int
main(int argc, char *argv[])
{
char *s, *d;
orc_uint8 *src, *dest;
OrcProfile prof;
OrcProfile prof_libc;
double ave, std;
double ave_libc, std_libc;
double null;
int i,j;
double cpufreq;
int unalign;
OrcProgram *p;
int level1, level2, level3;
int max;
/* const uint8_t zero = 0; */
orc_init ();
/* cpufreq = 2333e6; */
cpufreq = 1;
if (argc > 1) {
unalign = strtoul (argv[1], NULL, 0);
} else {
unalign = 0;
}
s = malloc(1024*1024*64+1024);
d = malloc(1024*1024*64+1024);
src = ORC_PTR_OFFSET(ALIGN(s,128),unalign);
dest = ALIGN(d,128);
orc_profile_init (&prof);
for(j=0;j<10;j++){
orc_profile_start(&prof);
orc_profile_stop(&prof);
}
orc_profile_get_ave_std (&prof, &null, &std);
{
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_name (p, "orc_memcpy");
/* orc_program_set_name (p, "orc_memset"); */
orc_program_add_destination (p, 1, "d1");
orc_program_add_source (p, 1, "s1");
/* orc_program_add_parameter (p, 1, "p1"); */
orc_program_append (p, "copyb", ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1);
result = orc_program_compile (p);
if (ORC_COMPILE_RESULT_IS_FATAL (result)) {
fprintf (stderr, "Failed to compile orc_memcpy\n");
return -1;
}
}
#ifndef M_LN2
#define M_LN2 0.69314718055994530942
#endif
orc_get_data_cache_sizes (&level1, &level2, &level3);
if (level3 > 0) {
max = (log(level3)/M_LN2 - 6.0) * 10 + 20;
} else if (level2 > 0) {
max = (log(level2)/M_LN2 - 6.0) * 10 + 20;
} else {
max = 140;
}
for(i=0;i<max;i++){
double x = i*0.1 + 6.0;
int size = pow(2.0, x);
if (flush_cache) {
touch (src, (1<<18));
}
if (hot_src) {
touch (src, size);
}
if (hot_dest) {
touch (dest, size);
}
orc_profile_init (&prof);
for(j=0;j<10;j++){
OrcExecutor _ex, *ex = &_ex;
void (*func) (OrcExecutor *);
orc_profile_start(&prof);
/* orc_memcpy (dest, src, size); */
ex->program = p;
ex->n = size;
ex->arrays[ORC_VAR_D1] = dest;
ex->arrays[ORC_VAR_S1] = (void *)src;
func = p->code_exec;
func (ex);
orc_profile_stop(&prof);
if (flush_cache) {
touch (src, (1<<18));
}
if (hot_src) {
touch (src, size);
}
if (hot_dest) {
touch (dest, size);
}
}
orc_profile_init (&prof_libc);
for(j=0;j<10;j++){
orc_profile_start(&prof_libc);
memcpy (dest, src, size);
orc_profile_stop(&prof_libc);
if (flush_cache) {
touch (src, (1<<18));
}
if (hot_src) {
touch (src, size);
}
if (hot_dest) {
touch (dest, size);
}
}
orc_profile_get_ave_std (&prof, &ave, &std);
orc_profile_get_ave_std (&prof_libc, &ave_libc, &std_libc);
ave -= null;
ave_libc -= null;
/* printf("%d: %10.4g %10.4g %10.4g %10.4g (libc %10.4g)\n", i, ave, std, */
/* ave/(1<<i), cpufreq/(ave/(1<<i)), */
/* cpufreq/(ave_libc/(1<<i))); */
printf("%g %10.4g %10.4g\n", x,
cpufreq/(ave/size), cpufreq/(ave_libc/size));
/* printf("%g %10.4g %10.4g\n", x, */
/* 32*(ave/(size)), 32*(ave_libc/(size))); */
fflush (stdout);
}
orc_program_free (p);
free (s);
free (d);
return 0;
}
double
orc_test_performance_full (OrcProgram *program, int flags,
const char *target_name)
{
OrcExecutor *ex;
int n;
int m;
OrcArray *dest_exec[4] = { NULL, NULL, NULL, NULL };
OrcArray *dest_emul[4] = { NULL, NULL, NULL, NULL };
OrcArray *src[8] = { NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL };
int i, j;
OrcCompileResult result;
OrcProfile prof;
double ave, std;
OrcTarget *target;
int misalignment;
ORC_DEBUG ("got here");
target = orc_target_get_by_name (target_name);
if (!(flags & ORC_TEST_FLAGS_BACKUP)) {
unsigned int flags;
flags = orc_target_get_default_flags (target);
result = orc_program_compile_full (program, target, flags);
if (!ORC_COMPILE_RESULT_IS_SUCCESSFUL(result)) {
//printf("compile failed\n");
orc_program_reset (program);
return 0;
}
}
if (program->constant_n > 0) {
n = program->constant_n;
} else {
//n = 64 + (orc_random(&rand_context)&0xf);
n = 1000;
}
ex = orc_executor_new (program);
orc_executor_set_n (ex, n);
if (program->is_2d) {
if (program->constant_m > 0) {
m = program->constant_m;
} else {
m = 8 + (orc_random(&rand_context)&0xf);
}
} else {
m = 1;
}
orc_executor_set_m (ex, m);
ORC_DEBUG("size %d %d", ex->n, ex->params[ORC_VAR_A1]);
misalignment = 0;
for(i=0;i<ORC_N_VARIABLES;i++){
if (program->vars[i].name == NULL) continue;
if (program->vars[i].vartype == ORC_VAR_TYPE_SRC) {
src[i-ORC_VAR_S1] = orc_array_new (n, m, program->vars[i].size,
misalignment);
orc_array_set_random (src[i-ORC_VAR_S1], &rand_context);
misalignment++;
} else if (program->vars[i].vartype == ORC_VAR_TYPE_DEST) {
dest_exec[i-ORC_VAR_D1] = orc_array_new (n, m, program->vars[i].size,
misalignment);
orc_array_set_pattern (dest_exec[i], ORC_OOB_VALUE);
dest_emul[i-ORC_VAR_D1] = orc_array_new (n, m, program->vars[i].size,
misalignment);
orc_array_set_pattern (dest_emul[i], ORC_OOB_VALUE);
misalignment++;
} else if (program->vars[i].vartype == ORC_VAR_TYPE_PARAM) {
orc_executor_set_param (ex, i, 2);
}
}
ORC_DEBUG ("running");
orc_profile_init (&prof);
for(i=0;i<10;i++){
orc_executor_set_n (ex, n);
orc_executor_set_m (ex, m);
for(j=0;j<ORC_N_VARIABLES;j++){
if (program->vars[j].vartype == ORC_VAR_TYPE_DEST) {
orc_executor_set_array (ex, j, dest_exec[j-ORC_VAR_D1]->data);
orc_executor_set_stride (ex, j, dest_exec[j-ORC_VAR_D1]->stride);
}
if (program->vars[j].vartype == ORC_VAR_TYPE_SRC) {
orc_executor_set_array (ex, j, src[j-ORC_VAR_S1]->data);
orc_executor_set_stride (ex, j, src[j-ORC_VAR_S1]->stride);
}
}
if (flags & ORC_TEST_FLAGS_BACKUP) {
orc_profile_start (&prof);
orc_executor_run_backup (ex);
orc_profile_stop (&prof);
} else if (flags & ORC_TEST_FLAGS_EMULATE) {
orc_profile_start (&prof);
orc_executor_emulate (ex);
orc_profile_stop (&prof);
} else {
orc_profile_start (&prof);
orc_executor_run (ex);
orc_profile_stop (&prof);
}
}
ORC_DEBUG ("done running");
orc_profile_get_ave_std (&prof, &ave, &std);
for(i=0;i<4;i++){
if (dest_exec[i]) orc_array_free (dest_exec[i]);
if (dest_emul[i]) orc_array_free (dest_emul[i]);
}
for(i=0;i<8;i++){
if (src[i]) orc_array_free (src[i]);
}
orc_executor_free (ex);
orc_program_reset (program);
return ave/(n*m);
}
int
main (int argc, char *argv[])
{
SchroFrame *dest;
SchroFrame *ref;
SchroFrame *addframe;
SchroParams params;
SchroVideoFormat video_format;
SchroMotionVector *motion_vectors;
int i;
int j;
OrcProfile prof;
double ave, std;
schro_init();
memset (&video_format, 0, sizeof(video_format));
memset (¶ms, 0, sizeof(params));
schro_video_format_set_std_video_format (&video_format,
SCHRO_VIDEO_FORMAT_CUSTOM);
video_format.width = 720;
video_format.height = 480;
video_format.chroma_format = SCHRO_CHROMA_420;
schro_video_format_validate (&video_format);
params.video_format = &video_format;
params.xbsep_luma = 8;
params.ybsep_luma = 8;
params.xblen_luma = 12;
params.yblen_luma = 12;
schro_params_calculate_mc_sizes(¶ms);
dest = schro_frame_new_and_alloc (NULL, SCHRO_FRAME_FORMAT_S16_420,
video_format.width, video_format.height);
schro_frame_clear(dest);
ref = schro_frame_new_and_alloc_extended (NULL, SCHRO_FRAME_FORMAT_U8_420,
video_format.width, video_format.height, 32);
schro_frame_clear(ref);
addframe = schro_frame_new_and_alloc (NULL, SCHRO_FRAME_FORMAT_S16_420,
video_format.width, video_format.height);
schro_frame_clear(addframe);
schro_upsampled_frame_upsample (ref);
motion_vectors = malloc(sizeof(SchroMotionVector) *
params.x_num_blocks * params.y_num_blocks);
memset (motion_vectors, 0, sizeof(SchroMotionVector) *
params.x_num_blocks * params.y_num_blocks);
printf("sizeof(SchroMotionVector) = %lu\n",(unsigned long) sizeof(SchroMotionVector));
printf("num blocks %d x %d\n", params.x_num_blocks, params.y_num_blocks);
for(i=0;i<params.x_num_blocks*params.y_num_blocks;i++){
motion_vectors[i].u.vec.dx[0] = 0;
motion_vectors[i].u.vec.dy[0] = 0;
motion_vectors[i].pred_mode = 1;
motion_vectors[i].split = 2;
}
for(i=0;i<10;i++){
orc_profile_init (&prof);
for(j=0;j<10;j++){
SchroMotion *motion;
void *mv_save;
motion = schro_motion_new (¶ms, ref, NULL);
mv_save = motion->motion_vectors;
motion->motion_vectors = motion_vectors;
orc_profile_start(&prof);
schro_motion_render (motion, dest, addframe, FALSE, NULL);
orc_profile_stop(&prof);
motion->motion_vectors = mv_save;
schro_motion_free (motion);
}
orc_profile_get_ave_std (&prof, &ave, &std);
printf("cycles %g %g\n", ave, std);
}
schro_frame_unref (ref);
schro_frame_unref (dest);
free (motion_vectors);
return 0;
}