Contents

Inference and Training Aspects

Propagation Kinds

oneDNN provides performance critical primitives to accelerate operations used both during training deep learning models and during the operations performed when the models are used for inference.

During inference, the input data is fed into the trained model which in turn produces a result (e.g. makes a prediction). This process is usually called forward propagation and corresponds to the dnnl::prop_kind::forward_inference propagation kind in oneDNN.

Training usually consists of the following steps.

  1. Make prediction based on the current state of the model. As in the case of inference, this step is called forward propagation, but corresponds to the dnnl::prop_kind::forward_training propagation kind. Note the difference in the names’ suffixes: _training here versus _inference mentioned above. The differences are covered below in the corresponding section below.

  2. Compute an error between predicted and the actual answer.

  3. Perform the backward propagation of errors to compute the weights (learnable parameters) gradient. For a given operation (layer) the backward propagation in turn can be split into two steps:

    • Propagating error with respect to data, i.e. computing diff_src from diff_dst (see Naming Conventions). This step corresponds to the dnnl::prop_kind::backward_data propagation kind;

    • Propagating error with respect to weights, i.e. computing diff_weights from diff_dst. This step makes sense only for the operations that have learnable parameters and corresponds to the dnnl::prop_kind::backward_weights propagation kind.

  4. Use computed gradients to modify the weights according to the chosen solver to improve the accuracy of the model.

Difference Between Forward Propagation on Training and Inference

Even though, mathematically, the forward propagation that happens during training and inference should be the same, in practice there are some differences mostly due to the performance considerations.

When executing inference, one may not care about values in the intermediate buffers during a model execution; hence one can reuse them as desired. However, if this is a forward propagation of a training it is beneficial to preserve input data, output data, or sometimes some intermediate data, that will later be used at the backward propagation to compute the gradients.

For example, let’s take max pooling (Pooling with algorithm kind dnnl::algorithm::pooling_max) as an example. The forward pass consists of computing the maximum values in the sliding window over the source tensor. Hence the output is just another tensor that contain these maximum values. However, in order to compute source gradient on backward propagation one needs to know the position of these maximum values in the source tensor. Of course, it is possible to use the original source tensor to locate the maximums again, but this might be more expensive compared to preserving the positions of the maximum values in another tensor, that will be then used during the backward propagation. oneDNN uses the latter approach: for max pooling primitive when the propagation kind is set to dnnl::prop_kind::forward_training the library produces one extra output called Workspace which will be covered later in this document.

Note

Key takeaways:

Different Backward Propagation Kinds

As mentioned above, oneDNN separates error back-propagation with respect to data and error back-propagation with respect to weights. The former corresponds to dnnl::prop_kind::backward_data, while the latter corresponds to dnnl::prop_kind::backward_weights (for example: Convolution).

Inference-Specific Aspects

The following list outlines the key specifics of running inference with oneDNN:

  1. As described above, always use dnnl::prop_kind::forward_inference as a propagation kind.

  2. To get maximum performance, consider performing operations in-place whenever possible (e.g. Eltwise and Batch Normalization). Check the primitives documentation pages to check which primitives support in-place operations.

  3. Create primitives once, and reuse them across multiple model invocations. This is especially relevant for the frameworks integration.

  4. Some primitives can be chained/fused with others using the post-ops attributes. This allows reducing memory bandwidth pressure and typically leads to better performance.

Most of these techniques are shown in the following examples:

Training-Specific Aspects

The following list outlines the key specifics of running training with oneDNN:

  1. During the forward propagation, use dnnl::prop_kind::forward_training as a propagation kind.

  2. During backward propagation, perform backward by data and backward by weights using the corresponding propagation kinds.

    • Note that some primitives may combine these operations if that is beneficial from a performance perspective. For example, RNN and Batch Normalization compute both diff_src and diff_weights at the same time. To highlight this behavior, the propagation kind is set to dnnl::prop_kind::backward.

  3. Create primitives once, and reuse them across multiple model invocations. This is especially relevant for the frameworks integration.

  4. The post-ops attributes in general are not applicable for training, because the fused computations they result in do not produce the intermediate tensors which may be required during the backward propagation. For example, if you fuse Convolution and Eltwise the direct output of the convolution would not be produced. However, it might be required during the backward propagation of the corresponding element-wise. (To compute diff_src, one must pass diff_dst memory and the original src memory, which was exactly the intermediate one.)

  5. To compute backward propagation, different primitives might require different tensors. The variety is caused by the mathematical formulas. For example, to compute backward propagation for Eltwise one needs to pass diff_dst and src, but to compute backward propagation for Softmax, one needs to pass diff_dst and dst. Check the documentation for each primitive to see what is required for each particular primitive.

  6. For the primitives that are created with dnnl::memory::format_tag::any memory format tag, there are no guarantees that the memory format on forward and backward propagations will match. So the robust integration should always be ready to emit Reorder when necessary. For example, it is not guaranteed that the src memory format of a convolution on forward propagation will always match the src memory format of the corresponding convolution on backward by weights propagation. Of course, the library tries to avoid unnecessary reorder, so in most cases the formats will be the same, but this would by no means always be true.

  7. For the memory bandwidth bound primitives like Eltwise, Pooling, and Batch Normalization, it is important to have diff_dst in the same memory format as the original dst. The mismatch of the formats would lead to significant performance issues. To ensure the proper format, users should always use dnnl::memory::format_tag::any memory format for gradient tensors (diff_dst, diff_src). If a primitive requires original data tensors (e.g. src in Eltwise or dst in Softmax) user must pass fully defined memory descriptor for these tensors. In other words src and dst memory descriptors cannot be initialized with dnnl::memory::format_tag::any for backward propagation. Based on the format of the original tensors, if any, and on forward primitive descriptor hint (see bullet 9 below) a primitive picks the proper format for the gradients. Occasionally, it might appear that the diff_dst that comes in is in other memory format than the primitive requires, hence robust integration code must be prepared to emit a reorder.

    • Alternatively, users may manually enforce diff_dst to have the same memory format as dst, though this is not recommended.

  8. Some primitives require an additional tensor to be passed between forward and backward propagation, which is called Workspace.

  9. When creating primitive descriptors on backward propagation, you might need to pass a primitive descriptor of the corresponding primitive from the forward propagation (in the API this primitive descriptor is typically called hint). This hint is required for the primitive to choose a proper implementation that would correspond to the one from the forward propagation. This is required only for the primitives that produce workspace, because it might be different for different implementations.

  10. When creating your working memory and memory descriptor, specify the type of memory you want to work with. This can be either 16-bit Brain Float (bf16) or 32-bit Floating Point (fp32). More details about using bf16 for training are detailed in the section Bfloat16 Training.

Most of these techniques are shown in the following examples:

Workspace

oneDNN uses the notion of workspace for some very particular cases. Specifically, the workspace is a tensor that the primitive fills in during forward propagation and that will then be used by the corresponding backward propagation operation. The example with max pooling was already discussed above.

The workflow for using workspace is:

  1. When creating a primitive for the forward propagation, query the primitive descriptor about the workspace requirement using .workspace_desc().

    • If the returned memory descriptor is essentially empty (i.e. is equal to dnnl::memory::desc() or for which dnnl::memory::desc::get_size() returns 0), no extra action is required the workspace is not required for this primitive in this configuration.

    • Otherwise, create a workspace memory based on the memory descriptor obtained and pass it to the execution function with DNNL_ARG_WORKSPACE tag.

  2. On backward propagation, attach that same workspace memory during the execution as well. The state of the workspace memory after backward computations are done is undefined.

Note

Even if workspace is not required, it is perfectly valid to create a workspace memory of zero size and follow the logic where the workspace is indeed required. Such an approach may simplify the integration because the common pass is used.

// FWD

auto forward_primitive_desc = ...::primitive_desc(); // create a primitive desc
auto workspace_md = forward_primitive_desc.workspace_desc(); // query workspace
memory workspace(workspace_md, engine); // create a memory (even if empty)

primitive_forward.execute(stream, {
        ...,
        {DNNL_ARG_WORKSPACE, workspace} // this is output
        });
// The workspace contains required information for the backward propagation,
// hence should not be used anywhere else.

// ...

// BWD
primitive_backward.execute(stream, {
        ...,
        {DNNL_ARG_WORKSPACE, workspace} // this input/output
        });
// The state of the workspace is undefined here

Warning

Do not confuse workspace with the Primitive Attributes: Scratchpad. The scratchpad is a temporary buffer that might be required by a primitive (no matter what propagation kind is) to perform an operation. It is used only during the primitive execution and should not be preserved across the calls.

Memory Format Propagation Inference