Release 0.2.0: armv7 neon, rv1106 rknn, int8, onnx ir · enthropy7/YSCV

Breaking release. Migration notes in CHANGELOG.md. Setup your agents and code.

onnx: optimizer runs on load (YSCV_ONNX_OPTIMIZE_OFF=1 or load_onnx_model_unoptimized to skip), def-use IR with dataflow fusion, attributes keyed by Attr, optimize_onnx_graph returns Result, non-SSA graphs rejected. Perf boost was seen on a few different models.
armv7: NEON across kernels, imgproc and video behind neon-v7 (nightly)
rknn: builds for 32-bit ARM, falls back to librknnmrt.so on RV1103/RV1106
int8: per-channel weights and asymmetric activations fixed, NHWC stream, fused QDQ boundaries, NEON widening GEMM for A53-class cores
new: crop_resize_bilinear*, IntegralImage, trace_pixel_polygons, ops::hog, KanLinear, V4L2 frame rate and controls
video: corrupt H.264/HEVC streams no longer panic or allocate gigabytes
now unsafe fn: AlignedVec::uninitialized, RknnBackend::wrap_fd/wrap_phys, RknnMem::as_slice/as_mut_slice, RgaBlender::blit/copy_at
renamed: YSCV_*AARCH64_LOW_K* → YSCV_*ARM_LOW_K*

etc. etc. etc.

one step closer to v1.0.0
few years left idk

Contributions

fix(yscv-video): use i32 widening multiply in NEON nv12_bt601_to_rgb by @Malinskiy in #7
Print out conv kernel and stride in profile_onnx_model_cpu by @Human9000-bit in #8
Fix onnx becnhmark job + add github actions lint job by @Human9000-bit in #10
Fix onnx becnh job #2 by @Human9000-bit in #12
Performance tuning for YOLO-like workloads by @Human9000-bit in #9
nix: initial commit by @r4v3n6101 in #15
Shape inference + deterministic graph evaluation by @Human9000-bit in #14
migrate to rustc's FxHasher by @Human9000-bit in #16
utilize load-time shape inference during runtime by @Human9000-bit in #17
onnx bench rework by @Human9000-bit in #18
nix: remove unused features from tests by @r4v3n6101 in #21
Document unlisted env vars + write down all env vars in a single file by @Human9000-bit in #23
Fix some things before yet another optimizer rework by @Human9000-bit in #24

thank you guys! ❤️