Expanded micro-operation parameters (added dedicated control bit to force the B input of

the modular multiplier to 1, this is necessary to bring numbers out of Montgomery domain).

5 files changed, 57 insertions, 16 deletions

diff --git a/rtl/modexpng_core_top.v b/rtl/modexpng_core_top.v
index eb6826c..e117e5d 100644
--- a/rtl/modexpng_core_top.v
+++ b/rtl/modexpng_core_top.v
@@ -578,6 +578,9 @@ module modexpng_core_top
     reg  [BANK_ADDR_W -1:0] mmm_sel_narrow_in_x;
     reg  [BANK_ADDR_W -1:0] mmm_sel_narrow_in_y;
     
+    reg                     mmm_force_unity_b_x;
+    reg                     mmm_force_unity_b_y;
+    
     wire                    rdct_ena_x;
     wire                    rdct_ena_y;
     wire                    rdct_rdy_x;
@@ -594,6 +597,7 @@ module modexpng_core_top
         .ladder_mode                (mmm_ladder_mode_x),
         .word_index_last            (mmm_word_index_last_x),
         .word_index_last_minus1     (mmm_word_index_last_minus1_x),
+        .force_unity_b              (mmm_force_unity_b_x),
         
         .sel_wide_in                (mmm_sel_wide_in_x),
         .sel_narrow_in              (mmm_sel_narrow_in_x),
@@ -648,6 +652,7 @@ module modexpng_core_top
         .ladder_mode                (mmm_ladder_mode_y),
         .word_index_last            (mmm_word_index_last_y),
         .word_index_last_minus1     (mmm_word_index_last_minus1_y),
+        .force_unity_b              (mmm_force_unity_b_y),
         
         .sel_wide_in                (mmm_sel_wide_in_y),
         .sel_narrow_in              (mmm_sel_narrow_in_y),
@@ -812,7 +817,9 @@ module modexpng_core_top
 
     //
     // Parameters
-    //    
+    //
+    wire uop_aux_is_1 = uop_data_aux == UOP_AUX_1;
+        
     always @(posedge clk)
         //
         if (uop_fsm_state == UOP_FSM_STATE_DECODE) begin
@@ -844,11 +851,12 @@ module modexpng_core_top
                         UOP_LADDER_PQ: {mmm_ladder_mode_x, mmm_ladder_mode_y} <= 2'bXX;
                     endcase
                     //
-                    {mmm_sel_wide_in_x,     mmm_sel_wide_in_y    } <= {2{uop_data_sel_wide_in   }};
-                    {mmm_sel_narrow_in_x,   mmm_sel_narrow_in_y  } <= {2{uop_data_sel_narrow_in }};
-                    {rdct_sel_wide_out_x,   rdct_sel_wide_out_y  } <= {2{uop_data_sel_wide_out  }}; 
-                    {rdct_sel_narrow_out_x, rdct_sel_narrow_out_y} <= {2{uop_data_sel_narrow_out}};
-
+                    {mmm_force_unity_b_x,   mmm_force_unity_b_y  } <= {2{uop_aux_is_1 ? 1'b0 : 1'b1}};
+                    {mmm_sel_wide_in_x,     mmm_sel_wide_in_y    } <= {2{uop_data_sel_wide_in      }};
+                    {mmm_sel_narrow_in_x,   mmm_sel_narrow_in_y  } <= {2{uop_data_sel_narrow_in    }};
+                    {rdct_sel_wide_out_x,   rdct_sel_wide_out_y  } <= {2{uop_data_sel_wide_out     }}; 
+                    {rdct_sel_narrow_out_x, rdct_sel_narrow_out_y} <= {2{uop_data_sel_narrow_out   }};
+                    
                     //
                 end
                 //
diff --git a/rtl/modexpng_microcode.vh b/rtl/modexpng_microcode.vh
index af21391..1465c48 100644
--- a/rtl/modexpng_microcode.vh
+++ b/rtl/modexpng_microcode.vh
@@ -17,15 +17,30 @@ localparam UOP_W = UOP_OPCODE_W + UOP_CRT_W + UOP_NPQ_W + UOP_AUX_W + UOP_LADDER
 
 // OPCODE
 localparam [UOP_OPCODE_W -1:0] UOP_OPCODE_STOP               = 4'd0;
+/* all fields are don't care
+ */
 
 localparam [UOP_OPCODE_W -1:0] UOP_OPCODE_INPUT_TO_WIDE      = 4'd1;
 localparam [UOP_OPCODE_W -1:0] UOP_OPCODE_INPUT_TO_NARROW    = 4'd2;
+/* CRT tells into which of the dual MMM to write
+ * NPQ specifies the width of the operand
+ * AUX specifies from which INPUT to read
+ * LADDER is don't care
+ * source WIDE is always don't care
+ * destination NARROW is don't care for _WIDE opcode and vice versa
+ * 
+*/
 
 localparam [UOP_OPCODE_W -1:0] UOP_OPCODE_OUTPUT_FROM_NARROW = 4'd3;
 
 //localparam [UOP_OPCODE_W -1:0] UOP_OPCODE_COPY_CRT_Y2X  = 4'd0;
 
 localparam [UOP_OPCODE_W -1:0] UOP_OPCODE_MODULAR_MULTIPLY   = 4'd8;
+/* CRT is don't care
+ * NPQ specifies the width of the operand
+ * AUX = AUX_2 forces B input to 1 (AUX_1 reads from source NARROW as usual)
+ * LADDER specifies Montgomery ladder mode 
+ */
 
 // CRT
 localparam [UOP_CRT_W -1:0] UOP_CRT_X   = 1'b0;
diff --git a/rtl/modexpng_mmm_dual.v b/rtl/modexpng_mmm_dual.v
index babd565..b9b41e8 100644
--- a/rtl/modexpng_mmm_dual.v
+++ b/rtl/modexpng_mmm_dual.v
@@ -7,6 +7,7 @@ module modexpng_mmm_dual
     ladder_mode,
     word_index_last,
     word_index_last_minus1,
+    force_unity_b,
     
     sel_wide_in, sel_narrow_in,
     
@@ -70,7 +71,8 @@ module modexpng_mmm_dual
     input                   ladder_mode;
     input [7:0] word_index_last;
     input [7:0] word_index_last_minus1;
-    
+    input       force_unity_b;
+        
     input [BANK_ADDR_W-1:0] sel_wide_in;
     input [BANK_ADDR_W-1:0] sel_narrow_in;
     
@@ -708,13 +710,22 @@ module modexpng_mmm_dual
     // DSP Feed Logic
     //
     reg dsp_merge_xy_b;
+    reg dsp_merge_xy_b_first;
     
-    always @(posedge clk)
+    always @(posedge clk) begin
         //
         case (fsm_state)
             FSM_STATE_MULT_SQUARE_COL_0_TRIG:   dsp_merge_xy_b <= 1'b1;
             FSM_STATE_MULT_TRIANGLE_COL_0_TRIG: dsp_merge_xy_b <= 1'b0;
         endcase
+        //
+        case (fsm_state)
+            FSM_STATE_MULT_SQUARE_COL_0_TRIG,
+            FSM_STATE_MULT_SQUARE_COL_N_TRIG: dsp_merge_xy_b_first <= 1'b1;
+            default:                          dsp_merge_xy_b_first <= 1'b0;
+        endcase
+        //
+    end
 
     //
     // On-the-fly Carry Recombination
@@ -723,6 +734,9 @@ module modexpng_mmm_dual
     wire [17:0] rd_narrow_y_dout_carry = rd_narrow_y_dout + {{16{1'b0}}, dsp_xy_b_carry};
     wire [17:0] rd_narrow_xy_dout_carry_mux = ladder_mode ? rd_narrow_y_dout_carry : rd_narrow_x_dout_carry;
   
