/// Select best launch configuration for the given shared memory requirements.
void config(const command_queue &q, std::function<size_t(size_t)> smem) {
// Select workgroup size that would fit into the device.
size_t ws = q.device().max_threads_per_block() / 2;
size_t max_ws = max_threads_per_block(q);
size_t max_smem = max_shared_memory_per_block(q);
// Reduce workgroup size until it satisfies resource requirements:
while( (ws > max_ws) || (smem(ws) > max_smem) )
ws /= 2;
config(num_workgroups(q), ws);
}
/// Select best launch configuration for the given shared memory requirements.
void config(const cl::CommandQueue &queue, std::function<size_t(size_t)> smem) {
cl::Device dev = queue.getInfo<CL_QUEUE_DEVICE>();
if ( is_cpu(queue) ) {
w_size = 1;
} else {
// Select workgroup size that would fit into the device.
w_size = dev.getInfo<CL_DEVICE_MAX_WORK_ITEM_SIZES>()[0] / 2;
size_t max_ws = max_threads_per_block(queue);
size_t max_smem = max_shared_memory_per_block(queue);
// Reduce workgroup size until it satisfies resource requirements:
while( (w_size > max_ws) || (smem(w_size) > max_smem) )
w_size /= 2;
}
g_size = w_size * num_workgroups(queue);
}
/// Select best launch configuration for the given shared memory requirements.
void config(const boost::compute::command_queue &queue, std::function<size_t(size_t)> smem) {
boost::compute::device dev = queue.get_device();
size_t ws;
if ( is_cpu(queue) ) {
ws = 1;
} else {
// Select workgroup size that would fit into the device.
ws = dev.get_info<std::vector<size_t>>(CL_DEVICE_MAX_WORK_ITEM_SIZES)[0] / 2;
size_t max_ws = max_threads_per_block(queue);
size_t max_smem = max_shared_memory_per_block(queue);
// Reduce workgroup size until it satisfies resource requirements:
while( (ws > max_ws) || (smem(ws) > max_smem) )
ws /= 2;
}
config(num_workgroups(queue), ws);
}
