//http://nesdev.parodius.com/bbs/viewtopic.php?p=4571&sid=db4c7e35316cc5d734606dd02f11dccb //todo - read http://wiki.nesdev.com/w/index.php/PPU_sprite_priority //TODO - correctly emulate PPU OFF state using BizHawk.Common; namespace BizHawk.Emulation.Cores.Nintendo.NES { sealed partial class PPU { const int kFetchTime = 2; struct BGDataRecord { public byte nt, at; public byte pt_0, pt_1; }; public short[] xbuf = new short[256*240]; int ppu_addr_temp; void Read_bgdata(ref BGDataRecord bgdata) { for (int i = 0; i < 8; i++) Read_bgdata(i,ref bgdata); } void Read_bgdata(int cycle, ref BGDataRecord bgdata) { switch (cycle) { case 0: ppu_addr_temp = ppur.get_ntread(); bgdata.nt = ppubus_read(ppu_addr_temp, true); runppu(1); break; case 1: runppu(1); break; case 2: { ppu_addr_temp = ppur.get_atread(); byte at = ppubus_read(ppu_addr_temp, true); //modify at to get appropriate palette shift if ((ppur.vt & 2) != 0) at >>= 4; if ((ppur.ht & 2) != 0) at >>= 2; at &= 0x03; at <<= 2; bgdata.at = at; //horizontal scroll clocked at cycle 3 and then //vertical scroll at 251 runppu(1); if (reg_2001.PPUON) { ppur.increment_hsc(); if (ppur.status.cycle == 251) ppur.increment_vs(); } break; } case 3: runppu(1); break; case 4: ppu_addr_temp = ppur.get_ptread(bgdata.nt); bgdata.pt_0 = ppubus_read(ppu_addr_temp, true); runppu(1); break; case 5: runppu(1); break; case 6: ppu_addr_temp |= 8; bgdata.pt_1 = ppubus_read(ppu_addr_temp, true); runppu(1); break; case 7: runppu(1); break; } //switch(cycle) } unsafe struct TempOAM { public fixed byte oam[4]; public fixed byte patterns[2]; public byte index; public byte present; } //TODO - check flashing sirens in werewolf short PaletteAdjustPixel(int pixel) { //tack on the deemph bits. THESE MAY BE ORDERED WRONG. PLEASE CHECK IN THE PALETTE CODE return (short)(pixel | reg_2001.intensity_lsl_6); } const int kLineTime = 341; public unsafe void FrameAdvance() { BGDataRecord *bgdata = stackalloc BGDataRecord[34]; //one at the end is junk, it can never be rendered //262 scanlines if (ppudead != 0) { FrameAdvance_ppudead(); return; } Reg2002_vblank_active_pending = true; ppuphase = PPUPHASE.VBL; ppur.status.sl = 241; //Not sure if this is correct. According to Matt Conte and my own tests, it is. Timing is probably off, though. //NOTE: Not having this here breaks a Super Donkey Kong game. reg_2003 = 0; //this was repeatedly finetuned from the fceux days thrugh the old cpu core and into the new one to pass 05-nmi_timing.nes //note that there is still some leniency. for instance, 4,2 will pass in addition to 3,3 const int delay = 6; runppu(3); bool nmi_destiny = reg_2000.vblank_nmi_gen && Reg2002_vblank_active; runppu(3); if (nmi_destiny) TriggerNMI(); nes.Board.AtVsyncNMI(); runppu(postNMIlines * kLineTime - delay); //this seems to run just before the dummy scanline begins clear_2002(); TempOAM* oams = stackalloc TempOAM[128]; int* oamcounts = stackalloc int[2]; int oamslot=0; int oamcount=0; idleSynch ^= true; //render 241 scanlines (including 1 dummy at beginning) for (int sl = 0; sl < 241; sl++) { ppur.status.cycle = 0; ppur.status.sl = sl; int yp = sl - 1; ppuphase = PPUPHASE.BG; if (NTViewCallback != null && yp == NTViewCallback.Scanline) NTViewCallback.Callback(); if (PPUViewCallback != null && yp == PPUViewCallback.Scanline) PPUViewCallback.Callback(); //twiddle the oam buffers int scanslot = oamslot ^ 1; int renderslot = oamslot; oamslot ^= 1; oamcount = oamcounts[renderslot]; //ok, we're also going to draw here. //unless we're on the first dummy scanline if (sl != 0) { //the main scanline rendering loop: //32 times, we will fetch a tile and then render 8 pixels. //two of those tiles were read in the last scanline. int yp_shift = yp << 8; for (int xt = 0; xt < 32; xt++) { int xstart = xt << 3; oamcount = oamcounts[renderslot]; int target = yp_shift + xstart; int rasterpos = xstart; //check all the conditions that can cause things to render in these 8px bool renderspritenow = reg_2001.show_obj && (xt > 0 || reg_2001.show_obj_leftmost); bool renderbgnow = reg_2001.show_bg && (xt > 0 || reg_2001.show_bg_leftmost); for (int xp = 0; xp < 8; xp++, rasterpos++) { //process the current clock's worth of bg data fetching //this needs to be split into 8 pieces or else exact sprite 0 hitting wont work due to the cpu not running while the sprite renders below Read_bgdata(xp, ref bgdata[xt + 2]); //bg pos is different from raster pos due to its offsetability. //so adjust for that here int bgpos = rasterpos + ppur.fh; int bgpx = bgpos & 7; int bgtile = bgpos >> 3; int pixel = 0, pixelcolor; //generate the BG data if (renderbgnow) { byte pt_0 = bgdata[bgtile].pt_0; byte pt_1 = bgdata[bgtile].pt_1; int sel = 7 - bgpx; pixel = ((pt_0 >> sel) & 1) | (((pt_1 >> sel) & 1) << 1); if (pixel != 0) pixel |= bgdata[bgtile].at; pixelcolor = PALRAM[pixel]; } else { if (!renderspritenow) { //according to qeed's doc, use palette 0 or $2006's value if it is & 0x3Fxx //EDIT - this requires corect emulation of PPU OFF state, and seems only to apply when the PPU is OFF // not sure why this was off, but having it on fixes full_nes_palette, and it's a behavior that's been // verified on the decapped PPU // if there's anything wrong with how we're doing this, someone please chime in int addr = ppur.get_2007access(); if ((addr & 0x3F00) == 0x3F00) { // System.Console.WriteLine("{0:X4}", addr); pixel = addr & 0x1F; } } pixelcolor = PALRAM[pixel]; pixelcolor |= 0x8000; //whats this? i think its a flag to indicate a hidden background to be used by the canvas filling logic later } if (!nes.Settings.DispBackground) pixelcolor = 0x8000; //whats this? i think its a flag to indicate a hidden background to be used by the canvas filling logic later //look for a sprite to be drawn bool havepixel = false; int renderslot_shift = renderslot << 6; for (int s = 0; s < oamcount; s++) { TempOAM* oam = &oams[renderslot_shift + s]; int x = oam->oam[3]; if (rasterpos >= x && rasterpos < x + 8) { //build the pixel. //fetch the LSB of the patterns int spixel = oam->patterns[0] & 1; spixel |= (oam->patterns[1] & 1) << 1; //shift down the patterns so the next pixel is in the LSB oam->patterns[0] >>= 1; oam->patterns[1] >>= 1; //bail out if we already have a pixel from a higher priority sprite. //notice that we continue looping anyway, so that we can shift down the patterns //transparent pixel bailout if (!renderspritenow || havepixel || spixel == 0) continue; havepixel = true; //TODO - make sure we dont trigger spritehit if the edges are masked for either BG or OBJ //spritehit: //1. is it sprite#0? //2. is the bg pixel nonzero? //then, it is spritehit. Reg2002_objhit |= (oam->index == 0 && pixel != 0 && rasterpos < 255); //priority handling, if in front of BG: bool drawsprite = !(((oam->oam[2] & 0x20) != 0) && ((pixel & 3) != 0)); if (drawsprite && nes.Settings.DispSprites) { //bring in the palette bits and palettize spixel |= (oam->oam[2] & 3) << 2; //save it for use in the framebuffer pixelcolor = PALRAM[0x10 + spixel]; } } //rasterpos in sprite range } //oamcount loop if (reg_2001.color_disable) pixelcolor &= 0x30; xbuf[target] = PaletteAdjustPixel(pixelcolor); target++; } //loop across 8 pixels } //loop across 32 tiles } else for (int xt = 0; xt < 32; xt++) Read_bgdata(ref bgdata[xt + 2]); //look for sprites (was supposed to run concurrent with bg rendering) oamcounts[scanslot] = 0; oamcount = 0; int spriteHeight = reg_2000.obj_size_16 ? 16 : 8; int scanslot_lshift = scanslot << 6; for (int i = 0; i < 64; i++) { oams[scanslot_lshift + i].present = 0; int spr = i * 4; if (yp >= OAM[spr] && yp < OAM[spr] + spriteHeight) { //if we already have maxsprites, then this new one causes an overflow, //set the flag and bail out. //should we set this flag anyway?? if (oamcount >= 8 && reg_2001.PPUON) { Reg2002_objoverflow = true; if(!nes.Settings.AllowMoreThanEightSprites) break; } //just copy some bytes into the internal sprite buffer TempOAM* oam = &oams[scanslot_lshift + oamcount]; for (int j = 0; j < 4; j++) oam->oam[j] = OAM[spr + j]; oam->present = 1; //note that we stuff the oam index into [6]. //i need to turn this into a struct so we can have fewer magic numbers oams[scanslot_lshift + oamcount].index = (byte)i; oamcount++; } } oamcounts[scanslot] = oamcount; ppuphase = PPUPHASE.OBJ; //fetch sprite patterns int oam_todo = oamcount; if (oam_todo < 8) oam_todo = 8; for (int s = 0; s < oam_todo; s++) { //if this is a real sprite sprite, then it is not above the 8 sprite limit. //this is how we support the no 8 sprite limit feature. //not that at some point we may need a virtual CALL_PPUREAD which just peeks and doesnt increment any counters //this could be handy for the debugging tools also bool realSprite = (s < 8); bool junksprite = (s >= oamcount || !reg_2001.PPUON); TempOAM* oam = &oams[scanslot_lshift + s]; int line = yp - oam->oam[0]; if ((oam->oam[2] & 0x80) != 0) //vflip line = spriteHeight - line - 1; int patternNumber = oam->oam[1]; int patternAddress; //create deterministic dummy fetch pattern. if (oam->present == 0) { //according to nintendulator: //* On the first empty sprite slot, read the Y-coordinate of sprite #63 followed by $FF for the remaining 7 cycles //* On all subsequent empty sprite slots, read $FF for all 8 reads //well, we shall just read $FF and that is good enough for now to make mmc3 work patternNumber = 0xFF; line = 0; } //8x16 sprite handling: if (reg_2000.obj_size_16) { int bank = (patternNumber & 1) << 12; patternNumber = patternNumber & ~1; patternNumber |= (line >> 3) & 1; patternAddress = (patternNumber << 4) | bank; } else patternAddress = (patternNumber << 4) | (reg_2000.obj_pattern_hi << 12); //offset into the pattern for the current line. //tricky: tall sprites have already had lines>8 taken care of by getting a new pattern number above. //so we just need the line offset for the second pattern patternAddress += line & 7; //garbage nametable fetches + scroll resets int garbage_todo = 2; ppubus_read(ppur.get_ntread(), true); if (reg_2001.PPUON) { if (sl == 0 && ppur.status.cycle == 304) { runppu(1); if (reg_2001.PPUON) ppur.install_latches(); runppu(1); garbage_todo = 0; } if ((sl != 0) && ppur.status.cycle == 256) { runppu(1); //at 257: 3d world runner is ugly if we do this at 256 if (reg_2001.PPUON) ppur.install_h_latches(); runppu(1); garbage_todo = 0; } } if (realSprite) runppu(garbage_todo); ppubus_read(ppur.get_atread(), true); //at or nt? if (realSprite) runppu(kFetchTime); // TODO - fake sprites should not come through ppubus_read but rather peek it // (at least, they should not probe it with AddressPPU. maybe the difference between peek and read is not necessary) if (junksprite) { if (realSprite) { ppubus_read(patternAddress, true); ppubus_read(patternAddress, true); runppu(kFetchTime * 2); } } else { int addr = patternAddress; oam->patterns[0] = ppubus_read(addr, true); if (realSprite) runppu(kFetchTime); addr += 8; oam->patterns[1] = ppubus_read(addr, true); if (realSprite) runppu(kFetchTime); // hflip if ((oam->oam[2] & 0x40) == 0) { oam->patterns[0] = BitReverse.Byte8[oam->patterns[0]]; oam->patterns[1] = BitReverse.Byte8[oam->patterns[1]]; } } } // sprite pattern fetch loop ppuphase = PPUPHASE.BG; // fetch BG: two tiles for next line for (int xt = 0; xt < 2; xt++) Read_bgdata(ref bgdata[xt]); // this sequence is tuned to pass 10-even_odd_timing.nes runppu(kFetchTime); bool evenOddDestiny = reg_2001.show_bg; runppu(kFetchTime); // After memory access 170, the PPU simply rests for 4 cycles (or the // equivelant of half a memory access cycle) before repeating the whole // pixel/scanline rendering process. If the scanline being rendered is the very // first one on every second frame, then this delay simply doesn't exist. if (sl == 0 && idleSynch && evenOddDestiny && chopdot) { } else runppu(1); } // scanline loop ppur.status.sl = 241; //idle for pre NMI lines runppu(preNMIlines * kLineTime); } //FrameAdvance void FrameAdvance_ppudead() { //not quite emulating all the NES power up behavior //since it is known that the NES ignores writes to some //register before around a full frame, but no games //should write to those regs during that time, it needs //to wait for vblank ppur.status.sl = 241; runppu(postNMIlines * kLineTime); ppur.status.sl = 0; runppu(241 * kLineTime); runppu(preNMIlines * kLineTime); --ppudead; } } }