.. index:: pair: page; Batch Normalization Primitive Example
.. _doxid-batch_normalization_example_cpp:

Batch Normalization Primitive Example
=====================================

This C++ API example demonstrates how to create and execute a :ref:`Batch Normalization <doxid-dev_guide_batch_normalization>` primitive in forward training propagation mode.

Key optimizations included in this example:

* In-place primitive execution;

* Source memory format for an optimized primitive implementation;

* Fused post-ops via operation descriptor flags;

.. ref-code-block:: cpp

	/*******************************************************************************
	* 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
	*
	*     http://www.apache.org/licenses/LICENSE-2.0
	*
	* Unless required by applicable law or agreed to in writing, software
	* distributed under the License is distributed on an "AS IS" BASIS,
	* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	* See the License for the specific language governing permissions and
	* limitations under the License.
	*******************************************************************************/
	
	
	#include <algorithm>
	#include <cmath>
	#include <iostream>
	#include <string>
	#include <vector>
	
	#include "example_utils.hpp"
	#include "oneapi/dnnl/dnnl.hpp"
	
	using namespace :ref:`dnnl <doxid-namespacednnl>`;
	
	using :ref:`tag <doxid-structdnnl_1_1memory_1a8e71077ed6a5f7fb7b3e6e1a5a2ecf3f>` = :ref:`memory::format_tag <doxid-structdnnl_1_1memory_1a8e71077ed6a5f7fb7b3e6e1a5a2ecf3f>`;
	using :ref:`dt <doxid-structdnnl_1_1memory_1a8e83474ec3a50e08e37af76c8c075dce>` = :ref:`memory::data_type <doxid-structdnnl_1_1memory_1a8e83474ec3a50e08e37af76c8c075dce>`;
	
	void batch_normalization_example(:ref:`dnnl::engine::kind <doxid-structdnnl_1_1engine_1a2635da16314dcbdb9bd9ea431316bb1a>` engine_kind) {
	
	    // Create execution dnnl::engine.
	    :ref:`dnnl::engine <doxid-structdnnl_1_1engine>` :ref:`engine <doxid-structdnnl_1_1engine>`(engine_kind, 0);
	
	    // Create dnnl::stream.
	    :ref:`dnnl::stream <doxid-structdnnl_1_1stream>` engine_stream(:ref:`engine <doxid-structdnnl_1_1engine>`);
	
	    // Tensor dimensions.
	    const :ref:`memory::dim <doxid-structdnnl_1_1memory_1a6ad818e4699872cc913474fa5f122cd5>` N = 3, // batch size
	            IC = 3, // channels
	            IH = 227, // tensor height
	            IW = 227; // tensor width
	
	    // Source (src) and destination (dst) tensors dimensions.
	    :ref:`memory::dims <doxid-structdnnl_1_1memory_1afdd20764d58c0b517d5a31276672aeb8>` src_dims = {N, IC, IH, IW};
	
	    // Scale/shift tensor dimensions.
	    :ref:`memory::dims <doxid-structdnnl_1_1memory_1afdd20764d58c0b517d5a31276672aeb8>` scale_shift_dims = {2, IC};
	
	    // Allocate buffers.
	    std::vector<float> src_data(product(src_dims));
	    std::vector<float> scale_shift_data(product(scale_shift_dims));
	
	    // Initialize src.
	    std::generate(src_data.begin(), src_data.end(), []() {
	        static int i = 0;
	        return std::cos(i++ / 10.f);
	    });
	
	    auto mid = scale_shift_data.begin() + IC;
	
	    // Initialize scale.
	    std::generate(scale_shift_data.begin(), mid, []() {
	        static int i = 0;
	        return std::sin(i++ * 2.f);
	    });
	
	    // Initialize shift.
	    std::generate(mid, scale_shift_data.end(), []() {
	        static int i = 0;
	        return std::tan(i++);
	    });
	
	    // Create src and scale/shift memory descriptors and memory objects.
	    auto :ref:`src_md <doxid-group__dnnl__api__primitives__common_1gga94efdd650364f4d9776cfb9b711cbdc1a90a729e395453e1d9411ad416c796819>` = :ref:`memory::desc <doxid-structdnnl_1_1memory_1_1desc>`(src_dims, dt::f32, tag::nchw);
	    auto scale_shift_md = :ref:`memory::desc <doxid-structdnnl_1_1memory_1_1desc>`(scale_shift_dims, dt::f32, tag::nc);
	
	    auto src_mem = :ref:`memory <doxid-structdnnl_1_1memory>`(src_md, :ref:`engine <doxid-structdnnl_1_1engine>`);
	    auto scale_shift_mem = :ref:`memory <doxid-structdnnl_1_1memory>`(scale_shift_md, :ref:`engine <doxid-structdnnl_1_1engine>`);
	
