Developer Guide

Abstract

This cuDNN Developer Guide provides an overview of cuDNN v7.6.4, and details about the types, enums, and routines within the cuDNN library API.

For previously released cuDNN developer documentation, see cuDNN Archives.

NVIDIA^® cuDNN is a GPU-accelerated library of primitives for deep neural networks. It provides highly tuned implementations of routines arising frequently in DNN applications:

• Convolution forward and backward, including cross-correlation
• Pooling forward and backward
• Softmax forward and backward
• Neuron activations forward and backward:
  • Rectified linear (ReLU)
  • Sigmoid
  • Hyperbolic tangent (TANH)
• Tensor transformation functions
• LRN, LCN and batch normalization forward and backward

cuDNN's convolution routines aim for a performance that is competitive with the fastest GEMM (matrix multiply)-based implementations of such routines, while using significantly less memory.

cuDNN features include customizable data layouts, supporting flexible dimension ordering, striding, and subregions for the 4D tensors used as inputs and outputs to all of its routines. This flexibility allows easy integration into any neural network implementation, and avoids the input/output transposition steps sometimes necessary with GEMM-based convolutions.

cuDNN offers a context-based API that allows for easy multithreading and (optional) interoperability with CUDA streams.

Basic concepts are described in this section.

2.1. Programming Model

The cuDNN Library exposes a Host API but assumes that for operations using the GPU, the necessary data is directly accessible from the device.

An application using cuDNN must initialize a handle to the library context by calling cudnnCreate(). This handle is explicitly passed to every subsequent library function that operates on GPU data. Once the application finishes using cuDNN, it can release the resources associated with the library handle using cudnnDestroy(). This approach allows the user to explicitly control the library's functioning when using multiple host threads, GPUs and CUDA Streams.

For example, an application can use cudaSetDevice() to associate different devices with different host threads, and in each of those host threads, use a unique cuDNN handle that directs the library calls to the device associated with it. Thus the cuDNN library calls made with different handles will automatically run on different devices.

The device associated with a particular cuDNN context is assumed to remain unchanged between the corresponding cudnnCreate() and cudnnDestroy() calls. In order for the cuDNN library to use a different device within the same host thread, the application must set the new device to be used by calling cudaSetDevice() and then create another cuDNN context, which will be associated with the new device, by calling cudnnCreate().

cuDNN API Compatibility

Beginning in cuDNN 7, the binary compatibility of patch and minor releases is maintained as follows:

• Any patch release x.y.z is forward- or backward-compatible with applications built against another cuDNN patch release x.y.w (i.e., of the same major and minor version number, but having w!=z)
• cuDNN minor releases beginning with cuDNN 7 are binary backward-compatible with applications built against the same or earlier patch release (i.e., an app built against cuDNN 7.x is binary compatible with cuDNN library 7.y, where y>=x)
• Applications compiled with a cuDNN version 7.y are not guaranteed to work with 7.x release when y > x.

2.2. Convolution Formulas

This section describes the various convolution formulas implemented in cuDNN convolution functions.

The convolution terms described in the table below apply to all the convolution formulas that follow.

TABLE OF CONVOLUTION TERMS

Normal Convolution (using cross-correlation mode)

$y_{\mathit{n,\; k,\; p,\; q}}=\underset{c}{\overset{C}{\sum }}\underset{r}{\overset{R}{\sum }}\underset{s}{\overset{S}{\sum }}{x}_{\mathit{n,\; c,\; p+r,\; q+s}}\times w_{\mathit{k,c,r,s}}$

Convolution with Padding

$x_{\mathit{<0,\; <0}}=0$

$x_{\mathit{>H,\; >W}}=0$

$y_{\mathit{n,\; k,\; p,\; q}}=\underset{c}{\overset{C}{\sum }}\underset{r}{\overset{R}{\sum }}\underset{s}{\overset{S}{\sum }}{x}_{\mathit{n,\; c,\; p+r-pad,\; q+s-pad}}\times w_{\mathit{k,c,r,s}}$

Convolution with Subsample-Striding

$y_{\mathit{n,\; k,\; p,\; q}}=\underset{c}{\overset{C}{\sum }}\underset{r}{\overset{R}{\sum }}\underset{s}{\overset{S}{\sum }}{x}_{\mathit{n,\; c,\; (p*u)\; +\; r,\; (q*v)\; +\; s}}\times w_{\mathit{k,c,r,s}}$

Convolution with Dilation

$y_{\mathit{n,\; k,\; p,\; q}}=\underset{c}{\overset{C}{\sum }}\underset{r}{\overset{R}{\sum }}\underset{s}{\overset{S}{\sum }}{x}_{\mathit{n,\; c,\; p\; +\; (r*dilh),\; q\; +\; (s*dilw)}}\times w_{\mathit{k,c,r,s}}$

Convolution using Convolution Mode

$y_{\mathit{n,\; k,\; p,\; q}}=\underset{c}{\overset{C}{\sum }}\underset{r}{\overset{R}{\sum }}\underset{s}{\overset{S}{\sum }}{x}_{\mathit{n,\; c,\; p\; +\; r,\; q\; +\; s}}\times w_{\mathit{k,\; c,\; R-r-1,\; S-s-1}}$

Convolution using Grouped Convolution

$C_g=\frac{C}{G}$

$K_g=\frac{K}{G}$

$y_{\mathit{n,\; k,\; p,\; q}}=\underset{c}{\overset{C_g}{\sum }}\underset{r}{\overset{R}{\sum }}\underset{s}{\overset{S}{\sum }}{x}_{\mathit{n,\; Cg*floor(k/Kg)+c,\; p+r,\; q+s}}\times w_{\mathit{k,c,r,s}}$

2.3. Notation

As of CUDNN v4 we have adopted a mathematicaly-inspired notation for layer inputs and outputs using x,y,dx,dy,b,w for common layer parameters. This was done to improve the readability and ease of understanding of the meaning of the parameters. All layers now follow a uniform convention as below:

During Inference:

y = layerFunction(x, otherParams).

During backpropagation:

(dx, dOtherParams) = layerFunctionGradient(x,y,dy,otherParams)

For convolution the notation is

y = x*w+b

where w is the matrix of filter weights, x is the previous layer's data (during inference), y is the next layer's data, b is the bias and * is the convolution operator.

In backpropagation routines the parameters keep their meanings.

The parameters dx,dy,dw,db always refer to the gradient of the final network error function with respect to a given parameter. So dy in all backpropagation routines always refers to error gradient backpropagated through the network computation graph so far. Similarly other parameters in more specialized layers, such as, for instance, dMeans or dBnBias refer to gradients of the loss function wrt those parameters.

2.4. Tensor Descriptor

The cuDNN Library describes data holding images, videos and any other data with contents with a generic n-D tensor defined with the following parameters :

• a dimension nbDims from 3 to 8
• a data type (32-bit floating point, 64 bit-floating point, 16 bit floating point...)
• dimA integer array defining the size of each dimension
• strideA integer array defining the stride of each dimension (e.g the number of elements to add to reach the next element from the same dimension)

The first dimension of the tensor defines the batch size n, and the second dimension defines the number of features maps c. This tensor definition allows for example to have some dimensions overlapping each others within the same tensor by having the stride of one dimension smaller than the product of the dimension and the stride of the next dimension. In cuDNN, unless specified otherwise, all routines will support tensors with overlapping dimensions for forward pass input tensors, however, dimensions of the output tensors cannot overlap. Even though this tensor format supports negative strides (which can be useful for data mirroring), cuDNN routines do not support tensors with negative strides unless specified otherwise.

2.4.1. WXYZ Tensor Descriptor

Tensor descriptor formats are identified using acronyms, with each letter referencing a corresponding dimension. In this document, the usage of this terminology implies :

• all the strides are strictly positive
• the dimensions referenced by the letters are sorted in decreasing order of their respective strides

2.4.2. 4-D Tensor Descriptor

A 4-D Tensor descriptor is used to define the format for batches of 2D images with 4 letters : N,C,H,W for respectively the batch size, the number of feature maps, the height and the width. The letters are sorted in decreasing order of the strides. The commonly used 4-D tensor formats are :

• NCHW
• NHWC
• CHWN

2.4.3. 5-D Tensor Description

A 5-D Tensor descriptor is used to define the format of batch of 3D images with 5 letters : N,C,D,H,W for respectively the batch size, the number of feature maps, the depth, the height and the width. The letters are sorted in descreasing order of the strides. The commonly used 5-D tensor formats are called :

• NCDHW
• NDHWC
• CDHWN

2.4.4. Fully-packed tensors

A tensor is defined as XYZ-fully-packed if and only if :

• the number of tensor dimensions is equal to the number of letters preceding the fully-packed suffix.
• the stride of the i-th dimension is equal to the product of the (i+1)-th dimension by the (i+1)-th stride.
• the stride of the last dimension is 1.

2.4.5. Partially-packed tensors

The partially 'XYZ-packed' terminology only applies in a context of a tensor format described with a superset of the letters used to define a partially-packed tensor. A WXYZ tensor is defined as XYZ-packed if and only if :

• the strides of all dimensions NOT referenced in the -packed suffix are greater or equal to the product of the next dimension by the next stride.
• the stride of each dimension referenced in the -packed suffix in position i is equal to the product of the (i+1)-st dimension by the (i+1)-st stride.
• if last tensor's dimension is present in the -packed suffix, its stride is 1.

For example a NHWC tensor WC-packed means that the c_stride is equal to 1 and w_stride is equal to c_dim x c_stride. In practice, the -packed suffix is usually with slowest changing dimensions of a tensor but it is also possible to refer to a NCHW tensor that is only N-packed.

2.4.6. Spatially packed tensors

Spatially-packed tensors are defined as partially-packed in spatial dimensions.

For example a spatially-packed 4D tensor would mean that the tensor is either NCHW HW-packed or CNHW HW-packed.

2.4.7. Overlapping tensors

A tensor is defined to be overlapping if a iterating over a full range of dimensions produces the same address more than once.

In practice an overlapped tensor will have stride[i-1] < stride[i]*dim[i] for some of the i from [1,nbDims] interval.

2.5. Data Layout Formats

This section describes how cuDNN Tensors are arranged in memory. See cudnnTensorFormat_t for enumerated Tensor format types.

2.5.1. Example

Consider a batch of images in 4D with the following dimensions:

• N, the batch size, is 1
• C, the number of feature maps (i.e., number of channels), is 64
• H, the image height, is 5, and
• W, the image width, is 4

To keep the example simple, the image pixel elements are expressed as a sequence of integers, 0, 1, 2, 3, and so on. See Figure 1.

Figure 1. Example with N=1, C=64, H=5, W=4.

2.5.2. NCHW Memory Layout

The above 4D Tensor is laid out in the memory in the NCHW format as below:

1. Beginning with the first channel (c=0), the elements are arranged contiguously in row-major order.
2. Continue with second and subsequent channels until the elements of all the channels are laid out.

   See Figure 2.

3. Proceed to the next batch (if N is > 1).

Figure 2. NCHW Memory Layout

2.5.3. NHWC Memory Layout

For the NHWC memory layout, the corresponding elements in all the C channels are laid out first, as below:

1. Begin with the first element of channel 0, then proceed to the first element of channel 1, and so on, until the first elements of all the C channels are laid out.
2. Next, select the second element of channel 0, then proceed to the second element of channel 1, and so on, until the second element of all the channels are laid out.
3. Follow the row-major order in channel 0 and complete all the elements. See Figure 3.
4. Proceed to the next batch (if N is > 1).

