void print_network(Network* network){
printf("####### NETWORK CONFIGURATION #######\n");
print_layer(network->input_layer);
for (size_t i = 0; i < network->hidden_layers_count; i++) {
print_layer(network->hidden_layers[i]);
}
print_layer(network->output_layer);
printf("#####################################\n\n");
}
static void DumpEdgeList(DirectedGraph<Layer*>& aGraph)
{
nsTArray<DirectedGraph<Layer*>::Edge> edges = aGraph.GetEdgeList();
for (uint32_t i = 0; i < edges.Length(); i++) {
fprintf(stderr, "From: ");
print_layer(stderr, edges.ElementAt(i).mFrom);
fprintf(stderr, ", To: ");
print_layer(stderr, edges.ElementAt(i).mTo);
fprintf(stderr, "\n");
}
}
void
print(int c) /* show everything (above draft layers) */
{
int n;
#ifndef NOWRITE
#if 0 /* 4/23/91--DBK */
/* Added 10/15/90--DBK. Is there any reason NOT to do this? */
redo_gbox = 1;
#else
redogbox(); /* just do it here */
#endif
if (pipefp) {
pclose(pipefp);
pipefp = NULL;
if ((eqnfp = fopen(psfname,"r")) == NULL)
fatal("cannot read eqn output");
}
flyback = (c == 'F');
if (c != 'N')
for (n = draftlayer; n <= top_layer; ++n)
print_layer(n);
if (eqnfp) {
fclose(eqnfp);
eqnfp = NULL;
eqn_count = 0;
unlink(psfname);
}
#else
fprintf(stderr,
"This version of picasso is for on-screen previewing only.\n");
exit(1);
#endif
}
static void DumpLayerList(nsTArray<Layer*>& aLayers)
{
for (uint32_t i = 0; i < aLayers.Length(); i++) {
print_layer(stderr, aLayers.ElementAt(i));
fprintf(stderr, " ");
}
fprintf(stderr, "\n");
}
void sastantua(int size)
{
int base;
int layer;
base = find_base(size);
layer = print_layer(base);
if (layer == size)
{
}
}
int do_layer (struct domain *dom, struct pdb_ATOM *model)
{
struct layer curr_layer;
int i, new_site=-1;
double z;
/* loop over layers in domain */
for (i = 0; i < dom->q; i++)
{
z = dom->z1 + i*dom->dz;
curr_layer = setup_layer(z, dom);
print_layer(&curr_layer);
new_site = do_ring(&curr_layer, model);
};
return new_site;
};
int main() {
LOG_LEVEL = 0;
// ==============
// = Unit Tests =
// ==============
if (!do_unittests()) {
printf("UNIT TESTS FAILED, ABORTING.");
return -1;
};
// =================
// = Setup Network =
// =================
// Generate + Load Network Config from network.hpp/network.cpp
network_t *net_CPU;
net_CPU = get_network_config();
// Assert that layer_t fits into a multiple of bus transactions:
// ONLY NECESSARY IF WE CAN MAP LAYER_T TRANSFER ONTO BUS_T AXI MASTER
// printf("size of layer_t: %d, size of bus_t: %d", (int)sizeof(layer_t),
// (int)sizeof(bus_t));
// assert((sizeof(layer_t) % sizeof(bus_t) == 0) &&
// "layert_t is not multiple of bus size. adjust size of
// layer_t.dummy!");
// ==========================
// = Setup FPGA Accelerator =
// ==========================
// Allocate Shared Memory for Config, Weights, Data.
// Copy Layer Config + Weights to FPGA.
setup_FPGA(net_CPU);
// ===========================
// = Load + Copy Input Image =
// ===========================
/* Structured: generate_structured_input_image(input_image,win,hin,chin);
PseudoRandom: generate_random_input_image(input_image, win, hin, chin, 1);
ReallyRandom: generate_random_input_image(input_image, win, hin, chin -1);
Prepared Input File (convert_image.py):
load_prepared_input_image(input_image, "./indata.bin", win, hin, chin);
JPG/PNG Input File (!not implemented!):
load_image_file(input_image, "./puppy-500x350.jpg", win, hin, chin);
do_preprocess(input_image, win, hin, chin); */
// Allocate Memory on CPU Side:
layer_t layer0 = net_CPU->layers[0];
int win = layer0.width;
int hin = layer0.height;
int chin = layer0.channels_in;
data_t *input_image = (data_t *)malloc(win * hin * chin * sizeof(data_t));
// Load Input Image
load_prepared_input_image(input_image, "./indata.bin", win, hin, chin);
// Copy onto FPGA
copy_input_image_to_FPGA(net_CPU, input_image);
// ============================
// = Execute FPGA Accelerator =
// ============================
L_LAYERS:
for (int layer_id = 0; layer_id < net_CPU->num_layers; layer_id++) {
layer_t layer = net_CPU->layers[layer_id];
LOG("Layer %2d: <%s>\n", layer_id, layer.name);
LOG_LEVEL_INCR;
// Calculate Memory Pointers
LOG("SHARED_DRAM is at address: %lu\n", (long)SHARED_DRAM);
int weights_offset =
((long)SHARED_DRAM_WEIGHTS - (long)SHARED_DRAM) / sizeof(data_t);
int input_offset =
((long)SHARED_DRAM_DATA - (long)SHARED_DRAM) / sizeof(data_t);
numfilterelems_t weights_per_filter = (layer.kernel == 3) ? 9 : 1;
weightaddr_t num_weights =
layer.channels_in * layer.channels_out * weights_per_filter;
// Print some Info on this Layer
printf("CPU: Offload CONV Layer ");
print_layer(&layer);
fflush(stdout);
// Offload Layer Calculation to FPGA
fpga_top(layer, (data_t *)SHARED_DRAM, weights_offset, num_weights,
input_offset);
LOG_LEVEL_DECR;
}
LOG_LEVEL = 0;
// ===============================
// = Copy Results back from FPGA =
// ===============================
layer_t *final = &net_CPU->layers[net_CPU->num_layers - 1];
int ch_out = (final->is_second_split_layer ? 2 : 1) * final->channels_out;
data_t *results = (data_t *)malloc(ch_out * sizeof(data_t));
copy_results_from_FPGA(net_CPU, results, ch_out);
// =====================
// = Calculate Softmax =
// =====================
std::vector<std::pair<data_t, int> > probabilities(ch_out);
calculate_softmax(net_CPU, results, probabilities);
// ==================
// = Report Results =
// ==================
printf("\nResult (top-5):\n====================\n");
for (int i = 0; i < std::min(5, ch_out); i++) {
printf(" %5.2f%%: class %3d (output %6.2f)\n",
100 * probabilities[i].first, probabilities[i].second,
results[probabilities[i].second]);
}
// ====================
// = TestBench Result =
// ====================
// Check if output is AS EXPECTED (+- 0.5%) (defined in network.hpp)
if (fabs(100 * probabilities[0].first - TEST_RESULT_EXPECTED) < 0.1) {
printf("\nTestBench Result: SUCCESS\n");
return 0;
} else {
printf("\nTestBench Result: FAILURE\n");
printf("Actual: %5.2f, Expected: %5.2f\n", 100 * probabilities[0].first,
TEST_RESULT_EXPECTED);
return -1;
}
}
int main(int argc, char **argv)
{
// Load database into memory. The file is not needed after this point.
rodb::Database db("example.rodb");
// Database::dump_yaml is helpful to see what's inside a compiled database.
if (argc == 2 && strcmp(argv[1], "--dump") == 0)
{
db.dump_yaml(std::cout);
}
// It's ok to make copies of values, they are very cheap.
rodb::Value root = db.root();
// Walk though an array.
rodb::Value layers = root["world"]["layers"];
for (size_t i = 0; i < layers.size(); ++i)
{
print_layer(layers[i]);
}
// Map element access is O(log N). If you don't care too much about performance it's
// ok to access same elements multiple times. It's fast enough. Otherwise it makes
// sense to store root["ball"] and root["ball"]["start_position"] into a temporary
// variable.
Point p1(root["ball"]["start_position"]["x"], root["ball"]["start_position"]["y"]);
Point p2(root["ball"]["start_position"]);
assert(p1.x == p2.x && p1.y == p2.y);
std::cout << "Point: {" << p1.x << ", " << p1.y << "}\n";
// Boolean values.
if (root["game"]["debug"])
{
std::cout << "Debug mode is on\n";
}
// Arrays don't have be uniform. They can contain mixed type values, including
// other arrays and maps.
rodb::Value allowed_levels = root["game"]["allowed_levels"];
for (size_t i = 0; i < allowed_levels.size(); ++i)
{
if (allowed_levels[i].is_int())
{
std::cout << "Level #" << (int)allowed_levels[i] << " is allowed\n";
}
else if (allowed_levels[i].is_string())
{
std::cout << "Level '" << (char const *)allowed_levels[i] << "' is allowed\n";
}
else if (allowed_levels[i].is_array() && allowed_levels[i].size() > 0)
{
std::cout << "Levels ";
rodb::Value group = allowed_levels[i];
for (size_t j = 0; j < group.size(); ++j)
{
if (j == group.size() - 1)
{
std::cout << " and ";
}
else if (j > 0)
{
std::cout << ",";
}
std::cout << (int)group[j];
}
std::cout << " are allowed\n";
}
}
return 0;
}
