Halide 19.0.0
Halide compiler and libraries
Loading...
Searching...
No Matches
cuda_functions.h File Reference

Go to the source code of this file.

Macros

#define CUDA_FN(ret, fn, args)
 
#define CUDA_FN_OPTIONAL(ret, fn, args)
 
#define CUDA_FN_3020(ret, fn, fn_3020, args)
 
#define CUDA_FN_4000(ret, fn, fn_4000, args)
 

Functions

 CUDA_FN (CUresult, cuInit,(unsigned int Flags))
 
 CUDA_FN (CUresult, cuDeviceGetCount,(int *count))
 
 CUDA_FN (CUresult, cuDeviceGet,(CUdevice *device, int ordinal))
 
 CUDA_FN (CUresult, cuDeviceGetAttribute,(int *, CUdevice_attribute attrib, CUdevice dev))
 
 CUDA_FN (CUresult, cuDeviceGetName,(char *, int len, CUdevice dev))
 
 CUDA_FN_3020 (CUresult, cuDeviceTotalMem, cuDeviceTotalMem_v2,(size_t *, CUdevice dev))
 
 CUDA_FN_3020 (CUresult, cuCtxCreate, cuCtxCreate_v2,(CUcontext *pctx, unsigned int flags, CUdevice dev))
 
 CUDA_FN_4000 (CUresult, cuCtxDestroy, cuCtxDestroy_v2,(CUcontext pctx))
 
 CUDA_FN (CUresult, cuProfilerStop,())
 
 CUDA_FN (CUresult, cuCtxGetApiVersion,(CUcontext ctx, unsigned int *version))
 
 CUDA_FN (CUresult, cuCtxGetDevice,(CUdevice *))
 
 CUDA_FN (CUresult, cuModuleLoadData,(CUmodule *module, const void *image))
 
 CUDA_FN (CUresult, cuModuleLoadDataEx,(CUmodule *module, const void *image, unsigned int numOptions, CUjit_option *options, void **optionValues))
 
 CUDA_FN (CUresult, cuModuleUnload,(CUmodule module))
 
 CUDA_FN (CUresult, cuModuleGetFunction,(CUfunction *hfunc, CUmodule hmod, const char *name))
 
 CUDA_FN_3020 (CUresult, cuMemAlloc, cuMemAlloc_v2,(CUdeviceptr *dptr, size_t bytesize))
 
 CUDA_FN_3020 (CUresult, cuMemFree, cuMemFree_v2,(CUdeviceptr dptr))
 
 CUDA_FN_3020 (CUresult, cuMemcpyHtoD, cuMemcpyHtoD_v2,(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount))
 
 CUDA_FN_3020 (CUresult, cuMemcpyDtoH, cuMemcpyDtoH_v2,(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount))
 
 CUDA_FN_3020 (CUresult, cuMemcpyDtoD, cuMemcpyDtoD_v2,(CUdeviceptr dstHost, CUdeviceptr srcDevice, size_t ByteCount))
 
 CUDA_FN_3020 (CUresult, cuMemcpyHtoDAsync, cuMemcpyHtoDAsync_v2,(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount, CUstream stream))
 
 CUDA_FN_3020 (CUresult, cuMemcpyDtoHAsync, cuMemcpyDtoHAsync_v2,(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream stream))
 
 CUDA_FN_3020 (CUresult, cuMemcpyDtoDAsync, cuMemcpyDtoDAsync_v2,(CUdeviceptr dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream stream))
 
 CUDA_FN_3020 (CUresult, cuMemcpy3D, cuMemcpy3D_v2,(const CUDA_MEMCPY3D *pCopy))
 
 CUDA_FN (CUresult, cuLaunchKernel,(CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void **kernelParams, void **extra))
 
 CUDA_FN (CUresult, cuCtxSynchronize,())
 
 CUDA_FN_4000 (CUresult, cuCtxPushCurrent, cuCtxPushCurrent_v2,(CUcontext ctx))
 
 CUDA_FN_4000 (CUresult, cuCtxPopCurrent, cuCtxPopCurrent_v2,(CUcontext *pctx))
 
 CUDA_FN (CUresult, cuPointerGetAttribute,(void *result, int query, CUdeviceptr ptr))
 
 CUDA_FN_OPTIONAL (CUresult, cuStreamSynchronize,(CUstream hStream))
 

Macro Definition Documentation

◆ CUDA_FN

#define CUDA_FN ( ret,
fn,
args )

Definition at line 7 of file cuda_functions.h.

◆ CUDA_FN_OPTIONAL

#define CUDA_FN_OPTIONAL ( ret,
fn,
args )

Definition at line 10 of file cuda_functions.h.

◆ CUDA_FN_3020

#define CUDA_FN_3020 ( ret,
fn,
fn_3020,
args )
Value:
CUDA_FN(ret, fn, args)
#define CUDA_FN(ret, fn, args)

Definition at line 13 of file cuda_functions.h.

◆ CUDA_FN_4000

#define CUDA_FN_4000 ( ret,
fn,
fn_4000,
args )
Value:
CUDA_FN(ret, fn, args)

Definition at line 16 of file cuda_functions.h.

Function Documentation

◆ CUDA_FN() [1/15]

CUDA_FN ( CUresult ,
cuInit ,
(unsigned int Flags)  )

◆ CUDA_FN() [2/15]

CUDA_FN ( CUresult ,
cuDeviceGetCount ,
(int *count)  )

◆ CUDA_FN() [3/15]

CUDA_FN ( CUresult ,
cuDeviceGet ,
(CUdevice *device, int ordinal)  )

◆ CUDA_FN() [4/15]

CUDA_FN ( CUresult ,
cuDeviceGetAttribute ,
(int *, CUdevice_attribute attrib, CUdevice dev)  )

◆ CUDA_FN() [5/15]

CUDA_FN ( CUresult ,
cuDeviceGetName ,
(char *, int len, CUdevice dev)  )

◆ CUDA_FN_3020() [1/11]

CUDA_FN_3020 ( CUresult ,
cuDeviceTotalMem ,
cuDeviceTotalMem_v2 ,
(size_t *, CUdevice dev)  )

◆ CUDA_FN_3020() [2/11]

CUDA_FN_3020 ( CUresult ,
cuCtxCreate ,
cuCtxCreate_v2 ,
(CUcontext *pctx, unsigned int flags, CUdevice dev)  )

◆ CUDA_FN_4000() [1/3]

CUDA_FN_4000 ( CUresult ,
cuCtxDestroy ,
cuCtxDestroy_v2 ,
(CUcontext pctx)  )

◆ CUDA_FN() [6/15]

CUDA_FN ( CUresult ,
cuProfilerStop ,
()  )

◆ CUDA_FN() [7/15]

CUDA_FN ( CUresult ,
cuCtxGetApiVersion ,
(CUcontext ctx, unsigned int *version)  )

◆ CUDA_FN() [8/15]

CUDA_FN ( CUresult ,
cuCtxGetDevice ,
(CUdevice *)  )

◆ CUDA_FN() [9/15]

CUDA_FN ( CUresult ,
cuModuleLoadData ,
(CUmodule *module, const void *image)  )

◆ CUDA_FN() [10/15]

CUDA_FN ( CUresult ,
cuModuleLoadDataEx ,
(CUmodule *module, const void *image, unsigned int numOptions, CUjit_option *options, void **optionValues)  )

◆ CUDA_FN() [11/15]

CUDA_FN ( CUresult ,
cuModuleUnload ,
(CUmodule module)  )

◆ CUDA_FN() [12/15]

CUDA_FN ( CUresult ,
cuModuleGetFunction ,
(CUfunction *hfunc, CUmodule hmod, const char *name)  )

◆ CUDA_FN_3020() [3/11]

CUDA_FN_3020 ( CUresult ,
cuMemAlloc ,
cuMemAlloc_v2 ,
(CUdeviceptr *dptr, size_t bytesize)  )

◆ CUDA_FN_3020() [4/11]

CUDA_FN_3020 ( CUresult ,
cuMemFree ,
cuMemFree_v2 ,
(CUdeviceptr dptr)  )

◆ CUDA_FN_3020() [5/11]

CUDA_FN_3020 ( CUresult ,
cuMemcpyHtoD ,
cuMemcpyHtoD_v2 ,
(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount)  )

◆ CUDA_FN_3020() [6/11]

CUDA_FN_3020 ( CUresult ,
cuMemcpyDtoH ,
cuMemcpyDtoH_v2 ,
(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount)  )

◆ CUDA_FN_3020() [7/11]

CUDA_FN_3020 ( CUresult ,
cuMemcpyDtoD ,
cuMemcpyDtoD_v2 ,
(CUdeviceptr dstHost, CUdeviceptr srcDevice, size_t ByteCount)  )

◆ CUDA_FN_3020() [8/11]

CUDA_FN_3020 ( CUresult ,
cuMemcpyHtoDAsync ,
cuMemcpyHtoDAsync_v2 ,
(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount, CUstream stream)  )

◆ CUDA_FN_3020() [9/11]

CUDA_FN_3020 ( CUresult ,
cuMemcpyDtoHAsync ,
cuMemcpyDtoHAsync_v2 ,
(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream stream)  )

◆ CUDA_FN_3020() [10/11]

CUDA_FN_3020 ( CUresult ,
cuMemcpyDtoDAsync ,
cuMemcpyDtoDAsync_v2 ,
(CUdeviceptr dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream stream)  )

◆ CUDA_FN_3020() [11/11]

CUDA_FN_3020 ( CUresult ,
cuMemcpy3D ,
cuMemcpy3D_v2 ,
(const CUDA_MEMCPY3D *pCopy)  )

◆ CUDA_FN() [13/15]

CUDA_FN ( CUresult ,
cuLaunchKernel ,
(CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void **kernelParams, void **extra)  )

◆ CUDA_FN() [14/15]

CUDA_FN ( CUresult ,
cuCtxSynchronize ,
()  )

◆ CUDA_FN_4000() [2/3]

CUDA_FN_4000 ( CUresult ,
cuCtxPushCurrent ,
cuCtxPushCurrent_v2 ,
(CUcontext ctx)  )

◆ CUDA_FN_4000() [3/3]

CUDA_FN_4000 ( CUresult ,
cuCtxPopCurrent ,
cuCtxPopCurrent_v2 ,
(CUcontext *pctx)  )

◆ CUDA_FN() [15/15]

CUDA_FN ( CUresult ,
cuPointerGetAttribute ,
(void *result, int query, CUdeviceptr ptr)  )

◆ CUDA_FN_OPTIONAL()

CUDA_FN_OPTIONAL ( CUresult ,
cuStreamSynchronize ,
(CUstream hStream)  )