Rewrite reduce mean/variance

[pfultz2]

Rewrites mean/variance to use reduce_mean(x) and reduce_mean(x*x) so it can be fused in the same reduction.

[migraphx-bot]

Test	Batch	Rate new 52fe8e	Rate old ee68f7	Diff	Compare
torchvision-resnet50	64	2,824.97	2,821.87	0.11%	✅
torchvision-resnet50_fp16	64	6,405.88	6,407.36	-0.02%	✅
torchvision-densenet121	32	2,092.78	2,096.07	-0.16%	✅
torchvision-densenet121_fp16	32	3,703.81	3,687.09	0.45%	✅
torchvision-inceptionv3	32	1,604.11	1,605.13	-0.06%	✅
torchvision-inceptionv3_fp16	32	2,555.85	2,551.45	0.17%	✅
cadene-inceptionv4	16	718.11	717.52	0.08%	✅
cadene-resnext64x4	16	680.69	680.72	-0.00%	✅
slim-mobilenet	64	5,947.46	5,944.92	0.04%	✅
slim-nasnetalarge	64	154.18	154.07	0.07%	✅
slim-resnet50v2	64	2,589.24	2,583.14	0.24%	✅
bert-mrpc-onnx	8	920.32	921.72	-0.15%	✅
bert-mrpc-tf	1	396.61	397.23	-0.16%	✅
pytorch-examples-wlang-gru	1	403.27	395.90	1.86%	✅
pytorch-examples-wlang-lstm	1	428.73	374.49	14.48%	🔆
torchvision-resnet50_1	1	608.81	609.69	-0.14%	✅
cadene-dpn92_1	1	390.69	390.37	0.08%	✅
cadene-resnext101_1	1	332.01	333.18	-0.35%	✅
onnx-taau-downsample	1	306.56	307.34	-0.25%	✅
dlrm-criteoterabyte	1	28.87	28.87	0.02%	✅
dlrm-criteoterabyte_fp16	1	48.27	48.32	-0.10%	✅
agentmodel	1	7,222.46	7,330.30	-1.47%	✅
unet_fp16	2	57.55	57.54	0.02%	✅
resnet50v1_fp16	1	906.34	911.10	-0.52%	✅
resnet50v1_int8	1	816.74	813.46	0.40%	✅
bert_base_cased_fp16	64	1,034.77	1,033.99	0.08%	✅
bert_large_uncased_fp16	32	300.48	300.56	-0.03%	✅
bert_large_fp16	1	159.95	160.04	-0.05%	✅
distilgpt2_fp16	16	1,854.32	1,854.56	-0.01%	✅
yolov5s	1	475.96	474.48	0.31%	✅
tinyllama	1	32.99	32.97	0.04%	✅
vicuna-fastchat	1	157.86	158.01	-0.10%	✅
whisper-tiny-encoder	1	348.54	346.15	0.69%	✅
whisper-tiny-decoder	1	396.66	396.01	0.16%	✅

Check results before 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

[TedThemistokleous]

Might be a slight nit-pick in naming but make "reduces" into something like "reductions" Just makes this a but more readable and then be clearer when you grab a reduce from the list of reductions.

Same goes for "preduce" make this into parallel_reductions since this might get interpreted as pointer-to-reduce with the name preduce.

[pfultz2]

I use the name reduce since the operator name is reduce. It seems confusing to call it reduction when the operator is not named reduction. I guess i can make these single letters and then you can interpret it however makes sense for you.

[TedThemistokleous]

This test case is good and I'm sure these can be massaged to create the other tests to handle cases where matching fails and some of the other coverage error warnings

[TedThemistokleous]

• Add additional test coverage rewrite_reduce.cpp
• Readability for prepare_reduce.cpp

[TedThemistokleous]

Looks good.