Fix mod op compile issue on GPU

[gyulaz-htec]

Lifting the failing int8, uint8 and int32 data types to float to workaround the problem described in #2268 (comment)

[migraphx-bot]

Test	Batch	Rate new a27482	Rate old c2167f	Diff	Compare
torchvision-resnet50	64	1,714.12	1,714.50	-0.02%	✅
torchvision-resnet50_fp16	64	3,810.36	3,812.07	-0.04%	✅
torchvision-densenet121	32	1,456.56	1,453.98	0.18%	✅
torchvision-densenet121_fp16	32	2,430.13	2,436.13	-0.25%	✅
torchvision-inceptionv3	32	883.76	883.63	0.02%	✅
torchvision-inceptionv3_fp16	32	1,382.20	1,417.34	-2.48%	✅
cadene-inceptionv4	16	408.08	407.89	0.05%	✅
cadene-resnext64x4	16	413.72	414.01	-0.07%	✅
slim-mobilenet	64	3,822.90	3,825.99	-0.08%	✅
slim-nasnetalarge	64	97.05	97.09	-0.03%	✅
slim-resnet50v2	64	1,651.52	1,651.41	0.01%	✅
bert-mrpc-onnx	8	589.51	590.42	-0.15%	✅
bert-mrpc-tf	1	288.82	280.89	2.82%	✅
pytorch-examples-wlang-gru	1	330.05	333.91	-1.16%	✅
pytorch-examples-wlang-lstm	1	264.58	305.41	-13.37%	🔴
torchvision-resnet50_1	1	434.55	452.26	-3.91%	🔴
cadene-dpn92_1	1	228.70	244.50	-6.46%	🔴
cadene-resnext101_1	1	189.05	188.04	0.54%	✅
onnx-taau-downsample	1	204.11	204.42	-0.15%	✅
dlrm-criteoterabyte	1	21.85	22.30	-2.01%	✅
dlrm-criteoterabyte_fp16	1	40.67	41.62	-2.29%	✅
agentmodel	1	7,250.47	5,802.65	24.95%	🔆
unet_fp16	2	33.74	34.27	-1.54%	✅
resnet50v1_fp16	1	569.50	553.42	2.91%	✅
resnet50v1_int8	1	464.01	463.00	0.22%	✅
bert_base_cased_fp16	64	620.80	623.70	-0.46%	✅
bert_large_uncased_fp16	32	193.81	194.38	-0.29%	✅
bert_large_fp16	1	103.96	104.06	-0.09%	✅
distilgpt2_fp16	16	1,188.43	1,192.82	-0.37%	✅
yolov5s	1	297.35	298.59	-0.41%	✅
tinyllama	1	23.34	23.34	-0.00%	✅
vicuna-fastchat	1	132.18	133.58	-1.06%	✅
whisper-tiny-encoder	1	240.92	240.98	-0.03%	✅
whisper-tiny-decoder	1	245.83	245.92	-0.04%	✅

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

All modified and coverable lines are covered by tests ✅

Project coverage is 91.82%. Comparing base (bceef13) to head (a274820).
Report is 158 commits behind head on develop.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #3086      +/-   ##
===========================================
- Coverage    91.82%   91.82%   -0.01%     
===========================================
  Files          486      486              
  Lines        18991    18990       -1     
===========================================
- Hits         17438    17437       -1     
  Misses        1553     1553              

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

[pfultz2]

We should add a function to do the mod on the GPU correctly instead of falling back to float. Maybe something like:

template <class T, class U>
constexpr auto mod(T a, U b)
{
    if constexpr(is_integral<T>{} and is_integral<U>{})
        return a % b;
    return vec_transform(a, b)([](auto x0, auto x1) { 
        return fmod(remainder(x0, x1) + x1, x1); 
    });
}

[gyulaz-htec]

@umangyadav @pfultz2 I've updated the code according to what @pfultz2 mentioned.
I had to go with ((a % b) + b) % b instead of a % b to match the rounding mode of modulos to what onnx expects, because rounding for modulus in c++ is implementation dependant.