void Context::Init(int device_id) {
device_id_ = device_id;
SwitchDevice();
#if defined(USE_CUDA)
if (blas_handle_ == nullptr) {
CUBLAS_CHECK(cublasCreate((cublasHandle_t*)&blas_handle_));
CHECK_NOTNULL(blas_handle_);
}
#endif
#if defined(USE_CUDNN)
if (cudnn_handle_ == nullptr) {
CUDNN_CHECK(cudnnCreate((cudnnHandle_t*)&cudnn_handle_));
CHECK_NOTNULL(cudnn_handle_);
}
#endif
#if defined(USE_NNPACK)
if (nnpack_handle_ == nullptr) {
CHECK_EQ(nnp_initialize(), nnp_status_success);
nnpack_handle_ = pthreadpool_create(0);
CHECK_NOTNULL(nnpack_handle_);
}
#endif
}
void CuDNNPoolingLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
PoolingLayer<Dtype>::LayerSetUp(bottom, top);
// stride
const int* kernel_shape_data = this->kernel_shape_.cpu_data();
// stride
const int* stride_data = this->stride_.cpu_data();
// padding
const int* pad_data = this->pad_.cpu_data();
int kernel_shape[this->num_spatial_axes_];
int stride[this->num_spatial_axes_];
int pad[this->num_spatial_axes_];
for (int i = 0; i < this->num_spatial_axes_; i++){
kernel_shape[i] = kernel_shape_data[i];
stride[i] = stride_data[i];
pad[i] = pad_data[i];
}
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensorDesc<Dtype>(&bottom_desc_);
cudnn::createTensorDesc<Dtype>(&top_desc_);
cudnn::createPoolingNdDesc<Dtype>(&pooling_desc_,
this->layer_param_.pooling_param().pool(), &mode_,
this->num_spatial_axes_, kernel_shape,
pad, stride);
handles_setup_ = true;
}
void CuDNNSoftmaxLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
SoftmaxLayer<Dtype>::LayerSetUp(bottom, top);
// Initialize CUDNN.
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
cudnn::createTensor4dDesc<Dtype>(&top_desc_);
handles_setup_ = true;
}
void CuDNNReLULayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
ReLULayer<Dtype>::LayerSetUp(bottom, top);
// initialize cuDNN
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensorNdDesc<Dtype>(&bottom_desc_);
cudnn::createTensorNdDesc<Dtype>(&top_desc_);
handles_setup_ = true;
}
void CuDNNTanHLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
TanHLayer<Dtype>::LayerSetUp(bottom, top);
// initialize cuDNN
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensorDesc<Dtype>(&bottom_desc_);
cudnn::createTensorDesc<Dtype>(&top_desc_);
cudnn::createActivationDescriptor<Dtype>(&activ_desc_, CUDNN_ACTIVATION_TANH);
handles_setup_ = true;
}
cudnnHandle_t cudnn_handle()
{
static int init = 0;
static cudnnHandle_t handle;
if(!init) {
cudnnCreate(&handle);
init = 1;
}
return handle;
}
GpuDevice::Impl::Impl(int d) : device(d) {
ActivateDevice();
for (size_t i = 0; i < kParallelism; ++i) {
CUDA_CALL(cudaStreamCreate(&stream[i]));
CUBLAS_CALL(cublasCreate(&cublas_handle[i]));
CUBLAS_CALL(cublasSetStream(cublas_handle[i], stream[i]));
CUDNN_CALL(cudnnCreate(&cudnn_handle[i]));
CUDNN_CALL(cudnnSetStream(cudnn_handle[i], stream[i]));
}
}
void CuDNNReLULayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
ReLULayer<Dtype>::LayerSetUp(bottom, top);
// initialize cuDNN
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
cudnn::createTensor4dDesc<Dtype>(&top_desc_);
handles_setup_ = true;
cudnnCreateActivationDescriptor(&activation_desc_);
cudnnSetActivationDescriptor(activation_desc_, CUDNN_ACTIVATION_RELU, CUDNN_PROPAGATE_NAN, 0.0);
}
void CuDNNPoolingLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
PoolingLayer<Dtype>::LayerSetUp(bottom, top);
