Skip to content

perf: bulk memcpy for primitive vector encoding on LE platforms - #715

Merged
lwshang merged 3 commits into
masterfrom
sat-perf-bulk-encode
Mar 15, 2026
Merged

perf: bulk memcpy for primitive vector encoding on LE platforms#715
lwshang merged 3 commits into
masterfrom
sat-perf-bulk-encode

Conversation

@sasa-tomic

@sasa-tomic sasa-tomic commented Mar 13, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Summary

On little-endian targets (including wasm32), Vec<i16/u32/f64/etc.> encoding now writes the backing memory directly instead of iterating element-by-element through the Serializer trait.

  • Added try_write_raw_elements default method to Compound trait
  • ValueSerializer::Compound implements it with direct byte write
  • [T]::idl_serialize detects fixed-width primitives via TypeId on #[cfg(target_endian = "little")]
  • Vec<T> and [T; N] delegate to the optimized slice path
  • Excludes usize/isize (platform-dependent size)

Benchmark (canbench, wasm32)

Metric Before After Change
vec_int16 Encoding 123.7M inst 8.4M inst 14.7x faster (-93.2%)
vec_int16 Total 817.9M inst 704.7M inst -13.8%

No regressions on any of the 9 benchmarks.

Compatibility

Wire format is unchanged — produces byte-identical output. Round-trip compatibility test added.

Relates to #710

@sasa-tomic
sasa-tomic requested a review from a team as a code owner March 13, 2026 14:04
@github-actions

github-actions Bot commented Mar 13, 2026 •
edited
Loading

Copy link
Copy Markdown
Name Max Mem (Kb) Encode Decode
blob 4_224 4_207_487 2_122_465 ($\textcolor{red}{0.00\%}$)
btreemap 73_856 531_975_943 ($\textcolor{red}{0.00\%}$) 13_058_092_273 ($\textcolor{red}{0.57\%}$)
nns 192 2_021_253 5_670_864 ($\textcolor{red}{0.12\%}$)
nns_list_proposal 1_216 7_018_096 ($\textcolor{red}{0.12\%}$) 64_361_414 ($\textcolor{red}{0.08\%}$)
option_list 64 715_981 ($\textcolor{green}{-0.00\%}$) 21_800_274 ($\textcolor{green}{-0.40\%}$)
text 6_336 4_204_384 7_877_792 ($\textcolor{red}{0.00\%}$)
variant_list 64 710_969 ($\textcolor{green}{-0.00\%}$) 20_596_350 ($\textcolor{red}{0.33\%}$)
vec_int16 12_480 ($\textcolor{green}{-25.29\%}$) 8_404_689 ($\textcolor{green}{-93.21\%}$) 635_462_524 ($\textcolor{red}{0.33\%}$)
  • Parser cost: 16_174_059
  • Extra args: 2_872_191
Click to see raw report 
---------------------------------------------------

Benchmark: blob
  total:
    instructions: 6.33 M (0.00%) (change within noise threshold)
    heap_increase: 66 pages (no change)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 4.21 M (no change)
    heap_increase: 66 pages (no change)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 2.12 M (0.00%) (change within noise threshold)
    heap_increase: 0 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: btreemap
  total:
    instructions: 13.59 B (0.54%) (change within noise threshold)
    heap_increase: 1154 pages (no change)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 531.98 M (0.00%) (change within noise threshold)
    heap_increase: 159 pages (no change)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 13.06 B (0.57%) (change within noise threshold)
    heap_increase: 995 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: extra_args
  total:
    instructions: 2.87 M (no change)
    heap_increase: 0 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: nns
  total:
    instructions: 24.70 M (0.03%) (change within noise threshold)
    heap_increase: 3 pages (no change)
    stable_memory_increase: 0 pages (no change)

  0. Parsing (scope):
    calls: 1 (no change)
    instructions: 16.17 M (no change)
    heap_increase: 3 pages (no change)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 2.02 M (no change)
    heap_increase: 0 pages (no change)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 5.67 M (0.12%) (change within noise threshold)
    heap_increase: 0 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: nns_list_proposal
  total:
    instructions: 71.38 M (0.08%) (change within noise threshold)
    heap_increase: 19 pages (no change)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 7.02 M (0.12%) (change within noise threshold)
    heap_increase: 5 pages (no change)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 64.36 M (0.08%) (change within noise threshold)
    heap_increase: 14 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: option_list
  total:
    instructions: 22.52 M (-0.39%) (change within noise threshold)
    heap_increase: 1 pages (no change)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 715.98 K (-0.00%) (change within noise threshold)
    heap_increase: 0 pages (no change)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 21.80 M (-0.40%) (change within noise threshold)
    heap_increase: 1 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: text
  total:
    instructions: 12.08 M (0.00%) (change within noise threshold)
    heap_increase: 99 pages (no change)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 4.20 M (no change)
    heap_increase: 66 pages (no change)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 7.88 M (0.00%) (change within noise threshold)
    heap_increase: 33 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: variant_list
  total:
    instructions: 21.31 M (0.32%) (change within noise threshold)
    heap_increase: 1 pages (no change)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 710.97 K (-0.00%) (change within noise threshold)
    heap_increase: 0 pages (no change)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 20.60 M (0.33%) (change within noise threshold)
    heap_increase: 1 pages (no change)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Benchmark: vec_int16
  total:
    instructions: 643.87 M (improved by 14.95%)
    heap_increase: 195 pages (improved by 25.29%)
    stable_memory_increase: 0 pages (no change)

  1. Encoding (scope):
    calls: 1 (no change)
    instructions: 8.40 M (improved by 93.21%)
    heap_increase: 130 pages (improved by 50.19%)
    stable_memory_increase: 0 pages (no change)

  2. Decoding (scope):
    calls: 1 (no change)
    instructions: 635.46 M (0.33%) (change within noise threshold)
    heap_increase: 65 pages (regressed from 0)
    stable_memory_increase: 0 pages (no change)

---------------------------------------------------

Summary:
  instructions:
    status:   Improvements detected 🟢
    counts:   [total 9 | regressed 0 | improved 1 | new 0 | unchanged 8]
    change:   [max +73.40M | p75 +57.17K | median +33 | p25 0 | min -113.19M]
    change %: [max +0.54% | p75 +0.08% | median 0.00% | p25 0.00% | min -14.95%]

  heap_increase:
    status:   Improvements detected 🟢
    counts:   [total 9 | regressed 0 | improved 1 | new 0 | unchanged 8]
    change:   [max 0 | p75 0 | median 0 | p25 0 | min -66]
    change %: [max 0.00% | p75 0.00% | median 0.00% | p25 0.00% | min -25.29%]

  stable_memory_increase:
    status:   No significant changes 👍
    counts:   [total 9 | regressed 0 | improved 0 | new 0 | unchanged 9]
    change:   [max 0 | p75 0 | median 0 | p25 0 | min 0]
    change %: [max 0.00% | p75 0.00% | median 0.00% | p25 0.00% | min 0.00%]

---------------------------------------------------

Only significant changes:
| status | name                   | calls |     ins |  ins Δ% |  HI |   HI Δ% | SMI |  SMI Δ% |
|--------|------------------------|-------|---------|---------|-----|---------|-----|---------|
|   +    | vec_int16::2. Decoding |     1 | 635.46M |  +0.33% |  65 |   +inf% |   0 |   0.00% |
|   -    | vec_int16              |       | 643.87M | -14.95% | 195 | -25.29% |   0 |   0.00% |
|   -    | vec_int16::1. Encoding |     1 |   8.40M | -93.21% | 130 | -50.19% |   0 |   0.00% |

ins = instructions, HI = heap_increase, SMI = stable_memory_increase, Δ% = percent change

---------------------------------------------------
Successfully persisted results to canbench_results.yml

@sasa-tomic @lwshang
perf: bulk memcpy for primitive vector encoding on LE platforms
cbf2f13 
On little-endian targets (including wasm32), Vec<i16/u32/f64/etc.>
encoding now writes the backing memory directly instead of iterating
element-by-element through the Serializer trait.

Benchmark: vec_int16 encoding 123M → 8.4M instructions (14.7x faster).

Wire format is unchanged — produces byte-identical output.

Made-with: Cursor
@sasa-tomic @lwshang
fmt
1952d76
@lwshang
lwshang force-pushed the sat-perf-bulk-encode branch from 33bbf0a to 1952d76 Compare March 15, 2026 17:37
@lwshang
lwshang changed the base branch from sat-perf-improvements to master March 15, 2026 17:45
@lwshang @claude
ci: trigger CI
7eb0068 
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@lwshang
lwshang merged commit eaf7e6c into master Mar 15, 2026
11 checks passed
@lwshang
lwshang deleted the sat-perf-bulk-encode branch March 15, 2026 17:51
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@lwshang lwshang lwshang approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@sasa-tomic @lwshang