Reference, highly optimized, masked C and ASM implementations of Ascon

Ascon is a family of lightweight cryptographic algorithms and consists of:

• Authenticated encryption schemes with associated data (AEAD)
• Hash functions (HASH) and extendible output functions (XOF)
• Pseudo-random functions (PRF) and message authentication codes (MAC)

All implementations use the "ECRYPT Benchmarking of Cryptographic Systems (eBACS)" interface:

• https://bench.cr.yp.to/call-aead.html for AEAD (Ascon-128, Ascon-128a, Ascon-80pq)
• https://bench.cr.yp.to/call-hash.html for HASH and XOF (Ascon-Hash, Ascon-Hasha, Ascon-Xof, Ascon-Xofa)
• https://nacl.cr.yp.to/auth.html for PRF and MAC (Ascon-Mac, Ascon-Prf, Ascon-PrfShort)

For more information on Ascon visit: https://ascon.iaik.tugraz.at/

TL;DR

If you do not know where to start, use the reference implementations (self-contained, portable, very fast):

• crypto_aead/ascon128v12/ref
• crypto_aead/ascon128av12/ref
• crypto_hash/asconxofv12/ref
• crypto_hash/asconxofav12/ref

Algorithms

This repository contains implementations of the following 10 Ascon v1.2 algorithms:

• crypto_aead/ascon128v12: Ascon-128 (authenticated encryption)
• crypto_aead/ascon128av12: Ascon-128a (authenticated encryption)
• crypto_aead/ascon80pqv12: Ascon-80pq (authenticated encryption)
• crypto_hash/asconhashv12: Ascon-Hash (hash function)
• crypto_hash/asconhashav12: Ascon-Hasha (hash function)
• crypto_hash/asconxofv12: Ascon-Xof (hash function with extendable output, XOF)
• crypto_hash/asconxofav12: Ascon-Xofa (hash function with extendable output, XOF)
• crypto_auth/asconmacv12: Ascon-Mac (message authentication code)
• crypto_auth/asconprfv12: Ascon-Prf (message authentication code with extendable output, PRF)
• crypto_auth/asconprfsv12: Ascon-PrfShort (message authentication code for very short messages)

We also provide two combined algorithm implementations supporting both AEAD and hashing:

• crypto_aead_hash/asconv12: Ascon-128 combined with Ascon-Hash
• crypto_aead_hash/asconav12: Ascon-128a combined with Ascon-Hasha

The following algorithms demonstrate the performance improvement of Ascon on 32-bit platforms without bit interleaving overhead. Bit interleaving could be performed externally on the host side or using a dedicated instruction (e.g. using the ARM Custom Datapath Extension). Note that a similar performance improvement could be achieved using funnel shift instructions (available on some 32-bit RISC-V extensions).

• crypto_aead/ascon128bi32v12: Ascon-128 (+17% on ARM1176JZF-S)
• crypto_aead/ascon128abi32v12: Ascon-128a (+23% on ARM1176JZF-S)
• crypto_hash/asconhashbi32v12: Ascon-Hash (+5% on ARM1176JZF-S)
• crypto_hash/asconhashabi32v12: Ascon-Hasha (+8% on ARM1176JZF-S)
• crypto_aead_hash/asconbi32v12: Ascon-128 combined with Ascon-Hash
• crypto_aead_hash/asconabi32v12: Ascon-128a combined with Ascon-Hasha

Implementations

For most algorithms, we provide the following pure C implementations:

• ref: reference implementation
• opt64: 64-bit speed-optimized
• opt32: 32-bit speed-optimized
• opt64_lowsize: 64-bit size-optimized
• opt32_lowsize: 32-bit size-optimized
• bi32: 32-bit speed-optimized bit-interleaved
• bi32_lowreg: 32-bit speed-optimized bit-interleaved (low register usage)
• bi32_lowsize: 32-bit size-optimized bit-interleaved
• esp32: 32-bit ESP32 optimized
• opt8: 8-bit size- and speed-optimized
• bi8: 8-bit optimized bit-interleaved

the following C with inline or partial ASM implementations:

