Eztrace_dev

Macros
#define	PastixKernelStop   0
	Main stop enum event for all the events in traces.
#define	PastixKernelsNbr   (PastixKernelLvl0Nbr + PastixKernelLvl1Nbr + PastixKernelLvl2Nbr)
	Total number of kernel events.

Typedefs
typedef enum pastix_ktype0_e	pastix_ktype0_t
	List of the Level 0 events that may be traced in PaStiX. More...
typedef enum pastix_ktype_e	pastix_ktype_t
	List of the Level 1 events that may be traced in PaStiX. More...
typedef enum pastix_ktype2_e	pastix_ktype2_t
	List of the Level 2 events that may be traced in PaStiX. More...

Enumerations
enum	pastix_ktype0_e { PastixKernelLvl0Facto , PastixKernelLvl0Solve , PastixKernelLvl0Diag , PastixKernelLvl0Nbr }
	List of the Level 0 events that may be traced in PaStiX. More...
enum	pastix_ktype_e {   PastixKernelGETRF , PastixKernelHETRF , PastixKernelPOTRF , PastixKernelPXTRF ,   PastixKernelSYTRF , PastixKernelSCALOCblk , PastixKernelSCALOBlok , PastixKernelTRSMForw ,   PastixKernelTRSMBack , PastixKernelTRSMCblk1d , PastixKernelTRSMCblk2d , PastixKernelTRSMCblkLR ,   PastixKernelTRSMBlok2d , PastixKernelTRSMBlokLR , PastixKernelGEMMCblk1d1d , PastixKernelGEMMCblk1d2d ,   PastixKernelGEMMCblk2d2d , PastixKernelGEMMCblkFRLR , PastixKernelGEMMCblkLRLR , PastixKernelGEMMBlok2d2d ,   PastixKernelGEMMBlokLRLR , PastixKernelGEADDCblkFRFR , PastixKernelGEADDCblkFRLR , PastixKernelGEADDCblkLRLR ,   PastixKernelLvl1Nbr }
	List of the Level 1 events that may be traced in PaStiX. More...
enum	pastix_ktype2_e {   PastixKernelLvl2GETRF , PastixKernelLvl2HETRF , PastixKernelLvl2POTRF , PastixKernelLvl2PXTRF ,   PastixKernelLvl2SYTRF , PastixKernelLvl2_FR_TRSM , PastixKernelLvl2_LR_TRSM , PastixKernelLvl2_FR_GEMM ,   PastixKernelLvl2_LR_FRFR2FR , PastixKernelLvl2_LR_FRLR2FR , PastixKernelLvl2_LR_LRFR2FR , PastixKernelLvl2_LR_LRLR2FR ,   PastixKernelLvl2_LR_FRFR2LR , PastixKernelLvl2_LR_FRLR2LR , PastixKernelLvl2_LR_LRFR2LR , PastixKernelLvl2_LR_LRLR2LR ,   PastixKernelLvl2_LR_FRFR2null , PastixKernelLvl2_LR_FRLR2null , PastixKernelLvl2_LR_LRFR2null , PastixKernelLvl2_LR_LRLR2null ,   PastixKernelLvl2_LR_init_compress , PastixKernelLvl2_LR_add2C_uncompress , PastixKernelLvl2_LR_add2C_recompress , PastixKernelLvl2_LR_add2C_updateCfr ,   PastixKernelLvl2_LR_add2C_orthou , PastixKernelLvl2_LR_add2C_rradd_orthogonalize , PastixKernelLvl2_LR_add2C_rradd_recompression , PastixKernelLvl2_LR_add2C_rradd_computeNewU ,   PastixKernelLvl2Nbr }
	List of the Level 2 events that may be traced in PaStiX. More...

Functions
static void	kernel_trace_start_lvl0 (pastix_ktype0_t ktype)
static void	kernel_trace_stop_lvl0 (double flops)
static void	kernel_trace_start_lvl2 (pastix_ktype2_t ktype)
static void	kernel_trace_stop_lvl2 (double flops)
static void	kernel_trace_stop_lvl2_rank (double flops, int rank)
void	kernelsTraceInit (const pastix_data_t *pastix_data, pastix_trace_t trace)
	Starts the trace module. More...
void	kernelsTraceFinalize (const pastix_data_t *pastix_data)
	Stops the trace module. More...
void	kernelsTraceStop (void)
	Pauses the trace module.
void	kernelsTraceStart (void)
	Resumes the trace module.
static double	kernel_trace_start (pastix_ktype_t ktype)
	Start the trace of a single kernel. More...
static void	kernel_trace_stop (int8_t inlast, pastix_ktype_t ktype, int m, int n, int k, double flops, double starttime)
	Stop the trace of a single kernel. More...

Variables
volatile double	kernels_flops [PastixKernelLvl1Nbr]
	Global array to store the number of flops executed per kernel.
pastix_atomic_lock_t	lock_flops
	Lock to accumulate flops.
double	overall_flops [3]
	Overall number of flops.

Detailed Description

Typedef Documentation

pastix_ktype0_t

typedef enum pastix_ktype0_e pastix_ktype0_t

List of the Level 0 events that may be traced in PaStiX.

This is only the high level steps.

pastix_ktype_t

typedef enum pastix_ktype_e pastix_ktype_t

List of the Level 1 events that may be traced in PaStiX.

This is the main information that traces all the major kernels during the factorization step.

pastix_ktype2_t

typedef enum pastix_ktype2_e pastix_ktype2_t

List of the Level 2 events that may be traced in PaStiX.

This is the low-level information that traces all the individual calls to blas/lapack routines in the code. It is used to compute the number of flops in low-rank compression, and to distinguish the amount of flops spent in each part of the low-rank updates.

Enumeration Type Documentation

pastix_ktype0_e

enum pastix_ktype0_e

List of the Level 0 events that may be traced in PaStiX.

This is only the high level steps.

Definition at line 34 of file kernels_enums.h.

pastix_ktype_e

enum pastix_ktype_e

List of the Level 1 events that may be traced in PaStiX.

This is the main information that traces all the major kernels during the factorization step.

Enumerator
PastixKernelGETRF	LU diagonal block kernel
PastixKernelHETRF	LDLh diagonal block kernel
PastixKernelPOTRF	Cholesky diagonal block kernel
PastixKernelPXTRF	Complex LL^t diagonal block kernel
PastixKernelSYTRF	LDLt diagonal block kernel
PastixKernelSCALOCblk	Scaling out-of-place of a panel
PastixKernelSCALOBlok	Scaling out-of-place of a block
PastixKernelTRSMForw	Solve forward TRSM block kernel
PastixKernelTRSMBack	Solve backward TRSM block kernel
PastixKernelTRSMCblk1d	TRSM applied to a panel in 1d layout
PastixKernelTRSMCblk2d	TRSM applied to a panel in 2d layout
PastixKernelTRSMCblkLR	TRSM applied to a panel in low-rank
PastixKernelTRSMBlok2d	TRSM applied to a block in 2d layout
PastixKernelTRSMBlokLR	TRSM applied to a block in low-rank
PastixKernelGEMMCblk1d1d	GEMM applied from a panel in 1d layout to a panel in 1d layout
PastixKernelGEMMCblk1d2d	GEMM applied from a panel in 1d layout to a panel in 2d layout
PastixKernelGEMMCblk2d2d	GEMM applied from a panel in 2d layout to a panel in 2d layout
PastixKernelGEMMCblkFRLR	GEMM applied from a panel in full-rank to a panel in low-rank
PastixKernelGEMMCblkLRLR	GEMM applied from a panel in low-rank to a panel in low-rank
PastixKernelGEMMBlok2d2d	GEMM applied from a block in 2d layout to a block in 2d layout
PastixKernelGEMMBlokLRLR	GEMM applied from a block in low-rank to a block in low-rank
PastixKernelGEADDCblkFRFR	GEADD applied from a panel in full-rank to a panel in full-rank
PastixKernelGEADDCblkFRLR	GEADD applied from a panel in full-rank to a panel in low-rank
PastixKernelGEADDCblkLRLR	GEADD applied from a panel in low-rank to a panel in low-rank

Definition at line 47 of file kernels_enums.h.

pastix_ktype2_e

enum pastix_ktype2_e

List of the Level 2 events that may be traced in PaStiX.

This is the low-level information that traces all the individual calls to blas/lapack routines in the code. It is used to compute the number of flops in low-rank compression, and to distinguish the amount of flops spent in each part of the low-rank updates.

Enumerator
PastixKernelLvl2GETRF	LU diagonal block kernel
PastixKernelLvl2HETRF	LDLh diagonal block kernel
PastixKernelLvl2POTRF	Cholesky diagonal block kernel
PastixKernelLvl2PXTRF	Complex LL^t diagonal block kernel
PastixKernelLvl2SYTRF	LDLt diagonal block kernel
PastixKernelLvl2_LR_add2C_rradd_orthogonalize	< CGS, partialQR or fullQR

Definition at line 84 of file kernels_enums.h.

Function Documentation

kernelsTraceInit()

void kernelsTraceInit	(	const pastix_data_t *	pastix_data,
		pastix_trace_t	trace
	)

Starts the trace module.

Parameters

[in]	pastix_data	The pastix_data structure of the problem to give input information to the different trace modes.
[in]	trace	Value of iparm_trace : PastixTraceNot = no trace PastixTraceNumfact = only traces the facto PastixTraceSolve = only traces the solve PastixTraceFactAndSolve = traces facto and solve

Definition at line 147 of file kernels_trace.c.

References solver_matrix_s::bloknbr, solver_matrix_s::cblkmaxblk, solver_matrix_s::cblkmin2d, solver_matrix_s::cblknbr, solver_matrix_s::cblktab, pastix_data_s::dir_global, solver_matrix_s::factotype, solver_cblk_s::fblokptr, kernels_flops, lock_flops, pastix_int_t, PastixFactLU, PastixTraceNot, and pastix_data_s::solvmatr.

Referenced by pastixInitWithAffinity().

kernelsTraceFinalize()

void kernelsTraceFinalize

(

const pastix_data_t *

pastix_data

)

Stops the trace module.

Parameters

[in]

pastix_data

The pastix_data structure of the problem to get input information for the different trace modes, and store output statistics.

Definition at line 259 of file kernels_trace.c.

References pastix_data_s::bcsc, pastix_data_s::dparm, DPARM_FACT_RLFLOPS, pastix_data_s::iparm, IPARM_GPU_NBR, IPARM_VERBOSE, lock_flops, overall_flops, pastix_int_t, PastixKernelGEMMBlokLRLR, PastixKernelGEMMCblk1d1d, PastixKernelGEMMCblkFRLR, PastixKernelGEMMCblkLRLR, PastixKernelGETRF, PastixKernelHETRF, PastixKernelPOTRF, PastixKernelPXTRF, PastixKernelSCALOBlok, PastixKernelSCALOCblk, PastixKernelSYTRF, PastixKernelTRSMBlokLR, PastixKernelTRSMCblk1d, PastixKernelTRSMCblk2d, PastixKernelTRSMCblkLR, and PastixVerboseNot.

Referenced by pastixFinalize().

kernel_trace_start()

static double kernel_trace_start ( pastix_ktype_t  ktype )

inlinestatic

Start the trace of a single kernel.

Parameters

[in]

ktype

Type of the kernel starting that need to be traced. With EZTrace mode, this call is empty if the environment variable PASTIX_EZTRACE_LEVEL is different from 1.

Returns

the starting time if PASTIX_GENERATE_MODEL is enabled, 0. otherwise.

Definition at line 87 of file kernels_trace.h.

Referenced by core_cgemmsp_block_frfr(), core_cgemmsp_block_frlr(), core_cgemmsp_block_lrlr(), core_ctrsmsp_2dsub(), core_ctrsmsp_lrsub(), core_dgemmsp_block_frfr(), core_dgemmsp_block_frlr(), core_dgemmsp_block_lrlr(), core_dtrsmsp_2dsub(), core_dtrsmsp_lrsub(), core_sgemmsp_block_frfr(), core_sgemmsp_block_frlr(), core_sgemmsp_block_lrlr(), core_strsmsp_2dsub(), core_strsmsp_lrsub(), core_zgemmsp_block_frfr(), core_zgemmsp_block_frlr(), core_zgemmsp_block_lrlr(), core_ztrsmsp_2dsub(), core_ztrsmsp_lrsub(), cpublok_cadd(), cpublok_dadd(), cpublok_sadd(), cpublok_zadd(), cpucblk_cadd(), cpucblk_cgemmsp(), cpucblk_cgetrfsp1d_getrf(), cpucblk_chetrfsp1d_hetrf(), cpucblk_cpotrfsp1d_potrf(), cpucblk_cpxtrfsp1d_pxtrf(), cpucblk_cscalo(), cpucblk_csytrfsp1d_sytrf(), cpucblk_dadd(), cpucblk_dgemmsp(), cpucblk_dgetrfsp1d_getrf(), cpucblk_dpotrfsp1d_potrf(), cpucblk_dscalo(), cpucblk_dsytrfsp1d_sytrf(), cpucblk_sadd(), cpucblk_sgemmsp(), cpucblk_sgetrfsp1d_getrf(), cpucblk_spotrfsp1d_potrf(), cpucblk_sscalo(), cpucblk_ssytrfsp1d_sytrf(), cpucblk_zadd(), cpucblk_zgemmsp(), cpucblk_zgetrfsp1d_getrf(), cpucblk_zhetrfsp1d_hetrf(), cpucblk_zpotrfsp1d_potrf(), cpucblk_zpxtrfsp1d_pxtrf(), cpucblk_zscalo(), cpucblk_zsytrfsp1d_sytrf(), solve_cblk_ctrsmsp_backward(), solve_cblk_ctrsmsp_forward(), solve_cblk_dtrsmsp_backward(), solve_cblk_dtrsmsp_forward(), solve_cblk_strsmsp_backward(), solve_cblk_strsmsp_forward(), solve_cblk_ztrsmsp_backward(), and solve_cblk_ztrsmsp_forward().

kernel_trace_stop()

static void kernel_trace_stop ( int8_t  inlast, pastix_ktype_t  ktype, int  m, int  n, int  k, double  flops, double  starttime  )

inlinestatic

Stop the trace of a single kernel.

Parameters

[in]	ktype	Type of the kernel starting that need to be traced. With EZTrace mode, this call is empty if the environment variable PASTIX_EZTRACE_LEVEL is different from 1.
[in]	m	The m parameter of the kernel (used by xxTRF, TRSM, and GEMM)
[in]	n	The n parameter of the kernel (used by TRSM, and GEMM)
[in]	k	The k parameter of the kernel (used by GEMM)
[in]	flops	The number of flops of the kernel
[in]	starttime	The stating time of the kernel. Used only if PASTIX_GENERATE_MODEL is enabled.

Definition at line 139 of file kernels_trace.h.

References lock_flops, overall_flops, and PastixKernelStop.

Referenced by core_cgemmsp_block_frfr(), core_cgemmsp_block_frlr(), core_cgemmsp_block_lrlr(), core_ctrsmsp_2dsub(), core_ctrsmsp_lrsub(), core_dgemmsp_block_frfr(), core_dgemmsp_block_frlr(), core_dgemmsp_block_lrlr(), core_dtrsmsp_2dsub(), core_dtrsmsp_lrsub(), core_sgemmsp_block_frfr(), core_sgemmsp_block_frlr(), core_sgemmsp_block_lrlr(), core_strsmsp_2dsub(), core_strsmsp_lrsub(), core_zgemmsp_block_frfr(), core_zgemmsp_block_frlr(), core_zgemmsp_block_lrlr(), core_ztrsmsp_2dsub(), core_ztrsmsp_lrsub(), cpublok_cadd(), cpublok_dadd(), cpublok_sadd(), cpublok_zadd(), cpucblk_cadd(), cpucblk_cgemmsp(), cpucblk_cscalo(), cpucblk_dadd(), cpucblk_dgemmsp(), cpucblk_dscalo(), cpucblk_sadd(), cpucblk_sgemmsp(), cpucblk_sscalo(), cpucblk_zadd(), cpucblk_zgemmsp(), and cpucblk_zscalo().