This is the artifact corresponding to the paper "Vectorized Falcon-Sign Implementations using SSE2, AVX2, AVX-512, NEON, and RVV" (IACR TCHES 2026).

Directory structure:

• help/: some helper scripts
• opt/: optimized code implementations, supporting three target platforms
• profiling/: benchmarks for some subroutines, such as BaseSampler, FFT/iFFT
• ref/: reference implementation, derived from the public domain C-FN-DSA project (https://github.com/pornin/c-fn-dsa at commit id 96e3b92)

The three target platforms used in our paper:

• The Intel i7-11700K CPU (Rocket Lake microarchitecture) operating at 3.6 GHz. Hyper-Threading and Turbo Boost are disabled. Ubuntu 24.04 with GCC 13.3.0.
• The Cortex-A72 processor in Raspberry Pi 4B running at 1.5 GHz. Ubuntu 20.04 with Clang 10.0.0.
• The SpacemiT X60 core in Milk-V Jupiter operating at 2.0 GHz, supporting the RV64GCBV instruction set with vector extension v1.0 (VLEN = 256 bits) and bit-manipulation extension v1.0.0. Bianbu 1.0.15 (Linux kernel 6.1.15) with GCC 13.2.0. The Bianbu 1.0.15 firmware can be found at jupiter-bianbu-build v1.0.15.

To precisely reproduce the performance data reported in our paper, ensure your hardware and software environment is as consistent as possible with our experimental environment (Section 2 of our paper).

Each of the directories opt/, profiling/, and ref/ contains three different Makefile files for compiling code for different platforms, namely Makefile, Makefile.armv8a, and Makefile.rv. You should specify the correct Makefile for compilation. For example, in the profiling/ directory:

• On the Intel i7-11700K, you can compile using: make all -j
• On the ARM Cortex-A72, you can compile using: make all -j -f Makefile.armv8a
• On the SpacemiT X60, you can compile using: make all -j -f Makefile.rv

On the three platforms mentioned in this project, you will need to perform some configuration to obtain the CPU clock cycle. You can find detailed explanations in the comments section of ref/speed_fndsa.c:

• For Intel CPU: Run the command as root: echo 2 > /sys/bus/event_source/devices/cpu/rdpmc and run sysctl -w kernel.perf_event_paranoid=-1
• For AArch64: Follow the instructions at https://github.com/jerinjacobk/armv8_pmu_cycle_counter_el0
• For RISC-V: Simply put, you need to run the executable file using perf stat ./speed_fndsa. For example, you can find the command in opt/Makefile.rv: perf stat ./out/speed_fndsa_rv64gc >>speed_fndsa_x60.txt 2>/dev/null

If you are using the Bianbu 1.0.15 mentioned above, you don't need to worry about this issue. However, if you are using a newer version, you may encounter the following problem.

Regarding obtaining CPU clock cycles on RISC-V, in opt/speed_fndsa.c you will see this code: __asm__ __volatile__("rdcycle %0" : "=r"(x)); This will produce an error in newer Linux kernel versions (e.g., 6.6.63) because rdcycle causes "Illegal Instruction", see https://forum.banana-pi.org/t/how-to-enable-rdinstret-and-rdcycle-on-bananapi-bpi-f3/19212 and camel-cdr/rvv-bench-results#1. Therefore, it is recommended that you use the perf tool to obtain the CPU clock cycles.

In fact, the relevant speed tests under profiling/ in this project are based on the perf tool, see profiling/cpucycles.c. Therefore, if you need to modify the method for obtaining clock cycles in opt/speed_fndsa.c, you can refer to the relevant speed tests in profiling/.

To reproduce "Table 3: The performance profiling of Falcon-1024’s signature generation", you first need to install gperftools, which can be done with the following commands:

# Install the dependencies
sudo apt install build-essential autoconf libtool
git clone https://github.com/gperftools/gperftools.git
cd gperftools
./autogen.sh
./configure
make
sudo make install

After running the above commands, the libraries will be installed in /usr/local/lib. Then install the pprof tool. We recommend setting up a Go language environment first, then installing pprof via go install github.com/google/pprof@latest, and ensure its path is added to your PATH.

For the Intel i7-11700K, go to the ref/ directory:

make all -j
make run_profiling

You will then get multiple txt files, such as gperf_sign_core_1024_avx2.txt, which contain the profiling results of the AVX2 version.

For the SpacemiT X60, the steps are similar: first install gperftools and pprof, then run make run_profiling -f Makefile.rv in the ref/ directory.

