How to pass vector parameter to OpenCL kernel in C?

问题

I'm having trouble passing a vector type (uint8) parameter to an OpenCL kernel function from the host code in C.

In the host I've got the data in an array:

cl_uint dataArr[8] = { 1, 2, 3, 4, 5, 6, 7, 8 };

(My real data is more than just [1, 8]; this is just for ease of explanation.)

I then transfer the data over to a buffer to be passed to the kernel:

cl_mem kernelInputData = clCreateBuffer(context,
    CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof(cl_uint)*8, dataArr, NULL);

Next, I pass this buffer into the kernel:

clSetKernelArg(kernel, 0, sizeof(cl_mem), &kernelInputData);

And the kernel function's signature looks something like this:

kernel void kernelFunction(constant uint8 *vectorPtr)

However, the kernel doesn't seem to be obtaining the correct input data from the pointer to kernelInputData. When I pass values back from within the kernel, I see that vectorPtr points to something with this structure: ( 1, 2, 3, 4, 5, ?, ?, ? ) where the question marks are usually 4293848814 but sometimes 0. Either way, not what they're supposed to be.

What am I doing wrong?

EDIT:

I've switched from using an array to cl_uint8 on the host side. I now have:

cl_uint8 dataVector = { 1, 2, 3, 4, 5, 6, 7, 8 };

And I pass this vector to the kernel like so:

clSetKernelArg(kernel, 0, sizeof(cl_uint8), &dataVector);

And the kernel function's signature looks something like this:

kernel void kernelFunction(constant uint8 *vectorPtr)

However, running this code gives me a CL_INVALID_ARG_SIZE error on clSetKernelArg(). This error goes away if I switch the ARG_SIZE paramater to sizeof(cl_uint8 *) but then I get an EXC_BAD_ACCESS error in __dynamic_cast within clSetKernelArg().

My device is:

Apple Macbook Pro (mid-2009)
OSX 10.8 Mountain Lion
NVIDIA GeForce 9400M
OpenCL 1.0
CLH 1.0

回答1:

You are defining an array of cl_uint of size 8. The creation of the cl_mem and the setting of kernel argument are right. But your kernel argument isn't correct: you try to read an array of cl_uint8 instead of cl_uint.

If you want to use a vector data type, you must declare: cl_uint8 dataArr of size 1. Or if you want to use an array of size 8: kernel void kernelFunction(constant uint *vectorPtr, uint size):

Edit:

The kernel parameter for cl_uint8 dataVector is not a pointer. So, the correct code is:

cl_uint8 dataVector = { 1, 2, 3, 4, 5, 6, 7, 8 };
clSetKernelArg(kernel, 0, sizeof(cl_uint8), &dataVector);

and

kernel void kernelFunction(constant uint8 vectorPtr)

回答2:

Minimal runnable example

int2 is passed to the kernel. It is initially created as an array of cl_int.

#include <assert.h>
#include <stdlib.h>

#include <CL/cl.h>

int main(void) {
    const char *source =
        "__kernel void main(__global int2 *out) {\n"
        "      out[get_global_id(0)]++;\n"
        "}\n";
    cl_command_queue command_queue;
    cl_context context;
    cl_device_id device;
    cl_int input[] = {0, 1, 2, 3};
    const size_t global_work_size = sizeof(input) / sizeof(cl_int2);
    cl_kernel kernel;
    cl_mem buffer;
    cl_platform_id platform;
    cl_program program;

    clGetPlatformIDs(1, &platform, NULL);
    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 1, &device, NULL);
    context = clCreateContext(NULL, 1, &device, NULL, NULL, NULL);
    command_queue = clCreateCommandQueue(context, device, 0, NULL);
    buffer = clCreateBuffer(context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, sizeof(input), &input, NULL);
    program = clCreateProgramWithSource(context, 1, &source, NULL, NULL);
    clBuildProgram(program, 1, &device, "", NULL, NULL);
    kernel = clCreateKernel(program, "main", NULL);
    clSetKernelArg(kernel, 0, sizeof(cl_mem), &buffer);
    clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, &global_work_size, NULL, 0, NULL, NULL);
    clFlush(command_queue);
    clFinish(command_queue);
    clEnqueueReadBuffer(command_queue, buffer, CL_TRUE, 0, sizeof(input), &input, 0, NULL, NULL);

    assert(input[0] == 1);
    assert(input[1] == 2);
    assert(input[2] == 3);
    assert(input[3] == 4);
    return EXIT_SUCCESS;
}

Tested on Ubuntu 15.10 OpenCL 1.2 NVIDIA 352.53.

来源：https://stackoverflow.com/questions/13118228/how-to-pass-vector-parameter-to-opencl-kernel-in-c

标签

vector

parameter-passing

opencl