Skip to content

O
opencv
Commit eef900e4 authored by Vadim Pisarevsky's avatar Vadim Pisarevsky
Browse files
Options

optimized cmp and cvtscale(16s->16s) using SSE2 (thanks to Victoria)

parent 677fc3a0
Hide whitespace changes
Inline Side-by-side
Showing with 214 additions and 3 deletions
modules/core/src/arithm.cpp
View file @ eef900e4
......@@ -2007,7 +2007,7 @@ cmp_(const T* src1, size_t step1, const T* src2, size_t step2,
for( ; x < size.width; x++ )
dst[x] = (uchar)(-(src1[x] > src2[x]) ^ m);
}
}
}
else if( code == CMP_EQ || code == CMP_NE )
{
......@@ -2036,7 +2036,79 @@ cmp_(const T* src1, size_t step1, const T* src2, size_t step2,
static void cmp8u(const uchar* src1, size_t step1, const uchar* src2, size_t step2,
uchar* dst, size_t step, Size size, void* _cmpop)
{
cmp_(src1, step1, src2, step2, dst, step, size, *(int*)_cmpop);
//vz optimized cmp_(src1, step1, src2, step2, dst, step, size, *(int*)_cmpop);
int code = *(int*)_cmpop;
step1 /= sizeof(src1[0]);
step2 /= sizeof(src2[0]);
if( code == CMP_GE || code == CMP_LT )
{
std::swap(src1, src2);
std::swap(step1, step2);
code = code == CMP_GE ? CMP_LE : CMP_GT;
}
if( code == CMP_GT || code == CMP_LE )
{
int m = code == CMP_GT ? 0 : 255;
#if CV_SSE2
__m128i m128, c128;
if( USE_SSE2 ){
m128 = code == CMP_GT ? _mm_setzero_si128() : _mm_set1_epi8 (0xff);
c128 = _mm_set1_epi8 (0x7f);
}
#endif
for( ; size.height--; src1 += step1, src2 += step2, dst += step )
{
int x =0;
#if CV_SSE2
if( USE_SSE2 ){
for( ; x <= size.width - 16; x += 16 )
{
__m128i r00 = _mm_loadu_si128((const __m128i*)(src1 + x));
__m128i r10 = _mm_loadu_si128((const __m128i*)(src2 + x));
// no simd for 8u comparison, that's why we need the trick
r00 = _mm_sub_epi8(r00,c128);
r10 = _mm_sub_epi8(r10,c128);
r00 =_mm_xor_si128(_mm_cmpgt_epi8(r00, r10), m128);
_mm_storeu_si128((__m128i*)(dst + x),r00);
}
}
#endif
for( ; x < size.width; x++ ){
dst[x] = (uchar)(-(src1[x] > src2[x]) ^ m);
}
}
}
else if( code == CMP_EQ || code == CMP_NE )
{
int m = code == CMP_EQ ? 0 : 255;
#if CV_SSE2
__m128i m128;
if( USE_SSE2 ){
m128 = code == CMP_EQ ? _mm_setzero_si128() : _mm_set1_epi8 (0xff);
}
#endif
for( ; size.height--; src1 += step1, src2 += step2, dst += step )
{
int x = 0;
#if CV_SSE2
if( USE_SSE2 ){
for( ; x <= size.width - 16; x += 16 )
{
__m128i r00 = _mm_loadu_si128((const __m128i*)(src1 + x));
__m128i r10 = _mm_loadu_si128((const __m128i*)(src2 + x));
r00 = _mm_xor_si128 ( _mm_cmpeq_epi8 (r00, r10), m128);
_mm_storeu_si128((__m128i*)(dst + x), r00);
}
}
#endif
for( ; x < size.width; x++ )
dst[x] = (uchar)(-(src1[x] == src2[x]) ^ m);
}
}
}
static void cmp8s(const schar* src1, size_t step1, const schar* src2, size_t step2,
......@@ -2054,7 +2126,102 @@ static void cmp16u(const ushort* src1, size_t step1, const ushort* src2, size_t
static void cmp16s(const short* src1, size_t step1, const short* src2, size_t step2,
uchar* dst, size_t step, Size size, void* _cmpop)
{
cmp_(src1, step1, src2, step2, dst, step, size, *(int*)_cmpop);
//vz optimized cmp_(src1, step1, src2, step2, dst, step, size, *(int*)_cmpop);
int code = *(int*)_cmpop;
step1 /= sizeof(src1[0]);
step2 /= sizeof(src2[0]);
if( code == CMP_GE || code == CMP_LT )
{
std::swap(src1, src2);
std::swap(step1, step2);
code = code == CMP_GE ? CMP_LE : CMP_GT;
}
if( code == CMP_GT || code == CMP_LE )
{
int m = code == CMP_GT ? 0 : 255;
#if CV_SSE2
__m128i m128;
if( USE_SSE2 ){
m128 = code == CMP_GT ? _mm_setzero_si128() : _mm_set1_epi16 (0xffff);
}
#endif
for( ; size.height--; src1 += step1, src2 += step2, dst += step )
{
int x =0;
#if CV_SSE2
if( USE_SSE2){//
for( ; x <= size.width - 16; x += 16 )
{
__m128i r00 = _mm_loadu_si128((const __m128i*)(src1 + x));
__m128i r10 = _mm_loadu_si128((const __m128i*)(src2 + x));
r00 = _mm_xor_si128 ( _mm_cmpgt_epi16 (r00, r10), m128);
__m128i r01 = _mm_loadu_si128((const __m128i*)(src1 + x + 8));
__m128i r11 = _mm_loadu_si128((const __m128i*)(src2 + x + 8));
r01 = _mm_xor_si128 ( _mm_cmpgt_epi16 (r01, r11), m128);
r11 = _mm_packs_epi16(r00, r01);
_mm_storeu_si128((__m128i*)(dst + x), r11);
}
if( x <= size.width-8)
{
__m128i r00 = _mm_loadu_si128((const __m128i*)(src1 + x));
__m128i r10 = _mm_loadu_si128((const __m128i*)(src2 + x));
r00 = _mm_xor_si128 ( _mm_cmpgt_epi16 (r00, r10), m128);
r10 = _mm_packs_epi16(r00, r00);
_mm_storel_epi64((__m128i*)(dst + x), r10);
x += 8;
}
}
#endif
for( ; x < size.width; x++ ){
dst[x] = (uchar)(-(src1[x] > src2[x]) ^ m);
}
}
}
else if( code == CMP_EQ || code == CMP_NE )
{
int m = code == CMP_EQ ? 0 : 255;
#if CV_SSE2
__m128i m128;
if( USE_SSE2 ){
m128 = code == CMP_EQ ? _mm_setzero_si128() : _mm_set1_epi16 (0xffff);
}
#endif
for( ; size.height--; src1 += step1, src2 += step2, dst += step )
{
int x = 0;
#if CV_SSE2
if( USE_SSE2 ){
for( ; x <= size.width - 16; x += 16 )
{
__m128i r00 = _mm_loadu_si128((const __m128i*)(src1 + x));
__m128i r10 = _mm_loadu_si128((const __m128i*)(src2 + x));
r00 = _mm_xor_si128 ( _mm_cmpeq_epi16 (r00, r10), m128);
__m128i r01 = _mm_loadu_si128((const __m128i*)(src1 + x + 8));
__m128i r11 = _mm_loadu_si128((const __m128i*)(src2 + x + 8));
r01 = _mm_xor_si128 ( _mm_cmpeq_epi16 (r01, r11), m128);
r11 = _mm_packs_epi16(r00, r01);
_mm_storeu_si128((__m128i*)(dst + x), r11);
}
if( x <= size.width - 8)
{
__m128i r00 = _mm_loadu_si128((const __m128i*)(src1 + x));
__m128i r10 = _mm_loadu_si128((const __m128i*)(src2 + x));
r00 = _mm_xor_si128 ( _mm_cmpeq_epi16 (r00, r10), m128);
r10 = _mm_packs_epi16(r00, r00);
_mm_storel_epi64((__m128i*)(dst + x), r10);
x += 8;
}
}
#endif
for( ; x < size.width; x++ )
dst[x] = (uchar)(-(src1[x] == src2[x]) ^ m);
}
}
}
static void cmp32s(const int* src1, size_t step1, const int* src2, size_t step2,
......
modules/core/src/convert.cpp
View file @ eef900e4
......@@ -606,6 +606,50 @@ cvtScale_( const T* src, size_t sstep,
}
}
//vz optimized template specialization
template<> static void
cvtScale_<short, short, float>( const short* src, size_t sstep,
short* dst, size_t dstep, Size size,
float scale, float shift )
{
sstep /= sizeof(src[0]);
dstep /= sizeof(dst[0]);
#if CV_SSE2
__m128 scale128, shift128;
if(USE_SSE2){
scale128 = _mm_set1_ps (scale);
shift128 = _mm_set1_ps (shift);
}
#endif
for( ; size.height--; src += sstep, dst += dstep )
{
int x = 0;
#if CV_SSE2
if(USE_SSE2)
{
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);
r0 = _mm_packs_epi32(r0, r1);
_mm_storeu_si128((__m128i*)(dst + x), r0);
}
}
#endif
for(; x < size.width; x++ )
dst[x] = saturate_cast<short>(src[x]*scale + shift);
}
}
template<typename T, typename DT> static void
cvt_( const T* src, size_t sstep,
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment