code_prob_summary.md

1.

~/InfiniTensor/Learning-CUDA$ make  VERBOSE=true

通过此来通过 nvidia的版本的检测。

/data/C500/Learning-CUDA# LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH make  VERBOSE=true  PLATFORM=metax

通过此来通过沐曦metax的版本的检测。

2.存在的问题和困惑

online softmax FA2的编写部分，出现了e-5级别的误差导致难以使用真正的FA2来去通过编译。

只能使用相对原始一些的attention计算逻辑。对于metax和nvidia都是类似的情况。

Case	Type	Avg Time (ms)	Max Diff	Max Tolerance	Result
1	float	0.452434	0.0000000	0.0000000	Pass
1	half	0.384831	0.0000000	0.0000000	Pass
2	float	0.408454	0.0000000	0.0000000	Pass
2	half	0.445536	0.0000000	0.0000000	Pass
3	float	0.387116	0.0000010	0.0007855	Pass
3	half	0.415170	0.0000000	0.0000000	Pass
4	float	0.450248	0.0000237	0.0001501	Pass
4	half	0.448288	0.0000076	0.0050435	Pass
5	float	0.584307	0.0000048	0.0009049	Pass
5	half	0.557689	0.0019531	0.0231250	Pass
6	float	2.688137	0.0000257	0.0000146	Fail
6	half	1.377309	0.0078125	0.0451562	Pass
7	float	0.391246	0.0000057	0.0008215	Pass
7	half	0.459092	0.0000000	0.0000000	Pass
8	float	0.438399	0.0000057	0.0008215	Pass
8	half	0.443422	0.0000000	0.0000000	Pass
9	float	0.429803	0.0000172	0.0002611	Pass
9	half	0.447457	0.0039062	0.0431250	Pass
10	float	0.558301	0.0000207	0.0002537	Pass
10	half	0.524745	0.0039062	0.0313086	Pass
11	float	0.400683	0.0000019	0.0009848	Pass
11	half	0.471791	0.0000305	0.0052132	Pass
12	float	0.427469	0.0000029	0.0009080	Pass
12	half	0.472397	0.0039062	0.0410937	Pass
13	float	15.444282	0.0000048	0.0008913	Pass
13	half	3.981441	0.0078125	0.0469141	Pass
14	float	63.248356	0.0000133	0.0000123	Fail
14	half	18.820628	0.0078125	0.0470703	Pass