[mlir][gpu][vector] Lower Vector dialect to GPU for element-wise ops only (#159091)

Current convertVectorToMMAOps starts from vector.contract and finds its
dependencies as the targets to convert. In GPU dialect, we have
gpu.subgroup_mma_elementwise operation. We should be able to lower
element-wise operations to GPU MMA operations without vector.contract.
This patch adds this case to the pattern.

Author: Hsiangkai

mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

@@ -355,11 +355,14 @@ static SetVector<Operation *> getOpToConvert(mlir::Operation *op,
   forwardSliceOptions.filter = hasVectorSrc;
 
   SetVector<Operation *> opToConvert;
-  op->walk([&](vector::ContractionOp contract) {
-    if (opToConvert.contains(contract.getOperation()))
+  op->walk([&](Operation *nestedOp) {
+    if (!isa<vector::ContractionOp>(nestedOp) &&
+        !elementwiseSupportsMMAMatrixType(nestedOp))
+      return;
+    if (opToConvert.contains(nestedOp))
       return;
     SetVector<Operation *> dependentOps =
-        getSliceContract(contract, backwardSliceOptions, forwardSliceOptions);
+        getSliceContract(nestedOp, backwardSliceOptions, forwardSliceOptions);
     // If any instruction cannot use MMA matrix type drop the whole
     // chain. MMA matrix are stored in an opaque type so they cannot be used
     // by all operations.

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

@@ -536,3 +536,22 @@ func.func @test_unsupported(%arg0: vector<4x4xi32>, %arg1: vector<4x4xi32>, %arg
                         %0, %1, %arg2 : vector<4x4xi64>, vector<4x4xi64> into vector<4x4xi64>
   return %2 : vector<4x4xi64>
 }
+
+// -----
+
+#map0 = affine_map<(d0, d1) -> (d1, d0)>
+
+// CHECK-LABEL: func @addf
+//       CHECK:   %[[A:.+]] = gpu.subgroup_mma_load_matrix {{.+}} {leadDimension = 16 : index} : memref<16x16xf16> -> !gpu.mma_matrix<16x16xf16, "COp">
+//       CHECK:   %[[B:.+]] = gpu.subgroup_mma_load_matrix {{.+}} {leadDimension = 16 : index, transpose} : memref<16x16xf16> -> !gpu.mma_matrix<16x16xf16, "COp">
+//       CHECK:   %[[C:.+]] = gpu.subgroup_mma_elementwise  addf %[[A]], %[[B]] : (!gpu.mma_matrix<16x16xf16, "COp">, !gpu.mma_matrix<16x16xf16, "COp">) -> !gpu.mma_matrix<16x16xf16, "COp">
+//       CHECK:   gpu.subgroup_mma_store_matrix %[[C]]
+func.func @addf(%arg0: memref<16x16xf16>, %arg1: memref<16x16xf16>, %arg2: memref<16x16xf16>) {
+  %c0 = arith.constant 0 : index
+  %cst = arith.constant 0.000000e+00 : f16
+  %A = vector.transfer_read %arg0[%c0, %c0], %cst {in_bounds = [true, true]} : memref<16x16xf16>, vector<16x16xf16>
+  %B = vector.transfer_read %arg1[%c0, %c0], %cst {permutation_map = #map0, in_bounds = [true, true]} : memref<16x16xf16>, vector<16x16xf16>
+  %C = arith.addf %A, %B : vector<16x16xf16>
+  vector.transfer_write %C, %arg2[%c0, %c0] {in_bounds = [true, true]} : vector<16x16xf16>, memref<16x16xf16>
+  return
+}