CuDNNConvolutionLayer<Dtype>::~CuDNNConvolutionLayer() {
// Check that handles have been setup before destroying.
if (!handles_setup_) { return; }
for (int_tp i = 0; i < bottom_descs_.size(); i++) {
cudnnDestroyTensorDescriptor(bottom_descs_[i]);
cudnnDestroyTensorDescriptor(top_descs_[i]);
cudnnDestroyConvolutionDescriptor(conv_descs_[i]);
}
if (this->bias_term_) {
cudnnDestroyTensorDescriptor(bias_desc_);
}
cudnnDestroyFilterDescriptor(filter_desc_);
for (int_tp g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
cudaStreamDestroy(stream_[g]);
cudnnDestroy(handle_[g]);
}
cudaFree(workspaceData);
delete [] stream_;
delete [] handle_;
delete [] fwd_algo_;
delete [] bwd_filter_algo_;
delete [] bwd_data_algo_;
delete [] workspace_fwd_sizes_;
delete [] workspace_bwd_data_sizes_;
delete [] workspace_bwd_filter_sizes_;
}
convolution_layer_updater_cuda::~convolution_layer_updater_cuda()
{
cudnnDestroyTensorDescriptor(input_data_desc);
cudnnDestroyTensorDescriptor(output_data_desc);
cudnnDestroyFilterDescriptor(weights_desc);
cudnnDestroyConvolutionDescriptor(convolution_desc);
cudnnDestroyTensorDescriptor(bias_desc);
}
static int c_make_filter(PyGpuArrayObject *var, cudnnFilterDescriptor_t *desc) {
cudnnStatus_t err;
err = cudnnCreateFilterDescriptor(desc);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could not create tensor descriptor: %s",
cudnnGetErrorString(err));
return -1;
}
if (c_set_filter(var, *desc, 1) != 0) {
cudnnDestroyFilterDescriptor(*desc);
return -1;
}
return 0;
}
CuDNNConvolutionLayer<Dtype>::~CuDNNConvolutionLayer() {
for (int i = 0; i < bottom_descs_.size(); i++) {
cudnnDestroyTensor4dDescriptor(bottom_descs_[i]);
cudnnDestroyTensor4dDescriptor(top_descs_[i]);
cudnnDestroyConvolutionDescriptor(conv_descs_[i]);
}
if (this->bias_term_) {
cudnnDestroyTensor4dDescriptor(bias_desc_);
}
cudnnDestroyFilterDescriptor(filter_desc_);
for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
cudaStreamDestroy(stream_[g]);
cudnnDestroy(handle_[g]);
}
delete [] stream_;
delete [] handle_;
}
CuDNNConvolutionLayer<Dtype>::~CuDNNConvolutionLayer() {
// Check that handles have been setup before destroying.
if (!handles_setup_) { return; }
for (int i = 0; i < bottom_descs_.size(); i++) {
cudnnDestroyTensorDescriptor(bottom_descs_[i]);
cudnnDestroyTensorDescriptor(top_descs_[i]);
cudnnDestroyConvolutionDescriptor(conv_descs_[i]);
}
if (this->bias_term_) {
cudnnDestroyTensorDescriptor(bias_desc_);
}
cudnnDestroyFilterDescriptor(filter_desc_);
delete [] fwd_algo_;
delete [] bwd_filter_algo_;
delete [] bwd_data_algo_;
delete [] workspace_fwd_sizes_;
delete [] workspace_bwd_data_sizes_;
delete [] workspace_bwd_filter_sizes_;
}
CudnnNdConvolutionLayer<Dtype>::~CudnnNdConvolutionLayer() {
// Check that handles have been setup before destroying.
