GHASH for high-end ARM cores

After years of using Cortex-A57 or A53 based systems as both my development machines as well as my optimization targets, I was recently given a ThunderX2 workstation, and after having moved my development environment to it, I have started  looking into whether the existing accelerated crypto drivers in the Linux kernel perform optimally on this micro-architecture.

Marvell’s ThunderX2 core (née Broadcom Vulcan) is a server class CPU with a deep pipeline. This typically means that it takes longer for the result of an instruction to become available to subsequent instructions, and instead of stalling the pipeline waiting for that result, we should try to do some other useful work in the meantime.

So for the simpler AES modes such as CBC and CTR, I have increased the interleave factor from 4x to 5x, resulting in a speedup of around 11% on ThunderX2. This result fueled my suspicion that the GCM code, which was using an interleave factor of only 2x, could benefit even more from this approach, and since I already added an implementation of GHASH (one of the constituent parts of GCM) that uses aggregation to operate on 4 blocks of input at a time, it should simply be a matter of combining this code with a 4-way interleaved implementation of CTR.

GHASH aggregation

GHASH is not a general purpose cryptographic hash function, but one that was specifically designed for (and should only be used in) AES in GCM mode. It is based on multiplication in GF(2128), and basically comes down to the following

X[0] = (H * I[0]) mod M
X[1] = (H * (X[0] ^ I[1])) mod M
X[n] = (H * (X[n-1] ^ I[n])) mod M

where I[] is an input array of 16 byte blocks, and the output of the hash function is the last value in the array X[]. H is a quantity that is derived from the AES key by encrypting a single AES block consisting of NUL bytes, and M is GHASH’s characteristic polynomial x128 + x7 + x2 + x + 1 over GF(2128).

Each line involves a multiplication in GF(2128), which means that it consists a multiplication step and a modulo division step, and it is possible the amortize the cost of the latter step over multiple multiplications, increasing performance. For example, we can process two blocks at a time like so

X[n] = (H * (((H * (X[n-2] ^ I[n-1])) mod M) ^ I[n])) mod M
     = ((H^2 * (X[n-2] ^ I[n-1])) ^ (H * I[n])) mod m

This is called GHASH aggregation, and was implemented for the core GHASH transform here. Note that this requires powers of H to be precomputed, and so we cannot apply this transformation an arbitrary number of times.

Another thing to note is that this algorithm combines the highest power of H with the first input block, which means that if we want to reuse the same code sequence to process fewer than n blocks, we need to populate the input blocks from right to left, and jump to the appropriate point in the routine so that only the lower powers of H are used in the computation.

Handling the tail

GCM is an AEAD (Authenticated Encryption with Associated Data) which can operate on inputs of any size. Since GCM is mainly used in the networking space, we should expect having to deal mostly with the typical size of a network packet, which is ~1500 bytes. This means that our 4-way aggregated code operating on 4 blocks or 64 bytes at a time will always have to deal with a tail block of less than 64 bytes in size.

Since out-of-order cores with deep pipelines perform poorly when running code with lots of branches and tight loops that load small quantities at a time, we should try to handle the tail with as few branches as possible, and use large, potentially overlapping loads to get the data into SIMD registers.

This is what the sequence below implements. It unconditionally performs four 16-byte loads, and manipulates the input address and the post-increment values so that we end up with I[] populated right to left with between 1 and 63 bytes of remaining input. (x0 contains the source address, and x2 the size of the remaining input minus 64)

mov     x15, #16
ands    x19, x0, #0xf
csel    x19, x19, x15, ne
adr_l   x17, .Lpermute_table + 16

sub     x11, x15, x19
add     x12, x17, x11
ld1     {T1.16b}, [x12]
sub     x11, x2, x11

cmp     x0, #-16
csel    x14, x15, xzr, gt
cmp     x0, #-32
csel    x15, x15, xzr, gt
cmp     x0, #-48
csel    x16, x19, xzr, gt
csel    x2, x2, x11, gt

ld1     {I0.16b}, [x2], x14
ld1     {I1.16b}, [x2], x15
ld1     {I2.16b}, [x2], x16
ld1     {I3.16b}, [x2]
tbl     I3.16b, {I3.16b}, T1.16b

Since we always load at least 16 bytes, it is up to the caller to present the input in a way that permits doing so if the entire input is less than 16 bytes in size (e.g, by copying the input to the end of a 16 byte buffer)

CTR encryption

While the 4-way GHASH code requires taking another code path through the 4-way routine when it is invoked with less than 4 blocks of input, the CTR code can be modified more easily to support this, by incrementing the overall counter by the remaining number of blocks, and then subtracting 4-n for each keystream block n. This naturally populates the keystream blocks from right to left as well, and instead of using a different encryption routine depending on the remaining input size, we can just unconditionally generate a 64 byte keystream and discard the parts that we don’t need.

Generating the tag

The final step in performing GCM is generating the authentication tag, which involves one final round of GHASH on a block of input containing the input size, and a final pass of CTR encryption.

While the existing code does not handle the tail or the tag directly, and jumps back to C code instead to copy the tail into a stack buffer and process it via the split GHASH and CTR routines, we have now incorporated this into the core GCM transform. This means we can go one step further, and fold the tag generation in as well.

ld1     {INP3.16b}, [x10]             // load lengths[]
mov     w9, #1                        // one block of input
bl      pmull_gcm_ghash_4x

mov     w11, #(0x1 << 24)             // BE '1U'
ld1     {KS0.16b}, [x5]               // load upper counter
mov     KS0.s[3], w11                 // set lower counter

enc_block KS0, x7, x6, x12

ext     X.16b, X.16b, X.16b, #8
rev64   X.16b, X.16b
eor     X.16b, X.16b, KS0.16b
st1     {XL.16b}, [x10]               // store tag


When combining all the changes above, we end up with a routine that can consume the entire encryption input buffer including the tail and the input block for the tag, and process it in a way that is up to 25% faster than the old code for 1500 byte blocks. Note that 4 KiB blocks are only 11-16% faster, and on other micro-architectures, performance regresses slightly for large inputs. This is not a use case worth obsessing about, though. More performance numbers below, the full implementation can be found here.

AES-128 AES-192 AES-256
#bytes 512 1500 4k 512 1500 4k 512 1500 4k
TX2 35% 23% 11% 34% 20% 9% 38% 25% 16%
EMAG 11% 6% 3% 12% 4% 2% 11% 4% 2%
A72 8% 5% -4% 9% 4% -5% 7% 4% -5%
A53 11% 6% -1% 10% 8% -1% 10% 8% -2%

GHASH for low-end ARM cores

The Galois hash algorithm (GHASH) is a fairly straight-forward keyed hash algorithm based on finite field multiplication, using the field GF(2128) with characteristic polynomial x128 + x7 + x2 + x + 1. (An excellent treatment of Galois fields can be found here)

The significance of GHASH is that it is used as the authentication component in the GCM algorithm, which is an implementation of authenticated encryption with associated data (AEAD), a cryptographic mode that combines authentication of data sent in the clear with authentication of data that is sent in encrypted form at the same time. It is widely used these days, primarily in the networking domain (IPsec, IEEE 802.11)

ISA support

Both the Intel and ARMv8 instruction sets now contain support for carry-less multiplication (also known as polynomial multiplication), primarily to allow for accelerated implementations of GHASH to be created, which formerly had to rely on unwieldy and less secure table based implementations. (The Linux implementation pre-computes a 4 KB lookup table for each instance of the hash algorithm that is in use, i.e., for each session having a different key. 4 KB per IPsec connection does not sound too bad in terms of memory usage, but the D-cache footprint may become a bottleneck when serving lots of concurrent connections.) In contrast, implementations based on these special instructions are time invariant, and are significantly faster (around 16x on high end ARMv8 cores).

Unfortunately, though, while ARMv8 specifies a range of polynomial multiplication instructions with various operand sizes, the one we are most interested in, which performs carry-less multiplication on two 64-bit operands to produce a 128-bit result, is optional in the architecture. So on low-end cores such as the Cortex-A53 (as can be found in the Raspberry Pi 3), the accelerated driver is not available because this particular instruction is not implemented.

Using vmull.p8 to implement vmull.p64

The other day, I stumbled upon the paper Fast Software Polynomial Multiplication on ARM Processors Using the NEON Engine by Danilo Camara, Conrado Gouvea, Julio Lopez and Ricardo Dahab, which describes how 64×64 to 128 bit polynomial multiplication (vmull.p64) can be composed using 8×8 to 16 bit polynomial multiplication (vmull.p8) combined with other SIMD arithmetic instructions. The nice thing about vmull.p8 is that it is a standard NEON instruction, which means all NEON capable CPUs implement it, including the Cortex-A53 on the Raspberry Pi 3.

Transliterating 32-bit ARM code to the 64-bit ISA

The algorithm as described in the paper is based on the 32-bit instruction set (retroactively named AArch32), which deviates significantly from the new 64-bit ISA called AArch64. The primary difference is that the number of SIMD registers has increased to 32, which is nice, but which has a downside as well: it is no longer possible to directly use the top half of a 128-bit register as a 64-bit register, which is something the polynomial multiplication algorithm relies on.

