mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-08-06 06:26:29 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Optimize exp32f with AVX2

Browse Source
This commit is contained in:
Woody Chow 2017-03-15 17:03:36 +09:00
parent a83a1cafa7
commit a8763c1fec
1 changed files with 90 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

91
modules/core/src/mathfuncs_core.cpp
View File

@ -551,7 +551,96 @@ void exp32f( const float *_x, float *y, int n )
const Cv32suf* x = (const Cv32suf*)_x; const Cv32suf* x = (const Cv32suf*)_x;
Cv32suf buf[4]; Cv32suf buf[4];
#if CV_SSE2 #if CV_AVX2
if( n >= 8 )
{
static const __m256d prescale4 = _mm256_set1_pd(exp_prescale);
static const __m256 postscale8 = _mm256_set1_ps((float)exp_postscale);
static const __m128 maxval4 = _mm_set1_ps((float)(exp_max_val/exp_prescale));
static const __m128 minval4 = _mm_set1_ps((float)(-exp_max_val/exp_prescale));
static const __m256 mA1 = _mm256_set1_ps(A1);
static const __m256 mA2 = _mm256_set1_ps(A2);
static const __m256 mA3 = _mm256_set1_ps(A3);
static const __m256 mA4 = _mm256_set1_ps(A4);
bool y_aligned = (size_t)(void*)y % 32 == 0;
ushort CV_DECL_ALIGNED(32) tab_idx[16];
for( ; i <= n - 8; i += 8 )
{
__m256 xf;
__m128i xi0, xi1;
__m256d xd0 = _mm256_cvtps_pd(_mm_min_ps(_mm_max_ps(_mm_loadu_ps(&x[i].f), minval4), maxval4));
__m256d xd1 = _mm256_cvtps_pd(_mm_min_ps(_mm_max_ps(_mm_loadu_ps(&x[i+4].f), minval4), maxval4));
xd0 = _mm256_mul_pd(xd0, prescale4);
xd1 = _mm256_mul_pd(xd1, prescale4);
xi0 = _mm256_cvtpd_epi32(xd0);
xi1 = _mm256_cvtpd_epi32(xd1);
xd0 = _mm256_sub_pd(xd0, _mm256_cvtepi32_pd(xi0));
xd1 = _mm256_sub_pd(xd1, _mm256_cvtepi32_pd(xi1));
// gcc does not support _mm256_set_m128
//xf = _mm256_set_m128(_mm256_cvtpd_ps(xd1), _mm256_cvtpd_ps(xd0));
xf = _mm256_insertf128_ps(xf, _mm256_cvtpd_ps(xd0), 0);
xf = _mm256_insertf128_ps(xf, _mm256_cvtpd_ps(xd1), 1);
xf = _mm256_mul_ps(xf, postscale8);
xi0 = _mm_packs_epi32(xi0, xi1);
_mm_store_si128((__m128i*)tab_idx, _mm_and_si128(xi0, _mm_set1_epi16(EXPTAB_MASK)));
xi0 = _mm_add_epi16(_mm_srai_epi16(xi0, EXPTAB_SCALE), _mm_set1_epi16(127));
xi0 = _mm_max_epi16(xi0, _mm_setzero_si128());
xi0 = _mm_min_epi16(xi0, _mm_set1_epi16(255));
xi1 = _mm_unpackhi_epi16(xi0, _mm_setzero_si128());
xi0 = _mm_unpacklo_epi16(xi0, _mm_setzero_si128());
__m256d yd0 = _mm256_set_pd(expTab[tab_idx[3]], expTab[tab_idx[2]], expTab[tab_idx[1]], expTab[tab_idx[0]]);
__m256d yd1 = _mm256_set_pd(expTab[tab_idx[7]], expTab[tab_idx[6]], expTab[tab_idx[5]], expTab[tab_idx[4]]);
// gcc does not support _mm256_set_m128
//__m256 yf = _mm256_set_m128(_mm256_cvtpd_ps(yd1), _mm256_cvtpd_ps(yd0));
__m256 yf;
yf = _mm256_insertf128_ps(yf, _mm256_cvtpd_ps(yd0), 0);
yf = _mm256_insertf128_ps(yf, _mm256_cvtpd_ps(yd1), 1);
//_mm256_set_m128i(xi1, xi0)
__m256i temp;
temp = _mm256_inserti128_si256(temp, xi0, 0);
temp = _mm256_inserti128_si256(temp, xi1, 1);
yf = _mm256_mul_ps(yf, _mm256_castsi256_ps(_mm256_slli_epi32(temp, 23)));
__m256 zf = _mm256_add_ps(xf, mA1);
#if CV_FMA3
zf = _mm256_fmadd_ps(zf, xf, mA2);
zf = _mm256_fmadd_ps(zf, xf, mA3);
zf = _mm256_fmadd_ps(zf, xf, mA4);
#else
zf = _mm256_add_ps(_mm256_mul_ps(zf, xf), mA2);
zf = _mm256_add_ps(_mm256_mul_ps(zf, xf), mA3);
zf = _mm256_add_ps(_mm256_mul_ps(zf, xf), mA4);
#endif
zf = _mm256_mul_ps(zf, yf);
if( y_aligned )
{
_mm256_store_ps(y + i, zf);
}
else
{
_mm256_storeu_ps(y + i, zf);
}
}
}
#elif CV_SSE2
if( n >= 8 ) if( n >= 8 )
{ {
static const __m128d prescale2 = _mm_set1_pd(exp_prescale); static const __m128d prescale2 = _mm_set1_pd(exp_prescale);