[Mlir-commits] [mlir] [mlir][gpu] Support extf before contract when converting to MMA ops (PR #91988)

[Quinn Dawkins]

================
@@ -490,3 +490,30 @@ func.func @fold_transpose_into_transfer_read(%alloc: memref<64x128xf16>, %vector
 }
 
 // -----
+
+#map1 = affine_map<(d0, d1, d2) -> (d0, d2)>
+#map2 = affine_map<(d0, d1, d2) -> (d1, d2)>
+#map3 = affine_map<(d0, d1, d2) -> (d0, d1)>
+
+// CHECK-LABEL: func @cast_f16_to_f32_read
+//       CHECK:    %[[A:.+]] = gpu.subgroup_mma_load_matrix {{.+}} {leadDimension = 16 : index} : memref<16x16xf16> -> !gpu.mma_matrix<16x16xf16, "AOp">
+//       CHECK:    %[[C:.+]] = gpu.subgroup_mma_load_matrix {{.+}} {leadDimension = 16 : index} : memref<16x16xf16> -> !gpu.mma_matrix<16x16xf16, "COp">
+//       CHECK:    %[[AE:.+]] = gpu.subgroup_mma_elementwise  extf %[[A]] : (!gpu.mma_matrix<16x16xf16, "AOp">) -> !gpu.mma_matrix<16x16xf32, "AOp">
+//       CHECK:    %[[CE:.+]] = gpu.subgroup_mma_elementwise  extf %[[C]] : (!gpu.mma_matrix<16x16xf16, "COp">) -> !gpu.mma_matrix<16x16xf32, "COp">
----------------
qedawkins wrote:

Is extf/si/ui the only valid unary elementwise operation on A/B fragments? Otherwise is there anything special about extends here?

https://github.com/llvm/llvm-project/pull/91988