The help/process_gperf.ipynb file might be helpful in processing the above profiling results.

To reproduce "Table 4: Benchmark results of various BaseSampler implementations" and "Table 5: Benchmark results of FFT/iFFT implementations on SpacemiT X60.", the main work is in the profiling/ directory.

For the Intel i7-11700K:

make all -j
make run_speed

You will then get speed_gaussian0_11700k.txt, which contains the experimental results of BaseSampler for SSE2, AVX2, and AVX-512F instruction sets.

If you want to test the correctness of our BaseSampler implementation, run make run_test. If no output from the diff command is observed, it indicates that the test passed.

For the Cortex-A72:

make all -j -f Makefile.armv8a
make run_speed -f Makefile.armv8a

The file speed_gaussian0_cortex_a72.txt you get contains the experimental results of BaseSampler for the NEON instruction set.

For the SpacemiT X60:

make all -j -f Makefile.rv
make run_speed -f Makefile.rv

The file speed_gaussian0_x60.txt you get contains the experimental results of BaseSampler for the RISC-V instruction set. The file speed_fft_rv64d_x60.txt you get contains the experimental results of FFT/iFFT for the RISC-V instruction set.

To reproduce "Table 6: Benchmark results of Falcon-{512,1024}’s signature generation (sign_core subroutine) on three target platforms (8 distinct instruction set configurations).", the main work is in the ref/ and opt/ directories.

First, reproduce the results of the reference implementations in the ref/ directory.

For the Intel i7-11700K:

make all -j
make run_speed

You will then get speed_fndsa_11700k.txt, which contains the experimental results of the reference implementations for the sign_core subroutine for SSE2, AVX2, and AVX-512F instruction sets.

For the Cortex-A72:

make all -j -f Makefile.armv8a
make run_speed -f Makefile.armv8a

The file speed_fndsa_cortex_a72.txt you get contains the experimental results of the reference implementations for the sign_core subroutine for the NEON instruction set.

For the SpacemiT X60:

make all -j -f Makefile.rv
make run_speed -f Makefile.rv

The file speed_fndsa_x60.txt you get contains the experimental results of the reference implementations for the sign_core subroutine for the RISC-V instruction set.

Then, reproduce the results of our optimized implementations in the opt/ directory. The commands and the filenames of the files you get are the same as those in the ref/ directory, so they are not repeated here.

In Section 7, we mentioned: "For implementations using NEON, the performance improvement is 17% compared to the reference implementation. If we exclude the 4-way hybrid Keccak and optimized FFT/iFFT, the improvement reduces to 9%. Integrating our BaseSampler with the 4-way hybrid Keccak results in a 13% improvement over the reference implementation."

If you want to reproduce the result on Cortex-A72 for "If we exclude the 4-way hybrid Keccak and optimized FFT/iFFT, the improvement reduces to 9%": In the opt/ directory, change Makefile.armv8a to -DFNDSA_NEON_HYBRID_SHA3=0 -DFNDSA_NEON_FFT_OPT=0, then:

make clean -f Makefile.armv8a
make all -j -f Makefile.armv8a
make run_speed -f Makefile.armv8a

If you want to reproduce the result on Cortex-A72 for "Integrating our BaseSampler with the 4-way hybrid Keccak results in a 13% improvement over the reference implementation": In the opt/ directory, change Makefile.armv8a to -DFNDSA_NEON_HYBRID_SHA3=1 -DFNDSA_NEON_FFT_OPT=0, then:

make clean -f Makefile.armv8a
make all -j -f Makefile.armv8a
make run_speed -f Makefile.armv8a

In Section 7, we mentioned: "All four versions on RISC-V show significant improvements. ... Without the optimized Keccak, the improvement is 41% compared to the reference implementation."

If you want to reproduce the above result on SpacemiT X60: In the opt/ directory, change Makefile.rv to -DKECCAK_OPT=0, then:

make clean -f Makefile.rv
make all -j -f Makefile.rv
make run_speed -f Makefile.rv

Section 7 mentions "our implementation using AVX2 increases the code size by approximately 2.7 KB compared to the reference implementation" To reproduce this result, run the following commands in the ref/ and opt/ directories respectively, and then compare the results:

nm out/speed_fndsa_avx2 --print-size --size-sort --radix=d | \
awk '{$1=""}1' | \
awk '{sum+=$1 ; print $0} END{print "Total size =", sum, "bytes =", sum/1024, "kB"}' > speed_fndsa_avx2_symbols_size.txt

We thank the artifact evaluation reviewers of TCHES 2026 for their valuable feedback.