feature: Enable support for cl_intel_subgroup_matrix_multiply_accumulate_tf32

aratajew · Compute-Runtime-Automation · commit 691f2f50204b · 2024-06-19T12:39:05.000+02:00
This change enables support for `cl_intel_subgroup_matrix_multiply_accumulate_tf32` extension for PVC B0 and later. Related-To: GSD-7825 Signed-off-by: Ratajewski, Andrzej <andrzej.ratajewski@intel.com> Source: f7888fa
diff --git a/opencl/test/unit_test/xe_hpc_core/pvc/test_cl_device_caps_pvc.cpp b/opencl/test/unit_test/xe_hpc_core/pvc/test_cl_device_caps_pvc.cpp
@@ -21,6 +21,7 @@ PVCTEST_F(PvcClDeviceCapsTests, givenPvcProductWhenDeviceCapsInitializedThenAddP
     EXPECT_TRUE(hasSubstr(dInfo.deviceExtensions, std::string("cl_intel_create_buffer_with_properties")));
     EXPECT_TRUE(hasSubstr(dInfo.deviceExtensions, std::string("cl_intel_subgroup_local_block_io")));
     EXPECT_TRUE(hasSubstr(dInfo.deviceExtensions, std::string("cl_intel_subgroup_matrix_multiply_accumulate")));
+    EXPECT_TRUE(hasSubstr(dInfo.deviceExtensions, std::string("cl_intel_subgroup_matrix_multiply_accumulate_tf32")));
     EXPECT_TRUE(hasSubstr(dInfo.deviceExtensions, std::string("cl_khr_subgroup_named_barrier")));
     EXPECT_TRUE(hasSubstr(dInfo.deviceExtensions, std::string("cl_intel_subgroup_extended_block_read")));
     EXPECT_TRUE(hasSubstr(dInfo.deviceExtensions, std::string("cl_intel_subgroup_2d_block_io")));
diff --git a/shared/source/helpers/compiler_product_helper.h b/shared/source/helpers/compiler_product_helper.h
@@ -56,6 +56,7 @@ class CompilerProductHelper {
     virtual bool isForceEmuInt32DivRemSPRequired() const = 0;
     virtual bool isStatelessToStatefulBufferOffsetSupported() const = 0;
     virtual bool isMatrixMultiplyAccumulateSupported(const ReleaseHelper *releaseHelper) const = 0;
+    virtual bool isMatrixMultiplyAccumulateTF32Supported(const HardwareInfo &hwInfo) const = 0;
     virtual bool isSplitMatrixMultiplyAccumulateSupported(const ReleaseHelper *releaseHelper) const = 0;
     virtual bool isBFloat16ConversionSupported(const ReleaseHelper *releaseHelper) const = 0;
     virtual bool isSubgroupLocalBlockIoSupported() const = 0;
@@ -106,6 +107,7 @@ class CompilerProductHelperHw : public CompilerProductHelper {
     bool isForceEmuInt32DivRemSPRequired() const override;
     bool isStatelessToStatefulBufferOffsetSupported() const override;
     bool isMatrixMultiplyAccumulateSupported(const ReleaseHelper *releaseHelper) const override;
+    bool isMatrixMultiplyAccumulateTF32Supported(const HardwareInfo &hwInfo) const override;
     bool isSplitMatrixMultiplyAccumulateSupported(const ReleaseHelper *releaseHelper) const override;
     bool isBFloat16ConversionSupported(const ReleaseHelper *releaseHelper) const override;
     bool isSubgroupLocalBlockIoSupported() const override;
diff --git a/shared/source/helpers/compiler_product_helper_base.inl b/shared/source/helpers/compiler_product_helper_base.inl
@@ -181,6 +181,10 @@ std::string CompilerProductHelperHw<gfxProduct>::getDeviceExtensions(const Hardw
         extensions += "cl_intel_subgroup_matrix_multiply_accumulate ";
     }
 
