Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-10 12:53:42 +08:00
Code Issues Packages Projects Releases Wiki Activity

Utilize shortened version of optimal HW thread count in nx4 scenarios.

Browse Source 
- also clean early return if simd size = 0

Change-Id: I9b01df091ab6dd6a3066d1a8762c7fb1530c2804
This commit is contained in:
Mrozek, Michal
2018-01-16 14:27:08 +01:00
parent 92ebb0c3c1
commit dd601ff73a
2 changed files with 84 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

127
runtime/command_queue/local_work_size.cpp
View File

@ -99,7 +99,7 @@ inline uint32_t factor<0>(size_t workItems, uint32_t workSize, uint32_t maxWorkG
}
void computePowerOfTwoLWS(const size_t workItems[3], size_t simdSize, uint32_t maxWorkGroupSize, size_t workGroupSize[3], const uint32_t workDim, bool canUseNx4) {
uint32_t targetIndex = 0;
uint32_t targetIndex = canUseNx4 ? 2 : 0;
while (optimalHardwareThreadCountGeneric[targetIndex] > 1 && maxWorkGroupSize < optimalHardwareThreadCountGeneric[targetIndex] * simdSize)
targetIndex++;
@ -330,76 +330,75 @@ void computeWorkgroupSizeND(WorkSizeInfo wsInfo, size_t workGroupSize[3], const
workGroupSize[i] = 1;
uint64_t totalNuberOfItems = workItems[0] * workItems[1] * workItems[2];
if (wsInfo.simdSize == 0) {
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) {
return computeWorkgroupSizeSquared(wsInfo.maxWorkGroupSize, workGroupSize, workItems, wsInfo.simdSize, workDim);
}
size_t itemsPowerOfTwoDivisors[3] = {1, 1, 1};
for (auto i = 0u; i < workDim; i++) {
uint32_t requiredWorkItemsCount = uint32_t(wsInfo.simdSize * optimalHardwareThreadCountGeneric[0]);
while (requiredWorkItemsCount > 1 && !(Math::isDivisableByPowerOfTwoDivisor(uint32_t(workItems[i]), requiredWorkItemsCount)))
requiredWorkItemsCount = requiredWorkItemsCount >> 1;
itemsPowerOfTwoDivisors[i] = requiredWorkItemsCount;
}
bool canUseNx4 = (wsInfo.imgUsed &&
(itemsPowerOfTwoDivisors[0] >= 4 || (itemsPowerOfTwoDivisors[0] >= 2 && wsInfo.simdSize == 8)) &&
itemsPowerOfTwoDivisors[1] >= 4);
//If computed dimension sizes which are powers of two are creating group which is
//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.simdSize, wsInfo.maxWorkGroupSize, workGroupSize, workDim, canUseNx4);
return;
}
//If coputed workgroup is at this point in correct size
else if (allItems >= wsInfo.simdSize) {
itemsPowerOfTwoDivisors[1] = canUseNx4 ? 4 : itemsPowerOfTwoDivisors[1];
for (auto i = 0u; i < workDim; i++)
workGroupSize[i] = itemsPowerOfTwoDivisors[i];
return;
}
}
//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 {
//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) {
return computeWorkgroupSizeSquared(wsInfo.maxWorkGroupSize, workGroupSize, workItems, wsInfo.simdSize, workDim);
}
if (workDim == 1)
computeWorkgroupSize1D(wsInfo.maxWorkGroupSize, workGroupSize, workItems, wsInfo.simdSize);
else {
uint32_t xyzFactors[3][1024];
uint32_t xyzFactorsLen[3] = {};
size_t itemsPowerOfTwoDivisors[3] = {1, 1, 1};
//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++) {
uint32_t requiredWorkItemsCount = uint32_t(wsInfo.simdSize * optimalHardwareThreadCountGeneric[0]);
while (requiredWorkItemsCount > 1 && !(Math::isDivisableByPowerOfTwoDivisor(uint32_t(workItems[i]), requiredWorkItemsCount)))
requiredWorkItemsCount = requiredWorkItemsCount >> 1;
itemsPowerOfTwoDivisors[i] = requiredWorkItemsCount;
}
bool canUseNx4 = (wsInfo.imgUsed &&
(itemsPowerOfTwoDivisors[0] >= 4 || (itemsPowerOfTwoDivisors[0] >= 2 && wsInfo.simdSize == 8)) &&
itemsPowerOfTwoDivisors[1] >= 4);
//If computed dimension sizes which are powers of two are creating group which is
//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.simdSize, wsInfo.maxWorkGroupSize, workGroupSize, workDim, canUseNx4);
return;
}
//If coputed workgroup is at this point in correct size
else if (allItems >= wsInfo.simdSize) {
itemsPowerOfTwoDivisors[1] = canUseNx4 ? 4 : itemsPowerOfTwoDivisors[1];
for (auto i = 0u; i < workDim; i++)
workGroupSize[i] = itemsPowerOfTwoDivisors[i];
return;
}
}
//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;
}
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 (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);
}
}
}

28
unit_tests/command_queue/local_work_size_tests.cpp
View File

@ -58,13 +58,6 @@ TEST(localWorkSizeTest, given1DimWorkGroupAndSimdEqual8WhenComputeCalledThenLoca
EXPECT_EQ(workGroupSize[0], 9u);
EXPECT_EQ(workGroupSize[1], 1u);
EXPECT_EQ(workGroupSize[2], 1u);
workGroup[0] = 128;
wsInfo.simdSize = 0;
OCLRT::computeWorkgroupSizeND(wsInfo, workGroupSize, workGroup, workDim);
EXPECT_EQ(workGroupSize[0], 1u);
EXPECT_EQ(workGroupSize[1], 1u);
EXPECT_EQ(workGroupSize[2], 1u);
}
TEST(localWorkSizeTest, given1DimWorkGroupAndSimdEqual32WhenComputeCalledThenLocalGroupComputed) {
@ -317,6 +310,27 @@ TEST(localWorkSizeTest, givenKernelWithTileYImagesAndNoBarriersWhenWorkgroupSize
EXPECT_EQ(workGroupSize[2], 1u);
}
TEST(localWorkSizeTest, givenSimd16KernelWithTileYImagesAndNoBarriersWhenWorkgroupSizeIsComputedThenItMimicsTilingPattern) {
WorkSizeInfo wsInfo(256, false, 16, 0u, platformDevices[0]->pPlatform->eRenderCoreFamily, 32u, 0u, true, true);
uint32_t workDim = 2;
size_t workGroup[3] = {1, 1, 1};
size_t workGroupSize[3];
workGroup[0] = 2048;
workGroup[1] = 2048;
OCLRT::computeWorkgroupSizeND(wsInfo, workGroupSize, workGroup, workDim);
EXPECT_EQ(workGroupSize[0], 32u);
EXPECT_EQ(workGroupSize[1], 4u);
EXPECT_EQ(workGroupSize[2], 1u);
workGroup[0] = 1920;
workGroup[1] = 1080;
OCLRT::computeWorkgroupSizeND(wsInfo, workGroupSize, workGroup, workDim);
EXPECT_EQ(workGroupSize[0], 32u);
EXPECT_EQ(workGroupSize[1], 4u);
EXPECT_EQ(workGroupSize[2], 1u);
}
TEST(localWorkSizeTest, givenKernelWithTwoDimensionalGlobalSizesWhenLwsIsComputedThenItHasMaxWorkgroupSize) {
WorkSizeInfo wsInfo(256, 0u, 32, 0u, platformDevices[0]->pPlatform->eRenderCoreFamily, 32u, 0u, false, false);
uint32_t workDim = 2;