| 56 | } |
| 57 | |
| 58 | void SyncMem::to_host() { |
| 59 | switch (head_) { |
| 60 | case UNINITIALIZED: |
| 61 | malloc_host(&host_ptr, size_); |
| 62 | memset(host_ptr, 0, size_); |
| 63 | head_ = HOST; |
| 64 | own_host_data = true; |
| 65 | total_memory_size += size_; |
| 66 | break; |
| 67 | case DEVICE: |
| 68 | #ifdef USE_CUDA |
| 69 | if (nullptr == host_ptr) { |
| 70 | CUDA_CHECK(cudaMallocHost(&host_ptr, size_)); |
| 71 | CUDA_CHECK(cudaMemset(host_ptr, 0, size_)); |
| 72 | own_host_data = true; |
| 73 | } |
| 74 | CUDA_CHECK(cudaMemcpy(host_ptr, device_ptr, size_, cudaMemcpyDeviceToHost)); |
| 75 | head_ = HOST; |
| 76 | #else |
| 77 | NO_GPU; |
| 78 | #endif |
| 79 | break; |
| 80 | case HOST:; |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | void SyncMem::to_device() { |
| 85 | #ifdef USE_CUDA |