oneAPI Deep Neural Network Library (oneDNN)
Performance library for Deep Learning

Annotated version: Primitive Example

* Copyright 2020 Intel Corporation
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* See the License for the specific language governing permissions and
* limitations under the License.
#include <algorithm>
#include <cmath>
#include <string>
#include <vector>
#include "dnnl.hpp"
#include "example_utils.hpp"
using namespace dnnl;
void prelu_example(dnnl::engine::kind engine_kind) {
// Create execution dnnl::engine.
dnnl::engine engine(engine_kind, 0);
// Create dnnl::stream.
dnnl::stream engine_stream(engine);
// Tensor dimensions.
const memory::dim N = 3, // batch size
IC = 3, // channels
IH = 227, // tensor height
IW = 227; // tensor width
// Source (src), weights and destination (dst) tensors dimensions.
const memory::dims src_dims = {N, IC, IH, IW};
const memory::dims weights_dims = {N, IC, IH, IW};
const memory::dims dst_dims = {N, IC, IH, IW};
// Allocate buffers. In this example, out-of-place primitive execution is
// demonstrated since both src and dst are required for later backward
// propagation.
std::vector<float> src_data(product(src_dims));
std::vector<float> weights_data(product(weights_dims));
std::vector<float> dst_data(product(dst_dims));
// Initialize src tensor.
std::generate(src_data.begin(), src_data.end(), []() {
static int i = 0;
return std::cos(i++ / 10.f);
// Initialize weights tensor.
std::fill(weights_data.begin(), weights_data.end(), 0.3f);
// Create memory objects for tensor data (src, weights, dst). In this
// example, NCHW layout is assumed for src, weights and dst.
auto user_src_mem = memory({src_dims, dt::f32, tag::nchw}, engine);
auto user_weights_mem = memory({weights_dims, dt::f32, tag::nchw}, engine);
auto user_dst_mem = memory({dst_dims, dt::f32, tag::nchw}, engine);
// Create memory descriptors for the primitive. Src tag is set
// to match src memory object. Setting weights tag to format_tag::any
// enables the PReLU primitive to choose memory layout for an optimized
// primitive implementation, and that layout may differ from the one
// provided by the user.
auto src_md = memory::desc(src_dims, dt::f32, tag::nchw);
auto weights_md = memory::desc(weights_dims, dt::f32, tag::any);
// Write data to memory object's handle.
write_to_dnnl_memory(, user_src_mem);
write_to_dnnl_memory(, user_weights_mem);
// Create operation descriptor.
auto prelu_d = prelu_forward::desc(
// Create primitive descriptor.
auto prelu_pd = prelu_forward::primitive_desc(prelu_d, engine);
// For now, assume that the weights memory layout generated
// by the primitive and the one provided by the user are identical.
auto prelu_weights_mem = user_weights_mem;
// Reorder the data in case the weights memory layout generated by
// the primitive and the one provided by the user are different. In this
// case, we create additional memory object with internal buffers that will
// contain the reordered data.
if (prelu_pd.weights_desc() != user_weights_mem.get_desc()) {
prelu_weights_mem = memory(prelu_pd.weights_desc(), engine);
reorder(user_weights_mem, prelu_weights_mem)
.execute(engine_stream, user_weights_mem, prelu_weights_mem);
// Create the primitive.
auto prelu_prim = prelu_forward(prelu_pd);
// Primitive arguments.
std::unordered_map<int, memory> prelu_args;
prelu_args.insert({DNNL_ARG_SRC, user_src_mem});
prelu_args.insert({DNNL_ARG_WEIGHTS, prelu_weights_mem});
prelu_args.insert({DNNL_ARG_DST, user_dst_mem});
// Primitive execution: PReLU.
prelu_prim.execute(engine_stream, prelu_args);
// Wait for the computation to finalize.
// Read data from memory object's handle.
read_from_dnnl_memory(, user_dst_mem);
int main(int argc, char **argv) {
return handle_example_errors(prelu_example, parse_engine_kind(argc, argv));