// Copyright (C) 2003 Dolphin Project. // This program is free software: you can redistribute it and/or modify // it under the terms of the GNU General Public License as published by // the Free Software Foundation, version 2.0. // This program is distributed in the hope that it will be useful, // but WITHOUT ANY WARRANTY; without even the implied warranty of // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the // GNU General Public License 2.0 for more details. // A copy of the GPL 2.0 should have been included with the program. // If not, see http://www.gnu.org/licenses/ // Official SVN repository and contact information can be found at // http://code.google.com/p/dolphin-emu/ // NOTES (mb2): // * GP/CPU sync can be done by several way: // - MP1 use BP (breakpoint) in movie-menus and mostly PEtoken in 3D // - ZWW as Crazy Taxi: PEfinish (GXSetDrawDone) // - SMS: BP, PEToken, PEfinish // - ZTP: seems to use PEfinish only // - Animal Crossing: PEfinish at start but there's a bug... // There's tons of HiWmk/LoWmk ping pong -> Another sync fashion? // - Super Monkey Ball Adventures: PEToken. Oddity: read&check-PEToken-value-loop stays // in its JITed block (never fall in Advance() until the game-watchdog's stuff). // That's why we can't let perform the AdvanceCallBack as usual. // The PEToken is volatile now and in the fifo struct. // - Super Monkey Ball: PEFinish. This game has the lamest way to deal with fifo sync for our MT's stuff. // A hack is mandatory. DONE and should be ok for other games. // *What I guess (thx to asynchronous DualCore mode): // PPC have a frame-finish watchdog. Handled by system timming stuff like the decrementer. // (DualCore mode): I have observed, after ZTP logos, a fifo-recovery start when DECREMENTER_EXCEPTION is throwned. // The frame setting (by GP) took too much time and didn't finish properly due to this watchdog. // Faster GX plugins required, indeed :p // * BPs are needed for some game GP/CPU sync. // But it could slowdown (MP1 at least) because our GP in DC is faster than "expected" in some area. // eg: in movie-menus in MP1, BP are reached quickly. // The bad thing is that involve too much PPC work (int ack, lock GP, reset BP, new BP addr, unlock BP...) hence the slowdown. // Anyway, emulation should more accurate like this and it emulate some sort of better load balancing. // Eather way in those area a more accurate GP timing could be done by slowing down the GP or something less stupid. // Not functional and not used atm (breaks MP2). // * funny, in revs before those with this note, BP irq wasn't cleared (a bug indeed) and MP1 menus was faster. // BP irq was raised and ack just once but never cleared. However it's sufficient for MP1 to work. // This hack is used atm. Known BPs handling doesn't work well (btw, BP irq clearing might be done by CPIntEnable raising edge). // The hack seems to be responsible of the movie stutering in MP1 menus. // TODO (mb2): // * raise watermark Ov/Un irq: POINTLESS since emulated GP timings can't be accuratly set. // Only 3 choices IMHO for a correct emulated load balancing in DC mode: // - make our own GP watchdog hack that can lock CPU if GP too slow. STARTED // - hack directly something in PPC timings (dunno how) // - boost GP so we can consider it as infinitely fast compared to CPU. // * raise ReadIdle/CmdIdle flags and observe behaviour of MP1 & ZTP (at least) // * Clean useless comments and debug stuff in Read16, Write16, GatherPipeBursted when sync will be fixed for DC // * (reminder) do the same in: // PeripheralInterface.cpp, PixelEngine.cpp, OGL->BPStructs.cpp, fifo.cpp... ok just check change log >> // TODO // * Kick GPU from dispatcher, not from writes // * Thunking framework // * Cleanup of messy now unnecessary safety code in jit #include "Common.h" #include "VideoCommon.h" #include "MathUtil.h" #include "Thread.h" #include "Atomic.h" #include "Fifo.h" #include "ChunkFile.h" #include "CommandProcessor.h" namespace CommandProcessor { // look for 1002 verts, breakpoint there, see why next draw is flushed // TODO(ector): Warn on bbox read/write // Fifo Status Register union UCPStatusReg { struct { unsigned OverflowHiWatermark : 1; unsigned UnderflowLoWatermark : 1; unsigned ReadIdle : 1; unsigned CommandIdle : 1; unsigned Breakpoint : 1; unsigned : 11; }; u16 Hex; UCPStatusReg() {Hex = 0; } UCPStatusReg(u16 _hex) {Hex = _hex; } }; // Fifo Control Register union UCPCtrlReg { struct { unsigned GPReadEnable : 1; unsigned CPIntEnable : 1; unsigned FifoOverflowIntEnable : 1; unsigned FifoUnderflowIntEnable : 1; unsigned GPLinkEnable : 1; unsigned BPEnable : 1; unsigned : 10; }; u16 Hex; UCPCtrlReg() {Hex = 0; } UCPCtrlReg(u16 _hex) {Hex = _hex; } }; // Fifo Control Register union UCPClearReg { struct { unsigned ClearFifoOverflow : 1; unsigned ClearFifoUnderflow : 1; unsigned ClearMetrices : 1; unsigned : 13; }; u16 Hex; UCPClearReg() {Hex = 0; } UCPClearReg(u16 _hex) {Hex = _hex; } }; // STATE_TO_SAVE // variables UCPStatusReg m_CPStatusReg; UCPCtrlReg m_CPCtrlReg; UCPClearReg m_CPClearReg; int m_bboxleft; int m_bboxtop; int m_bboxright; int m_bboxbottom; u16 m_tokenReg; SCPFifoStruct fifo; //This one is shared between gfx thread and emulator thread static u32 fake_GPWatchdogLastToken = 0; static Common::Event s_fifoIdleEvent; enum { GATHER_PIPE_SIZE = 32, INT_CAUSE_CP = 0x800 }; void DoState(PointerWrap &p) { p.Do(m_CPStatusReg); p.Do(m_CPCtrlReg); p.Do(m_CPClearReg); p.Do(m_bboxleft); p.Do(m_bboxtop); p.Do(m_bboxright); p.Do(m_bboxbottom); p.Do(m_tokenReg); p.Do(fifo); } // function void UpdateFifoRegister(); void UpdateInterrupts(); //inline void WriteLow (u32& _reg, u16 lowbits) {_reg = (_reg & 0xFFFF0000) | lowbits;} //inline void WriteHigh(u32& _reg, u16 highbits) {_reg = (_reg & 0x0000FFFF) | ((u32)highbits << 16);} inline void WriteLow (volatile u32& _reg, u16 lowbits) {Common::AtomicStore(_reg,(_reg & 0xFFFF0000) | lowbits);} inline void WriteHigh(volatile u32& _reg, u16 highbits) {Common::AtomicStore(_reg,(_reg & 0x0000FFFF) | ((u32)highbits << 16));} inline u16 ReadLow (u32 _reg) {return (u16)(_reg & 0xFFFF);} inline u16 ReadHigh (u32 _reg) {return (u16)(_reg >> 16);} int et_UpdateInterrupts; // for GP watchdog hack void IncrementGPWDToken() { Common::AtomicIncrement(fifo.Fake_GPWDToken); } // Check every FAKE_GP_WATCHDOG_PERIOD if a PE-frame-finish occured // if not then lock CPUThread until GP finish a frame. void WaitForFrameFinish() { while ((fake_GPWatchdogLastToken == fifo.Fake_GPWDToken) && fifo.bFF_GPReadEnable && (fifo.CPReadWriteDistance > 0) && !(fifo.bFF_BPEnable && fifo.bFF_Breakpoint)) s_fifoIdleEvent.MsgWait(); fake_GPWatchdogLastToken = fifo.Fake_GPWDToken; } void UpdateInterrupts_Wrapper(u64 userdata, int cyclesLate) { UpdateInterrupts(); } void Init() { m_CPStatusReg.Hex = 0; m_CPStatusReg.CommandIdle = 1; m_CPStatusReg.ReadIdle = 1; m_CPCtrlReg.Hex = 0; m_bboxleft = 0; m_bboxtop = 0; m_bboxright = 640; m_bboxbottom = 480; m_tokenReg = 0; fake_GPWatchdogLastToken = 0; memset(&fifo,0,sizeof(fifo)); fifo.CPCmdIdle = 1 ; fifo.CPReadIdle = 1; s_fifoIdleEvent.Init(); et_UpdateInterrupts = g_VideoInitialize.pRegisterEvent("UpdateInterrupts", UpdateInterrupts_Wrapper); } void Shutdown() { s_fifoIdleEvent.Shutdown(); } void Read16(u16& _rReturnValue, const u32 _Address) { DEBUG_LOG(COMMANDPROCESSOR, "(r): 0x%08x", _Address); switch (_Address & 0xFFF) { case STATUS_REGISTER: //TODO?: if really needed //m_CPStatusReg.CommandIdle = fifo.CPCmdIdle; // uncomment: change a bit the behaviour MP1. Not very useful though m_CPStatusReg.ReadIdle = fifo.CPReadIdle; //m_CPStatusReg.CommandIdle = fifo.CPReadIdle; // hack: CPU will always believe fifo is empty and on idle //m_CPStatusReg.ReadIdle = 1; //m_CPStatusReg.CommandIdle = 1; _rReturnValue = m_CPStatusReg.Hex; INFO_LOG(COMMANDPROCESSOR, "\t iBP %s | fREADIDLE %s | fCMDIDLE %s | iOvF %s | iUndF %s" , m_CPStatusReg.Breakpoint ? "ON" : "OFF" , m_CPStatusReg.ReadIdle ? "ON" : "OFF" , m_CPStatusReg.CommandIdle ? "ON" : "OFF" , m_CPStatusReg.OverflowHiWatermark ? "ON" : "OFF" , m_CPStatusReg.UnderflowLoWatermark ? "ON" : "OFF" ); return; case CTRL_REGISTER: _rReturnValue = m_CPCtrlReg.Hex; return; case CLEAR_REGISTER: _rReturnValue = m_CPClearReg.Hex; return; case FIFO_TOKEN_REGISTER: _rReturnValue = m_tokenReg; return; case FIFO_BOUNDING_BOX_LEFT: _rReturnValue = m_bboxleft; return; case FIFO_BOUNDING_BOX_RIGHT: _rReturnValue = m_bboxright; return; case FIFO_BOUNDING_BOX_TOP: _rReturnValue = m_bboxtop; return; case FIFO_BOUNDING_BOX_BOTTOM: _rReturnValue = m_bboxbottom; return; case FIFO_BASE_LO: _rReturnValue = ReadLow (fifo.CPBase); return; case FIFO_BASE_HI: _rReturnValue = ReadHigh(fifo.CPBase); return; case FIFO_END_LO: _rReturnValue = ReadLow (fifo.CPEnd); return; case FIFO_END_HI: _rReturnValue = ReadHigh(fifo.CPEnd); return; case FIFO_HI_WATERMARK_LO: _rReturnValue = ReadLow (fifo.CPHiWatermark); return; case FIFO_HI_WATERMARK_HI: _rReturnValue = ReadHigh(fifo.CPHiWatermark); return; case FIFO_LO_WATERMARK_LO: _rReturnValue = ReadLow (fifo.CPLoWatermark); return; case FIFO_LO_WATERMARK_HI: _rReturnValue = ReadHigh(fifo.CPLoWatermark); return; // TODO: cases cleanup case FIFO_RW_DISTANCE_LO: //_rReturnValue = ReadLow (fifo.CPReadWriteDistance); // hack: CPU will always believe fifo is empty and on idle _rReturnValue = 0; DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_RW_DISTANCE_LO : %04x", _rReturnValue); return; case FIFO_RW_DISTANCE_HI: //_rReturnValue = ReadHigh(fifo.CPReadWriteDistance); // hack: CPU will always believe fifo is empty and on idle _rReturnValue = 0; DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_RW_DISTANCE_HI : %04x", _rReturnValue); return; case