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

fix: Consider slm size in suggest work group cache 

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk 2023-06-23 10:20:15 +00:00 committed by Compute-Runtime-Automation
parent 97b4d8bab5
commit bc0a3a7eb5
4 changed files with 82 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
level_zero/core/source/kernel/kernel_imp.cpp
View File

@ -383,12 +383,13 @@ ze_result_t KernelImp::suggestGroupSize(uint32_t globalSizeX, uint32_t globalSiz
dim = (globalSizeZ > 1U) ? 3 : dim; dim = (globalSizeZ > 1U) ? 3 : dim;
auto cachedGroupSize = std::find_if(this->suggestGroupSizeCache.begin(), this->suggestGroupSizeCache.end(), [&](const auto &other) { auto cachedGroupSize = std::find_if(this->suggestGroupSizeCache.begin(), this->suggestGroupSizeCache.end(), [&](const auto &other) {
return other.first == workItems; return other.groupSize == workItems &&
other.slmArgsTotalSize == this->getSlmTotalSize();
}); });
if (cachedGroupSize != this->suggestGroupSizeCache.end()) { if (cachedGroupSize != this->suggestGroupSizeCache.end()) {
*groupSizeX = static_cast<uint32_t>(cachedGroupSize->second.x); *groupSizeX = static_cast<uint32_t>(cachedGroupSize->suggestedGroupSize.x);
*groupSizeY = static_cast<uint32_t>(cachedGroupSize->second.y); *groupSizeY = static_cast<uint32_t>(cachedGroupSize->suggestedGroupSize.y);
*groupSizeZ = static_cast<uint32_t>(cachedGroupSize->second.z); *groupSizeZ = static_cast<uint32_t>(cachedGroupSize->suggestedGroupSize.z);
return ZE_RESULT_SUCCESS; return ZE_RESULT_SUCCESS;
} }
@ -420,7 +421,7 @@ ze_result_t KernelImp::suggestGroupSize(uint32_t globalSizeX, uint32_t globalSiz
*groupSizeX = static_cast<uint32_t>(retGroupSize[0]); *groupSizeX = static_cast<uint32_t>(retGroupSize[0]);
*groupSizeY = static_cast<uint32_t>(retGroupSize[1]); *groupSizeY = static_cast<uint32_t>(retGroupSize[1]);
*groupSizeZ = static_cast<uint32_t>(retGroupSize[2]); *groupSizeZ = static_cast<uint32_t>(retGroupSize[2]);
this->suggestGroupSizeCache.push_back(std::make_pair(Vec3(workItems), Vec3(retGroupSize))); this->suggestGroupSizeCache.emplace_back(workItems, this->getSlmTotalSize(), retGroupSize);
return ZE_RESULT_SUCCESS; return ZE_RESULT_SUCCESS;
} }

10
level_zero/core/source/kernel/kernel_imp.h
View File

@ -241,8 +241,14 @@ struct KernelImp : Kernel {
std::unique_ptr<KernelExt> pExtension; std::unique_ptr<KernelExt> pExtension;
std::mutex printfLock; std::mutex printfLock;
using SuggestGroupSizeCacheT = std::vector<std::pair<Vec3<size_t>, Vec3<size_t>>>; struct SuggestGroupSizeCacheEntry {
SuggestGroupSizeCacheT suggestGroupSizeCache; Vec3<size_t> groupSize;
uint32_t slmArgsTotalSize = 0u;
Vec3<size_t> suggestedGroupSize;
SuggestGroupSizeCacheEntry(size_t groupSize[3], uint32_t slmArgsTotalSize, size_t suggestedGroupSize[3]) : groupSize(groupSize), slmArgsTotalSize(slmArgsTotalSize), suggestedGroupSize(suggestedGroupSize){};
};
std::vector<SuggestGroupSizeCacheEntry> suggestGroupSizeCache;
}; };
} // namespace L0 } // namespace L0

