Revert "refactor: change encoder for thread group over dispatch 2/n"

This reverts commit 0466317. Signed-off-by: Compute-Runtime-Validation <[email protected]>
intel · Oct 29, 2024 · 022f9e6 · 022f9e6
1 parent 7f81179
commit 022f9e6
Show file tree

Hide file tree

Showing 2 changed files with 19 additions and 15 deletions.
diff --git a/shared/source/command_container/command_encoder_xehp_and_later.inl b/shared/source/command_container/command_encoder_xehp_and_later.inl
@@ -1143,23 +1143,26 @@ void EncodeDispatchKernel<GfxFamily>::encodeThreadGroupDispatch(InterfaceDescrip
                 threadsPerXeCore /= 2;
             }
             auto tgDispatchSizeSelected = 8;
+            uint32_t numberOfThreadsInThreadGroup = interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup();
 
-            if (threadGroupDimensions[0] > 1 && (threadGroupDimensions[1] > 1 || threadGroupDimensions[2] > 1)) {
-                while (threadGroupDimensions[0] % tgDispatchSizeSelected != 0) {
+            if (walkerCmd.getThreadGroupIdXDimension() > 1 && (walkerCmd.getThreadGroupIdYDimension() > 1 || walkerCmd.getThreadGroupIdZDimension() > 1)) {
+                while (walkerCmd.getThreadGroupIdXDimension() % tgDispatchSizeSelected != 0) {
                     tgDispatchSizeSelected /= 2;
                 }
-            } else if (threadGroupDimensions[1] > 1 && threadGroupDimensions[2] > 1) {
-                while (threadGroupDimensions[1] % tgDispatchSizeSelected != 0) {
+            } else if (walkerCmd.getThreadGroupIdYDimension() > 1 && walkerCmd.getThreadGroupIdZDimension() > 1) {
+                while (walkerCmd.getThreadGroupIdYDimension() % tgDispatchSizeSelected != 0) {
                     tgDispatchSizeSelected /= 2;
                 }
             }
 
+            auto workgroupCount = walkerCmd.getThreadGroupIdXDimension() * walkerCmd.getThreadGroupIdYDimension() * walkerCmd.getThreadGroupIdZDimension();
+
             // make sure we fit all xe core
-            while (threadGroupCount / tgDispatchSizeSelected < hwInfo.gtSystemInfo.MaxSubSlicesSupported * tileCount && tgDispatchSizeSelected > 1) {
+            while (workgroupCount / tgDispatchSizeSelected < hwInfo.gtSystemInfo.MaxSubSlicesSupported * tileCount && tgDispatchSizeSelected > 1) {
                 tgDispatchSizeSelected /= 2;
             }
 
-            auto threadCountPerGrouping = tgDispatchSizeSelected * threadsPerThreadGroup;
+            auto threadCountPerGrouping = tgDispatchSizeSelected * numberOfThreadsInThreadGroup;
             // make sure we do not use more threads then present on each xe core
             while (threadCountPerGrouping > threadsPerXeCore && tgDispatchSizeSelected > 1) {
                 tgDispatchSizeSelected /= 2;
@@ -1184,25 +1187,26 @@ void EncodeDispatchKernel<GfxFamily>::encodeThreadGroupDispatch(InterfaceDescrip
                 uint32_t availableThreadCount = gfxCoreHelper.calculateAvailableThreadCount(hwInfo, grfCount);
                 availableThreadCount *= tileCount;
 
-                uint32_t dispatchedTotalThreadCount = threadsPerThreadGroup * threadGroupCount;
-                UNRECOVERABLE_IF(threadsPerThreadGroup == 0u);
+                uint32_t numberOfThreadsInThreadGroup = interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup();
+                uint32_t dispatchedTotalThreadCount = numberOfThreadsInThreadGroup * threadGroupCount;
+                UNRECOVERABLE_IF(numberOfThreadsInThreadGroup == 0u);
                 auto tgDispatchSizeSelected = 1u;
 
                 if (dispatchedTotalThreadCount <= availableThreadCount) {
                     tgDispatchSizeSelected = 1;
-                } else if (threadsPerThreadGroup <= maxThreadsInTGForTGDispatchSize8) {
+                } else if (numberOfThreadsInThreadGroup <= maxThreadsInTGForTGDispatchSize8) {
                     tgDispatchSizeSelected = 8;
-                } else if (threadsPerThreadGroup <= maxThreadsInTGForTGDispatchSize4) {
+                } else if (numberOfThreadsInThreadGroup <= maxThreadsInTGForTGDispatchSize4) {
                     tgDispatchSizeSelected = 4;
                 } else {
                     tgDispatchSizeSelected = 2;
                 }
-                if (threadGroupDimensions[0] > 1 && (threadGroupDimensions[1] > 1 || threadGroupDimensions[2] > 1)) {
-                    while (threadGroupDimensions[0] % tgDispatchSizeSelected != 0) {
+                if (walkerCmd.getThreadGroupIdXDimension() > 1 && (walkerCmd.getThreadGroupIdYDimension() > 1 || walkerCmd.getThreadGroupIdZDimension() > 1)) {
+                    while (walkerCmd.getThreadGroupIdXDimension() % tgDispatchSizeSelected != 0) {
                         tgDispatchSizeSelected /= 2;
                     }
-                } else if (threadGroupDimensions[1] > 1 && threadGroupDimensions[2] > 1) {
-                    while (threadGroupDimensions[1] % tgDispatchSizeSelected != 0) {
+                } else if (walkerCmd.getThreadGroupIdYDimension() > 1 && walkerCmd.getThreadGroupIdZDimension() > 1) {
+                    while (walkerCmd.getThreadGroupIdYDimension() % tgDispatchSizeSelected != 0) {
                         tgDispatchSizeSelected /= 2;
                     }
                 }

diff --git a/shared/source/xe_hpg_core/command_encoder_xe_hpg_core.cpp b/shared/source/xe_hpg_core/command_encoder_xe_hpg_core.cpp
@@ -33,7 +33,7 @@ void EncodeDispatchKernel<Family>::encodeThreadGroupDispatch(InterfaceDescriptor
                                                              const uint32_t *threadGroupDimensions, const uint32_t threadGroupCount, const uint32_t grfCount, const uint32_t threadsPerThreadGroup, WalkerType &walkerCmd) {
     const auto &productHelper = device.getProductHelper();
     if (productHelper.isDisableOverdispatchAvailable(hwInfo)) {
-        if (threadsPerThreadGroup == 1) {
+        if (interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup() == 1) {
             interfaceDescriptor.setThreadGroupDispatchSize(static_cast<INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE>(2u));
         } else {
             interfaceDescriptor.setThreadGroupDispatchSize(static_cast<INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE>(3u));