Skip to content

SIMD tracking issue. #64170

Open
Feature
Open
SIMD tracking issue.#64170
Feature
Assignees
modulovalue
Labels
area-sdkUse area-sdk for general purpose SDK issues (packaging, distribution, …).library-typed-data

Description

@modulovalue
modulovalue
opened on Aug 30, 2026

The following (incomplete, work in progress) tables track our SIMD value types and my progress on them so far. Many things below are aspirational and will need a separate discussion.

If you have a use case you would like to implement with SIMD, or if you run into problems, please open an issue and mention @modulovalue and I will take a look.

Int32x4

Operation API Codegen optimizations Benchmark Issues
splat ✅ cl/531960 ✅ cl/531960, cl/550200 ❓
zero (+ identity) ✅ cl/544840 🚧 ❓
one (* identity) 🚧 🚧 ❓
const constructor 🚧 🚧 ❓ #31487
+ (add) ✅ ✅ cl/497000 ✅ SimdInt32x4
- (subtract) ✅ ✅ cl/497000 ✅ SimdInt32x4
* (multiply) 🚧 cl/551260 🚧 ❓ #59816
unary - (sign negation) ✅ cl/545361 🚧 ❓
& (and) ✅ ✅ cl/497000 ✅ SimdInt32x4, BitArray32x4
| (or) ✅ ✅ cl/497000 ✅ SimdInt32x4, BitArray32x4
^ (xor) ✅ ✅ cl/497000 ✅ SimdInt32x4, BitArray32x4
~ (not) ✅ cl/532480 ✅ cl/532480 TODO extend BitArray32x4
andNot ✅ cl/551000 ✅ cl/558762 TODO extend BitArray32x4
<< (shiftLeft) ✅ cl/545480 ✅ dart2wasm cl/547880, VM cl/553360 ❓ #47114, #63680
>> (shiftRight) ✅ cl/545480 ✅ dart2wasm cl/547880, VM cl/553360 ❓ #47114, #63680
>>> (shiftRightUnsigned) ✅ cl/557620 ✅ dart2wasm cl/557620, VM cl/560562 ❓ #64345
equal ✅ cl/528620 ✅ cl/539060, cl/539520 ✅ SimdSearch
notEqual ✅ cl/541221 ✅ cl/541240 TODO extend SimdSearch
< (lessThan) ✅ cl/543500 🚧 dart2wasm ✅ cl/547880, VM in progress ❓ #63680
<= (lessThanOrEqual) ✅ cl/543500 🚧 dart2wasm ✅ cl/547880, VM in progress ❓ #63680
> (greaterThan) ✅ cl/543500 🚧 dart2wasm ✅ cl/547880, VM in progress ❓ #63680
>= (greaterThanOrEqual) ✅ cl/543500 🚧 dart2wasm ✅ cl/547880, VM in progress ❓ #63680
min ✅ cl/551040 🚧 dart2wasm ✅ cl/547880, VM in progress ❓
max ✅ cl/551040 🚧 dart2wasm ✅ cl/547880, VM in progress ❓
abs ✅ cl/545361 🚧 dart2wasm ✅ cl/547880, VM pending ❓
anyTrue ✅ cl/528620 ✅ cl/539400 ✅ SimdSearch #64129
allTrue ✅ cl/543480 ✅ cl/551641 ❓ #64129
signMask ✅ ✅ cl/521245, cl/542362, cl/542363 ✅ SimdInt32x4 #63680
x/y/z/w ✅ ✅ cl/507761 ✅ BitArray32x4
flagX/Y/Z/W ✅ ✅ ❓ #64129
withX/Y/Z/W ✅ ✅ cl/550201 ❓
withFlagX/Y/Z/W ✅ ✅ ❓ #64129
shuffle ✅ ✅ cl/559002, arm64 cl/562680, cl/562681, cl/563880, cl/563881 ❓ #63680
shuffleMix ✅ ✅ cl/559002, arm64 cl/562680, cl/562681, cl/563880, cl/563881 ❓
swizzle 🚧 🚧 ❓
select ✅ ✅ ❓
== / hashCode 🚧 in progress 🚧 in progress ❓ #43255
toString ✅ cl/547261 ✅ cl/547261 ❓ #63847
conversions (fromFloat32x4Bits) 🚧 ❓ ❓ #46182

