void fully_connected_layer_updater_cuda::enqueue_backward_weights_propagation(
cudaStream_t stream_id,
const std::vector<cuda_linear_buffer_device::const_ptr>& schema_data,
const std::vector<cuda_linear_buffer_device::ptr>& gradient,
const std::vector<cuda_linear_buffer_device::const_ptr>& data_custom,
const std::vector<cuda_linear_buffer_device::const_ptr>& input_neurons_buffers,
cuda_linear_buffer_device::const_ptr output_errors_buffer,
const std::vector<cuda_linear_buffer_device::const_ptr>& persistent_working_data,
cuda_linear_buffer_device::ptr temporary_working_fixed_buffer,
cuda_linear_buffer_device::ptr temporary_working_per_entry_buffer,
cuda_linear_buffer_device::const_ptr temporary_fixed_buffer,
cuda_linear_buffer_device::const_ptr temporary_per_entry_buffer,
unsigned int entry_count)
{
// Update weights
{
cublas_safe_call(cublasSetStream(cuda_config->get_cublas_handle(), stream_id));
float alpha = 1.0F;
float beta = 1.0F;
cublas_safe_call(cublasSgemm(
cuda_config->get_cublas_handle(),
CUBLAS_OP_N,
CUBLAS_OP_T,
input_elem_count_per_entry_list[0],
output_elem_count_per_entry,
entry_count,
&alpha,
*input_neurons_buffers[0],
input_elem_count_per_entry_list[0],
*output_errors_buffer,
output_elem_count_per_entry,
&beta,
*gradient[0],
input_elem_count_per_entry_list[0]));
}
// Update biases
if (bias)
{
cudnn_safe_call(cudnnSetStream(cuda_config->get_cudnn_handle(), stream_id));
cudnn_util::set_tensor_descriptor(
output_data_desc,
output_configuration_specific,
entry_count);
float alpha = 1.0F;
float beta = 1.0F;
cudnn_safe_call(cudnnConvolutionBackwardBias(
cuda_config->get_cudnn_handle(),
&alpha,
output_data_desc,
*output_errors_buffer,
&beta,
bias_desc,
*gradient[1]));
}
}
void convolution_layer_updater_cuda::enqueue_update_weights(
unsigned int offset_input_entry_id,
cudaStream_t stream_id,
const std::vector<cuda_linear_buffer_device_smart_ptr>& gradient,
const std::vector<cuda_linear_buffer_device_smart_ptr>& data_custom,
const std::vector<const_cuda_linear_buffer_device_smart_ptr>& schema_data,
cuda_linear_buffer_device_smart_ptr output_errors_buffer,
const_cuda_linear_buffer_device_smart_ptr input_neurons_buffer,
const std::vector<cuda_linear_buffer_device_smart_ptr>& additional_buffers,
std::vector<cuda_memobject_smart_ptr>& dynamic_memobjects,
unsigned int entry_count,
bool force_deterministic)
{
cudnn_safe_call(cudnnSetStream(cuda_config->get_cudnn_handle(), stream_id));
cudnn_safe_call(cudnnSetTensor4dDescriptor(
input_data_desc,
CUDNN_TENSOR_NCHW,
CUDNN_DATA_FLOAT,
entry_count,
input_configuration_specific.feature_map_count,
(input_configuration_specific.dimension_sizes.size() > 1) ? input_configuration_specific.dimension_sizes[1] : 1,
input_configuration_specific.dimension_sizes[0]));
cudnn_safe_call(cudnnSetTensor4dDescriptor(
output_data_desc,
CUDNN_TENSOR_NCHW,
CUDNN_DATA_FLOAT,
entry_count,
output_configuration_specific.feature_map_count,
(output_configuration_specific.dimension_sizes.size() > 1) ? output_configuration_specific.dimension_sizes[1] : 1,
output_configuration_specific.dimension_sizes[0]));
{
float alpha = 1.0F;
float beta = 1.0F;
cudnn_safe_call(cudnnConvolutionBackwardFilter(
cuda_config->get_cudnn_handle(),
&alpha,
input_data_desc,
(const float *)(*input_neurons_buffer) + input_elem_count_per_entry * offset_input_entry_id,
output_data_desc,
*output_errors_buffer,
convolution_desc,
&beta,
weights_desc,
*gradient[0]));
}
{
float alpha = 1.0F;
float beta = 1.0F;
cudnn_safe_call(cudnnConvolutionBackwardBias(
cuda_config->get_cudnn_handle(),
&alpha,
output_data_desc,
*output_errors_buffer,
&beta,
bias_desc,
*gradient[1]));
}
}
void convolution_layer_updater_cuda::enqueue_backward_weights_propagation(
cudaStream_t stream_id,
const std::vector<cuda_linear_buffer_device::const_ptr>& schema_data,
const std::vector<cuda_linear_buffer_device::ptr>& gradient,
const std::vector<cuda_linear_buffer_device::const_ptr>& data_custom,
const std::vector<cuda_linear_buffer_device::const_ptr>& input_neurons_buffers,
cuda_linear_buffer_device::const_ptr output_errors_buffer,
const std::vector<cuda_linear_buffer_device::const_ptr>& persistent_working_data,
cuda_linear_buffer_device::ptr temporary_working_fixed_buffer,
cuda_linear_buffer_device::ptr temporary_working_per_entry_buffer,
cuda_linear_buffer_device::const_ptr temporary_fixed_buffer,
cuda_linear_buffer_device::const_ptr temporary_per_entry_buffer,
unsigned int entry_count)
{
cudnn_safe_call(cudnnSetStream(cuda_config->get_cudnn_handle(), stream_id));
cudnn_util::set_tensor_descriptor(
input_data_desc,
input_configuration_specific_list[0],
entry_count);
cudnn_util::set_tensor_descriptor(
output_data_desc,
output_configuration_specific,
entry_count);
{
void * workspace = 0;
size_t workspace_size = 0;
if (temporary_working_fixed_buffer)
{
workspace = *temporary_working_fixed_buffer;
workspace_size = temporary_working_fixed_buffer->get_size();
}
cudnnConvolutionBwdFilterAlgo_t algo = cuda_config->cudnn_find_convolution_backward_weights_algo(
input_data_desc,
weights_desc,
convolution_desc,
output_data_desc,
*input_neurons_buffers[0],
*output_errors_buffer,
(unsigned char *)workspace,
(unsigned char *)workspace + update_weights_find_algo_working_buffer_size,
workspace_size - update_weights_find_algo_working_buffer_size);
float alpha = 1.0F;
float beta = 1.0F;
cudnn_safe_call(cudnnConvolutionBackwardFilter(
cuda_config->get_cudnn_handle(),
&alpha,
input_data_desc,
*input_neurons_buffers[0],
output_data_desc,
*output_errors_buffer,
convolution_desc,
algo,
workspace,
workspace_size,
&beta,
weights_desc,
*gradient[0]));
}
if (bias)
{
float alpha = 1.0F;
float beta = 1.0F;
cudnn_safe_call(cudnnConvolutionBackwardBias(
cuda_config->get_cudnn_handle(),
&alpha,
output_data_desc,
*output_errors_buffer,
&beta,
bias_desc,
*gradient[1]));
}
}
