f18298335152h / conv-relu-pooling-operator

used OpenMap, NEON, Inline Assembly speed on arm architecture platform

speed tech

1. OpenMap
2. NEON
3. Inline Assembly

(NEON Inline Assembly) make use of SMID

operator

1. Relu
2. Convolution2D
3. Pooling2D
4. FuseConvReluPooling

for the detail speed use OpenMap, NEON, Inline Assembly, please checkout code on src

do not use MKL for benchmark because coding on MAC

FLOPS please reference cost time on Mac Pro

1. Convolution2D cost time 21 microseconds:(inputs tensor: 5x5x1)
2. pooling cost time 20 microseconds :(inputs tensor: 5x5x1)
3. Relu cost time 12 microseconds :(inputs tensor: 5x5x1)
4. Fuse cost time 3 microseconds :(inputs tensor: 5x5x3)

FLOPS please reference cost time on Iphone or android

build-android-armv7a-with-neon contain executable file test , please run on arm architecture platform.

build-x86-64 contain executable file test, please run on pc platform such as Mac , Windows, Linux ect.

how to build

open the build.sh file to seeting cmake path for x86 or cross build cmake

1. ./build.sh build for armv7a or x86_64
2. pytorch_test.py can verify the result ,please run test print result compare with pytorch_test.py with the same input (install torch)

cost time will be printed on different platform (time: microseconds)

libintel.a will be builded you can depedency on you project