Enum ComputeBackendType

Namespace

DotCompute.Abstractions.Compute.Enums

Assembly

DotCompute.Abstractions.dll

Defines the available compute backend types for kernel execution. Each backend represents a different compute architecture and programming model.

public enum ComputeBackendType

Fields

CPU = 0

CPU backend using SIMD (Single Instruction, Multiple Data) instructions.

This backend leverages CPU vector instructions (SSE, AVX, NEON) for parallel computation. It provides good performance for algorithms that can be vectorized and is available on all systems. Memory access is through system RAM with cache hierarchy optimization.

CUDA = 1

CUDA backend for NVIDIA GPUs using the CUDA Compute Capability architecture.

Requires NVIDIA graphics hardware and CUDA drivers. Provides excellent performance for massively parallel workloads with thousands of threads. Features include shared memory, texture memory, and advanced memory coalescing optimizations. Only available on systems with compatible NVIDIA hardware.

DirectCompute = 5

DirectCompute backend for Windows platforms using DirectX compute shaders.

Microsoft's compute API integrated with DirectX graphics pipeline. Provides good integration with Windows graphics stack and supports interoperability with DirectX graphics resources. Features include structured buffers, unordered access views, and integration with Windows display drivers.

Metal = 3

Metal backend for Apple devices including iOS, macOS, and Apple Silicon.

Apple's proprietary compute API optimized for Apple hardware ecosystem. Provides deep integration with Apple's unified memory architecture and GPU designs. Features include Metal Performance Shaders integration and optimized memory management for Apple Silicon.

OpenCL = 2

OpenCL backend for cross-platform heterogeneous computing.

Industry-standard API supporting CPUs, GPUs, and other accelerators from multiple vendors (NVIDIA, AMD, Intel, ARM). Provides portable compute kernels that can run across different hardware platforms with consistent behavior and performance characteristics.

ROCm = 6

ROCm backend for AMD GPUs using the ROCm platform.

AMD's open-source platform for GPU computing with HIP (Heterogeneous-compute Interface for Portability). Provides excellent performance for AMD Radeon and Instinct GPUs with features including shared virtual memory, peer-to-peer transfers, and support for large memory workloads. Compatible with both consumer and enterprise AMD GPU hardware.

Vulkan = 4

Vulkan Compute backend using the Vulkan API's compute shader capabilities.

Modern low-level compute API providing explicit control over GPU resources and memory management. Offers excellent performance through reduced driver overhead and fine-grained resource control. Supports advanced features like subgroups, device-local memory, and multi-queue execution.

Remarks

The choice of backend affects performance characteristics, memory access patterns, and API compatibility. Some backends may not be available on all systems depending on hardware capabilities and driver availability.