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

fixed many bugs related to Mat::getUMat(), asynchronous kernel execution etc. Also, played a bit with ocl::cvtColor vs cv::cvtColor performance

Browse Source
This commit is contained in:
Vadim Pisarevsky
2013-12-01 03:12:19 +04:00
parent daefe6983f
commit 6da5d21331
10 changed files with 401 additions and 208 deletions
Download Patch File Download Diff File Expand all files Collapse all files

117
modules/core/src/ocl.cpp
View File

@@ -612,7 +612,7 @@ static void* initOpenCLAndLoad(const char* funcname)
return 0;
}
return funcname ? dlsym(handle, funcname) : 0;
return funcname && handle ? dlsym(handle, funcname) : 0;
}
#elif defined WIN32 || defined _WIN32
@@ -2002,7 +2002,7 @@ void* Queue::ptr() const
Queue& Queue::getDefault()
{
Queue& q = TLSData::get()->oclQueue;
if( !q.p )
if( !q.p && haveOpenCL() )
q.create(Context2::getDefault());
return q;
}
@@ -2251,22 +2251,32 @@ int Kernel::set(int i, const KernelArg& arg)
}
bool Kernel::run(int dims, size_t globalsize[], size_t localsize[],
bool Kernel::run(int dims, size_t _globalsize[], size_t _localsize[],
bool sync, const Queue& q)
{
if(!p || !p->handle || p->e != 0)
return false;
AutoBuffer<size_t> _globalSize(dims);
size_t * globalSizePtr = (size_t *)_globalSize;
for (int i = 0; i < dims; ++i)
globalSizePtr[i] = localsize == NULL ? globalsize[i] :
((globalsize[i] + localsize[i] - 1) / localsize[i]) * localsize[i];
cl_command_queue qq = getQueue(q);
size_t offset[CV_MAX_DIM] = {0};
size_t offset[CV_MAX_DIM] = {0}, globalsize[CV_MAX_DIM] = {1,1,1}, localsize[CV_MAX_DIM] = {1,1,1};
size_t total = 1;
for (int i = 0; i < dims; i++)
{
size_t val0 = _localsize ? _localsize[i] :
dims == 1 ? 64 : dims == 2 ? 16>>i : dims == 3 ? 8>>(i>0) : 1;
size_t val = 1;
while( val*2 < val0 )
val *= 2;
if( _localsize )
localsize[i] = val;
CV_Assert(_globalsize && _globalsize[i] >= 0);
total *= _globalsize[i];
globalsize[i] = ((_globalsize[i] + val - 1)/val)*val;
}
if( total == 0 )
return true;
cl_int retval = clEnqueueNDRangeKernel(qq, p->handle, (cl_uint)dims,
offset, globalSizePtr, localsize, 0, 0,
offset, globalsize, _localsize ? localsize : 0, 0, 0,
sync ? 0 : &p->e);
if( sync || retval < 0 )
{
@@ -2361,14 +2371,23 @@ struct Program::Impl
retval = clBuildProgram(handle, n,
(const cl_device_id*)deviceList,
buildflags.c_str(), 0, 0);
if( retval == CL_BUILD_PROGRAM_FAILURE )
if( retval < 0 )
{
char buf[1<<16];
size_t retsz = 0;
clGetProgramBuildInfo(handle, (cl_device_id)deviceList[0], CL_PROGRAM_BUILD_LOG,
sizeof(buf)-16, buf, &retsz);
errmsg = String(buf);
CV_Error_(Error::StsAssert, ("OpenCL program can not be built: %s", errmsg.c_str()));
retval = clGetProgramBuildInfo(handle, (cl_device_id)deviceList[0],
CL_PROGRAM_BUILD_LOG, 0, 0, &retsz);
if( retval >= 0 && retsz > 0 )
{
AutoBuffer<char> bufbuf(retsz + 16);
char* buf = bufbuf;
retval = clGetProgramBuildInfo(handle, (cl_device_id)deviceList[0],
CL_PROGRAM_BUILD_LOG, retsz+1, buf, &retsz);
if( retval >= 0 )
{
errmsg = String(buf);
CV_Error_(Error::StsAssert, ("OpenCL program can not be built: %s", errmsg.c_str()));
}
}
}
CV_Assert(retval >= 0);
}
@@ -2608,17 +2627,17 @@ ProgramSource2::hash_t ProgramSource2::hash() const
class OpenCLAllocator : public MatAllocator
{
public:
OpenCLAllocator() {}
OpenCLAllocator() { matStdAllocator = Mat::getStdAllocator(); }
UMatData* defaultAllocate(int dims, const int* sizes, int type, size_t* step) const
UMatData* defaultAllocate(int dims, const int* sizes, int type, void* data, size_t* step, int flags) const
{
UMatData* u = Mat::getStdAllocator()->allocate(dims, sizes, type, step);
UMatData* u = matStdAllocator->allocate(dims, sizes, type, data, step, flags);
u->urefcount = 1;
u->refcount = 0;
return u;
}
void getBestFlags(const Context2& ctx, int& createFlags, int& flags0) const
void getBestFlags(const Context2& ctx, int /*flags*/, int& createFlags, int& flags0) const
