OCB ARM CE: Move ocb_get_l handling to assembly part
5c418e597f0fUnpublished
Actions

Unpublished Commit · Learn More

Not On Permanent Ref: This commit is not an ancestor of any permanent ref.

Description

OCB ARM CE: Move ocb_get_l handling to assembly part

* cipher/rijndael-armv8-aarch32-ce.S: Add OCB 'L_{ntz(i)}' calculation.
* cipher/rijndael-armv8-aarch64-ce.S: Ditto.
* cipher/rijndael-armv8-ce.c (_gcry_aes_ocb_enc_armv8_ce)
(_gcry_aes_ocb_dec_armv8_ce, _gcry_aes_ocb_auth_armv8_ce)
(ocb_cryt_fn_t): Updated arguments.
(_gcry_aes_armv8_ce_ocb_crypt, _gcry_aes_armv8_ce_ocb_auth): Remove
'ocb_get_l' handling and splitting input to 32 block chunks, instead
pass full buffers to assembly.

Performance on Cortex-A53 (AArch32):

Before:
AES | nanosecs/byte mebibytes/sec cycles/byte

 OCB enc |      1.63 ns/B     583.8 MiB/s      1.88 c/B
 OCB dec |      1.67 ns/B     572.1 MiB/s      1.92 c/B
OCB auth |      1.33 ns/B     717.1 MiB/s      1.53 c/B

After (~12% faster):
AES | nanosecs/byte mebibytes/sec cycles/byte

 OCB enc |      1.47 ns/B     650.2 MiB/s      1.69 c/B
 OCB dec |      1.48 ns/B     644.5 MiB/s      1.70 c/B
OCB auth |      1.19 ns/B     798.2 MiB/s      1.38 c/B

Performance on Cortex-A53 (AArch64):

Before:
AES | nanosecs/byte mebibytes/sec cycles/byte

 OCB enc |      1.29 ns/B     738.5 MiB/s      1.49 c/B
 OCB dec |      1.32 ns/B     723.5 MiB/s      1.52 c/B
OCB auth |      1.15 ns/B     827.0 MiB/s      1.33 c/B

After (~8% faster):
AES | nanosecs/byte mebibytes/sec cycles/byte

 OCB enc |      1.21 ns/B     789.1 MiB/s      1.39 c/B
 OCB dec |      1.21 ns/B     789.2 MiB/s      1.39 c/B
OCB auth |      1.10 ns/B     867.0 MiB/s      1.27 c/B