| 1 | #pragma once |
| 2 | |
| 3 | #ifdef USE_ROCM |
| 4 | #include <hip/hip_runtime.h> |
| 5 | #endif |
| 6 | |
| 7 | #ifdef USE_ROCM |
| 8 | struct Utils { |
| 9 | static __host__ int get_warp_size() { |
| 10 | static bool is_cached = false; |
| 11 | static int result; |
| 12 | |
| 13 | if (!is_cached) { |
| 14 | int device_id; |
| 15 | cudaDeviceProp deviceProp; |
| 16 | cudaGetDevice(&device_id); |
| 17 | cudaGetDeviceProperties(&deviceProp, device_id); |
| 18 | |
| 19 | result = deviceProp.warpSize; |
| 20 | is_cached = true; |
| 21 | } |
| 22 | |
| 23 | return result; |
| 24 | } |
| 25 | |
| 26 | static __device__ constexpr int get_warp_size() { |
| 27 | #ifdef __GFX9__ |
| 28 | return 64; |
| 29 | #else |
| 30 | return 32; |
| 31 | #endif |
| 32 | } |
| 33 | }; |
| 34 | |
| 35 | #define WARP_SIZE Utils::get_warp_size() |
| 36 | #else |
| 37 | #define WARP_SIZE 32 |
| 38 | #endif |
| 39 | |
| 40 | #ifndef USE_ROCM |
| 41 | #define VLLM_LDG(arg) __ldg(arg) |
| 42 | #else |
| 43 | #define VLLM_LDG(arg) *(arg) |
| 44 | #endif |
| 45 | |
| 46 | #ifndef USE_ROCM |
| 47 | #define VLLM_SHFL_XOR_SYNC(var, lane_mask) \ |
| 48 | __shfl_xor_sync(uint32_t(-1), var, lane_mask) |
| 49 | #define VLLM_SHFL_XOR_SYNC_WIDTH(var, lane_mask, width) \ |
| 50 | __shfl_xor_sync(uint32_t(-1), var, lane_mask, width) |
| 51 | #else |
| 52 | #define VLLM_SHFL_XOR_SYNC(var, lane_mask) __shfl_xor(var, lane_mask) |
| 53 | #define VLLM_SHFL_XOR_SYNC_WIDTH(var, lane_mask, width) \ |
| 54 | __shfl_xor(var, lane_mask, width) |
| 55 | #endif |
| 56 | |
| 57 | #ifndef USE_ROCM |
| 58 | #define VLLM_SHFL_SYNC(var, src_lane) __shfl_sync(uint32_t(-1), var, src_lane) |
| 59 | #else |
| 60 | #define VLLM_SHFL_SYNC(var, src_lane) __shfl(var, src_lane) |
| 61 | #endif |
| 62 | |
| 63 | #ifndef USE_ROCM |
| 64 | #define VLLM_SHFL_DOWN_SYNC(var, lane_delta) \ |
| 65 | __shfl_down_sync(uint32_t(-1), var, lane_delta) |
| 66 | #else |
| 67 | #define VLLM_SHFL_DOWN_SYNC(var, lane_delta) __shfl_down(var, lane_delta) |
| 68 | #endif |
| 69 | |
| 70 | #ifndef USE_ROCM |
| 71 | #define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \ |
| 72 | cudaFuncSetAttribute(FUNC, cudaFuncAttributeMaxDynamicSharedMemorySize, VAL) |
| 73 | #else |
| 74 | #define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \ |
| 75 | hipFuncSetAttribute(FUNC, hipFuncAttributeMaxDynamicSharedMemorySize, VAL) |
| 76 | #endif |