FIFO_WRITE_POINTER_LO: _rReturnValue = ReadLow (fifo.CPWritePointer); DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_WRITE_POINTER_LO : %04x", _rReturnValue); return; case FIFO_WRITE_POINTER_HI: _rReturnValue = ReadHigh(fifo.CPWritePointer); DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_WRITE_POINTER_HI : %04x", _rReturnValue); return; case FIFO_READ_POINTER_LO: //_rReturnValue = ReadLow (fifo.CPReadPointer); // hack: CPU will always believe fifo is empty and on idle _rReturnValue = ReadLow (fifo.CPWritePointer); DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_READ_POINTER_LO : %04x", _rReturnValue); return; case FIFO_READ_POINTER_HI: //_rReturnValue = ReadHigh(fifo.CPReadPointer); // hack: CPU will always believe fifo is empty and on idle _rReturnValue = ReadHigh(fifo.CPWritePointer); DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_READ_POINTER_HI : %04x", _rReturnValue); return; case FIFO_BP_LO: _rReturnValue = ReadLow (fifo.CPBreakpoint); return; case FIFO_BP_HI: _rReturnValue = ReadHigh(fifo.CPBreakpoint); return; case CP_PERF0_L: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF0_L: %04x", _rReturnValue); break; // XF counters case CP_PERF0_H: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF0_H: %04x", _rReturnValue); break; case CP_PERF1_L: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF1_L: %04x", _rReturnValue); break; case CP_PERF1_H: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF1_H: %04x", _rReturnValue); break; case CP_PERF2_L: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF2_L: %04x", _rReturnValue); break; case CP_PERF2_H: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF2_H: %04x", _rReturnValue); break; case CP_PERF3_L: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF3_L: %04x", _rReturnValue); break; case CP_PERF3_H: _rReturnValue = 0; WARN_LOG(COMMANDPROCESSOR, "Read from PERF3_H: %04x", _rReturnValue); break; // case 0x64: // return 4; //Number of clocks per vertex.. todo: calculate properly //add all the other regs here? are they ever read? default: WARN_LOG(COMMANDPROCESSOR, "(r16) unknown CP reg @ %08x", _Address); _rReturnValue = 0; return; } } bool AllowIdleSkipping() { return !g_VideoInitialize.bOnThread || (!m_CPCtrlReg.CPIntEnable && !m_CPCtrlReg.BPEnable); } void Write16(const u16 _Value, const u32 _Address) { INFO_LOG(COMMANDPROCESSOR, "(write16): 0x%04x @ 0x%08x",_Value,_Address); //Spin until queue is empty - it WILL become empty because this is the only thread //that submits data if (g_VideoInitialize.bOnThread) { // Force complete fifo flush if we attempt to set/reset the fifo (API GXSetGPFifo or equivalent) // It's kind of an API hack but it works for lots of games... and I hope it's the same way for every games. // TODO: HLE for GX fifo's APIs? // Here is the hack: // - if (attempt to overwrite CTRL_REGISTER by 0x0000) // // then we assume CPReadWriteDistance will be overwrited very soon. // - if (fifo is not empty) // // (not 100% sure): shouln't happen unless PPC think having trouble with the sync // // and it attempt a fifo recovery (look for PI_FIFO_RESET in log). // // If we want to emulate self fifo recovery we need proper GX metrics emulation... yeah sure :p // - spin until fifo is empty // - else // - normal write16 if (((_Address&0xFFF) == CTRL_REGISTER) && (_Value == 0)) // API hack { // weird MP1 redo that right after linking fifo with GP... hmmm /*_dbg_assert_msg_(COMMANDPROCESSOR, fifo.CPReadWriteDistance == 0, "WTF! Something went wrong with GP/PPC the sync! -> CPReadWriteDistance: 0x%08X\n" " - The fifo is not empty but we are going to lock it anyway.\n" " - \"Normaly\", this is due to fifo-hang-so-lets-attempt-recovery.\n" " - The bad news is dolphin don't support special recovery features like GXfifo's metric yet.\n" " - The good news is, the time you read that message, the fifo should be empty now :p\n" " - Anyway, fifo flush will be forced if you press OK and dolphin might continue to work...\n" " - We aren't betting on that :)", fifo.CPReadWriteDistance); */ DEBUG_LOG(COMMANDPROCESSOR, "*********************** GXSetGPFifo very soon? ***********************"); // (mb2) We don't sleep here since it could be a perf issue for super monkey ball (yup only this game IIRC) // Touching that game is a no-go so I don't want to take the risk :p while (fifo.bFF_GPReadEnable && fifo.CPReadWriteDistance > 0 && !(fifo.bFF_BPEnable && fifo.bFF_Breakpoint) ) s_fifoIdleEvent.MsgWait(); } } switch (_Address & 0xFFF) { case STATUS_REGISTER: { UCPStatusReg tmpStatus(_Value); // set the flags to "all is okay" m_CPStatusReg.OverflowHiWatermark = 0; m_CPStatusReg.UnderflowLoWatermark = 0; // TOCHECK (mb2): could BP irq be cleared here too? //if (tmpStatus.Breakpoint!=m_CPStatusReg.Breakpoint) _asm int 3 // breakpoint /*if (tmpStatus.Breakpoint) { m_CPStatusReg.Breakpoint = 0; } //fifo.bFF_Breakpoint = m_CPStatusReg.Breakpoint; fifo.bFF_Breakpoint = m_CPStatusReg.Breakpoint ? true : false; //LOG(COMMANDPROCESSOR,"fifo.bFF_Breakpoint : %i",fifo.bFF_Breakpoint); */ // update interrupts UpdateInterrupts(); INFO_LOG(COMMANDPROCESSOR,"\t write to STATUS_REGISTER : %04x", _Value); } break; case CTRL_REGISTER: { UCPCtrlReg tmpCtrl(_Value); Common::AtomicStore(fifo.bFF_GPReadEnable, tmpCtrl.GPReadEnable); Common::AtomicStore(fifo.bFF_GPLinkEnable, tmpCtrl.GPLinkEnable); Common::AtomicStore(fifo.bFF_BPEnable, tmpCtrl.BPEnable); // TOCHECK (mb2): could BP irq be cleared with w16 to STATUS_REGISTER? // funny hack: eg in MP1 if we disable the clear breakpoint ability by commenting this block // the game is of course faster but looks stable too. // Well, the hack is more stable than the "proper" way actualy :p ... it breaks MP2 when ship lands // So I let the hack for now. // Checkmate re-enabled it, so please test // TODO (mb2): fix this! // BP interrupt is cleared here //if (m_CPCtrlReg.CPIntEnable && !tmpCtrl.Hex) // falling edge // raising edge or falling egde if ((!m_CPCtrlReg.CPIntEnable && tmpCtrl.CPIntEnable) || (m_CPCtrlReg.CPIntEnable && !tmpCtrl.Hex)) { m_CPStatusReg.Breakpoint = 0; Common::AtomicStore(fifo.bFF_Breakpoint, 0); } m_CPCtrlReg.Hex = tmpCtrl.Hex; UpdateInterrupts(); DEBUG_LOG(COMMANDPROCESSOR,"\t write to CTRL_REGISTER : %04x", _Value); DEBUG_LOG(COMMANDPROCESSOR, "\t GPREAD %s | CPULINK %s | BP %s || CPIntEnable %s | OvF %s | UndF %s" , fifo.bFF_GPReadEnable ? "ON" : "OFF" , fifo.bFF_GPLinkEnable ? "ON" : "OFF" , fifo.bFF_BPEnable ? "ON" : "OFF" , m_CPCtrlReg.CPIntEnable ? "ON" : "OFF" , m_CPCtrlReg.FifoOverflowIntEnable ? "ON" : "OFF" , m_CPCtrlReg.FifoUnderflowIntEnable ? "ON" : "OFF" ); } break; case PERF_SELECT: { WARN_LOG(COMMANDPROCESSOR, "write to PERF_SELECT: %04x", _Value); // Seems to select which set of perf counters should be exposed. } break; case CLEAR_REGISTER: { // ???? UCPClearReg tmpClearReg(_Value); m_CPClearReg.Hex = 0; INFO_LOG(COMMANDPROCESSOR,"\t write to CLEAR_REGISTER : %04x",_Value); } break; // Fifo Registers case FIFO_TOKEN_REGISTER: m_tokenReg = _Value; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_TOKEN_REGISTER : %04x", _Value); break; case FIFO_BASE_LO: WriteLow ((u32 &)fifo.CPBase, _Value); fifo.CPBase &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_BASE_LO : %04x", _Value); break; case FIFO_BASE_HI: WriteHigh((u32 &)fifo.CPBase, _Value); fifo.CPBase &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_BASE_HI : %04x", _Value); break; case FIFO_END_LO: WriteLow ((u32 &)fifo.CPEnd, _Value); fifo.CPEnd &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_END_LO : %04x", _Value); break; case FIFO_END_HI: WriteHigh((u32 &)fifo.CPEnd, _Value); fifo.CPEnd &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_END_HI : %04x", _Value); break; // Hm. Should we really & these with FFFFFFE0? // (mb2): never seen 32B not aligned values for those following regs. // fifo.CPEnd is the only value that could be not 32B aligned so far. case FIFO_WRITE_POINTER_LO: WriteLow ((u32 &)fifo.CPWritePointer, _Value); fifo.CPWritePointer &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_WRITE_POINTER_LO : %04x", _Value); break; case FIFO_WRITE_POINTER_HI: WriteHigh((u32 &)fifo.CPWritePointer, _Value); fifo.CPWritePointer &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_WRITE_POINTER_HI : %04x", _Value); break; case FIFO_READ_POINTER_LO: WriteLow ((u32 &)fifo.CPReadPointer, _Value); fifo.CPReadPointer &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_READ_POINTER_LO : %04x", _Value); break; case FIFO_READ_POINTER_HI: WriteHigh((u32 &)fifo.CPReadPointer, _Value); fifo.CPReadPointer &= 0xFFFFFFE0; DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_READ_POINTER_HI : %04x", _Value); break; case FIFO_HI_WATERMARK_LO: WriteLow ((u32 &)fifo.CPHiWatermark, _Value); DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_HI_WATERMARK_LO : %04x", _Value); break; case FIFO_HI_WATERMARK_HI: WriteHigh((u32 &)fifo.CPHiWatermark, _Value); DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_HI_WATERMARK_HI : %04x", _Value); break; case FIFO_LO_WATERMARK_LO: WriteLow ((u32 &)fifo.CPLoWatermark, _Value); DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_LO_WATERMARK_LO : %04x", _Value); break; case FIFO_LO_WATERMARK_HI: WriteHigh((u32 &)fifo.CPLoWatermark, _Value); DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_LO_WATERMARK_HI : %04x", _Value); break; case FIFO_BP_LO: WriteLow ((u32 &)fifo.CPBreakpoint, _Value); DEBUG_LOG(COMMANDPROCESSOR,"write to FIFO_BP_LO : %04x", _Value); break; case FIFO_BP_HI: WriteHigh((u32 &)fifo.CPBreakpoint, _Value); DEBUG_LOG(COMMANDPROCESSOR,"write to FIFO_BP_HI : %04x", _Value); break; // Super monkey try to overwrite CPReadWriteDistance by an old saved RWD value. Which is lame for us. // hack: We have to force CPU to think fifo is alway empty and on idle. // When we fall here CPReadWriteDistance should be always null and the game should always want to overwrite it by 0. // So, we can skip it. case FIFO_RW_DISTANCE_HI: //WriteHigh((u32 &)fifo.CPReadWriteDistance, _Value); DEBUG_LOG(COMMANDPROCESSOR,"try to write to FIFO_RW_DISTANCE_HI : %04x", _Value); break; case FIFO_RW_DISTANCE_LO: //WriteLow((u32 &)fifo.CPReadWriteDistance, _Value); DEBUG_LOG(COMMANDPROCESSOR,"try to write to FIFO_RW_DISTANCE_LO : %04x", _Value); break; default: WARN_LOG(COMMANDPROCESSOR, "(w16) unknown CP reg write %04x @ %08x", _Value, _Address); } // TODO(mb2): better. Check if it help: avoid CPReadPointer overwrites when stupidly done like in Super Monkey Ball if ((!fifo.bFF_GPReadEnable && fifo.CPReadIdle) || !g_VideoInitialize.bOnThread) // TOCHECK(mb2): check again if thread safe? UpdateFifoRegister(); } void Read32(u32& _rReturnValue, const u32 _Address) { _rReturnValue = 0; _dbg_assert_msg_(COMMANDPROCESSOR, 0, "Read32 from CommandProccessor at 0x%08x", _Address); } void Write32(const u32 _Data, const u32 _Address) { _dbg_assert_msg_(COMMANDPROCESSOR, 0, "Write32 at CommandProccessor at 0x%08x", _Address); } void STACKALIGN GatherPipeBursted() { // if we aren't linked, we don't care about gather pipe data if (!fifo.bFF_GPLinkEnable) return; if (g_VideoInitialize.bOnThread) { // update the fifo-pointer fifo.CPWritePointer += GATHER_PIPE_SIZE; if (fifo.CPWritePointer >= fifo.CPEnd) fifo.CPWritePointer = fifo.CPBase; Common::AtomicAdd(fifo.CPReadWriteDistance, GATHER_PIPE_SIZE); // High watermark overflow handling (hacked way) if (fifo.CPReadWriteDistance > fifo.CPHiWatermark) { // we should raise an Ov interrupt for an accurate fifo emulation and let PPC deal with it. // But it slowdowns things because of if(interrupt blah blah){} blocks for each 32B fifo transactions. // CPU would be a bit more loaded too by its interrupt handling... // Eather way, CPU would have the ability to resume another thread. // To be clear: this spin loop is like a critical section spin loop in the emulated GX thread hence "hacked way" // Yes, in real life, the only purpose of the low watermark interrupt is just for cooling down OV contention. // - @ game start -> watermark init: Overflow enabled, Underflow disabled // - if (OV is raised) // - CPU stop to write to fifo // - enable Underflow interrupt (this only happens if OV is raised) // - do other things // - if (Underflow is raised (implicite: AND if an OV has been raised)) // - CPU can write to fifo // - disable Underflow interrupt INFO_LOG(COMMANDPROCESSOR, "(GatherPipeBursted): CPHiWatermark reached"); // Wait for GPU to catch up while (!