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensorDesc<Dtype>(&bottom_desc_);
cudnn::createTensorDesc<Dtype>(&top_desc_);
cudnn::createNdPoolingDesc<Dtype>(&pooling_desc_,
this->layer_param_.pooling_param().pool(), &mode_,
this->kernel_shape_, this->pad_, this->stride_);
handles_setup_ = true;
}
cudnnHandle_t cudnn_handle()
{
static int init[16] = {0};
static cudnnHandle_t handle[16];
int i = cuda_get_device();
if(!init[i]) {
cudnnCreate(&handle[i]);
init[i] = 1;
}
return handle[i];
}
void CuDNNTanHLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
vector<Blob<Dtype>*>* top) {
TanHLayer<Dtype>::LayerSetUp(bottom, top);
// initialize cuDNN
CUDNN_CHECK(cudnnCreate(&handle_));
const int N = bottom[0]->num();
const int K = bottom[0]->channels();
const int H = bottom[0]->height();
const int W = bottom[0]->width();
cudnn::createTensor4dDesc<Dtype>(&bottom_desc_, N, K, H, W);
cudnn::createTensor4dDesc<Dtype>(&top_desc_, N, K, H, W);
}
void CuDNNTanHLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
TanHLayer<Dtype>::LayerSetUp(bottom, top);
// initialize cuDNN
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
cudnn::createTensor4dDesc<Dtype>(&top_desc_);
#if CUDNN_VERSION_MIN(5, 0, 0)
cudnnCreateActivationDescriptor(&activation_desc_);
cudnnSetActivationDescriptor(activation_desc_,
CUDNN_ACTIVATION_TANH, CUDNN_PROPAGATE_NAN, 0);
#endif
handles_setup_ = true;
}
void CuDNNPoolingLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
PoolingLayer<Dtype>::LayerSetUp(bottom, top);
// Sanity check: CUDNN currently only supports pad == 0.
CHECK_EQ(this->pad_h_, 0);
CHECK_EQ(this->pad_w_, 0);
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
cudnn::createTensor4dDesc<Dtype>(&top_desc_);
cudnn::createPoolingDesc<Dtype>(&pooling_desc_,
this->layer_param_.pooling_param().pool(), &mode_,
this->kernel_h_, this->kernel_w_, this->stride_h_, this->stride_w_);
handles_setup_ = true;
}
void CuDNNConvolutionLayer<Dtype>::LayerSetUp(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
ConvolutionLayer<Dtype>::LayerSetUp(bottom, top);
// Initialize CUDA streams and cuDNN.
stream_ = new cudaStream_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
handle_ = new cudnnHandle_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
workspaceSizeInBytes = 0;
workspace = NULL;
workspace = NULL;
workspaceSizeInBytes = (size_t)0;
for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
CUDA_CHECK(cudaStreamCreate(&stream_[g]));
CUDNN_CHECK(cudnnCreate(&handle_[g]));
CUDNN_CHECK(cudnnSetStream(handle_[g], stream_[g]));
}
// Set the indexing parameters.
weight_offset_ = (this->num_output_ / this->group_)
* (this->channels_ / this->group_) * this->kernel_h_ * this->kernel_w_;
bias_offset_ = (this->num_output_ / this->group_);
// Create filter descriptor.
cudnn::createFilterDesc<Dtype>(&filter_desc_,
this->num_output_ / this->group_, this->channels_ / this->group_,
this->kernel_h_, this->kernel_w_);
// Create tensor descriptor(s) for data and corresponding convolution(s).
for (int i = 0; i < bottom.size(); i++) {
cudnnTensorDescriptor_t bottom_desc;
cudnn::createTensor4dDesc<Dtype>(&bottom_desc);
bottom_descs_.push_back(bottom_desc);
cudnnTensorDescriptor_t top_desc;
cudnn::createTensor4dDesc<Dtype>(&top_desc);
top_descs_.push_back(top_desc);
cudnnConvolutionDescriptor_t conv_desc;
cudnn::createConvolutionDesc<Dtype>(&conv_desc);
conv_descs_.push_back(conv_desc);
}
// Tensor descriptor for bias.
