intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-20 00:24:58 +08:00
Code Issues Packages Projects Releases Wiki Activity

performance: Cache suggest group size

Browse Source 
Resolves: NEO-7968

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk
2023-06-16 10:17:51 +00:00
committed by Compute-Runtime-Automation
parent 1146f42bcb
commit 0cf975605b
4 changed files with 85 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -374,6 +374,16 @@ ze_result_t KernelImp::suggestGroupSize(uint32_t globalSizeX, uint32_t globalSiz
uint32_t dim = (globalSizeY > 1U) ? 2 : 1U;
dim = (globalSizeZ > 1U) ? 3 : dim;
auto cachedGroupSize = std::find_if(this->suggestGroupSizeCache.begin(), this->suggestGroupSizeCache.end(), [&](const auto &other) {
return other.first == workItems;
});
if (cachedGroupSize != this->suggestGroupSizeCache.end()) {
*groupSizeX = static_cast<uint32_t>(cachedGroupSize->second.x);
*groupSizeY = static_cast<uint32_t>(cachedGroupSize->second.y);
*groupSizeZ = static_cast<uint32_t>(cachedGroupSize->second.z);
return ZE_RESULT_SUCCESS;
}
if (NEO::DebugManager.flags.EnableComputeWorkSizeND.get()) {
auto usesImages = kernelDescriptor.kernelAttributes.flags.usesImages;
auto neoDevice = module->getDevice()->getNEODevice();
@@ -402,6 +412,7 @@ ze_result_t KernelImp::suggestGroupSize(uint32_t globalSizeX, uint32_t globalSiz
*groupSizeX = static_cast<uint32_t>(retGroupSize[0]);
*groupSizeY = static_cast<uint32_t>(retGroupSize[1]);
*groupSizeZ = static_cast<uint32_t>(retGroupSize[2]);
this->suggestGroupSizeCache.push_back(std::make_pair(Vec3(workItems), Vec3(retGroupSize)));
return ZE_RESULT_SUCCESS;
}

5
level_zero/core/source/kernel/kernel_imp.h
View File

@@ -8,6 +8,7 @@
#pragma once
#include "shared/source/command_stream/thread_arbitration_policy.h"
#include "shared/source/helpers/vec.h"
#include "shared/source/kernel/dispatch_kernel_encoder_interface.h"
#include "shared/source/memory_manager/unified_memory_manager.h"
#include "shared/source/unified_memory/unified_memory.h"
@@ -16,6 +17,7 @@
#include <memory>
#include <mutex>
#include <vector>
namespace L0 {
@@ -236,6 +238,9 @@ struct KernelImp : Kernel {
std::unique_ptr<KernelExt> pExtension;
std::mutex printfLock;
using SuggestGroupSizeCacheT = std::vector<std::pair<Vec3<size_t>, Vec3<size_t>>>;
SuggestGroupSizeCacheT suggestGroupSizeCache;
};
} // namespace L0

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

@@ -63,6 +63,7 @@ struct WhiteBox<::L0::Kernel> : public ::L0::KernelImp {
using ::L0::KernelImp::requiredWorkgroupOrder;
using ::L0::KernelImp::residencyContainer;
using ::L0::KernelImp::setAssertBuffer;
using ::L0::KernelImp::suggestGroupSizeCache;
using ::L0::KernelImp::surfaceStateHeapData;
using ::L0::KernelImp::surfaceStateHeapDataSize;
using ::L0::KernelImp::unifiedMemoryControls;

81
level_zero/core/test/unit_tests/sources/kernel/test_kernel_2.cpp
View File

@@ -112,6 +112,71 @@ TEST_F(KernelImp, WhenSuggestingGroupSizeThenClampToMaxGroupSize) {
EXPECT_EQ(1U, groupSize[2]);
}
TEST_F(KernelImp, WhenSuggestingGroupSizeThenCacheValues) {
DebugManagerStateRestore restorer;
WhiteBox<KernelImmutableData> kernelInfo = {};
NEO::KernelDescriptor descriptor;
kernelInfo.kernelDescriptor = &descriptor;
NEO::DebugManager.flags.EnableComputeWorkSizeND.set(false);
Mock<Module> module(device, nullptr);
module.getMaxGroupSizeResult = 8;
Mock<Kernel> kernel;
kernel.kernelImmData = &kernelInfo;
kernel.module = &module;
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 0u);
uint32_t groupSize[3];
kernel.KernelImp::suggestGroupSize(256, 1, 1, groupSize, groupSize + 1, groupSize + 2);
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[0], 256u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], groupSize[0]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], groupSize[1]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], groupSize[2]);
kernel.KernelImp::suggestGroupSize(256, 1, 1, groupSize, groupSize + 1, groupSize + 2);
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[0], 256u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], groupSize[0]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], groupSize[1]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], groupSize[2]);
kernel.KernelImp::suggestGroupSize(2048, 1, 1, groupSize, groupSize + 1, groupSize + 2);
EXPECT_EQ(kernel.suggestGroupSizeCache.size(), 2u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[0], 256u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].first[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].first[0], 2048u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].first[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].first[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].second[0], 8u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].second[1], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[1].second[2], 1u);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[0], groupSize[0]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[1], groupSize[1]);
EXPECT_EQ(kernel.suggestGroupSizeCache[0].second[2], groupSize[2]);
}
class KernelImpSuggestGroupSize : public DeviceFixture, public ::testing::TestWithParam<uint32_t> {
public:
void SetUp() override {
@@ -201,24 +266,14 @@ TEST_P(KernelImpSuggestGroupSize, WhenSlmSizeExceedsLocalMemorySizeAndSuggesting
function.kernelImmData = &funcInfo;
function.module = &module;
uint32_t groupSize[3];
EXPECT_EQ(ZE_RESULT_SUCCESS, function.KernelImp::suggestGroupSize(size, 1, 1, groupSize, groupSize + 1, groupSize + 2));
::testing::internal::CaptureStderr();
auto localMemSize = static_cast<uint32_t>(device->getNEODevice()->getDeviceInfo().localMemSize);
::testing::internal::CaptureStderr();
funcInfo.kernelDescriptor->kernelAttributes.slmInlineSize = localMemSize - 10u;
EXPECT_EQ(ZE_RESULT_SUCCESS, function.KernelImp::suggestGroupSize(size, 1, 1, groupSize, groupSize + 1, groupSize + 2));
std::string output = testing::internal::GetCapturedStderr();
EXPECT_EQ(std::string(""), output);
::testing::internal::CaptureStderr();
funcInfo.kernelDescriptor->kernelAttributes.slmInlineSize = localMemSize + 10u;
EXPECT_EQ(ZE_RESULT_ERROR_OUT_OF_DEVICE_MEMORY, function.KernelImp::suggestGroupSize(size, 1, 1, groupSize, groupSize + 1, groupSize + 2));
output = testing::internal::GetCapturedStderr();
auto output = testing::internal::GetCapturedStderr();
const auto &slmInlineSize = funcInfo.kernelDescriptor->kernelAttributes.slmInlineSize;
std::string expectedOutput = "Size of SLM (" + std::to_string(slmInlineSize) + ") larger than available (" + std::to_string(localMemSize) + ")\n";
EXPECT_EQ(expectedOutput, output);