warriors/trisha.md

Trisha

Triton VM warrior. Executes, proves, verifies, and deploys Trident programs on Triton VM with GPU-accelerated STARK proving.

source .tri → ProgramBundle → run / prove / verify / deploy
                                    ↓
                            GPU (Metal · Vulkan · DX12)
                            7 compute shaders · 4.3× speedup
                            24M H/s mining (M1 Max)

what it does

trisha run    program.tri                    # execute, print output
trisha prove  program.tri -o proof.toml      # STARK prove
trisha verify proof.toml                     # verify proof
trisha guess  program.tri --difficulty N     # GPU nonce mining
trisha deploy program.tri --state testnet    # on-chain (stub)

every command has a batch mode:

trisha prove batch a.tri b.tri c.tri --output proofs/ --max-parallel 4
trisha verify batch proofs/*.proof.toml

GPU backend

seven WGSL compute shaders. one pipeline. all platforms.

shader	operation
`goldilocks`	field arithmetic over 2⁶⁴ − 2³² + 1 via `vec2<u32>`
`ntt`	radix-2 Cooley-Tukey butterfly NTT
`poseidon2`	Poseidon2 permutation for Merkle trees
`fri`	FRI query folding and verification
`gemv`	matrix-vector multiply for polynomial evaluation
`tip5`	Tip5 batch hashing
`mine`	nonce search kernel

wgpu auto-selects the best native API: Metal on macOS, Vulkan on Linux, DX12 on Windows. GPU is on by default. CPU-only build: --no-default-features.

proving pipeline

ProgramBundle
    ↓
VM::trace_execution()      ← CPU, unavoidable
    ↓
Stark::prove()
    ├── iNTT    (interpolate) ←  GPU  shader: ntt.wgsl
    ├── NTT     (evaluate)   ←  GPU  shader: ntt.wgsl
    ├── Merkle               ←  GPU  shader: poseidon2.wgsl
    ├── FRI                  ←  GPU  shader: fri.wgsl
    ├── GEMV                 ←  GPU  shader: gemv.wgsl
    └── Tip5                 ←  GPU  shader: tip5.wgsl
    ↓
Proof → ProofFile → .proof.toml

NTT + Merkle share a single command encoder — zero per-layer CPU↔GPU sync.

proof file

.proof.toml: human-readable TOML envelope wrapping base64(bincode(Proof)).

[proof]
format = "stark-triton-v2"
program_name = "hello"
proving_time_ms = 280

[claim]
program_hash = ["1460305242624279511", ...]
public_output = ["42"]

[data]
proof = "AgIAAAAAAAD..."

patching

trisha patches triton-vm at build time rather than maintaining a fork. patches/apply.nu fetches upstream from crates.io, applies 5 surgical layers (GPU hooks), and writes .vendor/triton-vm/. one script, no diverging branches.

nu patches/apply.nu    # first time, or after version bump

build

nu patches/apply.nu    # vendor + patch triton-vm
cargo build --release
cargo test             # 15 integration tests
cargo install --path .

status

component	status
run / prove / verify	done
batch	done
GPU proving (7 shaders)	done
GPU mining (24M H/s)	done
proof file roundtrip	done
deploy (neptune)	stub
proof merging	planned
streaming NTT (2^25+)	planned

see roadmap/ for detailed plans.

workspace

repo	role
trident	compiler: source → TASM (ProgramBundle)
trisha	runtime: execute, prove, verify, deploy
triton-vm	STARK engine (vendored + patched)