Get sim working, make some changes to the final addition

ChaCha20_Poly1305_64/src/chacha20_block.sv

@@ -3,6 +3,7 @@ module chacha20_block #(
     parameter COUNTER_SIZE = 64,
     parameter NONCE_SIZE = 64,
     parameter STATE_SIZE = 512,
+    parameter ROUNDS = 20,
     parameter CONSTANT = 128'h657870616e642033322d62797465206b
 )(
     input  logic                    i_clk,
@@ -41,16 +42,19 @@ chacha20_qr u_chacha20_``name (     \
 )
 
 
-logic [31:0] state [21][16];
-logic [3:0] valid[21];
+logic [31:0] state [ROUNDS+1][16];
+logic [3:0] valid[ROUNDS+1];
 // logic [3:0] ready[21];
 
 
 // small fifo for storing the initial state.
 // better to store it in a memory than in flops
-logic [4:0] initial_state_wptr;
-logic [4:0] initial_state_rptr;
-logic [511:0] initial_states [20];
+logic [$clog2(ROUNDS)-1:0] initial_state_wptr;
+logic [$clog2(ROUNDS)-1:0] initial_state_rptr;
+logic [511:0] initial_states [ROUNDS];
+
+logic [511:0] state_pre_add;
+logic pre_add_valid;
 
 logic [511:0] write_initial_state, read_initial_state;
 
@@ -65,14 +69,23 @@ always_ff @(posedge i_clk) begin
             initial_states[initial_state_wptr] <= write_initial_state;
         end
 
-        if (valid[19][0]) begin
+        pre_add_valid <= valid[ROUNDS][0];
+
+        if (valid[ROUNDS][0]) begin
             read_initial_state <= initial_states[initial_state_rptr];
+            for (int i = 0; i < 16; i++) begin
+                state_pre_add[i*32 +: 32] <= state[ROUNDS][i];
+            end
         end
 
 
-            o_valid <= &valid[20];
+        o_valid <= pre_add_valid;
+
+        // We cannot just add state_pre_add and read_initial state
+        // because the addition needs to be done wordwise, with no
+        // carries between 32 bit groups.
         for (int i = 0; i < 16; i++) begin
-            o_state[i*32 +: 32] <= state[20][i] + read_initial_state[i*32 +: 32];
+            o_state[i*32 +: 32] <= state_pre_add[i*32 +: 32] + read_initial_state[i*32 +: 32];
         end
     end
 
@@ -111,7 +124,7 @@ end
 
 
 generate
-    for (genvar round = 0; round < 20; round+=2) begin : ROUND_LOOP
+    for (genvar round = 0; round < ROUNDS; round+=2) begin : ROUND_LOOP
         `QR(0, round, 0, 0, 4, 8,  12);
         `QR(1, round, 1, 1, 5, 9,  13);
         `QR(2, round, 2, 2, 6, 10, 14);