Standalone operator Convolution_backwards for MLIR

[lakhinderwalia]

Motivation

Support Operator Convolution_backwards via MLIR

Technical Details

This support is necessary without MIOPEN, e.g. on Windows platforms.

Changelog Category

• • Added: New functionality.
• • Changed: Changes to existing functionality.
• • Removed: Functionality or support that has been removed. (Compared to a previous release)
• • Optimized: Component performance that has been optimized or improved.
• • Resolved Issues: Known issues from a previous version that have been resolved.
• • Not Applicable: This PR is not to be included in the changelog.

[causten]

Can you add some unit tests

[Copilot]

Pull Request Overview

This PR adds MLIR support for the convolution_backwards operator to provide an alternative implementation when MIOPEN is not available (e.g., on Windows platforms).

• Adds mapping for convolution_backwards operator to MLIR dialect operation
• Implements predicate function to determine when convolution_backwards can use MLIR
• Integrates the new backward convolution matcher into the MLIR fusion pipeline

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/targets/gpu/mlir.cpp	Maps convolution_backwards operator name to MLIR dialect operation
src/targets/gpu/fuse_mlir.cpp	Adds predicate function and matcher for backward convolution operations in MLIR

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[pfultz2]

We do need to limit this to 2d convolution_backwards because I think thats the only one supported right now in MLIR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files

@@           Coverage Diff            @@
##           develop    #4292   +/-   ##
========================================
  Coverage    92.22%   92.22%           
========================================
  Files          557      557           
  Lines        25924    25924           
========================================
  Hits         23908    23908           
  Misses        2016     2016           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[pfultz2]

You need to add tests for fuse_mlir as well.

[TedThemistokleous]

Overall looks good. Have one ask and a question

Ask

• Add a test for batch size greater than one

Question - What do we do for datatypes like fp8/int8/ and smaller or is that not supported/intended for the backwards convolution since that should resolve to quant_convolution

[lakhinderwalia]

Overall looks good. Have one ask and a question

Ask

• Add a test for batch size greater than one

Question - What do we do for datatypes like fp8/int8/ and smaller or is that not supported/intended for the backwards convolution since that should resolve to quant_convolution

This PR is to bypass MIOpen when we cannot use it. The current MLIR support, as reflected here in the PR, is for only 2D kernels, and of only the float types that have been added: fp32, fp16, bf16. There is no MLIR support for fp8 or int8 right now.

[TedThemistokleous]

@lakhinderwalia link updated test to this PR when you upload

[pfultz2]

Overall looks good, just need to add fuse_mlir tests. You can see an example how we check the env variable in the test in #4261.

[lakhinderwalia]

Overall looks good, just need to add fuse_mlir tests. You can see an example how we check the env variable in the test in #4261.

It is not the case of a simple environment variable here. I was looking for an example to handle more of get_mode(), and that is what I was looking for a clarification on. Not sure if we want to replicate all that complexity in a simple test case, @pfultz2. Thanks.

[pfultz2]

Overall looks good, just need to add fuse_mlir tests. You can see an example how we check the env variable in the test in #4261.

It is not the case of a simple environment variable here. I was looking for an example to handle more of get_mode(), and that is what I was looking for a clarification on. Not sure if we want to replicate all that complexity in a simple test case, @pfultz2. Thanks.

convolution_backwards should be added to specific ops in the jenkins file so it gets tested on the CI, and then you can check it on the tests to disabled. Thats the way we currently do it.

Requested changes have been applied

[migraphx-bot]

