static int
do_profile( caddr_t start, unsigned long plength, unsigned scale, int thresh,
int bucket, unsigned int mask ) {
int i, retval;
unsigned long blength;
int num_buckets,j=0;
int num_bufs = num_events;
int event = num_events;
int events[MAX_TEST_EVENTS];
char header[BUFSIZ];
strncpy(header,"address\t\t",BUFSIZ);
//= "address\t\t\tcyc\tins\tfp_ins\n";
for(i=0;i<MAX_TEST_EVENTS;i++) {
if (mask & test_events[i].mask) {
events[j]=test_events[i].event;
if (events[j]==PAPI_TOT_CYC) {
strncat(header,"\tcyc",BUFSIZ-1);
}
if (events[j]==PAPI_TOT_INS) {
strncat(header,"\tins",BUFSIZ-1);
}
if (events[j]==PAPI_FP_INS) {
strncat(header,"\tfp_ins",BUFSIZ-1);
}
if (events[j]==PAPI_FP_OPS) {
strncat(header,"\tfp_ops",BUFSIZ-1);
}
if (events[j]==PAPI_L2_TCM) {
strncat(header,"\tl2_tcm",BUFSIZ-1);
}
j++;
}
}
strncat(header,"\n",BUFSIZ-1);
blength = prof_size( plength, scale, bucket, &num_buckets );
prof_alloc( num_bufs, blength );
if ( !TESTS_QUIET )
printf( "Overall event counts:\n" );
for ( i = 0; i < num_events; i++ ) {
if ( ( retval =
PAPI_profil( profbuf[i], ( unsigned int ) blength, start, scale,
EventSet, events[i], thresh,
PAPI_PROFIL_POSIX | bucket ) ) != PAPI_OK ) {
if (retval == PAPI_EINVAL) {
test_warn( __FILE__, __LINE__, "Trying to profile with derived event", 1);
num_events=i;
break;
}
else {
printf("Failed with event %d 0x%x\n",i,events[i]);
test_fail( __FILE__, __LINE__, "PAPI_profil", retval );
}
}
}
if ( ( retval = PAPI_start( EventSet ) ) != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_start", retval );
my_main( );
if ( ( retval = PAPI_stop( EventSet, values[0] ) ) != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_stop", retval );
if ( !TESTS_QUIET ) {
printf( TAB1, "PAPI_TOT_CYC:", ( values[0] )[--event] );
if ( strcmp( hw_info->model_string, "POWER6" ) != 0 ) {
printf( TAB1, "PAPI_TOT_INS:", ( values[0] )[--event] );
}
#if defined(__powerpc__)
printf( TAB1, "PAPI_FP_INS", ( values[0] )[--event] );
#else
if ( strcmp( hw_info->model_string, "Intel Pentium III" ) != 0 ) {
printf( TAB1, "PAPI_FP_OPS:", ( values[0] )[--event] );
printf( TAB1, "PAPI_L2_TCM:", ( values[0] )[--event] );
}
#endif
}
for ( i = 0; i < num_events; i++ ) {
if ( ( retval =
PAPI_profil( profbuf[i], ( unsigned int ) blength, start, scale,
EventSet, events[i], 0,
PAPI_PROFIL_POSIX ) ) != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_profil", retval );
}
prof_head( blength, bucket, num_buckets, header );
prof_out( start, num_events, bucket, num_buckets, scale );
retval = prof_check( num_bufs, bucket, num_buckets );
for ( i = 0; i < num_bufs; i++ ) {
free( profbuf[i] );
}
return retval;
}
void gb_run()
{
#ifdef PROFILE
prof_enter(PF_GB_RUN);
#endif
//if (rom_get_loaded()){
if (g_regs.LCDC&0x80){ // LCDC 起動時
g_regs.LY=(g_regs.LY+1)%154;
g_regs.STAT&=0xF8;
if (g_regs.LYC==g_regs.LY){
g_regs.STAT|=4;
if (g_regs.STAT&0x40)
cpu_irq(INT_LCDC);
}
if (g_regs.LY==0){
renderer_refresh();
if (now_frame>=skip){
//pgFillvram(0);
render_screen(vframe);
now_frame=0;
}
else
now_frame++;
lcd_clear_win_count();
// skip=skip_buf;
}
if (g_regs.LY>=144){ // VBlank 期間中
g_regs.STAT|=1;
if (g_regs.LY==144){
cpu_exec(72);
cpu_irq(INT_VBLANK);
if (g_regs.STAT&0x10)
cpu_irq(INT_LCDC);
cpu_exec(456-80);
}
else if (g_regs.LY==153){
cpu_exec(80);
g_regs.LY=0;
cpu_exec(456-80); // 前のラインのかなり早目から0になるようだ。
g_regs.LY=153;
}
else
cpu_exec(456);
}
else{ // VBlank 期間外
g_regs.STAT|=2;
if (g_regs.STAT&0x20)
cpu_irq(INT_LCDC);
cpu_exec(80); // state=2
g_regs.STAT|=3;
cpu_exec(169); // state=3
if (dma_executing){ // HBlank DMA
if (b_dma_first){
dma_dest_bank=vram_bank;
if (dma_src<0x4000)
dma_src_bank=get_rom();
else if (dma_src<0x8000)
dma_src_bank=mbc_get_rom();
else if (dma_src>=0xA000&&dma_src<0xC000)
dma_src_bank=mbc_get_sram()-0xA000;
else if (dma_src>=0xC000&&dma_src<0xD000)
dma_src_bank=ram-0xC000;
else if (dma_src>=0xD000&&dma_src<0xE000)
dma_src_bank=ram_bank-0xD000;
else dma_src_bank=NULL;
b_dma_first=false;
}
memcpy(dma_dest_bank+(dma_dest&0x1ff0),dma_src_bank+dma_src,16);
// fprintf(cpu_file,"%03d : dma exec %04X -> %04X rest %d\n",g_regs.LY,cpu_dma_src,cpu_dma_dest,cpu_dma_rest);
dma_src+=16;
dma_src&=0xfff0;
dma_dest+=16;
dma_dest&=0xfff0;
dma_rest--;
if (!dma_rest)
dma_executing=false;
// cpu_total_clock+=207*(cpu_speed?2:1);
// cpu_sys_clock+=207*(cpu_speed?2:1);
// cpu_div_clock+=207*(cpu_speed?2:1);
// g_regs.STAT|=3;
if (now_frame>=skip && !sgb_mask)
lcd_render(vframe,g_regs.LY);
g_regs.STAT&=0xfc;
cpu_exec(207); // state=3
}
else{
/* if (lcd_get_sprite_count()){
if (lcd_get_sprite_count()>=10){
cpu_exec(129);
if ((g_regs.STAT&0x08))
cpu_irq(INT_LCDC);
g_regs.STAT&=0xfc;
if (now_frame>=skip)
lcd_render(vframe,g_regs.LY);
cpu_exec(78); // state=0
}
else{
cpu_exec(129*lcd_get_sprite_count()/10);
if ((g_regs.STAT&0x08))
cpu_irq(INT_LCDC);
g_regs.STAT&=0xfc;
if (now_frame>=skip)
lcd_render(vframe,g_regs.LY);
cpu_exec(207-(129*lcd_get_sprite_count()/10)); // state=0
}
}
else{
*/ g_regs.STAT&=0xfc;
if (now_frame>=skip && !sgb_mask)
lcd_render(vframe,g_regs.LY);
if ((g_regs.STAT&0x08))
cpu_irq(INT_LCDC);
cpu_exec(207); // state=0
// }
}
}
}
else{ // LCDC 停止時
g_regs.LY=0;
// g_regs.LY=(g_regs.LY+1)%154;
re_render++;
if (re_render>=154){
gb_fill_vframe(0xff);
renderer_refresh();
if (now_frame>=skip){
//pgFillvram(0);
render_screen(vframe);
now_frame=0;
}
else
now_frame++;
lcd_clear_win_count();
re_render=0;
}
g_regs.STAT&=0xF8;
cpu_exec(456);
}
//}
#ifdef PROFILE
prof_out();
#endif
}
static int do_profile(caddr_t start, unsigned long plength, unsigned scale, int thresh, int bucket) {
int i, retval;
unsigned long blength;
int num_buckets;
char *profstr[5] = {"PAPI_PROFIL_POSIX",
"PAPI_PROFIL_RANDOM",
"PAPI_PROFIL_WEIGHTED",
"PAPI_PROFIL_COMPRESS",
"PAPI_PROFIL_<all>" };
int profflags[5] = {PAPI_PROFIL_POSIX,
PAPI_PROFIL_POSIX | PAPI_PROFIL_RANDOM,
PAPI_PROFIL_POSIX | PAPI_PROFIL_WEIGHTED,
PAPI_PROFIL_POSIX | PAPI_PROFIL_COMPRESS,
PAPI_PROFIL_POSIX | PAPI_PROFIL_WEIGHTED |
PAPI_PROFIL_RANDOM | PAPI_PROFIL_COMPRESS };
do_no_profile();
blength = prof_size(plength, scale, bucket, &num_buckets);
prof_alloc(5, blength);
for (i=0;i<5;i++) {
if (!TESTS_QUIET)
printf("Test type : \t%s\n", profstr[i]);
#ifndef SWPROFILE
if ((retval = PAPI_profil(profbuf[i], blength, start, scale,
EventSet, PAPI_event, thresh,
profflags[i] | bucket)) != PAPI_OK) {
test_fail(__FILE__, __LINE__, "PAPI_profil", retval);
}
#else
if ((retval = PAPI_profil(profbuf[i], blength, start, scale,
EventSet, PAPI_event, thresh,
profflags[i] | bucket | PAPI_PROFIL_FORCE_SW)) != PAPI_OK) {
test_fail(__FILE__, __LINE__, "PAPI_profil", retval);
}
#endif
if ((retval = PAPI_start(EventSet)) != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
do_flops(getenv("NUM_FLOPS") ? atoi(getenv("NUM_FLOPS")) : NUM_FLOPS);
if ((retval = PAPI_stop(EventSet, values[1])) != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
if (!TESTS_QUIET) {
printf(TAB1, event_name, (values[1])[0]);
printf(TAB1, "PAPI_TOT_CYC", (values[1])[1]);
}
if ((retval = PAPI_profil(profbuf[i], blength, start, scale,
EventSet, PAPI_event, 0, profflags[i])) != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_profil", retval);
}
prof_head(blength, bucket, num_buckets,
"address\t\t\tflat\trandom\tweight\tcomprs\tall\n");
prof_out(start, 5, bucket, num_buckets, scale);
retval = prof_check(5, bucket, num_buckets);
for (i=0;i<5;i++) {
free(profbuf[i]);
}
return(retval);
}
