Skip to content

O
opencv
Commit f1060ac5 authored by yao's avatar yao
Browse files
Options

a little simplify to pyrlk kernel

parent cdd9234f
Hide whitespace changes
Inline Side-by-side
Showing with 464 additions and 585 deletions
modules/ocl/perf/perf_opticalflow.cpp
View file @ f1060ac5
...@@ -48,8 +48,8 @@ ...@@ -48,8 +48,8 @@
///////////// PyrLKOpticalFlow //////////////////////// ///////////// PyrLKOpticalFlow ////////////////////////
PERFTEST(PyrLKOpticalFlow) PERFTEST(PyrLKOpticalFlow)
{ {
std::string images1[] = {"rubberwhale1.png", "basketball1.png"}; std::string images1[] = {"rubberwhale1.png", "aloeL.jpg"};
std::string images2[] = {"rubberwhale2.png", "basketball2.png"}; std::string images2[] = {"rubberwhale2.png", "aloeR.jpg"};
for (size_t i = 0; i < sizeof(images1) / sizeof(std::string); i++) for (size_t i = 0; i < sizeof(images1) / sizeof(std::string); i++)
{ {
......
modules/ocl/src/opencl/pyrlk.cl
View file @ f1060ac5
...@@ -17,6 +17,7 @@ ...@@ -17,6 +17,7 @@
// @Authors // @Authors
// Dachuan Zhao, dachuan@multicorewareinc.com // Dachuan Zhao, dachuan@multicorewareinc.com
// Yao Wang, bitwangyaoyao@gmail.com // Yao Wang, bitwangyaoyao@gmail.com
// Xiaopeng Fu, fuxiaopeng2222@163.com
// //
// Redistribution and use in source and binary forms, with or without modification, // Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met: // are permitted provided that the following conditions are met:
...@@ -47,6 +48,7 @@ ...@@ -47,6 +48,7 @@
//#pragma OPENCL EXTENSION cl_amd_printf : enable //#pragma OPENCL EXTENSION cl_amd_printf : enable
#define BUFFER 64 #define BUFFER 64
#define BUFFER2 BUFFER>>1
#ifndef WAVE_SIZE #ifndef WAVE_SIZE
#define WAVE_SIZE 1 #define WAVE_SIZE 1
#endif #endif
...@@ -58,53 +60,16 @@ void reduce3(float val1, float val2, float val3, __local float* smem1, __local ...@@ -58,53 +60,16 @@ void reduce3(float val1, float val2, float val3, __local float* smem1, __local
smem3[tid] = val3; smem3[tid] = val3;
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 32) for(int i = BUFFER2; i > 0; i >>= 1)
{
smem1[tid] += smem1[tid + 32];
smem2[tid] += smem2[tid + 32];
smem3[tid] += smem3[tid + 32];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16)
{
smem1[tid] += smem1[tid + 16];
smem2[tid] += smem2[tid + 16];
smem3[tid] += smem3[tid + 16];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 8)
{
smem1[tid] += smem1[tid + 8];
smem2[tid] += smem2[tid + 8];
smem3[tid] += smem3[tid + 8];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 4)
{
smem1[tid] += smem1[tid + 4];
smem2[tid] += smem2[tid + 4];
smem3[tid] += smem3[tid + 4];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 2)
{
smem1[tid] += smem1[tid + 2];
smem2[tid] += smem2[tid + 2];
smem3[tid] += smem3[tid + 2];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 1)
{ {
smem1[BUFFER] = smem1[tid] + smem1[tid + 1]; if(tid < i)
smem2[BUFFER] = smem2[tid] + smem2[tid + 1]; {
smem3[BUFFER] = smem3[tid] + smem3[tid + 1]; smem1[tid] += smem1[tid + i];
smem2[tid] += smem2[tid + i];
smem3[tid] += smem3[tid + i];
}
barrier(CLK_LOCAL_MEM_FENCE);
} }
barrier(CLK_LOCAL_MEM_FENCE);
} }
void reduce2(float val1, float val2, volatile __local float* smem1, volatile __local float* smem2, int tid) void reduce2(float val1, float val2, volatile __local float* smem1, volatile __local float* smem2, int tid)
...@@ -113,47 +78,15 @@ void reduce2(float val1, float val2, volatile __local float* smem1, volatile __l ...@@ -113,47 +78,15 @@ void reduce2(float val1, float val2, volatile __local float* smem1, volatile __l
smem2[tid] = val2; smem2[tid] = val2;
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 32) for(int i = BUFFER2; i > 0; i >>= 1)
{
smem1[tid] += smem1[tid + 32];
smem2[tid] += smem2[tid + 32];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16)
{ {
smem1[tid] += smem1[tid + 16]; if(tid < i)
smem2[tid] += smem2[tid + 16]; {
} smem1[tid] += smem1[tid + i];
barrier(CLK_LOCAL_MEM_FENCE); smem2[tid] += smem2[tid + i];
}
if (tid < 8) barrier(CLK_LOCAL_MEM_FENCE);
{
smem1[tid] += smem1[tid + 8];
smem2[tid] += smem2[tid + 8];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 4)
{
smem1[tid] += smem1[tid + 4];
smem2[tid] += smem2[tid + 4];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 2)
{
smem1[tid] += smem1[tid + 2];
smem2[tid] += smem2[tid + 2];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 1)
{
smem1[BUFFER] = smem1[tid] + smem1[tid + 1];
smem2[BUFFER] = smem2[tid] + smem2[tid + 1];
} }
barrier(CLK_LOCAL_MEM_FENCE);
} }
void reduce1(float val1, volatile __local float* smem1, int tid) void reduce1(float val1, volatile __local float* smem1, int tid)
