Skip to content

Latest commit

 

History

History
53 lines (44 loc) · 2.16 KB

File metadata and controls

53 lines (44 loc) · 2.16 KB

yscv — Project Context

Current state of the yscv framework.

Project priority is CPU inference on edge devices (Cortex-A SBCs, low-power x86, drone boards). Other backends — wgpu, MPSGraph, RKNN, BLAS — are opt-in extensions and keep widening. See README.md and AGENTS.md for the rules and PR workflow.

Architecture

18 library crates, 2 apps, 1 examples crate.

yscv (umbrella re-export)
├── yscv-tensor          ← N-dimensional tensor, f32/f16/bf16, SIMD ops
├── yscv-kernels         ← CPU + GPU compute backends, SIMD, RKNN NPU
├── yscv-threadpool      ← Work-stealing pool + rayon scope abstraction, opt-in PersistentSection
├── yscv-autograd        ← Reverse-mode autodiff, BackwardOps routing
├── yscv-optim           ← 8 optimizers + Lookahead, 11 LR schedulers
├── yscv-model           ← 39 layers, 17 architectures, LoRA, trainer
├── yscv-imgproc         ← 160 image ops, u8/f32 SIMD
├── yscv-video           ← H.264/HEVC/AV1 decode, V4L2, MJPEG, H.264 encode, MAVLink, overlay, framebuffer
├── yscv-video-mpp       ← Rockchip MPP hardware video encoder
├── yscv-detect          ← YOLOv8/v11, NMS, heatmap, RoI align
├── yscv-recognize       ← Cosine matching, VP-Tree ANN
├── yscv-track           ← DeepSORT, ByteTrack, Kalman
├── yscv-eval            ← Metrics, 8 dataset adapters
├── yscv-onnx            ← 121-op ONNX runtime, INT4/INT8 quantization, GPU, Metal, KV-cache, generation
├── yscv-pipeline        ← Pipeline framework: ONNX/Metal/RKNN dispatch, TOML config
└── yscv-cli             ← Inference CLI, camera diagnostics

Metrics

Metric Value
Tests 1,807
ONNX operators 121
Imgproc ops 160
Model architectures 17
WGSL + Metal shaders 61 + 4
SAFETY-commented unsafe blocks 220
Fuzz targets 5

Build

  • Edition 2024, MSRV 1.94
  • Release: lto = "thin", codegen-units = 1
  • BLAS: Accelerate (macOS), OpenBLAS (Linux)
  • GPU: wgpu (Vulkan/Metal/DX12) + Metal-native
  • NPU: RKNN (Rockchip, --features rknn)