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

Refactored NEON optimization usage

Browse Source
This commit is contained in:
Andrey Kamaev
2011-05-06 12:24:56 +00:00
parent 94a5bf88d0
commit 9f29506d2c
5 changed files with 45 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
modules/core/include/opencv2/core/internal.hpp
View File

@@ -122,8 +122,13 @@ CV_INLINE IppiSize ippiSize(int width, int height)
#if defined ANDROID && defined __ARM_NEON__
#include "arm_neon.h"
#define CV_NEON 1
#define CPU_HAS_NEON_FEATURE (true)
//TODO: make real check using stuff from "cpu-features.h"
//((bool)android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON)
#else
#define CV_NEON 0
#define CPU_HAS_NEON_FEATURE (false)
#endif
#ifndef IPPI_CALL

50
modules/features2d/src/brief.cpp
View File

@@ -44,11 +44,6 @@
#include <algorithm>
#include <vector>
#if ANDROID && HAVE_NEON
#include <cpu-features.h>
#include <arm_neon.h>
#endif
#include <iostream>
#include <iomanip>
@@ -115,9 +110,8 @@ Hamming::ResultType Hamming::operator()(const unsigned char* a, const unsigned c
{
#if __GNUC__
ResultType result = 0;
#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)
{
for (size_t i = 0; i < size; i += 16)
{
@@ -126,7 +120,7 @@ Hamming::ResultType Hamming::operator()(const unsigned char* a, const unsigned c
//uint8x16_t veorq_u8 (uint8x16_t, uint8x16_t)
uint8x16_t AxorB = veorq_u8 (A_vec, B_vec);
uint8x16_t bitsSet += vcntq_u8 (AxorB);
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);
@@ -138,25 +132,27 @@ Hamming::ResultType Hamming::operator()(const unsigned char* a, const unsigned c
}
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 (; a2 != a2_end; ++a2, ++b2)
result += __builtin_popcountll((*a2) ^ (*b2));
if (modulo)
{
//in the case where size is not divisible 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);
}
//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));
if (modulo)
{
//in the case where size is not divisible 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);
}
}
return result;
#else
return HammingLUT()(a,b,size);