Skip to content

perf(backend-cpu): primitive FP32 fast paths for the eager CPU ops; cache DirectCpuExecutionContext.ops (#949) #950

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
michalharakal merged 1 commit into from
Aug 11, 2026
Merged

perf(backend-cpu): primitive FP32 fast paths for the eager CPU ops; cache DirectCpuExecutionContext.ops (#949) #950

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 20 additions & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,26 @@

## [Unreleased]

### Performance

- **Primitive FP32 fast paths for the eager CPU ops** (`skainet-backend-cpu`,
[#949](https://github.com/SKaiNET-developers/SKaiNET/issues/949)). The generic paths in
`DefaultCpuOps` paid, per element: two `IntArray` allocations for broadcast index mapping, a
vararg-spread boxed `data.get`, a KClass `when (dtype)` comparison, and a boxed lambda
round-trip. On ART that overhead dominated LLM decode — 83% of end-to-end SmolLM2-135M decode
on a Pixel 8a was non-matmul overhead even with the NEON backend doing every matmul. The hot
ops now run flat primitive loops over the dense `FloatArray` buffer (falling back to the
generic path for other dtypes/layouts): binary/scalar arithmetic (incl. last-dim bias
broadcast, mirroring the JVM vector path's coverage), the activation family
(relu/sigmoid/silu/gelu/…), unary math (sqrt/exp/log/sin/cos/tanh/pow/…), softmax and
logSoftmax along the last dim (also removing an O(n²)-per-slice max/denominator recompute),
sum/mean reductions, concat (block copy), and reshape/flatten (buffer copy). Benefits every
non-JVM target — Android, Kotlin/Native, JS/Wasm; the JVM ops class keeps its Panama/FFM
specializations on top.
- **`DirectCpuExecutionContext.ops` is cached.** The getter previously constructed a fresh ops
instance on every access, re-running per-instance lazy kernel resolution in the eager hot
loop ([#949](https://github.com/SKaiNET-developers/SKaiNET/issues/949)).

## [0.39.0] - 2026-08-10

Headline: **on-device AI on Android becomes real.** A JNI NEON kernel backend
Expand Down
6 changes: 5 additions & 1 deletion ...s/skainet-backend-cpu/src/commonMain/kotlin/sk/ainet/context/DirectCpuExecutionContext.kt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,13 +34,17 @@ public class DirectCpuExecutionContext @kotlin.jvm.JvmOverloads constructor(
usagePercentage = 0.0
)
private val opsFactory = platformDefaultCpuOpsFactory()
// Cached: the getter used to build a fresh ops instance per access, which
// re-ran the per-instance lazy kernel resolution and allocated on every
// `ctx.ops` touch in the eager hot loop (#949).
private val cachedOps: TensorOps by lazy { opsFactory(tensorDataFactory) }
override val memoryInfo: MemoryInfo
get() = _memoryInfo
override val observers: ExecutionObserverRegistry
get() = observerRegistry

override val ops: TensorOps
get() = opsFactory(tensorDataFactory)
get() = cachedOps

override val hooks: sk.ainet.lang.nn.hooks.ForwardHooks?
get() = _hooks
Expand Down
Loading
Loading