Green Cycle — Mining Performance Improvement

All tests are passing. Choose ONE improvement activity and execute it. Always benchmark before and after to measure impact.

Priority Order (pick the highest-priority item you can make progress on)

Tier 1: Direct Performance Gains

1. Add on-GPU difficulty checking to CUDA backend

• Metal already has FC_PROOF_MODE — study gpu/metal.rs (search for FC_PROOF_MODE, found_flag)
• Port the concept to gpu/cuda.rs and the inline CUDA kernel
• GPU kernel writes found_flag + nonce + hash when difficulty is met
• Eliminates transferring ALL hashes back to CPU — huge bandwidth savings
• Verify with cargo test --workspace and benchmark with bench --backend cuda

2. Add on-GPU difficulty checking to OpenCL backend

• Same concept as above, port to kernels/uhash.cl and gpu/opencl.rs
• Add found_flag/found_data buffers, kernel writes result on match
• Verify and benchmark with bench --backend opencl

3. Add on-GPU difficulty checking to WGPU backend

• Port to kernels/uhash.wgsl and gpu/wgpu.rs
• WGSL has limitations (u32 only) — handle carefully
• Verify and benchmark with bench --backend wgpu

4. Replace per-batch thread spawning with persistent thread pool on CPU

• cpu/parallel.rs spawns fresh OS threads per find_proof_batch call
• Use std::thread::scope or a simple channel-based pool (no rayon — it was intentionally removed)
• Measure thread creation overhead with small batch sizes
• Benchmark with bench --backend cpu --threads N

Tier 2: Tuning and Testing

5. Add performance regression tests

• Create tests that measure hashrate and fail if it drops below a threshold
• Test each backend independently
• Useful for catching accidental regressions

6. Optimize CLI default batch_size

• CLI uses 4096 but ProverConfig default is 65536
• Profile the optimal batch size for CPU and each GPU backend
• Update the CLI default to match optimal

7. Improve auto-tuning coverage

• Profile additional parameter combinations
• Add tuning for CPU (optimal batch size per thread count)
• Improve WGPU memory detection (try adapter limits)

Tier 3: Algorithm-Level Optimizations

8. Parallelize chains within a single hash (CPU SIMD)

• 4 independent chains could use SIMD (4-wide AES on x86, NEON on ARM)
• Would require significant refactoring of hash.rs
• High risk, high reward — be very careful with correctness

9. Optimize GPU shader memory access patterns

• Profile and optimize scratchpad access in GPU kernels
• Consider coalesced memory access patterns for GPU architectures
• Minimize bank conflicts in shared memory

10. Add Metal shader optimizations

• Leverage Apple Family-specific features more aggressively
• Profile different threadgroup memory usage patterns
• Test SIMD group functions for intra-group communication

Rules

Pick exactly ONE activity per cycle

Always measure before AND after (run benchmarks, report numbers)

Follow existing code patterns in the backend you're modifying

Run cargo test --workspace to verify correctness

If optimizing a GPU kernel, verify hash output matches CPU reference

Rebuild and test after every change

Output: IMPROVED: <one-line summary with before/after numbers if applicable>