OpenCL: buffering inputs for the same kernel - buffer

Question

I have been trying to read a file and load it into a buffer of a kernel in OpenCL, while the kernel is processing another buffer. However, it seems to not like that: for some reason, the results are wrong.
First, I tried setting the Args for the same kernel every time before enqueueing a task. Then, I tried enqueuing tasks for 2 kernels of the same function like below, without changing the arguments:
krnl_1.setArg(0, buffer_a));
krnl_1.setArg(1, output_buffer));
krnl_2.setArg(0, buffer_b));
krnl_2.setArg(1, output_buffer));
void* ptr[2];
ptr[0] = q.enqueueMapBuffer(buffer_a, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, buffer_size_in_bytes, NULL, NULL, &err);
ptr[1] = q.enqueueMapBuffer(buffer_b, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, buffer_size_in_bytes, NULL, NULL, &err);
int sel = 0;
long long bytes_sent = 0;
// Fill buffer_a
bytes_sent += pread(myFd, (void*)ptr[sel], buffer_size_in_bytes, bytes_sent);
while (bytes_sent < total_size_in_bytes){
if (sel == 0){ // If buffer_a was just filled
q.enqueueTask(krnl_1);
sel = 1; // Fill buffer_b
} else { // If buffer_b was just filled
q.enqueueTask(krnl_2);
sel = 0; // Fill buffer_a
}
if (bytes_sent >= total_size_in_bytes) // If this is the last task
q.enqueueMigrateMemObjects({output_buffer},CL_MIGRATE_MEM_OBJECT_HOST);
else // Fill the buffer that is not being processed
bytes_sent += pread(myFd, (void*)ptr[sel], buffer_size_in_bytes, bytes_sent);
q.finish();
}
If I do it serially, it is working fine:
void* ptr[2];
ptr[0] = q.enqueueMapBuffer(buffer_a, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, buffer_size_in_bytes, NULL, NULL, &err);
ptr[1] = q.enqueueMapBuffer(buffer_b, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, buffer_size_in_bytes, NULL, NULL, &err);
int sel = 0;
long long bytes_sent = 0;
while (bytes_sent < total_size_in_bytes){
bytes_sent += pread(myFd, (void*)ptr[sel], buffer_size_in_bytes, bytes_sent);
if (sel == 0){
q.enqueueTask(krnl_1);
sel = 1;
} else {
q.enqueueTask(krnl_2);
sel = 0;
}
if (bytes_sent >= total_size_in_bytes) //if this is the last task
q.enqueueMigrateMemObjects({output_buffer},CL_MIGRATE_MEM_OBJECT_HOST);
q.finish();
}
I feel like I must have miscomprehended the way OpenCL treats arguments and enqueues tasks, but I cannot find any similar examples.

Answer 1

First, reads and writes by a kernel executing on a device to a memory region mapped for writing are undefined (for more information please see Accessing mapped regions of a memory object section for clEnqueueMapBuffer). Therefore, it is necessary to unmap the buffer before run the task.
Second, bytes_sent is incremented before it is checked in while() condition in the first example. If the first call of pread() reads all data, the loop won't be executed.
Therefore, I expect that the code should look something like this:
krnl_1.setArg(0, buffer_a);
krnl_1.setArg(1, output_buffer);
krnl_2.setArg(0, buffer_b);
krnl_2.setArg(1, output_buffer);
int sel = 0;
long long bytes_processed = 0;
void* buf_ptr = q.enqueueMapBuffer(buffer_a, CL_TRUE, CL_MAP_WRITE, 0, buffer_size_in_bytes, NULL, NULL, &err);
long long bytes_to_process = pread(myFd, buf_ptr, buffer_size_in_bytes, bytes_processed);
err = q.enqueueUnmapMemObject(buffer_a, buf_ptr);
while (bytes_processed < total_size_in_bytes)
{
if (sel == 0){ // If buffer_a was just filled
q.enqueueTask(krnl_1);
sel = 1; // Fill buffer_b
} else { // If buffer_b was just filled
q.enqueueTask(krnl_2);
sel = 0; // Fill buffer_a
}
bytes_processed += bytes_to_process;
if (bytes_processed < total_size_in_bytes) { // Fill the buffer that is not being processed
auto& buffer = sel ? buffer_b : buffer_a;
buf_ptr = q.enqueueMapBuffer(buffer, CL_TRUE, CL_MAP_WRITE, 0, buffer_size_in_bytes, NULL, NULL, &err);
bytes_to_process = pread(myFd, buf_ptr, buffer_size_in_bytes, bytes_processed);
err = q.enqueueUnmapMemObject(buffer, buf_ptr);
}
else { // If this is the last task
buf_ptr = q.enqueueMapBuffer(output_buffer, CL_TRUE, CL_MAP_READ, 0, output_buffer_size_in_bytes, NULL, NULL, &err);
}
}