INT8 & FP32 AiiOS ! 驱动万物智能 Alios TVM @ ARM CPU INT8 * Cache 芍四 Data FO Data FOData … QNNPACK Convolution 。，NHWC layout Cach， 浆百 FeU Cach- 区下 。， im2col + pack -35 1 129 中131 124有23152136 2 1.14 am omo oo Convolution Workload Performance AiOS 1驱动万物智能 Alios TVM @ ARM CPU INT8 Depthwise Convolution 。， NHWC layout 。 Using TVM schedule primitive completely 130 1.35 1.33. 1.15 116 111 09工08 工区 0.77 0.77 | | | Depthwise Convolution Workload Performance Alios TVM @ ARM CPU INT8 Performance Comparison @ rasp 3b+ AARCH64

batch size = 1 • 1.7x speedup on Inception asymmetric quantized model • Mobilenet requires depthwise convolution VNNI schedule • Symmetric model improves the speedup to 2.8x© 2019, Amazon Web Services,

����������� �� �������������������� 1. OpenVino a black box, can not deploy our network(with depthwise conv2d, ) 2. TVM can not only deploy our network, but also get a good performance gain by autotuning

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474 15.7.2 Convolution Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . from the mean. This implies that an alpha value of 0.0 gives the same sort of output as a normal convolution (ie. averaging or smoothing filter), where radius will determine the "strength" of the filter. placed elsewhere. You can find: • The Convolution Matrix filter which lets you build custom filters. • The Dilate filter. • The Erode filter. 15.7.2 Convolution Matrix 15.7.2.1 Overview You can find

complicated? See Convolution Matrix filter and you will understand better. Convolution matrix Overview You can find this filter via the image menu under Filters Generic Convolution Matrix Here domain. Most of filters are using convolution matrix. With the Convolution Matrix filter, if the fancy takes you, you can build a custom filter. What is a convolution matrix? It's possible to get a rough using mathematical tools that only a few ones know. Convolution is the treatment of a matrix by another one which is called "kernel". The Convolution Matrix filter uses a first matrix which is the Image

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 759 17.9.2 Convolution Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 760 17.9.3 Distance from the mean. This implies that an Alpha value of 0.0 gives the same sort of output as a normal convolution (i.e. averaging or smoothing filter), where Radius will determine the “strength” of the filter Generic filters are a catch-all for filters which can’t be placed elsewhere. You can find: • The Convolution Matrix filter which lets you build custom filters. • The Distance Map filter. • The GEGL graph

Unsharp Mask 498 Chapter 33: Generic Filters .......................................... 503 Convolution Matrix 504 Universal Filter 506 User Filter (Adobe Filter Factory Emulator) 507 i m p Us er ’ s M a n u a l Ge ner i c Fi lt er s CONVOLUTION MATRIX Convolution Matrix allows you to create simple custom filters. Convolution Matrix adds together the color values in the 5x5 pixel 0 0 0 0 0 1 1 0 0 0 1 0 -1 0 0 0 -1 -1 0 0 0 0 0 0 Figure 33.1 Embossing with the help of Convolution 506 T h e G i m p Us er ’ s M a n u a l Ge ner i c Fi lt er s its alpha value. Try this out

Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Trademark Effects of Convolution operators using TVM AWS DeepLens Acer aiSage NVIDIA Jetson Nano Speedup 0 2 4 6 8 SSD_MobileNet1

compute. @autotvm.register_ topi_schedule(schedule_conv2d_nchw，pvr, [direct]) convolution def schedule_conv2d_nchw_pvr(cfg, outs):

DPU Processor (xDNNv3) >> 3 ˃ Configurable Overlay Processor ˃ DNN Specific Instruction Set Convolution, Max Pool etc. ˃ Any Network, Any Image Size ˃ High Frequency & High Compute Efficiency ˃