Deep Neural Network Library (DNNL)  1.3.0
Performance library for Deep Learning
Reorder between CPU and GPU engines

This C API example demonstrates programming flow when reordering memory between CPU and GPU engines.

/*******************************************************************************
* Copyright 2019-2020 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 <stdio.h>
#include <stdlib.h>
#include "dnnl.h"
#include "example_utils.h"
size_t product(int n_dims, const dnnl_dim_t dims[]) {
size_t n_elems = 1;
for (int d = 0; d < n_dims; ++d) {
n_elems *= (size_t)dims[d];
}
return n_elems;
}
void fill(dnnl_memory_t mem, int n_dims, const dnnl_dim_t dims[]) {
const size_t n_elems = product(n_dims, dims);
float *array = (float *)malloc(n_elems * sizeof(float));
for (size_t e = 0; e < n_elems; ++e) {
array[e] = e % 7 ? 1.0f : -1.0f;
}
write_to_dnnl_memory(array, mem);
free(array);
}
int find_negative(dnnl_memory_t mem, int n_dims, const dnnl_dim_t dims[]) {
const size_t n_elems = product(n_dims, dims);
float *array = (float *)malloc(n_elems * sizeof(float));
read_from_dnnl_memory(array, mem);
int negs = 0;
for (size_t e = 0; e < n_elems; ++e) {
negs += array[e] < 0.0f;
}
free(array);
return negs;
}
void cross_engine_reorder() {
dnnl_engine_t engine_cpu, engine_gpu;
CHECK(dnnl_engine_create(&engine_cpu, dnnl_cpu, 0));
CHECK(dnnl_engine_create(&engine_gpu, dnnl_gpu, 0));
dnnl_dim_t tz[4] = {2, 16, 1, 1};
dnnl_memory_desc_t m_cpu_md, m_gpu_md;
CHECK(dnnl_memory_desc_init_by_tag(&m_cpu_md, 4, tz, dnnl_f32, dnnl_nchw));
CHECK(dnnl_memory_desc_init_by_tag(&m_gpu_md, 4, tz, dnnl_f32, dnnl_nchw));
dnnl_memory_t m_cpu, m_gpu;
&m_cpu, &m_cpu_md, engine_cpu, DNNL_MEMORY_ALLOCATE));
&m_gpu, &m_gpu_md, engine_gpu, DNNL_MEMORY_ALLOCATE));
fill(m_cpu, 4, tz);
if (find_negative(m_cpu, 4, tz) == 0)
COMPLAIN_EXAMPLE_ERROR_AND_EXIT(
"%s", "incorrect data fill, no negative values found");
/* reorder cpu -> gpu */
&r1_pd, &m_cpu_md, engine_cpu, &m_gpu_md, engine_gpu, NULL));
CHECK(dnnl_primitive_create(&r1, r1_pd));
/* relu gpu */
&relu_d, dnnl_forward, dnnl_eltwise_relu, &m_gpu_md, 0.0f, 0.0f));
&relu_pd, &relu_d, NULL, engine_gpu, NULL));
CHECK(dnnl_primitive_create(&relu, relu_pd));
/* reorder gpu -> cpu */
&r2_pd, &m_gpu_md, engine_gpu, &m_cpu_md, engine_cpu, NULL));
CHECK(dnnl_primitive_create(&r2, r2_pd));
dnnl_stream_t stream_gpu;
&stream_gpu, engine_gpu, dnnl_stream_default_flags));
dnnl_exec_arg_t r1_args[] = {{DNNL_ARG_FROM, m_cpu}, {DNNL_ARG_TO, m_gpu}};
CHECK(dnnl_primitive_execute(r1, stream_gpu, 2, r1_args));
dnnl_exec_arg_t relu_args[]
= {{DNNL_ARG_SRC, m_gpu}, {DNNL_ARG_DST, m_gpu}};
CHECK(dnnl_primitive_execute(relu, stream_gpu, 2, relu_args));
dnnl_exec_arg_t r2_args[] = {{DNNL_ARG_FROM, m_gpu}, {DNNL_ARG_TO, m_cpu}};
CHECK(dnnl_primitive_execute(r2, stream_gpu, 2, r2_args));
CHECK(dnnl_stream_wait(stream_gpu));
if (find_negative(m_cpu, 4, tz) != 0)
COMPLAIN_EXAMPLE_ERROR_AND_EXIT(
"%s", "found negative values after ReLU applied");
/* clean up */
dnnl_stream_destroy(stream_gpu);
dnnl_engine_destroy(engine_cpu);
dnnl_engine_destroy(engine_gpu);
}
int main() {
cross_engine_reorder();
printf("Example passed on CPU/GPU.\n");
return 0;
}