/gpu-post

Spacemesh proof of space time gpu optimized setup

Primary LanguageCGNU General Public License v3.0GPL-3.0

Warning

This repository is deprecated. Please refer to post-rs for the most recent implementation.

GPU Proof of Spacemesh Time Init (aka Smeshing Setup) Library

Build

Current functionality

A C library implementing the POST API setup method for general-purpose CPUs and for CUDA and Vulkan compute processors.

Runtime System Requirements

Windows 10/11, macOS or Linux. One or more of the following:

  • A GPU and drivers with CUDA support (minimum compute compatibility 5.0, maximum compute compatibility 9), such as a modern Nvidia GPU and Nvidia drivers version R525 or newer.
  • A GPU and drivers with Vulkan 1.3 support such as a modern AMD, Apple M1 processor, and Intel GPUs.
  • A x86-64 cpu such as AMD or Intel CPUs.
  • A ARM 64 bit cpu such as Apple Silicon or Ampere Altra
  • Both discrete and on-board GPUs are supported as long as they support the minimum CUDA or Vulkan runtime version.

We currently provide release binaries and build instructions for Windows, Mac and Ubuntu 22.04 but the library can be built on other Linux distros for usage on these systems.

GPU Memory Requirements

Minimum GPU RAM

  • 16 KiB per CUDA core for CUDA
  • 4 MiB per compute unit for Vulkan

Recommended GPU RAM

  • 2080 MiB

Runtime linux requirements

On Linux platforms with Hybrid Nvidia GPU setup please use Nvidia driver R525 or newer. Older ones are known to have compatibility issues. Non hybrid cards are confirmed to be working with R520 and older versions.

Build System Requirements

All Platforms

  • For building CUDA support: NVIDIA Cuda Toolkit 11, an NVIDIA GPU with CUDA support, and an Nvdia driver version R525 or newer.
    • If building on Linux you should refer to the distribution preferred method installation if available
  • For building Vulkan support: Vulkan SDK 1.3 and a GPU with Vulkan 1.3 runtime support.

Windows

  • Windows 10/11.
  • Microsoft Visual Studio 2022
  • You may also need to install specific versions of the Windows SDK when prompted when attempting to build the library for the first time.

Ubuntu

  • Ubuntu 22.04
  • Cmake, GCC 11+

Other linux distributions

  • Cmake, GCC 11+

macOS

  • Xcode
  • Xcode Command Line Dev Tools
  • Cmake, GCC 11+

macOS Dev Env Setup

  1. Install latest version of Xcode with the command line dev tools.
  2. Download the Vulkan 1.3 sdk installer for macOS from https://vulkan.lunarg.com/sdk/home#mac
  3. Install Vulkan SDK with the Vulkan installer.
  4. Change directory to the folder where the SDK is installed (default $ cd $HOME/VulkanSDK/1.3.xxx) and run the install script with $ sudo ./install_vulkan.py
  5. Add the Vulkan env vars to your .bash_profile file with the root location set to the sdk directory on your hard-drive. For example, if Vulkan sdk 1.2.154 is installed then the env vars should be set like this:
export VULKAN_SDK_VERSION="1.3.xxx"                   # Replace xxx with actual version
export VULKAN_ROOT_LOCATION="$HOME/VulkanSDK/1.3.xxx" # adapt to install location on your machine
export VULKAN_SDK="$VULKAN_ROOT_LOCATION/macOS"
export VK_ICD_FILENAMES="$VULKAN_SDK/share/vulkan/icd.d/MoltenVK_icd.json"
export VK_LAYER_PATH="$VULKAN_SDK/share/vulkan/explicit_layers.d"
export PATH="/usr/local/opt/python/libexec/bin:$VULKAN_SDK/bin:$PATH"
export DYLD_LIBRARY_PATH="$DYLD_LIBRARY_PATH:$VULKAN_SDK/lib/"

Build Configuration

Default build configuration:

Windows and Linux

SPACEMESHCUDA   "Build with CUDA support"   default: ON
SPACEMESHVULKAN "Build with Vulkan support" default: ON

macOS Build Configuration

SPACEMESHCUDA   "Build with CUDA support"   default: OFF
SPACEMESHVULKAN "Build with Vulkan support" default: ON

Building

To build the library with full support for both CUDA and Vulkan on Windows or on Linux use a system with an Nvidia GPU and drivers. Otherwise, turn off CUDA support and build for Vulkan only. Building on macOS only supports Vulkan.

Building on Windows

  1. Open project folder into Visual Studio 2022: File -> Open -> Folder.
  2. Set x64-Release Project Settings.
  3. Build: CMake -> Rebuild All.
  4. Run test: CMake -> Debug from Build Folder -> gpu-setup-test.exe

Ubuntu or macOS

If using VULKAN, make sure to clone the zlib submodule:

git submodule update --init

Configure your build using the default configuration:

cmake -B build

To disable CUDA use:

cmake -B build -DSPACEMESHCUDA=OFF

To disable VULKAN use:

cmake -B build -DSPACEMESHVULKAN=OFF

Build the project:

cmake --build build

Run the tests:

./build/test/gpu-setup-test -t
./build/test/gpu-setup-test -u
./build/test/gpu-setup-test -b

Running the Test App

macOS Configuration

  1. Since the test app is not notarized, you may need to enable it via spctl --add /path/to/gpu-setup-test or by right-click-open it and click open.
  2. Set execute permissions if not already set, e.g., chmod a+x gpu-setup-test
  3. Add the test app's path to the dynamic lib search path, e.g., export DYLD_LIBRARY_PATH=.

Linux Configuration

  1. Set execute permissions if not already set, e.g., chmod a+x gpu-setup-test
  2. Add the test app's path to the dynamic lib search path, e.g., export LD_LIBRARY_PATH=.

Run from the console to print usage:

$ gpu-setup-test
Usage:
--list               or -l                 print available providers
--benchmark          or -b                 run benchmark
--core               or -c                 test the core library use case
--test               or -t                 run basic test
--test-vector-check                        run a CPU test and compare with test-vector
--test-pow           or -tp                test pow computation
--test-leafs-pow     or -tlp               test pow computation while computing leafs
--unit-tests         or -u                 run unit tests
--integration-tests  or -i                 run integration tests
--label-size         or -s <1-256>         set label size [1-256]
--labels-count       or -n <1-32M>         set labels count [up to 32M]
--reference-provider or -r <id>            the result of this provider will be used as a reference [default - CPU]
--print              or -p                 print detailed data comparison report for incorrect results
--pow-diff           or -d <0-256>         count of leading zero bits in target D value [default - 16]
--srand-seed         or -ss <unsigned int> set srand seed value for POW test: 0 - use zero id/seed [default], -1 - use random value
--solution-idx       or -si <unsigned int> set solution index for POW test: index will be compared to be the found solution for Pow [default - unset]

Mixing CUDA and Vulkan

By default, the library does not detect supported Vulkan GPUs if CUDA GPUs are detected. This behavior can be changed using two environment variables:

SPACEMESH_DUAL_ENABLED
 empty or 0 - default behavior
 1 - detect Vulkan GPUs even if CUDA GPUs are detected

SPACEMESH_PROVIDERS_DISABLED
 empty - default behavior
 "cuda" - do not detect CUDA GPUs
 "vulkan" - do not detect Vulkan GPUs

Runtime Providers Recommendations

The library supports multiple compute providers at runtime. For best performance, use the following providers based on your OS and GPU:

OS / GPU Windows Linux macOS
Nvidia CUDA CUDA Vulkan
AMD Vulkan Vulkan Vulkan
Intel Vulkan Vulkan Vulkan
Apple M1 Vulkan Vulkan Vulkan

API

Compute leaves and/or pow solution:

int scryptPositions(
   uint32_t provider_id,      // POST compute provider ID
   const uint8_t *id,         // 32 bytes
   uint64_t start_position,   // e.g. 0
   uint64_t end_position,     // e.g. 49,999
   uint32_t hash_len_bits,    // (1...256) for each hash output, the number of prefix bits (not bytes) to copy into the buffer
   const uint8_t *salt,       // 32 bytes
   uint32_t options,          // compute leafs and/or compute pow
   uint8_t *out,              // memory buffer large enough to include hash_len_bits * number of requested hashes
   uint32_t N,                // scrypt N
   uint32_t R,                // scrypt r
   uint32_t P,                // scrypt p
   uint8_t *D,                // Target D for the POW computation. 256 bits.
   uint64_t *idx_solution,    // index of output where output < D if POW compute was on. MAX_UINT64 otherwise.
   uint64_t *hashes_computed, // The number of hashes computed, should be equal to the number of requested hashes.
   uint64_t *hashes_per_sec   // Performance
);

Supported scrypt parameters

The api currently only supports the following N, P, R scrypt params.

