[20190806]Atlas300性能实验

8/6/2019 - 10:47 AM

[20190806]Atlas300性能实验

20200306.md

Rendered
Source

像素	类型	NPU	进程数(线程数)	Batch	AI-CPU使用率（%）	ctrl-CPU使用率(%)	NPU使用率(%)	NPU显存(M)	带宽（%）	时间(ms)	每天处理图片数量(张）	cpu使用率（%）	内存使用（M）
200W	JPEG	1	1(8)	16	30-45	38-50	50	68	40-60	50.61		30	26000
	JPEG	1	1(7)	16				59		45.17			22000
	JPEG	1	1(6)	16	25-35	30-90	50	56	25-60	39.86
	JPEG	1	1(5)	16	32-36	25-35	50	56	30-50	29.56
	JPEG	1	1(3)	16	15-20	12-15	36	53	20-40	18.38
	H264		1(1)							41.75	不编码40.93/AiToMat删除后37.50
	H264	1	1(3)	16	15-40	3-40	20-36	56	5-54	47.47		150-250
	H264	1	1(6)	16						73.64
	H264	1	1(8)	16	15-30	89-95	48-50	76	50-60	147.22	不编码135.21/AiToMat删除后139.59	170-250	2500

原始预处理

PreProcess_Avg:Avg. of 3116 loops: 0.650000 ms
PreProcess_Normal_:Avg. of 4357 loops: 0.185000 ms
PreProcess_PlateRot_:Avg. of 79 loops: 2.083000 ms
PreProcess_Plate1v1_:Avg. of 157 loops: 0.581000 ms
PreProcess_Plate1v2_:Avg. of 1081 loops: 0.865000 ms

Dvpp预处理

PreProcess_Avg:Avg. of 3116 loops: 0.481000 ms
PreProcess_Normal_:Avg. of 4357 loops: 0.179000 ms
PreProcess_PlateRot_:Avg. of 79 loops: 2.120000 ms
PreProcess_Plate1v1_1v2_:Avg. of 1238 loops: 0.427000 ms

20191030.md

Rendered
Source

像素	类型	NPU	进程数(线程数)	Batch	NPU使用率(%)	NPU显存(M)	时间(ms)	每天处理图片数量(张）
200W	H264	1	1(6)	4			46.9279	11046733
			1(5)	8			35.53565	12156806
			1(4)	16			22.9951	15029288
	JPEG		1(8)	16			125	5529600
	H264		1(8)	16			175	3949714
20191226
200W	JPEG	4	1(9)		卡口目标少速度比视频目标多快50%，1.93个目标	8.342ms 改host为DMalloc申请buf后貌似有提速	单核9.162	37721972
500w			1(7)		6.44个目标		17.235	20051842
700w			1(5)		7.43个目标		20.672	16718322
200W	H264	4	1(9)		包括回传编码跟踪等		22.086	15648197
			1(9)		去掉回传编码跟踪	比上面速度快20%	17.918	19287894
			1(9)		目标少比目标多视频快35%		14.25	24251220

20190814.md

Rendered
Source

像素	类型	NPU	进程数(线程数)	Batch	ARM使用率(%)	NPU使用率(%)	NPU显存(M)	算法时间(ms)	解码时间（ms)	每天处理图片数量(张）	cpu使用率（%）	内存使用（M）
200万	1-3车	ascend310	1(host:18,device:46)	4	22	75(+-20)	1656	20.388*2	3.649	4237787/2	250	1585
			2		22*2	80	1761*2	23.026*2	4.384	7504560/2	225*2	1602+1341
		跑到1000多张崩了	3		22*3	115	1545*3	29.831*2	6.336	8688947/2	220*3	1484*3
上面的统计有误时间统一要乘以2
20190816		main中sendData改为200张循环一次	1	4	35-45	130	1000(参考，内存会上升)	23.736	3.649	3640040	60	517
			2	4	33*2		1000*2	35.256	5.512	4901293		522
			3	4	28*3		1000	52.755	6.411	4913278
20190817		发4张返回4张结果	8		(5-10)*8	40-120	246*8	135.867	3.6	5087328	(5-10)*8	513*8
20190820		不返回decodeBuf	8					103.608		6671299
20190823		8k内存池	5	4		(5-10)*8	639	64.446		6703286
			1	4				27.972		3088803
			8	4			246	93.261		7411458
			1	8				26.217		3295571
		4k内存池、内存管理	7	8			433	78.433		7711040		531
		fp16	5	16				44.115		9792657		519

		int8	6	16			574	43.941		11797571		519
		int8	6	16			559.8	42.667		12149816		527
500万
700万	5车	int8	6	16			569	47.719		10863617		541