+    if (isMatrixMultiplyAccumulateTF32Supported(hwInfo)) {
+        extensions += "cl_intel_subgroup_matrix_multiply_accumulate_tf32 ";
+    }
+
     if (isSplitMatrixMultiplyAccumulateSupported(releaseHelper)) {
         extensions += "cl_intel_subgroup_split_matrix_multiply_accumulate ";
     }
diff --git a/shared/source/helpers/compiler_product_helper_before_xe_hpc.inl b/shared/source/helpers/compiler_product_helper_before_xe_hpc.inl
@@ -16,6 +16,11 @@ bool CompilerProductHelperHw<gfxProduct>::isForceToStatelessRequired() const {
     return false;
 }
 
+template <PRODUCT_FAMILY gfxProduct>
+bool CompilerProductHelperHw<gfxProduct>::isMatrixMultiplyAccumulateTF32Supported(const HardwareInfo &hwInfo) const {
+    return false;
+}
+
 template <PRODUCT_FAMILY gfxProduct>
 bool CompilerProductHelperHw<gfxProduct>::isSubgroupNamedBarrierSupported() const {
     return false;
diff --git a/shared/source/helpers/compiler_product_helper_xe_hpc_and_later.inl b/shared/source/helpers/compiler_product_helper_xe_hpc_and_later.inl
@@ -20,6 +20,11 @@ bool CompilerProductHelperHw<gfxProduct>::isForceToStatelessRequired() const {
     return true;
 }
 
+template <PRODUCT_FAMILY gfxProduct>
+bool CompilerProductHelperHw<gfxProduct>::isMatrixMultiplyAccumulateTF32Supported(const HardwareInfo &hwInfo) const {
+    return true;
+}
+
 template <PRODUCT_FAMILY gfxProduct>
 bool CompilerProductHelperHw<gfxProduct>::isSubgroupNamedBarrierSupported() const {
     return true;
diff --git a/shared/source/xe_hpc_core/enable_compiler_product_helper_pvc.cpp b/shared/source/xe_hpc_core/enable_compiler_product_helper_pvc.cpp
@@ -77,6 +77,14 @@ bool CompilerProductHelperHw<IGFX_PVC>::isMatrixMultiplyAccumulateSupported(cons
     return true;
 }
 
+template <>
+bool CompilerProductHelperHw<IGFX_PVC>::isMatrixMultiplyAccumulateTF32Supported(const HardwareInfo &hwInfo) const {
+    auto config = getProductConfigFromHwInfo(hwInfo);
+    if (config >= AOT::PVC_XT_B0 && config < AOT::PVC_XT_C0_VG)
+        return true;
+    return false;
+}
+
 template <>
 bool CompilerProductHelperHw<IGFX_PVC>::isBFloat16ConversionSupported(const ReleaseHelper *releaseHelper) const {
     return true;
diff --git a/shared/test/unit_test/helpers/compiler_product_helper_tests.cpp b/shared/test/unit_test/helpers/compiler_product_helper_tests.cpp
@@ -168,6 +168,20 @@ HWTEST2_F(CompilerProductHelperFixture, GivenReleaseHelperThenMatrixMultiplyAccu
     }
 }
 
+HWTEST2_F(CompilerProductHelperFixture, GivenXeHpcAndLaterThenMatrixMultiplyAccumulateTF32IsSupported, IsAtLeastXeHpcCore) {
+    auto &compilerProductHelper = pDevice->getCompilerProductHelper();
+    auto hwInfo = *defaultHwInfo;
+
+    EXPECT_TRUE(compilerProductHelper.isMatrixMultiplyAccumulateTF32Supported(hwInfo));
+}
+
+HWTEST2_F(CompilerProductHelperFixture, GivenPreXeHpcThenMatrixMultiplyAccumulateTF32IsNotSupported, IsAtMostXeHpgCore) {
+    auto &compilerProductHelper = pDevice->getCompilerProductHelper();
+    auto hwInfo = *defaultHwInfo;
+
+    EXPECT_FALSE(compilerProductHelper.isMatrixMultiplyAccumulateTF32Supported(hwInfo));
+}
+
 HWTEST2_F(CompilerProductHelperFixture, GivenReleaseHelperThenDotProductAccumulateSystolicIsSupportedBasedOnReleaseHelper, IsNotXeHpcCore) {
     auto &compilerProductHelper = pDevice->getCompilerProductHelper();
     auto releaseHelper = pDevice->getReleaseHelper();
diff --git a/shared/test/unit_test/xe_hpc_core/pvc/compiler_product_helper_tests_pvc.cpp b/shared/test/unit_test/xe_hpc_core/pvc/compiler_product_helper_tests_pvc.cpp
@@ -1,12 +1,14 @@
 /*
- * Copyright (C) 2023 Intel Corporation
+ * Copyright (C) 2023-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
  */
 
 #include "shared/source/helpers/compiler_product_helper.h"
+#include "shared/source/os_interface/product_helper.h"
 #include "shared/source/xe_hpc_core/hw_cmds_pvc.h"
+#include "shared/source/xe_hpc_core/pvc/device_ids_configs_pvc.h"
 #include "shared/test/common/mocks/mock_execution_environment.h"
 #include "shared/test/common/test_macros/header/per_product_test_definitions.h"
 #include "shared/test/common/test_macros/test.h"
@@ -42,3 +44,40 @@ PVCTEST_F(CompilerProductHelperPvcTest, givenPvcWhenFailBuildProgramWithStateful
     auto &compilerProductHelper = rootDeviceEnvironment.getHelper<CompilerProductHelper>();
     EXPECT_FALSE(compilerProductHelper.failBuildProgramWithStatefulAccessPreference());
 }
+
+PVCTEST_F(CompilerProductHelperPvcTest, givenPvcB0AndLaterThenMatrixMultiplyAccumulateTF32IsSupported) {
+    MockExecutionEnvironment executionEnvironment{};
+    auto &rootDeviceEnvironment = *executionEnvironment.rootDeviceEnvironments[0];
+    auto &compilerProductHelper = rootDeviceEnvironment.getHelper<CompilerProductHelper>();
+    auto &productHelper = rootDeviceEnvironment.getHelper<ProductHelper>();
+
+    auto hwInfo = *rootDeviceEnvironment.getHardwareInfo();
+
+    for (auto revision : {REVISION_A0, REVISION_B}) {
+        for (auto deviceId : pvcXlDeviceIds) {
+            hwInfo.platform.usDeviceID = deviceId;
+            hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(revision, hwInfo);
+            EXPECT_FALSE(compilerProductHelper.isMatrixMultiplyAccumulateTF32Supported(hwInfo));
+        }
+    }
+
+    for (auto deviceId : pvcXtDeviceIds) {
+        hwInfo.platform.usDeviceID = deviceId;
+        hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_A0, hwInfo);
+        EXPECT_FALSE(compilerProductHelper.isMatrixMultiplyAccumulateTF32Supported(hwInfo));
+    }
+
+    for (auto revision : {REVISION_B, REVISION_C}) {
+        for (auto deviceId : pvcXtDeviceIds) {
+            hwInfo.platform.usDeviceID = deviceId;
+            hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(revision, hwInfo);
+            EXPECT_TRUE(compilerProductHelper.isMatrixMultiplyAccumulateTF32Supported(hwInfo));
+        }
+    }
+
+    for (auto deviceId : pvcXtVgDeviceIds) {
+        hwInfo.platform.usDeviceID = deviceId;
+        hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_C, hwInfo);
+        EXPECT_FALSE(compilerProductHelper.isMatrixMultiplyAccumulateTF32Supported(hwInfo));
+    }
+}

Original file line number	Diff line number	Diff line change
`@@ -181,6 +181,10 @@ std::string CompilerProductHelperHw<gfxProduct>::getDeviceExtensions(const Hardw`
`181`	`181`	`extensions += "cl_intel_subgroup_matrix_multiply_accumulate ";`
`182`	`182`	`}`
`183`	`183`
	`184`	`+ if (isMatrixMultiplyAccumulateTF32Supported(hwInfo)) {`
	`185`	`+ extensions += "cl_intel_subgroup_matrix_multiply_accumulate_tf32 ";`
	`186`	`+ }`
	`187`	`+`
`184`	`188`	`if (isSplitMatrixMultiplyAccumulateSupported(releaseHelper)) {`
`185`	`189`	`extensions += "cl_intel_subgroup_split_matrix_multiply_accumulate ";`
`186`	`190`	`}`