Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix: correct number of max work group count for concurrent kernel on PVC 

for single-CCS mode use all EUs

Related-To: NEO-8377
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
This commit is contained in:
Mateusz Jablonski 2024-02-26 09:35:17 +00:00 committed by Compute-Runtime-Automation
parent 5c4be8df54
commit 3e65e7bdba
3 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
level_zero/core/test/unit_tests/xe_hpc_core/pvc/test_kernel_pvc.cpp
View File

@ -1,5 +1,5 @@
/* /*
* Copyright (C) 2023 Intel Corporation * Copyright (C) 2023-2024 Intel Corporation
* *
* SPDX-License-Identifier: MIT * SPDX-License-Identifier: MIT
* *
@ -22,21 +22,21 @@ using KernelImpSuggestMaxCooperativeGroupCountTestsPvc = Test<L0::ult::KernelImp
PVCTEST_F(KernelImpSuggestMaxCooperativeGroupCountTestsPvc, GivenNoBarriersOrSlmUsedWhenCalculatingMaxCooperativeGroupCountThenResultIsCalculatedWithSimd) { PVCTEST_F(KernelImpSuggestMaxCooperativeGroupCountTestsPvc, GivenNoBarriersOrSlmUsedWhenCalculatingMaxCooperativeGroupCountThenResultIsCalculatedWithSimd) {
auto workGroupSize = lws[0] * lws[1] * lws[2]; auto workGroupSize = lws[0] * lws[1] * lws[2];
auto expected = (availableThreadCount / Math::divideAndRoundUp(workGroupSize, simd)) / 2; auto expected = availableThreadCount / Math::divideAndRoundUp(workGroupSize, simd);
EXPECT_EQ(expected, getMaxWorkGroupCount()); EXPECT_EQ(expected, getMaxWorkGroupCount());
} }
PVCTEST_F(KernelImpSuggestMaxCooperativeGroupCountTestsPvc, GivenBarriersWhenCalculatingMaxCooperativeGroupCountThenResultIsCalculatedWithRegardToBarriersCount) { PVCTEST_F(KernelImpSuggestMaxCooperativeGroupCountTestsPvc, GivenBarriersWhenCalculatingMaxCooperativeGroupCountThenResultIsCalculatedWithRegardToBarriersCount) {
usesBarriers = 1; usesBarriers = 1;
auto expected = (dssCount * (maxBarrierCount / usesBarriers)) / 2; auto expected = dssCount * (maxBarrierCount / usesBarriers);
EXPECT_EQ(expected, getMaxWorkGroupCount()); EXPECT_EQ(expected, getMaxWorkGroupCount());
} }
PVCTEST_F(KernelImpSuggestMaxCooperativeGroupCountTestsPvc, GivenUsedSlmSizeWhenCalculatingMaxCooperativeGroupCountThenResultIsCalculatedWithRegardToUsedSlmSize) { PVCTEST_F(KernelImpSuggestMaxCooperativeGroupCountTestsPvc, GivenUsedSlmSizeWhenCalculatingMaxCooperativeGroupCountThenResultIsCalculatedWithRegardToUsedSlmSize) {
usedSlm = 64 * MemoryConstants::kiloByte; usedSlm = 64 * MemoryConstants::kiloByte;
auto expected = (availableSlm / usedSlm) / 2; auto expected = availableSlm / usedSlm;
EXPECT_EQ(expected, getMaxWorkGroupCount()); EXPECT_EQ(expected, getMaxWorkGroupCount());
} }
} // namespace ult } // namespace ult
} // namespace L0 } // namespace L0

2
shared/source/xe_hpc_core/pvc/os_agnostic_product_helper_pvc.inl
View File

@ -218,7 +218,7 @@ std::optional<aub_stream::ProductFamily> ProductHelperHw<gfxProduct>::getAubStre
template <> template <>
uint32_t ProductHelperHw<gfxProduct>::getNumberOfPartsInTileForConcurrentKernel(uint32_t ccsCount) const { uint32_t ProductHelperHw<gfxProduct>::getNumberOfPartsInTileForConcurrentKernel(uint32_t ccsCount) const {
if (ccsCount == 1) { if (ccsCount == 1) {
return 2; return 1;
} else if (ccsCount == 2) { } else if (ccsCount == 2) {
return 4; return 4;
} }

2
shared/test/unit_test/xe_hpc_core/pvc/gfx_core_helper_tests_pvc.cpp
View File

@ -188,7 +188,7 @@ PVCTEST_F(GfxCoreHelperTestsPvc, GivenCooperativeEngineSupportedAndNotUsedWhenAd
} else { } else {
for (uint32_t ccsCount : {1, 2, 4}) { for (uint32_t ccsCount : {1, 2, 4}) {
hwInfo.gtSystemInfo.CCSInfo.NumberOfCCSEnabled = ccsCount; hwInfo.gtSystemInfo.CCSInfo.NumberOfCCSEnabled = ccsCount;
tilePartsForConcurrentKernels = ccsCount == 1 ? 2 tilePartsForConcurrentKernels = ccsCount == 1 ? 1
: ccsCount == 2 ? 4 : ccsCount == 2 ? 4
: 8; : 8;
EXPECT_EQ(passedMaxWorkGroupCount / tilePartsForConcurrentKernels, gfxCoreHelper.adjustMaxWorkGroupCount(passedMaxWorkGroupCount, engineGroupType, rootDeviceEnvironment, isEngineInstanced)); EXPECT_EQ(passedMaxWorkGroupCount / tilePartsForConcurrentKernels, gfxCoreHelper.adjustMaxWorkGroupCount(passedMaxWorkGroupCount, engineGroupType, rootDeviceEnvironment, isEngineInstanced));