Memory Format Propagation

This example demonstrates memory format propagation, which is critical for deep learning applications performance.

Example code: memory_format_propagation.cpp

Memory format propagation is one of the central notions that needs to be well-understood to use oneDNN correctly.

Convolution and inner product primitives choose the memory format when you create them with the placeholder memory format dnnl::memory::format_tag::any for input or output. The memory format chosen is based on different circumstances such as hardware and convolutional parameters. Using the placeholder memory format is the recommended practice for convolutions, since they are the most compute-intensive operations in most topologies where they are present.

Other primitives, such as Elementwise, LRN, batch normalization and other, on forward propagation should use the same memory format as the preceding layer thus propagating the memory format through multiple oneDNN primitives. This avoids unnecessary reorders which may be expensive and should be avoided unless a compute-intensive primitive requires a different format. For performance reasons, backward computations of such primitives requires consistent memory format with the corresponding forward computations. Hence, when initializing there primitives for backward computations you should use dnnl::memory::format_tag::any memory format tag as well.

Below is the short summary when to use and not to use memory format dnnl::memory::format_tag::any during operation description initialization:

Primitive Kinds

Forward Propagation

Backward Propagation

No Propagation

Compute intensive: (De-)convolution, Inner product, RNN

Use dnnl::memory::format_tag::any

Use dnnl::memory::format_tag::any


Memory-bandwidth limited: Pooling, Layer and Batch Normalization, Local Response Normalization, Elementwise, Shuffle, Softmax

Use memory format from preceding layer for inputs, and dnnl::memory::format_tag::any for outputs

Use dnnl::memory::format_tag::any for gradient tensors, and actual memory formats for data tensors


Memory-bandwidth limited: Reorder, Concat, Sum, Binary



Use memory format from preceding layer for inputs, and dnnl::memory::format_tag::any for outputs

Additional format synchronization is required between forward and backward computations when running training workloads. This topic is covered in Training-Specific Aspects.

For better understanding of the architecture and design of oneDNN as well as the concepts used in the library, please refer to Understanding Memory Formats.

Introduction to the tutorial

This C++ API example demonstrates how to use optimized memory formats supported by oneDNN:

  • How to configure primitives to use optimized memory formats.

  • How to determine whether data needs to be reordered from/to optimized memory formats.

This tutorial assumes that the reader has already reviewed the Getting started tutorial.

The example is built around a CNN consisting of a convolution followed by a pooling and consists of the following steps:

  1. Create a pooling primitive descriptor based on the memory format chosen by the convolution primitive.

  2. Create memory descriptors for input and output data in the NCHW memory format.

  3. Determine if input and output data needs to be reordered from/to the optimized memory format.

  4. Create memory objects; and necessary primitives and execute them.

These steps are implemented in the memory_format_propagation() function which in turn is called from main() which is also responsible for error handling.

memory_format_propagation() function


We start by creating an engine and a stream that we will use when creating primitive descriptors and executing primitives.

engine eng(engine_kind, 0);
stream s(eng);

Create convolution and pooling primitives

To specify that a primitive should pick an optimized format for the specified computation parameters, we create memory descriptors with memory format set to dnnl::memory::format_tag::any.

This approach works only for a limited set of primitives: convolutions and inner products. Additionally, dnnl::memory::format_tag::any can be specified for destination memory descriptors which implies that destination will have the same memory format as the source.

// Tensor and kernel dimensions. We use the same 3x3 kernel with padding=1
// for both convolution and pooling primitives, which means that the
// activation tensor shapes do not change.
const int N = 1, H = 14, W = 14, IC = 128, OC = 256, KH = 3, KW = 3;
auto conv_src_md = memory::desc({N, IC, H, W}, memory::data_type::f32,
        memory::format_tag::any // let convolution choose memory format
auto conv_weights_md = memory::desc(
        {OC, IC, KH, KW}, memory::data_type::f32,
        memory::format_tag::any // let convolution choose memory format
auto conv_dst_md = memory::desc({N, OC, H, W}, memory::data_type::f32,
        memory::format_tag::any // let convolution choose memory format
auto pool_dst_md = conv_dst_md; // shape does not change

Next, we pass the memory descriptors to primitive descriptors constructors.

auto conv_pd = convolution_forward::primitive_desc(
        {prop_kind::forward_inference, algorithm::convolution_auto,
                conv_src_md, conv_weights_md,
                conv_dst_md, // shape information
                {1, 1}, // strides
                {1, 1}, {1, 1}}, // left and right padding
auto pool_pd = pooling_forward::primitive_desc(
        {prop_kind::forward_inference, algorithm::pooling_max,
                conv_pd.dst_desc(), pool_dst_md, // shape information
                {1, 1}, {KH, KW}, // strides and kernel
                {1, 1}, {1, 1}}, // left and right padding

Create source and destination memory objects

We assume that the ‘user’ source and destination memory format is NCHW. Since there is no result validation in this tutorial, we do not bother with filling the data with some values and let oneDNN allocate the memory.

auto src_mem = memory(
        {{N, IC, H, W}, memory::data_type::f32, memory::format_tag::nchw},
auto weights_mem = memory({{OC, IC, KH, KW}, memory::data_type::f32,
auto dst_mem = memory(
        {{N, OC, H, W}, memory::data_type::f32, memory::format_tag::nchw},

Determine if source and destination need to be reordered

The idiomatic way to check if a reorder is necessary between the memory format expected a primitive (the convolution in our case) and the available memory format is to compare the corresponding memory descriptors.

bool need_reorder_src = conv_pd.src_desc() != src_mem.get_desc();


It is by design that it is not possible to just compare memory tags. The reason behind this is that a memory format tags only provide a partial description of how data is laid out in memory and do not, for example, describe memory objects obtained via sub-memory constructor.

