Skip to content

hk: reduce AGX2 reset work and G13X query overhead - #68

Closed
RyanTheTide wants to merge 3 commits into
aurora-wipfrom
perf/agx2-parity-20261024
Closed

RyanTheTide wants to merge 3 commits into
aurora-wipfrom
perf/agx2-parity-20261024

Conversation

@RyanTheTide

@RyanTheTide RyanTheTide commented Oct 24, 2026 •

Copy link
Copy Markdown

Remaining scope: G13X control-stream ioctl split only. The reset-work commits have been split into draft #69 at 3d00b6d8. This PR waits for the t-3018 kernel teardown fix and a fresh sustained qualification; do not merge it. Its existing combined head and historical qualification evidence are preserved below and have not been rewritten. The historical combined-head results do not establish the performance or stability of the G13X split by itself.


Draft: the M1 Pro firmware crash also reproduces on unchanged base 4911479 after accumulated CTS load. All 624,218 M1 Pro names now have complete candidate results across two boots with zero new bad results or skips. Sustained-run stability remains unresolved. G17P measurements are recorded below. M2 has zero deltas across 624,218 CTS cases. This is not ready to merge.

Proposed shipping scope: the two reset-work commits only. Hold the G13X split pending the kernel teardown fix and re-gating. The current combined head stays draft for review; exact-source qualification of a reset-only shipping head remains required.

The current-tip remeasurement still shows the M1 Pro occlusion round-trip overhead and M2 empty-submit regression against Fedora Honeykrisp. This changes command-buffer reset work on G13/G14 and restores separate control-stream ioctls on G13X.

The reset path leaves four native graphics caches and the native background/end-of-tile cache untouched on G13/G14. They start zeroed at allocation and are only populated by native graphics. All common recording state is still reset. On G13X, separate ioctls let rendering start while the following query-availability compute submission is prepared. G14G and native AGX3 keep batching. Query availability, waits, counts and barriers are unchanged.

Measurements

All values are microseconds, median of four independent repeat medians; brackets show the minimum and maximum repeat medians. Each comparison has its own interleaved stock runs. Base is 4911479b (including the occlusion completion and signalled-value cache fixes); candidate is af16d985. Lower is better.

Device Workload, queue depth Stock for base Base Stock for candidate Candidate Candidate/stock
M1 Pro G13X Occlusion overhead, d1 31.14 [24.64, 31.52] 69.1 [67.44, 78.61] 30.45 [13.25, 36.32] 24.51 [19.54, 28.79] 0.805
M1 Pro G13X Occlusion total, d1 349.9 [347.2, 353.1] 395.3 [390, 417.4] 350.9 [344.4, 366.6] 356.6 [351.6, 360.2] 1.016
M1 Pro G13X Occlusion total, d16 79.45 [79.16, 79.9] 78.89 [78.47, 79.52] 79.13 [79.11, 79.57] 79.08 [78.62, 79.31] 0.999
M1 Pro G13X Idle wake, 20 ms, d1 520.2 [518.4, 522.2] 527.9 [524.9, 534.5] 521.4 [512.2, 533] 530.8 [518.2, 534] 1.018
M1 Pro G13X Clear, d1 219.7 [211.5, 220.3] 225.1 [220.1, 230.2] 219.5 [209.7, 224.8] 226.8 [222.4, 232.6] 1.033
M1 Pro G13X Fill, 1080p, d1 1389 [1374, 1412] 1450 [1420, 1497] 1393 [1374, 1437] 1449 [1420, 1471] 1.041
M1 Pro G13X Empty submit, d32 1.517 [1.503, 1.521] 1.492 [1.489, 1.567] 1.526 [1.504, 1.559] 1.395 [1.392, 1.398] 0.914
M1 Pro G13X Empty submit, d64 1.614 [1.606, 1.624] 1.493 [1.488, 1.571] 1.575 [1.543, 1.623] 1.4 [1.396, 1.404] 0.889
M2 G14G Occlusion overhead, d1 72.93 [70.38, 73.33] 51.1 [46.9, 54.87] 75.16 [73.07, 81.18] 57.14 [47.74, 65.01] 0.760
M2 G14G Occlusion total, d1 321.3 [316.5, 323.4] 303.4 [295.2, 307.1] 325.5 [320.5, 328.9] 308.1 [307.1, 315.3] 0.946
M2 G14G Occlusion total, d16 88.76 [86.48, 89.48] 87.02 [86.9, 87.23] 88.98 [87.94, 89.29] 87.71 [87.08, 88.62] 0.986
M2 G14G Idle wake, 20 ms, d1 371.1 [367, 378.4] 367.4 [364.6, 372.6] 369 [365.8, 380.5] 368.3 [367.8, 369] 0.998
M2 G14G Clear, d1 172.1 [169.7, 175.3] 178.6 [175.6, 179.2] 173 [169, 174.3] 178.5 [175.6, 179.2] 1.032
M2 G14G Fill, 1080p, d1 860.6 [859.1, 965.3] 980.5 [902.5, 1078] 874.1 [871.1, 874.1] 941.5 [874.1, 1004] 1.077
M2 G14G Empty submit, d32 1.188 [1.183, 1.19] 1.262 [1.261, 1.265] 1.186 [1.185, 1.191] 1.178 [1.177, 1.179] 0.994
M2 G14G Empty submit, d64 1.183 [1.18, 1.184] 1.261 [1.26, 1.266] 1.182 [1.18, 1.183] 1.179 [1.177, 1.183] 0.997

