by Jason A. Donenfeld

[permalink] [raw]

Subject: Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements

Hi Martin,

On Tue, Nov 20, 2018 at 5:29 PM Martin Willi <[email protected]> wrote:
> Thanks for the offer, no need at this time. But I certainly would
> welcome if you could do some (Wireguard) benching with that code to see
> if it works for you.

I certainly will test it in a few different network circumstances,
especially since real testing like this is sometimes more telling than
busy-loop benchmarks.

> > Actually, similarly here, a 10nm Cannon Lake machine should be
> > arriving at my house this week, which should make for some
> > interesting testing ground for non-throttled zmm, if you'd like to
> > play with it.
>
> Maybe in a future iteration, thanks. In fact would it be interesting to
> know if Cannon Lake can handle that throttling better.

Everything I've read on the Internet seems to indicate that's the
case, so one of the first things I'll be doing is seeing if that's
true. There are also the AVX512 IFMA instructions to play with!

Jason

2018-11-11 19:24:51

by Martin Willi

[permalink] [raw]

Subject: [PATCH 5/6] crypto: x86/chacha20 - Add a 2-block AVX2 variant

This variant uses the same principle as the single block SSSE3 variant
by shuffling the state matrix after each round. With the wider AVX
registers, we can do two blocks in parallel, though.

This function can increase performance and efficiency significantly for
lengths that would otherwise require a 4-block function.

Signed-off-by: Martin Willi <[email protected]>
---
arch/x86/crypto/chacha20-avx2-x86_64.S | 197 +++++++++++++++++++++++++
arch/x86/crypto/chacha20_glue.c | 7 +
2 files changed, 204 insertions(+)

diff --git a/arch/x86/crypto/chacha20-avx2-x86_64.S b/arch/x86/crypto/chacha20-avx2-x86_64.S
index 7b62d55bee3d..8247076b0ba7 100644
--- a/arch/x86/crypto/chacha20-avx2-x86_64.S
+++ b/arch/x86/crypto/chacha20-avx2-x86_64.S
@@ -26,8 +26,205 @@ ROT16: .octa 0x0d0c0f0e09080b0a0504070601000302
CTRINC: .octa 0x00000003000000020000000100000000
.octa 0x00000007000000060000000500000004

+.section .rodata.cst32.CTR2BL, "aM", @progbits, 32
+.align 32
+CTR2BL: .octa 0x00000000000000000000000000000000
+ .octa 0x00000000000000000000000000000001
+
.text

+ENTRY(chacha20_2block_xor_avx2)
+ # %rdi: Input state matrix, s
+ # %rsi: up to 2 data blocks output, o
+ # %rdx: up to 2 data blocks input, i
+ # %rcx: input/output length in bytes
+
+ # This function encrypts two ChaCha20 blocks by loading the state
+ # matrix twice across four AVX registers. It performs matrix operations
+ # on four words in each matrix in parallel, but requires shuffling to
+ # rearrange the words after each round.
+
+ vzeroupper
+
+ # x0..3[0-2] = s0..3
+ vbroadcasti128 0x00(%rdi),%ymm0
+ vbroadcasti128 0x10(%rdi),%ymm1
+ vbroadcasti128 0x20(%rdi),%ymm2
+ vbroadcasti128 0x30(%rdi),%ymm3
+
+ vpaddd CTR2BL(%rip),%ymm3,%ymm3
+
+ vmovdqa %ymm0,%ymm8
+ vmovdqa %ymm1,%ymm9
+ vmovdqa %ymm2,%ymm10
+ vmovdqa %ymm3,%ymm11
+
+ vmovdqa ROT8(%rip),%ymm4
+ vmovdqa ROT16(%rip),%ymm5
+
+ mov %rcx,%rax
+ mov $10,%ecx
+
+.Ldoubleround:
+
+ # x0 += x1, x3 = rotl32(x3 ^ x0, 16)
+ vpaddd %ymm1,%ymm0,%ymm0
+ vpxor %ymm0,%ymm3,%ymm3
+ vpshufb %ymm5,%ymm3,%ymm3
+
+ # x2 += x3, x1 = rotl32(x1 ^ x2, 12)
+ vpaddd %ymm3,%ymm2,%ymm2
+ vpxor %ymm2,%ymm1,%ymm1
+ vmovdqa %ymm1,%ymm6
+ vpslld $12,%ymm6,%ymm6
+ vpsrld $20,%ymm1,%ymm1
+ vpor %ymm6,%ymm1,%ymm1
+
+ # x0 += x1, x3 = rotl32(x3 ^ x0, 8)
+ vpaddd %ymm1,%ymm0,%ymm0
+ vpxor %ymm0,%ymm3,%ymm3
+ vpshufb %ymm4,%ymm3,%ymm3
+
+ # x2 += x3, x1 = rotl32(x1 ^ x2, 7)
+ vpaddd %ymm3,%ymm2,%ymm2
+ vpxor %ymm2,%ymm1,%ymm1
+ vmovdqa %ymm1,%ymm7
+ vpslld $7,%ymm7,%ymm7
+ vpsrld $25,%ymm1,%ymm1
+ vpor %ymm7,%ymm1,%ymm1
+
+ # x1 = shuffle32(x1, MASK(0, 3, 2, 1))
+ vpshufd $0x39,%ymm1,%ymm1
+ # x2 = shuffle32(x2, MASK(1, 0, 3, 2))
+ vpshufd $0x4e,%ymm2,%ymm2
+ # x3 = shuffle32(x3, MASK(2, 1, 0, 3))
+ vpshufd $0x93,%ymm3,%ymm3
+
+ # x0 += x1, x3 = rotl32(x3 ^ x0, 16)
+ vpaddd %ymm1,%ymm0,%ymm0
+ vpxor %ymm0,%ymm3,%ymm3
+ vpshufb %ymm5,%ymm3,%ymm3
+
+ # x2 += x3, x1 = rotl32(x1 ^ x2, 12)
+ vpaddd %ymm3,%ymm2,%ymm2
+ vpxor %ymm2,%ymm1,%ymm1
+ vmovdqa %ymm1,%ymm6
+ vpslld $12,%ymm6,%ymm6
+ vpsrld $20,%ymm1,%ymm1
+ vpor %ymm6,%ymm1,%ymm1
+
+ # x0 += x1, x3 = rotl32(x3 ^ x0, 8)
+ vpaddd %ymm1,%ymm0,%ymm0
+ vpxor %ymm0,%ymm3,%ymm3
+ vpshufb %ymm4,%ymm3,%ymm3
+
+ # x2 += x3, x1 = rotl32(x1 ^ x2, 7)
+ vpaddd %ymm3,%ymm2,%ymm2
+ vpxor %ymm2,%ymm1,%ymm1
+ vmovdqa %ymm1,%ymm7
+ vpslld $7,%ymm7,%ymm7
+ vpsrld $25,%ymm1,%ymm1
+ vpor %ymm7,%ymm1,%ymm1
+
+ # x1 = shuffle32(x1, MASK(2, 1, 0, 3))
+ vpshufd $0x93,%ymm1,%ymm1
+ # x2 = shuffle32(x2, MASK(1, 0, 3, 2))
+ vpshufd $0x4e,%ymm2,%ymm2
+ # x3 = shuffle32(x3, MASK(0, 3, 2, 1))
+ vpshufd $0x39,%ymm3,%ymm3
+
+ dec %ecx
+ jnz .Ldoubleround
+
+ # o0 = i0 ^ (x0 + s0)
+ vpaddd %ymm8,%ymm0,%ymm7
+ cmp $0x10,%rax
+ jl .Lxorpart2
+ vpxor 0x00(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x00(%rsi)
+ vextracti128 $1,%ymm7,%xmm0
+ # o1 = i1 ^ (x1 + s1)
+ vpaddd %ymm9,%ymm1,%ymm7
+ cmp $0x20,%rax
+ jl .Lxorpart2
+ vpxor 0x10(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x10(%rsi)
+ vextracti128 $1,%ymm7,%xmm1
+ # o2 = i2 ^ (x2 + s2)
+ vpaddd %ymm10,%ymm2,%ymm7
+ cmp $0x30,%rax
+ jl .Lxorpart2
+ vpxor 0x20(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x20(%rsi)
+ vextracti128 $1,%ymm7,%xmm2
+ # o3 = i3 ^ (x3 + s3)
+ vpaddd %ymm11,%ymm3,%ymm7
+ cmp $0x40,%rax
+ jl .Lxorpart2
+ vpxor 0x30(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x30(%rsi)
+ vextracti128 $1,%ymm7,%xmm3
+
+ # xor and write second block
+ vmovdqa %xmm0,%xmm7
+ cmp $0x50,%rax
+ jl .Lxorpart2
+ vpxor 0x40(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x40(%rsi)
+
+ vmovdqa %xmm1,%xmm7
+ cmp $0x60,%rax
+ jl .Lxorpart2
+ vpxor 0x50(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x50(%rsi)
+
+ vmovdqa %xmm2,%xmm7
+ cmp $0x70,%rax
+ jl .Lxorpart2
+ vpxor 0x60(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x60(%rsi)
+
+ vmovdqa %xmm3,%xmm7
+ cmp $0x80,%rax
+ jl .Lxorpart2
+ vpxor 0x70(%rdx),%xmm7,%xmm6
+ vmovdqu %xmm6,0x70(%rsi)
+
+.Ldone2:
+ vzeroupper
+ ret
+
+.Lxorpart2:
+ # xor remaining bytes from partial register into output
+ mov %rax,%r9
+ and $0x0f,%r9
+ jz .Ldone2
+ and $~0x0f,%rax
+
+ mov %rsi,%r11
+
+ lea 8(%rsp),%r10
+ sub $0x10,%rsp
+ and $~31,%rsp
+
+ lea (%rdx,%rax),%rsi
+ mov %rsp,%rdi
+ mov %r9,%rcx
+ rep movsb
+
+ vpxor 0x00(%rsp),%xmm7,%xmm7
+ vmovdqa %xmm7,0x00(%rsp)
+
+ mov %rsp,%rsi
+ lea (%r11,%rax),%rdi
+ mov %r9,%rcx
+ rep movsb
+
+ lea -8(%r10),%rsp
+ jmp .Ldone2
+
+ENDPROC(chacha20_2block_xor_avx2)
+
ENTRY(chacha20_8block_xor_avx2)
# %rdi: Input state matrix, s
# %rsi: up to 8 data blocks output, o
diff --git a/arch/x86/crypto/chacha20_glue.c b/arch/x86/crypto/chacha20_glue.c
index b541da71f11e..82e46589a189 100644
--- a/arch/x86/crypto/chacha20_glue.c
+++ b/arch/x86/crypto/chacha20_glue.c
@@ -24,6 +24,8 @@ asmlinkage void chacha20_block_xor_ssse3(u32 *state, u8 *dst, const u8 *src,
asmlinkage void chacha20_4block_xor_ssse3(u32 *state, u8 *dst, const u8 *src,
unsigned int len);
#ifdef CONFIG_AS_AVX2
+asmlinkage void chacha20_2block_xor_avx2(u32 *state, u8 *dst, const u8 *src,
+ unsigned int len);
asmlinkage void chacha20_8block_xor_avx2(u32 *state, u8 *dst, const u8 *src,
unsigned int len);
static bool chacha20_use_avx2;
@@ -52,6 +54,11 @@ static void chacha20_dosimd(u32 *state, u8 *dst, const u8 *src,
state[12] += chacha20_advance(bytes, 8);
return;
}
+ if (bytes > CHACHA20_BLOCK_SIZE) {
+ chacha20_2block_xor_avx2(state, dst, src, bytes);
+ state[12] += chacha20_advance(bytes, 2);
+ return;
+ }
}
#endif
while (bytes >= CHACHA20_BLOCK_SIZE * 4) {
--
2.17.1