Test	Batch	Rate new 5bec67	Rate old 43d3f3	Diff	Compare
torchvision-resnet50	64	3,173.58	3,157.73	0.50%	✅
torchvision-resnet50_fp16	64	6,610.52	6,587.32	0.35%	✅
torchvision-densenet121	32	2,445.70	2,437.35	0.34%	✅
torchvision-densenet121_fp16	32	4,132.47	4,122.32	0.25%	✅
torchvision-inceptionv3	32	1,672.88	1,665.47	0.44%	✅
torchvision-inceptionv3_fp16	32	2,594.22	2,589.37	0.19%	✅
cadene-inceptionv4	16	797.71	794.82	0.36%	✅
cadene-resnext64x4	16	806.68	801.96	0.59%	✅
slim-mobilenet	64	8,233.95	8,209.01	0.30%	✅
slim-nasnetalarge	64	222.87	221.76	0.50%	✅
slim-resnet50v2	64	3,304.88	3,295.13	0.30%	✅
bert-mrpc-onnx	8	1,143.42	1,133.98	0.83%	✅
bert-mrpc-tf	1	484.76	490.52	-1.17%	✅
pytorch-examples-wlang-gru	1	313.52	317.20	-1.16%	✅
pytorch-examples-wlang-lstm	1	431.06	439.21	-1.86%	✅
torchvision-resnet50_1	1	799.36	802.49	-0.39%	✅
cadene-dpn92_1	1	453.65	436.73	3.88%	🔆
cadene-resnext101_1	1	369.64	368.11	0.41%	✅
onnx-taau-downsample	1	398.97	398.29	0.17%	✅
dlrm-criteoterabyte	1	32.05	31.92	0.39%	✅
dlrm-criteoterabyte_fp16	1	51.06	51.00	0.11%	✅
agentmodel	1	9,570.39	9,837.19	-2.71%	✅
unet_fp16	2	58.92	58.80	0.20%	✅
resnet50v1_fp16	1	988.06	995.41	-0.74%	✅
resnet50v1_int8	1	1,003.53	995.73	0.78%	✅
bert_base_cased_fp16	64	1,104.20	1,100.48	0.34%	✅
bert_large_uncased_fp16	32	345.70	344.24	0.42%	✅
bert_large_fp16	1	198.46	197.60	0.44%	✅
distilgpt2_fp16	16	2,087.76	2,076.97	0.52%	✅
yolov5s	1	586.29	587.20	-0.15%	✅
tinyllama	1	43.97	43.80	0.38%	✅
vicuna-fastchat	1	45.33	45.12	0.46%	✅
whisper-tiny-encoder	1	411.04	409.76	0.31%	✅
whisper-tiny-decoder	1	415.35	414.17	0.28%	✅
llama2_7b	1	19.16	19.12	0.17%	✅
qwen1.5-7b	1	23.52	23.43	0.38%	✅
phi3-3.8b	1	26.69	26.64	0.17%	✅
mask-rcnn	1	12.10	12.18	-0.65%	✅
llama3-8b	1	21.72	21.67	0.23%	✅
whisper-large-encoder	1	10.22	10.17	0.48%	✅
whisper-large-decoder	1	100.27	100.09	0.19%	✅
mistral-7b	1	23.73	23.65	0.36%	✅
FLUX.1-schnell	1	720.79	728.36	-1.04%	✅
nan	nan	nan	nan	nan%	❌

This build is not recommended to merge 🔴

[migraphx-bot]

     ✅ bert-mrpc-onnx: PASSED: MIGraphX meets tolerance

❌bert-mrpc-tf: ERROR - check error output

2025-09-19 12:05:27.610617: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
Traceback (most recent call last):
File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 359, in
main()
File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 306, in main
graph = load_tf_graph(model_name)
File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 300, in load_tf_graph
graph_def.ParseFromString(f.read())
File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/lib/io/file_io.py", line 116, in read
self._preread_check()
File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/lib/io/file_io.py", line 77, in _preread_check
self._read_buf = _pywrap_file_io.BufferedInputStream(
tensorflow.python.framework.errors_impl.UnimplementedError: File system scheme '[local]' not implemented (file: '/new-saved-models/tf-misc/bert_mrpc1.pb')

     ✅ pytorch-examples-wlang-gru: PASSED: MIGraphX meets tolerance

     ✅ pytorch-examples-wlang-lstm: PASSED: MIGraphX meets tolerance

     ✅ dlrm-criteoterabyte: PASSED: MIGraphX meets tolerance

     ✅ agentmodel: PASSED: MIGraphX meets tolerance

     ✅ unet: PASSED: MIGraphX meets tolerance

     ✅ resnet50v1: PASSED: MIGraphX meets tolerance

     ✅ bert_base_cased_fp16: PASSED: MIGraphX meets tolerance

🔴bert_large_uncased_fp16: FAILED: MIGraphX is not within tolerance - check verbose output

     ✅ bert_large: PASSED: MIGraphX meets tolerance

     ✅ yolov5s: PASSED: MIGraphX meets tolerance

     ✅ tinyllama: PASSED: MIGraphX meets tolerance

     ✅ vicuna-fastchat: PASSED: MIGraphX meets tolerance

     ✅ whisper-tiny-encoder: PASSED: MIGraphX meets tolerance

     ✅ whisper-tiny-decoder: PASSED: MIGraphX meets tolerance

     ✅ distilgpt2_fp16: PASSED: MIGraphX meets tolerance

     ✅ llama2_7b: PASSED: MIGraphX meets tolerance

     ✅ qwen1.5-7b: PASSED: MIGraphX meets tolerance

     ✅ phi3-3.8b: PASSED: MIGraphX meets tolerance

🔴mask-rcnn: FAILED: MIGraphX is not within tolerance - check verbose output

     ✅ llama3-8b: PASSED: MIGraphX meets tolerance

     ✅ whisper-large-decoder: PASSED: MIGraphX meets tolerance

     ✅ mistral-7b: PASSED: MIGraphX meets tolerance

     ✅ FLUX.1-schnell: PASSED: MIGraphX meets tolerance