[CoopVec] Add Linear Algebra common header with tests (#7350) (#7388)

This PR introduces the linear algebra header file, and places it in a
location that is by default included in all HLSL compilation. The
builtins in the API aren't yet defined, and depend on the #7290 PR
merging first.
The tests that have been added have temporary diagnostic messages while
7290 is in progress. They will need to be updated. Open to feedback on
better / suggested error messages, or whether there shouldn't be any
sema-level validation for these errors.

Fixes
[#7304](#7304)

Cherrypick of
#7350

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

Author: bob80905

tools/clang/lib/Headers/hlsl/dx/linalg.h

@@ -0,0 +1,182 @@
+// Header for linear algebra APIs.
+
+#if __spirv__
+#error "Cooperative vectors not (yet) supported for SPIRV"
+#endif
+
+#if ((__SHADER_TARGET_MAJOR > 6) ||                                            \
+     (__SHADER_TARGET_MAJOR == 6 && __SHADER_TARGET_MINOR >= 9)) &&            \
+    (__HLSL_VERSION >= 2021)
+
+namespace dx {
+namespace linalg {
+
+// NOTE: can't be an enum class because we get this error:
+//     error: non-type template argument of type 'dx::linalg::DataType' is not
+//     an integral constant expression
+//
+enum DataType {
+  DATA_TYPE_SINT16 = 2,           // ComponentType::I16
+  DATA_TYPE_UINT16 = 3,           // ComponentType::U16
+  DATA_TYPE_SINT32 = 4,           // ComponentType::I32
+  DATA_TYPE_UINT32 = 5,           // ComponentType::U32
+  DATA_TYPE_FLOAT16 = 8,          // ComponentType::F16
+  DATA_TYPE_FLOAT32 = 9,          // ComponentType::F32
+  DATA_TYPE_SINT8_T4_PACKED = 17, // ComponentType::PackedS8x32
+  DATA_TYPE_UINT8_T4_PACKED = 18, // ComponentType::PackedU8x32
+  DATA_TYPE_UINT8 = 19,           // ComponentType::U8
+  DATA_TYPE_SINT8 = 20,           // ComponentType::I8
+  DATA_TYPE_FLOAT8_E4M3 = 21,     // ComponentType::F8_E4M3
+                                  // (1 sign, 4 exp, 3 mantissa bits)
+  DATA_TYPE_FLOAT8_E5M2 = 22,     // ComponentType::F8_E5M2
+                                  // (1 sign, 5 exp, 2 mantissa bits)
+};
+
+enum MatrixLayout {
+  MATRIX_LAYOUT_ROW_MAJOR = 0,
+  MATRIX_LAYOUT_COLUMN_MAJOR = 1,
+  MATRIX_LAYOUT_MUL_OPTIMAL = 2,
+  MATRIX_LAYOUT_OUTER_PRODUCT_OPTIMAL = 3
+};
+
+//
+// Helper for signedness
+//
+namespace details {
+template <typename T> bool IsUnsigned() { return false; }
+
+#ifdef __HLSL_ENABLE_16_BIT
+template <> bool IsUnsigned<uint16_t>() { return true; }
+#endif
+
+template <> bool IsUnsigned<uint32_t>() { return true; }
+template <> bool IsUnsigned<uint64_t>() { return true; }
+} // namespace details
+
+//
+// (RW)MatrixRef
+//
+
+template <typename BufferTy, DataType DT, uint M, uint K, MatrixLayout ML,
+          bool Transpose>
+struct MatrixRefImpl {
+  BufferTy Buffer;
+  uint StartOffset;
+  uint Stride;
+};
+
+template <DataType DT, uint M, uint K, MatrixLayout ML, bool Transpose = false>
+using MatrixRef = MatrixRefImpl<ByteAddressBuffer, DT, M, K, ML, Transpose>;
+
+template <DataType DT, uint M, uint K, MatrixLayout ML, bool Transpose = false>
+using RWMatrixRef = MatrixRefImpl<RWByteAddressBuffer, DT, M, K, ML, Transpose>;
+
+//
+// (RW)VectorRef
+//
+
+template <typename BufferTy, DataType DT> struct VectorRefImpl {
+  BufferTy Buffer;
