oneDNN/tests/api.c

/*******************************************************************************
* Copyright 2017-2023 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 <math.h>
#include <stddef.h>
#include <stdio.h>
#include <stdlib.h>

#include "oneapi/dnnl/dnnl.h"

#define for_ for

#define CHECK(f) \
    do { \
        dnnl_status_t s = f; \
        if (s != dnnl_success) { \
            printf("[%s:%d] error: %s returns %d\n", __FILE__, __LINE__, #f, \
                    s); \
            exit(2); \
        } \
    } while (0)

#define CHECK_TRUE(expr) \
    do { \
        int e_ = expr; \
        if (!e_) { \
            printf("[%s:%d] %s failed\n", __FILE__, __LINE__, #expr); \
            exit(2); \
        } \
    } while (0)

typedef float real_t;

#define LENGTH_100 100

void test1() {
    dnnl_engine_t engine;
    CHECK(dnnl_engine_create(&engine, dnnl_cpu, 0));

    dnnl_dims_t dims = {LENGTH_100};
    real_t data[LENGTH_100];

    dnnl_memory_desc_t md;
    const_dnnl_memory_desc_t c_md_tmp;
    dnnl_memory_t m;

    CHECK(dnnl_memory_desc_create_with_tag(&md, 1, dims, dnnl_f32, dnnl_x));
    CHECK(dnnl_memory_create(&m, md, engine, NULL));

    void *req = NULL;

    CHECK(dnnl_memory_get_data_handle(m, &req));
    CHECK_TRUE(req == NULL);

#if DNNL_CPU_RUNTIME != DNNL_RUNTIME_SYCL
    CHECK(dnnl_memory_set_data_handle(m, data));
    CHECK(dnnl_memory_get_data_handle(m, &req));
    CHECK_TRUE(req == data);
#endif

    CHECK_TRUE(dnnl_memory_desc_get_size(md) == LENGTH_100 * sizeof(data[0]));

    CHECK(dnnl_memory_get_memory_desc(m, &c_md_tmp));
    CHECK_TRUE(dnnl_memory_desc_equal(md, c_md_tmp));

    CHECK(dnnl_memory_destroy(m));
    CHECK(dnnl_memory_desc_destroy(md));

    CHECK(dnnl_engine_destroy(engine));
}

#if DNNL_CPU_RUNTIME != DNNL_RUNTIME_SYCL

static size_t product(dnnl_dim_t *arr, size_t size) {
    size_t prod = 1;
    for (size_t i = 0; i < size; ++i)
        prod *= arr[i];
    return prod;
}

void test2() {
    /* AlexNet: c3
     * {2, 256, 13, 13} (x) {384, 256, 3, 3} -> {2, 384, 13, 13}
     * pad: {1, 1}
     * strides: {1, 1}
     */

    const dnnl_dim_t mb = 2;
    const dnnl_dim_t groups = 2;
    const int ndims = 4;
    dnnl_dims_t c3_src_sizes = {mb, 256, 13, 13};
    dnnl_dims_t c3_weights_sizes = {groups, 384 / groups, 256 / groups, 3, 3};
    dnnl_dims_t c3_bias_sizes = {384};
    dnnl_dims_t strides = {1, 1};
    dnnl_dims_t dilation = {0, 0};
    dnnl_dims_t padding = {0, 0}; // set proper values
    dnnl_dims_t c3_dst_sizes = {mb, 384,
            (c3_src_sizes[2] + 2 * padding[0] - c3_weights_sizes[3])
                            / strides[0]
                    + 1,
            (c3_src_sizes[3] + 2 * padding[1] - c3_weights_sizes[4])
                            / strides[1]
                    + 1};

    real_t *src
            = (real_t *)calloc(product(c3_src_sizes, ndims), sizeof(real_t));
    real_t *weights = (real_t *)calloc(
            product(c3_weights_sizes, ndims + 1), sizeof(real_t));
    real_t *bias = (real_t *)calloc(product(c3_bias_sizes, 1), sizeof(real_t));
    real_t *dst
            = (real_t *)calloc(product(c3_dst_sizes, ndims), sizeof(real_t));
    real_t *out_mem
            = (real_t *)calloc(product(c3_dst_sizes, ndims), sizeof(real_t));
    CHECK_TRUE(src && weights && bias && dst && out_mem);

    for (dnnl_dim_t i = 0; i < c3_bias_sizes[0]; ++i)
        bias[i] = i;

    dnnl_engine_t engine;
    CHECK(dnnl_engine_create(&engine, dnnl_cpu, 0));

    dnnl_stream_t stream;
    CHECK(dnnl_stream_create(&stream, engine, dnnl_stream_default_flags));

    /* first describe user data and create data descriptors for future
     * convolution w/ the specified format -- we do not want to do a reorder */
    dnnl_memory_desc_t c3_src_md, c3_weights_md, c3_bias_md, c3_dst_md, out_md;
    dnnl_memory_t c3_src, c3_weights, c3_bias, c3_dst, out;

    // src
    {
        CHECK(dnnl_memory_desc_create_with_tag(
                &c3_src_md, 4, c3_src_sizes, dnnl_f32, dnnl_nChw8c));
        CHECK(dnnl_memory_create(&c3_src, c3_src_md, engine, src));
    }

    // weights
    {
        CHECK(dnnl_memory_desc_create_with_tag(&c3_weights_md,
                4 + (groups != 1), c3_weights_sizes + (groups == 1), dnnl_f32,
                groups == 1 ? dnnl_OIhw8i8o : dnnl_gOIhw8i8o));
        CHECK(dnnl_memory_create(&c3_weights, c3_weights_md, engine, weights));
    }

    // bias
    {
        CHECK(dnnl_memory_desc_create_with_tag(
                &c3_bias_md, 1, c3_bias_sizes, dnnl_f32, dnnl_x));
        CHECK(dnnl_memory_create(&c3_bias, c3_bias_md, engine, bias));
    }

    // c3_dst
    {
        CHECK(dnnl_memory_desc_create_with_tag(
                &c3_dst_md, 4, c3_dst_sizes, dnnl_f32, dnnl_nChw8c));
        CHECK(dnnl_memory_create(&c3_dst, c3_dst_md, engine, dst));
    }

    // out
    {
        CHECK(dnnl_memory_desc_create_with_tag(
                &out_md, 4, c3_dst_sizes, dnnl_f32, dnnl_nchw));
        CHECK(dnnl_memory_create(&out, out_md, engine, out_mem));
    }

    /* create a convolution primitive descriptor */
    dnnl_primitive_desc_t c3_pd;
    dnnl_primitive_t c3;

    CHECK(dnnl_convolution_forward_primitive_desc_create(&c3_pd, engine,
            dnnl_forward_training, dnnl_convolution_direct, c3_src_md,
            c3_weights_md, c3_bias_md, c3_dst_md, strides, dilation, padding,
            NULL, NULL));

    CHECK_TRUE(dnnl_memory_desc_equal(c3_src_md,
            dnnl_primitive_desc_query_md(c3_pd, dnnl_query_src_md, 0)));
    CHECK_TRUE(dnnl_memory_desc_equal(c3_weights_md,
            dnnl_primitive_desc_query_md(c3_pd, dnnl_query_weights_md, 0)));
    CHECK_TRUE(dnnl_memory_desc_equal(c3_bias_md,
            dnnl_primitive_desc_query_md(c3_pd, dnnl_query_weights_md, 1)));
    CHECK_TRUE(dnnl_memory_desc_equal(c3_dst_md,
            dnnl_primitive_desc_query_md(c3_pd, dnnl_query_dst_md, 0)));

    CHECK_TRUE(dnnl_memory_desc_equal(c3_src_md,
            dnnl_primitive_desc_query_md(
                    c3_pd, dnnl_query_exec_arg_md, DNNL_ARG_SRC)));
    CHECK_TRUE(dnnl_memory_desc_equal(c3_weights_md,
            dnnl_primitive_desc_query_md(
                    c3_pd, dnnl_query_exec_arg_md, DNNL_ARG_WEIGHTS)));
    CHECK_TRUE(dnnl_memory_desc_equal(c3_bias_md,
            dnnl_primitive_desc_query_md(
                    c3_pd, dnnl_query_exec_arg_md, DNNL_ARG_BIAS)));
    CHECK_TRUE(dnnl_memory_desc_equal(c3_dst_md,
            dnnl_primitive_desc_query_md(
                    c3_pd, dnnl_query_exec_arg_md, DNNL_ARG_DST)));

    /* create a convolution and execute it */
    CHECK(dnnl_primitive_create(&c3, c3_pd));
    CHECK(dnnl_primitive_desc_destroy(c3_pd));

    dnnl_exec_arg_t c3_args[4] = {
            {DNNL_ARG_SRC, c3_src},
            {DNNL_ARG_WEIGHTS, c3_weights},
            {DNNL_ARG_BIAS, c3_bias},
            {DNNL_ARG_DST, c3_dst},
    };
    CHECK(dnnl_primitive_execute(c3, stream, 4, c3_args));
    CHECK(dnnl_primitive_destroy(c3));

    /* create a reorder primitive descriptor */
    dnnl_primitive_desc_t r_pd;
    CHECK(dnnl_reorder_primitive_desc_create(
            &r_pd, c3_dst_md, engine, out_md, engine, NULL));

    /* create a reorder and execute it */
    dnnl_primitive_t r;
    CHECK(dnnl_primitive_create(&r, r_pd));
    CHECK(dnnl_primitive_desc_destroy(r_pd));

    dnnl_exec_arg_t r_args[2] = {
            {DNNL_ARG_FROM, c3_dst},
            {DNNL_ARG_TO, out},
    };
    CHECK(dnnl_primitive_execute(r, stream, 2, r_args));
    CHECK(dnnl_primitive_destroy(r));

    CHECK(dnnl_stream_wait(stream));

    /* clean-up */
    CHECK(dnnl_memory_destroy(c3_src));
    CHECK(dnnl_memory_destroy(c3_weights));
    CHECK(dnnl_memory_destroy(c3_bias));
    CHECK(dnnl_memory_destroy(c3_dst));
    CHECK(dnnl_memory_destroy(out));
    CHECK(dnnl_stream_destroy(stream));
    CHECK(dnnl_engine_destroy(engine));

    CHECK(dnnl_memory_desc_destroy(c3_src_md));
    CHECK(dnnl_memory_desc_destroy(c3_weights_md));
    CHECK(dnnl_memory_desc_destroy(c3_bias_md));
    CHECK(dnnl_memory_desc_destroy(c3_dst_md));
    CHECK(dnnl_memory_desc_destroy(out_md));

    const dnnl_dim_t N = c3_dst_sizes[0], C = c3_dst_sizes[1],
                     H = c3_dst_sizes[2], W = c3_dst_sizes[3];
    for_(dnnl_dim_t n = 0; n < N; ++n)
    for_(dnnl_dim_t c = 0; c < C; ++c)
    for_(dnnl_dim_t h = 0; h < H; ++h)
    for (dnnl_dim_t w = 0; w < W; ++w) {
        dnnl_dim_t off = ((n * C + c) * H + h) * W + w;
        CHECK_TRUE(out_mem[off] == bias[c]);
    }

    free(src);
    free(weights);
    free(bias);
    free(dst);
    free(out_mem);
}

void test3() {
    const dnnl_dim_t mb = 2;
    const int ndims = 4;
    dnnl_dims_t l2_data_sizes = {mb, 256, 13, 13};

    real_t *src
            = (real_t *)calloc(product(l2_data_sizes, ndims), sizeof(real_t));
    real_t *dst
            = (real_t *)calloc(product(l2_data_sizes, ndims), sizeof(real_t));
    CHECK_TRUE(src && dst);

    for (size_t i = 0; i < product(l2_data_sizes, ndims); ++i)
        src[i] = (i % 13) + 1;

    dnnl_engine_t engine;
    CHECK(dnnl_engine_create(&engine, dnnl_cpu, 0));

    dnnl_stream_t stream;
    CHECK(dnnl_stream_create(&stream, engine, dnnl_stream_default_flags));

    dnnl_memory_desc_t l2_data_md;
    dnnl_memory_t l2_src, l2_dst;

    // src, dst
    {
        CHECK(dnnl_memory_desc_create_with_tag(
                &l2_data_md, ndims, l2_data_sizes, dnnl_f32, dnnl_nchw));
        CHECK(dnnl_memory_create(&l2_src, l2_data_md, engine, src));
        CHECK(dnnl_memory_create(&l2_dst, l2_data_md, engine, dst));
    }

    /* create an lrn */
    dnnl_primitive_desc_t l2_pd;
    dnnl_primitive_t l2;

    CHECK(dnnl_lrn_forward_primitive_desc_create(&l2_pd, engine,
            dnnl_forward_inference, dnnl_lrn_across_channels, l2_data_md,
            l2_data_md, 5, 1e-4f, 0.75f, 1.0f, NULL));

    CHECK_TRUE(dnnl_memory_desc_equal(l2_data_md,
            dnnl_primitive_desc_query_md(l2_pd, dnnl_query_src_md, 0)));
    CHECK_TRUE(dnnl_memory_desc_equal(l2_data_md,
            dnnl_primitive_desc_query_md(l2_pd, dnnl_query_dst_md, 0)));
    CHECK_TRUE(dnnl_primitive_desc_query_s32(
                       l2_pd, dnnl_query_num_of_inputs_s32, 0)
            == 1);
    CHECK_TRUE(dnnl_primitive_desc_query_s32(
                       l2_pd, dnnl_query_num_of_outputs_s32, 0)
            == 1);

    CHECK(dnnl_primitive_create(&l2, l2_pd));
    CHECK(dnnl_primitive_desc_destroy(l2_pd));

    dnnl_exec_arg_t l2_args[2] = {
            {DNNL_ARG_SRC, l2_src},
            {DNNL_ARG_DST, l2_dst},
    };
    CHECK(dnnl_primitive_execute(l2, stream, 2, l2_args));
    CHECK(dnnl_primitive_destroy(l2));

    CHECK(dnnl_stream_wait(stream));

    /* clean-up */
    CHECK(dnnl_memory_destroy(l2_src));
    CHECK(dnnl_memory_destroy(l2_dst));
    CHECK(dnnl_stream_destroy(stream));
    CHECK(dnnl_engine_destroy(engine));

    CHECK(dnnl_memory_desc_destroy(l2_data_md));

    const dnnl_dim_t N = l2_data_sizes[0], C = l2_data_sizes[1],
                     H = l2_data_sizes[2], W = l2_data_sizes[3];
    for_(dnnl_dim_t n = 0; n < N; ++n)
    for_(dnnl_dim_t c = 0; c < C; ++c)
    for_(dnnl_dim_t h = 0; h < H; ++h)
    for (dnnl_dim_t w = 0; w < W; ++w) {
        size_t off = ((n * C + c) * H + h) * W + w;
        real_t e = (off % 13) + 1;
        real_t diff = (real_t)fabs(dst[off] - e);
        if (diff / fabs(e) > 0.0125) printf("exp: %g, got: %g\n", e, dst[off]);
        CHECK_TRUE(diff / fabs(e) < 0.0125);
    }

    free(src);
    free(dst);
}
#endif

int main() {
    test1();
#if DNNL_CPU_RUNTIME != DNNL_RUNTIME_SYCL
    test2();
    test3();
#endif
    return 0;
}