[mlir][gpu] Add support for unsigned integer extend in vector to gpu.subgroup_mma lowering

[qedawkins]

Summary

Unsigned integer types are supported in subgroup mma ops by matching
against arith.extui ops. This allows for subgroup_mma_compute ops with
mixed signedness which requires later conversions to handle this. SPIR-V
cooperative matrix ops support this while the lowering to WMMA does not.

[antiagainst]

LG, thanks!

mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

543

Nit: use a ternary operator to select between IntegerType::Signed and IntegerType::Unsigned to avoid duplication?

[qedawkins]

mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

543

InferFragType is templated and would require a cast anyway so it's not clear to me that this is better. Would it be preferred to put this in its own templated function and update elType etc. by reference instead?

[qedawkins]

mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

543

@antiagainst gentle bump on this comment. If it's not a problem then I'll land this.

[ThomasRaoux]

mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

543

Feel free to remove the template from InferFragType and make it:

static const char *inferFragType(Operation* op) {
  assert(op->getNumResults() == 1);
  for (Operation *users : op->getUsers()) {
    auto contract = dyn_cast<vector::ContractionOp>(users);
    if (!contract)
      continue;
    if (contract.getLhs() == op.getResult(0))
      return "AOp";
    if (contract.getRhs() == op.getResult(0))
      return "BOp";
  }
  return "COp";
}

[qedawkins]

rebase and address comment

[ThomasRaoux]

mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

502

can you assert than op has exactly 1 result?

[qedawkins]

assert exactly one result

mlir/lib/Conversion/GPUToNVVM/WmmaOpsToNvvm.cpp

static NVVM::MMATypes getElementType(gpu::MMAMatrixType type) {
   if (type.getElementType().isF16())
     return NVVM::MMATypes::f16;
   if (type.getElementType().isF32())
     return type.getOperand().equals("COp") ? NVVM::MMATypes::f32
                                            : NVVM::MMATypes::tf32;
 
   if (type.getElementType().isSignedInteger(8))
     return NVVM::MMATypes::s8;
+  if (type.getElementType().isUnsignedInteger(8))
+    return NVVM::MMATypes::u8;
   // Accumulator type is signless and implies signed.
   if (type.getElementType().isInteger(32))
     return NVVM::MMATypes::s32;
   llvm_unreachable("Unsupported type");
 }
 
 /// This class implements the conversion of GPU MMA loadOp to wmma.load op
 /// in the NVVM dialect. The conversion not only emits the NVVM op but also
if (retType.getOperand().equals("AOp")) {
       m = NVVM::WMMALoadOp::inferMDimension(k, n, eltype);
     } else if (retType.getOperand().equals("COp")) {
       m = retTypeShape[0];
       n = retTypeShape[1];
       k = NVVM::WMMALoadOp::inferKDimension(m, n, eltype);
     }
     NVVM::MMAFrag frag = convertOperand(retType.getOperand());
     // Check that there is an exisiting instruction for the combination we need.
-    if (NVVM::WMMALoadOp::getIntrinsicID(m, n, k, layout, eltype, frag) == 0) {
-      llvm::errs() << "No matching intrinsic " << m << " " << n << " " << k
-                   << "\n";
+    if (NVVM::WMMALoadOp::getIntrinsicID(m, n, k, layout, eltype, frag) == 0)
       return rewriter.notifyMatchFailure(op, kInvalidCaseStr);
-    }
 
     Type resType = convertMMAToLLVMType(retType);
     Location loc = op->getLoc();
 
     // Create nvvm.mma_load op according to the operand types.
     Value dataPtr = getStridedElementPtr(
         loc,
         subgroupMmaLoadMatrixOp.getSrcMemref().getType().cast<MemRefType>(),
NVVM::MMALayout bLayout = subgroupMmaComputeOp.getBTranspose()
                                   ? NVVM::MMALayout::col
                                   : NVVM::MMALayout::row;
     NVVM::MMATypes sourceType = getElementType(aType);
     NVVM::MMATypes destType = getElementType(cType);
     if (NVVM::WMMAMmaOp::getIntrinsicID(m, n, k, aLayout, bLayout, sourceType,
                                         destType) == 0)
       return rewriter.notifyMatchFailure(op, kInvalidCaseStr);
 
