18#ifndef __DRIVER_CUDA_H__
19#define __DRIVER_CUDA_H__
23#include <common/config.h>
27#include <cuda_runtime_api.h>
28#ifdef STARPU_HAVE_LIBNVIDIA_ML
34#include <core/workers.h>
40void _starpu_cuda_init(
void);
41unsigned _starpu_get_cuda_device_count(
void);
42extern int _starpu_cuda_bus_ids[STARPU_MAXCUDADEVS+STARPU_MAXNUMANODES][STARPU_MAXCUDADEVS+STARPU_MAXNUMANODES];
44#if defined(STARPU_USE_CUDA) || defined(STARPU_SIMGRID)
46void _starpu_init_cuda(
void);
47void _starpu_init_cublas_v2_func(
void);
48void _starpu_shutdown_cublas_v2_func(
void);
49void _starpu_cublas_v2_init(
void);
50void _starpu_cublas_v2_shutdown(
void);
51void *_starpu_cuda_worker(
void *);
52#ifdef STARPU_HAVE_LIBNVIDIA_ML
53nvmlDevice_t _starpu_cuda_get_nvmldev(
struct cudaDeviceProp *props);
56# define _starpu_cuda_discover_devices(config) ((void) config)
60cudaStream_t starpu_cuda_get_local_in_transfer_stream(
void);
61cudaStream_t starpu_cuda_get_in_transfer_stream(
unsigned dst_node);
62cudaStream_t starpu_cuda_get_local_out_transfer_stream(
void);
63cudaStream_t starpu_cuda_get_out_transfer_stream(
unsigned src_node);
64cudaStream_t starpu_cuda_get_peer_transfer_stream(
unsigned src_node,
unsigned dst_node);
70int _starpu_cuda_copy_interface_from_cpu_to_cuda(starpu_data_handle_t handle,
void *src_interface,
unsigned src_node,
void *dst_interface,
unsigned dst_node,
struct _starpu_data_request *req);
71int _starpu_cuda_copy_interface_from_cuda_to_cuda(starpu_data_handle_t handle,
void *src_interface,
unsigned src_node,
void *dst_interface,
unsigned dst_node,
struct _starpu_data_request *req);
72int _starpu_cuda_copy_interface_from_cuda_to_cpu(starpu_data_handle_t handle,
void *src_interface,
unsigned src_node,
void *dst_interface,
unsigned dst_node,
struct _starpu_data_request *req);
74int _starpu_cuda_copy_data_from_cuda_to_cuda(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t size,
struct _starpu_async_channel *async_channel);
75int _starpu_cuda_copy_data_from_cuda_to_cpu(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t size,
struct _starpu_async_channel *async_channel);
76int _starpu_cuda_copy_data_from_cpu_to_cuda(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t size,
struct _starpu_async_channel *async_channel);
78int _starpu_cuda_copy2d_data_from_cuda_to_cuda(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t blocksize,
size_t numblocks,
size_t ld_src,
size_t ld_dst,
struct _starpu_async_channel *async_channel);
79int _starpu_cuda_copy2d_data_from_cuda_to_cpu(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t blocksize,
size_t numblocks,
size_t ld_src,
size_t ld_dst,
struct _starpu_async_channel *async_channel);
80int _starpu_cuda_copy2d_data_from_cpu_to_cuda(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t blocksize,
size_t numblocks,
size_t ld_src,
size_t ld_dst,
struct _starpu_async_channel *async_channel);
82int _starpu_cuda_copy3d_data_from_cuda_to_cuda(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t blocksize,
size_t numblocks_1,
size_t ld1_src,
size_t ld1_dst,
size_t numblocks_2,
size_t ld2_src,
size_t ld2_dst,
struct _starpu_async_channel *async_channel);
83int _starpu_cuda_copy3d_data_from_cuda_to_cpu(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t blocksize,
size_t numblocks_1,
size_t ld1_src,
size_t ld1_dst,
size_t numblocks_2,
size_t ld2_src,
size_t ld2_dst,
struct _starpu_async_channel *async_channel);
84int _starpu_cuda_copy3d_data_from_cpu_to_cuda(uintptr_t src,
size_t src_offset,
unsigned src_node, uintptr_t dst,
size_t dst_offset,
unsigned dst_node,
size_t blocksize,
size_t numblocks_1,
size_t ld1_src,
size_t ld1_dst,
size_t numblocks_2,
size_t ld2_src,
size_t ld2_dst,
struct _starpu_async_channel *async_channel);
86int _starpu_cuda_is_direct_access_supported(
unsigned node,
unsigned handling_node);
87uintptr_t _starpu_cuda_malloc_on_node(
unsigned dst_node,
size_t size,
int flags);
88void _starpu_cuda_free_on_node(
unsigned dst_node, uintptr_t addr,
size_t size,
int flags);
Definition copy_driver.h:127