Restore orphaned moe_fp32 solutions (reference-scalar + ggml baseline)

Tag q4_k_m defs correctness:sqnr + stable SiLU (moe)

Tag q4_k_m defs correctness:sqnr + stable SiLU (moe)

Tag q4_k_m defs correctness:sqnr + stable SiLU (moe)

Tag q4_k_m defs correctness:sqnr + stable SiLU (moe)

Tag q4_k_m defs correctness:sqnr + stable SiLU (moe)

Tag q4_k_m defs correctness:sqnr + stable SiLU (moe)

Tag q4_k_m defs correctness:sqnr + stable SiLU (moe)

Tag q8_0 MoE definitions correctness:sqnr

Tag q8_0 MoE definitions correctness:sqnr

Rewrite llama.cpp q8_0 reference-scalar solution signature, now compatible with bench/dataset/raw.py

Update all w8a8ch definitions' correctness reference, aligns with what llama.cpp low-bit kernel correctness reference style: dequant input, then calculate in full precision

Stable SiLU in moe references (clip before exp; prevents q8_0 overflow-driven inf)

Add kernel in different precision:

Definitions, solutions, and evaluation trace updates for ncnn kernels

Add reference-scalar llama.cpp solutions (naive scalar baseline, 16 defs)

Add llama.cpp baseline solutions (baseline-llamacpp-arm, 16 defs)

add baseline solution source at all neural network kernel definitions

Update definition and workload for CV model kernels:

Resolve conflict and recover nn kernel traces

local snapshot before syncing with HF

Replace pre-PR#15 conv2d-only trace with current simd-loop set (definitions + workloads + reference/autovec solutions)

reset convolution kernel definition, workload, and solution; cin now is an axe of input shape

Upload folder using huggingface_hub

Update bench-trace: sync definitions, solutions, workloads, traces (#1)

Add simd-loop reference-scalar solutions (self-contained, harness embedded)

Add simd-loop workloads (edge + perf sizes for 23 loops)

Add simd-loop definitions (23 loops: scalar-output, array-output, inplace-sort)

Add dataset card

Upload folder using huggingface_hub

initial commit