Skip to content

O
opencv
Commit 78a5aacb authored by Tomoaki Teshima's avatar Tomoaki Teshima
Browse files
Options

make NEON more similart to SSE2

parent 1b6e0279
Show whitespace changes
Inline Side-by-side
Showing with 64 additions and 7 deletions
modules/imgproc/src/smooth.cpp
View file @ 78a5aacb
...@@ -229,6 +229,8 @@ struct ColumnSum<int, uchar> : ...@@ -229,6 +229,8 @@ struct ColumnSum<int, uchar> :
#if CV_SSE2 #if CV_SSE2
bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2); bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
#elif CV_NEON
bool haveNEON = checkHardwareSupport(CV_CPU_NEON);
#endif #endif
if( width != (int)sum.size() ) if( width != (int)sum.size() )
...@@ -256,8 +258,11 @@ struct ColumnSum<int, uchar> : ...@@ -256,8 +258,11 @@ struct ColumnSum<int, uchar> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width - 4; i+=4 ) for( ; i <= width - 4; i+=4 )
vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i))); vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
SUM[i] += Sp[i]; SUM[i] += Sp[i];
...@@ -303,6 +308,8 @@ struct ColumnSum<int, uchar> : ...@@ -303,6 +308,8 @@ struct ColumnSum<int, uchar> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
float32x4_t v_scale = vdupq_n_f32((float)_scale); float32x4_t v_scale = vdupq_n_f32((float)_scale);
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
...@@ -318,6 +325,7 @@ struct ColumnSum<int, uchar> : ...@@ -318,6 +325,7 @@ struct ColumnSum<int, uchar> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
{ {
...@@ -351,6 +359,8 @@ struct ColumnSum<int, uchar> : ...@@ -351,6 +359,8 @@ struct ColumnSum<int, uchar> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)); int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
...@@ -362,6 +372,7 @@ struct ColumnSum<int, uchar> : ...@@ -362,6 +372,7 @@ struct ColumnSum<int, uchar> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
...@@ -404,6 +415,8 @@ struct ColumnSum<int, short> : ...@@ -404,6 +415,8 @@ struct ColumnSum<int, short> :
#if CV_SSE2 #if CV_SSE2
bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2); bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
#elif CV_NEON
bool haveNEON = checkHardwareSupport(CV_CPU_NEON);
#endif #endif
if( width != (int)sum.size() ) if( width != (int)sum.size() )
...@@ -411,6 +424,7 @@ struct ColumnSum<int, short> : ...@@ -411,6 +424,7 @@ struct ColumnSum<int, short> :
sum.resize(width); sum.resize(width);
sumCount = 0; sumCount = 0;
} }
SUM = &sum[0]; SUM = &sum[0];
if( sumCount == 0 ) if( sumCount == 0 )
{ {
...@@ -430,8 +444,11 @@ struct ColumnSum<int, short> : ...@@ -430,8 +444,11 @@ struct ColumnSum<int, short> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width - 4; i+=4 ) for( ; i <= width - 4; i+=4 )
vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i))); vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
SUM[i] += Sp[i]; SUM[i] += Sp[i];
...@@ -475,6 +492,8 @@ struct ColumnSum<int, short> : ...@@ -475,6 +492,8 @@ struct ColumnSum<int, short> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
float32x4_t v_scale = vdupq_n_f32((float)_scale); float32x4_t v_scale = vdupq_n_f32((float)_scale);
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
...@@ -488,6 +507,7 @@ struct ColumnSum<int, short> : ...@@ -488,6 +507,7 @@ struct ColumnSum<int, short> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
{ {
...@@ -520,6 +540,8 @@ struct ColumnSum<int, short> : ...@@ -520,6 +540,8 @@ struct ColumnSum<int, short> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)); int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
...@@ -530,6 +552,7 @@ struct ColumnSum<int, short> : ...@@ -530,6 +552,7 @@ struct ColumnSum<int, short> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
...@@ -570,8 +593,11 @@ struct ColumnSum<int, ushort> : ...@@ -570,8 +593,11 @@ struct ColumnSum<int, ushort> :
int* SUM; int* SUM;
bool haveScale = scale != 1; bool haveScale = scale != 1;
double _scale = scale; double _scale = scale;
#if CV_SSE2 #if CV_SSE2
bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2); bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
#elif CV_NEON
bool haveNEON = checkHardwareSupport(CV_CPU_NEON);
#endif #endif
if( width != (int)sum.size() ) if( width != (int)sum.size() )
...@@ -579,6 +605,7 @@ struct ColumnSum<int, ushort> : ...@@ -579,6 +605,7 @@ struct ColumnSum<int, ushort> :
sum.resize(width); sum.resize(width);
sumCount = 0; sumCount = 0;
} }
SUM = &sum[0]; SUM = &sum[0];
if( sumCount == 0 ) if( sumCount == 0 )
{ {
...@@ -590,16 +617,19 @@ struct ColumnSum<int, ushort> : ...@@ -590,16 +617,19 @@ struct ColumnSum<int, ushort> :
#if CV_SSE2 #if CV_SSE2
if(haveSSE2) if(haveSSE2)
{ {
for( ; i < width-4; i+=4 ) for( ; i <= width-4; i+=4 )
{ {
__m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i)); __m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i));
__m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i)); __m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i));
_mm_storeu_si128((__m128i*)(SUM+i), _mm_add_epi32(_sum, _sp)); _mm_storeu_si128((__m128i*)(SUM+i),_mm_add_epi32(_sum, _sp));
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width - 4; i+=4 ) for( ; i <= width - 4; i+=4 )
vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i))); vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
SUM[i] += Sp[i]; SUM[i] += Sp[i];
...@@ -642,6 +672,8 @@ struct ColumnSum<int, ushort> : ...@@ -642,6 +672,8 @@ struct ColumnSum<int, ushort> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
float32x4_t v_scale = vdupq_n_f32((float)_scale); float32x4_t v_scale = vdupq_n_f32((float)_scale);
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
...@@ -655,6 +687,7 @@ struct ColumnSum<int, ushort> : ...@@ -655,6 +687,7 @@ struct ColumnSum<int, ushort> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
{ {
...@@ -686,6 +719,8 @@ struct ColumnSum<int, ushort> : ...@@ -686,6 +719,8 @@ struct ColumnSum<int, ushort> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)); int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
...@@ -696,6 +731,7 @@ struct ColumnSum<int, ushort> : ...@@ -696,6 +731,7 @@ struct ColumnSum<int, ushort> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
...@@ -738,6 +774,8 @@ struct ColumnSum<int, int> : ...@@ -738,6 +774,8 @@ struct ColumnSum<int, int> :
#if CV_SSE2 #if CV_SSE2
bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2); bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
#elif CV_NEON
bool haveNEON = checkHardwareSupport(CV_CPU_NEON);
#endif #endif
if( width != (int)sum.size() ) if( width != (int)sum.size() )
...@@ -745,6 +783,7 @@ struct ColumnSum<int, int> : ...@@ -745,6 +783,7 @@ struct ColumnSum<int, int> :
sum.resize(width); sum.resize(width);
sumCount = 0; sumCount = 0;
} }
SUM = &sum[0]; SUM = &sum[0];
if( sumCount == 0 ) if( sumCount == 0 )
{ {
...@@ -764,8 +803,11 @@ struct ColumnSum<int, int> : ...@@ -764,8 +803,11 @@ struct ColumnSum<int, int> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width - 4; i+=4 ) for( ; i <= width - 4; i+=4 )
vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i))); vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
SUM[i] += Sp[i]; SUM[i] += Sp[i];
...@@ -803,6 +845,8 @@ struct ColumnSum<int, int> : ...@@ -803,6 +845,8 @@ struct ColumnSum<int, int> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
float32x4_t v_scale = vdupq_n_f32((float)_scale); float32x4_t v_scale = vdupq_n_f32((float)_scale);
for( ; i <= width-4; i+=4 ) for( ; i <= width-4; i+=4 )
{ {
...@@ -813,6 +857,7 @@ struct ColumnSum<int, int> : ...@@ -813,6 +857,7 @@ struct ColumnSum<int, int> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
{ {
...@@ -838,6 +883,8 @@ struct ColumnSum<int, int> : ...@@ -838,6 +883,8 @@ struct ColumnSum<int, int> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width-4; i+=4 ) for( ; i <= width-4; i+=4 )
{ {
int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)); int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
...@@ -845,6 +892,7 @@ struct ColumnSum<int, int> : ...@@ -845,6 +892,7 @@ struct ColumnSum<int, int> :
vst1q_s32(D + i, v_s0); vst1q_s32(D + i, v_s0);
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
...@@ -888,6 +936,8 @@ struct ColumnSum<int, float> : ...@@ -888,6 +936,8 @@ struct ColumnSum<int, float> :
#if CV_SSE2 #if CV_SSE2
bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2); bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
#elif CV_NEON
bool haveNEON = checkHardwareSupport(CV_CPU_NEON);
#endif #endif
if( width != (int)sum.size() ) if( width != (int)sum.size() )
...@@ -899,26 +949,27 @@ struct ColumnSum<int, float> : ...@@ -899,26 +949,27 @@ struct ColumnSum<int, float> :
SUM = &sum[0]; SUM = &sum[0];
if( sumCount == 0 ) if( sumCount == 0 )
{ {
memset((void *)SUM, 0, sizeof(int) * width); memset((void*)SUM, 0, width*sizeof(int));
for( ; sumCount < ksize - 1; sumCount++, src++ ) for( ; sumCount < ksize - 1; sumCount++, src++ )
{ {
const int* Sp = (const int*)src[0]; const int* Sp = (const int*)src[0];
i = 0; i = 0;
#if CV_SSE2 #if CV_SSE2
if(haveSSE2) if(haveSSE2)
{ {
for( ; i < width-4; i+=4 ) for( ; i <= width-4; i+=4 )
{ {
__m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i)); __m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i));
__m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i)); __m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i));
_mm_storeu_si128((__m128i*)(SUM+i), _mm_add_epi32(_sum, _sp)); _mm_storeu_si128((__m128i*)(SUM+i),_mm_add_epi32(_sum, _sp));
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width - 4; i+=4 ) for( ; i <= width - 4; i+=4 )
vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i))); vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
...@@ -956,6 +1007,8 @@ struct ColumnSum<int, float> : ...@@ -956,6 +1007,8 @@ struct ColumnSum<int, float> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
float32x4_t v_scale = vdupq_n_f32((float)_scale); float32x4_t v_scale = vdupq_n_f32((float)_scale);
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
...@@ -968,6 +1021,7 @@ struct ColumnSum<int, float> : ...@@ -968,6 +1021,7 @@ struct ColumnSum<int, float> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
...@@ -995,6 +1049,8 @@ struct ColumnSum<int, float> : ...@@ -995,6 +1049,8 @@ struct ColumnSum<int, float> :
} }
} }
#elif CV_NEON #elif CV_NEON
if(haveNEON)
{
for( ; i <= width-8; i+=8 ) for( ; i <= width-8; i+=8 )
{ {
int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)); int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
...@@ -1006,6 +1062,7 @@ struct ColumnSum<int, float> : ...@@ -1006,6 +1062,7 @@ struct ColumnSum<int, float> :
vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i))); vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4))); vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
} }
}
#endif #endif
for( ; i < width; i++ ) for( ; i < width; i++ )
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment