math.h

Go to the documentation of this file.

#ifndef GHOST_MATH_H
#define GHOST_MATH_H

#include "config.h"
#include "types.h"
#include "context.h"
#include "sparsemat.h"
#include "densemat.h"
#include "error.h"
#include "tsmm.h"
#include "tsmm_inplace.h"
#include "dot.h"

#include <stdarg.h>

#define SELECT_BLAS1_KERNEL(kernels,commonlocation,compute_at,storage,ret,...) \
    if (commonlocation == (GHOST_LOCATION_HOST | GHOST_LOCATION_DEVICE)) {\
        if (compute_at == GHOST_LOCATION_HOST) {\
            if (storage == GHOST_DENSEMAT_COLMAJOR) {\
                ret = kernels[GHOST_HOST_IDX][GHOST_CM_IDX](__VA_ARGS__);\
            } else {\
                ret = kernels[GHOST_HOST_IDX][GHOST_RM_IDX](__VA_ARGS__);\
            }\
        } else {\
            if (storage == GHOST_DENSEMAT_COLMAJOR) {\
                ret = kernels[GHOST_DEVICE_IDX][GHOST_CM_IDX](__VA_ARGS__);\
            } else {\
                ret = kernels[GHOST_DEVICE_IDX][GHOST_RM_IDX](__VA_ARGS__);\
            }\
        }\
    } else {\
        if (!kernels[ghost_idx_of_location(commonlocation)][ghost_idx_of_densemat_storage(storage)]) {\
            GHOST_ERROR_LOG("No kernel found!");\
            ret = GHOST_ERR_INVALID_ARG;\
        } else {\
            ret = kernels[ghost_idx_of_location(commonlocation)][ghost_idx_of_densemat_storage(storage)](__VA_ARGS__);\
        }\
    }

typedef enum {
    GHOST_GEMM_DEFAULT = 0,
    GHOST_GEMM_NOT_SPECIAL = 1,
    GHOST_GEMM_NOT_CLONE_ALIASED = 2,
    GHOST_GEMM_KAHAN = 4
} ghost_gemm_flags;

#ifdef __cplusplus
inline ghost_gemm_flags operator&=(ghost_gemm_flags &a, const ghost_gemm_flags &b) {
    a = static_cast<ghost_gemm_flags>(static_cast<int>(a) & static_cast<int>(b));
    return a;
}
inline ghost_gemm_flags operator~(const ghost_gemm_flags &a) {
    return static_cast<ghost_gemm_flags>(~static_cast<int>(a));
}
#endif

typedef struct {
    /* C = alpha(A*B) + beta(C)
     * C is mxn
     * A is mxk
     * B is kxn
     */ 
    ghost_gidx m,n,k;
    bool alphaisone;
    bool betaiszero;
    bool aisc;
    ghost_datatype dt;
}
ghost_gemm_perf_args;

#define GHOST_GEMM_ALL_REDUCE GHOST_ALLREDUCE
#define GHOST_GEMM_NO_REDUCE -2

typedef struct {
    ghost_lidx vecncols;
    ghost_lidx globalrows;
    ghost_gidx globalnnz;
    ghost_datatype dt;
}
ghost_kacz_perf_args;
#define GHOST_KACZ_PERF_UNIT "GF/s"
#define GHOST_KACZ_PERF_TAG "kacz"

typedef struct {
    ghost_lidx vecncols;
    ghost_lidx globalrows;
    ghost_gidx globalnnz;
    ghost_datatype dt;
    ghost_spmv_flags flags;
}
ghost_spmv_perf_args;
#define GHOST_SPMV_PERF_UNIT "GF/s"
#define GHOST_SPMV_PERF_TAG "spmv"

typedef struct {
    ghost_lidx ncols;
    ghost_gidx globnrows;
    ghost_datatype dt;
}
ghost_axpy_perf_args;
#define GHOST_AXPY_PERF_UNIT "GB/s"
#define GHOST_AXPY_PERF_TAG "axpy"

typedef struct {
    ghost_lidx ncols;
    ghost_gidx globnrows;
    ghost_datatype dt;
}
ghost_axpby_perf_args;
#define GHOST_AXPBY_PERF_UNIT "GB/s"
#define GHOST_AXPBY_PERF_TAG "axpby"

typedef struct {
    ghost_lidx ncols;
    ghost_gidx globnrows;
    ghost_datatype dt;
}
ghost_axpbypcz_perf_args;
#define GHOST_AXPBYPCZ_PERF_UNIT "GB/s"
#define GHOST_AXPBYPCZ_PERF_TAG "axpbypcz"

typedef struct {
    ghost_lidx ncols;
    ghost_gidx globnrows;
    ghost_datatype dt;
    bool samevec;
}
ghost_dot_perf_args;
#define GHOST_DOT_PERF_UNIT "GB/s"
#define GHOST_DOT_PERF_TAG "dot"

typedef struct {
    ghost_lidx ncols;
    ghost_gidx globnrows;
    ghost_datatype dt;
}
ghost_scale_perf_args;
#define GHOST_SCALE_PERF_UNIT "GB/s"
#define GHOST_SCALE_PERF_TAG "scale"


#ifdef __cplusplus
extern "C" {
#endif

    ghost_error ghost_spmv(ghost_densemat *res, ghost_sparsemat *mat, ghost_densemat *invec, ghost_spmv_opts opts);
    ghost_error ghost_spmv_nocomm(ghost_densemat *res, ghost_sparsemat *mat, ghost_densemat *invec, ghost_spmv_opts opts);
    ghost_error ghost_gemm_valid(ghost_densemat *x, ghost_densemat *v, const char * transv, 
    ghost_densemat *w, const char *transw, void *alpha, void *beta, int reduce,ghost_gemm_flags flags, int printerror); 
    ghost_error ghost_gemm(ghost_densemat *x, ghost_densemat *v, const char *transv, ghost_densemat *w, const char * transw, void *alpha, void *beta, int reduce,ghost_gemm_flags flags); 
    ghost_error ghost_mpi_operations_create();
    ghost_error ghost_mpi_operations_destroy();
    ghost_error ghost_mpi_op_densemat_sum(ghost_mpi_op * op, ghost_datatype datatype);
    ghost_error ghost_mpi_op_sum(ghost_mpi_op * op, ghost_datatype datatype);
    
    ghost_error ghost_spmv_nflops(int *nFlops, ghost_datatype m_t, ghost_datatype v_t);


    ghost_error ghost_axpy(ghost_densemat *y, ghost_densemat *x, void *a);
    ghost_error ghost_vaxpy(ghost_densemat *y, ghost_densemat *x, void *a);
    ghost_error ghost_axpby(ghost_densemat *y, ghost_densemat *x, void *a, void *b);
    ghost_error ghost_vaxpby(ghost_densemat *y, ghost_densemat *x, void *a, void *b);
    ghost_error ghost_axpbypcz(ghost_densemat *y, ghost_densemat *x, void *a, void *b, ghost_densemat *z, void *c);
    ghost_error ghost_vaxpbypcz(ghost_densemat *y, ghost_densemat *x, void *a, void *b, ghost_densemat *z, void *c); 

    ghost_error ghost_scale(ghost_densemat *x, void *s);
    ghost_error ghost_vscale(ghost_densemat *x, void *s);
    ghost_error ghost_normalize(ghost_densemat *x);
    ghost_error ghost_conj(ghost_densemat *x);
    
    ghost_error ghost_norm(void *norm, ghost_densemat *vec, void *pow);
    
    ghost_error ghost_nrm2(void *norm, ghost_densemat *vec);
    ghost_error ghost_mult(ghost_densemat *y, ghost_densemat *x, void *a);

    int ghost_kacz_perf(double *perf, double time, void *arg);
    int ghost_spmv_perf(double *perf, double time, void *arg);
    int ghost_axpy_perf(double *perf, double time, void *arg);
    int ghost_axpby_perf(double *perf, double time, void *arg);
    int ghost_scale_perf(double *perf, double time, void *arg);
    int ghost_dot_perf(double *perf, double time, void *arg);
    int ghost_axpbypcz_perf(double *perf, double time, void *arg);
    int ghost_gemm_perf_GFs(double *perf, double time, void *arg);
    int ghost_gemm_perf_GBs(double *perf, double time, void *arg);

    bool ghost_iszero(void *number, ghost_datatype dt);
    bool ghost_isone(void *vnumber, ghost_datatype dt);

#ifdef __cplusplus
} //extern "C"

#endif

#endif