[Mlir-commits] [mlir] [mlir][nvgpu] Add negative tests for invalid FP8 warpgroup MMA types (PR #208047)

[llvmlistbot at llvm.org]

https://github.com/Vaisman created https://github.com/llvm/llvm-project/pull/208047

Add negative tests for invalid FP8 type combinations in
nvgpu.warpgroup.mma.

FP8 inputs are supported for f16/f32 accumulators, but the verifier
should still reject unsupported combinations such as FP8 accumulators
and mixed FP8/non-FP8 input element types.

Tested:
  llvm-lit mlir/test/Dialect/NVGPU/invalid.mlir

>From a5371bfed110fea735f7ee892425331543a8fa5a Mon Sep 17 00:00:00 2001
From: Vasili Svirski <vasili.svirski at gmail.com>
Date: Mon, 6 Jul 2026 12:37:28 +0200
Subject: [PATCH] [mlir][nvgpu] Add negative tests for invalid FP8 warpgroup
 MMA types

Add verifier tests for unsupported FP8 type combinations in
nvgpu.warpgroup.mma.

FP8 inputs are supported for f16/f32 accumulators, but the verifier
should still reject unsupported combinations such as FP8 accumulators
and mixed FP8/non-FP8 input element types.

Tested:
  llvm-lit mlir/test/Dialect/NVGPU/invalid.mlir
---
 mlir/test/Dialect/NVGPU/invalid.mlir | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/mlir/test/Dialect/NVGPU/invalid.mlir b/mlir/test/Dialect/NVGPU/invalid.mlir
index 0341a08a2c778..49c471c881350 100644
--- a/mlir/test/Dialect/NVGPU/invalid.mlir
+++ b/mlir/test/Dialect/NVGPU/invalid.mlir
@@ -465,3 +465,25 @@ func.func @warpgroup_mma_store_mismatched_shape(
       to memref<64x64xf32, 3>
   return
 }
+
+// -----
+
+!tResult = !nvgpu.warpgroup.accumulator<fragmented = vector<128x128xf8E4M3FN>>
+!tDescA  = !nvgpu.warpgroup.descriptor<tensor = memref<128x64xf8E4M3FN, 3>>
+!tDescB  = !nvgpu.warpgroup.descriptor<tensor = memref<64x128xf8E4M3FN, 3>>
+func.func @warpgroup_mma_unsupported_fp8_accumulator(%descA: !tDescA, %descB: !tDescB, %acc: !tResult) {
+  // expected-error @+1 {{'f8E4M3FN' += 'f8E4M3FN' * 'f8E4M3FN'}}
+  %0 = nvgpu.warpgroup.mma %descA, %descB, %acc: !tDescA, !tDescB, !tResult -> !tResult
+  return
+}
+
+// -----
+
+!tResult = !nvgpu.warpgroup.accumulator<fragmented = vector<128x128xf32>>
+!tDescA  = !nvgpu.warpgroup.descriptor<tensor = memref<128x64xf8E4M3FN, 3>>
+!tDescB  = !nvgpu.warpgroup.descriptor<tensor = memref<64x128xf16, 3>>
+func.func @warpgroup_mma_unsupported_mixed_fp8_f16_inputs(%descA: !tDescA, %descB: !tDescB, %acc: !tResult) {
+  // expected-error @+1 {{'f32' += 'f8E4M3FN' * 'f16'}}
+  %0 = nvgpu.warpgroup.mma %descA, %descB, %acc: !tDescA, !tDescB, !tResult -> !tResult
+  return
+}