Remove the triage suspected machinery and break cycles with weak refs by teoxoy · Pull Request #5910 · gfx-rs/wgpu

teoxoy · 2024-07-04T20:20:02Z

Resolves #5560.
Resolves #5120.
Resolves #5592 & resolves #5583 by removing Device::temp_suspected.

PR doesn't need to be squashed, each commit builds by itself.

teoxoy · 2024-07-08T09:31:13Z

renderpass benchmark results of current trunk vs this PR (TLDR: 39-64% less time spent in .submit()):

Benchmarking Renderpass: Single Threaded/1 renderpasses x 10000 draws (Renderpass Time): Warming up for 3.0000 sAdapterInfo { name: "NVIDIA GeForce RTX 2080 with Max-Q Design", vendor: 4318, device: 7824, device_type: DiscreteGpu, driver: "NVIDIA", driver_info: "555.99", backend: Vulkan }
Renderpass: Single Threaded/1 renderpasses x 10000 draws (Renderpass Time)
                        time:   [45.386 ms 45.612 ms 45.862 ms]
                        thrpt:  [218.04 Kelem/s 219.24 Kelem/s 220.33 Kelem/s]
                 change:
                        time:   [-45.146% -43.159% -41.027%] (p = 0.00 < 0.05)
                        thrpt:  [+69.569% +75.928% +82.302%]
                        Performance has improved.
Found 5 outliers among 100 measurements (5.00%)
  4 (4.00%) high mild
  1 (1.00%) high severe
Renderpass: Single Threaded/2 renderpasses x 5000 draws (Renderpass Time)
                        time:   [45.995 ms 46.243 ms 46.515 ms]
                        thrpt:  [214.98 Kelem/s 216.25 Kelem/s 217.41 Kelem/s]
                 change:
                        time:   [-38.633% -36.379% -34.017%] (p = 0.00 < 0.05)
                        thrpt:  [+51.554% +57.181% +62.954%]
                        Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
  2 (2.00%) high mild
  1 (1.00%) high severe
Renderpass: Single Threaded/4 renderpasses x 2500 draws (Renderpass Time)
                        time:   [46.370 ms 46.602 ms 46.849 ms]
                        thrpt:  [213.45 Kelem/s 214.58 Kelem/s 215.66 Kelem/s]
                 change:
                        time:   [-36.293% -33.691% -30.883%] (p = 0.00 < 0.05)
                        thrpt:  [+44.681% +50.809% +56.969%]
                        Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
  6 (6.00%) high mild
Renderpass: Single Threaded/8 renderpasses x 1250 draws (Renderpass Time)
                        time:   [47.662 ms 47.883 ms 48.131 ms]
                        thrpt:  [207.77 Kelem/s 208.84 Kelem/s 209.81 Kelem/s]
                 change:
                        time:   [-48.640% -46.851% -45.073%] (p = 0.00 < 0.05)
                        thrpt:  [+82.058% +88.149% +94.703%]
                        Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
  1 (1.00%) high mild
  2 (2.00%) high severe
Renderpass: Single Threaded/1 renderpasses x 10000 draws (Submit Time)
                        time:   [7.6377 ms 7.6754 ms 7.7151 ms]
                        thrpt:  [1.2962 Melem/s 1.3029 Melem/s 1.3093 Melem/s]
                 change:
                        time:   [-66.548% -64.128% -61.611%] (p = 0.00 < 0.05)
                        thrpt:  [+160.49% +178.77% +198.93%]
                        Performance has improved.
Found 2 outliers among 100 measurements (2.00%)
  2 (2.00%) high mild
Renderpass: Single Threaded/2 renderpasses x 5000 draws (Submit Time)
                        time:   [9.2829 ms 9.3285 ms 9.3782 ms]
                        thrpt:  [1.0663 Melem/s 1.0720 Melem/s 1.0772 Melem/s]
                 change:
                        time:   [-52.002% -50.382% -48.791%] (p = 0.00 < 0.05)
                        thrpt:  [+95.276% +101.54% +108.34%]
                        Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
  2 (2.00%) high mild
  1 (1.00%) high severe
Renderpass: Single Threaded/4 renderpasses x 2500 draws (Submit Time)
                        time:   [10.535 ms 10.598 ms 10.669 ms]
                        thrpt:  [937.28 Kelem/s 943.61 Kelem/s 949.21 Kelem/s]
                 change:
                        time:   [-59.243% -56.854% -54.335%] (p = 0.00 < 0.05)
                        thrpt:  [+118.99% +131.77% +145.36%]
                        Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
  3 (3.00%) high mild
  1 (1.00%) high severe
Renderpass: Single Threaded/8 renderpasses x 1250 draws (Submit Time)
                        time:   [12.488 ms 12.657 ms 12.852 ms]
                        thrpt:  [778.10 Kelem/s 790.06 Kelem/s 800.78 Kelem/s]
                 change:
                        time:   [-40.483% -39.474% -38.497%] (p = 0.00 < 0.05)
                        thrpt:  [+62.594% +65.217% +68.020%]
                        Performance has improved.
Found 12 outliers among 100 measurements (12.00%)
  6 (6.00%) high mild
  6 (6.00%) high severe

