void forward_iseg_layer_gpu(const layer l, network net) {
copy_gpu(l.batch * l.inputs, net.input_gpu, 1, l.output_gpu, 1, net.st);
int b;
for (b = 0; b < l.batch; ++b) {
activate_array_gpu(l.output_gpu + b * l.outputs, l.classes * l.w * l.h,
LOGISTIC, net.st);
//if(l.extra) activate_array_gpu(l.output_gpu + b*l.outputs + l.classes*l.w*l.h, l.extra*l.w*l.h, LOGISTIC);
}
cuda_pull_array(l.output_gpu, net.input, l.batch * l.inputs);
forward_iseg_layer(l, net);
cuda_push_array(l.delta_gpu, l.delta, l.batch * l.outputs);
}
void forward_shortcut_layer_gpu(const layer l, network net)
{
copy_gpu(l.outputs*l.batch, net.input_gpu, 1, l.output_gpu, 1);
shortcut_gpu(l.batch, l.w, l.h, l.c, net.layers[l.index].output_gpu, l.out_w, l.out_h, l.out_c, l.output_gpu);
activate_array_gpu(l.output_gpu, l.outputs*l.batch, l.activation);
}
void backward_lstm_layer_gpu(layer l, network state) {
network s = { 0 };
s.train = state.train;
int i;
layer wf = *(l.wf);
layer wi = *(l.wi);
layer wg = *(l.wg);
layer wo = *(l.wo);
layer uf = *(l.uf);
layer ui = *(l.ui);
layer ug = *(l.ug);
layer uo = *(l.uo);
increment_layer(&wf, l.steps - 1);
increment_layer(&wi, l.steps - 1);
increment_layer(&wg, l.steps - 1);
increment_layer(&wo, l.steps - 1);
increment_layer(&uf, l.steps - 1);
increment_layer(&ui, l.steps - 1);
increment_layer(&ug, l.steps - 1);
increment_layer(&uo, l.steps - 1);
state.input_gpu += l.inputs * l.batch * (l.steps - 1);
if (state.delta_gpu)
state.delta_gpu += l.inputs * l.batch * (l.steps - 1);
l.output_gpu += l.outputs * l.batch * (l.steps - 1);
l.cell_gpu += l.outputs * l.batch * (l.steps - 1);
l.delta_gpu += l.outputs * l.batch * (l.steps - 1);
for (i = l.steps - 1; i >= 0; --i) {
if (i != 0)
copy_gpu(l.outputs * l.batch, l.cell_gpu - l.outputs * l.batch, 1,
l.prev_cell_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.cell_gpu, 1, l.c_gpu, 1, state.st);
if (i != 0)
copy_gpu(l.outputs * l.batch, l.output_gpu - l.outputs * l.batch, 1,
l.prev_state_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.output_gpu, 1, l.h_gpu, 1, state.st);
l.dh_gpu = (i == 0) ? 0 : l.delta_gpu - l.outputs * l.batch;
copy_gpu(l.outputs * l.batch, wf.output_gpu, 1, l.f_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, uf.output_gpu, 1, l.f_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, wi.output_gpu, 1, l.i_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, ui.output_gpu, 1, l.i_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, wg.output_gpu, 1, l.g_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, ug.output_gpu, 1, l.g_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, wo.output_gpu, 1, l.o_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, uo.output_gpu, 1, l.o_gpu, 1, state.st);
activate_array_gpu(l.f_gpu, l.outputs * l.batch, LOGISTIC, state.st);
activate_array_gpu(l.i_gpu, l.outputs * l.batch, LOGISTIC, state.st);
activate_array_gpu(l.g_gpu, l.outputs * l.batch, TANH, state.st);
activate_array_gpu(l.o_gpu, l.outputs * l.batch, LOGISTIC, state.st);
copy_gpu(l.outputs * l.batch, l.delta_gpu, 1, l.temp3_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.c_gpu, 1, l.temp_gpu, 1, state.st);
activate_array_gpu(l.temp_gpu, l.outputs * l.batch, TANH, state.st);
copy_gpu(l.outputs * l.batch, l.temp3_gpu, 1, l.temp2_gpu, 1, state.st);
mul_gpu(l.outputs * l.batch, l.o_gpu, 1, l.temp2_gpu, 1, state.st);
gradient_array_gpu(l.temp_gpu, l.outputs * l.batch, TANH, l.temp2_gpu, state.st);
axpy_gpu(l.outputs * l.batch, 1, l.dc_gpu, 1, l.temp2_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.c_gpu, 1, l.temp_gpu, 1, state.st);
activate_array_gpu(l.temp_gpu, l.outputs * l.batch, TANH, state.st);
mul_gpu(l.outputs * l.batch, l.temp3_gpu, 1, l.temp_gpu, 1, state.st);
gradient_array_gpu(l.o_gpu, l.outputs * l.batch, LOGISTIC, l.temp_gpu, state.st);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wo.delta_gpu, 1, state.st);
s.input_gpu = l.prev_state_gpu;
s.delta_gpu = l.dh_gpu;
backward_connected_layer_gpu(wo, s);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, uo.delta_gpu, 1, state.st);
s.input_gpu = state.input_gpu;
s.delta_gpu = state.delta_gpu;
backward_connected_layer_gpu(uo, s);
copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
mul_gpu(l.outputs * l.batch, l.i_gpu, 1, l.temp_gpu, 1, state.st);
gradient_array_gpu(l.g_gpu, l.outputs * l.batch, TANH, l.temp_gpu, state.st);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wg.delta_gpu, 1, state.st);
s.input_gpu = l.prev_state_gpu;
s.delta_gpu = l.dh_gpu;
backward_connected_layer_gpu(wg, s);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, ug.delta_gpu, 1, state.st);
s.input_gpu = state.input_gpu;
s.delta_gpu = state.delta_gpu;
backward_connected_layer_gpu(ug, s);
copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
mul_gpu(l.outputs * l.batch, l.g_gpu, 1, l.temp_gpu, 1, state.st);
gradient_array_gpu(l.i_gpu, l.outputs * l.batch, LOGISTIC, l.temp_gpu, state.st);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wi.delta_gpu, 1, state.st);
s.input_gpu = l.prev_state_gpu;
s.delta_gpu = l.dh_gpu;
backward_connected_layer_gpu(wi, s);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, ui.delta_gpu, 1, state.st);
s.input_gpu = state.input_gpu;
s.delta_gpu = state.delta_gpu;
backward_connected_layer_gpu(ui, s);
copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
mul_gpu(l.outputs * l.batch, l.prev_cell_gpu, 1, l.temp_gpu, 1, state.st);
gradient_array_gpu(l.f_gpu, l.outputs * l.batch, LOGISTIC, l.temp_gpu, state.st);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, wf.delta_gpu, 1, state.st);
s.input_gpu = l.prev_state_gpu;
s.delta_gpu = l.dh_gpu;
backward_connected_layer_gpu(wf, s);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, uf.delta_gpu, 1, state.st);
s.input_gpu = state.input_gpu;
s.delta_gpu = state.delta_gpu;
backward_connected_layer_gpu(uf, s);
copy_gpu(l.outputs * l.batch, l.temp2_gpu, 1, l.temp_gpu, 1, state.st);
mul_gpu(l.outputs * l.batch, l.f_gpu, 1, l.temp_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.temp_gpu, 1, l.dc_gpu, 1, state.st);
state.input_gpu -= l.inputs * l.batch;
if (state.delta_gpu)
state.delta_gpu -= l.inputs * l.batch;
l.output_gpu -= l.outputs * l.batch;
l.cell_gpu -= l.outputs * l.batch;
l.delta_gpu -= l.outputs * l.batch;
increment_layer(&wf, -1);
increment_layer(&wi, -1);
increment_layer(&wg, -1);
increment_layer(&wo, -1);
increment_layer(&uf, -1);
increment_layer(&ui, -1);
increment_layer(&ug, -1);
increment_layer(&uo, -1);
}
}
void forward_lstm_layer_gpu(layer l, network state) {
