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).
#include <algorithm>
#include <cmath>
#include <iostream>
#include <string>
#include <vector>
#include "dnnl.hpp"
#include "example_utils.hpp"
IC = 3,
IH = 150,
IW = 150;
std::vector<float> src_data(product(src_dims));
std::vector<float> mean_data(product(params_dims));
std::vector<float> variance_data(product(params_dims));
std::vector<float> scale_data(product(params_dims));
std::vector<float> shift_data(product(params_dims));
std::vector<float> oscale_data(product(params_dims));
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; });
auto src_md = memory::desc(src_dims, dt::u8, tag::nhwc);
auto mean_md = memory::desc(params_dims, dt::f32, tag::nhwc);
auto variance_md = memory::desc(params_dims, dt::f32, tag::nhwc);
auto scale_md = memory::desc(params_dims, dt::f32, tag::nhwc);
auto shift_md = memory::desc(params_dims, dt::f32, tag::nhwc);
auto oscale_md = memory::desc(params_dims, dt::f32, tag::nhwc);
auto src_mem = memory(src_md, engine);
auto mean_mem = memory(mean_md, engine);
auto variance_mem = memory(variance_md, engine);
auto scale_mem = memory(scale_md, engine);
auto shift_mem = memory(shift_md, engine);
auto oscale_mem = memory(oscale_md, engine);
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);
post_ops binary_ops;
primitive_attr binary_attr;
binary_attr.set_post_ops(binary_ops);
auto binary_pd = binary::primitive_desc(binary_d, binary_attr, engine);
auto binary_prim = binary(binary_pd);
std::unordered_map<int, memory> binary_args;
binary_args.insert(
binary_args.insert(
binary_args.insert(
binary_args.insert(
binary_prim.execute(engine_stream, binary_args);
engine_stream.wait();
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));
}
#define DNNL_ARG_ATTR_MULTIPLE_POST_OP(idx)
Arguments for a binary post operation.
Definition: dnnl_types.h:2592
#define DNNL_ARG_DST
A special mnemonic for destination argument for primitives that have a single destination.
Definition: dnnl_types.h:2422
#define DNNL_ARG_SRC_0
Source argument #0.
Definition: dnnl_types.h:2395
#define DNNL_ARG_SRC_1
Source argument #1.
Definition: dnnl_types.h:2407
@ src_md
source memory desc
@ binary_d
binary descriptor
oneDNN namespace
Definition: dnnl.hpp:74
An execution engine.
Definition: dnnl.hpp:895
kind
Kinds of engines.
Definition: dnnl.hpp:900
dnnl_dim_t dim
Integer type for representing dimension sizes and indices.
Definition: dnnl.hpp:1138
format_tag
Memory format tag specification.
Definition: dnnl.hpp:1237
data_type
Data type specification.
Definition: dnnl.hpp:1156
std::vector< dim > dims
Vector of dimensions.
Definition: dnnl.hpp:1141
An execution stream.
Definition: dnnl.hpp:1011