On 31 August 2018 at 17:56, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
> Hi Eric,
>
> On 31 August 2018 at 10:01, Eric Biggers <ebiggers@xxxxxxxxxx> wrote:
>> From: Eric Biggers <ebiggers@xxxxxxxxxx>
>>
>> Optimize ChaCha20 NEON performance by:
>>
>> - Implementing the 8-bit rotations using the 'vtbl.8' instruction.
>> - Streamlining the part that adds the original state and XORs the data.
>> - Making some other small tweaks.
>>
>> On ARM Cortex-A7, these optimizations improve ChaCha20 performance from
>> about 11.9 cycles per byte to 11.3.
>>
>> There is a tradeoff involved with the 'vtbl.8' rotation method since
>> there is at least one CPU where it's not fastest. But it seems to be a
>> better default; see the added comment.
>>
>> Signed-off-by: Eric Biggers <ebiggers@xxxxxxxxxx>
>> ---
>> arch/arm/crypto/chacha20-neon-core.S | 289 ++++++++++++++-------------
>> 1 file changed, 147 insertions(+), 142 deletions(-)
>>
>> diff --git a/arch/arm/crypto/chacha20-neon-core.S b/arch/arm/crypto/chacha20-neon-core.S
>> index 3fecb2124c35a..d381cebaba31d 100644
>> --- a/arch/arm/crypto/chacha20-neon-core.S
>> +++ b/arch/arm/crypto/chacha20-neon-core.S
>> @@ -18,6 +18,33 @@
>> * (at your option) any later version.
>> */
>>
>> + /*
>> + * NEON doesn't have a rotate instruction. The alternatives are, more or less:
>> + *
>> + * (a) vshl.u32 + vsri.u32 (needs temporary register)
>> + * (b) vshl.u32 + vshr.u32 + vorr (needs temporary register)
>> + * (c) vrev32.16 (16-bit rotations only)
>> + * (d) vtbl.8 + vtbl.8 (multiple of 8 bits rotations only,
>> + * needs index vector)
>> + *
>> + * ChaCha20 has 16, 12, 8, and 7-bit rotations. For the 12 and 7-bit
>> + * rotations, the only choices are (a) and (b). We use (a) since it takes
>> + * two-thirds the cycles of (b) on both Cortex-A7 and Cortex-A53.
>> + *
>> + * For the 16-bit rotation, we use vrev32.16 since it's consistently fastest
>> + * and doesn't need a temporary register.
>> + *
>> + * For the 8-bit rotation, we use vtbl.8 + vtbl.8. On Cortex-A7, this sequence
>> + * is twice as fast as (a), even when doing (a) on multiple registers
>> + * simultaneously to eliminate the stall between vshl and vsri. Also, it
>> + * parallelizes better when temporary registers are scarce.
>> + *
>> + * A disadvantage is that on Cortex-A53, the vtbl sequence is the same speed as
>> + * (a), so the need to load the rotation table actually makes the vtbl method
>> + * slightly slower overall on that CPU. Still, it seems to be a good
>> + * compromise to get a significant speed boost on some CPUs.
>> + */
>> +
>
> Thanks for sharing these results. I have been working on 32-bit ARM
> code under the assumption that the A53 pipeline more or less resembles
> the A7 one, but this is obviously not the case looking at your
> results. My contributions to arch/arm/crypto mainly involved Crypto
> Extensions code, which the A7 does not support in the first place, so
> it does not really matter, but I will keep this in mind going forward.
