Adding hold register and bit combine to create final s_mem data with fused add and div.

1 files changed, 14 insertions, 5 deletions

diff --git a/src/rtl/montprod.v b/src/rtl/montprod.v
index 12260ac..4b93e57 100644
--- a/src/rtl/montprod.v
+++ b/src/rtl/montprod.v
@@ -141,6 +141,10 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
   reg                  test_reg;
   reg                  test_new;
 
+  reg [(OPW - 2) : 0]  shr_data_out_reg;
+  reg                  shr_carry_out_reg;
+  reg                  shr_carry_out_new;
+
 
   //----------------------------------------------------------------
   // Wires.
@@ -167,6 +171,7 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
 
   reg [(OPW - 1) : 0]  sa_adder_data_in;
   reg [(OPW - 1) : 0]  muxed_s_mem_read_data;
+  reg [(OPW - 1) : 0]  shifted_s_mem_write_data;
 
   wire [(ADW - 1) : 0] length_m1;
 
@@ -221,7 +226,7 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
                                );
 
   shr #(.OPW(OPW)) shifter(
-                           .a(s_mem_read_data),
+                           .a(shr_data_in),
                            .carry_in(shr_carry_in_reg),
                            .adiv2(shr_data_out),
                            .carry_out(shr_carry_out)
@@ -246,6 +251,7 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
           word_index_prev_reg <= {ADW{1'b0}};
           add_carry_in_sa_reg <= 1'b0;
           add_carry_in_sm_reg <= 1'b0;
+          shr_data_out_reg    <= {(OPW - 1){1'b0}};
           shr_carry_in_reg    <= 1'b0;
           b_reg               <= 1'b0;
           q_reg               <= 1'b0;
@@ -269,6 +275,7 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
           shr_carry_in_reg    <= shr_carry_in_new;
           add_carry_in_sa_reg <= add_carry_in_sa_new;
           add_carry_in_sm_reg <= add_carry_in_sm_new;
+          shr_data_out_reg    <= shr_data_out[(OPW - 2) : 0];
 
           if (word_index_we)
             word_index_reg <= word_index_new;
@@ -321,7 +328,6 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
       tmp_s_mem_write_data  = {OPW{1'b0}};
       test_new              = 1'b0;
 
-
       case (montprod_ctrl_reg)
         CTRL_LOOP_ITER:
           begin
@@ -339,7 +345,7 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
         CTRL_CALC_SDIV2:
           begin
             //s = (s + q*M + b*A) >>> 1; s>>=1.  Takes (1..length) cycles.
-            s_mem_we_new  = 1'b1;
+            s_mem_we_new = 1'b1;
           end
 
         default:
@@ -364,18 +370,21 @@ module montprod #(parameter OPW = 32, parameter ADW = 8)
             else
               sa_adder_data_in = muxed_s_mem_read_data;
 
+
             if (b_reg)
               tmp_s_mem_write_data = add_result_sa;
             else if (q_reg)
               tmp_s_mem_write_data = add_result_sm;
             else if (first_iteration_reg)
               tmp_s_mem_write_data = {OPW{1'b0}};
-            else
-              tmp_s_mem_write_data = s_mem_read_data;
 
             s_mem_write_data = tmp_s_mem_write_data;
             add_carry_in_sa_new = add_carry_out_sa;
             add_carry_in_sm_new = add_carry_out_sm;
+
+
+            shr_data_in              = tmp_s_mem_write_data;
+            shifted_s_mem_write_data = {shr_carry_out ,shr_data_out_reg};
           end
 
         SMUX_SHR: