thrust/thrustArchived
Thrust
Thrust es una biblioteca de computación heterogénea y una biblioteca de plantillas de C++ que proporciona una colección de plantillas de alto nivel para ejecutar operaciones de datos en paralelo. Funciona como una biblioteca de algoritmos paralelos diseñada para trabajar en diferentes backends de hardware, incluyendo CPUs multinúcleo y hardware de GPU NVIDIA.
El framework utiliza una implementación de solo cabeceras y una interfaz de políticas de programación genérica para abstraer las diferencias entre los modelos de memoria y ejecución de CPU y GPU. Emplea una abstracción de datos basada en iteradores para proporcionar una interfaz uniforme para acceder a elementos a través de la RAM del host y la VRAM del dispositivo.
La biblioteca cubre capacidades de procesamiento paralelo, incluyendo la clasificación de datos en paralelo y el procesamiento de reducción agregada para calcular valores en grandes conjuntos de datos. Estas operaciones se gestionan a través de una biblioteca de programación paralela CUDA para computación de alto rendimiento en hardware GPU.
Features
- Parallel Algorithms - Provides a comprehensive collection of high-level parallel algorithms for data-parallel operations.
- Device Backends - Provides a specialized CUDA-based backend to offload data-parallel computations to NVIDIA GPUs.
- C++ Parallel Programming - Enables high-performance C++ programming for data-parallel operations across heterogeneous hardware.
- C++ Parallelism Libraries - Functions as a specialized C++ parallelism library offering high-level templates for data-parallel operations.
- Generic Programming - Utilizes generic programming with template parameters to decouple execution strategies from algorithm implementations.
- Iterator-Based Abstractions - Employs iterator-based abstractions to provide a uniform interface for accessing host and device memory.
- Template Libraries - Implemented as a C++ template library to ensure generic behavior across different hardware backends.
- Heterogeneous Computing Libraries - Acts as a heterogeneous computing library that abstracts the differences between CPU and GPU execution models.
- GPU-Accelerated Processing - Facilitates GPU-accelerated processing by moving large datasets to parallel hardware for fast computation.
- Parallel Sorting - Implements high-performance parallel data sorting by leveraging GPU and multicore CPU hardware.
- Header-Only Libraries - Ships as a header-only template library to enable aggressive compiler inlining and optimization.
- Parallel Reductions - Provides parallel reduction operations to calculate sums, minimums, or maximums across distributed cores.
- Compile-Time Type Dispatch - Uses C++ templates for compile-time type dispatch to eliminate runtime overhead in performance-critical paths.
- CUDA Libraries - Provides a robust set of CUDA-based algorithms and data structures for high-performance GPU computing.
- Aggregate Reductions - Provides parallel aggregate reduction processing for calculating values across large datasets.
- Asynchronous Execution - Implements non-blocking execution models to overlap data transfers with kernel execution.
- Procesamiento paralelo - Biblioteca de programación paralela en C++ para sistemas heterogéneos.
- GPU Acceleration - Parallel programming library for GPU acceleration.
- Parallel and High-Performance Computing - C++ parallel programming library for heterogeneous systems.
