Deep Neural Network Library (DNNL)  1.2.0
Performance library for Deep Learning
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages
Inference

DNNL includes primitives for operations throughout a deep learning network topology. However, it is important to note the scope of DNNL is limited to performance critical functionality and the library does not provide all the functions necessary to implement deep learning workloads, for instance data preprocessing or computing loss function. The soft-max classifier is the sole classifier included, but the application of other classifier types will require user's own implementations. The scope of the library is depicted in the following image:

img_inference_scope.jpg

Best Practices for Inference in DNNL

fp32 Inference

Use Forward Inference Primitives

DNNL provides a forward pass version of each primitive, that avoids storing information required for a backward pass (as in training).

Use the dnnl::prop_kind::forward_inference argument at creation of the operation descriptor, as in this convolution example:

auto conv_descr = convolution_forward::desc(prop_kind::forward_inference, ...);

Layout Propagation

Compute-intensive DNNL primitives execute with highest performance on CPU-friendly data formats. Please see description of data formats here.

Performance gains are maximized by reordering once, and then propagating the CPU-friendly format through as many layers as possible in your topology. DNNL provides the format_tag=any for memory descriptors that will be passed to compute-intensive primitives. The compute-intensive primitive types in DNNL are Convolution, Inner Product, and RNN.

To accomplish this propagation in a robust manner, its is recommended to follow these steps:

A. On compute-intensive operations:

B. On non-intensive operations:

Now let's take a look at the code syntax to accomplish the compute-intensive steps:

Pass the format_tag=any when creating DNNL memory descriptor for source, destination, and weights memory

source_mem_descr = memory::desc(args*, memory::format_tag::any);
dest_mem_descr = memory::desc(args*, memory::format_tag::any);
weights_mem_descr = memory::desc(args*, memory::format_tag::any);

Use these three memory descriptors with 'format _tag=any` to create operation descriptor

auto conv_descr = convolution_forward::desc(...,
source_mem_descr, weights_mem_descr, dest_mem_descr);

Use operation descriptor to create engine-aware primitive descriptor

auto conv_prim_descr = convolution_forward::primitive_desc(conv_descr, engine);

Query the primitive descriptor with .src_desc() method to get recommended format Write conditional reorder to execute only if user source data or weights don't match the recommended format (Note: Do this for weight_memory as well)

memory conv_source_memory = user_source_memory;
if (conv_prim_descr.src_desc() != user_source_memory.get_desc()) {
conv_source_memory = memory(conv_prim_descr.src_desc(), engine);
auto reorder_prim_descr = reorder::primitive_desc(user_source_memory, conv_source_memory);
reorder(reorder_prim_descr).execute(s, user_source_memory, conv_source_memory);
}

Create primitive and add it to stream with primitive.execute(stream, args)

auto conv = convolution_forward(conv_prim_descr);
conv.execute(s, {
{DNNL_ARG_SRC, conv_source_memory},
{DNNL_ARG_WEIGHTS, conv_weights_memory},
{DNNL_ARG_DST, conv_dest_memory}});

Cache Weights\ Weights are accessed many times during batched inference. At inference time these weights are essentially constants in the mapping function that the network is applying to the input data. As such, the weights should be reordered (if necessary) once and then used in the reorder form for the duration of the execution. This caching causes the computer to use them in a way similar to how a mathematical function applies a constant, i..e, "Grab-and-go" with no overhead for creation or reorder.

Primitive Reuse\ There is JIT compilation overhead associated with primitive creation. It is recommended to reuse any primitive that you can, and only create them once.

Fused Primitives\ DNNL provides fused versions of primitives that attach a non-intensive operation to the end of a compute-intensive operation and then executes both in a single pass, reducing the number of memory accesses needed for the combined operations. The non-intensive operation is added as a post-op attribute to the compute intensive primitive descriptor. Please note that post-ops do not change the number of inputs or outputs of the primitives. Please see the "Post-ops and Attributes" section of the doc for each primitive type in /docs/primitive/ for a list of available fused primitives.

A good example is adding ReLU as a post-op to convolution, which we will use as a demonstration below. The steps are

Create a post_op for fused ReLU

post_ops ops;
ops.append_eltwise(..., algorithm::eltwise_relu);

Create primitive attribute and add the post_op

primitive_attr attr;
attr.set_post_ops(ops);

Create a convolution descriptor

auto conv_descr = convolution_forward::desc(...);

Create a convolution primitive descriptor, passing the post-op infused attrs as an arg

auto conv_prim_descr = convolution_forward::primitive_desc(conv_descr, attrs, engine);

int8 Inference

DNNL supports low precision int8 for inference execution. Note that not all primitives have int8 versions. Sometimes the speed benefits would be minimal, or the loss in accuracy is not acceptable. Also the soft-max classifier only supports fp32, so int8 inference will require a reorder before executing this primitive.

By default, the DNNL reorder primitive does not scale upon casting to int8. In order to compress fp32 data to int8 precision while still preserving the entire shape of the distribution, a process called quantization must applied. Quantization will scale the data based on its range to efficiently fill the bits available for int8 type.

To achieve quantization upon casting, the user must provide a few inputs to DNNL in order to use int8 inference:

Please see the dedicated section on low precision computations in DNNL for a detailed discussion, including how to calculate the scaling factors.