CUDA Miner new hard fork

BTCW CUDA Miner — RTX 4080/4070 SUPER new Hard Fork

CUDA miner for the BTCW NO_EXT_WORK/NEWFORK signing-work algorithm. The
RTX 4080 SUPER build uses the W26 lookup-table profile and BTCW's original
224-byte shared-memory payload.

Algorithm

For each block template, the BTCW node supplies a 32-byte secp256k1 secret
key, 160 bytes of context, and a fixed 32-byte hash_no_sig message.

The GPU searches a 32-bit test_case. Each candidate is evaluated as follows:

1. Form RFC6979 extra entropy as LE32(test_case) || 28 zero bytes.
2. Generate deterministic ECDSA nonce k from the secret key, reduced
   message, and extra entropy. Test case zero is skipped because zero is
   reserved by the result mailbox.
3. Compute R = kG on secp256k1 and r = R.x mod n.
4. Compute s = k^-1 * (hash_no_sig + r * secret_key) mod n.
5. Normalize s to low-S form and DER-encode (r,s).
6. Accept only 70- or 71-byte DER signatures.
7. Compute SHA256(SHA256(DER_signature)).
8. Compare the result with the fixed Stage-2 target 2^228 - 1, equivalent
   to 28 leading zero bits in displayed hash order.
9. Return a successful test_case through the existing 64-bit POSIX shared
   memory nonce mailbox.

The BTCW node remains the final authority. It regenerates the signature using
CKey::Sign(hash_no_sig, ..., false, test_case) and validates it before block
submission.

CUDA design

Important optimizations include:

• fixed-layout, word-native RFC6979/HMAC-SHA256;
• reuse of secret-key and reduced-message SHA prefix state;
• secp256k1 GLV scalar splitting;
• a five-group signed W26 fixed-base lookup table of approximately 9 GiB;
• XYZZ mixed point addition;
• 128-candidate batched scalar inversion;
• two 64-candidate field-inversion sub-batches;
• PTX 32-bit Comba field and scalar multiplication;
• specialized DER SHA256d;
• direct comparison of the final SHA word with the fixed target.

The miner runs an RFC6979 GPU self-test at startup. It compares the optimized
extra-entropy implementation with the generic reference implementation for
known test cases and aborts if they differ.

Requirements

• Linux x86-64
• NVIDIA driver
• CUDA Toolkit with nvcc
• NVIDIA RTX 4080 SUPER
• substantially more than 12 GiB of free GPU memory for the W26 configuration
• a compatible BTCW NEWFORK node exposing /shared_mem