/**
* ufo_op_gradient_descent:
* @arg: A #UfoBuffer
* @out: A #UfoBuffer
* @resources: #UfoResources object
* @command_queue: A valid cl_command_queue
*
* Returns: (transfer full): Event of the POSC operation
*/
gpointer
ufo_op_gradient_descent (UfoBuffer *arg,
UfoBuffer *out,
UfoResources *resources,
gpointer command_queue)
{
UfoRequisition arg_requisition;
cl_event event;
GError *error = NULL;
static GMutex mutex;
ufo_buffer_get_requisition (arg, &arg_requisition);
ufo_buffer_resize (out, &arg_requisition);
cl_mem d_arg = ufo_buffer_get_device_image (arg, command_queue);
cl_mem d_out = ufo_buffer_get_device_image (out, command_queue);
cl_kernel kernel = ufo_resources_get_cached_kernel (resources, OPS_FILENAME, "descent_grad", &error);
if (error) {
g_error ("%s\n", error->message);
}
g_mutex_lock (&mutex);
UFO_RESOURCES_CHECK_CLERR (clSetKernelArg(kernel, 0, sizeof(void *), (void *) &d_arg));
UFO_RESOURCES_CHECK_CLERR (clSetKernelArg(kernel, 1, sizeof(void *), (void *) &d_out));
UFO_RESOURCES_CHECK_CLERR (clEnqueueNDRangeKernel (command_queue, kernel,
arg_requisition.n_dims, NULL, arg_requisition.dims,
NULL, 0, NULL, &event));
g_mutex_unlock (&mutex);
return event;
}
static void
handle_send_inputs (UfoDaemon *daemon, UfoMessage *request)
{
UfoDaemonPrivate *priv = UFO_DAEMON_GET_PRIVATE (daemon);
gpointer context = ufo_scheduler_get_context (priv->scheduler);
UfoRequisition *requisition = (UfoRequisition *) request->data;
// send the ack for that
UfoMessage *response = ufo_message_new (UFO_MESSAGE_ACK, 0);
ufo_messenger_send_blocking (priv->msger, response, NULL);
ufo_message_free (response);
if (priv->input == NULL) {
priv->input = ufo_buffer_new (requisition, NULL, context);
}
else {
if (ufo_buffer_cmp_dimensions (priv->input, requisition))
ufo_buffer_resize (priv->input, requisition);
}
UfoMessage *data_msg = ufo_messenger_recv_blocking (priv->msger, NULL);
response = ufo_message_new (UFO_MESSAGE_ACK, 0);
ufo_messenger_send_blocking (priv->msger, response, NULL);
ufo_message_free (response);
g_assert (ufo_buffer_get_size (priv->input) == data_msg->data_size);
ufo_buffer_set_host_array (priv->input, data_msg->data);
ufo_input_task_release_input_buffer (UFO_INPUT_TASK (priv->input_task), priv->input);
g_free (data_msg);
}
static gboolean
ufo_ir_sart_process_real (UfoMethod *method,
UfoBuffer *input,
UfoBuffer *output)
{
UfoIrSARTPrivate *priv = UFO_IR_SART_GET_PRIVATE (method);
UfoResources *resources = NULL;
UfoProjector *projector = NULL;
gpointer *cmd_queue = NULL;
gfloat relaxation_factor = 0;
guint max_iterations = 0;
g_object_get (method,
"ufo-resources", &resources,
"command-queue", &cmd_queue,
"projection-model", &projector,
"relaxation-factor", &relaxation_factor,
"max-iterations", &max_iterations,
NULL);
UfoGeometry *geometry = NULL;
g_object_get (projector, "geometry", &geometry, NULL);
//
// resize
UfoBuffer **method_buffers[4] = {
&priv->singular_volume,
&priv->singular_sino,
&priv->ray_weights,
&priv->b_temp
};
UfoBuffer *ref_buffers[4] = {
output,
input,
input,
input
};
for (guint i = 0; i < 4; ++i) {
UfoRequisition _req;
if (*method_buffers[i]) {
ufo_buffer_get_requisition (ref_buffers[i], &_req);
ufo_buffer_resize (*method_buffers[i], &_req);
} else {
*method_buffers[i] = ufo_buffer_dup (ref_buffers [i]);
}
}
ufo_op_set (priv->singular_volume, 1.0f, resources, cmd_queue);
ufo_op_set (priv->singular_sino, 1.0f, resources, cmd_queue);
ufo_op_set (priv->ray_weights, 0, resources, cmd_queue);
guint n_subsets = 0;
UfoProjectionsSubset *subset = generate_subsets (geometry, &n_subsets);
for (guint i = 0 ; i < n_subsets; ++i) {
ufo_projector_FP (projector,
priv->singular_volume,
priv->ray_weights,
&subset[i],
1.0f,
NULL);
}
ufo_op_inv (priv->ray_weights, resources, cmd_queue);
guint iteration = 0;
while (iteration < max_iterations) {
ufo_buffer_copy (input, priv->b_temp);
for (guint i = 0 ; i < n_subsets; i++) {
ufo_projector_FP (projector,
output,
priv->b_temp,
&subset[i],
-1.0f,
NULL);
ufo_op_mul_rows (priv->b_temp,
priv->ray_weights,
priv->b_temp,
subset[i].offset,
subset[i].n,
resources,
cmd_queue);
ufo_projector_BP (projector,
output,
priv->b_temp,
&subset[i],
relaxation_factor,
NULL);
}
iteration++;
}
return TRUE;
}