Implement BF16 using generic_float class

[richagadgil]

Uses generic float class (#3522) to create bf16 class.

BF16 has 1 sign bit, 8 bits for the exponent, and 7 bits for the mantissa: bf16 = migraphx::generic_float<7, 8>;

Summary of changes:

1. generic_float.cpp : Change the subnormal (when exponent==0) conversion to differentiate between FP16 and BF16 types
2. migraphx.h, shape.hpp, hip_gemm_impl.cpp, gemm_impl.cpp: Add BF16 shape type
3. type_traits.hpp: Add traits for BF16 type
4. tests/: Add tests for BF16 type

[TedThemistokleous]

Should this be an issue or is this tracked somewhere already?

[TedThemistokleous]

@richagadgil update your description. I think we've done a pass on this with some comments.

[pfultz2]

For the CI failures, just add an if statement to skip bf16 for now in the compile_math case in jit.cpp. That should get it passing and we can work on enabling it on a later PR.

[migraphx-bot]

Test	Batch	Rate new 47a181	Rate old 0f36aa	Diff	Compare
torchvision-resnet50	64	3,254.36	3,261.99	-0.23%	✅
torchvision-resnet50_fp16	64	6,989.58	6,984.41	0.07%	✅
torchvision-densenet121	32	2,435.66	2,434.46	0.05%	✅
torchvision-densenet121_fp16	32	4,085.94	4,068.77	0.42%	✅
torchvision-inceptionv3	32	1,628.48	1,630.14	-0.10%	✅
torchvision-inceptionv3_fp16	32	2,745.44	2,746.22	-0.03%	✅
cadene-inceptionv4	16	764.67	765.59	-0.12%	✅
cadene-resnext64x4	16	810.97	809.78	0.15%	✅
slim-mobilenet	64	7,467.12	7,474.57	-0.10%	✅
slim-nasnetalarge	64	208.49	208.58	-0.05%	✅
slim-resnet50v2	64	3,442.62	3,441.49	0.03%	✅
bert-mrpc-onnx	8	1,148.81	1,150.80	-0.17%	✅
bert-mrpc-tf	1	466.40	465.54	0.18%	✅
pytorch-examples-wlang-gru	1	419.06	420.06	-0.24%	✅
pytorch-examples-wlang-lstm	1	473.00	381.98	23.83%	🔆
torchvision-resnet50_1	1	761.83	750.44	1.52%	✅
cadene-dpn92_1	1	402.29	398.35	0.99%	✅
cadene-resnext101_1	1	382.31	382.96	-0.17%	✅
onnx-taau-downsample	1	345.79	346.08	-0.08%	✅
dlrm-criteoterabyte	1	33.34	33.35	-0.02%	✅
dlrm-criteoterabyte_fp16	1	52.76	52.68	0.15%	✅
agentmodel	1	8,177.28	8,091.53	1.06%	✅
unet_fp16	2	58.83	58.77	0.09%	✅
resnet50v1_fp16	1	926.85	943.16	-1.73%	✅
resnet50v1_int8	1	994.66	1,012.12	-1.72%	✅
bert_base_cased_fp16	64	1,170.84	1,169.97	0.07%	✅
bert_large_uncased_fp16	32	363.41	363.75	-0.10%	✅
bert_large_fp16	1	198.71	199.03	-0.16%	✅
distilgpt2_fp16	16	2,197.22	2,201.98	-0.22%	✅
yolov5s	1	535.63	539.79	-0.77%	✅
tinyllama	1	43.65	43.42	0.53%	✅
vicuna-fastchat	1	177.06	175.75	0.75%	✅
whisper-tiny-encoder	1	418.16	418.02	0.03%	✅
whisper-tiny-decoder	1	423.43	428.37	-1.15%	✅

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