// 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
// lgfgs expects a working implementation of this method
float THClStorage_get(THClState *state, const THClStorage *self, long index)
{
//// printf("THClStorage_get\n");
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() ) {
// if(state->trace) cout << "wrapper->copyToHost()" << endl;
// self->wrapper->copyToHost();
// }
// return self->data[index];
const char *uniqueName = __FILE__ ":get";
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__, getGetKernelSource(), "THClStorageGet" );
}
float res;
kernel->out(1, &res);
kernel->in(self->wrapper);
kernel->in((int64_t)index);
kernel->run_1d(1, 1);
if(state->addFinish) cl->finish();
return res;
}
// 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(testkernelstore, 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", "");
EXPECT_FALSE(cl->kernelExists("kernela"));
EXPECT_FALSE(cl->kernelExists("kernelb"));
cl->storeKernel("kernela", kernel);
EXPECT_TRUE(cl->kernelExists("kernela"));
EXPECT_FALSE(cl->kernelExists("kernelb"));
EXPECT_EQ(kernel, cl->getKernel("kernela"));
delete kernel;
delete cl;
}
void col2im(THClState *state, THClTensor* col, const int channels,
const int height, const int width, const int patch_h, const int patch_w, const int pad_h,
const int pad_w, const int stride_h, const int stride_w, THClTensor* im) {
int height_col = (height + 2 * pad_h - patch_h) / stride_h + 1;
int width_col = (width + 2 * pad_w - patch_w) / stride_w + 1;
int num_kernels = channels * height * width;
// To avoid involving atomic operations, we will launch one kernel per
// bottom dimension, and then in the kernel add up the top dimensions.
EasyCL *cl = im->storage->cl;
std::string uniqueName = "SpatialConvolutionMM::col2im";
CLKernel *kernel = 0;
if(cl->kernelExists(uniqueName)) {
kernel = cl->getKernel(uniqueName);
} else {
TemplatedKernel kernelBuilder(cl);
kernel = kernelBuilder.buildKernel(uniqueName, "SpatialConvolutionMM.cl",
SpatialConvolutionMM_getKernelTemplate(), "col2im_kernel");
}
THClKernels k(state, kernel);
k.in(num_kernels);
k.in(col);
k.in(height);
k.in(width);
k.in(channels);
k.in(patch_h);
k.in(patch_w);
k.in(pad_h);
k.in(pad_w);
k.in(stride_h);
k.in(stride_w);
k.in(height_col);
k.in(width_col);
k.out(im);
k.run(GET_BLOCKS(state, num_kernels), getNumThreads(state));
// col2im_kernel <<<GET_BLOCKS(num_kernels), CL_NUM_THREADS, 0, stream>>> (
// num_kernels, data_col, height, width, channels,
// patch_h, patch_w, pad_h, pad_w, stride_h, stride_w,
// height_col, width_col, data_im
// );
// THError("Not implemented");
}
void im2col(THClState *state, THClTensor* im, const int channels,
const int height, const int width, const int ksize_h, const int ksize_w, const int pad_h,
const int pad_w, const int stride_h, const int stride_w, THClTensor* col) {
// We are going to launch channels * height_col * width_col kernels, each
// kernel responsible for copying a single-channel grid.
int height_col = (height + 2 * pad_h - ksize_h) / stride_h + 1;
int width_col = (width + 2 * pad_w - ksize_w) / stride_w + 1;
int num_kernels = channels * height_col * width_col;
std::string uniqueName = "SpatialConvolutionMM::im2col";
EasyCL *cl = im->storage->cl;
CLKernel *kernel = 0;
if(cl->kernelExists(uniqueName)) {
kernel = cl->getKernel(uniqueName);
} else {
TemplatedKernel kernelBuilder(cl);
kernel = kernelBuilder.buildKernel(uniqueName, "SpatialConvolutionMM.cl",
SpatialConvolutionMM_getKernelTemplate(), "im2col_kernel");
}
THClKernels k(state, kernel);
k.in(num_kernels);
k.in(im);
k.in(height);
k.in(width);
k.in(ksize_h);
k.in(ksize_w);
k.in(pad_h);
k.in(pad_w);
k.in(stride_h);
k.in(stride_w);
k.in(height_col);
k.in(width_col);
k.out(col);
k.run(GET_BLOCKS(state, num_kernels), getNumThreads(state));
// Launch
// im2col_kernel <<<GET_BLOCKS(num_kernels), CL_NUM_THREADS, 0, stream>>> (
// num_kernels, data_im, height, width, ksize_h, ksize_w,
// pad_h, pad_w, stride_h, stride_w,
// height_col, width_col, data_col
// );
}
void kernelLaunch_THClTensor_reduceNoncontigDim(
THClState *state,
dim3 &grid,
dim3 &block,
int ADims,
int BDims,
TensorInfo<IndexType> out,
TensorInfo<IndexType> in,
IndexType reductionStride,
IndexType reductionSize,
IndexType totalSlices,
float init,
HasOperator2 const*modifyOp,
HasOperator3 const*reduceOp) {
// cl->finish();
StatefulTimer::timeCheck("Reduce-noncontig START");
std::string uniqueName = "THClTensor_reduceNoncontigDim_" + easycl::toString(ADims) + "_" + easycl::toString(BDims) + "_" +
TypeParseTraits<IndexType>::name + "_" + modifyOp->operator2() + "_" + reduceOp->operator3();
EasyCL *cl = in.wrapper->getCl();
CLKernel *kernel = 0;
if(cl->kernelExists(uniqueName)) {
kernel = cl->getKernel(uniqueName);
StatefulTimer::timeCheck("Apply3 1aa");
} else {
// launch kernel here....
