Ecosystem of libraries and tools for writing and executing fast GPU code fully in Rust.

Files	Last commit	Last update
.cargo	Feat: add axpy to blastoff and improve docs	4 years ago
.github	Feat: Tons of cust changes and start of OptiX (hardware rt) work * wip * bootstrap enough optix to get ex02 working * Add example 03 Generate an animated pattern in the raygen and display it in a window using glfw * add logging callback * remove ustr * Manually create OptixShaderBindingTable field-by-field instead of transmute * Switch Module and Pipeline methods to their structs Instead of having them on DeviceContext * Switch Module, Pipeline, ProgramGroup methods to their structs Instead of having them on DeviceContext * Refactor: remove dead imports * derive DeviceCopy * typo * Better error message * Move destroy to Drop impl * typo * rename OptixContext to DeviceContext * Make launch params variable name optional * Remove Clone from Module and ProgramGroup * Make log callback safe User catch_unwind to guard against panic in C. Remove note about lifetime of closure since it's 'static anyway. Have set_log_callback return a Result instead of panicking on error. * add wip glam support * dont panic in drop * Rework DevicePointer on top of CUdeviceptr This switches out T for CUdeviceptr in DevicePointer. This has the knock-on effect of removing a lot of "pretend we're a CPU pointer" stuff from downstream types like DeviceSlice. wip * bootstrap enough optix to get ex02 working * Add example 03 Generate an animated pattern in the raygen and display it in a window using glfw * add logging callback * remove ustr * Manually create OptixShaderBindingTable field-by-field instead of transmute * Switch Module and Pipeline methods to their structs Instead of having them on DeviceContext * Switch Module, Pipeline, ProgramGroup methods to their structs Instead of having them on DeviceContext * Refactor: remove dead imports * derive DeviceCopy * typo * Better error message * Move destroy to Drop impl * typo * rename OptixContext to DeviceContext * Make launch params variable name optional * Remove Clone from Module and ProgramGroup * Make log callback safe User catch_unwind to guard against panic in C. Remove note about lifetime of closure since it's 'static anyway. Have set_log_callback return a Result instead of panicking on error. * add wip glam support * dont panic in drop * wip accel support * Add accel wip Enough acceleration structure stuff to get example 04 running, and rebasing on top of deviceptr branch * Rework acceleration structure stuff Provide simple internally allocating API for Accel, but also allow creating one from raw parts to let user handle memory allocation. Original API kept as free functions and marked unsafe. Implement all build input types. Add mint support to cust and optix. * add lifetime bound on Instance to referenced Accel * Have DeviceCopy impl for lifetime markers use null type * Add unsaafe from_handle ctor * Add update for DynamicAccel * Hash build inputs to ensure update is sound * Add relocation info * Add remaning DeviceContext methods * Correct docstrings * Add doc comments * Add a prelude * Own the geometry flags array * Add prelude * own the geometry flags array and add support for pre_transform * Fill out context and add some module docs * Add some module docs * Update to latest library changes * Add more docs * Remove mut requirement for getting pointer * Add a simple memcpy_htod wrapper * Add back pointer offset methods * Big structure reorg and documentation push - Reorganized the module structure to something less fragmented. Modules are longer but more cohesive. - Integrated the optix programming guide inline into the module documentation. You need to build the docs with: RUSTDOCFLAGS="--html-in-header katex-header.html" cargo doc --no-deps To see the equations (this is done automatically on docs.rs) * Wrap SBT properly * Rename transform types * Simplify AccelBuildOptions creation Just take a build flags and move everything else to builders * Hide programming guide in details tags * Adapt to latest changes * Fix toolchain version * Fix name of DeviceContext * first optix rust test * Set ALLOW_COMPACTION in build options * Use find_cuda_helper to get cuda path * Handle differering enum representation on windows and linux * Add DeviceVariable * Add DeviceMemory trait Abstracts over different device storage representations * Add mem_get_info * Add external memory * Add a few more types to prelude * Add more types * Rework on top of new DeviceVariable * first optix rust test * tweak build * update to latest optix changes * Split DeviceCopy into cust_core * update to latest optix changes * trying to get print working * tweak test kernel * stop llvm optimizing out LaunchParams * Chore: update cargo.toml dep versions * Feat: second pass for fixing conflicts * Feat: delete as_ptr and as_mut_ptr on DeviceSlice * Revert "Feat: delete as_ptr and as_mut_ptr on DeviceSlice" This reverts commit e858fdcaf37c8beb0df74cc84dfb79300eff7eb3. * Feat: experiment with deleting as_ptr and as_mut_ptr * Fix issues and warnings * Chore: run formatting * Chore: exclude examples from building in CI * Feat: update changelog with changes, misc changes before merge Co-authored-by: rdambrosio <rdambrosio016@gmail.com>	4 years ago
crates	Fix: add Debian's installation to default paths (#72) When you install `nvidia-cuda-toolkit` on Debian, it installs the CUDA libraries into `/usr/lib/cuda`. Add this path to the default search paths.	4 years ago
examples	Fix: fix optix denoiser example	4 years ago
guide	Fix guide book to mention `rust-toolchain` and small build fix (#76)	4 years ago
scripts	Chore: export env through GITHUB_ENV	4 years ago
xtask	Initial commit	4 years ago
.gitignore	Initial commit	4 years ago
CODEOWNERS	Chore: Fix syntax	4 years ago
Cargo.toml	Feat: rust OptiX hardware raytracing MVP	4 years ago
Dockerfile	fix error - context creation failed (#31)	4 years ago
LICENSE-APACHE	Chore: add licenses	4 years ago
LICENSE-MIT	Chore: add licenses	4 years ago
README.md	Link to new rust-gpu home in README.md, tweak language (#125) Embark Studios transferred rust-gpu to the community.	1 year ago
katex-header.html	Feat: add axpy to blastoff and improve docs	4 years ago
rust-toolchain	Feat: update nightly to 2021-12-04 in anticipation for 0.2	4 years ago
rustfmt.toml	Add empty rustfmt	4 years ago