+  uint StartOffset;
+};
+
+template <DataType DT> using VectorRef = VectorRefImpl<ByteAddressBuffer, DT>;
+
+template <DataType DT>
+using RWVectorRef = VectorRefImpl<RWByteAddressBuffer, DT>;
+
+//
+// Vector
+//
+
+template <typename T, int N, DataType DT> struct InterpretedVector {
+  vector<T, N> Data;
+};
+
+template <DataType DT, typename T, int N>
+InterpretedVector<T, N, DT> MakeInterpretedVector(vector<T, N> Vec) {
+  InterpretedVector<T, N, DT> IV = {Vec};
+  return IV;
+}
+
+//
+// Mul
+//
+
+template <typename OutputElTy, typename InputElTy, int InputElCount,
+          typename MatrixBufferTy, DataType InputDT, DataType MatrixDT,
+          uint MatrixM, uint MatrixK, MatrixLayout MatrixLayout,
+          bool MatrixTranspose>
+vector<OutputElTy, MatrixM>
+Mul(MatrixRefImpl<MatrixBufferTy, MatrixDT, MatrixM, MatrixK, MatrixLayout,
+                  MatrixTranspose>
+        Matrix,
+    InterpretedVector<InputElTy, InputElCount, InputDT> InputVector) {
+
+  vector<OutputElTy, MatrixM> OutputVector;
+
+  __builtin_MatVecMul(
+      /*out*/ OutputVector, details::IsUnsigned<OutputElTy>(), InputVector.Data,
+      details::IsUnsigned<InputElTy>(), InputDT, Matrix.Buffer,
+      Matrix.StartOffset, MatrixDT, MatrixM, MatrixK, MatrixLayout,
+      MatrixTranspose, Matrix.Stride);
+
+  return OutputVector;
+}
+
+//
+// MulAdd
+//
+
+template <typename OutputElTy, typename InputElTy, int InputElCount,
+          typename MatrixBufferTy, DataType InputDT, DataType MatrixDT,
+          uint MatrixM, uint MatrixK, MatrixLayout MatrixLayout,
+          bool MatrixTranspose, typename BiasVectorBufferTy,
+          DataType BiasVectorDT>
+vector<OutputElTy, MatrixM>
+MulAdd(MatrixRefImpl<MatrixBufferTy, MatrixDT, MatrixM, MatrixK, MatrixLayout,
+                     MatrixTranspose>
+           Matrix,
+       InterpretedVector<InputElTy, InputElCount, InputDT> InputVector,
+       VectorRefImpl<BiasVectorBufferTy, BiasVectorDT> BiasVector) {
+
+  vector<OutputElTy, MatrixM> OutputVector;
+
+  __builtin_MatVecMulAdd(
+      /*out*/ OutputVector, details::IsUnsigned<OutputElTy>(), InputVector.Data,
+      details::IsUnsigned<InputElTy>(), InputDT, Matrix.Buffer,
+      Matrix.StartOffset, MatrixDT, MatrixM, MatrixK, MatrixLayout,
+      MatrixTranspose, Matrix.Stride, BiasVector.Buffer, BiasVector.StartOffset,
+      BiasVectorDT);
+
+  return OutputVector;
+}
+
+//
+// OuterProductAccumulate
+//
+
+template <typename ElTy, int MatrixM, int MatrixN, DataType MatrixDT,
+          MatrixLayout MatrixLayout>
+void OuterProductAccumulate(
+    vector<ElTy, MatrixM> InputVector1, vector<ElTy, MatrixN> InputVector2,
+    RWMatrixRef<MatrixDT, MatrixM, MatrixN, MatrixLayout, false> Matrix) {
+  __builtin_OuterProductAccumulate(InputVector1, InputVector2, Matrix.Buffer,
+                                   Matrix.StartOffset, MatrixDT, MatrixLayout,
+                                   Matrix.Stride);
+}
+
+//
+// VectorAccumulate
+//
+
+template <typename ElTy, int ElCount>
+void VectorAccumulate(vector<ElTy, ElCount> InputVector,
+                      RWByteAddressBuffer Buffer, uint Offset) {
+  __builtin_VectorAccumulate(InputVector, Buffer, Offset);
+}
+
+} // namespace linalg
+} // namespace dx
+
+#endif // SM 6.9 check and HV version check

