Check SSE/AVX/AVX2 at runtime to decline in number of release binaries

[gangliao]

Currently, we have four release binaries, no-avx CPU, avx-CPU, no-avx-GPU, avx-GPU., which looks quite miscellaneous. Actually, we can check CPU capabilities at runtime via

1. query cpuid

#ifdef _WIN32
/// MSVC CPUID
#define cpuid(info, x)    __cpuidex(info, x, 0)
#else
//  GCC Intrinsics
#include <cpuid.h>
/// gcc / clang CPUID
#define cpuid(info, x) __cpuid_count(x, 0, info[0], info[1], info[2], info[3])
#endif

2. Then, do some bit manipulation to check whether AVX is supported or not.

That can help us decline in number of release binaries from four to two.

• Add SIMD flags source code in paddle/utils Add SIMD flags for runtime check #800
• change macro #if in current PaddlePaddle to support AVX and SSE at runtime

[gangliao]

Reference

SIMD Extensions http://paginapessoal.utfpr.edu.br/gortan/aoc/transparencias/t09_mmx-simd/literatura/Wiki_SIMD_Extensions.pdf

[reyoung]

@gangliao To query CPUID will detect whether CPU supports AVX instincts during runtime. But Paddle code needs to be refactored to switch calculation between AVX and SSE.

The code in SIMDFunction.h and CUDA package only export one symbol for both AVX and SSE. Has this been done? Is there any plan?

[hedaoyuan]

这个可以搞定这个issue后在来做，不只是SIMDFunction.h和CUDA的问题。

1. SIMDFunction.h里面很多不用的函数可以删掉了
2. hl_cpu_lstm.cuh这种文件，把avx和非avx代码拆成两个文件分别编译，avx的加-mavx，符号名本来就是不同的两个。
3. BaseMatrix.cu需要考虑一下怎么同时编译avx和非avx两个版本。

[reyoung]

Maybe the first step is to add a cpuid.h/cpp in util package.

[gangliao]

@reyoung @hedaoyuan Yeah, util/simd.h almost done.
It will detect all CPUID flags related to SIMD for general-purpose utilities (SSE2, SSE3, SSSE3, AVX, AVX2, AVX512) to support switch calculation.

[typhoonzero]

I think this was already done.

Closing this issue, feel free to reopen it.