byuu says:
This will hopefully be a short-lived WIP, I just want to save
a breakpoint before I attempt something else.
NES, GB, GBC and GBA all load via const stream& now.
NES CPU core moved to Processor::RP2A03.
byuu says:
static vector<uint8_t> file::read(const string &filename); replaces:
static bool file::read(const string &filename, uint8_t *&data, unsigned
&size); This allows automatic deletion of the underlying data.
Added vectorstream, which is obviously a vector<uint8_t> wrapper for
a data stream. Plan is for all data accesses inside my emulation cores
to take stream objects, especially MSU1. This lets you feed the core
anything: memorystream, filestream, zipstream, gzipstream, httpstream,
etc. There will still be exceptions for link and serial, those need
actual library files on disk. But those aren't official hardware devices
anyway.
So to help with speed a bit, I'm rethinking the video rendering path.
Previous system:
- core outputs system-native samples (SNES = 19-bit LRGB, NES = 9-bit
emphasis+palette, DMG = 2-bit grayscale, etc.)
- interfaceSystem transforms samples to 30-bit via lookup table inside
the emulation core
- interfaceSystem masks off overscan areas, if enabled
- interfaceUI runs filter to produce new target buffer, if enabled
- interfaceUI transforms 30-bit video to native display depth (24-bit or
30-bit), and applies color-adjustments (gamma, etc) at the same time
New system:
- all cores now generate an internal palette, and call
Interface::videoColor(uint32_t source, uint16_t red, uint16_t green,
uint16_t blue) to get native display color post-adjusted (gamma, etc
applied already.)
- all cores output to uint32_t* buffer now (output video.palette[color]
instead of just color)
- interfaceUI runs filter to produce new target buffer, if enabled
- interfaceUI memcpy()'s buffer to the video card
videoColor() is pretty neat. source is the raw pixel (as per the
old-format, 19-bit SNES, 9-bit NES, etc), and you can create a color
from that if you really want to. Or return that value to get a buffer
just like v088 and below. red, green, blue are 16-bits per channel,
because why the hell not, right? Just lop off all the bits you don't
want. If you have more bits on your display than that, fuck you :P
The last step is extremely difficult to avoid. Video cards can and do
have pitches that differ from the width of the texture. Trying to make
the core account for this would be really awful. And even if we did
that, the emulation routine would need to write directly to a video card
RAM buffer. Some APIs require you to lock the video buffer while
writing, so this would leave the video buffer locked for a long time.
Probably not catastrophic, but still awful. And lastly, if the
emulation core tried writing directly to the display texture, software
filters would no longer be possible (unless you -really- jump through
hooks and divert to a memory buffer when a filter is enabled, but ...
fuck.)
Anyway, the point of all that work was to eliminate an extra video copy,
and the need for a really painful 30-bit to 24-bit conversion (three
shifts, three masks, three array indexes.) So this basically reverts us,
performance-wise, to where we were pre-30 bit support.
[...]
The downside to this is that we're going to need a filter for each
output depth. Since the array type is uint32_t*, and I don't intend to
support higher or lower depths, we really only need 24+30-bit versions
of each filter. Kinda shitty, but oh well.
byuu says:
Basically, the current implementation of nall/array is deprecated now.
The old method was for non-reference types, it acted like a vector for
POD types (raw memory allocation instead of placement new construction.)
And for reference types, it acted like an unordered set. Yeah, not good.
As of right now, nall/array is no longer used. The vector type usage was
replaced with actual vectors.
I've created nall/set, which now contains the specialization for
reference types.
nall/set basically acts much like std::unordered_set. No auto-sort, only
one of each type is allowed, automatic growth.
This will be the same both for reference and non-reference types ...
however, the non-reference type wasn't implemented just yet.
Future plans for nall/array are for it to be a statically allocated
block of memory, ala array<type, size>, which is meant for RAII memory
usage.
Have to work on the specifics, eg the size as a template parameter may
be problematic. I'd like to return allocated chunks of memory (eg
file::read) in this container so that I don't have to manually free the
data anymore.
I also removed nall/moduloarray, and moved that into the SNES DSP class,
since that's the only thing that uses it.
byuu says:
Changes to v088:
- OBJ mosaic Y fix
- Laevateinn compilation
- Remove filebrowser extra code
- Fix -march=native on Windows
- Fix purify mkdir
- GBA sound volume
- Add .gbb
- free firmware memory after file load
- Add GBA game to profile list (Yoshi's Island should work)
byuu says:
Changelog:
- DMA channel masks added (some are 27-bit source/target and some are
14-bit length -- hooray, varuint_t class.)
- No more state.pending flags. Instead, we set dma.pending flag when we
want a transfer (fixes GBA Video - Pokemon audio) [Cydrak]
- fixed OBJ Vmosaic [Cydrak, krom]
- OBJ cannot read <=0x13fff in BG modes 3-5 (fixes the garbled tile at
the top-left of some games)
- DMA timing should be much closer to hardware now, but probably not
perfect
- PPU frame blending uses blargg's bit-perfect, rounded method (slower,
but what can you do?)
