Shader shelf

WGSL kernels ported verbatim from the old llm/src/backend/wgpu/shaders/. None of them are wired into mr's op dispatch yet — they sit here as ready-to-use source so the implementation knowledge ships forward.

Wire-up prerequisites

Every kernel in this directory expects:

1. FrameAllocator for buffer lifetime (present: mr/src/wgpu_rs/alloc.rs).
2. dispatch_2d(n) for workgroup grid sizing when N > 65535 (present: mr/src/wgpu_rs/alloc.rs).
3. A graph executor that dispatches one op per compute pass (absent: mr/src/ir/executor.rs is a stub).

Once the executor lands, each shader moves one directory up (out of shelf/) and gets a Rust kernels/<name>.rs binding.

Inventory

Matmul kernels

• q4_matmul.wgsl — Q4_0 (32-value block, 18 bytes/block) matmul. Baseline. Sub-group reduce inside workgroup.
• q4k_matmul.wgsl — Q4_K (256-value superblock, 144 bytes/block) matmul. Reads raw superblocks directly on GPU, no CPU-side conversion. Inline decodes the packed 12-byte scale region using llama.cpp's get_scale_min_k4 bit extraction. Known needs-opt: superblock decode inside workgroup loop loses parallelism (audit: 13 tok/s vs Q4_0's 29).
• q6k_matmul.wgsl — Q6_K (256-value superblock, 210 bytes/block) matmul. Used for lm_head in most qwen models.
• ternary_matmul.wgsl — 2-bit packed BitNet weights ({-1, 0, +1}). Multiply-free inner loop: branches on encoding to fadd / fsub / skip. 8× less bandwidth than Q4 → memory-bound decode wins proportionally. Delivered 102 tok/s end-to-end on BitNet-2B in the old runtime.

Attention kernel

• attention.wgsl — single-pass fused SDPA decode. One workgroup per head, 256 threads. Online softmax via subgroupMax → 8-slot shared reduction → subgroupAdd for exp sum. All in registers; no second pass over the sequence. Requires enable subgroups; (WGSL spec-optional).

Fused ops

• fused_norm_q4.wgsl — RMSNorm inlined into Q4 matmul. Normalizes into shared memory, Q4 dequant reads from shared instead of DRAM. Saves 1 dispatch + 1 global write + 1 global read per linear layer. Limited to hidden_size ≤ 4096 (hardcoded shared-memory cap).
• fused_skip_norm.wgsl — Add + RMSNorm in one workgroup. Produces two outputs (normalized and raw residual) from one pass.

What did not make the shelf

From the old shaders/ dir, intentionally omitted:

• turbo_quant.wgsl — non-atomic f32 add masquerading as atomicAdd; correct only because each token gets its own workgroup. Needs a proper atomic path before shipping.
• q4k_embed.wgsl, q2k/q3k/q5k matmul — less commonly used quants; copy when a model needs them.
• activations.wgsl, elementwise.wgsl, argmax.wgsl, rms_norm.wgsl, rope.wgsl, layernorm.wgsl, norm.wgsl, conv.wgsl, spatial.wgsl, special.wgsl, adapter.wgsl, kv_cache.wgsl, f16_matmul.wgsl, f32_matmul.wgsl — straightforward ops whose mr/ native kernels already exist or are trivial to write fresh when needed.