opencl

name

opencl

description

OpenCL SDK (Khronos Group) for cross-platform GPU/CPU parallel computing in C and C++. Use when writing OpenCL kernels, managing devices/contexts/queues, allocating and transferring buffers or images, building and executing programs, or using the C++ wrapper (opencl.hpp / cl::CommandQueue, cl::Buffer, cl::KernelFunctor). Covers OpenCL C API, C++ bindings, and SDK utility libraries (OpenCLUtils, OpenCLSDK).

OpenCL SDK

Version: v2026.05.29 (Khronos Group OpenCL-SDK) Language: C (OpenCL 1.0–3.1) / C++ (opencl.hpp wrapper) License: Apache-2.0 Repo: https://github.com/KhronosGroup/OpenCL-SDK

Overview

OpenCL (Open Computing Language) is a framework for parallel programming across heterogeneous platforms — GPUs, CPUs, FPGAs, and DSPs — from a single API. The SDK bundles:

• OpenCL-Headers — C headers (<CL/cl.h>, <CL/cl_ext.h>)
• OpenCL-CLHPP — C++ wrapper (<CL/opencl.hpp>)
• OpenCL-ICD-Loader — runtime dispatch to installed platform drivers
• OpenCLUtils / OpenCLSDK — utility libraries (<CL/Utils/>, <CL/SDK/>)

Quick Start (C)

Kernel file saxpy.cl:

__kernel void saxpy(float a, __global float *x, __global float *y) {
    int i = get_global_id(0);
    y[i] = fma(a, x[i], y[i]);
}

Host:

#include <CL/cl.h>
cl_platform_id plat; cl_device_id dev;
clGetPlatformIDs(1, &plat, NULL);
clGetDeviceIDs(plat, CL_DEVICE_TYPE_DEFAULT, 1, &dev, NULL);
cl_context ctx = clCreateContext(NULL, 1, &dev, NULL, NULL, &err);
cl_command_queue q = clCreateCommandQueueWithProperties(ctx, dev, NULL, &err);
// ... load source, clCreateProgramWithSource, clBuildProgram,
//     clCreateKernel, clSetKernelArg, clEnqueueNDRangeKernel,
//     clEnqueueReadBuffer, clReleaseXxx ...

Quick Start (C++)

#define CL_HPP_ENABLE_EXCEPTIONS
#define CL_HPP_TARGET_OPENCL_VERSION 300   // default is now 310 (OpenCL 3.1); minimum 200
#include <CL/opencl.hpp>

cl::Context ctx{CL_DEVICE_TYPE_DEFAULT};
cl::Device  dev = ctx.getInfo<CL_CONTEXT_DEVICES>()[0];
cl::CommandQueue queue{ctx, dev};
cl::Program prog{ctx, source_string};
prog.build(dev);
auto saxpy = cl::KernelFunctor<cl_float, cl::Buffer, cl::Buffer>(prog, "saxpy");
saxpy(cl::EnqueueArgs{queue, cl::NDRange{N}}, a, buf_x, buf_y);

Core Concepts

• Work-item — one parallel execution unit; maps to one GPU thread
• Work-group — block of work-items sharing local memory and barriers
• NDRange — N-dimensional index space (up to 3D); defines total parallelism
• Context — owns devices, memory objects, programs, and queues
• Command Queue — ordered or OOO stream of commands to one device
• Memory object — buffer (linear) or image (typed, sampled); device-side
• Kernel — a __kernel function compiled from OpenCL C source or SPIR-V
• Event — synchronization token returned by every enqueue command
• Address spaces — __global (buffers), __local (shared), __constant (read-only), __private (per-item)

API Reference

Domain	Reference File	Key Functions / Types
Platform & Device	references/api-platform-device.md	clGetPlatformIDs, clGetDeviceIDs, clGetDeviceInfo, clCreateSubDevices, UUID/LUID & SPIR-V queries (3.1), timer APIs
Context & Queue	references/api-context-queue.md	clCreateContext, clCreateCommandQueueWithProperties, clFlush, clFinish, destructor callbacks
Memory Objects	references/api-memory.md	clCreateBuffer, clCreateImage, enqueue read/write/copy/fill, map/unmap, pipes, samplers, SVM
Programs & Kernels	references/api-program-kernel.md	clCreateProgramWithSource, clBuildProgram, clCompileProgram, clLinkProgram, clCreateKernel, clGetKernelSuggestedLocalWorkSize (3.1), sub-group queries
Execution & Events	references/api-execution.md	clEnqueueNDRangeKernel, clWaitForEvents, clSetEventCallback, profiling, extension access
C++ Wrapper	references/api-cpp-wrapper.md	cl::Context, cl::Buffer, cl::Pipe, cl::Sampler, cl::KernelFunctor, cl::SVMAllocator, exceptions
Workflows	references/workflows.md	Quick-start, vector add, image blur, async events, binary caching, error handling

Common Workflows

See references/workflows.md for complete, runnable examples:

• Vector add (C) — minimal host+kernel from scratch
• SAXPY (C++) — KernelFunctor pattern with RAII
• Device enumeration — iterate all platforms and devices
• Image blur — 2D image creation, read_imageui / write_imageui
• Async events — non-blocking enqueue chains
• Binary caching — save/restore compiled programs
• Error handling — C goto pattern vs. C++ exceptions

SDK Utility Libraries

Include <CL/Utils/Utils.h> (C) or <CL/Utils/Utils.hpp> (C++) and link OpenCLUtils / OpenCLUtilsCpp.

Header	API
<CL/Utils/Context.h>	cl_util_get_device, cl_util_get_context, cl_util_print_device_info
<CL/Utils/File.h>	cl_util_read_text_file, cl_util_read_exe_relative_text_file, cl_util_write_binaries
<CL/Utils/Error.h>	OCLERROR_RET, OCLERROR_PAR, MEM_CHECK macros, cl_util_print_error
<CL/Utils/Event.h>	cl_util_get_event_duration
<CL/Utils/Device.hpp>	cl::util::supports_extension, cl::util::supports_feature

SDK Library (samples only, not installed): <CL/SDK/CLI.h>, <CL/SDK/Random.h>, <CL/SDK/Image.h>.

Key Considerations

Release everything: Every clCreate* call must be paired with the corresponding clRelease*. Leak buffers or kernels and you exhaust device memory silently.

Blocking vs. non-blocking transfers: clEnqueueReadBuffer(..., CL_TRUE, ...) blocks the CPU. Use CL_FALSE + events for overlap. Always clFlush before blocking on an event from another thread.

Local work-group size: Must evenly divide global work size in each dimension. Query CL_KERNEL_WORK_GROUP_SIZE for the max; CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE for optimal alignment. Passing NULL lets the runtime choose (portable, not always optimal).

Build log on failure: clBuildProgram returns CL_BUILD_PROGRAM_FAILURE — always query CL_PROGRAM_BUILD_LOG to get the compiler error message. The SDK's cl_util_build_program does this automatically.

Image format validation: Not all cl_image_format combinations are supported on every device. Call clGetSupportedImageFormats before creating images.

Event callbacks must not block: Callbacks registered via clSetEventCallback are invoked from a runtime thread. Never call clFinish or clWaitForEvents inside a callback.

C++ exceptions: Enable with #define CL_HPP_ENABLE_EXCEPTIONS before including <CL/opencl.hpp>. Without it, check cl_int error parameters manually.

OpenCL version targeting: Set CL_HPP_TARGET_OPENCL_VERSION (e.g., 310, 300, 200, 120) to control which API surface is available in the C++ wrapper. As of this SDK the default is 310 (OpenCL 3.1) and the minimum is 200; the C headers default CL_TARGET_OPENCL_VERSION to 310 too. OpenCL 1.x deprecated clCreateCommandQueue; use clCreateCommandQueueWithProperties for 2.0+.

OpenCL 3.1 additions: clGetKernelSuggestedLocalWorkSize (query the runtime's optimal local size for an NDRange) is now core, along with device queries CL_DEVICE_UUID / CL_DRIVER_UUID, CL_DEVICE_LUID / CL_DEVICE_LUID_VALID / CL_DEVICE_NODE_MASK, integer dot-product capabilities, and SPIR-V extension/capability queries — all previously KHR extensions. In C++, use queue.getKernelSuggestedLocalWorkSize(...) and #define CL_HPP_CUSTOM_EXCEPTION_TYPE to plug your own exception class. CL_DEVICE_MAX_WORK_ITEM_SIZES is now spelled CL_DEVICE_MAX_WORK_GROUP_SIZES (same value; old name deprecated).

SVM requires OpenCL 2.0+: Shared Virtual Memory (clSVMAlloc) requires device support for CL_DEVICE_SVM_CAPABILITIES. Check before use.