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

Merge pull request #3603 from alalek:ocl_svm

Browse Source
This commit is contained in:
Vadim Pisarevsky
2015-01-29 05:35:46 +00:00
parent bdebf4cea7 0a07d780e0
commit 97aa8d3341
15 changed files with 1544 additions and 162 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
modules/core/src/matmul.cpp
View File

@@ -721,6 +721,16 @@ static bool ocl_gemm_amdblas( InputArray matA, InputArray matB, double alpha,
return false;
UMat A = matA.getUMat(), B = matB.getUMat(), D = matD.getUMat();
if (!ocl::internal::isCLBuffer(A) || !ocl::internal::isCLBuffer(B) || !ocl::internal::isCLBuffer(D))
{
return false;
}
if (haveC)
{
UMat C = matC.getUMat();
if (!ocl::internal::isCLBuffer(C))
return false;
}
if (haveC)
ctrans ? transpose(matC, D) : matC.copyTo(D);
else

3
modules/core/src/matrix.cpp
View File

@@ -159,8 +159,9 @@ void MatAllocator::copy(UMatData* usrc, UMatData* udst, int dims, const size_t s
memcpy(ptrs[1], ptrs[0], planesz);
}
BufferPoolController* MatAllocator::getBufferPoolController() const
BufferPoolController* MatAllocator::getBufferPoolController(const char* id) const
{
(void)id;
static DummyBufferPoolController dummy;
return &dummy;
}

1232
modules/core/src/ocl.cpp
View File

File diff suppressed because it is too large Load Diff

67
modules/core/src/opencl/runtime/opencl_core.cpp
View File

@@ -182,6 +182,65 @@ static void* opencl_check_fn(int ID);
#define CUSTOM_FUNCTION_ID 1000
#ifdef HAVE_OPENCL_SVM
#include "opencv2/core/opencl/runtime/opencl_svm_20.hpp"
#define SVM_FUNCTION_ID_START CUSTOM_FUNCTION_ID
#define SVM_FUNCTION_ID_END CUSTOM_FUNCTION_ID + 100
enum OPENCL_FN_SVM_ID
{
OPENCL_FN_clSVMAlloc = SVM_FUNCTION_ID_START,
OPENCL_FN_clSVMFree,
OPENCL_FN_clSetKernelArgSVMPointer,
OPENCL_FN_clSetKernelExecInfo,
OPENCL_FN_clEnqueueSVMFree,
OPENCL_FN_clEnqueueSVMMemcpy,
OPENCL_FN_clEnqueueSVMMemFill,
OPENCL_FN_clEnqueueSVMMap,
OPENCL_FN_clEnqueueSVMUnmap,
};
void* (CL_API_CALL *clSVMAlloc)(cl_context context, cl_svm_mem_flags flags, size_t size, unsigned int alignment) =
opencl_fn4<OPENCL_FN_clSVMAlloc, void*, cl_context, cl_svm_mem_flags, size_t, unsigned int>::switch_fn;
static const struct DynamicFnEntry _clSVMAlloc_definition = { "clSVMAlloc", (void**)&clSVMAlloc};
void (CL_API_CALL *clSVMFree)(cl_context context, void* svm_pointer) =
opencl_fn2<OPENCL_FN_clSVMFree, void, cl_context, void*>::switch_fn;
static const struct DynamicFnEntry _clSVMFree_definition = { "clSVMFree", (void**)&clSVMFree};
cl_int (CL_API_CALL *clSetKernelArgSVMPointer)(cl_kernel kernel, cl_uint arg_index, const void* arg_value) =
opencl_fn3<OPENCL_FN_clSetKernelArgSVMPointer, cl_int, cl_kernel, cl_uint, const void*>::switch_fn;
static const struct DynamicFnEntry _clSetKernelArgSVMPointer_definition = { "clSetKernelArgSVMPointer", (void**)&clSetKernelArgSVMPointer};
//void* (CL_API_CALL *clSetKernelExecInfo)(cl_kernel kernel, cl_kernel_exec_info param_name, size_t param_value_size, const void* param_value) =
// opencl_fn4<OPENCL_FN_clSetKernelExecInfo, void*, cl_kernel, cl_kernel_exec_info, size_t, const void*>::switch_fn;
//static const struct DynamicFnEntry _clSetKernelExecInfo_definition = { "clSetKernelExecInfo", (void**)&clSetKernelExecInfo};
//cl_int (CL_API_CALL *clEnqueueSVMFree)(...) =
// opencl_fn8<OPENCL_FN_clEnqueueSVMFree, cl_int, ...>::switch_fn;
//static const struct DynamicFnEntry _clEnqueueSVMFree_definition = { "clEnqueueSVMFree", (void**)&clEnqueueSVMFree};
cl_int (CL_API_CALL *clEnqueueSVMMemcpy)(cl_command_queue command_queue, cl_bool blocking_copy, void* dst_ptr, const void* src_ptr, size_t size, cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event) =
opencl_fn8<OPENCL_FN_clEnqueueSVMMemcpy, cl_int, cl_command_queue, cl_bool, void*, const void*, size_t, cl_uint, const cl_event*, cl_event*>::switch_fn;
static const struct DynamicFnEntry _clEnqueueSVMMemcpy_definition = { "clEnqueueSVMMemcpy", (void**)&clEnqueueSVMMemcpy};
cl_int (CL_API_CALL *clEnqueueSVMMemFill)(cl_command_queue command_queue, void* svm_ptr, const void* pattern, size_t pattern_size, size_t size, cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event) =
opencl_fn8<OPENCL_FN_clEnqueueSVMMemFill, cl_int, cl_command_queue, void*, const void*, size_t, size_t, cl_uint, const cl_event*, cl_event*>::switch_fn;
static const struct DynamicFnEntry _clEnqueueSVMMemFill_definition = { "clEnqueueSVMMemFill", (void**)&clEnqueueSVMMemFill};
cl_int (CL_API_CALL *clEnqueueSVMMap)(cl_command_queue command_queue, cl_bool blocking_map, cl_map_flags map_flags, void* svm_ptr, size_t size, cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event) =
opencl_fn8<OPENCL_FN_clEnqueueSVMMap, cl_int, cl_command_queue, cl_bool, cl_map_flags, void*, size_t, cl_uint, const cl_event*, cl_event*>::switch_fn;
static const struct DynamicFnEntry _clEnqueueSVMMap_definition = { "clEnqueueSVMMap", (void**)&clEnqueueSVMMap};
cl_int (CL_API_CALL *clEnqueueSVMUnmap)(cl_command_queue command_queue, void* svm_ptr, cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event) =
opencl_fn5<OPENCL_FN_clEnqueueSVMUnmap, cl_int, cl_command_queue, void*, cl_uint, const cl_event*, cl_event*>::switch_fn;
static const struct DynamicFnEntry _clEnqueueSVMUnmap_definition = { "clEnqueueSVMUnmap", (void**)&clEnqueueSVMUnmap};
static const struct DynamicFnEntry* opencl_svm_fn_list[] = {
&_clSVMAlloc_definition,
&_clSVMFree_definition,
&_clSetKernelArgSVMPointer_definition,
NULL/*&_clSetKernelExecInfo_definition*/,
NULL/*&_clEnqueueSVMFree_definition*/,
&_clEnqueueSVMMemcpy_definition,
&_clEnqueueSVMMemFill_definition,
&_clEnqueueSVMMap_definition,
&_clEnqueueSVMUnmap_definition,
};
#endif // HAVE_OPENCL_SVM
//
// END OF CUSTOM FUNCTIONS HERE
//
@@ -194,6 +253,14 @@ static void* opencl_check_fn(int ID)
assert(ID >= 0 && ID < (int)(sizeof(opencl_fn_list)/sizeof(opencl_fn_list[0])));
e = opencl_fn_list[ID];
}
#ifdef HAVE_OPENCL_SVM
else if (ID >= SVM_FUNCTION_ID_START && ID < SVM_FUNCTION_ID_END)
{
ID = ID - SVM_FUNCTION_ID_START;
assert(ID >= 0 && ID < (int)(sizeof(opencl_svm_fn_list)/sizeof(opencl_svm_fn_list[0])));
e = opencl_svm_fn_list[ID];
}
#endif
else
{
CV_ErrorNoReturn(cv::Error::StsBadArg, "Invalid function ID");

8
modules/core/src/umatrix.cpp
View File

@@ -55,7 +55,7 @@ UMatData::UMatData(const MatAllocator* allocator)
prevAllocator = currAllocator = allocator;
urefcount = refcount = 0;
data = origdata = 0;
size = 0; capacity = 0;
size = 0;
flags = 0;
handle = 0;
userdata = 0;
@@ -67,7 +67,7 @@ UMatData::~UMatData()
prevAllocator = currAllocator = 0;
urefcount = refcount = 0;
data = origdata = 0;
size = 0; capacity = 0;
size = 0;
flags = 0;
handle = 0;
userdata = 0;
@@ -221,7 +221,7 @@ UMat Mat::getUMat(int accessFlags, UMatUsageFlags usageFlags) const
temp_u = a->allocate(dims, size.p, type(), data, step.p, accessFlags, usageFlags);
temp_u->refcount = 1;
}
UMat::getStdAllocator()->allocate(temp_u, accessFlags, usageFlags);
UMat::getStdAllocator()->allocate(temp_u, accessFlags, usageFlags); // TODO result is not checked
hdr.flags = flags;
setSize(hdr, dims, size.p, step.p);
finalizeHdr(hdr);
@@ -575,7 +575,7 @@ Mat UMat::getMat(int accessFlags) const
{
if(!u)
return Mat();
u->currAllocator->map(u, accessFlags | ACCESS_READ);
u->currAllocator->map(u, accessFlags | ACCESS_READ); // TODO Support ACCESS_WRITE without unnecessary data transfers
CV_Assert(u->data != 0);
Mat hdr(dims, size.p, type(), u->data + offset, step.p);
hdr.flags = flags;