if (!handles_setup_) {
return;
}
for (int i = 0; i < bottom_descs_.size(); i++) {
cudnnDestroyTensorDescriptor(bottom_descs_[i]);
cudnnDestroyTensorDescriptor(top_descs_[i]);
cudnnDestroyConvolutionDescriptor(conv_descs_[i]);
}
if (this->bias_term_) {
cudnnDestroyTensorDescriptor(bias_desc_);
}
cudnnDestroyFilterDescriptor(filter_desc_);
for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
cudaStreamDestroy(stream_[g]);
cudnnDestroy(handle_[g]);
}
delete [] stream_;
delete [] handle_;
}
ConvBC01CuDNN<T>::~ConvBC01CuDNN() {
CUDNN_CHECK(cudnnDestroyTensorDescriptor(imgs_desc));
CUDNN_CHECK(cudnnDestroyTensorDescriptor(convout_desc));
CUDNN_CHECK(cudnnDestroyFilterDescriptor(filters_desc));
CUDNN_CHECK(cudnnDestroyConvolutionDescriptor(conv_desc));
}
THFloatTensor *cudnn_SpatialConvolution_updateOutput(struct module *module, THFloatTensor *input)
{
int kW = module->SpatialConvolution.kW;
int kH = module->SpatialConvolution.kH;
int dW = module->SpatialConvolution.dW;
int dH = module->SpatialConvolution.dH;
int padW = module->SpatialConvolution.padW;
int padH = module->SpatialConvolution.padH;
int nInputPlane = module->SpatialConvolution.nInputPlane;
int nOutputPlane = module->SpatialConvolution.nOutputPlane;
THFloatTensor *weight = module->SpatialConvolution.weight;
THFloatTensor *bias = module->SpatialConvolution.bias;
THFloatTensor *output = module->output;
int sizes[4];
int pad[2], filterStride[2], upscale[2];
cudnnTensorDescriptor_t dinput, dbias, doutput;
cudnnConvolutionDescriptor_t dconv;
cudnnFilterDescriptor_t dweight;
float one = 1, zero = 0;
size_t reqwssize;
static void *ws;
static size_t wssize;
static const int alg = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
pad[0] = padH;
pad[1] = padW;
filterStride[0] = dH;
filterStride[1] = dW;
upscale[0] = 1;
upscale[1] = 1;
if(input->nDimension <= 2)
{
// Here we use the SpatialConvolution module to perform a linear transformation
errcheck(cudnnCreateTensorDescriptor(&dinput));
if(input->nDimension == 1)
errcheck(cudnnSetTensor4dDescriptor(dinput, CUDNN_TENSOR_NCHW, floattype, 1, input->size[0], 1, 1));
else errcheck(cudnnSetTensor4dDescriptor(dinput, CUDNN_TENSOR_NCHW, floattype, input->size[0], input->size[1], 1, 1));
} else errcheck(THcudnn_TensorDescriptor(&dinput, input));
errcheck(cudnnCreateFilterDescriptor(&dweight));
errcheck(cudnnSetFilter4dDescriptor(dweight, floattype, nOutputPlane, nInputPlane, kH, kW));
errcheck(cudnnCreateTensorDescriptor(&dbias));
errcheck(cudnnSetTensor4dDescriptor(dbias, CUDNN_TENSOR_NCHW, floattype, 1, bias->size[0], 1, 1));
