Simplify select_module after simplify_dyn_ops pass

[CharlieL7]

• This pass recalculates the output_dyn_shapes attribute for select_module operators after the rest of simplify_dyn_ops.
• Recalculating the outputs shapes is required if the other matchers change the output shape of the submodules.
  • For example, the Resize operator can output {0, max_int} dynamic dimensions, but will be simplified for dynamic batch to a static shape.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (b30a447) 91.51% compared to head (affd0b8) 91.53%.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2714      +/-   ##
===========================================
+ Coverage    91.51%   91.53%   +0.02%     
===========================================
  Files          468      468              
  Lines        17599    17659      +60     
===========================================
+ Hits         16105    16165      +60     
  Misses        1494     1494              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[pfultz2]

We cant update compute_shape in select_module to compute the correct dynamic shape output?

[CharlieL7]

We cant update compute_shape in select_module to compute the correct dynamic shape output?

We could if we use shape.compute_shape(shape_inputs, module_inputs). Would have to change select_module to remove the output_dyn_shapes attribute and related changes to tests/IR.

[migraphx-bot]

Test	Batch	Rate new affd0b	Rate old 8e1705	Diff	Compare
torchvision-resnet50	64	2,837.70	2,839.27	-0.06%	✅
torchvision-resnet50_fp16	64	6,515.81	6,515.42	0.01%	✅
torchvision-densenet121	32	2,103.59	2,085.14	0.88%	✅
torchvision-densenet121_fp16	32	3,696.81	3,692.28	0.12%	✅
torchvision-inceptionv3	32	1,599.93	1,599.62	0.02%	✅
torchvision-inceptionv3_fp16	32	2,573.36	2,571.11	0.09%	✅
cadene-inceptionv4	16	724.19	724.04	0.02%	✅
cadene-resnext64x4	16	692.84	690.44	0.35%	✅
slim-mobilenet	64	6,881.67	6,900.95	-0.28%	✅
slim-nasnetalarge	64	177.05	177.02	0.01%	✅
slim-resnet50v2	64	2,665.23	2,666.38	-0.04%	✅
bert-mrpc-onnx	8	825.98	826.02	-0.01%	✅
bert-mrpc-tf	1	380.92	381.39	-0.12%	✅
pytorch-examples-wlang-gru	1	240.20	260.12	-7.66%	🔴
pytorch-examples-wlang-lstm	1	244.86	242.40	1.01%	✅
torchvision-resnet50_1	1	608.65	612.15	-0.57%	✅
cadene-dpn92_1	1	393.87	393.25	0.16%	✅
cadene-resnext101_1	1	333.52	333.56	-0.01%	✅
onnx-taau-downsample	1	306.17	306.50	-0.11%	✅
dlrm-criteoterabyte	1	21.57	21.57	0.02%	✅
dlrm-criteoterabyte_fp16	1	40.65	40.70	-0.11%	✅
agentmodel	1	4,025.44	4,368.82	-7.86%	🔴
unet_fp16	2	55.96	55.88	0.14%	✅
resnet50v1_fp16	1	859.56	879.94	-2.32%	✅
resnet50v1_int8	1	800.89	804.54	-0.45%	✅
bert_base_cased_fp16	64	936.35	936.58	-0.02%	✅
bert_large_uncased_fp16	32	292.65	292.69	-0.02%	✅
bert_large_fp16	1	184.15	184.27	-0.06%	✅
distilgpt2_fp16	16	1,642.97	1,642.05	0.06%	✅
yolov5s	1	490.69	495.36	-0.94%	✅
tinyllama	1	32.64	32.66	-0.06%	✅
vicuna-fastchat	1	157.25	159.46	-1.39%	✅
whisper-tiny-encoder	1	335.86	335.70	0.05%	✅
whisper-tiny-decoder	1	373.55	374.96	-0.38%	✅

This build is not recommended to merge 🔴

[migraphx-bot]

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

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

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

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

     ✅ torchvision-resnet50_1: PASSED: MIGraphX meets tolerance

     ✅ cadene-dpn92_1: PASSED: MIGraphX meets tolerance

     ✅ cadene-resnext101_1: 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

[CharlieL7]

Keeping this PR for now and pushing back the refactor to pass_manager for a later date. Tracking the pass_manager refactor as issue #2714 .