intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-29 09:03:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

Move grf size to HwInfo

Browse Source 
Change-Id: I65ee879644573586d63092b487f8b5ea0cedf1e3
Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>
This commit is contained in:
Maciej Plewka
2019-12-17 08:55:09 +01:00
committed by sys_ocldev
parent 2b0db66c52
commit 8803b4cd4e
31 changed files with 181 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
unit_tests/command_queue/local_id_tests.cpp
View File

@@ -40,37 +40,42 @@ TEST(LocalID, ThreadsPerWorkgroup) {
TEST(LocalID, PerThreadSizeLocalIDs_SIMD8) {
uint32_t simd = 8;
uint32_t grfSize = 32;
// 3 channels (x,y,z) * 1 GRFs per thread (@SIMD8)
EXPECT_EQ(3 * sizeof(GRF), getPerThreadSizeLocalIDs(simd));
EXPECT_EQ(3 * grfSize, getPerThreadSizeLocalIDs(simd, grfSize));
}
TEST(LocalID, PerThreadSizeLocalIDs_SIMD16) {
uint32_t simd = 16;
uint32_t grfSize = 32;
// 3 channels (x,y,z) * 1 GRFs per thread (@SIMD16)
EXPECT_EQ(3 * sizeof(GRF), getPerThreadSizeLocalIDs(simd));
EXPECT_EQ(3 * grfSize, getPerThreadSizeLocalIDs(simd, grfSize));
}
TEST(LocalID, PerThreadSizeLocalIDs_SIMD32) {
uint32_t simd = 32;
uint32_t grfSize = 32;
// 3 channels (x,y,z) * 2 GRFs per thread (@SIMD32)
EXPECT_EQ(6 * sizeof(GRF), getPerThreadSizeLocalIDs(simd));
EXPECT_EQ(6 * grfSize, getPerThreadSizeLocalIDs(simd, grfSize));
}
TEST(LocalID, PerThreadSizeLocalIDs_SIMD1) {
uint32_t simd = 1;
uint32_t grfSize = 32;
EXPECT_EQ(sizeof(GRF), getPerThreadSizeLocalIDs(simd));
EXPECT_EQ(grfSize, getPerThreadSizeLocalIDs(simd, grfSize));
}
struct LocalIDFixture : public ::testing::TestWithParam<std::tuple<int, int, int, int>> {
struct LocalIDFixture : public ::testing::TestWithParam<std::tuple<int, int, int, int, int>> {
void SetUp() override {
simd = std::get<0>(GetParam());
localWorkSizeX = std::get<1>(GetParam());
localWorkSizeY = std::get<2>(GetParam());
localWorkSizeZ = std::get<3>(GetParam());
grfSize = std::get<1>(GetParam());
localWorkSizeX = std::get<2>(GetParam());
localWorkSizeY = std::get<3>(GetParam());
localWorkSizeZ = std::get<4>(GetParam());
localWorkSize = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
if (localWorkSize > 256) {
@@ -232,6 +237,7 @@ struct LocalIDFixture : public ::testing::TestWithParam<std::tuple<int, int, int
uint32_t localWorkSizeZ;
uint32_t localWorkSize;
uint32_t simd;
uint32_t grfSize;
// Provide support for a max LWS of 256
// 32 threads @ SIMD8
@@ -242,20 +248,20 @@ struct LocalIDFixture : public ::testing::TestWithParam<std::tuple<int, int, int
TEST_P(LocalIDFixture, checkIDWithinLimits) {
generateLocalIDs(buffer, simd, std::array<uint16_t, 3>{{static_cast<uint16_t>(localWorkSizeX), static_cast<uint16_t>(localWorkSizeY), static_cast<uint16_t>(localWorkSizeZ)}},
std::array<uint8_t, 3>{{0, 1, 2}}, false);
std::array<uint8_t, 3>{{0, 1, 2}}, false, grfSize);
validateIDWithinLimits(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ);
}
TEST_P(LocalIDFixture, checkAllWorkItemsCovered) {
generateLocalIDs(buffer, simd, std::array<uint16_t, 3>{{static_cast<uint16_t>(localWorkSizeX), static_cast<uint16_t>(localWorkSizeY), static_cast<uint16_t>(localWorkSizeZ)}},