Figure 3. NHWC Memory Layout

2.5.4. NC/32HW32 Memory Layout

The NC/32HW32 is similar to NHWC, with a key difference. For the NC/32HW32 memory layout, the 64 channels are grouped into two groups of 32 channels each—first group consisting of channels c0 through c31, and the second group consisting of channels c32 through c63. Then each group is laid out using the NHWC format. See Figure 4.

Figure 4. NC/32HW32 Memory Layout

For the generalized NC/xHWx layout format, the following observations apply:

• Only the channel dimension, C, is grouped into x channels each.

• When x = 1, each group has only one channel. Hence, the elements of one channel (i.e, one group) are arranged contiguously (in the row-major order), before proceeding to the next group (i.e., next channel). This is the same as NCHW format.

• When x = C, then NC/xHWx is identical to NHWC, i.e., the entire channel depth C is considered as a single group. The case x = C can be thought of as vectorizing entire C dimension as one big vector, laying out all the Cs, followed by the remaining dimensions, just like NHWC.

• The tensor format CUDNN_TENSOR_NCHW_VECT_C can also be interpreted in the following way: The NCHW INT8x32 format is really N x (C/32) x H x W x 32 (32 Cs for every W), just as the NCHW INT8x4 format is N x (C/4) x H x W x 4 (4 Cs for every W). Hence the "VECT_C" name - each W is a vector (4 or 32) of Cs.

2.6. Thread Safety

The library is thread safe and its functions can be called from multiple host threads, as long as threads to do not share the same cuDNN handle simultaneously.

2.7. Reproducibility (determinism)

By design, most of cuDNN's routines from a given version generate the same bit-wise results across runs when executed on GPUs with the same architecture and the same number of SMs. However, bit-wise reproducibility is not guaranteed across versions, as the implementation of a given routine may change. With the current release, the following routines do not guarantee reproducibility because they use atomic operations:

• cudnnConvolutionBackwardFilter when CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0 or CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3 is used
• cudnnConvolutionBackwardData when CUDNN_CONVOLUTION_BWD_DATA_ALGO_0 is used
• cudnnPoolingBackward when CUDNN_POOLING_MAX is used
• cudnnSpatialTfSamplerBackward

2.8. Scaling Parameters

Many cuDNN routines like cudnnConvolutionForward accept pointers in host memory to scaling factors alpha and beta. These scaling factors are used to blend the computed values with the prior values in the destination tensor as follows (see Figure 5):

dstValue = alpha*computedValue + beta*priorDstValue.

Figure 5. Scaling Parameters for Convolution

When beta is zero, the output is not read and may contain uninitialized data (including NaN).

These parameters are passed using a host memory pointer. The storage data types for alpha and beta are:

• float for HALF and FLOAT tensors, and

• double for DOUBLE tensors.

When the data input x, the filter input w and the output y are all in INT8 data type, the function cudnnConvolutionBiasActivationForward() will perform the type conversion as shown in Figure 6:

Figure 6. INT8 for cudnnConvolutionBiasActivationForward

2.9. Tensor Core Operations

The cuDNN v7 library introduced the acceleration of compute-intensive routines using Tensor Core hardware on supported GPU SM versions. Tensor core operations are supported on the Volta and Turing GPU families.

2.9.1. Basics

Tensor core operations perform parallel floating point accumulation of multiple floating point product terms. Setting the math mode to CUDNN_TENSOR_OP_MATH via the cudnnMathType_t enumerator indicates that the library will use Tensor Core operations. This enumerator specifies the available options to enable the Tensor Core, and should be applied on a per-routine basis.

The default math mode is CUDNN_DEFAULT_MATH, which indicates that the Tensor Core operations will be avoided by the library. Because the CUDNN_TENSOR_OP_MATH mode uses the Tensor Cores, it is possible that these two modes generate slightly different numerical results due to different sequencing of the floating point operations.

For example, the result of multiplying two matrices using Tensor Core operations is very close to, but not always identical, the result achieved using a sequence of scalar floating point operations. For this reason, the cuDNN library requires an explicit user opt-in before enabling the use of Tensor Core operations.

However, experiments with training common deep learning models show negligible differences between using Tensor Core operations and scalar floating point paths, as measured by both the final network accuracy and the iteration count to convergence. Consequently, the cuDNN library treats both modes of operation as functionally indistinguishable, and allows for the scalar paths to serve as legitimate fallbacks for cases in which the use of Tensor Core operations is unsuitable.

Kernels using Tensor Core operations are available for both convolutions and RNNs.

See also Training with Mixed Precision.

2.9.2. Convolution Functions

2.9.2.1. Prerequisite

For the supported GPUs, the Tensor Core operations will be triggered for convolution functions only when cudnnSetConvolutionMathType is called on the appropriate convolution descriptor by setting the mathType to CUDNN_TENSOR_OP_MATH or CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION.

2.9.2.2. Supported Algorithms

When the prerequisite is met, the below convolution functions can be run as Tensor Core operations:

• cudnnConvolutionForward

• cudnnConvolutionBackwardData

• cudnnConvolutionBackwardFilter

See the table below for supported algorithms:

2.9.2.3. Data and Filter Formats

The cuDNN library may use padding, folding, and NCHW-to-NHWC transformations to call the Tensor Core operations. See Tensor Transformations.

For algorithms other than *_ALGO_WINOGRAD_NONFUSED, when the following requirements are met, the cuDNN library will trigger the Tensor Core operations:

• Input, filter, and output descriptors (xDesc, yDesc, wDesc, dxDesc, dyDesc and dwDesc as applicable) are of the dataType = CUDNN_DATA_HALF (i.e., FP16). For FP32 dataType see FP32-to-FP16 Conversion.

• The number of input and output feature maps (i.e., channel dimension C) is a multiple of 8. When the channel dimension is not a multiple of 8, see Padding.

• The filter is of type CUDNN_TENSOR_NCHW or CUDNN_TENSOR_NHWC.

• If using a filter of type CUDNN_TENSOR_NHWC, then: the input, filter, and output data pointers (X, Y, W, dX, dY, and dW as applicable) are aligned to 128-bit boundaries.

2.9.3. RNN Functions

2.9.3.1. Prerequisite

Tensor core operations will be triggered for these RNN functions only when cudnnSetRNNMatrixMathType is called on the appropriate RNN descriptor setting mathType to CUDNN_TENSOR_OP_MATH or CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION.

2.9.3.2. Supported Algorithms

When the above prerequisite is met, the RNN functions below can be run as Tensor Core operations:

• cudnnRNNForwardInference
• cudnnRNNForwardTraining
• cudnnRNNBackwardData
• cudnnRNNBackwardWeights
• cudnnRNNForwardInferenceEx
• cudnnRNNForwardTrainingEx
• cudnnRNNBackwardDataEx
• cudnnRNNBackwardWeightsEx

See the table below for the supported algorithms:

2.9.3.3. Data and Filter Formats

When the following requirements are met, then the cuDNN library will trigger the Tensor Core operations:

• For algo = CUDNN_RNN_ALGO_STANDARD:
  • The hidden state size, input size and the batch size is a multiple of 8.
  • All user-provided tensors, workspace, and reserve space are aligned to 128 bit boundaries.
  • For FP16 input/output, the CUDNN_TENSOR_OP_MATH or CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION is selected.
  • For FP32 input/output, CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION is selected.
• For algo = CUDNN_RNN_ALGO_PERSIST_STATIC:
  • The hidden state size and the input size is a multiple of 32.
  • The batch size is a multiple of 8.
  • If the batch size exceeds 96 (for forward training or inference) or 32 (for backward data), then the batch sizes constraints may be stricter, and large power-of-two batch sizes may be needed. (new for 7.1).
  • All user-provided tensors, workspace, and reserve space are aligned to 128 bit boundaries.
  • For FP16 input/output, CUDNN_TENSOR_OP_MATH or CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION is selected.
  • For FP32 input/output, CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION is selected.

See also Features of RNN Functions.

2.9.4. Tensor Transformations

A few functions in the cuDNN library will perform transformations such as folding, padding, and NCHW-to-NHWC conversion while performing the actual function operation. See below.

2.9.4.1. FP16 Data

Tensor Cores operate on FP16 input data with FP32 accumulation. The FP16 multiply leads to a full-precision result that is accumulated in FP32 operations with the other products in a given dot product for a matrix with m x n x k dimensions. See Figure 7.

Figure 7. Tensor Operation with FP16 Inputs

2.9.4.2. FP32-to-FP16 Conversion

The cuDNN API for allows the user to specify that FP32 input data may be copied and converted to FP16 data internally to use Tensor Core Operations for potentially improved performance. This can be achieved by selecting CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION enum for cudnnMathType_t. In this mode, the FP32 Tensors are internally down-converted to FP16, the Tensor Op math is performed, and finally up-converted to FP32 as outputs. See Figure 8.

Figure 8. Tensor Operation with FP32 Inputs

For Convolutions:

For convolutions, the FP32-to-FP16 conversion can be achieved by passing the CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION enum value to the cudnnSetConvolutionMathType() call. See the below code snippet:

            

            
// Set the math type to allow cuDNN to use Tensor Cores:
checkCudnnErr(cudnnSetConvolutionMathType(cudnnConvDesc, CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION));
        

For RNNs:

For RNNs, the FP32-to-FP16 conversion can be achieved by passing the CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION enum value to the cudnnSetRNNMatrixMathType() call to allow FP32 data to be converted for use in RNNs. See the below code snippet example:

            

            
// Set the math type to allow cuDNN to use Tensor Cores:
checkCudnnErr(cudnnSetRNNMatrixMathType(cudnnRnnDesc, CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION));
        

2.9.4.3. Padding

For packed NCHW data, when the channel dimension is not a multiple of 8, then the cuDNN library will pad the tensors as needed to enable Tensor Core operations. This padding is automatic for packed NCHW data in both the CUDNN_TENSOR_OP_MATH and the CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION cases.

The padding occurs with a negligible loss of performance. Hence, the NCHW Tensor dimensions such as below are allowed:

            

            
// Set NCHW Tensor dimensions, not necessarily as multiples of eight (only the input tensor is shown here):
int dimA[] = {1, 7, 32, 32};
int strideA[] = {7168, 1024, 32, 1};
        

2.9.4.4. Folding

In the folding operation the cuDNN library implicitly performs the formatting of input tensors and saves the input tensors in an internal workspace. This can lead to an acceleration of the call to Tensor Cores.

Folding enables the input Tensors to be transformed to a format that the Tensor Cores support (i.e., no strides).

2.9.4.5. Conversion Between NCHW and NHWC

Tensor Cores require that the Tensors be in NHWC data layout. Conversion between NCHW and NHWC is performed when the user requests Tensor Op math. However, as stated in Basics, a request to use Tensor Cores is just that, a request, and Tensor Cores may not be used in some cases. The cuDNN library converts between NCHW and NHWC if and only if Tensor Cores are requested and are actually used.

If your input (and output) are NCHW, then expect a layout change. See also for packed NCHW data.

Non-Tensor Op convolutions will not perform conversions between NCHW and NHWC.

In very rare, and difficult-to-qualify, cases that are a complex function of padding and filter sizes, it is possible that Tensor Ops are not enabled. In such cases, users should pre-pad.

2.9.5. Guidelines for a Deep Learning Compiler

For a deep learning compiler, the following are the key guidelines:

• Make sure that the convolution operation is eligible for Tensor Cores by avoiding any combinations of large padding and large filters.

• Transform the inputs and filters to NHWC, pre-pad channel and batch size to be a multiple of 8.

