Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

core: replace raw intrinsics with universal intrinsics in copy.cpp

Browse Source 
- use universal intrinsic instead of raw intrinsic
- add performance check for Mat::copyTo/setTo with mask
pull/10406/head
Sayed Adel 7 years ago
parent 1bc1f3d311
commit fd0ac962fb
2 changed files with 89 additions and 47 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 43
      modules/core/perf/perf_mat.cpp
  2. 93
      modules/core/src/copy.cpp

43
modules/core/perf/perf_mat.cpp
Unescape View File

@ -57,7 +57,7 @@ PERF_TEST_P(Size_MatType, Mat_Clone,
Size size = get<0>(GetParam()); Size size = get<0>(GetParam());
int type = get<1>(GetParam()); int type = get<1>(GetParam());
Mat source(size.height, size.width, type); Mat source(size.height, size.width, type);
Mat destination(size.height, size.width, type);; Mat destination(size.height, size.width, type);
declare.in(source, WARMUP_RNG).out(destination); declare.in(source, WARMUP_RNG).out(destination);
@ -97,6 +97,47 @@ PERF_TEST_P(Size_MatType, Mat_Clone_Roi,
SANITY_CHECK(destination, 1); SANITY_CHECK(destination, 1);
} }
PERF_TEST_P(Size_MatType, Mat_CopyToWithMask,
testing::Combine(testing::Values(TYPICAL_MAT_SIZES),
testing::Values(CV_8UC1, CV_8UC2))
)
{
const Size_MatType_t params = GetParam();
const Size size = get<0>(params);
const int type = get<1>(params);
Mat src(size, type), dst(size, type), mask(size, CV_8UC1);
declare.in(src, mask, WARMUP_RNG).out(dst);
TEST_CYCLE()
{
src.copyTo(dst, mask);
}
SANITY_CHECK(dst);
}
PERF_TEST_P(Size_MatType, Mat_SetToWithMask,
testing::Combine(testing::Values(TYPICAL_MAT_SIZES),
testing::Values(CV_8UC1, CV_8UC2))
)
{
const Size_MatType_t params = GetParam();
const Size size = get<0>(params);
const int type = get<1>(params);
const Scalar sc = Scalar::all(27);
Mat src(size, type), mask(size, CV_8UC1);
declare.in(src, mask, WARMUP_RNG).out(src);
TEST_CYCLE()
{
src.setTo(sc, mask);
}
SANITY_CHECK(src);
}
///////////// Transform //////////////////////// ///////////// Transform ////////////////////////
PERF_TEST_P(Size_MatType, Mat_Transform, PERF_TEST_P(Size_MatType, Mat_Transform,

