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

Do not enforce LWS in dec order when smaller than half simd 

When generating work group sizes first try with enforcing decremental
order X >= Y >= Z if generated work group size X * Y * Z is smaller
than half the kernel's SIMD size then generate again without
enforcing decremental order.

Signed-off-by: Krystian Chmielewski <krystian.chmielewski@intel.com>
This commit is contained in:
Krystian Chmielewski 2023-01-03 15:28:59 +00:00 committed by Compute-Runtime-Automation
parent 48247cc42f
commit 3731ee89d8
3 changed files with 117 additions and 89 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
opencl/test/unit_test/command_queue/local_work_size_tests.cpp
View File

@ -947,3 +947,31 @@ TEST_F(LocalWorkSizeTest, givenMaxWorkgroupSizeEqualToSimdSizeWhenLwsIsCalculate
EXPECT_EQ(workGroupSize[1], 1u);
EXPECT_EQ(workGroupSize[2], 1u);
}
TEST_F(LocalWorkSizeTest, givenGwsWithSmallXAndBigYWhenLwsIsCalculatedThenDescendingOrderIsNotEnforced) {
WorkSizeInfo wsInfo(256u, true, 32u, 0u, rootDeviceEnvironment, 0u, 0u, false, false, false);
uint32_t workDim = 3;
size_t workGroup[3] = {2, 1024, 1};
size_t workGroupSize[3];
NEO::choosePrefferedWorkgroupSize(wsInfo, workGroupSize, workGroup, workDim);
EXPECT_EQ(workGroupSize[0], 2u);
EXPECT_EQ(workGroupSize[1], 128u);
EXPECT_EQ(workGroupSize[2], 1u);
// Enforce strict ratio requirement
wsInfo.yTiledSurfaces = true;
NEO::choosePrefferedWorkgroupSize(wsInfo, workGroupSize, workGroup, workDim);
EXPECT_EQ(workGroupSize[0], 2u);
EXPECT_EQ(workGroupSize[1], 128u);
EXPECT_EQ(workGroupSize[2], 1u);
// Enforce ratio requirement
wsInfo.yTiledSurfaces = false;
wsInfo.slmTotalSize = 128U;
NEO::choosePrefferedWorkgroupSize(wsInfo, workGroupSize, workGroup, workDim);
EXPECT_EQ(workGroupSize[0], 2u);
EXPECT_EQ(workGroupSize[1], 128u);
EXPECT_EQ(workGroupSize[2], 1u);
}

