Skip to content

[HLSL] Add LinAlg matrix accumulation contention coverage #8818

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Jack Elliott (JoeCitizen) merged 2 commits into from
Aug 25, 2026
+228 −31
Merged

[HLSL] Add LinAlg matrix accumulation contention coverage #8818

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9346afb
[HLSL] Add LinAlg matrix accumulation contention coverage
Aug 19, 2026
2b56bb4
[HLSL] Weight LinAlg descriptor contention by a global Wave index
Aug 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
259 changes: 228 additions & 31 deletions tools/clang/unittests/HLSLExec/LinAlgTests.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3122,6 +3122,8 @@ class DxilConf_SM610_LinAlg {
TEST_METHOD(StoreDescriptorOOB_Wave_4x8_F16_OffsetPaddedPartialView);
TEST_METHOD(SplatStore_Wave_16x16_F16);
TEST_METHOD(AccumulateDescriptor_Wave_16x16_F16);
TEST_METHOD(AccumulateDescriptorContention_Wave_4x8_I32);
TEST_METHOD(AccumulateDescriptorContention_Wave_4x8_F32_OrderInvariant);

// Load/Store/Accumulate Memory
TEST_METHOD(LoadMemory_Wave_16x16_F16);
Expand All @@ -3130,6 +3132,8 @@ class DxilConf_SM610_LinAlg {
TEST_METHOD(LoadStoreMemory_Wave_4x8_F16_RowMajorOffsetPadded);
TEST_METHOD(LoadStoreMemory_Wave_4x8_F32_ColumnMajorOffsetPadded);
TEST_METHOD(LoadStoreMemory_ThreadGroup_4x8_F16);
TEST_METHOD(AccumulateMemoryContention_Wave_4x8_F16);
TEST_METHOD(AccumulateMemoryContention_Wave_4x8_I32);

// Element access
TEST_METHOD(ElementAccess_Wave_16x16_F16);
Expand Down Expand Up @@ -3928,6 +3932,8 @@ void DxilConf_SM610_LinAlg::SplatStore_Wave_16x16_F16() {
SelectedWaveSize);
}

static constexpr UINT DescriptorAccumulatesPerWave = 2;

static const char AccumulateDescriptorShader[] = R"(
#define USE_ACC 2

Expand All @@ -3940,15 +3946,27 @@ static const char AccumulateDescriptorShader[] = R"(
[WaveSize(4, 128)]
#endif
[numthreads(NUMTHREADS, 1, 1)]
void main() {
if (GetGroupWaveIndex() != 0)
void main(uint3 GroupID : SV_GroupID) {
if (GetGroupWaveIndex() >= ACTIVE_WAVE_COUNT)
return;

__builtin_LinAlgMatrix
[[__LinAlgMatrix_Attributes(COMP_TYPE, M_DIM, N_DIM, USE_ACC, SCOPE)]]
Mat;
__builtin_LinAlg_MatrixLoadFromDescriptor(
Mat, Input, 0, STRIDE, LAYOUT, 128);

// The index is distinct across every Wave in every group, so the total
// pins which Waves landed rather than only how many additions were
// applied. A Wave index alone would repeat in each group.
uint Weight = GroupID.x * ACTIVE_WAVE_COUNT + GetGroupWaveIndex() + 1;
for (uint I = 0; I < __builtin_LinAlg_MatrixLength(Mat); ++I) {
ELEM_TYPE Elem;
__builtin_LinAlg_MatrixGetElement(Elem, Mat, I);
__builtin_LinAlg_MatrixSetElement(
Mat, Mat, I, (ELEM_TYPE)(Weight * Elem));
}

__builtin_LinAlg_MatrixAccumulateToDescriptor(
Mat, Output, 0, STRIDE, LAYOUT, 128);
__builtin_LinAlg_MatrixAccumulateToDescriptor(
Expand All @@ -3959,11 +3977,18 @@ static const char AccumulateDescriptorShader[] = R"(
static void runAccumulateDescriptor(ID3D12Device *Device,
dxc::SpecificDllLoader &DxcSupport,
const MatrixParams &Params, int FillValue,
bool Verbose, UINT ForcedWaveSize = 0) {
bool Verbose, UINT ForcedWaveSize = 0,
UINT ActiveWaveCount = 1,
UINT DispatchX = 1) {
const size_t NumElements = Params.totalElements();
const size_t BufferSize = Params.totalBytes();
const UINT ContendingWaves = ActiveWaveCount * DispatchX;
const UINT AccumulationCount = DescriptorAccumulatesPerWave * ContendingWaves;
const UINT WeightSum = ContendingWaves * (ContendingWaves + 1) / 2;
const UINT WeightedAccumulation = DescriptorAccumulatesPerWave * WeightSum;

std::stringstream ExtraDefs;
ExtraDefs << " -DACTIVE_WAVE_COUNT=" << ActiveWaveCount;
if (ForcedWaveSize != 0)
ExtraDefs << " -DFORCED_WAVE_SIZE=" << ForcedWaveSize;

Expand All @@ -3973,10 +3998,16 @@ static void runAccumulateDescriptor(ID3D12Device *Device,
Verbose);

auto Expected = makeExpectedMat(Params.CompType, Params.M, Params.N,
static_cast<float>(FillValue) * 2, false);
static_cast<float>(FillValue) *
static_cast<float>(WeightedAccumulation),
false);

hlsl_test::LogCommentFmt(
L"Descriptor accumulation issues %u atomic matrix additions in total "
L"across %u Waves in each of %u groups, weighted to %u times the fill",
AccumulationCount, ActiveWaveCount, DispatchX, WeightedAccumulation);
auto Op = createComputeOp(AccumulateDescriptorShader, "cs_6_10",
"SRV(t0), UAV(u1)", Args.c_str());
"SRV(t0), UAV(u1)", Args.c_str(), DispatchX);
addSRVBuffer(Op.get(), "Input", BufferSize, "byname");
addUAVBuffer(Op.get(), "Output", BufferSize, true);
addRootView(Op.get(), 0, "Input");
Expand Down Expand Up @@ -4025,6 +4056,50 @@ void DxilConf_SM610_LinAlg::AccumulateDescriptor_Wave_16x16_F16() {
SelectedWaveSize);
}

static void runAccumulateDescriptorContention(
ID3D12Device *Device, dxc::SpecificDllLoader &DxcSupport,
ComponentType CompType, int FillValue, LPCWSTR CaseName, bool Verbose) {
MatrixParams Params = {};
Params.CompType = CompType;
Params.M = 4;
Params.N = 8;
Params.Use = MatrixUse::Accumulator;
Params.Scope = MatrixScope::Wave;
Params.Layout = MatrixLayout::RowMajor;
Params.NumThreads = 512;
Params.Enable16Bit = false;

UINT SelectedWaveSize = 0;
if (!matrixConstructionApplicable(Device, Params, {Params.Use}, CaseName,
SelectedWaveSize))
return;
if (!accumulateStoreApplicable(
Device, Params.CompType,
linalg_test::AtomicDestination::RWByteAddressBuffer, CaseName))
return;

constexpr UINT ActiveWaveCount = 4;
constexpr UINT DispatchX = 4;
Params.NumThreads = static_cast<int>(SelectedWaveSize * ActiveWaveCount);

runAccumulateDescriptor(Device, DxcSupport, Params, FillValue, Verbose,
SelectedWaveSize, ActiveWaveCount, DispatchX);
}

void DxilConf_SM610_LinAlg::AccumulateDescriptorContention_Wave_4x8_I32() {
runAccumulateDescriptorContention(
D3DDevice, DxcSupport, ComponentType::I32, /*FillValue=*/7,
L"AccumulateDescriptorContention_Wave_4x8_I32", VerboseLogging);
}

void DxilConf_SM610_LinAlg::
AccumulateDescriptorContention_Wave_4x8_F32_OrderInvariant() {
runAccumulateDescriptorContention(
D3DDevice, DxcSupport, ComponentType::F32, /*FillValue=*/1,
L"AccumulateDescriptorContention_Wave_4x8_F32_OrderInvariant",
VerboseLogging);
}

// Element access constructs a wave-scope matrix and then reads or writes its
// components, so applicability is exactly MatrixConstruction for the tile the
// case declares. D3D12LinearAlgebraRuntimeFeatureSupport.md guarantees only
Expand Down Expand Up @@ -7536,7 +7611,7 @@ static const char GroupSharedAccumulateShader[] = R"(

GroupMemoryBarrierWithGroupSync();

if (GetGroupWaveIndex() == 0) {
if (GetGroupWaveIndex() < ACTIVE_WAVE_COUNT) {
__builtin_LinAlgMatrix
[[__LinAlgMatrix_Attributes(COMP_TYPE, M_DIM, N_DIM, USE, SCOPE)]]
Mat;
Expand All @@ -7545,7 +7620,8 @@ static const char GroupSharedAccumulateShader[] = R"(
uint2 Coord = __builtin_LinAlg_MatrixGetCoordinate(Mat, I);
__builtin_LinAlg_MatrixSetElement(
Mat, Mat, I,
(ELEM_TYPE)(ACCUMULATE_START + Coord.x * N_DIM + Coord.y));
(ELEM_TYPE)((GetGroupWaveIndex() + 1) *
(ACCUMULATE_START + Coord.x * N_DIM + Coord.y)));
}
__builtin_LinAlg_MatrixAccumulateToMemory(
Mat, GsData, COMP_TYPE, MEM_OFFSET, MEM_STRIDE, MEM_LAYOUT);
Expand All @@ -7564,16 +7640,32 @@ static void runGroupSharedAccumulate(
ID3D12Device *Device, dxc::SpecificDllLoader &DxcSupport,
const MatrixParams &Params,
const cpu_oracle::MatrixBufferLayout &MemoryLayout, uint32_t InitialValue,
uint32_t AccumulateStartingValue, bool Verbose, UINT ForcedWaveSize = 0) {
if (!Device || Params.CompType != ComponentType::F16 ||
Params.Scope != MatrixScope::Wave ||
Params.Use != MatrixUse::Accumulator ||
InitialValue >
(std::numeric_limits<uint32_t>::max)() - AccumulateStartingValue) {
VERIFY_IS_TRUE(false, "Invalid group-shared accumulate parameters");
uint32_t AccumulateStartingValue, bool Verbose, UINT ForcedWaveSize = 0,
UINT ActiveWaveCount = 1) {
if (!Device) {
hlsl_test::LogErrorFmt(L"Group-shared accumulation has no device");
VERIFY_FAIL(L"Invalid group-shared accumulation device");
return;
}
if (Params.CompType != ComponentType::F16 &&
Params.CompType != ComponentType::I32) {
hlsl_test::LogErrorFmt(
L"Group-shared accumulation does not support component type %u",
static_cast<uint32_t>(Params.CompType));

@V-FEXrt Ashley Coleman (V-FEXrt) Aug 25, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: it would be a nice later cleanup to actually print the comp name here instead of a integer

VERIFY_FAIL(L"Unsupported group-shared accumulation component type");
return;
}
if (Params.Scope != MatrixScope::Wave) {
hlsl_test::LogErrorFmt(L"Group-shared accumulation requires Wave scope");

@V-FEXrt Ashley Coleman (V-FEXrt) Aug 25, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it makes sense to have this restriction here on this test but IIRC ThreadGroup is alos allowed so this message may be a bit confusing

@JoeCitizen Jack Elliott (JoeCitizen) Aug 25, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Noted for a NFC clean up change

VERIFY_FAIL(L"Invalid group-shared accumulation scope");
return;
}
if (Params.Use != MatrixUse::Accumulator) {
hlsl_test::LogErrorFmt(
L"Group-shared accumulation requires Accumulator use");
VERIFY_FAIL(L"Invalid group-shared accumulation matrix use");
return;
}

size_t BufferSize;
UINT NumElements;
if (!getGroupSharedBufferDescription(Params, MemoryLayout, BufferSize,
Expand All @@ -7583,24 +7675,74 @@ static void runGroupSharedAccumulate(
}

const size_t MatrixElements = Params.totalElements();
std::optional<cpu_oracle::TypedMatrix> InitialMatrix =
cpu_oracle::makeTypedMatrix<HLSLHalf_t>(
Params.M, Params.N,
std::vector<HLSLHalf_t>(
MatrixElements, HLSLHalf_t(static_cast<float>(InitialValue))));
std::optional<cpu_oracle::TypedMatrix> ExpectedMatrix =
cpu_oracle::makeSequentialMatrix(Params.CompType, Params.M, Params.N,
InitialValue + AccumulateStartingValue);
// Wave w contributes w + 1 times its element value, matching the shader, so
// the weighted total distinguishes a dropped update from a duplicated one.
const uint32_t WaveWeightSum = ActiveWaveCount * (ActiveWaveCount + 1) / 2;
std::optional<cpu_oracle::TypedMatrix> InitialMatrix;
std::optional<cpu_oracle::TypedMatrix> ExpectedMatrix;
if (Params.CompType == ComponentType::F16) {
std::vector<HLSLHalf_t> InitialValues(
MatrixElements, HLSLHalf_t(static_cast<float>(InitialValue)));
std::vector<HLSLHalf_t> ExpectedValues;
ExpectedValues.reserve(MatrixElements);
for (size_t I = 0; I < MatrixElements; ++I) {
const uint32_t Value =
InitialValue +
WaveWeightSum * (AccumulateStartingValue + static_cast<uint32_t>(I));
ExpectedValues.emplace_back(static_cast<float>(Value));
}
InitialMatrix = cpu_oracle::makeTypedMatrix(Params.M, Params.N,
std::move(InitialValues));
ExpectedMatrix = cpu_oracle::makeTypedMatrix(Params.M, Params.N,
std::move(ExpectedValues));
} else {
std::vector<int32_t> InitialValues(MatrixElements,
static_cast<int32_t>(InitialValue));
std::vector<int32_t> ExpectedValues;
ExpectedValues.reserve(MatrixElements);
for (size_t I = 0; I < MatrixElements; ++I) {
const uint32_t Value =
InitialValue +
WaveWeightSum * (AccumulateStartingValue + static_cast<uint32_t>(I));
ExpectedValues.push_back(static_cast<int32_t>(Value));
}
InitialMatrix = cpu_oracle::makeTypedMatrix(Params.M, Params.N,
std::move(InitialValues));
ExpectedMatrix = cpu_oracle::makeTypedMatrix(Params.M, Params.N,
std::move(ExpectedValues));
}
if (!InitialMatrix || !ExpectedMatrix) {
hlsl_test::LogErrorFmt(
L"Failed to build %ux%u group-shared accumulation matrices for "
L"component type %u",
Params.M, Params.N, static_cast<uint32_t>(Params.CompType));
VERIFY_FAIL(L"Invalid group-shared accumulation matrices");
return;
}

std::optional<std::vector<BYTE>> Initial =
makeGroupSharedTypedBuffer(Params.CompType, BufferSize, 90);
std::optional<std::vector<BYTE>> Expected =
makeGroupSharedTypedBuffer(Params.CompType, BufferSize, 90);
if (!InitialMatrix.has_value() || !ExpectedMatrix.has_value() ||
!Initial.has_value() || !Expected.has_value() ||
!cpu_oracle::writeMatrixBuffer(*InitialMatrix, MemoryLayout, *Initial) ||
!cpu_oracle::writeMatrixBuffer(*ExpectedMatrix, MemoryLayout,
if (!Initial || !Expected) {
hlsl_test::LogErrorFmt(
L"Failed to allocate %zu-byte group-shared accumulation buffers for "
L"component type %u",
BufferSize, static_cast<uint32_t>(Params.CompType));
VERIFY_FAIL(L"Invalid group-shared accumulation buffers");
return;
}
if (!cpu_oracle::writeMatrixBuffer(*InitialMatrix, MemoryLayout, *Initial)) {
hlsl_test::LogErrorFmt(
L"Failed to write the initial group-shared accumulation matrix");
VERIFY_FAIL(L"Invalid initial group-shared accumulation layout");
return;
}
if (!cpu_oracle::writeMatrixBuffer(*ExpectedMatrix, MemoryLayout,
*Expected)) {
VERIFY_IS_TRUE(false, "Failed to build group-shared accumulate oracle");
hlsl_test::LogErrorFmt(
L"Failed to write the expected group-shared accumulation matrix");
VERIFY_FAIL(L"Invalid expected group-shared accumulation layout");
return;
}

Expand All @@ -7611,6 +7753,7 @@ static void runGroupSharedAccumulate(
ExtraDefs << " -DMEM_STRIDE=" << MemoryLayout.StrideBytes / ElementBytes;
ExtraDefs << " -DMEM_LAYOUT=" << static_cast<UINT>(MemoryLayout.Layout);
ExtraDefs << " -DACCUMULATE_START=" << AccumulateStartingValue;
ExtraDefs << " -DACTIVE_WAVE_COUNT=" << ActiveWaveCount;
if (ForcedWaveSize != 0)
ExtraDefs << " -DFORCED_WAVE_SIZE=" << ForcedWaveSize;

Expand All @@ -7634,11 +7777,16 @@ static void runGroupSharedAccumulate(

MappedData OutData;
Result->Test->GetReadBackData("Output", &OutData);
std::wstringstream PublicRule;
PublicRule << L"MatrixAccumulateToMemory atomically adds each logical "
L"matrix value from "
<< ActiveWaveCount << L" active Wave";
if (ActiveWaveCount != 1)
PublicRule << L"s";
PublicRule << L", preserving every padding byte and guard";
VERIFY_IS_TRUE(verifyGroupSharedTypedBuffer(
Params.CompType, OutData.data(), OutData.size(), *Expected,
L"MatrixAccumulateToMemory adds each logical matrix value to 12 while "
L"preserving exact padding and guards",
Verbose));
PublicRule.str().c_str(), Verbose));
}

static void runPaddedGroupSharedAccumulateCase(
Expand Down Expand Up @@ -7714,6 +7862,55 @@ static void runPaddedGroupSharedAccumulateCase(
SelectedWaveSize);
}

static void runGroupSharedAccumulateContention(
ID3D12Device *Device, dxc::SpecificDllLoader &DxcSupport,
ComponentType CompType, LPCWSTR CaseName, bool Verbose) {
MatrixParams Params = {};
Params.CompType = CompType;
Params.M = 4;
Params.N = 8;
Params.Use = MatrixUse::Accumulator;
Params.Scope = MatrixScope::Wave;
Params.Layout = MatrixLayout::RowMajor;
Params.NumThreads = 512;
Params.Enable16Bit = CompType == ComponentType::F16;

UINT SelectedWaveSize = 0;
if (!matrixConstructionApplicable(Device, Params, {Params.Use}, CaseName,
SelectedWaveSize))
return;
if (!accumulateStoreApplicable(Device, Params.CompType,
linalg_test::AtomicDestination::GroupShared,
CaseName))
return;

constexpr UINT ActiveWaveCount = 4;
Params.NumThreads = static_cast<int>(SelectedWaveSize * ActiveWaveCount);

const size_t ElementBytes = elementSize(CompType);
const cpu_oracle::MatrixBufferLayout Memory = {
MatrixLayout::RowMajor,
/*OffsetBytes=*/4 * ElementBytes,
/*StrideBytes=*/Params.N * ElementBytes + 8,
};
runGroupSharedAccumulate(Device, DxcSupport, Params, Memory,
/*InitialValue=*/7,
/*AccumulateStartingValue=*/1, Verbose,
SelectedWaveSize, ActiveWaveCount);
Comment on lines +7896 to +7899
}

void DxilConf_SM610_LinAlg::AccumulateMemoryContention_Wave_4x8_F16() {
runGroupSharedAccumulateContention(D3DDevice, DxcSupport, ComponentType::F16,
L"AccumulateMemoryContention_Wave_4x8_F16",
VerboseLogging);
}

void DxilConf_SM610_LinAlg::AccumulateMemoryContention_Wave_4x8_I32() {
runGroupSharedAccumulateContention(D3DDevice, DxcSupport, ComponentType::I32,
L"AccumulateMemoryContention_Wave_4x8_I32",
VerboseLogging);
}

static const char ConvertShader[] = R"(
#define CT_F16 8
#define CT_F32 9
Expand Down
Loading