[Bug-Report|缺陷反馈]: UnsortedSegmentProd算子部分用例对比竞品性能不达标-ops-nn-AtomGit

已关闭

[Bug-Report|缺陷反馈]: UnsortedSegmentProd算子部分用例对比竞品性能不达标 #5636

zhangxiyan7创建于 14 天前关闭于 13 天前

14 天前创建

Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.

UnsortedSegmentProd算子部分用例对比竞品性能不达标

ascend950

ascend950环境下执行所属用例获取性能数据。

换算带宽后实现竞品耗时/NPU耗时 > 0.2，即对比竞品性能达标

部分性能不达标用例如下：

#	用例 / dtype	输入形状(seg数)	GPU耗时(µs)	初始NPU耗时(µs)	初始耗时比
1	000132 fp32/int32	[5410,77,4] (554)	19.35	381.08	0.051
2	000040 bf16/int32	[20,1,8,9,12] (906)	13.32	235.62	0.057
3	000047 bf16/int64	[2,2019,33,2,2,2,2,2] (231)	451.06	12101.55	0.037
4	000116 fp16/int32	[63,5,10044] (834)	28.21	518.22	0.054
5	000062 int32/int32	[3,2,1110,2,2,11,2,2] (224)	14.59	1248.42	0.012
6	000140 fp32/int32	[2,4,4,1,8,5,8] (308)	4.98	76.65	0.065
7	000131 fp32/int64	[60,5717] (262)	6.69	144.32	0.046
8	000163 int32/int64	[101,32383] (853)	70.59	1705.76	0.041
9	000130 fp32/int32	[1605,1054] (724)	620.00	34430.41	0.018
10	000171 int32/int64	[10090,3,3,3,5,3] (822)	490.38	26759.76	0.018
11	000017 fp32/int64	[770] (909)	8.04	1084.10	0.007
12	000007 fp32/int32	[52,9739,4,4] (97)	434.26	19934.99	0.022
13	000134 fp32/int32	[5,28,34,19] (66)	160.63	34359.26	0.005
14	000024 fp32/int32	[20,4,5,21,9] (769)	5.45	198.86	0.027
15	000006 fp16/int32	[68,4,4,13603] (208)	208.56	4464.44	0.047
16	000126 fp16/int32	[136,2,2,2,645,2,2,2] (965)	58.30	1237.38	0.047
17	000046 bf16/int32	[2,2,2,272,2,2,2,63] (406)	143.06	3622.09	0.039
18	000021 fp32/int64	[5,12,3171] (383)	23.28	152.96	0.152
19	000011 fp32/int64	[2,7081,2,2,2,6,3,2] (632)	57.57	4150.62	0.014
20	000037 bf16/int64	[145,4,8254] (1000)	78.42	1037.15	0.076
21	000135 fp32/int64	[5,5,4,41] (839)	45.12	3645.09	0.012
22	000146 bf16/int32	[23,47] (269)	14.94	480.06	0.031
23	000017 fp32/int64	[9718] (954)	44.76	2030.42	0.022
24	000056 int32/int32	[4,4,17,4,6535] (396)	3092.56	209483.03	0.015
25	000227 int32/int64	[116,317,10,446] (788)	1254.56	83651.63	0.015