- GBA carts really unload now
- added nall/gba/cartridge.hpp: used when there is no manifest. Scans
ROMs for library tags, and selects the first valid one found
- added EEPROM auto-detection when EEPROM size=0. Forces disk/save state
size to 8192 (otherwise states could crash between pre and post
detect.)
- detects first read after a set read address command when the size
is zero, and sets all subsequent bit-lengths to that value, prints
detected size to terminal
- added nall/nes/cartridge.hpp: moves iNES detection out of emulation
core.
Important to note: long-term goal is to remove all
nall/(system)/cartridge.hpp detections from the core and replace with
databases. All in good time.
Anyway, the GBA workarounds should work for ~98.5% of the library, if my
pre-scanning was correct (~40 games with odd tags. I reject ones without
numeric versions now, too.)
I think we're basically at a point where we can release a new version
now. Compatibility should be relatively high (at least for a first
release), and fixes are only going to affect one or two games at a time.
I'd like to start doing some major cleaning house internally (rename
NES->Famicom, SNES->SuperFamicom and such.) Would be much wiser to do
that on a .01 WIP to minimize regressions.
The main problems with a release now:
- speed is pretty bad, haven't really optimized much yet (not sure how
much we can improve it yet, this usually isn't easy)
- sound isn't -great-, but the GBA audio sucks anyway :P
- couple of known bugs (Sonic X video, etc.)
byuu says:
Changelog:
- revised NES XML tag nesting
- program.rom is going to refer to PRG+CHR combined. Split is going to
have to use different file names
- slot loader is gone (good riddance!)
- "Cartridge -> Load Game Boy Advance Cartridge ..." has become "Load ->
Game Boy Advance ..."
- Load Satellaview Slotted Cartridge is gone. If you load an SNES
cartridge and it sees <bsx><slot>, it asks if you want to load a BS-X
data pack
- If you load a Sufami Turbo cartridge with <cartridge linkable="true">,
it asks if you want to link in another Sufami Turbo cartridge
- if you try and load the same exact Sufami Turbo cartridge in both
slots, it yells at you for being an idiot :P
byuu says:
Be sure to run make install, and move required images to their appropriate system profile folders.
I still have no warnings in place if those images aren't present.
Changelog:
- OBJ mosaic should hopefully be emulated correctly now (thanks to krom
and Cydrak for testing the hardware behavior)
- emulated dummy serial registers, fixes Sonic Advance (you may still
need to specify 512KB FlashROM with an appropriate ID, I used
Panaonic's)
- GBA core exits scheduler (PPU thread) and calls
interface->videoRefresh() from main thread (not required, just nice)
- SRAM, FRAM, EEPROM and FlashROM initialized to 0xFF if it does not
exist (probably not needed, but FlashROM likes to reset to 0xFF
anyway)
- GBA manifest.xml for file-mode will now use "gamename.xml" instead of
"gamename.gba.xml"
- started renaming "NES" to "Famicom" and "SNES" to "Super Famicom" in
the GUI (may or may not change source code in the long-term)
- removed target-libsnes/
- added profile/
Profiles are the major new feature. So far we have:
Famicom.sys/{nothing (yet?)}
Super Famicom.sys/{ipl.rom}
Game Boy.sys/{boot.rom}
Game Boy Color.sys/{boot.rom}
Game Boy Advance.sys/{bios.rom[not included]}
Super Game Boy.sfc/{boot.rom,program.rom[not included]}
BS-X Satellaview.sfc/{program.rom,bsx.ram,bsx.pram}
Sufami Turbo.sfc/{program.rom}
The SGB, BSX and ST cartridges ask you to load GB, BS or ST cartridges
directly now. No slot loader for them. So the obvious downsides: you
can't quickly pick between different SGB BIOSes, but why would you want
to? Just use SGB2/JP. It's still possible, so I'll sacrifice a little
complexity for a rare case to make it a lot easier for the more common
case. ST cartridges currently won't let you load the secondary slot.
BS-X Town cart is the only useful game to load with nothing in the slot,
but only barely, since games are all seeded on flash and not on PSRAM
images. We can revisit a way to boot the BIOS directly if and when we
get the satellite uplink emulated and data can be downloaded onto the
PSRAM :P BS-X slotted cartridges still require the secondary slot.
My plan for BS-X slotted cartridges is to require a manifest.xml to
specify that it has the BS-X slot present. Otherwise, we have to load
the ROM into the SNES cartridge class, and parse its header before we
can find out if it has one. Screw that. If it's in the XML, I can tell
before loading the ROM if I need to present you with an optional slot
loading dialog. I will probably do something similar for Sufami Turbo.
Not all games even work with a secondary slot, so why ask you to load
a second slot for them? Let the XML request a second slot. A complete
Sufami Turbo ROM set will be trivial anyway. Not sure how I want to do
the sub dialog yet. We want basic file loading, but we don't want it to
look like the dialog 'didn't do anything' if it pops back open
immediately again. Maybe change the background color of the dialog to
a darker gray? Tacky, but it'd give you the visual cue without the need
for some subtle text changes.
byuu says:
Changelog:
- serialize processor.pc.data, not processor.pc
- call CPU processor.setMode() in ARM serialize
- serialize BIOS.mdr
- support SRAM > 32KB
- EEPROM, FlashROM serialize
- EEPROM lose nall/bitarray.hpp
- noise line feed after envelope
- space out PSR read
- ST018 needs byte reads fixed (don't align) [fixes HNMS2]
- flush sram/eeprom/flashrom to 0 on cartridge load
- APU/PPU dont sync back to CPU if syncing for state
- fixed APU sync problems in GB/GBC core that could possibly wreck save
states as well
Quite a lot of little problems there. I believe GBA save states are
fixed now.
byuu says:
Changelog:
- fixed FIFO[1] reset behavior (fixes audio in Sword of Mana)
- added FlashROM emulation (both sizes)
- GBA parses RAM settings from manifest.xml now
- save RAM is written to disk now
- added save state support (it's currently broken, though)
- fixed ROM/RAM access timings
- open bus should mostly work (we don't do the PC+12 stuff yet)
- emulated the undocumented memory control register (mirror IWRAM,
disable I+EWRAM, EWRAM wait state count)
- emulated keypad interrupts
- emulated STOP (freezes video, audio, DMA and timers; only breaks on
keypad IRQs)
- probably a lot more, it was a long night ...
Show stoppers, missing things, broken things, etc:
- ST018 is still completely broken
- GBC audio sequencer apparently needs work
- GBA audio FIFO buffer seems too quiet
- PHI / ROM prefetch needs to be emulated (no idea on how to do this,
especially PHI)
- SOUNDBIAS 64/128/256khz modes should output at that resolution
(really, we need to simulate PWM properly, no idea on how to do this)
- object mosaic top-left coordinates are wrong (minor, fixing will
actually make the effect look worse)
- need to emulate PPU greenswap and color palette distortion (no idea on
how do this)
- need GBA save type database (I would also LIKE to blacklist
/ patch-out trainers, but that's a discussion for another day.)
- some ARM ops advance the prefetch buffer, so you can read PC+12 in
some cases
byuu says:
(r24 was a point release during merging of changes.)
This release is almost entirely Cydrak's direct work:
- Added ARM::sequential() and some WAITCNT timings
- Added Bus::io(uint32 pc), intended for prefetch timing
- Added ARM::load() with data rotation (fixed Mother 3 graphics)
- Added ARM::store() for data mirroring
- LDM, STM, and instruction fetch still use read/write()
- ARM::vector() no longer unmasks FIQs
- Set THUMB shifter flags via bit(), consistent with ARM
- Replace shifter loops with conditional tests
My changes:
- fixed sprite clipping on left-edge of screen
- added first system folder, GBA.system
- sudo make install is now make install
- make install will create GBA.system for you in your home folder
Windows users, take data/GBA.system and put it in the same folder as
bsnes.exe, and give it a BIOS named bios.rom
Or place it in your home folder (%APPDATA%/bsnes)
Also note that this is highly experimental, I'll probably be changing
things a lot before release.
EDIT: I botched the cartridge timing change. Will fix in r26. It'll
still run a bit too fast for now, unfortunately.
byuu says:
Changelog:
- fixed cascading timers and readouts (speed hit from 320fps to 240fps;
would be 155fps with r20 timers) (fixes Spyro)
- OBJ mode 3 acts like OBJ mode 2 now (may not be correct, but nobody
has info on it)
- added background + object vertical+horizontal mosaic in all modes
(linear+affine+bitmap)
- object mosaic uses sprite (0,0) for start coordinates, not screen
(0,0) (again, nobody seems to have info on it)
- BIOS cannot be read by r(15)>=0x02000000; returns last BIOS read
instead (I can't believe games rely on this to work ... fixes SMA
Mario Bros.)
Mosaic is what concerns me the most, I've no idea if I'm doing it
correctly. But anything is probably better than nothing, so there's
that. I don't really notice the effect in Metroid Fusion. So either it's
broken, or it's really subtle.
byuu says:
Changelog:
- fixed below pixel green channel on color blending
- added semi-transparent objects [Exophase's method]
- added full support for windows (both inputs, OBJ windows, and output, with optional color effect disable)
- EEPROM uses nall::bitarray now to be friendlier to saving memory to disk
- removed incomplete mosaic support for now (too broken, untested)
- improved sprite priority. Hopefully it's right now.
Just about everything should look great now. It took 25 days, but we
finally have the BIOS rendering correctly.
In order to do OBJ windows, I had to drop my above/below buffers
entirely. I went with the nuclear option. There's separate layers for
all BGs and objects. I build the OBJ window table during object
rendering. So as a result, after rendering I go back and apply windows
(and the object window that now exists.) After that, I have to do
a painful Z-buffer select of the top two most important pixels. Since
I now know the layers, the blending enable tests are a lot nicer, at
least. But this obviously has quite a speed hit: 390fps to 325fps for
Mr. Driller 2 title screen.
TONC says that "bad" window coordinates do really insane things. GBAtek
says it's a simple y2 < y1 || y2 > 160 ? 160 : y2; x2 < x1 || x2 > 240
? 240 : x2; I like the GBAtek version more, so I went with that. I sure
hope it's right ... but my guess is the hardware does this with
a counter that wraps around or something. Also, say you have two OBJ
mode 2 sprites that overlap each other, but with different priorities.
The lower (more important) priority sprite has a clear pixel, but the
higher priority sprite has a set pixel. Do we set the "inside OBJ
window" flag to true here? Eg does the value OR, or does it hold the
most important sprite's pixel value? Cydrak suspects it's OR-based,
I concur from what I can see.
Mosaic, I am at a loss. I really need a lot more information in order to
implement it. For backgrounds, does it apply to the Vcounter of the
entire screen? Or does it apply post-scroll? Or does it even apply after
every adjust in affine/bitmap modes? I'm betting the hcounter
background mosaic starts at the leftmost edge of the screen, and repeats
previous pixels to apply the effect. Like SNES, very simple. For
sprites, the SNES didn't have this. Does the mosaic grid start at (0,0)
of the screen, or at (0,0) of each sprite? The latter will look a lot
nicer, but be a lot more complex. Is mosaic on affine objects any
different than mosaic of linear(tiled) objects?
With that out of the way, we still have to fix the CPU memory access
timing, add the rest of the CPU penalty cycles, the memory rotation
/ alignment / extend behavior needs to be fixed, the shifter desperately
needs to be moved from loops to single shift operations, and I need to
add flash memory support.
byuu says:
Timer speedup added. Boosts Mr. Driller 2 title from 170fps to 400fps.
Other games still benefit, but not as amazingly. I don't dip below
160fps ever here.
Reverted the memory speed to 2 for everything for now, to fix
Castlevania slowdown. We obviously need to add the N/S stuff before we
do that.
Added linear BG and linear OBJ mosaic-Y. Did not add mosaic-X, or any
mosaic to the affine/bitmap modes, because I'm not sure when to apply
the compensation.
Rewrote layer stuff. It now has two layers (above and below), and it
performs the four blending modes as needed.
Didn't add semi-transparent sprites because the docs are too confusing.
Added a blur filter directly into the PPU for now. This obviously
violates my interface, but F-Zero needed it for HUD display. We can
remove it when we have an official release with a blur filter available.
The filter still doesn't warp colors like a real GBA, because I don't
know the formula.
byuu says:
Changelog:
- HALT waits 16 cycles before testing IRQs instead of 1 (probably less
precise, but provides a massive speedup) [we will need to work on this
later]
- MMIO regs for CPU/PPU simplified by combining array accesses
- custom VRAM/PRAM/OAM read/write functions that emulate 8->16-bit
writes
- 16-bit PRAM data (decent speedup)
- emulated memory access speed (but don't handle non-sequential
penalties or PPU access penalties yet) [amazingly, doesn't help speed
at all]
- misc. code cleanups
For this WIP, FPS for Mr. Driller 2 went from 88fps to 172fps.
Compatibility should be unchanged. Timers are still an interesting
avenue to increase performance, but will be very tough to handle the
16MHz timers with eg a period of 65535 (overflow every single tick.) And
that's basically the last major speed boost we'll be able to get.
Blending and windowing is going to hurt performance, but it remains to
be seen how much.
byuu says:
Changelog:
- added FIFO buffer emulation (with DMA and all that jazz) [Cydrak]
- fixed timers and vcounter assign [Cydrak]
- emulated EEPROM (you have to change size manually for 14-bit mode, we
need a database badly now) [SMA runs now]
- removed OAM array, now decoding directly to struct Object {} [128] and
ObjectParam {} [32] (faster this way)
- check forceblank (still doesn't remove all garble between transitions,
though??)
- lots of other stuff
Delete your settings.cfg, or manually change frequencyGBA to 32768, or
bad things will happen (this may change back to 256KHz-4MHz later.)
15 of 16 games are fully playable now, and look and sound great.
The major missing detail right now is PPU blending support, and we
really need to optimize the hell out of the code.
byuu says:
Merged Cydrak's r17c changes:
- BG affine mode added
- BG bitmap mode added
- OBJ affine mode added
- fixed IRQ bug in THUMB mode (fixed almost every game)
- timers added (broke almost every game, whee.)
Cydrak is absolutely amazingly awesome and patient. This really wouldn't
be happening without him.
Also fixed some things from my end, including greatly improved sprite
priorities, and a much better priority sorter. Mr. Driller looks a lot
better now.
byuu says:
Emulated GBC sound plus the new extensions to it.
I am kind of surprised by how little developers utilized the GBC audio
portion.
Mr. Driller now has sound effects, and Pinobee no Daibouken has BGM.
I still have yet to emulate the GBA extra sound channels and PWM. Need
to emulate timers and DMA 2 refresh mode before I can do that.
Also, I moved both GBC and GBA audio to use length = data; if(++length
== 0); rather than length = 64 - data; if(--length == 0); so that
I could return literal values for register reads.
I thought there was a good reason we used the latter version, but
I can't hear any audible difference even in GBC games, so oh well.
Lastly, I think the pattern[++offset] in the wave channel was a bug in
the DMG/GBC only. I really, really hope it doesn't apply to the GBA,
because that will make bank selection a serious pain in the ass.
byuu says:
Fixed the r15 mask per Cydrak.
Added DMA support (immediate + Vblank + Hblank + HDMA) with IRQ support.
Basically only missing FIFO reload mode for the APU on channel 2.
Added background linear renderer (tilemap mode.)
Added really inefficient pixel priority selector, so that all BGs+OBJ
could be visible onscreen at the same time.
As a result of the above:
* Mr. Driller is our first fully playable game
* Bakunetsu Dodge Ball Fighters is also fully playable
* Pinobee no Daibouken is also fully playable
Most games (15 of 16 tested) are now showing *something*, many things
look really really good in fact.
Absolutely essential missing components:
- APU
- CPU timers and their interrupts
- DMA FIFO mode
- OBJ affine mode
- BG affine mode
- BG bitmap mode
- PPU windows (BG and OBJ)
- PPU mosaic
- PPU blending modes
- SRAM / EEPROM (going to rely on a database, not heuristics. Homebrew
will require a manifest file.)
byuu says:
Added linear (eg non-affine) sprite rendering, 4bpp and 8bpp with hflip
and vflip. Nothing else.
You can now see the Nintendo logo and Gameboy text at the end of the BIOS.
It's a start =)
byuu says:
Fixed aforementioned issues.
[From a previous post:
- mul was using r(d) instead of r(n) for accumulate.
- mull didn't remove c/v clear.
- APU register mask was broken, so SOUNDBIAS was reading out wrong.
- APU was only mapping 0x088 and not 0x089 as well.
- Halfword reads in CPU+PPU+APU were all reading from the low address
each time.]
All CPU+PPU registers are now hooked up (not that they do anything.)
SOUNDBIAS for APU was hooked up, got tired of working on it for the rest :P
I recall from the GB APU that you can't just assign values for the APU
MMIO regs. They do odd reload things as well.
Also, was using MMIO read code like this:
return (
(flaga << 0)
|| (flagb << 1)
|| (flagc << 2)
);
Logical or doesn't work so well with building flags :P
Bad habit from how I split multiple conditionals across several lines.
So ... r14 is basically what r13 should have been yesterday, delaying my
schedule by yet another day :(
byuu says:
Contains all of Cydrak's fixes, sans PPU.
On the PPU front, I've hooked up 100% of read and write registers.
All three DISPSTAT IRQs (Vblank, Hblank, Vcoincidence) are connected now
as well.
Super Mario Advance now runs without *appearing* to crash, although it's
hard to tell since I have no video or sound :P
ARM Wrestler is known to run, as is the BIOS.
byuu says:
Enough to get through the BIOS and into cartridge ROM.
I am a bit annoyed that I was basically told that the GBA PPU wasn't
that bad. Sprites are a clusterfuck, easily worse than Mode7, docs don't
even begin to explain them in enough detail.
This is going to be fun.
byuu says:
Added all of the above fixes and changes. [A lot of individual fixes for
the ARM core from Cydrak - Ed.] Also new is pipeline_decode() to fetch
data, and IME/IE/IF support, and an ARM::processor.irqline flag that
triggers IRQs at 0x18. Only Vblank is hooked up, which is what SWI 4 was
waiting on previously.
I'm sure my interrupt support is horribly broken and wrong. I was never
able to really figure out IE/IF on the Game Boy, so there's no question
this is even worse.
It's now going crazy and writing 0 to IE forever now after the Vblank
IRQ triggers.
byuu says:
Changelog:
- fixed THUMB hi immediate reads (immediate * 4)
- cartridge is properly mirrored to 32MB (eg 12mbit repeats as
lo8+hi4+hi4+lo8+hi4+hi4) [so it's a bit slower than a standard memcpy
fill]
- added ARM - load/store halfword register offset
- added ARM - load/store halfword immediate offset
- added ARM - load signed halfword/byte register offset
- added ARM - load signed halfword/byte immediate offset
- added decode() function to make opcode bit testing a lot clearer
(didn't apply it to the debugger yet)
All ARMv4M and all THUMBv4 instructions should now be implemented.
Although I'm not sure if my implementations of the new instructions are
correct.
byuu says:
Split apart necdsp: core is now in processor/upd96050 (wish I had
a better name for it, but there's no family name that is maskable.)
