Skip to content

OPENCL_BACKEND.md

Latest commit

 

History

History
85 lines (62 loc) · 4.01 KB

OPENCL_BACKEND.md

File metadata and controls

85 lines (62 loc) · 4.01 KB
Raw

OpenCL Compute Backend Support

GPUBench includes a cross-vendor OpenCL Compute Backend (--backend opencl), providing GPU compute and memory bandwidth benchmarking across AMD, Intel, NVIDIA, and mobile/embedded GPUs without requiring Vulkan 1.4 or ROCm/HIP drivers.

1. Architecture

The multi-backend architecture consists of:

  1. IComputeContext Interface: Backend-agnostic abstract interface for memory buffers, kernel compilation, argument binding, execution dispatches, and device queries.
  2. OpenCLContext Implementation: OpenCL driver interface utilizing dynamic loading (libOpenCL.so.1 on Linux, OpenCL.dll on Windows) to avoid hard link-time package dependencies.
  3. ComputeBackendFactory: Factory managing compile-time detection, runtime availability probes, and fallback backend selection (Vulkan -> OpenCL -> ROCm).
  4. ShaderCache Disk Binary Caching: Automatic caching of compiled OpenCL program binaries (clGetProgramInfo(CL_PROGRAM_BINARIES)) for instant subsequent execution without re-compilation overhead.
  5. Cross-Vendor Portable OpenCL C Kernels: High-throughput SIMD vector implementations written in standard OpenCL C 1.2+ for maximum cross-platform compatibility.

2. Benchmark Feature Matrix & Fallback Policy

GPUBench classifies tests cleanly into fully supported native workloads and transparent API limitations:

Benchmark OpenCL Status Capability / Extension Notes
FP64 Compute ✅ Supported cl_khr_fp64 / cl_amd_fp64 Native double-precision floating-point throughput.
FP32 Compute ✅ Supported Standard OpenCL C 1.2+ fma() Standard single-precision floating-point throughput.
FP16 Compute ✅ Supported cl_khr_fp16 (half / half2) Vector compute throughput. (Matrix mode requires cooperative matrix).
BF16 Compute ⚠️ API Limitation None (cl_khr_bfloat16 not standard) Reported as UNSUPPORTED (API Limitation) without error/crash.
FP8 Compute ⚠️ API Limitation None Reported as UNSUPPORTED (API Limitation).
FP4 / INT4 ⚠️ API / Toolchain Limitation None (Sub-byte 4-bit types not standardized in OpenCL C) Reported as UNSUPPORTED (Toolchain Limitation).
INT8 Compute ✅ Supported Standard vector math (char4) Vector throughput across all OpenCL vendors.
Device Memory Bandwidth ✅ Supported Standard OpenCL Buffers 128, 256, and 1024 threads/group sweep modes (Read, Write, R/W).
Cache Latency ✅ Supported Pointer-chasing buffer traversal Single-workitem pointer chasing across L0/L1/L2/L3 cache levels.
Pixel Fill Rate (ROPs) ⚠️ API Limitation Graphics pipeline required Reported as UNSUPPORTED (API Limitation) (OpenCL lacks rasterizer/ROPs).
Ray Tracing (All 9 Suites) ⚠️ API Limitation VK_KHR_ray_query / hardware BVH acceleration required Reported as UNSUPPORTED (API Limitation) (OpenCL lacks hardware BVH traversal and ray query primitives).

3. Building with OpenCL Support

Prerequisites

Install OpenCL ICD loader development packages:

  • Fedora / RHEL: 
    sudo dnf install ocl-icd-devel opencl-headers
  • Ubuntu / Debian: 
    sudo apt-get install opencl-headers ocl-icd-opencl-dev
  • Arch Linux: 
    sudo pacman -S opencl-headers ocl-icd

Build Steps

mkdir -p build && cd build
cmake ..
make -j16

4. Command-Line Usage

# List all available backends and their runtime availability
./gpubench --list-backends

# List all OpenCL devices detected on the system
./gpubench --list-devices --backend opencl

# Run all benchmarks on OpenCL backend (automatic fallback for unsupported suites)
./gpubench --backend opencl

# Run specific compute and memory benchmarks on OpenCL device 0
./gpubench -k opencl -d 0 -b FP32,FP16,FP64,INT8,Memory

# Export machine-readable results to JSON
./gpubench -k opencl -o results_opencl.json