Revert "#16356: Program Dispatch Modifications for MeshWorkload"

This reverts commit 3755428.

tt_metal/impl/CMakeLists.txt

@@ -17,7 +17,6 @@ set(IMPL_SRC
     ${CMAKE_CURRENT_SOURCE_DIR}/allocator/basic_allocator.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/allocator/l1_banking_allocator.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/program/program.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/program/program_dispatch_utils.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/dispatch/debug_tools.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/dispatch/command_queue.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/dispatch/worker_config_buffer.cpp

tt_metal/impl/device/device.cpp

@@ -938,7 +938,7 @@ void Device::init_command_queue_host() {
     this->sysmem_manager_ = std::make_unique<SystemMemoryManager>(this->id_, this->num_hw_cqs());
     hw_command_queues_.resize(num_hw_cqs());
     for (size_t cq_id = 0; cq_id < num_hw_cqs(); cq_id++) {
-        hw_command_queues_[cq_id] = std::make_unique<HWCommandQueue>(this, cq_id, dispatch_downstream_noc);
+        hw_command_queues_[cq_id] = std::make_unique<HWCommandQueue>(this, cq_id, NOC::NOC_0);
         // Need to do this since CommandQueue constructor is private
         sw_command_queues_.push_back(std::unique_ptr<CommandQueue>(new CommandQueue(this, cq_id)));
     }
@@ -1750,7 +1750,6 @@ LaunchMessageRingBufferState& Device::get_worker_launch_message_buffer_state(Sub
     return this->active_sub_device_manager_->get_worker_launch_message_buffer_state(sub_device_id);
 }
 
-// Main source to get NOC idx for dispatch core
 NOC Device::dispatch_go_signal_noc() const {
     return this->dispatch_s_enabled() ? NOC::NOC_1 : NOC::NOC_0;
 }

tt_metal/impl/device/device_pool.cpp

@@ -334,12 +334,12 @@ void DevicePool::add_devices_to_pool(const std::vector<chip_id_t>& device_ids) {
         }
     }
 
+    populate_fd_kernels(devices_to_activate, this->num_hw_cqs);
     for (const auto& device_id : devices_to_activate) {
         if (not this->is_device_active(device_id)) {
             this->activate_device(device_id);
         }
     }
-    populate_fd_kernels(devices_to_activate, this->num_hw_cqs);
 }
 
 void DevicePool::register_worker_thread_for_device(v1::DeviceHandle device, std::thread::id worker_thread_id) {

tt_metal/impl/dispatch/command_queue.hpp

@@ -331,6 +331,16 @@ class EnqueueProgramCommand : public Command {
         uint32_t unicast_cores_launch_message_wptr,
         SubDeviceId sub_device_id);
 
+    void assemble_preamble_commands(
+        ProgramCommandSequence& program_command_sequence, const tt::stl::Span<ConfigBufferEntry> kernel_config_addrs);
+    void assemble_stall_commands(ProgramCommandSequence& program_command_sequence, bool prefetch_stall);
+    void assemble_runtime_args_commands(ProgramCommandSequence& program_command_sequence);
+    void assemble_device_commands(
+        ProgramCommandSequence& program_command_sequence, const tt::stl::Span<ConfigBufferEntry> kernel_config_addrs);
+    void update_device_commands(
+        ProgramCommandSequence& cached_program_command_sequence,
+        const tt::stl::Span<ConfigBufferEntry> kernel_config_addrs);
+
     void write_program_command_sequence(
         const ProgramCommandSequence& program_command_sequence, bool stall_first, bool stall_before_program);
 
@@ -517,6 +527,7 @@ class HWCommandQueue {
     void set_num_worker_sems_on_dispatch(uint32_t num_worker_sems);
     void set_go_signal_noc_data_on_dispatch(const vector_memcpy_aligned<uint32_t>& go_signal_noc_data);
     void reset_worker_state(bool reset_launch_msg_state);
+
    private:
     uint32_t id;
     uint32_t size_B;

tt_metal/impl/dispatch/dispatch_core_common.hpp

@@ -6,7 +6,6 @@
 
 #include "common/core_descriptor.hpp"
 #include "tt_metal/common/core_coord.hpp"
-#include "tt_metal/impl/kernels/data_types.hpp"
 #include "tt_metal/llrt/get_platform_architecture.hpp"
 #include "tt_metal/tt_stl/reflection.hpp"
 
@@ -33,10 +32,6 @@ enum DispatchWorkerType : uint32_t {
     COUNT = 17
 };
 
-// NOC ID used by dispatch kernels to communicate with downstream cores. This parameter
-// is required when setting up Command Queue objects on host.
-static constexpr NOC dispatch_downstream_noc = NOC::NOC_0;
-
 enum class DispatchCoreType : uint32_t { WORKER, ETH, COUNT };
 
 enum class DispatchCoreAxis { ROW, COL, COUNT };

tt_metal/impl/dispatch/kernel_config/dispatch.cpp

@@ -26,6 +26,7 @@ void DispatchKernel::GenerateStaticConfigs() {
         uint32_t completion_queue_start_addr = issue_queue_start_addr + issue_queue_size;
         uint32_t completion_queue_size = device_->sysmem_manager().get_completion_queue_size(cq_id_);
 
+        logical_core_ = dispatch_core_manager::instance().dispatcher_core(device_->id(), channel, cq_id_);
         static_config_.dispatch_cb_base = my_dispatch_constants.dispatch_buffer_base();
         static_config_.dispatch_cb_log_page_size = dispatch_constants::DISPATCH_BUFFER_LOG_PAGE_SIZE;
         static_config_.dispatch_cb_pages = my_dispatch_constants.dispatch_buffer_pages();
@@ -76,6 +77,7 @@ void DispatchKernel::GenerateStaticConfigs() {
         uint32_t completion_queue_start_addr = issue_queue_start_addr + issue_queue_size;
         uint32_t completion_queue_size = device_->sysmem_manager().get_completion_queue_size(cq_id_);
 
+        logical_core_ = dispatch_core_manager::instance().dispatcher_core(servicing_device_id_, channel, cq_id_);
         static_config_.dispatch_cb_base = my_dispatch_constants.dispatch_buffer_base();
         static_config_.dispatch_cb_log_page_size = dispatch_constants::DISPATCH_BUFFER_LOG_PAGE_SIZE;
         static_config_.dispatch_cb_pages = my_dispatch_constants.dispatch_buffer_pages();
@@ -122,6 +124,7 @@ void DispatchKernel::GenerateStaticConfigs() {
         uint32_t completion_queue_start_addr = issue_queue_start_addr + issue_queue_size;
         uint32_t completion_queue_size = device_->sysmem_manager().get_completion_queue_size(cq_id_);
 
+        logical_core_ = dispatch_core_manager::instance().dispatcher_d_core(device_->id(), channel, cq_id_);
         static_config_.dispatch_cb_base = my_dispatch_constants.dispatch_buffer_base();
         static_config_.dispatch_cb_log_page_size = dispatch_constants::PREFETCH_D_BUFFER_LOG_PAGE_SIZE;
         static_config_.dispatch_cb_pages = my_dispatch_constants.dispatch_buffer_pages();

tt_metal/impl/dispatch/kernel_config/dispatch.hpp

@@ -65,24 +65,8 @@ class DispatchKernel : public FDKernel {
         bool h_variant,
         bool d_variant) :
         FDKernel(node_id, device_id, servicing_device_id, cq_id, noc_selection) {
-        TT_FATAL(
-            noc_selection.downstream_noc == dispatch_downstream_noc,
-            "Invalid downstream NOC specified for Dispatcher kernel");
-        TT_FATAL(
-            noc_selection.upstream_noc != noc_selection.downstream_noc,
-            "Dispatcher kernel cannot have identical upstream and downstream NOCs.");
         static_config_.is_h_variant = h_variant;
         static_config_.is_d_variant = d_variant;
-        uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device_id);
-        if (h_variant && d_variant) {
-            this->logical_core_ = dispatch_core_manager::instance().dispatcher_core(device_id, channel, cq_id);
-        } else if (h_variant) {
-            channel = tt::Cluster::instance().get_assigned_channel_for_device(servicing_device_id);
-            this->logical_core_ =
-                dispatch_core_manager::instance().dispatcher_core(servicing_device_id, channel, cq_id);
-        } else if (d_variant) {
-            this->logical_core_ = dispatch_core_manager::instance().dispatcher_d_core(device_id, channel, cq_id);
-        }
     }
     void CreateKernel() override;
     void GenerateStaticConfigs() override;

tt_metal/impl/dispatch/kernel_config/dispatch_s.cpp

@@ -27,7 +27,7 @@ void DispatchSKernel::GenerateStaticConfigs() {
             dispatch_s_buffer_base = dispatch_buffer_base;
         }
     }
-
+    logical_core_ = dispatch_core_manager::instance().dispatcher_s_core(device_->id(), channel, cq_id_);
     static_config_.cb_base = dispatch_s_buffer_base;
     static_config_.cb_log_page_size = dispatch_constants::DISPATCH_S_BUFFER_LOG_PAGE_SIZE;
     static_config_.cb_size = my_dispatch_constants.dispatch_s_buffer_size();

tt_metal/impl/dispatch/kernel_config/dispatch_s.hpp

@@ -29,10 +29,7 @@ class DispatchSKernel : public FDKernel {
 public:
     DispatchSKernel(
         int node_id, chip_id_t device_id, chip_id_t servicing_device_id, uint8_t cq_id, noc_selection_t noc_selection) :
-        FDKernel(node_id, device_id, servicing_device_id, cq_id, noc_selection) {
-        uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device_id);
-        this->logical_core_ = dispatch_core_manager::instance().dispatcher_s_core(device_id, channel, cq_id_);
-    }
+        FDKernel(node_id, device_id, servicing_device_id, cq_id, noc_selection) {}
     void CreateKernel() override;
     void GenerateStaticConfigs() override;
     void GenerateDependentConfigs() override;

tt_metal/impl/dispatch/kernel_config/eth_tunneler.cpp

@@ -54,7 +54,7 @@ void EthTunnelerKernel::GenerateDependentConfigs() {
             } else if (auto tk = dynamic_cast<EthTunnelerKernel*>(k)) {
                 tunneler_kernel = tk;
             } else {
-                TT_FATAL(false, "Unexpected kernel type upstream of TUNNELER");
+                TT_FATAL(false, "Unexpected kernelt tyoe downstream of TUNNELER");
             }
         }
         TT_ASSERT(tunneler_kernel && !tunneler_kernel->IsRemote());
@@ -175,7 +175,7 @@ void EthTunnelerKernel::GenerateDependentConfigs() {
             } else if (auto tk = dynamic_cast<EthTunnelerKernel*>(k)) {
                 ds_tunneler_kernel = tk;
             } else {
-                TT_FATAL(false, "Unexpected kernel type downstream of TUNNELER");
+                TT_FATAL(false, "Unexpected kernelt tyoe downstream of TUNNELER");
             }
         }
         TT_ASSERT(ds_tunneler_kernel && ds_tunneler_kernel == tunneler_kernel);

tt_metal/impl/dispatch/kernel_config/prefetch.cpp

@@ -23,6 +23,8 @@ void PrefetchKernel::GenerateStaticConfigs() {
         uint32_t issue_queue_start_addr = command_queue_start_addr + cq_start;
         uint32_t issue_queue_size = device_->sysmem_manager().get_issue_queue_size(cq_id_);
 
+        logical_core_ = dispatch_core_manager::instance().prefetcher_core(device_->id(), channel, cq_id_);
+
         dependent_config_.downstream_cb_base = my_dispatch_constants.dispatch_buffer_base();
         static_config_.downstream_cb_log_page_size = dispatch_constants::DISPATCH_BUFFER_LOG_PAGE_SIZE;
         static_config_.downstream_cb_pages = my_dispatch_constants.dispatch_buffer_pages();
@@ -81,6 +83,8 @@ void PrefetchKernel::GenerateStaticConfigs() {
         uint32_t issue_queue_start_addr = command_queue_start_addr + cq_start;
         uint32_t issue_queue_size = device_->sysmem_manager().get_issue_queue_size(cq_id_);
 
+        logical_core_ = dispatch_core_manager::instance().prefetcher_core(servicing_device_id_, channel, cq_id_);
+
         static_config_.downstream_cb_log_page_size = dispatch_constants::PREFETCH_D_BUFFER_LOG_PAGE_SIZE;
         if (tt::Cluster::instance().is_galaxy_cluster()) {  // TODO: whys is this hard-coded for galaxy?
             static_config_.downstream_cb_pages = my_dispatch_constants.mux_buffer_pages(1);
@@ -119,6 +123,8 @@ void PrefetchKernel::GenerateStaticConfigs() {
         static_config_.dispatch_s_buffer_size = 0;
         static_config_.dispatch_s_cb_log_page_size = 0;
     } else if (static_config_.is_d_variant.value()) {
+        logical_core_ = dispatch_core_manager::instance().prefetcher_d_core(device_->id(), channel, cq_id_);
+
         dependent_config_.downstream_cb_base = my_dispatch_constants.dispatch_buffer_base();
         static_config_.downstream_cb_log_page_size = dispatch_constants::PREFETCH_D_BUFFER_LOG_PAGE_SIZE;
         static_config_.downstream_cb_pages = my_dispatch_constants.dispatch_buffer_pages();

tt_metal/impl/dispatch/kernel_config/prefetch.hpp

@@ -66,19 +66,6 @@ class PrefetchKernel : public FDKernel {
         FDKernel(node_id, device_id, servicing_device_id, cq_id, noc_selection) {
         static_config_.is_h_variant = h_variant;
         static_config_.is_d_variant = d_variant;
-        uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device_id);
-        TT_FATAL(
-            noc_selection.downstream_noc == dispatch_downstream_noc,
-            "Invalid downstream NOC specified for Prefetcher kernel");
-        if (h_variant && d_variant) {
-            this->logical_core_ = dispatch_core_manager::instance().prefetcher_core(device_id, channel, cq_id);
-        } else if (h_variant) {
-            channel = tt::Cluster::instance().get_assigned_channel_for_device(servicing_device_id);
-            this->logical_core_ =
-                dispatch_core_manager::instance().prefetcher_core(servicing_device_id, channel, cq_id);
-        } else if (d_variant) {
-            this->logical_core_ = dispatch_core_manager::instance().prefetcher_d_core(device_id, channel, cq_id);
-        }
     }
     void CreateKernel() override;
     void GenerateStaticConfigs() override;

tt_metal/impl/dispatch/topology.cpp

@@ -79,18 +79,15 @@ static const std::vector<dispatch_kernel_node_t> two_chip_arch_1cq = {
 
     {3, 0, 1, 0, PREFETCH_H, {x, x, x, x}, {5, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
     {4, 0, 1, 0, DISPATCH_H, {6, x, x, x}, {3, x, x, x}, NOC::NOC_0, NOC::NOC_1, NOC::NOC_0},
-
     {5, 0, 1, 0, PACKET_ROUTER_MUX, {3, x, x, x}, {7, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
     {6, 0, 1, 0, DEMUX, {7, x, x, x}, {4, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {7, 0, 1, 0, US_TUNNELER_REMOTE, {11, 5, x, x}, {11, 6, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-
-    {8, 1, x, 0, PREFETCH_D, {13, x, x, x}, {9, 10, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {9, 1, x, 0, DISPATCH_D, {8, x, x, x}, {10, 12, x, x}, NOC::NOC_0, NOC::NOC_1, NOC::NOC_0},
-    {10, 1, x, 0, DISPATCH_S, {8, x, x, x}, {9, x, x, x}, NOC::NOC_1, NOC::NOC_1, NOC::NOC_1},
-
-    {11, 1, x, 0, US_TUNNELER_LOCAL, {7, 12, x, x}, {7, 13, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {12, 1, x, 0, MUX_D, {9, x, x, x}, {11, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {13, 1, x, 0, PACKET_ROUTER_DEMUX, {11, x, x, x}, {8, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {7, 0, 1, 0, US_TUNNELER_REMOTE, {8, 5, x, x}, {8, 6, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {8, 1, x, 0, US_TUNNELER_LOCAL, {7, 9, x, x}, {7, 10, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {9, 1, x, 0, MUX_D, {12, x, x, x}, {8, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {10, 1, x, 0, PACKET_ROUTER_DEMUX, {8, x, x, x}, {11, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {11, 1, x, 0, PREFETCH_D, {10, x, x, x}, {12, 13, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {12, 1, x, 0, DISPATCH_D, {11, x, x, x}, {13, 9, x, x}, NOC::NOC_0, NOC::NOC_1, NOC::NOC_0},
+    {13, 1, x, 0, DISPATCH_S, {11, x, x, x}, {12, x, x, x}, NOC::NOC_1, NOC::NOC_1, NOC::NOC_1},
 };
 
