1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418
//! Describes utility functionality for CUDA cuDNN.
use super::{
ActivationDescriptor, ConvolutionDescriptor, DropoutDescriptor, FilterDescriptor,
NormalizationDescriptor, PoolingDescriptor, RnnDescriptor,
};
use crate::cuda::CudaDeviceMemory;
use crate::ffi::*;
use num::traits::*;
#[derive(Debug, Copy, Clone)]
/// Defines the available data types for the CUDA cuDNN data representation.
pub enum DataType {
/// F32
Float,
/// F64
Double,
/// F16 (no native Rust support yet)
Half,
}
/// CuDnn type info for generic use.
pub trait DataTypeInfo {
/// Mostly internal.
fn cudnn_data_type() -> DataType;
fn size() -> usize;
}
impl DataTypeInfo for f32 {
fn cudnn_data_type() -> DataType {
DataType::Float
}
fn size() -> usize {
4_usize
}
}
impl DataTypeInfo for f64 {
fn cudnn_data_type() -> DataType {
DataType::Double
}
fn size() -> usize {
8_usize
}
}
// TODO f16
#[allow(missing_debug_implementations, missing_copy_implementations)]
/// Provides a convenient interface to access cuDNN's convolution parameters,
/// `algo` and `workspace` and `workspace_size_in_bytes`.
///
/// You woudn't use this struct yourself, but rather obtain it through `Cudnn.init_convolution()`.
pub struct ConvolutionConfig {
forward_algo: cudnnConvolutionFwdAlgo_t,
backward_filter_algo: cudnnConvolutionBwdFilterAlgo_t,
backward_data_algo: cudnnConvolutionBwdDataAlgo_t,
forward_workspace_size: usize,
backward_filter_workspace_size: usize,
backward_data_workspace_size: usize,
conv_desc: ConvolutionDescriptor,
filter_desc: FilterDescriptor,
}
impl ConvolutionConfig {
/// Returns a new ConvolutionConfig
#[allow(clippy::too_many_arguments)]
pub fn new(
algo_fwd: cudnnConvolutionFwdAlgo_t,
workspace_size_fwd: usize,
algo_filter_bwd: cudnnConvolutionBwdFilterAlgo_t,
workspace_filter_size_bwd: usize,
algo_data_bwd: cudnnConvolutionBwdDataAlgo_t,
workspace_data_size_bwd: usize,
conv_desc: ConvolutionDescriptor,
filter_desc: FilterDescriptor,
) -> ConvolutionConfig {
ConvolutionConfig {
forward_algo: algo_fwd,
forward_workspace_size: workspace_size_fwd,
backward_filter_algo: algo_filter_bwd,
backward_filter_workspace_size: workspace_filter_size_bwd,
backward_data_algo: algo_data_bwd,
backward_data_workspace_size: workspace_data_size_bwd,
conv_desc,
filter_desc,
}
}
/// Returns the largest workspace size out of the three.
///
/// Useful for creating a shared workspace.
pub fn largest_workspace_size(&self) -> usize {
if self.backward_data_workspace_size() >= self.backward_filter_workspace_size()
&& self.backward_data_workspace_size() >= self.forward_workspace_size()
{
self.backward_data_workspace_size()
} else if self.backward_filter_workspace_size() >= self.backward_data_workspace_size()
&& self.backward_filter_workspace_size() >= self.forward_workspace_size()
{
self.backward_filter_workspace_size()
} else {
self.forward_workspace_size()
}
}
/// Returns `forward_algo`.
pub fn forward_algo(&self) -> &cudnnConvolutionFwdAlgo_t {
&self.forward_algo
}
/// Returns `forward_workspace_size`.
pub fn forward_workspace_size(&self) -> usize {
self.forward_workspace_size
}
/// Returns `backward_filter_algo`.
pub fn backward_filter_algo(&self) -> &cudnnConvolutionBwdFilterAlgo_t {
&self.backward_filter_algo
}
/// Returns `backward_filter_workspace_size`.
pub fn backward_filter_workspace_size(&self) -> usize {
self.backward_filter_workspace_size
}
/// Returns `backward_data_algo`.
pub fn backward_data_algo(&self) -> &cudnnConvolutionBwdDataAlgo_t {
&self.backward_data_algo
}
/// Returns `backward_data_workspace_size`.
pub fn backward_data_workspace_size(&self) -> usize {
self.backward_data_workspace_size
}
/// Returns `conv_desc`.
pub fn conv_desc(&self) -> &ConvolutionDescriptor {
&self.conv_desc
}
/// Returns `filter_desc`.
pub fn filter_desc(&self) -> &FilterDescriptor {
&self.filter_desc
}
}
#[allow(missing_debug_implementations, missing_copy_implementations)]
/// Provides a convenient interface to access cuDNN's Normalization (LRN) Descriptor.
///
/// You woudn't use this struct yourself, but rather obtain it through `Cudnn.init_normalization()`.
pub struct NormalizationConfig {
lrn_desc: NormalizationDescriptor,
}
impl NormalizationConfig {
/// Returns a new LRN Config.
pub fn new(lrn_desc: NormalizationDescriptor) -> NormalizationConfig {
NormalizationConfig { lrn_desc }
}
/// Returns `lrn_desc`.
pub fn lrn_desc(&self) -> &NormalizationDescriptor {
&self.lrn_desc
}
}
#[allow(missing_debug_implementations, missing_copy_implementations)]
/// Provides a convenient interface to access cuDNN's Pooling Descriptor.
///
/// You woudn't use this struct yourself, but rather obtain it through `Cudnn.init_pooling()`.
pub struct PoolingConfig {
pooling_avg_desc: PoolingDescriptor,
pooling_max_desc: PoolingDescriptor,
}
impl PoolingConfig {
/// Returns a new PoolingConfig.
pub fn new(
pooling_avg_desc: PoolingDescriptor,
pooling_max_desc: PoolingDescriptor,
) -> PoolingConfig {
PoolingConfig {
pooling_avg_desc,
pooling_max_desc,
}
}
/// Returns `pooling_avg_desc`.
pub fn pooling_avg_desc(&self) -> &PoolingDescriptor {
&self.pooling_avg_desc
}
/// Returns `pooling_max_desc`.
pub fn pooling_max_desc(&self) -> &PoolingDescriptor {
&self.pooling_max_desc
}
}
#[allow(missing_debug_implementations, missing_copy_implementations)]
/// Provides a convenient interface to access cuDNN's Activation Descriptor.
///
/// You woudn't use this struct yourself, but rather obtain it through `Cudnn.init_activation()`.
pub struct ActivationConfig {
activation_sigmoid_desc: ActivationDescriptor,
activation_relu_desc: ActivationDescriptor,
activation_clipped_relu_desc: ActivationDescriptor,
activation_tanh_desc: ActivationDescriptor,
}
impl ActivationConfig {
/// Returns a new ActivationConfig.
pub fn new(
activation_sigmoid_desc: ActivationDescriptor,
activation_relu_desc: ActivationDescriptor,
activation_clipped_relu_desc: ActivationDescriptor,
activation_tanh_desc: ActivationDescriptor,
) -> ActivationConfig {
ActivationConfig {
activation_sigmoid_desc,
activation_relu_desc,
activation_clipped_relu_desc,
activation_tanh_desc,
}
}
/// Returns `activation_sigmoid_desc`.
pub fn activation_sigmoid_desc(&self) -> &ActivationDescriptor {
&self.activation_sigmoid_desc
}
/// Returns `activation_relu_desc`.
pub fn activation_relu_desc(&self) -> &ActivationDescriptor {
&self.activation_relu_desc
}
/// Returns `activation_clipped_relu_desc`.
pub fn activation_clipped_relu_desc(&self) -> &ActivationDescriptor {
&self.activation_clipped_relu_desc
}
/// Returns `activation_tanh_desc`.
pub fn activation_tanh_desc(&self) -> &ActivationDescriptor {
&self.activation_tanh_desc
}
}
#[allow(missing_debug_implementations, missing_copy_implementations)]
/// Provides a convenient interface to access cuDNN's Dropout Descriptor.
///
/// You wouldn't use this struct yourself, but rather obtain it through `Cudnn.init_dropout()`.
#[derive(Debug)]
pub struct DropoutConfig {
dropout_desc: DropoutDescriptor,
reserve_space: CudaDeviceMemory,
}
impl DropoutConfig {
/// Returns a new DropoutConfig.
pub fn new(dropout_desc: DropoutDescriptor, reserve: CudaDeviceMemory) -> DropoutConfig {
DropoutConfig {
dropout_desc,
reserve_space: reserve,
}
}
/// Returns `dropout_desc`.
pub fn dropout_desc(&self) -> &DropoutDescriptor {
&self.dropout_desc
}
/// Take the Reserve Memory of the DropoutDescriptor
pub fn take_mem(self) -> CudaDeviceMemory {
self.reserve_space
}
/// Returns the reserved space.
pub fn reserved_space(&self) -> &CudaDeviceMemory {
&self.reserve_space
}
}
#[allow(missing_debug_implementations, missing_copy_implementations)]
/// Provides an interfaces for CUDNN's Rnn Descriptor
/// # Arguments
/// * `rnn_desc` Previously created descriptor
/// * `hidden_size` Size of the hidden layer
/// * `num_layers` Number of layers
/// * `dropout_desc` Descriptor to a previously created & initialized dropout descriptor, applied
/// between layers.
