[MLIR][NVVM] Fixed assertion failure for insufficient parsing validation of nvvm dialect MMAOp #163432
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…ialect with MMAOp
|
@llvm/pr-subscribers-mlir-llvm @llvm/pr-subscribers-mlir Author: Stefan Mada (smada3) ChangesThe nvvm dialect instruction MmaOp will trigger an assertion failure if an invalid combination of row / col layouts for A and B are used, such as in this case:
I added a check to make sure the appropriate layouts are allowed as specified in the operation's description. Full diff: https://github.com/llvm/llvm-project/pull/163432.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
index 5edcc40bd2d32..d785a8918cc4a 100644
--- a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
+++ b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
@@ -787,6 +787,26 @@ LogicalResult MmaOp::verify() {
" attribute");
}
+ // Validate layout combinations. According to the operation description, most
+ // MMA operations require layoutA=row and layoutB=col. Only m8n8k4 with f16
+ // can use other layout combinations.
+ bool isM8N8K4_F16 =
+ (mmaShape[0] == 8 && mmaShape[1] == 8 && mmaShape[2] == 4 &&
+ getMultiplicandAPtxType() == MMATypes::f16);
+
+ if (!isM8N8K4_F16) {
+ // For all other shapes/types, layoutA must be row and layoutB must be col
+ if (getLayoutA() != MMALayout::row || getLayoutB() != MMALayout::col) {
+ return emitOpError("requires layoutA = #nvvm.mma_layout<row> and "
+ "layoutB = #nvvm.mma_layout<col> for shape <")
+ << mmaShape[0] << ", " << mmaShape[1] << ", " << mmaShape[2]
+ << "> with element types "
+ << stringifyEnum(*getMultiplicandAPtxType()) << " and "
+ << stringifyEnum(*getMultiplicandBPtxType())
+ << ". Only m8n8k4 with f16 supports other layouts.";
+ }
+ }
+
return success();
}
diff --git a/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir b/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir
index 0b3615487716d..cf5a9782c69ab 100644
--- a/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir
+++ b/mlir/test/Target/LLVMIR/nvvmir-invalid.mlir
@@ -559,3 +559,16 @@ llvm.func @clusterlaunchcontrol_query_cancel_get_first_cta_id_invalid_return_typ
%res = nvvm.clusterlaunchcontrol.query.cancel query = get_first_cta_id_x, %try_cancel_response : i1
llvm.return
}
+
+// -----
+
+// Test that ensures invalid row/col layouts for matrices A and B are not accepted
+llvm.func @nvvm_mma_m16n8k32_s4_s4(%a0 : i32, %a1 : i32, %b0 : i32, %c0 : i32, %c1 : i32, %c2 : i32, %c3 : i32) -> !llvm.struct<(i32,i32,i32,i32)> {
+ // expected-error@+1 {{Only m8n8k4 with f16 supports other layouts.}}
+ %0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3]
+ {layoutA = #nvvm.mma_layout<col>, layoutB = #nvvm.mma_layout<col>,
+ multiplicandAPtxType = #nvvm.mma_type<s4>, multiplicandBPtxType = #nvvm.mma_type<s4>,
+ intOverflowBehavior=#nvvm.mma_int_overflow<satfinite>,
+ shape = #nvvm.shape<m = 16, n = 8, k = 32>} : (i32, i32, i32) -> !llvm.struct<(i32,i32,i32,i32)>
+ llvm.return %0 : !llvm.struct<(i32,i32,i32,i32)>
+}
\ No newline at end of file
|
|
LGTM, |
schwarzschild-radius
approved these changes
Oct 15, 2025
|
@grypp If you think this change looks good, please merge after approving. I don't have commit permissions. |
grypp
approved these changes
Oct 15, 2025
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/169/builds/16035 Here is the relevant piece of the build log for the reference |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The nvvm dialect instruction MmaOp will trigger an assertion failure if an invalid combination of row / col layouts for A and B are used, such as in this case:
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3] {layoutA = #nvvm.mma_layout<col>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<s4>, multiplicandBPtxType = #nvvm.mma_type<s4>, intOverflowBehavior=#nvvm.mma_int_overflow<satfinite>, shape = #nvvm.shape<m = 16, n = 8, k = 32>} : (i32, i32, i32) -> !llvm.struct<(i32,i32,i32,i32)>I added a check to make sure the appropriate layouts are allowed as specified in the operation's description.