174
shared/source/helpers/local_work_size.cpp
View File

@ -1,5 +1,5 @@
/*
* Copyright (C) 2018-2022 Intel Corporation
* Copyright (C) 2018-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@ -124,18 +124,16 @@ void computePowerOfTwoLWS(const size_t workItems[3], WorkSizeInfo &workGroupInfo
}
}
void choosePreferredWorkGroupSizeWithRatio(uint32_t xyzFactors[3][1024], uint32_t xyzFactorsLen[3], size_t workGroupSize[3], const size_t workItems[3], WorkSizeInfo &wsInfo) {
float ratioDiff = 0;
float localRatio = float(0xffffffff);
uint64_t localWkgs = 0xffffffff;
uint64_t workGroups;
void choosePreferredWorkGroupSizeWithRatio(uint32_t xyzFactors[3][1024], uint32_t xyzFactorsLen[3], size_t workGroupSize[3], const size_t workItems[3], WorkSizeInfo &wsInfo, bool enforceDescendingOrder) {
float localRatio = std::numeric_limits<float>::max();
uint64_t localNumWorkgroups = std::numeric_limits<uint64_t>::max();
for (uint32_t xFactorsIdx = 0; xFactorsIdx < xyzFactorsLen[0]; ++xFactorsIdx) {
for (uint32_t yFactorsIdx = 0; yFactorsIdx < xyzFactorsLen[1]; ++yFactorsIdx) {
uint32_t xdim = xyzFactors[0][xyzFactorsLen[0] - 1 - xFactorsIdx];
uint32_t ydim = xyzFactors[1][yFactorsIdx];
if (ydim > xdim) {
if (enforceDescendingOrder && ydim > xdim) {
break;
}
@ -146,65 +144,57 @@ void choosePreferredWorkGroupSizeWithRatio(uint32_t xyzFactors[3][1024], uint32_
continue;
}
workGroups = Math::divideAndRoundUp(workItems[0], xdim);
workGroups *= Math::divideAndRoundUp(workItems[1], ydim);
uint64_t numWorkGroups = Math::divideAndRoundUp(workItems[0], xdim);
numWorkGroups *= Math::divideAndRoundUp(workItems[1], ydim);
ratioDiff = log((float)xdim) - log((float)ydim);
float ratioDiff = log(static_cast<float>(xdim)) - log(static_cast<float>(ydim));
ratioDiff = fabs(wsInfo.targetRatio - ratioDiff);
if (wsInfo.useStrictRatio == true) {
if (ratioDiff < localRatio) {
workGroupSize[0] = xdim;
workGroupSize[1] = ydim;
localRatio = ratioDiff;
localWkgs = workGroups;
}
} else {
if ((workGroups < localWkgs) ||
((workGroups == localWkgs) && (ratioDiff < localRatio))) {
workGroupSize[0] = xdim;
workGroupSize[1] = ydim;
localRatio = ratioDiff;
localWkgs = workGroups;
}
bool setWorkGroupSize = wsInfo.useStrictRatio
? (ratioDiff < localRatio)
: (numWorkGroups < localNumWorkgroups) || ((numWorkGroups == localNumWorkgroups) && (ratioDiff < localRatio));
if (setWorkGroupSize) {
workGroupSize[0] = xdim;
workGroupSize[1] = ydim;
localRatio = ratioDiff;
localNumWorkgroups = numWorkGroups;
}
}
}
}
void choosePreferredWorkGroupSizeWithOutRatio(uint32_t xyzFactors[3][1024], uint32_t xyzFactorsLen[3], size_t workGroupSize[3], const size_t workItems[3], WorkSizeInfo &wsInfo, uint32_t workdim) {
uint64_t localEuThrdsDispatched = 0xffffffffffffffff;
uint64_t workGroups;
void choosePreferredWorkGroupSizeWithOutRatio(uint32_t xyzFactors[3][1024], uint32_t xyzFactorsLen[3], size_t workGroupSize[3], const size_t workItems[3], WorkSizeInfo &wsInfo, bool enforceDescendingOrder) {
uint64_t localEuThrdsDispatched = std::numeric_limits<uint64_t>::max();
for (uint32_t xFactorsIdx = 0; xFactorsIdx < xyzFactorsLen[0]; ++xFactorsIdx) {
for (uint32_t zFactorsIdx = 0; zFactorsIdx < xyzFactorsLen[2]; ++zFactorsIdx) {
for (uint32_t yFactorsIdx = 0; yFactorsIdx < xyzFactorsLen[1]; ++yFactorsIdx) {
for (uint32_t yFactorsIdx = 0; yFactorsIdx < xyzFactorsLen[1]; ++yFactorsIdx) {
for (uint32_t zFactorsIdx = 0; zFactorsIdx < xyzFactorsLen[2]; ++zFactorsIdx) {
uint32_t xdim = xyzFactors[0][xyzFactorsLen[0] - 1 - xFactorsIdx];
uint32_t ydim = xyzFactors[1][yFactorsIdx];
uint32_t zdim = xyzFactors[2][zFactorsIdx];
uint32_t ydim = xyzFactors[1][xyzFactorsLen[1] - 1 - yFactorsIdx];
