debugging pipeline

scripts/neuron_net_test.py

@@ -13,15 +13,16 @@ WEIGTHS = [
         [-0.16126093, 0.5317601, 0.34316933, -0.7074082],
         [0.09219088, -0.624525, -0.61903083, -0.87057704]
     ],
-    [[0.36770403, -0.78046024, 0.3979908, 0.5494289, -0.13859335, 0.40053025, 0.08249452, -0.32528356],
-     [-0.17659009, 0.13901198, -0.45248222, -0.7894139, -0.81092286, -0.521815, 0.30632392, -0.3143816],
-     [-0.04314173, 0.14361085, 0.6259473, 0.3571782, -0.38011226, 0.01378736, 0.05794358, 0.09667788],
-     [-0.46864474, 0.36618456, -0.45595396, -0.39789405, 0.73964316, -0.30294785, 0.2482118, -0.2127953],
-     [-0.37941265, 0.45330787, -0.12066315, 0.5636705, 0.68990386, 0.6543718, 0.86367106, -0.5707757],
-     [-0.78606385, 0.24032554, -0.4472755, -0.24661142, -0.2698564, -0.8365823, -0.13674814, -0.39799848],
-     [0.11138931, 0.48950365, 0.12998834, 0.4947537, 0.516593, 0.82281274, 0.04789656, 0.30206403],
-     [0.23097174, 0.30290592, -0.596446, -0.40108407, 0.12246455, -0.47260976, -0.55030185, 0.44481543]
-     ]
+    [
+        [0.36770403, -0.78046024, 0.3979908, 0.5494289, -0.13859335, 0.40053025, 0.08249452, -0.32528356],
+        [-0.17659009, 0.13901198, -0.45248222, -0.7894139, -0.81092286, -0.521815, 0.30632392, -0.3143816],
+        [-0.04314173, 0.14361085, 0.6259473, 0.3571782, -0.38011226, 0.01378736, 0.05794358, 0.09667788],
+        [-0.46864474, 0.36618456, -0.45595396, -0.39789405, 0.73964316, -0.30294785, 0.2482118, -0.2127953],
+        [-0.37941265, 0.45330787, -0.12066315, 0.5636705, 0.68990386, 0.6543718, 0.86367106, -0.5707757],
+        [-0.78606385, 0.24032554, -0.4472755, -0.24661142, -0.2698564, -0.8365823, -0.13674814, -0.39799848],
+        [0.11138931, 0.48950365, 0.12998834, 0.4947537, 0.516593, 0.82281274, 0.04789656, 0.30206403],
+        [0.23097174, 0.30290592, -0.596446, -0.40108407, 0.12246455, -0.47260976, -0.55030185, 0.44481543]
+    ],
     [
         [0.5724262, 0.5853241, 0.3748752, -0.892384, -1.0270239, 0.2170913, -0.07271451, 0.14661156],
         [0.30391088, -0.92324615, 0.8088594, -1.0522624, 0.07374455, -0.550893, 0.8194236, -0.62796086]
@@ -36,10 +37,64 @@ BIAS = [
     [0.03787775, -0.03655371],
 ]
 
+RESULT = [
+    0x00000000, 0x3f800000,
+    0x3f800000, 0x00000000,
+    0x00000000, 0x00000000,
+    0x3f800000, 0x3f800000,
+    # 0x3f030126, 0x3f800000,
+    # 0x3e652010, 0x00000000,
+    # 0x3d2d25da, 0x3efea470,
+    # 0x3f1a4267, 0x3f06d0f4,
+]
+
 
 def generate_nn_values(dtype=np.float32):
     dsize = dtype_size(dtype)
 
 
+def sigm(x):
+    ex = 2.7182818 ** x
+    return ex / (ex + 1)
+
+
+def linr(x):
+    return x
+
+
+def process(value):
+    if value > 3 or value < 0:
+        raise ValueError("Value not in between 0 and 3")
+    onehot_mat = [
+        [0, 0, 0, 1],
+        [0, 0, 1, 0],
+        [0, 1, 0, 0],
+        [1, 0, 0, 0],
+    ]
+    L0 = onehot_mat[value]
+    L1_mult = [[L0[j] * WEIGTHS[0][i][j] for j in range(len(L0))] for i in range(8)]
+    L1 = [linr(sum(L1_mult[i]) + BIAS[0][i]) for i in range(8)]
+    L2 = [linr(sum([L1[j] * WEIGTHS[1][i][j] for j in range(len(L1))]) + BIAS[1][i]) for i in range(8)]
+    L3 = [sum([L2[j] * WEIGTHS[2][i][j] for j in range(len(L2))]) + BIAS[2][i] for i in range(2)]
+    # print(f"L0: {L0} \n\tL1mult: {L1_mult}\n\tL1: {L1} \n\tL2: {L2} \n\tL3: {L3}")
+    return [sigm(v) for v in L3]
+
+
+def show_results():
+    sim_data = np.array([process(0), process(1), process(2), process(3)]).T
+    hdl_data = np.frombuffer(b''.join([i.to_bytes(4, 'little') for i in RESULT]), dtype=np.float32).reshape(4, 2).T
+    print(hdl_data)
+
+    plt.plot(sim_data[0], sim_data[1], 'x', color='b')
+    plt.plot(hdl_data[0], hdl_data[1], 'x', color='r')
+    plt.xlabel('I')
+    plt.ylabel('Q')
+    # plt.xlim([0, 1])
+    # plt.ylim([0, 1])
+    plt.grid()
+    plt.show()
+
+
 if __name__ == '__main__':
-    generate_nn_values()
+    show_results()
+    # generate_nn_values()

scripts/number_conv.py

@@ -7,7 +7,7 @@ def float2verilog(lines, dtype=np.float32):
     dsize = dtype_size(dtype)
     print("")
     for i, line in enumerate(lines):
-        arr = line.replace('[', '').replace(']', '').split()
+        arr = line.replace('[', '').replace(']', '').replace('#', '').replace('/', '').split()
         nums = np.array([float(f.strip(',')) for f in arr if f], dtype=dtype)
         b = nums.tobytes()
         print(f'[{i}] = {{' +

src/FPA_module_test.sv

@@ -4,10 +4,10 @@ module floating_add #(parameter N=16, M=4)(input_1, input_2, sum, diff, clk, res
 	output logic [N-1:0] sum;
 	output logic [M:0] diff;
 
-//	logic flag_a;
-//	logic flag_b;
-//	logic [M:0] abs;
-//	logic [N-3-M:0] res;
+	logic flag_a;
+	logic flag_b;
+	logic [M:0] abs;
+	logic [N-3-M:0] res;
 	
 	logic [N-1:0] D0 [7:0];
 	logic [N-1:0] Q0 [7:0];
@@ -30,7 +30,7 @@ module floating_add #(parameter N=16, M=4)(input_1, input_2, sum, diff, clk, res
 			D0[6] = 0; // abs
 			D0[7] = 0; // res
 		end
-	pipe#(.N(N-1), .K(7)) pipe0(.clk(clk), .reset(reset), .D(D0), .Q(Q0));
+	pipe pipe0(.clk(clk), .reset(reset), .D(D0), .Q(Q0));
 	
 	
 	always_comb
@@ -72,7 +72,7 @@ module floating_add #(parameter N=16, M=4)(input_1, input_2, sum, diff, clk, res
 		end
 		
 		//Second pipeline stage 1
-		pipe#(.N(N-1), .K(7)) pipe1(.clk(clk), .reset(reset), .D(Q0), .Q(Q1));
+		pipe pipe1(.clk(clk), .reset(reset), .D(Q0), .Q(Q1));
 		
 	always_comb
 		begin
@@ -101,7 +101,7 @@ module floating_add #(parameter N=16, M=4)(input_1, input_2, sum, diff, clk, res
 							else
 								begin
 									Q1[7] = Q1[0][N-3-M:0] - (Q1[1][N-3-M:0] >> Q1[6]-1); // Subtract the mantissas
-									Q1[2][N-3-M:0] = Q1[7];
+									sum[N-3-M:0] = res;
 								end
 						end
 					else if (~Q1[4] & Q1[5])
@@ -145,29 +145,30 @@ module floating_add #(parameter N=16, M=4)(input_1, input_2, sum, diff, clk, res
 		end
 		
