cupti_common.h File Reference

Include dependency graph for cupti_common.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Macros
#define	DLSYM_AND_CHECK(dllib, name)
#define	CUDA_CALL(call, handleerror)
#define	CUDART_CALL(call, handleerror)
#define	CUPTI_CALL(call, handleerror)

Functions
void	cuptic_disabled_reason_set (const char *msg)
void	cuptic_disabled_reason_get (const char **pmsg)
void *	cuptic_load_dynamic_syms (const char *parent_path, const char *dlname, const char *search_subpaths[])
int	cuptic_shutdown (void)
int	cuptic_device_get_count (int *num_gpus)
int	cuptic_init (void)
int	cuptic_is_runtime_perfworks_api (void)
int	cuptic_is_runtime_events_api (void)
int	cuptic_ctxarr_create (cuptic_info_t *pinfo)
int	cuptic_ctxarr_update_current (cuptic_info_t info)
int	cuptic_ctxarr_get_ctx (cuptic_info_t info, int gpu_idx, CUcontext *ctx)
int	cuptic_ctxarr_destroy (cuptic_info_t *pinfo)
int	cuptic_device_acquire (cuptiu_event_table_t *evt_table)
int	cuptic_device_release (cuptiu_event_table_t *evt_table)

Variables
const char *	linked_cudart_path
void *	dl_cupti
unsigned int	_cuda_lock
CUresult(*	cuCtxGetCurrentPtr )(CUcontext *)
CUresult(*	cuCtxSetCurrentPtr )(CUcontext)
CUresult(*	cuCtxDestroyPtr )(CUcontext)
CUresult(*	cuCtxCreatePtr )(CUcontext *pctx, unsigned int flags, CUdevice dev)
CUresult(*	cuCtxGetDevicePtr )(CUdevice *)
CUresult(*	cuDeviceGetPtr )(CUdevice *, int)
CUresult(*	cuDeviceGetCountPtr )(int *)
CUresult(*	cuDeviceGetNamePtr )(char *, int, CUdevice)
CUresult(*	cuDevicePrimaryCtxRetainPtr )(CUcontext *pctx, CUdevice)
CUresult(*	cuDevicePrimaryCtxReleasePtr )(CUdevice)
CUresult(*	cuInitPtr )(unsigned int)
CUresult(*	cuGetErrorStringPtr )(CUresult error, const char **pStr)
CUresult(*	cuCtxPopCurrentPtr )(CUcontext *pctx)
CUresult(*	cuCtxPushCurrentPtr )(CUcontext pctx)
CUresult(*	cuCtxSynchronizePtr )()
CUresult(*	cuDeviceGetAttributePtr )(int *, CUdevice_attribute, CUdevice)
cudaError_t(*	cudaGetDeviceCountPtr )(int *)
cudaError_t(*	cudaGetDevicePtr )(int *)
cudaError_t(*	cudaSetDevicePtr )(int)
cudaError_t(*	cudaGetDevicePropertiesPtr )(struct cudaDeviceProp *prop, int device)
cudaError_t(*	cudaDeviceGetAttributePtr )(int *value, enum cudaDeviceAttr attr, int device)
cudaError_t(*	cudaFreePtr )(void *)
cudaError_t(*	cudaDriverGetVersionPtr )(int *)
cudaError_t(*	cudaRuntimeGetVersionPtr )(int *)
CUptiResult(*	cuptiGetVersionPtr )(uint32_t *)

Detailed Description

Author

Anustuv Pal anust.nosp@m.uv@i.nosp@m.cl.ut.nosp@m.k.ed.nosp@m.u

Definition in file cupti_common.h.

Macro Definition Documentation

CUDA_CALL

#define CUDA_CALL	(		call,
			handleerror
	)

Value:

    do {  \
        CUresult _status = (call);  \
        LOGCUDACALL("\t" #call "\n");  \
        if (_status != CUDA_SUCCESS) {  \
            ERRDBG("CUDA Error %d: Error in call to " #call "\n", _status);  \
            EXIT_OR_NOT; \
            handleerror;  \
        }  \
    } while (0);

Definition at line 58 of file cupti_common.h.

