Optimize OCB offset calculation
* cipher/cipher-internal.h (ocb_get_l): New. * cipher/cipher-ocb.c (_gcry_cipher_ocb_authenticate) (ocb_crypt): Use 'ocb_get_l' instead of '_gcry_cipher_ocb_get_l'. * cipher/camellia-glue.c (get_l): Remove. (_gcry_camellia_ocb_crypt, _gcry_camellia_ocb_auth): Precalculate offset array when block count matches parallel operation size; Use 'ocb_get_l' instead of 'get_l'. * cipher/rijndael-aesni.c (get_l): Add fast path for 75% most common offsets. (aesni_ocb_enc, aesni_ocb_dec, _gcry_aes_aesni_ocb_auth): Precalculate offset array when block count matches parallel operation size. * cipher/rijndael-ssse3-amd64.c (get_l): Add fast path for 75% most common offsets. * cipher/rijndael.c (_gcry_aes_ocb_crypt, _gcry_aes_ocb_auth): Use 'ocb_get_l' instead of '_gcry_cipher_ocb_get_l'. * cipher/serpent.c (get_l): Remove. (_gcry_serpent_ocb_crypt, _gcry_serpent_ocb_auth): Precalculate offset array when block count matches parallel operation size; Use 'ocb_get_l' instead of 'get_l'. * cipher/twofish.c (get_l): Remove. (_gcry_twofish_ocb_crypt, _gcry_twofish_ocb_auth): Use 'ocb_get_l' instead of 'get_l'.
Patch optimizes OCB offset calculation for generic code and
assembly implementations with parallel block processing.
Benchmark of OCB AES-NI on Intel Haswell:
$ tests/bench-slope --cpu-mhz 3201 cipher aes
Before:
AES | nanosecs/byte mebibytes/sec cycles/byte CTR enc | 0.274 ns/B 3483.9 MiB/s 0.876 c/B CTR dec | 0.273 ns/B 3490.0 MiB/s 0.875 c/B OCB enc | 0.289 ns/B 3296.1 MiB/s 0.926 c/B OCB dec | 0.299 ns/B 3189.9 MiB/s 0.957 c/B OCB auth | 0.260 ns/B 3670.0 MiB/s 0.832 c/B
After:
AES | nanosecs/byte mebibytes/sec cycles/byte CTR enc | 0.273 ns/B 3489.4 MiB/s 0.875 c/B CTR dec | 0.273 ns/B 3487.5 MiB/s 0.875 c/B OCB enc | 0.248 ns/B 3852.8 MiB/s 0.792 c/B OCB dec | 0.261 ns/B 3659.5 MiB/s 0.834 c/B OCB auth | 0.227 ns/B 4205.5 MiB/s 0.726 c/B
- Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>