colvar_gpu_support.h

#ifndef COLVAR_GPU_SUPPORT_H
#define COLVAR_GPU_SUPPORT_H
 
#include <vector>
#include <iostream>
#include <unordered_map>
 
class colvarmodule;
 
#define COLVARS_STRINGIFY(s) STRINGIFY_HELPER(s)
#define STRINGIFY_HELPER(s) #s
 
#if defined(COLVARS_CUDA)
#include <cuda_runtime.h>
#ifdef COLVARS_NVTX_PROFILING
#include <nvtx3/nvToolsExt.h>
#include <nvtx3/nvToolsExtCudaRt.h>
#endif
#define COLVARS_SYNC_WARP __syncwarp()
#endif // defined(COLVARS_CUDA)
 
#if defined(COLVARS_HIP)
#include <hip/hip_runtime.h>
#if defined(__HIP_PLATFORM_AMD__)
  #define COLVARS_SYNC_WARP do {\
    __builtin_amdgcn_fence(__ATOMIC_RELEASE, "wavefront"); \
    __builtin_amdgcn_wave_barrier(); \
    __builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "wavefront"); \
  } while (0)
#elif defined(__HIP_PLATFORM_NVIDIA__)
  #define COLVARS_SYNC_WARP __syncwarp()
#else
  #error "Unknown HIP platform"
#endif
#endif // defined(COLVARS_HIP)
 
#if defined(COLVARS_HIP)
#ifndef cudaError_t
#define cudaError_t hipError_t
#endif // cudaError_t
 
#ifndef cudaFree
#define cudaFree hipFree
#endif // cudaFree
 
#ifndef cudaFreeHost
#define cudaFreeHost hipFreeHost
#endif // cudaFreeHost
 
#ifndef cudaFreeAsync
#define cudaFreeAsync hipFreeAsync
#endif // cudaFreeAsync
 
#ifndef cudaGetErrorString
#define cudaGetErrorString hipGetErrorString
#endif // cudaGetErrorString
 
#ifndef cudaGraphAddChildGraphNode
#define cudaGraphAddChildGraphNode hipGraphAddChildGraphNode
#endif // cudaGraphAddChildGraphNode
 
#ifndef cudaGraphAddKernelNode
#define cudaGraphAddKernelNode hipGraphAddKernelNode
#endif // cudaGraphAddKernelNode
 
#ifndef cudaGraphAddMemcpyNode
#define cudaGraphAddMemcpyNode hipGraphAddMemcpyNode
#endif // cudaGraphAddMemcpyNode
 
#ifndef cudaGraphAddMemsetNode
#define cudaGraphAddMemsetNode hipGraphAddMemsetNode
#endif // cudaGraphAddMemsetNode
 
#ifndef cudaGraphCreate
#define cudaGraphCreate hipGraphCreate
#endif // cudaGraphCreate
 
#ifndef cudaGraphDestroy
#define cudaGraphDestroy hipGraphDestroy
#endif // cudaGraphDestroy
 
#ifndef cudaGraphExecDestroy
#define cudaGraphExecDestroy hipGraphExecDestroy
#endif // cudaGraphExecDestroy
 
#ifndef cudaGraphExecMemcpyNodeSetParams
#define cudaGraphExecMemcpyNodeSetParams hipGraphExecMemcpyNodeSetParams
#endif // cudaGraphExecMemcpyNodeSetParams
 
#ifndef cudaGraphExec_t
#define cudaGraphExec_t hipGraphExec_t
#endif // cudaGraphExec_t
 
#ifndef cudaGraphInstantiate
#define cudaGraphInstantiate hipGraphInstantiate
#endif // cudaGraphInstantiate
 
#ifndef cudaGraphInstantiateWithParams
#define cudaGraphInstantiateWithParams hipGraphInstantiateWithParams
#endif // cudaGraphInstantiateWithParams
 
#ifndef cudaGraphInstantiateParams
#define cudaGraphInstantiateParams hipGraphInstantiateParams
#endif // cudaGraphInstantiateParams
 
