frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

Performance improvement on convolution using SSE : Unrolled loop in add function in THVector.h

Browse Source
This commit is contained in:
Arnaud Chauveur
2012-04-08 10:27:30 -04:00
parent 5a0958d2f6
commit ae19f5ba96
1 changed files with 53 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

84
THVector.h
View File

@ -11,15 +11,15 @@
#ifdef USE_SSE2
#include <emmintrin.h>
#endif
#ifdef USE_SSE3
#include <pmmintrin.h>
#endif
#ifdef USE_SSSE3
#include <tmmintrin.h>
#endif
#if defined (USE_SSE4_2) || defined (USE_SSE4_1)
#include <smmintrin.h>
#endif
@ -41,20 +41,31 @@
#define THDoubleVector_add(y, x, c, n) { \
long i = 0; \
__m128d XMM7 = _mm_set1_pd(c); \
__m128d XMM0,XMM2; \
for (; i<=((n)-2); i+=2) { \
XMM0 = _mm_loadu_pd((x)+i); \
XMM2 = _mm_loadu_pd((y)+i); \
XMM0 = _mm_mul_pd(XMM0, XMM7); \
XMM2 = _mm_add_pd(XMM2, XMM0); \
_mm_storeu_pd((y)+i , XMM2); \
} \
for (; i<(n); i++) { \
y[i] += c * x[i]; \
} \
}
long i = 0; \
__m128d XMM7 = _mm_set1_pd(c); \
__m128d XMM0,XMM1,XMM2; \
__m128d XMM3,XMM4,XMM5; \
for (; i<=((n)-6); i+=6) { \
XMM0 = _mm_loadu_pd((x)+i); \
XMM1 = _mm_loadu_pd((x)+i+2); \
XMM2 = _mm_loadu_pd((x)+i+4); \
XMM3 = _mm_loadu_pd((y)+i); \
XMM4 = _mm_loadu_pd((y)+i+2); \
XMM5 = _mm_loadu_pd((y)+i+4); \
XMM0 = _mm_mul_pd(XMM0, XMM7); \
XMM1 = _mm_mul_pd(XMM1, XMM7); \
XMM2 = _mm_mul_pd(XMM2, XMM7); \
XMM3 = _mm_add_pd(XMM3, XMM0); \
XMM4 = _mm_add_pd(XMM4, XMM1); \
XMM5 = _mm_add_pd(XMM5, XMM2); \
_mm_storeu_pd((y)+i , XMM3); \
_mm_storeu_pd((y)+i+2, XMM4); \
_mm_storeu_pd((y)+i+4, XMM5); \
} \
for (; i<(n); i++) { \
y[i] += c * x[i]; \
} \
}
#define THDoubleVector_diff(z, x, y, n) { \
long i; \
@ -141,20 +152,31 @@
}
#define THFloatVector_add(y, x, c, n) { \
long i = 0; \
__m128 XMM7 = _mm_set_ps1(c); \
__m128 XMM0,XMM2; \
for (; i<=((n)-4); i+=4) { \
XMM0 = _mm_loadu_ps((x)+i); \
XMM2 = _mm_loadu_ps((y)+i); \
XMM0 = _mm_mul_ps(XMM0, XMM7); \
XMM2 = _mm_add_ps(XMM2, XMM0); \
_mm_storeu_ps((y)+i , XMM2); \
} \
for (; i<(n); i++) { \
y[i] += c * x[i]; \
} \
}
long i = 0; \
__m128 XMM7 = _mm_set_ps1(c); \
__m128 XMM0,XMM1,XMM2; \
__m128 XMM3,XMM4,XMM5; \
for (; i<=((n)-12); i+=12) { \
XMM0 = _mm_loadu_ps((x)+i); \
XMM1 = _mm_loadu_ps((x)+i+4); \
XMM2 = _mm_loadu_ps((x)+i+8); \
XMM3 = _mm_loadu_ps((y)+i); \
XMM4 = _mm_loadu_ps((y)+i+4); \
XMM5 = _mm_loadu_ps((y)+i+8); \
XMM0 = _mm_mul_ps(XMM0, XMM7); \
XMM1 = _mm_mul_ps(XMM1, XMM7); \
XMM2 = _mm_mul_ps(XMM2, XMM7); \
XMM3 = _mm_add_ps(XMM3, XMM0); \
XMM4 = _mm_add_ps(XMM4, XMM1); \
XMM5 = _mm_add_ps(XMM5, XMM2); \
_mm_storeu_ps((y)+i , XMM3); \
_mm_storeu_ps((y)+i+4, XMM4); \
_mm_storeu_ps((y)+i+8, XMM5); \
} \
for (; i<(n); i++) { \
y[i] += c * x[i]; \
} \
}
#define THFloatVector_diff(z, x, y, n) { \
long i; \