diff options
Diffstat (limited to 'arch/x86/crypto/aes-gcm-aesni-x86_64.S')
| -rw-r--r-- | arch/x86/crypto/aes-gcm-aesni-x86_64.S | 12 |
1 files changed, 6 insertions, 6 deletions
diff --git a/arch/x86/crypto/aes-gcm-aesni-x86_64.S b/arch/x86/crypto/aes-gcm-aesni-x86_64.S index 45940e2883a0..7c8a8a32bd3c 100644 --- a/arch/x86/crypto/aes-gcm-aesni-x86_64.S +++ b/arch/x86/crypto/aes-gcm-aesni-x86_64.S @@ -61,15 +61,15 @@ // for the *_aesni functions or AVX for the *_aesni_avx ones. (But it seems // there are no CPUs that support AES-NI without also PCLMULQDQ and SSE4.1.) // -// The design generally follows that of aes-gcm-avx10-x86_64.S, and that file is +// The design generally follows that of aes-gcm-vaes-avx512.S, and that file is // more thoroughly commented. This file has the following notable changes: // // - The vector length is fixed at 128-bit, i.e. xmm registers. This means // there is only one AES block (and GHASH block) per register. // -// - Without AVX512 / AVX10, only 16 SIMD registers are available instead of -// 32. We work around this by being much more careful about using -// registers, relying heavily on loads to load values as they are needed. +// - Without AVX512, only 16 SIMD registers are available instead of 32. We +// work around this by being much more careful about using registers, +// relying heavily on loads to load values as they are needed. // // - Masking is not available either. We work around this by implementing // partial block loads and stores using overlapping scalar loads and stores @@ -90,8 +90,8 @@ // multiplication instead of schoolbook multiplication. This saves one // pclmulqdq instruction per block, at the cost of one 64-bit load, one // pshufd, and 0.25 pxors per block. (This is without the three-argument -// XOR support that would be provided by AVX512 / AVX10, which would be -// more beneficial to schoolbook than Karatsuba.) +// XOR support that would be provided by AVX512, which would be more +// beneficial to schoolbook than Karatsuba.) // // As a rough approximation, we can assume that Karatsuba multiplication is // faster than schoolbook multiplication in this context if one pshufd and |