#ifndef cudaGraphInstantiateFlagUpload
#define cudaGraphInstantiateFlagUpload hipGraphInstantiateFlagUpload
#endif // cudaGraphInstantiateFlagUpload
 
#ifndef cudaGraphInstantiateSuccess
#define cudaGraphInstantiateSuccess hipGraphInstantiateSuccess
#endif // cudaGraphInstantiateSuccess
 
#ifndef cudaGraphLaunch
#define cudaGraphLaunch hipGraphLaunch
#endif // cudaGraphLaunch
 
#ifndef cudaGraphNode_t
#define cudaGraphNode_t hipGraphNode_t
#endif // cudaGraphNode_t
 
#ifndef cudaGraph_t
#define cudaGraph_t hipGraph_t
#endif // cudaGraph_t
 
#ifndef cudaGraphDebugDotPrint
#define cudaGraphDebugDotPrint hipGraphDebugDotPrint
#endif // cudaGraphDebugDotPrint
 
#ifndef cudaGraphDebugDotFlags
#define cudaGraphDebugDotFlags hipGraphDebugDotFlags
#endif // cudaGraphDebugDotFlags
 
#ifndef cudaGraphDebugDotFlagsVerbose
#define cudaGraphDebugDotFlagsVerbose hipGraphDebugDotFlagsVerbose
#endif // cudaGraphDebugDotFlagsVerbose
 
#ifndef cudaHostAllocMapped
#define cudaHostAllocMapped hipHostAllocMapped
#endif // cudaHostAllocMapped
 
#ifndef cudaHostAllocDefault
#define cudaHostAllocDefault hipHostAllocDefault
#endif // cudaHostAllocDefault
 
#ifndef cudaHostAlloc
#define cudaHostAlloc hipHostAlloc
#endif // cudaHostAlloc
 
#ifndef cudaLaunchKernel
#define cudaLaunchKernel hipLaunchKernel
#endif // cudaLaunchKernel
 
#ifndef cudaKernelNodeParams
#define cudaKernelNodeParams hipKernelNodeParams
#endif // cudaKernelNodeParams
 
#ifndef cudaMalloc
#define cudaMalloc hipMalloc
#endif // cudaMalloc
 
#ifndef cudaMallocAsync
#define cudaMallocAsync hipMallocAsync
#endif // cudaMallocAsync
 
#ifndef cudaMallocHost
#define cudaMallocHost hipMallocHost
#endif // cudaMallocHost
 
#ifndef cudaGraphAddMemcpyNode1D
#define cudaGraphAddMemcpyNode1D hipGraphAddMemcpyNode1D
#endif
 
#ifndef cudaMemcpy
#define cudaMemcpy hipMemcpy
#endif // cudaMemcpy
 
#ifndef cudaMemcpy3DParms
#define cudaMemcpy3DParms hipMemcpy3DParms
#endif // cudaMemcpy3DParms
 
#ifndef cudaMemcpyAsync
#define cudaMemcpyAsync hipMemcpyAsync
#endif // cudaMemcpyAsync
 
#ifndef cudaMemcpyDeviceToDevice
#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
#endif // cudaMemcpyDeviceToDevice
 
#ifndef cudaMemcpyDeviceToHost
#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
#endif // cudaMemcpyDeviceToHost
 
#ifndef cudaMemcpyHostToDevice
#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
#endif // cudaMemcpyHostToDevice
 
#ifndef cudaMemcpyKind
#define cudaMemcpyKind hipMemcpyKind
#endif // cudaMemcpyKind
 
#ifndef cudaMemset
#define cudaMemset hipMemset
#endif // cudaMemset
 
#ifndef cudaMemsetParams
#define cudaMemsetParams hipMemsetParams
#endif // cudaMemsetParams
 
#ifndef cudaMemsetAsync
#define cudaMemsetAsync hipMemsetAsync
#endif // cudaMemsetAsync
 
#ifndef cudaStreamCreate
#define cudaStreamCreate hipStreamCreate
#endif // cudaStreamCreate
 
#ifndef cudaStreamDestroy
#define cudaStreamDestroy hipStreamDestroy
#endif // cudaStreamDestroy
 
