v0.8.0 — Pure Go WebGPU (gogpu/wgpu)

Pure Go WebGPU — No Shared Libraries, No CGO

Born's GPU backend now uses gogpu/wgpu v0.26.8 — a pure Go WebGPU implementation. No more DLL/SO downloads. Just go build.

Highlights

• True single binary deployment — GPU support built into the executable
• Vulkan primary backend — stable on Windows, Linux, macOS
• Zero runtime dependencies — no wgpu-native, no shared libraries
• Validated — 105 GPU tests pass, real model training (20 epochs, 0 crashes)

What Changed

• Replaced go-webgpu/webgpu (Rust FFI) with gogpu/wgpu (pure Go)
• Fixed PipelineLayout lifetime for Vulkan SetBindGroup
• Fixed lazy ops buffer lifetime with immediate submit
• Fixed lazy chain data propagation
• Added runtime.KeepAlive guards for GC safety
• Added Poll(PollWait) in Release for clean shutdown
• Sign/Abs operators (#59 by @bennibbelink)

Breaking Changes

• wgpu_native shared library no longer needed (or used)
• IsAvailable() now verifies compute shader support (not just adapter presence)

Full changelog

See CHANGELOG.md

v0.7.16

v0.7.16 — Community Contributions, ONNX 49 Operators, Bugfixes

Third external contributor @gmohmad with 5 PRs! Plus continued work from @bennibbelink.

Added

• ONNX LayerNormalization operator (#47 by @gmohmad)
• BatchMatMul NumPy-style broadcasting — supports 2D×3D, singleton batch dims (#49 by @gmohmad)
• ONNX comparison operators: Greater, GreaterOrEqual, Less, LessOrEqual (#56 by @gmohmad)
• ONNX logical operators: Not, And, Or, Xor (#56 by @gmohmad)
• ONNX Erf operator (#56 by @gmohmad)
• Broadcasting for boolean and comparison ops in CPU backend (#56 by @gmohmad)
• tensor.BroadcastShapesMatMul public API
• ONNX AttributeProto tensor attribute parsing (#53 by @gmohmad)

Fixed

• CPU backend: prevent inplace mutation when operands alias — Mul(x,x) no longer corrupts input (#55, fixes #45)
• Squeeze scalar handling: returns Shape{} (scalar) instead of Shape{1} (#50 by @gmohmad)
• ONNX AttributeProto parser: correct protobuf field numbers, non-packed encoding support (#53 by @gmohmad)
• CI: added test gate job for branch protection (#52)

Refactored

• ConvDims/PoolDims parameter structs (#46 by @bennibbelink), moved to shared internal/tensor/ package (#48)
• 14 helper functions extracted from conv2d/maxpool2d inner loops (#17) — compiler-inlined, Conv2D batch path ~28% faster
• Resolved 86 gosec lint errors after golangci-lint v2.11.4 upgrade (#39)

Stats

• ONNX operators: 39 → 49
• PRs merged: 12
• Issues closed: 8 (#16, #17, #43, #44, #45, #48, #51, #54)
• Contributors: @gmohmad (5 PRs), @bennibbelink (1 PR)

Thank you! 🙏

A huge thanks to our contributors:

• @gmohmad — 5 PRs, 5 issues filed. Found real bugs (inplace aliasing, Squeeze scalar, AttributeProto parsing), added 10 new ONNX operators, and implemented broadcasting. Outstanding work!
• @bennibbelink — ConvDims/PoolDims refactoring that improved code quality across the conv2d/maxpool2d stack.

Community contributions make Born better. If you'd like to contribute, check our open issues!

Full Changelog: v0.7.15...v0.7.16

v0.7.15 — Erf Operator (Community Contribution)

Second External Contribution!

Thanks to @bennibbelink for a full vertical slice across the entire stack!

Added:

• Erf (error function) operator — element-wise error function
• Backend interface, CPU (math.Erf), WebGPU (Abramowitz & Stegun shader)
• Autodiff with correct backward pass: 2/√π · exp(-x²)
• Mock backend, Tensor API (tensor.Erf())
• Comprehensive tests: forward + backward, float32/float64, edge cases (Inf, NaN)

Links:

• PR: #37 by @bennibbelink

go get github.com/born-ml/born@v0.7.15

v0.7.14 — ONNX Equal Operator (Community Contribution)

First External Contribution!

Thanks to @jsully1720 for the first community PR!

Added:

• ONNX Equal operator — binary element-wise comparison returning bool tensor
• New comparison_ops.go category for ONNX comparison operators
• registerComparisonOps() wired into operator registry

ONNX operators: 38 → 39

Links:

• PR: #34 by @jsully1720
• Issue: #35

go get github.com/born-ml/born@v0.7.14

v0.7.13 - ABI Compliance Fixes

Changes

Dependencies Updated

Package	Old	New
go-webgpu/webgpu	v0.4.0	v0.4.1
go-webgpu/goffi	v0.4.0	v0.4.1 (indirect)

Upstream Bug Fixes (ABI compliance)

• Float32 encoding: correct XMM bit patterns via math.Float32bits
• AMD64 Unix stack: arguments beyond 6 GP registers properly pushed to stack
• ARM64 Unix stack: arguments beyond 8 GP registers correctly spilled to stack
• AMD64 struct returns (9-16 bytes): RAX+RDX register pair properly assembled
• AMD64 sret pointer: structs > 16 bytes use caller buffer as first argument (RDI)
• ARM64 HFA spilling: Homogeneous Floating-Point Aggregate overflow follows AAPCS64

Upstream Enhancements

• runtime.KeepAlive prevents GC of argument pointers during FFI calls
• ErrTooManyArguments overflow detection for calls exceeding 15 arguments

Impact

Critical ABI correctness fixes for multi-platform GPU backend reliability.

Full Changelog: v0.7.12...v0.7.13
Upstream: https://github.com/go-webgpu/webgpu/releases/tag/v0.4.1

v0.7.12 - FFI Hardening & Library Loading

Changes

Dependencies Updated

Package	Old	New
go-webgpu/webgpu	v0.3.2	v0.4.0
go-webgpu/goffi	v0.4.0	v0.4.0 (unchanged)

Upstream Improvements

• Null handle guards on 27 public FFI methods — prevents SIGSEGV on nil/released objects
• ptrFromUintptr helper eliminates all go vet unsafe.Pointer warnings
• WGPU_NATIVE_PATH env var for custom wgpu-native library path
• loadLibrary returns (Library, error) with proper error propagation
• Windows DLL eager loading — errors surface at init, not at first use
• Enhanced Init() error messages with library path and remediation suggestions
• 85 new null guard test cases upstream

Impact

Significantly improved safety and debuggability of GPU backend initialization.

Full Changelog: v0.7.11...v0.7.12
Upstream: https://github.com/go-webgpu/webgpu/releases/tag/v0.4.0

v0.7.11 - Crosscall2 Callback Integration

Changes

Dependencies Updated

Package	Old	New
go-webgpu/webgpu	v0.3.1	v0.3.2
go-webgpu/goffi	v0.3.9	v0.4.0 (indirect)

Upstream Improvements

• crosscall2 integration — callbacks now work from C-library-created threads (Metal, wgpu-native)
• fakecgo trampoline register fixes synced with purego v0.10.0

Bug Fixes

• Xavier init test: added float32 rounding tolerance to prevent false failures

Impact

Improved callback reliability on macOS Metal and native WebGPU implementations.

Full Changelog: v0.7.10...v0.7.11
Upstream: https://github.com/go-webgpu/webgpu/releases/tag/v0.3.2

v0.7.10

🔧 Dependencies Update + Lint Cleanup

Update WebGPU backend to v0.3.1 with critical ARM64 callback fix. Clean up 101 stale //nolint:gosec directives.

Updated Dependencies:

• go-webgpu/webgpu v0.3.0 → v0.3.1
• go-webgpu/goffi v0.3.8 → v0.3.9 (indirect)

Upstream Fixes:

• ARM64 callback trampoline rewrite — fixes LR corruption for callbacks at index > 0
• Symbol rename to prevent linker collision with purego

Code Quality:

• Removed 101 unused //nolint:gosec directives (gosec linter updated, no longer flags these)
• Standardized remaining nolint comments to short format
• 0 lint issues across all platforms

Impact: Critical fix for macOS Apple Silicon and Linux ARM64 users.

Links:

• Upstream release: go-webgpu v0.3.1
• PRs: #29, #30

v0.7.9

🔧 Dependencies Update

Update WebGPU backend to v0.3.0 with new capability-querying API and typed errors.

Updated Dependencies:

• go-webgpu/webgpu v0.2.1 → v0.3.0

New Upstream Features Available:

• Surface.GetCapabilities() — query supported formats, present modes, alpha modes
• Device.GetFeatures() / Device.HasFeature() — feature enumeration
• Device.GetLimits() — device limits (experimental)
• Typed errors with errors.Is() / errors.As() support (ErrValidation, ErrOutOfMemory, ErrInternal, ErrDeviceLost)
• Resource leak detection via SetDebugMode(true) / ReportLeaks()

Links:

• Upstream release: go-webgpu v0.3.0
• PR: #28

v0.7.8 - GoGPU Ecosystem Integration (Phase 1)

🔧 GoGPU Ecosystem Integration (Phase 1)

Migrate WebGPU backend to unified gputypes for future dual-backend support.

Updated Dependencies

Package	Old	New
go-webgpu/webgpu	v0.1.4	v0.2.1
go-webgpu/goffi	v0.3.7	v0.3.8
gogpu/gputypes	-	v0.2.0 (new)
dlclark/regexp2	v1.10.0	v1.11.5
google/uuid	v1.3.0	v1.6.0

Changes

• Migrated all WebGPU types from wgpu.* to gputypes.*:
  • BufferUsage, BufferUsageStorage, BufferUsageCopySrc, BufferUsageCopyDst
  • PowerPreferenceHighPerformance
• Updated 10 files in internal/backend/webgpu/
• Fixed 3 prealloc warnings in linter

Why This Matters

Prepares codebase for Pure Go WebGPU backend (gogpu/wgpu):

• Unified type system enables future dual-backend architecture
• Build tags will allow: go build (Rust FFI) vs go build -tags purego (Pure Go)

Links

• Upstream: go-webgpu v0.2.1
• GoGPU ecosystem: github.com/gogpu
• PR: #27

Installation

```bash
go get github.com/born-ml/born@v0.7.8
```