I would like to support the uPD7720 in the core as well, just for
completeness' sake, but I'll have to modify the decoder to drop one bit
from each mode.
So ... I'll do that later. Worst part is even if I do, I won't be able
to test it :(
Added all of Cydrak's changes. I also simplified LDMIA/STMIA and
PUSH/POP by merging the outer loops.
Probably infinitesimally slower, but less code is nicer. Maybe GCC
optimization will expand it, who knows.
byuu says:
Added some more ARM opcodes, hooked up MMIO. Bind it with mmio[(addr
000-3ff)] = this; inside CPU/PPU/APU, goes to read(), write().
Also moved the Hitachi HG51B core to processor/, and split it apart from
the snes/chip/hitachidsp implementation.
This one actually worked really well. Very clean split between MMIO/DMA
and the processor core. I may move a more generic DMA function inside
the core, not sure yet.
I still believe the HG51B169 to be a variant of the HG51BS family, but
given they're meant to be incredibly flexible microcontrollers, it's
possible that each variant gets its own instruction set.
So, who knows. We'll worry about it if we ever find another HG51B DSP,
I guess.
GBA BIOS is constantly reading from 04000300, but it never writes. If
I return prng()&1, I can get it to proceed until it hits a bad opcode
(stc opcode, which the GBA lacks a coprocessor so ... bad codepath.)
Without it, it just reads that register forever and keeps resetting the
system, or something ...
I guess we're going to have to try and get ARMwrestler working, because
the BIOS seems to need too much emulation code to do anything at all.
There was a "v087r07pre" release that I unfortunately missed.
byuu says (about v087r07pre):
Creates the bsnes/processor folder. This has a shared ARM core there
which both the GBA and ST018 inherit.
There are going to be separate decoders, and revision-specific checks,
to support the differences between v3+.
In the future, I also want to move the other processor cores here:
- GBZ80 (GB, GBC)
- 65816 (SNES CPU, SA-1)
- NEC uPD (7725, 96050, maybe 7720 just for fun)
- Hitachi HG51B169
- SuperFX
- SPC700
- 65(C?)02
Basically, the GBA/ST018 forces my hand to start coding a bit more like
a multi-system emulator.
Right now, the ST018 is broken. Hence the pre. Apparently the GBA core
being used now has some bugs. So this'll be a nice way to stress-test
the GBA core a bit before we make it to ARMwrestler.
byuu says (about v087r07):
Both snes/chip/armdsp and gba/cpu use processor/arm now.
Fixed THUMB to execute the BL prefix and suffix separately. I can now
get the GBA BIOS stuck in some kind of infinite loop. Hooray ...
I guess?
byuu says:
I believe I've implemented every THUMB instruction now, although I'm
sure there are dozens of bugs in the implementation.
It seems that the last jump taken is ending up being off-by-two. It's
probably due to not masking/adjusting PC correctly at certain points.
I don't know if any other bugs are being hit prior to this or not.
I don't implement any I/O registers yet, and the BIOS seems to be poking
at a few of them along the way, so ... who knows.
I could also be reading the log wrong, but it looks to me like there's
some PSR setting the mode flag register to 0, which is supposed to be an
undefined behavior mode ... perhaps mrs has no effect on the m/t bits,
and it just affects the i/f bits?
byuu says:
Implemented all of the ARMv3 instructions, and the bx rm instruction as
well. Already hit THUMB mode right at the start of the BIOS, sigh.
Implemented the first THUMB instruction to get that rolling. Also tried
to support the S flag to LDM/STM, but not sure how successful I was.
byuu says:
GBA stuff re-added. Only thing missing that was there before is the ARM
branch opcode.
Since we're going to be staring at it for a very long time, I added
a more interesting test video pattern.
Went from 6fps to 912fps. Amazing what being able to divide can do for
a frame rate.
byuu says:
Fixing the PPU stepping increased FPS to 250. Promising, at least, since
the ARM core is still severely overclocked.
However, I reverted back to r02. This one patches gameboy/ and GameBoy::
to gb/ and GB:: and that's it.
Sorry, I just couldn't shake this bad feeling about the code. There were
some poorly hacked-together constructs. I'd rather just redo two days of
work than feel bad about the codebase for the next several years. Going
to attempt the GBA bridge again. Third time's a charm, I suppose (there
was a pre-r03 WIP I abandoned as well.)
This isn't unprecedented, GB core took a few attempts like this as well.
byuu says:
Changelog:
- gameboy/ -> gb/
- GameBoy -> GB
- basic memory map for GBA
- enough code to execute the first BIOS instruction (b 0x68)
I have the code resetting r(15) to 0 on an exception just as a test.
Since that flushes the pipeline, that means we're basically executing "b
0x68" at 8MHz, and nothing else.
... and I am getting __6 motherfucking FPS__ at 4.4GHz on an i7.
Something is seriously, horribly, unfuckingbelievably wrong here, and
I can't figure out what it is.
My *fully complete* ARM core on the ST018 is even less efficient and
runs at 21.47MHz, and yet I get 60fps even after emulating the SNES
CPU+PPU @ 10+MHz each as well.
... I'm stuck. I can't proceed until we figure out what in the holy fuck
is going on here. So ... if anyone can help, please do. If we can't fix
this, the GBA emulation is dead.
I was able to profile on Windows, and I've included that in this WIP
under out/log.txt.
But it looks normal to me. But yeah, there's NO. FUCKING. WAY. This code
should be running this slowly.
byuu says:
I wanted to keep this a secret, but unlike other recent additions, this
will easily take several weeks, maybe months, to show anything.
Assuming I can even pull it off. Nothing technically overwhelming here,
I'm more worried about the near-impossibility of debugging the CPU.
byuu says:
Changelog:
- extended USART with quit(), readable(), writable() [both emulation and
hardware]
- quit() returns true on hardware when Ctrl+C (SIGINT) is generated
(breaks main loop); no effect under emulation yet (hard to
simulate)
- readable() returns true when data is ready to be read
(non-blocking support for read())
- writable() returns true when data can be written (non-blocking
support for write()) [always true under emulation, since we have
no buffer size limit]
byuu says:
Changelog:
- fixes ARM core unaligned memory reads (fixes HNMS2 AI, hopefully completely,
we'll see though) [Cydrak]
- ARM 40000010 writes are now connected to d2 rather than the timer
- ARM bus_readbyte() removed (would love to do the same for writebyte if
we can ... then we can drop back to bus_read + bus_write only)
- USART with IObit set acts as a regular gamepad now (don't have this
hooked up with real hardware, but oh well, it's technically possible
so there's that)
- OpenGL/GLX will use 30-bit when you have a 30-bit display; no need for
config file video.depth anymore
byuu says:
This release adds ST018 emulation. As this was the final unsupported
SNES coprocessor, this means that bsnes v087 is the first SNES emulator
to be able to claim 100% known compatibility with all officially
released games. And it does this with absolutely no hacks.
Again, I really have to stress the word known. No emulator is perfect.
No emulator ever really can be perfect for a system of this complexity.
The concept doesn't even really exist, since every SNES behaves subtly
different. What I mean by this, is that every single game ever
officially sold has been tested, and zero bugs (of any severity level)
are currently known.
It is of course extremely likely that bugs will be found in this
release, as well as in future releases. But this will always be
a problem for every emulator ever made: there is no way to test every
possible codepath of every single game to guarantee perfection. I will,
of course, continue to do my best to fix newfound bugs so long as I'm
around.
I'd really like to thank Cydrak and LostTemplar for their assistance in
emulating the ST018. I could not have done it without their help.
The ST018 ROM, like the other coprocessor ROMs, is copyrighted. This
means I am unable to distribute the image.
Changelog (since v086):
- emulated the 21.47MHz ST018 (ARMv3) coprocessor used by Hayazashi
Nidan Morita Shougi 2
- fixed PPU TM/TS edge case; fixes bottom scanline of text boxes in
Moryo Senki Madara 2
- fixed saving and loading of Super Game Boy save RAM
- NEC uPD7725,96050 ROMs now stored in little-endian format for
consistency
- cartridge folder concept has been reworked to use fixed file names
- added emulation of serial USART interface (replaces asynchronous UART
support previously)
byuu says:
Cydrak, I moved the step from the opcode decoder and opcodes themselves
into bus_(read,write)(byte,word), to minimize code.
If that's not feasible for some reason, please let me know and I'll
change it back to your latest WIP.
This has your carry flag fix, the timer skeleton (doesn't really work
yet), the Booth two-bit steps, and the carry flag clear thing inside
multiply ops.
Also added the aforementioned reset delay and reset bit stuff, and fixed
the steps to 21MHz for instructions and 64KHz for reset pulse.
I wasn't sure about the shifter extra cycles. I only saw it inside one
of the four (or was it three?) opcodes that have shifter functions.
Shouldn't it be in all of them?
The game does indeed appear to be fully playable now, but the AI doesn't
exactly match my real cartridge.
This could be for any number of reasons: ARM CPU bug, timer behavior
bug, oscillator differences between my real hardware and the emulator,
etc.
However ... the AI is 100% predictable every time, both under emulation
and on real hardware.
- For the first step, move 九-1 to 八-1.
- The opponent moves 三-3 to 四-3.
- Now move 七-1 to 六-1.
- The opponent moves 二-2 to 八-8.
However, on my real SNES, the opponent moves 一-3 to 二-4.
byuu says:
Most importantly ... I'm now using "st018.rom" which is the program ROM
+ data ROM in one "firmware" file. Since all three Seta DSPs have the
ST01N stamp, unlike some of the arcade variants, I'm just going to go
with ST01N from now on instead of ST-001N. I was using the latter as
that's what Overload called them.
Moving on ...
The memory map should match real hardware now, and I even match the open
bus read results.
I also return the funky 0x40404001 for 60000000-7fffffff, for whatever
that's worth.
The CPU-side registers are also mirrored correctly, as they were in the
last WIP, so we should be good there.
I also simulate the reset pulse now, and a 0->!0 transition of $3804
will destroy the ARM CPU thread.
It will wait until the value is set back to zero to resume execution.
At startup, the ARM CPU will sleep for a while, thus simulating the
reset delay behavior.
Still need to figure out the exact cycle length, but that's really not
important for emulation.
Note in registers.hpp, the |4 in status() is basically what allows the
CPU program to keep going, and hit the checkmate condition.
If we remove that, the CPU deadlocks. Still need to figure out how and
when d4 is set on $3804 reads.
I can run any test program on both real hardware and in my emulator and
compare results, so by all means ... if you can come up with a test,
I'll run it.
byuu says:
Attempted to fix the bugs pointed out by Cydrak for the shifter carry
and subtraction flags. No way to know if I was successful.
The memory map should exactly match real hardware now.
Also simplified bus reading/writing: we can get fancy when it works,
I suppose.
Reduced some of the code repetition to try and minimize the chances for
bugs.
I hopefully fixed up register-based ror shifting to what the docs were
saying.
And lastly, the disassembler should handle every opcode in every mode
now.
ldr rn,[pc,n] adds (pc,n) [absolute address] after opcode. I didn't want
to actually read from ROM here (in case it ever touches I/O or
something), but I suppose we could try anyway.
At startup, it will write out "disassembly.txt" which is a disassembly
of the entire program ROM.
If anyone wants to look for disassembly errors, I'll go ahead and fix
them. Just note that I won't do common substitutions like mov pc,lr ==
ret.
At this point, we can make two moves and then the game tells us that
we've won.
So ... I'm back to thinking the problem is with bugs in the ARM core,
and that our bidirectional communication is strong enough to play the
game.
Although that's not perfect. The game definitely looks at d4 (and
possibly others later), but my hardware tests can't get anything but
d0/d3 set.
byuu says:
That's my best implementation of the shifter carry. It's horribly
inefficient and possibly wrong (especially on ROR by register, but that
doesn't ever appear to be used in this program), but oh well. It's the
best I can do.
Game is basically getting stuck after a board upload and issuing another
command. It's sitting in a loop waiting on $3804.d0 to be set, meaning
the ARM is never writing anything for the CPU to read. There's some
chance that my $3804/r40000000 flags are wrong. Short of guessing
though, I'm not sure how we can get more info on how those work.
... I really can't debug this any better than I have. If no one else
sees anything, then we're going to have to give up and wait for MESS to
create opcode logs for us to compare against.
byuu says:
More ARM work. Can get in-game, and upload the board (0xaa) successfully.
Bug in checkmate command makes the CPU really difficult to defeat :P
byuu says:
Contains the fledgling beginnings of an ARM CPU core, which can execute
the first three and a half instructions of the ST-0018.
It's a start, I guess.
byuu says:
USART improvements. The clock pulse from reading data() drives both
reading and writing.
Also added a usart_init() to bind the initializer functions, so all you
need now is:
extern "C" usartproc void usart_main() { ... }
And inside, you use usart_read(), usart_write(), etc.
So we can add all the new functions we want (eg I'd like to have
usart_readable() to check if data is available) without changing the
entry point signature.
blargg enhanced his Teensy driver to ignore frame error reads, as well.
byuu says:
It is done. bsnes can now emulate sending and receiving data via USART.
As such, the UART code has been removed.
The final UART code can be downloaded here: http://byuu.org/snes/uart/
I won't maintain it going forward, because nobody ever used it, and
USART is superior in every way.
I've also verified both sending and receiving on the real SNES now :D
It's so easy ... a caveman with electrical engineering and computer
programming experience can do it.
byuu says:
USART implements reading and writing, but I don't yet have code to test
SNES reading yet.
So ... obviously I need to do that next.
Went ahead and required nall::function, so the modules will have to be
C++11. I don't see anyone else making these, and it avoids the annoyance
of deducing the correct controller port based on dynamic casting the
active thread.
Apparently a library can have a main() function to no ill effect, so
there's no need for USART_HARDWARE. Same exact code with different flags
will make the binary and the library.
byuu says:
There will probably be a series of small WIPs as I experiment here.
snes/controller/serial is now snes/controller/uart. Asynchronous serial
communications, typically capped at 57,600 baud.
snes/controller/usart is new. It aims to emulate the SNES connected to
a Teensy++ board, and can easily handle 524,288 baud.
And much more importantly, it's synchronous, so there are no timing
issues anymore. Just bit-bang as fast as you can.
Right now, the USART code is just enough for SNES->PC to transfer data
to ... well, nothing yet.
Unless anyone is actually using the UART stuff, I'll be removing it once
the USART is totally up and running.
No sense maintaining code that is 10x slower, more error prone, and used
by nobody.
Note: this is all thanks to blargg being absolutely amazing.
Tim Allen