Function TEST_CASE

test/fuse_horizontal_test.cpp:40–138

4 gathers with same embedding dim → should fuse into 1 batched gather

// 4 gathers with same embedding dim → should fuse into 1 batched gather
TEST_CASE(gather_horiz_fusion_basic)
{
    migraphx::module m1;
    {
        auto emb1 =
            m1.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {3, 2}}, 0));
        auto emb2 =
            m1.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {4, 2}}, 1));
        auto emb3 =
            m1.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {2, 2}}, 2));
        auto emb4 =
            m1.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {5, 2}}, 3));

        auto idx1 = m1.add_parameter("idx1", {migraphx::shape::int32_type, {2}});
        auto idx2 = m1.add_parameter("idx2", {migraphx::shape::int32_type, {3}});
        auto idx3 = m1.add_parameter("idx3", {migraphx::shape::int32_type, {1}});
        auto idx4 = m1.add_parameter("idx4", {migraphx::shape::int32_type, {2}});

        auto g1 = m1.add_instruction(migraphx::make_op("gather", {{"axis", 0}}), emb1, idx1);
        auto g2 = m1.add_instruction(migraphx::make_op("gather", {{"axis", 0}}), emb2, idx2);
        auto g3 = m1.add_instruction(migraphx::make_op("gather", {{"axis", 0}}), emb3, idx3);
        auto g4 = m1.add_instruction(migraphx::make_op("gather", {{"axis", 0}}), emb4, idx4);

        // Combine all outputs so every gather stays live through DCE
        m1.add_instruction(migraphx::make_op("concat", {{"axis", 0}}),
                           std::vector<migraphx::instruction_ref>{g1, g2, g3, g4});
    }
    run_pass(m1);

    migraphx::module m2;
    {
        // Embedding literals (added first → pushed to front → end up at the back of no-dep list)
        auto emb1 =
            m2.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {3, 2}}, 0));
        auto emb2 =
            m2.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {4, 2}}, 1));
        auto emb3 =
            m2.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {2, 2}}, 2));
        auto emb4 =
            m2.add_literal(migraphx::generate_literal({migraphx::shape::float_type, {5, 2}}, 3));

        // Parameters (added second → in middle of no-dep list)
        auto idx1 = m2.add_parameter("idx1", {migraphx::shape::int32_type, {2}});
        auto idx2 = m2.add_parameter("idx2", {migraphx::shape::int32_type, {3}});
        auto idx3 = m2.add_parameter("idx3", {migraphx::shape::int32_type, {1}});
        auto idx4 = m2.add_parameter("idx4", {migraphx::shape::int32_type, {2}});

        // Offset literals (added last → pushed to very front of no-dep list,
        // matching order of add_literal calls inside the pass's fuse loop)
        auto offset2 = m2.add_literal(
            migraphx::literal{migraphx::shape{migraphx::shape::int32_type}, {std::size_t(3)}});
        auto offset3 = m2.add_literal(
            migraphx::literal{migraphx::shape{migraphx::shape::int32_type}, {std::size_t(7)}});
        auto offset4 = m2.add_literal(
            migraphx::literal{migraphx::shape{migraphx::shape::int32_type}, {std::size_t(9)}});

        // Concatenated embedding table: [3+4+2+5, 2] = [14, 2]
        auto concat_emb =