• avx512: 320-bit speed-optimized AVX512
• neon: 64-bit speed-optimized ARM NEON
• armv6: 32-bit speed-optimized ARMv6
• armv6m: 32-bit speed-optimized ARMv6-M
• armv7m: 32-bit speed-optimized ARMv7-M
• armv6_lowsize: 32-bit size-optimized ARMv6
• armv6m_lowsize: 32-bit size-optimized ARMv6-M
• armv7m_lowsize: 32-bit size-optimized ARMv7-M
• armv7m_small: 32-bit small speed-optimized ARMv7-M
• bi32_armv6: 32-bit speed-optimized bit-interleaved ARMv6
• bi32_armv6m: 32-bit speed-optimized bit-interleaved ARMv6-M
• bi32_armv7m: 32-bit speed-optimized bit-interleaved ARMv7-M
• bi32_armv7m_small: 32-bit small bit-interleaved ARMv7-M
• avr: 8-bit size- and speed-optimized AVR
• avr_lowsize: 8-bit size-optimized AVR

the following ASM implementations:

• asm_esp32: 32-bit optimized ESP32 using funnel-shift instructions
• asm_rv32i: 32-bit optimized RV32I using the base instruction set
• asm_rv32b: 32-bit optimized RV32B using bitmanip base (Zbb)
• asm_fsr_rv32b: 32-bit optimized funnel-shift RV32B using bitmanip base and bitmanip terniary (ZbbZbt)
• asm_bi32_rv32b: 32-bit optimized bit-interleaved RV32B using bitmanip base and bitmanip permutations (ZbbZbp)

and the following high-level masked (shared) C with inline ASM implementations:

• protected_bi32_armv6: 32-bit masked bit-interleaved ARMv6
• protected_bi32_armv6_leveled: 32-bit masked and leveled bit-interleaved ARMv6

The masked C implementations can be used as a starting point to generate device specific C/ASM implementations. Note that the masked C implementations require a minimum amount of ASM instructions. Otherwise, the compiler may heavily optimize the code and even combine shares. Obviously, the output generated is very sensitive to compiler and environment changes and any generated output needs to be security evaluated. A preliminary evaluation of these implementations has been performed on some ChipWhisperer devices. The setup and preliminary results can found at: https://github.com/ascon/simpleserial-ascon

Performance results on different CPUs in cycles per byte

Ascon-128a

Message Length in Bytes	1	8	16	32	64	1536	long
AMD EPYC 7742*					7.4	4.4	4.2
AMD Ryzen 9 5950X*					8.1	5.3	5.2
Apple M1 (ARMv8)*					9.4	6.3	6.3
Cortex-A72 (ARMv8)*					10.9	7.2	7.0
Intel Xeon E5-2609 v4*					11.3	7.4	7.2
Intel Core i5-6300U	365	47	31	19	13.5	8.0	7.8
Intel Core i5-4200U	519	67	44	27	18.8	11.0	10.6
Cortex-A9 (ARMv7)*					42.8	24.6	24.0
Cortex-A7 (NEON)	2204	226	132	82	55.9	31.7	30.7
Cortex-A7 (ARMv7)*					55.5	38.2	37.5
ARM1176JZF-S (ARMv6)	1908	235	156	99	70.4	43.0	42.9

Ascon-128 and Ascon-80pq

Message Length in Bytes	1	8	16	32	64	1536	long
AMD EPYC 7742*					8.1	6.6	6.5
AMD Ryzen 9 5950X*					11.0	8.2	8.1
Apple M1 (ARMv8)*					12.5	9.5	9.3
Cortex-A72 (ARMv8)*					13.8	10.7	10.5
Intel Xeon E5-2609 v4*					14.9	10.8	10.6
Intel Core i5-6300U	367	58	35	23	17.6	11.9	11.4
Intel Core i5-4200U	521	81	49	32	23.9	16.2	15.8
Cortex-A9 (ARMv7)*					51.7	34.1	33.3
Cortex-A7 (NEON)	2182	249	148	97	71.7	47.5	46.5
Cortex-A7 (ARMv7)*					69.6	52.0	51.6
ARM1176JZF-S (ARMv6)	1921	277	167	112	83.7	57.2	56.8

Ascon-Hasha and Ascon-Xofa

Message Length in Bytes	1	8	16	32	64	1536	long
AMD EPYC 7742*
AMD Ryzen 7 1700*					22.0	12.1	11.7
Apple M1 (ARMv8)*
Cortex-A72 (ARMv8)*					22.2	14.5	14.2
Intel Xeon E5-2609 v4*					23.3	14.4	14.0
Intel Core i5-6300U	550	83	49	33	23.7	15.6	15.5
Intel Core i5-4200U	749	112	67	44	31.8	20.8	20.7
Cortex-A9 (ARMv7)*					87.5	45.6	44.0
Cortex-A7 (ARMv7)*					102.3	63.5	61.8
ARM1176JZF-S (ARMv6)	2390	356	211	138	100.7	65.7	65.3

Ascon-Hash and Ascon-Xof

Message Length in Bytes	1	8	16	32	64	1536	long
AMD EPYC 7742*					21.1	13.3	12.4
AMD Ryzen 9 5950X*					24.1	16.1	15.8
Apple M1 (ARMv8)*					29.2	19.6	18.5
Cortex-A72 (ARMv8)*					30.5	20.5	20.0
Intel Xeon E5-2609 v4*					31.9	21.4	21.2
Intel Core i5-6300U	747	114	69	46	34.2	23.2	23.1
Intel Core i5-4200U	998	153	92	61	45.5	30.9	30.7
Cortex-A9 (ARMv7)*					95.8	55.5	53.9
Cortex-A7 (ARMv7)*					138.1	89.9	88.8
ARM1176JZF-S (ARMv6)	3051	462	277	184	137.3	92.6	92.2

Ascon-Mac and Ascon-Prf

Message Length in Bytes	1	8	16	32	64	1536	long
Intel Core i5-6300U	369	46	24	18	11.7	6.4	6.3
Intel Core i5-4200U	506	63	32	24	16.2	8.8	8.7
ARM1176JZF-S (ARMv6)	1769	223	117	85	57.5	31.9	31.6

Ascon-PrfShort

Message Length in Bytes	1	8	16	32	64	1536	long
Intel Core i5-6300U	185	23	12	-	-	-	-
Intel Core i5-4200U	257	33	17	-	-	-	-
ARM1176JZF-S (ARMv6)	1057	132	69	-	-	-	-

* Results taken from eBACS: http://bench.cr.yp.to/

Build and test

Build and test all Ascon C targets using release flags (-O2 -fomit-frame-pointer -march=native -mtune=native):

mkdir build && cd build
cmake ..
cmake --build .
ctest

Build and test all Ascon C targets on Windows:

mkdir build && cd build
cmake ..
cmake --build . --config Release
ctest -C Release

Build and test all Ascon C targets using debug flags (with NIST defined flags and sanitizers):

mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Debug
cmake --build .
ctest

Manually set the compiler and/or release flags (e.g. to disable -march=native -mtune=native).

mkdir build && cd build
cmake .. -DCMAKE_C_COMPILER=clang -DREL_FLAGS="-O2;-fomit-frame-pointer"
cmake --build .
ctest

Build and run only specific algorithms, implementations and tests:

mkdir build && cd build
cmake .. -DALG_LIST="ascon128;asconhash" -DIMPL_LIST="opt64;bi32" -DTEST_LIST="genkat"
cmake --build .
ctest

Note that cmake stores variables in a cache. Therefore, variables can be set one-by-one, unset using e.g. cmake . -UIMPL_LIST and shown using cmake . -L:

mkdir build && cd build
cmake ..
cmake . -DALG_LIST="ascon128;asconhash"
cmake . -DIMPL_LIST="opt64;bi32"
cmake . -DTEST_LIST="genkat"
cmake . -L
cmake --build .
ctest

Cross compile and test with custom emulator using e.g. qemu-arm:

mkdir build && cd build
cmake .. -DCMAKE_C_COMPILER="arm-linux-gnueabi-gcc" \
         -DREL_FLAGS="-O2;-fomit-frame-pointer;-march=armv7;-mtune=cortex-m4" \
         -DEMULATOR="qemu-arm;-L;/usr/arm-linux-gnueabi" \
         -DALG_LIST="ascon128;ascon128a" -DIMPL_LIST="armv7m;bi32_armv7m"
cmake --build .
ctest

or using Intel SDE (use full path to sde or add to path variable):

mkdir build && cd build
cmake .. -DCMAKE_C_COMPILER=gcc -DIMPL_LIST=avx512 -DEMULATOR="sde;--" \
         -DREL_FLAGS="-O2;-fomit-frame-pointer;-march=icelake-client"
cmake --build .
ctest

Build and benchmark:

Build the getcycles test:

mkdir build && cd build
cmake .. -DALG_LIST="ascon128;asconhash" -DIMPL_LIST="opt32;opt32_lowsize" -DTEST_LIST="getcycles"
cmake --build .

