Implementation:Alibaba MNN OpenCL CL2 HPP Header

Knowledge Sources	Alibaba_MNN Khronos OpenCL
Domains	GPU_Computing, OpenCL
Last Updated	2026-02-10 12:00 GMT

Overview

Vendored modern Khronos OpenCL C++ bindings header (cl2.hpp) providing updated C++11/14 RAII wrappers for OpenCL 2.0+ features including SVM, pipes, and improved memory model for MNN's GPU backend.

Description

cl2.hpp is the modern replacement for the deprecated cl.hpp, redesigned to leverage C++11 and C++14 features such as move semantics, std::array, and variadic templates. It provides the same RAII wrapper classes (cl::Context, cl::CommandQueue, cl::Buffer, cl::Kernel, cl::Program) as its predecessor but adds support for OpenCL 2.0+ constructs including Shared Virtual Memory (SVM) allocators (cl::SVMAllocator, cl::SVMTraitReadWrite, cl::SVMTraitFine, cl::SVMTraitAtomic), pipe objects, and the unified memory model. MNN vendors this header to enable its OpenCL backend to take advantage of modern GPU capabilities when available while maintaining backward compatibility through compile-time version detection.

Usage

This header is included internally by MNN's OpenCL backend. It is not directly imported by end users.

Code Reference

Source Location

• Repository: Alibaba_MNN
• File: 3rd_party/OpenCLHeaders/CL/cl2.hpp
• Lines: 1-9685

Signature

namespace cl {

class Device : public detail::Wrapper<cl_device_id> {
public:
    Device() {}
    Device(cl_device_id device) { ... }
    static Device getDefault(cl_int *errResult = NULL);
};

class Context : public detail::Wrapper<cl_context> {
public:
    Context(const vector<Device>& devices, ...);
    static Context getDefault(cl_int *errResult = NULL);
};

class Buffer : public Memory {
public:
    Buffer(const Context& context, cl_mem_flags flags, size_t size, ...);
    Buffer(cl_mem_flags flags, size_t size, ...);  // uses default context
};

// OpenCL 2.0+ SVM support
template<class SVMTrait>
class SVMAllocator {
public:
    typedef typename SVMTrait::value_type value_type;
    pointer allocate(size_type size);
    void deallocate(pointer p, size_type size);
};

class SVMTraitReadWrite { ... };
class SVMTraitReadOnly  { ... };
class SVMTraitFine      { ... };
class SVMTraitAtomic    { ... };

} // namespace cl

Import

#include "3rd_party/OpenCLHeaders/CL/cl2.hpp"

I/O Contract

Inputs

Name	Type	Required	Description
N/A	N/A	N/A	Header-only; no runtime inputs

Outputs

Name	Type	Description
API declarations	C++ classes	Modern C++11/14 RAII wrapper classes for OpenCL objects, including SVM allocators and OpenCL 2.0+ constructs

Usage Examples

// MNN internal usage (not user-facing)
#define CL_HPP_TARGET_OPENCL_VERSION 200
#include "CL/cl2.hpp"

// Use default platform and device (C++11 style)
cl::Device device = cl::Device::getDefault();
cl::Context context = cl::Context::getDefault();
cl::CommandQueue queue = cl::CommandQueue::getDefault();

// Create buffer using default context
cl::Buffer buffer(CL_MEM_READ_WRITE, 1024);

// Build program and create kernel
cl::Program program(context, sourceCode);
program.build("-cl-std=CL2.0");
cl::Kernel kernel(program, "my_kernel");

// SVM allocation (OpenCL 2.0+)
cl::SVMAllocator<int, cl::SVMTraitReadWrite<>> svmAlloc(context);
int* svmPtr = svmAlloc.allocate(256);

// Execute kernel
queue.enqueueNDRangeKernel(kernel, cl::NullRange, cl::NDRange(256));
queue.finish();

Related Pages

• Environment:Alibaba_MNN_GPU_OpenCL_Environment
• Alibaba_MNN_OpenCL_CL_Header - Underlying C API that this header wraps
• Alibaba_MNN_OpenCL_CL_HPP_Header - Deprecated predecessor C++ bindings (cl.hpp)
• Alibaba_MNN_OpenCL_CL_Platform_Header - Platform type definitions