generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

avx2

Browse Source
This commit is contained in:
Ilya Lavrenov
2015-01-12 10:59:28 +03:00
parent 28833421ae
commit 8d48632ebe
7 changed files with 105 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

72
modules/core/src/convert.cpp
View File

@@ -2294,26 +2294,44 @@ cvtScale_<short, int, float>( const short* src, size_t sstep,
{
int x = 0;
#if CV_SSE2
if(USE_SSE2)//~5X
#if CV_AVX2
if (USE_AVX2)
{
__m256 scale256 = _mm256_set1_ps (scale);
__m256 shift256 = _mm256_set1_ps (shift);
__m256i zero = _mm256_setzero_si256();
for ( ; x <= size.width - 16; x += 16)
{
__m128 scale128 = _mm_set1_ps (scale);
__m128 shift128 = _mm_set1_ps (shift);
for(; x <= size.width - 8; x += 8 )
{
__m128i r0 = _mm_loadl_epi64((const __m128i*)(src + x));
__m128i r1 = _mm_loadl_epi64((const __m128i*)(src + x + 4));
__m128 rf0 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r0, r0), 16));
__m128 rf1 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r1, r1), 16));
rf0 = _mm_add_ps(_mm_mul_ps(rf0, scale128), shift128);
rf1 = _mm_add_ps(_mm_mul_ps(rf1, scale128), shift128);
r0 = _mm_cvtps_epi32(rf0);
r1 = _mm_cvtps_epi32(rf1);
_mm_storeu_si128((__m128i*)(dst + x), r0);
_mm_storeu_si128((__m128i*)(dst + x + 4), r1);
}
__m256i v_src = _mm256_loadu_si256((__m256i const *)(src + x));
__m256i v_src_lo = _mm256_unpacklo_epi16(v_src, zero);
__m256i v_src_hi = _mm256_unpackhi_epi16(v_src, zero);
__m256 v_dst0 = _mm256_add_ps( _mm256_mul_ps(_mm256_cvtepi32_ps (v_src_lo), scale256), shift256);
__m256 v_dst1 = _mm256_add_ps( _mm256_mul_ps(_mm256_cvtepi32_ps (v_src_hi), scale256), shift256);
_mm256_storeu_si256 ((__m256i *)(dst + x), _mm256_cvtps_epi32(v_dst0));
_mm256_storeu_si256 ((__m256i *)(dst + x + 8), _mm256_cvtps_epi32(v_dst1));
}
}
#endif
#if CV_SSE2
if (USE_SSE2)//~5X
{
__m128 scale128 = _mm_set1_ps (scale);
__m128 shift128 = _mm_set1_ps (shift);
for(; x <= size.width - 8; x += 8 )
{
__m128i r0 = _mm_loadl_epi64((const __m128i*)(src + x));
__m128i r1 = _mm_loadl_epi64((const __m128i*)(src + x + 4));
__m128 rf0 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r0, r0), 16));
__m128 rf1 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r1, r1), 16));
rf0 = _mm_add_ps(_mm_mul_ps(rf0, scale128), shift128);
rf1 = _mm_add_ps(_mm_mul_ps(rf1, scale128), shift128);
r0 = _mm_cvtps_epi32(rf0);
r1 = _mm_cvtps_epi32(rf1);
_mm_storeu_si128((__m128i*)(dst + x), r0);
_mm_storeu_si128((__m128i*)(dst + x + 4), r1);
}
}
#elif CV_NEON
float32x4_t v_shift = vdupq_n_f32(shift);
for(; x <= size.width - 8; x += 8 )
@@ -2330,24 +2348,6 @@ cvtScale_<short, int, float>( const short* src, size_t sstep,
}
#endif
//We will wait Haswell
/*
#if CV_AVX
if(USE_AVX)//2X - bad variant
{
////TODO:AVX implementation (optimization?) required
__m256 scale256 = _mm256_set1_ps (scale);
__m256 shift256 = _mm256_set1_ps (shift);
for(; x <= size.width - 8; x += 8 )
{
__m256i buf = _mm256_set_epi32((int)(*(src+x+7)),(int)(*(src+x+6)),(int)(*(src+x+5)),(int)(*(src+x+4)),(int)(*(src+x+3)),(int)(*(src+x+2)),(int)(*(src+x+1)),(int)(*(src+x)));
__m256 r0 = _mm256_add_ps( _mm256_mul_ps(_mm256_cvtepi32_ps (buf), scale256), shift256);
__m256i res = _mm256_cvtps_epi32(r0);
_mm256_storeu_si256 ((__m256i*)(dst+x), res);
}
}
#endif*/
for(; x < size.width; x++ )
dst[x] = saturate_cast<int>(src[x]*scale + shift);
}