+    wire [15:0] rd_narrow_xy_dout_carry_mux_or_unity = !force_unity_b ?
+        rd_narrow_xy_dout_carry_mux[15:0] : dsp_merge_xy_b_first ? WORD_ONE : WORD_ZERO;
+  
     always @(posedge clk)
         //
         if (narrow_xy_ena_dly2) begin // rewrite
@@ -732,15 +746,15 @@ module modexpng_mmm_dual
                 dsp_y_b <= rd_narrow_y_dout[15:0];
                 dsp_xy_b_carry <= 2'b00;
             end else begin
-                dsp_x_b <= rd_narrow_xy_dout_carry_mux[15:0];
-                dsp_y_b <= rd_narrow_xy_dout_carry_mux[15:0];
+                dsp_x_b <= rd_narrow_xy_dout_carry_mux_or_unity;
+                dsp_y_b <= rd_narrow_xy_dout_carry_mux_or_unity;
                 dsp_xy_b_carry <= rd_narrow_xy_dout_carry_mux[17:16];
             end                 
             //
         end else begin
             //
-            dsp_x_b <= {16{1'bX}};
-            dsp_y_b <= {16{1'bX}};
+            dsp_x_b <= WORD_DNC;
+            dsp_y_b <= WORD_DNC;
             //
             dsp_xy_b_carry <= 2'b00;
             //
diff --git a/rtl/modexpng_parameters.vh b/rtl/modexpng_parameters.vh
index 25fafb4..e610e47 100644
--- a/rtl/modexpng_parameters.vh
+++ b/rtl/modexpng_parameters.vh
@@ -87,10 +87,12 @@ localparam [OP_ADDR_W-1:0] OP_ADDR_ZERO = {OP_ADDR_W{1'b0}};
 localparam [OP_ADDR_W-1:0] OP_ADDR_ONE = {{(OP_ADDR_W-1){1'b0}}, 1'b1};
 localparam [OP_ADDR_W-1:0] OP_ADDR_DNC = {OP_ADDR_W{1'bX}};
 
-localparam [WORD_W-1:0] WORD_NULL = {WORD_W{1'b0}};
-localparam [WORD_EXT_W-1:0] WORD_EXT_NULL = {WORD_EXT_W{1'b0}};
+localparam [WORD_W-1:0] WORD_ZERO = {WORD_W{1'b0}};
+localparam [WORD_W-1:0] WORD_DNC  = {WORD_W{1'bX}};
+localparam [WORD_W-1:0] WORD_ONE  = {{(WORD_W-1){1'b0}}, 1'b1};
 
-localparam [WORD_EXT_W-1:0] WORD_EXT_DNC = {WORD_EXT_W{1'bX}};
+localparam [WORD_EXT_W-1:0] WORD_EXT_ZERO = {WORD_EXT_W{1'b0}};
+localparam [WORD_EXT_W-1:0] WORD_EXT_DNC  = {WORD_EXT_W{1'bX}};
 
 localparam [MAC_INDEX_W-1:0] MAC_INDEX_DNC = {MAC_INDEX_W{1'bX}};
 
diff --git a/rtl/modexpng_uop_rom.v b/rtl/modexpng_uop_rom.v
index d0b6253..73b3142 100644
--- a/rtl/modexpng_uop_rom.v
+++ b/rtl/modexpng_uop_rom.v
@@ -29,7 +29,9 @@ module modexpng_uop_rom
             6'd10:      data <= {UOP_OPCODE_INPUT_TO_NARROW,  UOP_CRT_X,   UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_DNC, BANK_DNC,    BANK_IN_1_M,        BANK_DNC,       BANK_NARROW_E    };
             6'd11:      data <= {UOP_OPCODE_INPUT_TO_NARROW,  UOP_CRT_Y,   UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_DNC, BANK_DNC,    BANK_IN_1_M,        BANK_DNC,       BANK_NARROW_E    };
             
-            6'd12:      data <= {UOP_OPCODE_MODULAR_MULTIPLY, UOP_CRT_DNC, UOP_NPQ_N,   UOP_AUX_DNC, UOP_LADDER_11,  BANK_WIDE_A, BANK_NARROW_A,      BANK_WIDE_B,    BANK_NARROW_B    };
+            6'd12:      data <= {UOP_OPCODE_MODULAR_MULTIPLY, UOP_CRT_DNC, UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_11,  BANK_WIDE_A, BANK_NARROW_A,      BANK_WIDE_B,    BANK_NARROW_B    };
+            6'd13:      data <= {UOP_OPCODE_MODULAR_MULTIPLY, UOP_CRT_DNC, UOP_NPQ_N,   UOP_AUX_1,   UOP_LADDER_11,  BANK_WIDE_B, BANK_NARROW_B,      BANK_WIDE_C,    BANK_NARROW_C    };
+            6'd14:      data <= {UOP_OPCODE_MODULAR_MULTIPLY, UOP_CRT_DNC, UOP_NPQ_N,   UOP_AUX_2,   UOP_LADDER_11,  BANK_WIDE_C, BANK_DNC,           BANK_WIDE_D,    BANK_NARROW_D    };
             
             default:    data <= {UOP_OPCODE_STOP,             UOP_CRT_DNC, UOP_NPQ_DNC, UOP_AUX_DNC, UOP_LADDER_DNC, UOP_SEL_DNC_ALL};
         endcase