This utility benchmarks an arbitrary workload before and after automatic tuning.
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdarg.h>
#include <oski/oski.h>
#include <oski/simplelist.h>
#include "abort_prog.h"
#include "testvec.h"
#include "timing.h"
#include "parse_opts.h"
#include "readhbpat.h"
#include "keropts.h"
Functions | |
static void | usage (const char *progname) |
Print command-line options. | |
static void | abort_option_error (const char *flag, const char *bad_value, const char *good_value_1,...) |
static simplelist_t * | process_options (int argc, char *argv[], char **p_matfile, int *p_is_workload_explicit, int *p_always_tune, int *p_expand_symm) |
static void | create_workload_data_MatMult (oski_index_t m, oski_index_t n, kerargs_MatMult_t *opt) |
Allocate temporary arrays for simulating an SpMV call. | |
static void | create_workload_data_MatTrisolve (oski_index_t m, oski_index_t n, kerargs_MatTrisolve_t *opt) |
Allocate temporary arrays for simulating a triangular solve call. | |
static void | create_workload_data_MatTransMatMult (oski_index_t m, oski_index_t n, kerargs_MatTransMatMult_t *opt) |
Allocate temporary arrays for simulating an A^T*A*x call. | |
static void | create_workload_data_MatMultAndMatTransMult (oski_index_t m, oski_index_t n, kerargs_MatMultAndMatTransMult_t *opt) |
Allocate temporary arrays for simulating two SpMV calls. | |
static void | create_workload_data_MatPowMult (oski_index_t m, oski_index_t n, kerargs_MatPowMult_t *opt) |
Allocate temporary arrays for simulating an SpMV call. | |
static void | create_workload_data (oski_index_t num_rows, oski_index_t num_cols, simplelist_t *workload) |
Allocate temporary arrays for workload simulation. | |
static void | set_workload_hints_MatMult (oski_matrix_t A, const kerargs_MatMult_t *opt) |
static void | set_workload_hints_MatTrisolve (oski_matrix_t A, const kerargs_MatTrisolve_t *opt) |
static void | set_workload_hints_MatTransMatMult (oski_matrix_t A, const kerargs_MatTransMatMult_t *opt) |
static void | set_workload_hints_MatMultAndMatTransMult (oski_matrix_t A, const kerargs_MatMultAndMatTransMult_t *opt) |
static void | set_workload_hints_MatPowMult (oski_matrix_t A, const kerargs_MatPowMult_t *opt) |
static int | change_numcalls (kernel_t *opt, int num_calls) |
static void | set_workload_hints (oski_matrix_t A, simplelist_t *workload, int always_tune) |
static void | run_workload_MatMult (const oski_matrix_t A, kerargs_MatMult_t *opt) |
static void | run_workload_MatTrisolve (const oski_matrix_t A, kerargs_MatTrisolve_t *opt) |
static void | run_workload_MatTransMatMult (const oski_matrix_t A, kerargs_MatTransMatMult_t *opt) |
static void | run_workload_MatMultAndMatTransMult (const oski_matrix_t A, kerargs_MatMultAndMatTransMult_t *opt) |
static void | run_workload_MatPowMult (const oski_matrix_t A, kerargs_MatPowMult_t *opt) |
static void | run_workload (const oski_matrix_t A, const simplelist_t *workload) |
static double | count_workload_flops_per_nz (const simplelist_t *workload) |
Given a workload, returns the number of flops executed per non-zero element. | |
int | main (int argc, char *argv[]) |
Main benchmark program. |