Float32x4

Operation API Codegen optimizations Benchmark Issues
splat ✅ ✅ ❓
zero (+ identity) ✅ ✅ ❓
one (* identity) 🚧 🚧 ❓
const constructor 🚧 🚧 ❓ #31487, #56363
+ (add) ✅ ✅ ❓
- (subtract) ✅ ✅ ❓
* (multiply) ✅ ✅ ❓
/ (divide) ✅ ✅ ❓
unary - (sign negation) ✅ ✅ ❓
abs ✅ ✅ ❓
min ✅ ✅ cl/543520, cl/562661 ❓ #63962, #64521
max ✅ ✅ cl/543520, cl/562661 ❓ #63962, #64521
sqrt ✅ ✅ ❓
reciprocal ✅ ✅ ❓ #39551
reciprocalSqrt ✅ ✅ ❓ #39551
scale ✅ ✅ ❓
clamp ✅ ✅ cl/549060 ❓ #40426
ceil 🚧 🚧 ❓
floor 🚧 🚧 ❓
trunc 🚧 🚧 ❓
nearest 🚧 🚧 ❓ #61240
pmin 🚧 🚧 ❓
pmax 🚧 🚧 ❓
equal ✅ ✅ ❓
notEqual ✅ ✅ ❓
< (lessThan) ✅ ✅ ❓
<= (lessThanOrEqual) ✅ ✅ ❓
> (greaterThan) ✅ ✅ ❓
>= (greaterThanOrEqual) ✅ ✅ ❓
signMask ✅ ✅ cl/542362, cl/542363 ❓
x/y/z/w ✅ ✅ ❓
withX/Y/Z/W ✅ ✅ ❓
shuffle ✅ ✅ cl/559002, arm64 cl/562680, cl/562681, cl/563880, cl/563881 ❓
shuffleMix ✅ ✅ cl/559002, arm64 cl/562680, cl/562681, cl/563880, cl/563881 ❓
== / hashCode 🚧 in progress 🚧 in progress ❓ #43255
toString ✅ cl/547261 ✅ cl/547261 ❓ #63847
conversions (fromInt32x4Bits, fromFloat64x2) 🚧 ❓ ❓ #46182

Float64x2

Operation API Codegen optimizations Benchmark Issues
splat ✅ ✅ ❓
zero (+ identity) ✅ ✅ ❓
one (* identity) 🚧 🚧 ❓
const constructor 🚧 🚧 ❓ #31487
+ (add) ✅ ✅ ❓
- (subtract) ✅ ✅ ❓
* (multiply) ✅ ✅ ❓
/ (divide) ✅ ✅ ❓
unary - (sign negation) ✅ ✅ ❓
abs ✅ ✅ ❓
min ✅ ✅ cl/543520, cl/562661 ❓ #63962, #31048, #64521
max ✅ ✅ cl/543520, cl/562661 ❓ #63962, #31048, #64521
sqrt ✅ ✅ ❓
scale ✅ ✅ ❓
clamp ✅ ✅ cl/549060 ❓ #40426
ceil 🚧 🚧 ❓
floor 🚧 🚧 ❓
trunc 🚧 🚧 ❓
nearest 🚧 🚧 ❓ #61240
pmin 🚧 🚧 ❓
pmax 🚧 🚧 ❓
equal 🚧 🚧 ❓
notEqual 🚧 🚧 ❓
< (lessThan) 🚧 🚧 ❓
<= (lessThanOrEqual) 🚧 🚧 ❓
> (greaterThan) 🚧 🚧 ❓
>= (greaterThanOrEqual) 🚧 🚧 ❓
signMask ✅ ✅ cl/542363 ❓
x/y ✅ ✅ ❓
withX/Y ✅ ✅ ❓
== / hashCode 🚧 in progress 🚧 in progress ❓ #43255
toString ✅ cl/547261 ✅ cl/547261 ❓ #63847
conversions (fromFloat32x4) 🚧 ❓ ❓ #46182

Int8x16 / Uint8x16, Int16x8 / Uint16x8, Uint32x4, Int64x2

They are on my todo list. We don't have types for those yet (#63471).

The wasm level groundwork has landed: WasmI8x16 cl/551701, WasmI16x8 cl/551702, WasmI64x2 cl/551700.

🚧 In progress: a Uint8x16 value type on the VM (unboxing, arithmetic, equal, anyTrue), not uploaded yet.

Priorities

Because there's a lot to do, I decided to focus on what's most important, and deprioritize certain parts:

  • pmin/pmax & relaxed_min/relaxed_max: pmin/pmax benefits x86 and hurts arm64. relaxed_min/relaxed_max restores performance on arm64 to min/max levels. Since float64 and float32 use cases are most relevant for Flutter, where most of our users are on arm64, min and max is already optimal for most of our users.
  • x86 would benefit from commutativity exposed in TFA for better code generation.
  • operations that are not part of the Wasm SIMD API surface (e.g., rotateRight/rotateLeft) will be investigated later.

Third-party blockers

Use cases

Activity

  1. kevmoo
    assigned
    modulovalue
    on Aug 30, 2026
  2. kevmoo
    added the
    Feature
    issue type on Aug 30, 2026
  3. lrhn
    marked Feature completeness for SIMD feature additions #64073 as a duplicate of this issue on Aug 30, 2026
  4. kevmoo
    added
    area-sdkUse area-sdk for general purpose SDK issues (packaging, distribution, …).
    library-typed-data
    on Aug 31, 2026
  5. copybara-service
    added 4 commits that reference this issue on Sep 2, 2026

    [typed_data] Add Int32x4.notEqual

    9634f6b

    [vm] Add vsshr_4s to the arm64 assembler

    6ecea5d

    [vm] Intrinsify Int32x4.notEqual

    e9780eb

    [benchmarks] Add SimdFields, covering SIMD values in instance fields

    b1d85a7

  6. krll-kov commented on Sep 4, 2026

    @krll-kov
    krll-kov
    on Sep 4, 2026

    @modulovalue

    A use case, since the issue asks for them. I'm porting libwebp's VP8 encoder to Dart. The inner kernels are a typical 8-bit image workload: forward and inverse 4x4 DCT, a Hadamard distortion metric, sums of squared differences, and a quantizer. All of them read Uint8List at a fixed row stride, and libwebp has NEON and SSE versions of all of them.

    Here is what I ran into trying to vectorize them with dart compile exe on macos_arm64.

    Arithmetic over a populated Float32x4List is fine, 2.5x faster than the same scalar loop. The problem is getting byte data into the lanes. Each lane has to be filled separately, and the int to double conversion is about 5x slower in AOT than in JIT. That alone made a vectorized 16x16 SSD 3x slower than the scalar version. I filed that as #64209. Writing the conversion as 0.0 + x instead of x.toDouble() avoids it, and then the same kernel is 14% faster than scalar.

    With that out of the way, what still blocks the rest:

    • Int32x4 has no * (Support multiplication on Int32x4 #59816), so an integer kernel can't stay in integer lanes. The DCT multiplies by 2217, 5352, 20091 and 35468, and the quantizer multiplies by a reciprocal, so both need it.
    • Float32x4 can't cover for that. It's only exact up to 2^24, and the second pass of the DCT reaches about 8.7e7 while the quantizer's coeff * iq reaches about 1e9. Only the Hadamard metric and the SSD stay in the exact range.
    • An 8-bit lane type ([SIMD] Support for Int8x16 and Int8x16List. #63471) would suit this workload much better than widening bytes to float. A 16-byte row would load in one read instead of sixteen byte reads and four vector constructions, and that path is where the time goes on small blocks. A 16-bit lane type would be even more useful for the transforms, since the intermediates outgrow 8 bits right away.

    So right now SIMD pays off in one kernel out of five and loses in the others.

  7. copybara-service
    added 5 commits that reference this issue on Sep 7, 2026

    [vm] Finish Int32x4 unboxing, symmetric to the float SIMD value types

    8f1c524

    [vm] Use an SSHR-based sign mask on arm64

    019cd35

    [vm] Bound the Simd sign mask result range

    7611654

    [typed_data] Add Int32x4.allTrue

    4e33b0a

    [typed_data] Add Int32x4 signed ordered comparisons

    20eb157

  8. 34 remaining items

  9. copybara-service
    added 15 commits that reference this issue on Oct 5, 2026

    [vm] x64: add pshufd

    1bbaa31

    [typed_data] Implement IEEE min/max for Float32x4/Float64x2 and add a…

    41464f0

    [vm] arm64: add tbl

    1f0ec8e

    [vm, dart2wasm] lower 32x4 shuffles with pshufd, tbl and i8x16.swizzle

    294ad6f

    [vm] x64: add psrld

    5e219b5

    [vm] arm64: add ushl in its 4S form

    7dd74a9

    [vm] Intrinsify Int32x4 operator >>>

    afd6754

    [vm] Implement Float32x4 and Float64x2 SIMD fallbacks in Dart

    f7cb0fa

    [typed_data] Implement Float32x4 and Float64x2 clamp with min and max

    cd6cf73

    [vm, compiler] Fold Float32x4/Float64x2 min and max with Dart semantics

    f04bdad

    [vm] arm64: give 32x4 shuffles their own emitters

    59c7b72

    [vm] arm64: use rev64 for 32x4 shuffles

    28ee0b0

    [vm] arm64: use ext for 32x4 shuffles

    a49cd0c

    [vm] Remove the Float32 and Float64 native fallbacks of typed lists

    2276e18

    [vm] arm64: use ins for one-lane 32x4 shuffles

    ca962be
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

area-sdkUse area-sdk for general purpose SDK issues (packaging, distribution, …).library-typed-data

Type

Feature

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions