Reduce size of x86-64 stitched Chacha20-Poly1305 implementations
664370ea02df
Actions

Description

Reduce size of x86-64 stitched Chacha20-Poly1305 implementations

* cipher/chacha20-amd64-avx2.c
(_gcry_chacha20_poly1305_amd64_avx2_blocks8): De-unroll round loop.
* cipher/chacha20-amd64-ssse3.c
(_gcry_chacha20_poly1305_amd64_ssse3_blocks4):
(_gcry_chacha20_poly1305_amd64_ssse3_blocks1): Ditto.

Object size before:

 text    data     bss     dec     hex filename
13428       0       0   13428    3474 cipher/.libs/chacha20-amd64-avx2.o
23175       0       0   23175    5a87 cipher/.libs/chacha20-amd64-ssse3.o

Object size after:

text    data     bss     dec     hex filename
4815       0       0    4815    12cf cipher/.libs/chacha20-amd64-avx2.o
9284       0       0    9284    2444 cipher/.libs/chacha20-amd64-ssse3.o

Benchmark on AMD Ryzen 3700X (AVX2 impl.):

Before:
CHACHA20 | nanosecs/byte mebibytes/sec cycles/byte auto Mhz

   STREAM enc |     0.266 ns/B      3581 MiB/s      1.15 c/B      4333
   STREAM dec |     0.265 ns/B      3598 MiB/s      1.15 c/B      4350
 POLY1305 enc |     0.313 ns/B      3046 MiB/s      1.35 c/B      4317
 POLY1305 dec |     0.296 ns/B      3222 MiB/s      1.29 c/B      4345
POLY1305 auth |     0.221 ns/B      4311 MiB/s     0.972 c/B      4394

After:
CHACHA20 | nanosecs/byte mebibytes/sec cycles/byte auto Mhz

   STREAM enc |     0.264 ns/B      3614 MiB/s      1.16 c/B      4380±2
   STREAM dec |     0.265 ns/B      3597 MiB/s      1.16 c/B      4374
 POLY1305 enc |     0.293 ns/B      3252 MiB/s      1.27 c/B      4326
 POLY1305 dec |     0.275 ns/B      3464 MiB/s      1.19 c/B      4323
POLY1305 auth |     0.219 ns/B      4360 MiB/s     0.963 c/B      4400

[v2]: Use two inner round loops with 3 and 2 iterations and different