The TIME_LOOP_HACKS shortcut skips the iterations of a
DEC rr / LD A,r / OR r / JR NZ (or JP NZ) delay loop by burning their
cycles in one go. Each skipped iteration runs four opcodes (R is
advanced by 4), but only LD, OR and the jump were counted: the DEC was
left out, so every iteration took 6 cycles too few (20 instead of 26
with JR) and such delays ran about 23% short.
Count the DEC as well, for the BC, DE and HL loops with JR and JP.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
The Mega Drive VDP was part of the System C/C2 video code and System 18 used
it through file-static state and a few flags, so a change for one board could
reach the others. It is now a module, vidhrdw/segavdp.c:
- all its state is in one struct, cleared at every start;
- each user starts it with its own configuration (struct segavdp_interface):
where the CRAM writes go (paletteram16, or the 64 System 18 colors set
directly at 0x1800), the initial palette bases, the scanline base of the
partial updates, the initial display enable, whether register 12 resizes
the screen, and the interrupt callback for writes to registers 0/1;
- the boards use accessors (segavdp_get_reg, segavdp_set_palbank, ...);
- vidhrdw/segac2_vidhrdw.c keeps only the C/C2, Mega-Tech, Mega Play and
bootleg video glue, vidhrdw/segas18.c the System 18 one.
Fixed on the way (state left over from a previous game in the same process):
the display enable and the Mega Play interrupt-acknowledge/HBLANK mode
(now also reset by the C/C2 machine init). The window X
position/width and the VBLANK flag are now in the save states. The dead
ddcrew/aquario VDP kludges of the old System 18 video are removed.
Save states loaded in a new instance (netplay, RunAhead with a second
instance) now continue exactly as the instance that saved them:
- Mega Drive based boards (drivers/segac2.c): the Z80 RAM and bus state
(running flag, bank latch) and the Mega-Tech / Mega Play BIOS registers
and RAM were not saved (mp_gunhe and mp_mazin went out of sync);
- sound/sn76496.c: the PSG had no save state at all;
- sound/fm.c: the YM2612 postload read the DAC enable from register 0x2d
instead of 0x2b (the DAC channel was lost or computed as FM after a load).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
Cotton (System 16B): the FD1094 sets cotton, cottonu, cottonj and cottonja
now run, with the decryption of machine/fd1094.c (from Moonwalker), and the
missing bootlegs cottonud, cottonjd and cottonjad are added. The dummy
cotton/cottona of system16.c are removed.
System 18: the games run on the 315-5195 mapper layout each of them sets
(read from MAME), the segaic16 video and the VDP layer, as in MAME's
segas18.cpp:
- Alien Storm (astorm, astorm3, astormu, astormj and bootlegs), Bloxeed
(bloxeed, bloxeeda, bloxeedd), Clutch Hitter (cltchitr, cltchitrj and
bootlegs), D. D. Crew (the six FD1094 sets and their bootlegs, 4 players
through the 8255), Desert Breaker (desertbr, desertbrj and bootlegs),
Laser Ghost (lghost, lghostu, lghostj and bootlegs, light guns through
the MSM6253, MAME's gun offset correction), Wally wo Sagase! (wwallyj,
wwallyja, wwallyja3p and bootlegs, uPD4701 trackballs), Shadow Dancer
(shdancer, shdancerj, shdancer1, were shdancer, shdancrj, shdancrb).
- 171-5987 ROM board: ROM 1 at 200000, tile and sprite banking, 1MB and
2MB program ROMs; fd1094.c decrypts several address windows.
- The VDP layer shows one frame after the other layers, as in MAME, and the
VDP sees the VBLANK (VIDEO_EOF of segac2) for its partial updates.
- astormbl, shdancbl, moonwlkb, aquario, hamaway and mwalk are unchanged.
Save states (RunAhead, netplay):
- System 16B: uPD7759, sound ROM bank, segaic16 tilemap and sprite banks,
flip, scroll latches, display enable.
- System 18: RF5C68 (registers and wave RAM), YM2612/YM3438 (envelope and
LFO counters, key, frequencies, algorithm memory, FM timer overflow time;
the postload also rewrites the b6/1b6 registers), sound ROM bank, VDP
layer, mapper and I/O registers.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
- AY-8910 of the Triple Punch board (Triple Punch, Knock Out!!, Mariner,
800 Fathoms) at 18432000/12 = 1.536 MHz, as current MAME ("verified on
hardware"): the sound was 16.7% too high and too fast. Mariner keeps
GAME_IMPERFECT_SOUND: its discrete sound circuits are not emulated, in
current MAME either; 800 Fathoms gets the flag, as in MAME
- new clones from current MAME: triplepa, knockoutb, knockoutc, 800fatha
(names and manufacturers of the sets as in MAME)
- save states of the Galaxian/Scramble/Frogger hardware (rollback, RunAhead,
netplay diverged):
- Galaxian video: star scroll/blink state, star timer start, flip, graphics
banks, background and bullet colors, rockclim scroll; a postload restores
the tilemap flip, column scroll and tiles
- 7474 flip-flops (machine/7474.c), registered by TTL7474_config
- Scramble sound: the sound CPU RAM (scobra map) in the saved memory, the
AY port B timer computed from the sound CPU cycles as current MAME
(no state, no 32-bit overflow; same values)
- RC filters of the sound streams (streams.c) and the global sound enable
(mixer.c)
- Galaxian sound statics (LFO, noise, tone generator)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
The driver of current MAME (Phil Bennett) in C89: the two 8086 CPUs and the
arithmetic unit, the road, object and character layers of the three screens,
the engine/noise sound with the AY-8910 / YM2149s, save states.
- 6 sets: tx1, tx1jb, tx1jc, buggyboy, buggyboyb, buggyboyjr
- the cockpit sets show their three screens side by side (768x240, 12:3), or
only the middle one with the new core option "TX-1 / Buggy Boy Screens"
(offered for those sets only)
- i86 core: the TEST input (WAIT instruction) and an opt-in 16-bit data bus
(an aligned word is one access for the handlers); both off by default, the
other i86/i186 games run as before
- the layer select and the TX-1 road pixel logic are lookup tables built
from MAME's logic: same pixels, TX-1 about 40% faster
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
None of the YMF278B (OPL4 PCM), YMF262 (OPL3) or Fuuki FG-3 state was in
the save states: after a load (RunAhead, netplay rollback, rewind) the
sound kept the voices of the frames run before the load, the YMF278B
timers and IRQ their old values, the sprites buffered by two frames and
the RAM shared by the two CPUs of the FG-3 too.
- YMF278B: every slot, the chip registers, timers and IRQ state.
- YMF262: every slot and channel, the chip registers; the slot outputs
(pointers to the channel outputs) are saved as indexes.
- Fuuki FG-3: the two buffered sprite RAMs and tile banks, the shared RAM.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
- i386 CPU core: the C core of MAME 0.139 (mame2010-libretro), on the
0.78 CPU interface, with save states. Fast memory: the driver gives
its plain RAM and ROM, which the CPU fetches, reads and writes
directly (a cached region per kind of access) instead of through
the memory handlers.
- Seibu SPI driver from current MAME: the SXX2C main board with its
cartridges, the SXX2E / SXX2F / SXX2G single boards and SYS386I.
Senkyu / Battle Balls, Viper Phase 1, Raiden Fighters, Raiden
Fighters 2 and Raiden Fighters Jet, 57 sets.
- Video, the CRTC, the tile and sprite decryption (SEI252, RISE10,
RISE11) as current MAME.
- DS2404 RTC from current MAME. Its 256 Hz tick keeps the time it had
left across a save state.
- intelfsh: 8-bit chips, the E28F008SA of the SPI main board.
- YMF271: the clock in the interface (16.384MHz on SXX2G).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
The bootleg has no rotary joystick: the game reads the direction the
player faces, and fires to, in the upper nibble of the joystick port, and
snk_rot8_ikarijpb() sets it from the joystick, so the player fires where
he walks. A new 'Aim with Right Stick' setting (a DIP switch that is not on
the board, Off by default) makes a pushed right stick set it instead: the
player walks with the left stick and aims with the right stick, in 8
directions. With the right stick released, the direction follows the
joystick again.
The right stick inputs (P1 Aim, P2 Aim) default to the right analog stick
of each player, and to I/J/K/L for player 1. They are read only by the
driver, not by the game. With the setting Off the game runs as before.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
The stream channels took the interface volumes in the wrong order:
channel 0 (the chip's left output) got the high 16 bits of
YM3012_VOL, the right one. Regression of #2100: before it the chip
output the same mix on both channels, so the order did not show.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
Sound: the sound program of Asura Buster and Asura Blade writes the key
on of a PCM slot again about once a second while the sound plays, with
the key already on. The 0.78 YMF278B restarted the sample on every key
on write, so voices and sound effects were repeated. As in current MAME,
the key on bit is now tracked per slot and a sample is only started on
a key off -> key on change. The YMF262 (the FM part of the OPL4) runs at
33.8688MHz / 2.375 = 14260547 Hz, as in MAME 0.143u3.