1
level_zero/core/test/unit_tests/mocks/mock_kernel.h
View File

@ -64,6 +64,7 @@ struct WhiteBox<::L0::Kernel> : public ::L0::KernelImp {
using ::L0::KernelImp::requiredWorkgroupOrder; using ::L0::KernelImp::requiredWorkgroupOrder;
using ::L0::KernelImp::residencyContainer; using ::L0::KernelImp::residencyContainer;
using ::L0::KernelImp::setAssertBuffer; using ::L0::KernelImp::setAssertBuffer;
using ::L0::KernelImp::slmArgsTotalSize;
using ::L0::KernelImp::suggestGroupSizeCache; using ::L0::KernelImp::suggestGroupSizeCache;
using ::L0::KernelImp::surfaceStateHeapData; using ::L0::KernelImp::surfaceStateHeapData;
using ::L0::KernelImp::surfaceStateHeapDataSize; using ::L0::KernelImp::surfaceStateHeapDataSize;

100
level_zero/core/test/unit_tests/sources/kernel/test_kernel_2.cpp
View File

@ -129,52 +129,86 @@ TEST_F(KernelImp, WhenSuggestingGroupSizeThenCacheValues) {
kernel.module = &module; kernel.module = &module;
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 0u); EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 0u);
EXPECT_EQ(kernel.getSlmTotalSize(), 0u);
uint32_t groupSize[3]; uint32_t groupSize[3];
kernel.KernelImp::suggestGroupSize(256, 1, 1, groupSize, groupSize + 1, groupSize + 2); kernel.KernelImp::suggestGroupSize(256, 1, 1, groupSize, groupSize + 1, groupSize + 2);
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 1u); EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[0], 256u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].slmArgsTotalSize, 0u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[0], 256u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], 8u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], groupSize[0]); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], groupSize[1]); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], groupSize[0]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], groupSize[2]); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], groupSize[1]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], groupSize[2]);
kernel.KernelImp::suggestGroupSize(256, 1, 1, groupSize, groupSize + 1, groupSize + 2); kernel.KernelImp::suggestGroupSize(256, 1, 1, groupSize, groupSize + 1, groupSize + 2);
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 1u); EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[0], 256u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].slmArgsTotalSize, 0u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[0], 256u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], 8u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], groupSize[0]); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], groupSize[1]); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], groupSize[0]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], groupSize[2]); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], groupSize[1]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], groupSize[2]);
kernel.KernelImp::suggestGroupSize(2048, 1, 1, groupSize, groupSize + 1, groupSize + 2); kernel.KernelImp::suggestGroupSize(2048, 1, 1, groupSize, groupSize + 1, groupSize + 2);
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 2u); EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 2u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[0], 256u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].slmArgsTotalSize, 0u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[0], 256u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], 8u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].first[0], 2048u); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].first[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[1].slmArgsTotalSize, 0u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].first[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[1].groupSize[0], 2048u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].second[0], 8u); EXPECT_EQ(kernel.suggestGroupSizeCache[1].groupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].second[1], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[1].groupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].second[2], 1u); EXPECT_EQ(kernel.suggestGroupSizeCache[1].suggestedGroupSize[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], groupSize[0]); EXPECT_EQ(kernel.suggestGroupSizeCache[1].suggestedGroupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], groupSize[1]); EXPECT_EQ(kernel.suggestGroupSizeCache[1].suggestedGroupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], groupSize[2]); EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], groupSize[0]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], groupSize[1]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], groupSize[2]);
kernel.slmArgsTotalSize = 1;
kernel.KernelImp::suggestGroupSize(2048, 1, 1, groupSize, groupSize + 1, groupSize + 2);
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 3u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].slmArgsTotalSize, 0u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[0], 256u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].groupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].slmArgsTotalSize, 0u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].groupSize[0], 2048u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].groupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].groupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].suggestedGroupSize[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].suggestedGroupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].suggestedGroupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[2].slmArgsTotalSize, 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[2].groupSize[0], 2048u);
EXPECT_EQ(kernel.suggestGroupSizeCache[2].groupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[2].groupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[2].suggestedGroupSize[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[2].suggestedGroupSize[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[2].suggestedGroupSize[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[0], groupSize[0]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[1], groupSize[1]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].suggestedGroupSize[2], groupSize[2]);
} }
class KernelImpSuggestGroupSize : public DeviceFixture, public ::testing::TestWithParam<uint32_t> { class KernelImpSuggestGroupSize : public DeviceFixture, public ::testing::TestWithParam<uint32_t> {