#0: seperate dispatch constants

- Rename dispatch_constants to DispatchMemMap - Delete old constants. New constants in DispatchConstants - DispatchMemMap to be configured by DispatchSettings
tenstorrent · Jan 29, 2025 · 59e073a · 59e073a
1 parent 5ebd7d8
commit 59e073a
Show file tree

Hide file tree

Showing 33 changed files with 404 additions and 417 deletions.
diff --git a/tests/tt_metal/tt_metal/device/test_device.cpp b/tests/tt_metal/tt_metal/device/test_device.cpp
@@ -243,7 +243,7 @@ TEST_F(DeviceFixture, TensixTestL1ToPCIeAt16BAlignedAddress) {
     uint32_t base_l1_src_address = device->get_base_allocator_addr(HalMemType::L1) + hal.get_alignment(HalMemType::L1);
     // This is a slow dispatch test dispatch core type is needed to query dispatch_constants
     uint32_t base_pcie_dst_address =
-        dispatch_constants::get(CoreType::WORKER).get_host_command_queue_addr(CommandQueueHostAddrType::UNRESERVED) +
+        DispatchMemMap::get(CoreType::WORKER).get_host_command_queue_addr(CommandQueueHostAddrType::UNRESERVED) +
         hal.get_alignment(HalMemType::L1);
 
     uint32_t size_bytes = 2048 * 128;

diff --git a/...metal/tt_metal/dispatch/dispatch_buffer/test_EnqueueWriteBuffer_and_EnqueueReadBuffer.cpp b/...metal/tt_metal/dispatch/dispatch_buffer/test_EnqueueWriteBuffer_and_EnqueueReadBuffer.cpp
@@ -194,7 +194,7 @@ void test_EnqueueWriteBuffer_and_EnqueueReadBuffer(IDevice* device, CommandQueue
     uint32_t cq_size = device->sysmem_manager().get_cq_size();
     CoreType dispatch_core_type = dispatch_core_manager::instance().get_dispatch_core_type(device->id());
     uint32_t cq_start =
-        dispatch_constants::get(dispatch_core_type).get_host_command_queue_addr(CommandQueueHostAddrType::UNRESERVED);
+        DispatchMemMap::get(dispatch_core_type).get_host_command_queue_addr(CommandQueueHostAddrType::UNRESERVED);
 
     std::vector<uint32_t> cq_zeros((cq_size - cq_start) / sizeof(uint32_t), 0);
 
@@ -549,7 +549,7 @@ TEST_F(CommandQueueSingleCardBufferFixture, TestPageLargerThanAndUnalignedToTran
     for (IDevice* device : devices_) {
         TestBufferConfig config = {
             .num_pages = num_round_robins * (device->num_banks(BufferType::DRAM)),
-            .page_size = dispatch_constants::TRANSFER_PAGE_SIZE + 32,
+            .page_size = DispatchConstants::TRANSFER_PAGE_SIZE + 32,
             .buftype = BufferType::DRAM};
         local_test_functions::test_EnqueueWriteBuffer_and_EnqueueReadBuffer(device, device->command_queue(), config);
     }
@@ -559,8 +559,7 @@ TEST_F(CommandQueueSingleCardBufferFixture, TestPageLargerThanMaxPrefetchCommand
     constexpr uint32_t num_round_robins = 1;
     for (IDevice* device : devices_) {
         CoreType dispatch_core_type = dispatch_core_manager::instance().get_dispatch_core_type(device->id());
-        const uint32_t max_prefetch_command_size =
-            dispatch_constants::get(dispatch_core_type).max_prefetch_command_size();
+        const uint32_t max_prefetch_command_size = DispatchMemMap::get(dispatch_core_type).max_prefetch_command_size();
         TestBufferConfig config = {
             .num_pages = 1, .page_size = max_prefetch_command_size + 2048, .buftype = BufferType::DRAM};
         local_test_functions::test_EnqueueWriteBuffer_and_EnqueueReadBuffer(device, device->command_queue(), config);
@@ -571,8 +570,7 @@ TEST_F(CommandQueueSingleCardBufferFixture, TestUnalignedPageLargerThanMaxPrefet
     constexpr uint32_t num_round_robins = 1;
     for (IDevice* device : devices_) {
         CoreType dispatch_core_type = dispatch_core_manager::instance().get_dispatch_core_type(device->id());
-        const uint32_t max_prefetch_command_size =
-            dispatch_constants::get(dispatch_core_type).max_prefetch_command_size();
+        const uint32_t max_prefetch_command_size = DispatchMemMap::get(dispatch_core_type).max_prefetch_command_size();
         uint32_t unaligned_page_size = max_prefetch_command_size + 4;
         TestBufferConfig config = {.num_pages = 1, .page_size = unaligned_page_size, .buftype = BufferType::DRAM};
         local_test_functions::test_EnqueueWriteBuffer_and_EnqueueReadBuffer(device, device->command_queue(), config);
@@ -613,8 +611,8 @@ TEST_F(CommandQueueSingleCardBufferFixture, TestWrapHostHugepageOnEnqueueReadBuf
         uint32_t page_size = 2048;
         uint32_t command_issue_region_size = device->sysmem_manager().get_issue_queue_size(0);
         CoreType dispatch_core_type = dispatch_core_manager::instance().get_dispatch_core_type(device->id());
-        uint32_t cq_start = dispatch_constants::get(dispatch_core_type)
-                                .get_host_command_queue_addr(CommandQueueHostAddrType::UNRESERVED);
+        uint32_t cq_start =
+            DispatchMemMap::get(dispatch_core_type).get_host_command_queue_addr(CommandQueueHostAddrType::UNRESERVED);
 
         uint32_t max_command_size = command_issue_region_size - cq_start;
         uint32_t buffer = 14240;

diff --git a/tests/tt_metal/tt_metal/dispatch/dispatch_event/test_EnqueueWaitForEvent.cpp b/tests/tt_metal/tt_metal/dispatch/dispatch_event/test_EnqueueWaitForEvent.cpp
@@ -193,7 +193,7 @@ TEST_F(MultiCommandQueueSingleDeviceEventFixture, TestEventsEnqueueWaitForEventC
     for (uint cq_id = 0; cq_id < cqs.size(); cq_id++) {
         for (size_t i = 0; i < num_cmds_per_cq * num_events_per_cq; i++) {
             uint32_t host_addr =
-                completion_queue_base[cq_id] + i * dispatch_constants::TRANSFER_PAGE_SIZE + sizeof(CQDispatchCmd);
+                completion_queue_base[cq_id] + i * DispatchConstants::TRANSFER_PAGE_SIZE + sizeof(CQDispatchCmd);
             tt::Cluster::instance().read_sysmem(&event, 4, host_addr, mmio_device_id, channel);
             log_debug(
                 tt::LogTest,

diff --git a/tests/tt_metal/tt_metal/dispatch/dispatch_event/test_events.cpp b/tests/tt_metal/tt_metal/dispatch/dispatch_event/test_events.cpp
@@ -14,7 +14,7 @@ using std::vector;
 using namespace tt::tt_metal;
 
 constexpr uint32_t completion_queue_event_offset = sizeof(CQDispatchCmd);
-constexpr uint32_t completion_queue_page_size = dispatch_constants::TRANSFER_PAGE_SIZE;
+constexpr uint32_t completion_queue_page_size = DispatchConstants::TRANSFER_PAGE_SIZE;
 
 enum class DataMovementMode : uint8_t { WRITE = 0, READ = 1 };
 

diff --git a/tests/tt_metal/tt_metal/perf_microbenchmark/3_pcie_transfer/test_pull_from_pcie.cpp b/tests/tt_metal/tt_metal/perf_microbenchmark/3_pcie_transfer/test_pull_from_pcie.cpp
@@ -226,7 +226,7 @@ int main(int argc, char** argv) {
         uint32_t host_write_ptr = 0;
 
         CoreType dispatch_core_type = dispatch_core_manager::instance().get_dispatch_core_type(device_id);
-        uint32_t prefetch_q_base = dispatch_constants::get(dispatch_core_type)
+        uint32_t prefetch_q_base = DispatchMemMap::get(dispatch_core_type)
                                        .get_device_command_queue_addr(CommandQueueDeviceAddrType::UNRESERVED);
 
         uint32_t reg_addr = prefetch_q_base;

diff --git a/tests/tt_metal/tt_metal/perf_microbenchmark/dispatch/test_bw_and_latency.cpp b/tests/tt_metal/tt_metal/perf_microbenchmark/dispatch/test_bw_and_latency.cpp
@@ -396,8 +396,8 @@ int main(int argc, char** argv) {
             vec.resize(page_size_g / sizeof(uint32_t));
 
             CoreType core_type = dispatch_core_manager::instance().get_dispatch_core_type(device->id());
-            uint32_t dispatch_l1_unreserved_base = dispatch_constants::get(core_type).get_device_command_queue_addr(
-                CommandQueueDeviceAddrType::UNRESERVED);
+            uint32_t dispatch_l1_unreserved_base =
+                DispatchMemMap::get(core_type).get_device_command_queue_addr(CommandQueueDeviceAddrType::UNRESERVED);
             for (int i = 0; i < warmup_iterations_g; i++) {
                 if (source_mem_g == 4) {
                     tt::Cluster::instance().read_core(

diff --git a/tests/tt_metal/tt_metal/perf_microbenchmark/dispatch/test_dispatcher.cpp b/tests/tt_metal/tt_metal/perf_microbenchmark/dispatch/test_dispatcher.cpp
@@ -436,8 +436,7 @@ int main(int argc, char** argv) {
 
         // Want different buffers on each core, instead use big buffer and self-manage it
         uint32_t dispatch_l1_unreserved_base =
-            dispatch_constants::get(CoreType::WORKER)
-                .get_device_command_queue_addr(CommandQueueDeviceAddrType::UNRESERVED);
+            DispatchMemMap::get(CoreType::WORKER).get_device_command_queue_addr(CommandQueueDeviceAddrType::UNRESERVED);
         uint32_t l1_buf_base = tt::align(dispatch_l1_unreserved_base, dispatch_buffer_page_size_g);
         TT_ASSERT((l1_buf_base & (dispatch_buffer_page_size_g - 1)) == 0);
 
@@ -526,13 +525,13 @@ int main(int argc, char** argv) {
         const uint32_t prefetch_sync_sem = spoof_prefetch_core_sem_1_id;
 
         const uint32_t host_completion_queue_wr_ptr =
-            dispatch_constants::get(CoreType::WORKER)
+            DispatchMemMap::get(CoreType::WORKER)
                 .get_host_command_queue_addr(CommandQueueHostAddrType::COMPLETION_Q_WR);
         const uint32_t dev_completion_queue_wr_ptr =
-            dispatch_constants::get(CoreType::WORKER)
+            DispatchMemMap::get(CoreType::WORKER)
                 .get_device_command_queue_addr(CommandQueueDeviceAddrType::COMPLETION_Q_WR);
         const uint32_t dev_completion_queue_rd_ptr =
-            dispatch_constants::get(CoreType::WORKER)
+            DispatchMemMap::get(CoreType::WORKER)
                 .get_device_command_queue_addr(CommandQueueDeviceAddrType::COMPLETION_Q_RD);
 
         std::vector<uint32_t> dispatch_compile_args = {
@@ -558,8 +557,8 @@ int main(int argc, char** argv) {
             0,                  // prefetch_downstream_buffer_pages
             num_compute_cores,  // max_write_packed_cores
             0,
-            dispatch_constants::DISPATCH_MESSAGE_ENTRIES,
-            dispatch_constants::DISPATCH_GO_SIGNAL_NOC_DATA_ENTRIES,
+            DispatchConstants::DISPATCH_MESSAGE_ENTRIES,
+            DispatchConstants::DISPATCH_GO_SIGNAL_NOC_DATA_ENTRIES,
             0,
             0,
             0,