if (this->bias_term_) {
cudnn::createTensor4dDesc<Dtype>(&bias_desc_);
}
handles_setup_ = true;
}
void CuDNNPoolingLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
vector<Blob<Dtype>*>* top, const bool init_ps, int* num_tables,
map<string, vector<int> >* layer_name_to_blob_global_idx) {
PoolingLayer<Dtype>::LayerSetUp(bottom, top, init_ps, num_tables,
layer_name_to_blob_global_idx);
CUDNN_CHECK(cudnnCreate(&handle_));
cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
cudnn::createTensor4dDesc<Dtype>(&top_desc_);
cudnn::createPoolingDesc<Dtype>(&pooling_desc_,
this->layer_param_.pooling_param().pool(), &mode_,
this->kernel_h_, this->kernel_w_, this->pad_h_, this->pad_w_,
this->stride_h_, this->stride_w_);
handles_setup_ = true;
}
void CuDNNLRNLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
LRNLayer<Dtype>::LayerSetUp(bottom, top);
CUDNN_CHECK(cudnnCreate(&handle_));
CUDNN_CHECK(cudnnCreateLRNDescriptor(&norm_desc_));
cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
cudnn::createTensor4dDesc<Dtype>(&top_desc_);
// create a LRN handle
handles_setup_ = true;
size_ = this->layer_param().lrn_param().local_size();
alpha_ = this->layer_param().lrn_param().alpha();
beta_ = this->layer_param().lrn_param().beta();
k_ = this->layer_param().lrn_param().k();
}
//Constructor
CDACN::CDACN(ActionDescriptor descriptor, float discount, int temporal_stride, int capacity, int burn_in, int batch_size, float base_learning_rate, float actor_learning_rate, float critic_learning_rate, float d, float advantage, unsigned int cycles) :
Agent(descriptor,temporal_stride,84,84,capacity,burn_in,batch_size),
discount_(discount),
exploration_rate_(1.0),
uniform_real(0.0,1.0),
base_learning_rate_(base_learning_rate),
actor_learning_rate_(actor_learning_rate),
critic_learning_rate_(critic_learning_rate),
d_(d),
advantage_(advantage),
cycles_(cycles)
{
cudnnCreate(&handle);
actor_critic = ActorCritic(handle,{1,batch_size},temporal_stride,descriptor_.action_dim(),base_learning_rate_,actor_learning_rate_,critic_learning_rate_);
actor_critic_n = ActorCritic(handle,{1,batch_size},temporal_stride,descriptor_.action_dim(),base_learning_rate_,actor_learning_rate_,critic_learning_rate_);
//Sync parameters
actor_critic.transfer(actor_critic_n);
}
inline void InitCuDNN() {
init_cudnn_ = false;
dtype_ = CUDNN_DATA_FLOAT;
switch(mode) {
case kMaxPooling: mode_ = CUDNN_POOLING_MAX; break;
// case kAvgPooling: mode_ = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; break;
default: utils::Error("This should not happen -,-"); break;
}
CUDA_CHECK(cudnnCreate(&handle_));
CUDA_CHECK(cudnnCreateTensorDescriptor(&in_desc_));
CUDA_CHECK(cudnnCreateTensorDescriptor(&out_desc_));
CUDA_CHECK(cudnnCreatePoolingDescriptor(&pooling_desc_));
CUDA_CHECK(cudnnSetPooling2dDescriptor(pooling_desc_, mode_,
Parent::param_.kernel_height,
Parent::param_.kernel_width,
0, 0,
Parent::param_.stride,
Parent::param_.stride));
}
GpuDevice::GpuDevice(uint64_t device_id, DeviceListener* l, int gpu_id) : ThreadedDevice(device_id, l, kParallelism), device_(gpu_id) {
CUDA_CALL(cudaSetDevice(device_));
cudaFree(0); // Initialize
auto allocator = [this](size_t len) -> void* {
void* ret;
CUDA_CALL(cudaSetDevice(device_));
CUDA_CALL(cudaMalloc(&ret, len));
return ret;
};
auto deallocator = [this](void* ptr) {
CUDA_CALL(cudaSetDevice(device_));
CUDA_CALL(cudaFree(ptr));
};
data_store_ = new PooledDataStore(DEFAULT_POOL_SIZE, allocator, deallocator);
for (size_t i = 0; i < kParallelism; ++i) {
CUDA_CALL(cudaStreamCreate(&stream_[i]));
CUBLAS_CALL(cublasCreate(&cublas_handle_[i]));
CUBLAS_CALL(cublasSetStream(cublas_handle_[i], stream_[i]));
CUDNN_CALL(cudnnCreate(&cudnn_handle_[i]));
CUDNN_CALL(cudnnSetStream(cudnn_handle_[i], stream_[i]));
}
}
CuDNNHandle::CuDNNHandle(cudaStream_t stream) {
if (cudnnCreate(&handle_) != CUDNN_STATUS_SUCCESS) {
LOG(ERROR) << "Cannot create cuDNN handle. cuDNN won't be available.";
}
CUDNN_CHECK(cudnnSetStream(handle_, stream));
}
void CuDNNConvolutionLayer<Dtype>::LayerSetUp(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
ConvolutionLayer<Dtype>::LayerSetUp(bottom, top);
// Initialize CUDA streams and cuDNN.
stream_ = new cudaStream_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
handle_ = new cudnnHandle_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
// Initialize algorithm arrays
fwd_algo_ = new cudnnConvolutionFwdAlgo_t[bottom.size()];
bwd_filter_algo_= new cudnnConvolutionBwdFilterAlgo_t[bottom.size()];
bwd_data_algo_ = new cudnnConvolutionBwdDataAlgo_t[bottom.size()];
// initialize size arrays
workspace_fwd_sizes_ = new uint_tp[bottom.size()];
workspace_bwd_filter_sizes_ = new uint_tp[bottom.size()];
workspace_bwd_data_sizes_ = new uint_tp[bottom.size()];
// workspace data
workspaceSizeInBytes = 0;
workspaceData = NULL;
workspace = new void*[this->group_ * CUDNN_STREAMS_PER_GROUP];
for (uint_tp i = 0; i < bottom.size(); ++i) {
// initialize all to default algorithms
fwd_algo_[i] = (cudnnConvolutionFwdAlgo_t)0;
bwd_filter_algo_[i] = (cudnnConvolutionBwdFilterAlgo_t)0;
bwd_data_algo_[i] = (cudnnConvolutionBwdDataAlgo_t)0;
// default algorithms don't require workspace
workspace_fwd_sizes_[i] = 0;
workspace_bwd_data_sizes_[i] = 0;
workspace_bwd_filter_sizes_[i] = 0;
}
for (int_tp g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
CUDA_CHECK(cudaStreamCreate(&stream_[g]));
CUDNN_CHECK(cudnnCreate(&handle_[g]));
CUDNN_CHECK(cudnnSetStream(handle_[g], stream_[g]));
workspace[g] = NULL;
}
// Set the indexing parameters.
bias_offset_ = (this->num_output_ / this->group_);
// Create filter descriptor.
const int_tp* kernel_shape_data = this->kernel_shape_.cpu_data();
const int_tp kernel_h = kernel_shape_data[0];
const int_tp kernel_w = kernel_shape_data[1];
cudnn::createFilterDesc<Dtype>(&filter_desc_,
this->num_output_ / this->group_, this->channels_ / this->group_,
kernel_h, kernel_w);
// Create tensor descriptor(s) for data and corresponding convolution(s).
