Tomba2 16:9 GL reproduced the reported collapse (12-14fps) and the GPU-side
split EXONERATED the GPU: scene_gpu ~10ms, mirror ~0.2ms, while emu_cpu is
~82ms (vs ~15ms in squash). gl_ws_ablate=1 (mirror passes fully off) leaves
it at ~75ms, so the wedge is CPU work outside the mirror draws. The GL
submission paths run INSIDE guest emulation (GP0/DMA processing), so driver
CPU cost lands in emu_cpu where no GPU timer sees it.
frame_perf now reports, per wide/4:3 frame class: CPU wall inside
flush_tex_batch (cpu_flush_ms_avg), CPU wall inside the glb_wide_* entry
points (cpu_wide_ms_avg), batches/frame, wide target sets/frame, and wide
FBO creations/frame (a nonzero steady value = surface recreation churn).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Native-wide perf attribution instrumentation (permanent):
- Each wide-mirror pass (tex-batch mirror, flat-geometry mirror, overlay
full-width rect, wide_clear) is bracketed with a GL_TIMESTAMP pair
(glQueryCounter; TIME_ELAPSED cannot nest inside the frame scene query).
Pairs pool per buffered frame (1024/frame), summed at readback into
mirror_gpu_ms + mirror_passes per sample; frame_perf now reports
mirror_gpu_ms_avg/max, canon_gpu_ms_avg (scene minus mirror),
mirror_passes_avg and mirror_pass_us for wide frames.
- New debug cmd gl_ws_ablate mode=<0..3>: 0 normal, 1 skip mirror passes
entirely, 2 mirror state churn without draws, 3 mirror draws stay on the
hr FBO (no per-pass FBO rebind; content-corrupting perf probe).
First MMX6 16:9 measurements with the split: mirror passes cost 20-125us
EACH (FBO ping-pong penalty per pass) but MMX6 2D batches to ~3-15
passes/frame so 16:9 attract holds ~60fps on this branch; the per-pass
cost x ~900 immediate flat prims is exactly the Tomba2 3D 60ms scene.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Root cause (proven via the new present ring + gl_coh_ring correlation):
flush_cpu_upload() merged all pending CPU->VRAM writes into ONE union
bounding box. A frame with two disjoint uploads (sprite column + tile row —
routine MMX6 texture streaming) produced a union spanning the display
framebuffers, which the flush painted from the STALE CPU VRAM mirror (the
FBO is authoritative under GL) — stomping live frames with black for
exactly two presents (one per double-buffer parity). Scene-dependence of
the uploads made it look config/timing dependent; the software renderer is
immune because its CPU array is authoritative.
Fix: exact pending-rect list (16 rects; merges only when zero uncovered
pixels are added — containment or same-band extension; wrap-aware GP0(A0)
transfers split into up to 4 exact rects instead of the old "wrapped: take
all" full-VRAM union; overflow -> order-preserving flush-all; pre-context
union retained where the CPU array is authoritative). Merge rule verified
by a 20k-randomized-rect host test (0 stale / 0 missing painted pixels).
Observability (always-on): gl_present_ring records every SwapWindow (path,
src/letterbox rects, drained glGetError, wall-ms, backbuffer + blit-source
pixel samples) — the instrument that gave the 1:1 black-capture
correlation. debug_server_send_fmt now heap-formats responses >64KB (they
previously truncated into unparseable JSON, breaking large ring dumps).
(debug_server.c also carries the Vulkan screenshot_file sync hunk from the
concurrent VK bring-up; see the next commit.)
Validated: ~18-min MMX6 GL attract soak, ~1600 window captures across 3+
attract cycles — zero isolated black frames, zero GL errors (pre-fix:
19/40 captures black at the intro). Tomba1 build-gl boots clean; its full
attract soak follows. ENHANCEMENTS.md added (framework enhancement ledger,
R1 OpenGL / R2 Vulkan state).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Vulkan backend alongside software + OpenGL: geometry/textured-prim
pipelines, GPU-authoritative VRAM image, deferred-submit + drain-at-present
perf model, GPU-direct present (vk_renderer_present_vram), and a
full-VRAM-pack readback path (ensure_cpu rect_adds the whole VRAM rect
before pack_flush so screenshots/guest VRAM reads reflect live content,
not just the incremental pack-dirty rect — readback-only, no per-frame cost).
Runtime plumbing: --debug-port / --renderer / --window-title CLI overrides
(multi-instance fleet), latency_ring, SPIR-V shaders + embed_spirv.py,
config_loader renderer selection, debug_server screenshot wiring.
All on feat/vulkan-renderer; snapshot in backup/vulkan-wip-20260617.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Measured Tomba village = 93% textured prims, so draw batching must target the
textured path (per-prim uv-limits/texpage/clut/depth move to vertex attributes
so consecutive textured prims coalesce into one draw). Flat batching alone
would touch only 7% of prims.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds prims/frame and per_prim_us to the frame_perf aggregate. Proved Tomba's
village is draw-call bound: ~1000 prims/frame at ~14us of GPU time PER
glDrawArrays (a 1-triangle draw should be <1us) — pipeline-flush overhead from
re-binding state between every primitive. 16:9 is 2x slower purely because the
native-wide double-draw issues 2x the draw calls. Points squarely at draw
batching (group consecutive same-state prims) as the fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Always-on per-frame ring in the GL renderer: two GL_TIME_ELAPSED queries
bracket each frame's scene draws vs the present clear+blit (true GPU time,
independent of CPU/GPU overlap), plus SDL perf-counter CPU wall time
(present-to-present total, present call). Exposed via the debug server
"frame_perf" command, aggregated over the recent ring and split by 4:3 vs
native-wide so a 16:9-vs-4:3 A/B reads off one query.
Built to answer where a frame goes without guessing. First use proved
Tomba's 16:9 village lag is the native-wide DOUBLE-DRAW (every primitive
re-issued into the wide FBO), NOT supersampling (SS2==SS1) and NOT the
compositor (present_gpu ~1ms): scene_gpu 16:9=28ms vs 4:3=17ms.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The GL native-wide present (Option B) did glFinish + glReadPixels of the whole
wide FBO to the CPU and re-uploaded it EVERY frame — a hard GPU pipeline stall +
round-trip that made OpenGL 16:9 run far slower than the software renderer (whose
wide path is pure CPU, no GPU sync). Confirmed by isolation: GL 4:3 smooth, GL
16:9 slow, independent of SS/AA.
Add gl_renderer_present_wide_fbo: blit the displayed buffer's wide FBO straight to
the window with glBlitFramebuffer (GPU-side, V-flipped, into the letterbox),
exactly like the canonical gl_renderer_present_vram — no readback, no glFinish,
no re-upload. main.cpp routes wide GL frames here (falling back to the CPU readout
only if the wide surface doesn't exist yet). USER-VALIDATED smooth. The CPU
readback path (glb_render_wide_display) stays as the fallback + the wide_shot
probe.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- FMV: 24-bit frames AND streamed 15-bit MDEC video (Tomba plays its
movies 15-bit; depth24 alone missed them — user-reported) present
pinned to native 4:3. Detector = colour MDEC decode within the last
30 vblanks (mdec_recently_active); texture-decompression luma
decodes do not stamp. Both GL present paths take force_4_3; SW path
centers a 4:3 dst rect. User-verified on the intro movies.
- Menu garble: full-2D screens (no tagged prims in flight for 2
frames) now squash lines, polylines, flat rects and mono sprites
too — dialog borders are drawn with these and stuck out past the
squashed SPRT boxes. Full-width TILEs (fades) stay exempt; in-game
non-textured prims stay untouched (may be world-anchored).
- HUD positioning: in-game untagged SPRTs pivot by thirds — outer-
third elements anchor to their screen edge (16:9 corner position,
native proportions), middle third to centre. Full-2D screens keep
the single centre pivot so composites stay coherent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds the DuckStation-style widescreen hack as a config option,
implemented in the runtime (no regen needed):
- gte.cpp: gte_set_display_aspect(num, den). RTPS/RTPT scale the
IR1*h/sz screen-X term by (4*den)/(3*num) around the game's own
OFX projection centre (identity at 4:3). Games cull against the
SXY values they read back from us, so post-projection screen-
bounds culls stay aligned with the visible frame.
- present: GL letterbox_rect generalised to the configured aspect
(gl_renderer_set_display_aspect); SDL_Renderer path gets an
aspect-derived logical size + dst rect. Window opens at the
configured aspect (clamp_window_4x3 -> clamp_window_aspect).
- config: [video] aspect_ratio = W:H (4:3 default, 16:9, up
to 32:9) in game.toml, settings.toml (load+save), launcher seed.
World geometry keeps correct proportions on the stretched present;
screen-space 2D (HUD, FMV, sprite quad widths) stretches — the
known limitation of the squash tier. EXPERIMENTAL.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Full rewrite of the GL backend around a single authoritative hr VRAM FBO
(RGBA8 + stencil mask-bit mirror, internal scale from [video]
supersampling): GPU fills/copies/draws rasterize into it, CPU uploads
batch through a pending-rect quad blit, a PACK pass re-encodes dirty
regions into the native R16UI raw mirror for texture sampling and
readback, and gpu.c routes GPUREAD/mask-check/peek reads through the
facade so the CPU array syncs before any CPU-side read.
Texture sampling adopts the Beetle-PSX model (structure only, studied
from beetle-psx-libretro):
- Per-primitive uv sampling bounds, clamped in the fragment shader when
no texture window is active. Filtered neighbours and S>1 interpolation
overshoot can never read outside the prim's own texture rect. Bounds
back off the exclusive max edge for axis-aligned (2D) uv mappings and
widen to the full page when the uv range crosses a 256 wrap boundary.
- Bilinear takes the NEAREST texel as base (cutout + STP authority),
neighbours toward the sub-texel offset, with per-texel opacity weights
and renormalised colour. The old "-0.5 then floor" base sampled one
texel OUTSIDE the prim on top/left edges; where that neighbour was
empty VRAM (e.g. tiled menu backgrounds with u0=0) the whole edge
column/row discarded — the vertical line on Tomba's title and the
cross seams on the memory-card screens.
The software renderer's bilinear gets the same reformulation so GL and
SW render filtered output identically; the SW nearest path (shipping
default) is untouched.
Also: always-on coherency event ring (flushes, fills, copies, draw
bboxes, packs, readbacks, presents, probe perturbations, with rects +
frame numbers) queryable via the debug server, and gl_vram_diff now
forces a full pack so it reads FBO truth even where the raw-mirror
invariant is broken.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Move the remaining 2D primitives (flat/textured rects, fill, lines) onto
the GPU, and present directly from the FBO color texture cropped to the
display region instead of reading the whole 1 MB FBO back to CPU VRAM
every frame.
Why both at once: the native present read the `vram` array directly and
relied on an incidental software op each frame (the HUD/sprites, drawn
last) to trigger the FBO->CPU sync. Moving every primitive to the GPU
removed that trigger, so a fully-GPU frame left `vram` stale -> black
present (the earlier FMV->black regression). The fix is to present from
the FBO when the GPU holds the freshest frame (gl_renderer_have_gpu_frame)
and to sync explicitly (gl_renderer_sync_cpu) on the software/24-bit path.
Result: the per-frame readback is gone for GPU scenes (the forest-lag
source), software/FMV frames keep the CPU present path (verified: the
MDEC intro renders correctly), and loads are faster.
KNOWN-FOLLOWUP cosmetic artifacts to clean up: the GPU textured shader
doesn't yet implement the texture window (wrap/mask), so tiled/wrapped
sprites show seams; bilinear texture_filtering isn't honored on the GPU
path; and the debug screenshot still reads CPU VRAM (stale under the
FBO-present path) so captures can lag the on-screen image.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The opengl backend now creates a genuine OpenGL context and presents through
it. Verified: Tomba boots and renders gameplay correctly via GL (context
"4.6.0 NVIDIA"), pixel-identical to software.
- gpu_gl_renderer.c: gl_renderer_init_context (SDL_GL context on a
SDL_WINDOW_OPENGL window), gl_renderer_present (upload ARGB image -> GL
texture -> full-screen quad -> SDL_GL_SwapWindow), present_blank, shutdown.
Legacy GL 1.x (immediate mode) so no GL loader needed yet (-lopengl32).
Backend vtable still delegates rasterization to sw_* this phase, so the GL
path is pixel-identical — this isolates/validates the context+present.
- main.cpp: window gains SDL_WINDOW_OPENGL when renderer=opengl; GL context
created with graceful fallback to the SDL_Renderer path on failure; present
+ disabled-frame + cleanup branch on g_gl_active. SW path untouched.
- runtime.cmake: link opengl32 (Win) / OpenGL::GL (other).
Phase 2b (next): move rasterization onto the GPU — VRAM as a GL texture/FBO,
modern-GL loader + shaders, primitives at internal resolution (the actual
scaling-without-perf-loss win).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Matthew Stanley