types.h

Go to the documentation of this file.

#pragma once
 
// Pulls in the active backend's runtime headers, and (under HIP) re-points the
// CUDA-spelled host API onto its HIP equivalents.
#include "backend/api.h"
 
#if defined(FZGMOD_BACKEND_HIP)
 
#include <cstddef>
#include <cstdint>
 
namespace fz {
 
using stream_t     = hipStream_t;
using event_t      = hipEvent_t;
using mempool_t    = hipMemPool_t;
using error_t      = hipError_t;
using graph_t      = hipGraph_t;
using graph_exec_t = hipGraphExec_t;
 
inline constexpr error_t kBackendSuccess = hipSuccess;
 
inline constexpr bool kBackendSupportsGraphCapture = true;
 
inline const char* getBackendErrorString(error_t err) {
    return hipGetErrorString(err);
}
 
inline size_t getPoolUsedMemCurrent(mempool_t pool) {
    if (!pool) return 0;
    uint64_t used = 0;
    hipMemPoolGetAttribute(pool, hipMemPoolAttrUsedMemCurrent, &used);
    return static_cast<size_t>(used);
}
 
inline size_t getPoolUsedMemHigh(mempool_t pool) {
    if (!pool) return 0;
    uint64_t high = 0;
    hipMemPoolGetAttribute(pool, hipMemPoolAttrUsedMemHigh, &high);
    return static_cast<size_t>(high);
}
 
} // namespace fz
 
#elif defined(FZGMOD_BACKEND_SYCL)
 
#error "FZGMOD_BACKEND=SYCL is not implemented yet"
 
#else // FZGMOD_BACKEND_CUDA (default)
 
#include <cstddef>
#include <cstdint>
 
namespace fz {
 
using stream_t     = cudaStream_t;
using event_t      = cudaEvent_t;
using mempool_t    = cudaMemPool_t;
using error_t      = cudaError_t;
using graph_t      = cudaGraph_t;
using graph_exec_t = cudaGraphExec_t;
 
inline constexpr error_t kBackendSuccess = cudaSuccess;
 
inline constexpr bool kBackendSupportsGraphCapture = true;
 
inline const char* getBackendErrorString(error_t err) {
    return cudaGetErrorString(err);
}
 
inline size_t getPoolUsedMemCurrent(mempool_t pool) {
    if (!pool) return 0;
    uint64_t used = 0;
    cudaMemPoolGetAttribute(pool, cudaMemPoolAttrUsedMemCurrent, &used);
    return static_cast<size_t>(used);
}
 
inline size_t getPoolUsedMemHigh(mempool_t pool) {
    if (!pool) return 0;
    uint64_t high = 0;
    cudaMemPoolGetAttribute(pool, cudaMemPoolAttrUsedMemHigh, &high);
    return static_cast<size_t>(high);
}
 
} // namespace fz
 
#endif