tools/clang/test/CodeGenDXIL/hlsl/linalg/mat-vec-mul.hlsl

@@ -0,0 +1,40 @@
+// RUN: %dxc -I %hlsl_headers -T lib_6_9 -enable-16bit-types %s | FileCheck %s
+
+#include <dx/linalg.h>
+
+ByteAddressBuffer Buf;
+
+export float4 Test1(vector<float, 4> Input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_FLOAT16, 4, 4, MATRIX_LAYOUT_MUL_OPTIMAL, true> Matrix = {
+      Buf, 0, 0};
+
+  // CHECK: %{{.+}} = call <4 x float> @dx.op.matVecMul.v4f32.v4f32(i32 305, <4 x float> %{{.+}}, i1 false, i32 8, %dx.types.Handle %{{.+}}, i32 0, i32 8, i32 4, i32 4, i32 2, i1 true, i32 0, i1 false)
+  return Mul<float>(    
+      Matrix, MakeInterpretedVector<DATA_TYPE_FLOAT16>(Input));
+}
+
+export vector<float, 8> Test2(vector<uint8_t4_packed, 6> Input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_UINT8, 8, 6 * 4, MATRIX_LAYOUT_MUL_OPTIMAL> Matrix = {
+      Buf, 0, 0};
+
+  // note the stride argument is dropped.
+  // CHECK: %{{.+}} = call <8 x float> @dx.op.matVecMul.v8f32.v6f32(i32 305, <6 x float> %{{.+}}, i1 false, i32 18, %dx.types.Handle %{{.+}}, i32 0, i32 19, i32 8, i32 24, i32 2, i1 false, i32 0, i1 false)
+  return Mul<float>(Matrix,
+                    MakeInterpretedVector<DATA_TYPE_UINT8_T4_PACKED>(Input));
+}
+
+// test that "stride" isn't ignored in non-optimal layouts
+export vector<float, 8> Test3(vector<uint8_t4_packed, 6> Input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_UINT8, 8, 6 * 4, MATRIX_LAYOUT_ROW_MAJOR> Matrix = {
+      Buf, 0, 6 * 4 * 8};
+
+  // CHECK: %{{.+}} = call <8 x float> @dx.op.matVecMul.v8f32.v6f32(i32 305, <6 x float> %{{.+}}, i1 false, i32 18, %dx.types.Handle %{{.+}}, i32 0, i32 19, i32 8, i32 24, i32 0, i1 false, i32 192, i1 false)
+  return Mul<float>(Matrix,
+                    MakeInterpretedVector<DATA_TYPE_UINT8_T4_PACKED>(Input));
+}

tools/clang/test/CodeGenDXIL/hlsl/linalg/mat-vec-muladd.hlsl

@@ -0,0 +1,90 @@
+// RUN: %dxc -I %hlsl_headers -T lib_6_9 %s | FileCheck %s
+
+#include <dx/linalg.h>
+
+ByteAddressBuffer Buf;
+
+export float4 Test1(float4 input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_FLOAT16, 4, 4, MATRIX_LAYOUT_MUL_OPTIMAL> matrix = {Buf,
+                                                                          0, 0};
+  VectorRef<DATA_TYPE_FLOAT16> biasVector = {Buf, 256};
+
+  InterpretedVector<float, 4, DATA_TYPE_FLOAT16> theVector = {input};
+
+  // CHECK: %{{.+}} = call <4 x float> @dx.op.matVecMulAdd.v4f32.v4f32(i32 306, <4 x float> %{{.+}}, i1 false, i32 8, %dx.types.Handle [[RES:%.+]], i32 0, i32 8, i32 4, i32 4, i32 2, i1 false, i32 0, %dx.types.Handle [[RES]], i32 256, i32 8, i1 false)
+  return MulAdd<float>(
+      matrix, theVector,
+      biasVector);
+}
+
+export float4 Test2(float4 input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_FLOAT16, 4, 4, MATRIX_LAYOUT_MUL_OPTIMAL, true> matrix = {
+      Buf, 0, 0};
+  VectorRef<DATA_TYPE_FLOAT16> biasVector = {Buf, 256};
+
+  InterpretedVector<float, 4, DATA_TYPE_FLOAT16> theVector = {input};
+
+  // CHECK: %{{.+}} = call <4 x float> @dx.op.matVecMulAdd.v4f32.v4f32(i32 306, <4 x float> %{{.+}}, i1 false, i32 8, %dx.types.Handle [[RES:%.+]], i32 0, i32 8, i32 4, i32 4, i32 2, i1 true, i32 0, %dx.types.Handle [[RES]], i32 256, i32 8, i1 false)
+  return MulAdd<float>(
+      matrix, theVector,
+      biasVector);
+}
+
+export float4 Test3(float4 input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_FLOAT16, 4, 4, MATRIX_LAYOUT_MUL_OPTIMAL, true> matrix = {
+      Buf, 0, 0};
+  VectorRef<DATA_TYPE_FLOAT16> biasVector = {Buf, 256};
+
+  // CHECK: %{{.+}} = call <4 x float> @dx.op.matVecMulAdd.v4f32.v4f32(i32 306, <4 x float> %{{.+}}, i1 false, i32 8, %dx.types.Handle [[RES:%.+]], i32 0, i32 8, i32 4, i32 4, i32 2, i1 true, i32 0, %dx.types.Handle [[RES]], i32 256, i32 8, i1 false)
+  return MulAdd<float>(
+      matrix, MakeInterpretedVector<DATA_TYPE_FLOAT16>(input),
+      biasVector);
+}
+
+namespace ProposalExample {
+
+ByteAddressBuffer model;
+
+vector<float, 3> ApplyNeuralMaterial(vector<half, 8> inputVector) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_FLOAT8_E4M3, 32, 8, MATRIX_LAYOUT_MUL_OPTIMAL> matrix0 = {
+      model, 0, 0};
+
+  VectorRef<DATA_TYPE_FLOAT16> biasVector0 = {model, 1024};
+
+  MatrixRef<DATA_TYPE_FLOAT8_E4M3, 32, 32, MATRIX_LAYOUT_MUL_OPTIMAL> matrix1 =
+      {model, 2048, 0};
+
+  VectorRef<DATA_TYPE_FLOAT16> biasVector1 = {model, 3072};
+
+  MatrixRef<DATA_TYPE_FLOAT8_E4M3, 3, 32, MATRIX_LAYOUT_MUL_OPTIMAL> matrix2 = {
+      model, 4096, 0};
+
+  VectorRef<DATA_TYPE_FLOAT16> biasVector2 = {model, 5120};
+
+  vector<half, 32> layer0 = MulAdd<half>(
+      matrix0, MakeInterpretedVector<DATA_TYPE_FLOAT8_E4M3>(inputVector),
+      biasVector0);
+  layer0 = max(layer0, 0);
+
+  vector<half, 32> layer1 = MulAdd<half>(
+      matrix1, MakeInterpretedVector<DATA_TYPE_FLOAT8_E4M3>(layer0),
+      biasVector1);
+  layer1 = max(layer1, 0);
+
+  vector<float, 3> output = MulAdd<float>(
+      matrix2, MakeInterpretedVector<DATA_TYPE_FLOAT8_E4M3>(layer1),
+      biasVector2);
+  output = exp(output);
+
+  return output;
+}
+
+} // namespace ProposalExample

tools/clang/test/CodeGenDXIL/hlsl/linalg/outerproductaccumulate.hlsl

@@ -0,0 +1,16 @@
+// RUN: %dxc -I %hlsl_headers -T lib_6_9 -enable-16bit-types %s | FileCheck %s
+
+#include <dx/linalg.h>
+
+RWByteAddressBuffer RWBuf;
+
+export void Test4(vector<half, 128> Input1, vector<half, 64> Input2) {
+  using namespace dx::linalg;
+
+  RWMatrixRef<DATA_TYPE_FLOAT16, 128, 64, MATRIX_LAYOUT_OUTER_PRODUCT_OPTIMAL>
+      matrix = {RWBuf, 0, 0};
+
+  // CHECK: call void @dx.op.outerProductAccumulate.v128f16.v64f16(i32 307, <128 x half> %{{.+}}, <64 x half> %{{.+}}, %dx.types.Handle %{{.+}}, i32 0, i32 8, i32 3, i32 0)
+
+  OuterProductAccumulate(Input1, Input2, matrix);  
+}

tools/clang/test/CodeGenDXIL/hlsl/linalg/vectoraccumulate.hlsl

@@ -0,0 +1,14 @@
+// RUN: %dxc -I %hlsl_headers -T lib_6_9 %s | FileCheck %s
+
+#include <dx/linalg.h>
+
+RWByteAddressBuffer RWBuf;
+
+export void Test5(vector<half, 128> Input) {
+  using namespace dx::linalg;
+
+  RWBuf.Store<vector<half, 128> >(0, Input);
+
+  // CHECK: call void @dx.op.vectorAccumulate.v128f32(i32 308, <128 x float> %{{.*}}, %dx.types.Handle %{{.*}}, i32 0)
+  VectorAccumulate(Input, RWBuf, 0);
+}

tools/clang/test/SemaHLSL/hlsl/linalg/make-interp-vec-errors.hlsl

@@ -0,0 +1,33 @@
+// RUN: %dxc -I %hlsl_headers -T lib_6_9 %s -verify
+
+#include <dx/linalg.h>
+ByteAddressBuffer Buf;
+
+export float4 Test1(vector<float, 4> Input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_UINT16, 4, 4, MATRIX_LAYOUT_MUL_OPTIMAL, true> Matrix = {
+      Buf, 0, 0};
+
+  // expected-error@+3{{no matching function for call to 'MakeInterpretedVector'}}
+  // expected-note@dx/linalg.h:97{{candidate template ignored: invalid explicitly-specified argument for template parameter 'DT'}}
+  return Mul<float>(    
+      Matrix, MakeInterpretedVector<2>(Input));
+}
+
+enum DataType {
+  DATA_TYPE_InvalidType = 40
+};
+
+export float4 Test2(vector<float, 4> Input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_UINT16, 4, 4, MATRIX_LAYOUT_MUL_OPTIMAL, true> Matrix = {
+      Buf, 0, 0};
+
+  // expected-error@+3{{no matching function for call to 'MakeInterpretedVector'}}
+  // expected-note@dx/linalg.h:97{{candidate template ignored: invalid explicitly-specified argument for template parameter 'DT'}}
+  return Mul<float>(    
+      Matrix, MakeInterpretedVector<DATA_TYPE_InvalidType>(Input));
+}
+

tools/clang/test/SemaHLSL/hlsl/linalg/mat-vec-mul-errors.hlsl

@@ -0,0 +1,16 @@
+// RUN: %dxc -I %hlsl_headers -T lib_6_9 %s -verify
+
+#include <dx/linalg.h>
+
+ByteAddressBuffer Buf;
+
+vector<float, 128> MixUpVectorAndMatrixArguments(vector<float, 128> Input) {
+  using namespace dx::linalg;
+
+  MatrixRef<DATA_TYPE_FLOAT16, 128, 128, MATRIX_LAYOUT_MUL_OPTIMAL> Matrix = {
+      Buf, 0, 0};
+
+  // expected-error@+2{{no matching function for call to 'Mul'}}
+  // expected-note@dx/linalg.h:111{{candidate template ignored: could not match 'MatrixRefImpl' against 'InterpretedVector'}}
+  return Mul<float>(MakeInterpretedVector<DATA_TYPE_FLOAT16>(Input), Matrix);
+}