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.
#define CL_HPP_MINIMUM_OPENCL_VERSION 100
#define CL_HPP_TARGET_OPENCL_VERSION 300
#include <CL/opencl.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_device.getInfo<CL_DEVICE_NAME>()); // device name
const string vendor = trim(cl_device.getInfo<CL_DEVICE_VENDOR>()); // device vendor
const uint compute_units = (uint)cl_device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>(); // compute units (CUs) can contain multiple cores depending on the microarchitecture
const uint clock_frequency = (uint)cl_device.getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>(); // in MHz
const bool is_gpu = cl_device.getInfo<CL_DEVICE_TYPE>()==CL_DEVICE_TYPE_GPU;
const int vendor_id = (int)cl_device.getInfo<CL_DEVICE_VENDOR_ID>(); // AMD=0x1002, Intel=0x8086, Nvidia=0x10DE, Apple=0x1027F00
float cores_per_cu = 1.0f;
if(vendor_id==0x1002) { // AMD GPU/CPU
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
cores_per_cu = is_gpu ? (amd_256_cores_per_dualcu ? 256.0f : amd_128_cores_per_dualcu ? 128.0f : 64.0f) : 0.5f; // 64 cores/CU (GCN, CDNA), 128 cores/dualCU (RDNA, RDNA2), 256 cores/dualCU (RDNA3), 1/2 core/CU (CPUs)
} else if(vendor_id==0x8086) { // Intel GPU/CPU
const bool intel_16_cores_per_cu = contains_any(to_lower(name), {"gpu max", "140v", "130v", "b580", "b570"}); // identify PVC/Xe2 GPUs
cores_per_cu = 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
} else if(vendor_id==0x10DE||vendor_id==0x13B5) { // Nvidia GPU/CPU
const uint nvidia_compute_capability = 10u*(uint)cl_device.getInfo<CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV>()+(uint)cl_device.getInfo<CL_DEVICE_COMPUTE_CAPABILITY_MINOR_NV>();
const bool nvidia__32_cores_per_cu = (nvidia_compute_capability <30); // identify Fermi GPUs
const bool nvidia_192_cores_per_cu = (nvidia_compute_capability>=30&&nvidia_compute_capability<50); // identify Kepler GPUs
const bool nvidia__64_cores_per_cu = (nvidia_compute_capability>=70&&nvidia_compute_capability<80)||contains_any(to_lower(name), {"p100", "a100", "a30"}); // identify Volta, Turing, P100, A100, A30
cores_per_cu = is_gpu ? (nvidia__32_cores_per_cu ? 32.0f : nvidia_192_cores_per_cu ? 192.0f : nvidia__64_cores_per_cu ? 64.0f : 128.0f) : 1.0f; // 32 (Fermi), 192 (Kepler), 64 (Volta, Turing, P100, A100, A30), 128 (Maxwell, Pascal, Ampere, Hopper, Ada, Blackwell) or 1 (CPUs)
} else if(vendor_id==0x1027F00) { // Apple iGPU
cores_per_cu = 128.0f; // Apple ARM GPUs usually have 128 cores/CU
} else if(vendor_id==0x1022||vendor_id==0x10006||vendor_id==0x6C636F70) { // x86 CPUs with PoCL runtime
cores_per_cu = 0.5f; // CPUs typically have 1/2 cores/CU due to SMT/hyperthreading
} else if(contains(to_lower(vendor), "arm")) { // ARM
cores_per_cu = is_gpu ? 8.0f : 1.0f; // ARM GPUs usually have 8 cores/CU, ARM CPUs have 1 core/CU
}
const uint ipc = is_gpu ? 2u : 32u; // IPC (instructions per cycle) is 2 for GPUs and 32 for most modern CPUs
const uint cores = to_uint((float)compute_units*cores_per_cu); // for CPUs, compute_units is the number of threads (twice the number of cores with hyperthreading)
const float 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_device }, "-cl-finite-math-only -cl-no-signed-zeros -cl-mad-enable -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;
}