std::array<uint8_t, 3>{{0, 1, 2}}, false);
std::array<uint8_t, 3>{{0, 1, 2}}, false, grfSize);
validateAllWorkItemsCovered(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ);
}
TEST_P(LocalIDFixture, WhenWalkOrderIsXyzThenProperLocalIdsAreGenerated) {
auto dimensionsOrder = std::array<uint8_t, 3>{{0, 1, 2}};
generateLocalIDs(buffer, simd, std::array<uint16_t, 3>{{static_cast<uint16_t>(localWorkSizeX), static_cast<uint16_t>(localWorkSizeY), static_cast<uint16_t>(localWorkSizeZ)}},
dimensionsOrder, false);
dimensionsOrder, false, grfSize);
validateAllWorkItemsCovered(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ);
validateWalkOrder(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ, dimensionsOrder);
}
@@ -263,7 +269,7 @@ TEST_P(LocalIDFixture, WhenWalkOrderIsXyzThenProperLocalIdsAreGenerated) {
TEST_P(LocalIDFixture, WhenWalkOrderIsYxzThenProperLocalIdsAreGenerated) {
auto dimensionsOrder = std::array<uint8_t, 3>{{1, 0, 2}};
generateLocalIDs(buffer, simd, std::array<uint16_t, 3>{{static_cast<uint16_t>(localWorkSizeX), static_cast<uint16_t>(localWorkSizeY), static_cast<uint16_t>(localWorkSizeZ)}},
dimensionsOrder, false);
dimensionsOrder, false, grfSize);
validateAllWorkItemsCovered(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ);
validateWalkOrder(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ, dimensionsOrder);
}
@@ -271,30 +277,30 @@ TEST_P(LocalIDFixture, WhenWalkOrderIsYxzThenProperLocalIdsAreGenerated) {
TEST_P(LocalIDFixture, WhenWalkOrderIsZyxThenProperLocalIdsAreGenerated) {
auto dimensionsOrder = std::array<uint8_t, 3>{{2, 1, 0}};
generateLocalIDs(buffer, simd, std::array<uint16_t, 3>{{static_cast<uint16_t>(localWorkSizeX), static_cast<uint16_t>(localWorkSizeY), static_cast<uint16_t>(localWorkSizeZ)}},
dimensionsOrder, false);
dimensionsOrder, false, grfSize);
validateAllWorkItemsCovered(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ);
validateWalkOrder(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ, dimensionsOrder);
}
TEST_P(LocalIDFixture, sizeCalculationLocalIDs) {
auto workItems = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
auto sizeTotalPerThreadData = getThreadsPerWG(simd, workItems) * getPerThreadSizeLocalIDs(simd);
auto sizeTotalPerThreadData = getThreadsPerWG(simd, workItems) * getPerThreadSizeLocalIDs(simd, grfSize);
// Should be multiple of GRFs
auto sizeGRF = sizeof(GRF);
EXPECT_EQ(0u, sizeTotalPerThreadData % sizeGRF);
EXPECT_EQ(0u, sizeTotalPerThreadData % grfSize);
auto numGRFsPerThread = (simd == 32) ? 2 : 1;
auto numThreadsExpected = Math::divideAndRoundUp(workItems, simd);
auto numGRFsExpected = 3 * numGRFsPerThread * numThreadsExpected;
EXPECT_EQ(numGRFsExpected * sizeGRF, sizeTotalPerThreadData);
EXPECT_EQ(numGRFsExpected * grfSize, sizeTotalPerThreadData);
}
struct LocalIdsLayoutForImagesTest : ::testing::TestWithParam<std::tuple<uint16_t, uint16_t, uint16_t>> {
struct LocalIdsLayoutForImagesTest : ::testing::TestWithParam<std::tuple<uint16_t, uint16_t, uint16_t, uint16_t>> {
void SetUp() override {
simd = std::get<0>(GetParam());
localWorkSize = {{std::get<1>(GetParam()),
std::get<2>(GetParam()),
grfSize = std::get<1>(GetParam());
localWorkSize = {{std::get<2>(GetParam()),
std::get<3>(GetParam()),
1u}};
rowWidth = simd == 32u ? 32u : 16u;
xDelta = simd == 8u ? 2u : 4u;
@@ -311,7 +317,7 @@ struct LocalIdsLayoutForImagesTest : ::testing::TestWithParam<std::tuple<uint16_
memset(memory.get(), 0xff, size);
buffer = reinterpret_cast<uint16_t *>(memory.get());
EXPECT_TRUE(isCompatibleWithLayoutForImages(localWorkSize, dimensionsOrder, simd));
generateLocalIDs(buffer, simd, localWorkSize, dimensionsOrder, true);
generateLocalIDs(buffer, simd, localWorkSize, dimensionsOrder, true, grfSize);
}
void validateGRF() {
uint32_t totalLocalIds = localWorkSize.at(0) * localWorkSize.at(1);
@@ -356,6 +362,7 @@ struct LocalIdsLayoutForImagesTest : ::testing::TestWithParam<std::tuple<uint16_
}
}
uint16_t simd;
uint16_t grfSize;
uint8_t rowWidth;
uint16_t xDelta;
std::array<uint16_t, 3> localWorkSize;
@@ -397,6 +404,7 @@ TEST_P(LocalIdsLayoutTest, givenLocalWorkgroupSize4x4x1WhenGenerateLocalIdsThenH
std::array<uint16_t, 3> localWorkSize{{xDelta, 4u, 1u}};
uint16_t totalLocalWorkSize = 4u * xDelta;
auto dimensionsOrder = std::array<uint8_t, 3>{{0u, 1u, 2u}};
uint32_t grfSize = 32;
auto elemsInBuffer = rowWidth * 3u;
auto size = elemsInBuffer * sizeof(uint16_t);
@@ -409,8 +417,8 @@ TEST_P(LocalIdsLayoutTest, givenLocalWorkgroupSize4x4x1WhenGenerateLocalIdsThenH
auto buffer2 = reinterpret_cast<uint16_t *>(alignedMemory2.get());
memset(buffer2, 0xff, size);
generateLocalIDs(buffer1, simd, localWorkSize, dimensionsOrder, false);
generateLocalIDs(buffer2, simd, localWorkSize, dimensionsOrder, true);
generateLocalIDs(buffer1, simd, localWorkSize, dimensionsOrder, false, grfSize);
generateLocalIDs(buffer2, simd, localWorkSize, dimensionsOrder, true, grfSize);
for (auto i = 0u; i < elemsInBuffer / rowWidth; i++) {
for (auto j = 0u; j < rowWidth; j++) {
@@ -439,15 +447,18 @@ TEST_P(LocalIdsLayoutForImagesTest, givenLocalWorkgroupSizeCompatibleWithLayoutF
#define LWSZParams ::testing::Values(1)
#endif
INSTANTIATE_TEST_CASE_P(AllCombinations, LocalIDFixture, ::testing::Combine(SIMDParams, LWSXParams, LWSYParams, LWSZParams));
#define GRFSizeParams ::testing::Values(32)
INSTANTIATE_TEST_CASE_P(AllCombinations, LocalIDFixture, ::testing::Combine(SIMDParams, GRFSizeParams, LWSXParams, LWSYParams, LWSZParams));
INSTANTIATE_TEST_CASE_P(LayoutTests, LocalIdsLayoutTest, SIMDParams);
INSTANTIATE_TEST_CASE_P(LayoutForImagesTests, LocalIdsLayoutForImagesTest, ::testing::Combine(SIMDParams, ::testing::Values(4, 8, 12, 20), ::testing::Values(4, 8, 12, 20)));
INSTANTIATE_TEST_CASE_P(LayoutForImagesTests, LocalIdsLayoutForImagesTest, ::testing::Combine(SIMDParams, GRFSizeParams, ::testing::Values(4, 8, 12, 20), ::testing::Values(4, 8, 12, 20)));
// To debug a specific configuration replace the list of Values with specific values.
// NOTE: You'll need a unique test prefix
INSTANTIATE_TEST_CASE_P(SingleTest, LocalIDFixture,
::testing::Combine(
::testing::Values(32), //SIMD
::testing::Values(32), //GRF
::testing::Values(5), //LWSX
::testing::Values(6), //LWSY
::testing::Values(7))); //LWSZ