Docs / Performance

Performance#

IEC plugin polling advances a scalar clock and skips work while line levels are unchanged and no device timer is due. Changed-line dispatch uses a dense array with index loops, no iterator or event objects are allocated per cycle. Plugin UI and printer page rendering stay outside the emulation loop.

This is a cycle-accurate C64: every master cycle it clocks the VIC-II, the 6510, two CIAs, the SID shadow voices and (when attached) the 1541's 6502. That per-cycle fidelity, not demo content, is what sets the cost, so the performance story is mostly about keeping the always-on per-cycle machinery cheap and allocation-free. Presentation switches live in src/switches.js; the rendering pipeline is summarised in the master overview ("Rendering pipeline") and detailed in the VIC-II §14.

1. Cost model#

A PAL frame contains 19,656 master cycles and has about 20 ms to complete. Across ten sampled demo runs, main-thread time was distributed as follows:

VIC-II capture, rendering and sprite work dominate. Most per-cycle work also runs on a static screen, so reducing hot-path cost matters more than optimizing individual effects. SID synthesis runs in the audio worklet. A second SID adds shadow and audio work only when enabled, while idle drive waits are skipped.

Monitor adds separately bounded collection and analysis costs:

Cycle observers use preallocated typed arrays and fixed nullable hooks, without per-cycle allocation or method wrapping. Storage rates use opt-in scalar counters. Closing or hiding Monitor stops sampling, cancels recording and detaches counters. Pause keeps the last complete recording and stops periodic snapshot copies and view rebuilding. Hidden sections do not repaint.

Sampled histories retain at most eight seconds and 401 observations. Main view presentation is capped at 10 Hz, retaining RAM changes between presentations. SID numeric text refreshes at most four times per second while running, with stable explanations and immediate paused readings. SID activity plots refresh at most ten times per emulated second, with prose every two seconds. Static screen and character/bitmap thumbnails redraw at most five times per second. The screen-byte grid redraws with visible Memory presentation, including text and fact rows. Its eight current-pointer sprite shapes follow every shared sample using reusable pixel buffers. Captured sprite cutouts are prepared once per shared frame when Memory is viewed. Rate windows retain 52. Tape decoding runs only on its visible page, capped at 32,768 source bytes per observation and 256 retained pulses. Replay polling remains at least 10 Hz independently of the selected live sampling rate.

A local Node/V8 Nine run measured median snapshot collection at 0.53 ms, versus 6.95 ms for a normal frame. At 10 samples/s, collection alone costs about 5.3 ms per second. These figures exclude UI rendering and audio deadlines. Repeat with node investigation/monitor-joined-sampling.mjs, and measure recording separately with node investigation/monitor-capture-cost.mjs. A detailed capture owns about 8.59 MiB across 99 buffers, including analysis and scratch arrays. The 8 MiB ceiling applies to provenance, not the whole capture. Fresh buffers are allocated per recording. Reusing them needs explicit ownership because inspection links can retain older frames. A synthetic Node/V8 CPU-loop probe measured about 6 ms/frame normally and 15–18 ms/frame across the two-frame capture window after warm-up. The first window averaged 30 ms/frame. A V8 trace showed core deoptimization at the first capture transition, without the same core pattern on the next three captures. These are local diagnostics, not browser or mobile latency guarantees.

See Monitor architecture for ownership and lifecycle.

2. Standing optimizations#

The VIC optimizations are permanent, with automatic live and diagnostic paths. WebGL and CRT presentation switches remain in src/switches.js: use ?NAME=0 in the browser or NAME=0 in Node harnesses.

• Automatic line batching. Defers a raster line's pixel emission and replays it in one burst, coalescing runs of unchanged cycles into wide segments through the same segment renderer; any mid-line event the CPU could observe triggers an immediate catch-up, so the output is byte-identical to per-cycle rendering. Tracing and armed collision IRQs select the live path. See the VIC-II §14.

• Compact render history. Register, matrix/color and sprite payloads record new snapshots only when their source versions change. Each cycle retains three byte indices instead of eleven snapshot references. Diagnostic array access materializes the dense view. Sprite snapshots use one contiguous 96-byte copy with typed field views; diagnostic recording retains separate field copies.

• Sprite interval scheduling. Deferred sprites with stable inputs wait for their next output interval. Input changes wake them immediately; collision drains and end-of-line passes keep their cycle ordering. Live rendering keeps the reference dispatch path.

• WebGL presenter (WEBGL_PRESENTER, default on). Uploads the finished framebuffer as a single texture per displayed frame instead of putImageData, saving a per-frame convert+upload on the main thread; most valuable on mobile GPUs. With CRT off the output is byte-exact (1:1 store, CSS scales it), and it falls back to the 2D path automatically when WebGL is unavailable; ?WEBGL_PRESENTER=0 forces the 2D path.

• CRT shader (CRT_SHADER, default on). Draws the CRT presets in the presenter's fragment shader at device resolution: four texel fetches per pixel, no per-frame allocation. Replaces the CSS path's filtered canvas under two blended overlays. ?CRT_SHADER=0 forces the CSS overlays, which browsers without WebGL and software rasterisers get anyway; ?CRT_SHADER_SOFTWARE=1 runs the shader on a software rasteriser (tooling).

• Prebuilt CPU opcode / micro-op table (always on). Each opcode's micro-op program is built once at construction and aliased by reference at dispatch, instead of allocating a fresh set of per-cycle closures every instruction. This removes what would otherwise be the largest source of hot-loop allocation. See the 6510 CPU §4.

• Allocation-free hot paths. The VIC per-cycle path reuses pooled scratch and flat typed arrays (pending border-FF transitions, sprite state, segment scratch) and aliases unchanged capture buffers rather than re-copying them through versioned capture and selective rendering; the interrupt sequence and the JAM-halt program are prebuilt like opcodes. In steady state the hot loop allocates essentially nothing. See the VIC-II §14 and the 6510 CPU §4.

• Cheap, monomorphic per-cycle work. Beyond avoiding allocation, the always-on path avoids dead work and unstable object shapes, both nearly free on desktop V8 but real cost on a phone (§4):

  • the VIC segment renderer scopes its per-column span-buffer fills to the few columns actually painted rather than the full 384-slot width;
  • graphics collision and sprite priority share one foreground buffer, so graphics rendering and fixups write each foreground value only once;
  • the VIC shares its BA sample after memory fetches with history capture, and its canonical AEC result with CPU arbitration, avoiding repeated bus-state calculations within the cycle;
  • the CIA read/peek path computes a timer's visible value only for the timer registers, not on every keyboard scan or ICR poll;
  • the SID sync stage skips its three oscillator probes when no voice has hard-sync enabled;
  • the two CIAs null-initialise their port callbacks in the constructor so both instances share one hidden class, keeping the shared hot I/O methods monomorphic.

• Lean audio-thread hot loop. The default reSID engine keeps the worklet branch-light and monomorphic:

  • an all-integer per-cycle pipeline: no libm in the loop; the transistor filter/mixer/volume stages are pre-solved lookup tables;
  • model-dependent tables and DAC references cached on the voice/filter objects.

  The heavyweight lever is the selectable WASM engine: the same model compiled from the Rust translation in rust/sid/, with bit-identical output. It renders a whole 128-sample block per call at ≈100 ms CPU per emulated second against ≈660 ms for the JS engine, which takes worklet-deadline pressure off phones entirely. See the SID §9.

• Drive idle-skip (automatic). When the IEC bus has been quiet long enough and the drive CPU is parked in a known ROM idle loop with the spindle stopped, the machine fast-forwards it to the next timer wake instead of clocking every cycle. Menus and load waits cost almost nothing; an actively streaming loader keeps the drive live. See the 1541 drive §10.

• Lazy three.js chunk. The 3D "Retro Vibes" scene and the model viewer import three.js only on first open, so the library stays out of the initial bundle. See Retro Vibes.

• Retro Vibes render reuse. Bedroom and Foggy Spotlight retain static shadow maps until casting geometry or the shadow rig changes. IK+ and Synthwave update their reflections at 30 Hz, rendered before the main pass so lit materials skip three's per-frame program re-check. IK+ wave shading keeps display cadence. IK+ merges its still, opaque scenery per material (mergeStatic), and the swaying maple within itself: about 170 draws a frame instead of 390, and half the per-draw uniform garbage. Its shadow maps refresh at 10 Hz (shadowHz), since only that slow sway moves a caster. The screen provider reuses its descriptor, and Starry pools its meteor geometry and material. See Retro Vibes §§4–5.

• Foggy Spotlight's extra passes. Smoke and contact shadows are built only off low-power devices, and the fog switch turns the smoke off. The smoke marches at most 800 px wide and the occlusion runs at half resolution with 8 samples. Depth of field is not used: it redrew the scene for its depth every frame, most of the scene's frame time, for a blur that barely showed. The floor's baked light is 1024 px and its detail a repeated 1024 px tile, so the scene adds about 23 MiB of textures and 448 KiB of downloads. Once warm, its per-frame code allocates nothing of its own: values that change every frame go into vector uniforms set in place, and the depth-pass hide list is filled by index rather than pushed and cleared.

The true PAL drive clock ratio and IEC read-side propagation are permanent parts of the hardware model. Disk writes respect each mounted image's write-protect state. See the 1541 drive and machine orchestrator. Mechanical timing experiments remain compile-time constants in drive1541.js; diagnostic console controls are documented in TESTING.

Assembly64 requests#

Catalog work stays outside the emulation loop. Each tab permits three concurrent requests, spaces request starts by 250 ms, shares identical in-flight JSON calls, and honors Retry-After without automatic retries. Search pages are cached for 30 seconds; presets, categories, metadata and file lists for five minutes. The in-memory cache is bounded to 64 responses and 4 MiB. Binaries are streamed with size limits and are never cached by the API client or service worker.

D71 and D81 images use the virtual drive with TDE off. Their filesystem work runs on file/channel operations, without adding per-cycle drive emulation.

3. Throughput & footprint#

The figures below are approximate and machine-dependent; treat them as orders of magnitude, not benchmarks.

• Throughput. Headless (machine.runFrame() with no presentation), the core runs comfortably above PAL realtime on an Apple M1-class machine; very roughly 1.5–3× depending on content, a near-static screen fastest and a sprite-heavy full-frame demo slowest. In the browser, presentation (the WebGL texture upload) and SID synthesis run on separate GPU/audio threads on top of this main-thread emulation cost.
• Garbage is negligible, and the JIT settles. A running demo allocates roughly 300 bytes per frame, a few kilobytes a second, costing well under a millisecond of collection per thousand frames. The optimiser's picture is equally quiet: it gives up on a handful of functions while warming up, then holds, with no recurring deoptimisation once a demo is underway.
• Memory is small and flat. All ArrayBuffers reachable from the machine total ~1.7 MB: the ~1 MB SharedArrayBuffer SID event ring (128k cycle-stamped entries), the ~408 KB framebuffer (384×272 RGBA), the 64 KB C64 RAM, and the per-cycle capture arrays. There is no per-frame growth; retained heap stays flat across long runs, with no leaks. Process RSS is dominated by the JS engine itself, not by these buffers.
• Bundle. three.js (~700 kB) is code-split into the lazy chunk described in §2, keeping the initial page load lean.

4. Platform & mobile#

Desktop V8 has generous heaps and an optimiser that erases most short-lived allocation, so garbage collection is a non-issue there. Phones are less forgiving:

• Tighter heaps and GC. JavaScriptCore (iOS Safari) does not eliminate short-lived allocation the way V8 does, and mobile nurseries are far smaller, so allocation that is effectively free on the desktop becomes real, main-thread GC pressure on a phone. This is exactly why the CPU dispatch, the VIC bookkeeping and the sprite paths are kept allocation-free (§2), a win that is nearly invisible in a desktop time profile but material on device.
• Shape and dead work, not just allocation. The same desktop-invisible, mobile-material asymmetry applies beyond GC: JavaScriptCore also pays for polymorphic property loads and per-cycle work that cannot affect the result, where V8's optimiser hides both. Keeping the always-on objects monomorphic (one stable hidden class per hot type) and their hot readers free of provably-dead work is therefore a mobile lever in its own right, one that, like allocation avoidance, barely registers in a desktop time profile.
• Throttling and pressure. Slower cores, thermal throttling and background memory pressure make mobile frame times less predictable; the heaviest full-frame demos may dip below realtime on a phone even where a laptop has ample headroom. The WebGL presenter (§2) is chosen partly to spare mobile GPUs the putImageData path.
• UI-thread jank on the play path. The on-screen touch joystick, the primary mobile input, does no synchronous layout, per-event allocation, or redundant control-port work while dragging: geometry is cached on pointerdown, the knob is a compositor-promoted transform coalesced to 60 Hz, and joystick-byte synchronization waits for the 50 Hz emulation-frame boundary. The base page is sized in dvh so iOS Safari's dynamic toolbar doesn't shift it, and the paint-driven CRT effects stop animating while the machine is paused.
• Background pause. When the tab or app loses focus or is backgrounded, emulation freezes and the audio context is suspended, then thaws on return, so a backgrounded tab neither burns CPU nor spikes when it comes back.

5. Requirements#

The SID audio ring is a SharedArrayBuffer, so the page must be cross-origin isolated. That requires two response headers on the document (and its assets):

Cross-Origin-Opener-Policy: same-origin
Cross-Origin-Embedder-Policy: require-corp

Without them SharedArrayBuffer is unavailable and the machine constructor throws. The Vite dev and preview servers set these headers automatically; a production host must send the same. See the machine orchestrator §7.

6. Measurement Harnesses#

Never trust a performance comparison across a different thermal window. On an M1-class laptop, a heavy build, test run, or screenshot pass can change the next number enough to make a neutral change look like a large regression. Run a same-thermal A/B, back-to-back:

1. Measure the candidate.
2. Stash only the files under test.
3. Measure the baseline.
4. Pop the stash and measure the candidate again.

Two traps are worth naming, because both produce confident nonsense:

• A sampling heap profile measures what survived, not what was allocated. Those profilers hold their samples weakly, so anything already collected is simply gone from the report; profile a longer run and the total can come out smaller. To measure how much garbage the machine makes, account for it at the collections themselves.
• Separate warm-up from steady state by changing the run length. A count that lands identically at three thousand frames and at twelve thousand is start-up cost, however alarming its size; only a figure that grows with the run is something the emulator pays for continuously.

The desktop V8 harness is tools/perf-workloads.mjs:

node --expose-gc tools/perf-workloads.mjs <workload> <mode> [frames]

The demo files the workloads load come from test/external-assets.json (orbit-untold-prg, raster-time-demo, the c64stuff collection); edit the paths there or set the per-entry environment variable.

Workloads are idle, orbit, rastertime, comaload, and comarun; modes are time, prof, alloc, allocsites, mem, and all. time is useful for throughput, but V8 can escape-analyze short-lived allocations away. Use allocsites to see allocation a non-EA engine, such as JavaScriptCore on iOS, would still pay for.

Compact render history, sprite interval scheduling and immutable graphics fetches are permanent parts of the renderer. Tracing and diagnostic history access retain dense recording; live rendering retains per-cycle sprite dispatch. Cartridge and tracing lines use the RAM-reading graphics path. Both graphics sources use the combined color/foreground decoder. Active fetches record packed samples once; live/deferred rendering and corrections consume them without rereading graphics RAM. RAM/DMA and fetch-configuration writes no longer force those lines to render early. Collision observers retain catch-up. The disk-trap workload explicitly disables true-drive mode and rejects a run whose pending LOAD/RUN never completes.

An AC-powered candidate/baseline/candidate check of compact history, sprite intervals and the interior pixel-copy path measured these median frame times (Node v25.6.0, 9 batches of 300 frames each):

These are modest throughput gains, about 3–5%, for the combined change, not isolated attribution to each optimization. Desktop WebKit idle measured 5.051 ms baseline against 4.996 / 4.999 ms candidate, effectively neutral. These figures do not establish a phone performance gain.

Next Round's first 6,000 PAL frames after RUN (true drive, SID 8580, loading excluded) were compared on frozen common sources with one change per candidate/baseline/candidate triplet. All 39 timed runs matched pixel and sampled-state hashes across Node 25.6.0 and desktop WebKit. AC power was used with low-power mode off; an inconsistent fetch triplet under background CPU load was repeated. Approximate reductions in mean frame time were:

These incremental gains are not additive. WebKit used JSC_useFTLJIT=false, verified in its option dump, with no extra warm-up beyond boot/loading. This limits the JIT tiers available; it does not emulate phone hardware or measure input latency. Input sampling, frame presentation scheduling and emulated cycle order are unchanged by contiguous copying.

V8 allocation sampling across all configurations measured 0.17–0.23 KiB of VIC allocation per frame, including runs with escape analysis disabled. There was no sustained allocation or deoptimization problem attributable to contiguous copying. Occasional missing-feedback bailouts occur as new sprite branches are reached during the demo; these are not JSC allocation or deoptimization results.

The combined package (sparse history, sprite intervals, contiguous copies and fetch-fed graphics using the combined decoder) was also measured together against dense history, per-cycle sprites and RAM-based graphics on the same source. Next Round used the same 6,000-frame window and candidate/baseline/candidate procedure:

All six runs matched every pixel and sampled-state hash. These package gains are measured directly, not added from the individual comparisons. WebKit's FTL tier was disabled in all three runs. The measurements exclude host input, presentation and audio synthesis, and do not establish actual phone latency. Fetch-fed graphics is permanent; tracing and cartridges retain the RAM-reading path.

The Safari/JSC harness is tools/jsc-perf.mjs plus tools/jsc-perf-harness.html:

npx playwright install webkit          # one-time
node tools/jsc-perf.mjs "label"

It starts a throwaway static server with the COOP/COEP headers required by the SID SharedArrayBuffer, launches Playwright WebKit, boots to READY, and times 9×300-frame idle batches through the real unbundled modules. Treat desktop WebKit time as only part of the story: allocation reductions can be neutral on a desktop core with an ample nursery and still remove visible jank on a phone.