uint32_t zdim = xyzFactors[2][xyzFactorsLen[2] - 1 - zFactorsIdx];
if (zdim > ydim) {
if (enforceDescendingOrder) {
if (ydim > xdim) {
break;
} else if (zdim > ydim) {
continue;
}
}
uint32_t numItemsInWorkGroup = xdim * ydim * zdim;
if (numItemsInWorkGroup > wsInfo.maxWorkGroupSize) {
continue;
}
if (ydim > xdim) {
if (numItemsInWorkGroup < wsInfo.minWorkGroupSize) {
break;
}
if ((xdim * ydim * zdim) > wsInfo.maxWorkGroupSize) {
break;
}
if ((xdim * ydim * zdim) < wsInfo.minWorkGroupSize) {
continue;
}
workGroups = Math::divideAndRoundUp(workItems[0], xdim);
workGroups *= Math::divideAndRoundUp(workItems[1], ydim);
workGroups *= Math::divideAndRoundUp(workItems[2], zdim);
uint64_t euThrdsDispatched;
euThrdsDispatched = Math::divideAndRoundUp(xdim * ydim * zdim, wsInfo.simdSize);
euThrdsDispatched *= workGroups;
uint64_t numWorkGroups = Math::divideAndRoundUp(workItems[0], xdim);
numWorkGroups *= Math::divideAndRoundUp(workItems[1], ydim);
numWorkGroups *= Math::divideAndRoundUp(workItems[2], zdim);
uint64_t numThreadsPerWorkGroup = Math::divideAndRoundUp(numItemsInWorkGroup, wsInfo.simdSize);
uint64_t euThrdsDispatched = numThreadsPerWorkGroup * numWorkGroups;
if (euThrdsDispatched < localEuThrdsDispatched) {
localEuThrdsDispatched = euThrdsDispatched;
workGroupSize[0] = xdim;
@ -216,10 +206,7 @@ void choosePreferredWorkGroupSizeWithOutRatio(uint32_t xyzFactors[3][1024], uint
}
}
void computeWorkgroupSize1D(uint32_t maxWorkGroupSize,
size_t workGroupSize[3],
const size_t workItems[3],
size_t simdSize) {
void computeWorkgroupSize1D(uint32_t maxWorkGroupSize, size_t workGroupSize[3], const size_t workItems[3], size_t simdSize) {
auto items = workItems[0];
// Determine the LSB set to quickly handle factors of 2
@ -240,6 +227,42 @@ void computeWorkgroupSize1D(uint32_t maxWorkGroupSize,
workGroupSize[2] = 1;
}
void choosePreferredWorkgroupSize(uint32_t xyzFactors[3][1024], uint32_t xyzFactorsLen[3], size_t workGroupSize[3], const size_t workItems[3], WorkSizeInfo &wsInfo, bool enforceDescendingOrder) {
// check if algorithm should use ratio
wsInfo.checkRatio(workItems);
if (wsInfo.useRatio) {
choosePreferredWorkGroupSizeWithRatio(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo, enforceDescendingOrder);
if (wsInfo.useStrictRatio && workGroupSize[0] * workGroupSize[1] * 2 <= wsInfo.simdSize) {
wsInfo.useStrictRatio = false;
choosePreferredWorkGroupSizeWithRatio(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo, enforceDescendingOrder);
}
} else {
choosePreferredWorkGroupSizeWithOutRatio(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo, enforceDescendingOrder);
}
}
void choosePrefferedWorkgroupSize(WorkSizeInfo &wsInfo, size_t workGroupSize[3], const size_t workItems[3], const uint32_t workDim) {
// find all divisors for all dimensions
uint32_t xyzFactors[3][1024];
uint32_t xyzFactorsLen[3] = {};
for (int i = 0; i < 3; i++)
xyzFactors[i][xyzFactorsLen[i]++] = 1;
for (auto i = 0u; i < workDim; i++) {
for (auto j = 2u; j < wsInfo.maxWorkGroupSize; ++j) {
if ((workItems[i] % j) == 0) {
xyzFactors[i][xyzFactorsLen[i]++] = j;
}
}
}
choosePreferredWorkgroupSize(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo, true);
size_t wgs = workGroupSize[0] * workGroupSize[1] * workGroupSize[2];
if (wgs * 2 <= wsInfo.simdSize) {
choosePreferredWorkgroupSize(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo, false);
}
}
void computeWorkgroupSize2D(uint32_t maxWorkGroupSize, size_t workGroupSize[3], const size_t workItems[3], size_t simdSize) {
uint32_t xFactors[1024];
uint32_t yFactors[1024];
@ -336,15 +359,12 @@ void computeWorkgroupSizeND(WorkSizeInfo &wsInfo, size_t workGroupSize[3], const
