generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Better NEON Hamming distance

Browse Source
This commit is contained in:
Andrey Kamaev 2011-10-26 13:04:53 +00:00
parent 509730c1cd
commit d7ff92439d
5 changed files with 42 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
modules/core/include/opencv2/core/internal.hpp
View File

@ -119,7 +119,7 @@ CV_INLINE IppiSize ippiSize(int width, int height)
#define CV_SSE3 0
#endif
#if defined ANDROID && defined __ARM_NEON__
#if defined ANDROID && defined __ARM_NEON__ && defined __GNUC__
#include "arm_neon.h"
#define CV_NEON 1

18
modules/core/src/stat.cpp
View File

@ -963,26 +963,22 @@ static const uchar popCountTable4[] =
int normHamming(const uchar* a, const uchar* b, int n)
{
int i = 0, result = 0;
#if defined __GNUC__ && CV_NEON
#if CV_NEON
if (CPU_HAS_NEON_FEATURE)
{
result = 0;
for( ; i <= n - 16; i += 16 )
{
uint32x4_t bits = vmovq_n_u32(0);
for (; i <= n - 16; i += 16) {
uint8x16_t A_vec = vld1q_u8 (a + i);
uint8x16_t B_vec = vld1q_u8 (b + i);
//uint8x16_t veorq_u8 (uint8x16_t, uint8x16_t)
uint8x16_t AxorB = veorq_u8 (A_vec, B_vec);
uint8x16_t bitsSet = vcntq_u8 (AxorB);
//uint16x8_t vpadalq_u8 (uint16x8_t, uint8x16_t)
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
uint64x2_t bitSet2 = vpaddlq_u32 (bitSet4);
result += vgetq_lane_u64 (bitSet2,0);
result += vgetq_lane_u64 (bitSet2,1);
bits = vaddq_u32(bits, bitSet4);
}
uint64x2_t bitSet2 = vpaddlq_u32 (bits);
result = vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),0);
result += vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),2);
}
else
#endif

49
modules/flann/include/opencv2/flann/dist.h
View File

@ -421,43 +421,42 @@ struct Hamming
{
ResultType result = 0;
#if __GNUC__
#if ANDROID && HAVE_NEON
static uint64_t features = android_getCpuFeatures();
if ((features& ANDROID_CPU_ARM_FEATURE_NEON)) {
#if CV_NEON
if (CPU_HAS_NEON_FEATURE) {
uint32x4_t bits = vmovq_n_u32(0);
for (size_t i = 0; i < size; i += 16) {
uint8x16_t A_vec = vld1q_u8 (a + i);
uint8x16_t B_vec = vld1q_u8 (b + i);
//uint8x16_t veorq_u8 (uint8x16_t, uint8x16_t)
uint8x16_t AxorB = veorq_u8 (A_vec, B_vec);
uint8x16_t bitsSet += vcntq_u8 (AxorB);
//uint16x8_t vpadalq_u8 (uint16x8_t, uint8x16_t)
uint8x16_t bitsSet = vcntq_u8 (AxorB);
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
uint64x2_t bitSet2 = vpaddlq_u32 (bitSet4);
result += vgetq_lane_u64 (bitSet2,0);
result += vgetq_lane_u64 (bitSet2,1);
bits = vaddq_u32(bits, bitSet4);
}
uint64x2_t bitSet2 = vpaddlq_u32 (bits);
result = vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),0);
result += vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),2);
}
else
#endif
//for portability just use unsigned long -- and use the __builtin_popcountll (see docs for __builtin_popcountll)
typedef unsigned long long pop_t;
const size_t modulo = size % sizeof(pop_t);
const pop_t* a2 = reinterpret_cast<const pop_t*> (a);
const pop_t* b2 = reinterpret_cast<const pop_t*> (b);
const pop_t* a2_end = a2 + (size / sizeof(pop_t));
{
//for portability just use unsigned long -- and use the __builtin_popcountll (see docs for __builtin_popcountll)
typedef unsigned long long pop_t;
const size_t modulo = size % sizeof(pop_t);
const pop_t* a2 = reinterpret_cast<const pop_t*> (a);
const pop_t* b2 = reinterpret_cast<const pop_t*> (b);
const pop_t* a2_end = a2 + (size / sizeof(pop_t));
for (; a2 != a2_end; ++a2, ++b2) result += __builtin_popcountll((*a2) ^ (*b2));
for (; a2 != a2_end; ++a2, ++b2) result += __builtin_popcountll((*a2) ^ (*b2));
if (modulo) {
//in the case where size is not dividable by sizeof(size_t)
//need to mask off the bits at the end
pop_t a_final = 0, b_final = 0;
memcpy(&a_final, a2, modulo);
memcpy(&b_final, b2, modulo);
result += __builtin_popcountll(a_final ^ b_final);
if (modulo) {
//in the case where size is not dividable by sizeof(size_t)
//need to mask off the bits at the end
pop_t a_final = 0, b_final = 0;
memcpy(&a_final, a2, modulo);
memcpy(&b_final, b2, modulo);
result += __builtin_popcountll(a_final ^ b_final);
}
}
#else
HammingLUT lut;

4
modules/flann/src/miniflann.cpp
View File

@ -312,7 +312,11 @@ buildIndex(void*& index, const Mat& data, const IndexParams& params, const Dista
buildIndex_<Distance, ::cvflann::Index<Distance> >(index, data, params, dist);
}
#if CV_NEON
typedef ::cvflann::Hamming<uchar> HammingDistance;
#else
typedef ::cvflann::HammingLUT HammingDistance;
#endif
typedef ::cvflann::LshIndex<HammingDistance> LshIndex;
Index::Index()

7
modules/flann/src/precomp.hpp
View File

@ -5,6 +5,12 @@
#include <cstdarg>
#include <sstream>
#ifdef HAVE_CVCONFIG_H
# include "cvconfig.h"
#endif
#include "opencv2/core/core.hpp"
#include "opencv2/core/internal.hpp"
#include "opencv2/flann/miniflann.hpp"
#include "opencv2/flann/dist.h"
#include "opencv2/flann/index_testing.h"
@ -15,7 +21,6 @@
// index types
#include "opencv2/flann/all_indices.h"
#include "opencv2/flann/flann_base.hpp"
#endif