CUDART_CALL

#define CUDART_CALL	(		call,
			handleerror
	)

Value:

    do {  \
        cudaError_t _status = (call);  \
        LOGCUDACALL("\t" #call "\n");  \
        if (_status != cudaSuccess) {  \
            ERRDBG("CUDART Error %d: Error in call to " #call "\n", _status);  \
            EXIT_OR_NOT; \
            handleerror;  \
        }  \
    } while (0);

Definition at line 68 of file cupti_common.h.

CUPTI_CALL

#define CUPTI_CALL	(		call,
			handleerror
	)

Value:

    do {  \
        CUptiResult _status = (call);  \
        LOGCUPTICALL("\t" #call "\n");  \
        if (_status != CUPTI_SUCCESS) {  \
            ERRDBG("CUPTI Error %d: Error in call to " #call "\n", _status);  \
            EXIT_OR_NOT; \
            handleerror;  \
        }  \
    } while (0);

Definition at line 78 of file cupti_common.h.

DLSYM_AND_CHECK

#define DLSYM_AND_CHECK	(		dllib,
			name
	)

Value:

    dlsym( dllib, name );  \
    if (dlerror() != NULL) {  \
        ERRDBG("A CUDA required function '%s' was not found in lib '%s'.\n", name, #dllib);  \
        return PAPI_EMISC;  \
    }

Definition at line 52 of file cupti_common.h.

Function Documentation

cuptic_ctxarr_create()

int cuptic_ctxarr_create

(

cuptic_info_t *

pinfo

)

Definition at line 520 of file cupti_common.c.

{
    COMPDBG("Entering.\n");
    int total_gpus;
    int papi_errno = cuptic_device_get_count(&total_gpus);
    if (papi_errno != PAPI_OK) {
        return PAPI_EMISC;
    }
    cuptic_info_t cuCtx = (cuptic_info_t) papi_calloc (total_gpus, sizeof(*pinfo));
    if (cuCtx == NULL) {
        return PAPI_ENOMEM;
    }
    *pinfo = cuCtx;
    return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_ctxarr_destroy()

int cuptic_ctxarr_destroy

(

cuptic_info_t *

pinfo

)

Definition at line 572 of file cupti_common.c.

{
    papi_free(*pinfo);
    *pinfo = NULL;
    return PAPI_OK;
}

Here is the caller graph for this function:

cuptic_ctxarr_get_ctx()

int cuptic_ctxarr_get_ctx	(	cuptic_info_t	info,
		int	gpu_idx,
		CUcontext *	ctx
	)

Definition at line 566 of file cupti_common.c.

{
    *ctx = info[gpu_idx].ctx;
    return PAPI_OK;
}

Here is the caller graph for this function:

cuptic_ctxarr_update_current()

int cuptic_ctxarr_update_current

(

cuptic_info_t

info

)

Definition at line 536 of file cupti_common.c.

{
    int papi_errno, gpu_id;
    CUcontext tempCtx;
    papi_errno = cudaGetDevicePtr(&gpu_id);
    if (papi_errno != cudaSuccess) {
        return PAPI_EMISC;
    }
    papi_errno = cuCtxGetCurrentPtr(&tempCtx);
    if (papi_errno != CUDA_SUCCESS) {
        return PAPI_EMISC;
    }
    if (info[gpu_id].ctx == NULL) {
        if (tempCtx != NULL) {
            LOGDBG("Registering device = %d with ctx = %p.\n", gpu_id, tempCtx);
            CUDA_CALL(cuCtxGetCurrentPtr(&info[gpu_id].ctx), return PAPI_EMISC);
        }
        else {
            CUDART_CALL(cudaFreePtr(NULL), return PAPI_EMISC);
            CUDA_CALL(cuCtxGetCurrentPtr(&info[gpu_id].ctx), return PAPI_EMISC);
            LOGDBG("Using primary device context %p for device %d.\n", info[gpu_id].ctx, gpu_id);
        }
    }
    /* If context has changed keep the first seen one but with warning */
    else if (info[gpu_id].ctx != tempCtx) {
        ERRDBG("Warning: cuda context for gpu %d has changed from %p to %p\n", gpu_id, info[gpu_id].ctx, tempCtx);
    }
    return PAPI_OK;
}

Here is the caller graph for this function:

cuptic_device_acquire()

int cuptic_device_acquire

(

cuptiu_event_table_t *

evt_table

)

Definition at line 629 of file cupti_common.c.