  • Supported N values: 1 - 28835
  • Supported R values: 1
  • Supported P values: 1

Gets the system's GPU capabilities. E.g. CUDA and/or NVIDIA or NONE:

int stats();

Stops all GPU work and don’t fill the passed-in buffer with any more results:

int stop(
 uint32_t ms_timeout   // timeout in milliseconds
);

Returns non-zero if stop in progress:

SPACEMESHAPI int spacemesh_api_stop_inprogress();

Returns POS compute providers info:

SPACEMESHAPI int spacemesh_api_get_providers(
 PostComputeProvider *providers, // out providers info buffer, if NULL - returns count of available providers
 int max_providers// buffer size
);

Linking

  1. Download release artifacts from a github release in this repo for your platform or build the artifacts from source code.
  2. Copy all artifacts to your project resources directory. The files should be included in your app's runtime resources.
  3. Use api.h to link the library from your code.

Testing

Integration test of the basic library use case in a Spacemesh full node to generate proof of space and find a pow solution:

/build/test/.gpu-setup-test -c -n 100663296 -d 20

Community Benchmarks

Disclaimer: these are community submitted benchmarks which haven't been verified. Your milage may vary. The library is also likely to have bugs, is in alpha quality and the gpu-post algorithm is likely to change before the release of the Spacemesh 0.2 testnet.

gpu-setup-test -b -n 2000000
Date Reporter Release Compute Provider OS & CPU Type Driver mh/s
06/21/2021 Obsidian v0.1.20 Geforce RTX 2080ti 11GB @ stock (1350 mhz / 7000 mhz) Windows 10 Pro v20H2, Build 19042.985, Intel i7-6700K @ 4.6ghz (HT enabled: 4c/8t) CUDA NVIDIA 466.11 2.56
06/22/2021 Scerbera v0.1.20 Geforce RTX 2060 SUPER Windows 10 CUDA NVIDIA 466.11 1.7
06/22/2021 Scerbera v0.1.20 AMD Radeon Pro WX 7100 Windows 10 CUDA NVIDIA 466.11 0.88
06/22/2021 Scerbera v0.1.20 RX VEGA 64 - Core Clock 1500 MHz - Memory Clock 960MHz Intel i7-8700K Windows 10 Vulkan Pro 20.Q4 0.9
06/22/2021 Scerbera v0.1.20 WX7100 - Core Clock 1250MHz - Memory Clock 1700 MHz Intel i7-8700K Windows 10 Vulkan Pro 20.Q4 0.87
06/28/2021 cmoetzing v0.1.20 MSI GeForce RTX 2060 VENTUS GP OC - Core Clock 1365MHz - Memory Clock 1750 MHz Ubuntu 20.04 Core i5-11600k CUDA NVIDIA 465.19.01 1.36
06/29/2021 avive v0.1.21 GeForce RTX 3090 Ubuntu 20.04 CUDA Nvidia 460.80 4.97
06/29/2021 avive v0.1.21 GeForce RTX 3080 Ubuntu 20.04 CUDA Nvidia 460.80 4.08
06/30/2021 shanyaa v0.1.21 GeForce RTX 3070 @ 1.9 Ghz core, 6.8 Ghz mem Windows 10 / AMD Ryzen 5800X CUDA Nvidia 466.63 2.7
06/30/2021 shanyaa v0.1.21 GeForce RTX 3070 @ 2 Ghz core, 8.08 Ghz mem Windows 10 / AMD Ryzen 5800X CUDA Nvidia 466.63 3.43
07/01/2021 avive v0.1.21 Nvdia CMP 30HX Ubuntu 20.04.2 LTS CUDA Nvidia 460.80 1.45
07/01/2021 avive v0.1.21 GeForce RTX 2060 Ubuntu 20.04.2 LTS CUDA Nvidia 465.27 1.56
07/01/2021 shanyaa v0.1.21 Intel Iris Xe (integrated graphics) Windows 10 / Intel core i7 1165G7 Vulkan Intel 27.20.100.9565 0.28
07/03/2021 neodied v0.1.21 Radeon 5700XT @ 1333 MHz core, 1824 MHz mem Windows 10 / Intel core i7 9700K Vulkan AMD Radeon Software 21.6.1 1.38
07/03/2021 neodied v0.1.21 Radeon 5700XT @ 2016 MHz core, 1748 MHz mem Windows 10 / Intel core i7 9700K Vulkan AMD Radeon Software 21.6.1 1.87
12/21/2022 lane v0.1.28 Apple M1 (built-in, 8 cores, Metal 3) macOS 13.1 Vulkan N/A 0.15
01/27/2023 lane v0.1.28 Apple M2 Pro (built-in, 16 GPU cores, Metal 3) macOS 13.2 Vulkan N/A 0.56
01/27/2023 nj v0.1.28 Apple M2 Pro (built-in, 19 GPU cores, Metal 3) macOS 13.2 Vulkan N/A 0.57

Prerelease Benchmarks

Scrypt Benchmarks (n=512, r=1, p=1) 1 byte per leaf, batch size leaves per API call.

Date Reporter impl cpu / gpu Host OS notes kh/s mh/s x factor over 1 4ghz cpu native thread x factor over 12 4ghz cpu native threads
11/19/2019 ae go-scrypt mbp + Intel i9 @ 2.9ghz - 1 core OS X go scrypt crypto lib (not scrypt-jane) 7 0.01 1 1
11/19/2019 ae sm-scrypt Ryzen 5 2600x @ 4ghz - 1 core Windows 10 scrypt-jane c code 7 0.01 1 1
11/19/2019 ae sm-scrypt Nvidia Geforce RTX 2070 8GB Windows 10 pre-optimized prototype 1,920 1.92 290 24.17
11/19/2019 ae sm-scrypt AMD Radeon RX 580 Windows 10 pre-optimized prototype 500 0.50 76 6.29
11/19/2019 ar sm-scrypt Nvidia GTX 1060 6G Windows 10 pre-optimized prototype 979 0.98 148 12.32
11/19/2019 ar sm-scrypt AMD Radeon 570 4GB Windows 10 pre-optimized prototype 355 0.36 54 4.47
11/12/2019 ae sm-scrypt AMD Radeon RX 580 Windows 10 optimized prototype 926 0.93 140 11.65
11/12/2019 ae sm-scrypt AMD Radeon RX 580 Ubuntu 18.0.4.3 LTS optimized prototype 893 0.89 135 11.24
11/12/2019 ae sm-scrypt Nvidia Geforce RTX 2070 8GB Ubuntu 19.10 LTS optimized prototype 1,923 1.92 292 24.37
01/22/2020 seagiv sm-scrypt Nvidia GTX 1060 6G Windows 10 vulkan pre-optimized prototype 276
01/22/2020 seagiv sm-scrypt AMD Radeon 570 4GB Windows 10 vulkan pre-optimized prototype 269
01/27/2020 seagiv sm-scrypt Nvidia GTX 1060 6G Windows 10 vulkan optimized prototype 642
01/27/2020 seagiv sm-scrypt AMD Radeon 570 4GB Windows 10 vulkan optimized prototype 966
01/29/2020 seagiv sm-scrypt AMD Radeon Pro 555x 4GB macOS 10.14.6 vulkan optimized prototype 266
01/31/2020 avive sm-scrypt AMD Radeon Pro 560x 4GB macOS 10.14.6 vulkan optimized prototype 406
01/31/2020 avive sm-scrypt Intel(R) UHD Graphics 630 1536MB macOS 10.14.6 vulkan optimized prototype 53
05/06/2020 avive sm-scrypt AMD Radeon RX 580 Windows 10 vulkan optimized prototype 1,074 1.074
09/08/2020 avive sm-scrypt Nvidia Tesla V 100 (16GB) Ubuntu 20.04 NVIDIA-SMI 450.51.06 CUDA Version: 11.0 CUDA optimized prototype 4,166 4.166
09/08/2020 avive sm-scrypt Nvidia Tesla T4 (16GB) Ubuntu 20.04 NVIDIA-SMI 450.51.06 CUDA Version: 11.0 CUDA optimized prototype 1,252 1.252
09/08/2020 avive sm-scrypt Nvidia Tesla P100-PCIE (32GB) Ubuntu 20.04 NVIDIA-SMI 450.51.06 CUDA Version: 11.0 CUDA optimized prototype 2,083 2.083
09/08/2020 avive sm-scrypt Nvidia Tesla P4 (32GB) Ubuntu 20.04 NVIDIA-SMI 450.51.06 CUDA Version: 11.0 CUDA optimized prototype 757 0.75
04/04/2020 avive sm-scrypt Apple M1 MacOS 11.2 vulkan optimized prototype 214 0.214
04/21/2020 avive sm-scrypt Nvidia RTX 2070 Super, 8GB Ubuntu 20.04, Driver 460.73.01 CUDA optimized prototype 2,038 2.038

3rd Party Vulkan and CUDA Benchmarks

The library performance on a GPU depends on the GPU's CUDA and Vulkan performance. The following benchmarks are available from geekbench: