GPU backend

trisha accelerates seven hot paths via custom WGSL compute shaders dispatched through wgpu. a single backend targets Metal (macOS), Vulkan (Linux/Windows), and DX12 (Windows) without platform-specific code.

seven shaders

Goldilocks field in WGSL

WGSL has no native u64. every field element is stored as vec2<u32> — (lo, hi) pair in little-endian 32-bit halves. Montgomery form is used for multiplication to avoid 128-bit intermediates.

// Goldilocks: p = 2^64 - 2^32 + 1
// R = 2^32 - 1, R^2 mod p = 0xFFFFFFFE00000001
struct Gl { lo: u32, hi: u32 }

fn gl_add(a: Gl, b: Gl) -> Gl { ... }
fn gl_mul(a: Gl, b: Gl) -> Gl { ... } // Montgomery
fn gl_inv(a: Gl) -> Gl { ... }        // Fermat: a^(p-2)

kernel fusion

NTT and Merkle hashing share a single command encoder with zero per-layer CPU sync. the CPU submits the full sequence — butterfly passes → hash layers — and the GPU executes without round-tripping to host between stages.

VRAM budget management

at init, wgpu queries the adapter for max_buffer_size and max_storage_buffers_per_shader_stage. trisha computes a VRAM budget and:

• buffers that fit → GPU dispatch
• buffers that overflow → chunked GPU dispatch or CPU fallback

this means trisha handles 4GB GPU laptops and 192GB M1 Ultra workstations with the same code path.

persistent buffers

three buffers are allocated once and reused across proofs:

• twiddle factors (NTT roots of unity)
• two_inverse (constant for butterfly normalization)
• Tip5 round constants (256 × u64)

allocation cost is paid once at WgpuBackend::new().

staging pool

GPU results must be readback to CPU for verification. eight staging buffers are pre-allocated in a pool. acquire() / release() prevent re-allocation across the 7 readback sites.

current speedup

on poseidon2.tri (2-op program, M1 Max):

large programs (2^25+ trace rows) fall back to CPU at NTT; gpu-proving-at-scale is the fix.

enabling / disabling GPU

GPU is on by default (--features gpu). CPU-only build:

cargo build --no-default-features

GpuBackend::try_new() returns None if no GPU is found; TrishaWarrior falls back to CpuBackend automatically.