| 82 | } |
| 83 | |
| 84 | void SyncMem::to_device() { |
| 85 | #ifdef USE_CUDA |
| 86 | switch (head_) { |
| 87 | case UNINITIALIZED: |
| 88 | CUDA_CHECK(cudaMalloc(&device_ptr, size_)); |
| 89 | CUDA_CHECK(cudaMemset(device_ptr, 0, size_)); |
| 90 | head_ = DEVICE; |
| 91 | own_device_data = true; |
| 92 | total_memory_size += size_; |
| 93 | break; |
| 94 | case HOST: |
| 95 | if (nullptr == device_ptr) { |
| 96 | CUDA_CHECK(cudaMalloc(&device_ptr, size_)); |
| 97 | CUDA_CHECK(cudaMemset(device_ptr, 0, size_)); |
| 98 | own_device_data = true; |
| 99 | } |
| 100 | CUDA_CHECK(cudaMemcpy(device_ptr, host_ptr, size_, cudaMemcpyHostToDevice)); |
| 101 | head_ = DEVICE; |
| 102 | break; |
| 103 | case DEVICE:; |
| 104 | } |
| 105 | #else |
| 106 | NO_GPU; |
| 107 | #endif |
| 108 | } |
| 109 | |
| 110 | void SyncMem::set_host_data(void *data) { |
| 111 | CHECK_NOTNULL(data); |