{
const Device& dev = ctx.device(0);
createFlags = CL_MEM_READ_WRITE;
@@ -2629,10 +2648,12 @@ public:
flags0 = UMatData::COPY_ON_MAP;
}
UMatData* allocate(int dims, const int* sizes, int type, size_t* step) const
UMatData* allocate(int dims, const int* sizes, int type,
void* data, size_t* step, int flags) const
{
if(!useOpenCL())
return defaultAllocate(dims, sizes, type, step);
return defaultAllocate(dims, sizes, type, data, step, flags);
CV_Assert(data == 0);
size_t total = CV_ELEM_SIZE(type);
for( int i = dims-1; i >= 0; i-- )
{
@@ -2643,13 +2664,13 @@ public:
Context2& ctx = Context2::getDefault();
int createFlags = 0, flags0 = 0;
getBestFlags(ctx, createFlags, flags0);
getBestFlags(ctx, flags, createFlags, flags0);
cl_int retval = 0;
void* handle = clCreateBuffer((cl_context)ctx.ptr(),
createFlags, total, 0, &retval);
if( !handle || retval < 0 )
return defaultAllocate(dims, sizes, type, step);
return defaultAllocate(dims, sizes, type, data, step, flags);
UMatData* u = new UMatData(this);
u->data = 0;
u->size = total;
@@ -2672,7 +2693,7 @@ public:
CV_Assert(u->origdata != 0);
Context2& ctx = Context2::getDefault();
int createFlags = 0, flags0 = 0;
getBestFlags(ctx, createFlags, flags0);
getBestFlags(ctx, accessFlags, createFlags, flags0);
cl_context ctx_handle = (cl_context)ctx.ptr();
cl_int retval = 0;
@@ -2697,19 +2718,41 @@ public:
return true;
}
void sync(UMatData* u) const
{
cl_command_queue q = (cl_command_queue)Queue::getDefault().ptr();
clFinish(q);
if( u->hostCopyObsolete() && u->handle &&
u->tempCopiedUMat() && u->refcount > 0 && u->origdata)
{
UMatDataAutoLock lock(u);
clEnqueueReadBuffer(q, (cl_mem)u->handle, CL_TRUE, 0,
u->size, u->origdata, 0, 0, 0);
u->markHostCopyObsolete(false);
}
else if( u->copyOnMap() && u->deviceCopyObsolete() && u->data )
{
UMatDataAutoLock lock(u);
clEnqueueWriteBuffer(q, (cl_mem)u->handle, CL_TRUE, 0,
u->size, u->data, 0, 0, 0);
}
}
void deallocate(UMatData* u) const
{
if(!u)
return;
UMatDataAutoLock lock(u);
// TODO: !!! when we add Shared Virtual Memory Support,
// this function (as well as the others should be corrected)
// this function (as well as the others) should be corrected
CV_Assert(u->handle != 0 && u->urefcount == 0);
if(u->tempUMat())
{
if( u->hostCopyObsolete() && u->refcount > 0 && u->tempCopiedUMat() )
{
clEnqueueWriteBuffer((cl_command_queue)Queue::getDefault().ptr(),
clEnqueueReadBuffer((cl_command_queue)Queue::getDefault().ptr(),
(cl_mem)u->handle, CL_TRUE, 0,
u->size, u->origdata, 0, 0, 0);
}
@@ -2717,7 +2760,7 @@ public:
clReleaseMemObject((cl_mem)u->handle);
u->handle = 0;
u->currAllocator = u->prevAllocator;
if(u->data && u->copyOnMap())
if(u->data && u->copyOnMap() && !(u->flags & UMatData::USER_ALLOCATED))
fastFree(u->data);
u->data = u->origdata;
if(u->refcount == 0)
@@ -2725,8 +2768,11 @@ public:
}
else
{
if(u->data && u->copyOnMap())
if(u->data && u->copyOnMap() && !(u->flags & UMatData::USER_ALLOCATED))
{
fastFree(u->data);
u->data = 0;
}
clReleaseMemObject((cl_mem)u->handle);
u->handle = 0;
delete u;
@@ -2793,15 +2839,18 @@ public:
UMatDataAutoLock autolock(u);
cl_command_queue q = (cl_command_queue)Queue::getDefault().ptr();
cl_int retval = 0;
if( !u->copyOnMap() && u->data )
{
CV_Assert( clEnqueueUnmapMemObject(q, (cl_mem)u->handle, u->data, 0, 0, 0) >= 0 );
CV_Assert( (retval = clEnqueueUnmapMemObject(q,
(cl_mem)u->handle, u->data, 0, 0, 0)) >= 0 );
clFinish(q);
u->data = 0;
}
else if( u->copyOnMap() && u->deviceCopyObsolete() )
{
CV_Assert( clEnqueueWriteBuffer(q, (cl_mem)u->handle, CL_TRUE, 0,
u->size, u->data, 0, 0, 0) >= 0 );
CV_Assert( (retval = clEnqueueWriteBuffer(q, (cl_mem)u->handle, CL_TRUE, 0,
u->size, u->data, 0, 0, 0)) >= 0 );
}
u->markDeviceCopyObsolete(false);
u->markHostCopyObsolete(false);
@@ -3033,6 +3082,8 @@ public:
if( sync )
clFinish(q);
}
MatAllocator* matStdAllocator;
};
MatAllocator* getOpenCLAllocator()