| 34 | } |
| 35 | |
| 36 | void TestCache(Cache* cache, uint32_t line_size) { |
| 37 | std::unique_ptr<ep::DeviceManagerRegistry> device_manager_registry( |
| 38 | new ep::DeviceManagerRegistry()); |
| 39 | auto device = device_manager_registry->GetDevice(DeviceType::kCUDA, 0); |
| 40 | ep::Stream* stream = device->CreateStream(); |
| 41 | |
| 42 | std::unordered_set<int64_t> in_cache; |
| 43 | const size_t n_iter = 32; |
| 44 | const uint32_t n_keys = 1024; |
| 45 | int64_t* d_keys; |
| 46 | int64_t* keys; |
| 47 | uint32_t* d_n_missing; |
| 48 | uint32_t* n_missing; |
| 49 | int64_t* d_missing_keys; |
| 50 | int64_t* missing_keys; |
| 51 | uint32_t* d_missing_indices; |
| 52 | uint32_t* missing_indices; |
| 53 | float* d_values; |
| 54 | float* values; |
| 55 | float* d_evicted_values; |
| 56 | float* evicted_values; |
| 57 | uint32_t* d_n_evicted; |
| 58 | uint32_t* n_evicted; |
| 59 | int64_t* d_evicted_keys; |
| 60 | int64_t* evicted_keys; |
| 61 | uint8_t* mask; |
| 62 | const size_t keys_size = n_keys * sizeof(int64_t); |
| 63 | OF_CUDA_CHECK(cudaMalloc(&d_keys, keys_size)); |
| 64 | OF_CUDA_CHECK(cudaMallocHost(&keys, keys_size)); |
| 65 | OF_CUDA_CHECK(cudaMalloc(&d_n_missing, sizeof(uint32_t))); |
| 66 | OF_CUDA_CHECK(cudaMallocHost(&n_missing, sizeof(uint32_t))); |
| 67 | OF_CUDA_CHECK(cudaMalloc(&d_missing_keys, keys_size)); |
| 68 | OF_CUDA_CHECK(cudaMallocHost(&missing_keys, keys_size)); |
| 69 | const size_t indices_size = n_keys * sizeof(uint32_t); |
| 70 | OF_CUDA_CHECK(cudaMalloc(&d_missing_indices, indices_size)); |
| 71 | OF_CUDA_CHECK(cudaMallocHost(&missing_indices, indices_size)); |
| 72 | const size_t values_size = n_keys * line_size * sizeof(float); |
| 73 | OF_CUDA_CHECK(cudaMalloc(&d_values, values_size)); |
| 74 | OF_CUDA_CHECK(cudaMallocHost(&values, values_size)); |
| 75 | OF_CUDA_CHECK(cudaMalloc(&d_evicted_values, values_size)); |
| 76 | OF_CUDA_CHECK(cudaMallocHost(&evicted_values, values_size)); |
| 77 | OF_CUDA_CHECK(cudaMalloc(&d_n_evicted, sizeof(uint32_t))); |
| 78 | OF_CUDA_CHECK(cudaMallocHost(&n_evicted, sizeof(uint32_t))); |
| 79 | OF_CUDA_CHECK(cudaMalloc(&d_evicted_keys, keys_size)); |
| 80 | OF_CUDA_CHECK(cudaMallocHost(&evicted_keys, keys_size)); |
| 81 | OF_CUDA_CHECK(cudaMalloc(&mask, n_keys)); |
| 82 | std::vector<int64_t> random_keys(n_keys * 32); |
| 83 | std::iota(random_keys.begin(), random_keys.end(), 1); |
| 84 | std::random_device rd; |
| 85 | std::mt19937 g(rd()); |
| 86 | for (size_t iter = 0; iter < n_iter; ++iter) { |
| 87 | std::shuffle(random_keys.begin(), random_keys.end(), g); |
| 88 | std::copy(random_keys.begin(), random_keys.begin() + n_keys, keys); |
| 89 | uint32_t expect_n_missing = 0; |
| 90 | std::unordered_set<int64_t> expect_missing_keys_set; |
| 91 | std::unordered_set<uint32_t> expect_missing_indices_set; |
| 92 | std::unordered_set<int64_t> keys_set; |
| 93 | for (size_t i = 0; i < n_keys; ++i) { |
no test coverage detected