OpenCL-Wrapper

OpenCL is the most powerful programming language ever created. Yet the OpenCL C++ bindings are cumbersome and the code overhead prevents many people from getting started. I created this lightweight OpenCL-Wrapper to greatly simplify OpenCL software development with C++ while keeping functionality and performance.

Works in Windows, Linux and Android with C++17.

Use-case example: FluidX3D builds entirely on top of this OpenCL-Wrapper.

Getting started:

▸Download+unzip the source code or git clone https://github.com/ProjectPhysX/OpenCL-Wrapper.git

Key simplifications:

1. select a Device with 1 line
   • automatically select fastest device / device with most memory / device with specified ID from a list of all devices
   • easily get device information (performance in TFLOPs/s, amount of memory and cache, FP64/FP16 capabilities, etc.)
   • automatic OpenCL C code compilation when creating the Device object
     • automatically enable FP64/FP16 capabilities in OpenCL C code
     • automatically print log to console if there are compile errors
     • easy option to generate PTX assembly for Nvidia GPUs and save that in a .ptx file
   • contains all device-specific workarounds/patches to make OpenCL fully cross-compatible
     • enable basic FP16 support on Nvidia Pascal and newer GPUs with driver 520 or newer
     • enable >4GB single buffer VRAM allocations on Intel Arc GPUs
     • fix for wrong VRAM capacity reporting on Intel Arc GPUs
     • fix for maximum buffer allocation size limit for AMD GPUs
     • fix for maximum buffer allocation size limit in Intel CPU Runtime for OpenCL
     • fix for terrible fma performance on ARM GPUs
2. create a Memory object with 1 line
   • one object for both host and device memory
   • easy host <-> device memory transfer (also for 1D/2D/3D grid domains)
   • easy handling of multi-dimensional vectors
   • can also be used to only allocate memory on host or only allocate memory on device
   • automatically tracks total global memory usage of device when allocating/deleting memory
3. create a Kernel with 1 line
   • Memory objects and constants are linked to OpenCL C kernel parameters during Kernel creation
   • a list of Memory objects and constants can be added to Kernel parameters in one line (add_parameters(...))
   • Kernel parameters can be edited (set_parameters(...))
   • easy Kernel execution: kernel.run();
   • Kernel function calls can be daisy chained, for example: kernel.set_parameters(3u, time).run();
   • failsafe: you'll get an error message if kernel parameters mismatch between C++ and OpenCL code
4. OpenCL C code is embedded into C++
   • syntax highlighting in the code editor is retained
   • notes / peculiarities of this workaround:
     • the #define R(...) string(" "#__VA_ARGS__" ") stringification macro converts its arguments to string literals; '\n' is converted to ' ' in the process
     • these string literals cannot be arbitrarily long, so interrupt them periodically with )+R(
     • to use unbalanced round brackets '('/')', exit the R(...) macro and insert a string literal manually: )+"void function("+R( and )+") {"+R(
     • to use preprocessor switch macros, exit the R(...) macro and insert a string literal manually: )+"#define TEST"+R( and )+"#endif"+R( // TEST
     • preprocessor replacement macros (for example #define VARIABLE 42) don't work; hand these to the Device constructor directly instead

No need to:

• have code overhead for selecting a platform/device, passing the OpenCL C code, etc.
• keep track of length and data type for buffers
• have duplicate code for host and device buffers
• keep track of total global memory usage
• keep track of global/local range for kernels
• bother with Queue, Context, Source, Program
• load a .cl file at runtime
• bother with device-specific workarounds/patches

Example (OpenCL vector addition)

main.cpp

#include "opencl.hpp"

int main() {
    Device device(select_device_with_most_flops()); // compile OpenCL C code for the fastest available device

    const uint N = 1024u; // size of vectors
    Memory<float> A(device, N); // allocate memory on both host and device
    Memory<float> B(device, N);
    Memory<float> C(device, N);

    Kernel add_kernel(device, N, "add_kernel", A, B, C); // kernel that runs on the device

    for(uint n=0u; n<N; n++) {
        A[n] = 3.0f; // initialize memory
        B[n] = 2.0f;
        C[n] = 1.0f;
    }

    print_info("Value before kernel execution: C[0] = "+to_string(C[0]));

    A.write_to_device(); // copy data from host memory to device memory
    B.write_to_device();
    add_kernel.run(); // run add_kernel on the device
    C.read_from_device(); // copy data from device memory to host memory

    print_info("Value after kernel execution: C[0] = "+to_string(C[0]));

    wait();
    return 0;
}

kernel.cpp

#include "kernel.hpp" // note: unbalanced round brackets () are not allowed and string literals can't be arbitrarily long, so periodically interrupt with )+R(
string opencl_c_container() { return R( // ########################## begin of OpenCL C code ####################################################################

kernel void add_kernel(global float* A, global float* B, global float* C) { // equivalent to "for(uint n=0u; n<N; n++) {", but executed in parallel
    const uint n = get_global_id(0);
    C[n] = A[n]+B[n];
}