for (int i = 0; i < 3; i++)
workGroupSize[i] = 1;
uint64_t totalNuberOfItems = workItems[0] * workItems[1] * workItems[2];
UNRECOVERABLE_IF(wsInfo.simdSize == 0);
// Find biggest power of two which devide each dimension size
if (wsInfo.slmTotalSize == 0 && !wsInfo.hasBarriers) {
if (DebugManager.flags.EnableComputeWorkSizeSquared.get() && workDim == 2 && !wsInfo.imgUsed) {
computeWorkgroupSizeSquared(wsInfo.maxWorkGroupSize, workGroupSize, workItems, wsInfo.simdSize, workDim);
return;
return computeWorkgroupSizeSquared(wsInfo.maxWorkGroupSize, workGroupSize, workItems, wsInfo.simdSize, workDim);
}
size_t itemsPowerOfTwoDivisors[3] = {1, 1, 1};
@ -363,8 +383,7 @@ void computeWorkgroupSizeND(WorkSizeInfo &wsInfo, size_t workGroupSize[3], const
// bigger than maxWorkGroupSize or this group would create more than optimal hardware threads then downsize it
uint64_t allItems = itemsPowerOfTwoDivisors[0] * itemsPowerOfTwoDivisors[1] * itemsPowerOfTwoDivisors[2];
if (allItems > wsInfo.simdSize && (allItems > wsInfo.maxWorkGroupSize || allItems > wsInfo.simdSize * optimalHardwareThreadCountGeneric[0])) {
computePowerOfTwoLWS(itemsPowerOfTwoDivisors, wsInfo, workGroupSize, workDim, canUseNx4);
return;
return computePowerOfTwoLWS(itemsPowerOfTwoDivisors, wsInfo, workGroupSize, workDim, canUseNx4);
}
// If coputed workgroup is at this point in correct size
else if (allItems >= wsInfo.simdSize) {
@ -374,41 +393,20 @@ void computeWorkgroupSizeND(WorkSizeInfo &wsInfo, size_t workGroupSize[3], const
return;
}
}
uint64_t totalNuberOfItems = workItems[0] * workItems[1] * workItems[2];
// If dimensions are not powers of two but total number of items is less than max work group size
if (totalNuberOfItems <= wsInfo.maxWorkGroupSize) {
for (auto i = 0u; i < workDim; i++)
workGroupSize[i] = workItems[i];
return;
} else {
if (workDim == 1)
computeWorkgroupSize1D(wsInfo.maxWorkGroupSize, workGroupSize, workItems, wsInfo.simdSize);
else {
uint32_t xyzFactors[3][1024];
uint32_t xyzFactorsLen[3] = {};
// check if algorithm should use ratio
wsInfo.checkRatio(workItems);
// find all divisors for all dimensions
for (int i = 0; i < 3; i++)
xyzFactors[i][xyzFactorsLen[i]++] = 1;
for (auto i = 0u; i < workDim; i++) {
for (auto j = 2u; j < wsInfo.maxWorkGroupSize; ++j) {
if ((workItems[i] % j) == 0) {
xyzFactors[i][xyzFactorsLen[i]++] = j;
}
}
}
if (wsInfo.useRatio) {
choosePreferredWorkGroupSizeWithRatio(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo);
if (wsInfo.useStrictRatio && workGroupSize[0] * workGroupSize[1] * 2 <= wsInfo.simdSize) {
wsInfo.useStrictRatio = false;
choosePreferredWorkGroupSizeWithRatio(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo);
}
} else
choosePreferredWorkGroupSizeWithOutRatio(xyzFactors, xyzFactorsLen, workGroupSize, workItems, wsInfo, workDim);
}
}
if (workDim == 1) {
return computeWorkgroupSize1D(wsInfo.maxWorkGroupSize, workGroupSize, workItems, wsInfo.simdSize);
}
choosePrefferedWorkgroupSize(wsInfo, workGroupSize, workItems, workDim);
}
Vec3<size_t> computeWorkgroupsNumber(const Vec3<size_t> &gws, const Vec3<size_t> &lws) {

4
shared/source/helpers/local_work_size.h
View File

@ -1,5 +1,5 @@
/*
* Copyright (C) 2018-2022 Intel Corporation
* Copyright (C) 2018-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@ -39,6 +39,8 @@ void computeWorkgroupSizeSquared(
size_t simdSize,
const uint32_t workDim);
void choosePrefferedWorkgroupSize(WorkSizeInfo &wsInfo, size_t workGroupSize[3], const size_t workItems[3], const uint32_t workDim);
Vec3<size_t> computeWorkgroupsNumber(
const Vec3<size_t> &gws,
const Vec3<size_t> &lws);