93
modules/core/src/copy.cpp
Unescape View File

@ -90,28 +90,27 @@ copyMask_<uchar>(const uchar* _src, size_t sstep, const uchar* mask, size_t mste
const uchar* src = (const uchar*)_src; const uchar* src = (const uchar*)_src;
uchar* dst = (uchar*)_dst; uchar* dst = (uchar*)_dst;
int x = 0; int x = 0;
#if CV_SSE4_2 #if CV_SIMD128
if(USE_SSE4_2)// if( hasSIMD128()
{ #if CV_SSE4_2
__m128i zero = _mm_setzero_si128 (); && USE_SSE4_2
#endif
for( ; x <= size.width - 16; x += 16 ) ) {
{ v_uint8x16 v_zero = v_setzero_u8();
const __m128i rSrc = _mm_lddqu_si128((const __m128i*)(src+x));
__m128i _mask = _mm_lddqu_si128((const __m128i*)(mask+x)); for( ; x <= size.width - 16; x += 16 )
__m128i rDst = _mm_lddqu_si128((__m128i*)(dst+x)); {
__m128i _negMask = _mm_cmpeq_epi8(_mask, zero); v_uint8x16 v_src = v_load(src + x),
rDst = _mm_blendv_epi8(rSrc, rDst, _negMask); v_dst = v_load(dst + x),
_mm_storeu_si128((__m128i*)(dst + x), rDst); v_nmask = v_load(mask + x) == v_zero;
}
} #if CV_SSE4_2
#elif CV_NEON v_dst = v_uint8x16(_mm_blendv_epi8(v_src.val, v_dst.val, v_nmask.val));
uint8x16_t v_one = vdupq_n_u8(1); #else
for( ; x <= size.width - 16; x += 16 ) v_dst = v_select(v_nmask, v_dst, v_src);
{ #endif
uint8x16_t v_mask = vcgeq_u8(vld1q_u8(mask + x), v_one); v_store(dst + x, v_dst);
uint8x16_t v_dst = vld1q_u8(dst + x), v_src = vld1q_u8(src + x); }
vst1q_u8(dst + x, vbslq_u8(v_mask, v_src, v_dst));
} }
#endif #endif
for( ; x < size.width; x++ ) for( ; x < size.width; x++ )
@ -130,31 +129,33 @@ copyMask_<ushort>(const uchar* _src, size_t sstep, const uchar* mask, size_t mst
const ushort* src = (const ushort*)_src; const ushort* src = (const ushort*)_src;
ushort* dst = (ushort*)_dst; ushort* dst = (ushort*)_dst;
int x = 0; int x = 0;
#if CV_SSE4_2 #if CV_SIMD128
if(USE_SSE4_2)// if( hasSIMD128()
{ #if CV_SSE4_2
__m128i zero = _mm_setzero_si128 (); && USE_SSE4_2
for( ; x <= size.width - 8; x += 8 ) #endif
) {
v_uint8x16 v_zero = v_setzero_u8();
for( ; x <= size.width - 16; x += 16 )
{ {
const __m128i rSrc =_mm_lddqu_si128((const __m128i*)(src+x)); v_uint16x8 v_src1 = v_load(src + x), v_src2 = v_load(src + x + 8),
__m128i _mask = _mm_loadl_epi64((const __m128i*)(mask+x)); v_dst1 = v_load(dst + x), v_dst2 = v_load(dst + x + 8);
_mask = _mm_unpacklo_epi8(_mask, _mask);
__m128i rDst = _mm_lddqu_si128((const __m128i*)(dst+x)); v_uint8x16 v_nmask1, v_nmask2;
__m128i _negMask = _mm_cmpeq_epi8(_mask, zero); v_uint8x16 v_nmask = v_load(mask + x) == v_zero;
rDst = _mm_blendv_epi8(rSrc, rDst, _negMask); v_zip(v_nmask, v_nmask, v_nmask1, v_nmask2);
_mm_storeu_si128((__m128i*)(dst + x), rDst);
} #if CV_SSE4_2
} v_dst1 = v_uint16x8(_mm_blendv_epi8(v_src1.val, v_dst1.val, v_nmask1.val));
#elif CV_NEON v_dst2 = v_uint16x8(_mm_blendv_epi8(v_src2.val, v_dst2.val, v_nmask2.val));
uint8x8_t v_one = vdup_n_u8(1); #else
for( ; x <= size.width - 8; x += 8 ) v_dst1 = v_select(v_reinterpret_as_u16(v_nmask1), v_dst1, v_src1);
{ v_dst2 = v_select(v_reinterpret_as_u16(v_nmask2), v_dst2, v_src2);
uint8x8_t v_mask = vcge_u8(vld1_u8(mask + x), v_one); #endif
uint8x8x2_t v_mask2 = vzip_u8(v_mask, v_mask); v_store(dst + x, v_dst1);
uint16x8_t v_mask_res = vreinterpretq_u16_u8(vcombine_u8(v_mask2.val[0], v_mask2.val[1])); v_store(dst + x + 8, v_dst2);
}
uint16x8_t v_src = vld1q_u16(src + x), v_dst = vld1q_u16(dst + x);
vst1q_u16(dst + x, vbslq_u16(v_mask_res, v_src, v_dst));
} }
#endif #endif
for( ; x < size.width; x++ ) for( ; x < size.width; x++ )

Write Preview
Loading…
Cancel
Save