intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-30 01:35:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add choose max row size parameter for local id generation.

Browse Source 
Change-Id: I77185b6c114092859c742236a4dfef01deb9ea21
This commit is contained in:
Michal Mrozek
2020-02-04 17:58:41 +01:00
committed by sys_ocldev
parent 66754c4849
commit 6cc9b9d125
9 changed files with 83 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
runtime/command_queue/local_id_gen.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2019 Intel Corporation
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -26,9 +26,9 @@ const uint16_t initialLocalID[] = {
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31};
// Lookup table for generating LocalIDs based on the SIMD of the kernel
void (*LocalIDHelper::generateSimd8)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder) = generateLocalIDsSimd<uint16x8_t, 8>;
void (*LocalIDHelper::generateSimd16)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder) = generateLocalIDsSimd<uint16x8_t, 16>;
void (*LocalIDHelper::generateSimd32)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder) = generateLocalIDsSimd<uint16x8_t, 32>;
void (*LocalIDHelper::generateSimd8)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize) = generateLocalIDsSimd<uint16x8_t, 8>;
void (*LocalIDHelper::generateSimd16)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize) = generateLocalIDsSimd<uint16x8_t, 16>;
void (*LocalIDHelper::generateSimd32)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize) = generateLocalIDsSimd<uint16x8_t, 32>;
// Initialize the lookup table based on CPU capabilities
LocalIDHelper::LocalIDHelper() {
@@ -49,11 +49,11 @@ void generateLocalIDs(void *buffer, uint16_t simd, const std::array<uint16_t, 3>
if (useLayoutForImages) {
generateLocalIDsWithLayoutForImages(buffer, localWorkgroupSize, simd);
} else if (simd == 32) {
LocalIDHelper::generateSimd32(buffer, localWorkgroupSize, threadsPerWorkGroup, dimensionsOrder);
LocalIDHelper::generateSimd32(buffer, localWorkgroupSize, threadsPerWorkGroup, dimensionsOrder, grfSize != 32);
} else if (simd == 16) {
LocalIDHelper::generateSimd16(buffer, localWorkgroupSize, threadsPerWorkGroup, dimensionsOrder);
LocalIDHelper::generateSimd16(buffer, localWorkgroupSize, threadsPerWorkGroup, dimensionsOrder, grfSize != 32);
} else if (simd == 8) {
LocalIDHelper::generateSimd8(buffer, localWorkgroupSize, threadsPerWorkGroup, dimensionsOrder);
LocalIDHelper::generateSimd8(buffer, localWorkgroupSize, threadsPerWorkGroup, dimensionsOrder, grfSize != 32);
} else {
generateLocalIDsForSimdOne(buffer, localWorkgroupSize, dimensionsOrder, grfSize);
}

12
runtime/command_queue/local_id_gen.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2019 Intel Corporation
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -43,9 +43,9 @@ inline uint32_t getPerThreadSizeLocalIDs(uint32_t simd, uint32_t grfSize, uint32
}
struct LocalIDHelper {
static void (*generateSimd8)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
static void (*generateSimd16)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
static void (*generateSimd32)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
static void (*generateSimd8)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
static void (*generateSimd16)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
static void (*generateSimd32)(void *buffer, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
static LocalIDHelper initializer;
@@ -57,7 +57,7 @@ extern const uint16_t initialLocalID[];
template <typename Vec, int simd>
void generateLocalIDsSimd(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup,
const std::array<uint8_t, 3> &dimensionsOrder);
const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
void generateLocalIDs(void *buffer, uint16_t simd, const std::array<uint16_t, 3> &localWorkgroupSize,
const std::array<uint8_t, 3> &dimensionsOrder, bool isImageOnlyKernel, uint32_t grfSize);
@@ -67,4 +67,4 @@ bool isCompatibleWithLayoutForImages(const std::array<uint16_t, 3> &localWorkgro
void generateLocalIDsForSimdOne(void *b, const std::array<uint16_t, 3> &localWorkgroupSize,
const std::array<uint8_t, 3> &dimensionsOrder, uint32_t grfSize);
} // namespace NEO
} // namespace NEO

6
runtime/command_queue/local_id_gen.inl
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2019 Intel Corporation
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -13,7 +13,7 @@ namespace NEO {
template <typename Vec, int simd>
inline void generateLocalIDsSimd(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup,
const std::array<uint8_t, 3> &dimensionsOrder) {
const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize) {
const int passes = simd / Vec::numChannels;
int pass = 0;
@@ -27,7 +27,7 @@ inline void generateLocalIDsSimd(void *b, const std::array<uint16_t, 3> &localWo
auto zero = Vec::zero();
auto one = Vec::one();
const auto threadSkipSize = (simd == 32 ? 32 : 16) * sizeof(uint16_t);
const auto threadSkipSize = ((simd == 32 || chooseMaxRowSize) ? 32 : 16) * sizeof(uint16_t);
Vec vSimdX(simd);
Vec vSimdY = zero;
Vec vSimdZ = zero;

8
runtime/command_queue/local_id_gen_avx2.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2019 Intel Corporation
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -12,7 +12,7 @@
#include <array>
namespace NEO {
template void generateLocalIDsSimd<uint16x16_t, 32>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
template void generateLocalIDsSimd<uint16x16_t, 16>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
template void generateLocalIDsSimd<uint16x16_t, 32>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
template void generateLocalIDsSimd<uint16x16_t, 16>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
} // namespace NEO
#endif
#endif

8
runtime/command_queue/local_id_gen_sse4.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2019 Intel Corporation
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -11,7 +11,7 @@
#include <array>
namespace NEO {
template void generateLocalIDsSimd<uint16x8_t, 32>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
template void generateLocalIDsSimd<uint16x8_t, 16>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
template void generateLocalIDsSimd<uint16x8_t, 8>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder);
template void generateLocalIDsSimd<uint16x8_t, 32>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
template void generateLocalIDsSimd<uint16x8_t, 16>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
template void generateLocalIDsSimd<uint16x8_t, 8>(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup, const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize);
} // namespace NEO