[Mlir-commits] [mlir] [MLIR][XeGPU] Add MXFP fp8 WG GEMM tests and use shared input set (PR #225944)

[Sang Ik Lee]

================
@@ -0,0 +1,254 @@
+// RUN: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup zebin-chip=cri"
+// RUN-DISABLED: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup zebin-chip=cri" \
+// RUN-DISABLED: | mlir-runner \
+// RUN-DISABLED:   --shared-libs=%mlir_levelzero_runtime \
+// RUN-DISABLED:   --shared-libs=%mlir_runner_utils \
+// RUN-DISABLED:   --shared-libs=%mlir_c_runner_utils \
+// RUN-DISABLED:   --entry-point-result=void \
+// RUN-DISABLED: | FileCheck --check-prefix=MISMATCH %s
+
+// Non-quantized bf16 counterpart of simple_mxfp_gemm_F4.mlir, for comparing a
+// quantized (dpas_mx) gemm against a plain dpas gemm at the same problem size.
+//
+// Identical to simple_mxfp_gemm_F4.mlir in:
+//   - problem size            M = 256, N = 256, K = 4096, C is f32
+//   - dispatch                blocks (8, 8, 1), threads (64, 1, 1)
+//   - workgroup tile of C     32x32, sg_layout = [2, 2], sg_data = [16, 16]
+//   - input set               the shared one built in @main below
+//   - reference result        host f32 GEMM over the same operand values
+//
+// Deliberately different: the K step is 256 rather than 1024. The mxfp kernel
----------------
silee2 wrote:

32 is optimum value but used K step closer to value used for MXFP usage cases. Will update in the future when MXFP test cases are updated.

https://github.com/llvm/llvm-project/pull/225944