• Make sure that all user-provided tensors, workspace and reserve space are aligned to 128 bit boundaries.

2.10. GPU and driver requirements

cuDNN v7.0 supports NVIDIA GPUs of compute capability 3.0 and higher. For x86_64 platform, cuDNN v7.0 comes with two deliverables: one requires a NVIDIA Driver compatible with CUDA Toolkit 8.0, the other requires a NVIDIA Driver compatible with CUDA Toolkit 9.0.

If you are using cuDNN with a Volta GPU, version 7 or later is required.

2.11. Backward compatibility and deprecation policy

When changing the API of an existing cuDNN function "foo" (usually to support some new functionality), first, a new routine "foo_v<n>" is created where n represents the cuDNN version where the new API is first introduced, leaving "foo" untouched. This ensures backward compatibility with the version n-1 of cuDNN. At this point, "foo" is considered deprecated, and should be treated as such by users of cuDNN. We gradually eliminate deprecated and suffixed API entries over the course of a few releases of the library per the following policy:

• In release n+1, the legacy API entry "foo" is remapped to a new API "foo_v<f>" where f is some cuDNN version anterior to n.
• Also in release n+1, the unsuffixed API entry "foo" is modified to have the same signature as "foo_<n>". "foo_<n>" is retained as-is.
• The deprecated former API entry with an anterior suffix _v<f> and new API entry with suffix _v<n> are maintained in this release.
• In release n+2, both suffixed entries of a given entry are removed.

As a rule of thumb, when a routine appears in two forms, one with a suffix and one with no suffix, the non-suffixed entry is to be treated as deprecated. In this case, it is strongly advised that users migrate to the new suffixed API entry to guarantee backwards compatibility in the following cuDNN release. When a routine appears with multiple suffixes, the unsuffixed API entry is mapped to the higher numbered suffix. In that case it is strongly advised to use the non-suffixed API entry to guarantee backward compatibiliy with the following cuDNN release.

2.12. Grouped Convolutions

cuDNN supports grouped convolutions by setting groupCount > 1 for the convolution descriptor convDesc, using cudnnSetConvolutionGroupCount().

Basic Idea

Conceptually, in grouped convolutions the input channels and the filter channels are split into groupCount number of independent groups, with each group having a reduced number of channels. Convolution operation is then performed separately on these input and filter groups.

For example, consider the following: if the number of input channels is 4, and the number of filter channels of 12. For a normal, ungrouped convolution, the number of computation operations performed are 12*4.

If the groupCount is set to 2, then there are now two input channel groups of two input channels each, and two filter channel groups of six filter channels each.

As a result, each grouped convolution will now perform 2*6 computation operations, and two such grouped convolutions are performed. Hence the computation savings are 2x: (12*4)/(2*(2*6))

cuDNN Grouped Convolution

• When using groupCount for grouped convolutions, you must still define all tensor descriptors so that they describe the size of the entire convolution, instead of specifying the sizes per group.
• Grouped convolutions are supported for all formats that are currently supported by the functions cuDNNConvolutionForward(), cudnnConvolutionBackwardData() and cudnnConvolutionBackwardFilter().
• The tensor stridings that are set for groupCount of 1 are also valid for any group count.
• By default the convolution descriptor convDesc is set to groupCount of 1.

Example

Below is an example showing the dimensions and strides for grouped convolutions for NCHW format, for 2D convolution.

xDesc or dxDesc:

• Dimensions: [batch_size, input_channel, x_height, x_width]
• Strides: [input_channels*x_height*x_width, x_height*x_width, x_width, 1]

wDesc or dwDesc:

• Dimensions: [output_channels, input_channels/groupCount, w_height, w_width]
• Format: NCHW

convDesc:

• Group Count: groupCount

yDesc or dyDesc:

• Dimensions: [batch_size, output_channels, y_height, y_width]
• Strides: [output_channels*y_height*y_width, y_height*y_width, y_width, 1]

2.13. API Logging

cuDNN API logging is a tool that records all input parameters passed into every cuDNN API function call. This functionality is disabled by default, and can be enabled through methods described in this section.

The log output contains variable names, data types, parameter values, device pointers, process ID, thread ID, cuDNN handle, cuda stream ID, and metadata such as time of the function call in microseconds.

When logging is enabled, the log output will be handled by the built-in default callback function. The user may also write their own callback function, and use the cudnnSetCallback to pass in the function pointer of their own callback function. The following is a sample output of the API log.

            

            
Function cudnnSetActivationDescriptor() called:
mode: type=cudnnActivationMode_t; val=CUDNN_ACTIVATION_RELU (1);
reluNanOpt: type=cudnnNanPropagation_t; val=CUDNN_NOT_PROPAGATE_NAN (0);
coef: type=double; val=1000.000000;
Time: 2017-11-21T14:14:21.366171 (0d+0h+1m+5s since start)
Process: 21264, Thread: 21264, cudnn_handle: NULL, cudnn_stream: NULL.
        

There are two methods to enable API logging.

Method 1: Using Environment Variables

To enable API logging using environment variables, follow these steps:

• Set the environment variable CUDNN_LOGINFO_DBG to “1”, and
• Set the environment varialbe CUDNN_LOGDEST_DBG to one of the following:
  • stdout, stderr, or a user-desired file path, for example, /home/userName1/log.txt.
• Include the conversion specifiers in the file name. For example:
  • To include date and time in the file name, use the date and time conversion specificers: log_%Y_%m_%d_%H_%M_%S.txt. The conversion specifiers will be automatically replaced with the date and time when the program is initiated, resulting in log_2017_11_21_09_41_00.txt.
  • To include the process id in the file name, use the %i conversion specifier: log_%Y_%m_%d_%H_%M_%S_%i.txt for the result: log_2017_11_21_09_41_00_21264.txt when the process id is 21264. When you have several processes running, using the process id conversion specifier will prevent these processes writing to the same file at the same time.
  Note:

  The supported conversion specifiers are similar to the strftime function.

  
  

  If the file already exists, the log will overwrite the existing file.

  Note:

  These environmental variables are only checked once at the initialization. Any subsequent changes in these environmental variables will not be effective in the current run. Also note that these environment settings can be overridden by the Method 2 below.

See also Table 1 for the impact on performance of API logging using environment variables.

Method 2

Method 2: To use API function calls to enable API logging, refer to the API description of cudnnSetCallback() and cudnnGetCallback().

2.14. Features of RNN Functions

The RNN functions are:

• cudnnRNNForwardInference
• cudnnRNNForwardTraining
• cudnnRNNBackwardData
• cudnnRNNBackwardWeights
• cudnnRNNForwardInferenceEx
• cudnnRNNForwardTrainingEx
• cudnnRNNBackwardDataEx
• cudnnRNNBackwardWeightsEx

See the table below for a list of features supported by each RNN function:

* Do not mix different algos for different steps of training. It’s also not recommended to mix non-extended and extended API for different steps of training.

** To use unpacked layout, user need to set CUDNN_RNN_PADDED_IO_ENABLED through cudnnSetRNNPaddingMode.

The following table provides the features supported by the algorithms referred in the above table: CUDNN_RNN_ALGO_STANDARD, CUDNN_RNN_ALGO_PERSIST_STATIC, and CUDNN_RNN_ALGO_PERSIST_DYNAMIC.

!CUDNN_TENSOR_OP_MATH or CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION can be set through cudnnSetRNNMatrixMathType.

2.15. Mixed Precision Numerical Accuracy

When the computation precision and the output precision are not the same, it is possible that the numerical accuracy will vary from one algorithm to the other.

For example, when the computation is performed in FP32 and the output is in FP16, the CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0 ("ALGO_0") has lower accuracy compared to the CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1 ("ALGO_1"). This is because ALGO_0 does not use extra workspace, and is forced to accumulate the intermediate results in FP16, i.e., half precision float, and this reduces the accuracy. The ALGO_1, on the other hand, uses additonal workspace to accumulate the intermediate values in FP32, i.e., full precision float.

This chapter describes all the types and enums of the cuDNN library API.

3.1. cudnnActivationDescriptor_t

cudnnActivationDescriptor_t is a pointer to an opaque structure holding the description of an activation operation. cudnnCreateActivationDescriptor is used to create one instance, and cudnnSetActivationDescriptor must be used to initialize this instance.

3.2. cudnnActivationMode_t

cudnnActivationMode_t is an enumerated type used to select the neuron activation function used in cudnnActivationForward(), cudnnActivationBackward() and cudnnConvolutionBiasActivationForward().

Values

CUDNN_ACTIVATION_SIGMOID

Selects the sigmoid function.

CUDNN_ACTIVATION_RELU

Selects the rectified linear function.

CUDNN_ACTIVATION_TANH

Selects the hyperbolic tangent function.

CUDNN_ACTIVATION_CLIPPED_RELU

Selects the clipped rectified linear function.

CUDNN_ACTIVATION_ELU

Selects the exponential linear function.

CUDNN_ACTIVATION_IDENTITY

Selects the identity function, intended for bypassing the activation step in cudnnConvolutionBiasActivationForward(). (The cudnnConvolutionBiasActivationForward() function must use CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM.) Does not work with cudnnActivationForward() or cudnnActivationBackward().

3.3. cudnnAttnDescriptor_t

cudnnAttnDescriptor_t is a pointer to an opaque structure holding parameters of the multi-head attention layer such as:

• weight and bias tensor shapes (vector lengths before and after linear projections)
• parameters that can be set in advance and do not change when invoking functions to evaluate forward responses and gradients (number of attention heads, softmax smoothing/sharpening coefficient)
• other settings that are necessary to compute temporary buffer sizes.

Use the cudnnCreateAttnDescriptor function to create an instance of the attention descriptor object and cudnnDestroyAttnDescriptor to delete the previously created descriptor. Use the cudnnSetAttnDescriptor function to configure the descriptor.

3.4. cudnnBatchNormMode_t

cudnnBatchNormMode_t is an enumerated type used to specify the mode of operation in cudnnBatchNormalizationForwardInference, cudnnBatchNormalizationForwardTraining, cudnnBatchNormalizationBackward and cudnnDeriveBNTensorDescriptor routines.

Values

CUDNN_BATCHNORM_PER_ACTIVATION

Normalization is performed per-activation. This mode is intended to be used after the non-convolutional network layers. In this mode, the tensor dimensions of bnBias and bnScale and the parameters used in the cudnnBatchNormalization* functions, are 1xCxHxW.

CUDNN_BATCHNORM_SPATIAL

Normalization is performed over N+spatial dimensions. This mode is intended for use after convolutional layers (where spatial invariance is desired). In this mode the bnBias and bnScale tensor dimensions are 1xCx1x1.

CUDNN_BATCHNORM_SPATIAL_PERSISTENT

This mode is similar to CUDNN_BATCHNORM_SPATIAL but it can be faster for some tasks.

An optimized path may be selected for CUDNN_DATA_FLOAT and CUDNN_DATA_HALF types, compute capability 6.0 or higher for the following two batch normalization API calls: cudnnBatchNormalizationForwardTraining, and cudnnBatchNormalizationBackward. In the case of cudnnBatchNormalizationBackward, the savedMean and savedInvVariance arguments should not be NULL.

The rest of this section applies to NCHW mode only:

This mode may use a scaled atomic integer reduction that is deterministic but imposes more restrictions on the input data range. When a numerical overflow occurs, the algorithm may produce NaN-s or Inf-s (infinity) in output buffers.

When Inf-s/NaN-s are present in the input data, the output in this mode is the same as from a pure floating-point implementation.

For finite but very large input values, the algorithm may encounter overflows more frequently due to a lower dynamic range and emit Inf-s/NaN-s while CUDNN_BATCHNORM_SPATIAL will produce finite results. The user can invoke cudnnQueryRuntimeError to check if a numerical overflow occurred in this mode.

3.5. cudnnBatchNormOps_t

cudnnBatchNormOps_t is an enumerated type used to specify the mode of operation in cudnnGetBatchNormalizationForwardTrainingExWorkspaceSize(), cudnnBatchNormalizationForwardTrainingEx(), cudnnGetBatchNormalizationBackwardExWorkspaceSize(), cudnnBatchNormalizationBackwardEx(), and cudnnGetBatchNormalizationTrainingExReserveSpaceSize() functions.

Values

CUDNN_BATCHNORM_OPS_BN

Only batch normalization is performed, per-activation.

CUDNN_BATCHNORM_OPS_BN_ACTIVATION

First, the batch normalization is performed, and then the activation is performed.

CUDNN_BATCHNORM_OPS_BN_ADD_ACTIVATION

Performs the batch normalization, then element-wise addition, followed by the activation operation.

3.6. cudnnConvolutionBwdDataAlgo_t

cudnnConvolutionBwdDataAlgo_t is an enumerated type that exposes the different algorithms available to execute the backward data convolution operation.

Values

CUDNN_CONVOLUTION_BWD_DATA_ALGO_0

This algorithm expresses the convolution as a sum of matrix product without actually explicitly form the matrix that holds the input tensor data. The sum is done using atomic adds operation, thus the results are non-deterministic.

CUDNN_CONVOLUTION_BWD_DATA_ALGO_1

This algorithm expresses the convolution as a matrix product without actually explicitly form the matrix that holds the input tensor data. The results are deterministic.

CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT

This algorithm uses a Fast-Fourier Transform approach to compute the convolution. A significant memory workspace is needed to store intermediate results. The results are deterministic.

CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING

This algorithm uses the Fast-Fourier Transform approach but splits the inputs into tiles. A significant memory workspace is needed to store intermediate results but less than CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT for large size images. The results are deterministic.

CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD

This algorithm uses the Winograd Transform approach to compute the convolution. A reasonably sized workspace is needed to store intermediate results. The results are deterministic.

CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED

This algorithm uses the Winograd Transform approach to compute the convolution. A significant workspace may be needed to store intermediate results. The results are deterministic.

3.7. cudnnConvolutionBwdDataAlgoPerf_t

cudnnConvolutionBwdDataAlgoPerf_t is a structure containing performance results returned by cudnnFindConvolutionBackwardDataAlgorithm() or heuristic results returned by cudnnGetConvolutionBackwardDataAlgorithm_v7().

Data Members

cudnnConvolutionBwdDataAlgo_t algo

The algorithm runs to obtain the associated performance metrics.

cudnnStatus_t status

If any error occurs during the workspace allocation or timing of cudnnConvolutionBackwardData(), this status will represent that error. Otherwise, this status will be the return status of cudnnConvolutionBackwardData().

• CUDNN_STATUS_ALLOC_FAILED if any error occurred during workspace allocation or if the provided workspace is insufficient.
• CUDNN_STATUS_INTERNAL_ERROR if any error occurred during timing calculations or workspace deallocation.
• Otherwise, this will be the return status of cudnnConvolutionBackwardData().

float time

The execution time of cudnnConvolutionBackwardData() (in milliseconds).

size_t memory

The workspace size (in bytes).

cudnnDeterminism_t determinism

The determinism of the algorithm.

cudnnMathType_t mathType

The math type provided to the algorithm.

int reserved[3]

Reserved space for future properties.

3.8. cudnnConvolutionBwdDataPreference_t

cudnnConvolutionBwdDataPreference_t is an enumerated type used by cudnnGetConvolutionBackwardDataAlgorithm() to help the choice of the algorithm used for the backward data convolution.

Values

CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE

In this configuration, the routine cudnnGetConvolutionBackwardDataAlgorithm() is guaranteed to return an algorithm that does not require any extra workspace to be provided by the user.

CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST

In this configuration, the routine cudnnGetConvolutionBackwardDataAlgorithm() will return the fastest algorithm regardless of how much workspace is needed to execute it.

CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT

In this configuration, the routine cudnnGetConvolutionBackwardDataAlgorithm() will return the fastest algorithm that fits within the memory limit that the user provided.

3.9. cudnnConvolutionBwdFilterAlgo_t

cudnnConvolutionBwdFilterAlgo_t is an enumerated type that exposes the different algorithms available to execute the backward filter convolution operation.

Values

CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0

This algorithm expresses the convolution as a sum of matrix product without actually explicitly form the matrix that holds the input tensor data. The sum is done using atomic adds operation, thus the results are non-deterministic.

CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1

This algorithm expresses the convolution as a matrix product without actually explicitly form the matrix that holds the input tensor data. The results are deterministic.

CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT

This algorithm uses the Fast-Fourier Transform approach to compute the convolution. A significant workspace is needed to store intermediate results. The results are deterministic.

CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3

This algorithm is similar to CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0 but uses some small workspace to precomputes some indices. The results are also non-deterministic.

CUDNN_CONVOLUTION_BWD_FILTER_WINOGRAD_NONFUSED

This algorithm uses the Winograd Transform approach to compute the convolution. A significant workspace may be needed to store intermediate results. The results are deterministic.

CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT_TILING

This algorithm uses the Fast-Fourier Transform approach to compute the convolution but splits the input tensor into tiles. A significant workspace may be needed to store intermediate results. The results are deterministic.

3.10. cudnnConvolutionBwdFilterAlgoPerf_t

cudnnConvolutionBwdFilterAlgoPerf_t is a structure containing performance results returned by cudnnFindConvolutionBackwardFilterAlgorithm() or heuristic results returned by cudnnGetConvolutionBackwardFilterAlgorithm_v7().

Data Members

cudnnConvolutionBwdFilterAlgo_t algo

The algorithm runs to obtain the associated performance metrics.

cudnnStatus_t status

If any error occurs during the workspace allocation or timing of cudnnConvolutionBackwardFilter(), this status will represent that error. Otherwise, this status will be the return status of cudnnConvolutionBackwardFilter().

• CUDNN_STATUS_ALLOC_FAILED if any error occurred during workspace allocation or if the provided workspace is insufficient.
• CUDNN_STATUS_INTERNAL_ERROR if any error occurred during timing calculations or workspace deallocation.
• Otherwise, this will be the return status of cudnnConvolutionBackwardFilter().

float time

The execution time of cudnnConvolutionBackwardFilter() (in milliseconds).

size_t memory

The workspace size (in bytes).

cudnnDeterminism_t determinism

The determinism of the algorithm.

cudnnMathType_t mathType

The math type provided to the algorithm.

int reserved[3]

Reserved space for future properties.

3.11. cudnnConvolutionBwdFilterPreference_t

cudnnConvolutionBwdFilterPreference_t is an enumerated type used by cudnnGetConvolutionBackwardFilterAlgorithm() to help the choice of the algorithm used for the backward filter convolution.

Values

CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE

In this configuration, the routine cudnnGetConvolutionBackwardFilterAlgorithm() is guaranteed to return an algorithm that does not require any extra workspace to be provided by the user.

CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST

In this configuration, the routine cudnnGetConvolutionBackwardFilterAlgorithm() will return the fastest algorithm regardless of how much workspace is needed to execute it.

CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT

In this configuration, the routine cudnnGetConvolutionBackwardFilterAlgorithm() will return the fastest algorithm that fits within the memory limit that the user provided.

3.12. cudnnConvolutionDescriptor_t

cudnnConvolutionDescriptor_t is a pointer to an opaque structure holding the description of a convolution operation. cudnnCreateConvolutionDescriptor() is used to create one instance, and cudnnSetConvolutionNdDescriptor() or cudnnSetConvolution2dDescriptor() must be used to initialize this instance.

3.13. cudnnConvolutionFwdAlgo_t

cudnnConvolutionFwdAlgo_t is an enumerated type that exposes the different algorithms available to execute the forward convolution operation.

Values

CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM

This algorithm expresses the convolution as a matrix product without actually explicitly form the matrix that holds the input tensor data.

CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM

This algorithm expresses convolution as a matrix product without actually explicitly form the matrix that holds the input tensor data, but still needs some memory workspace to precompute some indices in order to facilitate the implicit construction of the matrix that holds the input tensor data.

CUDNN_CONVOLUTION_FWD_ALGO_GEMM

This algorithm expresses the convolution as an explicit matrix product. A significant memory workspace is needed to store the matrix that holds the input tensor data.

CUDNN_CONVOLUTION_FWD_ALGO_DIRECT

This algorithm expresses the convolution as a direct convolution (for example, without implicitly or explicitly doing a matrix multiplication).

CUDNN_CONVOLUTION_FWD_ALGO_FFT

This algorithm uses the Fast-Fourier Transform approach to compute the convolution. A significant memory workspace is needed to store intermediate results.

CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING

This algorithm uses the Fast-Fourier Transform approach but splits the inputs into tiles. A significant memory workspace is needed to store intermediate results but less than CUDNN_CONVOLUTION_FWD_ALGO_FFT for large size images.

CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD

This algorithm uses the Winograd Transform approach to compute the convolution. A reasonably sized workspace is needed to store intermediate results.

CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED

This algorithm uses the Winograd Transform approach to compute the convolution. A significant workspace may be needed to store intermediate results.

3.14. cudnnConvolutionFwdAlgoPerf_t

cudnnConvolutionFwdAlgoPerf_t is a structure containing performance results returned by cudnnFindConvolutionForwardAlgorithm() or heuristic results returned by cudnnGetConvolutionForwardAlgorithm_v7().

Data Members

cudnnConvolutionFwdAlgo_t algo

The algorithm runs to obtain the associated performance metrics.

cudnnStatus_t status

If any error occurs during the workspace allocation or timing of cudnnConvolutionForward(), this status will represent that error. Otherwise, this status will be the return status of cudnnConvolutionForward().

• CUDNN_STATUS_ALLOC_FAILED if any error occurred during workspace allocation or if the provided workspace is insufficient.
• CUDNN_STATUS_INTERNAL_ERROR if any error occurred during timing calculations or workspace deallocation.
• Otherwise, this will be the return status of cudnnConvolutionForward().

float time

The execution time of cudnnConvolutionForward() (in milliseconds).

size_t memory

The workspace size (in bytes).

cudnnDeterminism_t determinism

The determinism of the algorithm.

cudnnMathType_t mathType

The math type provided to the algorithm.

int reserved[3]

Reserved space for future properties.

3.15. cudnnConvolutionFwdPreference_t

cudnnConvolutionFwdPreference_t is an enumerated type used by cudnnGetConvolutionForwardAlgorithm() to help the choice of the algorithm used for the forward convolution.

Values

CUDNN_CONVOLUTION_FWD_NO_WORKSPACE

In this configuration, the routine cudnnGetConvolutionForwardAlgorithm() is guaranteed to return an algorithm that does not require any extra workspace to be provided by the user.

CUDNN_CONVOLUTION_FWD_PREFER_FASTEST

In this configuration, the routine cudnnGetConvolutionForwardAlgorithm() will return the fastest algorithm regardless of how much workspace is needed to execute it.

CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT

In this configuration, the routine cudnnGetConvolutionForwardAlgorithm() will return the fastest algorithm that fits within the memory limit that the user provided.