errcheck(cudnnCreateConvolutionDescriptor(&dconv));
errcheck(cudnnSetConvolutionNdDescriptor(dconv, 2, pad, filterStride, upscale, CUDNN_CROSS_CORRELATION, floattype));
errcheck(cudnnGetConvolutionNdForwardOutputDim(dconv, dinput, dweight, 4, sizes));
THCudaTensor_resize4d(output, sizes[0], sizes[1], sizes[2], sizes[3]);
errcheck(THcudnn_TensorDescriptor(&doutput, output));
if(alg == CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM || alg == CUDNN_CONVOLUTION_FWD_ALGO_GEMM ||
alg == CUDNN_CONVOLUTION_FWD_ALGO_FFT || alg == CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING)
{
errcheck(cudnnGetConvolutionForwardWorkspaceSize(THcudnn_getHandle(), dinput, dweight, dconv, doutput, alg, &reqwssize));
if(reqwssize > wssize)
{
wssize = reqwssize;
errcheck(cudaMalloc(&ws, reqwssize));
}
}
errcheck(cudnnConvolutionForward(THcudnn_getHandle(), &one, dinput, THFloatTensor_data(input),
dweight, THFloatTensor_data(weight), dconv, alg, ws, wssize, &zero,
doutput, THFloatTensor_data(output)));
errcheck(cudnnAddTensor_v3(THcudnn_getHandle(), &one, dbias, THFloatTensor_data(bias),
&one, doutput, THFloatTensor_data(output)));
cudnnDestroyTensorDescriptor(dinput);
cudnnDestroyFilterDescriptor(dweight);
cudnnDestroyTensorDescriptor(dbias);
cudnnDestroyTensorDescriptor(doutput);
cudnnDestroyConvolutionDescriptor(dconv);
return output;
}
int dnn_rnn_fwd(cudnnRNNDescriptor_t desc,
PyGpuArrayObject *w, PyGpuArrayObject *x,
PyGpuArrayObject *hx, PyGpuArrayObject *cx,
gpudata **reserve, PyGpuArrayObject **y,
PyGpuArrayObject **hy, PyGpuArrayObject **cy,
cudnnHandle_t _handle) {
PyGpuContextObject *c = x->context;
cudnnTensorDescriptor_t xdesc = NULL;
cudnnTensorDescriptor_t hxdesc = NULL;
cudnnTensorDescriptor_t cxdesc = NULL;
cudnnTensorDescriptor_t ydesc = NULL;
cudnnTensorDescriptor_t hydesc = NULL;
cudnnTensorDescriptor_t cydesc = NULL;
cudnnFilterDescriptor_t wdesc = NULL;
cudnnTensorDescriptor_t *xl = NULL;
cudnnTensorDescriptor_t *yl = NULL;
gpudata *workspace = NULL;
size_t worksize, ressize;
size_t seqLength = PyGpuArray_DIM(x, 0);
size_t miniBatch = PyGpuArray_DIM(x, 1);
size_t inputSize = PyGpuArray_DIM(x, 2);
size_t hiddenSizeDir = PyGpuArray_DIM(hx, 2);
size_t shape[3];
int strs[3], dims[3];
cudnnStatus_t err;
cudnnDataType_t dt;
int res = -1;
switch (x->ga.typecode) {
case GA_FLOAT:
dt = CUDNN_DATA_FLOAT;
break;
case GA_DOUBLE:
dt = CUDNN_DATA_DOUBLE;
break;
case GA_HALF:
dt = CUDNN_DATA_HALF;
break;
default:
PyErr_SetString(PyExc_TypeError, "Unsupported data type for x");
return -1;
}
// This is early to match the exit() in the fail label.
