byuu says:
So this turned out to be a rather unproductive ten-hour rabbit hole, but
...
I reworked nall/primitives.hpp a lot. And because the changes are
massive, testing of this WIP for regressions is critically important. I
really can't stress that enough, we're almost certainly going to have
some hidden regressions here ...
We now have a nall/primitives/ subfolder that splits up the classes into
manageable components. The bit-field support is now shared between both
Natural and Integer. All of the assignment operator overloads are now
templated and take references instead of values. Things like the
GSU::Register class are non-copyable on account of the function<>
object inside of it, and previously only operator= would work with
classes like that.
The big change is nall/primitives/operators.hpp, which is a really
elaborate system to compute the minimum number of bits needed for any
operation, and to return a Natural<T> or Integer<T> when one or both of
the arguments are such a type.
Unfortunately, it doesn't really work yet ... Kirby's Dream Land 3
breaks if we include operators.hpp. Zelda 3 runs fine with this, but I
had to make a huge amount of core changes, including introducing a new
ternary(bool, lhs, rhs) function to nall/algorithm to get past
Natural<X> and Natural<Y> not being equivalent (is_integral types get a
special exemption to ternary ?: type equivalence, yet it's impossible to
simulate with our own classes, which is bullshit.) The horrifying part
is that ternary() will evaluate both lhs and rhs, unlike ?:
I converted some of the functions to test ? uint(x) : uint(y), and
others to ternary(test, x, y) ... I don't have a strong preference
either way yet.
But the part where things may have gotten broken is in the changes to
where ternary() was placed. Some cases like in the GBA PPU renderer, it
was rather unclear the order of evaluations, so I may have made a
mistake somewhere.
So again, please please test this if you can. Or even better, look over
the diff.
Longer-term, I'd really like the enable nall/primitives/operators.hpp,
but right now I'm not sure why Kirby's Dream Land 3 is breaking. Help
would be appreciated, but ... it's gonna be really complex and difficult
to debug, so I'm probably gonna be on my own here ... sigh.
byuu says:
This is probably the largest code-change diff I've done in years.
I spent four days working 10-16 hours a day reworking layouts in hiro
completely.
The result is we now have TableLayout, which will allow for better
horizontal+vertical combined alignment.
Windows, GTK2, and now GTK3 are fully supported.
Windows is getting the initial window geometry wrong by a bit.
GTK2 and GTK3 work perfectly. I basically abandoned trying to detect
resize signals, and instead keep a list of all hiro windows that are
allocated, and every time the main loop runs, it will query all of them
to see if they've been resized. I'm disgusted that I have to do this,
but after fighting with GTK for years, I'm about sick of it. GTK was
doing this crazy thing where it would trigger another size-allocate
inside of a previous size-allocate, and so my layouts would be halfway
through resizing all the widgets, and then the size-allocate would kick
off another one. That would end up leaving the rest of the first layout
loop with bad widget sizes. And if I detected a second re-entry and
blocked it, then the entire window would end up with the older geometry.
I started trying to build a message queue system to allow the second
layout resize to occur after the first one completed, but this was just
too much madness, so I went with the simpler solution.
Qt4 has some geometry problems, and doesn't show tab frame layouts
properly yet.
Qt5 causes an ICE error and tanks my entire Xorg display server, so ...
something is seriously wrong there, and it's not hiro's fault. Creating
a dummy Qt5 application without even using hiro, just int main() {
TestObject object; } with object performing a dynamic\_cast to a derived
type segfaults. Memory is getting corrupted where GCC allocates the
vtables for classes, just by linking in Qt. Could be somehow related to
the -fPIC requirement that only Qt5 has ... could just be that FreeBSD
10.1 has a buggy implementation of Qt5. I don't know. It's beyond my
ability to debug, so this one's going to stay broken.
The Cocoa port is busted. I'll fix it up to compile again, but that's
about all I'm going to do.
Many optimizations mean bsnes and higan open faster. GTK2 and GTK3 both
resize windows very quickly now.
higan crashes when you load a game, so that's not good. bsnes works
though.
bsnes also has the start of a localization engine now. Still a long way
to go.
The makefiles received a rather substantial restructuring. Including the
ruby and hiro makefiles will add the necessary compilation rules for
you, which also means that moc will run for the qt4 and qt5 targets, and
windres will run for the Windows targets.
byuu says:
Changelog:
- hiro/Windows: use `WS_CLIPSIBLINGS` on Label to prevent resize
drawing issues
- bsnes: correct viewport resizing
- bsnes: speed up window resizing a little bit
- bsnes: fix the cheat editor list enable checkbox
- bsnes: fix the state manager filename display in game ROM mode
- bsnes: fix the state manager save/rename/remove functionality in
game ROM mode
- bsnes: correct path searching for IPS and BPS patches in game ROM
mode
- bsnes: patch BS-X town cartridge to disable play limits
- bsnes: do not load (program,data,expansion).(rom,flash) from disk in
game pak mode
- this is required to support soft-patching and ROM hacks
- bsnes: added speed mode selection (50%, 75%, 100%, 150%, 200%);
maintains proper pitch
- bsnes: added icons to the menubar
- this is particularly useful to tell game ROMs from game paks in
the load recent game menu
- bsnes: added emblem at bottom left of status bar to indicate if a
game is verified or not
- verified means it is in the icarus verified game dump database
- the verified diamond is orange; the unverified diamond is blue
- bsnes: added an option (which defaults to off) to warn when loading
unverified games
- working around a bug in GTK, I have to use the uglier
MessageWindow instead of MessageDialog
- bsnes: added (non-functional) link to <https://doc.byuu.org/bsnes/>
to the help menu
- bsnes: added GUI setting to toggle memory auto-save feature
- bsnes: added GUI setting to toggle capturing a backup save state
when closing the emulator
- bsnes: made auto-saving states on exit an option
- bsnes: added an option to auto-load the auto-saved state on load
- basically, the two combined implements auto-resume
- bsnes: when firmware is missing, offer to take the user to the
online help documentation
- bsnes: added fast PPU option to disable the sprite limit
- increase from 32 items/line + 34 tiles/line to 128 items/line +
128 tiles/line
- technically, 1024 tiles/line are possible with 128 sprites at
64-width
- but this is just a waste of cache locality and worst-case
performance; it'll never happen
Errata:
- hiro/Windows: fallthrough on Canvas `WM_ERASEBKGND` to prevent
startup flicker
byuu says:
Changelog:
- nall: renamed array to adaptive_array; marked it as deprecated
- nall: created new array class; which is properly static (ala
std::array) with optional bounds-checking
- sfc/ppu-fast: converted unmanaged arrays to use nall/array (no speed
penalty)
- bsnes: rewrote the cheat code editor to a new design
- nall: string class can stringify pointer types directly now, so
pointer() was removed
- nall: added array_view and pointer types (still unsure if/how I'll
use pointer)
byuu says:
Changelog:
- sfc/ppu-fast: fixed overscan crash
- sfc/ppu-fast: fixed direct color mode
- sfc: reconnected MSU1 support
- higan: game.sfc/msu1/data.rom, game.sfc/msu1/track-#.pcm
- bsnes: game.msu, game-#.pcm
- bsnes: added cheat code editor
- bsnes: added cheat code database support
- sfc/ppu-fast: clear overscan lines when overscan disabled
- sfc: output 223/239 lines instead of 224/240 lines
- bsnes: fix aspect correction calculation
- bsnes: crop line 224 when overscan masking is enabled
- bsnes: exposed Expansion Port menu; but hid “21fx” from the list of
devices
- bsnes: tools menu is hidden until a game is loaded
- ruby/input/keyboard/quartz: fixed compilation error
So only bsnes the automated overscan cropping option. In higan, you can
crop however many lines you like from the top or bottom of the image.
But for bsnes, it automatically eats sixteen lines. My view right now is
that if bsnes is meant to be the casual gaming emulator, that it should
eat line 224 in this mode. Most games show content here, but because of
the way the SNES PPU works, the very last line ends up on its very own
tile row (line 0 isn't rendered), if the scroll registers don't account
for it. There's a small number of games that will draw junk data to the
very last scanline of the frame as a result of this. So I chose, at
least for now, to hide it. Users can obviously disable overscan cropping
to see this scanline. I'm open to being convinced not to do this, if
someone has a compelling reason.
We're pretty much screwed one way or the other with no overscan masking.
If we output 239 lines, then most games will render 7 blank lines + 224
drawn lines + 8 blank lines, and the black top and bottom aren't
centered. But if we output 240 lines to get 8 + 224 + 8, then games that
do use overscan will have a blank line at the very bottom of the window.