{
    gpu_occupancy_t bitmask;
    int papi_errno = _devmask_events_get(evt_table, &bitmask);
    if (papi_errno != PAPI_OK)
        return papi_errno;
    if (bitmask & global_gpu_bitmask) {
        return PAPI_ECNFLCT;
    }
    _papi_hwi_lock(_cuda_lock);
    global_gpu_bitmask |= bitmask;
    _papi_hwi_unlock(_cuda_lock);
    return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_device_get_count()

int cuptic_device_get_count

(

int *

num_gpus

)

Definition at line 303 of file cupti_common.c.

{
    cudaError_t cuda_errno = cudaGetDeviceCountPtr(num_gpus);
    if (cuda_errno != cudaSuccess) {
        cuptic_disabled_reason_set(cudaGetErrorStringPtr(cuda_errno));
        return PAPI_EMISC;
    }
    return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_device_release()

int cuptic_device_release

(

cuptiu_event_table_t *

evt_table

)

Definition at line 644 of file cupti_common.c.

{
    gpu_occupancy_t bitmask;
    int papi_errno = _devmask_events_get(evt_table, &bitmask);
    if (papi_errno != PAPI_OK) {
        return papi_errno;
    }
    if ((bitmask & global_gpu_bitmask) != bitmask) {
        return PAPI_EMISC;
    }
    _papi_hwi_lock(_cuda_lock);
    global_gpu_bitmask ^= bitmask;
    _papi_hwi_unlock(_cuda_lock);
    return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_disabled_reason_get()

void cuptic_disabled_reason_get

(

const char **

pmsg

)

Definition at line 390 of file cupti_common.c.

{
    *pmsg = cuptic_disabled_reason_g;
}

Here is the caller graph for this function:

cuptic_disabled_reason_set()

void cuptic_disabled_reason_set

(

const char *

msg

)

Definition at line 385 of file cupti_common.c.

{
    cuptic_disabled_reason_g = msg;
}

Here is the caller graph for this function:

cuptic_init()

int cuptic_init

(

void

)

Definition at line 417 of file cupti_common.c.

{
    int papi_errno = get_user_cudart_path();
    if (papi_errno == PAPI_OK) {
        LOGDBG("Linked cudart root: %s\n", linked_cudart_path);
    }
    else {
        LOGDBG("Target application not linked with cuda runtime libraries.\n");
    }
    papi_errno = util_load_cuda_sym();
    if (papi_errno != PAPI_OK) {
        cuptic_disabled_reason_set("Unable to load CUDA library functions.");
        goto fn_exit;
    }
 
    gpu_collection_e kind;
    papi_errno = util_gpu_collection_kind(&kind);
    if (papi_errno != PAPI_OK) {
        goto fn_exit;
    }
 
    if (kind == GPU_COLLECTION_MIXED) {
        cuptic_disabled_reason_set("No support for systems with mixed compute capabilities, such as CC < 7.0 and CC > 7.0 GPUS.");
        papi_errno = PAPI_ECMP;
        goto fn_exit;
    }
fn_exit:
    return papi_errno;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_is_runtime_events_api()

int cuptic_is_runtime_events_api

(

void

)

Definition at line 488 of file cupti_common.c.

{
    static int is_events_api = -1;
    if (is_events_api != -1) {
        goto fn_exit;
    }
 
    gpu_collection_e gpus_kind;
    int papi_errno = util_gpu_collection_kind(&gpus_kind);
    if (papi_errno != PAPI_OK) {
        goto fn_exit;
    }
 
    /*
     * See cupti_config.h: When NVIDIA removes the events API add a check in the following condition
     * to check the `util_dylib_cupti_version()` is also <= CUPTI_EVENTS_API_MAX_SUPPORTED_VERSION.
     */
    if ((gpus_kind == GPU_COLLECTION_ALL_EVENTS || gpus_kind == GPU_COLLECTION_ALL_CC70)) {
        is_events_api = 1;
        goto fn_exit;
    } else {
        is_events_api = 0;
        goto fn_exit;
    }
fn_exit:
    return is_events_api;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_is_runtime_perfworks_api()

int cuptic_is_runtime_perfworks_api

(

void

)

Definition at line 447 of file cupti_common.c.