Sets as in current MAME:
- asurabus: USA set (new, parent)
- asurabusj: Japan set 1 (the set that was `asurabus`)
- asurabusja: Japan set 2 (new)
- asurabusjr: Japan, ARCADIA review build (new); P1 button 4 pauses
the game, so its inputs have a 4th button per player
DIP switches as in current MAME, for Asura Blade too:
- "Demo Sounds & Music" (SW1:3,4), was two "Unknown" switches
- Difficulty, Damage and Max Rounds labels and defaults fixed: the old
Max Rounds default ($0040) is shown as "Error" in the service mode
- Coin A/B: 1C_2C to 1C_5C and "2C Start / 1C Continue" added
- DIP locations as comments (the core has no PORT_DIPLOCATION)
Metadata: XML entries from the core's generator, hiscore and catver
entries for the new sets, CHANGELOG.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01N3fzdsnzkQVtgYtYzqepjX
New: Intel/Sharp 16-bit flash memories (intelfsh.c), the RF5C296 PC card
controller with the ATA flash card and the three Taito locked cards
(ataflash.c), the G-NET board in zn.c (flash bank, control registers,
COM_DELAY, ZN-2 SPU toggle, the MCU protocol for analog and trackball
inputs, mahjong panel, Mawasunda handles), the Zoom board in flash mode
(samples read from the flashes), CHD metadata (CIS, IDNT, KEY), and a
PC hook in the PSX CPU for the IRQ hack of Space Invaders Anniversary
and Kollon. JP1 "Auto" for the mod BIOS and Compact Flash games.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
NBA Jam Extreme (ver. 1.10I and 1.04) as in current MAME, with the RAX
sound board: an ADSP-2181 (new in the ADSP-2100 core: its I/O space,
memory overlays, interrupts and instructions), byte DMA from the sound
ROM and two DACs fed by SPORT0 autobuffering.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
- ADSP recompiler: at exit, adsp_jit_exit() freed the translations
through adsp_jit_flush(), which then scans the ADSP program memory
again through OP_ROM. There is no active CPU at that point: closing
Killer Instinct or Killer Instinct 2 with the ADSP recompiler on
crashed. The exit now only frees the translations.
- MIPS3: its registers were not in the save state, so after a load the
CPU went on with the registers of the moment of the load and crashed.
They are saved now, and the Compare interrupt timer is set again
after a load.
- Killer Instinct: the control registers and the end of the IRQ 0 pulse
are saved too; the displayed buffer is set again after a load.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
- the Atlus BIOS set coh1001l and hvnsgate, with the names and files of current MAME
- 4 banks of 8MB of game ROM, selected by the byte at 0x1fb00002
- sound: a 68000 at 10MHz with a YMZ280B; 16-bit command latch at 0x1fb00000
holding 68000 IRQ 3, YMZ280B on IRQ 2; and the PSX SPU
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
- the 45 sets and 2 BIOS sets of current MAME, with its names and files
- the CAT702 keys are read from the dumps in the sets (CP01/CP10, CPnn)
- inputs of current MAME (capcom6b, capcom4b, glpracr, plsmaswd, strider2, tgmj)
- game ROM bank: 16 banks, as in MAME
- QSound Z80: 250Hz interrupt, as in MAME; its phase and the Z80 ROM bank
are in the save state
- GAME_IMPERFECT_GRAPHICS and GAME_NO_SOUND (glpracr) removed
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
osd_get_joy_list() rebuilt the whole list (4 players x all codes) at every
call, and the input code calls it at every key check: about 100 times a
frame, some 750k instructions a frame whatever the game. It is now built
again only when a port's layout changes.
doapp, 600 frames (callgrind): retro_run 15.02 G -> 14.57 G instructions
(-3%), video and audio identical frame by frame. The saving is per frame,
so it is a larger share on lighter games.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
The MIPS Dynarec option was offered and turned on only for Namco System
11/12; zn.c already calls psxcpu_set_jit() when the option is on, but the
option stayed off, so the ZN games always ran the interpreter.
Video and audio identical frame by frame, recompiler vs interpreter, 4000
frames: the 17 Tecmo TPS sets and the 9 Taito FX-1A sets (branch
taito-fx1a).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
The first displayed line goes at disstart - 16 and the lines out of the
range are black. Super Football Champ sets 24..248 (224 lines): the picture
was 8 lines too high with 16 lines of off-screen VRAM below it. With the
usual 16..256 range nothing changes: doapp, glpracr2, tekken3, psyforce
identical frame by frame over 2500 frames.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
YM2610_save_state leaves out the global envelope and LFO counters, the key
state, the SSG-EG negation, the envelope output and the algorithm memory
of the FM part, and the counters, envelope and noise generator of the SSG
(the AY8910 has no save state, and the YM postload rewrites its registers,
which restarts the envelope). With YM2610_timer_state_register() these are
now kept in a copy and put back after the chip postload. Other YM2610
drivers: unchanged.
