mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-06-20 01:41:14 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #15828 from ChipKerchner:momentsToHal

Browse Source 
* Convert moments in tile algorithms to HAL (1.3x faster for VSX).

* Adding NEON code back in for non 64-bit platforms.

* Remove floats from post processing.
This commit is contained in:
Chip Kerchner 2019-11-05 10:52:35 -05:00 committed by Alexander Alekhin
parent 53139e6ebe
commit 2112aa31e6
1 changed files with 35 additions and 125 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

160
modules/imgproc/src/moments.cpp
View File

@ -38,8 +38,10 @@
// the use of this software, even if advised of the possibility of such damage. // the use of this software, even if advised of the possibility of such damage.
// //
//M*/ //M*/
#include "precomp.hpp" #include "precomp.hpp"
#include "opencl_kernels_imgproc.hpp" #include "opencl_kernels_imgproc.hpp"
#include "opencv2/core/hal/intrin.hpp"
namespace cv namespace cv
{ {
@ -211,7 +213,7 @@ struct MomentsInTile_SIMD
} }
}; };
#if CV_SSE2 #if CV_SIMD128
template <> template <>
struct MomentsInTile_SIMD<uchar, int, int> struct MomentsInTile_SIMD<uchar, int, int>
@ -226,115 +228,33 @@ struct MomentsInTile_SIMD<uchar, int, int>
int x = 0; int x = 0;
{ {
__m128i dx = _mm_set1_epi16(8); v_int16x8 dx = v_setall_s16(8), qx = v_int16x8(0, 1, 2, 3, 4, 5, 6, 7);
__m128i z = _mm_setzero_si128(), qx0 = z, qx1 = z, qx2 = z, qx3 = z, qx = _mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7); v_uint32x4 z = v_setzero_u32(), qx0 = z, qx1 = z, qx2 = z, qx3 = z;
for( ; x <= len - 8; x += 8 ) for( ; x <= len - 8; x += 8 )
{ {
__m128i p = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(ptr + x)), z); v_int16x8 p = v_reinterpret_as_s16(v_load_expand(ptr + x));
__m128i sx = _mm_mullo_epi16(qx, qx); v_int16x8 sx = v_mul_wrap(qx, qx);
qx0 = _mm_add_epi16(qx0, p); qx0 += v_reinterpret_as_u32(p);
qx1 = _mm_add_epi32(qx1, _mm_madd_epi16(p, qx)); qx1 = v_reinterpret_as_u32(v_dotprod(p, qx, v_reinterpret_as_s32(qx1)));
qx2 = _mm_add_epi32(qx2, _mm_madd_epi16(p, sx)); qx2 = v_reinterpret_as_u32(v_dotprod(p, sx, v_reinterpret_as_s32(qx2)));
qx3 = _mm_add_epi32(qx3, _mm_madd_epi16( _mm_mullo_epi16(p, qx), sx)); qx3 = v_reinterpret_as_u32(v_dotprod(v_mul_wrap(p, qx), sx, v_reinterpret_as_s32(qx3)));
qx = _mm_add_epi16(qx, dx); qx += dx;
} }
__m128i qx01_lo = _mm_unpacklo_epi32(qx0, qx1); x0 = v_reduce_sum(qx0);
__m128i qx23_lo = _mm_unpacklo_epi32(qx2, qx3); x0 = (x0 & 0xffff) + (x0 >> 16);
__m128i qx01_hi = _mm_unpackhi_epi32(qx0, qx1); x1 = v_reduce_sum(qx1);
__m128i qx23_hi = _mm_unpackhi_epi32(qx2, qx3); x2 = v_reduce_sum(qx2);
qx01_lo = _mm_add_epi32(qx01_lo, qx01_hi); x3 = v_reduce_sum(qx3);
qx23_lo = _mm_add_epi32(qx23_lo, qx23_hi);
__m128i qx0123_lo = _mm_unpacklo_epi64(qx01_lo, qx23_lo);
__m128i qx0123_hi = _mm_unpackhi_epi64(qx01_lo, qx23_lo);
qx0123_lo = _mm_add_epi32(qx0123_lo, qx0123_hi);
_mm_store_si128((__m128i*)buf, qx0123_lo);
x0 = (buf[0] & 0xffff) + (buf[0] >> 16);
x1 = buf[1];
x2 = buf[2];
x3 = buf[3];
} }
return x; return x;
} }
int CV_DECL_ALIGNED(16) buf[4];
}; };
#elif CV_NEON
template <>
struct MomentsInTile_SIMD<uchar, int, int>
{
MomentsInTile_SIMD()
{
ushort CV_DECL_ALIGNED(8) init[4] = { 0, 1, 2, 3 };
qx_init = vld1_u16(init);
v_step = vdup_n_u16(4);
}
int operator() (const uchar * ptr, int len, int & x0, int & x1, int & x2, int & x3)
{
int x = 0;
uint32x4_t v_z = vdupq_n_u32(0), v_x0 = v_z, v_x1 = v_z,
v_x2 = v_z, v_x3 = v_z;