I'm also trying out a modified cheat code file format. It's been forever
since I bothered to look at it, and the “cartridge” parent node doesn't
match what I'm doing with trying to rename “cartridge” to “game” in
manifests. And indeed, the idea of requiring a root node is rather
superfluous for a cheat code file. Current format looks like this:
cheat
description: foo
code: 7e2000=20+7e2001=30?40
enabled
cheat
description: bar
code: 7e4000=80
Open to discussing this, and I'd like to sync up with Snes9X before they
push out a new release, and I'll agree to finalize and never change this
format again.
I chose to use .cht for the extension when using game files (eg
gamename.cht)
byuu says:
Changelog:
- sfc/ppu-fast:
- don't use mosaicSize unless mosaicEnable is set
- fix background tiles that aren't 8x8 in size
- flush (render) queued lines whenever VRAM or OAM are modified
mid-frame
- queue tile outputs to buffer for object rendering final pass
- fix object window mask indexing
- disable color bleed when output width is 256 pixels
- handle reset(bool) events
- implemented save states
- icarus: fixed SPC7110-RAM-EPSONRTC mapping typo [hex_usr]
- bsnes: fixed overscan masking mode when output height is 240
Todo:
- sfc/ppu-fast: should not have deleted the tilecache freeing in
~PPU()
- ruby/input/carbon: change setPath() call to setPathID()
Errata:
- Rendering Ranger R2 crashes at startup, seems to be an issue with
the expansion port device
Bug reports on the new fast SNES PPU are now welcome.
byuu says:
Changelog:
- nall/GNUmakefile: added `openmp=(true,false)` option; can be toggled
when building higan/bsnes
- defaults to disabled on macOS, because Xcode doesn't stupidly
doesn't ship with support for it
- higan/GNUmakefile: forgot to switch target,profile back from
bsnes,fast to higan,accurate
- this is just gonna happen from time to time, sorry
- sfc/dsp: when using the fast profile, the DSP syncs per sample
instead of per clock
- should only negatively impact Koushien 2, but is a fairly
significant speedup otherwise
- sfc/ppc,ppu-fast: optimized the code a bit (ppu 130fps to 133fps)
- sfc/ppu-fast: basic vertical mosaic support (not accurate, but
should look okay hopefully)
- sfc/ppu-fast: added missing mode7 hflip support
- sfc/ppu-fast: added support to render at 256-width and/or 240-height
- gives a decent speed boost, and also allows all of the older
quark shaders to work nicely again
- it does violate the contract of Emulator::Interface, but oh
well, it works fine in the bsnes GUI
- sfc/ppu-fast: use cached CGRAM values for mode7 and sprites
- sfc/ppu-fast: use global range/time over flags in object rendering
- may not actually work as we intended since it's a race condition
even if it's only ORing the flags
- really don't want to have to make those variables atomic if I
don't have to
- sfc/ppu-fast: should fully support interlace and overscan modes now
- hiro/cocoa: updated macOS Gatekeeper disable support to work on
10.13+
- ruby: forgot to fix macOS input driver, sorry
- nall/GNUmakefile: if uname is present, then just default to rm
instead of del (fixes Msys)
Note: blur emulation option will break pretty badly in 256x240 output
mode. I'll fix it later.
byuu says:
Changelog:
- sfc/ppu-fast: everything other than vertical mosaic and interlace
support is in
Games are quite playable now, and you're welcome to try things out, but
please don't report bugs yet. It's still too early for that.
byuu says:
Changelog:
- sfc/ppu: collapsed folders to a single directory to match all other
emulated processors
- sfc/ppu-fast: implemented I/O registers
byuu says:
Changelog:
- SNES: started on skeleton of the new parallel PPU core
To build the new PPU core, set profile=fast via GNU make. The old core
is profile=accurate.
The names of the profiles, and the name of the folder for the fast PPU
are subject to change.
The new PPU core doesn't do anything but demonstrate the proof of
concept: every scanline, make a copy of all the PPU registers and CGRAM.
Share the VRAM and OAM. Batch render all scanlines at once using OpenMP
at the end of each frame and blit the result.
With no PPU core at all, bsnes runs 91% faster than with the accuracy
PPU (230fps vs 120fps.) That's the absolute theoretical best-case
scenario. With the skeleton in place, we're already around 220fps. It'll
go down more as the PPU line renderer starts to do real work. I don't
know where things will end up yet. I suppose we'll find out in time.
My own copy of TDM/GCC can't use OpenMP on Windows, so ... it won't
parallelize if you build with that. I'm going to have to switch to a
different MinGW distribution once this is complete, I suppose.
Tim Allen