simple profiler added

git-svn-id: file:///home/notaz/opt/svn/PicoDrive@850 be3aeb3a-fb24-0410-a615-afba39da0efa

pico/32x/32x.c

@@ -165,6 +165,8 @@ static void p32x_start_blank(void)
 
 static __inline void run_m68k(int cyc)
 {
+  pprof_start(m68k);
+
 #if defined(EMU_C68K)
   PicoCpuCM68k.cycles = cyc;
   CycloneRun(&PicoCpuCM68k);
@@ -174,6 +176,8 @@ static __inline void run_m68k(int cyc)
 #elif defined(EMU_F68K)
   SekCycleCnt += fm68k_emulate(cyc+1, 0, 0);
 #endif
+
+  pprof_end(m68k);
 }
 
 // ~1463.8, but due to cache misses and slow mem
@@ -194,10 +198,18 @@ static __inline void run_m68k(int cyc)
     slice = SekCycleCnt - slice; /* real count from 68k */ \
     if (SekCycleCnt < SekCycleAim) \
       elprintf(EL_32X, "slice %d", slice); \
-    if (!(Pico32x.emu_flags & (P32XF_SSH2POLL|P32XF_SSH2VPOLL))) \
+    if (!(Pico32x.emu_flags & (P32XF_SSH2POLL|P32XF_SSH2VPOLL))) { \
+      pprof_start(ssh2); \
       sh2_execute(&ssh2, CYCLES_M68K2SH2(slice)); \
-    if (!(Pico32x.emu_flags & (P32XF_MSH2POLL|P32XF_MSH2VPOLL))) \
+      pprof_end(ssh2); \
+    } \
+    if (!(Pico32x.emu_flags & (P32XF_MSH2POLL|P32XF_MSH2VPOLL))) { \
+      pprof_start(msh2); \
       sh2_execute(&msh2, CYCLES_M68K2SH2(slice)); \
+      pprof_end(msh2); \
+    } \
+    pprof_start(dummy); \
+    pprof_end(dummy); \
   } \
 }
 

pico/cd/pico.c

@@ -66,6 +66,9 @@ PICO_INTERNAL int PicoResetMCD(void)
 static __inline void SekRunM68k(int cyc)
 {
   int cyc_do;
+
+  pprof_start(m68k);
+
   SekCycleAim+=cyc;
   if ((cyc_do=SekCycleAim-SekCycleCnt) <= 0) return;
 #if defined(EMU_CORE_DEBUG)
@@ -81,6 +84,7 @@ static __inline void SekRunM68k(int cyc)
   g_m68kcontext=&PicoCpuFM68k;
   SekCycleCnt+=fm68k_emulate(cyc_do, 0, 0);
 #endif
+  pprof_end(m68k);
 }
 
 static __inline void SekRunS68k(int cyc)

pico/draw.c

@@ -1485,6 +1485,8 @@ void PicoDrawSync(int to, int blank_last_line)
   int sh = (Pico.video.reg[0xC] & 8) >> 3; // shadow/hilight?
   int bgc = Pico.video.reg[7];
 
+  pprof_start(draw);
+
   if (rendlines != 240)
     offs = 8;
 
@@ -1516,6 +1518,8 @@ void PicoDrawSync(int to, int blank_last_line)
     line++;
   }
   DrawScanline = line;
+
+  pprof_end(draw);
 }
 
 void PicoDrawSetColorFormat(int which)

pico/memory.c

@@ -257,8 +257,11 @@ void NOINLINE ctl_write_z80busreq(u32 d)
     else
     {
       z80stopCycle = SekCyclesDone();
-      if ((PicoOpt&POPT_EN_Z80) && !Pico.m.z80_reset)
+      if ((PicoOpt&POPT_EN_Z80) && !Pico.m.z80_reset) {
+        pprof_start(m68k);
         PicoSyncZ80(z80stopCycle);
+        pprof_end_sub(m68k);
+      }
     }
     Pico.m.z80Run = d;
   }
@@ -272,8 +275,11 @@ void NOINLINE ctl_write_z80reset(u32 d)
   {
     if (d)
     {
-      if ((PicoOpt&POPT_EN_Z80) && Pico.m.z80Run)
+      if ((PicoOpt&POPT_EN_Z80) && Pico.m.z80Run) {
+        pprof_start(m68k);
         PicoSyncZ80(SekCyclesDone());
+        pprof_end_sub(m68k);
+      }
       YM2612ResetChip();
       timers_reset();
     }

pico/pico.c

@@ -272,6 +272,8 @@ PICO_INTERNAL int CheckDMA(void)
 static __inline void SekRunM68k(int cyc)
 {
   int cyc_do;
+  pprof_start(m68k);
+
   SekCycleAim+=cyc;
   if ((cyc_do=SekCycleAim-SekCycleCnt) <= 0) return;
 #if defined(EMU_CORE_DEBUG)
@@ -286,6 +288,8 @@ static __inline void SekRunM68k(int cyc)
 #elif defined(EMU_F68K)
   SekCycleCnt+=fm68k_emulate(cyc_do+1, 0, 0);
 #endif
+
+  pprof_end(m68k);
 }
 
 #include "pico_cmn.c"
@@ -303,38 +307,47 @@ PICO_INTERNAL void PicoSyncZ80(int m68k_cycles_done)
   z80_cycle_aim = cycles_68k_to_z80(m68k_cycles_done);
   cnt = z80_cycle_aim - z80_cycle_cnt;
 
+  pprof_start(z80);
+
   elprintf(EL_BUSREQ, "z80 sync %i (%i|%i -> %i|%i)", cnt, z80_cycle_cnt, z80_cycle_cnt / 228,
     z80_cycle_aim, z80_cycle_aim / 228);
 
   if (cnt > 0)
     z80_cycle_cnt += z80_run(cnt);
+
+  pprof_end(z80);
 }
 
 
 void PicoFrame(void)
 {
+  pprof_start(frame);
+
   Pico.m.frame_count++;
 
   if (PicoAHW & PAHW_SMS) {
     PicoFrameMS();
-    return;
+    goto end;
   }
 
   // TODO: MCD+32X
   if (PicoAHW & PAHW_MCD) {
     PicoFrameMCD();
-    return;
+    goto end;
   }
 
   if (PicoAHW & PAHW_32X) {
     PicoFrame32x();
-    return;
+    goto end;
   }
 
   //if(Pico.video.reg[12]&0x2) Pico.video.status ^= 0x10; // change odd bit in interlace mode
 
   PicoFrameStart();
   PicoFrameHints();
+
+end:
+  pprof_end(frame);
 }
 
 void PicoFrameDrawOnly(void)

pico/pico_int.h

@@ -779,6 +779,18 @@ extern void lprintf(const char *fmt, ...);
 #define elprintf(w,f,...)
 #endif
 
+// profiling
+#ifdef PPROF
+#include <platform/linux/pprof.h>
+#else
+#define pprof_init()
+#define pprof_finish()
+#define pprof_start(x)
+#define pprof_end(...)
+#define pprof_end_sub(...)
+#endif
+
+// misc
 #ifdef _MSC_VER
 #define cdprintf
 #else

pico/sound/sound.c

@@ -308,6 +308,8 @@ static int PsndRender(int offset, int length)
 		(Pico_mcd->pcm.control & 0x80) && Pico_mcd->pcm.enabled;
   offset <<= stereo;
 
+  pprof_start(sound);
+
 #if !SIMPLE_WRITE_SOUND
   if (offset == 0) { // should happen once per frame
     // compensate for float part of PsndLen
@@ -362,6 +364,8 @@ static int PsndRender(int offset, int length)
   // convert + limit to normal 16bit output
   PsndMix_32_to_16l(PsndOut+offset, buf32, length);
 
+  pprof_end(sound);
+
   return length;
 }
 

platform/common/common.mak

@@ -7,6 +7,10 @@ endif
 ifeq "$(profile)" "2"
 CFLAGS += -fprofile-use
 endif
+ifeq "$(pprof)" "1"
+DEFINES += PPROF
+OBJS += platform/linux/pprof.o
+endif
 
 # === Pico core ===
 # Pico

platform/common/emu.c

@@ -1347,6 +1347,8 @@ void emu_init(void)
 	mkdir_path(path, pos, "srm");
 	mkdir_path(path, pos, "brm");
 
+	pprof_init();
+
 	make_config_cfg(path);
 	config_readlrom(path);
 
@@ -1373,6 +1375,8 @@ void emu_finish(void)
 #endif
 	}
 
+	pprof_finish();
+
 	PicoExit();
 }
 
@@ -1432,6 +1436,8 @@ void emu_loop(void)
 		unsigned int timestamp;
 		int diff, diff_lim;
 
+		pprof_start(main);
+
 		timestamp = get_ticks();
 		if (reset_timing) {
 			reset_timing = 0;
@@ -1537,7 +1543,7 @@ void emu_loop(void)
 		PicoFrame();
 		pemu_finalize_frame(fpsbuff, notice_msg);
 
-		//plat_video_flip();
+		// plat_video_flip();
 
 		/* frame limiter */
 		if (!reset_timing && !(currentConfig.EmuOpt & (EOPT_NO_FRMLIMIT|EOPT_EXT_FRMLIMIT)))
@@ -1560,6 +1566,8 @@ void emu_loop(void)
 		plat_video_flip();
 
 		pframes_done++; frames_done++; frames_shown++;
+
+		pprof_end(main);
 	}
 
 	emu_set_fastforward(0);

platform/linux/Makefile

@@ -88,6 +88,9 @@ PicoDrive : $(OBJS)
 	@echo ">>>" $@
 	$(CC) $(CFLAGS) $^ $(LDFLAGS) -Wl,-Map=PicoDrive.map -o $@
 