network s = { 0 };
s.train = state.train;
int i;
layer wf = *(l.wf);
layer wi = *(l.wi);
layer wg = *(l.wg);
layer wo = *(l.wo);
layer uf = *(l.uf);
layer ui = *(l.ui);
layer ug = *(l.ug);
layer uo = *(l.uo);
fill_gpu(l.outputs * l.batch * l.steps, 0, wf.delta_gpu, 1, state.st);
fill_gpu(l.outputs * l.batch * l.steps, 0, wi.delta_gpu, 1, state.st);
fill_gpu(l.outputs * l.batch * l.steps, 0, wg.delta_gpu, 1, state.st);
fill_gpu(l.outputs * l.batch * l.steps, 0, wo.delta_gpu, 1, state.st);
fill_gpu(l.outputs * l.batch * l.steps, 0, uf.delta_gpu, 1, state.st);
fill_gpu(l.outputs * l.batch * l.steps, 0, ui.delta_gpu, 1, state.st);
fill_gpu(l.outputs * l.batch * l.steps, 0, ug.delta_gpu, 1, state.st);
fill_gpu(l.outputs * l.batch * l.steps, 0, uo.delta_gpu, 1, state.st);
if (state.train) {
fill_gpu(l.outputs * l.batch * l.steps, 0, l.delta_gpu, 1, state.st);
}
for (i = 0; i < l.steps; ++i) {
s.input_gpu = l.h_gpu;
forward_connected_layer_gpu(wf, s);
forward_connected_layer_gpu(wi, s);
forward_connected_layer_gpu(wg, s);
forward_connected_layer_gpu(wo, s);
s.input_gpu = state.input_gpu;
forward_connected_layer_gpu(uf, s);
forward_connected_layer_gpu(ui, s);
forward_connected_layer_gpu(ug, s);
forward_connected_layer_gpu(uo, s);
copy_gpu(l.outputs * l.batch, wf.output_gpu, 1, l.f_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, uf.output_gpu, 1, l.f_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, wi.output_gpu, 1, l.i_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, ui.output_gpu, 1, l.i_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, wg.output_gpu, 1, l.g_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, ug.output_gpu, 1, l.g_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, wo.output_gpu, 1, l.o_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, uo.output_gpu, 1, l.o_gpu, 1, state.st);
activate_array_gpu(l.f_gpu, l.outputs * l.batch, LOGISTIC, state.st);
activate_array_gpu(l.i_gpu, l.outputs * l.batch, LOGISTIC, state.st);
activate_array_gpu(l.g_gpu, l.outputs * l.batch, TANH, state.st);
activate_array_gpu(l.o_gpu, l.outputs * l.batch, LOGISTIC, state.st);
copy_gpu(l.outputs * l.batch, l.i_gpu, 1, l.temp_gpu, 1, state.st);
mul_gpu(l.outputs * l.batch, l.g_gpu, 1, l.temp_gpu, 1, state.st);
mul_gpu(l.outputs * l.batch, l.f_gpu, 1, l.c_gpu, 1, state.st);
axpy_gpu(l.outputs * l.batch, 1, l.temp_gpu, 1, l.c_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.c_gpu, 1, l.h_gpu, 1, state.st);
activate_array_gpu(l.h_gpu, l.outputs * l.batch, TANH, state.st);
mul_gpu(l.outputs * l.batch, l.o_gpu, 1, l.h_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.c_gpu, 1, l.cell_gpu, 1, state.st);
copy_gpu(l.outputs * l.batch, l.h_gpu, 1, l.output_gpu, 1, state.st);
state.input_gpu += l.inputs * l.batch;
l.output_gpu += l.outputs * l.batch;
l.cell_gpu += l.outputs * l.batch;
increment_layer(&wf, 1);
increment_layer(&wi, 1);
increment_layer(&wg, 1);
increment_layer(&wo, 1);
increment_layer(&uf, 1);
increment_layer(&ui, 1);
increment_layer(&ug, 1);
increment_layer(&uo, 1);
}
}
void forward_activation_layer_gpu(layer l, network net) {
copy_gpu(l.outputs * l.batch, net.input_gpu, 1, l.output_gpu, 1, net.st);
activate_array_gpu(l.output_gpu, l.outputs * l.batch, l.activation, net.st);
}
