intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-21 01:04:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor: add param rootDeviceEnvironment to calculateNumThreadsPerThreadGroup

Browse Source 
Signed-off-by: Katarzyna Cencelewska <katarzyna.cencelewska@intel.com>
This commit is contained in:
Katarzyna Cencelewska
2024-03-21 18:54:41 +00:00
committed by Compute-Runtime-Automation
parent ec009cf9e3
commit dd1d52259e
30 changed files with 231 additions and 180 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
shared/source/kernel/implicit_args_helper.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2022-2023 Intel Corporation
* Copyright (C) 2022-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -7,6 +7,7 @@
#include "shared/source/kernel/implicit_args_helper.h"
#include "shared/source/execution_environment/root_device_environment.h"
#include "shared/source/helpers/aligned_memory.h"
#include "shared/source/helpers/basic_math.h"
#include "shared/source/helpers/hw_walk_order.h"
@@ -43,7 +44,7 @@ uint32_t getGrfSize(uint32_t simd) {
return 32u;
}
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor, bool isHwLocalIdGeneration, const GfxCoreHelper &gfxCoreHelper) {
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor, bool isHwLocalIdGeneration, const RootDeviceEnvironment &rootDeviceEnvironment) {
if (!pImplicitArgs) {
return 0;
}
@@ -58,15 +59,15 @@ uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const
auto itemsInGroup = Math::computeTotalElementsCount(localWorkSize);
uint32_t localIdsSizeNeeded =
alignUp(static_cast<uint32_t>(NEO::PerThreadDataHelper::getPerThreadDataSizeTotal(
simdSize, grfSize, 3u, itemsInGroup, isHwLocalIdGeneration, gfxCoreHelper)),
simdSize, grfSize, 3u, itemsInGroup, isHwLocalIdGeneration, rootDeviceEnvironment)),
MemoryConstants::cacheLineSize);
return implicitArgsSize + localIdsSizeNeeded;
}
}
void *patchImplicitArgs(void *ptrToPatch, const ImplicitArgs &implicitArgs, const KernelDescriptor &kernelDescriptor, std::optional<std::pair<bool, uint32_t>> hwGenerationOfLocalIdsParams, const GfxCoreHelper &gfxCoreHelper) {
void *patchImplicitArgs(void *ptrToPatch, const ImplicitArgs &implicitArgs, const KernelDescriptor &kernelDescriptor, std::optional<std::pair<bool, uint32_t>> hwGenerationOfLocalIdsParams, const RootDeviceEnvironment &rootDeviceEnvironment) {
auto localIdsGeneratedByHw = hwGenerationOfLocalIdsParams.has_value() ? hwGenerationOfLocalIdsParams.value().first : false;
auto totalSizeToProgram = getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor, localIdsGeneratedByHw, gfxCoreHelper);
auto totalSizeToProgram = getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor, localIdsGeneratedByHw, rootDeviceEnvironment);
auto retVal = ptrOffset(ptrToPatch, totalSizeToProgram);
auto patchImplicitArgsBufferInCrossThread = NEO::isValidOffset<>(kernelDescriptor.payloadMappings.implicitArgs.implicitArgsBuffer);
@@ -82,7 +83,7 @@ void *patchImplicitArgs(void *ptrToPatch, const ImplicitArgs &implicitArgs, cons
static_cast<uint16_t>(implicitArgs.localSizeY),
static_cast<uint16_t>(implicitArgs.localSizeZ)}},
dimensionOrder,
false, grfSize, gfxCoreHelper);
false, grfSize, rootDeviceEnvironment);
auto sizeForLocalIdsProgramming = totalSizeToProgram - ImplicitArgs::getSize();
ptrToPatch = ptrOffset(ptrToPatch, sizeForLocalIdsProgramming);
}

8
shared/source/kernel/implicit_args_helper.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2021-2023 Intel Corporation
* Copyright (C) 2021-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -17,14 +17,14 @@
namespace NEO {
struct KernelDescriptor;
class GfxCoreHelper;
struct RootDeviceEnvironment;
inline constexpr const char *implicitArgsRelocationSymbolName = "__INTEL_PATCH_CROSS_THREAD_OFFSET_OFF_R0";
namespace ImplicitArgsHelper {
std::array<uint8_t, 3> getDimensionOrderForLocalIds(const uint8_t *workgroupDimensionsOrder, std::optional<std::pair<bool /* localIdsGeneratedByRuntime */, uint32_t /* walkOrderForHwGenerationOfLocalIds */>> hwGenerationOfLocalIdsParams);
uint32_t getGrfSize(uint32_t simd);
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor, bool localIdsGeneratedByRuntime, const GfxCoreHelper &gfxCoreHelper);
void *patchImplicitArgs(void *ptrToPatch, const ImplicitArgs &implicitArgs, const KernelDescriptor &kernelDescriptor, std::optional<std::pair<bool /* localIdsGeneratedByRuntime */, uint32_t /* walkOrderForHwGenerationOfLocalIds */>> hwGenerationOfLocalIdsParams, const GfxCoreHelper &gfxCoreHelper);
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor, bool localIdsGeneratedByRuntime, const RootDeviceEnvironment &rootDeviceEnvironment);
void *patchImplicitArgs(void *ptrToPatch, const ImplicitArgs &implicitArgs, const KernelDescriptor &kernelDescriptor, std::optional<std::pair<bool /* localIdsGeneratedByRuntime */, uint32_t /* walkOrderForHwGenerationOfLocalIds */>> hwGenerationOfLocalIdsParams, const RootDeviceEnvironment &rootDeviceEnvironment);
} // namespace ImplicitArgsHelper
} // namespace NEO

18
shared/source/kernel/local_ids_cache.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2022-2023 Intel Corporation
* Copyright (C) 2022-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -7,6 +7,7 @@
#include "shared/source/kernel/local_ids_cache.h"
#include "shared/source/execution_environment/root_device_environment.h"
#include "shared/source/helpers/aligned_memory.h"
#include "shared/source/helpers/basic_math.h"
#include "shared/source/helpers/gfx_core_helper.h"
@@ -34,12 +35,13 @@ std::unique_lock<std::mutex> LocalIdsCache::lock() {
return std::unique_lock<std::mutex>(setLocalIdsMutex);
}
size_t LocalIdsCache::getLocalIdsSizeForGroup(const Vec3<uint16_t> &group, const GfxCoreHelper &gfxCoreHelper) const {
size_t LocalIdsCache::getLocalIdsSizeForGroup(const Vec3<uint16_t> &group, const RootDeviceEnvironment &rootDeviceEnvironment) const {
const auto numElementsInGroup = static_cast<uint32_t>(Math::computeTotalElementsCount({group[0], group[1], group[2]}));
if (isSimd1(simdSize)) {
return static_cast<size_t>(numElementsInGroup * localIdsSizePerThread);
}
const auto numberOfThreads = gfxCoreHelper.calculateNumThreadsPerThreadGroup(simdSize, numElementsInGroup, grfSize, false);
auto &gfxCoreHelper = rootDeviceEnvironment.getHelper<NEO::GfxCoreHelper>();
const auto numberOfThreads = gfxCoreHelper.calculateNumThreadsPerThreadGroup(simdSize, numElementsInGroup, grfSize, false, rootDeviceEnvironment);
return static_cast<size_t>(numberOfThreads * localIdsSizePerThread);
}
@@ -52,7 +54,7 @@ void LocalIdsCache::setLocalIdsForEntry(LocalIdsCacheEntry &entry, void *destina
std::memcpy(destination, entry.localIdsData, entry.localIdsSize);
}
void LocalIdsCache::setLocalIdsForGroup(const Vec3<uint16_t> &group, void *destination, const GfxCoreHelper &gfxCoreHelper) {
void LocalIdsCache::setLocalIdsForGroup(const Vec3<uint16_t> &group, void *destination, const RootDeviceEnvironment &rootDeviceEnvironment) {
auto setLocalIdsLock = lock();
LocalIdsCacheEntry *leastAccessedEntry = &cache[0];
for (auto &cacheEntry : cache) {
@@ -65,12 +67,12 @@ void LocalIdsCache::setLocalIdsForGroup(const Vec3<uint16_t> &group, void *desti
}
}
commitNewEntry(*leastAccessedEntry, group, gfxCoreHelper);
commitNewEntry(*leastAccessedEntry, group, rootDeviceEnvironment);
setLocalIdsForEntry(*leastAccessedEntry, destination);
}
void LocalIdsCache::commitNewEntry(LocalIdsCacheEntry &entry, const Vec3<uint16_t> &group, const GfxCoreHelper &gfxCoreHelper) {
entry.localIdsSize = getLocalIdsSizeForGroup(group, gfxCoreHelper);
void LocalIdsCache::commitNewEntry(LocalIdsCacheEntry &entry, const Vec3<uint16_t> &group, const RootDeviceEnvironment &rootDeviceEnvironment) {
entry.localIdsSize = getLocalIdsSizeForGroup(group, rootDeviceEnvironment);
entry.groupSize = group;
entry.accessCounter = 0U;
if (entry.localIdsSize > entry.localIdsSizeAllocated) {
@@ -79,7 +81,7 @@ void LocalIdsCache::commitNewEntry(LocalIdsCacheEntry &entry, const Vec3<uint16_
entry.localIdsSizeAllocated = entry.localIdsSize;
}
NEO::generateLocalIDs(entry.localIdsData, static_cast<uint16_t>(simdSize),
{group[0], group[1], group[2]}, wgDimOrder, usesOnlyImages, grfSize, gfxCoreHelper);
{group[0], group[1], group[2]}, wgDimOrder, usesOnlyImages, grfSize, rootDeviceEnvironment);
}
} // namespace NEO

10
shared/source/kernel/local_ids_cache.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2022-2023 Intel Corporation
* Copyright (C) 2022-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -12,7 +12,7 @@
#include <mutex>
namespace NEO {
class GfxCoreHelper;
struct RootDeviceEnvironment;
class LocalIdsCache {
public:
struct LocalIdsCacheEntry {
@@ -30,13 +30,13 @@ class LocalIdsCache {
LocalIdsCache(size_t cacheSize, std::array<uint8_t, 3> wgDimOrder, uint8_t simdSize, uint8_t grfSize, bool usesOnlyImages = false);
~LocalIdsCache();
void setLocalIdsForGroup(const Vec3<uint16_t> &group, void *destination, const GfxCoreHelper &gfxCoreHelper);
size_t getLocalIdsSizeForGroup(const Vec3<uint16_t> &group, const GfxCoreHelper &gfxCoreHelper) const;
void setLocalIdsForGroup(const Vec3<uint16_t> &group, void *destination, const RootDeviceEnvironment &rootDeviceEnvironment);
size_t getLocalIdsSizeForGroup(const Vec3<uint16_t> &group, const RootDeviceEnvironment &rootDeviceEnvironment) const;
size_t getLocalIdsSizePerThread() const;
protected:
void setLocalIdsForEntry(LocalIdsCacheEntry &entry, void *destination);
void commitNewEntry(LocalIdsCacheEntry &entry, const Vec3<uint16_t> &group, const GfxCoreHelper &gfxCoreHelper);
void commitNewEntry(LocalIdsCacheEntry &entry, const Vec3<uint16_t> &group, const RootDeviceEnvironment &rootDeviceEnvironment);
std::unique_lock<std::mutex> lock();
StackVec<LocalIdsCacheEntry, 4> cache;