Deep Neural Network Library (DNNL)  1.1.3
Performance library for Deep Learning
Getting started

Example code: getting_started.cpp

This C++ API example demonstrates basics of DNNL programming model:

The example uses the ReLU operation and consists of the following steps:

  1. Creating Engine and stream to execute a primitive.
  2. Performing Data preparation (code outside of DNNL).
  3. Wrapping data into DNNL memory object (using different flavors).
  4. Creating a ReLU primitive.
  5. Executing the ReLU primitive.
  6. Obtaining the result and validation (checking that the resulting image does not contain negative values).

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

Public headers

To start using DNNL we should first include dnnl.hpp header file in the program. We also include dnnl_debug.h in example_utils.hpp that contains some debugging facilities like returning a string representation for common DNNL C types.

getting_started_tutorial() function

Engine and stream

All DNNL primitives and memory objects are attached to a particular dnnl::engine, which is an abstraction of an computational device (see also Basic Concepts). The primitives are created and optimized for the device they are attached to and the memory objects refer to memory residing on the corresponding device. In particular, that means neither memory objects nor primitives that were created for one engine can be used on another.

To create an engine we should specify the dnnl::engine::kind and the index of the device of the given kind.

engine eng(engine_kind, 0);

In addition to an engine, all primitives require a dnnl::stream for the execution. The stream encapsulates an execution context and is tied to a particular engine.

The creation is pretty straightforward:

stream engine_stream(eng);














In the simple cases, when a program works with one device only (e.g. only on CPU), an engine and a stream can be created once and used throughout the program. Some frameworks create singleton objects that hold DNNL engine and stream and are use them throughout the code.











Data preparation (code outside of DNNL)

Now that the preparation work is done, let's create some data to work with. We will create a 4D tensor in NHWC format, which is quite popular in many frameworks.

Note that even though we work with one image only, the image tensor is still 4D. The extra 4th dimension (here N) corresponds to the batch, and, in case of a single image, equals to 1. This is pretty typical to have the batch dimension even when working with a single image.

In DNNL all CNN primitives assume tensors have batch dimension, which is always the first logical dimension (see also Naming Conventions).

const int N = 1, H = 13, W = 13, C = 3;
// Compute physical strides for each dimension
const int stride_N = H * W * C;
const int stride_H = W * C;
const int stride_W = C;
const int stride_C = 1;
// An auxiliary function that maps logical index to the physical offset
auto offset = [=](int n, int h, int w, int c) {
return n * stride_N + h * stride_H + w * stride_W + c * stride_C;
};
// The image size
const int image_size = N * H * W * C;
// Allocate a buffer for the image
std::vector<float> image(image_size);
// Initialize the image with some values
for (int n = 0; n < N; ++n)
for (int h = 0; h < H; ++h)
for (int w = 0; w < W; ++w)
for (int c = 0; c < C; ++c) {
int off = offset(
n, h, w, c); // Get the physical offset of a pixel
image[off] = -std::cos(off / 10.f);
}












Wrapping data into DNNL memory object

Now, having the image ready, let's wrap it in an dnnl::memory object to be able to pass the data to DNNL primitives.

Creating dnnl::memory consists of 2 steps:

  1. Initializing the dnnl::memory::desc struct (also referred as memory descriptor) that only describes the tensor data, but doesn't contain the data itself. Memory descriptors are used to create dnnl::memory objects and to initialize primitive descriptors (shown later in the example);
  2. Creating the dnnl::memory object itself (also referred as a memory object), based on the memory descriptor initialized in step 1, an engine, and, optionally, a handle to a data. The memory object is used when a primitive is executed.

Thanks to the list initialization introduced in C++11 it is possible to combine these two steps whenever a memory descriptor is not used anywhere else but in creating an dnnl::memory object.

However, for the sake of demonstration, we will show both steps explicitly.









Memory descriptor

To initialize the dnnl::memory::desc we need to pass:

  1. The tensor's dimensions, the semantic order of which is defined by the primitive that will use this memory (descriptor). Which leads to the following:
    Warning
    Memory descriptors and objects are not aware of any meaning of the data they describe or contain.
  2. The data type for the tensor (dnnl::memory::data_type).
  3. The memory format tag (dnnl::memory::format_tag) that describes how the data is going to be laid out in device's memory. The memory format is required for the primitive to correctly handle the data.

The code:

auto src_md = memory::desc(
{N, C, H, W}, // logical dims, the order is defined by a primitive
memory::data_type::f32, // tensor's data type
memory::format_tag::nhwc // memory format, NHWC in this case
);










The first thing to notice here is that we pass dimensions as {N, C, H, W} while it might seem more natural to pass {N, H, W, C}, which better corresponds to the user's code. This is because DNNL CNN primitives like ReLU always expect tensors in the following form:

Spatial dim Ten
0D \(N \times C\)
1D \(N \times C \times W\)
2D \(N \times C \times H \times W\)
3D \(N \times C \times D \times H \times W\)

where:

Now that the logical order of dimension is defined, we need to specify the memory format (the third parameter), which describes how logical indices map to the offset in memory. This is the place where user's format NHWC comes into play. DNNL has different dnnl::memory::format_tag values that covers the most popular memory formats like NCHW, NHWC, CHWN, and some others.

The memory descriptor for the image is called src_md. The src part comes from the fact that the image will be a source for the ReLU primitive (i.e. we formulate memory names from the primitive perspective, hence we will use dst to name the output memory). The md is an acronym for Memory Descriptor.







Alternative way to create a memory descriptor

Before we continue with memory creation, let us show the alternative way to create the same memory descriptor: instead of using the dnnl::memory::format_tag we can directly specify the strides of each tensor dimension:

auto alt_src_md = memory::desc(
{N, C, H, W}, // logical dims, the order is defined by a primitive
memory::data_type::f32, // tensor's data type
{stride_N, stride_C, stride_H, stride_W} // the strides
);
// Sanity check: the memory descriptors should be the same
if (src_md != alt_src_md)
throw std::string("memory descriptor initialization mismatch");








Just as before, the tensor's dimensions come in the N, C, H, W order as required by CNN primitives. To define the physical memory format the strides are passed as the third parameter. Note that the order of the strides corresponds to the order of the tensor's dimensions.

Warning
Using the wrong order might lead to incorrect results or even a crash.





Creating a memory object

Having a memory descriptor and an engine prepared let's create an input and an output memory objects for ReLU primitive

// src_mem contains a copy of image after write_to_dnnl_memory function
auto src_mem = memory(src_md, eng);
write_to_dnnl_memory(image.data(), src_mem);
// For dst_mem the library allocates buffer
auto dst_mem = memory(src_md, eng);






We already have a memory buffer for the source memory object. We pass it to the dnnl::memory::memory(const desc &, const engine &, void *) constructor that takes a buffer pointer with its last argument.

Let's use a constructor that instructs the library to allocate a memory buffer for the dst_mem for educational purposes.

The key difference between these two are:

  1. The library will own the memory for dst_mem and will deallocate it when dst_mem is destroyed. That means the memory buffer can only be used while dst_mem is alive.
  2. Library-allocated buffers have good alignment which typically results in better performance.
Note
Memory allocated outside of the library and passed to Intel DNNL should have good alignment for better performance.

In subsequent section we will show how to get the buffer (pointer) from the dst_mem memory object.

Creating a ReLU primitive

Let's now create a ReLU primitive.

The library implements ReLU primitive as a particular algorithm of a more general Eltwise primitive which applies specified function to each and every element of the source tensor.

Just like in case of dnnl::memory a user should always go through (at least) 3 creation steps (which however, can be sometimes combined thanks to C++11):

  1. Initialize operation descriptor (in case of this example, dnnl::eltwise_forward::desc), which defines the operation parameters.
  2. Create an operation primitive descriptor (here dnnl::eltwise_forward::primitive_desc), which is a lightweight descriptor of the actual algorithm that implements given operation. User can query different characteristics of the chosen implementation like memory consumptions and some others that will be covered in the next topic (Memory format propagation).
  3. Create a primitive (here dnnl::eltwise_forward) that can be executed on memory objects to compute the operation.

DNNL separates the steps 2 and 3 to allow user to inspect details of a primitive implementation prior to creating the primitive which may be expensive because, for example, DNNL generates the optimized computational code on the fly.

Note
Primitive creation might be a very expensive operation, so consider creating primitive objects once and executing them multiple times.

The code:

// ReLU op descriptor (no engine- or implementation-specific information)
auto relu_d = eltwise_forward::desc(
prop_kind::forward_inference, algorithm::eltwise_relu,
src_md, // the memory descriptor for an operation to work on
0.f, // alpha parameter means negative slope in case of ReLU
0.f // beta parameter is ignored in case of ReLU
);
// ReLU primitive descriptor, which corresponds to a particular
// implementation in the library
auto relu_pd
= eltwise_forward::primitive_desc(relu_d, // an operation descriptor
eng // an engine the primitive will be created for
);
// ReLU primitive
auto relu = eltwise_forward(relu_pd); // !!! this can take quite some time





A note about variable names. Similar to the _md suffix used for memory descriptor, we use _d for the operation descriptor names, _pd for the primitive descriptors, and no suffix for primitives themselves.

It is worth mentioning that we specified the exact tensor and its memory format when we were initializing the relu_d. That means relu primitive would perform computations with memory objects that correspond to this description. This is the one and only one way of creating non-compute-intensive primitives like Eltwise, Batch Normalization, and others.

Compute-intensive primitives (like Convolution) have an ability to define the appropriate memory format on their own. This is one of the key features of the library and will be discussed in detail in the next topic: Memory format propagation.


Executing the ReLU primitive

Finally, let's execute the primitive and wait for its completion.

The input and output memory objects are passed to the execute() method using a <tag, memory> map. Each tag specifies what kind of tensor each memory object represents. All Eltwise primitives require the map to have two elements: a source memory object (input) and a destination memory (output).

A primitive is executed in a stream (the first parameter of the execute() method). Depending on a stream kind an execution might be blocking or non-blocking. This means that we need to call dnnl::stream::wait before accessing the results.

// Execute ReLU (out-of-place)
relu.execute(engine_stream, // The execution stream
{
// A map with all inputs and outputs
{DNNL_ARG_SRC, src_mem}, // Source tag and memory obj
{DNNL_ARG_DST, dst_mem}, // Destination tag and memory obj
});
// Wait the stream to complete the execution
engine_stream.wait();



The Eltwise is one of the primitives that support in-place operations, meaning the source and destination memory can be the same. To perform in-place transformation user needs to pass the same memory object for the both DNNL_ARG_SRC and DNNL_ARG_DST tags:

// Execute ReLU (in-place)
// relu.execute(engine_stream, {
// {DNNL_ARG_SRC, src_mem},
// {DNNL_ARG_DST, src_mem},
// });


Obtaining the result and validation

Now that we have the computed result let's validate that it is actually correct. The result is stored in the dst_mem memory object. So we need to obtain the C++ pointer to a buffer with data via dnnl::memory::get_data_handle() and cast it to the proper data type as shown below.

Warning
The dnnl::memory::get_data_handle() returns a raw handle to the buffer which type is engine specific. For CPU engine the buffer is always a pointer to void which can safely be used. However, for engines other than CPU the handle might be runtime-specific type, such as cl_mem in case of GPU/OpenCL.
// Obtain a buffer for the `dst_mem` and cast it to `float *`.
// This is safe since we created `dst_mem` as f32 tensor with known
// memory format.
std::vector<float> relu_image(image_size);
read_from_dnnl_memory(relu_image.data(), dst_mem);
// Check the results
for (int n = 0; n < N; ++n)
for (int h = 0; h < H; ++h)
for (int w = 0; w < W; ++w)
for (int c = 0; c < C; ++c) {
int off = offset(
n, h, w, c); // get the physical offset of a pixel
float expected = image[off] < 0
? 0.f
: image[off]; // expected value
if (relu_image[off] != expected) {
std::stringstream ss;
ss << "Unexpected output at index(" << n << ", " << c
<< ", " << h << ", " << w << "): "
<< "Expect " << expected << " "
<< "Got " << relu_image[off];
throw ss.str();
}
}

main() function

We now just call everything we prepared earlier.

Since we are using DNNL C++ API we use exception to handle errors (see C and C++ APIs). The DNNL C++ API throws exceptions of type dnnl::error, which contains the error status (of type dnnl_status_t) and a human-readable error message accessible through regular what() method.

int main(int argc, char **argv) {
try {
engine::kind engine_kind = parse_engine_kind(argc, argv);
getting_started_tutorial(engine_kind);
std::cout << "Example passes" << std::endl;
} catch (dnnl::error &e) {
std::cerr << "DNNL error: " << e.what() << std::endl
<< "Error status: " << dnnl_status2str(e.status) << std::endl;
return 1;
} catch (std::string &e) {
std::cerr << "Error in the example: " << e << std::endl;
return 2;
}
return 0;
}

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

Example passes

Users are encouraged to experiment with the code to familiarize themselves with the concepts. In particular, one of the changes that might be of interest is to spoil some of the library calls to check how error handling happens. For instance, if we replace

relu.execute(engine_stream, {
{DNNL_ARG_SRC, src_mem},
{DNNL_ARG_DST, dst_mem},
});

with

relu.execute(engine_stream, {
{DNNL_ARG_SRC, src_mem},
// {DNNL_ARG_DST, dst_mem}, // Oops, forgot about this one
});

we should get the following output:

DNNL error: could not execute a primitive
Error status: invalid_arguments