Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-20 13:11:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add L1 cache flag in hw_helper.

Browse Source 
Related-To: NEO-4611
Change-Id: Ie135a2d5ff4f18f5d3b26202b2b6e53781e97501
This commit is contained in:
Sebastian Luzynski
2020-05-07 11:52:25 +02:00
committed by sys_ocldev
parent 8ce44b0689
commit c38f95592e
3 changed files with 23 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
opencl/test/unit_test/helpers/hw_helper_tests.cpp
View File

@ -294,7 +294,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenNoAllocationProvidedThe
size_t offset = 0x1000;
uint32_t pitch = 0x40;
SURFACE_TYPE type = RENDER_SURFACE_STATE::SURFACE_TYPE_SURFTYPE_BUFFER;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, offset, pitch, nullptr, false, type, true);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, offset, pitch, nullptr, false, type, true, false);
RENDER_SURFACE_STATE *state = reinterpret_cast<RENDER_SURFACE_STATE *>(stateBuffer);
EXPECT_EQ(length.SurfaceState.Depth + 1u, state->getDepth());
@ -310,7 +310,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenNoAllocationProvidedThe
size = 0x1003;
length.Length = static_cast<uint32_t>(alignUp(size, 4) - 1);
bool isReadOnly = false;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, nullptr, isReadOnly, type, true);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, nullptr, isReadOnly, type, true, false);
EXPECT_EQ(gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED), state->getMemoryObjectControlState());
EXPECT_EQ(length.SurfaceState.Depth + 1u, state->getDepth());
EXPECT_EQ(length.SurfaceState.Width + 1u, state->getWidth());
@ -320,7 +320,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenNoAllocationProvidedThe
size = 0x1000;
addr = 0x2001;
length.Length = static_cast<uint32_t>(size - 1);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, nullptr, isReadOnly, type, true);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, nullptr, isReadOnly, type, true, false);
EXPECT_EQ(gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED), state->getMemoryObjectControlState());
EXPECT_EQ(length.SurfaceState.Depth + 1u, state->getDepth());
EXPECT_EQ(length.SurfaceState.Width + 1u, state->getWidth());
@ -331,7 +331,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenNoAllocationProvidedThe
size = 0x1005;
length.Length = static_cast<uint32_t>(alignUp(size, 4) - 1);
isReadOnly = true;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, nullptr, isReadOnly, type, true);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, nullptr, isReadOnly, type, true, false);
EXPECT_EQ(gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER), state->getMemoryObjectControlState());
EXPECT_EQ(length.SurfaceState.Depth + 1u, state->getDepth());
EXPECT_EQ(length.SurfaceState.Width + 1u, state->getWidth());
@ -366,7 +366,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenAllocationProvidedThenU
GraphicsAllocation allocation(0, GraphicsAllocation::AllocationType::UNKNOWN, cpuAddr, gpuAddr, 0u, allocSize, MemoryPool::MemoryNull);
allocation.setDefaultGmm(new Gmm(pDevice->getGmmClientContext(), allocation.getUnderlyingBuffer(), allocation.getUnderlyingBufferSize(), false));
SURFACE_TYPE type = RENDER_SURFACE_STATE::SURFACE_TYPE_SURFTYPE_BUFFER;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, true);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, true, false);
EXPECT_EQ(length.SurfaceState.Depth + 1u, state->getDepth());
EXPECT_EQ(length.SurfaceState.Width + 1u, state->getWidth());
EXPECT_EQ(length.SurfaceState.Height + 1u, state->getHeight());
@ -404,7 +404,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenGmmAndAllocationCompres
allocation.setDefaultGmm(new Gmm(rootDeviceEnvironment.getGmmClientContext(), allocation.getUnderlyingBuffer(), allocation.getUnderlyingBufferSize(), false));
allocation.getDefaultGmm()->isRenderCompressed = true;
SURFACE_TYPE type = RENDER_SURFACE_STATE::SURFACE_TYPE_SURFTYPE_BUFFER;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, false);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, false, false);
EXPECT_EQ(RENDER_SURFACE_STATE::COHERENCY_TYPE_GPU_COHERENT, state->getCoherencyType());
EXPECT_EQ(AUXILIARY_SURFACE_MODE::AUXILIARY_SURFACE_MODE_AUX_CCS_E, state->getAuxiliarySurfaceMode());
@ -436,7 +436,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenGmmCompressionEnabledAn
allocation.setDefaultGmm(new Gmm(rootDeviceEnvironment.getGmmClientContext(), allocation.getUnderlyingBuffer(), allocation.getUnderlyingBufferSize(), false));
allocation.getDefaultGmm()->isRenderCompressed = true;
SURFACE_TYPE type = RENDER_SURFACE_STATE::SURFACE_TYPE_SURFTYPE_BUFFER;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, false);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, false, false);
EXPECT_EQ(RENDER_SURFACE_STATE::COHERENCY_TYPE_IA_COHERENT, state->getCoherencyType());
EXPECT_EQ(AUXILIARY_SURFACE_MODE::AUXILIARY_SURFACE_MODE_AUX_NONE, state->getAuxiliarySurfaceMode());
@ -467,7 +467,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenGmmCompressionDisabledA
GraphicsAllocation allocation(0, GraphicsAllocation::AllocationType::BUFFER_COMPRESSED, cpuAddr, gpuAddr, 0u, allocSize, MemoryPool::MemoryNull);
allocation.setDefaultGmm(new Gmm(rootDeviceEnvironment.getGmmClientContext(), allocation.getUnderlyingBuffer(), allocation.getUnderlyingBufferSize(), false));
SURFACE_TYPE type = RENDER_SURFACE_STATE::SURFACE_TYPE_SURFTYPE_BUFFER;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, false);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, false, false);
EXPECT_EQ(RENDER_SURFACE_STATE::COHERENCY_TYPE_IA_COHERENT, state->getCoherencyType());
EXPECT_EQ(AUXILIARY_SURFACE_MODE::AUXILIARY_SURFACE_MODE_AUX_NONE, state->getAuxiliarySurfaceMode());
@ -499,7 +499,7 @@ HWTEST_F(HwHelperTest, givenCreatedSurfaceStateBufferWhenGmmAndAllocationCompres
allocation.setDefaultGmm(new Gmm(rootDeviceEnvironment.getGmmClientContext(), allocation.getUnderlyingBuffer(), allocation.getUnderlyingBufferSize(), false));
allocation.getDefaultGmm()->isRenderCompressed = true;
SURFACE_TYPE type = RENDER_SURFACE_STATE::SURFACE_TYPE_SURFTYPE_BUFFER;
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, true);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, stateBuffer, size, addr, 0, pitch, &allocation, false, type, true, false);
EXPECT_EQ(RENDER_SURFACE_STATE::COHERENCY_TYPE_IA_COHERENT, state->getCoherencyType());
EXPECT_EQ(AUXILIARY_SURFACE_MODE::AUXILIARY_SURFACE_MODE_AUX_NONE, state->getAuxiliarySurfaceMode());
@ -552,7 +552,7 @@ HWTEST_F(HwHelperTest, DISABLED_profilingCreationOfRenderSurfaceStateVsMemcpyOfC
for (uint32_t i = 0; i < maxLoop; ++i) {
auto t1 = std::chrono::high_resolution_clock::now();
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, surfaceStates[i], size, addr, 0, pitch, nullptr, false, type, true);
helper.setRenderSurfaceStateForBuffer(rootDeviceEnvironment, surfaceStates[i], size, addr, 0, pitch, nullptr, false, type, true, false);
auto t2 = std::chrono::high_resolution_clock::now();
timesCreate.push_back(t1);
timesCreate.push_back(t2);