The original code looks something like this (note the use of ‘high’ and ‘low’ registers in the same instruction)

.macro          vmull_p64, rq, ad, bd
vext.8          t0l, \ad, \ad, #1       @ A1
vmull.p8        t0q, t0l, \bd           @ F = A1*B
vext.8          \rq\()_L, \bd, \bd, #1  @ B1
vmull.p8        \rq, \ad, \rq\()_L      @ E = A*B1
vext.8          t1l, \ad, \ad, #2       @ A2
vmull.p8        t1q, t1l, \bd           @ H = A2*B
vext.8          t3l, \bd, \bd, #2       @ B2
vmull.p8        t3q, \ad, t3l           @ G = A*B2
vext.8          t2l, \ad, \ad, #3       @ A3
vmull.p8        t2q, t2l, \bd           @ J = A3*B
veor            t0q, t0q, \rq           @ L = E + F
vext.8          \rq\()_L, \bd, \bd, #3  @ B3
vmull.p8        \rq, \ad, \rq\()_L      @ I = A*B3
veor            t1q, t1q, t3q           @ M = G + H
vext.8          t3l, \bd, \bd, #4       @ B4
vmull.p8        t3q, \ad, t3l           @ K = A*B4
veor            t0l, t0l, t0h           @ t0 = (L) (P0 + P1) << 8
vand            t0h, t0h, k48
veor            t1l, t1l, t1h           @ t1 = (M) (P2 + P3) << 16
vand            t1h, t1h, k32
veor            t2q, t2q, \rq           @ N = I + J
veor            t0l, t0l, t0h
veor            t1l, t1l, t1h
veor            t2l, t2l, t2h           @ t2 = (N) (P4 + P5) << 24
vand            t2h, t2h, k16
veor            t3l, t3l, t3h           @ t3 = (K) (P6 + P7) << 32
vmov.i64        t3h, #0
vext.8          t0q, t0q, t0q, #15
veor            t2l, t2l, t2h
vext.8          t1q, t1q, t1q, #14
vmull.p8        \rq, \ad, \bd           @ D = A*B
vext.8          t2q, t2q, t2q, #13
vext.8          t3q, t3q, t3q, #12
veor            t0q, t0q, t1q
veor            t2q, t2q, t3q
veor            \rq, \rq, t0q
veor            \rq, \rq, t2q

However, things like veor t1l, t1l, t1h or using ext with upper halves of registers are not possible in AArch64, and so we need to transpose the contents of some of registers using the tbl and/or zip/unzip instructions. Also, the vmull.p8 instruction now exists in two variants: pmull operating on the lower halves and pmull2 operating on the upper halves of the input operands.

We end up with the following sequence, which is 3 instructions longer than the original:

.macro          __pmull_p8, rq, ad, bd, i
.ifb            \i
ext             t4.8b, \ad\().8b, \ad\().8b, #1         // A1
ext             t8.8b, \bd\().8b, \bd\().8b, #1         // B1
ext             t5.8b, \ad\().8b, \ad\().8b, #2         // A2
ext             t7.8b, \bd\().8b, \bd\().8b, #2         // B2
ext             t6.8b, \ad\().8b, \ad\().8b, #3         // A3
ext             t9.8b, \bd\().8b, \bd\().8b, #3         // B3
ext             t3.8b, \bd\().8b, \bd\().8b, #4         // B4

pmull           t4.8h, t4.8b, \bd\().8b                 // F = A1*B
pmull           t8.8h, \ad\().8b, t8.8b                 // E = A*B1
pmull           t5.8h, t5.8b, \bd\().8b                 // H = A2*B
pmull           t7.8h, \ad\().8b, t7.8b                 // G = A*B2
pmull           t6.8h, t6.8b, \bd\().8b                 // J = A3*B
pmull           t9.8h, \ad\().8b, t9.8b                 // I = A*B3
pmull           t3.8h, \ad\().8b, t3.8b                 // K = A*B4
pmull           \rq\().8h, \ad\().8b, \bd\().8b         // D = A*B
tbl             t4.16b, {\ad\().16b}, perm1.16b         // A1
tbl             t8.16b, {\bd\().16b}, perm1.16b         // B1
tbl             t5.16b, {\ad\().16b}, perm2.16b         // A2
tbl             t7.16b, {\bd\().16b}, perm2.16b         // B2
tbl             t6.16b, {\ad\().16b}, perm3.16b         // A3
tbl             t9.16b, {\bd\().16b}, perm3.16b         // B3
tbl             t3.16b, {\bd\().16b}, perm4.16b         // B4

pmull2          t4.8h, t4.16b, \bd\().16b               // F = A1*B
pmull2          t8.8h, \ad\().16b, t8.16b               // E = A*B1
pmull2          t5.8h, t5.16b, \bd\().16b               // H = A2*B
pmull2          t7.8h, \ad\().16b, t7.16b               // G = A*B2
pmull2          t6.8h, t6.16b, \bd\().16b               // J = A3*B
pmull2          t9.8h, \ad\().16b, t9.16b               // I = A*B3
pmull2          t3.8h, \ad\().16b, t3.16b               // K = A*B4
pmull2          \rq\().8h, \ad\().16b, \bd\().16b       // D = A*B

eor             t4.16b, t4.16b, t8.16b                  // L = E + F
eor             t5.16b, t5.16b, t7.16b                  // M = G + H
eor             t6.16b, t6.16b, t9.16b                  // N = I + J

uzp1            t8.2d, t4.2d, t5.2d
uzp2            t4.2d, t4.2d, t5.2d
uzp1            t7.2d, t6.2d, t3.2d
uzp2            t6.2d, t6.2d, t3.2d

// t4 = (L) (P0 + P1) << 8
// t5 = (M) (P2 + P3) << 16
eor             t8.16b, t8.16b, t4.16b
and             t4.16b, t4.16b, k32_48.16b

// t6 = (N) (P4 + P5) << 24
// t7 = (K) (P6 + P7) << 32
eor             t7.16b, t7.16b, t6.16b
and             t6.16b, t6.16b, k00_16.16b

eor             t8.16b, t8.16b, t4.16b
eor             t7.16b, t7.16b, t6.16b

zip2            t5.2d, t8.2d, t4.2d
zip1            t4.2d, t8.2d, t4.2d
zip2            t3.2d, t7.2d, t6.2d
zip1            t6.2d, t7.2d, t6.2d

ext             t4.16b, t4.16b, t4.16b, #15
ext             t5.16b, t5.16b, t5.16b, #14
ext             t6.16b, t6.16b, t6.16b, #13
ext             t3.16b, t3.16b, t3.16b, #12

eor             t4.16b, t4.16b, t5.16b
eor             t6.16b, t6.16b, t3.16b
eor             \rq\().16b, \rq\().16b, t4.16b
eor             \rq\().16b, \rq\().16b, t6.16b

On the Raspberry Pi 3, this code runs 2.8x faster than the generic, table based C code. This is a nice improvement, but we can do even better.

GHASH reduction

The accelerated GHASH implementation that uses the vmull.p64 instruction looks like this:

ext		T2.16b, XL.16b, XL.16b, #8
ext		IN1.16b, T1.16b, T1.16b, #8
eor		T1.16b, T1.16b, T2.16b
eor		XL.16b, XL.16b, IN1.16b

pmull2		XH.1q, XL.2d, SHASH.2d		// a1 * b1
eor		T1.16b, T1.16b, XL.16b
pmull	 	XL.1q, XL.1d, SHASH.1d		// a0 * b0
pmull		XM.1q, T1.1d, SHASH2.1d		// (a1 + a0)(b1 + b0)

eor		T2.16b, XL.16b, XH.16b
ext		T1.16b, XL.16b, XH.16b, #8
eor		XM.16b, XM.16b, T2.16b

pmull		T2.1q, XL.1d, MASK.1d
eor		XM.16b, XM.16b, T1.16b

mov		XH.d[0], XM.d[1]
mov		XM.d[1], XL.d[0]

eor		XL.16b, XM.16b, T2.16b
ext		T2.16b, XL.16b, XL.16b, #8
pmull		XL.1q, XL.1d, MASK.1d

eor		T2.16b, T2.16b, XH.16b
eor		XL.16b, XL.16b, T2.16b

What should be noted here is that the finite field multiplication consists of a multiplication step and a reduction step, where the latter essentially performs the modulo division involving the characteristic polynomial (which is how we normalize the result, i.e., ensure that it remains inside the field)

So while this sequence is optimal for cores that implement vmull.p64 natively, we can switch to a reduction step that does not involve polynomial multiplication at all, and avoid two copies of the fallback vmull.p64 sequence consisting of 40 instructions each.

ext		T2.16b, XL.16b, XL.16b, #8
ext		IN1.16b, T1.16b, T1.16b, #8
eor		T1.16b, T1.16b, T2.16b
eor		XL.16b, XL.16b, IN1.16b

__pmull_p8	XH, XL, SHASH, 2		// a1 * b1
eor		T1.16b, T1.16b, XL.16b
__pmull_p8 	XL, XL, SHASH			// a0 * b0
__pmull_p8	XM, T1, SHASH2			// (a1 + a0)(b1 + b0)

eor		T2.16b, XL.16b, XH.16b
ext		T1.16b, XL.16b, XH.16b, #8
eor		XM.16b, XM.16b, T2.16b

eor		XM.16b, XM.16b, T1.16b

mov		XL.d[1], XM.d[0]
mov		XH.d[0], XM.d[1]

shl		T1.2d, XL.2d, #57
shl		T2.2d, XL.2d, #62
eor		T2.16b, T2.16b, T1.16b
shl		T1.2d, XL.2d, #63
eor		T2.16b, T2.16b, T1.16b
ext		T1.16b, XL.16b, XH.16b, #8
eor		T2.16b, T2.16b, T1.16b

mov		XL.d[1], T2.d[0]
mov		XH.d[0], T2.d[1]

ushr		T2.2d, XL.2d, #1
eor		XH.16b, XH.16b, XL.16b
eor		XL.16b, XL.16b, T2.16b
ushr		T2.2d, T2.2d, #6
ushr		XL.2d, XL.2d, #1

eor		T2.16b, T2.16b, XH.16b
eor		XL.16b, XL.16b, T2.16b

Loop invariants

Another observation one can make when looking at this code is that the vmull.p64 calls that remain all involve right hand sides that are invariants during the execution of the loop. For the version that uses the native vmull.p64, this does not matter much, but for our fallback sequence, it means that some instructions essentially calculate the same value each time, and the computation can be taken out of the loop instead.

Since we have plenty of spare registers on AArch64, we can dedicate 8 of them to prerotated B1/B2/B3/B4 values of SHASH and SHASH2. With that optimization folded in as well, this implementation runs at 4x the speed of the generic GHASH driver. When combined with the bit-sliced AES driver, GCM performance on the Cortex-A53 increases twofold, from 58 to 29 cycles per byte.

The patches implementing this for AArch64 and for AArch32 can be found here.

Time invariant AES

Rule #1 of crypto club: don’t roll your own

Kernel hackers are usually self-righteous bastards who think that they are smarter than everyone else (and I am no exception). Sometimes, it’s hard to fight the urge to reimplement something simply because you think you would have done a slightly better job. On the enterprise scale, this is usually referred to as not invented here syndrome (NIH), and in the past, I have worked for companies where this resulted in entire remote procedure call (RPC) protocol stacks being developed, including a home cooked interface definition language (IDL) and a compiler (yes, I am looking at you, TomTom).

The EDK2 open source project is also full of reinvented wheels, where everything from the build tools to string libraries have been implemented and even invented from scratch. But when it came to incorporating crypto into the code base, they did the right thing, and picked the OpenSSL library, even if this meant putting the burden on the developer to go and find the correct tarball and unpack it in the right place. (Due to license incompatibilities, merging the OpenSSL code into the EDK2 tree would render it undistributable.)

The bottom line, of course, is that you are not smarter than everyone else, and in fact, that there are very smart people out there whose livelihood depends on breaking your supposedly secure system. So instead of reimplementing existing crypto algorithms, or, god forbid, inventing ‘better’ ones, you can spend your time more wisely and learn about existing algorithms and how to use them correctly.

Rule #2 of crypto club: read the manual

Not all encryption modes are suitable for all purposes. For instance, symmetric stream ciphers such as RC4, or AES in CTR mode, should never reuse the same combination of key and initialization vector (IV). This makes stream ciphers mostly unsuitable for disk encryption, which typically derives its IV from the sector number, and sectors are typically written to more than once. (The reason is that, since the key stream is xor’ed with the plaintext to obtain the ciphertext, two ciphertexts encrypted with the same key and IV xor’ed with each other will produce the same value as the two plaintexts xor’ed together, which means updates to disk blocks are essentially visible in the clear. Ouch.)

Many other algorithms have similar limitations: DES had its weak keys, RSA needs padding to be safe, and DSA (as well as ElGamal encryption) should not reuse its k parameter, or its key can be trivially factored out.

Algorithm versus implementation

Unfortunately, we are not there yet. Even after having ticked all the boxes, we may still end up with a system that is insecure. One notable example is AES, which is superb in all other aspects, but, as Daniel J. Bernstein claimed in this paper in 2005, its implementation may be vulnerable to attacks.

In a nutshell, Daniel J. Bernstein’s paper shows that there is an exploitable correlation between the key and the response time of a network service that involves AES encryption, but only when the plaintext is known. This is due to the fact that the implementation performs data dependent lookups in precomputed tables, which are typically 4 – 8 KB in size (i.e., much larger than a typical cacheline), resulting in a variance in the response time.

This may sound peculiar, i.e., if the plaintext is known, what is there to attack, right? But the key itself is also confidential, and AES is also used in a number of MAC algorithms where the plaintext is usually not secret to begin with. Also, the underlying structure of the network protocol may allow the plaintext to be predicted with a reasonable degree of certainty.

For this reason, OpenSSL (which was the implementation under attack in the paper), has switched to time invariant AES implementations as much as possible.

Time invariant AES

On 64-bit ARM, we now have three separate time invariant implementations of AES, one based on the ARMv8 Crypto Extensions and two that are NEON based. On 32-bit ARM, however, the only time invariant AES implementation is the bit sliced NEON one, which is very inefficient when operating in sequential modes such as CBC encryption or CCM/CMAC authentication. (There is an ARMv8 Crypto Extensions implementation for 32-bit ARM as well, but that is currently only relevant for 32-bit kernels running on 64-bit hardware.)

So for Linux v4.11, I have implemented a generic, [mostly] time invariant AES cipher, that should eliminate variances in AES processing time that are correlated with the key. It achieves this by choosing a slightly slower algorithm that is equivalent to the table based AES, but uses only 256 bytes of lookup data (the actual AES S-box), and mixes some S-box values at fixed offsets with the first round key. Every time the key is used, these values need to be xor’ed again, which will pull the entire S-box into the D-cache, hiding the lookup latency of subsequent data dependent accesses.

So if you care more about security than about performance when it comes to networking, for instance, for unmonitored IoT devices that listen for incoming network connections all day, my recommendation is to disable the table based AES, and use the fixed time flavour instead.


The priority based selection rules will still select the much faster NEON code when possible (provided that the CPU has a NEON unit), but this is dependent on the choice of chaining mode.

Algorithm Resolved as
Disk encryption xts(aes) xts-aes-neonbs
mac80211 CMAC cmac(aes) cmac(aes-fixed-time)
VPN ccm(aes) ccm_base(ctr-aes-neonbs,cbcmac(aes-fixed-time))

Accelerated AES for the arm64 Linux kernel

The ARMv8 architecture extends the AArch64 and AArch32 instruction sets with dedicated instructions for AES encryption, SHA-1 and SHA-256 cryptographic hashing, and 64×64 to 128 polynomial multiplication, and implementations of the various algorithms that use these instructions have been added to the ARM and arm64 ports of the Linux kernel over the past couple of years. Given that my main focus is on enterprise class systems, which typically use high end SoCs, I have never felt the urge to spend too much time on accelerated implementations for systems that lack these optional instructions (although I did contribute a plain NEON version of AES in ECB/CBC/CTR/XTS modes back in 2013). Until recently, that is, when I received a Raspberry Pi 3 from my esteemed colleague Joakim Bech, the tech lead of the Linaro Security Working Group. This system is built around a Broadcom SoC containing 4 Cortex-A53 cores that lack the ARMv8 Crypto Extensions, and as it turns out, its AES performance was dreadful.

AES primer

The Advanced Encryption Standard (AES) is a variant of the Rijndael cipher with a fixed block size of 16 bytes, and supports key sizes of 16, 24 and 32 bytes, referred to as AES-128, AES-192 and AES-256, respectively. It consists of a sequence of rounds (10, 12, or 14 for the respective key sizes) that operate on a state that can be expressed in matrix notation as follows:

| a0, b0, c0, d0 |
| a1, b1, c1, d1 |
| a2, b2, c2, d2 |
| a3, b3, c3, d3 |

where each element represents one byte, in column major order (i.e., the elements are assigned from the input in the order a0, a1, a2, a3, b0, b1, etc)

Each round consists of a sequence of operations performed on the state, called AddRoundKey, SubBytes, ShiftRows and MixColumns. All rounds are identical, except for the last one, which omits the MixColumns operation, and performs a final AddRoundKey operation instead.


AES defines a key schedule generation algorithm, which turns the input key into a key schedule consisting of 11, 13 or 15 round keys (depending on key size), of 16 bytes each. The AddRoundKey operation simply xor’s the round key of the current round with the AES state, i.e.,

| a0 ^ rk0, b0 ^ rk4, c0 ^ rk8,  d0 ^ rk12 |
| a1 ^ rk1, b1 ^ rk5, c1 ^ rk9,  d1 ^ rk13 |
| a2 ^ rk2, b2 ^ rk6, c2 ^ rk10, d2 ^ rk14 |
| a3 ^ rk3, b3 ^ rk7, c3 ^ rk11, d3 ^ rk15 |

where rkN refers to byte N of the round key of the current round.


The SubBytes operation is a byte wise substitution, using one of two S-boxes defined by AES, one for encryption and one for decryption. It simply maps each possible 8-bit value onto another 8-bit value, like below

| sub(a0), sub(b0), sub(c0), sub(d0) |
| sub(a1), sub(b1), sub(c1), sub(d1) |
| sub(a2), sub(b2), sub(c2), sub(d2) |
| sub(a3), sub(b3), sub(c3), sub(d3) |


The ShiftRows operation is a transposition step, where all rows of the state except the first one are shifted left or right (for encryption or decryption, respectively), by 1, 2 or 3 positions (depending on the row). For encryption, it looks like this:

| a0, b0, c0, d0 |
| b1, c1, d1, a1 |
| c2, d2, a2, b2 |
| d3, a3, b3, c3 |


The MixColumns operation is also essentially a transposition step, but in a somewhat more complicated manner. It involves the following matrix multiplication, which is carried out in GF(28) using the characteristic polynomial 0x11b. (An excellent treatment of Galois fields can be found here.)

| 2, 3, 1, 1 |   | a0, b0, c0, d0 |
| 1, 2, 3, 1 |   | a1, b1, c1, d1 |
| 1, 1, 2, 3 | x | a2, b2, c2, d2 |
| 3, 1, 1, 2 |   | a3, b3, c3, d3 |

Table based AES

The MixColumns operation is computationally costly when executed sequentially, so it is typically implemented using lookup tables when coded in C. This turns the operation from a transposition into a substitution, which means it can be merged with the SubBytes operation. Even the ShiftRows operation can be folded in as well, resulting in the following transformation:

| 2, 3, 1, 1 |   | sub(a0), sub(b0), sub(c0), sub(d0) |
| 1, 2, 3, 1 |   | sub(b1), sub(c1), sub(d1), sub(a1) |
| 1, 1, 2, 3 | x | sub(c2), sub(d2), sub(a2), sub(b2) |
| 3, 1, 1, 2 |   | sub(d3), sub(a3), sub(b3), sub(c3) |

The generic AES implementation in the Linux kernel implements this by using 4 lookup tables of 256 32-bit words each, where each of those tables corresponds with a column in the matrix on the left, and each element N contains the product of that column with the vector { sub(N) }. (A separate set of 4 lookup tables based on the identity matrix is used in the last round, since it omits the MixColumns operation.)