	    // Write data to memory object's handle.
	    write_to_dnnl_memory(src_data.data(), src_mem);
	    write_to_dnnl_memory(scale_shift_data.data(), scale_shift_mem);
	
	    // Create operation descriptor.
	    auto bnorm_d = :ref:`batch_normalization_forward::desc <doxid-structdnnl_1_1batch__normalization__forward_1_1desc>`(
	            :ref:`prop_kind::forward_training <doxid-group__dnnl__api__attributes_1ggac7db48f6583aa9903e54c2a39d65438fa24775787fab8f13aa4809e1ce8f82aeb>`, src_md, 1.e-10f,
	            :ref:`normalization_flags::use_scale_shift <doxid-group__dnnl__api__primitives__common_1ggad8ef0fcbb7b10cae3d67dd46892002bea0cbad8a346551ec47da7ff1afb1e1f76>`
	                    | :ref:`normalization_flags::fuse_norm_relu <doxid-group__dnnl__api__primitives__common_1ggad8ef0fcbb7b10cae3d67dd46892002bea898ce555425ee54271096bc9c8e0400c>`);
	
	    // Create primitive descriptor.
	    auto bnorm_pd
	            = :ref:`batch_normalization_forward::primitive_desc <doxid-structdnnl_1_1batch__normalization__forward_1_1primitive__desc>`(bnorm_d, :ref:`engine <doxid-structdnnl_1_1engine>`);
	
	    // Create memory objects using memory descriptors created by the primitive
	    // descriptor: mean, variance, workspace.
	    // NOTE: Here, the ReLU post-ops require a workspace for later usage in
	    // backward propagation mode.
	    auto mean_mem = :ref:`memory <doxid-structdnnl_1_1memory>`(bnorm_pd.mean_desc(), :ref:`engine <doxid-structdnnl_1_1engine>`);
	    auto variance_mem = :ref:`memory <doxid-structdnnl_1_1memory>`(bnorm_pd.variance_desc(), :ref:`engine <doxid-structdnnl_1_1engine>`);
	    auto workspace_mem = :ref:`memory <doxid-structdnnl_1_1memory>`(bnorm_pd.workspace_desc(), :ref:`engine <doxid-structdnnl_1_1engine>`);
	
	    // Create the primitive.
	    auto bnorm_prim = :ref:`batch_normalization_forward <doxid-structdnnl_1_1batch__normalization__forward>`(bnorm_pd);
	
	    // Primitive arguments. Set up in-place execution by assigning src as DST.
	    std::unordered_map<int, memory> bnorm_args;
	    bnorm_args.insert({:ref:`DNNL_ARG_SRC <doxid-group__dnnl__api__primitives__common_1gac37ad67b48edeb9e742af0e50b70fe09>`, src_mem});
	    bnorm_args.insert({:ref:`DNNL_ARG_MEAN <doxid-group__dnnl__api__primitives__common_1ga9bcff7f442a5d6a0ac1183533e721066>`, mean_mem});
	    bnorm_args.insert({:ref:`DNNL_ARG_VARIANCE <doxid-group__dnnl__api__primitives__common_1gaa0e60e8d129936ba29555e17efb82581>`, variance_mem});
	    bnorm_args.insert({:ref:`DNNL_ARG_SCALE_SHIFT <doxid-group__dnnl__api__primitives__common_1ga6d529c1a672c080d4b13f6929e915de4>`, scale_shift_mem});
	    bnorm_args.insert({:ref:`DNNL_ARG_WORKSPACE <doxid-group__dnnl__api__primitives__common_1ga550c80e1b9ba4f541202a7ac98be117f>`, workspace_mem});
	    bnorm_args.insert({:ref:`DNNL_ARG_DST <doxid-group__dnnl__api__primitives__common_1ga3ca217e4a06d42a0ede3c018383c388f>`, src_mem});
	
	    // Primitive execution: batch normalization with ReLU.
	    bnorm_prim.execute(engine_stream, bnorm_args);
	
	    // Wait for the computation to finalize.
	    engine_stream.wait();
	
	    // Read data from memory object's handle.
	    read_from_dnnl_memory(src_data.data(), src_mem);
	}
	
	int main(int argc, char **argv) {
	    return handle_example_errors(
	            batch_normalization_example, parse_engine_kind(argc, argv));
	}