Renderpass: Multi Threaded/2 threads x 5000 draws
                        time:   [26.509 ms 26.626 ms 26.747 ms]
                        thrpt:  [373.88 Kelem/s 375.57 Kelem/s 377.22 Kelem/s]
                 change:
                        time:   [-1.3415% -0.7862% -0.2772%] (p = 0.00 < 0.05)
                        thrpt:  [+0.2779% +0.7924% +1.3597%]
                        Change within noise threshold.
Found 1 outliers among 100 measurements (1.00%)
  1 (1.00%) high mild
Renderpass: Multi Threaded/4 threads x 2500 draws
                        time:   [17.354 ms 17.457 ms 17.562 ms]
                        thrpt:  [569.42 Kelem/s 572.83 Kelem/s 576.23 Kelem/s]
                 change:
                        time:   [-0.5265% +0.2546% +1.0503%] (p = 0.52 > 0.05)
                        thrpt:  [-1.0394% -0.2539% +0.5293%]
                        No change in performance detected.
Found 2 outliers among 100 measurements (2.00%)
  1 (1.00%) low mild
  1 (1.00%) high mild
Renderpass: Multi Threaded/8 threads x 1250 draws
                        time:   [13.877 ms 13.933 ms 13.991 ms]
                        thrpt:  [714.74 Kelem/s 717.70 Kelem/s 720.60 Kelem/s]
                 change:
                        time:   [-1.8826% -1.1951% -0.5450%] (p = 0.00 < 0.05)
                        thrpt:  [+0.5480% +1.2096% +1.9188%]
                        Change within noise threshold.
Found 1 outliers among 100 measurements (1.00%)
  1 (1.00%) high mild

Renderpass: Bindless/10000 draws
                        time:   [18.025 ms 18.092 ms 18.160 ms]
                        thrpt:  [550.67 Kelem/s 552.73 Kelem/s 554.79 Kelem/s]
                 change:
                        time:   [-3.7818% -3.2353% -2.7117%] (p = 0.00 < 0.05)
                        thrpt:  [+2.7873% +3.3435% +3.9305%]
                        Performance has improved.

Renderpass: Empty Submit with 90000 Resources
                        time:   [4.1057 µs 4.1180 µs 4.1325 µs]
                        change: [-99.905% -99.904% -99.903%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 8 outliers among 100 measurements (8.00%)
  5 (5.00%) high mild
  3 (3.00%) high severe

nical

Great work! This makes significant changes to some pretty complicated systems. The changes look good to me although I usually have a firmer understanding of the potential side effects when I approve a PR.

So this calls for some more thorough testing than usual or a second pair of eyes.

teoxoy · 2024-07-08T14:56:51Z

I modified the benchmark to check the poll time to see what impact removing triage_suspected and dropping the Trackers in triage_submissions had. Times improved for 1 and 2 command buffers, remained the same for 4 and increased for 8; I don't think this is surprising, but I was actually expecting perf regressions in poll across the board.

Benchmarking Renderpass: Single Threaded/1 renderpasses x 10000 draws (Poll Time): Warming up for 3.0000 sAdapterInfo { name: "NVIDIA GeForce RTX 2080 with Max-Q Design", vendor: 4318, device: 7824, device_type: DiscreteGpu, driver: "NVIDIA", driver_info: "555.99", backend: Vulkan }
Renderpass: Single Threaded/1 renderpasses x 10000 draws (Poll Time)
                        time:   [4.3303 ms 4.3876 ms 4.4523 ms]
                        thrpt:  [2.2460 Melem/s 2.2792 Melem/s 2.3093 Melem/s]
                 change:
                        time:   [-34.959% -27.738% -19.927%] (p = 0.00 < 0.05)
                        thrpt:  [+24.886% +38.386% +53.748%]
                        Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
  4 (4.00%) high mild
  2 (2.00%) high severe
Renderpass: Single Threaded/2 renderpasses x 5000 draws (Poll Time)
                        time:   [4.4162 ms 4.4597 ms 4.5056 ms]
                        thrpt:  [2.2195 Melem/s 2.2423 Melem/s 2.2644 Melem/s]
                 change:
                        time:   [-49.601% -45.172% -40.151%] (p = 0.00 < 0.05)
                        thrpt:  [+67.087% +82.389% +98.418%]
                        Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
  4 (4.00%) high mild
Renderpass: Single Threaded/4 renderpasses x 2500 draws (Poll Time)
                        time:   [5.3340 ms 5.3993 ms 5.4708 ms]
                        thrpt:  [1.8279 Melem/s 1.8521 Melem/s 1.8748 Melem/s]
                 change:
                        time:   [-8.7991% -1.7697% +5.5161%] (p = 0.64 > 0.05)
                        thrpt:  [-5.2277% +1.8016% +9.6481%]
                        No change in performance detected.
Found 2 outliers among 100 measurements (2.00%)
  1 (1.00%) high mild
  1 (1.00%) high severe
Renderpass: Single Threaded/8 renderpasses x 1250 draws (Poll Time)
                        time:   [7.1084 ms 7.1824 ms 7.2583 ms]
                        thrpt:  [1.3777 Melem/s 1.3923 Melem/s 1.4068 Melem/s]
                 change:
                        time:   [+43.896% +52.281% +60.210%] (p = 0.00 < 0.05)
                        thrpt:  [-37.582% -34.332% -30.505%]
                        Performance has regressed.

teoxoy · 2024-07-08T15:09:11Z

All in all, perf of submit + poll has improved by 22-32%.

Benchmarking Renderpass: Single Threaded/1 renderpasses x 10000 draws (Submit + Poll Time): Warming up for 3.0000 sAdapterInfo { name: "NVIDIA GeForce RTX 2080 with Max-Q Design", vendor: 4318, device: 7824, device_type: DiscreteGpu, driver: "NVIDIA", driver_info: "555.99", backend: Vulkan }
Renderpass: Single Threaded/1 renderpasses x 10000 draws (Submit + Poll Time)
                        time:   [11.894 ms 11.955 ms 12.020 ms]
                        thrpt:  [831.93 Kelem/s 836.46 Kelem/s 840.79 Kelem/s]
                 change:
                        time:   [-33.133% -32.534% -31.944%] (p = 0.00 < 0.05)
                        thrpt:  [+46.938% +48.223% +49.551%]
                        Performance has improved.
Found 2 outliers among 100 measurements (2.00%)
  1 (1.00%) high mild
  1 (1.00%) high severe
Renderpass: Single Threaded/2 renderpasses x 5000 draws (Submit + Poll Time)
                        time:   [13.566 ms 13.656 ms 13.757 ms]
                        thrpt:  [726.90 Kelem/s 732.26 Kelem/s 737.13 Kelem/s]
                 change:
                        time:   [-30.535% -29.948% -29.234%] (p = 0.00 < 0.05)
                        thrpt:  [+41.311% +42.752% +43.958%]
                        Performance has improved.
Found 2 outliers among 100 measurements (2.00%)
  1 (1.00%) high mild
  1 (1.00%) high severe
Renderpass: Single Threaded/4 renderpasses x 2500 draws (Submit + Poll Time)
                        time:   [15.506 ms 15.593 ms 15.685 ms]
                        thrpt:  [637.56 Kelem/s 641.30 Kelem/s 644.91 Kelem/s]
                 change:
                        time:   [-27.790% -27.327% -26.847%] (p = 0.00 < 0.05)
                        thrpt:  [+36.699% +37.603% +38.485%]
                        Performance has improved.
Found 2 outliers among 100 measurements (2.00%)
  2 (2.00%) high mild
Renderpass: Single Threaded/8 renderpasses x 1250 draws (Submit + Poll Time)
                        time:   [19.174 ms 19.330 ms 19.534 ms]
                        thrpt:  [511.92 Kelem/s 517.34 Kelem/s 521.55 Kelem/s]
                 change:
                        time:   [-23.052% -22.296% -21.462%] (p = 0.00 < 0.05)
                        thrpt:  [+27.326% +28.693% +29.958%]
                        Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
  2 (2.00%) high mild
  1 (1.00%) high severe

cwfitzgerald

This looks great! I have one comment about the trackers. Did you do much changes in the trackers, or just moving code around, it's a bit hard to tell.

Love the diff numbers!

wgpu-core/src/track/buffer.rs

teoxoy · 2024-07-09T07:52:43Z

Did you do much changes in the trackers, or just moving code around, it's a bit hard to tell.