 		// Final pipeline stage 
-		pipe#(.N(N-1), .K(7)) pipe2(.clk(clk), .reset(reset), .D(Q1), .Q(Q2));
+		pipe pipe2(.clk(clk), .reset(reset), .D(Q1), .Q(Q2));
 		assign sum = Q2[2];
 		assign diff = Q2[3];
 endmodule : floating_add
 
 
 
-module floating_product #(parameter N=16, M=4)(input_1, input_2, product);
+module floating_product #(parameter N=16, M=4)(input_1, input_2, product, clk, reset);
 	input logic [N-1:0] input_1, input_2;
+	input logic clk, reset;
 	output logic [N-1:0] product;
 
 	// sign_x = x[N-1]
 	// exponent_x = x[N-2:N-2-M]
 	// mantissa_x = x[N-3-M:0]
 
-//	logic [N-2:N-2-M] sum;
-//	logic [2*(N-3-M):0] mult;
+	logic [N-2:N-2-M] sum;
+	logic [2*(N-3-M):0] mult;
 	logic [2*(N-3-M):0] D0 [4:0];
 	logic [2*(N-3-M):0] Q0 [4:0];
 	logic [2*(N-3-M):0] Q1 [4:0];
 	logic [2*(N-3-M):0] Q2 [4:0];
 	
-	//First pipeline stage
+	// First pipeline stage
 	always_comb
 		begin
 			D0[0] = input_1;
@@ -176,20 +177,21 @@ module floating_product #(parameter N=16, M=4)(input_1, input_2, product);
 			D0[3] = 0; // sum
 			D0[4] = 0; // mult
 		end
-	pipe#(.N(2*N-3-M), .K(4)) pipe0(.clk(clk), .reset(reset), .D(D0), .Q(Q0));
+		
+	pipe pipe0(.clk(clk), .reset(reset), .D(D0), .Q(Q0));
 
 	// We have assigned an {M+1} bit exponent so we must have a 2^{M} offset
 	assign Q0[3] = Q0[0][N-2:N-2-M] + Q0[1][N-2:N-2-M];
 	assign Q0[2][N-2:N-2-M] = Q0[3] - (1'b1 << M) + 2;
 	
-	//Second Pipeline stage
-	pipe#(.N(2*N-3-M), .K(4)) pipe1(.clk(clk), .reset(reset), .D(Q0), .Q(Q1));
+	// Second pipeline stage
+	pipe pipe1(.clk(clk), .reset(reset), .D(Q0), .Q(Q1));
 
 	always_comb
 		begin
 				// Setting the mantissa of the output
 				Q1[4] = Q1[0][N-3-M:0] * Q1[1][N-3-M:0];
-				if (Q1[4][N-3-M]) Q0[2][N-3-M:0] = Q1[4][2*(N-3-M):2*(N-3-M)-9];
+				if (Q1[4][N-3-M]) Q1[2][N-3-M:0] = Q1[4][2*(N-3-M):2*(N-3-M)-9];
 				else Q1[2][N-3-M:0] = Q1[4][2*(N-3-M):2*(N-3-M)-9] << 1;
 				Q1[2][N-1] = Q1[0][N-1] ^ Q1[1][N-1];
 		end
@@ -201,24 +203,22 @@ endmodule : floating_product
 
 
 
-module pipe #(parameter N, K)(clk, reset, Q, D);
+module pipe #(parameter N=16)(clk, reset, Q, D);
 	input logic clk, reset;
-	input reg [N:0] D [K:0];
-	output reg [N:0] Q [K:0];
-	logic [N:0] in_pipe [K:0];
+	input logic [N-1:0] D;
+	output reg [N-1:0] Q;
+	reg [N-1:0] in_pipe;
 	
