Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[MNN:Sync] Sync Internal 2.9.1 #2896

Merged
merged 1 commit into from
Jun 3, 2024
Merged

[MNN:Sync] Sync Internal 2.9.1 #2896

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 12 additions & 1 deletion docs/faq.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -225,8 +225,19 @@ OpenCL / Vulkan 采用静态变量自注册的方式往 MNN 主库注册后端.
- 一般是直接访问了 tensor 的 host
- 按 [输入数据](./inference/session.html#id8) 和[获取输出](./inference/session.html#id21) 里面的方式建host tensor 并 copy ，参考相关文档修改使用代码
- 是否可基于 deviceId 直接传 GPU 地址？
- 可以，需要理解 MNN GPU 内存布局并传上下文给 MNN ，但相关实现较复杂
- 可以，可以通过setDevicePtr设置输入VARP的GPU地址,通过copyToDevicePtr设置输出VARP拷贝到的GPU地址
- 相关使用参考tools/cpp/GpuInterTest.cpp
- 目前OPENCL推理支持OPENCL/OPENGL内存做输入输出。CUDA推理支持CUDA内存做输入输出
- 采用 MNN_Express 系列接口，可以支持模型之间的内存直接传递不做拷贝

### 多卡GPU上，用户指定特定GPU做推理问题

- 通过设置MNNDeviceContext结构体参数来指定特定GPU
- 通过设置platformSize、platformId、deviceId参数来进行指定
- 目前支持OpenCL和CUDA后端进行设置
- 具体可以参考：tools/cpp/testModel.cpp


## 性能相关
### 使用 GPU 时，调用 copyToHostTensor / copyFromHostTensor 非常慢
GPU 后端调用 copy 的时间包含两个部分
Expand Down
8 changes: 7 additions & 1 deletion express/Executor.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -595,7 +595,13 @@ void Executor::_makeCache(const std::vector<EXPRP>& expr, bool forceCPU) {
Session::ModeGroup group;
group.inputMode = Interpreter::Session_Input_User;
group.outputMode = Interpreter::Session_Output_User;
group.callBackMode = Interpreter::Session_Release;
auto globalExecutor = ExecutorScope::Current();
auto debug = globalExecutor->getDebugTools();
if (debug->after != nullptr && debug->before != nullptr) {
group.callBackMode = Interpreter::Session_Debug;
} else {
group.callBackMode = Interpreter::Session_Release;
}
group.memoryUsageMode = Interpreter::Session_Memory_Cache;
std::shared_ptr<ComputeCache> cahce(new ComputeCache);
for (auto& iter : dstExpr) {
Expand Down
16 changes: 15 additions & 1 deletion express/Utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@
#include <map>
#include <set>
#include <stack>
#include <MNN/expr/ExecutorScope.hpp>
#include "MNN_generated.h"
#include "core/TensorUtils.hpp"
#include "core/Session.hpp"
Expand Down Expand Up @@ -188,6 +189,9 @@ void* Executor::ComputeCache::mapOutput(int offset, Tensor* dest) {
//MNN_ASSERT(nullptr != ptr);
return ptr;
}
if (0 == tensor->usize()) {
return nullptr;
}
Utils::allocMemoryForHostTensor(dest);
tensor->copyToHostTensor(dest);
MNN_ASSERT(nullptr != dest->host<void>());
Expand All @@ -213,6 +217,9 @@ ErrorCode Executor::ComputeCache::compute() {
std::stack<ComputeCache*> dfsStack;
std::set<ComputeCache*> visited;
dfsStack.push(this);
ErrorCode code = NO_ERROR;
auto globalExecutor = ExecutorScope::Current();
auto debug = globalExecutor->getDebugTools();
while (!dfsStack.empty()) {
//printf("stcak = %d\n", dfsStack.size());
auto cache = dfsStack.top();
Expand Down Expand Up @@ -248,7 +255,14 @@ ErrorCode Executor::ComputeCache::compute() {
} else {
visited.insert(cache);
dfsStack.pop();
cache->mSession->run();
if (debug->after != nullptr && debug->before != nullptr) {
code = mSession->runWithCallBack(debug->before, debug->after);
} else {
code = mSession->run();
}
if (NO_ERROR != code) {
return code;
}
cache->mContentDirty = false;
}
}
Expand Down
24 changes: 18 additions & 6 deletions express/module/StaticModule.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -185,7 +185,8 @@ void StaticModule::resetInputOutputs() {
des->usage = Tensor::InsideDescribe::INPUT;
}
pipelineInfo.first.inputTensorCopyCache.insert(std::make_pair(mInputTensors[i], std::make_tuple(nullptr, nullptr, true, true)));
mPrevInputTensor[i] = nullptr;
mPrevInputTensor[i].first = nullptr;
mPrevInputTensor[i].second = nullptr;
}
mOutputTensors.resize(mResource->mOutputFromTensor.size());
for (int i = 0; i < mResource->mOutputFromTensor.size(); ++i) {
Expand Down Expand Up @@ -221,9 +222,15 @@ StaticModule::StaticModule(std::vector<int> inputs,
if (bnCache.cache.first->type() == MNN_FORWARD_CPU) {
bnCache.cache.second = bnCache.cache.first;
} else {
// Use Multi-thread if user has set numberthread > 1
BackendConfig defaultConfig;
defaultConfig.flags = 4;
bnCache.cache.second.reset(rt.second->onCreate(&defaultConfig));
auto cpurt = rt.first.find(MNN_FORWARD_CPU);
if (cpurt != rt.first.end()) {
bnCache.cache.second.reset(cpurt->second->onCreate(&defaultConfig));
} else {
bnCache.cache.second.reset(rt.second->onCreate(&defaultConfig));
}
}
if (config.rearrange) {
mResource->mBuffer = preRearrangeWeights(scheduleInfo, bnCache.cache.first.get(), bnCache.cache.second.get());
Expand Down Expand Up @@ -285,7 +292,8 @@ StaticModule::~StaticModule() {
void StaticModule::onClearCache() {
if (nullptr != mSession) {
for (int i=0; i<mPrevInputTensor.size(); ++i) {
mPrevInputTensor[i] = nullptr;
mPrevInputTensor[i].first = nullptr;
mPrevInputTensor[i].second = nullptr;
}
for (auto& iter : mSession->getPipelineInfo(0).first.inputTensorCopyCache) {
std::get<3>(iter.second) = true;
Expand Down Expand Up @@ -318,19 +326,23 @@ std::vector<Express::VARP> StaticModule::onForward(const std::vector<Express::VA
MNN_ASSERT(inputs.size() == mInputTensors.size());
auto& pipelineInfo = mSession->getPipelineInfo(0);
if (mResource->mModes.inputMode == Interpreter::Session_Input_User) {
pipelineInfo.first.inputBackendChange = false;
for (int i = 0; i < inputs.size(); ++i) {
if (nullptr == mInputTensors[i]) {
continue;
}
auto inputTensor = Utils::getTensor(inputs[i]);
Schedule::TENSORCACHE* cacheTensor = nullptr;

