The core and the headless tools no longer use mutexes, condition variables,
spinlocks, std::atomic, std::thread or volatile flags. Threads hand off through
libretro-common's retro_atomic and rthreads plus a few small pieces on top of
them in Common/Thread: Thread, MpscQueue, EventCounter, ParkingLot waits and
ReaderGate. Every wait blocks until it's woken; nothing polls on a timer.
State that only one thread touches just loses its lock. The PSP filesystem,
sceIo and the savedata dialog now belong to the emulation thread: async reads
and writes run when they're queued and complete later in emulated time, and
the IO and savedata threads are gone. The friend finder only talks to the
adhoc server and hands what it gets to the emulation thread. GPU init always
runs on the thread that starts the game.
The Vulkan render and compile queues, pipeline and render pass caches, delete
list and SPIR-V cache, the GL render queue, the software renderer's jit caches,
the file loader caches, the web debugger, UPnP, input, savestates, logging and
the GE debugger all hand off through queues and atomics. The IR jit's spinning
profiler thread goes.
A CPU thread with nothing to run now sleeps until work is queued instead of
spinning. Headless with --debugger waits for a client while stopped, and ends
the run once every client that connected has left. Web server shutdown keeps
servicing the CPU queue until the server has finished.
The libretro core is the frontend, and the headless tools are the test and
debugging hosts. The Windows, UWP, Android, iOS, macOS and SDL apps, the UI
framework and imgui go, along with what only they used: the Win32 debugger
hooks (g_frameMutex, the shutdown lock), JNI thread attachment, the Android
text drawer, RAIntegration, Discord, recent files, game installs, tilt input,
the audio resamplers, the control mapper, memstick moving and the remote ISO
and upload server modes. The web server keeps the debugger. The MSVC, Gradle
and ndk-build projects go too.
The core no longer sleeps to pace itself: the frontend runs retro_run at
its own rate, so the frame throttle, the idle-vblank delay and the real
clock sync event (and its core option) are gone, and a stepping CPU hands
straight back instead of waiting 2 ms. Waits that polled now park on a
lock-free eventcount through Common/Thread/ParkingLot and are woken by
whoever changes the state: a paused GL emu thread, which also queues a
marker behind its last frame so pausing runs the render queue up to it
instead of polling; async I/O results; the software renderer when its
queue is full; Vulkan pipeline compiles; and debugger requests once
stepping ends. Boot joins the loader thread, cache shutdown joins its
read-ahead thread, and the slow-I/O simulation is removed.
The path browser, the HTTP and WebSocket servers, the Windows console and
the game info cache are left out of the core build. libretro/test gains
lrgltest, which drives the GL backend on a surfaceless EGL context through
every path that pauses or stops the emu thread; run.sh and the image
checks CI job run it on llvmpipe.
Deflate, zlib and gzip streams go through rinflate and rdeflate, zstd
through rzstd, CHD images through rchd, zip archives through
rzip_archive and PNG through rpng. The bundled libraries, their build
glue and the system-library options are gone; adler32 and uncompress
keep zlib's ABI for the prebuilt FFmpeg archives that still call them.
libretro-common is one CMake library that the core, the headless tool
and the unit tests all link.
The reads take shorter paths on the way. CHD decodes straight from the
loader's reads and a range of blocks at a time, and a file read at any
alignment decodes into place; CSO keeps one inflate stream instead of
one per block; a plain ISO serves a file read with one read of its exact
bytes; a stored zip member is read in place and a deflated one through
restart points rather than a copy of the whole image in memory; zip
members decode into the caller's buffer; sequential reads through the
frontend's VFS skip the seek, which also keeps its buffering.
libretro/test reads one image as ISO, CSO, CHD and zip through the block
devices, boots each, and checks the codecs against Python-made vectors;
a CI job runs it.
Linux's uint64_t is unsigned long, which neither of its u32 and u64
(unsigned long long) overloads takes exactly, so the call was ambiguous.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The range check was written as tests for outside, which are all false for
NaN, so a NaN vertex passed and was converted to an integer. That differs
by CPU: arm64 gives 0, which lands off screen, while x86-64 gives INT_MIN,
which wraps, and the triangle's area overflowed into a huge sliver. In the
recording-opengl and recording-vulkan dumps that drew a beam across the
top right that a PSP replay of the dump doesn't show. Now such a vertex is
flagged, and its triangles dropped, on both.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A full state ring flushed the binner, and the next state change filled
it again in the same way. Now the flush also doubles the ring, up to
32768 states, so it happens at most a few times. Nothing draws right
after a flush, so the ring can move then. In the benchmark games, no
more than 512 states were queued between flushes, so 4096 stays the
starting size.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The binner kept a bounding box of every queued item's range, updated for
each item, only to tell whether anything had been queued since the last
flush. A flag says the same. Expand, which grew the box, is now
ItemQueued.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each queued primitive adds its vertex flags (alpha, fog, colors) to its
state, and states are optimized for those flags before their items are
distributed. A state can still gain primitives after some of its items
are in tiles. Re-optimizing it then changed its pixel and sampler
functions and IDs while threads drew with them. With the x86 pixel JIT,
the span function could turn null between checking and calling it, as
in Coded Arms' frame dump. God of War hit this about four times a frame.
