Tuning module implementation. More...

#include <math.h>
#include <oski/common.h>
#include <oski/simplelist.h>
#include <oski/heur_internal.h>
#include <oski/heur_typedep.h>
#include <oski/heurexport.h>
#include <oski/trace.h>
#include <oski/tune.h>
#include <oski/timer.h>
#include <oski/matrix.h>
#include <oski/vecview.h>
#include <oski/xforms.h>

Defines
#define	BEST_FRAC_OBSERVED .25
	Fraction of observed workload available for tuning.
#define	BEST_FRAC_WORKLOAD .25
	Fraction of hint workload available for tuning.
Functions
static double	ComputeTuningBudget (oski_matrix_t A_tunable)
	Returns the estimated number of seconds available for tuning.
static int	ChooseFastest (oski_matrix_t A_tunable)
	Determines whether the heuristic-selected data structure leads to faster execution times than the input data structure.
static void	FreeTunedMat (oski_matrix_t A_tunable)
int	oski_TuneMat (oski_matrix_t A_tunable)
	Basic outline of this routine's implementation:

Detailed Description

Tuning module implementation.

Current list of heuristics:

Cache blocked SpMV
Register blocked SpMV
Register blocked sparse $A^TA\cdot x$ .
Register blocked, switch-to-dense triangular solve

Basic structure of a heuristic:

Has two associated cost estimates
1. time to run heuristic,
2. time to convert.

Here's a typical implementation of a heuristic:

FUNCTION HeurType :: EvaluateHeuristic( heur:HeurType, A:matrix )
    IF this heuristic does not apply to A and its trace THEN
        RETURN NULL;
    ENDIF

    LET results = Choose tuning parameters for A
    RETURN results

For example, here is the heuristic for register blocked SpMV:

FUNCTION RBSpMV :: EvaluateHeuristic( heur:HeurType, A:matrix )
    IF A->trace not "dominated" by calls to SpMV THEN
        RETURN NULL;
    ENDIF

    LET EstFill[MAX_R, MAX_C] = EstimateFill( A )
    LET Mflops_dense[MAX_R, MAX_C] = LoadRegProfile( A )

    LET r, c = max_{r,c} Mflops_dense[r,c] / EstFill[r, c]

    IF r == 1 AND c == 1 THEN
        RETURN NULL;
    ELSE
        LET results = new RbSpMV_results( r, c )
        RETURN results
    ENDIF