[Mlir-commits] [mlir] [MLIR][NVGPU] Introduce `nvgpu.wargroup.mma.store` Op for Hopper GPUs (PR #65441)

[llvmlistbot at llvm.org]

================
@@ -1141,11 +1150,246 @@ struct NVGPUTmaCreateDescriptorOpLowering
   }
 };
 
+struct NVGPUWarpgroupMmaOpLowering
+    : public ConvertOpToLLVMPattern<nvgpu::WarpgroupMmaOp> {
+  using ConvertOpToLLVMPattern<nvgpu::WarpgroupMmaOp>::ConvertOpToLLVMPattern;
+
+  LogicalResult getWgmmaShape(int64_t sizeM, int64_t sizeN, Type inputElemType,
----------------
qcolombet wrote:

May be interesting to reference NVIDIA's documentation here or in the commit message:
https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#asynchronous-warpgroup-level-matrix-shape

https://github.com/llvm/llvm-project/pull/65441