Auto Translate

Rust CUDA 项目

一个用于编写和执行极快 GPU 代码的生态系统，完全在 Rust 语言中实现

指南 | 入门 | 特性

⚠️ 该项目仍处于早期开发阶段，可能会遇到错误、安全问题以及不工作的情况 ⚠️

Important

这个项目已经不再沉寂，并且正在重启。请贡献力量！

目标

Rust CUDA 项目的目标是使 Rust 成为使用 CUDA Toolkit 进行极快 GPU 计算的一流语言。它提供了将 Rust 编译为极快 PTX 代码的工具，以及与现有 CUDA 库一起使用的库。

背景

在历史上，通用高性能 GPU 计算主要使用 CUDA Toolkit 进行。CUDA Toolkit 主要提供了一种使用 Fortran/C/C++ 代码进行 GPU 计算的方法，与 CPU 代码配合使用单一源代码。它还提供了许多库、工具、论坛和文档来补充单一源 CPU/GPU 代码。

CUDA 是专属于 NVIDIA 的工具包。已经提出了许多跨平台 GPU 计算的工具，例如 OpenCL、Vulkan Computing 和 HIP。然而，CUDA 仍然是最常用的工具包，因此迫切需要使 Rust 成为 CUDA Toolkit 的可行选项。

但是，Rust 与 CUDA 一起使用在历史上一直是一条非常坎坷的道路。直到现在，唯一可行的选择是使用 LLVM PTX 后端，但是，LLVM PTX 后端并不总是有效，并且会为许多常见的 Rust 操作生成无效的 PTX 代码，近年来，随着诸如 rust-gpu（用于 Rust -> SPIR-V）等项目的发展，已经一次又一次地证明，Rust 在 GPU 上需要一个专门的解决方案。

我们的希望是，通过这个项目，我们可以推动 Rust GPU 计算行业的发展，并使 Rust 成为这类任务的优秀语言。Rust 提供了许多好处，例如每个内核的 __restrict__ 性能优势，出色的模块/包管理系统，使用 unsafe 限定 CPU/GPU 代码的不安全区域，对底层 CUDA 库的高级封装等。

结构

Rust CUDA 项目的范围非常广泛，它涵盖了 CUDA 生态系统的方方面面，包括用 Rust 使其可用的库和工具。因此，项目包含了许多针对 CUDA 生态系统各个角落的包。

当前的库阵容如下：

rustc_codegen_nvvm 是一个 rustc 后端，目标为 NVVM IR（LLVM IR 的一个子集），用于 libnvvm 库。
- 生成高度优化的 PTX 代码，可以被 CUDA 驱动 API 加载到 GPU 上执行。
- 在不久的将来，它将仅支持 CUDA，但未来可能用于目标 amdgpu。
cuda_std 用于 GPU 边的函数和实用工具，例如线程索引查询、内存分配、warp 内联函数等。
- 不是一个底层库，提供了许多实用函数，以简化编写更清洁、更可靠的 GPU 内核。
- 与 rustc_codegen_nvvm 密切相关，后者通过它内部暴露 GPU 功能。
cudnn 提供了用于深度神经网络的 GPU 加速原语集合。
cust 用于 CPU 边的 CUDA 功能，例如启动 GPU 内核、GPU 内存分配、设备查询等。
- 高级封装，具有 RAII 和 Rust Results 等特性，简化了 GPU 接口的管理。
- CUDA 驱动 API 的高级封装，是 C++ 中更常见的 CUDA 运行时 API 的较低级别版本。
- 比较 CUDA 运行时 API 提供了更细粒度的控制，例如内核并发和模块加载。
gpu_rand 用于 GPU 友好的随机数生成，目前仅实现了 rand_xoshiro 的 xoroshiro 随机数生成器。
optix 用于 CPU 边的硬件光线追踪和去噪，使用 CUDA OptiX 库。