[meh7an]

What

This PR builds on #10423 (masked result memo). On a masked hit there, materializePreparedMemoized makes a fresh copy of the donor result (std::make_shared<RecompileResult>(*donor)) and puts it into the exact table. The insert evicts the set's least recently used entry, and when that drops the evicted result's last reference, the result is destroyed on the worker, under the source's lock.

A RecompileResult holds dozens of vectors, and every DescriptorBinding in it holds about twenty more. So each patched hit costs one full allocation of a result and one full free of another.

On Sonic Origins Plus's Sonic 2 menu page with #10423 applied, the result memo evicts about 31,000 results per second. Sampling the queue worker (APS5_SAMPLE_WORKER, 80 s on the page) attributes:

• 13 % of its samples to destroying results, most of them in that eviction;
• 15 % to operator delete, mostly the Windows heap's frees;
• 4 % to the copy.

This PR keeps the storage instead:

• When the memo evicts a result that nothing else references (use_count() == 1 under the source's lock, followed by an acquire fence), it moves the result into a small list owned by the calling thread (ThreadOwned, at most 8 results) instead of freeing it.
• The next masked hit on that thread copy-assigns the donor into a kept result (*result = donor) instead of allocating one. The vectors keep their capacity, so most of the copy allocates nothing.
• applyMemoPatch then writes the snapshot's words as before. ResultKeyMemo's assignment already clears the draw-key memo, so nothing of the old result survives.

Nothing holds a weak_ptr to a RecompileResult, so a result whose count is 1 under the lock cannot gain a new owner while it is reused. Every other owner (a recorded draw, the other table) keeps the count above 1, and that result is freed as before when its last owner drops it.

• APS5_NO_RESULT_RECYCLE=1 frees every evicted result, as perf(shader): reuse a memoized result by patching its copied words #10423 does.
• ShaderRecompiler::RecycledResults() counts the reuses.
• agc_shader_table_snapshot_reuse now ends by patching the program for 640 distinct output buffers, enough to fill the memo and evict.
  • Each patched result must hold its own buffer's descriptor, recycled storage or not.
  • The counter must rise exactly when recycling is on.
  • agc_shader_table_snapshot_reuse_unrecycled (new) runs the test with APS5_NO_RESULT_RECYCLE=1. It expects no reuse, and the two existing variants (default and _unmasked) still pass.

Overlap.

• tools/pr_overlap.py lists perf(shader): reuse an evicted memo result's bindings for the next miss #9049, perf(agc): GPU descriptor heap on VK_EXT_descriptor_heap (APS5_GPU_HEAP) #9508 and fix(shader): bound the bytes the recompiler's result memos hold #9989. Their conflicts are with lines perf(shader): reuse a memoized result by patching its copied words #10423 changes in Recompiler.cpp (the memo entries and materializePreparedMemoized). There is also the ENVIRONMENT_MODIFICATION test list in the AGC CMakeLists.txt, where perf(shader): reuse a memoized result by patching its copied words #10423 and this PR add test names and perf(agc): GPU descriptor heap on VK_EXT_descriptor_heap (APS5_GPU_HEAP) #9508 edits the same line.
• Of the PRs it could not compare, feat(agc): look up table images at run time through a device image table on the GPU heap (APS5_GPU_PATH) #7642 and feat(agc): serve lane-keyed, pointer and run-time-address image table columns through the device image table #10193 add a line to the miss path of materializePreparedMemoized, which perf(shader): reuse a memoized result by patching its copied words #10423 rewrites. The result fields they add are copied by the copy-assignment like every other member.
• Proposed order: perf(shader): reuse an evicted memo result's bindings for the next miss #9049 and fix(shader): bound the bytes the recompiler's result memos hold #9989 go first (comment, comment). After they merge, perf(shader): reuse a memoized result by patching its copied words #10423 and this PR rebase onto them and take on the integration:
  • one reuse path for evicted results, feeding perf(shader): reuse an evicted memo result's bindings for the next miss #9049's miss path from the pool;
  • the masked table and the recycle pool counted in ResultMemoBytes() and trimmed with the memo, with a test.

Tested

Windows 11 Pro 10.0.26300, NVIDIA GeForce RTX 2080 Ti, driver 617.14, i9-9900K.

Sonic Origins Plus (PPSA05324), Sonic 2 page of the main menu, main's defaults. Two clean builds on main e7f46f5: #10423's head 54d7c78 and this PR's head bdca440. Each got one warm-up boot, then two interleaved traced rounds (APS5_TRACE_SECONDS=200:15):

	frame p50 (ms)	p90 (ms)	Driver.Packet (ms/frame)	Driver.Draw (ms/frame)	title FPS (median of six)
#10423 54d7c78	80.1 / 85.1	86.8 / 116.7	80.1 / 89.2	59.6 / 68.0	12.21 / 12.70
this PR bdca440	84.6 / 80.1	116.3 / 87.7	88.3 / 80.5	65.6 / 59.9	12.53 / 12.55

• No frame-time difference within this page's noise. The page switches between a fast mode (p90 about 87 ms) and a slow one (about 116 ms), and each build landed in each mode once. Averaged, frame p50 goes 82.6 -> 82.4 ms and Driver.Draw 63.8 -> 62.8 ms.

• Why: the queue worker now waits on the draw thread. A sampled profile of the queue worker (APS5_SAMPLE_WORKER, 80 s) shows:

  • materializePreparedMemoized going from 19 % to 14 % of its samples and operator delete from 15 % to 9 %;
  • its wait in DrawPipeline::Enqueue on a full draw queue growing from 11 % to 19 %.

  The CPU this PR saves turns into waiting, so on this page it cannot show as frame time until the draw thread gets faster. I'm keeping the PR as a draft until then and will re-measure.

• The earlier flag A/B was noise. It ran APS5_NO_RESULT_RECYCLE=1 against the default in one uncommitted build (the traces say 2ef1dd3beb, dirty) and read p50 89.2 / 84.2 -> 81.8 / 82.8 ms. Its first "off" round landed in the slow mode, which this clean rerun shows is noise. Its traces: recycle-s2-round1.zip, recycle-s2-round2.zip.

• No mismatches. APS5_VERIFY_MASKED_RESULT_MEMO=1 with recycling on: 0 patch mismatches over about 440,000 patched results.

Tests

• ctest (2064 tests) on main e7f46f5 plus perf(shader): reuse a memoized result by patching its copied words #10423 plus this PR: everything passes except agc_driver_sample_order, agc_driver_sample_order_grid_reference and agc_driver_sample_order_recipe_depth_locations, which fail the same way on plain main e7f46f5 on this machine (as in perf(shader): reuse a memoized result by patching its copied words #10423).
• ctest --test-dir build -R "^agc_shader_|^agc_driver_draw_cache" -V with ANYPS5_REQUIRE_VULKAN=1, ANYPS5_NO_SHADER_CACHE=1 and VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation: 48/48 on main e7f46f5, 49/49 with perf(shader): reuse a memoized result by patching its copied words #10423, and 50/50 with this PR (the extra one is agc_shader_table_snapshot_reuse_unrecycled). None of the logs has was not found but was requested by env var VK_INSTANCE_LAYERS or a VUID- line.

The summaries of all four clean rounds follow, then the full digests of round 1. All four traces are attached:

recycle-clean-s2-round1.zip
recycle-clean-s2-round2.zip

round	build	summary
1	#10423 54d7c78	seconds 200:15 from present 3354: 182 presents, frame p50 80.1 / p90 86.8 / p99 116.5 / max 137.9 ms, GPU busy 74.6 ms per present
1	this PR bdca440	seconds 200:15 from present 3333: 165 presents, frame p50 84.6 / p90 116.3 / p99 192.6 / max 193.5 ms, GPU busy 74.7 ms per present
2	#10423 54d7c78	seconds 200:15 from present 3427: 164 presents, frame p50 85.1 / p90 116.7 / p99 126.9 / max 128.3 ms, GPU busy 77.4 ms per present
2	this PR bdca440	seconds 200:15 from present 3408: 181 presents, frame p50 80.1 / p90 87.7 / p99 114.6 / max 115.5 ms, GPU busy 74.3 ms per present

Round 1, #10423 54d7c78

seconds 200:15 from present 3354: 182 presents, frame p50 80.1 / p90 86.8 / p99 116.5 / max 137.9 ms, GPU busy 74.6 ms per present
NVIDIA GeForce RTX 2080 Ti, NVIDIA 617.14, Windows 10.0.26300, 3840x2160, commit 54d7c78, dropped: yes

{"commit":"54d7c780a583","dirty":false,"env":{"APS5_TRACE_DIR":"<path>","APS5_TRACE_SECONDS":"200:15"},"gpu":"NVIDIA GeForce RTX 2080 Ti","driver":"NVIDIA 617.14","os":"Windows 10.0.26300","cpu":"Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz","cpu_threads":16,"memory_mib":32684,"title_id":"PPSA05324","title":"Sonic Origins","content_version":"02.002.000","resolution":"3840x2160","window":{"spec":"seconds 200:15","first_present":3354,"ms":15050.577},"presents":182,"frame_ms":{"p50":80.127,"p90":86.84,"p99":116.533,"max":137.889},"gpu_busy_ms_per_present":74.62,"worker_busy_pct":{"queue worker 0x0":97.001},"top_phases_ms_per_frame":{"Driver.Packet":80.115,"Graphics.Draw":62.677,"Driver.Packet.draw":60.628,"Driver.Draw":59.584,"Graphics.Draw.draw_and_resource_release":52.265,"Driver.Draw.shader_compile_and_link":37.047,"VideoOut.Flip":32.487,"Driver.Present":32.397},"trace_dropped":true,"trace_dropped_threads":{"queue worker 0x0":{"events_full_after_s":6.102,"events_lost":386130,"totals_full":false},"draw committer":{"events_full_after_s":12.778,"events_lost":47537,"totals_full":false}},"trace":"aps5-trace-54d7c780a583-9420.json.gz"}

Slowest frames (largest slices inside each, inclusive)

• present 3373: 137.9 ms at 1605 ms: batch (gpu) 131.7, Driver.Packet (queue worker 0x0) 96.4, Graphics.Draw (draw committer) 90.7, Graphics.Draw.draw_and_resource_release (draw committer) 85.4
• present 3451: 116.5 ms at 7947 ms: batch (gpu) 75.4, Graphics.Draw (draw committer) 68.6, Graphics.Draw.draw_and_resource_release (draw committer) 63.6, VideoOut.Flip (presenter) 32.3
• present 3403: 113.4 ms at 4064 ms: batch (gpu) 75.0, Driver.Packet (queue worker 0x0) 72.2, Graphics.Draw (draw committer) 66.5, Graphics.Draw.draw_and_resource_release (draw committer) 62.0
• present 3495: 113.1 ms at 11690 ms: batch (gpu) 75.1, Graphics.Draw (draw committer) 68.8, Graphics.Draw.draw_and_resource_release (draw committer) 62.8, VideoOut.Flip (presenter) 31.6
• present 3485: 113.0 ms at 10850 ms: batch (gpu) 74.2, Graphics.Draw (draw committer) 68.1, Graphics.Draw.draw_and_resource_release (draw committer) 61.9, VideoOut.Flip (presenter) 30.0

Threads (ms per frame)

thread	busy	wait	untraced	in slices < 20 us	largest
title	1.4 (2 %)	0.0 (0 %)	81.3	0.7	Shader.ResolveGraphicsStagesAbi 0.9, Shader.ResolveAbi 0.5
queue worker 0x0	17.5 (21 %)	0.0 (0 %)	65.2	75.6	Driver.Packet 17.5, Driver.EndSubmission 0.0
draw committer	33.3 (40 %)	0.6 (1 %)	48.8	48.3	Graphics.Draw 31.7, Recorder.Submit 1.1
title	0.4 (0 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.1
presenter	0.7 (1 %)	31.9 (39 %)	50.0	0.0	VideoOut.Flip 32.5, Driver.Present 0.2
title	0.5 (1 %)	0.0 (0 %)	82.2	0.1	Shader.ResolveGraphicsStagesAbi 0.3, Shader.ResolveAbi 0.2
title	0.3 (0 %)	0.0 (0 %)	82.4	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.1
title	0.4 (0 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (1 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (1 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (1 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	82.3	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.5 (1 %)	0.0 (0 %)	82.2	0.1	Shader.ResolveGraphicsStagesAbi 0.3, Shader.ResolveAbi 0.2
gpu	74.5 (90 %)	0.0 (0 %)	8.2	0.0	batch 74.5

Largest slices (ms per frame, inclusive, all threads): batch 74.5, VideoOut.Flip 32.5, Driver.Present 32.4, VideoOut.Flip.present 32.2, Driver.Present.render_fence_wait 31.9, Vulkan.Retire 31.9, Vulkan.Retire.wait 31.9, Graphics.Draw 31.7, Graphics.Draw.draw_and_resource_release 26.1, Driver.Packet 17.5

Longest waits

• Driver.Present.render_fence_wait (presenter) 87.9 ms at 1417 ms, present 3371
• Driver.Present.render_fence_wait (presenter) 82.5 ms at 1330 ms, present 3370
• Driver.Present.render_fence_wait (presenter) 78.7 ms at 1605 ms, present 3373
• Driver.Present.render_fence_wait (presenter) 70.6 ms at 1524 ms, present 3372
• Driver.Present.render_fence_wait (presenter) 68.2 ms at 1248 ms, present 3369
• Driver.Present.render_fence_wait (presenter) 60.5 ms at 1160 ms, present 3368
• Driver.Present.render_fence_wait (presenter) 52.3 ms at 1074 ms, present 3367
• Driver.Present.render_fence_wait (presenter) 40.8 ms at 992 ms, present 3366
• Driver.Present.render_fence_wait (presenter) 35.8 ms at 7399 ms, present 3444
• Driver.Present.render_fence_wait (presenter) 35.4 ms at 11605 ms, present 3494

Largest GPU batches (GPU ranges inside, and the CPU slices that recorded each, inclusive)

• batch 1354: 17.6 ms at 1649 ms, present 3373: GPU no ranges; recorded by Driver.Packet (queue worker 0x0) 1.3, Driver.Packet.dispatch (queue worker 0x0) 0.6, Driver.Dispatch (queue worker 0x0) 0.5, Driver.Packet.pipeline_drain (queue worker 0x0) 0.3
• batch 7443: 15.5 ms at 8544 ms, present 3457: GPU no ranges; recorded by no CPU slices
• batch 10023: 15.3 ms at 11604 ms, present 3493: GPU no ranges; recorded by no CPU slices
• batch 9453: 15.3 ms at 10961 ms, present 3485: GPU no ranges; recorded by no CPU slices
• batch 3280: 15.0 ms at 3818 ms, present 3399: GPU no ranges; recorded by Driver.Packet (queue worker 0x0) 1.3, Driver.Packet.dispatch (queue worker 0x0) 0.6, Driver.Dispatch (queue worker 0x0) 0.5, Driver.Packet.pipeline_drain (queue worker 0x0) 0.4

Round 1, this PR bdca440

seconds 200:15 from present 3333: 165 presents, frame p50 84.6 / p90 116.3 / p99 192.6 / max 193.5 ms, GPU busy 74.7 ms per present
NVIDIA GeForce RTX 2080 Ti, NVIDIA 617.14, Windows 10.0.26300, 3840x2160, commit bdca440, dropped: yes

{"commit":"bdca440d390c","dirty":false,"env":{"APS5_TRACE_DIR":"<path>","APS5_TRACE_SECONDS":"200:15"},"gpu":"NVIDIA GeForce RTX 2080 Ti","driver":"NVIDIA 617.14","os":"Windows 10.0.26300","cpu":"Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz","cpu_threads":16,"memory_mib":32684,"title_id":"PPSA05324","title":"Sonic Origins","content_version":"02.002.000","resolution":"3840x2160","window":{"spec":"seconds 200:15","first_present":3333,"ms":15011.098},"presents":165,"frame_ms":{"p50":84.575,"p90":116.336,"p99":192.64,"max":193.492},"gpu_busy_ms_per_present":74.693,"worker_busy_pct":{"queue worker 0x0":97.218},"top_phases_ms_per_frame":{"Driver.Packet":88.322,"Graphics.Draw":69.74,"Driver.Packet.draw":66.729,"Driver.Draw":65.594,"Graphics.Draw.draw_and_resource_release":58.618,"Driver.Draw.shader_compile_and_link":38.525,"VideoOut.Flip":30.932,"Driver.Present":30.913},"trace_dropped":true,"trace_dropped_threads":{"queue worker 0x0":{"events_full_after_s":6.173,"events_lost":385570,"totals_full":false},"draw committer":{"events_full_after_s":13.162,"events_lost":38828,"totals_full":false}},"trace":"aps5-trace-bdca440d390c-24144.json.gz"}

Slowest frames (largest slices inside each, inclusive)

• present 3344: 193.5 ms at 947 ms: batch (gpu) 191.9, Driver.Packet (queue worker 0x0) 155.2, VideoOut.Flip (presenter) 122.9, VideoOut.Flip.present (presenter) 122.8
• present 3333: 192.6 ms at 0 ms: batch (gpu) 156.2, Driver.Packet (queue worker 0x0) 118.8, Graphics.Draw (draw committer) 79.7, Driver.Packet.draw (queue worker 0x0) 76.3
• present 3390: 133.7 ms at 5212 ms: Driver.Packet (queue worker 0x0) 108.2, Graphics.Draw (draw committer) 95.6, Driver.Packet.draw (queue worker 0x0) 93.9, Driver.Draw (queue worker 0x0) 91.9
• present 3413: 121.8 ms at 7357 ms: Graphics.Draw (draw committer) 90.2, batch (gpu) 77.3, Graphics.Draw.draw_and_resource_release (draw committer) 61.1, VideoOut.Flip (presenter) 32.7
• present 3478: 120.8 ms at 13148 ms: batch (gpu) 73.9, VideoOut.Flip (presenter) 23.8, VideoOut.Flip.present (presenter) 23.7, Driver.Present (presenter) 23.7

Threads (ms per frame)

thread	busy	wait	untraced	in slices < 20 us	largest
title	1.5 (2 %)	0.0 (0 %)	89.5	0.7	Shader.ResolveGraphicsStagesAbi 0.9, Shader.ResolveAbi 0.6
queue worker 0x0	22.2 (24 %)	0.1 (0 %)	68.6	74.1	Driver.Packet 22.3, Driver.EndSubmission 0.0
draw committer	41.0 (45 %)	0.5 (1 %)	49.5	51.2	Graphics.Draw 38.8, Recorder.Submit 1.6
title	0.4 (0 %)	0.0 (0 %)	90.6	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	90.5	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.5 (1 %)	0.0 (0 %)	90.5	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	90.6	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
presenter	0.7 (1 %)	30.4 (33 %)	59.9	0.0	VideoOut.Flip 30.9, Driver.Present 0.2
title	0.5 (1 %)	0.0 (0 %)	90.5	0.1	Shader.ResolveGraphicsStagesAbi 0.3, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	90.6	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.5 (0 %)	0.0 (0 %)	90.5	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	90.6	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	90.5	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
title	0.3 (0 %)	0.0 (0 %)	90.7	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.1
title	0.5 (1 %)	0.0 (0 %)	90.5	0.1	Shader.ResolveGraphicsStagesAbi 0.3, Shader.ResolveAbi 0.2
title	0.4 (0 %)	0.0 (0 %)	90.5	0.1	Shader.ResolveGraphicsStagesAbi 0.2, Shader.ResolveAbi 0.2
gpu	74.7 (82 %)	0.0 (0 %)	16.3	0.0	batch 74.7

