MegEngine/dnn/test/rocm/indexing_multi_axis_vec.cpp

/**
 * \file dnn/test/rocm/indexing_multi_axis_vec.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 "hcc_detail/hcc_defs_prologue.h"
#include "test/rocm/benchmarker.h"
#include "test/rocm/fixture.h"

#include "megdnn/oprs.h"
#include "test/common/checker.h"
#include "test/common/index.h"
#include "test/common/indexing_multi_axis_vec.h"

#include <random>

using namespace megdnn;
using namespace test;

namespace {

class OrderedRNG final : public RNG {
public:
    void gen(const TensorND& tensor) override {
        auto span = tensor.layout.span();
        if (tensor.layout.dtype == dtype::Float32()) {
            auto ptr = tensor.ptr<float>() + span.low_elem;
            for (size_t i = 0, it = span.dist_elem(); i < it; ++i) {
                ptr[i] = i;
            }
        } else {
            auto ptr = tensor.ptr<int>() + span.low_elem;
            for (size_t i = 0, it = span.dist_elem(); i < it; ++i) {
                ptr[i] = i;
            }
        }
    }
};

template <class Opr>
void run_check(Handle* handle) {
    // see OprProxyIndexingMultiAxisVecHelper for more details
    // execs() give input, output and index layouts

    Checker<Opr> checker(handle);
    size_t idx_size0, idx_size1;
    OrderedRNG rng_inp;
    IndexRNG rng0{idx_size0, 2}, rng1{idx_size1, 3};
    checker.set_dtype(0, dtype::Float32())
            .  // data
            set_dtype(1, dtype::Float32())
            .  // value
            set_dtype(2, dtype::Int32())
            .  // idx0
            set_dtype(3, dtype::Int32())
            .  // idx1
            set_rng(0, &rng_inp)
            .set_rng(1, &rng_inp)
            .set_rng(2, &rng0)
            .set_rng(3, &rng1);

    idx_size0 = 23;
    checker.set_proxy({{0}})
            .execs({{23}, {100}, {100}})
            .execs({{23, 5}, {100, 5}, {100}});

    idx_size0 = 2;
    idx_size1 = 3;
    checker.set_proxy({{0, 1}})
            .execs({{2, 3}, {10}, {10}, {10}})
            .execs({{2, 3, 5}, {10, 5}, {10}, {10}});

    idx_size0 = 4;
    idx_size1 = 6;
    TensorLayout inp_layout{{3, 4, 5, 6}, dtype::Float32()};
    inp_layout.stride[0] *= 8;
    inp_layout.stride[1] *= 2;
    checker.set_proxy({{1, 3}}).execl({
            inp_layout,
            {{7, 3, 5}, dtype::Float32()},
            {{7}, dtype::Int32()},
            {{1}, dtype::Int32()},
    });

    idx_size0 = 4;
    idx_size1 = 5;
    checker.set_proxy({{2, 3}}).execs(
            {{2, 3, 4, 5, 6, 7}, {2, 3, 10, 6, 7}, {10}, {10}});

    idx_size0 = 4;
    checker.set_proxy({{1}}).execs({{1, 4}, {1, 1024 * 1024}, {1024 * 1024}});

    if (std::is_same<Opr, IndexingIncrMultiAxisVec>::value) {
        idx_size0 = 4;
        TensorLayout val_layout{{23}, dtype::Float32()};
        val_layout.stride[0] = 0;
        checker.set_proxy({{0}}).execl(
                {{{4}, dtype::Float32()}, val_layout, {{23}, dtype::Int32()}});
    }
}
}  // namespace

TEST_F(ROCM, INDEXING_MULTI_AXIS_VEC) {
    run_check<IndexingMultiAxisVec>(handle_rocm());
    Checker<IndexingMultiAxisVec> checker(handle_rocm());
    size_t idx_size0;
    OrderedRNG rng_inp;
    IndexRNG rng0{idx_size0, 2};
    checker.set_dtype(0, dtype::Float32())
            .  // data
            set_dtype(1, dtype::Float32())
            .  // value
            set_dtype(2, dtype::Int32())
            .  // idx0
            set_rng(0, &rng_inp)
            .set_rng(1, &rng_inp)
            .set_rng(2, &rng0);

    idx_size0 = 20;
    checker.set_proxy({{0}}).execl(
            {TensorLayout{{20}, dtype::Float32()}, TensorLayout{{9}, dtype::Float32()},
             TensorLayout{TensorShape{9}, {-1}, dtype::Int32()}});
}

TEST_F(ROCM, INDEXING_INCR_MULTI_AXIS_VEC) {
    run_check<IndexingIncrMultiAxisVec>(handle_rocm());
}

TEST_F(ROCM, INDEXING_SET_MULTI_AXIS_VEC) {
    Checker<IndexingSetMultiAxisVec> checker(handle_rocm());
    OrderedRNG rng;
    checker.set_dtype(0, dtype::Float32())
            .  // data
            set_dtype(1, dtype::Float32())
            .  // value
            set_dtype(2, dtype::Int32())
            .  // idx0
            set_rng(0, &rng)
            .set_rng(1, &rng)
            .set_rng(2, &rng);

    checker.set_proxy({{1}}).execs({{5, 8, 3}, {5, 2, 3}, {2}});
}

TEST_F(ROCM_ERROR_INFO, INDEXING_MULTI_AXIS_VEC) {
    Checker<IndexingMultiAxisVec> checker(handle_rocm());
    UniformIntRNG idx_rng{-5, 5};
    checker.set_dtype(0, dtype::Float32())
            .  // data
            set_dtype(1, dtype::Float32())
            .  // value
            set_dtype(2, dtype::Int32())
            .  // idx
            set_rng(2, &idx_rng);

    bool failed = false;
    ASSERT_EQ(0u, get_error_info().nr_error);
    auto on_fail = [&failed, this]() {
        failed = true;
        auto info = get_error_info();
        ASSERT_GE(info.nr_error, 1u);
        printf("error msg: ");
        printf(info.msg, info.msg_args[0], info.msg_args[1], info.msg_args[2],
               info.msg_args[3]);
        printf("\n");
    };

    checker.set_proxy({{0}}).execs({{23}, {100}, {100}});

    idx_rng = {-500, 500};
    checker.set_expect_exec_fail(on_fail).execs({{23}, {100}, {100}});

    ASSERT_TRUE(failed);
}

TEST_F(ROCM, INDEXING_MULTI_AXIS_VEC_BENCHMARK) {
    ROCMBenchmarker<IndexingMultiAxisVec> benchmarker(
            handle_rocm(), handle_naive(false));
    benchmarker.set_display(true);
    OrderedRNG rng_inp;
    size_t idx_size = 10000;
    IndexRNG rng0{idx_size, 3}, rng1{idx_size, 1};
    benchmarker.set_dtype(0, dtype::Float32())
            .set_dtype(1, dtype::Float32())
            .set_dtype(2, dtype::Int32())
            .set_dtype(3, dtype::Int32())
            .set_rng(0, &rng_inp)
            .set_rng(1, &rng_inp)
            .set_rng(2, &rng0)
            .set_rng(3, &rng1)
            .set_proxy({{0, 1}});
    auto time_ms =
            benchmarker.execs({{1000, 1000, 1000}, {1000, 1000}, {1000}, {1000}});
    long io = 2 * 1000 * 1000 * dtype::Float32().size();
    printf("io = %.3f GB, random access bandwidth = %.3f GB/s\n", (float)(io / 1e9),
           (float)(io / (time_ms * 1e6)));
}

// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}