potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-31 00:03:03 +04:00
Code Packages Releases Wiki Activity

imgproc: speed up threshold of 64F version using NEON and SSE

Browse Source 
* use NEON under aarch64 only
  * check 64F version correctly
This commit is contained in:
Tomoaki Teshima
2016-06-21 11:36:52 +09:00
committed by Tomoaki Teshima
parent 0cb1e96472
commit 1095076d7f
2 changed files with 197 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/imgproc/src/thresh.cpp
+188
View File
@@ -915,6 +915,10 @@ thresh_64f(const Mat& _src, Mat& _dst, double thresh, double maxval, int type)
size_t src_step = _src.step / sizeof(src[0]);
size_t dst_step = _dst.step / sizeof(dst[0]);
#if CV_SSE2
volatile bool useSIMD = checkHardwareSupport(CV_CPU_SSE);
#endif
if (_src.isContinuous() && _dst.isContinuous())
{
roi.width *= roi.height;
@@ -927,6 +931,45 @@ thresh_64f(const Mat& _src, Mat& _dst, double thresh, double maxval, int type)
for (i = 0; i < roi.height; i++, src += src_step, dst += dst_step)
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128d thresh2 = _mm_set1_pd(thresh), maxval2 = _mm_set1_pd(maxval);
for( ; j <= roi.width - 8; j += 8 )
{
__m128d v0, v1, v2, v3;
v0 = _mm_loadu_pd( src + j );
v1 = _mm_loadu_pd( src + j + 2 );
v2 = _mm_loadu_pd( src + j + 4 );
v3 = _mm_loadu_pd( src + j + 6 );
v0 = _mm_cmpgt_pd( v0, thresh2 );
v1 = _mm_cmpgt_pd( v1, thresh2 );
v2 = _mm_cmpgt_pd( v2, thresh2 );
v3 = _mm_cmpgt_pd( v3, thresh2 );
v0 = _mm_and_pd( v0, maxval2 );
v1 = _mm_and_pd( v1, maxval2 );
v2 = _mm_and_pd( v2, maxval2 );
v3 = _mm_and_pd( v3, maxval2 );
_mm_storeu_pd( dst + j, v0 );
_mm_storeu_pd( dst + j + 2, v1 );
_mm_storeu_pd( dst + j + 4, v2 );
_mm_storeu_pd( dst + j + 6, v3 );
}
}
#elif CV_NEON && defined(__aarch64__)
float64x2_t v_thresh = vdupq_n_f64(thresh);
uint64x2_t v_maxval = vreinterpretq_u64_f64(vdupq_n_f64(maxval));
for( ; j <= roi.width - 4; j += 4 )
{
float64x2_t v_src0 = vld1q_f64(src + j);
float64x2_t v_src1 = vld1q_f64(src + j + 2);
uint64x2_t v_dst0 = vandq_u64(vcgtq_f64(v_src0, v_thresh), v_maxval);
uint64x2_t v_dst1 = vandq_u64(vcgtq_f64(v_src1, v_thresh), v_maxval);
vst1q_f64(dst + j, vreinterpretq_f64_u64(v_dst0));
vst1q_f64(dst + j + 2, vreinterpretq_f64_u64(v_dst1));
}
#endif
for (; j < roi.width; j++)
dst[j] = src[j] > thresh ? maxval : 0;
@@ -938,6 +981,45 @@ thresh_64f(const Mat& _src, Mat& _dst, double thresh, double maxval, int type)
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128d thresh2 = _mm_set1_pd(thresh), maxval2 = _mm_set1_pd(maxval);
for( ; j <= roi.width - 8; j += 8 )
{
__m128d v0, v1, v2, v3;
v0 = _mm_loadu_pd( src + j );
v1 = _mm_loadu_pd( src + j + 2 );
v2 = _mm_loadu_pd( src + j + 4 );
v3 = _mm_loadu_pd( src + j + 6 );
v0 = _mm_cmple_pd( v0, thresh2 );
v1 = _mm_cmple_pd( v1, thresh2 );
v2 = _mm_cmple_pd( v2, thresh2 );
v3 = _mm_cmple_pd( v3, thresh2 );
v0 = _mm_and_pd( v0, maxval2 );
v1 = _mm_and_pd( v1, maxval2 );
v2 = _mm_and_pd( v2, maxval2 );
v3 = _mm_and_pd( v3, maxval2 );
_mm_storeu_pd( dst + j, v0 );
_mm_storeu_pd( dst + j + 2, v1 );
_mm_storeu_pd( dst + j + 4, v2 );
_mm_storeu_pd( dst + j + 6, v3 );
}
}
#elif CV_NEON && defined(__aarch64__)
float64x2_t v_thresh = vdupq_n_f64(thresh);
uint64x2_t v_maxval = vreinterpretq_u64_f64(vdupq_n_f64(maxval));
for( ; j <= roi.width - 4; j += 4 )
{
float64x2_t v_src0 = vld1q_f64(src + j);
float64x2_t v_src1 = vld1q_f64(src + j + 2);
uint64x2_t v_dst0 = vandq_u64(vcleq_f64(v_src0, v_thresh), v_maxval);
uint64x2_t v_dst1 = vandq_u64(vcleq_f64(v_src1, v_thresh), v_maxval);
vst1q_f64(dst + j, vreinterpretq_f64_u64(v_dst0));
vst1q_f64(dst + j + 2, vreinterpretq_f64_u64(v_dst1));
}
#endif
