...

demo/lua2vgg16.lua

@@ -0,0 +1,50 @@
+loadfile("./util/lua2cc.lua")() 
+
+network = {
+    networkName  = "vgg16",
+    createScope  = "vgg16",
+    parameterLv  = 0,
+    inputLayers  = {"in"},
+    outputLayers = {"out"},
+    l1   = conv2d          ({input = "in",
+                              stride = 1, padding = 1}),
+    l2   = relu             ({}),
+    l3   = conv2d           ({stride = 1, padding = 1}),
+    l4   = relu             ({}),
+    l5   = maxPool2d        ({stride = 2}),
+    l6   = conv2d           ({stride = 1, padding = 1}),
+    l7   = relu             ({}),
+    l8   = conv2d           ({stride = 1, padding = 1}),
+    l9   = relu             ({}),
+    l10  = maxPool2d        ({stride = 2}),
+    l11  = conv2d           ({stride = 1, padding = 1}),
+    l12  = relu             ({}),
+    l13  = conv2d           ({stride = 1, padding = 1}),
+    l14  = relu             ({}),
+    l15  = conv2d           ({stride = 1, padding = 1}),
+    l16  = relu             ({}),
+    l17  = maxPool2d        ({stride = 2}),
+    l18  = conv2d           ({stride = 1, padding = 1}),
+    l19  = relu             ({}),
+    l20  = conv2d           ({stride = 1, padding = 1}),
+    l21  = relu             ({}),
+    l22  = conv2d           ({stride = 1, padding = 1}),
+    l23  = relu             ({}),
+    l24  = maxPool2d        ({stride = 2}),
+    l25  = conv2d           ({stride = 1, padding = 1}),
+    l26  = relu             ({}),
+    l27  = conv2d           ({stride = 1, padding = 1}),
+    l28  = relu             ({}),
+    l29  = conv2d           ({stride = 1, padding = 1}),
+    l30  = relu             ({}),
+    l31  = maxPool2d        ({stride = 2}),
+    l32  = reshape          ({shape = {-1, 1, 1}}),
+    l33  = fullyConnected   ({}),
+    l34  = relu             ({}),
+    l35  = fullyConnected   ({}),
+    l36  = relu             ({}),
+    l37  = fullyConnected   ({}),
+    out  = softmax          ({input = "l37"})
+}
+
+ccCodeTranslator(network, {file = "vgg16.c"})

makefile

@@ -15,7 +15,7 @@ DFLAG += # -g -fsanitize=address -fno-omit-frame-pointer
 CFLAG += # -std=c89
 CFLAG += -Wall # -Wpedantic
 
-OFLAG += -O3
+OFLAG += -O3 -march=native
 
 # Enable OpenMP
 OFLAG += -DENABLE_OPENMP -fopenmp
@@ -127,7 +127,7 @@ $(OBJS_PATH)/build:
 all: $(APPS)
 
 $(CATCOON_A): $(ALL_O)
-	cd $(OBJS_PATH) && $(AR) $@ $^ && $(MV) $@ ..
+	cd $(OBJS_PATH) && $(AR) $@ $(ALL_O) && $(MV) $@ ..
 
 %.o: ./src/%.c
 	$(CC) -c -o $(OBJS_PATH)/$@ $< $(CFLAG) $(INC)