Add correct amount of memory

should be 160, not 20. There are 8 cycles per stage and 20 stages

ChaCha20_Poly1305_64/src/chacha20_block.sv

@@ -21,6 +21,9 @@ module chacha20_block #(
     input  logic                    i_ready
 );
 
+// each round is 8 stages
+localparam PIPE_STAGES = ROUNDS * 8;
+
 `define QR(name, i, n, a, b, c, d)  \
 chacha20_qr u_chacha20_``name (     \
     .i_clk      (i_clk),            \
@@ -49,9 +52,9 @@ logic [3:0] valid[ROUNDS+1];
 
 // small fifo for storing the initial state.
 // better to store it in a memory than in flops
-logic [$clog2(ROUNDS)-1:0] initial_state_wptr;
-logic [$clog2(ROUNDS)-1:0] initial_state_rptr;
-logic [511:0] initial_states [ROUNDS];
+logic [$clog2(PIPE_STAGES)-1:0] initial_state_wptr;
+logic [$clog2(PIPE_STAGES)-1:0] initial_state_rptr;
+logic [511:0] initial_states [PIPE_STAGES];
 
 logic [511:0] state_pre_add;
 logic pre_add_valid;
@@ -67,12 +70,22 @@ always_ff @(posedge i_clk) begin
     end else begin
         if (i_valid) begin
             initial_states[initial_state_wptr] <= write_initial_state;
+            if (initial_state_wptr < PIPE_STAGES-1) begin
+                initial_state_wptr <= initial_state_wptr + 1;
+            end else begin
+                initial_state_wptr <= '0;
+            end
         end
 
         pre_add_valid <= valid[ROUNDS][0];
 
         if (valid[ROUNDS][0]) begin
             read_initial_state <= initial_states[initial_state_rptr];
+            if (initial_state_rptr < PIPE_STAGES-1) begin
+                initial_state_rptr <= initial_state_rptr + 1;
+            end else begin
+                initial_state_rptr <= '0;
+            end
             for (int i = 0; i < 16; i++) begin
                 state_pre_add[i*32 +: 32] <= state[ROUNDS][i];
             end