Add missing ready, fix constant endianness

ChaCha20_Poly1305_64/sim/chacha20_pipelined_block.py

@@ -57,7 +57,7 @@ class TB:
         data_in.extend(struct.unpack("8I", key.to_bytes(32, "little")))
         data_in.extend(struct.unpack("2I", counter.to_bytes(8, "little")))
         data_in.extend(struct.unpack("2I", nonce.to_bytes(8, "little")))
-        
+
         data_out = chacha_block(data_in)
 
         await self.expected_queue.put(data_out)
@@ -72,6 +72,8 @@ class TB:
             self.dut.i_ready.value = 1
             self.dut.i_valid.value = 1
             await RisingEdge(self.dut.i_clk)
+            while not self.dut.o_ready.value == 1:
+                await RisingEdge(self.dut.i_clk)
             self.dut.i_valid.value = 0
 
     async def run_output(self):
@@ -91,7 +93,7 @@ async def test_sanity(dut):
 
     await tb.cycle_reset()
 
-    count = 1
+    count = 200
 
     for i in range(count):
         key = random.randint(0, 2**256-1)

ChaCha20_Poly1305_64/src/chacha20_pipelined_block.sv

@@ -46,7 +46,7 @@ always_ff @(posedge i_clk) begin
         initial_state_rptr <= '0;
         initial_state_wptr <= '0;
     end else begin
-        if (i_valid) begin
+        if (i_valid & o_ready) begin
             initial_states[initial_state_wptr] <= write_initial_state;
             if (initial_state_wptr < PIPE_STAGES-1) begin
                 initial_state_wptr <= initial_state_wptr + 1;
@@ -73,18 +73,18 @@ always_ff @(posedge i_clk) begin
         // 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
+        for (int i = 0; i < 16; i++) begin : final_add
             o_state[i*32 +: 32] <= state_pre_add[i*32 +: 32] + read_initial_state[i*32 +: 32];
         end
     end
 end
 
 always_comb begin
-    for (int i = 0; i < 4; i++) begin
-        state[0][32*(3-i) +: 32] = CONSTANT[32*(3-i) +: 32];    // constant is big endian
+    for (int i = 0; i < 4; i++) begin : constant
+        state[0][32*i +: 32] = CONSTANT[32*(3-i) +: 32];    // constant is big endian
     end
     
-    for (int i = 0; i < 8; i++) begin
+    for (int i = 0; i < 8; i++) begin : key
         state[0][32*(i+4) +: 32] = i_key[32*i +: 32];
     end
 
@@ -103,6 +103,8 @@ always_comb begin
     original_initial_state = read_initial_state;
 end
 
+assign ready[ROUNDS] = i_ready;
+
 
 generate
     for (genvar round = 0; round < ROUNDS; round+=2) begin : ROUND_LOOP

ChaCha20_Poly1305_64/src/chacha20_pipelined_round.sv

@@ -3,9 +3,7 @@ module chacha20_pipelined_round #(
     parameter KEY_SIZE = 256,
     parameter COUNTER_SIZE = 64,
     parameter NONCE_SIZE = 64,
-    parameter STATE_SIZE = 512,
-    parameter ROUNDS = 20,
-    parameter CONSTANT = 128'h657870616e642033322d62797465206b
+    parameter STATE_SIZE = 512
 )(
     input  logic                    i_clk,
     input  logic                    i_rst,
@@ -55,15 +53,15 @@ always_comb begin
             end
 
             1: begin
-                `QR_IN(1, 5, 9, 13);
+                `QR_IN(1, 6, 11, 12);
             end
 
             2: begin
-                `QR_IN(2, 6, 10, 14);
+                `QR_IN(2, 7, 8, 13);
             end
 
             3: begin
-                `QR_IN(3, 7, 11, 15);
+                `QR_IN(3, 4, 9, 14);
             end
         endcase
     end else begin