Module q8_0_reference 

The scalar twin of crate::q8_0_shader, and the byte→word packing both the twin and the real upload path share.

What this is for

This repo’s rule for a kernel it cannot run is the one every unsafe SIMD arm here already follows and that ferrox-cuda’s mul_mm_ref follows for CUDA: a scalar twin implementing identical arithmetic, held against an independent reference.

matvec_reference is not “a Q8_0 matvec that should agree”. It is a transcription of the shader: same word-indexed byte extraction, same integer f16 decode, same (d * q) * x term shape, same block-ascending then element-ascending accumulation order, same row < rows guard. Read it beside emit_main; a line here with no counterpart there is a bug in one of them.

Two independent references check it, and neither shares any code with it:

• f16_to_f32 against the half crate, over all 65,536 f16 bit patterns.
• the whole matvec against ferrox_quant::dequant_q8_0 followed by a plain dot product – a different unpacker, a different f16 decoder, and the same numbers.

What that does not establish is that the emitted SPIR-V says what this file says; the two are hand-transcribed from each other. On a machine with a Vulkan driver, crate::device closes that gap by running the real shader and comparing. See the crate docs for whether that has happened.

Functions

f16_to_f32

The shader’s integer f16 decode, in Rust.

matvec_reference

Host emulation of the Q8_0 matvec shader.

pack_words

Pack raw bytes into the uint[] a storage buffer holds, zero-padding the tail.