...@@ -161,45 +94,18 @@ void reduce1(float val1, volatile __local float* smem1, int tid) ...@@ -161,45 +94,18 @@ void reduce1(float val1, volatile __local float* smem1, int tid)
smem1[tid] = val1; smem1[tid] = val1;
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 32) for(int i = BUFFER2; i > 0; i >>= 1)
{ {
smem1[tid] += smem1[tid + 32]; if(tid < i)
} {
barrier(CLK_LOCAL_MEM_FENCE); smem1[tid] += smem1[tid + i];
}
if (tid < 16) barrier(CLK_LOCAL_MEM_FENCE);
{
smem1[tid] += smem1[tid + 16];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 8)
{
smem1[tid] += smem1[tid + 8];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 4)
{
smem1[tid] += smem1[tid + 4];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 2)
{
smem1[tid] += smem1[tid + 2];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 1)
{
smem1[BUFFER] = smem1[tid] + smem1[tid + 1];
} }
barrier(CLK_LOCAL_MEM_FENCE);
} }
#else #else
void reduce3(float val1, float val2, float val3, void reduce3(float val1, float val2, float val3,
__local volatile float* smem1, __local volatile float* smem2, __local volatile float* smem3, int tid) __local volatile float* smem1, __local volatile float* smem2, __local volatile float* smem3, int tid)
{ {
smem1[tid] = val1; smem1[tid] = val1;
smem2[tid] = val2; smem2[tid] = val2;
...@@ -212,15 +118,19 @@ __local volatile float* smem1, __local volatile float* smem2, __local volatile f ...@@ -212,15 +118,19 @@ __local volatile float* smem1, __local volatile float* smem2, __local volatile f
smem2[tid] += smem2[tid + 32]; smem2[tid] += smem2[tid + 32];
smem3[tid] += smem3[tid + 32]; smem3[tid] += smem3[tid + 32];
#if WAVE_SIZE < 32 #if WAVE_SIZE < 32
} barrier(CLK_LOCAL_MEM_FENCE); }
if (tid < 16) { barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16)
{
#endif #endif
smem1[tid] += smem1[tid + 16]; smem1[tid] += smem1[tid + 16];
smem2[tid] += smem2[tid + 16]; smem2[tid] += smem2[tid + 16];
smem3[tid] += smem3[tid + 16]; smem3[tid] += smem3[tid + 16];
#if WAVE_SIZE <16 #if WAVE_SIZE <16
} barrier(CLK_LOCAL_MEM_FENCE); }
if (tid < 8) { barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 8)
{
#endif #endif
smem1[tid] += smem1[tid + 8]; smem1[tid] += smem1[tid + 8];
smem2[tid] += smem2[tid + 8]; smem2[tid] += smem2[tid + 8];
...@@ -238,6 +148,7 @@ __local volatile float* smem1, __local volatile float* smem2, __local volatile f ...@@ -238,6 +148,7 @@ __local volatile float* smem1, __local volatile float* smem2, __local volatile f
smem2[tid] += smem2[tid + 1]; smem2[tid] += smem2[tid + 1];
smem3[tid] += smem3[tid + 1]; smem3[tid] += smem3[tid + 1];
} }
barrier(CLK_LOCAL_MEM_FENCE);
} }
void reduce2(float val1, float val2, __local volatile float* smem1, __local volatile float* smem2, int tid) void reduce2(float val1, float val2, __local volatile float* smem1, __local volatile float* smem2, int tid)
...@@ -251,14 +162,18 @@ void reduce2(float val1, float val2, __local volatile float* smem1, __local vola ...@@ -251,14 +162,18 @@ void reduce2(float val1, float val2, __local volatile float* smem1, __local vola
smem1[tid] += smem1[tid + 32]; smem1[tid] += smem1[tid + 32];
smem2[tid] += smem2[tid + 32]; smem2[tid] += smem2[tid + 32];
#if WAVE_SIZE < 32 #if WAVE_SIZE < 32
} barrier(CLK_LOCAL_MEM_FENCE); }
if (tid < 16) { barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16)
{
#endif #endif
smem1[tid] += smem1[tid + 16]; smem1[tid] += smem1[tid + 16];
smem2[tid] += smem2[tid + 16]; smem2[tid] += smem2[tid + 16];
#if WAVE_SIZE <16 #if WAVE_SIZE <16
} barrier(CLK_LOCAL_MEM_FENCE); }
if (tid < 8) { barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 8)
{
#endif #endif
smem1[tid] += smem1[tid + 8]; smem1[tid] += smem1[tid + 8];
smem2[tid] += smem2[tid + 8]; smem2[tid] += smem2[tid + 8];
...@@ -272,6 +187,7 @@ void reduce2(float val1, float val2, __local volatile float* smem1, __local vola ...@@ -272,6 +187,7 @@ void reduce2(float val1, float val2, __local volatile float* smem1, __local vola
smem1[tid] += smem1[tid + 1]; smem1[tid] += smem1[tid + 1];
smem2[tid] += smem2[tid + 1]; smem2[tid] += smem2[tid + 1];
} }
barrier(CLK_LOCAL_MEM_FENCE);
} }
void reduce1(float val1, __local volatile float* smem1, int tid) void reduce1(float val1, __local volatile float* smem1, int tid)
...@@ -283,19 +199,24 @@ void reduce1(float val1, __local volatile float* smem1, int tid) ...@@ -283,19 +199,24 @@ void reduce1(float val1, __local volatile float* smem1, int tid)
{ {
smem1[tid] += smem1[tid + 32]; smem1[tid] += smem1[tid + 32];
#if WAVE_SIZE < 32 #if WAVE_SIZE < 32
} barrier(CLK_LOCAL_MEM_FENCE); }
if (tid < 16) { barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16)
{
#endif #endif
smem1[tid] += smem1[tid + 16]; smem1[tid] += smem1[tid + 16];
#if WAVE_SIZE <16 #if WAVE_SIZE <16
} barrier(CLK_LOCAL_MEM_FENCE); }
if (tid < 8) { barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 8)
{
#endif #endif
smem1[tid] += smem1[tid + 8]; smem1[tid] += smem1[tid + 8];
smem1[tid] += smem1[tid + 4]; smem1[tid] += smem1[tid + 4];
smem1[tid] += smem1[tid + 2]; smem1[tid] += smem1[tid + 2];
smem1[tid] += smem1[tid + 1]; smem1[tid] += smem1[tid + 1];
} }
barrier(CLK_LOCAL_MEM_FENCE);
} }
#endif #endif
...@@ -306,106 +227,100 @@ void reduce1(float val1, __local volatile float* smem1, int tid) ...@@ -306,106 +227,100 @@ void reduce1(float val1, __local volatile float* smem1, int tid)
__constant sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_LINEAR; __constant sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_LINEAR;
void SetPatch(image2d_t I, float x, float y, void SetPatch(image2d_t I, float x, float y,
float* Pch, float* Dx, float* Dy, float* Pch, float* Dx, float* Dy,
float* A11, float* A12, float* A22) float* A11, float* A12, float* A22)
{ {
*Pch = read_imagef(I, sampler, (float2)(x, y)).x; *Pch = read_imagef(I, sampler, (float2)(x, y)).x;
float dIdx = 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x + 1, y)).x + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)).x - float dIdx = 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x + 1, y)).x + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)).x -
(3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x - 1, y)).x + 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1)).x); (3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x - 1, y)).x + 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1)).x);
float dIdy = 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x, y + 1)).x + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)).x - float dIdy = 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x, y + 1)).x + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)).x -
(3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x, y - 1)).x + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1)).x); (3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)).x + 10.0f * read_imagef(I, sampler, (float2)(x, y - 1)).x + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1)).x);
*Dx = dIdx; *Dx = dIdx;
*Dy = dIdy; *Dy = dIdy;
*A11 += dIdx * dIdx; *A11 += dIdx * dIdx;
*A12 += dIdx * dIdy; *A12 += dIdx * dIdy;
*A22 += dIdy * dIdy; *A22 += dIdy * dIdy;
} }
void GetPatch(image2d_t J, float x, float y, void GetPatch(image2d_t J, float x, float y,
float* Pch, float* Dx, float* Dy, float* Pch, float* Dx, float* Dy,
float* b1, float* b2) float* b1, float* b2)
{ {
float J_val = read_imagef(J, sampler, (float2)(x, y)).x; float J_val = read_imagef(J, sampler, (float2)(x, y)).x;
float diff = (J_val - *Pch) * 32.0f; float diff = (J_val - *Pch) * 32.0f;
*b1 += diff**Dx; *b1 += diff**Dx;
*b2 += diff**Dy; *b2 += diff**Dy;
} }
void GetError(image2d_t J, const float x, const float y, const float* Pch, float* errval) void GetError(image2d_t J, const float x, const float y, const float* Pch, float* errval)
{ {
float diff = read_imagef(J, sampler, (float2)(x,y)).x-*Pch; float diff = read_imagef(J, sampler, (float2)(x,y)).x-*Pch;
*errval += fabs(diff); *errval += fabs(diff);
} }
void SetPatch4(image2d_t I, const float x, const float y, void SetPatch4(image2d_t I, const float x, const float y,
float4* Pch, float4* Dx, float4* Dy, float4* Pch, float4* Dx, float4* Dy,
float* A11, float* A12, float* A22) float* A11, float* A12, float* A22)
{ {
*Pch = read_imagef(I, sampler, (float2)(x, y)); *Pch = read_imagef(I, sampler, (float2)(x, y));
float4 dIdx = 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1)) + 10.0f * read_imagef(I, sampler, (float2)(x + 1, y)) + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)) - float4 dIdx = 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1)) + 10.0f * read_imagef(I, sampler, (float2)(x + 1, y)) + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)) -
(3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)) + 10.0f * read_imagef(I, sampler, (float2)(x - 1, y)) + 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1))); (3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)) + 10.0f * read_imagef(I, sampler, (float2)(x - 1, y)) + 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1)));
float4 dIdy = 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1)) + 10.0f * read_imagef(I, sampler, (float2)(x, y + 1)) + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)) - float4 dIdy = 3.0f * read_imagef(I, sampler, (float2)(x - 1, y + 1)) + 10.0f * read_imagef(I, sampler, (float2)(x, y + 1)) + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y + 1)) -
(3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)) + 10.0f * read_imagef(I, sampler, (float2)(x, y - 1)) + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1))); (3.0f * read_imagef(I, sampler, (float2)(x - 1, y - 1)) + 10.0f * read_imagef(I, sampler, (float2)(x, y - 1)) + 3.0f * read_imagef(I, sampler, (float2)(x + 1, y - 1)));
*Dx = dIdx; *Dx = dIdx;
*Dy = dIdy; *Dy = dIdy;
float4 sqIdx = dIdx * dIdx; float4 sqIdx = dIdx * dIdx;
*A11 += sqIdx.x + sqIdx.y + sqIdx.z; *A11 += sqIdx.x + sqIdx.y + sqIdx.z;
sqIdx = dIdx * dIdy; sqIdx = dIdx * dIdy;
*A12 += sqIdx.x + sqIdx.y + sqIdx.z; *A12 += sqIdx.x + sqIdx.y + sqIdx.z;
sqIdx = dIdy * dIdy; sqIdx = dIdy * dIdy;
*A22 += sqIdx.x + sqIdx.y + sqIdx.z; *A22 += sqIdx.x + sqIdx.y + sqIdx.z;
} }
void GetPatch4(image2d_t J, const float x, const float y, void GetPatch4(image2d_t J, const float x, const float y,
const float4* Pch, const float4* Dx, const float4* Dy, const float4* Pch, const float4* Dx, const float4* Dy,
float* b1, float* b2) float* b1, float* b2)
{ {
float4 J_val = read_imagef(J, sampler, (float2)(x, y)); float4 J_val = read_imagef(J, sampler, (float2)(x, y));
float4 diff = (J_val - *Pch) * 32.0f; float4 diff = (J_val - *Pch) * 32.0f;
float4 xdiff = diff* *Dx; float4 xdiff = diff* *Dx;
*b1 += xdiff.x + xdiff.y + xdiff.z; *b1 += xdiff.x + xdiff.y + xdiff.z;
xdiff = diff* *Dy; xdiff = diff* *Dy;
*b2 += xdiff.x + xdiff.y + xdiff.z; *b2 += xdiff.x + xdiff.y + xdiff.z;
} }
void GetError4(image2d_t J, const float x, const float y, const float4* Pch, float* errval) void GetError4(image2d_t J, const float x, const float y, const float4* Pch, float* errval)
{ {
float4 diff = read_imagef(J, sampler, (float2)(x,y))-*Pch; float4 diff = read_imagef(J, sampler, (float2)(x,y))-*Pch;
*errval += fabs(diff.x) + fabs(diff.y) + fabs(diff.z); *errval += fabs(diff.x) + fabs(diff.y) + fabs(diff.z);
} }
#define GRIDSIZE 3 #define GRIDSIZE 3
__kernel void lkSparse_C1_D5(image2d_t I, image2d_t J, __kernel void lkSparse_C1_D5(image2d_t I, image2d_t J,
__global const float2* prevPts, int prevPtsStep, __global float2* nextPts, int nextPtsStep, __global uchar* status, __global float* err, __global const float2* prevPts, int prevPtsStep, __global float2* nextPts, int nextPtsStep, __global uchar* status, __global float* err,
const int level, const int rows, const int cols, int PATCH_X, int PATCH_Y, int cn, int c_winSize_x, int c_winSize_y, int c_iters, char calcErr) const int level, const int rows, const int cols, int PATCH_X, int PATCH_Y, int cn, int c_winSize_x, int c_winSize_y, int c_iters, char calcErr)
{ {
#ifdef CPU
__local float smem1[BUFFER+1];
__local float smem2[BUFFER+1];
__local float smem3[BUFFER+1];
#else
__local float smem1[BUFFER]; __local float smem1[BUFFER];
__local float smem2[BUFFER]; __local float smem2[BUFFER];
__local float smem3[BUFFER]; __local float smem3[BUFFER];
#endif
unsigned int xid=get_local_id(0); unsigned int xid=get_local_id(0);
unsigned int yid=get_local_id(1); unsigned int yid=get_local_id(1);
unsigned int gid=get_group_id(0); unsigned int gid=get_group_id(0);
unsigned int xsize=get_local_size(0); unsigned int xsize=get_local_size(0);
unsigned int ysize=get_local_size(1); unsigned int ysize=get_local_size(1);
int xBase, yBase, i, j, k; int xBase, yBase, i, j, k;
float2 c_halfWin = (float2)((c_winSize_x - 1)>>1, (c_winSize_y - 1)>>1); float2 c_halfWin = (float2)((c_winSize_x - 1)>>1, (c_winSize_y - 1)>>1);
const int tid = mad24(yid, xsize, xid); const int tid = mad24(yid, xsize, xid);
...@@ -432,77 +347,71 @@ __kernel void lkSparse_C1_D5(image2d_t I, image2d_t J, ...@@ -432,77 +347,71 @@ __kernel void lkSparse_C1_D5(image2d_t I, image2d_t J,
float dIdx_patch[GRIDSIZE][GRIDSIZE]; float dIdx_patch[GRIDSIZE][GRIDSIZE];
float dIdy_patch[GRIDSIZE][GRIDSIZE]; float dIdy_patch[GRIDSIZE][GRIDSIZE];
yBase=yid; yBase=yid;
{ {
xBase=xid; xBase=xid;
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][0], &dIdx_patch[0][0], &dIdy_patch[0][0], &I_patch[0][0], &dIdx_patch[0][0], &dIdy_patch[0][0],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][1], &dIdx_patch[0][1], &dIdy_patch[0][1], &I_patch[0][1], &dIdx_patch[0][1], &dIdy_patch[0][1],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][2], &dIdx_patch[0][2], &dIdy_patch[0][2], &I_patch[0][2], &dIdx_patch[0][2], &dIdy_patch[0][2],
&A11, &A12, &A22); &A11, &A12, &A22);
} }
yBase+=ysize; yBase+=ysize;
{ {
xBase=xid; xBase=xid;
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][0], &dIdx_patch[1][0], &dIdy_patch[1][0], &I_patch[1][0], &dIdx_patch[1][0], &dIdy_patch[1][0],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][1], &dIdx_patch[1][1], &dIdy_patch[1][1], &I_patch[1][1], &dIdx_patch[1][1], &dIdy_patch[1][1],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][2], &dIdx_patch[1][2], &dIdy_patch[1][2], &I_patch[1][2], &dIdx_patch[1][2], &dIdy_patch[1][2],
&A11, &A12, &A22); &A11, &A12, &A22);
} }
yBase+=ysize; yBase+=ysize;
if(yBase<c_winSize_y) if(yBase<c_winSize_y)
{ {
xBase=xid; xBase=xid;
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][0], &dIdx_patch[2][0], &dIdy_patch[2][0], &I_patch[2][0], &dIdx_patch[2][0], &dIdy_patch[2][0],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][1], &dIdx_patch[2][1], &dIdy_patch[2][1], &I_patch[2][1], &dIdx_patch[2][1], &dIdy_patch[2][1],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, SetPatch(I, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][2], &dIdx_patch[2][2], &dIdy_patch[2][2], &I_patch[2][2], &dIdx_patch[2][2], &dIdy_patch[2][2],
&A11, &A12, &A22); &A11, &A12, &A22);
} }
reduce3(A11, A12, A22, smem1, smem2, smem3, tid); reduce3(A11, A12, A22, smem1, smem2, smem3, tid);
barrier(CLK_LOCAL_MEM_FENCE);
#ifdef CPU
A11 = smem1[BUFFER];
A12 = smem2[BUFFER];
A22 = smem3[BUFFER];
#else
A11 = smem1[0]; A11 = smem1[0];
A12 = smem2[0]; A12 = smem2[0];
A22 = smem3[0]; A22 = smem3[0];
#endif barrier(CLK_LOCAL_MEM_FENCE);
float D = A11 * A22 - A12 * A12; float D = A11 * A22 - A12 * A12;
...@@ -532,81 +441,76 @@ __kernel void lkSparse_C1_D5(image2d_t I, image2d_t J, ...@@ -532,81 +441,76 @@ __kernel void lkSparse_C1_D5(image2d_t I, image2d_t J,
float b1 = 0; float b1 = 0;
float b2 = 0; float b2 = 0;
yBase=yid; yBase=yid;
{ {
xBase=xid; xBase=xid;
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][0], &dIdx_patch[0][0], &dIdy_patch[0][0], &I_patch[0][0], &dIdx_patch[0][0], &dIdy_patch[0][0],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][1], &dIdx_patch[0][1], &dIdy_patch[0][1], &I_patch[0][1], &dIdx_patch[0][1], &dIdy_patch[0][1],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][2], &dIdx_patch[0][2], &dIdy_patch[0][2], &I_patch[0][2], &dIdx_patch[0][2], &dIdy_patch[0][2],
&b1, &b2); &b1, &b2);
} }
yBase+=ysize; yBase+=ysize;
{ {
xBase=xid; xBase=xid;
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][0], &dIdx_patch[1][0], &dIdy_patch[1][0], &I_patch[1][0], &dIdx_patch[1][0], &dIdy_patch[1][0],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][1], &dIdx_patch[1][1], &dIdy_patch[1][1], &I_patch[1][1], &dIdx_patch[1][1], &dIdy_patch[1][1],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][2], &dIdx_patch[1][2], &dIdy_patch[1][2], &I_patch[1][2], &dIdx_patch[1][2], &dIdy_patch[1][2],
&b1, &b2); &b1, &b2);
} }
yBase+=ysize; yBase+=ysize;
if(yBase<c_winSize_y) if(yBase<c_winSize_y)
{ {
xBase=xid; xBase=xid;
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][0], &dIdx_patch[2][0], &dIdy_patch[2][0], &I_patch[2][0], &dIdx_patch[2][0], &dIdy_patch[2][0],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][1], &dIdx_patch[2][1], &dIdy_patch[2][1], &I_patch[2][1], &dIdx_patch[2][1], &dIdy_patch[2][1],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetPatch(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][2], &dIdx_patch[2][2], &dIdy_patch[2][2], &I_patch[2][2], &dIdx_patch[2][2], &dIdy_patch[2][2],
&b1, &b2); &b1, &b2);
} }
reduce2(b1, b2, smem1, smem2, tid); reduce2(b1, b2, smem1, smem2, tid);
barrier(CLK_LOCAL_MEM_FENCE);
#ifdef CPU
b1 = smem1[BUFFER];
b2 = smem2[BUFFER];
#else
b1 = smem1[0]; b1 = smem1[0];
b2 = smem2[0]; b2 = smem2[0];
#endif barrier(CLK_LOCAL_MEM_FENCE);
float2 delta; float2 delta;
delta.x = A12 * b2 - A22 * b1; delta.x = A12 * b2 - A22 * b1;
delta.y = A12 * b1 - A11 * b2; delta.y = A12 * b1 - A11 * b2;
prevPt += delta; prevPt += delta;
if (fabs(delta.x) < THRESHOLD && fabs(delta.y) < THRESHOLD) if (fabs(delta.x) < THRESHOLD && fabs(delta.y) < THRESHOLD)
break; break;
...@@ -615,97 +519,87 @@ __kernel void lkSparse_C1_D5(image2d_t I, image2d_t J, ...@@ -615,97 +519,87 @@ __kernel void lkSparse_C1_D5(image2d_t I, image2d_t J,
D = 0.0f; D = 0.0f;
if (calcErr) if (calcErr)
{ {
yBase=yid; yBase=yid;
{ {
xBase=xid; xBase=xid;
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][0], &D); &I_patch[0][0], &D);
xBase+=xsize; xBase+=xsize;
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][1], &D); &I_patch[0][1], &D);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[0][2], &D); &I_patch[0][2], &D);
} }
yBase+=ysize; yBase+=ysize;
{ {
xBase=xid; xBase=xid;
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][0], &D); &I_patch[1][0], &D);
xBase+=xsize; xBase+=xsize;
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][1], &D); &I_patch[1][1], &D);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[1][2], &D); &I_patch[1][2], &D);
} }
yBase+=ysize; yBase+=ysize;
if(yBase<c_winSize_y) if(yBase<c_winSize_y)
{ {
xBase=xid; xBase=xid;
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][0], &D); &I_patch[2][0], &D);
xBase+=xsize; xBase+=xsize;
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][1], &D); &I_patch[2][1], &D);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f, GetError(J, prevPt.x + xBase + 0.5f, prevPt.y + yBase + 0.5f,
&I_patch[2][2], &D); &I_patch[2][2], &D);
} }
reduce1(D, smem1, tid); reduce1(D, smem1, tid);
} }
if (tid == 0) if (tid == 0)
{ {
prevPt += c_halfWin; prevPt += c_halfWin;
nextPts[gid] = prevPt; nextPts[gid] = prevPt;
if (calcErr) if (calcErr)
#ifdef CPU
err[gid] = smem1[BUFFER] / (float)(c_winSize_x * c_winSize_y);
#else
err[gid] = smem1[0] / (float)(c_winSize_x * c_winSize_y); err[gid] = smem1[0] / (float)(c_winSize_x * c_winSize_y);
#endif
} }
} }
__kernel void lkSparse_C4_D5(image2d_t I, image2d_t J, __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J,
__global const float2* prevPts, int prevPtsStep, __global float2* nextPts, int nextPtsStep, __global uchar* status, __global float* err, __global const float2* prevPts, int prevPtsStep, __global float2* nextPts, int nextPtsStep, __global uchar* status, __global float* err,
const int level, const int rows, const int cols, int PATCH_X, int PATCH_Y, int cn, int c_winSize_x, int c_winSize_y, int c_iters, char calcErr) const int level, const int rows, const int cols, int PATCH_X, int PATCH_Y, int cn, int c_winSize_x, int c_winSize_y, int c_iters, char calcErr)
{ {
#ifdef CPU __local float smem1[BUFFER];
__local float smem1[BUFFER+1]; __local float smem2[BUFFER];
__local float smem2[BUFFER+1]; __local float smem3[BUFFER];
__local float smem3[BUFFER+1];
#else
__local float smem1[BUFFER];
__local float smem2[BUFFER];
__local float smem3[BUFFER];
#endif
unsigned int xid=get_local_id(0); unsigned int xid=get_local_id(0);
unsigned int yid=get_local_id(1); unsigned int yid=get_local_id(1);
unsigned int gid=get_group_id(0); unsigned int gid=get_group_id(0);
unsigned int xsize=get_local_size(0); unsigned int xsize=get_local_size(0);
unsigned int ysize=get_local_size(1); unsigned int ysize=get_local_size(1);
int xBase, yBase, i, j, k; int xBase, yBase, i, j, k;
float2 c_halfWin = (float2)((c_winSize_x - 1)>>1, (c_winSize_y - 1)>>1); float2 c_halfWin = (float2)((c_winSize_x - 1)>>1, (c_winSize_y - 1)>>1);
const int tid = mad24(yid, xsize, xid); const int tid = mad24(yid, xsize, xid);
...@@ -721,7 +615,7 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J, ...@@ -721,7 +615,7 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J,
return; return;
} }
nextPt -= c_halfWin; nextPt -= c_halfWin;
// extract the patch from the first image, compute covariation matrix of derivatives // extract the patch from the first image, compute covariation matrix of derivatives
...@@ -732,80 +626,74 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J, ...@@ -732,80 +626,74 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J,
float4 I_patch[8]; float4 I_patch[8];
float4 dIdx_patch[8]; float4 dIdx_patch[8];
float4 dIdy_patch[8]; float4 dIdy_patch[8];
float4 I_add,Dx_add,Dy_add; float4 I_add,Dx_add,Dy_add;
yBase=yid; yBase=yid;
{ {
xBase=xid; xBase=xid;
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[0], &dIdx_patch[0], &dIdy_patch[0], &I_patch[0], &dIdx_patch[0], &dIdy_patch[0],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[1], &dIdx_patch[1], &dIdy_patch[1], &I_patch[1], &dIdx_patch[1], &dIdy_patch[1],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[2], &dIdx_patch[2], &dIdy_patch[2], &I_patch[2], &dIdx_patch[2], &dIdy_patch[2],
&A11, &A12, &A22); &A11, &A12, &A22);
} }
yBase+=ysize; yBase+=ysize;
{ {
xBase=xid; xBase=xid;
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[3], &dIdx_patch[3], &dIdy_patch[3], &I_patch[3], &dIdx_patch[3], &dIdy_patch[3],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[4], &dIdx_patch[4], &dIdy_patch[4], &I_patch[4], &dIdx_patch[4], &dIdy_patch[4],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[5], &dIdx_patch[5], &dIdy_patch[5], &I_patch[5], &dIdx_patch[5], &dIdy_patch[5],
&A11, &A12, &A22); &A11, &A12, &A22);
} }
yBase+=ysize; yBase+=ysize;
if(yBase<c_winSize_y) if(yBase<c_winSize_y)
{ {
xBase=xid; xBase=xid;
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[6], &dIdx_patch[6], &dIdy_patch[6], &I_patch[6], &dIdx_patch[6], &dIdy_patch[6],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[7], &dIdx_patch[7], &dIdy_patch[7], &I_patch[7], &dIdx_patch[7], &dIdy_patch[7],
&A11, &A12, &A22); &A11, &A12, &A22);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, SetPatch4(I, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_add, &Dx_add, &Dy_add, &I_add, &Dx_add, &Dy_add,
&A11, &A12, &A22); &A11, &A12, &A22);
} }
reduce3(A11, A12, A22, smem1, smem2, smem3, tid); reduce3(A11, A12, A22, smem1, smem2, smem3, tid);
barrier(CLK_LOCAL_MEM_FENCE);
#ifdef CPU
A11 = smem1[BUFFER];
A12 = smem2[BUFFER];
A22 = smem3[BUFFER];
#else
A11 = smem1[0]; A11 = smem1[0];
A12 = smem2[0]; A12 = smem2[0];
A22 = smem3[0]; A22 = smem3[0];
#endif barrier(CLK_LOCAL_MEM_FENCE);
float D = A11 * A22 - A12 * A12; float D = A11 * A22 - A12 * A12;
...@@ -821,7 +709,7 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J, ...@@ -821,7 +709,7 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J,
A12 /= D; A12 /= D;
A22 /= D; A22 /= D;
nextPt = nextPts[gid] * 2.0f - c_halfWin; nextPt = nextPts[gid] * 2.0f - c_halfWin;
for (k = 0; k < c_iters; ++k) for (k = 0; k < c_iters; ++k)
{ {
...@@ -835,81 +723,76 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J, ...@@ -835,81 +723,76 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J,
float b1 = 0; float b1 = 0;
float b2 = 0; float b2 = 0;
yBase=yid; yBase=yid;
{ {
xBase=xid; xBase=xid;
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[0], &dIdx_patch[0], &dIdy_patch[0], &I_patch[0], &dIdx_patch[0], &dIdy_patch[0],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[1], &dIdx_patch[1], &dIdy_patch[1], &I_patch[1], &dIdx_patch[1], &dIdy_patch[1],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[2], &dIdx_patch[2], &dIdy_patch[2], &I_patch[2], &dIdx_patch[2], &dIdy_patch[2],
&b1, &b2); &b1, &b2);
} }
yBase+=ysize; yBase+=ysize;
{ {
xBase=xid; xBase=xid;
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[3], &dIdx_patch[3], &dIdy_patch[3], &I_patch[3], &dIdx_patch[3], &dIdy_patch[3],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[4], &dIdx_patch[4], &dIdy_patch[4], &I_patch[4], &dIdx_patch[4], &dIdy_patch[4],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[5], &dIdx_patch[5], &dIdy_patch[5], &I_patch[5], &dIdx_patch[5], &dIdy_patch[5],
&b1, &b2); &b1, &b2);
} }
yBase+=ysize; yBase+=ysize;
if(yBase<c_winSize_y) if(yBase<c_winSize_y)
{ {
xBase=xid; xBase=xid;
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[6], &dIdx_patch[6], &dIdy_patch[6], &I_patch[6], &dIdx_patch[6], &dIdy_patch[6],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[7], &dIdx_patch[7], &dIdy_patch[7], &I_patch[7], &dIdx_patch[7], &dIdy_patch[7],
&b1, &b2); &b1, &b2);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetPatch4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_add, &Dx_add, &Dy_add, &I_add, &Dx_add, &Dy_add,
&b1, &b2); &b1, &b2);
} }
reduce2(b1, b2, smem1, smem2, tid); reduce2(b1, b2, smem1, smem2, tid);
barrier(CLK_LOCAL_MEM_FENCE);
#ifdef CPU
b1 = smem1[BUFFER];
b2 = smem2[BUFFER];
#else
b1 = smem1[0]; b1 = smem1[0];
b2 = smem2[0]; b2 = smem2[0];
#endif barrier(CLK_LOCAL_MEM_FENCE);
float2 delta; float2 delta;
delta.x = A12 * b2 - A22 * b1; delta.x = A12 * b2 - A22 * b1;
delta.y = A12 * b1 - A11 * b2; delta.y = A12 * b1 - A11 * b2;
nextPt +=delta; nextPt +=delta;
if (fabs(delta.x) < THRESHOLD && fabs(delta.y) < THRESHOLD) if (fabs(delta.x) < THRESHOLD && fabs(delta.y) < THRESHOLD)
break; break;
...@@ -918,78 +801,74 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J, ...@@ -918,78 +801,74 @@ __kernel void lkSparse_C4_D5(image2d_t I, image2d_t J,
D = 0.0f; D = 0.0f;
if (calcErr) if (calcErr)
{ {
yBase=yid; yBase=yid;
{ {
xBase=xid; xBase=xid;
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[0], &D); &I_patch[0], &D);
xBase+=xsize; xBase+=xsize;
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[1], &D); &I_patch[1], &D);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[2], &D); &I_patch[2], &D);
} }
yBase+=ysize; yBase+=ysize;
{ {
xBase=xid; xBase=xid;
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[3], &D); &I_patch[3], &D);
xBase+=xsize; xBase+=xsize;
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[4], &D); &I_patch[4], &D);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[5], &D); &I_patch[5], &D);
} }
yBase+=ysize; yBase+=ysize;
if(yBase<c_winSize_y) if(yBase<c_winSize_y)
{ {
xBase=xid; xBase=xid;
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[6], &D); &I_patch[6], &D);
xBase+=xsize; xBase+=xsize;
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_patch[7], &D); &I_patch[7], &D);
xBase+=xsize; xBase+=xsize;
if(xBase<c_winSize_x) if(xBase<c_winSize_x)
GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f, GetError4(J, nextPt.x + xBase + 0.5f, nextPt.y + yBase + 0.5f,
&I_add, &D); &I_add, &D);
} }
reduce1(D, smem1, tid); reduce1(D, smem1, tid);
} }
if (tid == 0) if (tid == 0)
{ {
nextPt += c_halfWin; nextPt += c_halfWin;
nextPts[gid] = nextPt; nextPts[gid] = nextPt;
if (calcErr) if (calcErr)
#ifdef CPU
err[gid] = smem1[BUFFER] / (float)(3 * c_winSize_x * c_winSize_y);
#else
err[gid] = smem1[0] / (float)(3 * c_winSize_x * c_winSize_y); err[gid] = smem1[0] / (float)(3 * c_winSize_x * c_winSize_y);
#endif
} }
} }
__kernel void lkDense_C1_D0(image2d_t I, image2d_t J, __global float* u, int uStep, __global float* v, int vStep, __global const float* prevU, int prevUStep, __global const float* prevV, int prevVStep, __kernel void lkDense_C1_D0(image2d_t I, image2d_t J, __global float* u, int uStep, __global float* v, int vStep, __global const float* prevU, int prevUStep, __global const float* prevV, int prevVStep,
const int rows, const int cols, /*__global float* err, int errStep, int cn,*/ int c_winSize_x, int c_winSize_y, int c_iters, char calcErr) const int rows, const int cols, /*__global float* err, int errStep, int cn,*/ int c_winSize_x, int c_winSize_y, int c_iters, char calcErr)
{ {
int c_halfWin_x = (c_winSize_x - 1) / 2; int c_halfWin_x = (c_winSize_x - 1) / 2;
int c_halfWin_y = (c_winSize_y - 1) / 2; int c_halfWin_y = (c_winSize_y - 1) / 2;
const int patchWidth = get_local_size(0) + 2 * c_halfWin_x; const int patchWidth = get_local_size(0) + 2 * c_halfWin_x;
const int patchHeight = get_local_size(1) + 2 * c_halfWin_y; const int patchHeight = get_local_size(1) + 2 * c_halfWin_y;
...@@ -1003,7 +882,7 @@ __kernel void lkDense_C1_D0(image2d_t I, image2d_t J, __global float* u, int uSt ...@@ -1003,7 +882,7 @@ __kernel void lkDense_C1_D0(image2d_t I, image2d_t J, __global float* u, int uSt
const int xBase = get_group_id(0) * get_local_size(0); const int xBase = get_group_id(0) * get_local_size(0);
const int yBase = get_group_id(1) * get_local_size(1); const int yBase = get_group_id(1) * get_local_size(1);
sampler_t sampleri = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST; sampler_t sampleri = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
for (int i = get_local_id(1); i < patchHeight; i += get_local_size(1)) for (int i = get_local_id(1); i < patchHeight; i += get_local_size(1))
{ {
...@@ -1017,10 +896,10 @@ __kernel void lkDense_C1_D0(image2d_t I, image2d_t J, __global float* u, int uSt ...@@ -1017,10 +896,10 @@ __kernel void lkDense_C1_D0(image2d_t I, image2d_t J, __global float* u, int uSt
// Sharr Deriv // Sharr Deriv
dIdx_patch[i * patchWidth + j] = 3 * read_imagei(I, sampleri, (float2)(x+1, y-1)).x + 10 * read_imagei(I, sampleri, (float2)(x+1, y)).x + 3 * read_imagei(I, sampleri, (float2)(x+1, y+1)).x - dIdx_patch[i * patchWidth + j] = 3 * read_imagei(I, sampleri, (float2)(x+1, y-1)).x + 10 * read_imagei(I, sampleri, (float2)(x+1, y)).x + 3 * read_imagei(I, sampleri, (float2)(x+1, y+1)).x -
(3 * read_imagei(I, sampleri, (float2)(x-1, y-1)).x + 10 * read_imagei(I, sampleri, (float2)(x-1, y)).x + 3 * read_imagei(I, sampleri, (float2)(x-1, y+1)).x); (3 * read_imagei(I, sampleri, (float2)(x-1, y-1)).x + 10 * read_imagei(I, sampleri, (float2)(x-1, y)).x + 3 * read_imagei(I, sampleri, (float2)(x-1, y+1)).x);
dIdy_patch[i * patchWidth + j] = 3 * read_imagei(I, sampleri, (float2)(x-1, y+1)).x + 10 * read_imagei(I, sampleri, (float2)(x, y+1)).x + 3 * read_imagei(I, sampleri, (float2)(x+1, y+1)).x - dIdy_patch[i * patchWidth + j] = 3 * read_imagei(I, sampleri, (float2)(x-1, y+1)).x + 10 * read_imagei(I, sampleri, (float2)(x, y+1)).x + 3 * read_imagei(I, sampleri, (float2)(x+1, y+1)).x -
(3 * read_imagei(I, sampleri, (float2)(x-1, y-1)).x + 10 * read_imagei(I, sampleri, (float2)(x, y-1)).x + 3 * read_imagei(I, sampleri, (float2)(x+1, y-1)).x); (3 * read_imagei(I, sampleri, (float2)(x-1, y-1)).x + 10 * read_imagei(I, sampleri, (float2)(x, y-1)).x + 3 * read_imagei(I, sampleri, (float2)(x+1, y-1)).x);
} }
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
......
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