The only behavioral change was .insert_single() will no longer panic if the resource at the tracker index already exists. This was done because we recycle tracker indices on resource drop and I found no issue with just overwriting the data in the tracker at a given index.

…ources. It's worth noting that `suspected_resources` never contained those resources.

The Vec only ever contained 0 or 1 command buffers. We now acquire an encoder on every submit for pending writes but that's not a problem since those are coming from a pool anyway.

…ive in `LifetimeTracker.handle_mapping` This change doesn't change behavior as `Global.buffer_drop` already unmaps the buffer.

teoxoy requested a review from a team as a code owner July 4, 2024 20:20

nical approved these changes Jul 8, 2024

View reviewed changes

cwfitzgerald reviewed Jul 8, 2024

View reviewed changes

wgpu-core/src/track/buffer.rs Outdated Show resolved Hide resolved

teoxoy added 20 commits July 9, 2024 10:01

remove unused TempResource::Texture

df18fc5

change prepare_staging_buffer to return a non null u8 pointer

7675aa1

use StagingBuffer instead of Buffer for mapped_at_creation Buffers

f529af1

remove unused TempResource::Buffer

1a10ef4

remove Device.create_buffer_impl

0eec33e

remove Arcs around TempResource variants

839d469

add ActiveSubmission.temp_resources that contains all temporary res…

1a4675b

…ources. It's worth noting that `suspected_resources` never contained those resources.

remove device_maintain_ids

00a6486

move trackers into EncoderInFlight on submit

da338bf

remove PendingWrites.executing_command_buffers

725c838

The Vec only ever contained 0 or 1 command buffers. We now acquire an encoder on every submit for pending writes but that's not a problem since those are coming from a pool anyway.

remove LifetimeTracker.future_suspected_{buffers,textures}

2243a61

don't check if the buffer is still "present" from the user's perspect…

52a89cf

…ive in `LifetimeTracker.handle_mapping` This change doesn't change behavior as `Global.buffer_drop` already unmaps the buffer.

remove the triage suspected machinery

d2fbf9b

replace the tracker in Device with a new DeviceTracker

77f196a

make resource metadata generic over a T: Clone

7fa17ff

remove unnecessary PhantomData

3deb5cb

remove unused BufferTracker.get()

c45c2b2

don't call drain_transitions() of we drop the results

b3559cd

introduce DeviceBufferTracker which holds weak references to buffers

ae33a88

introduce DeviceTextureTracker which holds weak references to textures

7ba4498

teoxoy force-pushed the submit-tracking branch from 20a8e35 to 7ba4498 Compare July 9, 2024 08:35

teoxoy merged commit 4255268 into gfx-rs:trunk Jul 9, 2024

teoxoy deleted the submit-tracking branch July 9, 2024 08:48

teoxoy mentioned this pull request Jul 11, 2024

[wgpu-core] fix trying to create 0 sized staging buffers when creating mapped_at_creation buffers #5945

Merged

teoxoy self-assigned this Jul 15, 2024

This was referenced Jul 16, 2024

Early frees on CPU Implementations #3193

Closed

Vulkan Validation Error: Cannot free VkBuffer that is in use by a command buffer. #1689

Closed

[test] remove molten vk failure #5964

Closed

This was referenced Jul 26, 2024

Boids Fails Validation on MoltenVK #2774

Closed

Buffers Leaked when Device Handle is Dropped Before Buffers #5529

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Remove the triage suspected machinery and break cycles with weak refs#5910

Remove the triage suspected machinery and break cycles with weak refs#5910
teoxoy merged 20 commits intogfx-rs:trunkfrom
teoxoy:submit-tracking

teoxoy commented Jul 4, 2024 •

edited

Loading

Uh oh!

teoxoy commented Jul 8, 2024 •

edited

Loading

Uh oh!

nical left a comment

Uh oh!

teoxoy commented Jul 8, 2024

Uh oh!

teoxoy commented Jul 8, 2024

Uh oh!

cwfitzgerald left a comment

Uh oh!

Uh oh!

teoxoy commented Jul 9, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

teoxoy commented Jul 4, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

teoxoy commented Jul 8, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nical left a comment

Choose a reason for hiding this comment

Uh oh!

teoxoy commented Jul 8, 2024

Uh oh!

teoxoy commented Jul 8, 2024

Uh oh!

cwfitzgerald left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

teoxoy commented Jul 9, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

teoxoy commented Jul 4, 2024 •

edited

Loading

teoxoy commented Jul 8, 2024 •

edited

Loading