cannon.c File Reference

#include "config-f90.h"
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <complex.h>
#include <stdint.h>
#include "../helpers/scalapack_interfaces.h"
#include "./gpu_vendor_agnostic_layer.h"
#include <mpi.h>
#include "../general/precision_macros.h"
#include "cannon_forw_template.h"
#include "cannon_back_template.h"

Macros
#define	C_INT_TYPE_PTR   int*
#define	C_INT_TYPE   int
#define	BLAS_KIND   c_int
#define	C_INT_MPI_TYPE_PTR   int*
#define	C_INT_MPI_TYPE   int
#define	MPI_KIND   c_int
#define	NVTX_RANGE_PUSH(msg)   ((void)0)
#define	NVTX_RANGE_POP()   ((void)0)
#define	gpuErrCheck(ans)   { gpuAssert((ans), __FILE__, __LINE__); }
#define	REALCASE   1
#define	DOUBLE_PRECISION   1
#define	cublasXgemm   cublasDgemm
#define	gpublasXgemm   gpublasDgemm
#define	REALCASE   1
#define	SINGLE_PRECISION   1
#define	cublasXgemm   cublasSgemm
#define	gpublasXgemm   gpublasSgemm
#define	COMPLEXCASE   1
#define	DOUBLE_PRECISION   1
#define	cublasXgemm   cublasZgemm
#define	gpublasXgemm   gpublasZgemm
#define	COMPLEXCASE   1
#define	SINGLE_PRECISION   1
#define	cublasXgemm   cublasCgemm
#define	gpublasXgemm   gpublasCgemm

Functions
static void	gpuAssert (int status, const char *file, int line)
void	cannons_reduction_c_d (double *A, double *U, int local_rowsCast, int local_colsCast, int *a_desc, double *Res, int ToStore, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)
void	cannons_triang_rectangular_c_d (double *U, double *B, int local_rowsCast, int local_colsCast, int *u_desc, int *b_desc, double *Res, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)
void	cannons_reduction_c_f (float *A, float *U, int local_rowsCast, int local_colsCast, int *a_desc, float *Res, int ToStore, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)
void	cannons_triang_rectangular_c_f (float *U, float *B, int local_rowsCast, int local_colsCast, int *u_desc, int *b_desc, float *Res, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)
void	cannons_reduction_c_dc (double complex *A, double complex *U, int local_rowsCast, int local_colsCasr, int *a_desc, double complex *Res, int ToStore, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)
void	cannons_triang_rectangular_c_dc (double complex *U, double complex *B, int local_rowsCast, int local_colsCast, int *u_desc, int *b_desc, double complex *Res, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)
void	cannons_reduction_c_fc (float complex *A, float complex *U, int local_rowsCast, int local_colsCast, int *a_desc, float complex *Res, int ToStore, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)
void	cannons_triang_rectangular_c_fc (float complex *U, float complex *B, int local_rowsCast, int local_colsCast, int *u_desc, int *b_desc, float complex *Res, int row_comm, int col_comm, int wantDebug, int useGPU, intptr_t *gpublasHandle)

Variables
int	gpuMemcpyHostToDevice
int	gpuMemcpyDeviceToHost

Macro Definition Documentation

BLAS_KIND

#define BLAS_KIND   c_int

C_INT_MPI_TYPE

#define C_INT_MPI_TYPE   int

C_INT_MPI_TYPE_PTR

#define C_INT_MPI_TYPE_PTR   int*

C_INT_TYPE

#define C_INT_TYPE   int

C_INT_TYPE_PTR

#define C_INT_TYPE_PTR   int*

COMPLEXCASE [1/2]

#define COMPLEXCASE   1

COMPLEXCASE [2/2]

#define COMPLEXCASE   1

cublasXgemm [1/4]

#define cublasXgemm   cublasDgemm

cublasXgemm [2/4]

#define cublasXgemm   cublasSgemm

cublasXgemm [3/4]

#define cublasXgemm   cublasZgemm

cublasXgemm [4/4]

#define cublasXgemm   cublasCgemm

DOUBLE_PRECISION [1/2]

#define DOUBLE_PRECISION   1

DOUBLE_PRECISION [2/2]

#define DOUBLE_PRECISION   1

gpublasXgemm [1/4]

#define gpublasXgemm   gpublasDgemm

gpublasXgemm [2/4]

#define gpublasXgemm   gpublasSgemm

gpublasXgemm [3/4]

#define gpublasXgemm   gpublasZgemm

gpublasXgemm [4/4]

#define gpublasXgemm   gpublasCgemm

gpuErrCheck

#define gpuErrCheck

(

ans

)

{ gpuAssert((ans), __FILE__, __LINE__); }

MPI_KIND

#define MPI_KIND   c_int

NVTX_RANGE_POP

#define NVTX_RANGE_POP

(

)

((void)0)

NVTX_RANGE_PUSH

#define NVTX_RANGE_PUSH

(

msg

)

((void)0)

REALCASE [1/2]

#define REALCASE   1

REALCASE [2/2]

#define REALCASE   1

SINGLE_PRECISION [1/2]

#define SINGLE_PRECISION   1

SINGLE_PRECISION [2/2]

#define SINGLE_PRECISION   1

Function Documentation

cannons_reduction_c_d()

void cannons_reduction_c_d	(	double *	A,
		double *	U,
		int	local_rowsCast,
		int	local_colsCast,
		int *	a_desc,
		double *	Res,
		int	ToStore,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

cannons_reduction_c_dc()

void cannons_reduction_c_dc	(	double complex *	A,
		double complex *	U,
		int	local_rowsCast,
		int	local_colsCasr,
		int *	a_desc,
		double complex *	Res,
		int	ToStore,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

cannons_reduction_c_f()

void cannons_reduction_c_f	(	float *	A,
		float *	U,
		int	local_rowsCast,
		int	local_colsCast,
		int *	a_desc,
		float *	Res,
		int	ToStore,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

cannons_reduction_c_fc()

void cannons_reduction_c_fc	(	float complex *	A,
		float complex *	U,
		int	local_rowsCast,
		int	local_colsCast,
		int *	a_desc,
		float complex *	Res,
		int	ToStore,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

cannons_triang_rectangular_c_d()

void cannons_triang_rectangular_c_d	(	double *	U,
		double *	B,
		int	local_rowsCast,
		int	local_colsCast,
		int *	u_desc,
		int *	b_desc,
		double *	Res,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

cannons_triang_rectangular_c_dc()

void cannons_triang_rectangular_c_dc	(	double complex *	U,
		double complex *	B,
		int	local_rowsCast,
		int	local_colsCast,
		int *	u_desc,
		int *	b_desc,
		double complex *	Res,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

cannons_triang_rectangular_c_f()

void cannons_triang_rectangular_c_f	(	float *	U,
		float *	B,
		int	local_rowsCast,
		int	local_colsCast,
		int *	u_desc,
		int *	b_desc,
		float *	Res,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

cannons_triang_rectangular_c_fc()

void cannons_triang_rectangular_c_fc	(	float complex *	U,
		float complex *	B,
		int	local_rowsCast,
		int	local_colsCast,
		int *	u_desc,
		int *	b_desc,
		float complex *	Res,
		int	row_comm,
		int	col_comm,
		int	wantDebug,
		int	useGPU,
		intptr_t *	gpublasHandle )

gpuAssert()

static void gpuAssert ( int status, const char * file, int line )

inlinestatic

Variable Documentation

gpuMemcpyDeviceToHost

int gpuMemcpyDeviceToHost

gpuMemcpyHostToDevice

int gpuMemcpyHostToDevice