mirror of
https://github.com/vllm-project/vllm.git
synced 2025-10-20 14:53:52 +08:00
58 lines
1.7 KiB
C
58 lines
1.7 KiB
C
#pragma once
|
|
|
|
#ifdef USE_ROCM
|
|
#include <hip/hip_runtime.h>
|
|
#endif
|
|
|
|
#ifndef USE_ROCM
|
|
#define WARP_SIZE 32
|
|
#else
|
|
#define WARP_SIZE warpSize
|
|
#endif
|
|
|
|
#ifndef USE_ROCM
|
|
#define VLLM_LDG(arg) __ldg(arg)
|
|
#else
|
|
#define VLLM_LDG(arg) *(arg)
|
|
#endif
|
|
|
|
#ifndef USE_ROCM
|
|
#define VLLM_SHFL_XOR_SYNC(var, lane_mask) \
|
|
__shfl_xor_sync(uint32_t(-1), var, lane_mask)
|
|
#define VLLM_SHFL_XOR_SYNC_WIDTH(var, lane_mask, width) \
|
|
__shfl_xor_sync(uint32_t(-1), var, lane_mask, width)
|
|
#else
|
|
#define VLLM_SHFL_XOR_SYNC(var, lane_mask) __shfl_xor(var, lane_mask)
|
|
#define VLLM_SHFL_XOR_SYNC_WIDTH(var, lane_mask, width) \
|
|
__shfl_xor(var, lane_mask, width)
|
|
#endif
|
|
|
|
#ifndef USE_ROCM
|
|
#define VLLM_SHFL_SYNC(var, src_lane) __shfl_sync(uint32_t(-1), var, src_lane)
|
|
#else
|
|
#define VLLM_SHFL_SYNC(var, src_lane) __shfl(var, src_lane)
|
|
#endif
|
|
|
|
#ifndef USE_ROCM
|
|
#define VLLM_SHFL_DOWN_SYNC(var, lane_delta) \
|
|
__shfl_down_sync(uint32_t(-1), var, lane_delta)
|
|
#else
|
|
#define VLLM_SHFL_DOWN_SYNC(var, lane_delta) __shfl_down(var, lane_delta)
|
|
#endif
|
|
|
|
#ifndef USE_ROCM
|
|
#define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \
|
|
cudaFuncSetAttribute(FUNC, cudaFuncAttributeMaxDynamicSharedMemorySize, VAL)
|
|
#else
|
|
#define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \
|
|
hipFuncSetAttribute(FUNC, hipFuncAttributeMaxDynamicSharedMemorySize, VAL)
|
|
#endif
|
|
|
|
// #ifndef USE_ROCM
|
|
// #define VLLM_cudaHostGetDevicePointer(device_ptr, host_ptr, flags) \
|
|
// cudaHostGetDevicePointer(device_ptr, host_ptr, flags)
|
|
// #else
|
|
// #define VLLM_cudaHostGetDevicePointer(device_ptr, host_ptr, flags) \
|
|
// hipHostGetDevicePointer(device_ptr, host_ptr, flags)
|
|
// #endif
|