void ConvBC01CuDNN<T>::fprop(const T *imgs, const T *filters, int n_imgs,
int n_channels, int n_filters, int img_h, int img_w, int filter_h,
int filter_w, T *convout) {
bool set_conv_desc = false;
if (n_imgs != this->n_imgs || n_channels != this->n_channels ||
img_h != this->img_h || img_w != this->img_w) {
CUDNN_CHECK(cudnnSetTensor4dDescriptor(
imgs_desc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, n_imgs, n_channels,
img_h, img_w
));
this->n_imgs = n_imgs;
this->n_channels = n_channels;
this->img_h = img_h;
this->img_w = img_w;
set_conv_desc = true;
}
if (n_filters != this->n_filters || n_channels != this->n_channels ||
filter_h != this->filter_h || filter_w != this->filter_w) {
CUDNN_CHECK(cudnnSetFilter4dDescriptor(
filters_desc, CUDNN_DATA_FLOAT, n_filters, n_channels, filter_h,
filter_w
));
this->n_filters = n_filters;
this->n_channels = n_channels;
this->filter_h = filter_h;
this->filter_w = filter_w;
set_conv_desc = true;
}
if (set_conv_desc) {
CUDNN_CHECK(cudnnSetConvolution2dDescriptor(
conv_desc, pad_y, pad_x, stride_y, stride_x, 1, 1, CUDNN_CONVOLUTION
));
int n, c, h, w;
CUDNN_CHECK(cudnnGetConvolution2dForwardOutputDim(
conv_desc, imgs_desc, filters_desc, &n, &c, &h, &w
));
CUDNN_CHECK(cudnnSetTensor4dDescriptor(
convout_desc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, n, c, h, w
));
CUDNN_CHECK(cudnnGetConvolutionForwardAlgorithm(
CUDNN::handle(), imgs_desc, filters_desc, conv_desc, convout_desc,
CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, WORKSPACE_LIMIT,
&fwd_algo
));
CUDNN_CHECK(cudnnGetConvolutionForwardWorkspaceSize(
CUDNN::handle(), imgs_desc, filters_desc, conv_desc, convout_desc,
fwd_algo, &workspace_size
));
}
void *workspace = NULL;
if (workspace_size > 0) {
workspace = CUDA::buffer(workspace_size);
}
CUDNN_CHECK(cudnnConvolutionForward(
CUDNN::handle(), &CUDNN::one, imgs_desc, imgs, filters_desc, filters,
conv_desc, fwd_algo, workspace, workspace_size, &CUDNN::zero,
convout_desc, convout
));
}
void ConvBC01CuDNN<T>::fprop(const T *imgs, const T *filters, int n_imgs,
int n_channels, int n_filters, int img_h, int img_w, int filter_h,
int filter_w, T *convout) {
bool set_conv_desc = false;
if (n_imgs != this->n_imgs || n_channels != this->n_channels ||
img_h != this->img_h || img_w != this->img_w) {
CUDNN_CHECK(cudnnSetTensor4dDescriptor(
imgs_desc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, n_imgs, n_channels,
img_h, img_w
));
this->n_imgs = n_imgs;
this->n_channels = n_channels;
this->img_h = img_h;
this->img_w = img_w;
set_conv_desc = true;
}
if (n_filters != this->n_filters || n_channels != this->n_channels ||
filter_h != this->filter_h || filter_w != this->filter_w) {
CUDNN_CHECK(cudnnSetFilter4dDescriptor(
filters_desc, CUDNN_DATA_FLOAT, n_filters, n_channels, filter_h,
filter_w
));
this->n_filters = n_filters;
this->n_channels = n_channels;
this->filter_h = filter_h;
this->filter_w = filter_w;
set_conv_desc = true;
}
if (set_conv_desc) {
CUDNN_CHECK(cudnnSetConvolution2dDescriptor(
conv_desc, pad_y, pad_x, stride_y, stride_x, 1, 1, CUDNN_CONVOLUTION
));
int n, c, h, w;
CUDNN_CHECK(cudnnGetConvolution2dForwardOutputDim(
conv_desc, imgs_desc, filters_desc, &n, &c, &h, &w
));
CUDNN_CHECK(cudnnSetTensor4dDescriptor(
convout_desc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, n, c, h, w
));
const int n_requestedAlgo = 10;
int n_returnedAlgo;
cudnnConvolutionFwdAlgoPerf_t fwd_algo_perf[n_requestedAlgo];
CUDNN_CHECK(cudnnFindConvolutionForwardAlgorithm(
CUDNN::handle(), imgs_desc, filters_desc, conv_desc, convout_desc,
n_requestedAlgo, &n_returnedAlgo, fwd_algo_perf
));
if (n_returnedAlgo == 0) {
throw std::runtime_error("No cudnnConvolutionFwdAlgoPerf_t found");
}
fwd_algo = fwd_algo_perf[0].algo;
cudnnConvolutionBwdDataAlgoPerf_t bwd_data_algo_perf[n_requestedAlgo];
CUDNN_CHECK(cudnnFindConvolutionBackwardDataAlgorithm(
CUDNN::handle(), filters_desc, convout_desc, conv_desc, imgs_desc,
n_requestedAlgo, &n_returnedAlgo, bwd_data_algo_perf
));
if (n_returnedAlgo == 0) {
throw std::runtime_error("No cudnnConvolutionBwdDataAlgoPerf_t found");
}
bwd_imgs_algo = bwd_data_algo_perf[0].algo;
cudnnConvolutionBwdFilterAlgoPerf_t bwd_filters_algo_perf[n_requestedAlgo];
CUDNN_CHECK(cudnnFindConvolutionBackwardFilterAlgorithm(
CUDNN::handle(), imgs_desc, convout_desc, conv_desc, filters_desc,
n_requestedAlgo, &n_returnedAlgo, bwd_filters_algo_perf
));
if (n_returnedAlgo == 0) {
throw std::runtime_error("No cudnnConvolutionBwdFilterAlgoPerf_t found");
}
bwd_filters_algo = bwd_filters_algo_perf[0].algo;
size_t fwd_workspace_size;
size_t bwd_imgs_workspace_size;
size_t bwd_filters_workspace_size;
CUDNN_CHECK(cudnnGetConvolutionForwardWorkspaceSize(
CUDNN::handle(), imgs_desc, filters_desc, conv_desc, convout_desc,
fwd_algo, &fwd_workspace_size
));
CUDNN_CHECK(cudnnGetConvolutionBackwardDataWorkspaceSize(
CUDNN::handle(), filters_desc, convout_desc, conv_desc, imgs_desc,
bwd_imgs_algo, &bwd_imgs_workspace_size
));
CUDNN_CHECK(cudnnGetConvolutionBackwardFilterWorkspaceSize(
CUDNN::handle(), imgs_desc, convout_desc, conv_desc, filters_desc,
bwd_filters_algo, &bwd_filters_workspace_size
));
workspace_size = std::max(fwd_workspace_size, bwd_imgs_workspace_size);
workspace_size = std::max(workspace_size, bwd_filters_workspace_size);
}
void *workspace = NULL;
if (workspace_size > 0) {
workspace = CUDA::buffer(workspace_size);
}
CUDNN_CHECK(cudnnConvolutionForward(
CUDNN::handle(), &CUDNN::one, imgs_desc, imgs, filters_desc, filters,
conv_desc, fwd_algo, workspace, workspace_size, &CUDNN::zero,
convout_desc, convout
));
}