20190823.md

Rendered
Source

模块（单进程）	batch=4耗时	Batch=8耗时	Batch=16耗时
DvppJpegDecode	2.878*4	2.873*8	2.859
ObjectDetectStage1_v3_Input	0.028	0.036	0.036
ObjectDetectStage1_v3_PreProcess	6.018	11.945	11.985
ObjectDetectStage1_v3_Predict	24.847	48.648	97.002
ObjectDetectStage1_v3_GetLayer	0.743	0.77	0.771
ObjectDetectStage1_v3_PostProcess	0.664*4	0.661*8	0.661
ObjectDetectStage1_v3_Output	0.011*4	0.011*8	0.01
VehicleDetectStage2_Input	0.028	0.034	0.034
VehicleDetectStage2_Index	0.005	0.007	0.009
VehicleDetectStage2_PreProcess	0.842	1.563	2.643
VehicleDetectStage2_Predict	4.516	7.334	13.754
VehicleDetectStage2_GetLayer	0.206	0.201	0.188
VehicleDetectStage2_PostProcess	0.172*4	0.163*8	0.159
VehicleDetectStage2_Output	0.003*4	0.003*8	0.021
VehicleDetectStage2_Segment_Index	0.005	0.009	0.01
VehicleDetectStage2_Segment_PreProcess	0.354	0.578	0.544
VehicleDetectStage2_Segment_Predict	0.784	0.947	1.02
VehicleDetectStage2_Segment_Output	0.002	0.002	0.002
Vehicle_Input	0.023	0.027	0.027
Vehicle_Index	0.002	0.005	0.006
Vehicle_PreProcess	0.759	1.526	2.741
Vehicle_Predict	1.555	2.247	3.327
Vehicle_Classification	0.011*4	0.02*8	0.021
Vehicle_ExtractFeature	0.008*4	0.016*8	0.018
Vehicle_FeatureCode	0.487	0.996	1.084
VehicleDriver_Input	0.025	0.031	0.028
VehicleDriver_Index	0.006	0.01	0.012
VehicleDriver_PreProcess	0.289	0.354	0.323
VehicleDriver_Predict	4.38	6.736	11.514
VehicleDriver_ArgMax	0.001*4	0.001*8	0.001
VehicleDriver_Output	0.001*4	0.001*8	0.001
VehiclePlate_Input	0.024	0.03	0.027
VehiclePlateAlign_Index	0.063	0.066	0.07
VehiclePlateAlign_PreProcess	0.644	0.646	0.934
VehiclePlateAlign_Predict	1.614	2.774	6.691
VehiclePlateAlign_Output	0.001*4	0.001*8	0.001
VehiclePlate_ProcessPlate	9.818	16.827	17.364
VehiclePlate_Index	5.442	10.142	10.578
VehiclePlate_CropToMat	0.186	0.152	0.154
VehiclePlate_ImageRotate	1.21	1.178	1.251
VehiclePlate_PreProcess	0.05	0.098	0.163
VehiclePlate_Predict	1.461	2.154	3.594
VehiclePlate_Output	0.094	0.172	0.279
VehiclePlate_PostProcess	0.054	0.098	0.1
VehicleSpecial_Input	0.019	0.019	0.018
VehicleSpecial_Index	0.018	0.026	0.027
VehicleSpecial_PreProcess	1.127	2.012	2.71
VehicleSpecial_Predict	2.053	3.078	5.366
VehicleSpecial_GetLayer	0.022	0.02	0.019
VehicleSpecial_PostProcess	0.039*4	0.035*8	0.035
VehicleSpecial_Output	0.001*4	0.001*8	0.001
其他（框架消耗？）	111.888-76.774=35.114

20190806.md

Rendered
Source

网络	耗时(batch=1)(3.4GB)	耗时(Batch=4)
ObjectDetectStageCentG320x384	18.99000
ObjectDetectStageV160x128	4.001000
VehiclePlateSegmentCH	0.879000
VehicleDriverGeneral	3.446000
VehiclePlateNo	2.013000
VehiclePlateAlignCH	1.655000
VehiclePlateNameCH	1.576000
VehiclePlateExceptionPlate	1.012000
VehiclePlateExceptionHead	0.895000
VehicleLabel	2.778000
VehicleColor	1.679000
VehicleType	1.083000
ObjectDetectStageT160x96	1.498000
Graph整体运行时间	60ms

Cacher is the code snippet organizer for pro developers

We empower you and your team to get more done, faster

[20190806]Atlas300性能实验