The combined SubBytes/ShiftRows/MixColumns encryption operation can now be summarized as

| tbl0[a0]  tbl0[b0]  tbl0[c0]  tbl0[d0] |
|   (+)       (+)       (+)       (+)    |
| tbl1[b1]  tbl1[c1]  tbl1[d1]  tbl1[a1] |
|   (+)       (+)       (+)       (+)    |
| tbl2[c2]  tbl2[d2]  tbl2[a2]  tbl2[b2] |
|   (+)       (+)       (+)       (+)    |
| tbl3[d3]  tbl3[a3]  tbl3[b3]  tbl3[c3] |

where tblN refers to each of the lookup tables, (+) refers to exclusive-or, and the AES state columns are represented using 32-bit words.

Note that lookup table based AES is sensitive to cache timing attacks, due to the fact that the memory access pattern during the first round is strongly correlated with the key xor’ed with the plaintext, allowing an attacker to discover key bits if it can observe the cache latencies of the memory accesses.

Please refer to this link for more information about the AES algorithm.

Scalar AES for arm64

The first observation one can make when looking at the structure of the lookup tables is that the 4 tables are identical under rotation of each element by a constant. Since rotations are cheap on arm64, it makes sense to use only a single table, and derive the other values by rotation. Note that this does not reduce the number of lookups performed, but it does reduce the D-cache footprint by 75%.

So for the v4.11 release of the Linux kernel, a scalar implementation of AES has been queued for arm64 that uses just 4 of the the 16 lookup tables from the generic driver. On the Raspberry Pi 3, this code manages 31.8 cycles per byte (down from 34.5 cycles per byte for the generic code). However, this is still a far cry from the 12.9 cycles per byte measured on Cortex-A57 (down from 18.0 cycles per byte), so perhaps we can do better using the NEON. (Note that the dedicated AES instructions manage 0.9 cycles per byte on recent Cortex-A57 versions.)

Accelerated AES using the NEON

The AArch64 version of the NEON instruction set has one huge advantage over other SIMD implementations: it has 32 registers, each 128 bits wide. (Other SIMD ISAs typically have 16 such registers). This means we can load the entire AES S-box (256 bytes) into 16 SIMD registers, and still have plenty of registers left to perform the actual computation, where the tbl/tbx NEON instructions can be used to perform the S-box substitutions on all bytes of the AES state in parallel.

This does imply that we will not be able to implement the MixColumns operation using table lookups, and instead, we will need to perform the matrix multiplication in GF(28) explicitly. Fortunately, this is not as complicated as it sounds: with some shifting, masking and xor’ing, and using a table lookup (using a permute vector in v14) to perform the 32-bit rotation, we can perform the entire matrix multiplication in 9 NEON instructions. The SubBytes operation takes another 8 instructions, since we need to split the 256 byte S-box lookup into 4 separate tbl/tbx instructions. This gives us the following sequence for a single inner round of encryption, where the input AES state is in register v0. (See below for a breakdown of the MixColumns transformation)

/* AddRoundKey (round key pointer in x0) */
ld1	{v15.4s}, [x0], #16
eor	v0.16b, v0.16b, v15.16b

/* SubBytes (S-box in registers v16 - v31) */
movi	v15.16b, #0x40
sub	v9.16b, v0.16b, v15.16b
tbl	v0.16b, {v16.16b-v19.16b}, v0.16b
sub	v10.16b, v9.16b, v15.16b
tbx	v0.16b, {v20.16b-v23.16b}, v9.16b
sub	v11.16b, v10.16b, v15.16b
tbx	v0.16b, {v24.16b-v27.16b}, v10.16b
tbx	v0.16b, {v28.16b-v31.16b}, v11.16b

/* ShiftRows (permutation vector in v13) */
tbl	v0.16b, {v0.16b}, v13.16b

/* MixColumns (v12.16b = { 0x1b, 0x1b, ... }) */
sshr	v8.16b, v0.16b, #7
add	v9.16b, v0.16b, v0.16b
and	v8.16b, v8.16b, v12.16b
eor	v9.16b, v9.16b, v8.16b
rev32	v8.8h, v0.8h
eor	v8.16b, v8.16b, v9.16b
eor	v0.16b, v0.16b, v8.16b
tbl	v0.16b, {v0.16b}, v14.16b	/* ror	v0.4s, v0.4s, #8 */
eor	v0.16b, v0.16b, v8.16b

Looking at the instruction count, one would expect the performance of this algorithm to be around 15 cycles per byte when interleaved 2x or 4x (i.e., the code above, but operating on 2 or 4 AES states in parallel, to eliminate data dependencies between adjacent instructions). However, on the Raspberry Pi 3, this code manages only 22.0 cycles per byte, which is still a huge improvement over the scalar code, but not as fast as we had hoped. This is due to the micro-architectural properties of the tbl/tbx instructions, which take 4 cycles to complete on the Cortex-A53 when using the 4 register variant. And indeed, if we base the estimation on the cycle count, by taking 4 cycles for each such tbl/tbx instruction, and 1 cycle for all other instructions, we get the more realistic number of 21.25 cycles per byte.

As a bonus, this code is not vulnerable to cache timing attacks, given that the memory access patterns are not correlated with the input data or the key.

This code has been part of the arm64 Linux kernel since 2013, but some improvements to it have been queued for v4.11 as well.

Bit sliced AES using the NEON

The AES S-box is not an arbitrary bijective mapping, it has a carefully chosen structure, based again on finite field arithmetic. So rather than performing 16 lookups each round, it is possible to calculate the subsitution values, and one way to do this is described in the paper Faster and Timing-Attack Resistant AES-GCM by Emilia Kaesper and Peter Schwabe. It is based on bit slicing, which is a method to make hardware algorithms suitable for implementation in software. In the AES case, this involves bit slicing 8 blocks of input, i.e., collecting all bits N of each of the 128 bytes of input into NEON register qN. Subsequently, a sequence of logic operations is executed on those 8 AES states in parallel, which mimics the network of logic gates in a hardware implementation of the AES S-box. While software is usually orders of magnitude slower than hardware, the fact that the operations are performed on 128 bits at a time compensates for this.

An implementation of AES using bit slicing is queued for v4.11 as well, which manages 19.8 cycles per byte on the Raspberry Pi 3, which makes it the preferred option for parallelizable modes such as CTR or XTS. It is based on the ARM implementation, which I ported from OpenSSL to the kernel back in 2013, in collaboration with Andy Polyakov, who authored the ARM version of the code originally. However, it has been modified to reuse the key schedule generation routines of the generic AES code, and to use the same expanded key schedule both for encryption and decryption, which reduces the size of the per-key data structure by 1696 bytes.

The code can be found here.


For the Raspberry Pi 3 (as well as any other system using version r0p4 of the Cortex-A53), we can summarize the AES performance as follows:

AES performance (cycles per byte) CBC CTR XTS
encryption 24.7 19.8 20.1
decryption 22.2 19.8 22.7

Appendix: Breakdown of the MixColumns transform using NEON instructions

v0 =  |            a0,            b0,            c0,            d0 |
      |            a1,            b1,            c1,            d1 |
      |            a2,            b2,            c2,            d2 |
      |            a3,            b3,            c3,            d3 |

/* multiply by 'x' (0b10) in GF(2^8) */
sshr	v8.16b, v0.16b, #7
add	v9.16b, v0.16b, v0.16b
and	v8.16b, v8.16b, v12.16b
eor	v9.16b, v9.16b, v8.16b

v9 = |           2a0,           2b0,           2c0,           2d0 |
     |           2a1,           2b1,           2c1,           2d1 |
     |           2a2,           2b2,           2c2,           2d2 |
     |           2a3,           2b3,           2c3,           2d3 |

rev32	v8.8h, v0.8h

v8 = |            a2,            b2,            c2,            d2 |
     |            a3,            b3,            c3,            d3 |
     |            a0,            b0,            c0,            d0 |
     |            a1,            b1,            c1,            d1 |

eor	v8.16b, v8.16b, v9.16b

v8 = |        2a0^a2,        2b0^b2,        2c0^c2,        2d0^d2 |
     |        2a1^a3,        2b1^b3,        2c1^c3,        2d1^d3 |
     |        2a2^a0,        2b2^b0,        2c2^c0,        2d2^d0 |
     |        2a3^a1,        2b3^b1,        2c3^c1,        2d3^d1 |

eor	v0.16b, v0.16b, v8.16b

v0 = |        3a0^a2,        3b0^b2,        3c0^c2,        3d0^d2 |
     |        3a1^a3,        3b1^b3,        3c1^c3,        3d1^d3 |
     |        3a2^a0,        3b2^b0,        3c2^c0,        3d2^d0 |
     |        3a3^a1,        3b3^b1,        3c3^c1,        3d3^d1 |

tbl	v0.16b, {v0.16b}, v14.16b	/* ror	v0.4s, v0.4s, #8 */

v0 = |        3a1^a3,        3b1^b3,        3c1^c3,        3d1^d3 |
     |        3a2^a0,        3b2^b0,        3c2^c0,        3d2^d0 |
     |        3a3^a1,        3b3^b1,        3c3^c1,        3d3^d1 |
     |        3a0^a2,        3b0^b2,        3c0^c2,        3d0^d2 |

eor	v0.16b, v0.16b, v8.16b
v0 =  | 2a0^3a1^a2^a3, 2b0^3b1^b2^b3, 2c0^3c1^c2^c3, 2d0^3d1^d2^d3 |
      | 2a1^3a2^a3^a0, 2b1^3b2^b3^b0, 2c1^3c2^c3^c0, 2d1^3d2^d3^d0 |
      | 2a2^3a3^a0^a1, 2b2^3b3^b0^b1, 2c2^3c3^c0^c1, 2d2^3d3^d0^d1 |
      | 2a3^3a0^a1^a2, 2b3^3b0^b1^b2, 2c3^3c0^c1^c2, 2d3^3d0^d1^d2 |