Benchmark v1.1.1

Usage:

Usage: ./randomx-benchmark [OPTIONS]
Supported options:
  --help        shows this message
  --mine        mining mode: 2080 MiB
  --verify      verification mode: 256 MiB
  --jit         JIT compiled mode (default: interpreter)
  --secure      W^X policy for JIT pages (default: off)
  --largePages  use large pages (default: small pages)
  --softAes     use software AES (default: hardware AES)
  --threads T   use T threads (default: 1)
  --affinity A  thread affinity bitmask (default: 0)
  --init Q      initialize dataset with Q threads (default: 1)
  --nonces N    run N nonces (default: 1000)
  --seed S      seed for cache initialization (default: 0)

Changes:

• fix MSYS thread handle cast #121
• Optimized loading from scratchpad #124
• JIT compiler for ARMv8 #125

Benchmark v1.1.0

Performance may be slightly lower than v1.0.4 on some CPUs due to #118.

Using the new --secure option may decrease mining performance by up to 30%, but provides slightly more secure memory permission management.

Usage: ./randomx-benchmark [OPTIONS]
Supported options:
  --help        shows this message
  --mine        mining mode: 2080 MiB
  --verify      verification mode: 256 MiB
  --jit         x86-64 JIT compiled mode (default: interpreter)
  --secure      W^X policy for JIT pages (default: off)
  --largePages  use large pages (default: small pages)
  --softAes     use software AES (default: hardware AES)
  --threads T   use T threads (default: 1)
  --affinity A  thread affinity bitmask (default: 0)
  --init Q      initialize dataset with Q threads (default: 1)
  --nonces N    run N nonces (default: 1000)
  --seed S      seed for cache initialization (default: 0)

Changes:

• Fix undefined behavior in rotr/rotl (#86)
• Fix: possible overflow when using non-standard Dataset size (#87)
• Sanity checks (#88)
• Use cmake for building (#90)
• Vector intrinsics for ARMv8 (#93)
• Refactoring (#95)
• benchmark: affinity cast for mingw (#96)
• Do not attempt to allocate >= 4 GiB on 32-bit systems (#99)
• Fix: hardcoded JIT code buffer size (#98)
• Fix a compiler error on Apple OSX (#110)
• Fix use-after-free in benchmark (debug build)
• Changes recommended by Quarkslab (#111)
• Optional W^X policy for JIT pages (#112)
• Increase the frequency of CBRANCH (#118)

v1.0.4

Benchmark usage:

RandomX benchmark v1.0.4
Usage: bin/randomx-benchmark [OPTIONS]
Supported options:
  --help        shows this message
  --mine        mining mode: 2080 MiB
  --verify      verification mode: 256 MiB
  --jit         x86-64 JIT compiled mode (default: interpreter)
  --largePages  use large pages
  --softAes     use software AES (default: x86 AES-NI)
  --threads T   use T threads (default: 1)
  --affinity A  thread affinity bitmask (default: 0)
  --init Q      initialize dataset with Q threads (default: 1)
  --nonces N    run N nonces (default: 1000)
  --seed S      seed for cache initialization (default: 0)

Changes:

• Dataset prefetch in interpreted VM #52
• Support building a shared library #53
• Use strongly typed enums #55
• Fix cmake standalone build #56
• Fix Windows detection #58
• Configuration guidelines #59
• Fix possible overflow in static_assert #67
• benchmark: add thread affinity mask option #68
• Regression tests #73
• Large page support for FreeBSD #74
• Different round keys for columns 0,1 and 2,3 in AesGenerator4R #76
• Decrease the frequency of FADD/FSUB in favor of FMUL #77

Improved program generation

RandomX benchmark
Usage: benchmark.exe [OPTIONS]
Supported options:
  --help        shows this message
  --mine        mining mode: 2080 MiB
  --verify      verification mode: 256 MiB
  --jit         x86-64 JIT compiled mode (default: interpreter)
  --largePages  use large pages
  --softAes     use software AES (default: x86 AES-NI)
  --threads T   use T threads (default: 1)
  --init Q      initialize dataset with Q threads (default: 1)
  --nonces N    run N nonces (default: 1000)
  --seed S      seed for cache initialization (default: 0)

Changes in this release:

• the number of AES rounds for program generation increased from 1 to 4

Minor changes

RandomX benchmark
Usage: benchmark.exe [OPTIONS]
Supported options:
  --help        shows this message
  --mine        mining mode: 2080 MiB
  --verify      verification mode: 256 MiB
  --jit         x86-64 JIT compiled mode (default: interpreter)
  --largePages  use large pages
  --softAes     use software AES (default: x86 AES-NI)
  --threads T   use T threads (default: 1)
  --init Q      initialize dataset with Q threads (default: 1)
  --nonces N    run N nonces (default: 1000)
  --seed S      seed for cache initialization (default: 0)

Minor changes in this release:

• improved condition register selection for CBRANCH instruction
• added support for building under MinGW
• fixed a possible memory leak

Legacy 32-bit support

RandomX benchmark
Usage: benchmark.exe [OPTIONS]
Supported options:
  --help        shows this message
  --mine        mining mode: 2080 MiB
  --verify      verification mode: 256 MiB
  --jit         x86-64 JIT compiled mode (default: interpreter)
  --largePages  use large pages
  --softAes     use software AES (default: x86 AES-NI)
  --threads T   use T threads (default: 1)
  --init Q      initialize dataset with Q threads (default: 1)
  --nonces N    run N nonces (default: 1000)
  --seed S      seed for cache initialization (default: 0)

Changes:

• Added support for legacy x87 floating point on 32-bit systems
• Fixed FSCAL instruction causing group F registers to exceed their intended maximum value

Benchmark

Benchmark to test RandomX.

RandomX benchmark
Usage: benchmark [OPTIONS]
Supported options:
  --help        shows this message
  --mine        mining mode: 2 GiB
  --verify      verification mode: 256 MiB
  --jit         x86-64 JIT compiled mode (default: interpreter)
  --largePages  use large pages
  --softAes     use software AES (default: x86 AES-NI)
  --threads T   use T threads (default: 1)
  --init Q      initialize dataset with Q threads (default: 1)
  --nonces N    run N nonces (default: 1000)
  --seed S      seed for cache initialization (default: 0)

Faster verification

Major improvement of verification speed. Mining performance should be mostly unaffected.

Changes:

• Dataset size decreased from 4 GiB to 2 GiB
• Added new verification option --jit for x86 systems (2-3 times faster verification)
• Added random branches in generated programs

Floating point portability fix

Reworked int -> float conversion for group E registers to avoid a small chance of denormal results.

Reworked division instructions

Issue #26 : Division instructions were replaced by a single multiplication by a 64-bit reciprocal. Performance is increased by 1-5% depending on the CPU.