TemplatedKernel kernelBuilder(cl);
std::set<int> dims_set;
if(ADims >= 0) {
dims_set.insert(ADims);
}
if(BDims >= 0) {
dims_set.insert(BDims);
}
std::vector<int> dims;
for( std::set<int>::iterator it = dims_set.begin(); it != dims_set.end(); it++ ) {
dims.push_back(*it);
}
std::string indexType = TypeParseTraits<IndexType>::name;
kernelBuilder
.set("include_THClReduceApplyUtils", THClReduceApplyUtils_getKernelTemplate())
.set("dims", dims)
.set("dim1", ADims)
.set("dim2", BDims)
.set("defreduceblock", 1)
.set("WarpSize", 32) // probably can do like 'if nvidia 32 else 64' ?
.set("MAX_CLTORCH_DIMS", MAX_CLTORCH_DIMS)
.set("IndexType", indexType)
.set("modify_operation", modifyOp->operator2())
.set("reduce_operation", reduceOp->operator3())
;
kernel = kernelBuilder.buildKernel(uniqueName, "THClReduce.cl", THClReduce_getKernelSource(), "THClTensor_reduceNoncontigDim");
}
THClKernels k(state, kernel);
k.out(out);
k.in(in);
k.in((int)reductionStride);
k.in((int)reductionSize);
k.in((int)totalSlices);
k.in(init);
k.run(grid, block);
if(state->addFinish) cl->finish();
StatefulTimer::timeCheck("Reduce-noncontig END");
}
static int clnn_SpatialMaxPooling_updateOutput(lua_State *L)
{
THClState *state = getCltorchState(L);
THClTensor *input = (THClTensor *)luaT_checkudata(L, 2, "torch.ClTensor");
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int padW = luaT_getfieldcheckint(L, 1, "padW");
int padH = luaT_getfieldcheckint(L, 1, "padH");
bool ceil_mode = luaT_getfieldcheckboolean(L, 1, "ceil_mode");
THClTensor *output = (THClTensor *)luaT_getfieldcheckudata(L, 1, "output", "torch.ClTensor");
THClTensor *indices = (THClTensor *)luaT_getfieldcheckudata(L, 1, "indices", "torch.ClTensor");
THAssert(THClTensor_checkGPU(state, 3, input, output, indices));
luaL_argcheck(L, input->nDimension == 3 || input->nDimension == 4, 2, "3D or 4D (batch) tensor expected");
long nInputCols, nInputRows, nInputPlane, batchSize;
long nOutputCols, nOutputRows;
if (input->nDimension == 3) {
nInputCols = input->size[2];
nInputRows = input->size[1];
nInputPlane = input->size[0];
batchSize = 1;
}
else
{
nInputCols = input->size[3];
nInputRows = input->size[2];
nInputPlane = input->size[1];
batchSize = input->size[0];
}
luaL_argcheck(L, nInputCols >= kW - padW && nInputRows >= kH - padH, 2, "input image smaller than kernel size");
luaL_argcheck(L, kW/2 >= padW && kH/2 >= padH, 2, "pad should be smaller than half of kernel size");
if(ceil_mode) {
nOutputCols = ceil(float(nInputCols - kW + 2*padW) / float(dW)) + 1;
nOutputRows = ceil(float(nInputRows - kH + 2*padH) / float(dH)) + 1;
}
else {
nOutputCols = floor(float(nInputCols - kW + 2*padW) / float(dW)) + 1;
nOutputRows = floor(float(nInputRows - kH + 2*padH) / float(dH)) + 1;
}
input = THClTensor_newContiguous(state, input);
//float* input_data = THClTensor_data(state, input);
THClTensor_resize4d(state, output, batchSize, nInputPlane, nOutputRows, nOutputCols);
THClTensor_resizeAs(state, indices, output);
//float* indices_data = THClTensor_data(state, indices);
// float* output_data = THClTensor_data(state, output);
int count = THClTensor_nElement(state, output);
EasyCL *cl = input->storage->cl;
std::string uniqueName = __FILE__ "MaxPoolForward";
CLKernel *kernel = 0;
if(cl->kernelExists(uniqueName)) {
kernel = cl->getKernel(uniqueName);
} else {
TemplatedKernel kernelBuilder(cl);
kernelBuilder.set("forward", 1);
kernelBuilder.set("Dtype", "float");
kernel = kernelBuilder.buildKernel(uniqueName, __FILE__,
SpatialMaxPooling_getKernelTemplate(), "MaxPoolForward");
}
THClKernels k(state, kernel);
k.in((int)count);
k.in(input);
k.in((int)batchSize);
k.in((int)nInputPlane);
k.in((int)nInputRows);
k.in((int)nInputCols);
k.in((int)nOutputRows);
k.in((int)nOutputCols);
k.in((int)kH);
k.in((int)kW);
k.in((int)dH);
k.in((int)dW);
k.in((int)padH);
k.in((int)padW);
k.out(output);
k.out(indices);
dim3 blocks(GET_BLOCKS(state, count));
dim3 threads(GET_CL_NUM_THREADS(state));
k.run(blocks, threads);
// maxpool <<<blocks, threads, 0, THClState_getCurrentStream(state)>>> (
// input_data, output_data,
// indices_data+nbatch*nInputPlane*nOutputCols*nOutputRows, indices_data,
// nInputPlane, nInputRows, nInputCols, kH, kW, dH, dW);
// THError("not implemented");
if(input->nDimension == 3)
THClTensor_resize3d(state, output, nInputPlane, nOutputRows, nOutputCols);
THClTensor_free(state, input);
return 1;
}
static int clnn_SpatialMaxPooling_updateGradInput(lua_State *L)
{
THClState *state = getCltorchState(L);
THClTensor *input = (THClTensor *)luaT_checkudata(L, 2, "torch.ClTensor");
THClTensor *gradOutput = (THClTensor *)luaT_checkudata(L, 3, "torch.ClTensor");
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int padW = luaT_getfieldcheckint(L, 1, "padW");
int padH = luaT_getfieldcheckint(L, 1, "padH");
bool ceil_mode = luaT_getfieldcheckboolean(L, 1, "ceil_mode");
THClTensor *gradInput = (THClTensor *)luaT_getfieldcheckudata(L, 1, "gradInput", "torch.ClTensor");
THClTensor *indices = (THClTensor *)luaT_getfieldcheckudata(L, 1, "indices", "torch.ClTensor");
THAssert(THClTensor_checkGPU(state, 4, input, gradOutput, indices, gradInput));
input = THClTensor_newContiguous(state, input);
gradOutput = THClTensor_newContiguous(state, gradOutput);
long nInputCols, nInputRows, nInputPlane, batchSize;
long nOutputCols, nOutputRows;
if (input->nDimension == 3) {
nInputCols = input->size[2];
nInputRows = input->size[1];
nInputPlane = input->size[0];
batchSize = 1;
}
else
{
nInputCols = input->size[3];
nInputRows = input->size[2];
nInputPlane = input->size[1];
batchSize = input->size[0];
}
if(ceil_mode) {
nOutputCols = ceil(float(nInputCols - kW + 2*padW) / float(dW)) + 1;
nOutputRows = ceil(float(nInputRows - kH + 2*padH) / float(dH)) + 1;
}
else {
nOutputCols = floor(float(nInputCols - kW + 2*padW) / float(dW)) + 1;
nOutputRows = floor(float(nInputRows - kH + 2*padH) / float(dH)) + 1;
}
gradOutput = THClTensor_newContiguous(state, gradOutput);
THClTensor_resizeAs(state, gradInput, input);
int count = THClTensor_nElement(state, input);
EasyCL *cl = input->storage->cl;
std::string uniqueName = __FILE__ "MaxPoolBackward";
CLKernel *kernel = 0;
if(cl->kernelExists(uniqueName)) {
kernel = cl->getKernel(uniqueName);
} else {
TemplatedKernel kernelBuilder(cl);
kernelBuilder.set("backward", 1);
kernelBuilder.set("Dtype", "float");
kernel = kernelBuilder.buildKernel(uniqueName, __FILE__,
SpatialMaxPooling_getKernelTemplate(), "MaxPoolBackward");
}
THClKernels k(state, kernel);
k.in((int)count);
k.in(gradOutput);
k.in(indices);
k.in((int)batchSize);
k.in((int)nInputPlane);
k.in((int)nInputRows);
k.in((int)nInputCols);
k.in((int)nOutputRows);
k.in((int)nOutputCols);
k.in((int)kH);
k.in((int)kW);
k.in((int)dH);
k.in((int)dW);
k.in((int)padH);
k.in((int)padW);
k.out(gradInput);
dim3 blocks(GET_BLOCKS(state, count));
dim3 threads(GET_CL_NUM_THREADS(state));
k.run(blocks, threads);
// MaxPoolBackward <<< GET_BLOCKS(count), CL_NUM_THREADS, 0, THClState_getCurrentStream(state) >>>
// (count,
// THClTensor_data(state, gradOutput),
// THClTensor_data(state, indices),
// batchSize, nInputPlane, nInputRows, nInputCols, nOutputRows, nOutputCols,
// kH, kW, dH, dW, padH, padW,
// THClTensor_data(state, gradInput));
// clean
THClTensor_free(state, input);
THClTensor_free(state, gradOutput);
return 1;
}
void THNN_ClSpatialAveragePooling_updateOutput(THClState *state, THClTensor *input, THClTensor *output, int kW, int kH, int dW, int dH, int padW, int padH, bool ceil_mode, bool count_include_pad)
{
THAssert(THClTensor_checkGPU(state, 2, input, output));
THArgCheck(input->nDimension == 3 || input->nDimension == 4, 2, "3D or 4D (batch) tensor expected");
long nInputCols, nInputRows, nInputPlane, batchSize;
long nOutputCols, nOutputRows;
if (input->nDimension == 3) {
nInputCols = input->size[2];
nInputRows = input->size[1];
nInputPlane = input->size[0];
batchSize = 1;
}
else
{
nInputCols = input->size[3];
nInputRows = input->size[2];
nInputPlane = input->size[1];
batchSize = input->size[0];
}
THArgCheck(nInputCols >= kW - 2*padW && nInputRows >= kH - 2*padH, 2, "input image smaller than kernel size");
THArgCheck(kW/2 >= padW && kH/2 >= padH, 2, "pad should be smaller than half of kernel size");
if(ceil_mode) {
nOutputCols = ceil(float(nInputCols - kW + 2*padW) / float(dW)) + 1;
nOutputRows = ceil(float(nInputRows - kH + 2*padH) / float(dH)) + 1;
}
else {
nOutputCols = floor(float(nInputCols - kW + 2*padW) / float(dW)) + 1;
nOutputRows = floor(float(nInputRows - kH + 2*padH) / float(dH)) + 1;
}
if (padW || padH)
{
// ensure that the last pooling starts inside the image
// needed to avoid problems in ceil mode
if ((nOutputRows - 1)*dH >= nInputRows + padH)
--nOutputRows;
if ((nOutputCols - 1)*dW >= nInputCols + padW)
--nOutputCols;
}
input = THClTensor_newContiguous(state, input);
// input_data = THClTensor_data(state, input);
THClTensor_resize4d(state, output, batchSize, nInputPlane, nOutputRows, nOutputCols);
// float* output_data = THClTensor_data(state, output);
int count = THClTensor_nElement(state, output);
EasyCL *cl = input->storage->cl;
const char *uniqueName = 0;
if(count_include_pad) {
uniqueName = __FILE__ "forward_includepad";
} else {
uniqueName = __FILE__ "forward_not_includepad";
}
CLKernel *kernel = 0;
if(cl->kernelExists(uniqueName)) {
kernel = cl->getKernel(uniqueName);
} else {
TemplatedKernel kernelBuilder(cl);
kernelBuilder.set("forward", 1);
kernelBuilder.set("Dtype", "float");
kernelBuilder.set("COUNT_INCLUDE_PAD", count_include_pad);
kernel = kernelBuilder.buildKernel(uniqueName, __FILE__,
getKernelTemplate(), "AvePoolForward");
}
THClKernels k(state, kernel);
k.in((int)count);
k.in(input);
k.in((int)batchSize);
k.in((int)nInputPlane);
k.in((int)nInputRows);
k.in((int)nInputCols);
k.in((int)nOutputRows);
k.in((int)nOutputCols);
k.in((int)kH);
k.in((int)kW);
k.in((int)dH);
k.in((int)dW);
k.in((int)padH);
k.in((int)padW);
k.out(output);
dim3 blocks(GET_BLOCKS(state, count));
dim3 threads(GET_CL_NUM_THREADS(state));
k.run(blocks, threads);
// AvePoolForward<float, true>
// <<<GET_BLOCKS(count), CL_NUM_THREADS, 0, THClState_getCurrentStream(state) >>>(
// count, input_data,
// batchSize, nInputPlane, nInputRows, nInputCols, nOutputRows, nOutputCols,
// kH, kW, dH, dW, padH, padW, output_data);
if(input->nDimension == 3)
THClTensor_resize3d(state, output, nInputPlane, nOutputRows, nOutputCols);
THClTensor_free(state, input);
}