[Mhlo] BatchNorm & Concat #1539
Conversation
|
Can one of the admins verify this patch? |
|
Can one of the admins verify this patch? |
| // RUN: onnx-mlir-opt --convert-onnx-to-mhlo %s -split-input-file | FileCheck %s | ||
|
|
||
| // Test when output shape is unkown | ||
| func @test_concat_unknown_dims(%arg0 : tensor<5x5x1x32xf32>, %arg1 : tensor<5x5x3x32xf32>) -> tensor<?x?x?x?xf32> { |
There was a problem hiding this comment.
Why is this unknown? The shape could be inferred?
What you want to demo is <5x5x?x32> , <5x5x?x32> -> <5x5x?x32> ?
There was a problem hiding this comment.
Should we call --shape-inference before --convert-onnx-to-mhlo? It depends on whether MHLO ops require ranked tensors not not.
|
@jenkins-droid test this please |
| // RUN: onnx-mlir-opt --convert-onnx-to-mhlo %s -split-input-file | FileCheck %s | ||
|
|
||
| // Test when output shape is unkown | ||
| func @test_concat_unknown_dims(%arg0 : tensor<5x5x1x32xf32>, %arg1 : tensor<5x5x3x32xf32>) -> tensor<?x?x?x?xf32> { |
There was a problem hiding this comment.
Should we call --shape-inference before --convert-onnx-to-mhlo? It depends on whether MHLO ops require ranked tensors not not.
| if (op->getNumResults() < 1) { | ||
| op->emitError() << "ONNXConcatOp Has No Output\n"; | ||
| return failure(); | ||
| } |
There was a problem hiding this comment.
Isn't this catched by the operation verifier?
| return failure(); | ||
| } | ||
| RankedTensorType resultType = | ||
| op->getResult(0).getType().dyn_cast_or_null<RankedTensorType>(); |
There was a problem hiding this comment.
FYI, we have isRankedShapedType in TypeUtilities` for this purpose. Please feel free to use it.
| op->getResult(0).getType().dyn_cast_or_null<RankedTensorType>(); | ||
| if (resultType == nullptr) { | ||
| op->emitError() << "Concat Output Is Not Ranked\n"; | ||
| return failure(); |
There was a problem hiding this comment.
Not specific to this PR, I am a bit concerned about failures during rewrite patterns that can cause the whole compilation pipeline terminated. This would work if all ONNX ops can be lowered to MHLO. However, if an op does not satisfy the condition here, there is no a fallback path so that the op will be lowered to other dialects such as Krnl for CPU or ZHigh for NNPA.
One way to solve this is to use DynamicallyLegalOp in --convert--onnx-to-mhlo pass to selectively add operations into the conversion target. We did this for --convert-onnx-to-zhigh: https://github.com/onnx/onnx-mlir/blob/main/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHigh.cpp#L323
There was a problem hiding this comment.
Thanks for the comment! Will discuss with co-workers about the fallback path for mhlo lowering. @yaochengji
There was a problem hiding this comment.
@tungld Currently -convert-onnx-to-hlo uses applyPartialConversion and it doesn't make ONNX Dialect or any ONNX operations illegal. Therefore it is fine if there're some ONNX ops couldn't be converted to MHLO.
There was a problem hiding this comment.
@yaochengji I see. Thanks for clarifying that!
Signed-off-by: chongsong.chen <[email protected]>
Signed-off-by: chongsong.chen <[email protected]>
Signed-off-by: chongsong.chen <[email protected]>
chenchongsong
force-pushed
the
upstream-main
branch
from
322c4ee to
e3ed189
Compare
|
Can one of the admins verify this patch? |
1 similar comment
|
Can one of the admins verify this patch? |
| assert(op->getNumResults() == 1 && "ONNXConcatOp shoule have 1 result"); | ||
| Type resultType = op->getResult(0).getType(); | ||
| if (!onnx_mlir::isRankedShapedType(resultType)) { | ||
| op->emitError() << "Concat Output Is Not Ranked\n"; |
There was a problem hiding this comment.
I don't think it's an error. Suggest using llvm::dbgs() and DEBUG_TYPE here.
Signed-off-by: chongsong.chen <[email protected]>
|
Can one of the admins verify this patch? |
|
@chenchongsong merged with the latest before starting to make a run, hope its ok with you. |
|
Jenkins Linux amd64 Build #6671 [push] [Mhlo] BatchNorm & Conca... started at 03:42 |
|
Jenkins Linux ppc64le Build #5783 [push] [Mhlo] BatchNorm & Conca... started at 04:45 |
|
Jenkins Linux s390x Build #6687 [push] [Mhlo] BatchNorm & Conca... started at 04:42 |
|
Jenkins Linux s390x Build #6687 [push] [Mhlo] BatchNorm & Conca... passed after 1 hr 43 min |
|
Jenkins Linux ppc64le Build #5783 [push] [Mhlo] BatchNorm & Conca... passed after 1 hr 45 min |
|
Jenkins Linux amd64 Build #6671 [push] [Mhlo] BatchNorm & Conca... passed after 1 hr 59 min |
This PR serves the following purposes:
ONNXBatchNormalizationInferenceModeOpandONNXConcatOpfrom ONNX-MLIR to MHLO.