raspberry pi4B ncnn cpu vulkan benchmark
杨利
env: MANJARO ARM aarch64
commit id: fff16a025d21feb11ae51e86365abd8bfd86e900
2022 Apr 28
写在前面的话：
raspberry os在64位同样也做了测试，历经坎坷，但是收效甚微，而且vulkan驱动不好找。在MANJARO系统使用vulkan更容易一些，这里放出来NCNN在raspberry pi4上的benchnark。先说结论：分配512MB的显存情况下，树莓派GPU+vulkan的表现明显弱于CPU的算力。应该是树莓派的Vulkan支持尚不完善的原因。
另一个vulkan:https://qengineering.eu/install-vulkan-on-raspberry-pi.html
thread 1 cpu:
loop_count = 10
num_threads = 1
powersave = 0
gpu_device = -1
cooling_down = 1
squeezenet min = 87.23 max = 88.53 avg = 87.95
squeezenet_int8 min = 76.35 max = 77.11 avg = 76.71
mobilenet min = 140.10 max = 142.05 avg = 140.97
mobilenet_int8 min = 95.47 max = 95.79 avg = 95.63
mobilenet_v2 min = 101.54 max = 102.40 avg = 101.99
mobilenet_v3 min = 82.33 max = 83.70 avg = 82.99
shufflenet min = 50.17 max = 51.85 avg = 50.94
shufflenet_v2 min = 48.87 max = 49.48 avg = 49.10
mnasnet min = 92.76 max = 93.17 avg = 92.96
proxylessnasnet min = 111.36 max = 112.04 avg = 111.67
efficientnet_b0 min = 178.04 max = 178.42 avg = 178.24
efficientnetv2_b0 min = 202.40 max = 203.09 avg = 202.80
regnety_400m min = 122.74 max = 123.09 avg = 122.91
blazeface min = 15.64 max = 15.91 avg = 15.79
googlenet min = 271.19 max = 272.28 avg = 271.75
googlenet_int8 min = 239.69 max = 241.40 avg = 240.20
resnet18 min = 216.32 max = 217.22 avg = 216.87
resnet18_int8 min = 179.47 max = 179.86 avg = 179.68
alexnet min = 202.26 max = 202.81 avg = 202.54
vgg16 min = 1286.14 max = 1291.54 avg = 1287.90
vgg16_int8 min = 994.59 max = 1002.22 avg = 999.48
resnet50 min = 613.59 max = 628.67 avg = 618.64
resnet50_int8 min = 487.12 max = 489.30 avg = 488.22
squeezenet_ssd min = 201.68 max = 202.58 avg = 202.05
squeezenet_ssd_int8 min = 174.25 max = 176.63 avg = 175.01
mobilenet_ssd min = 280.41 max = 281.18 avg = 280.76
mobilenet_ssd_int8 min = 192.00 max = 192.72 avg = 192.36
mobilenet_yolo min = 631.44 max = 642.08 avg = 635.85
mobilenetv2_yolov3 min = 346.23 max = 347.11 avg = 346.83
yolov4-tiny min = 430.36 max = 432.57 avg = 431.51
nanodet_m min = 118.20 max = 118.70 avg = 118.47
yolo-fastest-1.1 min = 59.48 max = 60.90 avg = 60.00
yolo-fastestv2 min = 49.94 max = 50.71 avg = 50.22
thread 1 gpu with vulkan
[0 V3D 4.2] queueC=0[1] queueG=0[1] queueT=0[1]
[0 V3D 4.2] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[0 V3D 4.2] fp16-p/s/a=1/1/0 int8-p/s/a=1/1/0
[0 V3D 4.2] subgroup=16 basic=1 vote=0 ballot=0 shuffle=0
loop_count = 10
num_threads = 1
powersave = 0
gpu_device = 0
cooling_down = 1
squeezenet min = 308.47 max = 309.27 avg = 308.87
squeezenet_int8 min = 83.22 max = 83.79 avg = 83.47
mobilenet min = 345.16 max = 345.46 avg = 345.27
mobilenet_int8 min = 99.83 max = 101.45 avg = 100.58
mobilenet_v2 min = 244.97 max = 245.23 avg = 245.08
mobilenet_v3 min = 231.20 max = 231.39 avg = 231.28
shufflenet min = 143.88 max = 144.13 avg = 144.01
shufflenet_v2 min = 192.31 max = 192.94 avg = 192.42
mnasnet min = 249.60 max = 249.76 avg = 249.70
proxylessnasnet min = 265.37 max = 265.52 avg = 265.44
efficientnet_b0 min = 374.82 max = 375.16 avg = 374.99
efficientnetv2_b0 min = 625.12 max = 626.87 avg = 625.78
regnety_400m min = 318.96 max = 319.88 avg = 319.25
blazeface min = 52.95 max = 53.52 avg = 53.13
googlenet min = 803.00 max = 803.40 avg = 803.20
googlenet_int8 min = 245.09 max = 247.60 avg = 246.22
resnet18 min = 895.02 max = 896.33 avg = 895.97
resnet18_int8 min = 181.80 max = 183.03 avg = 182.34
alexnet min = 499.71 max = 500.70 avg = 500.26
vgg16 min = 4311.12 max = 4312.92 avg = 4312.02
vgg16_int8 min = 998.62 max = 1003.33 avg = 1001.12
resnet50 min = 2022.53 max = 2023.51 avg = 2023.05
resnet50_int8 min = 490.57 max = 494.13 avg = 492.33
squeezenet_ssd min = 1143.39 max = 1144.38 avg = 1143.78
squeezenet_ssd_int8 min = 177.60 max = 180.83 avg = 179.01
mobilenet_ssd min = 816.07 max = 816.63 avg = 816.43
mobilenet_ssd_int8 min = 195.79 max = 196.89 avg = 196.37
mobilenet_yolo min = 1622.98 max = 1623.30 avg = 1623.19
mobilenetv2_yolov3 min = 808.26 max = 808.45 avg = 808.37
yolov4-tiny min = 1704.12 max = 1704.79 avg = 1704.52
nanodet_m min = 389.94 max = 390.18 avg = 390.07
yolo-fastest-1.1 min = 200.23 max = 200.50 avg = 200.36
yolo-fastestv2 min = 164.09 max = 164.36 avg = 164.19
thread 4 cpu:
loop_count = 10
num_threads = 4
powersave = 0
gpu_device = -1
cooling_down = 1
squeezenet min = 51.93 max = 52.47 avg = 52.14
squeezenet_int8 min = 42.81 max = 43.33 avg = 43.07
mobilenet min = 62.97 max = 66.78 avg = 63.72
mobilenet_int8 min = 36.24 max = 39.39 avg = 36.69
mobilenet_v2 min = 61.20 max = 62.30 avg = 61.62
mobilenet_v3 min = 48.56 max = 75.32 avg = 51.63
shufflenet min = 34.52 max = 54.34 avg = 36.62
shufflenet_v2 min = 27.39 max = 27.79 avg = 27.52
mnasnet min = 52.07 max = 54.51 avg = 52.62
proxylessnasnet min = 54.93 max = 56.66 avg = 55.43
efficientnet_b0 min = 81.97 max = 82.88 avg = 82.32
efficientnetv2_b0 min = 89.38 max = 90.46 avg = 89.88
regnety_400m min = 75.65 max = 76.17 avg = 75.81
blazeface min = 10.88 max = 11.08 avg = 10.98
googlenet min = 129.04 max = 131.39 avg = 129.72
googlenet_int8 min = 106.56 max = 107.41 avg = 106.93
resnet18 min = 152.15 max = 166.36 avg = 158.47
resnet18_int8 min = 85.29 max = 86.35 avg = 85.82
alexnet min = 130.07 max = 132.20 avg = 130.82
vgg16 min = 812.36 max = 1004.54 avg = 903.46
vgg16_int8 min = 437.49 max = 1657.19 avg = 726.93