We repeat the process for the weights and destination memory format descriptors as well.

bool need_reorder_weights
        = conv_pd.weights_desc() != weights_mem.get_desc();
bool need_reorder_dst = conv_pd.dst_desc() != dst_mem.get_desc();

Allocate intermediate buffers if necessary

Based on the flags computed before, we can now decide if we need extra intermediate buffers to hold the source and weights data for the convolution and the output of the pooling.

Memory objects for the intermediate buffers are created based on the memory descriptors obtained from the primitive descriptors to ensure consistency.

auto conv_src_mem
        = need_reorder_src ? memory(conv_pd.src_desc(), eng) : src_mem;
auto conv_weights_mem = need_reorder_weights
        ? memory(conv_pd.weights_desc(), eng)
        : weights_mem;
auto conv_dst_mem = memory(conv_pd.dst_desc(), eng);
auto pool_dst_mem
        = need_reorder_dst ? memory(pool_pd.dst_desc(), eng) : dst_mem;

Perform reorders for source data if necessary

Now we get to the part where we actually start executing things. We check if reorders are necessary based on the flags computed before and create and execute them immediately.


We call dnnl::stream::wait() before reorder primitives get out of scope and destroyed to accommodate for potentially asynchronous execution.

if (need_reorder_src) {
    auto reorder_src = reorder(src_mem, conv_src_mem);
            s, {{DNNL_ARG_FROM, src_mem}, {DNNL_ARG_TO, conv_src_mem}});
    s.wait(); // wait for the reorder to complete

if (need_reorder_weights) {
    auto reorder_weights = reorder(weights_mem, conv_weights_mem);
            {{DNNL_ARG_FROM, weights_mem},
                    {DNNL_ARG_TO, conv_weights_mem}});
    s.wait(); // wait for the reorder to complete

Create and execute convolution and pooling primitives

After the reorders, we are now ready to compute convolution and pooling.

auto conv_scratchpad_mem = memory(conv_pd.scratchpad_desc(), eng);
auto conv = convolution_forward(conv_pd);
        {{DNNL_ARG_SRC, conv_src_mem}, {DNNL_ARG_WEIGHTS, conv_weights_mem},
                {DNNL_ARG_DST, conv_dst_mem}});
auto pool_scratchpad_mem = memory(pool_pd.scratchpad_desc(), eng);
auto pool = pooling_forward(pool_pd);
        s, {{DNNL_ARG_SRC, conv_dst_mem}, {DNNL_ARG_DST, pool_dst_mem}});

Reorder destination data if necessary

The only potentially remaining operation is a reorder from the pooling destination memory object to the user’s one. Similarly to the reorders for the source and weights memory objects, it is performed depending on the value of the previously computed flag.

if (need_reorder_dst) {
    auto reorder_dst = reorder(pool_dst_mem, dst_mem);
            s, {{DNNL_ARG_FROM, pool_dst_mem}, {DNNL_ARG_TO, dst_mem}});


Upon compiling and run the example the output should be just:

Example passed.

It may be interesting to check what really happens during the run. We can use DNNL_VERBOSE environment variable for that (see also Verbose Mode). Here’s an example output:

$ DNNL_VERBOSE=1 ./memory-format-propagation-cpp
dnnl_verbose,info,oneDNN <ver> (Git Hash <hash>)
dnnl_verbose,info,cpu,isa:Intel AVX2
    src_f32::blocked:abcd:f0 dst_f32::blocked:aBcd8b:f0,,,1x128x14x14,0.326904
    src_f32::blocked:abcd:f0 dst_f32::blocked:ABcd8b8a:f0,,,256x128x3x3,0.244141
    src_f32::blocked:aBcd8b:f0 wei_f32::blocked:ABcd8b8a:f0 bia_undef::undef::f0 dst_f32::blocked:aBcd8b:f0,,
    src_f32::blocked:aBcd8b:f0 dst_f32::blocked:aBcd8b:f0 ws_undef::undef::f0,,
    src_f32::blocked:aBcd8b:f0 dst_f32::blocked:abcd:f0,,,1x256x14x14,0.0419922
Example passed on CPU.

From this output we can deduce that:

  • The convolution primitive picked up dnnl::memory::format_tag::aBcd8b optimized memory format for activations. In this format the channels dimension (denoted by letter B since it is the second dimension; see also Naming Conventions) is blocked by a factor of 8. Because of this memory format is different from the NCHW format the tutorial uses, the source and destination had to be reordered to and from this optimized memory layout.

  • The convolution primitive picked up dnnl::memory::format_tag::ABcd8b8a optimized memory format (output (A) and input (B) channel dimensions blocked by 8) which we also had to reorder the initial weights to since they are in the OIHW memory format.