3.16. cudnnConvolutionMode_t

cudnnConvolutionMode_t is an enumerated type used by cudnnSetConvolutionDescriptor() to configure a convolution descriptor. The filter used for the convolution can be applied in two different ways, corresponding mathematically to a convolution or to a cross-correlation. (A cross-correlation is equivalent to a convolution with its filter rotated by 180 degrees.)

Values

CUDNN_CONVOLUTION

In this mode, a convolution operation will be done when applying the filter to the images.

CUDNN_CROSS_CORRELATION

In this mode, a cross-correlation operation will be done when applying the filter to the images.

3.17. cudnnCTCLossAlgo_t

cudnnCTCLossAlgo_t is an enumerated type that exposes the different algorithms available to execute the CTC loss operation.

Values

CUDNN_CTC_LOSS_ALGO_DETERMINISTIC

Results are guaranteed to be reproducible

CUDNN_CTC_LOSS_ALGO_NON_DETERMINISTIC

Results are not guaranteed to be reproducible

3.18. cudnnCTCLossDescriptor_t

cudnnCTCLossDescriptor_t is a pointer to an opaque structure holding the description of a CTC loss operation. cudnnCreateCTCLossDescriptor() is used to create one instance, cudnnSetCTCLossDescriptor() is used to initialize this instance, and cudnnDestroyCTCLossDescriptor() is used to destroy this instance.

3.19. cudnnDataType_t

cudnnDataType_t is an enumerated type indicating the data type to which a tensor descriptor or filter descriptor refers.

Values

CUDNN_DATA_FLOAT

The data is a 32-bit single-precision floating-point (float).

CUDNN_DATA_DOUBLE

The data is a 64-bit double-precision floating-point (double).

CUDNN_DATA_HALF

The data is a 16-bit floating-point.

CUDNN_DATA_INT8

The data is an 8-bit signed integer.

CUDNN_DATA_UINT8

The data is an 8-bit unsigned integer.

CUDNN_DATA_INT32

The data is a 32-bit signed integer.

CUDNN_DATA_INT8x4

The data is 32-bit elements each composed of 4 8-bit signed integers. This data type is only supported with tensor format CUDNN_TENSOR_NCHW_VECT_C.

CUDNN_DATA_INT8x32

The data is 32-element vectors, each element being an 8-bit signed integer. This data type is only supported with the tensor format CUDNN_TENSOR_NCHW_VECT_C. Moreover, this data type can only be used with algo 1, meaning, CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM. For more information, see cudnnConvolutionFwdAlgo_t.

CUDNN_DATA_UINT8x4

The data is 32-bit elements each composed of 4 8-bit unsigned integers. This data type is only supported with tensor format CUDNN_TENSOR_NCHW_VECT_C.

3.20. cudnnDeterminism_t

cudnnDeterminism_t is an enumerated type used to indicate if the computed results are deterministic (reproducible). For more information, see Reproducibility (determinism).

Values

CUDNN_NON_DETERMINISTIC

Results are not guaranteed to be reproducible.

CUDNN_DETERMINISTIC

Results are guaranteed to be reproducible.

3.21. cudnnDirectionMode_t

cudnnDirectionMode_t is an enumerated type used to specify the recurrence pattern in the cudnnRNNForwardInference(), cudnnRNNForwardTraining(), cudnnRNNBackwardData() and cudnnRNNBackwardWeights() routines.

Values

CUDNN_UNIDIRECTIONAL

The network iterates recurrently from the first input to the last.

CUDNN_BIDIRECTIONAL

Each layer of the network iterates recurrently from the first input to the last and separately from the last input to the first. The outputs of the two are concatenated at each iteration giving the output of the layer.

3.22. cudnnDivNormMode_t

cudnnDivNormMode_t is an enumerated type used to specify the mode of operation in cudnnDivisiveNormalizationForward() and cudnnDivisiveNormalizationBackward().

Values

CUDNN_DIVNORM_PRECOMPUTED_MEANS

The means tensor data pointer is expected to contain means or other kernel convolution values precomputed by the user. The means pointer can also be NULL, in that case, it's considered to be filled with zeroes. This is equivalent to spatial LRN.

Note:

In the backward pass, the means are treated as independent inputs and the gradient over means is computed independently. In this mode, to yield a net gradient over the entire LCN computational graph, the destDiffMeans result should be backpropagated through the user's means layer (which can be implemented using average pooling) and added to the destDiffData tensor produced by cudnnDivisiveNormalizationBackward().

3.23. cudnnDropoutDescriptor_t

cudnnDropoutDescriptor_t is a pointer to an opaque structure holding the description of a dropout operation. cudnnCreateDropoutDescriptor() is used to create one instance, cudnnSetDropoutDescriptor() is used to initialize this instance, cudnnDestroyDropoutDescriptor() is used to destroy this instance, cudnnGetDropoutDescriptor() is used to query fields of a previously initialized instance, cudnnRestoreDropoutDescriptor() is used to restore an instance to a previously saved off state.

3.24. cudnnErrQueryMode_t

cudnnErrQueryMode_t is an enumerated type passed to cudnnQueryRuntimeError() to select the remote kernel error query mode.

Values

CUDNN_ERRQUERY_RAWCODE

Read the error storage location regardless of the kernel completion status.

CUDNN_ERRQUERY_NONBLOCKING

Report if all tasks in the user stream of the cuDNN handle were completed. If that is the case, report the remote kernel error code.

CUDNN_ERRQUERY_BLOCKING

Wait for all tasks to complete in the user stream before reporting the remote kernel error code.

3.25. cudnnFilterDescriptor_t

cudnnFilterDescriptor_t is a pointer to an opaque structure holding the description of a filter dataset. cudnnCreateFilterDescriptor() is used to create one instance, and cudnnSetFilter4dDescriptor() or cudnnSetFilterNdDescriptor() must be used to initialize this instance.

3.26. cudnnFoldingDirection_t

cudnnFoldingDirection_t is an enumerated type used to select the folding direction. For more information, see cudnnTensorTransformDescriptor_t.

3.27. cudnnFusedOps_t

The cudnnFusedOps_t type is an enumerated type to select a specific sequence of computations to perform in the fused operations.

Member	Description
CUDNN_FUSED_SCALE_BIAS_ACTIVATION_CONV_BNSTATS = 0	On a per-channel basis, performs these operations in this order: scale, add bias, activation, convolution, and generate batchnorm statistics.
CUDNN_FUSED_SCALE_BIAS_ACTIVATION_WGRAD = 1	On a per-channel basis, performs these operations in this order: scale, add bias, activation, convolution backward weights, and generate batchnorm statistics.
CUDNN_FUSED_BN_FINALIZE_STATISTICS_TRAINING = 2	Computes the equivalent scale and bias from ySum, ySqSum and learned scale, bias. Optionally update running statistics and generate saved stats
CUDNN_FUSED_BN_FINALIZE_STATISTICS_INFERENCE = 3	Computes the equivalent scale and bias from the learned running statistics and the learned scale, bias.
CUDNN_FUSED_CONV_SCALE_BIAS_ADD_ACTIVATION = 4	On a per-channel basis, performs these operations in this order: convolution, scale, add bias, element-wise addition with another tensor, and activation.
CUDNN_FUSED_SCALE_BIAS_ADD_ACTIVATION_GEN_BITMASK = 5	On a per-channel basis, performs these operations in this order: scale and bias on one tensor, scale, and bias on a second tensor, element-wise addition of these two tensors, and on the resulting tensor perform activation, and generate activation bit mask.
CUDNN_FUSED_DACTIVATION_FORK_DBATCHNORM = 6	On a per-channel basis, performs these operations in this order: backward activation, fork (meaning, write out gradient for the residual branch), and backward batch norm.

3.28. cudnnFusedOpsConstParamLabel_t

The cudnnFusedOpsConstParamLabel_t is an enumerated type for the selection of the type of the cudnnFusedOps descriptor. For more information, see cudnnSetFusedOpsConstParamPackAttribute.

            

            
typedef enum {
	CUDNN_PARAM_XDESC                          = 0,
	CUDNN_PARAM_XDATA_PLACEHOLDER              = 1,
	CUDNN_PARAM_BN_MODE                        = 2,
	CUDNN_PARAM_BN_EQSCALEBIAS_DESC            = 3,
	CUDNN_PARAM_BN_EQSCALE_PLACEHOLDER         = 4,
	CUDNN_PARAM_BN_EQBIAS_PLACEHOLDER          = 5,
	CUDNN_PARAM_ACTIVATION_DESC                = 6,
	CUDNN_PARAM_CONV_DESC                      = 7,
	CUDNN_PARAM_WDESC                          = 8,
	CUDNN_PARAM_WDATA_PLACEHOLDER              = 9,
	CUDNN_PARAM_DWDESC                         = 10,
	CUDNN_PARAM_DWDATA_PLACEHOLDER             = 11,
	CUDNN_PARAM_YDESC                          = 12,
	CUDNN_PARAM_YDATA_PLACEHOLDER              = 13,
	CUDNN_PARAM_DYDESC                         = 14,
	CUDNN_PARAM_DYDATA_PLACEHOLDER             = 15,
	CUDNN_PARAM_YSTATS_DESC                    = 16,
	CUDNN_PARAM_YSUM_PLACEHOLDER               = 17,
	CUDNN_PARAM_YSQSUM_PLACEHOLDER             = 18,
	CUDNN_PARAM_BN_SCALEBIAS_MEANVAR_DESC      = 19,
	CUDNN_PARAM_BN_SCALE_PLACEHOLDER           = 20,
	CUDNN_PARAM_BN_BIAS_PLACEHOLDER            = 21,
	CUDNN_PARAM_BN_SAVED_MEAN_PLACEHOLDER      = 22,
	CUDNN_PARAM_BN_SAVED_INVSTD_PLACEHOLDER    = 23,
	CUDNN_PARAM_BN_RUNNING_MEAN_PLACEHOLDER    = 24,
	CUDNN_PARAM_BN_RUNNING_VAR_PLACEHOLDER     = 25,
	CUDNN_PARAM_ZDESC                          = 26,
	CUDNN_PARAM_ZDATA_PLACEHOLDER              = 27,
	CUDNN_PARAM_BN_Z_EQSCALEBIAS_DESC          = 28,
	CUDNN_PARAM_BN_Z_EQSCALE_PLACEHOLDER       = 29,
	CUDNN_PARAM_BN_Z_EQBIAS_PLACEHOLDER        = 30,
	CUDNN_PARAM_ACTIVATION_BITMASK_DESC        = 31,
	CUDNN_PARAM_ACTIVATION_BITMASK_PLACEHOLDER = 32,
	CUDNN_PARAM_DXDESC                         = 33,
	CUDNN_PARAM_DXDATA_PLACEHOLDER             = 34,
	CUDNN_PARAM_DZDESC                         = 35,
	CUDNN_PARAM_DZDATA_PLACEHOLDER             = 36,
	CUDNN_PARAM_BN_DSCALE_PLACEHOLDER          = 37,
	CUDNN_PARAM_BN_DBIAS_PLACEHOLDER           = 38,
	} cudnnFusedOpsConstParamLabel_t;
        

As of cuDNN 7.6.0, if the conditions in Table 3 are met, then the fully fused fast path will be triggered. Otherwise, a slower partially fused path will be triggered.

As of cuDNN 7.6.0, if the conditions in Table 5 are met, then the fully fused fast path will be triggered. Otherwise a slower partially fused path will be triggered.

3.29. cudnnFusedOpsConstParamPack_t

cudnnFusedOpsConstParamPack_t is a pointer to an opaque structure holding the description of the cudnnFusedOps constant parameters. Use the function cudnnCreateFusedOpsConstParamPack ​to create one instance of this structure, and the function cudnnDestroyFusedOpsConstParamPack to destroy a previously-created descriptor.