);} // ############################################################### end of OpenCL C code #####################################################################

For comparison, the very same OpenCL vector addition example looks like this when directly using the OpenCL C++ bindings:

#include <CL/cl.hpp>
#include "utilities.hpp"

#define WORKGROUP_SIZE 64

int main() {

    // 1. select device

    vector<cl::Device> cl_devices; // get all devices of all platforms
    {
        vector<cl::Platform> cl_platforms; // get all platforms (drivers)
        cl::Platform::get(&cl_platforms);
        for(uint i=0u; i<(uint)cl_platforms.size(); i++) {
            vector<cl::Device> cl_devices_available;
            cl_platforms[i].getDevices(CL_DEVICE_TYPE_ALL, &cl_devices_available);
            for(uint j=0u; j<(uint)cl_devices_available.size(); j++) {
                cl_devices.push_back(cl_devices_available[j]);
            }
        }
    }
    cl::Device cl_device; // select fastest available device
    {
        float best_value = 0.0f;
        uint best_i = 0u; // index of fastest device
        for(uint i=0u; i<(uint)cl_devices.size(); i++) { // find device with highest (estimated) floating point performance
            const string name = trim(cl_devices[i].getInfo<CL_DEVICE_NAME>()); // device name
            const string vendor = trim(cl_devices[i].getInfo<CL_DEVICE_VENDOR>()); // device vendor
            const uint compute_units = (uint)cl_devices[i].getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>(); // compute units (CUs) can contain multiple cores depending on the microarchitecture
            const uint clock_frequency = (uint)cl_devices[i].getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>(); // in MHz
            const bool is_gpu = cl_devices[i].getInfo<CL_DEVICE_TYPE>()==CL_DEVICE_TYPE_GPU;
            const uint ipc = is_gpu?2u:32u; // IPC (instructions per cycle) is 2 for GPUs and 32 for most modern CPUs
            const bool nvidia_192_cores_per_cu = contains_any(to_lower(name), {"gt 6", "gt 7", "gtx 6", "gtx 7", "quadro k", "tesla k"}) || (clock_frequency<1000u&&contains(to_lower(name), "titan")); // identify Kepler GPUs
            const bool nvidia_64_cores_per_cu = contains_any(to_lower(name), {"p100", "v100", "a100", "a30", " 16", " 20", "titan v", "titan rtx", "quadro t", "tesla t", "quadro rtx"}) && !contains(to_lower(name), "rtx a"); // identify P100, Volta, Turing, A100, A30
            const bool amd_128_cores_per_dualcu = contains(to_lower(name), "gfx10"); // identify RDNA/RDNA2 GPUs where dual CUs are reported
            const bool amd_256_cores_per_dualcu = contains(to_lower(name), "gfx11"); // identify RDNA3 GPUs where dual CUs are reported
            const bool intel_16_cores_per_cu = contains(to_lower(name), "gpu max"); // identify PVC GPUs
            const float nvidia = (float)(contains(to_lower(vendor), "nvidia"))*(nvidia_64_cores_per_cu?64.0f:nvidia_192_cores_per_cu?192.0f:128.0f); // Nvidia GPUs have 192 cores/CU (Kepler), 128 cores/CU (Maxwell, Pascal, Ampere, Hopper, Ada) or 64 cores/CU (P100, Volta, Turing, A100, A30)
            const float amd = (float)(contains_any(to_lower(vendor), {"amd", "advanced"}))*(is_gpu?(amd_256_cores_per_dualcu?256.0f:amd_128_cores_per_dualcu?128.0f:64.0f):0.5f); // AMD GPUs have 64 cores/CU (GCN, CDNA), 128 cores/dualCU (RDNA, RDNA2) or 256 cores/dualCU (RDNA3), AMD CPUs (with SMT) have 1/2 core/CU
            const float intel = (float)(contains(to_lower(vendor), "intel"))*(is_gpu?(intel_16_cores_per_cu?16.0f:8.0f):0.5f); // Intel GPUs have 16 cores/CU (PVC) or 8 cores/CU (integrated/Arc), Intel CPUs (with HT) have 1/2 core/CU
            const float apple = (float)(contains(to_lower(vendor), "apple"))*(128.0f); // Apple ARM GPUs usually have 128 cores/CU
            const float arm = (float)(contains(to_lower(vendor), "arm"))*(is_gpu?8.0f:1.0f); // ARM GPUs usually have 8 cores/CU, ARM CPUs have 1 core/CU
            cores = to_uint((float)compute_units*(nvidia+amd+intel+apple+arm)); // for CPUs, compute_units is the number of threads (twice the number of cores with hyperthreading)
            tflops = 1E-6f*(float)cores*(float)ipc*(float)clock_frequency; // estimated device floating point performance in TeraFLOPs/s
            if(tflops>best_value) {
                best_value = tflops;
                best_i = i;
            }
        }
        const string name = trim(cl_devices[best_i].getInfo<CL_DEVICE_NAME>()); // device name
        cl_device = cl_devices[best_i];
        print_info(name); // print device name
    }

