cudacomp.h

Go to the documentation of this file.

#ifndef _CUDACOMP_H
#define _CUDACOMP_H


#ifdef HAVE_CUDA

#include <cuda_runtime_api.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <device_types.h>
#include <pthread.h>

#endif

#ifdef HAVE_CUDA

// data passed to each thread
typedef struct
{
    int thread_no;
    long numl0;
    int cindex; // computation index
    int_fast8_t *status; // where to white status

    // timers
    struct timespec *t0;
    struct timespec *t1;
    struct timespec *t2;
    struct timespec *t3;
    struct timespec *t4;
    struct timespec *t5;

} THDATA;



typedef struct
{
    int8_t    init;                   
    int8_t   *refWFSinit;            
    int8_t    alloc;                  
    imageID   CM_ID;
    uint64_t  CM_cnt;
    long      timerID;

    uint32_t M;
    uint32_t N;


    int_fast8_t
    sem;                    
    int_fast8_t gpuinit;

    sem_t **semptr1;                   
    sem_t **semptr2;                   
    sem_t **semptr3;                   
    sem_t **semptr4;                   
    sem_t **semptr5;                   
    // computer memory (host)
    float *cMat;
    float **cMat_part;
    float *wfsVec;
    float **wfsVec_part;
    float *wfsRef;
    float **wfsRef_part;
    float *dmVec;
    float *dmVecTMP;
    float **dmVec_part;
    float **dmRef_part;

    // GPU memory (device)
    float **d_cMat;
    float **d_wfsVec;
    float **d_dmVec;
    float **d_wfsRef;
    float **d_dmRef;

    // threads
    THDATA *thdata;
    int *iret;
    pthread_t *threadarray;
    int      NBstreams;
    cudaStream_t *stream;
    cublasHandle_t *handle;

    // splitting limits
    uint_fast32_t *Nsize;
    uint_fast32_t *Noffset;

    int *GPUdevice;

    int_fast8_t orientation;

    long IDout;


} GPUMATMULTCONF;
#endif






/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */


void __attribute__((constructor)) libinit_cudacomp();

//errno_t init_cudacomp();


int CUDACOMP_init();

int CUDACOMP_printGPUMATMULTCONF(int index);

errno_t GPUcomp_test(long NBact, long NBmodes, long WFSsize, long GPUcnt);






#ifdef HAVE_CUDA

/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */



void matrixMulCPU(float *cMat, float *wfsVec, float *dmVec, int M, int N);


void *compute_function(void *ptr);


int GPUloadCmat(int index);


int GPU_loop_MultMat_setup(
    int         index,
    const char *IDcontrM_name,
    const char *IDwfsim_name,
    const char *IDoutdmmodes_name,
    long        NBGPUs,
    int        *GPUdevice,
    int         orientation,
    int         USEsem,
    int         initWFSref,
    long        loopnb
);


int GPU_loop_MultMat_execute(
    int          index,
    int_fast8_t *status,
    int_fast8_t *GPUstatus,
    float        alpha,
    float        beta,
    int          timing,
    int          TimerOffsetIndex
);


int GPU_loop_MultMat_free(int index);




#ifdef HAVE_MAGMA
/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */


long CUDACOMP_MatMatMult_testPseudoInverse(
    const char *IDmatA_name,
    const char *IDmatAinv_name,
    const char *IDmatOut_name
);



int CUDACOMP_magma_compute_SVDpseudoInverse_SVD(
    const char  *ID_Rmatrix_name,
    const char  *ID_Cmatrix_name,
    double       SVDeps,
    long         MaxNBmodes,
    const char  *ID_VTmatrix_name
);



int CUDACOMP_magma_compute_SVDpseudoInverse(
    const char *ID_Rmatrix_name,
    const char *ID_Cmatrix_name,
    double SVDeps,
    long MaxNBmodes,
    const char *ID_VTmatrix_name,
    int LOOPmode,
    int PSINV_MODE,
    double qdwh_s,
    float qdwh_tol,
    int testmode
);



int GPU_SVD_computeControlMatrix(
    int         device,
    const char *ID_Rmatrix_name,
    const char *ID_Cmatrix_name,
    double      SVDeps,
    const char *ID_VTmatrix_name
);


#endif



/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */
/* =============================================================================================== */


int CUDACOMP_Coeff2Map_Loop(const char *IDmodes_name, const char *IDcoeff_name,
                            int GPUindex, const char *IDoutmap_name, int offsetmode,
                            const char *IDoffset_name);


errno_t CUDACOMP_MVMextractModesLoop_FPCONF();
errno_t CUDACOMP_MVMextractModesLoop_RUN();

int CUDACOMP_MVMextractModesLoop(
    const char *in_stream,
    const char *intot_stream,
    const char *IDmodes_name,
    const char *IDrefin_name,
    const char *IDrefout_name,
    const char *IDmodes_val_name,
    int         GPUindex,
    int         PROCESS,
    int         TRACEMODE,
    int         MODENORM,
    int         insem,
    int         axmode,
    long        twait,
    int         semwarn
);



#endif


#endif