layer_hessian_cuda::buffer_set layer_hessian_cuda::allocate_all_buffers(unsigned int max_entry_count) const
{
buffer_set res;
std::vector<size_t> sizes = get_sizes_of_additional_buffers_per_entry();
for(std::vector<size_t>::const_iterator it = sizes.begin(); it != sizes.end(); ++it)
{
size_t sz = *it * max_entry_count;
res.additional_buffers.push_back(cuda_linear_buffer_device_smart_ptr(new cuda_linear_buffer_device(sz)));
}
std::vector<size_t> fixed_sizes = get_sizes_of_additional_buffers_fixed();
for(std::vector<size_t>::const_iterator it = fixed_sizes.begin(); it != fixed_sizes.end(); ++it)
{
res.additional_buffers.push_back(cuda_linear_buffer_device_smart_ptr(new cuda_linear_buffer_device(*it)));
}
{
size_t sz = output_elem_count_per_entry * sizeof(float) * max_entry_count;
res.output_neurons_buffer = cuda_linear_buffer_device_smart_ptr(new cuda_linear_buffer_device(sz));
}
if (backprop_required && !is_in_place_backprop())
{
size_t sz = input_elem_count_per_entry * sizeof(float) * max_entry_count;
res.input_errors_buffer = cuda_linear_buffer_device_smart_ptr(new cuda_linear_buffer_device(sz));
}
fill_additional_buffers(res.additional_buffers);
return res;
}
void weight_vector_bound_cuda::update_buffer_configuration(
buffer_cuda_size_configuration& buffer_configuration,
unsigned int updater_entry_count) const
{
std::vector<size_t> per_entry_sizes = get_sizes_of_additional_buffers_per_entry();
for(std::vector<size_t>::const_iterator it = per_entry_sizes.begin(); it != per_entry_sizes.end(); ++it)
buffer_configuration.add_constant_buffer(*it * updater_entry_count);
}
std::vector<cuda_linear_buffer_device_smart_ptr> weight_vector_bound_cuda::allocate_additional_buffers(unsigned int max_entry_count)
{
std::vector<cuda_linear_buffer_device_smart_ptr> res;
std::vector<size_t> per_entry_sizes = get_sizes_of_additional_buffers_per_entry();
for(std::vector<size_t>::const_iterator it = per_entry_sizes.begin(); it != per_entry_sizes.end(); ++it)
res.push_back(cuda_linear_buffer_device_smart_ptr(new cuda_linear_buffer_device(*it * max_entry_count)));
return res;
}
void weight_vector_bound_cuda::update_buffer_configuration(buffer_cuda_size_configuration& buffer_configuration) const
{
std::vector<size_t> per_entry_sizes = get_sizes_of_additional_buffers_per_entry();
for(std::vector<size_t>::const_iterator it = per_entry_sizes.begin(); it != per_entry_sizes.end(); ++it)
buffer_configuration.add_per_entry_buffer(*it);
std::vector<unsigned int> tex_per_entry = get_linear_addressing_through_texture_per_entry();
for(std::vector<unsigned int>::const_iterator it = tex_per_entry.begin(); it != tex_per_entry.end(); ++it)
buffer_configuration.add_per_entry_linear_addressing_through_texture(*it);
}
void layer_hessian_cuda::update_buffer_configuration(buffer_cuda_size_configuration& buffer_configuration) const
{
std::vector<size_t> sizes = get_sizes_of_additional_buffers_per_entry();
for(std::vector<size_t>::const_iterator it = sizes.begin(); it != sizes.end(); ++it)
buffer_configuration.add_per_entry_buffer(*it);
std::vector<size_t> fixed_sized = get_sizes_of_additional_buffers_fixed();
for(std::vector<size_t>::const_iterator it = fixed_sized.begin(); it != fixed_sized.end(); ++it)
buffer_configuration.add_constant_buffer(*it);
buffer_configuration.add_per_entry_buffer(output_elem_count_per_entry * sizeof(float));
if (backprop_required && !is_in_place_backprop())
buffer_configuration.add_per_entry_buffer(input_elem_count_per_entry * sizeof(float));
std::vector<unsigned int> tex_per_entry = get_linear_addressing_through_texture_per_entry();
for(std::vector<unsigned int>::const_iterator it = tex_per_entry.begin(); it != tex_per_entry.end(); ++it)
buffer_configuration.add_per_entry_linear_addressing_through_texture(*it);
}