    // 2. embed OpenCL C code (raw string literal breaks syntax highlighting)

    string opencl_c_code = R"(
        kernel void add_kernel(global float* A, global float* B, global float* C) { // equivalent to "for(uint n=0u; n<N; n++) {", but executed in parallel
            const uint n = get_global_id(0);
            C[n] = A[n]+B[n];
        }
    )";

    // 3. compile OpenCL C code

    cl::Context cl_context;
    cl::Program cl_program;
    cl::CommandQueue cl_queue;
    {
        cl_context = cl::Context(cl_device);
        cl_queue = cl::CommandQueue(cl_context, cl_device);
        cl::CommandQueue cl_queue(cl_context, cl_device); // queue to push commands for the device
        cl::Program::Sources cl_source;
        cl_source.push_back({ opencl_c_code.c_str(), opencl_c_code.length() });
        cl_program = cl::Program(cl_context, cl_source);
        int error = cl_program.build("-cl-fast-relaxed-math -w"); // compile OpenCL C code, disable warnings
        if(error) print_warning(cl_program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(cl_device)); // print build log
        if(error) print_error("OpenCL C code compilation failed.");
        else print_info("OpenCL C code successfully compiled.");
    }

    // 4. allocate memory on host and device

    const uint N = 1024u;
    float* host_A;
    float* host_B;
    float* host_C;
    cl::Buffer device_A;
    cl::Buffer device_B;
    cl::Buffer device_C;
    {
        host_A = new float[N];
        host_B = new float[N];
        host_C = new float[N];
        for(uint i=0u; i<N; i++) {
            host_A[i] = 0.0f; // zero all buffers
            host_B[i] = 0.0f;
            host_C[i] = 0.0f;
        }
        int error = 0;
        device_A = cl::Buffer(cl_context, CL_MEM_READ_WRITE, N*sizeof(float), nullptr, &error);
        if(error) print_error("OpenCL Buffer allocation failed with error code "+to_string(error)+".");
        device_B = cl::Buffer(cl_context, CL_MEM_READ_WRITE, N*sizeof(float), nullptr, &error);
        if(error) print_error("OpenCL Buffer allocation failed with error code "+to_string(error)+".");
        device_C = cl::Buffer(cl_context, CL_MEM_READ_WRITE, N*sizeof(float), nullptr, &error);
        if(error) print_error("OpenCL Buffer allocation failed with error code "+to_string(error)+".");
        cl_queue.enqueueWriteBuffer(device_A, true, 0u, N*sizeof(float), (void*)host_A); // have to keep track of buffer range and buffer data type
        cl_queue.enqueueWriteBuffer(device_B, true, 0u, N*sizeof(float), (void*)host_B);
        cl_queue.enqueueWriteBuffer(device_C, true, 0u, N*sizeof(float), (void*)host_C);
    }

    // 5. create Kernel object and link input parameters

    cl::NDRange cl_range_global, cl_range_local;
    cl::Kernel cl_kernel;
    {
        cl_kernel = cl::Kernel(cl_program, "add_kernel");
        cl_kernel.setArg(0, device_A);
        cl_kernel.setArg(1, device_B);
        cl_kernel.setArg(2, device_C);
        cl_range_local = cl::NDRange(WORKGROUP_SIZE);
        cl_range_global = cl::NDRange(((N+WORKGROUP_SIZE-1)/WORKGROUP_SIZE)*WORKGROUP_SIZE); // make global range a multiple of local range
    }

    // 6. finally run the actual program

    {
        for(uint i=0u; i<N; i++) {
            host_A[i] = 3.0f; // initialize buffers on host
            host_B[i] = 2.0f;
            host_C[i] = 1.0f;
        }

        print_info("Value before kernel execution: C[0] = "+to_string(host_C[0]));

        cl_queue.enqueueWriteBuffer(device_A, true, 0u, N*sizeof(float), (void*)host_A); // copy A and B to device
        cl_queue.enqueueWriteBuffer(device_B, true, 0u, N*sizeof(float), (void*)host_B); // have to keep track of buffer range and buffer data type
        cl_queue.enqueueNDRangeKernel(cl_kernel, cl::NullRange, cl_range_global, cl_range_local); // have to keep track of kernel ranges
        cl_queue.finish(); // don't forget to finish the queue
        cl_queue.enqueueReadBuffer(device_C, true, 0u, N*sizeof(float), (void*)host_C);

        print_info("Value after kernel execution: C[0] = "+to_string(host_C[0]));
    }

    wait();
    return 0;
}