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

ocl: workaround for getUMat()

Browse Source
This commit is contained in:
Alexander Alekhin 2015-09-08 04:06:04 +03:00
parent cea2dafa0f
commit ad70ab404c
6 changed files with 159 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
modules/core/include/opencv2/core/mat.hpp
View File

@ -497,6 +497,7 @@ struct CV_EXPORTS UMatData
void* userdata;
int allocatorFlags_;
int mapcount;
UMatData* originalUMatData;
};

15
modules/core/src/matrix.cpp
View File

@ -208,17 +208,14 @@ public:
if(!u)
return;
CV_Assert(u->urefcount >= 0);
CV_Assert(u->refcount >= 0);
if(u->refcount == 0)
CV_Assert(u->urefcount == 0);
CV_Assert(u->refcount == 0);
if( !(u->flags & UMatData::USER_ALLOCATED) )
{
if( !(u->flags & UMatData::USER_ALLOCATED) )
{
fastFree(u->origdata);
u->origdata = 0;
}
delete u;
fastFree(u->origdata);
u->origdata = 0;
}
delete u;
}
};

42
modules/core/src/ocl.cpp
View File

@ -4453,8 +4453,11 @@ public:
#endif
{
tempUMatFlags = UMatData::TEMP_UMAT;
handle = clCreateBuffer(ctx_handle, CL_MEM_USE_HOST_PTR|createFlags,
u->size, u->origdata, &retval);
if (u->origdata == cv::alignPtr(u->origdata, 4)) // There are OpenCL runtime issues for less aligned data
{
handle = clCreateBuffer(ctx_handle, CL_MEM_USE_HOST_PTR|createFlags,
u->size, u->origdata, &retval);
}
if((!handle || retval < 0) && !(accessFlags & ACCESS_FAST))
{
handle = clCreateBuffer(ctx_handle, CL_MEM_COPY_HOST_PTR|CL_MEM_READ_WRITE|createFlags,
@ -4510,17 +4513,17 @@ public:
if(!u)
return;
CV_Assert(u->urefcount >= 0);
CV_Assert(u->refcount >= 0);
CV_Assert(u->urefcount == 0);
CV_Assert(u->refcount == 0 && "UMat deallocation error: some derived Mat is still alive");
CV_Assert(u->handle != 0 && u->urefcount == 0);
CV_Assert(u->handle != 0);
CV_Assert(u->mapcount == 0);
if(u->tempUMat())
{
CV_Assert(u->origdata);
// UMatDataAutoLock lock(u);
if( u->hostCopyObsolete() && u->refcount > 0 )
if (u->hostCopyObsolete())
{
#ifdef HAVE_OPENCL_SVM
if ((u->allocatorFlags_ & svm::OPENCL_SVM_BUFFER_MASK) != 0)
@ -4579,7 +4582,12 @@ public:
void* data = clEnqueueMapBuffer(q, (cl_mem)u->handle, CL_TRUE,
(CL_MAP_READ | CL_MAP_WRITE),
0, u->size, 0, 0, 0, &retval);
CV_Assert(u->origdata == data);
CV_OclDbgAssert(retval == CL_SUCCESS);
if (u->originalUMatData)
{
CV_Assert(u->originalUMatData->data == data);
}
CV_OclDbgAssert(clEnqueueUnmapMemObject(q, (cl_mem)u->handle, data, 0, 0, 0) == CL_SUCCESS);
CV_OclDbgAssert(clFinish(q) == CL_SUCCESS);
}
@ -4587,6 +4595,10 @@ public:
}
u->markHostCopyObsolete(false);
}
else
{
// nothing
}
#ifdef HAVE_OPENCL_SVM
if ((u->allocatorFlags_ & svm::OPENCL_SVM_BUFFER_MASK) != 0)
{
@ -4612,16 +4624,12 @@ public:
if(u->data && u->copyOnMap() && u->data != u->origdata)
fastFree(u->data);
u->data = u->origdata;
if(u->refcount == 0)
{
u->currAllocator->deallocate(u);
u = NULL;
}
u->currAllocator->deallocate(u);
u = NULL;
}
else
{
CV_Assert(u->origdata == NULL);
CV_Assert(u->refcount == 0);
if(u->data && u->copyOnMap() && u->data != u->origdata)
{
fastFree(u->data);
@ -4670,17 +4678,13 @@ public:
delete u;
u = NULL;
}
CV_Assert(u == NULL || u->refcount);
CV_Assert(u == NULL);
}
// synchronized call (external UMatDataAutoLock, see UMat::getMat)
void map(UMatData* u, int accessFlags) const
{
if(!u)
return;
CV_Assert( u->handle != 0 );
UMatDataAutoLock autolock(u);
CV_Assert(u && u->handle);
if(accessFlags & ACCESS_WRITE)
u->markDeviceCopyObsolete(true);

94
modules/core/src/umatrix.cpp
View File

@ -67,6 +67,7 @@ UMatData::UMatData(const MatAllocator* allocator)
handle = 0;
userdata = 0;
allocatorFlags_ = 0;
originalUMatData = NULL;
}
UMatData::~UMatData()
@ -80,6 +81,50 @@ UMatData::~UMatData()
handle = 0;
userdata = 0;
allocatorFlags_ = 0;
if (originalUMatData)
{
UMatData* u = originalUMatData;
CV_XADD(&(u->urefcount), -1);
CV_XADD(&(u->refcount), -1);
bool showWarn = false;
if (u->refcount == 0)
{
if (u->urefcount > 0)
showWarn = true;
// simulate Mat::deallocate
if (u->mapcount != 0)
{
(u->currAllocator ? u->currAllocator : /* TODO allocator ? allocator :*/ Mat::getStdAllocator())->unmap(u);
}
else
{
// we don't do "map", so we can't do "unmap"
}
}
if (u->refcount == 0 && u->urefcount == 0) // oops, we need to free resources
{
showWarn = true;
// simulate UMat::deallocate
u->currAllocator->deallocate(u);
}
#ifndef NDEBUG
if (showWarn)
{
static int warn_message_showed = 0;
if (warn_message_showed++ < 100)
{
fflush(stdout);
fprintf(stderr, "\n! OPENCV warning: getUMat()/getMat() call chain possible problem."
"\n! Base object is dead, while nested/derived object is still alive or processed."
"\n! Please check lifetime of UMat/Mat objects!\n");
fflush(stderr);
}
}
#else
(void)showWarn;
#endif
originalUMatData = NULL;
}
}
void UMatData::lock()
@ -222,20 +267,34 @@ UMat Mat::getUMat(int accessFlags, UMatUsageFlags usageFlags) const
UMat hdr;
if(!data)
return hdr;
CV_Assert((!u || u->mapcount==0) && "Don't get UMat from temp-Mat!");
Size wholeSize;
Point ofs;
locateROI(wholeSize, ofs);
Size sz(cols, rows);
if (ofs.x != 0 || ofs.y != 0)
{
Mat src = *this;
int dtop = ofs.y;
int dbottom = wholeSize.height - src.rows - ofs.y;
int dleft = ofs.x;
int dright = wholeSize.width - src.cols - ofs.x;
src.adjustROI(dtop, dbottom, dleft, dright);
return src.getUMat(accessFlags, usageFlags)(cv::Rect(ofs.x, ofs.y, sz.width, sz.height));
}
CV_Assert(data == datastart);
accessFlags |= ACCESS_RW;
UMatData* temp_u = u;
if(!temp_u)
UMatData* new_u = NULL;
{
MatAllocator *a = allocator, *a0 = getStdAllocator();
if(!a)
a = a0;
temp_u = a->allocate(dims, size.p, type(), data, step.p, accessFlags, usageFlags);
new_u = a->allocate(dims, size.p, type(), data, step.p, accessFlags, usageFlags);
}
bool allocated = false;
try
{
allocated = UMat::getStdAllocator()->allocate(temp_u, accessFlags, usageFlags);
allocated = UMat::getStdAllocator()->allocate(new_u, accessFlags, usageFlags);
}
catch (const cv::Exception& e)
{
@ -243,14 +302,26 @@ UMat Mat::getUMat(int accessFlags, UMatUsageFlags usageFlags) const
}
if (!allocated)
{
allocated = getStdAllocator()->allocate(temp_u, accessFlags, usageFlags);
allocated = getStdAllocator()->allocate(new_u, accessFlags, usageFlags);
CV_Assert(allocated);
}
if (u != NULL)
{
#ifdef HAVE_OPENCL
if (ocl::useOpenCL() && new_u->currAllocator == ocl::getOpenCLAllocator())
{
CV_Assert(new_u->tempUMat());
}
#endif
new_u->originalUMatData = u;
CV_XADD(&(u->refcount), 1);
CV_XADD(&(u->urefcount), 1);
}
hdr.flags = flags;
setSize(hdr, dims, size.p, step.p);
finalizeHdr(hdr);
hdr.u = temp_u;
hdr.offset = data - datastart;
hdr.u = new_u;
hdr.offset = 0; //data - datastart;
hdr.addref();
return hdr;
}
@ -639,7 +710,6 @@ Mat UMat::getMat(int accessFlags) const
{
if(!u)
return Mat();
CV_Assert(!u->tempUMat() && "Don't get Mat from temp UMat! Use copyTo().");
// TODO Support ACCESS_READ (ACCESS_WRITE) without unnecessary data transfers
accessFlags |= ACCESS_RW;
UMatDataAutoLock autolock(u);
@ -668,10 +738,10 @@ void* UMat::handle(int accessFlags) const
if( !u )
return 0;
// check flags: if CPU copy is newer, copy it back to GPU.
if( u->deviceCopyObsolete() )
CV_Assert(u->refcount == 0);
CV_Assert(!u->deviceCopyObsolete() || u->copyOnMap());
if (u->deviceCopyObsolete())
{
CV_Assert(u->refcount == 0 || u->origdata);
u->currAllocator->unmap(u);
}

48
modules/core/test/test_umat.cpp
View File

@ -263,7 +263,7 @@ TEST_P(UMatBasicTests, GetUMat)
}
}
INSTANTIATE_TEST_CASE_P(UMat, UMatBasicTests, Combine(testing::Values(CV_8U), testing::Values(1, 2),
INSTANTIATE_TEST_CASE_P(UMat, UMatBasicTests, Combine(testing::Values(CV_8U, CV_64F), testing::Values(1, 2),
testing::Values(cv::Size(1, 1), cv::Size(1, 128), cv::Size(128, 1), cv::Size(128, 128), cv::Size(640, 480)), Bool()));
//////////////////////////////////////////////////////////////// Reshape ////////////////////////////////////////////////////////////////////////
@ -1084,7 +1084,7 @@ TEST(UMat, unmap_in_class)
Mat dst;
m.convertTo(dst, CV_32FC1);
// some additional CPU-based per-pixel processing into dst
intermediateResult = dst.getUMat(ACCESS_READ);
intermediateResult = dst.getUMat(ACCESS_READ); // this violates lifetime of base(dst) / derived (intermediateResult) objects. Use copyTo?
std::cout << "data processed..." << std::endl;
} // problem is here: dst::~Mat()
std::cout << "leave ProcessData()" << std::endl;
@ -1285,4 +1285,48 @@ TEST(UMat, mat_umat_sync)
ASSERT_EQ(0, countNonZero(uDiff));
}
TEST(UMat, testTempObjects_UMat)
{
UMat u(10, 10, CV_8UC1, Scalar(1));
{
UMat u2 = u.getMat(ACCESS_RW).getUMat(ACCESS_RW);
u2.setTo(Scalar(255));
}
UMat uDiff;
compare(u, 255, uDiff, CMP_NE);
ASSERT_EQ(0, countNonZero(uDiff));
}
TEST(UMat, testTempObjects_Mat)
{
Mat m(10, 10, CV_8UC1, Scalar(1));
{
Mat m2;
ASSERT_ANY_THROW(m2 = m.getUMat(ACCESS_RW).getMat(ACCESS_RW));
}
}
TEST(UMat, testWrongLifetime_UMat)
{
UMat u(10, 10, CV_8UC1, Scalar(1));
{
UMat u2 = u.getMat(ACCESS_RW).getUMat(ACCESS_RW);
u.release(); // base object
u2.release(); // derived object, should show warning message
}
}
TEST(UMat, testWrongLifetime_Mat)
{
Mat m(10, 10, CV_8UC1, Scalar(1));
{
UMat u = m.getUMat(ACCESS_RW);
Mat m2 = u.getMat(ACCESS_RW);
m.release(); // base object
m2.release(); // map of derived object
u.release(); // derived object, should show warning message
}
}
} } // namespace cvtest::ocl

2
modules/video/src/lkpyramid.cpp
View File

@ -1009,7 +1009,7 @@ namespace cv
idxArg = kernel.set(idxArg, (int)winSize.height); // int c_winSize_y
idxArg = kernel.set(idxArg, (int)iters); // int c_iters
idxArg = kernel.set(idxArg, (char)calcErr); //char calcErr
return kernel.run(2, globalThreads, localThreads, false);
return kernel.run(2, globalThreads, localThreads, true); // sync=true because ocl::Image2D lifetime is not handled well for temp UMat
}
private:
inline static bool isDeviceCPU()