TEST(testinout, main) {
if(!EasyCL::isOpenCLAvailable()) {
cout << "opencl library not found" << endl;
exit(-1);
}
cout << "found opencl library" << endl;
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
CLKernel *kernel = cl->buildKernelFromString(getKernel(), "test", "");
float inout[5];
for(int i = 0; i < 5; i++) {
inout[i] = i * 3;
}
kernel->inout(5, inout);
size_t global = 5;
size_t local = 5;
kernel->run(1, &global, &local);
assertEquals(inout[0] , 7);
assertEquals(inout[1] , 10);
assertEquals(inout[2] , 13);
assertEquals(inout[3] , 16);
assertEquals(inout[4] , 19);
cout << "tests completed ok" << endl;
delete kernel;
delete cl;
}
// this runs an entire kernel to get one value. Clearly this is going to be pretty slow, but
// at least it's more or less compatible, and comparable, to how cutorch does it
void THClStorage_set(THClState *state, THClStorage *self, long index, float value)
{
//// cout << "set size=" << self->size << " index=" << index << " value=" << value << endl;
THArgCheck((index >= 0) && (index < self->size), 2, "index out of bounds");
THArgCheck(self->wrapper != 0, 1, "storage argument not initialized, is empty");
// if( self->wrapper->isDeviceDirty() ) { // we have to do this, since we're going to copy it all back again
// // although I suppose we could set via a kernel perhaps
// // either way, this function is pretty inefficient right now :-P
// if(state->trace) cout << "wrapper->copyToHost() size " << self->size << endl;
// self->wrapper->copyToHost();
// }
// self->data[index] = value;
// if(state->trace) cout << "wrapper->copyToDevice() size " << self->size << endl;
// self->wrapper->copyToDevice();
const char *uniqueName = __FILE__ ":set";
EasyCL *cl = self->cl; // cant remember if this is a good idea or not :-P
CLKernel *kernel = 0;
if(cl->kernelExists(uniqueName)) {
kernel = cl->getKernel(uniqueName);
} else {
TemplatedKernel kernelBuilder(cl);
kernel = kernelBuilder.buildKernel( uniqueName, __FILE__, getSetKernelSource(), "THClStorageSet" );
}
kernel->inout(self->wrapper);
kernel->in((int64_t)index);
kernel->in(value);
kernel->run_1d(1, 1);
if(state->addFinish) cl->finish();
}
TEST(testlocal, notUselocal) {
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
CLKernel *kernel = cl->buildKernel("testlocal.cl", "notUseLocal");
int workgroupSize = 64;
float *myarray = new float[workgroupSize];
kernel->in(workgroupSize);
kernel->inout(workgroupSize, myarray);
kernel->run_1d(workgroupSize, workgroupSize);
delete[]myarray;
delete kernel;
delete cl;
}
TEST(testfloatarray, main) {
if(!EasyCL::isOpenCLAvailable()) {
cout << "opencl library not found" << endl;
exit(-1);
}
cout << "found opencl library" << endl;
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
CLKernel *kernel = cl->buildKernelFromString(getKernel(), "test", "");
float in[5];
float inout[5];
float out[5];
for(int i = 0; i < 5; i++) {
in[i] = i * 3;
inout[i] = i * 3;
}
kernel->in(5, in);
kernel->out(5, out);
kernel->inout(5, inout);
kernel->run_1d(5, 5);
for(int i = 0; i < 5; i++) {
cout << out[i] << " ";
}
cout << endl;
for(int i = 0; i < 5; i++) {
cout << inout[i] << " ";
}
cout << endl;
assertEquals(inout[0], 7);
assertEquals(inout[1] , 10);
assertEquals(inout[2] , 34);
assertEquals(inout[3] , 16);
assertEquals(inout[4], 19);
assertEquals(out[0] , 5);
assertEquals(out[1] , 8);
assertEquals(out[2] , 26);
assertEquals(out[3] , 14);
assertEquals(out[4] , 17);
cout << "tests completed ok" << endl;
delete kernel;
delete cl;
}
VIRTUAL void GpuOp::apply1_inplace( int N, CLWrapper*destinationWrapper, Op1 *op ) {
StatefulTimer::instance()->timeCheck("GpuOp::apply inplace start" );
string kernelName = "GpuOp::" + op->getName() + "_inplace";
if( !cl->kernelExists( kernelName ) ) {
buildKernel( kernelName, op, true );
}
CLKernel *kernel = cl->getKernel( kernelName );
kernel->in( N );
kernel->inout( destinationWrapper );
int globalSize = N;
int workgroupSize = 64;
int numWorkgroups = ( globalSize + workgroupSize - 1 ) / workgroupSize;
kernel->run_1d( numWorkgroups * workgroupSize, workgroupSize );
cl->finish();
StatefulTimer::instance()->timeCheck("GpuOp::apply inplace end" );
}
TEST(testlocal, localreduce) {
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
CLKernel *kernel = cl->buildKernel("testlocal.cl", "reduceViaScratch");
int workgroupSize = min(512, cl->getMaxWorkgroupSize());
float *myarray = new float[workgroupSize];
// Timer timer;
for(int i = 0; i < 2000; i++) {
float sumViaCpu = 0;
for(int i = 0; i < workgroupSize; i++) {
myarray[i] = (i + 7) * 3;
sumViaCpu += myarray[i];
}
EXPECT_NE(myarray[0], sumViaCpu);
kernel->inout(workgroupSize, myarray)->localFloats(workgroupSize);
kernel->run_1d(workgroupSize, workgroupSize);
EXPECT_EQ(myarray[0], sumViaCpu);
}
// timer.timeCheck("after iterations");
delete[]myarray;
delete kernel;
delete cl;
}
