Add concat fusions

[pfultz2]

This will fuse the inputs to concat if they are pointwise operators.

[migraphx-bot]

Test	Batch	Rate new 525826	Rate old c3049e	Diff	Compare
torchvision-resnet50	64	2,835.40	nan	nan%	❌
torchvision-resnet50_fp16	64	6,502.28	6,508.68	-0.10%	✅
torchvision-densenet121	32	2,083.41	2,096.89	-0.64%	✅
torchvision-densenet121_fp16	32	3,665.74	3,663.69	0.06%	✅
torchvision-inceptionv3	32	1,598.15	1,597.69	0.03%	✅
torchvision-inceptionv3_fp16	32	2,565.21	2,570.26	-0.20%	✅
cadene-inceptionv4	16	723.06	722.53	0.07%	✅
cadene-resnext64x4	16	690.05	692.47	-0.35%	✅
slim-mobilenet	64	8,326.62	8,340.32	-0.16%	✅
slim-nasnetalarge	64	231.93	230.68	0.54%	✅
slim-resnet50v2	64	2,665.83	2,666.08	-0.01%	✅
bert-mrpc-onnx	8	813.52	812.68	0.10%	✅
bert-mrpc-tf	1	387.87	388.68	-0.21%	✅
pytorch-examples-wlang-gru	1	302.51	307.81	-1.72%	✅
pytorch-examples-wlang-lstm	1	318.59	321.36	-0.86%	✅
torchvision-resnet50_1	1	608.24	600.35	1.31%	✅
torchvision-inceptionv3_1	1	343.47	344.92	-0.42%	✅
cadene-dpn92_1	1	402.30	402.49	-0.05%	✅
cadene-resnext101_1	1	328.54	327.84	0.21%	✅
slim-vgg16_1	1	459.46	459.46	0.00%	✅
slim-mobilenet_1	1	2,178.00	2,148.27	1.38%	✅
slim-inceptionv4_1	1	213.93	214.50	-0.27%	✅
onnx-taau-downsample	1	305.39	305.58	-0.06%	✅
dlrm-criteoterabyte	1	21.62	21.58	0.15%	✅
dlrm-criteoterabyte_fp16	1	40.66	40.69	-0.06%	✅
agentmodel	1	6,037.26	6,010.48	0.45%	✅
unet_fp16	2	54.79	54.81	-0.04%	✅
resnet50v1_fp16	1	931.58	932.09	-0.05%	✅
bert_base_cased_fp16	64	924.67	924.64	0.00%	✅
bert_large_uncased_fp16	32	290.67	290.67	0.00%	✅
bert_large_fp16	1	171.81	171.98	-0.10%	✅
distilgpt2_fp16	16	1,280.85	1,281.64	-0.06%	✅

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

     ✅ torchvision-inceptionv3_1: PASSED: MIGraphX meets tolerance

     ✅ cadene-dpn92_1: PASSED: MIGraphX meets tolerance

     ✅ cadene-resnext101_1: PASSED: MIGraphX meets tolerance

     ✅ slim-vgg16_1: PASSED: MIGraphX meets tolerance

     ✅ slim-mobilenet_1: PASSED: MIGraphX meets tolerance

     ✅ slim-inceptionv4_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: PASSED: MIGraphX meets tolerance

     ✅ bert_large: PASSED: MIGraphX meets tolerance

🔴distilgpt2_fp16: FAILED: MIGraphX is not within tolerance - check verbose output

[kahmed10]

Is the old license stamper script being run? The date should be 2023

[kahmed10]

same as the other file

[kahmed10]

same as the other file

[kahmed10]

same as the other file

[TedThemistokleous]

Asking the question here more out of curiosity. I assume order doesn't matter here for fuse_pointwise vs fuse_concat? Is there a benefit to swapping order ever around the fuse_reduce?

[pfultz2]

It needs to run after fuse_pointwise otherwise there will be no pointwise modules to fuse with.

Is there a benefit to swapping order ever around the fuse_reduce?

I dont think so. In either case we will have two kernels that need to be run if these two passes overlap.

[TedThemistokleous]

Fix CI but otherwise I get the idea.