 static const std::vector<dispatch_kernel_node_t> two_chip_arch_2cq = {
@@ -106,17 +103,15 @@ static const std::vector<dispatch_kernel_node_t> two_chip_arch_2cq = {
 
     {8, 0, 1, 0, PACKET_ROUTER_MUX, {4, 5, x, x}, {10, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
     {9, 0, 1, 0, DEMUX, {10, x, x, x}, {6, 7, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {10, 0, 1, 0, US_TUNNELER_REMOTE, {15, 8, x, x}, {15, 9, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-
-    {11, 1, x, 0, PREFETCH_D, {17, x, x, x}, {13, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {12, 1, x, 1, PREFETCH_D, {17, x, x, x}, {14, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {13, 1, x, 0, DISPATCH_D, {11, x, x, x}, {16, x, x, x}, NOC::NOC_0, NOC::NOC_1, NOC::NOC_0},
-    {14, 1, x, 1, DISPATCH_D, {12, x, x, x}, {16, x, x, x}, NOC::NOC_0, NOC::NOC_1, NOC::NOC_0},
-
-    {15, 1, x, 0, US_TUNNELER_LOCAL, {10, 16, x, x}, {10, 17, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {16, 1, x, 0, MUX_D, {13, 14, x, x}, {15, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-    {17, 1, x, 0, PACKET_ROUTER_DEMUX, {15, x, x, x}, {11, 12, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
-
+    {10, 0, 1, 0, US_TUNNELER_REMOTE, {11, 8, x, x}, {11, 9, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {11, 1, x, 0, US_TUNNELER_LOCAL, {10, 12, x, x}, {10, 13, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {12, 1, x, 0, MUX_D, {16, 17, x, x}, {11, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {13, 1, x, 0, PACKET_ROUTER_DEMUX, {11, x, x, x}, {14, 15, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+
+    {14, 1, x, 0, PREFETCH_D, {13, x, x, x}, {16, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {15, 1, x, 1, PREFETCH_D, {13, x, x, x}, {17, x, x, x}, NOC::NOC_0, NOC::NOC_0, NOC::NOC_0},
+    {16, 1, x, 0, DISPATCH_D, {14, x, x, x}, {12, x, x, x}, NOC::NOC_0, NOC::NOC_1, NOC::NOC_0},
+    {17, 1, x, 1, DISPATCH_D, {15, x, x, x}, {12, x, x, x}, NOC::NOC_0, NOC::NOC_1, NOC::NOC_0},
 };
 
 static const std::vector<dispatch_kernel_node_t> galaxy_nine_chip_arch_1cq = {

tt_metal/impl/kernels/kernel_types.hpp

@@ -15,7 +15,7 @@
 
 namespace tt::tt_metal {
 
-using KernelHandle = std::uint32_t;
+using KernelHandle = std::uint16_t;
 
 struct DataMovementConfig {
     DataMovementProcessor processor = DataMovementProcessor::RISCV_0;  // For data transfer kernels: NCRISC & BRISC