/// * `input_mode` Specifies behaviour at the input to the first layer
/// * `direction_mode` Specifies the recurrence pattern - i.e bidirectional
/// * `rnn_mode` Type of network used in routines ForwardInference, ForwardTraining, BackwardData,
/// BackwardWeights. Can be ReLU, tanh, LSTM (Long Short Term Memory), or GRU (Gated Recurrent Unit).
/// * `algo` - Only required in v6 implementation FIXME: Should this be checked in compilation?
/// * `data_type` Math Precision - default f32
///
/// The LSTM network offered by CUDNN is a four-gate network that does not use peephole connections.
/// Greff, et al. (2015)[1] suggests it doesn't matter what kind of network it is, although
/// Jozefowicz, et al. (2015)[2] suggests that the most important gates are the forget and input,
/// followed by the output gate, so the peephole connection isn't as important to be concerned with.
/// A positive bias, as encouraged in the paper, can be achieved by setting `bias_mode` to
/// CUDNN_RNN_DOUBLE_BIAS, which is the default, or CUDN_RNN_SINGLE_INP_BIAS or
/// CUDNN_RNN_SINGLE_REC_BIAS
///
/// [1]: arxiv.org/pdf/1503.04069.pdf
/// [2]: jmlr.org/proceedings/papers/v37/jozefowicz15.pdf
pub struct RnnConfig {
rnn_desc: RnnDescriptor,
/// Size of Hidden Layer
pub hidden_size: ::libc::c_int,
/// Number of Hidden Layers
pub num_layers: ::libc::c_int,
/// Length of Sequence
pub sequence_length: ::libc::c_int,
dropout_desc: cudnnDropoutDescriptor_t,
input_mode: cudnnRNNInputMode_t,
direction_mode: cudnnDirectionMode_t,
rnn_mode: cudnnRNNMode_t,
algo: cudnnRNNAlgo_t,
data_type: cudnnDataType_t,
workspace_size: usize,
training_reserve_size: usize,
training_reserve: CudaDeviceMemory,
}
impl RnnConfig {
/// Initialise a RNN Config
#[allow(clippy::too_many_arguments)]
pub fn new(
rnn_desc: RnnDescriptor,
hidden_size: i32,
num_layers: i32,
sequence_length: i32,
dropout_desc: cudnnDropoutDescriptor_t,
input_mode: cudnnRNNInputMode_t,
direction_mode: cudnnDirectionMode_t,
rnn_mode: cudnnRNNMode_t,
// Requires v6
algo: cudnnRNNAlgo_t,
data_type: cudnnDataType_t,
workspace_size: usize,
training_reserve_size: usize,
training_reserve: CudaDeviceMemory,
) -> RnnConfig {
RnnConfig {
rnn_desc,
hidden_size,
num_layers,
sequence_length,
dropout_desc,
input_mode,
direction_mode,
rnn_mode,
algo,
data_type,
workspace_size,
training_reserve_size,
training_reserve,
}
}
/// Workspace Size required for RNN Operations
pub fn rnn_workspace_size(&self) -> usize {
self.workspace_size
}
/// Largest Workspace Size for RNN
pub fn largest_workspace_size(&self) -> usize {
self.rnn_workspace_size()
}
/// Training Reserve Size for RNN
pub fn training_reserve_size(&self) -> usize {
self.training_reserve_size
}
/// Training Reserve Space on GPU for RNN
pub fn training_reserve(&self) -> &CudaDeviceMemory {
&self.training_reserve
}
/// Accessor function for Rnn Descriptor
pub fn rnn_desc(&self) -> &RnnDescriptor {
&self.rnn_desc
}
/// Accessor function for Sequence Length
pub fn sequence_length(&self) -> &i32 {
&self.sequence_length
}
}
#[allow(missing_debug_implementations, missing_copy_implementations)]
/// Provides a convenient interface for cuDNN's scaling parameters `alpha` and `beta`.
///
/// Scaling parameters lend the source value with prior value in the destination
/// tensor as follows: dstValue = alpha[0]*srcValue + beta[0]*priorDstValue. When beta[0] is
/// zero, the output is not read and can contain any uninitialized data (including NaN). The
/// storage data type for alpha[0], beta[0] is float for HALF and SINGLE tensors, and double
/// for DOUBLE tensors. These parameters are passed using a host memory pointer.
///
/// For improved performance it is advised to use beta[0] = 0.0. Use a non-zero value for
/// beta[0] only when blending with prior values stored in the output tensor is needed.
pub struct ScalParams<T>
where
T: Float + DataTypeInfo,
{
/// Alpha
pub a: T,
/// Beta
pub b: T,
}
impl<T> Default for ScalParams<T>
where
T: Float + Zero + One + DataTypeInfo,
{
/// Provides default values for ScalParams<f32>.
fn default() -> ScalParams<T> {
ScalParams {
a: One::one(),
b: Zero::zero(),
}
}
}