+pprof: pprof.c
+	$(CROSS)gcc -O2 -ggdb -DPPROF -DPPROF_TOOL -I../../ -I. $^ -o $@ -lrt
+
 %.o : %.asm
 	@echo ">>>" $<
 	nasm -f elf $< -o $@

platform/linux/pprof.c

@@ -0,0 +1,126 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
+
+#include <pico/pico_int.h>
+
+struct pp_counters *pp_counters;
+static int shmemid;
+
+void pprof_init(void)
+{
+	int this_is_new_shmem = 1;
+	key_t shmemkey;
+	void *shmem;
+
+#ifndef PPROF_TOOL
+	unsigned int tmp = pprof_get_one();
+	printf("pprof: measured diff is %u\n", pprof_get_one() - tmp);
+#endif
+
+	shmemkey = ftok(".", 0x02ABC32E);
+	if (shmemkey == -1)
+	{
+		perror("pprof: ftok failed");
+		return;
+	}
+
+#ifndef PPROF_TOOL
+	shmemid = shmget(shmemkey, sizeof(*pp_counters),
+		IPC_CREAT | IPC_EXCL | 0644);
+	if (shmemid == -1)
+#endif
+	{
+		shmemid = shmget(shmemkey, sizeof(*pp_counters),
+				0644);
+		if (shmemid == -1)
+		{
+			perror("pprof: shmget failed");
+			return;
+		}
+		this_is_new_shmem = 0;
+	}
+
+	shmem = shmat(shmemid, NULL, 0);
+	if (shmem == (void *)-1)
+	{
+		perror("pprof: shmat failed");
+		return;
+	}
+
+	pp_counters = shmem;
+	if (this_is_new_shmem) {
+		memset(pp_counters, 0, sizeof(*pp_counters));
+		printf("pprof: pp_counters cleared.\n");
+	}
+}
+
+void pprof_finish(void)
+{
+	shmdt(pp_counters);
+	shmctl(shmemid, IPC_RMID, NULL);
+}
+
+#ifdef PPROF_TOOL
+
+#define IT(n) { pp_##n, #n }
+static const struct {
+	enum pprof_points pp;
+	const char *name;
+} pp_tab[] = {
+	IT(main),
+	IT(frame),
+	IT(draw),
+	IT(sound),
+	IT(m68k),
+	IT(z80),
+	IT(msh2),
+	IT(ssh2),
+	IT(dummy),
+};
+
+int main(int argc, char *argv[])
+{
+	unsigned long long old[pp_total_points], new[pp_total_points];
+	int base = 0;
+	int l, i;
+
+	pprof_init();
+	if (pp_counters == NULL)
+		return 1;
+
+	if (argc >= 2)
+		base = atoi(argv[1]);
+
+	memset(old, 0, sizeof(old));
+	for (l = 0; ; l++)
+	{
+		if ((l & 0x1f) == 0) {
+			for (i = 0; i < ARRAY_SIZE(pp_tab); i++)
+				printf("%6s ", pp_tab[i].name);
+			printf("\n");
+		}
+
+		memcpy(new, pp_counters->counter, sizeof(new));
+		for (i = 0; i < ARRAY_SIZE(pp_tab); i++)
+		{
+			unsigned long long idiff = new[i] - old[i];
+			unsigned long long bdiff = (new[base] - old[base]) | 1;
+			printf("%6.2f ", (double)idiff * 100.0 / bdiff);
+		}
+		printf("\n");
+		memcpy(old, new, sizeof(old));
+
+		if (argc < 3)
+			break;
+		usleep(atoi(argv[2]));
+	}
+
+	return 0;
+}
+
+#endif // PPROF_TOOL
+

platform/linux/pprof.h

@@ -0,0 +1,54 @@
+#ifndef __PPROF_H__
+#define __PPROF_H__
+
+enum pprof_points {
+  pp_main,
+  pp_frame,
+  pp_draw,
+  pp_sound,
+  pp_m68k,
+  pp_z80,
+  pp_msh2,
+  pp_ssh2,
+  pp_dummy,
+  pp_total_points
+};
+
+struct pp_counters
+{
+	unsigned long long counter[pp_total_points];
+};
+
+extern struct pp_counters *pp_counters;
+
+#ifdef __i386__
+static __attribute__((always_inline)) inline unsigned int pprof_get_one(void)
+{
+  unsigned long long ret;
+  __asm__ __volatile__ ("rdtsc" : "=A" (ret));
+  return (unsigned int)ret;
+}
+
+#elif defined(__GP2X__)
+// XXX: MMSP2 only
+extern volatile unsigned long *gp2x_memregl;
+#define pprof_get_one() (unsigned int)gp2x_memregl[0x0a00 >> 2]
+
+#else
+#error no timer
+#endif
+
+#define pprof_start(point) { \
+    unsigned int pp_start_##point = pprof_get_one()
+#define pprof_end(point) \
+    pp_counters->counter[pp_##point] += pprof_get_one() - pp_start_##point; \
+  }
+// subtract for recursive stuff
+#define pprof_end_sub(point) \
+    pp_counters->counter[pp_##point] -= pprof_get_one() - pp_start_##point; \
+  }
+
+extern void pprof_init(void);
+extern void pprof_finish(void);
+
+#endif // __PPROF_H__