performance: optimize surface state programming

- eliminate read-modify-write on gfx memory when global bindless mode
enabled

Signed-off-by: Mateusz Hoppe <[email protected]>
Source: d5e275c

Author: HoppeMateusz

level_zero/core/source/kernel/patch_with_implicit_surface.inl

@@ -42,6 +42,7 @@ inline void patchImplicitArgBindlessOffsetAndSetSurfaceState(ArrayRef<uint8_t> c
     auto &gfxCoreHelper = device.getGfxCoreHelper();
     void *surfaceStateAddress = nullptr;
     auto surfaceStateSize = gfxCoreHelper.getRenderSurfaceStateSize();
+    bool useTempBuffer = false;
 
     if (NEO::isValidOffset(ptr.bindless)) {
         if (device.getBindlessHeapsHelper()) {
@@ -50,6 +51,7 @@ inline void patchImplicitArgBindlessOffsetAndSetSurfaceState(ArrayRef<uint8_t> c
             auto patchLocation = ptrOffset(crossThreadData.begin(), ptr.bindless);
             auto patchValue = gfxCoreHelper.getBindlessSurfaceExtendedMessageDescriptorValue(static_cast<uint32_t>(ssInHeap.surfaceStateOffset));
             patchWithRequiredSize(const_cast<uint8_t *>(patchLocation), sizeof(patchValue), patchValue);
+            useTempBuffer = true;
         } else {
             auto index = std::numeric_limits<uint32_t>::max();
             const auto &iter = kernelDescriptor.getBindlessOffsetToSurfaceState().find(ptr.bindless);
@@ -64,12 +66,18 @@ inline void patchImplicitArgBindlessOffsetAndSetSurfaceState(ArrayRef<uint8_t> c
     }
 
     if (surfaceStateAddress) {
+        std::unique_ptr<uint64_t[]> surfaceState;
+
+        if (useTempBuffer) {
+            surfaceState = std::make_unique<uint64_t[]>(surfaceStateSize / sizeof(uint64_t));
+        }
+
         auto addressToPatch = allocation->getGpuAddress();
         size_t sizeToPatch = allocation->getUnderlyingBufferSize();
         auto isDebuggerActive = device.getDebugger() != nullptr;
 
         NEO::EncodeSurfaceStateArgs args;
-        args.outMemory = surfaceStateAddress;
+        args.outMemory = useTempBuffer ? surfaceState.get() : surfaceStateAddress;
         args.graphicsAddress = addressToPatch;
         args.size = sizeToPatch;
         args.mocs = gfxCoreHelper.getMocsIndex(*device.getGmmHelper(), true, false) << 1;
@@ -81,5 +89,10 @@ inline void patchImplicitArgBindlessOffsetAndSetSurfaceState(ArrayRef<uint8_t> c
         args.isDebuggerActive = isDebuggerActive;
 
         gfxCoreHelper.encodeBufferSurfaceState(args);
+
+        if (useTempBuffer) {
+            memcpy_s(surfaceStateAddress, surfaceStateSize,
+                     surfaceState.get(), surfaceStateSize);
+        }
     }
 }

level_zero/core/test/unit_tests/sources/kernel/test_kernel_2.cpp

@@ -595,6 +595,7 @@ HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalVarBufferAndBindlessExplic
         void encodeBufferSurfaceState(EncodeSurfaceStateArgs &args) const override {
             savedSurfaceStateArgs = args;
             ++encodeBufferSurfaceStateCalled;
+            NEO::GfxCoreHelperHw<FamilyType>::encodeBufferSurfaceState(args);
         }
     };
 
@@ -651,6 +652,7 @@ HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalVarBufferAndBindlessExplic
 }
 
 HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalConstBufferAndBindlessExplicitAndImplicitArgsAndBindlessHeapsHelperWhenInitializeKernelImmutableDataThenSurfaceStateIsSetAndImplicitArgBindlessOffsetIsPatched, IsAtLeastXeHpgCore) {
+    using RENDER_SURFACE_STATE = typename FamilyType::RENDER_SURFACE_STATE;
     HardwareInfo hwInfo = *defaultHwInfo;
 
     auto device = std::unique_ptr<NEO::MockDevice>(NEO::MockDevice::createWithNewExecutionEnvironment<NEO::MockDevice>(&hwInfo, 0));
@@ -667,6 +669,7 @@ HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalConstBufferAndBindlessExpl
         void encodeBufferSurfaceState(EncodeSurfaceStateArgs &args) const override {
             savedSurfaceStateArgs = args;
             ++encodeBufferSurfaceStateCalled;
+            NEO::GfxCoreHelperHw<FamilyType>::encodeBufferSurfaceState(args);
         }
     };
 
@@ -724,12 +727,17 @@ HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalConstBufferAndBindlessExpl
         EXPECT_EQ(allocSize, savedSurfaceStateArgs.size);
         EXPECT_EQ(gpuAddress, savedSurfaceStateArgs.graphicsAddress);
 
-        EXPECT_EQ(globalConstBuffer.getBindlessInfo().ssPtr, savedSurfaceStateArgs.outMemory);
+        EXPECT_NE(globalConstBuffer.getBindlessInfo().ssPtr, savedSurfaceStateArgs.outMemory);
+
+        const auto surfState = reinterpret_cast<RENDER_SURFACE_STATE *>(globalConstBuffer.getBindlessInfo().ssPtr);
+        ASSERT_NE(nullptr, surfState);
+        EXPECT_EQ(gpuAddress, surfState->getSurfaceBaseAddress());
         EXPECT_EQ(&globalConstBuffer, savedSurfaceStateArgs.allocation);
     }
 }
 
 HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalVarBufferAndBindlessExplicitAndImplicitArgsAndBindlessHeapsHelperWhenInitializeKernelImmutableDataThenSurfaceStateIsSetAndImplicitArgBindlessOffsetIsPatched, IsAtLeastXeHpgCore) {
+    using RENDER_SURFACE_STATE = typename FamilyType::RENDER_SURFACE_STATE;
     HardwareInfo hwInfo = *defaultHwInfo;
 
     auto device = std::unique_ptr<NEO::MockDevice>(NEO::MockDevice::createWithNewExecutionEnvironment<NEO::MockDevice>(&hwInfo, 0));
@@ -746,6 +754,7 @@ HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalVarBufferAndBindlessExplic
         void encodeBufferSurfaceState(EncodeSurfaceStateArgs &args) const override {
             savedSurfaceStateArgs = args;
             ++encodeBufferSurfaceStateCalled;
+            NEO::GfxCoreHelperHw<FamilyType>::encodeBufferSurfaceState(args);
         }
     };
 
@@ -803,7 +812,11 @@ HWTEST2_F(KernelImmutableDataBindlessTest, givenGlobalVarBufferAndBindlessExplic
         EXPECT_EQ(allocSize, savedSurfaceStateArgs.size);
         EXPECT_EQ(gpuAddress, savedSurfaceStateArgs.graphicsAddress);
 
-        EXPECT_EQ(globalVarBuffer.getBindlessInfo().ssPtr, savedSurfaceStateArgs.outMemory);
+        EXPECT_NE(globalVarBuffer.getBindlessInfo().ssPtr, savedSurfaceStateArgs.outMemory);
+
+        const auto surfState = reinterpret_cast<RENDER_SURFACE_STATE *>(globalVarBuffer.getBindlessInfo().ssPtr);
+        ASSERT_NE(nullptr, surfState);
+        EXPECT_EQ(gpuAddress, surfState->getSurfaceBaseAddress());
         EXPECT_EQ(&globalVarBuffer, savedSurfaceStateArgs.allocation);
     }
 }