Commit Graph

9 Commits

Author SHA1 Message Date
155d7b0fbd SPMD API overhaul: gather/scatter, job & packaging updates
- ISPMDLane: add MaskGather, MaskStore, Scatter, MaskScatter; update MaskLoad/Gather signatures for hardware parity
- WideLane/ScalarLane: implement new methods with HW/fallback logic
- MathV: gather/mask-gather now delegate to lane methods
- Vector2/3/4: add CompressStore, Scatter, MaskScatter
- SPMD jobs/tests/README: migrate to new APIs for correctness
- Use Unsafe.BitCast instead of Unsafe.As/AsRef
- Add SPMDUtility for gather index extraction
- Job system: add ICustomJob<TSelf>, ScheduleCustom overload
- FreeList concurrency obsolete; always thread-safe
- NuGet: include LICENSE/README, set license/readme in .csproj
- Docs: update SPMD usage, clarify safety notes
- Minor: doc fixes, CompressStore test improvements
2026-05-04 13:56:49 +09:00
99fcbec753 Refactor SPMD jobs for true vectorized/masked execution
- Change IJobSPMD.Execute to (indices, mask, ctx) signature for all arities, enabling proper vectorized/masked SPMD execution.
- Update all SPMD job wrappers, extension methods, and test jobs to use new interface.
- Add AVX2 gather/masked gather support to MathV.GatherVector2/3/4 and related methods; use [ConstantExpected] byte scale.
- Improve gather/select logic, pointer arithmetic, and overloads for ref/int* index access.
- Refactor GGXMipGenerationBenchmark and jobs for SPMD, with per-mip-level vectorized jobs and improved memory access.
- Clean up code, fix naming, update comments, and bump version to 1.3.6.
2026-05-03 23:32:04 +09:00
eb01e557d5 Relax SPMD job constraints, improve safety and docs
Removed unmanaged struct requirement from SPMD job wrappers and extension methods, allowing managed types. Updated all wrappers and extension methods to require only the interface constraint. Refactored SPMD test jobs to use safe ref-based Store overloads. Improved README and docs with clearer debug/mimalloc instructions and a better SPMD example. Cleaned up Program.cs by removing obsolete experimental code. Enhanced math precision in GGXMipGenerationBenchmark. Updated T4 template to generate new constraints and APIs.
2026-05-01 12:39:37 +09:00
9f7507ba71 Refactor SPMD lane abstraction and add gather support
- Rename ISPMD interfaces to ISPMDLane for clarity
- Add gather and mask load methods to ISPMDLane, implement for ScalarLane and WideLane
- Add GetUnsafePtr() for direct pointer access
- Update MathV and vector types to use new interface and gather methods
- Update SPMD job interfaces and implementations to ISPMDLane
- Improve hash codes, range checks, and safety checks in vector types
- Update codegen templates for new interface/methods
- Refactor SPMD jobs to use gather methods for efficient vectorized access
2026-04-25 11:50:51 +09:00
cfd01eb9b6 Refactor SPMD job system, add GGX mipmap benchmark
- Replace IJobSPMD with T4-generated, multi-type SPMD job interfaces and wrappers (up to 8 numeric types)
- Extend ISPMD with Cast/BitCast; implement for ScalarLane and WideLane (SIMD-aware)
- Add unary minus, scalar-lane, and lane-scalar operators to Vector2/3/4; improve Select methods
- WideLane now partial with T4-generated Cast/BitCast (SIMD conversions)
- SPMD job Execute now requires unmanaged TLane; update all usages and benchmarks
- Add GGXMipGenerationBenchmark with vectorized and scalar paths, SkiaSharp output
- Update project files: add generated code, SkiaSharp, bump version to 1.3.0
- Misc: fix formatting, method signatures, FreeList logic
2026-04-25 01:50:06 +09:00
8231d6df60 backup 2026-03-30 12:47:29 +09:00
37d548085e Refactor job API: add JobExecutionContext, update tests
Major breaking change: job interfaces now use JobExecutionContext
instead of threadIndex, enabling thread-aware and dynamic job
dispatching. Updated all job system, SPMD, and test code to match.
Collections improved with new methods and clearer enumerators.
Renamed IJobScheduler.WaitComplete to Wait. Incremented project
versions. Includes bug fixes, documentation, and style updates.
2026-03-04 11:43:39 +09:00
4f964b2d2a Add Vector type in SPMD to total of load vector size * lane width of number into memory for simd calculation. 2026-02-13 21:47:05 +09:00
75d33d0763 Add Vector3<TLane, TNumber> 2026-02-13 12:51:22 +09:00