Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

Merge pull request #12829 from tomoaki0705:fixTestVideoOpenCLPyrLK

Browse Source
pull/12905/head
Alexander Alekhin 6 years ago
parent 113793fee7 803f86b9ca
commit ab2c16b2e5
2 changed files with 56 additions and 66 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 25
      modules/video/src/lkpyramid.cpp
  2. 97
      modules/video/src/opencl/pyrlk.cl

25
modules/video/src/lkpyramid.cpp
Unescape View File

@ -814,7 +814,7 @@ namespace
double minEigThreshold_ = 1e-4) : double minEigThreshold_ = 1e-4) :
winSize(winSize_), maxLevel(maxLevel_), criteria(criteria_), flags(flags_), minEigThreshold(minEigThreshold_) winSize(winSize_), maxLevel(maxLevel_), criteria(criteria_), flags(flags_), minEigThreshold(minEigThreshold_)
#ifdef HAVE_OPENCL #ifdef HAVE_OPENCL
, iters(criteria_.maxCount), derivLambda(criteria_.epsilon), useInitialFlow(0 != (flags_ & OPTFLOW_LK_GET_MIN_EIGENVALS)), waveSize(0) , iters(criteria_.maxCount), derivLambda(criteria_.epsilon), useInitialFlow(0 != (flags_ & OPTFLOW_LK_GET_MIN_EIGENVALS))
#endif #endif
{ {
} }
@ -856,8 +856,6 @@ namespace
calcPatchSize(); calcPatchSize();
if (patch.x <= 0 || patch.x >= 6 || patch.y <= 0 || patch.y >= 6) if (patch.x <= 0 || patch.x >= 6 || patch.y <= 0 || patch.y >= 6)
return false; return false;
if (!initWaveSize())
return false;
return true; return true;
} }
@ -926,19 +924,6 @@ namespace
int iters; int iters;
double derivLambda; double derivLambda;
bool useInitialFlow; bool useInitialFlow;
int waveSize;
bool initWaveSize()
{
waveSize = 1;
if (isDeviceCPU())
return true;
ocl::Kernel kernel;
if (!kernel.create("lkSparse", cv::ocl::video::pyrlk_oclsrc, ""))
return false;
waveSize = (int)kernel.preferedWorkGroupSizeMultiple();
return true;
}
dim3 patch; dim3 patch;
void calcPatchSize() void calcPatchSize()
{ {
@ -977,8 +962,8 @@ namespace
if (isDeviceCPU()) if (isDeviceCPU())
build_options = " -D CPU"; build_options = " -D CPU";
else else
build_options = cv::format("-D WAVE_SIZE=%d -D WSX=%d -D WSY=%d", build_options = cv::format("-D WSX=%d -D WSY=%d",
waveSize, wsx, wsy); wsx, wsy);
ocl::Kernel kernel; ocl::Kernel kernel;
if (!kernel.create("lkSparse", cv::ocl::video::pyrlk_oclsrc, build_options)) if (!kernel.create("lkSparse", cv::ocl::video::pyrlk_oclsrc, build_options))
@ -1064,7 +1049,9 @@ namespace
_status.create((int)npoints, 1, CV_8UC1); _status.create((int)npoints, 1, CV_8UC1);
UMat umatNextPts = _nextPts.getUMat(); UMat umatNextPts = _nextPts.getUMat();
UMat umatStatus = _status.getUMat(); UMat umatStatus = _status.getUMat();
return sparse(_prevImg.getUMat(), _nextImg.getUMat(), _prevPts.getUMat(), umatNextPts, umatStatus, umatErr); UMat umatPrevPts;
_prevPts.getMat().copyTo(umatPrevPts);
return sparse(_prevImg.getUMat(), _nextImg.getUMat(), umatPrevPts, umatNextPts, umatStatus, umatErr);
} }
#endif #endif

97
modules/video/src/opencl/pyrlk.cl
Unescape View File

@ -53,9 +53,6 @@
#define LM_H (LSy*GRIDSIZE+2) #define LM_H (LSy*GRIDSIZE+2)
#define BUFFER (LSx*LSy) #define BUFFER (LSx*LSy)
#define BUFFER2 BUFFER>>1 #define BUFFER2 BUFFER>>1
#ifndef WAVE_SIZE
#define WAVE_SIZE 1
#endif
#ifdef CPU #ifdef CPU
@ -78,7 +75,7 @@ inline void reduce3(float val1, float val2, float val3, __local float* smem1,
} }
} }
inline void reduce2(float val1, float val2, volatile __local float* smem1, volatile __local float* smem2, int tid) inline void reduce2(float val1, float val2, __local float* smem1, __local float* smem2, int tid)
{ {
smem1[tid] = val1; smem1[tid] = val1;
smem2[tid] = val2; smem2[tid] = val2;
@ -95,7 +92,7 @@ inline void reduce2(float val1, float val2, volatile __local float* smem1, volat
} }
} }
inline void reduce1(float val1, volatile __local float* smem1, int tid) inline void reduce1(float val1, __local float* smem1, int tid)
{ {
smem1[tid] = val1; smem1[tid] = val1;
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
@ -111,7 +108,7 @@ inline void reduce1(float val1, volatile __local float* smem1, int tid)
} }
#else #else
inline void reduce3(float val1, float val2, float val3, inline void reduce3(float val1, float val2, float val3,
__local volatile float* smem1, __local volatile float* smem2, __local volatile float* smem3, int tid) __local float* smem1, __local float* smem2, __local float* smem3, int tid)
{ {
smem1[tid] = val1; smem1[tid] = val1;
smem2[tid] = val2; smem2[tid] = val2;
@ -123,38 +120,39 @@ inline void reduce3(float val1, float val2, float val3,
smem1[tid] += smem1[tid + 32]; smem1[tid] += smem1[tid + 32];
smem2[tid] += smem2[tid + 32]; smem2[tid] += smem2[tid + 32];
smem3[tid] += smem3[tid + 32]; smem3[tid] += smem3[tid + 32];
#if WAVE_SIZE < 32
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16) if (tid < 16)
{ {
#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
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid<1) if (tid < 8)
{ {
#endif smem1[tid] += smem1[tid + 8];
local float8* m1 = (local float8*)smem1; smem2[tid] += smem2[tid + 8];
local float8* m2 = (local float8*)smem2; smem3[tid] += smem3[tid + 8];
local float8* m3 = (local float8*)smem3; }
float8 t1 = m1[0]+m1[1]; barrier(CLK_LOCAL_MEM_FENCE);
float8 t2 = m2[0]+m2[1]; if (tid < 4)
float8 t3 = m3[0]+m3[1]; {
float4 t14 = t1.lo + t1.hi; smem1[tid] += smem1[tid + 4];
float4 t24 = t2.lo + t2.hi; smem2[tid] += smem2[tid + 4];
float4 t34 = t3.lo + t3.hi; smem3[tid] += smem3[tid + 4];
smem1[0] = t14.x+t14.y+t14.z+t14.w; }
smem2[0] = t24.x+t24.y+t24.z+t24.w; barrier(CLK_LOCAL_MEM_FENCE);
smem3[0] = t34.x+t34.y+t34.z+t34.w; if (tid == 0)
{
smem1[0] = (smem1[0] + smem1[1]) + (smem1[2] + smem1[3]);
smem2[0] = (smem2[0] + smem2[1]) + (smem2[2] + smem2[3]);
smem3[0] = (smem3[0] + smem3[1]) + (smem3[2] + smem3[3]);
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
} }
inline void reduce2(float val1, float val2, __local volatile float* smem1, __local volatile float* smem2, int tid) inline void reduce2(float val1, float val2, __local float* smem1, __local float* smem2, int tid)
{ {
smem1[tid] = val1; smem1[tid] = val1;
smem2[tid] = val2; smem2[tid] = val2;
@ -164,33 +162,35 @@ inline void reduce2(float val1, float val2, __local volatile float* smem1, __loc
{ {
smem1[tid] += smem1[tid + 32]; smem1[tid] += smem1[tid + 32];
smem2[tid] += smem2[tid + 32]; smem2[tid] += smem2[tid + 32];
#if WAVE_SIZE < 32
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16) if (tid < 16)
{ {
#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
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid<1) if (tid < 8)
{ {
#endif smem1[tid] += smem1[tid + 8];
local float8* m1 = (local float8*)smem1; smem2[tid] += smem2[tid + 8];
local float8* m2 = (local float8*)smem2; }
float8 t1 = m1[0]+m1[1]; barrier(CLK_LOCAL_MEM_FENCE);
float8 t2 = m2[0]+m2[1]; if (tid < 4)
float4 t14 = t1.lo + t1.hi; {
float4 t24 = t2.lo + t2.hi; smem1[tid] += smem1[tid + 4];
smem1[0] = t14.x+t14.y+t14.z+t14.w; smem2[tid] += smem2[tid + 4];
smem2[0] = t24.x+t24.y+t24.z+t24.w; }
barrier(CLK_LOCAL_MEM_FENCE);
if (tid == 0)
{
smem1[0] = (smem1[0] + smem1[1]) + (smem1[2] + smem1[3]);
smem2[0] = (smem2[0] + smem2[1]) + (smem2[2] + smem2[3]);
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
} }
inline void reduce1(float val1, __local volatile float* smem1, int tid) inline void reduce1(float val1, __local float* smem1, int tid)
{ {
smem1[tid] = val1; smem1[tid] = val1;
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
@ -198,23 +198,26 @@ inline void reduce1(float val1, __local volatile float* smem1, int tid)
if (tid < 32) if (tid < 32)
{ {
smem1[tid] += smem1[tid + 32]; smem1[tid] += smem1[tid + 32];
#if WAVE_SIZE < 32
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16) if (tid < 16)
{ {
#endif
smem1[tid] += smem1[tid + 16]; smem1[tid] += smem1[tid + 16];
#if WAVE_SIZE <16
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (tid<1) if (tid < 8)
{ {
#endif smem1[tid] += smem1[tid + 8];
local float8* m1 = (local float8*)smem1; }
float8 t1 = m1[0]+m1[1]; barrier(CLK_LOCAL_MEM_FENCE);
float4 t14 = t1.lo + t1.hi; if (tid < 4)
smem1[0] = t14.x+t14.y+t14.z+t14.w; {
smem1[tid] += smem1[tid + 4];
}
barrier(CLK_LOCAL_MEM_FENCE);
if (tid == 0)
{
smem1[0] = (smem1[0] + smem1[1]) + (smem1[2] + smem1[3]);
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
} }

Write Preview
Loading…
Cancel
Save