Function test_load_mpp_cases

tests/ds4_test.c:1513–1540

static void test_metal_compressor_ape_add_exact(void) {
    test_metal_compressor_ape_add_exact_case(128, 4, 3, 16, 1, 7, false);
    test_metal_compressor_ape_add_exact_case(512, 4, 0, 16, 1, 13, false);
    test_metal_compressor_ape_add_exact_case(512, 4, 1, 14, 0, 19, false);
    test_metal_compressor_ape_add_exact_case(512, 128, 127, 257, 1, 31, false);
}

static void test_metal_compressor_ratio4_pack_exact(void) {
    test_metal_compressor_ape_add_exact_case(128, 4, 0, 4, 1, 37, true);
    test_metal_compressor_ape_add_exact_case(512, 4, 0, 8, 1, 41, true);
    test_metal_compressor_ape_add_exact_case(512, 4, 1, 14, 0, 43, true);
}

static void test_metal_compressor_ratio4_replay_pack_exact_case(
        uint32_t head_dim,
        uint32_t n_tokens,
        uint32_t seed) {
    const uint32_t width = 2u * head_dim;
    const uint32_t n_comp = n_tokens / 4u;
    const uint64_t input_count = (uint64_t)n_tokens * width;
    const uint64_t state_count = (uint64_t)8u * width;
    const uint64_t comp_count = (uint64_t)n_comp * head_dim;
    const uint64_t input_bytes = input_count * sizeof(float);
    const uint64_t state_bytes = state_count * sizeof(float);
    const uint64_t comp_bytes = comp_count * sizeof(float);
    const uint64_t ape_bytes = (uint64_t)4u * width * sizeof(uint16_t);
    const uint64_t page = (uint64_t)getpagesize();
    const uint64_t norm_offset = test_round_up_u64(ape_bytes, page);
    const uint64_t model_bytes = test_round_up_u64(
        norm_offset + (uint64_t)head_dim * sizeof(float), page);

    ds4_gpu_tensor *kv = ds4_gpu_tensor_alloc(input_bytes);