[X86] improve conv+bias+act profiler #6704

chenjiaoAngel · 2021-08-17T04:48:14Z

1x1s1p0	1,3,112,112; out=32	1,32,112,112; out=32	1,32,224,224; out=32
blas.gemm	0.92254	1.01716	4.32123
now_blas.gemm	0.25896	0.34941	1.89072
improve	256.25%	191.11%	128.55%

…into x86_sgemm

paddle-bot-old · 2021-08-17T04:48:16Z

Thanks for your contribution!

zhangjun · 2021-08-23T03:38:29Z

lite/kernels/x86/conv_compute.cc

+    col_data = static_cast<float*>(
+        TargetMalloc(TARGET(kX86), col_size * sizeof(float)));


后面可以考虑复用下

好，这样可以减少内存占用

chenjiaoAngel added 3 commits August 16, 2021 21:16

improve conv_compute profiler. test=develop

576b043

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle-Lite …

2fba2ed

…into x86_sgemm

fix im2col compute error. test=develop

ea2c10a

chenjiaoAngel added 4 commits August 17, 2021 15:08

remove unsed var. test=develop

5d0875c

fix profiler error. test=develop

0b56632

add gemv implementation. test=develop

2cf7e1d

fix im2col_s1 error. test=develop

b8fef5c

zhangjun approved these changes Aug 23, 2021

View reviewed changes

chenjiaoAngel merged commit 5f77258 into PaddlePaddle:develop Aug 23, 2021

Provide feedback

		col_data = static_cast<float*>(
		TargetMalloc(TARGET(kX86), col_size * sizeof(float)));