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

ocl: fix compilation warnings, update openCLExecuteKernelInterop

Browse Source
This commit is contained in:
Alexander Alekhin
2013-10-09 16:57:52 +04:00
parent e3b42ed137
commit d26f62569d
6 changed files with 133 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
modules/ocl/src/cl_operations.cpp
View File

@@ -302,28 +302,27 @@ void openCLExecuteKernel(Context *ctx, const cv::ocl::ProgramEntry* source, stri
total_kernel_time = 0;
cout << "-------------------------------------" << endl;
cout << setiosflags(ios::left) << setw(15) << "excute time";
cout << setiosflags(ios::left) << setw(15) << "lauch time";
cout << setiosflags(ios::left) << setw(15) << "execute time";
cout << setiosflags(ios::left) << setw(15) << "launch time";
cout << setiosflags(ios::left) << setw(15) << "kernel time" << endl;
int i = 0;
for(i = 0; i < RUN_TIMES; i++)
openCLExecuteKernel_(ctx, source, kernelName, globalThreads, localThreads, args, channels, depth,
build_options);
cout << "average kernel excute time: " << total_execute_time / RUN_TIMES << endl; // "ms" << endl;
cout << "average kernel execute time: " << total_execute_time / RUN_TIMES << endl; // "ms" << endl;
cout << "average kernel total time: " << total_kernel_time / RUN_TIMES << endl; // "ms" << endl;
#endif
}
double openCLExecuteKernelInterop(Context *ctx, const cv::ocl::ProgramEntry* source, string kernelName,
void openCLExecuteKernelInterop(Context *ctx, const cv::ocl::ProgramSource& source, string kernelName,
size_t globalThreads[3], size_t localThreads[3],
vector< pair<size_t, const void *> > &args, int channels, int depth, const char *build_options,
bool finish, bool measureKernelTime, bool cleanUp)
vector< pair<size_t, const void *> > &args, int channels, int depth, const char *build_options)
{
//construct kernel name
//The rule is functionName_Cn_Dn, C represent Channels, D Represent DataType Depth, n represent an integer number
//for exmaple split_C2_D2, represent the split kernel with channels =2 and dataType Depth = 2(Data type is char)
//for example split_C2_D2, represent the split kernel with channels = 2 and dataType Depth = 2 (Data type is char)
stringstream idxStr;
if(channels != -1)
idxStr << "_C" << channels;
@@ -331,63 +330,27 @@ double openCLExecuteKernelInterop(Context *ctx, const cv::ocl::ProgramEntry* sou
idxStr << "_D" << depth;
kernelName += idxStr.str();
cl_kernel kernel;
kernel = openCLGetKernelFromSource(ctx, source, kernelName, build_options);
std::string name = std::string("custom_") + source.name;
ProgramEntry program = { name.c_str(), source.programStr, source.programHash };
cl_kernel kernel = openCLGetKernelFromSource(ctx, &program, kernelName, build_options);
double kernelTime = 0.0;
if( globalThreads != NULL)
CV_Assert(globalThreads != NULL);
if ( localThreads != NULL)
{
if ( localThreads != NULL)
{
globalThreads[0] = divUp(globalThreads[0], localThreads[0]) * localThreads[0];
globalThreads[1] = divUp(globalThreads[1], localThreads[1]) * localThreads[1];
globalThreads[2] = divUp(globalThreads[2], localThreads[2]) * localThreads[2];
globalThreads[0] = roundUp(globalThreads[0], localThreads[0]);
globalThreads[1] = roundUp(globalThreads[1], localThreads[1]);
globalThreads[2] = roundUp(globalThreads[2], localThreads[2]);
//size_t blockSize = localThreads[0] * localThreads[1] * localThreads[2];
cv::ocl::openCLVerifyKernel(ctx, kernel, localThreads);
}
for(size_t i = 0; i < args.size(); i ++)
openCLSafeCall(clSetKernelArg(kernel, i, args[i].first, args[i].second));
if(measureKernelTime == false)
{
openCLSafeCall(clEnqueueNDRangeKernel(getClCommandQueue(ctx), kernel, 3, NULL, globalThreads,
localThreads, 0, NULL, NULL));
}
else
{
cl_event event = NULL;
openCLSafeCall(clEnqueueNDRangeKernel(getClCommandQueue(ctx), kernel, 3, NULL, globalThreads,
localThreads, 0, NULL, &event));
cl_ulong end_time, queue_time;
openCLSafeCall(clWaitForEvents(1, &event));
openCLSafeCall(clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_END,
sizeof(cl_ulong), &end_time, 0));
openCLSafeCall(clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_QUEUED,
sizeof(cl_ulong), &queue_time, 0));
kernelTime = (double)(end_time - queue_time) / (1000 * 1000);
clReleaseEvent(event);
}
cv::ocl::openCLVerifyKernel(ctx, kernel, localThreads);
}
for(size_t i = 0; i < args.size(); i ++)
openCLSafeCall(clSetKernelArg(kernel, i, args[i].first, args[i].second));
if(finish)
{
clFinish(getClCommandQueue(ctx));
}
openCLSafeCall(clEnqueueNDRangeKernel(getClCommandQueue(ctx), kernel, 3, NULL, globalThreads,
localThreads, 0, NULL, NULL));
if(cleanUp)
{
openCLSafeCall(clReleaseKernel(kernel));
}
return kernelTime;
clFinish(getClCommandQueue(ctx));
openCLSafeCall(clReleaseKernel(kernel));
}
cl_mem load_constant(cl_context context, cl_command_queue command_queue, const void *value,

10
modules/ocl/src/cl_programcache.cpp
View File

@@ -67,7 +67,6 @@
namespace cv { namespace ocl {
#define MAX_PROG_CACHE_SIZE 1024
/*
* The binary caching system to eliminate redundant program source compilation.
* Strictly, this is not a cache because we do not implement evictions right now.
@@ -291,7 +290,7 @@ struct ProgramFileCache
bool writeConfigurationToFile(const string& options, std::vector<char>& buf)
{
if (hash_ == NULL)
return true; // don't save dynamic kernels
return true; // don't save programs without hash
if (!f.is_open())
{
@@ -469,7 +468,7 @@ cl_program ProgramCache::getProgram(const Context *ctx, const cv::ocl::ProgramEn
{
stringstream src_sign;
src_sign << (int64)(source->programStr);
src_sign << source->name;
src_sign << getClContext(ctx);
if (NULL != build_options)
{
@@ -514,15 +513,10 @@ cl_program ProgramCache::getProgram(const Context *ctx, const cv::ocl::ProgramEn
cl_program program = programFileCache.getOrBuildProgram(ctx, source, all_build_options);
//Cache the binary for future use if build_options is null
if( (this->cacheSize += 1) < MAX_PROG_CACHE_SIZE)
{
cv::AutoLock lockCache(mutexCache);
this->addProgram(src_sign.str(), program);
}
else
{
cout << "Warning: code cache has been full.\n";
}
return program;
}