+    NVVM::MMATypes bElementType = getElementType(
+        subgroupMmaComputeOp.getOpB().getType().cast<gpu::MMAMatrixType>());
+    if (bElementType != sourceType)
+      return rewriter.notifyMatchFailure(
+          op, "WMMA compute op input matrix element types must match.");
+
     unpackOp(adaptor.getOpA());
     unpackOp(adaptor.getOpB());
     unpackOp(adaptor.getOpC());
 
     rewriter.replaceOpWithNewOp<NVVM::WMMAMmaOp>(
         op, adaptor.getOpC().getType(), m, n, k, aLayout, bLayout, sourceType,
         destType, unpackedOps);
     return success();

mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

static bool transferReadSupportsMMAMatrixType(vector::TransferReadOp readOp,
   if (readOp.getMask() || readOp.hasOutOfBoundsDim() ||
       readOp.getVectorType().getRank() != 2)
     return false;
   if (!getMemrefConstantHorizontalStride(readOp.getShapedType()))
     return false;
 
   // Only allow integer types if the signedness can be inferred.
   if (!useNvGpu && readOp.getVectorType().getElementType().isInteger(8))
-    if (!readOp->hasOneUse() || !isa<arith::ExtSIOp>(*readOp->user_begin()))
+    if (!readOp->hasOneUse() || (!isa<arith::ExtSIOp>(*readOp->user_begin()) &&
+                                 !isa<arith::ExtUIOp>(*readOp->user_begin())))
       return false;
 
   AffineMap map = readOp.getPermutationMap();
   OpBuilder b(readOp.getContext());
   AffineExpr innerDim = b.getAffineDimExpr(map.getNumDims() - 1);
   AffineExpr zero = b.getAffineConstantExpr(0);
   auto broadcastInnerDim = AffineMap::get(map.getNumDims(), 0, {zero, innerDim},
                                           readOp.getContext());
static bool constantSupportsMMAMatrixType(arith::ConstantOp constantOp) {
   return constantOp.getValue().isa<SplatElementsAttr>();
 }
 
 /// Return true if this is a broadcast from scalar to a 2D vector.
 static bool broadcastSupportsMMAMatrixType(vector::BroadcastOp broadcastOp) {
   return broadcastOp.getVectorType().getRank() == 2;
 }
 
-/// Return true if this signed extend op can be folded into a contract op.
-static bool signedExtendSupportsMMAMatrixType(arith::ExtSIOp extOp) {
+/// Return true if this integer extend op can be folded into a contract op.
+template <typename ExtOpTy>
+static bool integerExtendSupportsMMAMatrixType(ExtOpTy extOp) {
   if (!isa<vector::TransferReadOp>(extOp.getOperand().getDefiningOp()))
     return false;
   return llvm::all_of(extOp->getUsers(), [](Operation *user) {
     return isa<vector::ContractionOp>(user);
   });
 }
 
 /// Return the MMA elementwise enum associated with `op` if it is supported.
if (auto extractStridedSlice = dyn_cast<vector::ExtractStridedSliceOp>(op))
     return useNvGpu &&
            extractStridedSliceSupportsMMAMatrixType(extractStridedSlice);
   if (auto contract = dyn_cast<vector::ContractionOp>(op))
     return contractSupportsMMAMatrixType(contract, useNvGpu);
   if (auto constant = dyn_cast<arith::ConstantOp>(op))
     return constantSupportsMMAMatrixType(constant);
   if (auto broadcast = dyn_cast<vector::BroadcastOp>(op))
     return broadcastSupportsMMAMatrixType(broadcast);
-  if (auto extend = dyn_cast<arith::ExtSIOp>(op))
-    return signedExtendSupportsMMAMatrixType(extend);
+  if (auto signedExtend = dyn_cast<arith::ExtSIOp>(op))
+    return integerExtendSupportsMMAMatrixType<arith::ExtSIOp>(signedExtend);
+  if (auto unsignedExtend = dyn_cast<arith::ExtUIOp>(op))
+    return integerExtendSupportsMMAMatrixType<arith::ExtUIOp>(unsignedExtend);
   return elementwiseSupportsMMAMatrixType(op);
 }
 
