[mlir][nvgpu]add dim check test to nvgpu.mma op. (#122864)

add shape checks of matrixA, matrixB, and matrixC to the nvgpu.mma's verify.

Author: linuxlonelyeagle

mlir/lib/Dialect/NVGPU/IR/NVGPUDialect.cpp

@@ -203,6 +203,18 @@ static LogicalResult verifyMmaSyncOp(Operation *op,
   // Basic verification
   //
 
+  if (aShape.size() != 2) {
+    return op->emitError() << "matrixA must be 2 dimensional vector";
+  }
+
+  if (bShape.size() != 2) {
+    return op->emitError() << "matrixB must be 2 dimensional vector";
+  }
+
+  if (cShape.size() != 2) {
+    return op->emitError() << "matrixC must be 2 dimensional vector";
+  }
+
   auto [m, n, k] = mmaShape;
 
   // verify warp-wide size for vector a

mlir/test/Dialect/NVGPU/invalid.mlir

@@ -354,3 +354,27 @@ func.func @rcp_unsupported_ftz(%in : vector<16xf32>) {
   // expected-error @+1 {{'nvgpu.rcp' op has a limitation. #nvgpu<rcp_rounding_mode approx> or non-ftz is not supported yet.}}
   %out = nvgpu.rcp %in {rounding = approx} : vector<16xf32>
 }
+
+// -----
+
+func.func @check_matrixA_dim(%arg0: vector<16xf16>, %arg1: vector<2x2xf16>, %arg2: vector<2x2xf16>) -> vector<2x2xf16> {
+  // expected-error @+1 {{matrixA must be 2 dimensional vector}}
+  %d = nvgpu.mma.sync (%arg0, %arg1, %arg2) {mmaShape = [16, 8, 16]} : (vector<16xf16>, vector<2x2xf16>, vector<2x2xf16>) -> vector<2x2xf16>
+  return %d : vector<2x2xf16>
+}
+
+// -----
+
+func.func @check_matrixB_dim(%arg0: vector<4x4xf16>, %arg1: vector<4xf16>, %arg2: vector<2x2xf16>) -> vector<2x2xf16> {
+  // expected-error @+1 {{matrixB must be 2 dimensional vector}}
+  %d = nvgpu.mma.sync (%arg0, %arg1, %arg2) {mmaShape = [16, 8, 16]} : (vector<4x4xf16>, vector<4xf16>, vector<2x2xf16>) -> vector<2x2xf16>
+  return %d : vector<2x2xf16>
+}
+
+// -----
+
+func.func @check_matrixC_dim(%arg0: vector<4x4xf16>, %arg1: vector<2x2xf16>, %arg2: vector<4xf16>) -> vector<2x2xf16> {
+  // expected-error @+1 {{matrixC must be 2 dimensional vector}}
+  %d = nvgpu.mma.sync (%arg0, %arg1, %arg2) {mmaShape = [16, 8, 16]} : (vector<4x4xf16>, vector<2x2xf16>, vector<4xf16>) -> vector<2x2xf16>
+  return %d : vector<2x2xf16>
+}