potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #14210 from terfendail:wui_512

Browse Source 
AVX512 wide universal intrinsics (#14210)

* Added implementation of 512-bit wide universal intrinsics(WIP)

* Added implementation of 512-bit wide universal intrinsics: implemented WUI vector types(WIP)

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented load/store

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented fp16 load/store

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented recombine and zip, implemented non-saturating and saturating arithmetics

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented bit operations

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented comparisons

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented lane shifts and reduction

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented absolute values

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented rounding and cast to float

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented LUT

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented type extension/narrowing and matrix operations

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented load_deinterleave for 2 and 3 channels images

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented load_deinterleave for 2- and implemented for 4-channel images

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented store_interleave

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented signmask and checks

* Added implementation of 512-bit wide universal intrinsics(WIP): build fixes

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented popcount in case AVX512_BITALG is unavailable

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented zip

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented rotate for s8 and s16

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented interleave/deinterleave for s8 and s16

* Added implementation of 512-bit wide universal intrinsics(WIP): updated v512_set macros

* Added implementation of 512-bit wide universal intrinsics(WIP): fix for GCC wrong _mm512_abs_pd definition

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked v_zip to avoid AVX512_VBMI intrinsics

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked v_invsqrt to avoid AVX512_ER intrinsics

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked v_rotate, v_popcount and interleave/deinterleave for U8 to avoid AVX512_VBMI intrinsics

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed integral image SIMD part

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed warnings

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed load_deinterleave for u8 and u16

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed v_invsqrt accuracy for f64

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed interleave/deinterleave for u32 and u64

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed interleave_pairs, interleave_quads and pack_triplets

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed rotate_left

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed rotate_left/right, part 2

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed 512-wide universal intrinsics based resize

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed findContours by avoiding use of uint64 dependent 512-wide v_signmask()

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed trailing whitespaces

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked specific intrinsic sets dependent parts to check availability of intrinsics based on CPU feature group defines

* Added implementation of 512-bit wide universal intrinsics(WIP):Updated AVX512 implementation of v_popcount to avoid AVX512VPOPCNTDQ intrinsics if unavailable.

* Added implementation of 512-bit wide universal intrinsics(WIP): Fixed universal intrinsics data initialisation, v_mul_wrap, v_floor, v_ceil and v_signmask.

* Added implementation of 512-bit wide universal intrinsics(WIP): Removed hasSIMD512()

* Added implementation of 512-bit wide universal intrinsics(WIP): Fixes for gcc build

* Added implementation of 512-bit wide universal intrinsics(WIP): Reworked v_signmask, v_check_any() and v_check_all() implementation.
This commit is contained in:
Vitaly Tuzov
2019-06-03 18:05:35 +03:00
committed by Alexander Alekhin
parent 3289a0aff9
commit 3b015dfc7d
10 changed files with 2902 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/imgproc/src/contours.cpp
+26 -8
View File
@@ -1061,10 +1061,16 @@ cvFindNextContour( CvContourScanner scanner )
}
else
{
v_uint8 v_prev = vx_setall_u8((uchar)prev);
for (; x <= width - v_uint8::nlanes; x += v_uint8::nlanes)
#if CV_SIMD_WIDTH > 16
v_uint8 vx_prev = vx_setall_u8((uchar)prev);
while (x <= width - v_uint8::nlanes &&
v_check_all(vx_load((uchar*)(img + x)) == vx_prev))
x += v_uint8::nlanes;
#endif
v_uint8x16 v_prev = v_setall_u8((uchar)prev);
for (; x <= width - v_uint8x16::nlanes; x += v_uint8x16::nlanes)
{
unsigned int mask = (unsigned int)v_signmask(vx_load((uchar*)(img + x)) != v_prev);
unsigned int mask = (unsigned int)v_signmask(v_load((uchar*)(img + x)) != v_prev);
if (mask)
{
p = img[(x += cv::trailingZeros32(mask))];
@@ -1328,10 +1334,16 @@ CvLinkedRunPoint;
inline int findStartContourPoint(uchar *src_data, CvSize img_size, int j)
{
#if CV_SIMD
v_uint8 v_zero = vx_setzero_u8();
for (; j <= img_size.width - v_uint8::nlanes; j += v_uint8::nlanes)
#if CV_SIMD_WIDTH > 16
v_uint8 vx_zero = vx_setzero_u8();
while (j <= img_size.width - v_uint8::nlanes &&
v_check_all(vx_load((uchar*)(src_data + j)) == vx_zero))
j += v_uint8::nlanes;
#endif
v_uint8x16 v_zero = v_setzero_u8();
for (; j <= img_size.width - v_uint8x16::nlanes; j += v_uint8x16::nlanes)
{
unsigned int mask = (unsigned int)v_signmask(vx_load((uchar*)(src_data + j)) != v_zero);
unsigned int mask = (unsigned int)v_signmask(v_load((uchar*)(src_data + j)) != v_zero);
if (mask)
{
j += cv::trailingZeros32(mask);
@@ -1353,10 +1365,16 @@ inline int findEndContourPoint(uchar *src_data, CvSize img_size, int j)
}
else
{
v_uint8 v_zero = vx_setzero_u8();
#if CV_SIMD_WIDTH > 16
v_uint8 vx_zero = vx_setzero_u8();
while (j <= img_size.width - v_uint8::nlanes &&
v_check_all(vx_load((uchar*)(src_data + j)) != vx_zero))
j += v_uint8::nlanes;
#endif
v_uint8x16 v_zero = v_setzero_u8();
for (; j <= img_size.width - v_uint8::nlanes; j += v_uint8::nlanes)
{
unsigned int mask = (unsigned int)v_signmask(vx_load((uchar*)(src_data + j)) == v_zero);
unsigned int mask = (unsigned int)v_signmask(v_load((uchar*)(src_data + j)) == v_zero);
if (mask)
{
j += cv::trailingZeros32(mask);