Get the CPU cycle performance:

./getcycles_crypto_aead_ascon128v12_opt32
./getcycles_crypto_aead_ascon128v12_opt32_lowsize
./getcycles_crypto_hash_asconhashv12_opt32
./getcycles_crypto_hash_asconhashv12_opt32_lowsize

Get the implementation size:

size -t libcrypto_aead_ascon128v12_opt32.a
size -t libcrypto_aead_ascon128v12_opt32_lowsize.a
size -t libcrypto_hash_asconhashv12_opt32.a
size -t libcrypto_hash_asconhashv12_opt32_lowsize.a

Manually build and run a single Ascon target:

Build example for AEAD algorithms:

gcc -march=native -O3 -Icrypto_aead/ascon128v12/opt64 crypto_aead/ascon128v12/opt64/*.c -Itests tests/genkat_aead.c -o genkat
gcc -march=native -O3 -Icrypto_aead/ascon128v12/opt64 crypto_aead/ascon128v12/opt64/*.c -DCRYPTO_AEAD -Itests tests/getcycles.c -o getcycles

Build example for HASH algorithms:

gcc -march=native -O3 -Icrypto_hash/asconhashv12/opt64 crypto_hash/asconhashv12/opt64/*.c -Itests tests/genkat_hash.c -o genkat
gcc -march=native -O3 -Icrypto_hash/asconhashv12/opt64 crypto_hash/asconhashv12/opt64/*.c -DCRYPTO_HASH -Itests tests/getcycles.c -o getcycles

Generate KATs and get CPU cycles:

./genkat
./getcycles

Manually build and run an RV32 target:

Setup:

sudo apt install gcc-riscv64-unknown-elf picolibc-riscv64-unknown-elf qemu-system-misc

Example to build, run and test an AEAD/HASH algorithm using gcc, picolibc and qemu:

riscv64-unknown-elf-gcc -O2 -march=rv32i -mabi=ilp32 --specs=picolibc.specs --oslib=semihost --crt0=hosted -Ttests/rv32.ld \
    -Icrypto_aead/ascon128v12/asm_rv32i crypto_aead/ascon128v12/asm_rv32i/*.[cS] -Itests tests/genkat_aead.c -o genkat
qemu-system-riscv32 -semihosting-config enable=on -monitor none -serial none -nographic -machine virt,accel=tcg -cpu rv32 -bios none -kernel genkat
diff LWC_AEAD_KAT_128_128.txt crypto_aead/ascon128v12/LWC_AEAD_KAT_128_128.txt

riscv64-unknown-elf-gcc -O2 -march=rv32i -mabi=ilp32 --specs=picolibc.specs --oslib=semihost --crt0=hosted -Ttests/rv32.ld \
    -Icrypto_hash/asconhashv12/opt32 crypto_hash/asconhashv12/opt32/*.[cS] -Itests tests/genkat_hash.c -o genkat
qemu-system-riscv32 -semihosting-config enable=on -monitor none -serial none -nographic -machine virt,accel=tcg -cpu rv32 -bios none -kernel genkat
diff LWC_HASH_KAT_256.txt crypto_hash/asconhashv12/LWC_HASH_KAT_256.txt

Manually build and run an AVR target:

Example to build, run and test an AEAD algorithm using avr-gcc, avr-libc and simavr.

Setup:

sudo apt install gcc-avr avr-libc simavr
git clone https://github.com/JohannCahier/avr_uart.git

Single test vector using demo and performance measurement using getcycles:

avr-gcc -mmcu=atmega128 -std=c99 -Os -Icrypto_aead/ascon128v12/opt8 crypto_aead/ascon128v12/opt8/*.[cS] \
    -DAVR_UART -Iavr_uart avr_uart/avr_uart.c -Wno-incompatible-pointer-types -Wno-cpp \
    -DCRYPTO_AEAD -Itests tests/demo.c -o demo
simavr -m atmega128 ./demo

avr-gcc -mmcu=atmega128 -std=c99 -Os -Icrypto_aead/ascon128v12/opt8 crypto_aead/ascon128v12/opt8/*.[cS] \
    -DAVR_UART -Iavr_uart avr_uart/avr_uart.c -Wno-incompatible-pointer-types -Wno-cpp \
    -DCRYPTO_AEAD -Itests tests/getcycles.c -o getcycles
simavr -t -m atmega128 ./getcycles

Generate all test vectors for AEAD/HASH and write result to a file. Press Ctrl-C to quit simavr after about a minute:

avr-gcc -mmcu=atmega128 -std=c99 -Os -Icrypto_aead/ascon128v12/opt8 crypto_aead/ascon128v12/opt8/*.[cS] \
    -DAVR_UART -Iavr_uart avr_uart/avr_uart.c -Wno-incompatible-pointer-types -Wno-cpp \
    -Itests tests/genkat_aead.c -o genkat_aead
echo "Press Ctrl-C to quit simavr after about a minute"
simavr -t -m atmega128 ./genkat_aead 2> LWC_AEAD_KAT_128_128.txt
sed -i -e 's/\x1b\[[0-9;]*m//g' -e 's/\.\.$//' LWC_AEAD_KAT_128_128.txt
diff LWC_AEAD_KAT_128_128.txt crypto_aead/ascon128v12/LWC_AEAD_KAT_128_128.txt

avr-gcc -mmcu=atmega128 -std=c99 -Os -Icrypto_hash/asconhashv12/opt8 crypto_hash/asconhashv12/opt8/*.[cS] \
    -DAVR_UART -Iavr_uart avr_uart/avr_uart.c -Wno-incompatible-pointer-types -Wno-cpp \
    -Itests tests/genkat_hash.c -o genkat_hash
echo "Press Ctrl-C to quit simavr after about a minute"
simavr -t -m atmega128 ./genkat_hash 2> LWC_HASH_KAT_256.txt
sed -i -e 's/\x1b\[[0-9;]*m//g' -e 's/\.\.$//' LWC_HASH_KAT_256.txt
diff LWC_HASH_KAT_256.txt crypto_hash/asconhashv12/LWC_HASH_KAT_256.txt

Benchmarking

Hints to get more reliable getcycles results on Intel/AMD CPUs:

• Determine the processor base frequency (also called design frequency):

  • e.g. using the Intel/AMD website
  • or using lscpu listed under model name

• Disable turbo boost (this should lock the frequency to the next value below the processor base frequency):

  echo 1 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo
  

• If the above does not work, manually set the frequency using e.g. cpufreq-set.

• Determine the actual frequency (under load):

  • e.g. by watching the frequency using lscpu or cpufreq-info

• Determine the scaling factor between the actual and base frequency:

  • factor = actual frequency / base frequency

• Run a getcycles program using the frequency factor and watch the results:

  while true; do ./getcycles_crypto_aead_ascon128v12_opt64 $factor; done
  

• Run the benchmark-getcycles.sh script with the frequency factor and a specific algorithm to benchmark all corresponding getcycles implementations:

  scripts/benchmark-getcycles.sh $factor ascon128
  

Hints to activate the performance monitor unit (PMU) on ARM CPUs:

• First try to install linux-tools and see if it works.

• On many ARM platforms, the PMU has to be enabled using a kernel module:

  • Source code for Armv6 (32-bit): http://sandsoftwaresound.net/raspberry-pi/raspberry-pi-gen-1/performance-counter-kernel-module/
  • Source code for Armv7 (32-bit): https://github.com/thoughtpolice/enable_arm_pmu
  • Source code for Armv8/Aarch64 (64-bit): https://github.com/rdolbeau/enable_arm_pmu

• Steps to compile the kernel module on the raspberry pi:

  • Find out the kernel version using uname -a
  • Download the kernel header files, e.g. raspberrypi-kernel-header
  • Download the source code for the Armv6 kernel module
  • Build, install and load the kernel module

Benchmark Ascon v1.2 using supercop

Download supercop according to the website: http://bench.cr.yp.to/supercop.html

To test only Ascon, just run the following commands:

./do-part init
./do-part crypto_aead ascon128v12
./do-part crypto_aead ascon128av12
./do-part crypto_aead ascon80pqv12
./do-part crypto_hash asconhashv12
./do-part crypto_hash asconxofv12

Show the cycles/Byte for a 1536 Byte long message:

cat bench/*/data | grep '_cycles 1536 ' | awk '{printf "%.1f\t%s\t%s\n", $9/$8,
$6, $7}' | sort -nr

Evaluate and optimize Ascon on constraint devices:

• The ascon-c code allows to set compile-time parameters ASCON_INLINE_MODE (IM), ASCON_INLINE_PERM (IP), ASCON_UNROLL_LOOPS (UL), ASCON_INLINE_BI (IB), via command line or in the crypto_*/ascon*/*/config.h files.
• Use the benchmark-config.sh script to evaluate all combinations of these parameters for a given list of Ascon implementations. The script is called with an output file, frequency factor, the algorithm, and the list of implementations to test:

  scripts/benchmark-config.sh results-config.md $factor ascon128 ref opt64 opt64_lowsize
  

• The results-config.md file then contains a markup table with size and cycles for each implementation and parameter set to evaluate several time-area trade-offs.
• The benchmark-all.sh and benchmark-size.sh scripts provides a time/size and size-only table of all currently compiled implementations:

  scripts/benchmark-all.sh results-all.md
  scripts/benchmark-size.sh results-size.md