Save state, 600 frames after same-session and cross-session load:
- sfchamp: video and audio identical
- psyforce: audio identical (was: all 600 frames differ)
- mgcldtex, mgcldate: video identical; audio differs only in the first
26-27 frames, by at most 103/32768, decaying to 0 (was: 470 frames)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
The core does not save its timers, so after a load the two FM timers of
the YM2610 kept the state of the running session and the sound CPU drifted
(sfchamp: audio differed 2 s after a reload). A driver can now ask for the
timers to be saved with YM2610_timer_state_register() in its init; the
other YM2610 drivers keep their save state layout.
sfchamp: video and audio identical over 600 frames after same-session and
cross-session load.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
The Taito FX-1A games on the ZN-1 board code from #2081, with the sets
and file names of current MAME. Replaces the old non-working sets.
- zn.c: FX-1A board (1 MB VRAM, 8 MB game ROM banks at 0x1f000000, bank
register at 0x1fb40000), CAT702 keys read from the TT01 / TTnn dumps in
the sets, Z80 with a YM2610B and the TC0140SYT at 0x1fb80000, plus the
PSX SPU.
- taitosnd.c: optional reset and save state registration for the
TC0140SYT (used by the FX-1A only; other drivers unchanged).
- Sets: coh1000t (BIOS), sfchamp/o/u/j, psyforce/j/x, mgcldtex, mgcldate
(MAME names: the old mgcldate is now mgcldtex, the old mgcldtea is now
mgcldate). Old sound player and FX-1A machine removed.
- FX-1B sets (raystorm, ftimpcta, gdarius, gdarius2) unchanged for now.
- metadata/mame2003-plus.xml: FX-1A entries regenerated.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
CarnEvil (#1659): MIPS III recompiler and native FPU on the Seattle games
that were checked (CarnEvil, SF Rush, The Rock: identical frames to the
interpreter), Voodoo blitters for their common modes, IDE and Galileo DMA
straight into RAM, VBLANK latch, Galileo timers and IDE seek timing as
current MAME, 54.123711 Hz. Same frame cadence as current MAME.
Seattle driver toward current MAME: machine configurations and refresh
rates, I/O ASIC values, Voodoo clipping, LFB formats, dithering, alpha
blending, depth bias and alpha test, wheel board and latched shifter,
SMC91C94 ethernet on the Flagstaff board (SF Rush's race waited for it),
49-way joysticks on Blitz '99/2000 (I40) and Blitz, ROM names, disk SHA1s,
descriptions and clones of current MAME (mace 1.0ce, macea, wg3dhfe,
blitz99a), San Francisco Rush: The Rock.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016PwPGTj4SDjmvPp3DRm9j5
A store into a 64-byte granule holding translated code flushed the whole
translation cache. 1 on 1 Government keeps a variable next to code: 285
flushes and about 300000 recompiled blocks in 3000 frames, so the
recompiler was slower than the interpreter (220 vs 300 fps). The code map
now has one entry per 4-byte word (4 MB table, 16 MB of host addresses):
only a store over an instruction that was actually translated flushes.
A flush is still never missed, an aliased entry only flushes more.
1on1gov with the recompiler: 220 -> 360 fps; frames identical to the
interpreter on all the Tecmo TPS sets.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
The Model 2 games spend 80-86% of their i960 instructions in short wait
loops: read a RAM word or the video control register, maybe count, branch
back. Nothing else runs during the i960 timeslice, so when two iterations
in a row only load, compute and branch, read the same side-effect-free
addresses, cost the same cycles and change each register by the same
amount (none of them compared), the remaining iterations are skipped: their
cycles taken, the counters advanced. The last iteration of the timeslice
still runs, so the CPU stops exactly where the interpreter would.
The driver lists what an idle loop may read (i960_set_idle_read_callback):
ROM, RAM and the video control register. Off for other drivers.
Frame by frame identical video and audio on Daytona USA, The House of the
Dead, Virtua Cop and Virtua Cop 2; i960 time down by about 70%.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
- the mipmap levels and the microtexture of a polygon are worked out once
per polygon, not at every texel fetch
- texel addresses with unsigned shifts (the signed divisions by 2 were the
hottest line), the column and row terms shared by the four texels
- four transparent texels return at once; the microtexture blend is
skipped for a translucent texel with no alpha (it is discarded anyway)
Output identical frame by frame on Daytona USA, The House of the Dead,
Virtua Cop and Virtua Cop 2.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
The 3D frame was cleared and composited over the whole screen every frame,
and the 64K z buckets cleared at every new frame. Now the buffer is its own
fill map (a drawn pixel is never 0), only the box of the pixels drawn since
the last clear is cleared and composited, and only the z buckets in use are
cleared. Output identical frame by frame.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MEgow1uRCv8pxMErKhmoHZ
Claude