(fifo.bFF_BPEnable && fifo.bFF_Breakpoint) && fifo.CPReadWriteDistance > fifo.CPLoWatermark) s_fifoIdleEvent.MsgWait(); } // check if we are in sync _assert_msg_(COMMANDPROCESSOR, fifo.CPWritePointer == *(g_VideoInitialize.Fifo_CPUWritePointer), "FIFOs linked but out of sync"); _assert_msg_(COMMANDPROCESSOR, fifo.CPBase == *(g_VideoInitialize.Fifo_CPUBase), "FIFOs linked but out of sync"); _assert_msg_(COMMANDPROCESSOR, fifo.CPEnd == *(g_VideoInitialize.Fifo_CPUEnd), "FIFOs linked but out of sync"); } else { fifo.CPWritePointer += GATHER_PIPE_SIZE; if (fifo.CPWritePointer >= fifo.CPEnd) fifo.CPWritePointer = fifo.CPBase; // check if we are in sync _assert_msg_(COMMANDPROCESSOR, fifo.CPWritePointer == *(g_VideoInitialize.Fifo_CPUWritePointer), "FIFOs linked but out of sync"); _assert_msg_(COMMANDPROCESSOR, fifo.CPBase == *(g_VideoInitialize.Fifo_CPUBase), "FIFOs linked but out of sync"); _assert_msg_(COMMANDPROCESSOR, fifo.CPEnd == *(g_VideoInitialize.Fifo_CPUEnd), "FIFOs linked but out of sync"); UpdateFifoRegister(); } } // This is mostly used in single core mode void CatchUpGPU() { // check if we are able to run this buffer if ((fifo.bFF_GPReadEnable) && !(fifo.bFF_BPEnable && fifo.bFF_Breakpoint)) { // HyperIris: Memory_GetPtr is an expensive call, call it less, run faster u8 *ptr = Memory_GetPtr(fifo.CPReadPointer); while (fifo.CPReadWriteDistance > 0) { // check if we are on a breakpoint if (fifo.bFF_BPEnable) { //MessageBox(0,"Breakpoint enabled",0,0); if ((fifo.CPReadPointer & ~0x1F) == (fifo.CPBreakpoint & ~0x1F)) { //_assert_msg_(POWERPC,0,"BP: %08x",fifo.CPBreakpoint); //LOG(COMMANDPROCESSOR,"!!! BP irq raised"); fifo.bFF_Breakpoint = 1; m_CPStatusReg.Breakpoint = 1; UpdateInterrupts(); break; } } // read the data and send it to the VideoPlugin fifo.CPReadPointer += 32; // We are going to do FP math on the main thread so have to save the current state SaveSSEState(); LoadDefaultSSEState(); Fifo_SendFifoData(ptr,32); LoadSSEState(); // adjust ptr += 32; fifo.CPReadWriteDistance -= 32; // increase the ReadPtr if (fifo.CPReadPointer >= fifo.CPEnd) { fifo.CPReadPointer = fifo.CPBase; // adjust, take care ptr = Memory_GetPtr(fifo.CPReadPointer); INFO_LOG(COMMANDPROCESSOR, "BUFFER LOOP"); } } } } // __________________________________________________________________________________________________ // !!! Temporary (I hope): re-used in DC mode // UpdateFifoRegister // It's no problem if the gfx falls behind a little bit. Better make sure to stop the cpu thread // when the distance is way huge, though. // So: // CPU thread /// 0. Write data (done before entering this) // 1. Compute distance // 2. If distance > threshold, sleep and goto 1 // GPU thread // 1. Compute distance // 2. If distance < threshold, sleep and goto 1 (or wait for trigger?) // 3. Read and use a bit of data, goto 1 void UpdateFifoRegister() { // update the distance int wp = fifo.CPWritePointer; int rp = fifo.CPReadPointer; int dist; if (wp >= rp) dist = wp - rp; else dist = (wp - fifo.CPBase) + (fifo.CPEnd - rp); Common::AtomicStore(fifo.CPReadWriteDistance, dist); if (!g_VideoInitialize.bOnThread) CatchUpGPU(); } void UpdateInterrupts() { if (m_CPCtrlReg.CPIntEnable && (fifo.bFF_BPEnable && fifo.bFF_Breakpoint)) { g_VideoInitialize.pSetInterrupt(INT_CAUSE_CP, true); } else { g_VideoInitialize.pSetInterrupt(INT_CAUSE_CP, false); } } void UpdateInterruptsFromVideoPlugin() { if (fifo.bFF_Breakpoint) // implicit since only BP trigger (see fifo.cpp) can call this m_CPStatusReg.Breakpoint = 1; g_VideoInitialize.pScheduleEvent_Threadsafe(0, et_UpdateInterrupts, 0); } void SetFifoIdleFromVideoPlugin() { s_fifoIdleEvent.Set(); } } // end of namespace CommandProcessor