Multiple fixes for WinRT

Fixed flann build with NEON;
Fixed Haming distance with NEON;
Honest cvRound for WinRT added;
cvRound test added;
Video IO with direct show disabled;
This commit is contained in:
Alexander Smorkalov
2013-03-26 17:19:52 -07:00
parent 7ec2b6bad0
commit 6f68640d4d
9 changed files with 59 additions and 54 deletions

View File

@@ -136,7 +136,6 @@ CV_INLINE IppiSize ippiSize(int width, int height)
#ifdef __ARM_NEON__
# include <arm_neon.h>
# define CV_NEON 1
# define CPU_HAS_NEON_FEATURE (true)
#endif
#ifndef CV_SSE

View File

@@ -323,7 +323,12 @@ CV_INLINE int cvRound( double value )
# endif
#else
// while this is not IEEE754-compliant rounding, it's usually a good enough approximation
return (int)(value + (value >= 0 ? 0.5 : -0.5));
double intpart, fractpart;
fractpart = modf(value, &intpart);
if ((abs(fractpart) != 0.5) || ((((int)intpart) % 2) != 0))
return (int)(value + (value >= 0 ? 0.5 : -0.5));
else
return (int)intpart;
#endif
}

View File

@@ -999,25 +999,22 @@ static int normHamming(const uchar* a, int n)
{
int i = 0, result = 0;
#if CV_NEON
if (CPU_HAS_NEON_FEATURE)
{
uint32x4_t bits = vmovq_n_u32(0);
for (; i <= n - 16; i += 16) {
uint8x16_t A_vec = vld1q_u8 (a + i);
uint8x16_t bitsSet = vcntq_u8 (A_vec);
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
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);
uint32x4_t bits = vmovq_n_u32(0);
for (; i <= n - 16; i += 16) {
uint8x16_t A_vec = vld1q_u8 (a + i);
uint8x16_t bitsSet = vcntq_u8 (A_vec);
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
bits = vaddq_u32(bits, bitSet4);
}
else
#endif
for( ; i <= n - 4; i += 4 )
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
for( ; i <= n - 4; i += 4 )
result += popCountTable[a[i]] + popCountTable[a[i+1]] +
popCountTable[a[i+2]] + popCountTable[a[i+3]];
#endif
for( ; i < n; i++ )
result += popCountTable[a[i]];
return result;
@@ -1027,27 +1024,24 @@ int normHamming(const uchar* a, const uchar* b, int n)
{
int i = 0, result = 0;
#if CV_NEON
if (CPU_HAS_NEON_FEATURE)
{
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 AxorB = veorq_u8 (A_vec, B_vec);
uint8x16_t bitsSet = vcntq_u8 (AxorB);
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
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);
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 AxorB = veorq_u8 (A_vec, B_vec);
uint8x16_t bitsSet = vcntq_u8 (AxorB);
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
bits = vaddq_u32(bits, bitSet4);
}
else
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
for( ; i <= n - 4; i += 4 )
result += popCountTable[a[i] ^ b[i]] + popCountTable[a[i+1] ^ b[i+1]] +
popCountTable[a[i+2] ^ b[i+2]] + popCountTable[a[i+3] ^ b[i+3]];
#endif
for( ; i <= n - 4; i += 4 )
result += popCountTable[a[i] ^ b[i]] + popCountTable[a[i+1] ^ b[i+1]] +
popCountTable[a[i+2] ^ b[i+2]] + popCountTable[a[i+3] ^ b[i+3]];
for( ; i < n; i++ )
result += popCountTable[a[i] ^ b[i]];
return result;

View File

@@ -1551,3 +1551,16 @@ TEST(Core_Add, AddToColumnWhen4Rows)
ASSERT_EQ(0, countNonZero(m1 - m2));
}
TEST(Core_round, CvRound)
{
ASSERT_EQ(2, cvRound(2.0));
ASSERT_EQ(2, cvRound(2.1));
ASSERT_EQ(-2, cvRound(-2.1));
ASSERT_EQ(3, cvRound(2.8));
ASSERT_EQ(-3, cvRound(-2.8));
ASSERT_EQ(2, cvRound(2.5));
ASSERT_EQ(4, cvRound(3.5));
ASSERT_EQ(-2, cvRound(-2.5));
ASSERT_EQ(-4, cvRound(-3.5));
}

View File

@@ -456,7 +456,6 @@ struct Hamming
ResultType operator()(Iterator1 a, Iterator2 b, size_t size, ResultType /*worst_dist*/ = -1) const
{
ResultType result = 0;
#ifdef __GNUC__
#ifdef __ARM_NEON__
{
uint32x4_t bits = vmovq_n_u32(0);
@@ -473,7 +472,7 @@ struct Hamming
result = vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),0);
result += vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),2);
}
#else
#elif __GNUC__
{
//for portability just use unsigned long -- and use the __builtin_popcountll (see docs for __builtin_popcountll)
typedef unsigned long long pop_t;
@@ -493,8 +492,8 @@ struct Hamming
result += __builtin_popcountll(a_final ^ b_final);
}
}
#endif //NEON
#else
#else // NO NEON and NOT GNUC
typedef unsigned long long pop_t;
HammingLUT lut;
result = lut(reinterpret_cast<const unsigned char*> (a),
reinterpret_cast<const unsigned char*> (b), size * sizeof(pop_t));

View File

@@ -89,10 +89,8 @@ if(HAVE_QT)
if(${_have_flag})
set_source_files_properties(${_RCC_OUTFILES} PROPERTIES COMPILE_FLAGS -Wno-missing-declarations)
endif()
elseif(WIN32)
if (NOT ARM)
list(APPEND highgui_srcs src/window_w32.cpp)
endif()
elseif(WIN32 AND NOT ARM)
list(APPEND highgui_srcs src/window_w32.cpp)
elseif(HAVE_GTK)
list(APPEND highgui_srcs src/window_gtk.cpp)
elseif(APPLE)
@@ -107,12 +105,9 @@ elseif(APPLE)
endif()
endif()
if(WIN32)
list(APPEND highgui_srcs src/cap_dshow.cpp)
if (NOT ARM)
list(APPEND highgui_srcs src/cap_vfw.cpp src/cap_cmu.cpp)
endif()
endif(WIN32)
if(WIN32 AND NOT ARM)
list(APPEND highgui_srcs src/cap_dshow.cpp src/cap_vfw.cpp src/cap_cmu.cpp)
endif()
if(HAVE_XINE)
list(APPEND highgui_srcs src/cap_xine.cpp)