MegEngine/dnn/test/cpu/matrix_mul_int_8x8x16.cpp

/**
 * \file dnn/test/cpu/matrix_mul_int_8x8x16.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 */
#include "test/cpu/fixture.h"

#include "test/common/benchmarker.h"
#include "test/common/checker.h"
#include "test/common/convolution.h"

namespace megdnn {
namespace test {

TEST_F(CPU, MATRIX_MUL_INT_8_8_16) {
    Checker<MatrixMul> checker(handle());
    param::MatrixMul param;
    checker.set_dtype(0, dtype::Int8());
    checker.set_dtype(1, dtype::Int8());
    checker.set_dtype(2, dtype::Int16());
    checker.set_param(param);
    for (size_t b : {1, 2, 3})
        for (size_t i : {10, 20})
            for (size_t o : {11, 22}) {
                checker.exec({{b, i}, {i, o}, {}});
            }
    for (size_t m = 16; m <= 512; m *= 4)
        for (size_t n = 16; n <= 512; n *= 4)
            for (size_t k = 16; k <= 512; k *= 4) {
                checker.exec({{m, k}, {k, n}, {}});

                checker.exec({{m + 1, k}, {k, n}, {}});
                checker.exec({{m + 5, k}, {k, n}, {}});
                checker.exec({{m + 7, k}, {k, n}, {}});

                checker.exec({{m, k}, {k, n + 15}, {}});
                checker.exec({{m, k}, {k, n + 9}, {}});
                checker.exec({{m, k}, {k, n + 8}, {}});
                checker.exec({{m, k}, {k, n + 7}, {}});
                checker.exec({{m, k}, {k, n + 1}, {}});

                checker.exec({{m + 1, k}, {k, n + 9}, {}});
                checker.exec({{m + 7, k}, {k, n + 15}, {}});
                checker.exec({{m + 7, k}, {k, n + 7}, {}});
            }
    // test transpose scenerio
    {
        for (int mask = 0; mask < 4; ++mask) {
            param::MatrixMul param;
            param.transposeA = (mask & 1);
            param.transposeB = (mask & 2);
            checker.set_param(param);
            size_t m = 100, n = 101, k = 102;
            TensorShape A =
                    param.transposeA ? TensorShape({k, m}) : TensorShape({m, k});
            TensorShape B =
                    param.transposeB ? TensorShape({n, k}) : TensorShape({k, n});
            checker.exec({A, B, {}});
        }
    }
}

#if MEGDNN_WITH_BENCHMARK

TEST_F(CPU, BENCHMARK_MATRIX_MUL_INT8_INT8_INT16) {
    bool verbose = getenv("MEGDNN_BENCH_VERBOSE");
    using Param = param::MatrixMul;
    double speedup_sum = 0, speedup_wsum = 0;
    auto run = [&](const TensorShapeArray& shapes, const Param& param) {
        TensorLayoutArray layouts;
        layouts.emplace_back(shapes[0], dtype::Int8());
        layouts.emplace_back(shapes[1], dtype::Int8());
        layouts.emplace_back(shapes[2], dtype::Int16());
        Benchmarker<MatrixMul> benchmarker_cpu(handle());
        param::MatrixMul param_int(param);
        benchmarker_cpu.set_param(param_int);
        Benchmarker<MatrixMul> benchmarker_float(handle());
        benchmarker_float.set_param(param);
        auto t2 = benchmarker_cpu.set_display(false).set_adaptive_benchmark(0.01).execl(
                layouts);
        auto t4 =
                benchmarker_float.set_display(false).set_adaptive_benchmark(0.01).exec(
                        shapes);
        if (t2 > t4 || verbose) {
            std::cout << "MatA=" << shapes[0].to_string()
                      << " MatB=" << shapes[1].to_string() << " float=" << t4 << "ms"
                      << " int=" << t2 << "ms"
                      << " speedup=" << t4 / t2 << std::endl;
        }
        speedup_sum += t4 / t2;
        speedup_wsum += 1;
    };
    for (size_t m = 16; m <= 256; m *= 4)
        for (size_t k = 16; k <= 256; k *= 4)
            for (size_t n = 16; n <= 1024; n *= 4) {
                Param param;
                run({{m, k}, {k, n}, {}}, param);
                run({{m, k}, {k, n + 8}, {}}, param);
                run({{m, k}, {k, n + 15}, {}}, param);

                run({{m + 5, k}, {k, n}, {}}, param);
                run({{m + 7, k}, {k, n}, {}}, param);
            }
    printf("average speedup: %.3f\n", speedup_sum / speedup_wsum);
}

#endif

}  // namespace test
}  // namespace megdnn
// vim: syntax=cpp.doxygen