{
    static int is_perfworks_api = -1;
    if (is_perfworks_api != -1) {
        goto fn_exit;
    }
    char *papi_cuda_110_cc70_perfworks_api = getenv("PAPI_CUDA_110_CC_70_PERFWORKS_API");
 
    gpu_collection_e gpus_kind;
    int papi_errno = util_gpu_collection_kind(&gpus_kind);
    if (papi_errno != PAPI_OK) {
        goto fn_exit;
    }
 
    unsigned int cuptiVersion = util_dylib_cupti_version();
 
    if (gpus_kind == GPU_COLLECTION_ALL_CC70 && 
        (cuptiVersion == CUPTI_PROFILER_API_MIN_SUPPORTED_VERSION || util_dylib_cu_runtime_version() == 11000))
    {
        if (papi_cuda_110_cc70_perfworks_api != NULL) {
            is_perfworks_api = 1;
            goto fn_exit;
        }
        else {
            is_perfworks_api = 0;
            goto fn_exit;
        }
    }
 
    if ((gpus_kind == GPU_COLLECTION_ALL_PERF || gpus_kind == GPU_COLLECTION_ALL_CC70) && cuptiVersion >= CUPTI_PROFILER_API_MIN_SUPPORTED_VERSION) {
        is_perfworks_api = 1;
        goto fn_exit;
    } else {
        is_perfworks_api = 0;
        goto fn_exit;
    }
 
fn_exit:
    return is_perfworks_api;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_load_dynamic_syms()

void * cuptic_load_dynamic_syms	(	const char *	parent_path,
		const char *	dlname,
		const char *	search_subpaths[]
	)

Definition at line 110 of file cupti_common.c.

{
    void *dl = NULL;
    char lookup_path[PATH_MAX];
    char *found_files[CUPTIU_MAX_FILES];
    int i, count;
    for (i = 0; search_subpaths[i] != NULL; i++) {
        sprintf(lookup_path, search_subpaths[i], parent_path, dlname);
        dl = dlopen(lookup_path, RTLD_NOW | RTLD_GLOBAL);
        if (dl) {
            return dl;
        }
    }
    count = cuptiu_files_search_in_path(dlname, parent_path, found_files);
    for (i = 0; i < count; i++) {
        dl = dlopen(found_files[i], RTLD_NOW | RTLD_GLOBAL);
        if (dl) {
            break;
        }
    }
    for (i = 0; i < count; i++) {
        papi_free(found_files[i]);
    }
    return dl;
}

Here is the call graph for this function:

Here is the caller graph for this function:

cuptic_shutdown()

int cuptic_shutdown

(

void

)

Definition at line 280 of file cupti_common.c.

{
    unload_cuda_sym();
    unload_cudart_sym();
    unload_cupti_common_sym();
    unload_linked_cudart_path();
    return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

Variable Documentation

_cuda_lock

unsigned int _cuda_lock

extern

Definition at line 21 of file cupti_common.c.

cuCtxCreatePtr

CUresult(* cuCtxCreatePtr) (CUcontext *pctx, unsigned int flags, CUdevice dev) ( CUcontext *  pctx, unsigned int  flags, CUdevice  dev  )

extern

Definition at line 26 of file cupti_common.c.

cuCtxDestroyPtr

CUresult(* cuCtxDestroyPtr) (CUcontext) ( CUcontext  )

extern

Definition at line 25 of file cupti_common.c.

cuCtxGetCurrentPtr

CUresult(* cuCtxGetCurrentPtr) (CUcontext *) ( CUcontext *  )

extern

Definition at line 23 of file cupti_common.c.

cuCtxGetDevicePtr

CUresult(* cuCtxGetDevicePtr) (CUdevice *) ( CUdevice *  )

extern

Definition at line 27 of file cupti_common.c.

cuCtxPopCurrentPtr

CUresult(* cuCtxPopCurrentPtr) (CUcontext *pctx) ( CUcontext *  pctx )

extern

Definition at line 35 of file cupti_common.c.

cuCtxPushCurrentPtr

CUresult(* cuCtxPushCurrentPtr) (CUcontext pctx) ( CUcontext  pctx )

extern

Definition at line 36 of file cupti_common.c.

cuCtxSetCurrentPtr

CUresult(* cuCtxSetCurrentPtr) (CUcontext) ( CUcontext  )

extern

Definition at line 24 of file cupti_common.c.

cuCtxSynchronizePtr

CUresult(* cuCtxSynchronizePtr) () ( )

extern

Definition at line 37 of file cupti_common.c.

cudaDeviceGetAttributePtr

cudaError_t(* cudaDeviceGetAttributePtr) (int *value, enum cudaDeviceAttr attr, int device) ( int *  value, enum cudaDeviceAttr  attr, int  device  )

extern

Definition at line 45 of file cupti_common.c.

cudaDriverGetVersionPtr

cudaError_t(* cudaDriverGetVersionPtr) (int *) ( int *  )

extern

Definition at line 47 of file cupti_common.c.

cudaFreePtr

cudaError_t(* cudaFreePtr) (void *) ( void *  )

extern

Definition at line 46 of file cupti_common.c.

cudaGetDeviceCountPtr

cudaError_t(* cudaGetDeviceCountPtr) (int *) ( int *  )

extern

Definition at line 40 of file cupti_common.c.

cudaGetDevicePropertiesPtr

cudaError_t(* cudaGetDevicePropertiesPtr) (struct cudaDeviceProp *prop, int device) ( struct cudaDeviceProp *  prop, int  device  )

extern

Definition at line 44 of file cupti_common.c.

cudaGetDevicePtr

cudaError_t(* cudaGetDevicePtr) (int *) ( int *  )

extern

Definition at line 41 of file cupti_common.c.

cudaRuntimeGetVersionPtr

cudaError_t(* cudaRuntimeGetVersionPtr) (int *) ( int *  )

extern

Definition at line 48 of file cupti_common.c.

cudaSetDevicePtr

cudaError_t(* cudaSetDevicePtr) (int) ( int  )

extern

Definition at line 43 of file cupti_common.c.

cuDeviceGetAttributePtr

CUresult(* cuDeviceGetAttributePtr) (int *, CUdevice_attribute, CUdevice) ( int *  , CUdevice_attribute  , CUdevice    )

extern

Definition at line 38 of file cupti_common.c.

cuDeviceGetCountPtr

CUresult(* cuDeviceGetCountPtr) (int *) ( int *  )

extern

Definition at line 29 of file cupti_common.c.

cuDeviceGetNamePtr

CUresult(* cuDeviceGetNamePtr) (char *, int, CUdevice) ( char *  , int  , CUdevice    )

extern

Definition at line 30 of file cupti_common.c.

cuDeviceGetPtr

CUresult(* cuDeviceGetPtr) (CUdevice *, int) ( CUdevice *  , int    )

extern

Definition at line 28 of file cupti_common.c.

cuDevicePrimaryCtxReleasePtr

CUresult(* cuDevicePrimaryCtxReleasePtr) (CUdevice) ( CUdevice  )

extern

Definition at line 32 of file cupti_common.c.

cuDevicePrimaryCtxRetainPtr

CUresult(* cuDevicePrimaryCtxRetainPtr) (CUcontext *pctx, CUdevice) ( CUcontext *  pctx, CUdevice    )

extern

Definition at line 31 of file cupti_common.c.

cuGetErrorStringPtr

CUresult(* cuGetErrorStringPtr) (CUresult error, const char **pStr) ( CUresult  error, const char **  pStr  )

extern

Definition at line 34 of file cupti_common.c.

cuInitPtr

CUresult(* cuInitPtr) (unsigned int) ( unsigned int  )

extern

Definition at line 33 of file cupti_common.c.

cuptiGetVersionPtr

CUptiResult(* cuptiGetVersionPtr) (uint32_t *) ( uint32_t *  )

extern

Definition at line 50 of file cupti_common.c.

dl_cupti

void* dl_cupti

extern

Definition at line 19 of file cupti_common.c.

linked_cudart_path

const char* linked_cudart_path

extern

Definition at line 18 of file cupti_common.c.