for (; j < roi.width; j++)
dst[j] = src[j] <= thresh ? maxval : 0;
}
@@ -948,6 +1030,40 @@ thresh_64f(const Mat& _src, Mat& _dst, double thresh, double maxval, int type)
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128d thresh2 = _mm_set1_pd(thresh);
for( ; j <= roi.width - 8; j += 8 )
{
__m128d v0, v1, v2, v3;
v0 = _mm_loadu_pd( src + j );
v1 = _mm_loadu_pd( src + j + 2 );
v2 = _mm_loadu_pd( src + j + 4 );
v3 = _mm_loadu_pd( src + j + 6 );
v0 = _mm_min_pd( v0, thresh2 );
v1 = _mm_min_pd( v1, thresh2 );
v2 = _mm_min_pd( v2, thresh2 );
v3 = _mm_min_pd( v3, thresh2 );
_mm_storeu_pd( dst + j, v0 );
_mm_storeu_pd( dst + j + 2, v1 );
_mm_storeu_pd( dst + j + 4, v2 );
_mm_storeu_pd( dst + j + 6, v3 );
}
}
#elif CV_NEON && defined(__aarch64__)
float64x2_t v_thresh = vdupq_n_f64(thresh);
for( ; j <= roi.width - 4; j += 4 )
{
float64x2_t v_src0 = vld1q_f64(src + j);
float64x2_t v_src1 = vld1q_f64(src + j + 2);
float64x2_t v_dst0 = vminq_f64(v_src0, v_thresh);
float64x2_t v_dst1 = vminq_f64(v_src1, v_thresh);
vst1q_f64(dst + j, v_dst0);
vst1q_f64(dst + j + 2, v_dst1);
}
#endif
for (; j < roi.width; j++)
dst[j] = std::min(src[j], thresh);
}
@@ -958,6 +1074,42 @@ thresh_64f(const Mat& _src, Mat& _dst, double thresh, double maxval, int type)
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128d thresh2 = _mm_set1_pd(thresh);
for( ; j <= roi.width - 8; j += 8 )
{
__m128d v0, v1, v2, v3;
v0 = _mm_loadu_pd( src + j );
v1 = _mm_loadu_pd( src + j + 2 );
v2 = _mm_loadu_pd( src + j + 4 );
v3 = _mm_loadu_pd( src + j + 6 );
v0 = _mm_and_pd( v0, _mm_cmpgt_pd(v0, thresh2));
v1 = _mm_and_pd( v1, _mm_cmpgt_pd(v1, thresh2));
v2 = _mm_and_pd( v2, _mm_cmpgt_pd(v2, thresh2));
v3 = _mm_and_pd( v3, _mm_cmpgt_pd(v3, thresh2));
_mm_storeu_pd( dst + j, v0 );
_mm_storeu_pd( dst + j + 2, v1 );
_mm_storeu_pd( dst + j + 4, v2 );
_mm_storeu_pd( dst + j + 6, v3 );
}
}
#elif CV_NEON && defined(__aarch64__)
float64x2_t v_thresh = vdupq_n_f64(thresh);
for( ; j <= roi.width - 4; j += 4 )
{
float64x2_t v_src0 = vld1q_f64(src + j);
float64x2_t v_src1 = vld1q_f64(src + j + 2);
uint64x2_t v_dst0 = vandq_u64(vcgtq_f64(v_src0, v_thresh),
vreinterpretq_u64_f64(v_src0));
uint64x2_t v_dst1 = vandq_u64(vcgtq_f64(v_src1, v_thresh),
vreinterpretq_u64_f64(v_src1));
vst1q_f64(dst + j, vreinterpretq_f64_u64(v_dst0));
vst1q_f64(dst + j + 2, vreinterpretq_f64_u64(v_dst1));
}
#endif
for (; j < roi.width; j++)
{
double v = src[j];
@@ -971,6 +1123,42 @@ thresh_64f(const Mat& _src, Mat& _dst, double thresh, double maxval, int type)
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128d thresh2 = _mm_set1_pd(thresh);
for( ; j <= roi.width - 8; j += 8 )
{
__m128d v0, v1, v2, v3;
v0 = _mm_loadu_pd( src + j );
v1 = _mm_loadu_pd( src + j + 2 );
v2 = _mm_loadu_pd( src + j + 4 );
v3 = _mm_loadu_pd( src + j + 6 );
v0 = _mm_and_pd( v0, _mm_cmple_pd(v0, thresh2));
v1 = _mm_and_pd( v1, _mm_cmple_pd(v1, thresh2));
v2 = _mm_and_pd( v2, _mm_cmple_pd(v2, thresh2));
v3 = _mm_and_pd( v3, _mm_cmple_pd(v3, thresh2));
_mm_storeu_pd( dst + j, v0 );
_mm_storeu_pd( dst + j + 2, v1 );
_mm_storeu_pd( dst + j + 4, v2 );
_mm_storeu_pd( dst + j + 6, v3 );
}
}
#elif CV_NEON && defined(__aarch64__)
float64x2_t v_thresh = vdupq_n_f64(thresh);
for( ; j <= roi.width - 4; j += 4 )
{
float64x2_t v_src0 = vld1q_f64(src + j);
float64x2_t v_src1 = vld1q_f64(src + j + 2);
uint64x2_t v_dst0 = vandq_u64(vcleq_f64(v_src0, v_thresh),
vreinterpretq_u64_f64(v_src0));
uint64x2_t v_dst1 = vandq_u64(vcleq_f64(v_src1, v_thresh),
vreinterpretq_u64_f64(v_src1));
vst1q_f64(dst + j, vreinterpretq_f64_u64(v_dst0));
vst1q_f64(dst + j + 2, vreinterpretq_f64_u64(v_dst1));
}
#endif
for (; j < roi.width; j++)
{
double v = src[j];