#ifndef cudaStreamSynchronize
#define cudaStreamSynchronize hipStreamSynchronize
#endif // cudaStreamSynchronize
 
#ifndef cudaStream_t
#define cudaStream_t hipStream_t
#endif // cudaStream_t
 
#ifndef cudaSuccess
#define cudaSuccess hipSuccess
#endif // cudaSuccess
 
#ifndef make_cudaExtent
#define make_cudaExtent make_hipExtent
#endif // make_cudaExtent
 
#ifndef make_cudaPitchedPtr
#define make_cudaPitchedPtr make_hipPitchedPtr
#endif // make_cudaPitchedPtr
 
#ifndef make_cudaPos
#define make_cudaPos make_hipPos
#endif // make_cudaPos
 
#ifndef cudaEvent_t
#define cudaEvent_t hipEvent_t
#endif // cudaEvent_t
 
#ifndef cudaEventCreateWithFlags
#define cudaEventCreateWithFlags hipEventCreateWithFlags
#endif // cudaEventCreate
 
#ifndef cudaEventDisableTiming
#define cudaEventDisableTiming hipEventDisableTiming
#endif // cudaEventDisableTiming
 
#ifndef cudaEventDestroy
#define cudaEventDestroy hipEventDestroy
#endif // cudaEventDestroy
 
#ifndef cudaEventRecord
#define cudaEventRecord hipEventRecord
#endif // cudaEventRecord
 
#ifndef cudaEventSynchronize
#define cudaEventSynchronize hipEventSynchronize
#endif // cudaEventSynchronize
 
#ifndef cudaStreamWaitEvent
#define cudaStreamWaitEvent hipStreamWaitEvent
#endif // cudaStreamWaitEvent
 
#ifndef cudaGraphAddEventRecordNode
#define cudaGraphAddEventRecordNode hipGraphAddEventRecordNode
#endif // cudaGraphAddEventRecordNode
 
#ifndef cudaGetDevice
#define cudaGetDevice hipGetDevice
#endif // cudaGetDevice
 
#ifndef cudaDeviceProp
#define cudaDeviceProp hipDeviceProp_t
#endif // cudaDeviceProp
 
#ifndef cudaDeviceGetPCIBusId
#define cudaDeviceGetPCIBusId hipDeviceGetPCIBusId
#endif // cudaDeviceGetPCIBusId
 
#ifndef cudaGetDeviceProperties
#define cudaGetDeviceProperties hipGetDeviceProperties
#endif // cudaGetDeviceProperties
 
#ifndef cudaDeviceGetAttribute
#define cudaDeviceGetAttribute hipDeviceGetAttribute
#endif // cudaDeviceGetAttribute
 
#ifndef cudaDevAttrWarpSize
#define cudaDevAttrWarpSize hipDeviceAttributeWarpSize
#endif // cudaDevAttrWarpSize
 
#endif // defined(COLVARS_HIP)
 
namespace colvars_gpu {
 
#if defined(COLVARS_CUDA) || defined(COLVARS_HIP)
constexpr static unsigned int default_block_size = 128;
constexpr static unsigned int default_reduce_max_num_blocks = 64;
// static unsigned int default_atom_wise_num_blocks = 64;
#endif
 
#if defined(COLVARS_CUDA) || defined(COLVARS_HIP)
#define COLVARS_HOST_DEVICE __device__ __host__
#define COLVARS_DEVICE __device__
#else
#define COLVARS_HOST_DEVICE
#define COLVARS_DEVICE
#endif
 
// HIP does not have cuda::std::array since libhipcxx is not a part of the ROCm distribution,
// so reinvent the wheel...
#if defined(COLVARS_CUDA) || defined(COLVARS_HIP)
template <typename T, unsigned long N>
class array1d {
public:
  T m_data[N];
  using value_type = T;
  using size_type = decltype(N);
  using reference = value_type&;
  using const_reference = const value_type&;
  using pointer = T*;
  using const_pointer = const T*;
  COLVARS_HOST_DEVICE constexpr size_type size() const {return N;}
  COLVARS_HOST_DEVICE reference operator[](size_type pos) {return m_data[pos];}
  COLVARS_HOST_DEVICE const_reference operator[](size_type pos) const {return m_data[pos];}
};
#endif
 
// TODO: What about SYCL?
#if ( defined(COLVARS_CUDA) || defined(COLVARS_HIP) )
template <typename T>
class CudaHostAllocator {
public:
  using value_type = T;
 
  CudaHostAllocator() = default;
 
  template<typename U>
  constexpr CudaHostAllocator(const CudaHostAllocator<U>&) noexcept {}
 
  friend bool operator==(const CudaHostAllocator&, const CudaHostAllocator&) { return true; }
  friend bool operator!=(const CudaHostAllocator&, const CudaHostAllocator&) { return false; }
 
  T* allocate(size_t n) {
    T* ptr;
    if (cudaHostAlloc(&ptr, n * sizeof(T), cudaHostAllocMapped) != cudaSuccess) {
      throw std::bad_alloc();
    }
    return ptr;
  }
  void deallocate(T* ptr, size_t n) noexcept {
    (void)cudaFreeHost(ptr);
  }
  template<typename U, typename... Args>
  void construct(U* p, Args&&... args) {
      new(p) U(std::forward<Args>(args)...);
  }
 
  template<typename U>
  void destroy(U* p) noexcept {
      p->~U();
  }
};
#endif
 
 
#if defined(COLVARS_CUDA) || defined (COLVARS_HIP)
int gpuAssert(cudaError_t code, const char *file, int line);
#endif
 
} // namespace colvars_gpu
 
#if defined(COLVARS_CUDA) || defined (COLVARS_HIP)
#define checkGPUError(ans) colvars_gpu::gpuAssert((ans), __FILE__, __LINE__);
#endif
 
namespace colvars_gpu {
#if defined(COLVARS_CUDA) || defined (COLVARS_HIP)
 
class gpu_graph_t {
public:
  gpu_graph_t() {}
  int reset();
  ~gpu_graph_t();
  int dump_graph(const std::string& filename);
  int init_graph_exec(colvarmodule* cvmodule, cudaStream_t stream);
  bool graph_exec_initialized = false;
  cudaGraph_t graph = nullptr;
  cudaGraphExec_t graph_exec = nullptr;
  std::unordered_map<std::string, cudaGraphNode_t> nodes = {};
};
 
int add_clear_array_node_impl(
  void* dst, const size_t num_elements, const size_t sizeofT,
  cudaGraphNode_t& node_out, cudaGraph_t& graph,
  const std::vector<cudaGraphNode_t>& dependencies);
 
int add_copy_node_impl(
  const void* src, void* dst, const size_t num_elements, const size_t sizeofT,
  cudaMemcpyKind kind, cudaGraphNode_t& node_out, cudaGraph_t& graph,
  const std::vector<cudaGraphNode_t>& dependencies);
 
template <typename T>
int add_clear_array_node(
  T* dst, const size_t num_elements,
  cudaGraphNode_t& node_out, cudaGraph_t& graph,
  const std::vector<cudaGraphNode_t>& dependencies) {
  return add_clear_array_node_impl(
    dst, num_elements, sizeof(T), node_out, graph, dependencies);
}
 
template <typename T>
int add_copy_node(
  const T* src, T* dst, size_t num_elements,
  cudaMemcpyKind kind, cudaGraphNode_t& node_out, cudaGraph_t& graph,
  const std::vector<cudaGraphNode_t>& dependencies) {
  return add_copy_node_impl(src, dst, num_elements, sizeof(T),
                            kind, node_out, graph, dependencies);
}
 
int prepare_dependencies(
  const std::vector<std::pair<std::string, bool>>& node_names,
  std::vector<cudaGraphNode_t>& dependencies,
  const std::unordered_map<std::string, cudaGraphNode_t>& map,
  const std::string& caller_operation_name = "");
 
#endif // defined(COLVARS_CUDA) || defined (COLVARS_HIP)
}
 
#endif // COLVAR_GPU_SUPPORT_H