eliminate_pad_test.cpp

#include <migraphx/dead_code_elimination.hpp>
#include <migraphx/eliminate_pad.hpp>
#include <migraphx/pass_manager.hpp>
#include <migraphx/instruction.hpp>
#include <basic_ops.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/make_op.hpp>

#include <test.hpp>

void run_pass(migraphx::module& m)
{
    migraphx::run_passes(m, {migraphx::eliminate_pad{}, migraphx::dead_code_elimination{}});
}

migraphx::instruction_ref
create_im2col(migraphx::instruction_ref& l_img, size_t channels, migraphx::module& m)
{
    size_t f[2] = {1, 1};
    std::vector<int32_t> weights(channels * f[0] * f[1]);
    migraphx::shape s_weights{migraphx::shape::int32_type, {1, channels, f[0], f[1]}};
    auto l_weights = m.add_literal(migraphx::literal{s_weights, weights});
    return m.add_instruction(migraphx::make_op("im2col"), l_img, l_weights);
}

migraphx::instruction_ref
create_conv(migraphx::instruction_ref& l_img,
            size_t channels,
            migraphx::module& m,
            migraphx::op::padding_mode_t padding_mode = migraphx::op::padding_mode_t::default_)
{
    migraphx::shape s_weights{migraphx::shape::int32_type, {4, channels, 3, 3}};
    std::vector<int32_t> weights(4 * channels * 3 * 3);
    auto l_weights = m.add_literal(migraphx::literal{s_weights, weights});
    migraphx::op::convolution op;
    op.padding_mode = padding_mode;
    return m.add_instruction(op, l_img, l_weights);
}

TEST_CASE(rewrite_pad)
{
    migraphx::module m;
    size_t img_dim[2] = {2, 2};
    size_t channels   = 1;
    std::vector<int32_t> input(channels * img_dim[0] * img_dim[1]);
    std::iota(input.begin(), input.end(), 0);

    migraphx::shape s_img{migraphx::shape::int32_type, {1, channels, img_dim[0], img_dim[1]}};
    auto l_img = m.add_literal(migraphx::literal{s_img, input});
    auto padded_img =
        m.add_instruction(migraphx::make_op("pad", {{"pads", {0, 0, 1, 1, 0, 0, 1, 1}}}), l_img);

    auto l0 = create_im2col(padded_img, channels, m);
    auto l1 = create_conv(padded_img, channels, m);
    auto l2 = m.add_instruction(migraphx::make_op("pooling", {{"mode", "max"}}), padded_img);
    m.add_instruction(migraphx::make_op("identity"), l0, l1, l2);

    auto s0 = l0->get_shape();
    auto s1 = l1->get_shape();
    auto s2 = l2->get_shape();
    run_pass(m);
    EXPECT(l0->get_shape() == s0);
    EXPECT(l1->get_shape() == s1);
    EXPECT(l2->get_shape() == s2);
    auto op0 = l0->get_operator().to_value();
    auto om1 = l1->get_operator().to_value();
    auto om2 = l2->get_operator().to_value();

    EXPECT(op0["padding"].to_vector<std::size_t>() == std::vector<std::size_t>{1, 1});
    EXPECT(om1["padding"].to_vector<std::size_t>() == std::vector<std::size_t>{1, 1});
    EXPECT(om2["padding"].to_vector<std::size_t>() == std::vector<std::size_t>{1, 1});

    EXPECT(std::none_of(
        m.begin(), m.end(), [](const migraphx::instruction& ins) { return ins.name() == "pad"; }));
}

TEST_CASE(rewrite_pad_im2col_asymetric)
{
    migraphx::module m;

    size_t img_dim[2] = {2, 2};
    size_t channels   = 1;
    std::vector<int32_t> input(channels * img_dim[0] * img_dim[1]);
    std::iota(input.begin(), input.end(), 0);

    migraphx::shape s_img{migraphx::shape::int32_type, {1, channels, img_dim[0], img_dim[1]}};
    auto l_img = m.add_literal(migraphx::literal{s_img, input});
    auto padded_img =
        m.add_instruction(migraphx::make_op("pad", {{"pads", {0, 0, 0, 0, 0, 0, 2, 2}}}), l_img);

    auto l0 = create_im2col(padded_img, channels, m);

    auto s0 = l0->get_shape();
    run_pass(m);
    EXPECT(l0->get_shape() == s0);
    auto op0 = l0->get_operator().to_value();

    EXPECT(op0["padding"].to_vector<std::size_t>() == std::vector<std::size_t>{0, 0});

    run_pass(m);
    EXPECT(std::any_of(
        m.begin(), m.end(), [](const migraphx::instruction& ins) { return ins.name() == "pad"; }));
}

int main(int argc, const char* argv[]) { test::run(argc, argv); }