The M1 Pro query overhead changes from 2.22x stock on the base to 0.805x stock here; total query time is 1.016x stock with overlapping ranges. M2 empty-submit time is at stock levels. The earlier idle-wake regression does not reproduce on the base or candidate. Clear remains about 3% slower; the M1 Pro fill median is about 4% slower with overlapping ranges.

A separate M2 fill stability check uses a 1 s warmup and five 2 s samples with timestamps disabled, four repeats per build. Raw us/submit repeat medians: stock 1009.5911, 1048.7677, 998.5022, 1003.0090; candidate 1020.4082, 1012.1457, 1032.5245, 1041.1244. Stock is 1006 [998.5, 1049] us and candidate 1026 [1012, 1041] us (1.020x). The larger difference in the short-window M2 fill table does not persist under that method. These are measured residual differences, not a claim that every workload is faster.

G17P base/candidate check

Same median-of-four-repeat method, in microseconds with repeat ranges. This baseline is Borealis 4911479b, rather than Fedora Honeykrisp. Four alternating pairs, the same allowed CPU affinity on both builds, inactive display, initial cooldown plus at least 60 s between builds, and process-local ICDs. All 64 benchmark rows completed successfully on the same 7.3.0-aurora-73w-dj423 boot.

G17P workload Depth Base 4911479b Candidate af16d985 Candidate/base
Occlusion overhead 1 30.41 [29.17, 30.77] 32.62 [30.91, 36.6] 1.073
Occlusion total 1 243.8 [242.5, 244.1] 248.6 [243, 253.8] 1.020
Occlusion total 16 59.88 [59.67, 60.42] 59.77 [59.68, 60.11] 0.998
Fill 1080p 1 1040 [1025, 1050] 1027 [1025, 1037] 0.988
Clear 1 168.7 [166.2, 173.3] 164.2 [160.7, 167.8] 0.974
Idle wake, 20 ms 1 579.4 [576.9, 580.2] 578.4 [578, 582.5] 0.998
Empty submit 32 2.418 [2.413, 2.433] 2.393 [2.386, 2.399] 0.989
Empty submit 64 2.436 [2.426, 2.447] 2.414 [2.4, 2.431] 0.991
Compute light 1 163.8 [162.2, 167.2] 160.6 [157.8, 162.8] 0.981

Most end-to-end medians are flat or faster. Occlusion d1 is an open concern: total time is 2.0% higher with overlapping repeat ranges, while the derived overhead is 7.3% higher and its ranges narrowly do not overlap. Fill CPU recording/submission medians also increased (8.009 to 10.04 us and 28.37 to 36.33 us), with a substantial change among the base repeats. These measurements do not establish a clean G17P non-regression result; a focused repeat has been requested before drawing a conclusion.

G17P raw repeat medians (JSON)
[
  {
    "workload": "Occlusion overhead",
    "test": "query",
    "depth": 1,
    "metric": "overhead_us",
    "build": "base491",
    "repeat_medians_us": [
      30.7175,
      30.7737,
      29.1724,
      30.0995
    ]
  },
  {
    "workload": "Occlusion overhead",
    "test": "query",
    "depth": 1,
    "metric": "overhead_us",
    "build": "candidateaf16",
    "repeat_medians_us": [
      32.3914,
      30.9102,
      32.8482,
      36.6047
    ]
  },
  {
    "workload": "Occlusion total",
    "test": "query",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "base491",
    "repeat_medians_us": [
      243.6647,
      243.9024,
      242.483,
      244.1406
    ]
  },
  {
    "workload": "Occlusion total",
    "test": "query",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "candidateaf16",
    "repeat_medians_us": [
      245.3386,
      242.9543,
      251.7623,
      253.8071
    ]
  },
  {
    "workload": "Occlusion total",
    "test": "query",
    "depth": 16,
    "metric": "us_per_submit",
    "build": "base491",
    "repeat_medians_us": [
      60.4157,
      59.6659,
      59.9664,
      59.7943
    ]
  },
  {
    "workload": "Occlusion total",
    "test": "query",
    "depth": 16,
    "metric": "us_per_submit",
    "build": "candidateaf16",
    "repeat_medians_us": [
      60.1106,
      59.8372,
      59.7086,
      59.6801
    ]
  },
  {
    "workload": "Fill 1080p",
    "test": "render.heavy",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "base491",
    "repeat_medians_us": [
      1024.5902,
      1041.6667,
      1037.3444,
      1050.4202
    ]
  },
  {
    "workload": "Fill 1080p",
    "test": "render.heavy",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "candidateaf16",
    "repeat_medians_us": [
      1024.5902,
      1024.5902,
      1028.8066,
      1037.3444
    ]
  },
  {
    "workload": "Clear",
    "test": "render.light",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "base491",
    "repeat_medians_us": [
      173.2502,
      169.3767,
      166.2234,
      168.0108
    ]
  },
  {
    "workload": "Clear",
    "test": "render.light",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "candidateaf16",
    "repeat_medians_us": [
      160.6684,
      161.8123,
      166.6667,
      167.7852
    ]
  },
  {
    "workload": "Idle wake, 20 ms",
    "test": "idle.wake",
    "depth": 1,
    "metric": "wake_latency_us",
    "build": "base491",
    "repeat_medians_us": [
      580.2095,
      576.9385,
      580.229,
      578.562
    ]
  },
  {
    "workload": "Idle wake, 20 ms",
    "test": "idle.wake",
    "depth": 1,
    "metric": "wake_latency_us",
    "build": "candidateaf16",
    "repeat_medians_us": [
      582.543,
      578.167,
      578.6455,
      577.958
    ]
  },
  {
    "workload": "Empty submit",
    "test": "submit.empty",
    "depth": 32,
    "metric": "us_per_submit",
    "build": "base491",
    "repeat_medians_us": [
      2.4184,
      2.4184,
      2.4135,
      2.4326
    ]
  },
  {
    "workload": "Empty submit",
    "test": "submit.empty",
    "depth": 32,
    "metric": "us_per_submit",
    "build": "candidateaf16",
    "repeat_medians_us": [
      2.3928,
      2.3928,
      2.3987,
      2.386
    ]
  },
  {
    "workload": "Empty submit",
    "test": "submit.empty",
    "depth": 64,
    "metric": "us_per_submit",
    "build": "base491",
    "repeat_medians_us": [
      2.4393,
      2.4323,
      2.4257,
      2.4467
    ]
  },
  {
    "workload": "Empty submit",
    "test": "submit.empty",
    "depth": 64,
    "metric": "us_per_submit",
    "build": "candidateaf16",
    "repeat_medians_us": [
      2.4038,
      2.3999,
      2.4245,
      2.4307
    ]
  },
  {
    "workload": "Compute light",
    "test": "compute.light",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "base491",
    "repeat_medians_us": [
      163.292,
      162.2323,
      167.2241,
      164.2576
    ]
  },
  {
    "workload": "Compute light",
    "test": "compute.light",
    "depth": 1,
    "metric": "us_per_submit",
    "build": "candidateaf16",
    "repeat_medians_us": [
      158.5289,
      157.8283,
      162.6545,
      162.7604
    ]
  }
]

Method

  • gpubench 070c2579fcc2156f5ffef50eab06a8f41d3eab0e, fixed AArch64 binary per device. Standard run: 300 ms warmup, five 250 ms samples, three timestamped samples. Four repeats in AB, BA, AB, BA order, at least 60 s cooldown between builds, waiting up to 240 s to return within 2 C of the initial reference temperature.
  • Process-local VK_DRIVER_FILES overlays, same boot and kernel 7.3.0-aurora-73w, display inactive and keyboard backlight zero. No system package changes.
  • Exact Fedora Honeykrisp baselines: M1 Pro 26.1.3-1.fc44, M2 26.2.3-1.fc44, extracted into scratch from verified rollback RPMs. System Borealis packages are not used as the stock baseline.
  • Candidate and base use the rolling release recipe: plain build, Fedora compiler flags, b_ndebug=true, AArch64 offline SDK. Candidate overlay SHA256 2454a97d0433a08f3fa51aa68404bd5b9b6b4b0eca00842d270b55cd21f92a3f.
  • CPU profiles identify extra reset/cache clearing. Bounded captures of the public benchmark's submit packets identify the stock render/compute split. A two-repeat batching diagnostic improves G13X query latency and worsens G14G, motivating the chip-specific policy. The diagnostic timings are not used as acceptance measurements.

Validation

  • Release build passes. Existing hk_sync_prior_test: 2,817 checks, zero failures.
  • Affected Vulkan CTS 1.4.6.2 on M2: 624,218 cases per build, identical result sets (241,612 Pass / 382,606 Skip), zero new failures or skips against 4911479b. Frozen groups: command buffers, queries, synchronization/synchronization2, render passes, image clears, blend, dynamic state, copies/blits.
  • M1 Pro case coverage: 624,218 exact names across candidate chunks 0–60 and 61–124 on two fresh boots, with 241,612 Pass / 382,606 Skip. Compared with the complete fresh base (241,611 Pass / 382,606 Skip / 1 Flake), there are zero new bad results and zero new skips; the sole difference is the base implicit binary-semaphore copy Flake becoming Pass. Kit, source/library, jobs, grouping, kernel and every selected chunk list are verified. This is case coverage assembled across boots, not a successful uninterrupted full candidate run. Both earlier candidate crashes remain part of the stability result.
  • G17P: all four alternating repeat pairs completed; the measurements and raw medians above leave an occlusion d1 concern open. No clean non-regression claim is made.

A pre-existing M1 Pro firmware crash occurs under sustained CTS, and the candidate may make it arrive sooner. This remains a merge blocker; keep this PR draft.

CTS boot / order Completed chunks before crash Crash group Teardown bursts Bursts / completed chunk
Earlier base, then candidate 125 base + 54 candidate = 179 Candidate image clearing Not recorded as bursts —
Fresh candidate first 61 Candidate blend 14 0.230
Fresh base twice 125 first pass + 65 second pass = 190 Base blend 27 0.142
Fresh candidate, remaining chunks 61–124 64 No crash in this segment 10 0.156
Fresh reset-only, early chunks 0–60 61 No crash in this segment 8 0.131
Fresh ioctl-only, early chunks 58 Ioctl-only blend 20 0.345

The fresh base's first complete pass had 16 bursts over 125 chunks (0.128/chunk), with no firmware/GPU-crashed messages. Its second pass added 11 teardown messages in 11 bursts, then 1 firmware-crash message and 8 GPU-crashed messages. The fresh candidate recorded 48 teardown messages in 14 bursts, then 1 firmware-crash message and 8 GPU-crashed messages. A burst starts after a gap over one second. These are descriptive rates from few boots with different prior load, not an estimate of crash probability.

A matched comparison of identical completed chunks 0–60 (305,000 cases), excluding both crash chunks, finds 12 teardown messages across 8 affected chunks on base versus 48 messages across 11 chunks on candidate: 0.197 versus 0.787 messages per chunk. Case lists, CTS binary, grouping, jobs and kernel match. The 4× message count keeps the acceleration concern open; this single pair does not assign it to either change.

The reset changes leave the submitted GPU work unchanged by design: they skip CPU stores to native-only caches on G13/G14. They can still change CPU timing and how teardown overlaps pending work, so an increase in cancellations has not been excluded. The G13X ioctl split is a separate change to submission timing. The component split test below isolates the large increase to the G13X ioctl split in these samples; reset-only does not reproduce it. Failed runs and their incomplete chunks remain preserved; successful subsets do not establish sustained-run stability.

Component split test and recommendation

Each component ran from a separate fresh boot with the same kit, kernel, jobs, grouping and frozen list. Reset-only 3d00b6d8 includes both reset commits and retains base batching. Ioctl-only 0cd978b2 applies the G13X separate-ioctl policy to the base without either reset commit. Reset-only completed chunks 0–60 (305,000 cases; 114,346 Pass / 190,654 Skip) without a firmware crash and recorded 9 teardown messages. Ioctl-only completed chunks 0–57, then crashed in chunk 58, recording 68 teardown messages and 20 bursts over that boot.

For a fair four-way comparison, use only identical complete chunks 0–57 (290,000 cases). The ioctl-only interrupted chunk is excluded from this table and remains reported above.

Build Teardown messages Affected chunks / 58 Messages / chunk
Base 4911479b 10 7 0.172
Reset-only 3d00b6d8 9 7 0.155
Ioctl-only 0cd978b2 59 14 1.017
Combined af16d985 33 10 0.569

The reset-work part alone did not increase cancellations in this sample. The ioctl split reproduces the increase (5.9× base over the complete common prefix) and an early firmware crash with the original reset behavior. This supports the ioctl split as added exposure to the kernel teardown fragility, consistent with more work remaining in flight at context teardown. Live job depth was not directly measured, and these are single-boot component samples; a smaller reset-timing effect is not ruled out.

Recommendation: select the reset-work part only for shipping, and hold the G13X ioctl split until the kernel teardown fix is available and has passed a sustained-load re-gate. The reset changes address the measured M2 empty-submit gap while preserving the GPU submission shape. A default-off shipping switch would retain an unqualified path without a default performance benefit.

The proposed reset-only head is 3d00b6d8 (commits 51c0bfd9 and 3d00b6d8 on 4911479b). The current combined draft head remains af16d985. This is a scope recommendation for review, not approval to merge the current combined head. Keep this PR draft. Before shipping a reset-only head, finish its exact-source qualification, including the focused G17P query timing recheck; the original G17P window above is not a clean non-regression result. No PR head rewrite or merge has been performed.

Raw repeat medians for the table (JSON)
[
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      31.515,
      30.9424,
      31.3326,
      24.6416
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      78.6089,
      68.9714,
      69.2242,
      67.435
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      36.3179,
      29.8872,
      31.0212,
      13.2492
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      19.5432,
      24.209,
      24.8201,
      28.7854
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      351.6174,
      353.1073,
      348.1894,
      347.2222
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      417.3623,
      396.1965,
      394.3218,
      390.0156
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      366.5689,
      344.3526,
      351.1236,
      350.6311
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      355.1136,
      360.2305,
      351.6174,
      358.1662
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      79.164,
      79.8977,
      79.7194,
      79.1891
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      78.4683,
      79.5165,
      79.2142,
      78.5669
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      79.5672,
      79.1139,
      79.139,
      79.1139
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      79.3147,
      79.2142,
      78.6164,
      78.9391
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      518.3935,
      522.2485,
      518.436,
      521.936
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      524.9155,
      534.54,
      527.707,
      528.1645
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      533.0195,
      512.186,
      515.1655,
      527.5405
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      518.1655,
      534.0195,
      529.0195,
      532.665
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      220.2643,
      211.5059,
      220.0704,
      219.2982
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      220.0704,
      230.2026,
      230.2026,
      220.0704
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      223.4138,
      209.7315,
      215.5172,
      224.8201
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      222.4199,
      232.5581,
      229.1476,
      224.4165
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      1388.8889,
      1388.8889,
      1373.6264,
      1412.4294
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      1420.4545,
      1470.5882,
      1428.5714,
      1497.006
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      1404.4944,
      1373.6264,
      1381.2155,
      1436.7816
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      1470.5882,
      1420.4545,
      1453.4884,
      1445.0867
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.517,
      1.5162,
      1.5033,
      1.5209
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.4932,
      1.49,
      1.5674,
      1.489
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.5052,
      1.5477,
      1.504,
      1.559
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.3981,
      1.3946,
      1.3956,
      1.3916
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.624,
      1.6136,
      1.6148,
      1.606
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "base",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.4951,
      1.4909,
      1.571,
      1.4881
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.6232,
      1.5428,
      1.5957,
      1.5543
    ]
  },
  {
    "device": "M1 Pro G13X",
    "comparison": "candidate",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.4044,
      1.3998,
      1.4009,
      1.396
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      72.5533,
      70.3788,
      73.3284,
      73.3107
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      51.2399,
      54.8712,
      50.9546,
      46.8972
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      81.1773,
      73.0715,
      75.6548,
      74.659
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "query",
    "metric": "overhead_us",
    "depth": 1,
    "repeat_medians_us": [
      47.7361,
      65.0083,
      53.0603,
      61.2258
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      316.4557,
      323.4153,
      322.5806,
      320.1024
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      307.1253,
      305.6235,
      301.2048,
      295.1594
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      328.9474,
      327.654,
      320.5128,
      323.4153
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      307.8818,
      315.2585,
      307.1253,
      308.2614
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      86.4753,
      89.0313,
      88.4956,
      89.4775
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      86.9868,
      87.0474,
      86.8961,
      87.2296
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      87.9353,
      89.2857,
      89.1266,
      88.8415
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "query",
    "metric": "us_per_submit",
    "depth": 16,
    "repeat_medians_us": [
      88.6211,
      88.1213,
      87.0777,
      87.2905
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      378.418,
      374.293,
      366.9805,
      367.9805
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      364.6475,
      367.3975,
      367.335,
      372.564
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      365.772,
      370.085,
      368.001,
      380.5425
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "idle.wake",
    "metric": "wake_latency_us",
    "depth": 1,
    "repeat_medians_us": [
      367.939,
      368.7515,
      367.772,
      368.9805
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      170.5321,
      175.3156,
      173.7318,
      169.7217
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      178.0627,
      179.0831,
      179.2115,
      175.5618
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      173.6111,
      169.0331,
      174.3375,
      172.295
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "render.light",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      175.5618,
      179.2115,
      178.0627,
      178.9549
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      859.1065,
      859.1065,
      862.069,
      965.251
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      968.9922,
      1077.5862,
      902.5271,
      992.0635
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      871.0801,
      874.1259,
      874.1259,
      874.1259
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "render.heavy",
    "metric": "us_per_submit",
    "depth": 1,
    "repeat_medians_us": [
      980.3922,
      874.1259,
      1004.0161,
      902.5271
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.1874,
      1.1901,
      1.1883,
      1.183
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.2608,
      1.2616,
      1.2647,
      1.2617
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.185,
      1.1845,
      1.1861,
      1.1909
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 32,
    "repeat_medians_us": [
      1.1785,
      1.1787,
      1.1782,
      1.1773
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.1815,
      1.184,
      1.1836,
      1.1798
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "base",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.2607,
      1.2663,
      1.2617,
      1.2604
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "stock",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.183,
      1.1804,
      1.1822,
      1.1825
    ]
  },
  {
    "device": "M2 G14G",
    "comparison": "candidate",
    "build": "tip",
    "test": "submit.empty",
    "metric": "us_per_submit",
    "depth": 64,
    "repeat_medians_us": [
      1.1834,
      1.1789,
      1.179,
      1.1771
    ]
  }
]

AGX2 never populates the native graphics loader, launch, combined-sampler or spilled-render-target caches. Their storage is already zeroed when the command buffer is allocated. Leave that contiguous storage untouched on G13 and G14 resets, and reset all common state on either side. Native graphics resets keep the full clear.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
The per-recording background/end-of-tile cache is only populated by the native graphics path. G13 and G14 otherwise invalidate sixteen entries spread over 13.5 KiB on every command-buffer reset, despite all of them remaining unused. Keep the existing invalidation on other generations.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
G13X serializes render and compute work. Submit each control stream when it is ready so rendering can start while the following compute submission is prepared. This restores the upstream submission shape for post-graphics query availability. Keep batching on the other chips, including G14G and native AGX3.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
@RyanTheTide

Copy link
Copy Markdown
Author

Superseded by #87, merged into aurora-wip in the 2026-10-28 consolidation.

@RyanTheTide
RyanTheTide deleted the perf/agx2-parity-20261024 branch October 28, 2026 04:26
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant