.. index:: pair: example; bnorm_u8_via_binary_postops.cpp .. _doxid-bnorm_u8_via_binary_postops_8cpp-example: bnorm_u8_via_binary_postops.cpp =============================== The example implements the Batch normalization u8 via the following operations: binary_sub(src, mean), binary_div(tmp_dst, variance), binary_mul(tmp_dst, scale), binary_add(tmp_dst, shift). Annotated version: :ref:`Bnorm u8 by binary post-ops example ` The example implements the Batch normalization u8 via the following operations: binary_sub(src, mean), binary_div(tmp_dst, variance), binary_mul(tmp_dst, scale), binary_add(tmp_dst, shift). Annotated version: :ref:`Bnorm u8 by binary post-ops example ` .. 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 #include #include #include #include #include "dnnl.hpp" #include "example_utils.hpp" using namespace :ref:`dnnl `; using :ref:`tag ` = :ref:`memory::format_tag `; using :ref:`dt ` = :ref:`memory::data_type `; void bnorm_u8_via_binary_postops(:ref:`dnnl::engine::kind ` engine_kind) { // Create execution dnnl::engine. :ref:`dnnl::engine ` :ref:`engine `(engine_kind, 0); // Create dnnl::stream. :ref:`dnnl::stream ` engine_stream(:ref:`engine `); // Tensor dimensions. const :ref:`memory::dim ` N = 3, // batch size IC = 3, // channels IH = 150, // tensor height IW = 150; // tensor width // Tensors dimensions. :ref:`memory::dims ` src_dims = {N, IC, IH, IW}; :ref:`memory::dims ` params_dims = {1, IC, 1, 1}; // Allocate buffers. std::vector src_data(product(src_dims)); std::vector mean_data(product(params_dims)); std::vector variance_data(product(params_dims)); std::vector scale_data(product(params_dims)); std::vector shift_data(product(params_dims)); std::vector oscale_data(product(params_dims)); // Initialize std::generate(src_data.begin(), src_data.end(), []() { static int i = 0; return std::cos(i++ / 10.f); }); std::generate(mean_data.begin(), mean_data.end(), []() { static int i = 0; return std::sin(i++ * 2.f); }); std::generate(variance_data.begin(), variance_data.end(), []() { static int i = 0; return std::sin(i++ * 4.f); }); std::generate(scale_data.begin(), scale_data.end(), []() { static int i = 0; return std::sin(i++ * 6.f); }); std::generate(shift_data.begin(), shift_data.end(), []() { static int i = 0; return std::sin(i++ * 8.f); }); std::generate(oscale_data.begin(), oscale_data.end(), []() { return 0.5; }); // Create descriptors. auto :ref:`src_md ` = :ref:`memory::desc `(src_dims, dt::u8, tag::nhwc); auto mean_md = :ref:`memory::desc `(params_dims, dt::f32, tag::nhwc); auto variance_md = :ref:`memory::desc `(params_dims, dt::f32, tag::nhwc); auto scale_md = :ref:`memory::desc `(params_dims, dt::f32, tag::nhwc); auto shift_md = :ref:`memory::desc `(params_dims, dt::f32, tag::nhwc); auto oscale_md = :ref:`memory::desc `(params_dims, dt::f32, tag::nhwc); // Create src memory objects. auto src_mem = :ref:`memory `(src_md, :ref:`engine `); auto mean_mem = :ref:`memory `(mean_md, :ref:`engine `); auto variance_mem = :ref:`memory `(variance_md, :ref:`engine `); auto scale_mem = :ref:`memory `(scale_md, :ref:`engine `); auto shift_mem = :ref:`memory `(shift_md, :ref:`engine `); auto oscale_mem = :ref:`memory `(oscale_md, :ref:`engine `); // Write data to memory object's handle. write_to_dnnl_memory(src_data.data(), src_mem); write_to_dnnl_memory(mean_data.data(), mean_mem); write_to_dnnl_memory(variance_data.data(), variance_mem); write_to_dnnl_memory(scale_data.data(), scale_mem); write_to_dnnl_memory(shift_data.data(), shift_mem); write_to_dnnl_memory(oscale_data.data(), oscale_mem); // Create operation descriptor. // dst_tmp = src - mean auto :ref:`binary_d ` = :ref:`binary::desc `(:ref:`algorithm::binary_sub `, src_md, mean_md, src_md); // Bnorm operation with scale and shift :ref:`post_ops ` binary_ops; // dst_tmp = dst_tmp / variance binary_ops.:ref:`append_binary `(:ref:`algorithm::binary_div `, variance_md); // dst_tmp = dst_tmp * scale binary_ops.:ref:`append_binary `(:ref:`algorithm::binary_mul `, scale_md); // dst_tmp = dst_tmp + shift binary_ops.:ref:`append_binary `(:ref:`algorithm::binary_add `, shift_md); // dst = dst_tmp * output_scale (only for re-quantization) binary_ops.:ref:`append_binary `(:ref:`algorithm::binary_mul `, oscale_md); primitive_attr binary_attr; binary_attr.set_post_ops(binary_ops); // Create primitive descriptor. auto binary_pd = :ref:`binary::primitive_desc `(binary_d, binary_attr, :ref:`engine `); // Create the primitive. auto binary_prim = :ref:`binary `(binary_pd); // Primitive arguments. std::unordered_map binary_args; binary_args.insert({:ref:`DNNL_ARG_SRC_0 `, src_mem}); binary_args.insert({:ref:`DNNL_ARG_SRC_1 `, mean_mem}); // In-place mode (dst is src) binary_args.insert({:ref:`DNNL_ARG_DST `, src_mem}); binary_args.insert( {:ref:`DNNL_ARG_ATTR_MULTIPLE_POST_OP `(0) | :ref:`DNNL_ARG_SRC_1 `, variance_mem}); binary_args.insert( {:ref:`DNNL_ARG_ATTR_MULTIPLE_POST_OP `(1) | :ref:`DNNL_ARG_SRC_1 `, scale_mem}); binary_args.insert( {:ref:`DNNL_ARG_ATTR_MULTIPLE_POST_OP `(2) | :ref:`DNNL_ARG_SRC_1 `, shift_mem}); binary_args.insert( {:ref:`DNNL_ARG_ATTR_MULTIPLE_POST_OP `(3) | :ref:`DNNL_ARG_SRC_1 `, oscale_mem}); // Primitive execution binary_prim.execute(engine_stream, binary_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( bnorm_u8_via_binary_postops, parse_engine_kind(argc, argv)); }