if (mPrevInputTensor[i] != inputTensor) {
if (mPrevInputTensor[i].first != inputTensor) {
auto newBackend = TensorUtils::getDescribeOrigin(inputTensor)->getBackend();
if (mPrevInputTensor[i].second != newBackend) {
pipelineInfo.first.inputBackendChange = true;
}
auto cacheIter = pipelineInfo.first.inputTensorCopyCache.find(mInputTensors[i]);
cacheTensor = &cacheIter->second;
MNN_ASSERT(cacheIter != pipelineInfo.first.inputTensorCopyCache.end());
std::get<3>(cacheIter->second) = true;
mPrevInputTensor[i] = inputTensor;
mPrevInputTensor[i] = std::make_pair(inputTensor, newBackend);
if (std::get<1>(*cacheTensor) != nullptr) {
if (!WrapExecution::needWrap(inputTensor, TensorUtils::getDescribeOrigin(std::get<0>(*cacheTensor))->getBackend())) {
// No need copy now, reset it
Expand Down
2 changes: 1 addition & 1 deletion express/module/StaticModule.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@ class StaticModule : public Module {
};
std::shared_ptr<Session> mSession;
std::vector<Tensor*> mInputTensors;
std::vector<Tensor*> mPrevInputTensor;
std::vector<std::pair<Tensor*, Backend*>> mPrevInputTensor;
std::vector<Tensor*> mOutputTensors;
std::shared_ptr<Resource> mResource;
};
Expand Down
2 changes: 1 addition & 1 deletion include/MNN/MNNDefine.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,6 +69,6 @@ MNN_ERROR("Check failed: %s ==> %s\n", #success, #log); \
#define STR(x) STR_IMP(x)
#define MNN_VERSION_MAJOR 2
#define MNN_VERSION_MINOR 9
#define MNN_VERSION_PATCH 0
#define MNN_VERSION_PATCH 1
#define MNN_VERSION STR(MNN_VERSION_MAJOR) "." STR(MNN_VERSION_MINOR) "." STR(MNN_VERSION_PATCH)
#endif /* MNNDefine_h */
10 changes: 7 additions & 3 deletions source/backend/arm82/Arm82Unary.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,8 +114,10 @@ struct _Exp {
void operator()(void* outRaw, const void* inpRaw, int realSize) const {
auto out = (float*)outRaw;
auto inp = (const float*)inpRaw;
float offset[2] = {
float offset[4] = {
1.0f,
0.0f,
0.0f,
0.0f
};
MNNExp(out, inp, offset, realSize);
Expand All @@ -125,9 +127,11 @@ struct _ExpM1 {
void operator()(void* outRaw, const void* inpRaw, int realSize) const {
auto out = (float*)outRaw;
auto inp = (const float*)inpRaw;
float offset[2] = {
float offset[4] = {
1.0f,
-1.0f
-1.0f,
0.0f,
0.0f
};
MNNExp(out, inp, offset, realSize);
}
Expand Down
61 changes: 54 additions & 7 deletions source/backend/cpu/BinaryUtils.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,7 +198,26 @@ void executeVec(void* outputRaw, const void* inputRaw0, const void* inputRaw1, i

if (-1 == needBroadcastIndex) {
if (sizeDivUnit > 0) {
for (int i = 0; i < sizeDivUnit; ++i) {
int sizeDivC4 = sizeDivUnit / 4;
int sizeDivUnitRemain = sizeDivUnit % 4;
for (int i = 0; i < sizeDivC4; ++i) {
V a0 = V::load(src0);
V b0 = V::load(src1);
V a1 = V::load(src0 + 1 * pack);
V b1 = V::load(src1 + 1 * pack);
V a2 = V::load(src0 + 2 * pack);
V b2 = V::load(src1 + 2 * pack);
V a3 = V::load(src0 + 3 * pack);
V b3 = V::load(src1 + 3 * pack);
V::save(dst, compute(a0, b0));
V::save(dst+1*pack, compute(a1, b1));
V::save(dst+2*pack, compute(a2, b2));
V::save(dst+3*pack, compute(a3, b3));
src0 += 4*pack;
src1 += 4*pack;
dst += 4*pack;
}
for (int i = 0; i < sizeDivUnitRemain; ++i) {
V a = V::load(src0);
V b = V::load(src1);
V::save(dst, compute(a, b));
Expand All @@ -222,11 +241,23 @@ void executeVec(void* outputRaw, const void* inputRaw0, const void* inputRaw1, i
const U srcValue0 = src0[0];
V a = V(srcValue0);
if (sizeDivUnit > 0) {
for (int i = 0; i < sizeDivUnit; ++i) {
const auto src1Ptr = src1;
auto dstPtr = dst;
V b = V::load(src1Ptr);
V::save(dstPtr, compute(a, b));
int sizeDivC4 = sizeDivUnit / 4;
int sizeUnitRemain = sizeDivUnit % 4;
for (int i = 0; i < sizeDivC4; ++i) {
V b0 = V::load(src1);
V b1 = V::load(src1 + 1*pack);
V b2 = V::load(src1 + 2*pack);
V b3 = V::load(src1 + 3*pack);
V::save(dst, compute(a, b0));
V::save(dst+1*pack, compute(a, b1));
V::save(dst+2*pack, compute(a, b2));
V::save(dst+3*pack, compute(a, b3));
src1 += 4*pack;
dst += 4*pack;
}
for (int i = 0; i < sizeUnitRemain; ++i) {
V b = V::load(src1);
V::save(dst, compute(a, b));
src1 += pack;
dst += pack;
}
Expand All @@ -243,7 +274,23 @@ void executeVec(void* outputRaw, const void* inputRaw0, const void* inputRaw1, i
const auto srcValue1 = static_cast<U>(src1[0]);
V b = V(srcValue1);
if (sizeDivUnit > 0) {
for (int i = 0; i < sizeDivUnit; ++i) {
int sizeDivC4 = sizeDivUnit / 4;
int sizeUnitRemain = sizeDivUnit % 4;
for (int i = 0; i < sizeDivC4; ++i) {
const auto src0Ptr = src0;
auto dstPtr = dst;
V a0 = V::load(src0Ptr);
V a1 = V::load(src0Ptr + 1*pack);
V a2 = V::load(src0Ptr + 2*pack);
V a3 = V::load(src0Ptr + 3*pack);
V::save(dstPtr, compute(a0, b));
V::save(dstPtr+1*pack, compute(a1, b));
V::save(dstPtr+2*pack, compute(a2, b));
V::save(dstPtr+3*pack, compute(a3, b));
src0 += 4*pack;
dst += 4*pack;
}
for (int i = 0; i < sizeUnitRemain; ++i) {
const auto src0Ptr = src0;
auto dstPtr = dst;
V a = V::load(src0Ptr);
Expand Down
26 changes: 18 additions & 8 deletions source/backend/cpu/CPUAttention.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,14 +113,23 @@ static void prefill_unpack(char* pack_qkv, char* unpack_qkv, int mNumHead, int m

template <typename T>
static void prefill_softmax(int* mask_ptr, float* mask_qk, float* softmax_qk, char* unpack_qk, char* pack_qk,
float mScale, int eP, int query_e, int seq_len, float min_val) {
float mScale, int eP, int query_e, int seq_len, float min_val, bool float_mask) {
T* qk_src = reinterpret_cast<T*>(unpack_qk);
T* qk_dst = reinterpret_cast<T*>(pack_qk);
for (int i = 0; i < seq_len * seq_len; i++) {
if (mask_ptr[i]) {
mask_qk[i] = qk_src[i] * mScale;
} else {
mask_qk[i] = min_val;
if (float_mask) {
T* fpmask_ptr = reinterpret_cast<T*>(mask_ptr);
// float mask
for (int i = 0; i < seq_len * seq_len; i++) {
mask_qk[i] = qk_src[i] * mScale + fpmask_ptr[i];
}
} else {
// int mask
for (int i = 0; i < seq_len * seq_len; i++) {
if (mask_ptr[i]) {
mask_qk[i] = qk_src[i] * mScale;
} else {
mask_qk[i] = min_val;
}
}
}
for (int i = 0; i < seq_len; i++) {
Expand Down Expand Up @@ -258,6 +267,7 @@ ErrorCode CPUAttentionImpl::onExecute(Backend* _backend, const std::vector<Tenso
auto key = inputs[1];
auto value = inputs[2];
auto mask = inputs[3];
bool float_mask = (mask->getType() == halide_type_of<float>());
auto shape = query->shape();
int seq_len = shape[1];
mThreadNum = ((CPUBackend *)backend())->threadNumber();
Expand Down Expand Up @@ -324,9 +334,9 @@ ErrorCode CPUAttentionImpl::onExecute(Backend* _backend, const std::vector<Tenso
// div scale and mask
auto mask_ptr = mask->host<int>();
if (bytes == 2) {
prefill_softmax<FLOAT16_T>(mask_ptr, mask_qk, softmax_qk, unpack_qk, pack_qk, mScale, eP, query_e, seq_len, -65504.0);
prefill_softmax<FLOAT16_T>(mask_ptr, mask_qk, softmax_qk, unpack_qk, pack_qk, mScale, eP, query_e, seq_len, -65504.0, float_mask);
} else {
prefill_softmax<float>(mask_ptr, mask_qk, softmax_qk, unpack_qk, pack_qk, mScale, eP, query_e, seq_len, std::numeric_limits<float>::lowest());
prefill_softmax<float>(mask_ptr, mask_qk, softmax_qk, unpack_qk, pack_qk, mScale, eP, query_e, seq_len, std::numeric_limits<float>::lowest(), float_mask);
}
// qk @ v
for (int i = 0 ; i < loop_e; i++) {
Expand Down
10 changes: 5 additions & 5 deletions source/backend/cpu/CPUCast.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,31 +14,31 @@
#include <cmath>

namespace MNN {
ErrorCode CPUCastCreator::cast(void* const inputRaw, void* outputRaw, ConvertType type,
ErrorCode CPUCastCreator::cast(const void* inputRaw, void* outputRaw, ConvertType type,
int number, float scale, float zero, float min, float max, const CPUBackend* bn) {
auto pack = bn->functions()->pack;
int c4Size = number / pack;
int remain = number % pack;
if (type == FlOAT_TO_INT8) {
scale = (scale == 0.f ? 0.f : 1.f / scale);
std::vector<float> scales(pack, scale);
bn->int8Functions()->MNNFloat2Int8(static_cast<float*>(inputRaw), static_cast<int8_t*>(outputRaw), c4Size, scales.data(), min, max, zero);
bn->int8Functions()->MNNFloat2Int8((float*)(inputRaw), (int8_t*)(outputRaw), c4Size, scales.data(), min, max, zero);
if (remain > 0) {
std::vector<float> tempSrc(pack);
std::vector<int8_t> tempDst(pack);
::memcpy(tempSrc.data(), static_cast<float* const>(inputRaw) + c4Size * pack, remain * sizeof(float));
::memcpy(tempSrc.data(), (float*)(inputRaw) + c4Size * pack, remain * sizeof(float));
bn->int8Functions()->MNNFloat2Int8(tempSrc.data(), tempDst.data(), 1, scales.data(), min, max, zero);
::memcpy(static_cast<int8_t*>(outputRaw) + c4Size * pack, tempDst.data(), remain * sizeof(int8_t));
}
return NO_ERROR;
}
if (type == INT8_TO_FlOAT) {
std::vector<float> scales(pack, scale);
bn->int8Functions()->MNNInt8ScaleToFloat(static_cast<float*>(outputRaw), static_cast<int8_t*>(inputRaw), scales.data(), c4Size, zero);
bn->int8Functions()->MNNInt8ScaleToFloat((float*)(outputRaw), (int8_t*)(inputRaw), scales.data(), c4Size, zero);
if (remain > 0) {
std::vector<float> tempDst(pack);
std::vector<int8_t> tempSrc(pack);
::memcpy(tempSrc.data(), static_cast<int8_t* const>(inputRaw) + c4Size * pack, remain * sizeof(int8_t));
::memcpy(tempSrc.data(), (int8_t*)(inputRaw) + c4Size * pack, remain * sizeof(int8_t));
bn->int8Functions()->MNNInt8ScaleToFloat(tempDst.data(), tempSrc.data(), scales.data(), 1, zero);
::memcpy(static_cast<float*>(outputRaw) + c4Size * pack, tempDst.data(), remain * sizeof(float));
}
Expand Down
2 changes: 1 addition & 1 deletion source/backend/cpu/CPUCast.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,7 @@ class CPUCastCreator : public CPUBackend::Creator {
virtual Execution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
const MNN::Op* op, Backend* backend) const override;
static ErrorCode cast(const Tensor* input, const Tensor* output, const CPUBackend* bn, ConvertType type);
static ErrorCode cast(void* const inputRaw, void* outputRaw, ConvertType type, int number, float scale, float zero, float min, float max, const CPUBackend* bn);
static ErrorCode cast(const void* inputRaw, void* outputRaw, ConvertType type, int number, float scale, float zero, float min, float max, const CPUBackend* bn);
};
} // namespace MNN
#endif /* CPUCast_hpp */
Loading
Loading