resnet50 min = 315.49 max = 391.08 avg = 348.88
resnet50_int8 min = 258.68 max = 396.38 avg = 286.31
squeezenet_ssd min = 177.35 max = 242.16 avg = 199.35
squeezenet_ssd_int8 min = 119.77 max = 123.66 avg = 122.09
mobilenet_ssd min = 151.96 max = 176.89 avg = 162.62
mobilenet_ssd_int8 min = 82.95 max = 98.27 avg = 87.34
mobilenet_yolo min = 336.06 max = 364.58 avg = 347.83
mobilenetv2_yolov3 min = 194.39 max = 254.23 avg = 208.75
yolov4-tiny min = 250.72 max = 263.13 avg = 254.51
nanodet_m min = 71.37 max = 72.80 avg = 71.86
yolo-fastest-1.1 min = 47.95 max = 57.21 avg = 49.25
yolo-fastestv2 min = 38.46 max = 38.71 avg = 38.57
thread 4 gpu with vulkan
[0 V3D 4.2] queueC=0[1] queueG=0[1] queueT=0[1]
[0 V3D 4.2] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[0 V3D 4.2] fp16-p/s/a=1/1/0 int8-p/s/a=1/1/0
[0 V3D 4.2] subgroup=16 basic=1 vote=0 ballot=0 shuffle=0
loop_count = 10
num_threads = 4
powersave = 0
gpu_device = 0
cooling_down = 1
squeezenet min = 305.52 max = 306.38 avg = 305.92
squeezenet_int8 min = 44.80 max = 54.47 avg = 46.48
mobilenet min = 342.60 max = 342.87 avg = 342.69
mobilenet_int8 min = 37.23 max = 37.81 avg = 37.42
mobilenet_v2 min = 245.07 max = 247.07 avg = 245.34
mobilenet_v3 min = 230.95 max = 231.27 avg = 231.07
shufflenet min = 143.72 max = 144.89 avg = 144.03
shufflenet_v2 min = 192.21 max = 192.48 avg = 192.31
mnasnet min = 249.26 max = 250.15 avg = 249.53
proxylessnasnet min = 265.09 max = 265.51 avg = 265.26
efficientnet_b0 min = 374.56 max = 376.18 avg = 374.95
efficientnetv2_b0 min = 624.94 max = 637.84 avg = 627.64
regnety_400m min = 318.95 max = 319.99 avg = 319.19
blazeface min = 53.02 max = 53.14 avg = 53.06
googlenet min = 803.82 max = 804.77 avg = 804.07
googlenet_int8 min = 107.19 max = 119.33 avg = 109.47
resnet18 min = 895.64 max = 897.14 avg = 896.53
resnet18_int8 min = 86.94 max = 87.82 avg = 87.40
alexnet min = 499.26 max = 501.15 avg = 500.33
vgg16 min = 4315.99 max = 4317.85 avg = 4316.88
vgg16_int8 min = 412.25 max = 438.12 avg = 418.59
resnet50 min = 2024.29 max = 2025.05 avg = 2024.64
resnet50_int8 min = 223.42 max = 272.70 avg = 230.76
squeezenet_ssd min = 1144.16 max = 1144.95 avg = 1144.46
squeezenet_ssd_int8 min = 112.33 max = 122.58 avg = 114.04
mobilenet_ssd min = 816.72 max = 817.11 avg = 816.89
mobilenet_ssd_int8 min = 77.19 max = 77.76 avg = 77.53
mobilenet_yolo min = 1623.28 max = 1623.88 avg = 1623.53
mobilenetv2_yolov3 min = 808.65 max = 808.88 avg = 808.77
yolov4-tiny min = 1704.79 max = 1706.01 avg = 1705.30
nanodet_m min = 389.68 max = 390.61 avg = 389.88
yolo-fastest-1.1 min = 199.75 max = 200.16 avg = 199.87
yolo-fastestv2 min = 163.96 max = 164.44 avg = 164.05
类似资料：