Largest slices (ms per frame, inclusive, all threads): batch 74.7, Graphics.Draw 38.8, Graphics.Draw.draw_and_resource_release 31.6, VideoOut.Flip 30.9, Driver.Present 30.9, VideoOut.Flip.present 30.8, Driver.Present.render_fence_wait 30.4, Vulkan.Retire 30.4, Vulkan.Retire.wait 30.3, Driver.Packet 22.3

Longest waits

• Driver.Present.render_fence_wait (presenter) 122.6 ms at 947 ms, present 3344
• Driver.Present.render_fence_wait (presenter) 118.0 ms at 1141 ms, present 3345
• Driver.Present.render_fence_wait (presenter) 71.0 ms at 1260 ms, present 3346
• Driver.Present.render_fence_wait (presenter) 47.5 ms at 1355 ms, present 3347
• Driver.Present.render_fence_wait (presenter) 34.7 ms at 7811 ms, present 3418
• Driver.Present.render_fence_wait (presenter) 34.6 ms at 14200 ms, present 3489
• Driver.Present.render_fence_wait (presenter) 34.5 ms at 6822 ms, present 3407
• Driver.Present.render_fence_wait (presenter) 34.4 ms at 14286 ms, present 3490
• Driver.Present.render_fence_wait (presenter) 34.4 ms at 1438 ms, present 3348
• Driver.Present.render_fence_wait (presenter) 34.2 ms at 8639 ms, present 3427

Largest GPU batches (GPU ranges inside, and the CPU slices that recorded each, inclusive)

• batch 853: 32.2 ms at 1185 ms, present 3345: GPU no ranges; recorded by Graphics.Draw (draw committer) 0.8, Graphics.Draw.draw_and_resource_release (draw committer) 0.7, Resources.Prepare (draw committer) 0.1, Recorder.Submit (draw committer) 0.1
• batch 851: 30.3 ms at 1149 ms, present 3345: GPU no ranges; recorded by Graphics.Draw (draw committer) 1.3, Graphics.Draw.draw_and_resource_release (draw committer) 1.2, Recorder.Submit (draw committer) 0.0, Vulkan.DrawFromRecipe (draw committer) 0.0
• batch 782: 29.8 ms at 975 ms, present 3344: GPU no ranges; recorded by Driver.Packet (queue worker 0x0) 0.5, Driver.Packet.dispatch (queue worker 0x0) 0.3, Driver.Dispatch (queue worker 0x0) 0.3, Vulkan.Dispatch (queue worker 0x0) 0.2
• batch 780: 27.7 ms at 936 ms, present 3343: GPU no ranges; recorded by Graphics.Draw (draw committer) 2.4, Graphics.Draw.draw_and_resource_release (draw committer) 2.2, Recorder.Submit (draw committer) 0.1, Resources.Prepare (draw committer) 0.1
• batch 836: 23.2 ms at 1101 ms, present 3344: GPU no ranges; recorded by Driver.Packet (queue worker 0x0) 0.4, Driver.Packet.draw (queue worker 0x0) 0.2, Driver.Draw (queue worker 0x0) 0.2, Driver.Packet.pipeline_drain (queue worker 0x0) 0.2

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: perf(shader): reuse a memoized result by patching its copied words #10423, perf(shader): reuse an evicted memo result's bindings for the next miss #9049, fix(shader): bound the bytes the recompiler's result memos hold #9989 (proposed order, see Overlap)
• Fixes: none
• Unblocks: none
• GPUs tested: NVIDIA GeForce RTX 2080 Ti, driver 617.14, Windows 11 Pro 10.0.26300
• AI-assisted: yes

[github-actions]

This pull request is labeled perf, so it needs the summary line of a capture on main and one with the change, from the same title, scene and machine, as the Performance report form asks. Please paste both aps5-summary.jsonl lines, attach the two traces and add the output of tools/trace_digest.py.

[MarcZX10]

Reviewed exact head bdca440d390c7b833a14e15ee38664e16a902033 against its stated dependency #10423. The feature-flag comparison and attached traces are useful, but I found integration and validation gaps that need resolution before this can merge:

• The performance summaries compare APS5_NO_RESULT_RECYCLE=1 with recycling enabled in the same dirty build, and both report commit 2ef1dd3bebef. That is a useful feature-flag ablation, but it is not the required two-commit base/change comparison. Please run the same scene on the #10423 dependency head and this exact PR head (or the agreed current integration base and this head), attach both trace captures, and put both distinct commit IDs in the summaries. The hosted product build/test jobs are currently skipped, so they do not supply this comparison.
• Please coordinate the new retention path with #9989. That PR's current budget/trim walk counts source->memo entries; this change adds a second 256-entry source.maskedMemo table that owns full RecompileResults and a per-thread pool retaining up to eight more full results. Those references and their vector capacities are outside the current byte accounting, so the configured memo budget would not bound the retained memory after these changes are combined. Update the accounting/trim behavior and add a test that exercises the bound with both tables and the recycle pool, or agree on an explicit integration order that resolves this before merge.
• Please also reconfirm the interaction with #9049. Its current change reuses an evicted exact entry's bindings only when use_count() == 1; the added masked table can keep a second reference to that same result, while this PR separately recycles the complete result when an entry is evicted. This can change when #9049's reuse path is eligible, contrary to the earlier expectation that the two optimizations add up. Coordinate with that author, declare the agreed dependency/order, and validate the combined path.

I did not run the GPU/AGC tests locally: this checkout has no usable configured full build, and the latest hosted Linux/Windows product jobs are skipped. I therefore have not independently validated runtime equivalence or performance.

[github-actions]

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 1368	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 1395	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 1896	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 2434	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 5021	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Recompiler.cpp
PR 5371	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp
PR 5609	core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp
PR 5925	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Recompiler.hpp
PR 6187	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6188	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Recompiler.hpp
PR 6192	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6209	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6235	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.hpp
PR 6280	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6300	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6412	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp
PR 6481	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6488	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp
PR 6539	core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp
PR 6642	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6808	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6924	core/shader/recompiler/Recompiler.cpp docs/dev/ENVIRONMENT.md
PR 7048	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 7201	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp
PR 7275	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 7449	docs/dev/ENVIRONMENT.md
PR 7457	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 7484	core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 7604	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 7642	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 7684	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp
PR 7689	core/shader/recompiler/Recompiler.hpp
PR 8004	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 8014	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 8077	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/ENVIRONMENT.md
PR 8161	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 8246	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 8484	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 8542	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 8592	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 8896	docs/dev/ENVIRONMENT.md
PR 9176	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 9248	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 9287	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 9289	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 9345	docs/dev/ENVIRONMENT.md
PR 9356	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 9387	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 9402	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 9508	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 9669	docs/dev/ENVIRONMENT.md
PR 9676	docs/dev/ENVIRONMENT.md
PR 9748	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp
PR 9852	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 10193	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md
PR 10442	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/ENVIRONMENT.md
PR 10547	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp core/shader/recompiler/Recompiler.cpp core/shader/recompiler/Recompiler.hpp docs/dev/ENVIRONMENT.md

[meh7an]

@MarcZX10 Thanks, all three points hold, and the first one changed the picture.

• Comparison. I reran the same scene on two clean builds on main e7f46f5: perf(shader): reuse a memoized result by patching its copied words #10423's head 54d7c78 and this head bdca440, one warm-up boot each and two interleaved rounds. The summaries carry both commit IDs, and all four traces are attached. Within this page's noise there is no frame-time difference: the page alternates between a fast and a slow mode, each build landed in each once, and p50 averages 82.6 -> 82.4 ms. The earlier flag A/B's "gain" came from its first "off" round landing in the slow mode. The worker's CPU saving is real in a sampled profile, but on this page it turns into waiting on the draw thread (DrawPipeline::Enqueue 11 % -> 19 % of its samples). So I've converted the PR to a draft until the draw-thread work lands, and will re-measure then. The description now says all of this.
• fix(shader): bound the bytes the recompiler's result memos hold #9989. I proposed that it goes first (comment). Once it merges, perf(shader): reuse a memoized result by patching its copied words #10423 and this PR rebase onto it, count the masked table and the recycle pool in ResultMemoBytes(), trim them with the memo, and extend agc_shader_result_memo_budget to cover both.
• perf(shader): reuse an evicted memo result's bindings for the next miss #9049. I proposed that it goes first too (comment). After it, there is one reuse path for evicted results: the recycle pool, which also feeds perf(shader): reuse an evicted memo result's bindings for the next miss #9049's miss path. The combined path gets validated then.

The checklist now says Depends on: #10423, #9049, #9989, pending both authors' confirmation.

[QDarkRex]

A second title and a second vendor for this change: ASTRO BOT's Crash Site hub on Windows / AMD. The worker-side saving shows up clearly there too. The frame does not move, because on this title the draw committer bounds the frame rather than the queue 0 worker.

• Host: Windows 11 (26300), Ryzen 7 5700X, 32 GB RAM, Radeon RX 9060 XT, AMD driver 26.10.1, WinLibs GCC 15.2.0 r7, Release.
• Builds: this PR's head bdca440 against its parent 54d7c78 (the perf(shader): reuse a memoized result by patching its copied words #10423 commit), so the A/B isolates the second commit.
• Runs: ASTRO BOT PPSA21564 01.007.000, -lvl hub_crashsite, no input, 300 s each. One warm-up per arm, then two rounds in the order A B / B A. APS5_HOST_IMPORT_MIB=14336, APS5_TRACE_SUMMARY_EVERY=10, APS5_SAMPLE_WORKER and APS5_SAMPLE_COMMITTER were set in every run. Frame figures come from the 10 s windows from 120 s on. The worker shares are inclusive, as a fraction of all of the queue 0 worker's samples (about 19,000 per run).

Run	Build	FPS	frame p50	Destroying a RecompileResult	operator delete	materializePreparedMemoized
1	parent	6.43	154.2 ms	5.43 %	5.83 %	10.18 %
2	this PR	6.39	154.8 ms	0.36 %	3.56 %	8.18 %
3	this PR	6.44	154.7 ms	0.42 %	4.27 %	9.34 %
4	parent	6.39	155.4 ms	6.15 %	6.54 %	11.52 %

• Result destruction on the worker drops from 5.4–6.2 % to 0.4 %. operator delete drops by about 2 points, and the whole memoized materialize by about 2 points. As a share of the worker's non-blocked samples, materializePreparedMemoized goes from 28 % to 22 %.
• On this scene the memo hits rarely before perf(shader): reuse a memoized result by patching its copied words #10423: on main 5d3ab77 with APS5_PROFILE_DRAW=1, the line reads 11739 hits, 173897 misses (6.3% hits) … 173878 evictions per 10 s. So nearly every draw stage takes the eviction this PR reuses.
• FPS, frame p50 and GPU busy (100.7–101.0 ms per present) are the same in both arms. The queue 0 worker spends 29–30 % of its samples in DrawPipeline::Enqueue, waiting for the committer, so a worker saving does not reach the frame on this scene.
• No errors in any run.

AI-assisted: yes (Claude Code). The runs were made on my machine.