ConvInteger: fix parsing for x_zero_point and w_zero_point

[kahmed10]

• The previous test did not properly expose the bug when parsing ConvInteger found in SD3. Updated the parser and updated the unit test.
• I also updated some function and types for clarity.
• I also needed to add an exception for layout in find_inner_broadcast. The unit test otherwise would not run because the zero points were broadcasted, and applying layout to a 1d tensor with 4d permutation would not work.
• Feel free to suggest better variable names.

[kahmed10]

Some further details on the bug:
The previous logic after checking if x_zp and w_zp were not symmetric was faulty. There was already an implicit multibroadcast in the logic by using add_common_op, so that's why it didn't throw any errors. And the shapes of the test inputs happened to be in a way that the broadcasting rules did not break.

Previous input shapes:

x -> [1,3,5,5]
w -> [1,3,2,2]
conv(x,w) -> [1,1,3,3]

The previous logic would then look at x_zp and find it's not symmetric.
After which it would use add_common_op to do conv(x_zp,w), but that would multibroadcast x_zp to shape [1,3,2,2].
The resulting shape of conv(x_zp,w) would be (1,1,1,1).
ret=conv(x,w)-conv(x_zp,w) -> [1,1,3,3] - [1,1,1,1] -> [1,1,3,3] - [1,1,3,3] technically works.
Similarly, using add_common_op to conv(x,w_zp) would multibroadcast w_zp to shape [1,3,5,5].
The resulting shape of conv(x,w_zp) would also be (1,1,1,1).
ret=ret-conv(x,w_zp) -> [1,1,3,3] - [1,1,1,1] -> [1,1,3,3] - [1,1,3,3] also technically works.
Then the final add_common_op for adding conv(x_zp,w_zp) would also technically work:
ret=ret-conv(x_zp,w_zp) -> [1,1,3,3] - [1,1,1,1] -> [1,1,3,3] - [1,1,3,3]

But now if you have the new input shapes:

x_new -> [2,3,10,10]
w_new -> [4,3,3,3]
conv(x_new,w_w) -> [2,4,8,8]

Now if you tried to use add_common_op for conv(x_zp,w_new), you'd get the following:

ret=conv(x_new,w_new)-conv(z_zp,w_new) -> [2,4,8,8] - [4,4,1,1] -> incompatible shapes

TL;DR: The previous logic of using add_common_op was wrong. The test case just so happened to produce broadcastable shapes. Bad test case did not uncover the bug in the logic.

[TedThemistokleous]

Comment: Good to add the comment here!

[lakhinderwalia]

Thanks for adding the general description, and it helps a lot. And for modifying the test_cases. Approved.

[migraphx-bot]

Test	Batch	Rate new 1ce1f9	Rate old 150105	Diff	Compare
torchvision-resnet50	64	3,253.57	3,254.33	-0.02%	✅
torchvision-resnet50_fp16	64	6,927.88	6,927.70	0.00%	✅
torchvision-densenet121	32	2,454.74	2,454.03	0.03%	✅
torchvision-densenet121_fp16	32	4,182.86	4,166.00	0.40%	✅
torchvision-inceptionv3	32	1,630.27	1,630.12	0.01%	✅
torchvision-inceptionv3_fp16	32	2,714.81	2,716.54	-0.06%	✅
cadene-inceptionv4	16	763.30	763.13	0.02%	✅
cadene-resnext64x4	16	813.38	813.13	0.03%	✅
slim-mobilenet	64	7,463.06	7,460.93	0.03%	✅
slim-nasnetalarge	64	208.67	208.67	-0.00%	✅
slim-resnet50v2	64	3,444.86	3,446.18	-0.04%	✅
bert-mrpc-onnx	8	1,148.90	1,146.26	0.23%	✅
bert-mrpc-tf	1	489.20	481.46	1.61%	✅
pytorch-examples-wlang-gru	1	484.53	470.60	2.96%	✅
pytorch-examples-wlang-lstm	1	454.03	443.44	2.39%	✅
torchvision-resnet50_1	1	807.66	804.59	0.38%	✅
cadene-dpn92_1	1	429.19	430.86	-0.39%	✅
cadene-resnext101_1	1	385.86	386.12	-0.07%	✅
onnx-taau-downsample	1	373.45	372.92	0.14%	✅
dlrm-criteoterabyte	1	33.31	33.33	-0.07%	✅
dlrm-criteoterabyte_fp16	1	52.36	52.63	-0.51%	✅
agentmodel	1	8,700.12	8,589.15	1.29%	✅
unet_fp16	2	58.56	58.37	0.33%	✅
resnet50v1_fp16	1	1,028.12	1,036.76	-0.83%	✅
resnet50v1_int8	1	1,022.10	1,039.07	-1.63%	✅
bert_base_cased_fp16	64	1,180.85	1,182.01	-0.10%	✅
bert_large_uncased_fp16	32	365.23	365.27	-0.01%	✅
bert_large_fp16	1	201.65	202.67	-0.50%	✅
distilgpt2_fp16	16	2,227.31	2,227.86	-0.02%	✅
yolov5s	1	537.79	523.22	2.79%	✅
tinyllama	1	43.60	43.57	0.06%	✅
vicuna-fastchat	1	157.73	173.92	-9.31%	🔴
whisper-tiny-encoder	1	417.40	418.33	-0.22%	✅
whisper-tiny-decoder	1	429.83	430.88	-0.24%	✅

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

[codecov]

Codecov Report

Attention: Patch coverage is 92.30769% with 1 line in your changes missing coverage. Please review.

Project coverage is 92.28%. Comparing base (976ae75) to head (314b6ba).
Report is 1 commits behind head on develop.

Files with missing lines	Patch %	Lines
src/onnx/parse_convolution.cpp	90.90%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##           develop    #3763   +/-   ##
========================================
  Coverage    92.28%   92.28%           
========================================
  Files          519      519           
  Lines        22222    22227    +5     
========================================
+ Hits         20507    20512    +5     
  Misses        1715     1715           

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