>
>> #include <linux/linkage.h>
>>
>> .text
>> @@ -46,6 +73,9 @@ ENTRY(chacha20_block_xor_neon)
>> vmov q10, q2
>> vmov q11, q3
>>
>> + ldr ip, =.Lrol8_table
>> + vld1.8 {d10}, [ip, :64]
>> +
>
> I usually try to avoid the =literal ldr notation, because it involves
> an additional load via the D-cache. Could you use a 64-bit literal
> instead of a byte array and use vldr instead? Or switch to adr? (and
> move the literal in range, I suppose)
>
>> mov r3, #10
>>
>> .Ldoubleround:
>> @@ -63,9 +93,9 @@ ENTRY(chacha20_block_xor_neon)
>>
>> // x0 += x1, x3 = rotl32(x3 ^ x0, 8)
>> vadd.i32 q0, q0, q1
>> - veor q4, q3, q0
>> - vshl.u32 q3, q4, #8
>> - vsri.u32 q3, q4, #24
>> + veor q3, q3, q0
>> + vtbl.8 d6, {d6}, d10
>> + vtbl.8 d7, {d7}, d10
>>
>> // x2 += x3, x1 = rotl32(x1 ^ x2, 7)
>> vadd.i32 q2, q2, q3
>> @@ -94,9 +124,9 @@ ENTRY(chacha20_block_xor_neon)
>>
>> // x0 += x1, x3 = rotl32(x3 ^ x0, 8)
>> vadd.i32 q0, q0, q1
>> - veor q4, q3, q0
>> - vshl.u32 q3, q4, #8
>> - vsri.u32 q3, q4, #24
>> + veor q3, q3, q0
>> + vtbl.8 d6, {d6}, d10
>> + vtbl.8 d7, {d7}, d10
>>
>> // x2 += x3, x1 = rotl32(x1 ^ x2, 7)
>> vadd.i32 q2, q2, q3
>> @@ -143,11 +173,11 @@ ENDPROC(chacha20_block_xor_neon)
>>
>> .align 5
>> ENTRY(chacha20_4block_xor_neon)
>> - push {r4-r6, lr}
>> - mov ip, sp // preserve the stack pointer
>> - sub r3, sp, #0x20 // allocate a 32 byte buffer
>> - bic r3, r3, #0x1f // aligned to 32 bytes
>> - mov sp, r3
>> + push {r4}
>
> The ARM EABI mandates 8 byte stack alignment, and if you take an
> interrupt right at this point, you will enter the interrupt handler
> with a misaligned stack. Whether this could actually cause any
> problems is a different question, but it is better to keep it 8-byte
> aligned to be sure.
>
> I'd suggest you just push r4 and lr, so you can return from the
> function by popping r4 and pc, as is customary on ARM.
>
>> + mov r4, sp // preserve the stack pointer
>> + sub ip, sp, #0x20 // allocate a 32 byte buffer
>> + bic ip, ip, #0x1f // aligned to 32 bytes
>> + mov sp, ip
>>
>
> Is there a reason for preferring ip over r3 for preserving the
> original value of sp?
>
>> // r0: Input state matrix, s
>> // r1: 4 data blocks output, o
>> @@ -157,25 +187,24 @@ ENTRY(chacha20_4block_xor_neon)
>> // This function encrypts four consecutive ChaCha20 blocks by loading
>> // the state matrix in NEON registers four times. The algorithm performs
>> // each operation on the corresponding word of each state matrix, hence
>> - // requires no word shuffling. For final XORing step we transpose the
>> - // matrix by interleaving 32- and then 64-bit words, which allows us to
>> - // do XOR in NEON registers.
>> + // requires no word shuffling. The words are re-interleaved before the
>> + // final addition of the original state and the XORing step.
>> //
>>
>> - // x0..15[0-3] = s0..3[0..3]
>> - add r3, r0, #0x20
>> + // x0..15[0-3] = s0..15[0-3]
>> + add ip, r0, #0x20
>> vld1.32 {q0-q1}, [r0]
>> - vld1.32 {q2-q3}, [r3]
>> + vld1.32 {q2-q3}, [ip]
>>
>> - adr r3, CTRINC
>> + adr ip, .Lctrinc1
>> vdup.32 q15, d7[1]
>> vdup.32 q14, d7[0]
>> - vld1.32 {q11}, [r3, :128]
>> + vld1.32 {q4}, [ip, :128]
>> vdup.32 q13, d6[1]
>> vdup.32 q12, d6[0]
>> - vadd.i32 q12, q12, q11 // x12 += counter values 0-3
>> vdup.32 q11, d5[1]
>> vdup.32 q10, d5[0]
>> + vadd.u32 q12, q12, q4 // x12 += counter values 0-3
>> vdup.32 q9, d4[1]
>> vdup.32 q8, d4[0]
>> vdup.32 q7, d3[1]
>> @@ -187,6 +216,7 @@ ENTRY(chacha20_4block_xor_neon)
>> vdup.32 q1, d0[1]
>> vdup.32 q0, d0[0]
>>
>> + adr ip, .Lrol8_table
>> mov r3, #10
>>
>> .Ldoubleround4:
>> @@ -238,24 +268,25 @@ ENTRY(chacha20_4block_xor_neon)
>> // x1 += x5, x13 = rotl32(x13 ^ x1, 8)
>> // x2 += x6, x14 = rotl32(x14 ^ x2, 8)
>> // x3 += x7, x15 = rotl32(x15 ^ x3, 8)
>> + vld1.8 {d16}, [ip, :64]
Also, would it perhaps be more efficient to keep the rotation vector
in a pair of GPRs, and use something like
vmov d16, r4, r5
here?
>> vadd.i32 q0, q0, q4
>> vadd.i32 q1, q1, q5
>> vadd.i32 q2, q2, q6
>> vadd.i32 q3, q3, q7
>>
>> - veor q8, q12, q0
>> - veor q9, q13, q1
>> - vshl.u32 q12, q8, #8
>> - vshl.u32 q13, q9, #8
>> - vsri.u32 q12, q8, #24
>> - vsri.u32 q13, q9, #24
>> + veor q12, q12, q0
>> + veor q13, q13, q1
>> + veor q14, q14, q2
>> + veor q15, q15, q3
>>
>> - veor q8, q14, q2
>> - veor q9, q15, q3
>> - vshl.u32 q14, q8, #8
>> - vshl.u32 q15, q9, #8
>> - vsri.u32 q14, q8, #24
>> - vsri.u32 q15, q9, #24
>> + vtbl.8 d24, {d24}, d16
>> + vtbl.8 d25, {d25}, d16
>> + vtbl.8 d26, {d26}, d16
>> + vtbl.8 d27, {d27}, d16
>> + vtbl.8 d28, {d28}, d16
>> + vtbl.8 d29, {d29}, d16
>> + vtbl.8 d30, {d30}, d16
>> + vtbl.8 d31, {d31}, d16
>>
>> vld1.32 {q8-q9}, [sp, :256]
>>
>> @@ -334,24 +365,25 @@ ENTRY(chacha20_4block_xor_neon)
>> // x1 += x6, x12 = rotl32(x12 ^ x1, 8)
>> // x2 += x7, x13 = rotl32(x13 ^ x2, 8)
>> // x3 += x4, x14 = rotl32(x14 ^ x3, 8)
>> + vld1.8 {d16}, [ip, :64]
>> vadd.i32 q0, q0, q5
>> vadd.i32 q1, q1, q6
>> vadd.i32 q2, q2, q7
>> vadd.i32 q3, q3, q4
>>
>> - veor q8, q15, q0
>> - veor q9, q12, q1
>> - vshl.u32 q15, q8, #8
>> - vshl.u32 q12, q9, #8
>> - vsri.u32 q15, q8, #24
>> - vsri.u32 q12, q9, #24
>> + veor q15, q15, q0
>> + veor q12, q12, q1
>> + veor q13, q13, q2
>> + veor q14, q14, q3
>>
>> - veor q8, q13, q2
>> - veor q9, q14, q3
>> - vshl.u32 q13, q8, #8
>> - vshl.u32 q14, q9, #8
>> - vsri.u32 q13, q8, #24
>> - vsri.u32 q14, q9, #24
>> + vtbl.8 d30, {d30}, d16
>> + vtbl.8 d31, {d31}, d16
>> + vtbl.8 d24, {d24}, d16
>> + vtbl.8 d25, {d25}, d16
>> + vtbl.8 d26, {d26}, d16
>> + vtbl.8 d27, {d27}, d16
>> + vtbl.8 d28, {d28}, d16
>> + vtbl.8 d29, {d29}, d16
>>
>> vld1.32 {q8-q9}, [sp, :256]
>>
>> @@ -381,104 +413,74 @@ ENTRY(chacha20_4block_xor_neon)
>> vsri.u32 q6, q9, #25
>>
>> subs r3, r3, #1
>> - beq 0f
>> -
>> - vld1.32 {q8-q9}, [sp, :256]
>> - b .Ldoubleround4
>> -
>> - // x0[0-3] += s0[0]
>> - // x1[0-3] += s0[1]
>> - // x2[0-3] += s0[2]
>> - // x3[0-3] += s0[3]
>> -0: ldmia r0!, {r3-r6}
>> - vdup.32 q8, r3
>> - vdup.32 q9, r4
>> - vadd.i32 q0, q0, q8
>> - vadd.i32 q1, q1, q9
>> - vdup.32 q8, r5
>> - vdup.32 q9, r6
>> - vadd.i32 q2, q2, q8
>> - vadd.i32 q3, q3, q9
>> -
>> - // x4[0-3] += s1[0]
>> - // x5[0-3] += s1[1]
>> - // x6[0-3] += s1[2]
>> - // x7[0-3] += s1[3]
>> - ldmia r0!, {r3-r6}
>> - vdup.32 q8, r3
>> - vdup.32 q9, r4
>> - vadd.i32 q4, q4, q8
>> - vadd.i32 q5, q5, q9
>> - vdup.32 q8, r5
>> - vdup.32 q9, r6
>> - vadd.i32 q6, q6, q8
>> - vadd.i32 q7, q7, q9
>> -
>> - // interleave 32-bit words in state n, n+1
>> - vzip.32 q0, q1
>> - vzip.32 q2, q3
>> - vzip.32 q4, q5
>> - vzip.32 q6, q7
>> -
>> - // interleave 64-bit words in state n, n+2
>> - vswp d1, d4
>> - vswp d3, d6
>> - vswp d9, d12
>> - vswp d11, d14
>> -
>> - // xor with corresponding input, write to output
>> - vld1.8 {q8-q9}, [r2]!
>> - veor q8, q8, q0
>> - veor q9, q9, q4
>> - vst1.8 {q8-q9}, [r1]!
>> -
>> vld1.32 {q8-q9}, [sp, :256]
>> -
>> - // x8[0-3] += s2[0]
>> - // x9[0-3] += s2[1]
>> - // x10[0-3] += s2[2]
>> - // x11[0-3] += s2[3]
>> - ldmia r0!, {r3-r6}
>> - vdup.32 q0, r3
>> - vdup.32 q4, r4
>> - vadd.i32 q8, q8, q0
>> - vadd.i32 q9, q9, q4
>> - vdup.32 q0, r5
>> - vdup.32 q4, r6
>> - vadd.i32 q10, q10, q0
>> - vadd.i32 q11, q11, q4
>> -
>> - // x12[0-3] += s3[0]
>> - // x13[0-3] += s3[1]
>> - // x14[0-3] += s3[2]
>> - // x15[0-3] += s3[3]
>> - ldmia r0!, {r3-r6}
>> - vdup.32 q0, r3
>> - vdup.32 q4, r4
>> - adr r3, CTRINC
>> - vadd.i32 q12, q12, q0
>> - vld1.32 {q0}, [r3, :128]
>> - vadd.i32 q13, q13, q4
>> - vadd.i32 q12, q12, q0 // x12 += counter values 0-3
>> -
>> - vdup.32 q0, r5
>> - vdup.32 q4, r6
>> - vadd.i32 q14, q14, q0
>> - vadd.i32 q15, q15, q4
>> -
>> - // interleave 32-bit words in state n, n+1
>> - vzip.32 q8, q9
>> - vzip.32 q10, q11
>> - vzip.32 q12, q13
>> - vzip.32 q14, q15
>> -
>> - // interleave 64-bit words in state n, n+2
>> - vswp d17, d20
>> - vswp d19, d22
>> - vswp d25, d28
>> + bne .Ldoubleround4
>> +
>> + // re-interleave the 32-bit words of the four blocks
>> + vzip.32 q14, q15 // => (14 15 14 15) (14 15 14 15)
>> + vzip.32 q12, q13 // => (12 13 12 13) (12 13 12 13)
>> + vzip.32 q10, q11 // => (10 11 10 11) (10 11 10 11)
>> + vzip.32 q8, q9 // => (8 9 8 9) (8 9 8 9)
>> + vzip.32 q6, q7 // => (6 7 6 7) (6 7 6 7)
>> + vzip.32 q4, q5 // => (4 5 4 5) (4 5 4 5)
>> + vzip.32 q2, q3 // => (2 3 2 3) (2 3 2 3)
>> + vzip.32 q0, q1 // => (0 1 0 1) (0 1 0 1)
>> vswp d27, d30
>> + vswp d25, d28
>> + vswp d19, d22
>> + vswp d17, d20
>> + vswp d11, d14
>> + vst1.32 {q14-q15}, [sp, :256] // free up two registers
>> + vswp d9, d12
>> + vswp d3, d6
>> + vld1.32 {q14-q15}, [r0]! // load s0..7
>> + vswp d1, d4
>>
>> - vmov q4, q1
>> + // Swap q1 and q4 so that we'll free up consecutive registers (q0-q1)
>> + // after XORing the first 32 bytes.
>> + vswp q1, q4
>> +
>> + // blocks are: (q0 q1 q8 q12) (q2 q6 q10 q14) (q4 q5 q9 q13) (q3 q7 q11 q15)
>> +
>> + // x0..3[0-3] += s0..3[0-3] (add orig state to 1st row of each block)
>> + vadd.u32 q0, q0, q14
>> + vadd.u32 q2, q2, q14
>> + vadd.u32 q4, q4, q14
>> + vadd.u32 q3, q3, q14
>> +
>> + // x4..7[0-3] += s4..7[0-3] (add orig state to 2nd row of each block)
>> + vadd.u32 q1, q1, q15
>> + vadd.u32 q6, q6, q15
>> + vadd.u32 q5, q5, q15
>> + vadd.u32 q7, q7, q15
>> +
>> + // XOR first 32 bytes using keystream from first two rows of first block
>> + vld1.8 {q14-q15}, [r2]!
>> + veor q14, q14, q0
>> + veor q15, q15, q1
>> + vld1.32 {q0-q1}, [r0] // load s8..s15
>> + vst1.8 {q14-q15}, [r1]!
>> +
>> + // x8..11[0-3] += s8..11[0-3] (add orig state to 3rd row of each block)
>> + vadd.u32 q8, q8, q0
>> + vadd.u32 q10, q10, q0
>> + vld1.32 {q14-q15}, [sp, :256]
>> + vadd.u32 q9, q9, q0
>> + adr ip, .Lctrinc2
>> + vadd.u32 q11, q11, q0
>> +
>> + // x12..15[0-3] += s12..15[0-3] (add orig state to 4th row of each block)
>> + vld1.32 {q0}, [ip, :128] // load (1, 0, 2, 0)
>
> Would something like
>
> vmov.i32 d0, #1
> vmovl.u32 q0, d0
> vadd.u64 d1, d1
>
> (untested) work as well?
>
>> + vadd.u32 q15, q15, q1
>> + vadd.u32 q13, q13, q1
>> + vadd.u32 q14, q14, q1
>> + vadd.u32 q12, q12, q1
>> + vadd.u32 d30, d30, d1 // x12[3] += 2
>> + vadd.u32 d26, d26, d1 // x12[2] += 2
>> + vadd.u32 d28, d28, d0 // x12[1] += 1
>> + vadd.u32 d30, d30, d0 // x12[3] += 1
>> +
>> + // XOR the rest of the data with the keystream
>>
>> vld1.8 {q0-q1}, [r2]!
>> veor q0, q0, q8
>> @@ -511,13 +513,16 @@ ENTRY(chacha20_4block_xor_neon)
>> vst1.8 {q0-q1}, [r1]!
>>
>> vld1.8 {q0-q1}, [r2]
>> + mov sp, r4 // restore original stack pointer
>> veor q0, q0, q11
>> veor q1, q1, q15
>> vst1.8 {q0-q1}, [r1]
>>
>> - mov sp, ip
>> - pop {r4-r6, pc}
>> + pop {r4}
>> + bx lr
>> ENDPROC(chacha20_4block_xor_neon)
>>
>> .align 4
>> -CTRINC: .word 0, 1, 2, 3
>> +.Lctrinc1: .word 0, 1, 2, 3
>> +.Lctrinc2: .word 1, 0, 2, 0
>> +.Lrol8_table: .byte 3, 0, 1, 2, 7, 4, 5, 6
>> --
>> 2.18.0
>>