uint16x4_t qx = qx_init;
for( ; x <= len - 8; x += 8 )
{
uint16x8_t v_src = vmovl_u8(vld1_u8(ptr + x));
// first part
uint32x4_t v_qx = vmovl_u16(qx);
uint16x4_t v_p = vget_low_u16(v_src);
uint32x4_t v_px = vmull_u16(qx, v_p);
v_x0 = vaddw_u16(v_x0, v_p);
v_x1 = vaddq_u32(v_x1, v_px);
v_px = vmulq_u32(v_px, v_qx);
v_x2 = vaddq_u32(v_x2, v_px);
v_x3 = vaddq_u32(v_x3, vmulq_u32(v_px, v_qx));
qx = vadd_u16(qx, v_step);
// second part
v_qx = vmovl_u16(qx);
v_p = vget_high_u16(v_src);
v_px = vmull_u16(qx, v_p);
v_x0 = vaddw_u16(v_x0, v_p);
v_x1 = vaddq_u32(v_x1, v_px);
v_px = vmulq_u32(v_px, v_qx);
v_x2 = vaddq_u32(v_x2, v_px);
v_x3 = vaddq_u32(v_x3, vmulq_u32(v_px, v_qx));
qx = vadd_u16(qx, v_step);
}
vst1q_u32(buf, v_x0);
x0 = buf[0] + buf[1] + buf[2] + buf[3];
vst1q_u32(buf, v_x1);
x1 = buf[0] + buf[1] + buf[2] + buf[3];
vst1q_u32(buf, v_x2);
x2 = buf[0] + buf[1] + buf[2] + buf[3];
vst1q_u32(buf, v_x3);
x3 = buf[0] + buf[1] + buf[2] + buf[3];
return x;
}
uint CV_DECL_ALIGNED(16) buf[4];
uint16x4_t qx_init, v_step;
};
#endif
#if CV_SSE4_1
template <> template <>
struct MomentsInTile_SIMD<ushort, int, int64> struct MomentsInTile_SIMD<ushort, int, int64>
{ {
@ -348,49 +268,39 @@ struct MomentsInTile_SIMD<ushort, int, int64>
int x = 0; int x = 0;
{ {
__m128i v_delta = _mm_set1_epi32(4), v_zero = _mm_setzero_si128(), v_x0 = v_zero, v_int32x4 v_delta = v_setall_s32(4), v_ix0 = v_int32x4(0, 1, 2, 3);
v_x1 = v_zero, v_x2 = v_zero, v_x3 = v_zero, v_ix0 = _mm_setr_epi32(0, 1, 2, 3); v_uint32x4 z = v_setzero_u32(), v_x0 = z, v_x1 = z, v_x2 = z;
v_uint64x2 v_x3 = v_reinterpret_as_u64(z);
for( ; x <= len - 4; x += 4 ) for( ; x <= len - 4; x += 4 )
{ {
__m128i v_src = _mm_loadl_epi64((const __m128i *)(ptr + x)); v_int32x4 v_src = v_reinterpret_as_s32(v_load_expand(ptr + x));
v_src = _mm_unpacklo_epi16(v_src, v_zero);
v_x0 = _mm_add_epi32(v_x0, v_src); v_x0 += v_reinterpret_as_u32(v_src);
v_x1 = _mm_add_epi32(v_x1, _mm_mullo_epi32(v_src, v_ix0)); v_x1 += v_reinterpret_as_u32(v_src * v_ix0);
__m128i v_ix1 = _mm_mullo_epi32(v_ix0, v_ix0); v_int32x4 v_ix1 = v_ix0 * v_ix0;
v_x2 = _mm_add_epi32(v_x2, _mm_mullo_epi32(v_src, v_ix1)); v_x2 += v_reinterpret_as_u32(v_src * v_ix1);
v_ix1 = _mm_mullo_epi32(v_ix0, v_ix1); v_ix1 = v_ix0 * v_ix1;
v_src = _mm_mullo_epi32(v_src, v_ix1); v_src = v_src * v_ix1;
v_x3 = _mm_add_epi64(v_x3, _mm_add_epi64(_mm_unpacklo_epi32(v_src, v_zero), _mm_unpackhi_epi32(v_src, v_zero))); v_uint64x2 v_lo, v_hi;
v_expand(v_reinterpret_as_u32(v_src), v_lo, v_hi);
v_x3 += v_lo + v_hi;
v_ix0 = _mm_add_epi32(v_ix0, v_delta); v_ix0 += v_delta;
} }
__m128i v_x01_lo = _mm_unpacklo_epi32(v_x0, v_x1); x0 = v_reduce_sum(v_x0);
__m128i v_x22_lo = _mm_unpacklo_epi32(v_x2, v_x2); x1 = v_reduce_sum(v_x1);
__m128i v_x01_hi = _mm_unpackhi_epi32(v_x0, v_x1); x2 = v_reduce_sum(v_x2);
__m128i v_x22_hi = _mm_unpackhi_epi32(v_x2, v_x2); v_store_aligned(buf64, v_reinterpret_as_s64(v_x3));
v_x01_lo = _mm_add_epi32(v_x01_lo, v_x01_hi);
v_x22_lo = _mm_add_epi32(v_x22_lo, v_x22_hi);
__m128i v_x0122_lo = _mm_unpacklo_epi64(v_x01_lo, v_x22_lo);
__m128i v_x0122_hi = _mm_unpackhi_epi64(v_x01_lo, v_x22_lo);
v_x0122_lo = _mm_add_epi32(v_x0122_lo, v_x0122_hi);
_mm_store_si128((__m128i*)buf64, v_x3);
_mm_store_si128((__m128i*)buf, v_x0122_lo);
x0 = buf[0];
x1 = buf[1];
x2 = buf[2];
x3 = buf64[0] + buf64[1]; x3 = buf64[0] + buf64[1];
} }
return x; return x;
} }
int CV_DECL_ALIGNED(16) buf[4];
int64 CV_DECL_ALIGNED(16) buf64[2]; int64 CV_DECL_ALIGNED(16) buf64[2];
}; };