape, reverse_transpose(new_transpose))
- info(f'{node.name} broadcasted input {in_edge.to_idx} requires reshape {from_shape}->{to_shape}')
- cur_actions.append(
- InsertReshapeAction(
- node, direction='in', idx=in_edge.to_idx,
- in_shape=from_shape,
- out_shape=to_shape))
+ b_axes = broadcasted_axes(in_shape, max_shape)
+
+ new_transpose = expand_axes_in_transpose(transpose, len(b_axes))
+ new_shape = tuple(([1] * len(b_axes)) + in_shape)
+ transpose_history += [
+ TransposeHistory(node, in_shape,
+ new_transpose,
+ new_shape)
+ ]
transpose = new_transpose
in_shape = new_shape
# on nodes where inputs must have the same transpose applied i.e. Binary arithmetic ops etc
- # TODO - Expand to handle multiple outputs on expressions
if any(isinstance(node, cls) for cls in NODES_TO_EXPLORE_UP):
check_for_null_transpose(node, transpose)
max_shape = compute_max_shape(node.out_dims)
@@ -374,24 +410,39 @@ def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, tr
if check_continue(visited_nodes, cur_visited_nodes, exclude_nodes, edge.from_node, 'up', edge.from_idx):
continue
# if other edges are broadcasted they may need to be reshaped
- edge_in_shape, new_shape, new_transpose = broadcast_reduce(
- max_shape, node.in_dims[edge.to_idx].shape, transpose)
+ edge_in_shape = node.in_dims[edge.to_idx].shape
+ # different rank so broadcasted
+ if len(edge_in_shape) != len(max_shape):
+ # strip the broadcasted axis from the transpose
+ b_axes = broadcasted_axes(edge_in_shape, max_shape)
+ transpose_without_broadcast = strip_axes_from_transpose(
+ reverse_transpose(transpose), b_axes)
+ # from shape will be the old shape with the unbroadcasted transpose
+ from_shape = apply_transpose(
+ edge_in_shape, transpose_without_broadcast)
+ # to shape is the broadcasted input shape with the transpose with the leading ones removed
+ broadcasted_shape = ([1] * len(b_axes)) + list(edge_in_shape)
+ to_shape = strip_leading_ones(apply_transpose(
+ broadcasted_shape, reverse_transpose(transpose)), len(from_shape))
+ # if they are not equal insert a reshape
+ if from_shape != to_shape:
+ info(
+ f'{node.name} broadcasted input {edge.to_idx} requires reshape {from_shape}->{to_shape}')
+ cur_actions.append(
+ InsertReshapeAction(
+ node, direction='in', idx=edge.to_idx,
+ in_shape=from_shape,
+ out_shape=to_shape
+ ))
+ new_transpose = transpose_without_broadcast
+ else:
+ new_transpose = transpose
+
new_history = [
- TransposeHistory(node, apply_transpose(
- node.in_dims[edge.to_idx].shape, new_transpose), reverse_transpose(new_transpose))
+ TransposeHistory(node, edge_in_shape,
+ new_transpose, edge_in_shape)
]
- from_shape = apply_transpose(
- edge_in_shape, reverse_transpose(new_transpose))
- to_shape = apply_transpose(new_shape, reverse_transpose(transpose))
- if from_shape != to_shape:
- info(
- f'{node.name} broadcasted input {edge.to_idx} requires reshape {from_shape}->{to_shape}')
- cur_actions.append(
- InsertReshapeAction(
- node, direction='in', idx=edge.to_idx,
- in_shape=from_shape,
- out_shape=strip_leading_ones(to_shape, len(from_shape))
- ))
+
new_actions, visited_up_nodes = search_up(
G, edge.from_node, exclude_nodes, visited_nodes | cur_visited_nodes, edge, new_history)
cur_visited_nodes |= visited_up_nodes
@@ -457,11 +508,12 @@ def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, tr
if identity_transpose(new_transpose):
return cur_actions + [EndActionDown(node)], cur_visited_nodes
- from_shape = do_transpose(reverse_transpose(transpose), node.slice_shape) if transpose is not None else None
+ from_shape = do_transpose(reverse_transpose(
+ transpose), node.slice_shape) if transpose is not None else None
transpose_history = transpose_history + \
[TransposeHistory(node, node.slice_shape,
- transpose, from_shape, new_transpose)]
+ new_transpose, node.out_shape)]
transpose = new_transpose
elif node.__class__ in TRANSIENT_ACTIONS:
check_for_null_transpose(node, transpose)
@@ -475,26 +527,46 @@ def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, tr
# the order wrong in this case
old_transpose = get_reshape_transpose(
node.old_shape.shape, node.shape.shape)
- new_transpose = apply_transpose(transpose, old_transpose)
+ if reverses_transpose(transpose, old_transpose):
+ cur_actions += [
+ DeleteReshapeAction(
+ node
+ )
+ ]
+ return cur_actions + [
+ DeleteReshapeAction(
+ node
+ ),
+ EndActionDown(node)], cur_visited_nodes
+ new_transpose = apply_transpose(
+ transpose, old_transpose)
+ info(
+ f"pass reshape that is transpose {node.name} down trans: old {transpose} new {new_transpose} shape: old {node.old_shape} new {node.shape}")
+ # insert an action to rewrite the reshape shapes
+ from_shape = apply_transpose(
+ node.old_shape.shape, reverse_transpose(transpose))
+ to_shape = apply_transpose(
+ node.shape.shape, reverse_transpose(transpose))
else:
# the transpose that we are actually applying is the reverse of the transpose that we are propagating down
# So we reverse the transpose before evaluating the reshape and then reverse the result
new_transpose = reverse_transpose(reverse_reshape(
reverse_transpose(transpose), node.old_shape, node.shape))
-
- info(
- f"pass reshape {node.name} down trans: old {transpose} new {new_transpose} shape: old {node.old_shape} new {node.shape}")
- if new_transpose is None and len(node.shape) > 1:
info(
- f"rejected {node.name} - transpose out - does not reverse - inserting transpose {transpose}")
- return [InsertTransposeAction(node, direction='in', idx=in_edge.to_idx, transpose=transpose), EndActionDown(node)], cur_visited_nodes
- # insert an action to rewrite the reshape shapes
- from_shape = node.old_shape.calc_transpose(
- reverse_transpose(transpose)) if transpose is not None else None
- to_shape = node.shape.calc_transpose(reverse_transpose(
- new_transpose)) if new_transpose is not None else None
+ f"pass reshape {node.name} down trans: old {transpose} new {new_transpose} shape: old {node.old_shape} new {node.shape}")
+
+ if new_transpose is None and len(node.shape) > 1:
+ info(
+ f"rejected {node.name} - transpose out - does not reverse - inserting transpose {transpose}")
+ return [InsertTransposeAction(node, direction='in', idx=in_edge.to_idx, transpose=transpose), EndActionDown(node)], cur_visited_nodes
+
+ # insert an action to rewrite the reshape shapes
+ from_shape = apply_transpose(node.old_shape.shape,
+ reverse_transpose(transpose)) if transpose is not None else None
+ to_shape = apply_transpose(node.shape.shape, reverse_transpose(
+ new_transpose)) if new_transpose is not None else None
info(f"rewrite reshape to {from_shape}->{to_shape}")
- if from_shape is None or to_shape is None or from_shape.shape != to_shape.shape:
+ if from_shape is None or to_shape is None or from_shape != to_shape:
cur_actions += [
SetReshapeAction(
node,
@@ -514,7 +586,7 @@ def search_down(G, node, exclude_nodes, visited_nodes: VisitedNodes, in_edge, tr
transpose_history = transpose_history + \
[TransposeHistory(node, node.old_shape.shape,
- transpose, from_shape, new_transpose)]
+ new_transpose, node.shape.shape)]
if new_transpose is None:
try:
@@ -542,19 +614,27 @@ def continue_down(G, node, exclude_nodes, visited_nodes, cur_visited_nodes, cur_
def search_up(G, node, exclude_nodes, visited_nodes, out_edge, transpose_history):
- transpose = transpose_history[-1].to_transpose
+ transpose = transpose_history[-1].transpose
+ debug(f'up at {node.name} trans {transpose}')
cur_visited_nodes = VisitedNodes()
cur_visited_nodes.visit_up(node, out_edge.from_idx)
if transpose is not None and len(transpose) == 1:
+ info(
+ f'accepted {node.name} - single dimension transpose')
return [EndActionUp(node)], cur_visited_nodes
- if isinstance(node, SensitiveToOrder) and transpose_does_nothing(transpose, node.out_dims[out_edge.from_idx].shape):
+ if isinstance(node, SensitiveToOrder) and transpose_does_nothing(reverse_transpose(transpose), node.out_dims[out_edge.from_idx].shape):
new_shape = apply_transpose(
- node.out_dims[out_edge.from_idx].shape, transpose)
+ node.out_dims[out_edge.from_idx].shape, reverse_transpose(transpose))
# could be that the transpose does nothing to the data layout but still changes the positions of
# the ones in the shape
if new_shape == node.out_dims[out_edge.from_idx].shape:
+ info(
+ f'accepted {node.name} - transpose does nothing')
return [EndActionUp(node)], cur_visited_nodes
+ info(
+ f'accepted {node.name} - transpose does nothing with reshape '
+ f'{node.out_dims[out_edge.from_idx].shape} -> {new_shape}')
return [
InsertReshapeAction(node, direction='out', idx=out_edge.from_idx, out_edge=out_edge,
in_shape=node.out_dims[out_edge.from_idx].shape, out_shape=new_shape),
@@ -564,7 +644,7 @@ def search_up(G, node, exclude_nodes, visited_nodes, out_edge, transpose_history
check_for_null_transpose(node, transpose)
info(
f'rejected {node.name} - sensitive to order - inserting transpose {transpose}')
- return [InsertTransposeAction(node, direction='out', idx=out_edge.from_idx, out_edge=out_edge, transpose=transpose), EndActionUp(node)], cur_visited_nodes
+ return [InsertTransposeAction(node, direction='out', idx=out_edge.from_idx, out_edge=out_edge, transpose=reverse_transpose(transpose)), EndActionUp(node)], cur_visited_nodes
cur_actions = []
@@ -585,7 +665,7 @@ def search_up(G, node, exclude_nodes, visited_nodes, out_edge, transpose_history
exclude_nodes,
visited_nodes | cur_visited_nodes,
edge,
- [TransposeHistory(node, apply_transpose(node.out_dims[edge.from_idx], transpose), reverse_transpose(transpose))])
+ [TransposeHistory(node, node.out_dims[edge.from_idx], transpose, apply_transpose(node.out_dims[edge.from_idx], transpose))])
cur_visited_nodes |= visited_down_nodes
cur_actions += new_actions
@@ -610,9 +690,9 @@ def search_up(G, node, exclude_nodes, visited_nodes, out_edge, transpose_history
if isinstance(node, TransposeParameters):
check_for_null_transpose(node, transpose)
# TODO - in_dims or out_dims - 99% sure in_dims
- if reverses_transpose(node.transpose, transpose, node.in_dims[0]):
+ if tuple(node.transpose) == tuple(transpose):
info(
- f"accepted {node.name} - transpose {node.transpose} reversed out by {transpose} on {node.in_dims[0]}")
+ f"accepted {node.name} - transpose {node.transpose} equals {transpose} on {node.in_dims[0]}")
reshape = requires_reshape(
node.transpose, transpose, node.out_dims[0])
if reshape:
@@ -629,32 +709,46 @@ def search_up(G, node, exclude_nodes, visited_nodes, out_edge, transpose_history
check_for_null_transpose(node, transpose)
if node.fixed_order:
info(f"rejected {node.name} - fixed order input")
- return [InsertTransposeAction(node, direction='out', idx=out_edge.from_idx, out_edge=out_edge, transpose=reverse_transpose(transpose)), EndActionUp(node)], cur_visited_nodes
+ return [InsertTransposeAction(node, direction='out', idx=out_edge.from_idx, out_edge=out_edge, transpose=transpose), EndActionUp(node)], cur_visited_nodes
info(
- f"accepted {node.name} - input without fixed order - transpose input {transpose}")
- return cur_actions + [ReorderInputDims.from_history(node, transpose_history, transpose=transpose), EndActionUp(node)], cur_visited_nodes
+ f"accepted {node.name} - input without fixed order - transpose input {reverse_transpose(transpose)}")
+ return cur_actions + [ReorderInputDims.from_history(node, transpose_history, transpose=reverse_transpose(transpose)), EndActionUp(node)], cur_visited_nodes
# Constant can be reordered
if isinstance(node, ConstantInputParameters):
check_for_null_transpose(node, transpose)
info(
f"accepted {node.name} - constant input - transpose constant {transpose}")
- return cur_actions + [ReorderConstantInput.from_history(node, transpose_history, transpose=transpose), EndActionUp(node)], cur_visited_nodes
+ return cur_actions + [ReorderConstantInput.from_history(node, transpose_history, transpose=reverse_transpose(transpose)), EndActionUp(node)], cur_visited_nodes
# Conditions that can pass through the Transpose
- if node.__class__ in TRANSIENT_ACTIONS:
+ if isinstance(node, StridedSliceParameters) and node.changes_shape:
+ reversed_below = reverse_transpose(transpose)
+ reversed_above = reverse_broadcast(
+ node.out_shape, node.post_slice_shape, reversed_below)
+ new_transpose = reverse_transpose(reversed_above)
+ transpose_history = transpose_history + \
+ [TransposeHistory(node, node.out_shape,
+ new_transpose, node.post_slice_shape)]
cur_actions.append(
- TRANSIENT_ACTIONS[node.__class__](node, transpose, "up"))
+ TransposeSlidedSlice(node, reversed_above, "up", transpose))
+ transpose = new_transpose
+ elif node.__class__ in TRANSIENT_ACTIONS:
+ check_for_null_transpose(node, transpose)
+
+ cur_actions.append(
+ TRANSIENT_ACTIONS[node.__class__](node, reverse_transpose(transpose), "up"))
elif isinstance(node, ReshapeParameters):
check_for_null_transpose(node, transpose) # TODO - may eliminate
- new_transpose = reverse_reshape(transpose, node.shape, node.old_shape)
+ new_transpose = reverse_reshape(reverse_transpose(
+ transpose), node.shape, node.old_shape)
# if the upwards shape has one dimension we keep going since we want to find
# nodes such as a linear layer that can reorder their output filters
# This could be extended to recurrent layers for the inner dimension
info(
- f"pass reshape {node.name} up trans: old {transpose} new {new_transpose} shape: old {node.old_shape} new {node.shape}")
+ f"pass reshape {node.name} up trans: old {transpose} new {new_transpose} shape: {node.old_shape} -> {node.shape}")
if new_transpose is None and len(node.old_shape) > 1:
info(f"rejected {node.name} - transpose in - does not reverse")
return [InsertTransposeAction(node, direction='out', idx=out_edge.from_idx, out_edge=out_edge, transpose=reverse_transpose(transpose)), EndActionUp(node)], cur_visited_nodes
@@ -663,10 +757,9 @@ def search_up(G, node, exclude_nodes, visited_nodes, out_edge, transpose_history
from_shape = node.old_shape.calc_transpose(
new_transpose) if new_transpose is not None else None
to_shape = node.shape.calc_transpose(
- transpose) if transpose is not None else None
+ reverse_transpose(transpose)) if transpose is not None else None
transpose_history = transpose_history + \
- [TransposeHistory(node, node.shape, transpose,
- to_shape, new_transpose)]
+ [TransposeHistory(node, node.shape, new_transpose, node.old_shape)]
info(f"rewrite reshape to {from_shape}->{to_shape}")
if from_shape is None or to_shape is None or from_shape.shape != to_shape.shape:
cur_actions.extend([
@@ -709,22 +802,37 @@ def continue_up(G, node, exclude_nodes, visited_nodes, cur_visited_nodes, cur_ac
if check_continue(visited_nodes, cur_visited_nodes, exclude_nodes, edge.from_node, 'up', edge.from_idx):
continue
-
- if isinstance(node, Broadcastable) and node.in_dims[edge.to_idx].rank != node.out_dims[0].rank:
+ edge_in_shape = node.in_dims[edge.to_idx].shape
+ if isinstance(node, Broadcastable) and len(edge_in_shape) != node.out_dims[0].rank:
max_shape = compute_max_shape(node.out_dims)
- old_shape, shape, new_transpose = broadcast_reduce(
- max_shape, node.in_dims[edge.to_idx].shape, transpose)
- extra_history = [
- TransposeHistory(
- node, node.in_dims[edge.to_idx].shape, transpose, shape, new_transpose)
- ]
- old_shape = apply_transpose(old_shape, new_transpose)
- shape = apply_transpose(shape, transpose)
- cur_actions.append(InsertReshapeAction(
- node, direction='in', idx=edge.to_idx,
- in_shape=Dim.unnamed(old_shape), out_shape=Dim.unnamed(shape)))
+ b_axes = broadcasted_axes(edge_in_shape, max_shape)
+
+ transpose_without_broadcast = strip_axes_from_transpose(
+ reverse_transpose(transpose), b_axes)
+ # from shape will be the old shape with the unbroadcasted transpose
+ from_shape = apply_transpose(
+ edge_in_shape, transpose_without_broadcast)
+ # to shape is the broadcasted input shape with the transpose with the leading ones removed
+ broadcasted_shape = ([1] * len(b_axes)) + list(edge_in_shape)
+ to_shape = strip_leading_ones(apply_transpose(
+ broadcasted_shape, reverse_transpose(transpose)), len(from_shape))
+ # if they are not equal insert a reshape
+ if from_shape != to_shape:
+ info(
+ f'{node.name} broadcasted input {edge.to_idx} requires reshape {from_shape}->{to_shape}')
+ cur_actions.append(
+ InsertReshapeAction(
+ node, direction='in', idx=edge.to_idx,
+ in_shape=from_shape,
+ out_shape=to_shape
+ ))
+ extra_history = [
+ TransposeHistory(
+ node, broadcasted_shape, transpose_without_broadcast, edge_in_shape)
+ ]
+ else:
+ extra_history = []
else:
- new_transpose = transpose
extra_history = []
new_actions, visited_up_nodes = search_up(
G, edge.from_node, exclude_nodes, visited_nodes | cur_visited_nodes, edge,
@@ -749,17 +857,62 @@ def apply_actions(G, results: Sequence[Action]):
action.execute(G)
+def search_edge_down(G, edge, pass_classes, stop_class):
+ node = edge.to_node
+ if isinstance(node, stop_class):
+ return node
+ elif isinstance(node, pass_classes):
+ edges = G.out_edges(node)
+ if len(edges) != 1:
+ return None
+ return search_edge_down(G, edges[0], pass_classes, stop_class)
+ else:
+ return None
+
+
+def find_sequences(G, end_node_class, middle_node_classes):
+ nodes = set(G.nodes(node_classes=end_node_class))
+ pairs = []
+ while nodes:
+ node = nodes.pop()
+ edges = G.out_edges(node)
+ if len(edges) != 1:
+ continue
+ end_node = search_edge_down(
+ G, edges[0], middle_node_classes, end_node_class)
+ if end_node is None:
+ continue
+ pairs.append((node, end_node))
+ if end_node in nodes:
+ nodes.remove(end_node)
+ return pairs
+
+
def combine_transposes(G):
- transpose_seqs = [trans for trans in G.nodes(node_classes=TransposeParameters)
- if len(G.out_edges(trans.name)) == 1 and
- isinstance(G.out_edges(trans.name)[0].to_node, TransposeParameters)]
- for trans in transpose_seqs:
- to_trans = G.out_edges(trans.name)[0].to_node
- new_transpose = apply_transpose(trans.transpose, to_trans.transpose)
+ trans_pairs = find_sequences(
+ G,
+ TransposeParameters,
+ (CopyParameters, UnaryOpParameters, ActivationParameters))
+
+ for tstart, tend in trans_pairs:
+ new_transpose = apply_transpose(tstart.transpose, tend.transpose)
info(
- f'combine transposes {trans.name} and {to_trans.name} {trans.transpose} & {to_trans.transpose} -> {new_transpose}')
- trans.transpose = new_transpose
- G.remove_and_reconnect(to_trans, edge_class=NNEdge)
+ f'combine transposes {tstart.name} and {tend.name} {tstart.transpose} & {tend.transpose} -> {new_transpose}')
+ tstart.transpose = new_transpose
+ G.remove_and_reconnect(tend, edge_class=NNEdge)
+
+
+def combine_reshapes(G):
+ reshape_pairs = find_sequences(
+ G,
+ ReshapeParameters,
+ (CopyParameters, UnaryOpParameters, ActivationParameters))
+
+ for rstart, rend in reshape_pairs:
+ info(
+ f'combine reshapes {rstart.name} and {rend.name} {rstart.shape} & {rend.shape}')
+ rstart.shape = rend.shape
+ G.remove_and_reconnect(rend, edge_class=NNEdge)
def remove_silly_reshapes(G):
@@ -796,19 +949,19 @@ def delete_step_idx(G, action: DeleteTransposeAction):
return G.in_edges(action.node)[0].from_node.step_idx
-def eliminate_transposes(G, debug_function=None, one_cycle=False):
+def eliminate_transposes(G, debug_function=None, one_cycle=False, single_step=False, do_silly=True):
info("eliminating unnecessary transposes")
found_results = True
pass_count = 0
while found_results:
pass_count += 1
- if pass_count > 50:
+ if pass_count > (200 if single_step else 50):
raise ValueError(
"Sorry, eliminate transposes is stuck in a loop. Please report to GreenWaves.")
found_results = False
visited_nodes = set()
actions = []
- info("search for transposes")
+ info(f"search for transposes +++ PASS {pass_count}")
transposes = G.nodes(node_classes=TransposeParameters)
while transposes:
transpose_node = transposes.pop(0)
@@ -830,7 +983,11 @@ def eliminate_transposes(G, debug_function=None, one_cycle=False):
visited_nodes,
cur_visited_up,
in_edge,
- [TransposeHistory(transpose_node, transpose_node.in_dims[0], transpose_node.transpose)])
+ [TransposeHistory( # When going up transpose is reversed
+ transpose_node,
+ transpose_node.out_dims[0].shape,
+ reverse_transpose(transpose_node.transpose),
+ transpose_node.in_dims[0].shape)])
# cur_visited_up.append(transpose_node)
except CantContinueError:
cur_actions_up = cur_visited_up = None
@@ -853,7 +1010,11 @@ def eliminate_transposes(G, debug_function=None, one_cycle=False):
visited_nodes,
cur_visited_down,
edge,
- [TransposeHistory(transpose_node, transpose_node.out_dims[0], transpose_node.transpose)])
+ [TransposeHistory(
+ transpose_node,
+ transpose_node.in_dims[0].shape,
+ transpose_node.transpose,
+ transpose_node.out_dims[0].shape)])
cur_actions_down += this_actions_down
cur_visited_down |= this_visited_down
except CantContinueError:
@@ -876,6 +1037,8 @@ def eliminate_transposes(G, debug_function=None, one_cycle=False):
actions += cur_actions_up
visited_nodes |= set(cur_visited_up.nodes)
visited_nodes.add(transpose_node)
+ if single_step:
+ break
# if transpose cannot be removed upwards movement push the transpose down if it actually moved
elif down_count > 0 or (down_count == 0 and transpose_moved(G, cur_actions_down)):
info(
@@ -884,14 +1047,21 @@ def eliminate_transposes(G, debug_function=None, one_cycle=False):
actions += cur_actions_down
visited_nodes |= set(cur_visited_down.nodes)
visited_nodes.add(transpose_node)
+ if single_step:
+ break
+ else:
+ info(
+ f'no elimination for {transpose_node.name} found')
if found_results:
info("eliminate transposes")
apply_actions(G, actions)
else:
info("no transposes to eliminate found")
- remove_silly_reshapes(G)
- combine_transposes(G)
+ if do_silly:
+ remove_silly_reshapes(G)
+ combine_reshapes(G)
+ combine_transposes(G)
G.add_dimensions()
if debug_function:
debug_function(G)
diff --git a/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes_actions.py b/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes_actions.py
index 22410ee25..55927a224 100644
--- a/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes_actions.py
+++ b/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes_actions.py
@@ -27,8 +27,12 @@
from utils.node_id import NodeId
LOG = logging.getLogger("nntool." + __name__)
-LOGL = LOG.info
+def info(msg):
+ LOG.info(msg)
+
+def debug(msg):
+ LOG.debug(msg)
class Action(ABC):
def __init__(self, node) -> None:
@@ -50,7 +54,7 @@ def __init__(self, node, message=None) -> None:
self.message = message
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
class StartAction(DebugActionBase):
@@ -116,7 +120,7 @@ def __init__(self, node, transpose=None, reshape_from=None, reshape_to=None, **k
self.reshape_to = reshape_to
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
direction = self.direction
if self.reshape_from is not None:
params = ReshapeParameters(G.unique_name(
@@ -150,7 +154,7 @@ def __init__(self, node, in_shape=None, out_shape=None, **kwargs) -> None:
self.out_shape = out_shape.clone() if out_shape is not None else None
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
params = ReshapeParameters(G.unique_name(
f'{node.name}_reshape'), old_shape=self.in_shape, shape=self.out_shape)
self.do_insert(node, G, params)
@@ -158,15 +162,21 @@ def _execute(self, node, G):
def __str__(self) -> str:
return f"insert reshape at {self.node.name}:{self.direction}_{self.idx} in {self.in_shape} out {self.out_shape}"
+def make_dim(shape):
+ if shape is None:
+ return shape
+ if isinstance(shape, Dim):
+ return shape.clone()
+ return Dim.unnamed(shape)
class SetReshapeAction(Action):
def __init__(self, node, in_shape=None, out_shape=None) -> None:
super(SetReshapeAction, self).__init__(node)
- self.in_shape = in_shape.clone() if in_shape is not None else None
- self.out_shape = out_shape.clone() if out_shape is not None else None
+ self.in_shape = make_dim(in_shape)
+ self.out_shape = make_dim(out_shape)
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
if self.in_shape is not None:
node.old_shape = self.in_shape
if self.out_shape is not None:
@@ -178,7 +188,7 @@ def __str__(self) -> str:
class SwitchBatchLinearAction(Action):
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
self.node.batch_minor = not self.node.batch_minor
def __str__(self) -> str:
@@ -195,7 +205,7 @@ def __init__(self, node, transpose_in, dir=None, transpose_out=None) -> None:
self.transpose_out = tuple(transpose_out)
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
node.act_slice = [node.act_slice[idx] for idx in self.transpose_in]
node.out_shape = [node.out_shape[idx] for idx in self.transpose_out]
@@ -209,7 +219,7 @@ def __init__(self, node, transpose, dir=None) -> None:
self.transpose = tuple(transpose)
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
node.padding = [node.padding[idx] for idx in self.transpose]
node.pad_vals = [node.pad_vals[idx] for idx in self.transpose]
@@ -223,7 +233,7 @@ def __init__(self, node, transpose, dir=None) -> None:
self.transpose = tuple(transpose)
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
node.axis = self.transpose[node.axis]
def __str__(self) -> str:
@@ -254,7 +264,7 @@ def from_history(cls, node, history, transpose=None):
class ReorderInputDims(TransposeInputBase):
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
node.dims.transpose(self.transpose)
if node.out_dims_hint:
node.out_dims_hint[0] = [node.out_dims_hint[0][idx]
@@ -267,7 +277,7 @@ def __str__(self) -> str:
class ReorderConstantInput(TransposeInputBase):
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
if G.quantization:
qrec = G.quantization.get(NodeId(node), None)
else:
@@ -293,7 +303,7 @@ def __str__(self) -> str:
class DeleteReshapeAction(Action):
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
if node.name not in G:
return
G.remove_and_reconnect(node, edge_class=NNEdge)
@@ -308,7 +318,7 @@ def __init__(self, node, reshape=None) -> None:
self.reshape = reshape
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
if node.name not in G:
return
if self.reshape:
@@ -342,7 +352,7 @@ def __move_transpose(self, from_dir, to_dir):
pass
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
node.transpose = self.transpose
def __str__(self) -> str:
@@ -358,31 +368,28 @@ def __init__(self, node, direction, transpose, shape, qrec=None) -> None:
self.qrec = qrec
@classmethod
- def out_from_history(cls, node, history, qrec):
+ def from_history(cls, node, history, qrec, dir):
# Find the first entry in the transpose history that actually has a transpose
first_valid_entry = next(iter([rec
for rec in reversed(history)
- if rec.from_transpose]))
- # in outwards direction the from_shape is the shape before the transpose and we
- # want to apply the transpose to get to the shape
- return cls(node, "out", tuple(first_valid_entry.from_transpose),
- tuple(first_valid_entry.from_shape), qrec=qrec)
+ if rec.transpose]))
+ # arriving from the top the transpose is in the down direction and from the
+ # bottom in the up direction so in both cases we need to reverse it
+ transpose = tuple(reverse_transpose(first_valid_entry.transpose))
+ # shape closest to the node
+ shape = tuple(first_valid_entry.to_shape)
+ return cls(node, dir, transpose, shape, qrec=qrec)
+
+ @classmethod
+ def out_from_history(cls, node, history, qrec):
+ return cls.from_history(node, history, qrec, "out")
@classmethod
def in_from_history(cls, node, history, qrec):
- # Find the first entry in the transpose history that actually has a transpose
- first_valid_entry = next(iter([rec
- for rec in reversed(history)
- if rec.from_transpose]))
- # in down direction we are pushing the transpose into the FC so
- # the from_shape is the shape with the transpose applied and
- # we want to reverse the from_transpose to get back to the shape before it
- transpose = tuple(reverse_transpose(first_valid_entry.from_transpose))
- shape = tuple(first_valid_entry.from_shape)
- return cls(node, "in", transpose, shape, qrec=qrec)
+ return cls.from_history(node, history, qrec, "in")
def _execute(self, node, G):
- LOGL("%s", str(self))
+ info(f"{self}")
filter_node = node.contained_filters()[0] if isinstance(node, LinearFusionParameters) else node
in_edges = G.indexed_in_edges(node.name)
weights_node = in_edges[1].from_node
diff --git a/tools/nntool/graph/matches/matchers/move_node_up.py b/tools/nntool/graph/matches/matchers/move_node_up.py
index 49c08b4bc..76b1545ec 100644
--- a/tools/nntool/graph/matches/matchers/move_node_up.py
+++ b/tools/nntool/graph/matches/matchers/move_node_up.py
@@ -18,7 +18,8 @@
Conv2DParameters, FcParameters, GlobalPoolingParameters,
MatrixAddParameters, MatrixMulParameters, NNEdge,
PoolingParameters, ReluActivationParameters,
- ReshapeParameters, TransposeParameters)
+ ReshapeParameters, TransposeParameters, MatMulTransposedParameters)
+from graph.types.others import ReverseParameters, StridedSliceParameters
from graph.types.tensor_arithmetic import MatMulOpParameters
from utils.graph import GraphView
from utils.node_id import NodeId
@@ -128,36 +129,22 @@ def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
return has_modified_graph
-@groups('scaled')
-@match_name("move_activations_scale8")
+@groups('*')
+@match_name("move_activations_up")
@description("Tries to move activations so they are after layers that they can be fused with."
- "Should be run before match_gap_ * fusions. Compatible with AutoTiler SQ8 kernels.")
+ "Should be run before match_gap_ * fusions.")
@needs_valid_dimension(True)
-@run_before('fuse_gap_convs', 'fuse_gap_linear', 'fuse_gap_pool', 'fuse_op_activation_scale8')
+@run_before('fuse_gap_convs', 'fuse_gap_linear', 'fuse_gap_pool', 'fuse_op_activation_scale8', 'fuse_op_activation_pow2')
class MoveActivationsMatcherScale8(MoveNodeUpMatcher):
- ValidNodesToPass = (ReshapeParameters,
+ ValidNodesToPass = (ReshapeParameters, StridedSliceParameters, ReverseParameters,
TransposeParameters, ConcatParameters)
- ValidFusions = (Conv2DParameters, FcParameters, PoolingParameters, PoolingParameters,
- GlobalPoolingParameters, MatrixAddParameters, MatrixMulParameters, MatMulOpParameters)
+ ValidFusions = (Conv2DParameters, FcParameters, PoolingParameters,
+ GlobalPoolingParameters, MatrixAddParameters, MatrixMulParameters,
+ MatMulOpParameters, MatMulTransposedParameters)
ValidNodes = (ActivationParameters,)
-
-@groups('symmetric')
-@match_name("move_activations_pow2")
-@description("Tries to move activations so they are after layers that they can be fused with."
- "Should be run before match_gap_ * fusions. Compatible with AutoTiler POW2 kernels.")
-@needs_valid_dimension(True)
-@run_before('fuse_gap_convs', 'fuse_gap_linear', 'fuse_gap_pool', 'fuse_op_activation_pow2')
-class MoveActivationsMatcherPow2(MoveNodeUpMatcher):
-
- ValidNodesToPass = (ReshapeParameters,
- TransposeParameters, ConcatParameters)
- ValidFusions = (Conv2DParameters, FcParameters, PoolingParameters)
- ValidNodes = (ActivationParameters,)
-
-
@groups('scaled')
@match_name("move_pooling_scale8")
@description("Tries to move pooling layers so they are after layers that they can be fused with."
diff --git a/tools/nntool/graph/matches/matchers/slice_to_split.py b/tools/nntool/graph/matches/matchers/slice_to_split.py
index e6414e0aa..20f601e88 100644
--- a/tools/nntool/graph/matches/matchers/slice_to_split.py
+++ b/tools/nntool/graph/matches/matchers/slice_to_split.py
@@ -202,7 +202,7 @@ def _match(self, G: GraphView, set_identity: bool = True, **kwargs) -> bool:
diff_slices = [sl for idx, sl in enumerate(
slices) if idx in diff_axes]
axis_lengths = in_edge[0].out_dims[in_edge[1]].shape
- if min(not_diff_axes) < max(diff_axes):
+ if not_diff_axes and min(not_diff_axes) < max(diff_axes):
transpose_from = tuple(range(len(slices)))
transpose_to = tuple(diff_axes + not_diff_axes)
axis_lengths = [axis_lengths[idx] for idx in transpose_to]
diff --git a/tools/nntool/graph/nngraph.py b/tools/nntool/graph/nngraph.py
index eb826daa1..2d4353ef1 100644
--- a/tools/nntool/graph/nngraph.py
+++ b/tools/nntool/graph/nngraph.py
@@ -290,9 +290,9 @@ def nodes_iterator(self, yield_fusions=True):
else:
yield (step_idx, node, None, None)
- def adjust_order(self, reshape_weights=True, postprocess=True, debug_function=None, one_cycle=False):
+ def adjust_order(self, reshape_weights=True, postprocess=True, debug_function=None, one_cycle=False, single_step=False):
adjust_order(self, reshape_weights=reshape_weights,
- postprocess=postprocess, debug_function=debug_function, one_cycle=one_cycle)
+ postprocess=postprocess, debug_function=debug_function, one_cycle=one_cycle, single_step=single_step)
LOG.info("adjusted order")
self.graph_identity.is_adjusted = True
@@ -376,3 +376,10 @@ def __repr__(self):
renderer = TextTableRenderer(150)
tab.render(renderer)
return renderer.get_output()
+
+ def total_ops(self):
+ tot_ops = 0
+ for node in self.nodes():
+ ops = node.compute_load()
+ tot_ops += ops if ops else 0
+ return tot_ops
diff --git a/tools/nntool/graph/types/conv2d.py b/tools/nntool/graph/types/conv2d.py
index bb17608d9..de85d7e01 100644
--- a/tools/nntool/graph/types/conv2d.py
+++ b/tools/nntool/graph/types/conv2d.py
@@ -15,7 +15,7 @@
import logging
-from ..dim import DilationDim, Dim2D
+from ..dim import DilationDim, Dim, Dim2D
from .base import (ComparableParameters, FilterLikeParameters, FilterParameters,
MultiplicativeBiasParameters, NoSizeChangeParameters,
SingleInputAndOutput, cls_op_name, SensitiveToOrder, nargs)
@@ -167,8 +167,11 @@ def get_output_size(self, in_dims):
pad = self.padding.height_width()
- out_dim = ((in_dim - filter_d + pad)//self.stride) + 1
- out_dim.c = self.filter.out_c
+ h = ((in_dim.h - filter_d.h + pad.h)//self.stride.h) + 1
+ w = ((in_dim.w - filter_d.w + pad.w)//self.stride.w) + 1
+ if h < 0 or w < 0:
+ raise ValueError(f'{self.name}dimension calculation invalid {h}, {w}')
+ out_dim = Dim.named_ordered(h=h, w=w, c=self.filter.out_c)
out_dim.impose_order(in_dim.order)
if self.batch is not None:
out_dim.insert_axis(0, new_name='n')
diff --git a/tools/nntool/graph/types/dsp_preprocessing.py b/tools/nntool/graph/types/dsp_preprocessing.py
index 177430e3c..237d743dc 100644
--- a/tools/nntool/graph/types/dsp_preprocessing.py
+++ b/tools/nntool/graph/types/dsp_preprocessing.py
@@ -155,11 +155,11 @@ def gen_fft_twiddles(self):
else:
win_lut = None
fft_twiddles = ConstantInputParameters(self.name + "_FFT_Twiddles", value=gen_fft_twiddles_lut(
- self.n_cfft, rad4=is_rad4), dims=Dim.unnamed([2, n_fft_lut]))
+ self.n_cfft, rad4=is_rad4), dims=Dim.unnamed([2*n_fft_lut]))
swaptable = ConstantInputParameters(self.name + "_SwapTable", value=gen_fft_swaptable_lut(
self.n_cfft, rad4=is_rad4), dims=Dim.unnamed([self.n_cfft]))
rfft_twiddles = ConstantInputParameters(self.name + "_RFFT_Twiddles", value=gen_rfft_twiddles_lut(
- self.n_fft), dims=Dim.unnamed([2, self.n_fft//2]))
+ self.n_fft), dims=Dim.unnamed([2*self.n_fft//2]))
return win_lut, fft_twiddles, swaptable, rfft_twiddles
@cls_op_name("RFFT2D")
@@ -352,8 +352,8 @@ def gen_melfilter(self):
return melfilt_coeff_sparse_node, melfilt_sparsity_node
def get_melfilter_size(self):
- melfilt_sparsity, melfilt_coeff = self.gen_melfilter()
- return melfilt_sparsity.dqvalue.size, melfilt_coeff.dqvalue.size
+ melfilt_coeff, melfilt_sparsity = self.gen_melfilter()
+ return melfilt_coeff.dqvalue.size, melfilt_sparsity.dqvalue.size
def gen_dct_matrix(self):
norm_factor = np.ones((self.n_dct, self.n_dct))
diff --git a/tools/nntool/graph/types/input_output.py b/tools/nntool/graph/types/input_output.py
index f39356396..9d25014a4 100644
--- a/tools/nntool/graph/types/input_output.py
+++ b/tools/nntool/graph/types/input_output.py
@@ -24,6 +24,7 @@
LOG = logging.getLogger("nntool." + __name__)
+
@not_generated
class InputOutputParameters(Parameters):
@@ -44,7 +45,8 @@ def graph_anon_label(self):
@property
def graph_label(self):
- shape = self.out_dims[0] if self.out_dims and self.out_dims[0] else "No Shape!"
+ shape = (self.out_dims[0] if self.out_dims and
+ self.out_dims[0] is not None else "No Shape!")
if self.fixed_order:
return [self.name, f'{shape}', "Frozen Order"]
return [self.name, f'{shape}']
@@ -129,7 +131,6 @@ def __call__(self, graph):
raise ValueError('expecting NNGraph as parameter')
return NNNodeRef(self, 0, graph)
-
def verify(self, G):
problems = []
for edge in G.in_edges(self.name):
@@ -153,9 +154,6 @@ def set_input(self, value):
self.output_value = value
-
-
-
@cls_op_name('output')
class OutputParameters(InputOutputParameters, InsensitiveToQuantization):
diff --git a/tools/nntool/graph/types/pooling.py b/tools/nntool/graph/types/pooling.py
index cac79088b..df274d955 100644
--- a/tools/nntool/graph/types/pooling.py
+++ b/tools/nntool/graph/types/pooling.py
@@ -13,11 +13,11 @@
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
-from copy import deepcopy
import logging
+from copy import deepcopy
-from ..dim import PoolFilterDim
-from .base import FilterLikeParameters, cls_op_name, SensitiveToOrder
+from ..dim import Dim, PoolFilterDim
+from .base import FilterLikeParameters, SensitiveToOrder, cls_op_name
LOG = logging.getLogger("nntool." + __name__)
@@ -70,9 +70,13 @@ def get_output_size(self, in_dims):
self.padding.calculate_same(in_dims, self.filter, self.stride)
pad = self.padding.height_width()
- out_dim = ((in_dims - self.filter + pad)//self.stride) + 1
- out_dim.c = in_dims.c
+ h = ((in_dims.h - self.filter.h + pad.h)//self.stride.h) + 1
+ w = ((in_dims.w - self.filter.w + pad.w)//self.stride.w) + 1
+ if h < 0 or w < 0:
+ raise ValueError(f'{self.name}dimension calculation invalid {h}, {w}')
+
+ out_dim = Dim.named_ordered(c=in_dims.c, h=h, w=w)
out_dim.impose_order(in_dims.order)
return [out_dim]
diff --git a/tools/nntool/graph/types/tensor_arithmetic.py b/tools/nntool/graph/types/tensor_arithmetic.py
index f21dce53d..f83ac516f 100644
--- a/tools/nntool/graph/types/tensor_arithmetic.py
+++ b/tools/nntool/graph/types/tensor_arithmetic.py
@@ -215,7 +215,11 @@ def get_parameter_size(self):
return 0
def compute_load(self):
- return self.out_dims[0].size() * 2
+ line_m1 = self.in_dims[0].shape[-2]
+ col_m1 = self.in_dims[0].shape[-1]
+ col_m2 = self.in_dims[1].shape[-2] if isinstance(self, MatMulTransposedParameters) else self.in_dims[1].shape[-1]
+ n_mat = np.prod(self.in_dims[1].shape[:-2])
+ return n_mat * (line_m1 * col_m1 * col_m2)
def get_output_size(self, in_dims):
x_shape = list(in_dims[0].shape).copy()
@@ -260,7 +264,7 @@ def get_output_size(self, in_dims):
remove_last = True
else:
remove_last = False
- y_shape = y_shape[::-1]
+ y_shape = y_shape[:-2] + y_shape[-2:][::-1]
x_chans = x_shape[:-2:]
y_chans = y_shape[:-2:]
out_chans = Dim.npbroadcast([x_chans, y_chans])
diff --git a/tools/nntool/importer/onnx/handlers/backend/conv_mixin.py b/tools/nntool/importer/onnx/handlers/backend/conv_mixin.py
index f77b5d261..ce52ae936 100644
--- a/tools/nntool/importer/onnx/handlers/backend/conv_mixin.py
+++ b/tools/nntool/importer/onnx/handlers/backend/conv_mixin.py
@@ -47,22 +47,42 @@ def conv(cls, node, quantized=False, **kwargs):
x = inputs[0]
x_rank = len(x[2].shape)
x_shape = x[2].shape
- if x_shape[0] is not None and x_shape[0] > 1:
- batch = x_shape[0]
- logger.warning(
- f"{valid_name} has a non 1 batch dimension of {batch} -"
- " this is not supported by nntool or autotiler kernels")
+
+ if x_shape[0] is not None:
+ real_in_shape = tuple(x_shape.copy())
+ if x_shape[0] > 1:
+ # support for multi batch is very limited
+ batch = x_shape[0]
+ logger.warning(
+ f"{valid_name} has a non 1 batch dimension of {batch} -"
+ " this is not supported by nntool or autotiler kernels")
+ else:
+ # if the batch is specified but is 1 then the input will be reshaped
+ # and the output will have the batch dim set as unknown.
+ batch = None
else:
+ real_in_shape = tuple(x_shape[1:])
batch = None
- real_in_shape = deepcopy(x_shape)
- #conv_shape = [x if idx > 0 and x is not None else 1 for idx, x in enumerate(x_shape)]
- conv_shape = x_shape
- if None in x_shape:
- real_in_shape.remove(None)
+
spatial_size = x_rank - 2
assert spatial_size == 2 or spatial_size == 1, "only 1D and 2D convolutions supported"
+
+ # Input error checking
+ undefined = []
+ if x_shape[1] is None:
+ # cope with swapped batch and channel due to bad initial reshape
+ if x_shape[0] == 1:
+ batch = None
+ x_shape = [x_shape[1], x_shape[0]] + list(x_shape[2:])
+ real_in_shape = x_shape[1:]
+ else:
+ undefined.append(f"input channel size of filter {valid_name} must be defined.")
+
if not all(dim is not None for dim in x_shape[-spatial_size:]):
- raise ValueError(f"input spatial size {x_shape} of filter {valid_name} must be defined. You may need to override input dimensions.")
+ undefined.append(f"input spatial size {x_shape} of filter {valid_name} must be defined.")
+ if undefined:
+ raise ValueError(f"{' '.join(undefined)}. You may need to override input dimensions.")
+
# M x C/group x kH x kW
weights_idx = 3 if quantized else 1
weights_node = inputs[weights_idx][0]
@@ -70,15 +90,15 @@ def conv(cls, node, quantized=False, **kwargs):
weights = cls.get_constant(inputs[weights_idx])
out_c = weights.shape[0]
group = node.attrs.get("group", 1)
- in_c = conv_shape[-spatial_size -
- 1] if conv_shape[-spatial_size-1] is not None else 1
+ in_c = x_shape[1]
filt_in_c = in_c // group
if in_c != weights.shape[1] * group:
raise ValueError(f'node {valid_name} has incorrect input channel '
f'dimension {in_c} expecting {weights.shape[1] * group}')
if spatial_size == 1:
filt_w = weights.shape[-1]
- filt_h = 1
+ filt_h = h = 1
+ w = x_shape[-1]
# create a new constant node since we are changing the shape
weights = np.reshape(weights, (out_c, filt_in_c, filt_h, filt_w))
weights_node = ConstantInputParameters(f'{valid_name}_weights', value=weights,
@@ -88,9 +108,14 @@ def conv(cls, node, quantized=False, **kwargs):
else:
filt_h = weights.shape[-2]
filt_w = weights.shape[-1]
- h = 1 if spatial_size == 1 else (
- conv_shape[-2] if conv_shape[-2] is not None else 1)
- w = conv_shape[-1] if conv_shape[-1] is not None else 1
+ h = x_shape[-2]
+ w = x_shape[-1]
+
+ conv_in_shape = (in_c, h, w)
+
+ # h = 1 if spatial_size == 1 else (
+ # x_shape[-2] if x_shape[-2] is not None else 1)
+ # w = x_shape[-1] if x_shape[-1] is not None else 1
filt_dim = Conv2DFilterDim(filt_h, filt_w,
out_c, in_c=filt_in_c)
@@ -174,48 +199,86 @@ def conv(cls, node, quantized=False, **kwargs):
to_node=params, from_idx=0, to_idx=1))
G.add_edge(NNEdge(from_node=biases_node,
to_node=params, from_idx=0, to_idx=2))
- if conv_shape != real_in_shape:
- # insert reshape from [xx,None,xx,xx] -> [None, xx, xx, xx]
- rbatch_params = ReshapeParameters(f'{valid_name}_reshape_batchdim',
- old_shape=Dim.unnamed(
- conv_shape),
- shape=Dim.unnamed(real_in_shape))
+
+ # check if input needs a reshape
+ if conv_in_shape != real_in_shape:
+ r1_params = ReshapeParameters(f'{valid_name}_reshape_in',
+ old_shape=Dim.unnamed(real_in_shape),
+ shape=Dim.unnamed(conv_in_shape))
G.add_edge(
- NNEdge(from_node=x[0], to_node=rbatch_params, from_idx=x[1], to_idx=0))
- prev_node = rbatch_params
- prev_idx = 0
+ NNEdge(from_node=x[0], to_node=r1_params, from_idx=x[1], to_idx=0))
+ G.add_edge(NNEdge(from_node=r1_params,
+ to_node=params, from_idx=0, to_idx=0))
else:
- prev_node = x[0]
- prev_idx = x[1]
+ G.add_edge(
+ NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
+ # check if output needs a reshape
if spatial_size == 1:
if batch is not None:
- oned_in_shape = [batch, in_c, w]
- twod_in_shape = [batch, in_c, 1, w]
oned_out_shape = [batch, out_dims[0].c, out_dims[0].w]
pout_dims = ProvisionalDim(oned_out_shape)
else:
- oned_in_shape = [in_c, w]
- twod_in_shape = [in_c, 1, w]
oned_out_shape = [out_dims[0].c, out_dims[0].w]
- pout_dims = ProvisionalDim([conv_shape[0]] + oned_out_shape)
- r1_params = ReshapeParameters(f'{valid_name}_reshape2d',
- old_shape=Dim.unnamed(oned_in_shape),
- shape=Dim.unnamed(twod_in_shape))
- r2_params = ReshapeParameters(f'{valid_name}_reshape1d',
+ pout_dims = ProvisionalDim([None] + oned_out_shape)
+
+ r2_params = ReshapeParameters(f'{valid_name}_reshape_out',
old_shape=out_dims[0],
shape=Dim.unnamed(oned_out_shape))
- G.add_edge(
- NNEdge(from_node=prev_node, to_node=r1_params, from_idx=prev_idx, to_idx=0))
- G.add_edge(NNEdge(from_node=r1_params,
- to_node=params, from_idx=0, to_idx=0))
G.add_edge(NNEdge(from_node=params,
to_node=r2_params, from_idx=0, to_idx=0))
- all_nodes[node.output[0]] = (r2_params, 0, pout_dims, o_qtype)
- return r2_params
+ params = r2_params
else:
- pout_dims = ProvisionalDim([conv_shape[0]] + out_dims[0].shape)
- G.add_edge(
- NNEdge(from_node=prev_node, to_node=params, from_idx=prev_idx, to_idx=0))
- all_nodes[node.output[0]] = (params, 0, pout_dims, o_qtype)
- return params
+ pout_dims = ProvisionalDim([batch] + out_dims[0].shape)
+
+ all_nodes[node.output[0]] = (params, 0, pout_dims, o_qtype)
+ return params
+
+
+
+
+
+ # # # insert reshape from [xx,None,xx,xx] -> [None, xx, xx, xx]
+ # # rbatch_params = ReshapeParameters(f'{valid_name}_reshape_batchdim',
+ # # old_shape=Dim.unnamed(
+ # # conv_shape),
+ # # shape=Dim.unnamed(real_in_shape))
+ # # G.add_edge(
+ # # NNEdge(from_node=x[0], to_node=rbatch_params, from_idx=x[1], to_idx=0))
+ # # prev_node = rbatch_params
+ # # prev_idx = 0
+ # # else:
+ # # prev_node = x[0]
+ # # prev_idx = x[1]
+
+ # if spatial_size == 1:
+ # if batch is not None:
+ # oned_in_shape = [batch, in_c, w]
+ # twod_in_shape = [batch, in_c, 1, w]
+ # oned_out_shape = [batch, out_dims[0].c, out_dims[0].w]
+ # pout_dims = ProvisionalDim(oned_out_shape)
+ # else:
+ # oned_in_shape = [in_c, w]
+ # twod_in_shape = [in_c, 1, w]
+ # oned_out_shape = [out_dims[0].c, out_dims[0].w]
+ # pout_dims = ProvisionalDim([conv_shape[0]] + oned_out_shape)
+ # r1_params = ReshapeParameters(f'{valid_name}_reshape2d',
+ # old_shape=Dim.unnamed(oned_in_shape),
+ # shape=Dim.unnamed(twod_in_shape))
+ # r2_params = ReshapeParameters(f'{valid_name}_reshape1d',
+ # old_shape=out_dims[0],
+ # shape=Dim.unnamed(oned_out_shape))
+ # G.add_edge(
+ # NNEdge(from_node=prev_node, to_node=r1_params, from_idx=prev_idx, to_idx=0))
+ # G.add_edge(NNEdge(from_node=r1_params,
+ # to_node=params, from_idx=0, to_idx=0))
+ # G.add_edge(NNEdge(from_node=params,
+ # to_node=r2_params, from_idx=0, to_idx=0))
+ # all_nodes[node.output[0]] = (r2_params, 0, pout_dims, o_qtype)
+ # return r2_params
+ # else:
+ # pout_dims = ProvisionalDim([conv_shape[0]] + out_dims[0].shape)
+ # G.add_edge(
+ # NNEdge(from_node=prev_node, to_node=params, from_idx=prev_idx, to_idx=0))
+ # all_nodes[node.output[0]] = (params, 0, pout_dims, o_qtype)
+ # return params
diff --git a/tools/nntool/importer/onnx/handlers/backend/gather.py b/tools/nntool/importer/onnx/handlers/backend/gather.py
index 9a822dcfa..1a2761e3d 100644
--- a/tools/nntool/importer/onnx/handlers/backend/gather.py
+++ b/tools/nntool/importer/onnx/handlers/backend/gather.py
@@ -13,6 +13,8 @@
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
+import copy
+from graph.types.others import StridedSliceParameters
import numpy as np
from graph.types import ConstantInputParameters, GatherParameters, NNEdge
from importer.common.constant_mixin import ConstantMixin
@@ -44,8 +46,14 @@ def _common(cls, node, **kwargs):
logger.info(f"reducing {valid_name} to a constant {cls.print_small(x_val)}")
params = ConstantInputParameters(valid_name, value=np.take(x_val, indices, axis=axis))
else:
- axis = cls._trim_axis(axis, x_shape)
- params = GatherParameters(valid_name, axis=axis, indices=indices)
+ if np.ndim(indices) <= 1:
+ idx = np.asscalar(indices)
+ act_slice = tuple([(0, dim, 1) if i != axis else (idx, idx+1, 1) for i, dim in enumerate(x_shape) if dim is not None])
+ out_shape = pshape.known_shape.copy()
+ params = StridedSliceParameters(valid_name, act_slice=act_slice, out_shape=out_shape)
+ else:
+ axis = cls._trim_axis(axis, x_shape)
+ params = GatherParameters(valid_name, axis=axis, indices=indices)
G.add_edge(NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
all_nodes[node.output[0]] = (params, 0, pshape, x[3])
return params
diff --git a/tools/nntool/importer/onnx/handlers/backend/mat_mul_mixin.py b/tools/nntool/importer/onnx/handlers/backend/mat_mul_mixin.py
index b2d6a0c6f..72c18169e 100644
--- a/tools/nntool/importer/onnx/handlers/backend/mat_mul_mixin.py
+++ b/tools/nntool/importer/onnx/handlers/backend/mat_mul_mixin.py
@@ -90,9 +90,13 @@ def _handle(cls, node, quantized=False, **kwargs):
)
else:
params = MatMulTransposedParameters(valid_name)
- trans2 = TransposeParameters(f'{valid_name}_tin2', transpose=(1, 0))
+ trans_shape = [i for i in range(len(y_shape))]
+ temp = trans_shape[-1]
+ trans_shape[-1] = trans_shape[-2]
+ trans_shape[-2] = temp
+ trans2 = TransposeParameters(f'{valid_name}_tin2', transpose=tuple(trans_shape))
out_dims = params.get_output_size(
- [Dim.unnamed(x_shape), Dim.unnamed(y_shape[::-1])])
+ [Dim.unnamed(x_shape), Dim.unnamed(y_shape[:-2] + y_shape[-2:][::-1])])
G.add_edge(
NNEdge(from_node=y[0], to_node=trans2, from_idx=y[1], to_idx=0))
G.add_edge(
diff --git a/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py b/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py
index 6a3aef956..77c796f0a 100644
--- a/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py
+++ b/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py
@@ -24,6 +24,7 @@ def pad_start_with(cls, val, pad_val, dlen):
def calc_pad_dim(cls, node, expected_len):
if "auto_pad" not in node.attrs or node.attrs["auto_pad"] == "NOTSET":
pads = cls.pad_start_with(node.attrs.get("pads", []), [0], expected_len)
+ pads = pads if len(pads) < 4 else [pads[0], pads[2], pads[1], pads[3]]
pad_dim = PadDim(*pads)
elif node.attrs["auto_pad"] == "VALID":
pad_dim = PadDim.valid()
diff --git a/tools/nntool/quantization/float/kernels/matrix_operations.py b/tools/nntool/quantization/float/kernels/matrix_operations.py
index 15789ea37..5ee185f17 100644
--- a/tools/nntool/quantization/float/kernels/matrix_operations.py
+++ b/tools/nntool/quantization/float/kernels/matrix_operations.py
@@ -137,7 +137,8 @@ def execute(cls, params,
if len(in_tensors) > 2:
biases = in_tensors[2]
if len(biases.shape) == 1:
- biases = np.expand_dims(biases, 1 if mat2.shape[1] == 1 else 0)
+ if biases.shape[0] == mat1.shape[0]:
+ biases = np.expand_dims(biases, -1)
else:
biases = 0
@@ -201,6 +202,7 @@ def execute(cls, params,
class BinaryOpFloat32(KernelBase):
+ @staticmethod
def FUNC(x, y): return x
@classmethod
@@ -293,4 +295,5 @@ class SqrtFloat32(UnaryOpFloat32):
@params_type(RSqrtOpParameters)
@qrec_type('float')
class RSqrtFloat32(UnaryOpFloat32):
+ @staticmethod
def FUNC(x): return 1.0/np.sqrt(x)
diff --git a/tools/nntool/quantization/float/quantizers/conv_fusion_float.py b/tools/nntool/quantization/float/quantizers/conv_fusion_float.py
index fe714462a..8226ae742 100644
--- a/tools/nntool/quantization/float/quantizers/conv_fusion_float.py
+++ b/tools/nntool/quantization/float/quantizers/conv_fusion_float.py
@@ -14,6 +14,8 @@
# along with this program. If not, see <https://www.gnu.org/licenses/>.
from copy import deepcopy
+from quantization.float.quantizers.filter_float import AT_HWC_KER_IN_ORDER, AT_HWC_KER_OUT_ORDER, AT_CHW_KER_IN_ORDER, AT_CHW_KER_OUT_ORDER
+from quantization.quantizer_options import HWC_OPTION
import numpy as np
from bfloat16 import bfloat16
@@ -24,9 +26,11 @@
from quantization.unified_quantization_handler import (fusion_handler,
in_qs_constraint,
out_qs_constraint,
- params_type)
-
+ params_type, options)
+@options(
+ HWC_OPTION
+)
@params_type(ConvFusionParameters, LinearFusionParameters)
@in_qs_constraint({'dtype': set([np.float16, np.float32, bfloat16])})
@out_qs_constraint({'dtype': set([np.float16, np.float32, bfloat16])})
@@ -36,4 +40,9 @@ class FilterFusionFloat(FloatQuantizionHandler):
def _quantize(cls, params, in_qs, stats, **kwargs):
_, dtype = cls.get_float_opts(**kwargs)
out_qs = [deepcopy(in_qs[0])]
+ opts = kwargs['opts']
+ if opts['hwc']:
+ cls.check_order(params, AT_HWC_KER_IN_ORDER, AT_HWC_KER_OUT_ORDER)
+ else:
+ cls.check_order(params, AT_CHW_KER_IN_ORDER, AT_CHW_KER_OUT_ORDER)
return QRec.float(in_qs=in_qs, out_qs=out_qs, float_dtype=dtype)
diff --git a/tools/nntool/quantization/multiplicative/quantizers/filter_mult.py b/tools/nntool/quantization/multiplicative/quantizers/filter_mult.py
index 3bcf75f22..f28a6cac5 100644
--- a/tools/nntool/quantization/multiplicative/quantizers/filter_mult.py
+++ b/tools/nntool/quantization/multiplicative/quantizers/filter_mult.py
@@ -52,22 +52,11 @@
def can_ne16(fusion, params, G):
if not isinstance(params, (Conv2DParameters, FcParameters, MatMulTransposedParameters)):
return False
- # if fusion:
- # if fusion.fusion_type in ['conv_active_pool', 'conv_active']:
- # if any(not isinstance(node, (Conv2DParameters, ReluActivationParameters, PoolingParameters))
- # for node in fusion.contained_nodes()):
- # return False
- # else:
- # return False
if isinstance(params, Conv2DParameters):
- # if (params.filter.w != params.filter.h or (params.filter.w != 1 and params.filter.w != 3)):
- # return False
if (params.is_depthwise_conv() and (params.filter.w != 3 or params.filter.h != 3)):
return False
- if (params.stride.size() != 1 and params.stride.shape != [2, 2]) and not ((params.filter.w == 1 or params.filter.h == 1)):
- return False
elif isinstance(params, MatMulTransposedParameters):
- in_nodes = [edge.from_node for edge in G.in_edges(params)]
+ in_nodes = [edge.from_node for edge in G.indexed_in_edges(params)]
if not isinstance(in_nodes[1], ConstantInputParameters):
return False
return True
diff --git a/tools/nntool/quantization/symmetric/kernels/dsp_preprocessing.py b/tools/nntool/quantization/symmetric/kernels/dsp_preprocessing.py
index 765c3a543..e6ba670bd 100644
--- a/tools/nntool/quantization/symmetric/kernels/dsp_preprocessing.py
+++ b/tools/nntool/quantization/symmetric/kernels/dsp_preprocessing.py
@@ -130,9 +130,9 @@ def execute(cls, params,
qrec: QRec,
**kwargs):
in_data = in_tensors[0]
- fft_twiddles = in_tensors[2]
+ fft_twiddles = np.stack([in_tensors[2][::2], in_tensors[2][1::2]], axis=0)
swap_table = in_tensors[3]
- rfft_twiddles = in_tensors[4]
+ rfft_twiddles = np.stack([in_tensors[4][::2], in_tensors[4][1::2]], axis=0)
spectrograms = []
for frame_idx in range(params.n_frames):
@@ -164,9 +164,9 @@ def execute(cls, params,
in_tensors,
qrec: QRec,
**kwargs):
- fft_twiddles = in_tensors[2]
+ fft_twiddles = np.stack([in_tensors[2][::2], in_tensors[2][1::2]], axis=0)
swap_table = in_tensors[3]
- rfft_twiddles = in_tensors[4]
+ rfft_twiddles = np.stack([in_tensors[4][::2], in_tensors[4][1::2]], axis=0)
mel_filterbank_sparsity_mat = in_tensors[5]
mel_filterbank_coeff = in_tensors[6]
diff --git a/tools/nntool/quantization/symmetric/kernels/matrix_operations.py b/tools/nntool/quantization/symmetric/kernels/matrix_operations.py
index 25b3df930..fad901498 100644
--- a/tools/nntool/quantization/symmetric/kernels/matrix_operations.py
+++ b/tools/nntool/quantization/symmetric/kernels/matrix_operations.py
@@ -204,7 +204,8 @@ def execute(cls, params,
if len(in_tensors) > 2:
biases = in_tensors[2]
if len(biases.shape) == 1:
- biases = np.expand_dims(biases, 1 if mat2.shape[1] == 1 else 0)
+ if biases.shape[0] == mat1.shape[0]:
+ biases = np.expand_dims(biases, -1)
else:
biases = 0
@@ -235,7 +236,8 @@ def execute(cls, params,
if len(in_tensors) > 2:
biases = in_tensors[2]
if len(biases.shape) == 1:
- biases = np.expand_dims(biases, 1 if mat2.shape[1] == 1 else 0)
+ if biases.shape[0] == mat1.shape[0]:
+ biases = np.expand_dims(biases, -1)
else:
biases = 0
diff --git a/tools/nntool/utils/gen_twiddles.py b/tools/nntool/utils/gen_twiddles.py
index 786aedf0c..60020a925 100644
--- a/tools/nntool/utils/gen_twiddles.py
+++ b/tools/nntool/utils/gen_twiddles.py
@@ -9,14 +9,23 @@ def gen_fft_twiddles_lut(Nfft, Inverse=False, rad4=False):
Twiddles_real = np.cos(-Phi)
Twiddles_imag = np.sin(-Phi)
if rad4:
- return np.stack([Twiddles_real[:int(3*Nfft/4)], Twiddles_imag[:int(3*Nfft/4)]], axis=0)
- return np.stack([Twiddles_real[:Nfft//2], Twiddles_imag[:Nfft//2]], axis=0)
+ twid = np.empty((2 * int(3*Nfft/4), ), dtype=Twiddles_real.dtype)
+ twid[0::2] = Twiddles_real[:int(3*Nfft/4)]
+ twid[1::2] = Twiddles_imag[:int(3*Nfft/4)]
+ return twid
+ twid = np.empty((2 * int(Nfft//2), ), dtype=Twiddles_real.dtype)
+ twid[0::2] = Twiddles_real[:int(Nfft//2)]
+ twid[1::2] = Twiddles_imag[:int(Nfft//2)]
+ return twid
def gen_rfft_twiddles_lut(Nfft):
Phi = (np.pi * 2 / Nfft) * np.arange(0, Nfft//2)
Twiddles_real = np.sin(Phi)
Twiddles_imag = np.cos(Phi)
- return np.stack([Twiddles_real, Twiddles_imag], axis=0)
+ twid = np.empty((Twiddles_real.size + Twiddles_imag.size, ), dtype=Twiddles_real.dtype)
+ twid[0::2] = Twiddles_real
+ twid[1::2] = Twiddles_imag
+ return twid
def gen_fft_swaptable_lut(Ni, rad4=False):
if rad4:
diff --git a/utils/gapy/run.py b/utils/gapy/run.py
index bb68ccf00..e30abc17c 100644
--- a/utils/gapy/run.py
+++ b/utils/gapy/run.py
@@ -125,6 +125,10 @@ def appendArgs(top_parser: argparse.ArgumentParser, parser: argparse.ArgumentPar
action = "store_true",
help = "Launch gtkwave")
+ parser.add_argument("--wsl", dest = "wsl",
+ action = None,
+ help = "Launch command in wsl environment")
+
[args, otherArgs] = top_parser.parse_known_args()
if args.platform is not None:
@@ -164,6 +168,7 @@ def operationFunc(args, config = None, system = None):
#if args.target is None:
# raise InputError('The target must be specified')
config.set('runner/platform', args.platform)
+ config.set('runner/wsl', args.wsl)
if args.binary is not None:
config.set('runner/boot-loader', args.binary)
diff --git a/utils/gapy/runner/board/board_runner.py b/utils/gapy/runner/board/board_runner.py
index b833da654..5b9580d86 100644
--- a/utils/gapy/runner/board/board_runner.py
+++ b/utils/gapy/runner/board/board_runner.py
@@ -95,6 +95,15 @@ def exec(self):
script = self.config.get_str('openocd/script')
binary = self.config.get_str('runner/boot-loader')
+ wsl = self.config.get_str('runner/wsl')
+ if wsl is None:
+ wsl_bin = binary
+ else:
+ path_header = '\\"//wsl$/' + wsl
+ path_footer = '\\"'
+ wsl_bin = path_header + binary + path_footer
+ script = os.environ.get('OPENOCD_CHIP_TARGET')
+
with open(binary, 'rb') as file:
elffile = ELFFile(file)
entry = elffile.header['e_entry']
@@ -138,9 +147,9 @@ def exec(self):
else:
platform = self.config.get_str('runner/platform')
if chip_family == 'vega' or chip_family == 'gap9_v2':
- cmd = '%s -d0 -c "gdb_port disabled; telnet_port disabled; tcl_port disabled" -c "script %s; script %s; load_and_start_binary %s 0x%x"' % (openocd, cable, script, binary, entry)
+ cmd = '%s -d0 -c "gdb_port disabled; telnet_port disabled; tcl_port disabled" -f "%s" -f "%s" -c "load_and_start_binary %s 0x%x"' % (openocd, cable, script, wsl_bin, entry)
else:
- cmd = "%s -d0 -c 'gdb_port disabled; telnet_port disabled; tcl_port disabled' -f %s -f %s -f tcl/jtag_boot_entry.tcl -c 'gap8_jtag_load_binary_and_start \"%s\" elf 0x%x'" % (openocd, cable, script, binary, entry)
+ cmd = "%s -d0 -c 'gdb_port disabled; telnet_port disabled; tcl_port disabled' -f %s -f %s -f tcl/jtag_boot_entry.tcl -c 'gap8_jtag_load_binary_and_start \"%s\" elf 0x%x'" % (openocd, cable, script, wsl_bin, entry)
os.chdir(self.config.get_str('gapy/work_dir'))
diff --git a/utils/rules/pulp_rules.mk b/utils/rules/pulp_rules.mk
index b7af23603..10d3fce53 100644
--- a/utils/rules/pulp_rules.mk
+++ b/utils/rules/pulp_rules.mk
@@ -233,8 +233,12 @@ ifdef LFS_ROOT_DIR
override config_args += --config-opt=flash/content/partitions/lfs/root_dir=$(LFS_ROOT_DIR)
endif
+ifneq ($(wsl),)
+WSL_ENV="--wsl=$(wsl)"
+endif
+
flash:
- gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --flash --force --binary=$(BIN) $(runner_args)
+ gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --flash --force --binary=$(BIN) $(runner_args)
flash_fs:
gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --flash --binary=$(BIN) $(runner_args)
@@ -246,10 +250,10 @@ run.prepare:
gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec-prepare --binary=$(BIN) $(runner_args)
run.exec:
- gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec --binary=$(BIN) $(runner_args)
+ gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec --binary=$(BIN) $(runner_args) $(WSL_ENV)
run:
- gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec-prepare --exec --binary=$(BIN) $(runner_args)
+ gapy --target=$(GAPY_TARGET) --platform=$(platform) --work-dir=$(BUILDDIR) $(config_args) $(gapy_args) run --exec-prepare --exec --binary=$(BIN) $(runner_args) $(WSL_ENV)
#$(INSTALL_DIR)/runner/run_gapuino.sh $(BUILDDIR) $(BIN) $(RAW_IMAGE_PLPBRIDGE_FLAGS) $(PLPBRIDGE_FLAGS) $(PLPBRIDGE_EXTRA_FLAGS)
From bf0b89d1089237e5b94fa8ad4e401e3780cda95a Mon Sep 17 00:00:00 2001
From: yao <yao.zhang@greenwaves-technologies.com>
Date: Mon, 20 Dec 2021 16:08:13 +0100
Subject: [PATCH 2/3] Release 4.9.0
---
.../pmsis/periph/i2s/wav_out_long/wav_out.c | 4 +-
gvsoc/gvsoc/bin/pulp-pc-info | 6 +
gvsoc/gvsoc/engine/include/gv/gvsoc.hpp | 2 +-
gvsoc/gvsoc/engine/python/gv/gvsoc.py | 1 +
gvsoc/gvsoc/engine/src/trace/trace.cpp | 26 ++-
gvsoc/gvsoc/engine/vp/trace_domain_impl.cpp | 3 +-
.../models/cpu/iss/vp/include/iss_wrapper.hpp | 1 +
.../models/cpu/iss/vp/src/iss_wrapper.cpp | 8 +
.../models/pulp/udma/hyper/udma_hyper_v3.cpp | 5 +
.../models/pulp/udma/i2c/v4/udma_i2c.cpp | 3 +
.../models/pulp/udma/mram/udma_mram_v2.cpp | 3 +
.../models/pulp/udma/udma_v4_impl.cpp | 8 +
libs/gap_lib/testbench/testlib.c | 7 +-
.../cluster/cluster_sync/fc_to_cl_delegate.h | 2 +-
.../pmsis_api/include/pmsis/drivers/i2s.h | 27 ++-
rtos/pulp/pulpos-2/kernel/mem_slab.c | 72 +++++++
tools/autotiler_v3/Makefile | 2 +-
tools/autotiler_v3_get/Makefile | 2 +-
tools/jenkins/gap_sdk_version.txt | 2 +-
.../new_generators/float/pool_float.py | 189 ++++++++++++++++++
.../graph/manipulations/adjust_order.py | 55 ++---
.../eliminate_transposes.py | 21 +-
tools/nntool/graph/nngraph.py | 5 +-
tools/nntool/graph/types/others.py | 2 +-
.../importer/onnx/handlers/backend/gru.py | 7 +-
.../importer/onnx/handlers/backend/lstm.py | 4 +-
.../onnx/handlers/backend/pad_mixin.py | 23 ++-
.../importer/onnx/handlers/backend/rnn.py | 5 +-
.../onnx/handlers/backend/rnn_mixin.py | 8 +-
.../tflite2/handlers/backend/concatenation.py | 17 +-
.../handlers/backend/fully_connected.py | 48 +++--
tools/nntool/interpreter/commands/adjust.py | 29 ++-
.../float/quantizers/pool_float.py | 52 +++++
utils/bin/binary-size | 1 +
utils/gap_configs/python/ips/iss/iss.py | 6 +-
utils/rules/pmsis_rules.mk | 2 +-
36 files changed, 560 insertions(+), 98 deletions(-)
create mode 100644 rtos/pulp/pulpos-2/kernel/mem_slab.c
create mode 100644 tools/nntool/generation/new_generators/float/pool_float.py
create mode 100644 tools/nntool/quantization/float/quantizers/pool_float.py
diff --git a/examples/pmsis/periph/i2s/wav_out_long/wav_out.c b/examples/pmsis/periph/i2s/wav_out_long/wav_out.c
index cb946ec1f..70b964eca 100644
--- a/examples/pmsis/periph/i2s/wav_out_long/wav_out.c
+++ b/examples/pmsis/periph/i2s/wav_out_long/wav_out.c
@@ -50,7 +50,7 @@ static int fs_write_from_L3(void *file, void *data, int size_total, struct pi_de
if(rest_size >= INTER_BUFF_SIZE)
size = INTER_BUFF_SIZE;
else
- size = INTER_BUFF_SIZE - rest_size;
+ size = rest_size;
pi_ram_read(ram, (_l3_buff+l3_index), _tmp_buffer, (uint32_t) size);
pi_fs_write(file, _tmp_buffer, size);
@@ -59,6 +59,8 @@ static int fs_write_from_L3(void *file, void *data, int size_total, struct pi_de
rest_size = rest_size - size;
} while (rest_size);
+ pmsis_l2_malloc_free(_tmp_buffer, (uint32_t) INTER_BUFF_SIZE);
+
return 0;
}
diff --git a/gvsoc/gvsoc/bin/pulp-pc-info b/gvsoc/gvsoc/bin/pulp-pc-info
index b86a4b8d8..62bff025e 100755
--- a/gvsoc/gvsoc/bin/pulp-pc-info
+++ b/gvsoc/gvsoc/bin/pulp-pc-info
@@ -3,6 +3,7 @@
import argparse
import os
from subprocess import Popen, PIPE
+import sys
parser = argparse.ArgumentParser(description='Generate PC debug info')
@@ -125,6 +126,11 @@ for f in functions:
# And finally generate the output files
+if args.allFile is None and args.pcFile is None and args.debugFile is None and args.inlineFile is None:
+ for f in functions:
+ f.dumpAll(sys.stdout)
+
+
# PC oriented file
if args.allFile != None:
with open(args.allFile, 'w') as file:
diff --git a/gvsoc/gvsoc/engine/include/gv/gvsoc.hpp b/gvsoc/gvsoc/engine/include/gv/gvsoc.hpp
index bd9f07568..a5819c6fa 100644
--- a/gvsoc/gvsoc/engine/include/gv/gvsoc.hpp
+++ b/gvsoc/gvsoc/engine/include/gv/gvsoc.hpp
@@ -200,7 +200,7 @@ namespace gv {
* @param id ID of the VCD event.
* @param value The new value.
*/
- virtual void event_update_logical(int64_t timestamp, int id, uint8_t value) = 0;
+ virtual void event_update_logical(int64_t timestamp, int id, uint64_t value) = 0;
/**
* Called by GVSOC to update the value of a bitfield VCD event.
diff --git a/gvsoc/gvsoc/engine/python/gv/gvsoc.py b/gvsoc/gvsoc/engine/python/gv/gvsoc.py
index 329df9389..04f900465 100644
--- a/gvsoc/gvsoc/engine/python/gv/gvsoc.py
+++ b/gvsoc/gvsoc/engine/python/gv/gvsoc.py
@@ -120,6 +120,7 @@ def gen_config(args, config):
for binary in debug_binaries:
full_config.set('**/debug_binaries', binary + '.debugInfo')
+ full_config.set('**/binaries', binary)
gvsoc_config_path = os.path.join(config.get_str('gapy/work_dir'), 'gvsoc_config.json')
diff --git a/gvsoc/gvsoc/engine/src/trace/trace.cpp b/gvsoc/gvsoc/engine/src/trace/trace.cpp
index 3ef9c4602..26605db01 100644
--- a/gvsoc/gvsoc/engine/src/trace/trace.cpp
+++ b/gvsoc/gvsoc/engine/src/trace/trace.cpp
@@ -83,6 +83,8 @@ void vp::component_trace::new_trace_event_string(std::string name, trace *trace)
trace->name = name;
trace->path = top.get_path() + "/" + name;
+ trace->width = 0;
+ trace->is_real = false;
trace->is_string = true;
trace->pending_timestamp = -1;
trace->bytes = 0;
@@ -409,15 +411,33 @@ void vp::trace_engine::flush_event_traces(int64_t timestamp)
}
else if (current->is_string)
{
- //this->vcd_user->event_update_logical(int id, uint8_t *value, uint8_t *flags);
+ this->vcd_user->event_update_string(timestamp, current->id, (char *)current->buffer);
}
- else if (current->width > 1)
+ else if (current->width > 8)
{
+ if (current->width <= 16)
+ {
+ this->vcd_user->event_update_logical(timestamp, current->id, *(uint16_t *)current->buffer);
+ }
+ else if (current->width <= 32)
+ {
+ this->vcd_user->event_update_logical(timestamp, current->id, *(uint32_t *)current->buffer);
+ }
+ else if (current->width <= 64)
+ {
+ this->vcd_user->event_update_logical(timestamp, current->id, *(uint64_t *)current->buffer);
+ }
+ else
+ {
+ // Use bitfield
+ }
}
else
{
- this->vcd_user->event_update_logical(timestamp, current->id, *current->buffer);
+ uint64_t value = (uint64_t)*(current->buffer);
+
+ this->vcd_user->event_update_logical(timestamp, current->id, value);
}
}
else
diff --git a/gvsoc/gvsoc/engine/vp/trace_domain_impl.cpp b/gvsoc/gvsoc/engine/vp/trace_domain_impl.cpp
index 182e9a408..76f4757b3 100644
--- a/gvsoc/gvsoc/engine/vp/trace_domain_impl.cpp
+++ b/gvsoc/gvsoc/engine/vp/trace_domain_impl.cpp
@@ -178,7 +178,7 @@ void trace_domain::check_trace_active(vp::trace *trace, int event)
{
for (auto &x : events_path_regex)
{
- if (x.second->is_path || regexec(x.second->regex, full_path.c_str(), 0, NULL, 0) == 0)
+ if ((x.second->is_path && x.second->path == full_path) || regexec(x.second->regex, full_path.c_str(), 0, NULL, 0) == 0)
{
std::string file_path = x.second->file_path;
vp::Event_trace *event_trace;
@@ -462,6 +462,7 @@ void trace_domain::conf_trace(int event, std::string path_str, bool enabled)
const char *file_path = "all.vcd";
const char *path = path_str.c_str();
char *delim = (char *)::index(path, '@');
+
if (delim)
{
*delim = 0;
diff --git a/gvsoc/gvsoc/models/cpu/iss/vp/include/iss_wrapper.hpp b/gvsoc/gvsoc/models/cpu/iss/vp/include/iss_wrapper.hpp
index ac739ef20..f7e234e1c 100644
--- a/gvsoc/gvsoc/models/cpu/iss/vp/include/iss_wrapper.hpp
+++ b/gvsoc/gvsoc/models/cpu/iss/vp/include/iss_wrapper.hpp
@@ -151,6 +151,7 @@ class iss_wrapper : public vp::component, vp::Gdbserver_core
vp::trace inline_trace_event;
vp::trace line_trace_event;
vp::trace file_trace_event;
+ vp::trace binaries_trace_event;
vp::trace pcer_trace_event[32];
vp::trace insn_trace_event;
diff --git a/gvsoc/gvsoc/models/cpu/iss/vp/src/iss_wrapper.cpp b/gvsoc/gvsoc/models/cpu/iss/vp/src/iss_wrapper.cpp
index 4559223c9..69df9a4e0 100644
--- a/gvsoc/gvsoc/models/cpu/iss/vp/src/iss_wrapper.cpp
+++ b/gvsoc/gvsoc/models/cpu/iss/vp/src/iss_wrapper.cpp
@@ -1300,6 +1300,7 @@ int iss_wrapper::build()
traces.new_trace_event_string("func", &func_trace_event);
traces.new_trace_event_string("inline_func", &inline_trace_event);
traces.new_trace_event_string("file", &file_trace_event);
+ traces.new_trace_event_string("binaries", &binaries_trace_event);
traces.new_trace_event("line", &line_trace_event, 32);
traces.new_trace_event_real("ipc_stat", &ipc_stat_event);
@@ -1423,6 +1424,13 @@ void iss_wrapper::start()
iss_register_debug_info(this, x->get_str().c_str());
}
+ if (this->get_js_config()->get("**/binaries") != NULL)
+ {
+ for (auto x:this->get_js_config()->get("**/binaries")->get_elems())
+ {
+ this->binaries_trace_event.event_string("static enable " + x->get_str());
+ }
+ }
trace.msg("ISS start (fetch: %d, is_active: %d, boot_addr: 0x%lx)\n", fetch_enable_reg.get(), is_active_reg.get(), get_config_int("boot_addr"));
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/hyper/udma_hyper_v3.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/hyper/udma_hyper_v3.cpp
index 135ddc6cb..7de6736f5 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/hyper/udma_hyper_v3.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/hyper/udma_hyper_v3.cpp
@@ -83,6 +83,8 @@ Hyper_periph::Hyper_periph(udma *top, int id, int itf_id) : Udma_periph(top, id)
this->read_req_waiting = new Udma_queue<Hyper_read_request>(-1);
this->read_req_ready = new Udma_queue<Hyper_read_request>(-1);
+ memset((void *)&this->ca, 0, sizeof(this->ca));
+
}
@@ -191,6 +193,9 @@ void Hyper_periph::reset(bool active)
{
Udma_periph::reset(active);
+ this->rx_channel->reset(active);
+ this->tx_channel->reset(active);
+
if (!active)
{
this->pending_tx = false;
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/i2c/v4/udma_i2c.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/i2c/v4/udma_i2c.cpp
index a1a71b451..8a589314f 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/i2c/v4/udma_i2c.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/i2c/v4/udma_i2c.cpp
@@ -141,6 +141,9 @@ void I2c_periph::reset(bool active)
{
Udma_periph::reset(active);
+ this->foll_rx_channel->reset(active);
+ this->foll_tx_channel->reset(active);
+
if (active)
{
}
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/mram/udma_mram_v2.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/mram/udma_mram_v2.cpp
index d8d98a74c..0a944fd4e 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/mram/udma_mram_v2.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/mram/udma_mram_v2.cpp
@@ -508,6 +508,9 @@ void Mram_periph::reset(bool active)
this->read_fifo->reset(active);
+ this->rx_channel->reset(active);
+ this->tx_channel->reset(active);
+
if (!active)
{
this->pending_transfer_size = 0;
diff --git a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_impl.cpp b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_impl.cpp
index e58dad15b..127890b20 100644
--- a/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_impl.cpp
+++ b/gvsoc/gvsoc_gap/models/pulp/udma/udma_v4_impl.cpp
@@ -709,6 +709,14 @@ void udma::start()
void udma::reset(bool active)
{
+ for (auto x: this->periphs)
+ {
+ if (x)
+ {
+ x->reset(active);
+ }
+ }
+
for (auto x: this->addrgen_linear)
{
x->reset(active);
diff --git a/libs/gap_lib/testbench/testlib.c b/libs/gap_lib/testbench/testlib.c
index 1619bdc2a..063221819 100644
--- a/libs/gap_lib/testbench/testlib.c
+++ b/libs/gap_lib/testbench/testlib.c
@@ -185,6 +185,7 @@ void i2s_slot_deinit(i2s_slot_test_t *i2s_slot)
struct pi_i2s_channel_conf i2s_conf;
pi_i2s_channel_conf_init(&i2s_conf);
i2s_conf.options = PI_I2S_OPT_DISABLED | (i2s_slot->is_rx ? PI_I2S_OPT_IS_RX: PI_I2S_OPT_IS_TX);
+
if (i2s_slot->frame)
{
pi_i2s_frame_channel_conf_set(i2s_slot->i2s, i2s_slot->frame, i2s_slot->slot, &i2s_conf);
@@ -237,11 +238,11 @@ int i2s_slot_init(i2s_test_t *test, i2s_slot_test_t *i2s_slot, struct pi_device
if (slot_config->is_rx)
{
- i2s_conf.options = PI_I2S_OPT_PINGPONG | PI_I2S_OPT_IS_RX | PI_I2S_OPT_ENABLED;
+ i2s_conf.options = PI_I2S_OPT_IS_RX | PI_I2S_OPT_ENABLED;
}
else
{
- i2s_conf.options = PI_I2S_OPT_PINGPONG | PI_I2S_OPT_IS_TX | PI_I2S_OPT_ENABLED;
+ i2s_conf.options = PI_I2S_OPT_IS_TX | PI_I2S_OPT_ENABLED;
}
if (slot_config->slab)
@@ -315,10 +316,12 @@ int i2s_slot_init(i2s_test_t *test, i2s_slot_test_t *i2s_slot, struct pi_device
if (slot_config->slab)
{
+ i2s_conf.options |= PI_I2S_OPT_MEM_SLAB;
i2s_conf.mem_slab = &i2s_slot->slab;
}
else
{
+ i2s_conf.options |= PI_I2S_OPT_PINGPONG;
i2s_conf.pingpong_buffers[0] = i2s_slot->buffers[0];
i2s_conf.pingpong_buffers[1] = i2s_slot->buffers[1];
}
diff --git a/rtos/pmsis/pmsis_api/include/pmsis/cluster/cluster_sync/fc_to_cl_delegate.h b/rtos/pmsis/pmsis_api/include/pmsis/cluster/cluster_sync/fc_to_cl_delegate.h
index 04a238380..a4bf4a5ab 100644
--- a/rtos/pmsis/pmsis_api/include/pmsis/cluster/cluster_sync/fc_to_cl_delegate.h
+++ b/rtos/pmsis/pmsis_api/include/pmsis/cluster/cluster_sync/fc_to_cl_delegate.h
@@ -207,7 +207,7 @@ static inline int pi_cluster_send_task_async(struct pi_device *device,
int pi_cluster_send_task_to_cl(struct pi_device *device, struct pi_cluster_task *task);
int pi_cluster_send_task_to_cl_async(struct pi_device *device, struct pi_cluster_task *cluster_task, pi_task_t *task);
-int pi_cluster_send_tasklet_to_cl_async(struct pi_device *device, struct pi_cluster_task *cluster_task, pi_task_t *task);
+void pi_cluster_send_tasklet_to_cl(struct pi_device *device, struct pi_cluster_task *cluster_task);
static inline int pi_cluster_send_task(struct pi_device *device, struct pi_cluster_task *task)
{
diff --git a/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h b/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h
index 62c560de7..2a170bcc5 100644
--- a/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h
+++ b/rtos/pmsis/pmsis_api/include/pmsis/drivers/i2s.h
@@ -524,25 +524,38 @@ struct pi_i2s_conf
* \struct pi_i2s_channel_conf
*
* \brief Interface channel configuration options. This configuration has to be
- * used when configuring a channel in TDM mode.
+ * used when configuring a channel in TDM mode. This can also be used to configure
+ * channels when they are part of a frame. Be careful in that case that some fields
+ * are not specific restrictions in this case.
*/
struct pi_i2s_channel_conf
{
size_t block_size; /*!< Size of one RX/TX memory block(buffer) in bytes.
On some chips, this size may have to be set under a
- maximum size, check the chip-specific section. */
- pi_mem_slab_t *mem_slab; /*!< memory slab to store RX/TX data. */
+ maximum size, check the chip-specific section.
+ In frame-based mode, this field should be the same for
+ all channels. */
+ pi_mem_slab_t *mem_slab; /*!< memory slab to store RX/TX data.
+ In frame-based mode, this field should be the same for
+ all channels. */
void *pingpong_buffers[2]; /*!< Pair of buffers used in double-bufferin
- mode to capture the incoming samples. */
+ mode to capture the incoming samples.
+ In frame-based mode, this field should be the same for
+ all channels. */
pi_i2s_fmt_t format; /*!< Data stream format as defined by PI_I2S_FMT_* constants. */
uint8_t word_size; /*!< Number of bits representing one data word. */
int8_t mem_word_size; /*!< Number of bits representing one data word in memory.
- If it is -1, this is equal to word_size. */
+ If it is -1, this is equal to word_size.
+ In frame-based mode, this field should be the same for
+ all channels. */
pi_i2s_opt_t options; /*!< Configuration options as defined by PI_I2S_OPT_* constants. */
int8_t asrc_channel; /*!< If different from -1, this redirect the specified
stream(can be input or output) to/from the ASRC block with
- the channel specified here. */
- uint8_t ts_evt_id; /*!< UDMA Config Event ID for generating the timestamp */
+ the channel specified here.
+ In frame-based mode, this field should be -1.*/
+ uint8_t ts_evt_id; /*!< UDMA Config Event ID for generating the timestamp.
+ In frame-based mode, this field should be the same for
+ all channels. */
uint8_t slot_enable; /*!< Specifies if the corresponding slot must be enabled or not.
It is by default set to 1. */
};
diff --git a/rtos/pulp/pulpos-2/kernel/mem_slab.c b/rtos/pulp/pulpos-2/kernel/mem_slab.c
new file mode 100644
index 000000000..e9457b4b2
--- /dev/null
+++ b/rtos/pulp/pulpos-2/kernel/mem_slab.c
@@ -0,0 +1,72 @@
+/*
+ * Copyright (c) 2016 Wind River Systems, Inc.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "pmsis.h"
+
+static void create_free_list(struct pi_mem_slab *slab)
+{
+ uint32_t j;
+ char *p;
+
+ slab->free_list = NULL;
+ p = slab->buffer + slab->block_size*(slab->num_blocks - 1);
+
+ for (j = 0U; j < slab->num_blocks; j++) {
+ *(char **)p = slab->free_list;
+ slab->free_list = p;
+ p -= slab->block_size;
+ }
+}
+
+
+void pi_mem_slab_init(pi_mem_slab_t *slab, void *buffer,
+ size_t block_size, uint32_t num_blocks)
+{
+ slab->num_blocks = num_blocks;
+ slab->block_size = block_size;
+ slab->buffer = buffer;
+ slab->num_used = 0U;
+ create_free_list(slab);
+}
+
+
+int pi_mem_slab_alloc(pi_mem_slab_t *slab, void **mem, int32_t timeout)
+{
+ int result;
+
+ int irq = hal_irq_disable();
+
+ if (slab->free_list != NULL)
+ {
+ /* take a free block */
+ *mem = slab->free_list;
+ slab->free_list = *(char **)(slab->free_list);
+ slab->num_used++;
+ result = 0;
+ }
+ else
+ {
+ /* don't wait for a free block to become available */
+ *mem = NULL;
+ result = -1;
+ }
+
+ hal_irq_restore(irq);
+
+ return result;
+}
+
+
+void pi_mem_slab_free(pi_mem_slab_t *slab, void **mem)
+{
+ int irq = hal_irq_disable();
+
+ **(char ***)mem = slab->free_list;
+ slab->free_list = *(char **)mem;
+ slab->num_used--;
+
+ hal_irq_restore(irq);
+}
diff --git a/tools/autotiler_v3/Makefile b/tools/autotiler_v3/Makefile
index c6f308357..dc12e7fd7 100644
--- a/tools/autotiler_v3/Makefile
+++ b/tools/autotiler_v3/Makefile
@@ -1,4 +1,4 @@
-TILER_VER=4.1.0
+TILER_VER=4.3.0
export TILER_LIB=libtile.${TILER_VER}.a
ifdef GAP_SDK_HOME
export TILER_URL=$(GAP_SDK_HOME)/.tiler_url
diff --git a/tools/autotiler_v3_get/Makefile b/tools/autotiler_v3_get/Makefile
index c6f308357..dc12e7fd7 100644
--- a/tools/autotiler_v3_get/Makefile
+++ b/tools/autotiler_v3_get/Makefile
@@ -1,4 +1,4 @@
-TILER_VER=4.1.0
+TILER_VER=4.3.0
export TILER_LIB=libtile.${TILER_VER}.a
ifdef GAP_SDK_HOME
export TILER_URL=$(GAP_SDK_HOME)/.tiler_url
diff --git a/tools/jenkins/gap_sdk_version.txt b/tools/jenkins/gap_sdk_version.txt
index 1f8776c61..dbbe1c752 100644
--- a/tools/jenkins/gap_sdk_version.txt
+++ b/tools/jenkins/gap_sdk_version.txt
@@ -1 +1 @@
-0addfcd309a48d5a5dbd3e058de0cf1ffc660a2b
+1f3f3d29ae92199246b03f00e4c207540022b502
diff --git a/tools/nntool/generation/new_generators/float/pool_float.py b/tools/nntool/generation/new_generators/float/pool_float.py
new file mode 100644
index 000000000..40aac877d
--- /dev/null
+++ b/tools/nntool/generation/new_generators/float/pool_float.py
@@ -0,0 +1,189 @@
+# Copyright (C) 2021 GreenWaves Technologies, SAS
+
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as
+# published by the Free Software Foundation, either version 3 of the
+# License, or (at your option) any later version.
+
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU Affero General Public License for more details.
+
+# You should have received a copy of the GNU Affero General Public License
+# along with this program. If not, see <https://www.gnu.org/licenses/>.
+
+from functools import reduce
+from utils.largest_factor import balanced_divisors
+from quantization.multiplicative.mulbias import compute_in_out_scale
+from generation.new_generators.helpers.in_out_bindings_mixin import InOutBindingsMixin
+from graph.types.global_pooling import GlobalAveragePoolParameters, GlobalSumPoolParameters
+from graph.dim import PadDim
+from generation.new_generators.helpers.act_infos import gen_act_infos
+from generation.at_types.at_params import NO_POOL, gen_activation_op, gen_globalpool_at_params, gen_pool_at_params
+import logging
+from utils.node_id import NodeId
+
+import numpy as np
+from generation.at_types.constant_info import ConstantInfo
+from generation.at_types.gen_ctrl import GenCtrl
+from generation.at_types.tc_arg_info import GlobalArgInfo
+from generation.generators.globals.global_names import INFOS
+from generation.generators.kernels.autotiler_kernel import NewAutoTilerKernel
+from generation.helpers.gen_constant import gen_constant
+from generation.new_generators.generator_base import (GeneratorBase,
+ paramstype, ktype)
+from graph.types import GlobalPoolingParameters, PoolingParameters, ActivationFusion
+from quantization.qtype import QType
+
+LOG = logging.getLogger("nntool." + __name__)
+
+@paramstype(ActivationFusion, GlobalPoolingParameters, PoolingParameters)
+@ktype('float')
+class PoolActGenerator(GeneratorBase, InOutBindingsMixin):
+
+ @classmethod
+ def globals_generator(cls, gen, node, qrec, pnode, fnode) -> bool:
+ return True
+
+ @classmethod
+ def bindings_generator(cls, gen, node, qrec, in_eparams, out_eparams, cname) -> bool:
+ if isinstance(node, ActivationFusion):
+ cnodes = node.contained_nodes()
+ if isinstance(cnodes[0], (GlobalPoolingParameters, PoolingParameters)):
+ cls.set_in_out_bindings(gen, in_eparams, out_eparams, cname, cnodes[0], qrec)
+ return True
+ return False
+ elif isinstance(node, (GlobalPoolingParameters, PoolingParameters)):
+ cls.set_in_out_bindings(gen, in_eparams, out_eparams, cname, node, qrec)
+ else:
+ return False
+ return True
+
+ @classmethod
+ def kernel_generator(cls, gen, node, qrec, in_eparams, out_eparams, cname) -> bool:
+ if isinstance(node, (GlobalPoolingParameters, PoolingParameters)):
+ pool_kernel = PoolActKernel if isinstance(node, PoolingParameters) else GlobalPoolActKernel
+ gen.kernels.append(pool_kernel(node.name, cname, node, qrec, None, None,
+ force_relu=gen.force_relu, gen_ctrl=node.get_gen_ctrl()))
+ elif isinstance(node, ActivationFusion):
+ cnodes = node.contained_nodes()
+ quants = [gen.G.quantization[NodeId(node, fnode)] for fnode in cnodes]
+ pool_kernel = PoolActKernel if isinstance(cnodes[0], PoolingParameters) else GlobalPoolActKernel
+ gen.kernels.append(pool_kernel(node.name, cname, cnodes[0], quants[0], cnodes[1], quants[1],
+ force_relu=gen.force_relu, gen_ctrl=node.get_gen_ctrl()))
+ return True
+
+class PoolActKernel(NewAutoTilerKernel):
+ CALL_TEMPLATE = """// generator for {node_name}
+CNN_PoolAct_fp16("{cname}", {gen_ctrl},
+ {feat}, {feat}, {width}, {height},
+ {kop_pool}, {fpx}, {fpy}, {dpx}, {dpy}, {spx}, {spy}, {pool_pad},
+ {kop_act});
+"""
+ def __init__(self, node_name, cname, pool_params, pool_q, act_params, act_q, force_relu, gen_ctrl=None):
+ if gen_ctrl is None:
+ self.gen_ctrl = gen_ctrl = GenCtrl(None, cname=cname)
+ else:
+ gen_ctrl.cname = cname
+ self.gen_ctrl = gen_ctrl
+ if pool_params.ker_in_order and pool_params.ker_in_order[0] == ["h", "w", "c"]:
+ hwc = True
+ gen_ctrl.hwc = 1
+
+ pad_compatibilities = []
+ at_pool_params = gen_pool_at_params(
+ pool_params, pad_compatibilities)
+ in_dim = pool_params.in_dims[0]
+ out_dim = pool_params.out_dims[0]
+ in_q = pool_q.in_qs[0]
+ out_q = pool_q.out_qs[0]
+
+ if act_params is not None:
+ act_op = gen_activation_op(
+ act_params.activation, force_relu=force_relu, asymmetric=act_q.in_qs[0].zero_point != 0)
+ if out_dim is None:
+ out_dim = act_params.out_dims[0].expand_to_chw()
+ out_q = act_q.out_qs[0]
+ else:
+ act_op = "KOP_NONE"
+
+ if pad_compatibilities:
+ reduction = PadDim.pad_compatibility_reduce(*pad_compatibilities,
+ "convolution padding is not compatible with pool padding")
+ if not reduction[2]: # default is balanced pad left
+ at_pad_ctrl = next(i for i, v in enumerate(reduction) if v)
+ LOG.debug("%s: generating pad control block", node_name)
+ self.gen_ctrl.PadType = at_pad_ctrl
+
+ attrs = {
+ 'in_size': in_q.dtype_bits//8 if in_q.signed else -in_q.dtype_bits//8,
+ 'out_size': out_q.dtype_bits//8 if out_q.signed else -out_q.dtype_bits//8,
+ 'feat': in_dim.c,
+ 'width': in_dim.w,
+ 'height': in_dim.h,
+ 'kop_pool': at_pool_params.PoolOper,
+ 'fpx': at_pool_params.Fpx,
+ 'fpy': at_pool_params.Fpy,
+ 'dpx': at_pool_params.Dpx,
+ 'dpy': at_pool_params.Dpy,
+ 'spx': at_pool_params.Spx,
+ 'spy': at_pool_params.Spy,
+ 'pool_pad': at_pool_params.PoolPad,
+ 'kop_act': act_op
+ }
+
+ extra_attrs = {
+ 'cname': cname,
+ 'node_name': node_name
+ }
+ super().__init__(attrs, extra_attrs, gen_ctrl=gen_ctrl)
+
+class GlobalPoolActKernel(NewAutoTilerKernel):
+ CALL_TEMPLATE = """// generator for {node_name}
+CNN_GlobalPoolAct_fp16("{cname}", {gen_ctrl},
+ {feat}, {feat}, {width}, {height},
+ {kop_pool}, {kop_act});
+"""
+ def __init__(self, node_name, cname, pool_params, pool_q, act_params, act_q, force_relu, gen_ctrl=None):
+ if gen_ctrl is None:
+ self.gen_ctrl = gen_ctrl = GenCtrl(None, cname=cname)
+ else:
+ gen_ctrl.cname = cname
+ self.gen_ctrl = gen_ctrl
+
+ at_pool_params = gen_globalpool_at_params(pool_params)
+ in_dim = pool_params.in_dims[0]
+ out_dim = pool_params.out_dims[0]
+ in_q = pool_q.in_qs[0]
+ out_q = pool_q.out_qs[0]
+ reduce_sz = reduce(lambda x, y: x * y, (sz for idx, sz in enumerate(in_dim.shape)
+ if idx not in pool_params.axis), 1)
+ #self.c = in_dim.size()/reduce_sz
+ feat = reduce_sz
+ height, width = balanced_divisors(in_dim.size()/reduce_sz)
+
+ if act_params is not None:
+ act_op = gen_activation_op(
+ act_params.activation, force_relu=force_relu, asymmetric=act_q.in_qs[0].zero_point != 0)
+ if out_dim is None:
+ out_dim = act_params.out_dims[0].expand_to_chw()
+ out_q = act_q.out_qs[0]
+ else:
+ act_op = "KOP_NONE"
+
+ attrs = {
+ 'in_size': in_q.dtype_bits//8 if in_q.signed else -in_q.dtype_bits//8,
+ 'out_size': out_q.dtype_bits//8 if out_q.signed else -out_q.dtype_bits//8,
+ 'feat': feat,
+ 'width': width,
+ 'height': height,
+ 'kop_pool': at_pool_params.GlobalPoolOper,
+ 'kop_act': act_op
+ }
+
+ extra_attrs = {
+ 'cname': cname,
+ 'node_name': node_name
+ }
+ super().__init__(attrs, extra_attrs, gen_ctrl=gen_ctrl)
diff --git a/tools/nntool/graph/manipulations/adjust_order.py b/tools/nntool/graph/manipulations/adjust_order.py
index 4067c3397..ef489fb3d 100644
--- a/tools/nntool/graph/manipulations/adjust_order.py
+++ b/tools/nntool/graph/manipulations/adjust_order.py
@@ -26,34 +26,35 @@
LOG = logging.getLogger("nntool." + __name__)
-def adjust_order(G, reshape_weights=True, postprocess=True, debug_function=None, one_cycle=False, single_step=False):
- opts = {'reshape_weights': reshape_weights}
- selector = AdjusterBase.get_all_handlers(opts)
- LOG.info("adding transposes to correct tensor order for AT kernels")
- ConstantInputParameters.clear_compression_state(G)
- for node in G.nodes(node_classes=tuple(selector)):
- adjusters = selector[node.__class__]
- for adjuster, attrs in adjusters:
- if attrs:
- not_selected = False
- for attr, val in attrs.items():
- if not hasattr(node, attr):
- not_selected = True
- break
- if callable(val):
- if not val(getattr(node, attr)):
+def adjust_order(G, reshape_weights=True, postprocess=True, debug_function=None, steps=None, single_step=False):
+ if steps is None:
+ opts = {'reshape_weights': reshape_weights}
+ selector = AdjusterBase.get_all_handlers(opts)
+ LOG.info("adding transposes to correct tensor order for AT kernels")
+ ConstantInputParameters.clear_compression_state(G)
+ for node in G.nodes(node_classes=tuple(selector)):
+ adjusters = selector[node.__class__]
+ for adjuster, attrs in adjusters:
+ if attrs:
+ not_selected = False
+ for attr, val in attrs.items():
+ if not hasattr(node, attr):
not_selected = True
break
- elif getattr(node, attr) != val:
- not_selected = True
- break
- if not_selected:
- continue
- adjuster.adjust(G, node)
- break
- add_dimensions(G)
+ if callable(val):
+ if not val(getattr(node, attr)):
+ not_selected = True
+ break
+ elif getattr(node, attr) != val:
+ not_selected = True
+ break
+ if not_selected:
+ continue
+ adjuster.adjust(G, node)
+ break
+ add_dimensions(G)
if debug_function:
debug_function(G)
- if postprocess:
- eliminate_transposes(G, debug_function=debug_function, one_cycle=one_cycle, single_step=single_step)
- add_dimensions(G)
+ if steps is not None or postprocess:
+ eliminate_transposes(G, debug_function=debug_function, steps=steps, single_step=single_step)
+ # add_dimensions(G)
diff --git a/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py b/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py
index 3be578e8d..86174f2f4 100644
--- a/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py
+++ b/tools/nntool/graph/manipulations/eliminate_transposes/eliminate_transposes.py
@@ -949,19 +949,23 @@ def delete_step_idx(G, action: DeleteTransposeAction):
return G.in_edges(action.node)[0].from_node.step_idx
-def eliminate_transposes(G, debug_function=None, one_cycle=False, single_step=False, do_silly=True):
+def eliminate_transposes(G, debug_function=None, steps=None, single_step=False, do_silly=True):
info("eliminating unnecessary transposes")
found_results = True
pass_count = 0
while found_results:
+ if steps is not None:
+ if pass_count >= steps:
+ break
+ else:
+ if pass_count >= 50:
+ raise ValueError(
+ "Sorry, eliminate transposes seems to be stuck in a loop. Please report to GreenWaves.")
pass_count += 1
- if pass_count > (200 if single_step else 50):
- raise ValueError(
- "Sorry, eliminate transposes is stuck in a loop. Please report to GreenWaves.")
found_results = False
visited_nodes = set()
actions = []
- info(f"search for transposes +++ PASS {pass_count}")
+ info(f"search for transposes +++ STEP {pass_count}")
transposes = G.nodes(node_classes=TransposeParameters)
while transposes:
transpose_node = transposes.pop(0)
@@ -1037,7 +1041,7 @@ def eliminate_transposes(G, debug_function=None, one_cycle=False, single_step=Fa
actions += cur_actions_up
visited_nodes |= set(cur_visited_up.nodes)
visited_nodes.add(transpose_node)
- if single_step:
+ if single_step or steps is not None:
break
# if transpose cannot be removed upwards movement push the transpose down if it actually moved
elif down_count > 0 or (down_count == 0 and transpose_moved(G, cur_actions_down)):
@@ -1047,7 +1051,7 @@ def eliminate_transposes(G, debug_function=None, one_cycle=False, single_step=Fa
actions += cur_actions_down
visited_nodes |= set(cur_visited_down.nodes)
visited_nodes.add(transpose_node)
- if single_step:
+ if single_step or steps is not None:
break
else:
info(
@@ -1065,7 +1069,4 @@ def eliminate_transposes(G, debug_function=None, one_cycle=False, single_step=Fa
G.add_dimensions()
if debug_function:
debug_function(G)
- if one_cycle:
- LOG.info("cycle complete")
- break
LOG.info("no further transpose sequences found")
diff --git a/tools/nntool/graph/nngraph.py b/tools/nntool/graph/nngraph.py
index 2d4353ef1..abf34a24d 100644
--- a/tools/nntool/graph/nngraph.py
+++ b/tools/nntool/graph/nngraph.py
@@ -290,9 +290,10 @@ def nodes_iterator(self, yield_fusions=True):
else:
yield (step_idx, node, None, None)
- def adjust_order(self, reshape_weights=True, postprocess=True, debug_function=None, one_cycle=False, single_step=False):
+ def adjust_order(self, reshape_weights=True, postprocess=True, debug_function=None, steps=None, single_step=False):
adjust_order(self, reshape_weights=reshape_weights,
- postprocess=postprocess, debug_function=debug_function, one_cycle=one_cycle, single_step=single_step)
+ postprocess=postprocess, debug_function=debug_function,
+ steps=steps, single_step=single_step)
LOG.info("adjusted order")
self.graph_identity.is_adjusted = True
diff --git a/tools/nntool/graph/types/others.py b/tools/nntool/graph/types/others.py
index b9c11d385..2d033305c 100644
--- a/tools/nntool/graph/types/others.py
+++ b/tools/nntool/graph/types/others.py
@@ -668,7 +668,7 @@ def __init__(self, *args, old_shape=None, shape=None, **kwargs):
if not isinstance(shape, Dim):
shape = Dim.unnamed(shape)
if old_shape is not None and not isinstance(old_shape, Dim):
- old_shape = Dim.unnamed(shape)
+ old_shape = Dim.unnamed(old_shape)
assert shape.is_ordered and (old_shape is None or old_shape.is_ordered)
self._shape = shape
self._old_shape = old_shape
diff --git a/tools/nntool/importer/onnx/handlers/backend/gru.py b/tools/nntool/importer/onnx/handlers/backend/gru.py
index 2b363dc9f..6e6bb8c2d 100644
--- a/tools/nntool/importer/onnx/handlers/backend/gru.py
+++ b/tools/nntool/importer/onnx/handlers/backend/gru.py
@@ -36,11 +36,10 @@ def _common(cls, node, **kwargs):
input_shapes = [inp[2].shape if inp else None for inp in inputs]
x = inputs[0]
- seq_len = input_shapes[0][0]
+ seq_len = input_shapes[0][1] if node.attrs.get("layout", 0) else input_shapes[0][0]
if seq_len is None:
- logger.warning("sequence length is variable in size. forcing to 20. "
- "reexport your graph with sequence length set to the maxmimum sequence size")
- seq_len = 20
+ seq_len = 1
+
input_size = input_shapes[0][2]
hidden_size = node.attrs["hidden_size"]
direction = node.attrs.get("direction", "forward")
diff --git a/tools/nntool/importer/onnx/handlers/backend/lstm.py b/tools/nntool/importer/onnx/handlers/backend/lstm.py
index 3d4eff864..c043fd030 100644
--- a/tools/nntool/importer/onnx/handlers/backend/lstm.py
+++ b/tools/nntool/importer/onnx/handlers/backend/lstm.py
@@ -35,7 +35,9 @@ def _common(cls, node, **kwargs):
input_shapes = [inp[2].shape if inp else None for inp in inputs]
x = inputs[0]
- seq_len = input_shapes[0][0]
+ seq_len = input_shapes[0][1] if node.attrs.get("layout", 0) else input_shapes[0][0]
+ if seq_len is None:
+ seq_len = 1
input_size = input_shapes[0][2]
hidden_size = node.attrs["hidden_size"]
direction = node.attrs.get("direction", "forward")
diff --git a/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py b/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py
index 77c796f0a..0d6c3ddd0 100644
--- a/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py
+++ b/tools/nntool/importer/onnx/handlers/backend/pad_mixin.py
@@ -13,9 +13,28 @@
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
+from itertools import chain
+
from graph.dim import PadDim
+
class PadMixin(object):
+ @classmethod
+ def mix_pads(cls, val):
+ if len(val) == 0:
+ pads_start = [0, 0]
+ pads_end = [0, 0]
+ elif len(val) == 2:
+ pads_start = [0, val[0]]
+ pads_end = [0, val[1]]
+ elif len(val) == 4:
+ pads_start = val[:2]
+ pads_end = val[2:]
+ else:
+ raise ValueError('unexpected pad lenght')
+
+ return list(chain(*zip(pads_start, pads_end)))
+
@classmethod
def pad_start_with(cls, val, pad_val, dlen):
return pad_val * (dlen - len(val)) + val
@@ -23,9 +42,7 @@ def pad_start_with(cls, val, pad_val, dlen):
@classmethod
def calc_pad_dim(cls, node, expected_len):
if "auto_pad" not in node.attrs or node.attrs["auto_pad"] == "NOTSET":
- pads = cls.pad_start_with(node.attrs.get("pads", []), [0], expected_len)
- pads = pads if len(pads) < 4 else [pads[0], pads[2], pads[1], pads[3]]
- pad_dim = PadDim(*pads)
+ pad_dim = PadDim(*cls.mix_pads(node.attrs.get("pads", [])))
elif node.attrs["auto_pad"] == "VALID":
pad_dim = PadDim.valid()
elif node.attrs["auto_pad"] == "SAME_UPPER":
diff --git a/tools/nntool/importer/onnx/handlers/backend/rnn.py b/tools/nntool/importer/onnx/handlers/backend/rnn.py
index 0d51005cb..ce22a6340 100644
--- a/tools/nntool/importer/onnx/handlers/backend/rnn.py
+++ b/tools/nntool/importer/onnx/handlers/backend/rnn.py
@@ -35,7 +35,10 @@ def _common(cls, node, **kwargs):
input_shapes = [inp[2].shape if inp else None for inp in inputs]
x = inputs[0]
- seq_len = input_shapes[0][0]
+ seq_len = input_shapes[0][1] if node.attrs.get("layout", 0) else input_shapes[0][0]
+ if seq_len is None:
+ seq_len = 1
+
input_size = input_shapes[0][2]
hidden_size = node.attrs["hidden_size"]
direction = node.attrs.get("direction", "forward")
diff --git a/tools/nntool/importer/onnx/handlers/backend/rnn_mixin.py b/tools/nntool/importer/onnx/handlers/backend/rnn_mixin.py
index 2d8ddbd0c..6f8a17329 100644
--- a/tools/nntool/importer/onnx/handlers/backend/rnn_mixin.py
+++ b/tools/nntool/importer/onnx/handlers/backend/rnn_mixin.py
@@ -36,20 +36,20 @@ def deep_update(cls, x, y):
@classmethod
def extract_weights(cls, weights, hidden_size, keys, num_directions):
return {
- 'forward' if dir == 0 else 'backward':
+ 'forward' if direction == 0 else 'backward':
{keys[i]: arr.reshape((hidden_size, -1))
for i, arr in enumerate(np.split(dir_arr, len(keys), axis=1))}
- for dir, dir_arr in enumerate(np.split(weights, num_directions, axis=0))
+ for direction, dir_arr in enumerate(np.split(weights, num_directions, axis=0))
}
@classmethod
def extract_biases(cls, biases, hidden_size, keys, num_directions):
biases = biases.reshape((num_directions, len(keys), hidden_size))
return {
- 'forward' if dir == 0 else 'backward':
+ 'forward' if direction == 0 else 'backward':
{keys[i]: arr.reshape((hidden_size))
for i, arr in enumerate(np.split(dir_arr, len(keys), axis=1))}
- for dir, dir_arr in enumerate(np.split(biases, num_directions, axis=0))
+ for direction, dir_arr in enumerate(np.split(biases, num_directions, axis=0))
}
@classmethod
diff --git a/tools/nntool/importer/tflite2/handlers/backend/concatenation.py b/tools/nntool/importer/tflite2/handlers/backend/concatenation.py
index 8bea05cd9..dd10ed7fb 100644
--- a/tools/nntool/importer/tflite2/handlers/backend/concatenation.py
+++ b/tools/nntool/importer/tflite2/handlers/backend/concatenation.py
@@ -13,12 +13,13 @@
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
+from collections import Counter
from functools import reduce
import numpy as np
from graph.dim import Dim
from graph.types import ConcatParameters, ConstantInputParameters, NNEdge
-from graph.types.others import ReshapeParameters
+from graph.types.others import CopyParameters, ReshapeParameters
from importer.common.constant_mixin import ConstantMixin
from importer.common.provisional_dim import ProvisionalDim
from importer.tflite2.common import LOG
@@ -63,10 +64,16 @@ def _common(cls, node: TFLiteNode, **kwargs):
buffer_idxes = [tensor.buffer_idx for tensor in node.input]
non_zero_idxes = [idx for idx in buffer_idxes if idx != 0]
- if len(set(non_zero_idxes)) != len(non_zero_idxes):
- raise NotImplementedError(
- "concats with multiple versions of the same input are not supported. "
- "This is normally a graph design problem.")
+ duplicates = [idx for idx, count in Counter(non_zero_idxes).items() if count > 1]
+ if duplicates:
+ LOG.warning(f'concat {node.name} has duplicate inputs. Inserting copies but this is not very efficient.')
+ for idx in duplicates:
+ dup_idxes = [i for i, x in enumerate(buffer_idxes) if x == idx]
+ for dup_idx in dup_idxes[1:]:
+ cparams = CopyParameters(G.unique_name(f'{node.name}_dup_{dup_idxes[0]}_{dup_idx}'))
+ dup_inp = inputs[dup_idx]
+ G.add_edge(NNEdge(from_node=dup_inp[0], from_idx=dup_inp[1], to_node=cparams))
+ inputs[dup_idx] = tuple([cparams, 0] + list(dup_inp[2:]))
axis = node_opts.Axis()
if any(inp_shape[axis] is None for inp_shape in inp_shapes):
diff --git a/tools/nntool/importer/tflite2/handlers/backend/fully_connected.py b/tools/nntool/importer/tflite2/handlers/backend/fully_connected.py
index ebe001ab5..7593d3a69 100644
--- a/tools/nntool/importer/tflite2/handlers/backend/fully_connected.py
+++ b/tools/nntool/importer/tflite2/handlers/backend/fully_connected.py
@@ -13,11 +13,10 @@
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
-from graph.types.tensor_arithmetic import MatMulTransposedParameters
import numpy as np
from graph.dim import Dim, FcFilterDim
from graph.types import (ConstantInputParameters, FcParameters,
- MatMulOpParameters, NNEdge, ReshapeParameters,
+ MatMulTransposedParameters, NNEdge, ReshapeParameters,
TransposeParameters)
from importer.common.provisional_dim import ProvisionalDim
from importer.tflite2.common import check
@@ -42,13 +41,14 @@ def _common(cls, node, **kwargs):
all_nodes = kwargs['all_nodes']
keep_dims = node_opts.KeepNumDims()
- check(not keep_dims,
- f'keep dims on Fully Connected {node.name} is not supported')
+ # check(not keep_dims,
+ # f'keep dims on Fully Connected {node.name} is not supported')
inputs = [all_nodes[t] if t is not None else None for t in node.input]
x = inputs[0]
- x_known_shape = x[2].known_shape
+ x_shape = x[2]
+ x_known_shape = x_shape.known_shape
inp_sz = np.prod(np.array(x_known_shape))
weights = inputs[1]
weights_node = weights[0]
@@ -57,6 +57,20 @@ def _common(cls, node, **kwargs):
f'bad filter shape {weights_shape} in {node.name}')
out_c = weights_shape[0]
batch_size = inp_sz // weights_shape[1]
+
+ keep_dims = node_opts.KeepNumDims()
+ if keep_dims:
+ if x_shape.shape[-1] != weights_shape[1]:
+ raise ValueError(
+ f'Keep dims set on {node.name} but last input dimension does not match weights')
+ out_shape = x_shape.shape.copy()
+ out_shape[-1] = out_c
+ elif batch_size > 1:
+ out_shape = (batch_size, out_c)
+ else:
+ out_shape = (None, out_c)
+ real_out_shape = tuple(dim for dim in out_shape if dim is not None)
+
filt_dim = FcFilterDim(weights_shape[0], weights_shape[1])
node.input[1].used = True
@@ -70,7 +84,7 @@ def _common(cls, node, **kwargs):
bias_node = ConstantInputParameters(f'{node.name}_bias',
dims=Dim.unnamed([out_c]),
value=np.zeros([out_c],
- dtype=np.float32))
+ dtype=np.float32))
if batch_size > 1:
# add a reshape to force the size of the input to batch * in_c
@@ -91,12 +105,15 @@ def _common(cls, node, **kwargs):
cls.new_load_filter_parameters(G, params, weights_shape, 0,
node.input[0], weights_node,
bias_node, node.output[0], opts)
- trans2 = TransposeParameters(G.unique_name(f'{node.name}_tin2'), transpose=(1, 0))
- G.add_edge(NNEdge(from_node=link[0], to_node=params, from_idx=link[1]))
- G.add_edge(NNEdge(from_node=weights_node, to_node=params, to_idx=1))
+ trans2 = TransposeParameters(G.unique_name(
+ f'{node.name}_tin2'), transpose=(1, 0))
+ G.add_edge(
+ NNEdge(from_node=link[0], to_node=params, from_idx=link[1]))
+ G.add_edge(NNEdge(from_node=weights_node,
+ to_node=params, to_idx=1))
#G.add_edge(NNEdge(from_node=trans2, to_node=params, to_idx=1))
G.add_edge(NNEdge(from_node=bias_node, to_node=params, to_idx=2))
- out_shape = [batch_size, out_c]
+ fc_shape = (batch_size, out_c)
else:
ker_in_order = None
ker_out_order = None
@@ -116,12 +133,17 @@ def _common(cls, node, **kwargs):
G.add_edge(NNEdge(from_node=bias_node, to_node=params, to_idx=2))
G.add_edge(NNEdge(from_node=link[0], to_node=params,
from_idx=link[1], to_idx=0))
-
- out_shape = [None, out_c]
+ fc_shape = (out_c,)
pout_dims = ProvisionalDim(out_shape)
-
aparams = cls.fuse_activation(node_opts, node.name, params, **kwargs)
+
+ if real_out_shape != fc_shape:
+ rparams = ReshapeParameters(G.unique_name(f'{node.name}_keepdims'),
+ old_shape=fc_shape, shape=real_out_shape)
+ G.add_edge(NNEdge(from_node=aparams, to_node=rparams))
+ aparams = rparams
+
all_nodes[node.output[0]] = (aparams, 0, pout_dims)
return params
diff --git a/tools/nntool/interpreter/commands/adjust.py b/tools/nntool/interpreter/commands/adjust.py
index 73e04b5f9..f03975ac8 100644
--- a/tools/nntool/interpreter/commands/adjust.py
+++ b/tools/nntool/interpreter/commands/adjust.py
@@ -16,19 +16,36 @@
from cmd2 import Cmd2ArgumentParser, with_argparser
from interpreter.nntool_shell_base import NNToolShellBase
+
class AdjustCommand(NNToolShellBase):
# ADJUST COMMAND
parser_adjust = Cmd2ArgumentParser()
- parser_adjust.add_argument('-n', '--no_postprocess',
- action='store_true', help='Don\'t try to eliminate transposes')
- parser_adjust.add_argument('-o', '--one_cycle',
- action='store_true', help='Do one cycle of post processing')
+ parser_adjust_choices = parser_adjust.add_mutually_exclusive_group()
+ parser_adjust_choices.add_argument(
+ '-n', '--no_postprocess',
+ action='store_true', help='Don\'t try to eliminate transposes')
+ parser_adjust_choices.add_argument(
+ '-o', '--one_step',
+ action='store_true', help='Do one step of post processing')
+ parser_adjust_choices.add_argument(
+ '-s', '--steps',
+ type=int, help='Do fixed number of steps of post processing')
+ parser_adjust_choices.add_argument(
+ '-i', '--individual_step',
+ action='store_true', help='Do the post processing step by step')
+
@with_argparser(parser_adjust)
def do_adjust(self, args):
"""
Adjust activation and parameter tensors to match AutoTiler order.
Must be run before generating code."""
self._check_graph()
- self.G.adjust_order(postprocess=not args.no_postprocess, one_cycle=args.one_cycle)
+ if args.one_step:
+ steps = 1
+ elif args.steps:
+ steps = args.steps
+ else:
+ steps = None
+ self.G.adjust_order(
+ postprocess=not args.no_postprocess, steps=steps, single_step=args.individual_step)
self.G.add_dimensions()
-
\ No newline at end of file
diff --git a/tools/nntool/quantization/float/quantizers/pool_float.py b/tools/nntool/quantization/float/quantizers/pool_float.py
new file mode 100644
index 000000000..aeb32042d
--- /dev/null
+++ b/tools/nntool/quantization/float/quantizers/pool_float.py
@@ -0,0 +1,52 @@
+# Copyright (C) 2020 GreenWaves Technologies, SAS
+
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as
+# published by the Free Software Foundation, either version 3 of the
+# License, or (at your option) any later version.
+
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU Affero General Public License for more details.
+
+# You should have received a copy of the GNU Affero General Public License
+# along with this program. If not, see <https://www.gnu.org/licenses/>.
+
+
+from quantization.quantizer_options import HWC_OPTION
+import numpy as np
+from bfloat16 import bfloat16
+from graph.types import PoolingParameters
+from quantization.float.float_quantization_handler import \
+ FloatQuantizionHandler
+from quantization.new_qrec import QRec
+from quantization.qtype import QType
+from quantization.qtype_constraint import MatchAll
+from quantization.unified_quantization_handler import (in_qs_constraint,
+ out_qs_constraint,
+ params_type,
+ options)
+
+
+@options(
+ HWC_OPTION
+)
+@params_type(PoolingParameters)
+@in_qs_constraint(MatchAll({'dtype': set([np.float32, np.float16, bfloat16])}))
+@out_qs_constraint(MatchAll({'dtype': set([np.float32, np.float16, bfloat16])}))
+class FloatDefault(FloatQuantizionHandler):
+ @classmethod
+ def _quantize(cls, params, in_qs, stats, **kwargs):
+ force_out_qs, dtype = cls.get_float_opts(**kwargs)
+ if force_out_qs and any(qtype.dtype != dtype for qtype in force_out_qs if qtype is not None):
+ return None
+ opts = kwargs['opts']
+ if opts['hwc']:
+ cls.check_order(params, [['h', 'w', 'c']], [['h', 'w', 'c']])
+ elif params.in_dims_hint:
+ cls.check_order(params, [['c', 'h', 'w']], [['c', 'h', 'w']])
+ # all inputs and outputs are set to the required float type
+ return QRec.float(in_qs=[QType(dtype=dtype)],
+ out_qs=[QType(dtype=dtype)],
+ float_dtype=dtype)
diff --git a/utils/bin/binary-size b/utils/bin/binary-size
index 1e05ee214..8918afc8d 100755
--- a/utils/bin/binary-size
+++ b/utils/bin/binary-size
@@ -110,6 +110,7 @@ groups = {
Group('PulpOS:kernel', ['pos_kernel.*', 'pos_soc_.*', 'pos_cbsys_.*']),
Group('PulpOS:fll', ['pos_fll.*', 'pos_freq.*', 'pi_freq_.*']),
Group('PulpOS:cpi', ['pos_cpi.*', 'pi_cpi_.*']),
+ Group('PulpOS:i2s', ['pos_i2s.*', 'pi_i2s_.*']),
Group('PulpOS:mram', ['pos_mram.*']),
Group('PulpOS:rtc', ['pos_rtc.*']),
Group('PulpOS:cluster', ['pos_cluster.*', 'pi_cluster_.*']),
diff --git a/utils/gap_configs/python/ips/iss/iss.py b/utils/gap_configs/python/ips/iss/iss.py
index ade8b4ad1..d19fbe1ee 100644
--- a/utils/gap_configs/python/ips/iss/iss.py
+++ b/utils/gap_configs/python/ips/iss/iss.py
@@ -34,7 +34,9 @@ class Iss(st.Component):
riscv_dbg_unit : bool, optional
True if a riscv debug unit should be included, False otherwise (default: False).
debug_binaries : list, optional
- A list of path to riscv binaries which can be used to get debug symbols for the assembly trace (default: []).
+ A list of path to riscv binaries debug info which can be used to get debug symbols for the assembly trace (default: []).
+ binaries : list, optional
+ A list of path to riscv binaries (default: []).
debug_handler : int, optional
The address where the core should jump when switching to debug mode (default: 0).
power_models : dict, optional
@@ -62,6 +64,7 @@ def __init__(self,
first_external_pcer: int=0,
riscv_dbg_unit: bool=False,
debug_binaries: list=[],
+ binaries: list=[],
debug_handler: int=0,
power_models: dict={},
power_models_file: str=None,
@@ -82,6 +85,7 @@ def __init__(self,
'first_external_pcer': first_external_pcer,
'riscv_dbg_unit': riscv_dbg_unit,
'debug_binaries': debug_binaries,
+ 'binaries': binaries,
'debug_handler': debug_handler,
'power_models': power_models,
'cluster_id': cluster_id,
diff --git a/utils/rules/pmsis_rules.mk b/utils/rules/pmsis_rules.mk
index ce9288edc..73c6c59be 100644
--- a/utils/rules/pmsis_rules.mk
+++ b/utils/rules/pmsis_rules.mk
@@ -421,4 +421,4 @@ profiler:
cd $(BUILDDIR) && export PULP_CONFIG_FILE=$(BUILDDIR)/gvsoc_config.json && profiler $(BUILDDIR) $(BIN) gvsoc_config.json --signal-tree-file=$(PROFILER_SIGNAL_TREE)
size:
- @$(GAP_SDK_HOME)/utils/bin/binary-size --binary=$(BIN) --depth=10 --groups=$(PMSIS_OS) >> $(BIN).size
+ $(GAP_SDK_HOME)/utils/bin/binary-size --binary=$(BIN) --depth=10 --groups=$(PMSIS_OS)
From db984bc652efaaca0caaece17dc1270bf8a13685 Mon Sep 17 00:00:00 2001
From: yao <yao.zhang@greenwaves-technologies.com>
Date: Mon, 20 Dec 2021 16:08:53 +0100
Subject: [PATCH 3/3] Remove unused files
---
tools/autotiler_v3_get/LICENSE | 204 --------------
tools/autotiler_v3_get/Makefile | 25 --
tools/autotiler_v3_get/download_tiler.sh | 33 ---
tools/autotiler_v3_get/get_tiler.py | 335 -----------------------
4 files changed, 597 deletions(-)
delete mode 100644 tools/autotiler_v3_get/LICENSE
delete mode 100644 tools/autotiler_v3_get/Makefile
delete mode 100755 tools/autotiler_v3_get/download_tiler.sh
delete mode 100644 tools/autotiler_v3_get/get_tiler.py
diff --git a/tools/autotiler_v3_get/LICENSE b/tools/autotiler_v3_get/LICENSE
deleted file mode 100644
index e8567dbda..000000000
--- a/tools/autotiler_v3_get/LICENSE
+++ /dev/null
@@ -1,204 +0,0 @@
-LICENSE AGREEMENT FOR USE OF THE AUTOTILER
-
-Definitions
-“Affiliate(s)” of a person means any entity that Controls, is Controlled by, or
-is under common Control with such person, where “Control” of a person or
-entity (for purposes of this definition only) means the possession, directly
-or indirectly, of the power to direct or cause the direction of the
-management, operating policies, or assets of that person or entity, whether by
-way of ownership of more than 50% of its voting or equity securities or
-assets or by way of contract, management agreement, voting trust, or
-otherwise.
-“Authorized User(s)” means the individuals You authorize to access the Software,
-including Your employees or third parties that access the Software solely on
-Your behalf for Your internal and/or personal operations.
-“GreenWaves Technologies” “we,” “our” or “us” means GreenWaves Technologies SAS,
-a company incorporated under the laws of France, and registered with the Trade
-and Companies Registry of Grenoble, under number 808 076 582, having its
-principal place of business at 28 cours Jean Jaurès, 38000 Grenoble, France.
-“Documentation” means the GreenWaves Technologies user or technical manuals,
-specifications, privacy data sheets or other information applicable to the
-Software.
-“Personal Data” means any information that can be used to identify an individual
-and is required in order to enable You to download the Software, i.e Your name
-and email address, as well as Your company name.
-“Software” means Greenwaves Technologies’ software Autotiler which is a
-downloadable file enabling easier programming of GreenWaves Technologies’ GAP
-products by automatic generation of code for memory tiling and transfers between
-memory levels.
-“You” and “Your” means the individual or legal entity downloading the Software.
-
-1. Delivery
-The Software is deemed to be delivered and accepted by You at the date it is
-made available for download by You. By downloading the Software, You agree to
-be bound by the terms of this EULA. If You do not have the authority to enter
-into this EULA or You do not agree with its terms, do not Use the Software and
-uninstall it.
-
-2. Scope of the license
-Subject to Your compliance with this EULA, GreenWaves Technologies grants You a
-worldwide, non-exclusive and non-transferable license to use the Software and
-related Documentation, as well as the code automatically generated by the
-Software for memory tiling and transfers between memory levels, for Your
-internal and/or personal use in accordance with the terms set forth, solely
-in order to program GreenWaves Technologies’ products, or to execute or simulate
-software on another system with a final objective of developing code to be used
-on GreenWaves Technologies' product. GreenWaves Technologies also grants You a
-worldwide, non-exclusive and non-transferable license to compile the code
-automatically generated by the Software, for the exclusive purpose of being run
-on a GreenWaves Technologies’ product as embedded in Your product. Free or open
-source software not owned by GreenWaves Technologies is subject to separate
-license terms made available with the Software Development Kit or at
-https://github.com/GreenWaves-Technologies/gap_sdk.
-
-3. Ownership
-GreenWaves Technologies retains exclusive ownership in all intellectual property
-rights in and to the Software and all underlying technologies and associated
-Documentation related thereto.
-
-4. Limitations and Restrictions
-Unless expressly authorized by GreenWaves Technologies in writing or otherwise
-permitted under applicable law, You shall not:
-sell, resell, transfer, sublicense, or assign Your rights under this license;
-modify, adapt or create derivative works;
-reverse engineer, decompile, decrypt, disassemble or otherwise attempt to derive
-the source code;
-make the functionality available to third parties, whether as an application
-service provider, or on an outsourcing, membership or subscription, rental,
-service bureau, cloud service, managed or hosted service, or other similar
-basis;
-use Software that is licensed for the development of code to run products other
-than GreenWaves Technologies’ GAP products;
-remove, modify, or conceal any product identification, copyright, proprietary,
-intellectual property notices or other marks from the Software and related
-Documentation, as well as from the code automatically generated by the Software.
-
-5. Use by Authorized Users
-You may allow Authorized Users to use the Software solely on Your behalf for
-Your internal and/or personal operations. You are responsible for ensuring that
-Authorized Users comply with the terms of this EULA and You are liable for any
-breach of the same by such Authorized Users.
-
-6. Limited Warranty
-To the extent permitted by applicable law, You expressly acknowledge and agree
-that the Software, as well as the code it automatically generates, are provided
-“AS-IS”, used at Your sole risk and without support or any express or implied
-warranty of any kind or indemnity for any problems or issues. GreenWaves
-Technologies hereby disclaims all warranties and conditions with respect to the
-Software, either express, implied or statutory, including, but not limited to,
-the implied warranties and/or conditions of merchantability, satisfactory
-quality, fitness for a particular purpose, accuracy, quiet enjoyment, and
-non-infringement of third party rights.
-GreenWaves Technologies does not warrant against interference with your
-enjoyment of the Software, that the functions contained in, or services
-performed or provided by, the Software will meet your requirements, that the
-operation of the Software will be uninterrupted or error-free, that defects in
-the Software will be corrected, or that the Software will be compatible or work
-with any third party software, applications or third party services other than
-those expressly identified by GreenWaves Technologies.
-You further acknowledge that the Software is not intended or suitable for use
-in situations or environments where the failure or time delays of, or errors or
-inaccuracies in, the content, data or information provided by the Software could
-lead to death, personal injury, or severe physical or environmental damage,
-including without limitation the operation of nuclear facilities, aircraft
-navigation or communication systems, air traffic control, life support or
-weapons systems.