-	always_ff @(posedge clk)
+	always @(posedge clk or negedge reset)
 		begin
-			if(reset) 
-				begin
-					in_pipe <= 0;
-					Q <= 0;
-				end
-			else 
-				begin
-					in_pipe <= D;
-					Q <= in_pipe;
-				end
+			if(reset) in_pipe = 0;
+			else in_pipe = D;
+		end
+	
+	always @(posedge clk or negedge reset)
+		begin
+			if(reset) Q = 0;
+			else Q = in_pipe;
 		end
 endmodule : pipe
 
@@ -229,9 +229,9 @@ module floating_tb;
 	logic [15:0] input_a, input_b, result_add, result_mult;
 	logic [4:0] diff;
 
-	floating_add adder1(.input_1(input_a), .input_2(input_b), .sum(result_add), .diff(diff), .clk(clk), .reset(reset));
+	floating_add adder1(.input_1(input_a), .input_2(input_b), .sum(result_add), .diff(diff));
 
-	floating_product multiplier1(.input_1(input_a), .input_2(input_b), .product(result_mult), .clk(clk), .reset(reset));
+	floating_product multiplier1(.input_1(input_a), .input_2(input_b), .product(result_mult));
 
 
 	reg [15:0] test_mem [29:0][3:0];

src/neural/layer.sv

@@ -20,15 +20,15 @@ module neuron_layer#(parameter C, K, N=32)(clk, rst, x, y, w, b, left, right);
 
     genvar i, j;
     generate
-        for(i=0; i<CONNS; i++) begin
+        for(i=0; i<CONNS; i++) begin: gen_signal_conn
             assign stb[i] = left[i].stb;
             assign left[i].ack = &ack_t[i];
         end
     endgenerate
 
     generate
-        for(i=0; i<NEURONS; i++) begin
-            for(j=0; j<CONNS; j++) begin
+        for(i=0; i<NEURONS; i++) begin: gen_neruons
+            for(j=0; j<CONNS; j++) begin: gen_cross
                 assign ack_t[j][i] = ack[i][j];
             end
             neuron#(.K(C), .N(N)) n(
@@ -68,21 +68,24 @@ module neuron_network_tb;
     abus_io left[3:0]();
     abus_io right[1:0]();
 
-    reg [31:0] layer1_w [7:0][3:0];
-    reg [31:0] layer1_b [7:0];
+    reg [31:0] layer1_w [0:7][3:0];
+    reg [31:0] layer1_b [0:7];
     reg [31:0] layer1_o [7:0];
     abus_io layer1_io [7:0]();
 
-    reg [31:0] layer2_w [7:0][7:0];
+    reg [31:0] layer2_w [0:7][7:0];
     reg [31:0] layer2_b [7:0];
     reg [31:0] layer2_o [7:0];
     abus_io layer2_io [7:0]();
 
-    reg [31:0] layer3_w [1:0][7:0];
+    reg [31:0] layer3_w [0:1][7:0];
     reg [31:0] layer3_b [1:0];
     reg [31:0] layer3_o [1:0];
     abus_io layer3_io [1:0]();
 
+    logic y_stb;
+    assign y_stb = right[0].stb & right[1].stb;
+
     neuron_layer#(.C(2), .K(3)) layer1(
         .clk(clk),
         .rst(rst),
@@ -174,19 +177,34 @@ module neuron_network_tb;
 
         #15;
         rst = 0;
-        x = {0, 0, 0 , 'h3f800000};
+        x = {0, 0, 0, 'h3f800000};
+        read_value();
+        x = {0, 0, 'h3f800000, 0};
+        read_value();
+        x = {0, 'h3f800000, 0, 0};
+        read_value();
+        x = {'h3f800000, 0, 0, 0};
+        read_value();
+    end
+
+    task read_value;
         left[0].stb = 1;
         left[1].stb = 1;
         left[2].stb = 1;
         left[3].stb = 1;
-
         #15;
         left[0].stb = 0;
         left[1].stb = 0;
         left[2].stb = 0;
         left[3].stb = 0;
-
-    end
+        wait(y_stb == 1);
+        right[0].ack = 1;
+        right[1].ack = 1;
+        #15;
+        right[0].ack = 0;
+        right[1].ack = 0;
+        $display("0x%H, 0x%H", y[0], y[1]);
+    endtask : read_value
 
 
 endmodule : neuron_network_tb