8
shared/source/helpers/hw_helper.h
View File

@ -66,7 +66,8 @@ class HwHelper {
GraphicsAllocation *gfxAlloc,
bool isReadOnly,
uint32_t surfaceType,
bool forceNonAuxMode) = 0;
bool forceNonAuxMode,
bool useL1Cache) = 0;
virtual const EngineInstancesContainer getGpgpuEngineInstances(const HardwareInfo &hwInfo) const = 0;
virtual const StackVec<size_t, 3> getDeviceSubGroupSizes() const = 0;
virtual bool getEnableLocalMemory(const HardwareInfo &hwInfo) const = 0;
@ -194,7 +195,10 @@ class HwHelperHw : public HwHelper {
GraphicsAllocation *gfxAlloc,
bool isReadOnly,
uint32_t surfaceType,
bool forceNonAuxMode) override;
bool forceNonAuxMode,
bool useL1Cache) override;
void setL1CachePolicy(bool useL1Cache, typename GfxFamily::RENDER_SURFACE_STATE *surfaceState, const HardwareInfo *hwInfo);
const EngineInstancesContainer getGpgpuEngineInstances(const HardwareInfo &hwInfo) const override;

8
shared/source/helpers/hw_helper_base.inl
View File

@ -103,7 +103,8 @@ void HwHelperHw<Family>::setRenderSurfaceStateForBuffer(const RootDeviceEnvironm
GraphicsAllocation *gfxAlloc,
bool isReadOnly,
uint32_t surfaceType,
bool forceNonAuxMode) {
bool forceNonAuxMode,
bool useL1Cache) {
using RENDER_SURFACE_STATE = typename Family::RENDER_SURFACE_STATE;
using SURFACE_FORMAT = typename RENDER_SURFACE_STATE::SURFACE_FORMAT;
using AUXILIARY_SURFACE_MODE = typename RENDER_SURFACE_STATE::AUXILIARY_SURFACE_MODE;
@ -158,8 +159,13 @@ void HwHelperHw<Family>::setRenderSurfaceStateForBuffer(const RootDeviceEnvironm
state.setAuxiliarySurfaceMode(AUXILIARY_SURFACE_MODE::AUXILIARY_SURFACE_MODE_AUX_NONE);
}
*surfaceState = state;
setL1CachePolicy(useL1Cache, surfaceState, rootDeviceEnvironment.getHardwareInfo());
}
template <typename GfxFamily>
void NEO::HwHelperHw<GfxFamily>::setL1CachePolicy(bool useL1Cache, typename GfxFamily::RENDER_SURFACE_STATE *surfaceState, const HardwareInfo *hwInfo) {}
template <typename Family>
bool HwHelperHw<Family>::getEnableLocalMemory(const HardwareInfo &hwInfo) const {
if (DebugManager.flags.EnableLocalMemory.get() != -1) {