biquad.vhd: multiply and resize in different cycles

The long combinatorial logic to implement resize chained with DPS48s may make timing closure difficult. So, break these operations in two different cycles.
lnls-dig · Mar 28, 2024 · 80ce66e · 80ce66e
1 parent 25f90e6
commit 80ce66e
Showing 1 changed file with 46 additions and 22 deletions.
diff --git a/modules/common/biquad/biquad.vhd b/modules/common/biquad/biquad.vhd
@@ -85,7 +85,12 @@ ARCHITECTURE behave OF biquad IS
                   -(g_COEFF_FRAC_WIDTH + g_X_FRAC_WIDTH + g_EXTRA_BITS)) :=
              (OTHERS => '0');
 
-  SIGNAL state : NATURAL RANGE 0 TO 3 := 0;
+  SIGNAL b0_times_w_tmp, a1_times_w_d1_tmp, a2_times_w_d2_tmp,
+         b1_times_w_d1_tmp, b2_times_w_d2_tmp :
+           SFIXED((g_COEFF_INT_WIDTH + w'LEFT + 1)-1 DOWNTO
+                  -g_COEFF_FRAC_WIDTH + w'RIGHT) := (OTHERS => '0');
+
+  SIGNAL state : NATURAL RANGE 0 TO 4 := 0;
 BEGIN
   PROCESS(clk_i) IS
   BEGIN
@@ -101,50 +106,68 @@ BEGIN
         b2_times_w_d2 <= (OTHERS => '0');
         aux_a <= (OTHERS => '0');
         aux_b <= (OTHERS => '0');
+        b0_times_w_tmp <= (OTHERS => '0');
+        a1_times_w_d1_tmp <= (OTHERS => '0');
+        a2_times_w_d2_tmp <= (OTHERS => '0');
+        b1_times_w_d1_tmp <= (OTHERS => '0');
+        b2_times_w_d2_tmp <= (OTHERS => '0');
         state <= 0;
         y_valid_o <= '0';
       ELSE
         y_valid_o <= '0';
 
         CASE state IS
-          -- Computes: w[n] = x[n] - a1*w[n - 1] - a2*w[n - 2]
+          -- Computes: w[n] = x[n] - a1*w[n - 1] - a2*w[n - 2] (resized)
           WHEN 0 =>
             IF x_valid_i THEN
               w <= resize(x_i + aux_a, w'LEFT, w'RIGHT);
               state <= 1;
             END IF;
 
-          -- Computes: b0*w[n]
-          --           w[n - 1] (for the next iteration)
-          --           w[n - 2] (for the next iteration)
+          -- Computes: b0*w[n] (full precision)
+          --           w[n - 1] for the next iteration
+          --           w[n - 2] for the next iteration
           WHEN 1 =>
-            b0_times_w <= resize(coeffs_i.b0*w, b0_times_w'LEFT,
-                                 b0_times_w'RIGHT);
+            b0_times_w_tmp <= coeffs_i.b0*w;
             w_d1 <= w;
             w_d2 <= w_d1;
             state <= 2;
 
-          -- Computes: y[n] = w[n] + b1*w[n - 1] + b2*w[n - 2]
-          --           a1*w[n - 1] (for the next iteration)
-          --           a2*w[n - 2] (for the next iteration)
-          --           b1*w[n - 1] (for the next iteration)
-          --           b2*w[n - 2] (for the next iteration)
+          -- Computes: b0*w[n] (resized)
+          --           a1*w[n - 1] for the next iteration (full precision)
+          --           a2*w[n - 2] for the next iteration (full precision)
+          --           b1*w[n - 1] for the next iteration (full precision)
+          --           b2*w[n - 2] for the next iteration (full precision)
           WHEN 2 =>
-            y_o <= resize(b0_times_w + aux_b, y_o'LEFT, y_o'RIGHT);
-            y_valid_o <= '1';
-            a1_times_w_d1 <= resize(coeffs_i.a1*w_d1, a1_times_w_d1'LEFT,
+            b0_times_w <= resize(b0_times_w_tmp, b0_times_w'LEFT,
+                                 b0_times_w'RIGHT);
+            a1_times_w_d1_tmp <= coeffs_i.a1*w_d1;
+            a2_times_w_d2_tmp <= coeffs_i.a2*w_d2;
+            b1_times_w_d1_tmp <= coeffs_i.b1*w_d1;
+            b2_times_w_d2_tmp <= coeffs_i.b2*w_d2;
+            state <= 3;
+
+          -- Computes:  a1*w[n - 1] for the next iteration (resized)
+          --            a2*w[n - 2] for the next iteration (resized)
+          --            b1*w[n - 1] for the next iteration (resized)
+          --            b2*w[n - 2] for the next iteration (resized)
+          --            y[n] = w[n] + b1*w[n - 1] + b2*w[n - 2] (resized)
+          WHEN 3 =>
+            a1_times_w_d1 <= resize(a1_times_w_d1_tmp, a1_times_w_d1'LEFT,
                                     a1_times_w_d1'RIGHT);
-            a2_times_w_d2 <= resize(coeffs_i.a2*w_d2, a2_times_w_d2'LEFT,
+            a2_times_w_d2 <= resize(a2_times_w_d2_tmp, a2_times_w_d2'LEFT,
                                     a2_times_w_d2'RIGHT);
-            b1_times_w_d1 <= resize(coeffs_i.b1*w_d1, b1_times_w_d1'LEFT,
+            b1_times_w_d1 <= resize(b1_times_w_d1_tmp, b1_times_w_d1'LEFT,
                                     b1_times_w_d1'RIGHT);
-            b2_times_w_d2 <= resize(coeffs_i.b2*w_d2, b2_times_w_d2'LEFT,
+            b2_times_w_d2 <= resize(b2_times_w_d2_tmp, b2_times_w_d2'LEFT,
                                     b2_times_w_d2'RIGHT);
-            state <= 3;
+            y_o <= resize(b0_times_w + aux_b, y_o'LEFT, y_o'RIGHT);
+            y_valid_o <= '1';
+            state <= 4;
 
-          -- Computes: -a1*w[n - 1] - a2*w[n - 2] (for the next iteration)
-          --            b1*w[n - 1] + b2*w[n - 2] (for the next iteration)
-          WHEN 3 =>
+          -- Computes: -a1*w[n - 1] - a2*w[n - 2] for the next iteration (resized)
+          --            b1*w[n - 1] + b2*w[n - 2] for the next iteration (resized)
+          WHEN 4 =>
             aux_a <= resize(-a1_times_w_d1 - a2_times_w_d2, aux_a'LEFT,
                             aux_a'RIGHT);
             aux_b <= resize(b1_times_w_d1 + b2_times_w_d2, aux_b'LEFT,
@@ -158,5 +181,6 @@ BEGIN
   busy_o <= '1' WHEN (state = 0 AND x_valid_i = '1') ELSE
             '1' WHEN state = 1 ELSE
             '1' WHEN state = 2 ELSE
+            '1' WHEN state = 3 ELSE
             '0';
 END ARCHITECTURE behave;