Cluster set power state for Blackhole

tenstorrent · Feb 28, 2025 · f08d12a · f08d12a
1 parent da0cc30
commit f08d12a
Show file tree

Hide file tree

Showing 3 changed files with 33 additions and 8 deletions.
diff --git a/device/api/umd/device/cluster.h b/device/api/umd/device/cluster.h
@@ -1195,6 +1195,8 @@ class Cluster : public tt_device {
     static std::unique_ptr<tt_ClusterDescriptor> create_cluster_descriptor(
         const std::unordered_map<chip_id_t, std::unique_ptr<tt::umd::Chip>>& chips);
 
+    void initialize_arc_communication();
+
     // State variables
     std::vector<tt::ARCH> archs_in_cluster = {};
     std::set<chip_id_t> all_chip_ids_ = {};
@@ -1205,6 +1207,8 @@ class Cluster : public tt_device {
 
     std::shared_ptr<tt_ClusterDescriptor> cluster_desc;
 
+    std::unique_ptr<tt::umd::BlackholeArcMessageQueue> bh_arc_msg_queue = nullptr;
+
     // remote eth transfer setup
     static constexpr std::uint32_t NUM_ETH_CORES_FOR_NON_MMIO_TRANSFERS = 6;
     static constexpr std::uint32_t NON_EPOCH_ETH_CORES_FOR_NON_MMIO_TRANSFERS = 4;
@@ -1229,6 +1233,7 @@ class Cluster : public tt_device {
     std::unordered_map<chip_id_t, std::unordered_set<tt_xy_pair>> workers_per_chip = {};
     std::unordered_set<tt_xy_pair> eth_cores = {};
     std::unordered_set<tt_xy_pair> dram_cores = {};
+    std::unordered_map<chip_id_t, std::unique_ptr<BlackholeArcMessageQueue>> bh_arc_msg_queues = {};
 
     std::map<std::set<chip_id_t>, std::unordered_map<chip_id_t, std::vector<std::vector<int>>>> bcast_header_cache = {};
     bool perform_harvesting_on_sdesc = false;

diff --git a/device/cluster.cpp b/device/cluster.cpp
@@ -429,6 +429,8 @@ void Cluster::construct_cluster(
             }
         }
     }
+
+    initialize_arc_communication();
 }
 
 std::unique_ptr<Chip> Cluster::construct_chip_from_cluster(
@@ -3230,6 +3232,17 @@ void Cluster::broadcast_tensix_risc_reset_to_cluster(const TensixSoftResetOption
     }
 }
 
+void Cluster::initialize_arc_communication() {
+    if (arch_name == tt::ARCH::BLACKHOLE) {
+        for (auto& chip : all_chip_ids_) {
+            bh_arc_msg_queues.insert(
+                {chip,
+                 BlackholeArcMessageQueue::get_blackhole_arc_message_queue(
+                     this, chip, BlackholeArcMessageQueueIndex::APPLICATION)});
+        }
+    }
+}
+
 void Cluster::set_power_state(tt_DevicePowerState device_state) {
     // MT Initial BH - ARC messages not supported in Blackhole
     if (arch_name != tt::ARCH::BLACKHOLE) {
@@ -3242,6 +3255,16 @@ void Cluster::set_power_state(tt_DevicePowerState device_state) {
                     exit_code == 0, "Failed to set power state to {} with exit code: {}", (int)device_state, exit_code);
             }
         }
+    } else {
+        for (auto& chip : all_chip_ids_) {
+            std::unique_ptr<BlackholeArcMessageQueue>& bh_arc_msg_queue = bh_arc_msg_queues.at(chip);
+
+            if (device_state == tt_DevicePowerState::BUSY) {
+                bh_arc_msg_queue->send_message(tt::umd::blackhole::ArcMessageType::AICLK_GO_BUSY);
+            } else {
+                bh_arc_msg_queue->send_message(tt::umd::blackhole::ArcMessageType::AICLK_GO_LONG_IDLE);
+            }
+        }
     }
 }
 
@@ -3302,9 +3325,10 @@ void Cluster::deassert_resets_and_set_power_state() {
             }
             enable_ethernet_queue(30);
         }
-        // Set power state to busy
-        set_power_state(tt_DevicePowerState::BUSY);
     }
+
+    // Set power state to busy
+    set_power_state(tt_DevicePowerState::BUSY);
 }
 
 void Cluster::verify_eth_fw() {

diff --git a/tests/blackhole/test_cluster_bh.cpp b/tests/blackhole/test_cluster_bh.cpp
@@ -121,8 +121,6 @@ TEST(SiliconDriverBH, CreateDestroy) {
 //         true,
 //         simulated_harvesting_masks);
 
-//     ASSERT_EQ(cluster.using_harvested_soc_descriptors(), true) << "Expected Driver to have performed harvesting";
-
 //     for (const auto& chip : sdesc_per_chip) {
 //         ASSERT_EQ(chip.second.workers.size(), 48)
 //             << "Expected SOC descriptor with harvesting to have 48 workers for chip" << chip.first;
@@ -156,8 +154,6 @@ TEST(SiliconDriverBH, CreateDestroy) {
 //         false,
 //         simulated_harvesting_masks);
 
-//     ASSERT_EQ(cluster.using_harvested_soc_descriptors(), false)
-//         << "SOC descriptors should not be modified when harvesting is disabled";
 //     for (const auto& chip : sdesc_per_chip) {
 //         ASSERT_EQ(chip.second.workers.size(), 1) << "Expected 1x1 SOC descriptor to be unmodified by driver";
 //     }
@@ -767,8 +763,8 @@ TEST(SiliconDriverBH, DISABLED_VirtualCoordinateBroadcast) {  // same problem as
     tt_device_params default_params;
     cluster.start_device(default_params);
     auto eth_version = cluster.get_ethernet_fw_version();
-    bool virtual_bcast_supported =
-        (eth_version >= tt_version(6, 8, 0) || eth_version == tt_version(6, 7, 241)) && cluster.translation_tables_en;
+    bool virtual_bcast_supported = (eth_version >= tt_version(6, 8, 0) || eth_version == tt_version(6, 7, 241)) &&
+                                   cluster.get_soc_descriptor(*target_devices.begin()).noc_translation_id_enabled;
     if (!virtual_bcast_supported) {
         cluster.close_device();
         GTEST_SKIP() << "SiliconDriverWH.VirtualCoordinateBroadcast skipped since ethernet version does not support "