[Mlir-commits] [mlir] [mlir] GEMM Hopper Tensor Core Integration Test (PR #81478)

[Guray Ozen]

================
@@ -0,0 +1,266 @@
+# RUN: env SUPPORT_LIB=%mlir_cuda_runtime \
+# RUN:   %PYTHON %s | FileCheck %s
+# CHECK: PASS
+
+# ===--- GEMM Hopper Tensor Core Integration Test ---===
+#
+# This test aims to validate the correctness of the supported GEMM kernels in
+# NVGPU dialects, with current support for Multistage and Warp Specialization
+# kernels.
+# The test constructs and metaprograms IR using Python bindings, allowing
+# generic IR building. This flexibility enables changes to the shape,
+# tile size, or data type of the GEMM for testing purposes.
+# The entry function is `matmul`, where one can specify GEMM shape, tile size,
+# data type, GEMM algorithm (Multistage or Warp Specialization), and the maximum
+# number of stages.
+# Verification is done via numpy's matmul operation.
+#
+# Example:
+# matmul(input_type=np.float16,                # input types
+#        output_type=np.float32,               # output type
----------------
grypp wrote:

This makes sense. This is also valid for the input types, we can have mixed types.

However, these things are not supported having different `accumulation_type` so I will leave it as it is. When we have it, we can change the code.

https://github.com/llvm/llvm-project/pull/81478