-
-7. Limitations of Liability
-To the extent not prohibited by applicable law, in no event shall GreenWaves
-Technologies, its Affiliates, agents, employees or principals, be liable for the
-following, regardless of the theory of liability or whether relating to or
-arising out of this EULA, the Software or otherwise, even if a party has been
-advised of the possibility of such damages:
-indirect, incidental, exemplary, special or consequential damages;
-loss or corruption of data or interrupted or loss of business; or
-loss of revenue, profits, goodwill or anticipated sales or savings.
-Some jurisdictions do not allow the exclusion or limitation of liability for
-personal injury, or of incidental or consequential damages, so this limitation
-may not apply to You.
-To the extent permitted by applicable law, all liability of GreenWaves
-Technologies, its Affiliates, agents, employees, collectively, to You, whether
-based in warranty, contract, tort (including negligence), or otherwise (other
-than as may be required by applicable law in cases involving personal injury),
-shall not exceed, in the aggregate, the total amount of one thousand euros
-(€1,000). This limitation of liability for the Software is cumulative and not
-per incident.
-Nothing in this EULA limits or excludes any liability that cannot be limited or
-excluded under applicable law.
-
-8. Term and Termination
-Your license begins on the date the Software is downloaded and continues until
-terminated.
-To the extent permitted under applicable law, GreenWaves Technologies has
-the right to immediately suspend or terminate Your use of the Software if
-You breach Section 2, 4 or 5 of this EULA. Upon termination of the EULA,
-You must cease any further use of the Software, and destroy any copies of
-Software within Your control.
-
-9. Personal Data
-GreenWaves Technologies processes and uses Personal Data in order to enable You
-to download the Software and keep a track record of the persons who are bound
-by the present EULA. If You have accepted to receive news on GreenWaves
-Technologies products, You will receive emails with updates, new versions and
-new products available.
-Your Personal Data is not sold or otherwise transmitted to any third party. It
-is stored by GreenWaves Technologies for a period of 10 years further to Your
-download of the Software. If You have subscribed to GreenWaves Technologies
-newsletter, Your Personal Data will be stored as long as You do not unsubscribe
-from this service.
-GreenWaves Technologies will maintain appropriate administrative, physical and
-technical safeguards, which are designed to protect the security,
-confidentiality and integrity of Personal Data processed by GreenWaves
-Technologies.
-To the extent required by applicable law, in particular EU Regulation 2016/679
-relating to General Data Protection Rules, You may have the right to access
-certain Personal Data we process, request that we update or modify such Personal
-Data when inaccurate, object or restrict to our use of You Personal Data,
-withdraw Your consent at any time when GreenWaves Technologies processes Your
-Personal Data based on your consent, and request that we delete You Personal
-Data, by writing to websales@greenwaves-technologies.com.
-If You think that the way GreenWaves Technologies processes Your Personal Data
-does not comply with applicable data protection laws, you can contact the
-relevant competent data protection authority. You can obtain the information by
-contacting EU data protection authorities at
-http://ec.europa.eu/justice/article-29/structure/data-protection-authorities/index_en.htm.
-
-10. Export Controls
-You may not use or otherwise export or re-export the Software except as
-authorized by applicable laws and the laws of the jurisdictions in which the
-Software was obtained.
-
-11. Force Majeure
-Neither party will be responsible for failure of performance due to an event
-that is unforeseeable and beyond the affected party’s reasonable control,
-including accidents, severe weather events, acts of God, actions of any
-government agency, pandemic, acts of terrorism, or the stability or availability
-of the Internet or portions thereof.
-
-12. Complete Agreement
-If any portion of this EULA is found to be void or unenforceable, the remaining
-provisions of the EULA shall remain in full force and effect. Except as
-expressly stated or as expressly amended in a signed agreement, the EULA is the
-complete agreement between the parties with respect to the Software and
-supersedes all prior or contemporaneous communications, understandings or
-agreements (whether written or oral) regarding this subject matter.
-To the extent permitted by applicable law, the parties agree that the English
-version of the EULA will govern in the event of a conflict between it and any
-version translated into another language.
-
-
-
-13. Notices
-GreenWaves Technologies may provide You with notice via email and/or postings on
-the GreenWaves Technologies.com website.
-Notices to GreenWaves Technologies should be sent to GreenWaves Technologies, 28
-cours Jean Jaurès, 38000 GRENOBLE websales@greenwaves-technologies.com.
-
-14. Dispute Settlement and Governing Laws
-The validation, interpretation, modification, fulfillment, and dispute
-settlement of this EULA are governed by the laws of France, without regard
-to conflict of law principles. If any dispute arises concerning the content or
-performance of this EULA, the dispute shall be settled through amicable
-settlement. In the event that the dispute cannot be settled through negotiation,
-either party could submit the dispute to the French courts.
-
diff --git a/tools/autotiler_v3_get/Makefile b/tools/autotiler_v3_get/Makefile
deleted file mode 100644
index dc12e7fd7..000000000
--- a/tools/autotiler_v3_get/Makefile
+++ /dev/null
@@ -1,25 +0,0 @@
-TILER_VER=4.3.0
-export TILER_LIB=libtile.${TILER_VER}.a
-ifdef GAP_SDK_HOME
-export TILER_URL=$(GAP_SDK_HOME)/.tiler_url
-else
-export TILER_URL=.tiler_url
-endif
-
-all: Autotiler/LibTile.a
-
-clean:
- rm -rf Autotiler/LibTile*
- rm -f $(TILER_URL)
-
-ifeq (,$(wildcard $(TILER_URL)))
-$(TILER_URL): get_tiler.py
- python3 get_tiler.py
-endif
-
-Autotiler/LibTile.a: $(TILER_URL)
- mkdir -p Autotiler
- rm -rf Autotiler/LibTile*
- ./download_tiler.sh
-
-.PHONY: all clean
diff --git a/tools/autotiler_v3_get/download_tiler.sh b/tools/autotiler_v3_get/download_tiler.sh
deleted file mode 100755
index 0ca2e7b46..000000000
--- a/tools/autotiler_v3_get/download_tiler.sh
+++ /dev/null
@@ -1,33 +0,0 @@
-#!/usr/bin/env bash
-
-if [ -z "$TILER_LICENSE_AGREED" ]
-then
- echo "Please read carefully the autotiler agreement before proceeding with the download."
- read -n 1 -s -r -p "Press enter when you are ready to display the license."
-
- more LICENSE
-
- while [ true ] ; do
- read -n 1 -p "Do you agree with the license (y/n) ? " key
-
- echo
-
- if [ "$key" == "y" ]
- then
- break
- fi
-
- if [ "$key" == "n" ]
- then
- exit 0
- fi
-
- done
-fi
-
-echo ${TILER_LIB} | wget --no-use-server-timestamps --base=`cat $TILER_URL` --input-file=- -O Autotiler/LibTile.a
-if [ $? != 0 ]; then
- rm $TILER_URL
- rm -f Autotiler/LibTile.a
- exit 1
-fi
diff --git a/tools/autotiler_v3_get/get_tiler.py b/tools/autotiler_v3_get/get_tiler.py
deleted file mode 100644
index 34f1ec57c..000000000
--- a/tools/autotiler_v3_get/get_tiler.py
+++ /dev/null
@@ -1,335 +0,0 @@
-import requests, json, sys, re, os
-
-class bcolors:
- HEADER = '\033[95m'
- OKBLUE = '\033[94m'
- OKGREEN = '\033[92m'
- WARNING = '\033[93m'
- FAIL = '\033[91m'
- ENDC = '\033[0m'
- BOLD = '\033[1m'
- UNDERLINE = '\033[4m'
-
-Countries = [
- ('US', 'United States'),
- ('AF', 'Afghanistan'),
- ('AL', 'Albania'),
- ('DZ', 'Algeria'),
- ('AS', 'American Samoa'),
- ('AD', 'Andorra'),
- ('AO', 'Angola'),
- ('AI', 'Anguilla'),
- ('AQ', 'Antarctica'),
- ('AG', 'Antigua And Barbuda'),
- ('AR', 'Argentina'),
- ('AM', 'Armenia'),
- ('AW', 'Aruba'),
- ('AU', 'Australia'),
- ('AT', 'Austria'),
- ('AZ', 'Azerbaijan'),
- ('BS', 'Bahamas'),
- ('BH', 'Bahrain'),
- ('BD', 'Bangladesh'),
- ('BB', 'Barbados'),
- ('BY', 'Belarus'),
- ('BE', 'Belgium'),
- ('BZ', 'Belize'),
- ('BJ', 'Benin'),
- ('BM', 'Bermuda'),
- ('BT', 'Bhutan'),
- ('BO', 'Bolivia'),
- ('BA', 'Bosnia And Herzegowina'),
- ('BW', 'Botswana'),
- ('BV', 'Bouvet Island'),
- ('BR', 'Brazil'),
- ('BN', 'Brunei Darussalam'),
- ('BG', 'Bulgaria'),
- ('BF', 'Burkina Faso'),
- ('BI', 'Burundi'),
- ('KH', 'Cambodia'),
- ('CM', 'Cameroon'),
- ('CA', 'Canada'),
- ('CV', 'Cape Verde'),
- ('KY', 'Cayman Islands'),
- ('CF', 'Central African Rep'),
- ('TD', 'Chad'),
- ('CL', 'Chile'),
- ('CN', 'China'),
- ('CX', 'Christmas Island'),
- ('CC', 'Cocos Islands'),
- ('CO', 'Colombia'),
- ('KM', 'Comoros'),
- ('CG', 'Congo'),
- ('CK', 'Cook Islands'),
- ('CR', 'Costa Rica'),
- ('CI', 'Cote D`ivoire'),
- ('HR', 'Croatia'),
- ('CU', 'Cuba'),
- ('CY', 'Cyprus'),
- ('CZ', 'Czech Republic'),
- ('DK', 'Denmark'),
- ('DJ', 'Djibouti'),
- ('DM', 'Dominica'),
- ('DO', 'Dominican Republic'),
- ('TP', 'East Timor'),
- ('EC', 'Ecuador'),
- ('EG', 'Egypt'),
- ('SV', 'El Salvador'),
- ('GQ', 'Equatorial Guinea'),
- ('ER', 'Eritrea'),
- ('EE', 'Estonia'),
- ('ET', 'Ethiopia'),
- ('FK', 'Falkland Islands (Malvinas)'),
- ('FO', 'Faroe Islands'),
- ('FJ', 'Fiji'),
- ('FI', 'Finland'),
- ('FR', 'France'),
- ('GF', 'French Guiana'),
- ('PF', 'French Polynesia'),
- ('TF', 'French S. Territories'),
- ('GA', 'Gabon'),
- ('GM', 'Gambia'),
- ('GE', 'Georgia'),
- ('DE', 'Germany'),
- ('GH', 'Ghana'),
- ('GI', 'Gibraltar'),
- ('GR', 'Greece'),
- ('GL', 'Greenland'),
- ('GD', 'Grenada'),
- ('GP', 'Guadeloupe'),
- ('GU', 'Guam'),
- ('GT', 'Guatemala'),
- ('GN', 'Guinea'),
- ('GW', 'Guinea-bissau'),
- ('GY', 'Guyana'),
- ('HT', 'Haiti'),
- ('HN', 'Honduras'),
- ('HK', 'Hong Kong'),
- ('HU', 'Hungary'),
- ('IS', 'Iceland'),
- ('IN', 'India'),
- ('ID', 'Indonesia'),
- ('IR', 'Iran'),
- ('IQ', 'Iraq'),
- ('IE', 'Ireland'),
- ('IL', 'Israel'),
- ('IT', 'Italy'),
- ('JM', 'Jamaica'),
- ('JP', 'Japan'),
- ('JO', 'Jordan'),
- ('KZ', 'Kazakhstan'),
- ('KE', 'Kenya'),
- ('KI', 'Kiribati'),
- ('KP', 'Korea (North)'),
- ('KR', 'Korea (South)'),
- ('KW', 'Kuwait'),
- ('KG', 'Kyrgyzstan'),
- ('LA', 'Laos'),
- ('LV', 'Latvia'),
- ('LB', 'Lebanon'),
- ('LS', 'Lesotho'),
- ('LR', 'Liberia'),
- ('LY', 'Libya'),
- ('LI', 'Liechtenstein'),
- ('LT', 'Lithuania'),
- ('LU', 'Luxembourg'),
- ('MO', 'Macau'),
- ('MK', 'Macedonia'),
- ('MG', 'Madagascar'),
- ('MW', 'Malawi'),
- ('MY', 'Malaysia'),
- ('MV', 'Maldives'),
- ('ML', 'Mali'),
- ('MT', 'Malta'),
- ('MH', 'Marshall Islands'),
- ('MQ', 'Martinique'),
- ('MR', 'Mauritania'),
- ('MU', 'Mauritius'),
- ('YT', 'Mayotte'),
- ('MX', 'Mexico'),
- ('FM', 'Micronesia'),
- ('MD', 'Moldova'),
- ('MC', 'Monaco'),
- ('MN', 'Mongolia'),
- ('MS', 'Montserrat'),
- ('MA', 'Morocco'),
- ('MZ', 'Mozambique'),
- ('MM', 'Myanmar'),
- ('NA', 'Namibia'),
- ('NR', 'Nauru'),
- ('NP', 'Nepal'),
- ('NL', 'Netherlands'),
- ('AN', 'Netherlands Antilles'),
- ('NC', 'New Caledonia'),
- ('NZ', 'New Zealand'),
- ('NI', 'Nicaragua'),
- ('NE', 'Niger'),
- ('NG', 'Nigeria'),
- ('NU', 'Niue'),
- ('NF', 'Norfolk Island'),
- ('MP', 'Northern Mariana Islands'),
- ('NO', 'Norway'),
- ('OM', 'Oman'),
- ('PK', 'Pakistan'),
- ('PW', 'Palau'),
- ('PA', 'Panama'),
- ('PG', 'Papua New Guinea'),
- ('PY', 'Paraguay'),
- ('PE', 'Peru'),
- ('PH', 'Philippines'),
- ('PN', 'Pitcairn'),
- ('PL', 'Poland'),
- ('PT', 'Portugal'),
- ('PR', 'Puerto Rico'),
- ('QA', 'Qatar'),
- ('RE', 'Reunion'),
- ('RO', 'Romania'),
- ('RU', 'Russian Federation'),
- ('RW', 'Rwanda'),
- ('KN', 'Saint Kitts And Nevis'),
- ('LC', 'Saint Lucia'),
- ('VC', 'St Vincent/Grenadines'),
- ('WS', 'Samoa'),
- ('SM', 'San Marino'),
- ('ST', 'Sao Tome'),
- ('SA', 'Saudi Arabia'),
- ('SN', 'Senegal'),
- ('SC', 'Seychelles'),
- ('SL', 'Sierra Leone'),
- ('SG', 'Singapore'),
- ('SK', 'Slovakia'),
- ('SI', 'Slovenia'),
- ('SB', 'Solomon Islands'),
- ('SO', 'Somalia'),
- ('ZA', 'South Africa'),
- ('ES', 'Spain'),
- ('LK', 'Sri Lanka'),
- ('SH', 'St. Helena'),
- ('PM', 'St.Pierre'),
- ('SD', 'Sudan'),
- ('SR', 'Suriname'),
- ('SZ', 'Swaziland'),
- ('SE', 'Sweden'),
- ('CH', 'Switzerland'),
- ('SY', 'Syrian Arab Republic'),
- ('TW', 'Taiwan'),
- ('TJ', 'Tajikistan'),
- ('TZ', 'Tanzania'),
- ('TH', 'Thailand'),
- ('TG', 'Togo'),
- ('TK', 'Tokelau'),
- ('TO', 'Tonga'),
- ('TT', 'Trinidad And Tobago'),
- ('TN', 'Tunisia'),
- ('TR', 'Turkey'),
- ('TM', 'Turkmenistan'),
- ('TV', 'Tuvalu'),
- ('UG', 'Uganda'),
- ('UA', 'Ukraine'),
- ('AE', 'United Arab Emirates'),
- ('UK', 'United Kingdom'),
- ('UY', 'Uruguay'),
- ('UZ', 'Uzbekistan'),
- ('VU', 'Vanuatu'),
- ('VA', 'Vatican City State'),
- ('VE', 'Venezuela'),
- ('VN', 'Viet Nam'),
- ('VG', 'Virgin Islands (British)'),
- ('VI', 'Virgin Islands (U.S.)'),
- ('EH', 'Western Sahara'),
- ('YE', 'Yemen'),
- ('YU', 'Yugoslavia'),
- ('ZR', 'Zaire'),
- ('ZM', 'Zambia'),
- ('ZW', 'Zimbabwe')
-]
-
-def fatal_error(error):
- print (bcolors.FAIL + 'FATAL ERROR: ' + str(error) + bcolors.ENDC)
- sys.exit(1)
-
-def download_file(url):
- local_filename = url.split('/')[-1]
- r = requests.get(url, stream=True, timeout=3)
- r.raise_for_status()
- with open(local_filename, 'wb') as f:
- for chunk in r.iter_content(chunk_size=1024):
- if chunk:
- f.write(chunk)
- return local_filename
-
-ri = vars(__builtins__).get('raw_input',input)
-print("Registration is required to load the GAP8 AutoTiler library\n")
-
-print("In case you have already registered, you can directly enter the link")
-print("to the AutoTiler library that you received from an email.")
-print("Otherwise just press Enter.")
-url = ri("Enter URL from email: ")
-
-if url == '':
- print("You will be prompted for your name, company and email address and the")
- print("link for the AutoTiler libray will be sent to your email address.")
- print("This information is used uniquely to keep track of AutoTiler users.")
- forename = ri("Enter your first name: ")
- surname = ri("Enter your last name: ")
- company = ri("Enter your company name: ")
-
- while True:
- country = ri("Enter your country: ")
- country = country.upper()
- matches = []
- for c in Countries:
- if c[0] == country or c[1].upper() == country:
- matches = [c]
- break
- elif c[1].upper().startswith(country):
- matches.append(c)
- if len(matches) == 1:
- country = matches[0][0]
- break
- elif len(matches) > 1:
- print("Do you mean:")
- for c in matches:
- print("{} ({})".format(c[1], c[0]))
- print()
- else:
- print("I don't know that country. Please enter a valid country name or ISO code.")
-
- print("Country ", country)
-
- while True:
- email = ri("Enter your email address: ")
- if re.match(r"[^@]+@[^@]+\.[^@]+", email):
- break
- else:
- print("Please enter a valid email address")
-
- url = 'https://hooks.zapier.com/hooks/catch/2624512/e6qico/'
- payload = { 'forename': forename, 'surname': surname, 'company': company, 'email': email, 'country': country }
- headers = { 'content-type': 'application/json' }
-
- print("Triggering email ... please wait")
- try:
- response = requests.post(url, data=json.dumps(payload), headers=headers, timeout=3)
- response.raise_for_status()
- except requests.exceptions.RequestException as ex:
- fatal_error(ex)
-
-
- print("Please check your email and copy and paste the URL in the email below")
- print("Please keep this URL, you will ask to enter it in case you")
- print("install again the SDK.")
- url = ri("Enter URL from email: ")
-try:
- while True:
- if re.match(r"^https://.*/$", url):
- break
- else:
- print("Please enter a valid URL, it must start with https:// and")
- print("end with /.")
- url = ri("Enter URL from email: ")
- f = open(os.environ.get('TILER_URL'),"w+")
- f.write(url)
- f.close()
-except:
- fatal_error("problem writing file " + os.environ.get('TILER_URL'))
\ No newline at end of file