CODEWITHSUNDEEP
multithreadinggpuopenclread-write
José Eduardo Bueno
José Eduardo Bueno

Reputation: 147

Write data "by chunks" into OpenCL buffer

I'm working on porting a CUDA application to OpenCL and I noticed that CUDA offers the functionality of writing data to its "buffers" by chunks. What I mean by this is the following:

int *vals = new int[N/4];
int *d_vec = nullptr;
cudaMalloc((void**)&d_vec, sizeof(int) * N);
for(int i = 0; i < 4; i++){
    cudaMemcpy(d_vec + i*(N/4), vals, sizeof(int) * N/4, cudaMemcpyHostToDevice);
}

What the code above does is to write the vals array (which has 1/4 of the d_vec buffer) sequentially to d_vec. So my question is, is it possible to do the same with OpenCL? That is, allocate a buffer and write values sequentially to it, without having to write an array with the full buffer size?

Thank you!

Upvotes: 0

Views: 100

Answers (2)

noma
noma

Reputation: 1229

Yes, you can specify a size and offset for clEnqueueWriteBuffer, which would be your replacement for the cudaMemcpy.

cl_int clEnqueueWriteBuffer(
    cl_command_queue command_queue,
    cl_mem buffer,
    cl_bool blocking_write,
    size_t offset, // from your example: i*(N/4)
    size_t size,   // from your example: sizeof(int) * N/4
    const void* ptr,
    cl_uint num_events_in_wait_list,
    const cl_event* event_wait_list,
    cl_event* event);

Upvotes: 2

ProjectPhysX
ProjectPhysX

Reputation: 5746

Yes, this is indeed possible with enqueueWriteBuffer:

cl_int *vals = new cl_int[N/4];
Buffer d_vec;
d_vec = Buffer(context, CL_MEM_READ_WRITE, N*sizeof(cl_int));
for(int i = 0; i < 4; i++){
    queue.enqueueWriteBuffer(d_vec, true, i*(N/4), sizeof(cl_int)*N/4, (void*)vals);
    queue.finish();
}

Upvotes: 2

Related Questions