cuda_enter(c->ctx);
err = cudnnCreateTensorDescriptor(&xdesc);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could not create xdesc: %s",
cudnnGetErrorString(err));
goto fail;
}
dims[0] = PyGpuArray_DIM(x, 1);
dims[1] = PyGpuArray_DIM(x, 2);
dims[2] = 1;
strs[0] = dims[1] * dims[2];
strs[1] = dims[2];
strs[2] = 1;
err = cudnnSetTensorNdDescriptor(xdesc, dt, 3, dims, strs);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could not set xdesc: %s",
cudnnGetErrorString(err));
goto fail;
}
if (c_make_tensorNd(hx, &hxdesc) != 0)
goto fail;
if (cx != NULL)
if (c_make_tensorNd(cx, &cxdesc) != 0)
goto fail;
if (c_make_filter(w, &wdesc) != 0)
goto fail;
shape[0] = seqLength;
shape[1] = miniBatch;
shape[2] = hiddenSizeDir;
if (theano_prep_output(y, 3, shape, x->ga.typecode, GA_C_ORDER, c) != 0)
goto fail;
err = cudnnCreateTensorDescriptor(&ydesc);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could not create ydesc: %s",
cudnnGetErrorString(err));
goto fail;
}
dims[0] = shape[1];
dims[1] = shape[2];
dims[2] = 1;
strs[0] = dims[2] * dims[1];
strs[1] = dims[2];
strs[2] = 1;
err = cudnnSetTensorNdDescriptor(ydesc, dt, 3, dims, strs);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could not set ydesc: %s",
cudnnGetErrorString(err));
goto fail;
}
if (theano_prep_output(hy, 3, PyGpuArray_DIMS(hx),
hx->ga.typecode, GA_C_ORDER, c) != 0)
goto fail;
if (c_make_tensorNd(*hy, &hydesc) != 0)
goto fail;
if (cy != NULL) {
if (theano_prep_output(cy, 3, PyGpuArray_DIMS(cx),
cx->ga.typecode, GA_C_ORDER, c) != 0)
goto fail;
if (c_make_tensorNd(*cy, &cydesc) != 0)
goto fail;
}
xl = (cudnnTensorDescriptor_t *)calloc(sizeof(cudnnTensorDescriptor_t), seqLength);
if (xl == NULL) {
PyErr_NoMemory();
goto fail;
}
for (size_t i = 0; i < seqLength; i++)
xl[i] = xdesc;
yl = (cudnnTensorDescriptor_t *)calloc(sizeof(cudnnTensorDescriptor_t), seqLength);
if (yl == NULL) {
PyErr_NoMemory();
goto fail;
}
for (size_t i = 0; i < seqLength; i++)
yl[i] = ydesc;
err = cudnnGetRNNWorkspaceSize(_handle, desc, (int)seqLength,
xl, &worksize);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could not get worksize: %s",
cudnnGetErrorString(err));
goto fail;
}
workspace = gpudata_alloc(c->ctx, worksize, NULL, 0, NULL);
if (workspace == NULL) {
PyErr_Format(PyExc_RuntimeError, "Could not allocate workspace");
goto fail;
}
err = cudnnGetRNNTrainingReserveSize(_handle, desc, (int)seqLength,
xl, &ressize);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could not get reserve size: %s",
cudnnGetErrorString(err));
goto fail;
}
*reserve = gpudata_alloc(c->ctx, ressize, NULL, 0, NULL);
if (*reserve == NULL) {
PyErr_Format(PyExc_RuntimeError, "Could not allocate reserve");
goto fail;
}
err = cudnnRNNForwardTraining(_handle, desc, (int)seqLength,
xl, PyGpuArray_DEV_DATA(x),
hxdesc, PyGpuArray_DEV_DATA(hx),
cxdesc, cx ? PyGpuArray_DEV_DATA(cx) : NULL,
wdesc, PyGpuArray_DEV_DATA(w),
yl, PyGpuArray_DEV_DATA(*y),
hydesc, PyGpuArray_DEV_DATA(*hy),
cydesc, cy ? PyGpuArray_DEV_DATA(*cy) : NULL,
*(void **)workspace, worksize,
*(void **)(*reserve), ressize);
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
"Could run RNN: %s",
cudnnGetErrorString(err));
goto fail;
}
res = 0;
fail:
if (xdesc != NULL)
cudnnDestroyTensorDescriptor(xdesc);
if (hxdesc != NULL)
cudnnDestroyTensorDescriptor(hxdesc);
if (cxdesc != NULL)
cudnnDestroyTensorDescriptor(cxdesc);
if (wdesc != NULL)
cudnnDestroyFilterDescriptor(wdesc);
if (ydesc != NULL)
cudnnDestroyTensorDescriptor(ydesc);
if (hydesc != NULL)
cudnnDestroyTensorDescriptor(hydesc);
if (cydesc != NULL)
cudnnDestroyTensorDescriptor(cydesc);
free(xl);
free(yl);
if (workspace != NULL)
gpudata_release(workspace);
cuda_exit(c->ctx);
return res;
}
