Module buffer 

Buffer arena for the wgpu backend. Mirrors the rlx-metal arena shape: pre-plan one big storage buffer at compile time, sub-allocate per-node offsets at known positions, treat I/O as write_buffer / read_buffer against those offsets.

wgpu’s storage buffers are fine for both reads and writes from compute shaders; there’s no shared-memory requirement at the API level (unlike Metal where StorageModeShared matters). On Apple Silicon wgpu’s Metal backend gives us unified memory automatically.

Structs

Arena

One contiguous arena buffer + per-node byte offsets. Lives for the entire executable graph’s lifetime.

ReadbackLayout

Layout for batched output readback into a staging buffer.

ReadbackStaging

Reusable MAP_READ staging buffer for output readback.

TinyReadbackStaging

Fixed 256 B MAP_READ staging for scalar (≤16 B) readback — avoids map_buffer_on_submit + full-layout decode on MoltenVK hot paths.

Functions

decode_mapped_readback_f32

Decode f32 outputs after submit + map callback (used with schedule_readback_map).

decode_tiny_mapped_f32

After submit: decode one f32 vector from an already-mapped tiny staging buffer.

encode_readback_copies

Append arena→staging copies to an encoder (no submit).

map_readback_f32

Map staging after submit and decode f32 outputs (one poll).

plan_f32_uniform

Plan memory using f32-sized slots regardless of declared IR dtype, with liveness-aware slot reuse (see rlx_compile::memory::plan_memory_f32_uniform).

read_f32_many_pooled

Read several nodes with one submit + one poll (contiguous staging layout).

read_f32_pooled

Read one node via a reused staging buffer (one submit + one poll).

read_tiny_f32_after_submit

After submit: map only len bytes and decode one f32 vector.

schedule_readback_map

Schedule map_async on the encoder so mapping starts with submit (wgpu 29+).

use_tiny_readback

True when fused readback can use the tiny scalar fast path.

wait_readback_map

Poll until a readback map callback completes (fast path for tiny outputs).