3.30. cudnnFusedOpsPlan_t

cudnnFusedOpsPlan_t is a pointer to an opaque structure holding the description of the cudnnFusedOpsPlan. This descriptor contains the plan information, including the problem type and size, which kernels should be run, and the internal workspace partition. Use the function cudnnCreateFusedOpsPlan to create one instance of this structure, and the function cudnnDestroyFusedOpsPlan to destroy a previously-created descriptor.

3.31. cudnnFusedOpsPointerPlaceHolder_t

cudnnFusedOpsPointerPlaceHolder_t is an enumerated type used to select the alignment type of the cudnnFusedOps descriptor pointer.

3.32. cudnnFusedOpsVariantParamLabel_t

The cudnnFusedOpsVariantParamLabel_t is an enumerated type that is used to set the buffer pointers. These buffer pointers can be changed in each iteration.

            

            
typedef enum {
	CUDNN_PTR_XDATA                              = 0,
	CUDNN_PTR_BN_EQSCALE                         = 1,
	CUDNN_PTR_BN_EQBIAS                          = 2,
	CUDNN_PTR_WDATA                              = 3,
	CUDNN_PTR_DWDATA                             = 4,
	CUDNN_PTR_YDATA                              = 5,
	CUDNN_PTR_DYDATA                             = 6,
	CUDNN_PTR_YSUM                               = 7,
	CUDNN_PTR_YSQSUM                             = 8,
	CUDNN_PTR_WORKSPACE                          = 9,
	CUDNN_PTR_BN_SCALE                           = 10,
	CUDNN_PTR_BN_BIAS                            = 11,
	CUDNN_PTR_BN_SAVED_MEAN                      = 12,
	CUDNN_PTR_BN_SAVED_INVSTD                    = 13,
	CUDNN_PTR_BN_RUNNING_MEAN                    = 14,
	CUDNN_PTR_BN_RUNNING_VAR                     = 15,
	CUDNN_PTR_ZDATA                              = 16,
	CUDNN_PTR_BN_Z_EQSCALE                       = 17,
	CUDNN_PTR_BN_Z_EQBIAS                        = 18,
	CUDNN_PTR_ACTIVATION_BITMASK                 = 19,
	CUDNN_PTR_DXDATA                             = 20,
	CUDNN_PTR_DZDATA                             = 21,
	CUDNN_PTR_BN_DSCALE                          = 22,
	CUDNN_PTR_BN_DBIAS                           = 23,
	CUDNN_SCALAR_SIZE_T_WORKSPACE_SIZE_IN_BYTES  = 100,
	CUDNN_SCALAR_INT64_T_BN_ACCUMULATION_COUNT   = 101,
	CUDNN_SCALAR_DOUBLE_BN_EXP_AVG_FACTOR        = 102,
	CUDNN_SCALAR_DOUBLE_BN_EPSILON               = 103,
	} cudnnFusedOpsVariantParamLabel_t;
        

3.33. cudnnFusedOpsVariantParamPack_t

cudnnFusedOpsVariantParamPack_t is a pointer to an opaque structure holding the description of the cudnnFusedOps variant parameters. Use the function cudnnCreateFusedOpsVariantParamPack to create one instance of this structure, and the function cudnnDestroyFusedOpsVariantParamPack to destroy a previously-created descriptor.

3.34. cudnnHandle_t

cudnnHandle_t is a pointer to an opaque structure holding the cuDNN library context. The cuDNN library context must be created using cudnnCreate() and the returned handle must be passed to all subsequent library function calls. The context should be destroyed at the end using cudnnDestroy(). The context is associated with only one GPU device, the current device at the time of the call to cudnnCreate(). However, multiple contexts can be created on the same GPU device.

3.35. cudnnIndicesType_t

cudnnIndicesType_t is an enumerated type used to indicate the data type for the indices to be computed by the cudnnReduceTensor() routine. This enumerated type is used as a field for the cudnnReduceTensorDescriptor_t descriptor.

Values

CUDNN_32BIT_INDICES

Compute unsigned int indices.

CUDNN_64BIT_INDICES

Compute unsigned long indices.

CUDNN_16BIT_INDICES

Compute unsigned short indices.

CUDNN_8BIT_INDICES

Compute unsigned char indices.

3.36. cudnnLossNormalizationMode_t

cudnnLossNormalizationMode_t is an enumerated type that controls the input normalization mode for a loss function. This type can be used with cudnnSetCTCLossDescriptorEx.

Values

CUDNN_LOSS_NORMALIZATION_NONE

The input probs of cudnnCTCLoss function is expected to be the normalized probability, and the output gradients is the gradient of loss with respect to the unnormalized probability.

CUDNN_LOSS_NORMALIZATION_SOFTMAX

The input probs of cudnnCTCLoss function is expected to be the unnormalized activation from the previous layer, and the output gradients is the gradient with respect to the activation. Internally the probability is computed by softmax normalization.

3.37. cudnnLRNMode_t

cudnnLRNMode_t is an enumerated type used to specify the mode of operation in cudnnLRNCrossChannelForward() and cudnnLRNCrossChannelBackward().

Values

CUDNN_LRN_CROSS_CHANNEL_DIM1

LRN computation is performed across tensor's dimension dimA[1].

3.38. cudnnMathType_t

cudnnMathType_t is an enumerated type used to indicate if the use of Tensor Core operations is permitted a given library routine.

Values

CUDNN_DEFAULT_MATH

Tensor Core operations are not used.

CUDNN_TENSOR_OP_MATH

The use of Tensor Core operations is permitted.

CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION

Enables the use of FP32 tensors for both input and output.

3.39. cudnnMultiHeadAttnWeightKind_t

cudnnMultiHeadAttnWeightKind_t is an enumerated type that specifies a group of weights or biases in the cudnnGetMultiHeadAttnWeights() function.

Values

CUDNN_MH_ATTN_Q_WEIGHTS

Selects the input projection weights for queries.

CUDNN_MH_ATTN_K_WEIGHTS

Selects the input projection weights for keys.

CUDNN_MH_ATTN_V_WEIGHTS

Selects the input projection weights for values.

CUDNN_MH_ATTN_O_WEIGHTS

Selects the output projection weights.

CUDNN_MH_ATTN_Q_BIASES

Selects the input projection biases for queries.

CUDNN_MH_ATTN_K_BIASES

Selects the input projection biases for keys.

CUDNN_MH_ATTN_V_BIASES

Selects the input projection biases for values.

CUDNN_MH_ATTN_O_BIASES

Selects the output projection biases.

3.40. cudnnNanPropagation_t

cudnnNanPropagation_t is an enumerated type used to indicate if a given routine should propagate Nan numbers. This enumerated type is used as a field for the cudnnActivationDescriptor_t descriptor and cudnnPoolingDescriptor_t descriptor.

Values

CUDNN_NOT_PROPAGATE_NAN

Nan numbers are not propagated.

CUDNN_PROPAGATE_NAN

Nan numbers are propagated.

3.41. cudnnOpTensorDescriptor_t

cudnnOpTensorDescriptor_t is a pointer to an opaque structure holding the description of a Tensor Core operation, used as a parameter to cudnnOpTensor(). cudnnCreateOpTensorDescriptor() is used to create one instance, and cudnnSetOpTensorDescriptor() must be used to initialize this instance.

3.42. cudnnOpTensorOp_t

cudnnOpTensorOp_t is an enumerated type used to indicate the Tensor Core operation to be used by the cudnnOpTensor() routine. This enumerated type is used as a field for the cudnnOpTensorDescriptor_t descriptor.

Values

CUDNN_OP_TENSOR_ADD

The operation to be performed is addition.

CUDNN_OP_TENSOR_MUL

The operation to be performed is multiplication.

CUDNN_OP_TENSOR_MIN

The operation to be performed is a minimum comparison.

CUDNN_OP_TENSOR_MAX

The operation to be performed is a maximum comparison.

CUDNN_OP_TENSOR_SQRT

The operation to be performed is square root, performed on only the A tensor.

CUDNN_OP_TENSOR_NOT

The operation to be performed is negation, performed on only the A tensor.

3.43. cudnnPersistentRNNPlan_t

cudnnPersistentRNNPlan_t is a pointer to an opaque structure holding a plan to execute a dynamic persistent RNN. cudnnCreatePersistentRNNPlan() is used to create and initialize one instance.

3.44. cudnnPoolingDescriptor_t

cudnnPoolingDescriptor_t is a pointer to an opaque structure holding the description of a pooling operation. cudnnCreatePoolingDescriptor() is used to create one instance, and cudnnSetPoolingNdDescriptor() or cudnnSetPooling2dDescriptor() must be used to initialize this instance.

3.45. cudnnPoolingMode_t

cudnnPoolingMode_t is an enumerated type passed to cudnnSetPoolingDescriptor() to select the pooling method to be used by cudnnPoolingForward() and cudnnPoolingBackward().

Values

CUDNN_POOLING_MAX

The maximum value inside the pooling window is used.

CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING

Values inside the pooling window are averaged. The number of elements used to calculate the average includes spatial locations falling in the padding region.

CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING

Values inside the pooling window are averaged. The number of elements used to calculate the average excludes spatial locations falling in the padding region.

CUDNN_POOLING_MAX_DETERMINISTIC

The maximum value inside the pooling window is used. The algorithm used is deterministic.

3.46. cudnnReduceTensorDescriptor_t

cudnnReduceTensorDescriptor_t is a pointer to an opaque structure holding the description of a tensor reduction operation, used as a parameter to cudnnReduceTensor(). cudnnCreateReduceTensorDescriptor() is used to create one instance, and cudnnSetReduceTensorDescriptor() must be used to initialize this instance.

cudnnReduceTensorIndices_t

cudnnReduceTensorIndices_t is an enumerated type used to indicate whether indices are to be computed by the cudnnReduceTensor() routine. This enumerated type is used as a field for the cudnnReduceTensorDescriptor_t descriptor.

Values

CUDNN_REDUCE_TENSOR_NO_INDICES

Do not compute indices.

CUDNN_REDUCE_TENSOR_FLATTENED_INDICES

Compute indices. The resulting indices are relative, and flattened.

3.48. cudnnReduceTensorOp_t

cudnnReduceTensorOp_t is an enumerated type used to indicate the Tensor Core operation to be used by the cudnnReduceTensor() routine. This enumerated type is used as a field for the cudnnReduceTensorDescriptor_t descriptor.

Values

CUDNN_REDUCE_TENSOR_ADD

The operation to be performed is addition.

CUDNN_REDUCE_TENSOR_MUL

The operation to be performed is multiplication.

CUDNN_REDUCE_TENSOR_MIN

The operation to be performed is a minimum comparison.

CUDNN_REDUCE_TENSOR_MAX

The operation to be performed is a maximum comparison.

CUDNN_REDUCE_TENSOR_AMAX

The operation to be performed is a maximum comparison of absolute values.

CUDNN_REDUCE_TENSOR_AVG

The operation to be performed is averaging.

CUDNN_REDUCE_TENSOR_NORM1

The operation to be performed is addition of absolute values.

CUDNN_REDUCE_TENSOR_NORM2

The operation to be performed is a square root of sum of squares.

CUDNN_REDUCE_TENSOR_MUL_NO_ZEROS

The operation to be performed is multiplication, not including elements of value zero.

3.49. cudnnReorderType_t

            

            
typedef enum {
	CUDNN_DEFAULT_REORDER = 0,
	CUDNN_NO_REORDER      = 1,
	} cudnnReorderType_t;
        

cudnnReorderType_t is an enumerated type to set the convolution reordering type. The reordering type can be set by cudnnSetConvolutionReorderType and its status can be read by cudnnGetConvolutionReorderType.

3.50. cudnnRNNAlgo_t

cudnnRNNAlgo_t is an enumerated type used to specify the algorithm used in the cudnnRNNForwardInference(), cudnnRNNForwardTraining(), cudnnRNNBackwardData() and cudnnRNNBackwardWeights() routines.

Values

CUDNN_RNN_ALGO_STANDARD

Each RNN layer is executed as a sequence of operations. This algorithm is expected to have robust performance across a wide range of network parameters.

CUDNN_RNN_ALGO_PERSIST_STATIC

The recurrent parts of the network are executed using a persistent kernel approach. This method is expected to be fast when the first dimension of the input tensor is small (meaning, a small minibatch).

CUDNN_RNN_ALGO_PERSIST_STATIC is only supported on devices with compute capability >= 6.0.

CUDNN_RNN_ALGO_PERSIST_DYNAMIC

The recurrent parts of the network are executed using a persistent kernel approach. This method is expected to be fast when the first dimension of the input tensor is small (ie. a small minibatch). When using CUDNN_RNN_ALGO_PERSIST_DYNAMIC persistent kernels are prepared at runtime and are able to optimize using the specific parameters of the network and active GPU. As such, when using CUDNN_RNN_ALGO_PERSIST_DYNAMIC a one-time plan preparation stage must be executed. These plans can then be reused in repeated calls with the same model parameters.

The limits on the maximum number of hidden units supported when using CUDNN_RNN_ALGO_PERSIST_DYNAMIC are significantly higher than the limits when using CUDNN_RNN_ALGO_PERSIST_STATIC, however throughput is likely to significantly reduce when exceeding the maximums supported by CUDNN_RNN_ALGO_PERSIST_STATIC. In this regime, this method will still outperform CUDNN_RNN_ALGO_STANDARD for some cases.

CUDNN_RNN_ALGO_PERSIST_DYNAMIC is only supported on devices with compute capability >= 6.0 on Linux machines.

3.51. cudnnRNNBiasMode_t

cudnnRNNBiasMode_t is an enumerated type used to specify the number of bias vectors for RNN functions. See the description of the cudnnRNNMode_t enumerated type for the equations for each cell type based on the bias mode.

Values

CUDNN_RNN_NO_BIAS

Applies RNN cell formulas that do not use biases.

CUDNN_RNN_SINGLE_INP_BIAS

Applies RNN cell formulas that use one input bias vector in the input GEMM.

CUDNN_RNN_DOUBLE_BIAS

Applies RNN cell formulas that use two bias vectors.

CUDNN_RNN_SINGLE_REC_BIAS

Applies RNN cell formulas that use one recurrent bias vector in the recurrent GEMM.

3.52. cudnnRNNClipMode_t

cudnnRNNClipMode_t is an enumerated type used to select the LSTM cell clipping mode. It is used with cudnnRNNSetClip(), cudnnRNNGetClip() functions, and internally within LSTM cells.

Values

CUDNN_RNN_CLIP_NONE

Disables LSTM cell clipping.

CUDNN_RNN_CLIP_MINMAX

Enables LSTM cell clipping.

3.53. cudnnRNNDataDescriptor_t

cudnnRNNDataDescriptor_t is a pointer to an opaque structure holding the description of an RNN data set. The function cudnnCreateRNNDataDescriptor() is used to create one instance, and cudnnSetRNNDataDescriptor() must be used to initialize this instance.

3.54. cudnnRNNDataLayout_t

cudnnRNNDataLayout_t is an enumerated type used to select the RNN data layout. It is used used in the API calls cudnnGetRNNDataDescriptor and cudnnSetRNNDataDescriptor.

Values

CUDNN_RNN_DATA_LAYOUT_SEQ_MAJOR_UNPACKED

Data layout is padded, with outer stride from one time-step to the next.

CUDNN_RNN_DATA_LAYOUT_SEQ_MAJOR_PACKED

The sequence length is sorted and packed as in basic RNN API.

CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED

Data layout is padded, with outer stride from one batch to the next.

3.55. cudnnRNNDescriptor_t

cudnnRNNDescriptor_t is a pointer to an opaque structure holding the description of an RNN operation. cudnnCreateRNNDescriptor() is used to create one instance, and cudnnSetRNNDescriptor() must be used to initialize this instance.

3.56. cudnnRNNInputMode_t

cudnnRNNInputMode_t is an enumerated type used to specify the behavior of the first layer in the cudnnRNNForwardInference(), cudnnRNNForwardTraining(), cudnnRNNBackwardData() and cudnnRNNBackwardWeights() routines.

Values

CUDNN_LINEAR_INPUT

A biased matrix multiplication is performed at the input of the first recurrent layer.

CUDNN_SKIP_INPUT

No operation is performed at the input of the first recurrent layer. If CUDNN_SKIP_INPUT is used the leading dimension of the input tensor must be equal to the hidden state size of the network.

3.57. cudnnRNNMode_t

cudnnRNNMode_t is an enumerated type used to specify the type of network used in the cudnnRNNForwardInference, cudnnRNNForwardTraining, cudnnRNNBackwardData and cudnnRNNBackwardWeights routines.

Values

CUDNN_RNN_RELU

A single-gate recurrent neural network with a ReLU activation function.

In the forward pass, the output ht for a given iteration can be computed from the recurrent input ht-1 and the previous layer input xt, given the matrices W, R and the bias vectors, where ReLU(x) = max(x, 0).

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_DOUBLE_BIAS (default mode), then the following equation with biases bW and bR applies:

Copy

Copied!

            

            
ht = ReLU(Wixt + Riht-1 + bWi + bRi)
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_SINGLE_INP_BIAS or CUDNN_RNN_SINGLE_REC_BIAS, then the following equation with bias b applies:

Copy

Copied!

            

            
ht = ReLU(Wixt + Riht-1 + bi)
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_NO_BIAS, then the following equation applies:

Copy

Copied!

            

            
ht = ReLU(Wixt + Riht-1)
        

CUDNN_RNN_TANH

A single-gate recurrent neural network with a tanh activation function.

In the forward pass, the output ht for a given iteration can be computed from the recurrent input ht-1 and the previous layer input xt, given the matrices W, R and the bias vectors, and where tanh is the hyperbolic tangent function.

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_DOUBLE_BIAS (default mode), then the following equation with biases bW and bR applies:

Copy

Copied!

            

            
ht = tanh(Wixt + Riht-1 + bWi + bRi)
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_SINGLE_INP_BIAS or CUDNN_RNN_SINGLE_REC_BIAS, then the following equation with bias b applies:

Copy

Copied!

            

            
ht = tanh(Wixt + Riht-1 + bi)
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_NO_BIAS, then the following equation applies:

Copy

Copied!

            

            
ht = tanh(Wixt + Riht-1)
        

CUDNN_LSTM

A four-gate Long Short-Term Memory network with no peephole connections.

In the forward pass, the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1, the cell input ct-1 and the previous layer input xt, given the matrices W, R and the bias vectors.

In addition, the following applies:

• σ is the sigmoid operator such that: σ(x) = 1 / (1 + e-x),
• ◦ represents a point-wise multiplication,
• tanh is the hyperbolic tangent function, and
• it, ft, ot, c't represent the input, forget, output and new gates respectively.

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_DOUBLE_BIAS (default mode), then the following equations with biases bW and bR apply:

Copy

Copied!

            

            
it = σ(Wixt + Riht-1 + bWi + bRi)
ft = σ(Wfxt + Rfht-1 + bWf + bRf)
ot = σ(Woxt + Roht-1 + bWo + bRo)
c't = tanh(Wcxt + Rcht-1 + bWc + bRc)
ct = ft ◦ ct-1 + it ◦ c't
ht = ot ◦ tanh(ct)
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_SINGLE_INP_BIAS or CUDNN_RNN_SINGLE_REC_BIAS, then the following equations with bias b apply:

Copy

Copied!

            

            
it = σ(Wixt + Riht-1 + bi)
ft = σ(Wfxt + Rfht-1 + bf)
ot = σ(Woxt + Roht-1 + bo)
c't = tanh(Wcxt + Rcht-1 + bc)
ct = ft ◦ ct-1 + it ◦ c't
ht = ot ◦ tanh(ct)
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_NO_BIAS, then the following equations apply:

Copy

Copied!

            

            
it = σ(Wixt + Riht-1)
ft = σ(Wfxt + Rfht-1)
ot = σ(Woxt + Roht-1)
c't = tanh(Wcxt + Rcht-1)
ct = ft ◦ ct-1 + it ◦ c't
ht = ot◦tanh(ct)
        

CUDNN_GRU

A three-gate network consisting of Gated Recurrent Units.

In the forward pass, the output ht for a given iteration can be computed from the recurrent input ht-1 and the previous layer input xt given matrices W, R and the bias vectors.

In addition, the following applies:

• σ is the sigmoid operator such that: σ(x) = 1 / (1 + e-x),
• ◦ represents a point-wise multiplication,
• tanh is the hyperbolic tangent function, and
• it, rt, h't represent the input, reset, new gates respectively.

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_DOUBLE_BIAS (default mode), then the following equations with biases bW and bR apply:

Copy

Copied!

            

            
it = σ(Wixt + Riht-1 + bWi + bRu)
rt = σ(Wrxt + Rrht-1 + bWr + bRr)
h't = tanh(Whxt + rt◦(Rhht-1 + bRh) + bWh)
ht = (1 - it) ◦ h't + it ◦ ht-1
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_SINGLE_INP_BIAS, then the following equations with bias b apply:

Copy

Copied!

            

            
it = σ(Wixt + Riht-1 + bi)
rt = σ(Wrxt + Rrht-1 + br)
h't = tanh(Whxt + rt ◦ (Rhht-1) + bWh)
ht = (1 - it) ◦ h't + it ◦ ht-1
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_SINGLE_REC_BIAS, then the following equations with bias b apply:

Copy

Copied!

            

            
it = σ(Wixt + Riht-1 + bi)
rt = σ(Wrxt + Rrht-1 + br)
h't = tanh(Whxt + rt ◦ (Rhht-1 + bRh))
ht = (1 - it) ◦ h't + it ◦ ht-1
        

If cudnnRNNBiasMode_t biasMode in rnnDesc is CUDNN_RNN_NO_BIAS, then the following equations apply:

Copy

Copied!

            

            
it = σ(Wixt + Riht-1)
rt = σ(Wrxt + Rrht-1)
h't = tanh(Whxt + rt ◦ (Rhht-1))
ht = (1 - it) ◦ h't + it ◦ ht-1
        

3.58. cudnnRNNPaddingMode_t

cudnnRNNPaddingMode_t is an enumerated type used to enable or disable the padded input/output.

Values

CUDNN_RNN_PADDED_IO_DISABLED

Disables the padded input/output.

CUDNN_RNN_PADDED_IO_ENABLED

Enables the padded input/output.

3.59. cudnnSamplerType_t

cudnnSamplerType_t is an enumerated type passed to cudnnSetSpatialTransformerNdDescriptor() to select the sampler type to be used by cudnnSpatialTfSamplerForward() and cudnnSpatialTfSamplerBackward().

Values

CUDNN_SAMPLER_BILINEAR

Selects the bilinear sampler.

3.60. cudnnSeqDataAxis_t

cudnnSeqDataAxis_t is an enumerated type that indexes active dimensions in the dimA[] argument that is passed to the cudnnSetSeqDataDescriptor() function to configure the sequence data descriptor of type cudnnSeqDataDescriptor_t.

cudnnSeqDataAxis_t constants are also used in the axis[] argument of the cudnnSetSeqDataDescriptor() call to define the layout of the sequence data buffer in memory.

See cudnnSetSeqDataDescriptor()for a detailed description on how to use the cudnnSeqDataAxis_t enumerated type.

The CUDNN_SEQDATA_DIM_COUNT macro defines the number of constants in the cudnnSeqDataAxis_t enumerated type. This value is currently set to 4.

Values

CUDNN_SEQDATA_TIME_DIM

Identifies the TIME (sequence length) dimension or specifies the TIME in the data layout.

CUDNN_SEQDATA_BATCH_DIM

Identifies the BATCH dimension or specifies the BATCH in the data layout.

CUDNN_SEQDATA_BEAM_DIM

Identifies the BEAM dimension or specifies the BEAM in the data layout.

CUDNN_SEQDATA_VECT_DIM

Identifies the VECT (vector) dimension or specifies the VECT in the data layout.

3.61. cudnnSeqDataDescriptor_t

cudnnSeqDataDescriptor_t is a pointer to an opaque structure holding parameters of the sequence data container or buffer. The sequence data container is used to store fixed size vectors defined by the VECT dimension. Vectors are arranged in additional three dimensions: TIME, BATCH and BEAM.

The TIME dimension is used to bundle vectors into sequences of vectors. The actual sequences can be shorter than the TIME dimension, therefore, additional information is needed about each sequence length and how unused (padding) vectors should be saved.

It is assumed that the sequence data container is fully packed. The TIME, BATCH and BEAM dimensions can be in any order when vectors are traversed in the ascending order of addresses. Six data layouts (permutation of TIME, BATCH and BEAM) are possible. The cudnnSeqDataDescriptor_t object holds the following parameters:

• data type used by vectors
• TIME, BATCH, BEAM and VECTdimensions
• data layout
• the length of each sequence along the TIME dimension
• an optional value to be copied to output padding vectors

Use the cudnnCreateSeqDataDescriptor() function to create one instance of the sequence data descriptor object and cudnnDestroySeqDataDescriptor() to delete a previously created descriptor. Use the cudnnSetSeqDataDescriptor() function to configure the descriptor.

This descriptor is used by multi-head attention API functions.

3.62. cudnnSoftmaxAlgorithm_t

cudnnSoftmaxAlgorithm_t is used to select an implementation of the softmax function used in cudnnSoftmaxForward() and cudnnSoftmaxBackward().

Values

CUDNN_SOFTMAX_FAST

This implementation applies the straightforward softmax operation.

CUDNN_SOFTMAX_ACCURATE

This implementation scales each point of the softmax input domain by its maximum value to avoid potential floating point overflows in the softmax evaluation.

CUDNN_SOFTMAX_LOG

This entry performs the log softmax operation, avoiding overflows by scaling each point in the input domain as in CUDNN_SOFTMAX_ACCURATE.

3.63. cudnnSoftmaxMode_t

cudnnSoftmaxMode_t is used to select over which data the cudnnSoftmaxForward() and cudnnSoftmaxBackward() are computing their results.

Values

CUDNN_SOFTMAX_MODE_INSTANCE

The softmax operation is computed per image (N) across the dimensions C,H,W.

CUDNN_SOFTMAX_MODE_CHANNEL

The softmax operation is computed per spatial location (H,W) per image (N) across the dimension C.

3.64. cudnnSpatialTransformerDescriptor_t

cudnnSpatialTransformerDescriptor_t is a pointer to an opaque structure holding the description of a spatial transformation operation. cudnnCreateSpatialTransformerDescriptor() is used to create one instance, cudnnSetSpatialTransformerNdDescriptor() is used to initialize this instance, and cudnnDestroySpatialTransformerDescriptor() is used to destroy this instance.

3.65. cudnnStatus_t

cudnnStatus_t is an enumerated type used for function status returns. All cuDNN library functions return their status, which can be one of the following values:

Values

CUDNN_STATUS_SUCCESS

The operation completed successfully.

CUDNN_STATUS_NOT_INITIALIZED

The cuDNN library was not initialized properly. This error is usually returned when a call to cudnnCreate() fails or when cudnnCreate() has not been called prior to calling another cuDNN routine. In the former case, it is usually due to an error in the CUDA Runtime API called by cudnnCreate() or by an error in the hardware setup.

CUDNN_STATUS_ALLOC_FAILED

Resource allocation failed inside the cuDNN library. This is usually caused by an internal cudaMalloc() failure.

To correct: prior to the function call, deallocate previously allocated memory as much as possible.

CUDNN_STATUS_BAD_PARAM

An incorrect value or parameter was passed to the function.

To correct: ensure that all the parameters being passed have valid values.

CUDNN_STATUS_ARCH_MISMATCH

The function requires a feature absent from the current GPU device. Note that cuDNN only supports devices with compute capabilities greater than or equal to 3.0.

To correct: compile and run the application on a device with appropriate compute capability.

CUDNN_STATUS_MAPPING_ERROR

An access to GPU memory space failed, which is usually caused by a failure to bind a texture.

To correct: prior to the function call, unbind any previously bound textures.

Otherwise, this may indicate an internal error/bug in the library.

CUDNN_STATUS_EXECUTION_FAILED

The GPU program failed to execute. This is usually caused by a failure to launch some cuDNN kernel on the GPU, which can occur for multiple reasons.

To correct: check that the hardware, an appropriate version of the driver, and the cuDNN library are correctly installed.

Otherwise, this may indicate an internal error/bug in the library.

CUDNN_STATUS_INTERNAL_ERROR

An internal cuDNN operation failed.

CUDNN_STATUS_NOT_SUPPORTED

The functionality requested is not presently supported by cuDNN.

CUDNN_STATUS_LICENSE_ERROR

The functionality requested requires some license and an error was detected when trying to check the current licensing. This error can happen if the license is not present or is expired or if the environment variable NVIDIA_LICENSE_FILE is not set properly.

CUDNN_STATUS_RUNTIME_PREREQUISITE_MISSING

Runtime library required by RNN calls (libcuda.so or nvcuda.dll) cannot be found in predefined search paths.

CUDNN_STATUS_RUNTIME_IN_PROGRESS

Some tasks in the user stream are not completed.

CUDNN_STATUS_RUNTIME_FP_OVERFLOW

Numerical overflow occurred during the GPU kernel execution.

3.66. cudnnTensorDescriptor_t

cudnnCreateTensorDescriptor_t is a pointer to an opaque structure holding the description of a generic n-D dataset. cudnnCreateTensorDescriptor() is used to create one instance, and one of the routines cudnnSetTensorNdDescriptor(), cudnnSetTensor4dDescriptor() or cudnnSetTensor4dDescriptorEx() must be used to initialize this instance.

3.67. cudnnTensorFormat_t

cudnnTensorFormat_t is an enumerated type used by cudnnSetTensor4dDescriptor() to create a tensor with a pre-defined layout. For a detailed explanation of how these tensors are arranged in memory, see Data Layout Formats.

Values

CUDNN_TENSOR_NCHW

This tensor format specifies that the data is laid out in the following order: batch size, feature maps, rows, columns. The strides are implicitly defined in such a way that the data are contiguous in memory with no padding between images, feature maps, rows, and columns; the columns are the inner dimension and the images are the outermost dimension.

CUDNN_TENSOR_NHWC

This tensor format specifies that the data is laid out in the following order: batch size, rows, columns, feature maps. The strides are implicitly defined in such a way that the data are contiguous in memory with no padding between images, rows, columns, and feature maps; the feature maps are the inner dimension and the images are the outermost dimension.

CUDNN_TENSOR_NCHW_VECT_C

This tensor format specifies that the data is laid out in the following order: batch size, feature maps, rows, columns. However, each element of the tensor is a vector of multiple feature maps. The length of the vector is carried by the data type of the tensor. The strides are implicitly defined in such a way that the data are contiguous in memory with no padding between images, feature maps, rows, and columns; the columns are the inner dimension and the images are the outermost dimension. This format is only supported with tensor data types CUDNN_DATA_INT8x4, CUDNN_DATA_INT8x32, and CUDNN_DATA_UINT8x4.

The CUDNN_TENSOR_NCHW_VECT_C can also be interpreted in the following way: The NCHW INT8x32 format is really N x (C/32) x H x W x 32 (32 Cs for every W), just as the NCHW INT8x4 format is N x (C/4) x H x W x 4 (4 Cs for every W). Hence, the VECT_C name - each W is a vector (4 or 32) of Cs.

3.68. cudnnTensorTransformDescriptor_t

cudnnTensorTransformDescriptor_t is an opaque structure containing the description of the tensor transform. Use the cudnnCreateTensorTransformDescriptor function to create an instance of this descriptor, and cudnnDestroyTensorTransformDescriptor function to destroy a previously created instance.

3.69. cudnnWgradMode_t

cudnnWgradMode_t is an enumerated type that selects how buffers holding gradients of the loss function, computed with respect to trainable parameters, are updated. Currently, this type is used by the cudnnGetMultiHeadAttnWeights() function only. Values

CUDNN_WGRAD_MODE_ADD

A weight gradient component corresponding to a new batch of inputs is added to previously evaluated weight gradients. Before using this mode, the buffer holding weight gradients should be initialized to zero. Alternatively, the first API call outputting to an uninitialized buffer should use the CUDNN_WGRAD_MODE_SET option.

CUDNN_WGRAD_MODE_SET

A weight gradient component, corresponding to a new batch of inputs, overwrites previously stored weight gradients in the output buffer.

This chapter describes the API of all the routines of the cuDNN library.

4.1. cudnnActivationBackward

            

            
cudnnStatus_t cudnnActivationBackward(
    cudnnHandle_t                    handle,
    cudnnActivationDescriptor_t      activationDesc,
    const void                      *alpha,
    const cudnnTensorDescriptor_t    yDesc,
    const void                      *y,
    const cudnnTensorDescriptor_t    dyDesc,
    const void                      *dy,
    const cudnnTensorDescriptor_t    xDesc,
    const void                      *x,
    const void                      *beta,
    const cudnnTensorDescriptor_t    dxDesc,
    void                            *dx)
        

This routine computes the gradient of a neuron activation function.

handle

Input. Handle to a previously created cuDNN context. For more information, see cudnnHandle_t.

activationDesc

Input. Activation descriptor. See cudnnActivationDescriptor_t.

alpha, beta

Input. Pointers to scaling factors (in host memory) used to blend the computation result with prior value in the output layer as follows:

Copy

Copied!

            

            
dstValue = alpha[0]*result + beta[0]*priorDstValue
        

For more information, see Scaling Parameters in the cuDNN Developer Guide.

yDesc

Input. Handle to the previously initialized input tensor descriptor. For more information, see cudnnTensorDescriptor_t.

y

Input. Data pointer to GPU memory associated with the tensor descriptor yDesc.

dyDesc

Input. Handle to the previously initialized input differential tensor descriptor.

dy

Input. Data pointer to GPU memory associated with the tensor descriptor dyDesc.

xDesc

Input. Handle to the previously initialized output tensor descriptor.

x

Input. Data pointer to GPU memory associated with the output tensor descriptor xDesc.

dxDesc

Input. Handle to the previously initialized output differential tensor descriptor.

dx

Output. Data pointer to GPU memory associated with the output tensor descriptor dxDesc.