just some simple profiling for the LLE DSP

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@901 8ced0084-cf51-0410-be5f-012b33b47a6e

Docs/DSP/Crazy Taxi.txt

@@ -87,7 +87,7 @@ Init-Code()
 0021 00fe 0e31 SR	@0x0e31, $30
 0023 16fc dcd1 SI	@DMBH, #0xdcd1
 0025 16fd 0000 SI	@DMBL, #0x0000
-0027 16fb 0001 SI	@DIRQ, #0x0001
+0027 16fb 0001 SI	@DIRQ, #0x0001		// write mail: DSP_TASK_STATE_RUN(0xDCD10000) and calls initCB of the task
 
 // wait for mail
 0029 26fc      LRS	$30, @DMBH
@@ -109,7 +109,9 @@ Init-Code()
 0038 0092 00ff LRI	$18, #0x00ff
 003a 16fc dcd1 SI	@DMBH, #0xdcd1
 003c 16fd 0001 SI	@DMBL, #0x0001
-003e 16fb 0001 SI	@DIRQ, #0x0001
+003e 16fb 0001 SI	@DIRQ, #0x0001		// write mail: DSP_TASK_STATE_RUN(0xDCD10001) and calls resCB of the task
+
+
 0040 26fc      LRS	$30, @DMBH
 0041 02a0 8000 ANDCF	$30, #0x8000
 0043 029c 0040 JZR	0x0040
@@ -121,9 +123,15 @@ Main()
 0046 8100      CLR	$30
 0047 8900      CLR	$31
 0048 009f babe LRI	$31, #0xbabe
+
+
+// wait for dma timewasting :D
 004a 26fe      LRS	$30, @CMBH
 004b 02c0 8000 ANDF	$30, #0x8000
 004d 029c 004a JZR	0x004a
+
+
+
 004f 8200      CMP	
 0050 0294 004a JNE	0x004a
 0052 23ff      LRS	$27, @CMBL
@@ -727,6 +735,8 @@ Main()
 033b 1c03      MRR	$0, $3
 033c 1ff5      MRR	$31, $21
 033d 191a      LRRI	$26, @$0
+
+
 033e f858      ADDPAXZ.L	$30, $26 : $27, @$0
 033f fba0      ADDPAXZ.LS	$31, $27 : $26, $30
 0340 f8b1      ADDPAXZ.LS	$30, $26 : $27, $31
@@ -736,6 +746,10 @@ Main()
 0344 f8b1      ADDPAXZ.LS	$30, $26 : $27, $31
 0345 fba0      ADDPAXZ.LS	$31, $27 : $26, $30
 0346 f83b      ADDPAXZ.S	$30, $26 : @$3, $31
+
+
+
+
 0347 1b7e      SRRI	@$3, $30
 0348 0083 0e04 LRI	$3, #0x0e04
 034a 8100      CLR	$30
@@ -818,6 +832,11 @@ Main()
 03b3 16cd 03c0 SI	@DSPA, #0x03c0
 03b5 16c9 0000 SI	@DSCR, #0x0000
 03b7 16cb 0080 SI	@DSBL, #0x0080    // Don't know w.at this dma xfer is
+
+
+
+
+
 03b9 8100      CLR	$30
 03ba 8900      CLR	$31
 03bb 00de 0b84 LR	$30, @0x0b84
@@ -2598,11 +2617,11 @@ Main()
 0c90 0cdd      LRIS	$28, #0xdd			// normal case to return to the main-loop
 
 
-// Mix and write back to RAM??? It is called by AXLIST_END command only
+// Wait for new message from the CPU
 0c91 8e00      S40	
 0c92 8100      CLR	$30
 0c93 8900      CLR	$31
-0c94 02bf 0ce0 CALL	0x0ce0    // wait for CMBH
+0c94 02bf 0ce0 CALL	0x0ce0    			// wait for answer for our "0xDCD10002"-message... and we have to wait really long :(
 0c96 27ff      LRS	$31, @CMBL
 0c97 009e 0c8d LRI	$30, #0x0c8d
 0c99 4c00      ADD	$30, $31
@@ -2612,6 +2631,8 @@ Main()
 0c9d 176f      JMPR	$3
 0c9e 0021      HALT	
 
+//////////////////////////////////////////////////////////////////////////7
+
 // case 0x00:
 0c9f 029f 0030 JMP	0x0030
 0ca1 0021      HALT	

Source/Plugins/Plugin_DSP_LLE/Src/Globals.cpp

@@ -88,4 +88,48 @@ u32 Memory_Read_U32(u32 _uAddress)
 	_uAddress &= RAM_MASK;
 	return Common::swap32(*(u32*)&g_dsp.cpu_ram[_uAddress]);
 }
-// =============
+
+#if PROFILE
+
+#define PROFILE_MAP_SIZE 0x10000
+
+u64 g_profileMap[PROFILE_MAP_SIZE];
+bool g_profile = false;
+
+void ProfilerStart()
+{
+	g_profile = true;
+}
+
+void ProfilerAddDelta(int _addr, int _delta)
+{
+	if (g_profile)
+	{
+		g_profileMap[_addr] += _delta;
+	}	
+}
+
+void ProfilerInit()
+{
+	memset(g_profileMap, 0, sizeof(g_profileMap));
+}
+
+void ProfilerDump(uint64 count)
+{
+	FILE* pFile = fopen("c:\\_\\DSP_Prof.txt", "wt");
+	if (pFile != NULL)
+	{
+		fprintf(pFile, "Number of DSP steps: %i\n\n", count);
+		for (int i=0; i<PROFILE_MAP_SIZE;i++)
+		{
+			if (g_profileMap[i] > 0)
+			{
+				fprintf(pFile, "0x%04X: %u\n", i, g_profileMap[i]);
+			}
+		}
+
+		fclose(pFile);
+	}
+}
+
+#endif

Source/Plugins/Plugin_DSP_LLE/Src/Globals.h

@@ -65,5 +65,13 @@ typedef const uint32 cuint32;
 u16 Memory_Read_U16(u32 _uAddress); // For PB address detection
 u32 Memory_Read_U32(u32 _uAddress);
 
+#if PROFILE	
+	void ProfilerDump(uint64 _count);
+	void ProfilerInit();
+	void ProfilerAddDelta(int _addr, int _delta);
+	void ProfilerStart();
+#endif
+
+
 #endif
 

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_interpreter.cpp

@@ -228,11 +228,25 @@ void gdsp_step()
 
 	if (g_dsp.pc == 0x80e7)
 	{
-		g_dsp.pc = HLE_ROM_80E7_81F8();
+		//g_dsp.pc = HLE_ROM_80E7_81F8();
 	}
 
 	g_dsp.err_pc = g_dsp.pc;
 
+#if PROFILE
+	ProfilerAddDelta(g_dsp.err_pc, 1);
+	if (g_dsp.step_counter == 1)
+	{
+		ProfilerInit();
+	}
+
+	if ((g_dsp.step_counter & 0xFFFFF) == 0)
+	{
+		ProfilerDump(g_dsp.step_counter);
+	}
+
+#endif
+
 	uint16 opc = dsp_fetch_code();
 	dsp_op[opc >> 12](opc);
 

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_interpreter.h

@@ -91,5 +91,4 @@ uint16* gdsp_get_irom(void);
 uint16* gdsp_get_dram(void);
 uint16* gdsp_get_drom(void);
 
-
 #endif

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_opcodes_helper.h

@@ -107,6 +107,10 @@ inline void dsp_op_write_reg(uint8 reg, uint16 val)
 
 inline sint64 dsp_get_long_prod()
 {
+#if PROFILE
+	ProfilerAddDelta(g_dsp.err_pc, 1);
+#endif
+
 	sint64 val;
 	sint64 low_prod;
 	val   = (sint8)g_dsp.r[0x16];
@@ -122,6 +126,10 @@ inline sint64 dsp_get_long_prod()
 
 inline void dsp_set_long_prod(sint64 val)
 {
+#if PROFILE
+	ProfilerAddDelta(g_dsp.err_pc, 1);
+#endif
+
 	g_dsp.r[0x14] = (uint16)val;
 	val >>= 16;
 	g_dsp.r[0x15] = (uint16)val;
@@ -139,6 +147,10 @@ inline void dsp_set_long_prod(sint64 val)
 
 inline sint64 dsp_get_long_acc(uint8 reg)
 {
+#if PROFILE
+	ProfilerAddDelta(g_dsp.err_pc, 1);
+#endif
+
 	_dbg_assert_(reg < 2);
 	sint64 val;
 	sint64 low_acc;
@@ -154,6 +166,10 @@ inline sint64 dsp_get_long_acc(uint8 reg)
 
 inline uint64 dsp_get_ulong_acc(uint8 reg)
 {
+#if PROFILE
+	ProfilerAddDelta(g_dsp.err_pc, 1);
+#endif
+
 	_dbg_assert_(reg < 2);
 	uint64 val;
 	uint64 low_acc;
@@ -169,6 +185,10 @@ inline uint64 dsp_get_ulong_acc(uint8 reg)
 
 inline void dsp_set_long_acc(uint8 _reg, sint64 val)
 {
+#if PROFILE
+	ProfilerAddDelta(g_dsp.err_pc, 1);
+#endif
+
 	_dbg_assert_(_reg < 2);
 	g_dsp.r[0x1c + _reg] = (uint16)val;
 	val >>= 16;
@@ -208,6 +228,10 @@ inline sint16 dsp_get_acc_h(uint8 _reg)
 
 inline sint64 dsp_get_long_acx(uint8 _reg)
 {
+#if PROFILE
+	ProfilerAddDelta(g_dsp.err_pc, 1);
+#endif
+
 	_dbg_assert_(_reg < 2);
 	sint64 val = (sint16)g_dsp.r[0x1a + _reg];
 	val <<= 16;

Source/Plugins/Plugin_DSP_LLE/Src/main.cpp

@@ -66,7 +66,7 @@ bool AXTask(u32& _uMail);
 bool bCanWork = false;
 
 // Set this if you want to log audio. search for log_ai in this file to see the filename.
-static bool log_ai = false;
+static bool log_ai = true;
 WaveFileWriter g_wave_writer;
 
 // ==============
@@ -258,10 +258,11 @@ void DSP_Initialize(DSPInitialize _dspInitialize)
 // --------------
 //		Dump UCode to file...
 
-   	FILE* t = fopen("DSP_UC_09CD143F.txt", "wb");
+   	FILE* t = fopen("C:\\_\\DSP_UC_09CD143F.txt", "wb");
-	if(t) {
+	if (t != NULL)
+	{
    		gd_globals_t gdg;
-   		gd_dis_file(&gdg, "DSP_UC_09CD143F.bin", t);
+   		gd_dis_file(&gdg, "C:\\_\\DSP_UC_09CD143F.bin", t);
    		fclose(t);   
 	}
 // --------------
@@ -353,6 +354,13 @@ void DSP_WriteMailboxHigh(bool _CPUMailbox, u16 _uHighMail)
 			ErrorLog("Mailbox isnt empty ... strange");
 		}
 
+#if PROFILE
+		if ((_uHighMail) == 0xBABE)
+		{
+			ProfilerStart();
+		}
+#endif
+
 		gdsp_mbox_write_h(GDSP_MBOX_CPU, _uHighMail);
 	}
 	else