int clPeak::runComputeDP(cl::CommandQueue &queue, cl::Program &prog, device_info_t &devInfo)
{
float timed, gflops;
cl_uint workPerWI;
cl::NDRange globalSize, localSize;
cl_double A = 1.3f;
int iters = devInfo.computeIters;
if(!isComputeDP)
return 0;
if(!devInfo.doubleSupported)
{
cout << NEWLINE TAB TAB "No double precision support! Skipped" << endl;
return 0;
}
try
{
cl::Context ctx = queue.getInfo<CL_QUEUE_CONTEXT>();
uint globalWIs = (devInfo.numCUs) * (devInfo.computeWgsPerCU) * (devInfo.maxWGSize);
uint t = MIN((globalWIs * sizeof(cl_double)), devInfo.maxAllocSize);
t = roundToPowOf2(t);
globalWIs = t / sizeof(cl_double);
cl::Buffer outputBuf = cl::Buffer(ctx, CL_MEM_WRITE_ONLY, (globalWIs * sizeof(cl_double)));
globalSize = globalWIs;
localSize = devInfo.maxWGSize;
cl::Kernel kernel_v1(prog, "compute_dp_v1");
kernel_v1.setArg(0, outputBuf), kernel_v1.setArg(1, A);
cl::Kernel kernel_v2(prog, "compute_dp_v2");
kernel_v2.setArg(0, outputBuf), kernel_v2.setArg(1, A);
cl::Kernel kernel_v4(prog, "compute_dp_v4");
kernel_v4.setArg(0, outputBuf), kernel_v4.setArg(1, A);
cl::Kernel kernel_v8(prog, "compute_dp_v8");
kernel_v8.setArg(0, outputBuf), kernel_v8.setArg(1, A);
cl::Kernel kernel_v16(prog, "compute_dp_v16");
kernel_v16.setArg(0, outputBuf), kernel_v16.setArg(1, A);
cout << NEWLINE TAB TAB "Double-precision compute (GFLOPS)" << endl;
cout << setprecision(2) << fixed;
///////////////////////////////////////////////////////////////////////////
// Vector width 1
cout << TAB TAB TAB "double : "; cout.flush();
workPerWI = 4096; // Indicates flops executed per work-item
timed = run_kernel(queue, kernel_v1, globalSize, localSize, iters);
gflops = ((float)globalWIs * workPerWI) / timed / 1e3f;
cout << gflops << endl;
///////////////////////////////////////////////////////////////////////////
// Vector width 2
cout << TAB TAB TAB "double2 : "; cout.flush();
workPerWI = 4096;
timed = run_kernel(queue, kernel_v2, globalSize, localSize, iters);
gflops = ((float)globalWIs * workPerWI) / timed / 1e3f;
cout << gflops << endl;
///////////////////////////////////////////////////////////////////////////
// Vector width 4
cout << TAB TAB TAB "double4 : "; cout.flush();
workPerWI = 4096;
timed = run_kernel(queue, kernel_v4, globalSize, localSize, iters);
gflops = ((float)globalWIs * workPerWI) / timed / 1e3f;
cout << gflops << endl;
///////////////////////////////////////////////////////////////////////////
// Vector width 8
cout << TAB TAB TAB "double8 : "; cout.flush();
workPerWI = 4096;
timed = run_kernel(queue, kernel_v8, globalSize, localSize, iters);
gflops = ((float)globalWIs * workPerWI) / timed / 1e3f;
cout << gflops << endl;
///////////////////////////////////////////////////////////////////////////
// Vector width 16
cout << TAB TAB TAB "double16 : "; cout.flush();
workPerWI = 4096;
timed = run_kernel(queue, kernel_v16, globalSize, localSize, iters);
gflops = ((float)globalWIs * workPerWI) / timed / 1e3f;
cout << gflops << endl;
///////////////////////////////////////////////////////////////////////////
}
catch(cl::Error error)
{
cerr << error.what() << "(" << error.err() << ")" << endl;
cerr << TAB TAB TAB "Tests skipped" << endl;
return -1;
}
return 0;
}
int clPeak::runTransferBandwidthTest(cl::CommandQueue &queue, cl::Program &prog, device_info_t &devInfo)
{
if(!isTransferBW)
return 0;
float timed, gbps;
cl::NDRange globalSize, localSize;
cl::Context ctx = queue.getInfo<CL_QUEUE_CONTEXT>();
int iters = devInfo.transferBWIters;
Timer timer;
float *arr = NULL;
cl_uint maxItems = devInfo.maxAllocSize / sizeof(float) / 2;
cl_uint numItems;
// Set an upper-limit for cpu devies
if(devInfo.deviceType & CL_DEVICE_TYPE_CPU) {
numItems = roundToPowOf2(maxItems, 26);
} else {
numItems = roundToPowOf2(maxItems);
}
try
{
arr = new float[numItems];
cl::Buffer clBuffer = cl::Buffer(ctx, (CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR), (numItems * sizeof(float)));
log->print(NEWLINE TAB TAB "Transfer bandwidth (GBPS)" NEWLINE);
log->xmlOpenTag("transfer_bandwidth");
log->xmlAppendAttribs("unit", "gbps");
///////////////////////////////////////////////////////////////////////////
// enqueueWriteBuffer
log->print(TAB TAB TAB "enqueueWriteBuffer : ");
// Dummy warm-up
queue.enqueueWriteBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
queue.enqueueWriteBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr, NULL, &timeEvent);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
Timer timer;
timer.start();
for(int i=0; i<iters; i++)
{
queue.enqueueWriteBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
}
queue.finish();
timed = timer.stopAndTime();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
log->print(gbps); log->print(NEWLINE);
log->xmlRecord("enqueuewritebuffer", gbps);
///////////////////////////////////////////////////////////////////////////
// enqueueReadBuffer
log->print(TAB TAB TAB "enqueueReadBuffer : ");
// Dummy warm-up
queue.enqueueReadBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
queue.enqueueReadBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr, NULL, &timeEvent);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
Timer timer;
timer.start();
for(int i=0; i<iters; i++)
{
queue.enqueueReadBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
}
queue.finish();
timed = timer.stopAndTime();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
log->print(gbps); log->print(NEWLINE);
log->xmlRecord("enqueuereadbuffer", gbps);
///////////////////////////////////////////////////////////////////////////
// enqueueMapBuffer
log->print(TAB TAB TAB "enqueueMapBuffer(for read) : ");
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_READ, 0, (numItems * sizeof(float)), NULL, &timeEvent);
queue.finish();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
for(int i=0; i<iters; i++)
{
Timer timer;
void *mapPtr;
timer.start();
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_READ, 0, (numItems * sizeof(float)));
queue.finish();
timed += timer.stopAndTime();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
}
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
log->print(gbps); log->print(NEWLINE);
log->xmlRecord("enqueuemapbuffer", gbps);
///////////////////////////////////////////////////////////////////////////
// memcpy from mapped ptr
log->print(TAB TAB TAB TAB "memcpy from mapped ptr : ");
queue.finish();
timed = 0;
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_READ, 0, (numItems * sizeof(float)));
queue.finish();
timer.start();
memcpy(arr, mapPtr, (numItems * sizeof(float)));
timed += timer.stopAndTime();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
log->print(gbps); log->print(NEWLINE);
log->xmlRecord("memcpy_from_mapped_ptr", gbps);
///////////////////////////////////////////////////////////////////////////
// enqueueUnmap
log->print(TAB TAB TAB "enqueueUnmap(after write) : ");
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_WRITE, 0, (numItems * sizeof(float)));
queue.finish();
queue.enqueueUnmapMemObject(clBuffer, mapPtr, NULL, &timeEvent);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
for(int i=0; i<iters; i++)
{
Timer timer;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_WRITE, 0, (numItems * sizeof(float)));
queue.finish();
timer.start();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
timed += timer.stopAndTime();
}
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
log->print(gbps); log->print(NEWLINE);
log->xmlRecord("enqueueunmap", gbps);
///////////////////////////////////////////////////////////////////////////
// memcpy to mapped ptr
log->print(TAB TAB TAB TAB "memcpy to mapped ptr : ");
queue.finish();
timed = 0;
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_WRITE, 0, (numItems * sizeof(float)));
queue.finish();
timer.start();
memcpy(mapPtr, arr, (numItems * sizeof(float)));
timed += timer.stopAndTime();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
log->print(gbps); log->print(NEWLINE);
log->xmlRecord("memcpy_to_mapped_ptr", gbps);
///////////////////////////////////////////////////////////////////////////
log->xmlCloseTag(); // transfer_bandwidth
if(arr) delete [] arr;
}
catch(cl::Error error)
{
stringstream ss;
ss << error.what() << " (" << error.err() << ")" NEWLINE
<< TAB TAB TAB "Tests skipped" NEWLINE;
log->print(ss.str());
if(arr) delete [] arr;
return -1;
}
return 0;
}
int clPeak::runTransferBandwidthTest(cl::CommandQueue &queue, cl::Program &prog, device_info_t &devInfo)
{
if(!isTransferBW)
return 0;
float timed, gbps;
cl::NDRange globalSize, localSize;
cl::Context ctx = queue.getInfo<CL_QUEUE_CONTEXT>();
int iters = devInfo.transferBWIters;
Timer timer;
cl_uint maxItems = devInfo.maxAllocSize / sizeof(float) / 2;
cl_uint numItems;
// Set an upper-limit for cpu devies
if(devInfo.deviceType & CL_DEVICE_TYPE_CPU) {
numItems = roundToPowOf2(maxItems, 26);
} else {
numItems = roundToPowOf2(maxItems);
}
float *arr = new float[numItems];
try
{
cl::Buffer clBuffer = cl::Buffer(ctx, (CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR), (numItems * sizeof(float)));
cout << NEWLINE TAB TAB "Transfer bandwidth (GBPS)" << endl;
cout << setprecision(2) << fixed;
///////////////////////////////////////////////////////////////////////////
// enqueueWriteBuffer
cout << TAB TAB TAB "enqueueWriteBuffer : "; cout.flush();
// Dummy warm-up
queue.enqueueWriteBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
queue.enqueueWriteBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr, NULL, &timeEvent);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
Timer timer;
timer.start();
for(int i=0; i<iters; i++)
{
queue.enqueueWriteBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
}
queue.finish();
timed = timer.stopAndTime();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
cout << gbps << endl;
///////////////////////////////////////////////////////////////////////////
// enqueueReadBuffer
cout << TAB TAB TAB "enqueueReadBuffer : "; cout.flush();
// Dummy warm-up
queue.enqueueReadBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
queue.enqueueReadBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr, NULL, &timeEvent);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
Timer timer;
timer.start();
for(int i=0; i<iters; i++)
{
queue.enqueueReadBuffer(clBuffer, CL_TRUE, 0, (numItems * sizeof(float)), arr);
}
queue.finish();
timed = timer.stopAndTime();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
cout << gbps << endl;
///////////////////////////////////////////////////////////////////////////
// enqueueMapBuffer
cout << TAB TAB TAB "enqueueMapBuffer(for read) : "; cout.flush();
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_READ, 0, (numItems * sizeof(float)), NULL, &timeEvent);
queue.finish();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
for(int i=0; i<iters; i++)
{
Timer timer;
void *mapPtr;
timer.start();
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_READ, 0, (numItems * sizeof(float)));
queue.finish();
timed += timer.stopAndTime();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
}
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
cout << gbps << endl;
///////////////////////////////////////////////////////////////////////////
// memcpy from mapped ptr
cout << TAB TAB TAB TAB "memcpy from mapped ptr : "; cout.flush();
queue.finish();
timed = 0;
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_READ, 0, (numItems * sizeof(float)));
queue.finish();
timer.start();
memcpy(arr, mapPtr, (numItems * sizeof(float)));
timed += timer.stopAndTime();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
cout << gbps << endl;
///////////////////////////////////////////////////////////////////////////
// enqueueUnmap
cout << TAB TAB TAB "enqueueUnmap(after write) : "; cout.flush();
queue.finish();
timed = 0;
if(useEventTimer)
{
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_WRITE, 0, (numItems * sizeof(float)));
queue.finish();
queue.enqueueUnmapMemObject(clBuffer, mapPtr, NULL, &timeEvent);
queue.finish();
timed += timeInUS(timeEvent);
}
} else
{
for(int i=0; i<iters; i++)
{
Timer timer;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_WRITE, 0, (numItems * sizeof(float)));
queue.finish();
timer.start();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
timed += timer.stopAndTime();
}
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
cout << gbps << endl;
///////////////////////////////////////////////////////////////////////////
// memcpy to mapped ptr
cout << TAB TAB TAB TAB "memcpy to mapped ptr : "; cout.flush();
queue.finish();
timed = 0;
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
void *mapPtr;
mapPtr = queue.enqueueMapBuffer(clBuffer, CL_TRUE, CL_MAP_WRITE, 0, (numItems * sizeof(float)));
queue.finish();
timer.start();
memcpy(mapPtr, arr, (numItems * sizeof(float)));
timed += timer.stopAndTime();
queue.enqueueUnmapMemObject(clBuffer, mapPtr);
queue.finish();
}
timed /= iters;
gbps = ((float)numItems * sizeof(float)) / timed / 1e3f;
cout << gbps << endl;
///////////////////////////////////////////////////////////////////////////
}
catch(cl::Error error)
{
cerr << error.what() << "(" << error.err() << ")" << endl;
cerr << TAB TAB TAB "Tests skipped" << endl;
if(arr) delete [] arr;
return -1;
}
if(arr) delete [] arr;
return 0;
}