 /// Return an unsorted slice handling scf.for region differently than
 /// `getSlice`. In scf.for we only want to include as part of the slice elements
 /// that are part of the use/def chain.
 static SetVector<Operation *> getSliceContract(Operation *op,
                                                TransitiveFilter backwardFilter,

 struct CombineTransferReadOpTranspose final
     : public OpRewritePattern<vector::TransposeOp> {
   using OpRewritePattern<vector::TransposeOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(vector::TransposeOp op,
                                 PatternRewriter &rewriter) const override {
     // Look through integer extend ops.
     Value source = op.getVector();
-    auto extOp = source.getDefiningOp<arith::ExtSIOp>();
     auto resultType = op.getVectorType();
-    if (extOp) {
-      source = extOp.getOperand();
+    Operation *extOp;
+    if ((extOp = source.getDefiningOp<arith::ExtSIOp>()) ||
+        (extOp = source.getDefiningOp<arith::ExtUIOp>())) {
+      source = extOp->getOperand(0);
       resultType =
           VectorType::get(resultType.getShape(),
                           source.getType().cast<VectorType>().getElementType());
     }
 
     auto transferReadOp = source.getDefiningOp<vector::TransferReadOp>();
     if (!transferReadOp)
       return failure();
Value result =
             .create<vector::TransferReadOp>(
                 loc, resultType, transferReadOp.getSource(),
                 transferReadOp.getIndices(), AffineMapAttr::get(newMap),
                 transferReadOp.getPadding(), transferReadOp.getMask(),
                 transferReadOp.getInBoundsAttr())
             .getResult();
 
     // Fuse through the integer extend op.
-    if (extOp)
+    if (extOp) {
+      if (isa<arith::ExtSIOp>(extOp))
         result = rewriter.create<arith::ExtSIOp>(loc, op.getType(), result)
                      .getResult();
+      else
+        result = rewriter.create<arith::ExtUIOp>(loc, op.getType(), result)
+                     .getResult();
+    }
 
     rewriter.replaceOp(op, result);
     return success();
   }
 };
 
 } // namespace
 
 // MMA types have different layout based on how they are used in matmul ops.
 // Figure the right layout to use by looking at op uses.
 // TODO: Change the GPU dialect to abstract the layout at the this level and
 // only care about it during lowering to NVVM.
-template <typename OpTy>
-static const char *inferFragType(OpTy op) {
+static const char *inferFragType(Operation *op) {
   for (Operation *users : op->getUsers()) {
     auto contract = dyn_cast<vector::ContractionOp>(users);
     if (!contract)
       continue;
-    if (contract.getLhs() == op.getResult())
+    if (contract.getLhs() == op->getResult(0))
       return "AOp";
-    if (contract.getRhs() == op.getResult())
+    if (contract.getRhs() == op->getResult(0))
       return "BOp";
   }
   return "COp";
 }
 
 static void convertTransferReadOp(vector::TransferReadOp op,
                                   llvm::DenseMap<Value, Value> &valueMapping) {
   assert(op.getTransferRank() > 0 && "unexpected 0-d transfer");
if (auto cstExpr =
     assert(cstExpr.getValue() == 0);
     stride = 0;
   }
   assert(stride);
   Value mappingResult = op.getResult();
   auto elType = op.getVectorType().getElementType();
   const char *fragType = inferFragType(op);
   if (op->hasOneUse()) {
-    auto extOp = dyn_cast<arith::ExtSIOp>(*op->user_begin());
-    // Infer the signedness of the mma type from the signed extend.
-    if (extOp) {
-      elType = IntegerType::get(op.getContext(),
-                                elType.cast<IntegerType>().getWidth(),
-                                IntegerType::Signed);
-      mappingResult = extOp.getResult();
-      fragType = inferFragType(extOp);
+    auto user = *op->user_begin();
+    // Infer the signedness of the mma type from the integer extend.
+    bool isSignedExtend = isa<arith::ExtSIOp>(user);
+    if (isSignedExtend || isa<arith::ExtUIOp>(user)) {
+      elType = IntegerType::get(
+          op.getContext(), elType.cast<IntegerType>().getWidth(),
+          isSignedExtend ? IntegerType::Signed : IntegerType::Unsigned);
+      mappingResult = user->getResult(0);
+      fragType = inferFragType(user);
     }
   }
   gpu::MMAMatrixType type =
       gpu::MMAMatrixType::get(op.getVectorType().getShape(), elType, fragType);
   Value load = b.create<gpu::SubgroupMmaLoadMatrixOp>(
       op.getLoc(), type, op.getSource(), op.getIndices(),
       b.getIndexAttr(*stride), isTranspose ? b.getUnitAttr() : UnitAttr());
   valueMapping[mappingResult] = load;
 }
 

mlir/lib/Dialect/SPIRV/IR/SPIRVOps.cpp

verifyCoopMatrixMulAdd(spirv::NVCooperativeMatrixMulAddOp op) {
   if (typeA.getRows() != typeR.getRows() ||
       typeA.getColumns() != typeB.getRows() ||
       typeB.getColumns() != typeR.getColumns())
     return op.emitOpError("matrix size must match");
   if (typeR.getScope() != typeA.getScope() ||
       typeR.getScope() != typeB.getScope() ||
       typeR.getScope() != typeC.getScope())
     return op.emitOpError("matrix scope must match");
-  if (typeA.getElementType() != typeB.getElementType() ||
-      typeR.getElementType() != typeC.getElementType())
-    return op.emitOpError("matrix element type must match");
+  auto elementTypeA = typeA.getElementType();
+  auto elementTypeB = typeB.getElementType();
+  if (isa<IntegerType>(elementTypeA) && isa<IntegerType>(elementTypeB)) {
+    if (elementTypeA.cast<IntegerType>().getWidth() !=
+        elementTypeB.cast<IntegerType>().getWidth())
+      return op.emitOpError(
+          "matrix A and B integer element types must be the same bit width");
+  } else if (elementTypeA != elementTypeB) {
+    return op.emitOpError(
+        "matrix A and B non-integer element types must match");
+  }
+  if (typeR.getElementType() != typeC.getElementType())
+    return op.emitOpError("matrix accumulator element type must match");
   return success();
 }
 
 LogicalResult spirv::NVCooperativeMatrixMulAddOp::verify() {
   return verifyCoopMatrixMulAdd(*this);
 }
 
 static LogicalResult

mlir/test/Conversion/VectorToGPU/vector-to-mma-ops.mlir

func.func @matmul_int8(%arg0: memref<16x16xi8>, %arg1: memref<16x16xi8>, %arg2: memref<16x16xi32>) {
   %Br = vector.transfer_read %arg1[%c0, %c0], %cst_i8 {permutation_map = #map0, in_bounds = [true, true]} : memref<16x16xi8>, vector<16x16xi8>
   %C = vector.transfer_read %arg2[%c0, %c0], %cst_i32 {in_bounds = [true, true]} : memref<16x16xi32>, vector<16x16xi32>
   %Ae = arith.extsi %Ar : vector<16x16xi8> to vector<16x16xi32>
   %Be = arith.extsi %Br : vector<16x16xi8> to vector<16x16xi32>
   %D = vector.contract {indexing_maps = [#map1, #map2, #map3], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %Ae, %Be, %C : vector<16x16xi32>, vector<16x16xi32> into vector<16x16xi32>
   vector.transfer_write %D, %arg2[%c0, %c0] {in_bounds = [true, true]} : vector<16x16xi32>, memref<16x16xi32>
   return
 }
+
+// CHECK-LABEL: func @matmul_mixed_signedness_int8
+//   CHECK-DAG:   %[[A:.+]] = gpu.subgroup_mma_load_matrix %{{.*}}[%{{.*}}, %{{.*}}] {leadDimension = 16 : index} : memref<16x16xi8> -> !gpu.mma_matrix<16x16xui8, "AOp">
+//   CHECK-DAG:   %[[B:.+]] = gpu.subgroup_mma_load_matrix %{{.*}}[%{{.*}}, %{{.*}}] {leadDimension = 16 : index} : memref<16x16xi8> -> !gpu.mma_matrix<16x16xsi8, "BOp">
+//   CHECK-DAG:   %[[C:.+]] = gpu.subgroup_mma_load_matrix %{{.*}}[%{{.*}}, %{{.*}}] {leadDimension = 16 : index} : memref<16x16xi32> -> !gpu.mma_matrix<16x16xi32, "COp">
+//       CHECK:   %[[D:.+]] = gpu.subgroup_mma_compute %[[A]], %[[B]], %[[C]] : !gpu.mma_matrix<16x16xui8, "AOp">, !gpu.mma_matrix<16x16xsi8, "BOp"> -> !gpu.mma_matrix<16x16xi32, "COp">
+//       CHECK:   gpu.subgroup_mma_store_matrix %[[D]], %{{.*}}[%{{.*}}, %{{.*}}] {leadDimension = 16 : index} : !gpu.mma_matrix<16x16xi32, "COp">, memref<16x16xi32>
+func.func @matmul_mixed_signedness_int8(%arg0: memref<16x16xi8>, %arg1: memref<16x16xi8>, %arg2: memref<16x16xi32>) {
+  %cst_0 = arith.constant dense<0> : vector<16x16xi8>
+  %c0 = arith.constant 0 : index
+  %cst_i8 = arith.constant 0 : i8
+  %cst_i32 = arith.constant 0 : i32
+  %Ar = vector.transfer_read %arg0[%c0, %c0], %cst_i8 {in_bounds = [true, true]} : memref<16x16xi8>, vector<16x16xi8>
+  %Br = vector.transfer_read %arg1[%c0, %c0], %cst_i8 {permutation_map = #map0, in_bounds = [true, true]} : memref<16x16xi8>, vector<16x16xi8>
+  %C = vector.transfer_read %arg2[%c0, %c0], %cst_i32 {in_bounds = [true, true]} : memref<16x16xi32>, vector<16x16xi32>
+  %Ae = arith.extui %Ar : vector<16x16xi8> to vector<16x16xi32>
+  %Be = arith.extsi %Br : vector<16x16xi8> to vector<16x16xi32>
+  %D = vector.contract {indexing_maps = [#map1, #map2, #map3], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %Ae, %Be, %C : vector<16x16xi32>, vector<16x16xi32> into vector<16x16xi32>
+  vector.transfer_write %D, %arg2[%c0, %c0] {in_bounds = [true, true]} : vector<16x16xi32>, memref<16x16xi32>
+  return
+}

mlir/test/Dialect/SPIRV/IR/cooperative-matrix-ops.mlir

spirv.func @cooperative_matrix_muladd(%a : !spirv.coopmatrix<8x16xi32, Subgroup>, %b : !spirv.coopmatrix<16x8xi32, Workgroup>, %c : !spirv.coopmatrix<8x8xi32, Subgroup>) "None" {
   // expected-error @+1 {{'spirv.NV.CooperativeMatrixMulAdd' op matrix scope must match}}
   %r = spirv.NV.CooperativeMatrixMulAdd %a, %b, %c : !spirv.coopmatrix<8x16xi32, Subgroup>, !spirv.coopmatrix<16x8xi32, Workgroup> -> !spirv.coopmatrix<8x8xi32, Subgroup>
   spirv.Return
 }
 
 // -----
 
 spirv.func @cooperative_matrix_muladd(%a : !spirv.coopmatrix<8x16xf32, Subgroup>, %b : !spirv.coopmatrix<16x8xi32, Subgroup>, %c : !spirv.coopmatrix<8x8xi32, Subgroup>) "None" {
-  // expected-error @+1 {{matrix element type must match}}
+  // expected-error @+1 {{matrix A and B non-integer element types must match}}
   %r = spirv.NV.CooperativeMatrixMulAdd %a, %b, %c : !spirv.coopmatrix<8x16xf32, Subgroup>, !spirv.coopmatrix<16x8xi32, Subgroup> -> !spirv.coopmatrix<8x8xi32, Subgroup>
   spirv.Return
 }
 
 // -----
 
+spirv.func @cooperative_matrix_muladd(%a : !spirv.coopmatrix<8x16xui8, Subgroup>, %b : !spirv.coopmatrix<16x8xsi32, Subgroup>, %c : !spirv.coopmatrix<8x8xi32, Subgroup>) "None" {
+  // expected-error @+1 {{matrix A and B integer element types must be the same bit width}}
+  %r = spirv.NV.CooperativeMatrixMulAdd %a, %b, %c : !spirv.coopmatrix<8x16xui8, Subgroup>, !spirv.coopmatrix<16x8xsi32, Subgroup> -> !spirv.coopmatrix<8x8xi32, Subgroup>
+  spirv.Return
+}
+
+// -----
+
 spirv.func @cooperative_matrix_load_memaccess(%ptr : !spirv.ptr<!spirv.struct<(f32 [0])>, StorageBuffer>, %stride : i32, %b : i1) "None" {
   // expected-error @+1 {{Pointer must point to a scalar or vector type}}
   %0 = spirv.NV.CooperativeMatrixLoad %ptr, %stride, %b : !spirv.ptr<!spirv.struct<(f32 [0])>, StorageBuffer> as !spirv.coopmatrix<8x16xi32, Subgroup>
   spirv.Return
 }
 
 // -----
 
 spirv.func @cooperative_matrix_load_memaccess(%ptr : !spirv.ptr<i32, Function>, %stride : i32, %b : i1) "None" {
   // expected-error @+1 {{Pointer storage class must be Workgroup, StorageBuffer or PhysicalStorageBufferEXT}}
   %0 = spirv.NV.CooperativeMatrixLoad %ptr, %stride, %b : !spirv.ptr<i32, Function> as !spirv.coopmatrix<8x16xi32, Subgroup>
   spirv.Return
 }