Now a state is live once any of its items is in a tile, until the tiles
are drawn and reset, and a live state isn't changed. A primitive that
would change a live state's flags starts a new, unoptimized state from
the same registers instead, optimized for its own primitives only.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The pixel and sampler JITs only exist on x86-64, but elsewhere a cache
miss still flushed the binner before the compile that did nothing, and
the cache never filled. So on ARM every switch to a new pixel or sampler
state was a full flush: hundreds per frame.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Expanding a pending range by a rectangle starting below its start added
a negative row count to an unsigned height, which wrapped to millions of
rows. Overlap tests then overflowed and missed real overlaps. Now a
merged range is whole rows from the lower start to the higher end.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A draw whose texture rectangle overlapped what queued primitives write
flushed, once per draw. NeedsOrder already orders each primitive by the
texels it actually reads, so let it. God of War's bloom passes texture
from the main framebuffer declared as 512x512, which reaches into the
buffer they draw to, and flushed every 30 primitives: 6% faster there.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every tile a triangle touches sets it up again, so larger tiles do less
of that. The tile size is now picked at startup, as large as it can be
with at least sixteen tiles per thread over the screen. With fewer, a
scene whose drawing is concentrated on part of the screen leaves threads
idle: on eight threads, 128x32 tiles (four per thread) took LocoRoco a
third longer than 64x16, and God of War 5% longer. So eight threads keep
64x16, and four or fewer get larger tiles.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Returning to the CPU flushed all queued drawing, also when a list only reached its stall address.
Wipeout Pure builds its lists as it goes and stalls about a hundred times a frame, so the main thread
spent a third of its time finishing the drawing at those flushes. A stalled list isn't done, so now
only a finished one flushes there, which makes Wipeout Pure 15% faster.
What relied on that flush now flushes itself: memory copies, sets, readbacks and uploads that overlap
the queued drawing, and save states, before memory is saved or loaded (FlushPendingDrawing). Syncs and
the display already did. Queued drawing still reads textures from PSP memory when it runs, so a game
that rewrites a texture between stalls without syncing would see the new one; on hardware the GE has
usually drawn with the old one by then.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Waking a worker is a condition variable signal, a system call, and the main thread paid one per idle
worker every 48 queued items. With Tekken 6's tiny draws (about 5 vertices each) the workers keep
running dry, and those signals were about 10% of the main thread, which is what limits it there.
Now no more threads are woken than tiles have work, and only the first is woken from the main thread;
it wakes the rest.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A state that textures from what it draws to, or whose scissor reaches
past the stride, used to switch the binner to one thread, and every
primitive was drawn alone, copying the whole texture for its snapshot.
In Wipeout, God of War and GTA that was a third to half of the time.
Now it's decided per primitive. One is drawn after everything queued
and before what follows only if it needs that: it textures from texels
a queued primitive (or itself) writes, it writes texels a queued one
still has to read, or it reaches past the stride. Reads are bounded by
the vertices' texture coordinates plus the filter's neighbors and
mapped to tiles of the target; queued writes and reads are tracked per
tile. Everything else goes to the tiles as usual. The in-order ones are
split into bands of rows when large, since each pixel reads only the
snapshot, and the snapshot now only copies the previous primitive's
columns.
Same results on 263 frame dumps and 8 games.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each thread used to own a band of the screen, fixed when the work was
first split, and got its own copies of the primitives touching it. Bands
were rarely balanced (edge bands had half the work of middle ones, and
on big.LITTLE CPUs half the threads are slower), and primitives queued
up to the 2048 limit before the threads started.
Now primitives go to 64x16 tiles, as indices into the shared queue, in
batches of 16. A thread takes any tile with work and draws its items in
order; only one thread holds a tile at a time, so per-pixel order is
kept. A flush or a full queue has the main thread draw tiles too instead
of waiting, and when there's only a little work it does that alone
rather than wake the threads. Items are reclaimed once every tile has
drawn them. This also drops the per-thread queues, up to 23 MB.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
HLSL types an unsuffixed hex literal as uint, so asint(x) - 0x3F800000
went unsigned and wrapped for every x below 1. The huge t then turned
into a NaN, and anything lit with specular or powered diffuse came out
black on D3D11 (the flags in Everybody's Golf's language select, #22418).
GLSL types it as int and was unaffected.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Texture pack installs took the destination directory from the archive's
textures.ini, and PKG update installs from the PBOOT's DISC_ID, both
unchecked, so a crafted archive or package could write outside
PSP/TEXTURES or PSP/GAME. Add IsSafePathComponent() to PathUtil and check
both, replacing the WebServer's own upload filename check with it too.
DISC_ID from any PARAM.SFO also names configs, cheats, textures, exdata
and savestates, so ReadSFO now drops one that isn't a safe name and the
game gets a fake ID like homebrew. The same goes for a frame dump's ID.
Combines #22419 and #22420 by Acts1631.
Co-Authored-By: Acts1631 <acts1631kjv@proton.me>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The GE keeps the specular and spot exponents to 5 significant bits,
truncated, and from 512 up (infinity and NaN included) uses 496. Only
the specular one was truncated, and neither saturated. A huge or NaN
exponent overflowed pspPow's bits: highlights wrong on every backend,
and on D3D11, where the overflowing float to int conversion gives a
NaN, black (Everybody's Golf, #22418). With the exponent clamped, t
stays in range for every base the backends produce.
Measured, now bit exact in the software renderer and within 4 levels
on Vulkan.
Fixes#22418
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The x86 sampler JIT returned zero for a whole sample when a texture level
pointer was null, and for a mipmapped texture it checked the next level
too, even when it isn't blended in. Tactics Ogre draws its font from a
mipmapped texture whose next level has an invalid address, at level 0,
so all its text was missing on x86-64. The generic samplers read a null
level's texels as zero only where they use that level, and still apply
the texture function. Now samplers for such textures aren't compiled,
and the JIT's null checks are gone. Test: gpu/textures/mipinvalid.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
IsLargerThanHalf's non-SIMD loop assigned instead of or-ing, so only blue
counted: a yellow or red light gave no diffuse or specular on riscv64 and
loongarch64 (gpu/commands/light, gpu/exact/lighting and curves failed
there). The NEON version took a channel of 1 for none; it now agrees with
SSE, which counts any channel above zero.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The exact path evaluates its own columns, so the float SampleU results went
unused. Pursuit Force's frame dump replays 16% faster in the software
renderer.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Like z, a primitive whose vertices are all beyond the same x or y plane
(|x| > w) isn't drawn, for every primitive type and with clipping on or off;
a vertex on the plane is inside (gpu/probe exp176, gpu/clipping/xycull).
Immediate-mode vertices carry screen coordinates in their clip position, so
only z culls them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each quad is its left edge down, the diagonal from the bottom left up to the
top right, then its right edge down (gpu/probe exp175, exact; fixes
gpu/primitives/bezier and spline, whose end pixels depend on direction and
order).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Transform-mode lines take the triangle's per-vertex s, t, q (15 bits at the
larger exponent) along the line's gradient, like their color and depth, then
u = s * R(q). A pixel just inside a line's end samples short of the end
vertex's coordinate rather than at it (gpu/texmtx/prims: texel row 255, not a
wrap to 0).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
NaN and +inf directions light like (1, 1, 1), -NaN and -inf like (-1, -1, -1)
(gpu/commands/light).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A vertex with z exactly -w isn't clipped, so with clipping on the GE discards
a triangle whose such vertex lands outside the 4096 guard band. The check
skipped it (gpu/clipping/guardband).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A texture that fits in the GE's 8 KB texture cache keeps what was first
read until TEXFLUSH, across primitives and draws, so a later self-textured
draw doesn't see an earlier one's writes (gpu/probe exp174: three passes
over a 16x16 4444 buffer differ in 177 pixels with and without a flush
between them, both now exact). FF Type-0's (NPJH50443) blur; two of its
PSP replays become exact.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
It used (0, 0, 1) instead. The PSP keeps it zero, under every light kind
and type (gpu/probe exp173). SOCOM (UCES01242) lights meshes whose s8
normals are all zero; its PSP replay becomes exact.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The same way as Gouraud color: a gradient along the major axis from the
setup reciprocal, and each pixel's value from its center's signed distance
from v0, so an endpoint's pixel extrapolates past the endpoint's depth
(gpu/probe exp170, exact; it had 435 wrong pixels). Coded Arms' (ULUS10019)
last pixel, where a short line passed a GEQUAL test, and SOCOM's line depth.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Which side an edge is on decides whether a pixel center exactly on it is
inside. The test divided with truncation, so a vertex 0.2 subpixels left of
the opposite edge counted as on it, and a thin triangle's right edge became
inclusive. Peace Walker (ULUS10509), ULJM05302 and mali_missing_circle PSP
replays become exact.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Two triangles forming a rectangle each have their own UV planes. Scaled, a
sample can land right on a texel edge, where the plane's anchor decides the
side: Tokiden's (NPJH50878) 480x33 bar over 280x20 texels took the row
above in its second triangle (gpu/probe exp167). 1:1 samples are half a
texel from any edge.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Pixels past the framebuffer's stride land at the start of the next row,
which another binning task may be drawing, so the result raced. Tokimeki
Memorial (NPJH50127) blurs a 128 wide buffer under a full screen scissor;
its PSP replay is now exact (gpu/probe exp166).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With no steps along the line, both endpoint weights were zero, so fog (and
UVs) interpolated to zero: full fog. SOCOM's (UCUS98649) radar lines, a few
pixels long in an orthographic view, came out in the fog color.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
It used (0, 0, 1) instead, lighting everything facing the camera. The PSP
leaves the direction zero: no diffuse, and specular from the eye vector
alone (gpu/probe exp164). ULJM05302 has such a light on its cars, and the
Yu-Gi-Oh (ULJM05940) replay becomes exact.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The GE sets up UV planes per sprite, and their gradient is only exact for a
power of two area, so a 49 wide 1:1 sprite samples just below the texel
centers while the 128 wide sprite it was joined into hit them (gpu/probe
exp163, the ULJM05302 track map). Nearest sampling picks the same texels
either way.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The jump back to the start of the list buffer sets BASE to the buffer's,
but vertex and index addresses already queued for the next draw were made
for the previous one, so they pointed 16 MB away. Tiger Woods (ULUS10420)
read a skinned draw's indices as zeros and lost it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The software renderer has the real contents of what a replay drew, and a
recorded framebuffer texture can be stale or run over other buffers at its
recorded size. So a copy leaves out the areas the replay's draws could reach
(region and scissor, or the vertex bounds of a 2D draw), instead of skipping
it only when it starts at a drawn buffer:
- Princess Maker 5 textures from the right half of its 1024 wide
framebuffer, and the copy blanked the whole frame.
- Rainbow Six (ULUS10206) copies its 512x512 frame texture, which covers its
display buffer, before each strip it draws into that buffer.
- Burnout keeps a texture beside its frame in columns 480-511, which still
comes through.
- Depth writes count as whole rows in 8-row bands, since the PSP stores
depth swizzled (SOCOM UCUS98649's frame texture copy ran into its depth
buffer's columns 480-511).
Copies also stop at the end of VRAM, and flush the software renderer's
queued draws before writing.
Memsets were also never applied with the software renderer, which leaves
them to the caller.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The GE samples through its texture cache, so a primitive that textures from
its own render target mostly reads the old contents (geprobe exp81, exp159,
exp160). Give each such primitive a snapshot, refreshing only the rows the
previous one drew.
3rd Birthday (ULUS10567) bloom becomes exact on PSP replays, Burnout
Dominator's glare 93% -> 98% and Ridge Racer's blur 40% -> 94% of pixels.
The remaining differences are rows where the cache's lazily loaded 8-row
blocks already see the primitive's own writes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The GE never forms a world position for lighting: the vector to a point
or spot light is the light position minus the world translation, minus
the model position times the world matrix, summed like a matrix row.
gpu/probe exp153-158 (point lights 0.02-20 units away, spot exponents up
to 40, game-like world matrices; exp153 is a Syphon Filter vertex) are
exact, and Syphon Filter (#13568) becomes exact.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An endpoint exactly on the left or right corner of a pixel's diamond (top
or bottom for y-major lines) doesn't count as inside it: the pixel is lit
only if the line crosses the diamond (gpu/probe exp149). Makes exp72,
Virtua Tennis and echochrome exact.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The minor offset walks from v0 with a slope quantized like a triangle
plane's gradient (setup reciprocal, 14 fraction bits per subpixel,
floored), rather than one product per pixel. Makes gpu/probe exp112
exact, and Virtua Tennis and echochrome lose their AA line misses.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Next to an edge between two z = 0 vertices, the floored depth plane can
come out -1. The GE stores and range-tests 0 there, in through and
transform mode, with depth clamp on or off (gpu/probe exp148). Fixes a
seam across the sky in Frontier Gate Boost.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
De Boor at t = k/256 with the Bezier path's lerps. Each blending factor
(t - knot) / span is quantized in 1/256 from the nearer of its two knots,
and a segment's last point is evaluated as the next one's first. Colors,
texture coordinates, generated UVs and normals work as for Bezier
patches. Bit exact on gpu/probe exp144-146; software renderer only.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The cross product of two tangents, each the difference of the de
Casteljau's last two points: the row's along u, and the columns' last
points evaluated along u for v. The lerps pass their operand through at
the patch edges. Vertex normals stay unused (gpu/probe exp142-143).
Patch normals are now also generated for shade mapping with lighting
off, on every backend: the GE uses them there too.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