for (int_tp i = 0; i < bottom.size(); i++) {
cudnnTensorDescriptor_t bottom_desc;
cudnn::createTensor4dDesc<Dtype>(&bottom_desc);
bottom_descs_.push_back(bottom_desc);
cudnnTensorDescriptor_t top_desc;
cudnn::createTensor4dDesc<Dtype>(&top_desc);
top_descs_.push_back(top_desc);
cudnnConvolutionDescriptor_t conv_desc;
cudnn::createConvolutionDesc<Dtype>(&conv_desc);
conv_descs_.push_back(conv_desc);
}
// Tensor descriptor for bias.
if (this->bias_term_) {
cudnn::createTensor4dDesc<Dtype>(&bias_desc_);
}
handles_setup_ = true;
}
void CudnnNdConvolutionLayer<Dtype>::LayerSetUp(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
ConvolutionParameter conv_param =
this->layer_param_.convolution_param();
// Configure the kernel size, padding, stride, and inputs.
CHECK(conv_param.has_kernel_shape())
<< "Kernel shape is required.";
if (conv_param.has_pad_shape()) {
CHECK_EQ(conv_param.kernel_shape().dim_size(),
conv_param.pad_shape().dim_size())
<< "Kernel and Pad shape don't match !";
}
if (conv_param.has_stride_shape()) {
CHECK_EQ(conv_param.kernel_shape().dim_size(),
conv_param.stride_shape().dim_size())
<< "Kernel and Stride shape don't match !";
}
for (int i = 0; i < conv_param.kernel_shape().dim_size(); ++i) {
kernel_shape_.push_back(conv_param.kernel_shape().dim(i));
CHECK_GT(kernel_shape_[i], 0) << "Filter dimensions cannot be zero.";
}
if (conv_param.has_pad_shape()) {
for (int i = 0; i < conv_param.kernel_shape().dim_size(); ++i) {
pad_shape_.push_back(conv_param.pad_shape().dim(i));
}
} else {
pad_shape_ = std::vector<int>(kernel_shape_.size(), 0);
}
if (conv_param.has_stride_shape()) {
for (int i = 0; i < conv_param.kernel_shape().dim_size(); ++i) {
stride_shape_.push_back(conv_param.stride_shape().dim(i));
}
} else {
stride_shape_ = std::vector<int>(kernel_shape_.size(), 1);
}
// Configure output channels and groups.
channels_ = bottom[0]->shape(1);
num_output_ = this->layer_param_.convolution_param().num_output();
CHECK_GT(num_output_, 0);
group_ = this->layer_param_.convolution_param().group();
CHECK_EQ(channels_ % group_, 0);
CHECK_EQ(num_output_ % group_, 0)
<< "Number of output should be multiples of group.";
// Handle the parameters: weights and biases.
// - blobs_[0] holds the filter weights
// - blobs_[1] holds the biases (optional)
bias_term_ = this->layer_param_.convolution_param().bias_term();
vector<int> weight_shape(kernel_shape_);
weight_shape.insert(weight_shape.begin(), channels_ / group_);
weight_shape.insert(weight_shape.begin(), num_output_);
if (this->blobs_.size() > 0) {
LOG(INFO) << "Skipping parameter initialization";
} else {
if (bias_term_) {
this->blobs_.resize(2);
} else {
this->blobs_.resize(1);
}
// Initialize and fill the weights:
// output channels x input channels per-group x kernel height x kernel width
this->blobs_[0].reset(new Blob<Dtype>(weight_shape));
shared_ptr<Filler<Dtype> > weight_filler(GetFiller<Dtype>(
this->layer_param_.convolution_param().weight_filler()));
weight_filler->Fill(this->blobs_[0].get());
// If necessary, initialize and fill the biases.
if (bias_term_) {
vector<int> bias_shape(1, num_output_);
this->blobs_[1].reset(new Blob<Dtype>(bias_shape));
shared_ptr<Filler<Dtype> > bias_filler(GetFiller<Dtype>(
this->layer_param_.convolution_param().bias_filler()));
bias_filler->Fill(this->blobs_[1].get());
}
}
// Propagate gradients to the parameters (as directed by backward pass).
this->param_propagate_down_.resize(this->blobs_.size(), true);
// Initialize CUDA streams and cuDNN.
stream_ = new cudaStream_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
handle_ = new cudnnHandle_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
workspaceSizeInBytes = 0;
workspace_data_ = NULL;
for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
CUDA_CHECK(cudaStreamCreate(&stream_[g]));
CUDNN_CHECK(cudnnCreate(&handle_[g]));
CUDNN_CHECK(cudnnSetStream(handle_[g], stream_[g]));
}
// Set the indexing parameters.
weight_shape[0] /= group_;
weight_offset_ = 1;
for (int i = 0; i < weight_shape.size(); ++i) {
weight_offset_ *= weight_shape[i];
}
bias_offset_ = weight_shape[0];
// Create filter descriptor.
cudnn::createNdFilterDesc<Dtype>(&filter_desc_, weight_shape);
bwd_filter_algo_= new cudnnConvolutionBwdFilterAlgo_t[bottom.size()];
bwd_data_algo_ = new cudnnConvolutionBwdDataAlgo_t[bottom.size()];
workspace_bwd_filter_sizes_ = new size_t[bottom.size()];
workspace_bwd_data_sizes_ = new size_t[bottom.size()];
workspace_ = new void*[this->group_ * CUDNN_STREAMS_PER_GROUP];
// Create tensor descriptor(s) for data and corresponding convolution(s).
for (int i = 0; i < bottom.size(); i++) {
cudnnTensorDescriptor_t bottom_desc;
cudnn::createTensorDesc<Dtype>(&bottom_desc);
bottom_descs_.push_back(bottom_desc);
cudnnTensorDescriptor_t top_desc;
cudnn::createTensorDesc<Dtype>(&top_desc);
top_descs_.push_back(top_desc);
cudnnConvolutionDescriptor_t conv_desc;
cudnn::createConvolutionDesc<Dtype>(&conv_desc);
conv_descs_.push_back(conv_desc);
workspace_bwd_data_sizes_[i] = 0;
workspace_bwd_filter_sizes_[i] = 0;
}
// Tensor descriptor for bias.
if (this->bias_term_) {
cudnn::createTensorDesc<Dtype>(&bias_desc_);
}
handles_setup_ = true;
}
SaberStatus VenderConv2DActPooling<NV, AK_FLOAT, AK_FLOAT, AK_FLOAT, NCHW, NCHW, NCHW>::\
create(const std::vector<DataTensor_in *>& inputs,
std::vector<DataTensor_out *>& outputs,
ConvActivePoolingParam<OpTensor>& param, Context<NV> &ctx) {
if (!(ctx == this->_ctx)) {
if (_handle != NULL) {
CUDNN_CHECK(cudnnDestroy(_handle));
}
this->_ctx = ctx;
cudaStream_t cuda_stream;
cuda_stream = ctx.get_compute_stream();
CUDNN_CHECK(cudnnCreate(&_handle));
CUDNN_CHECK(cudnnSetStream(_handle, cuda_stream));
}
int input_num = inputs[0]->num();
int input_channel = inputs[0]->channel();
int input_height = inputs[0]->height();
int input_width = inputs[0]->width();
int output_channel = outputs[0]->channel();
int output_height = outputs[0]->height();
int output_width = outputs[0]->width();
{
_inner_shape = inputs[0]->shape();
_inner_shape[0] = input_num;
_inner_shape[1] = param.conv_param.weight()->num();
int kernel_exten = param.conv_param.dilation_h *
(param.conv_param.weight()->height() - 1) + 1;
int output_dim = (input_height + 2 * param.conv_param.pad_h - kernel_exten)
/ param.conv_param.stride_h + 1;
_inner_shape[2] = output_dim;
kernel_exten = param.conv_param.dilation_w *
(param.conv_param.weight()->width() - 1) + 1;
output_dim = (input_width + 2 * param.conv_param.pad_w - kernel_exten)
/ param.conv_param.stride_w + 1;
_inner_shape[3] = output_dim;
_inner_tensor.re_alloc(_inner_shape);
}
int kernel_h = param.conv_param.weight()->height();
int kernel_w = param.conv_param.weight()->width();
int filter_dim_a[] = {output_channel,
input_channel / param.conv_param.group,
kernel_h, kernel_w};
cudnn::setNDFilterDesc<OpDataType>(&_filter_desc,
param.conv_param.weight()->dims(),
filter_dim_a, CUDNN_TENSOR_NCHW);
Shape in_stride = inputs[0]->get_stride();
Shape inner_stride = _inner_tensor.get_stride();
Shape out_stride = outputs[0]->get_stride();
int dim_a[] = {input_num, input_channel,
input_height, input_width};
int dim_inner[] = {_inner_shape[0], _inner_shape[1],
_inner_shape[2], _inner_shape[3]};
int dim_b[] = {input_num, output_channel,
output_height, output_width};
cudnn::setTensorNdDesc<InDataType >(&_input_descs,
inputs[0]->dims(), dim_a, &in_stride[0]);
cudnn::setTensorNdDesc<InDataType >(&_inner_descs,
4, dim_inner,
&inner_stride[0]);
cudnn::setTensorNdDesc<InDataType>(&_output_descs,
outputs[0]->dims(), dim_b, &out_stride[0]);
int pad_a[] = {param.conv_param.pad_h, param.conv_param.pad_w};
int filter_stride_a[] = {param.conv_param.stride_h, param.conv_param.stride_w};
int dilation_a[] = {param.conv_param.dilation_h, param.conv_param.dilation_w};
cudnn::setConvolutionNdDesc<OpDataType >(&_conv_descs,
inputs[0]->dims() - 2, pad_a,
filter_stride_a, dilation_a);
// set activation descriptor
if (param.has_activation) {
cudnn::set_activation_des<OpDataType>(&_active_descs, param.activation_param.active);
}
if (param.has_pooling) {
int windowHeight[] = {param.pooling_param.window_h,
param.pooling_param.window_w};
int padding[] = {param.pooling_param.pad_h,
param.pooling_param.pad_w};
int stride[] = {param.pooling_param.stride_h,
param.pooling_param.stride_w};
cudnn::set_nd_pooling_des<OpDataType >(&_pooling_descs,
param.pooling_param.pooling_type,
_inner_tensor.dims() - 2,
windowHeight,
padding,stride);
}
// true: use tensor core
// false: disable tensor core
cudnn::set_math_type<OpDataType>(&_conv_descs, _use_tensor_core);
cudnn::set_group_count<OpDataType>(&_conv_descs, param.conv_param.group);
// Get fastest implement of cudnn
// set up algo and workspace size
if (param.conv_param.group == inputs[0]->channel() && \
inputs[0]->channel() == outputs[0]->channel()) {
_fwd_algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;//CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
} else {
CUDNN_CHECK(cudnnGetConvolutionForwardAlgorithm(_handle, \
_input_descs, _filter_desc, _conv_descs, _inner_descs, \
_preference, _workspace_limit_bytes, &_fwd_algo));
}
CUDNN_CHECK(cudnnGetConvolutionForwardWorkspaceSize(_handle,
_input_descs, _filter_desc,
_conv_descs, _inner_descs,
_fwd_algo, &_workspace_fwd_sizes));
if (_workspace_fwd_sizes > _workspaceSizeInBytes) {
_workspaceSizeInBytes = _workspace_fwd_sizes;
if (_workspaceData != NULL) {
cudaFree(_workspaceData);
}
cudaMalloc(&_workspaceData, _workspaceSizeInBytes);
_workspace = reinterpret_cast<char*>(_workspaceData);
}
if (param.conv_param.bias()->size()> 0) {
int dim_bias[] = {1, output_channel, 1, 1};
int stride_bias[] = {output_channel, 1, 1, 1};
cudnn::setTensorNdDesc<OpDataType >(&_bias_desc,
4, dim_bias, stride_bias);
}
return SaberSuccess;
}
CUDNN() {
CUDNN_CHECK(cudnnCreate(&handle_));
}