frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 12:54:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

use fp16<->fp32 intrinsic (#17496)

Browse Source 
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17496

As title.

Reviewed By: hyuen

Differential Revision: D14222907

fbshipit-source-id: d5d6c032e725ca8b52aca2be7401ec3c59f6a242
This commit is contained in:
Jongsoo Park
2019-03-07 02:17:42 -08:00
committed by Facebook Github Bot
parent f8778aef78
commit 1d522598fb
2 changed files with 22 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
caffe2/perfkernels/adagrad_avx.cc
View File

@ -104,9 +104,12 @@ void adagrad_fp16_update_prefetch__avx_f16c(
for (; i < N; ++i) {
float gi = g[i];
float hi = h[i] + gi * gi;
nh[i] = hi;
nw[i] = w[i] + lr * gi / (std::sqrt(hi) + epsilon);
float nhi =
_cvtsh_ss(reinterpret_cast<const unsigned short*>(h)[i]) + gi * gi;
reinterpret_cast<unsigned short*>(nh)[i] = _cvtss_sh(nhi, 0);
float nwi = _cvtsh_ss(reinterpret_cast<const unsigned short*>(w)[i]) +
lr * gi / (std::sqrt(nhi) + epsilon);
reinterpret_cast<unsigned short*>(nw)[i] = _cvtss_sh(nwi, 0);
}
}

16
caffe2/perfkernels/cvtsh_ss_bugfix.h
View File

@ -12,6 +12,7 @@
#if __APPLE_NEED_FIX || __CLANG_NEED_FIX
#include <c10/util/Half.h>
#include <emmintrin.h>
// This version of clang has a bug that _cvtsh_ss is not defined, see
@ -25,6 +26,14 @@ _cvtsh_ss(unsigned short a)
return r[0];
}
static __inline unsigned short
__attribute__((__always_inline__, __nodebug__, __target__("f16c")))
_cvtss_sh(float a, int imm8) {
unsigned short ret;
*reinterpret_cast<at::Half*>(&ret) = a;
return ret;
}
#endif // __APPLE_NEED_FIX || __CLANG_NEED_FIX
#undef __APPLE_NEED_FIX
@ -32,6 +41,7 @@ _cvtsh_ss(unsigned short a)
#ifdef _MSC_VER
#include <c10/util/Half.h>
#include <cstdint>
// It seems that microsoft msvc does not have a _cvtsh_ss implementation so
@ -54,4 +64,10 @@ static inline float _cvtsh_ss(unsigned short x) {
return t1.floatval;
}
static inline unsigned short _cvtss_sh(float x, int imm8) {
unsigned short ret;
*reinterpret_cast<at::Half*>(&ret) = x;
return ret;
}
#endif // _MSC_VER