This repo provides:

• A 32-way bitsliced AES-128 implementation in CUDA
• Compile-time round keys (no runtime key storage) using templates/constexpr
• Implicit ShiftRows (via column remapping) in the main rounds
• Tools for input generation, unpacking, and verification

cmake -S . -B build -DGPU_ARCH=90 # adjust SM (e.g. 86, 89) cmake --build build -j

Build prints ptxas info; expect 0 bytes lmem for kernels.

python3 tools/make_inputs.py --grid 2x1x1 --block 256x1x1 --seed 1 # writes inputs/run_YYYYmmdd_HHMMSS/{plaintexts.bin, plaintexts.hex, slices_u32_le.bin, meta.json}

• groups = grid.x*grid.y*grid.z * block.x*block.y*block.z
• Each thread processes 32 plaintexts (one 128-bit state in bitslice).
• Bitsliced input layout: groups * 128 little-endian uint32_t slices.

IN=inputs/run_*/slices_u32_le.bin OUT=outputs/run_full_slices_u32_le.bin ./build/cuda-aes-full "$IN" "$OUT" 2x1x1 256x1x1

• Stores bitsliced ciphertext to OUT.

python3 tools/verify_outputs.py \ --meta inputs/run_*/meta.json \ --slices_out "$OUT" \ --keyhex 2b7e151628aed2a6abf7158809cf4f3c

• Also writes:
  • outputs/ciphertexts_from_cuda.bin (unpacked CUDA output)
  • outputs/ciphertexts_from_python.bin (Python AES-128 reference)

• S-box only:

  ./build/cuda-aes-sbox-only inputs/.../slices_u32_le.bin outputs/sbox_only.bin 256

• MixColumns only:

  ./build/cuda-aes-mix-only inputs/.../slices_u32_le.bin outputs/mix_only.bin 256

• Round keys are computed in templates (include/aes_keys.hpp).
• AddRoundKey is emitted as compile-time ~reg for key-bit=1 (xor with all-ones), using no registers or memory for keys.
• Edit the key in src/aes_full_kernel.cu inside run_aes_bs_full():

  using MyKey = StaticKey< 0x2B,0x7E,0x15,0x16, 0x28,0xAE,0xD2,0xA6, 0xAB,0xF7,0x15,0x88, 0x09,0xCF,0x4F,0x3C >;

• ShiftRows in the main rounds is handled implicitly by calling MixColumns with bytes {0,5,10,15}, {4,9,14,3}, {8,13,2,7}, {12,1,6,11}.
• Final round does SubBytes + ShiftRows only, then applies the last round key. A register-only permutation is used once.