Added input data, matched 256-cycle Bayer CLT in all modes

dsp/mclt16x16.v

@@ -550,7 +550,8 @@ D11 - negate for mode 3 (SS)
         // are these shift OK? Will need to be valis only @ dtt_start_out
         .shift_h       (x_shft_r4),     // input[6:0] signed 
         .shift_v       (y_shft_r4),     // input[6:0] signed
-        .inv_checker    (1'b0),          // input only used for Bayer mosaic data 
+//        .inv_checker    (1'b0),          // input only used for Bayer mosaic data 
+        .inv           (3'b0),          // input only used for Bayer mosaic data 
         .fd_din        (dtt_rd_data),   // input[24:0] signed. Expected latency = 3 from start  
         .fd_out        (dout),          // output[24:0] reg signed 
         .pre_first_out (pre_first_out), // output reg 

dsp/mclt16x16_bayer.v

@@ -458,7 +458,8 @@ module  mclt16x16_bayer#(
         // are these shift OK? Will need to be valis only @ dtt_start_out
         .shift_h       (x_shft_r5),     // input[6:0] signed 
         .shift_v       (y_shft_r5),     // input[6:0] signed
-        .inv_checker   (inv_checker_r5),// input only used for Bayer mosaic data 
+//        .inv_checker   (inv_checker_r5),// input only used for Bayer mosaic data 
+        .inv           ({inv_checker_r5,1'b0,inv_checker_r5}),// input only used for Bayer mosaic data 
         .fd_din        (dtt_rd_data0),  // input[24:0] signed. Expected latency = 3 from start  
         .fd_out        (dout0),         // output[24:0] reg signed 
         .pre_first_out (pre_first_out), // output reg 
@@ -480,7 +481,9 @@ module  mclt16x16_bayer#(
         // are these shift OK? Will need to be valis only @ dtt_start_out
         .shift_h       (x_shft_r5),     // input[6:0] signed 
         .shift_v       (y_shft_r5),     // input[6:0] signed
-        .inv_checker   (inv_checker_r5),// input only used for Bayer mosaic data 
+//        .inv_checker   (inv_checker_r5),// input only used for Bayer mosaic data
+        .inv           ({inv_checker_r5,1'b0,inv_checker_r5}),// input only used for Bayer mosaic data 
+         
         .fd_din        (dtt_rd_data1),  // input[24:0] signed. Expected latency = 3 from start  
         .fd_out        (dout1),         // output[24:0] reg signed 
         .pre_first_out (),              // output reg 

dsp/mclt16x16_bayer3.v

@@ -53,7 +53,8 @@ module  mclt16x16_bayer3#(
     parameter DSP_B_WIDTH =     18, // signed, output from sin/cos ROM
     parameter DSP_A_WIDTH =     25,
     parameter DSP_P_WIDTH =     48,
-    parameter DEAD_CYCLES =     14  // start next block immedaitely, or with longer pause
+    parameter DEAD_CYCLES =     14,  // start next block immedaitely, or with longer pause
+    parameter OUTS_AT_ONCE =     1  // 0: outputs with lowest latency, 1: all at once (with green)
 )(
     input                             clk,          //!< system clock, posedge
     input                             rst,          //!< sync reset
@@ -102,9 +103,11 @@ module  mclt16x16_bayer3#(
     localparam  DTT_IN_DELAY =  63; // 69; // wa -ra min = 1
 //    localparam  DTT_OUT_DELAY = 128; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
     // May be tweaked so outputs will appear simultaneously
-    localparam  DTT_OUT_DELAY_R = 64; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
-    localparam  DTT_OUT_DELAY_B = 64; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
-    localparam  DTT_OUT_DELAY_G = 128; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
+    
+    
+    localparam  DTT_OUT_DELAY_G = 128-17; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
+    localparam  DTT_OUT_DELAY_R = OUTS_AT_ONCE ? (DTT_OUT_DELAY_G + 128) : 64-19; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
+    localparam  DTT_OUT_DELAY_B = OUTS_AT_ONCE ? (DTT_OUT_DELAY_G + 64) :  64-19; // 191; // start output to sin/cos rotator, with checker - 2*64 +/=?
 
     reg                      [7:0] in_cntr; //
     reg                            run_r;
@@ -468,6 +471,18 @@ module  mclt16x16_bayer3#(
 
 
 // Three of 2 page buffers after dtt (feeding two phase rotators), address MSB is not needed
+    reg  [8:0] dbg_prerot_bufwr_r, dbg_prerot_bufwr_b, dbg_prerot_bufwr_g;
+    always @(posedge clk) begin
+        if (dtt_out_we_r) dbg_prerot_bufwr_r <= dtt_out_ram_wa_rb;
+        if (dtt_out_we_b) dbg_prerot_bufwr_b <= dtt_out_ram_wa_rb;
+        if (dtt_out_we_g) dbg_prerot_bufwr_g <= dtt_out_ram_wa_g;
+    end
+//    wire [8:0] dbg_prerot_buf_r = dtt_rd_regen_r[0]?(dtt_out_ram_wa_rb - dtt_rd_ra_r):'bz; // SuppressThisWarning VEditor : debug output
+//    wire [8:0] dbg_prerot_buf_b = dtt_rd_regen_b[0]?(dtt_out_ram_wa_rb - dtt_rd_ra_b):'bz; // SuppressThisWarning VEditor : debug output
+//    wire [8:0] dbg_prerot_buf_g = dtt_rd_regen_g[0]?(dtt_out_ram_wa_g -  dtt_rd_ra_g):'bz; // SuppressThisWarning VEditor : debug output
+    wire [8:0] dbg_prerot_buf_r = dtt_rd_regen_r[0]?(dbg_prerot_bufwr_r -  dtt_rd_ra_r):'bz; // SuppressThisWarning VEditor : debug output
+    wire [8:0] dbg_prerot_buf_b = dtt_rd_regen_b[0]?(dbg_prerot_bufwr_b -  dtt_rd_ra_b):'bz; // SuppressThisWarning VEditor : debug output
+    wire [8:0] dbg_prerot_buf_g = dtt_rd_regen_g[0]?(dbg_prerot_bufwr_g -  dtt_rd_ra_g):'bz; // SuppressThisWarning VEditor : debug output
     ram18p_var_w_var_r #(
         .REGISTERS(1),
         .LOG2WIDTH_WR(5),
@@ -537,7 +552,7 @@ module  mclt16x16_bayer3#(
         .shift_h       (x_shft_rot_ram_reg),      // input[6:0] signed 
         .shift_v       (y_shft_rot_ram_reg),      // input[6:0] signed 
         .inv_checker   (inv_checker_rot_ram_reg), // input
-        .inv_rows      (valid_odd_rot_ram_reg),   // input
+        .odd_rows      (valid_odd_rot_ram_reg),   // input
         .in_addr       (dtt_rd_ra_r),             // output[7:0] 
         .in_re         (dtt_rd_regen_r),          // output[1:0] 
         .fd_din        (dtt_rd_data_r),           // input[24:0] signed 
@@ -566,7 +581,7 @@ module  mclt16x16_bayer3#(
         .shift_h       (x_shft_rot_ram_reg),      // input[6:0] signed 
         .shift_v       (y_shft_rot_ram_reg),      // input[6:0] signed 
         .inv_checker   (inv_checker_rot_ram_reg), // input
-        .inv_rows      (valid_odd_rot_ram_reg),   // input
+        .odd_rows      (valid_odd_rot_ram_reg),   // input
         .in_addr       (dtt_rd_ra_b),             // output[7:0] 
         .in_re         (dtt_rd_regen_b),          // output[1:0] 
         .fd_din        (dtt_rd_data_b),           // input[24:0] signed 
@@ -594,7 +609,7 @@ module  mclt16x16_bayer3#(
         .shift_h       (x_shft_rot_ram_reg),      // input[6:0] signed 
         .shift_v       (y_shft_rot_ram_reg),      // input[6:0] signed 
         .inv_checker   (inv_checker_rot_ram_reg), // input
-        .inv_rows      (valid_odd_rot_ram_reg),   // input
+        .odd_rows      (valid_odd_rot_ram_reg),   // input
         .in_addr       (dtt_rd_ra_g),             // output[7:0] 
         .in_re         (dtt_rd_regen_g),          // output[1:0] 
         .fd_din        (dtt_rd_data_g),           // input[24:0] signed 

dsp/mclt_test_05.tf

@@ -76,6 +76,8 @@ module  mclt_test_05 ();
     parameter DSP_A_WIDTH =     25;
     parameter DSP_P_WIDTH =     48;
     parameter DEAD_CYCLES =     14;  // start next block immedaitely, or with longer pause
+//    parameter OUTS_AT_ONCE =     0;  // 0: outputs with lowest latency, 1: all at once (with green)
+    parameter OUTS_AT_ONCE =     1;  // 0: outputs with lowest latency, 1: all at once (with green)
     
     reg              RST = 1'b1;
     reg              CLK = 1'b0;
@@ -170,7 +172,8 @@ module  mclt_test_05 ();
     reg                   [1:0] byr_index; // [0:2]; // bayer index of top-left 16x16 tile
 
     initial begin
-        $readmemh("input_data/mclt_dtt_all_00_x1489_y951.dat",  java_all);
+//        $readmemh("input_data/mclt_dtt_all_00_x1489_y951.dat",  java_all);
+        $readmemh("input_data/mclt_dtt_all_02_x1489_y951.dat",  java_all);
 
         $display("000c:  %h", java_all['h000c]);
 
@@ -531,25 +534,32 @@ module  mclt_test_05 ();
     );
     
 
-    reg FIRST_OUT;
-    always @(posedge CLK) FIRST_OUT <= mclt16x16_bayer_i.pre_first_out;
+//dout_r, dout_b, dout_g
+    integer n7r, n7b, n7g;
+    reg [7:0] cntr7r, cntr7b, cntr7g;
+    always @ (posedge CLK) begin
+        if      (RST)           n7r <= -1; else if (pre_first_out_r) n7r <= n7r + 1;
+        if (pre_first_out_r) cntr7r <= 0; else if (dv_r)            cntr7r <= cntr7r + 1; 
+
+        if      (RST)           n7b <= -1; else if (pre_first_out_b) n7b <= n7b + 1;
+        if (pre_first_out_b) cntr7b <= 0; else if (dv_b)            cntr7b <= cntr7b + 1; 
+
+        if      (RST)           n7g <= -1; else if (pre_first_out_g) n7g <= n7g + 1;
+        if (pre_first_out_g) cntr7g <= 0; else if (dv_g)            cntr7g <= cntr7g + 1; 
+    end
+
+    integer diff7r, diff7b, diff7g; // SuppressThisWarning VEditor : assigned in $readmem() system task
+///    wire [OUT_WIDTH-1:0] java_dout_r0 = jav_dtt_rot['h300*out_addr_r[8] + 'h000 + {out_addr_r[7:6], out_addr_r[2:0], out_addr_r[5:3]}];
+///    wire [OUT_WIDTH-1:0] java_dout_b0 = jav_dtt_rot['h300*out_addr_b[8] + 'h100 + {out_addr_b[7:6], out_addr_b[2:0], out_addr_b[5:3]}];
+///    wire [OUT_WIDTH-1:0] java_dout_g0 = jav_dtt_rot['h300*out_addr_g[8] + 'h200 + {out_addr_g[7:6], out_addr_g[2:0], out_addr_g[5:3]}];
+    wire [OUT_WIDTH-1:0] java_dout_r = jav_dtt_rot['h300*n7r + 'h000 + {cntr7r[1:0], cntr7r[7:2]}];
+    wire [OUT_WIDTH-1:0] java_dout_b = jav_dtt_rot['h300*n7b + 'h100 + {cntr7b[1:0], cntr7b[7:2]}];
+    wire [OUT_WIDTH-1:0] java_dout_g = jav_dtt_rot['h300*n7g + 'h200 + {cntr7g[1:0], cntr7g[7:2]}];
     
-    integer n7, cntr7, diff70, diff71; // SuppressThisWarning VEditor : assigned in $readmem() system task
-    wire [OUT_WIDTH-1:0] java_data_dtt_rot0 = jav_dtt_rot[{n7[2:0], cntr7[1],cntr7[0],cntr7[6:2],1'b0}];  //java_dtt_rot0[{cntr7[1],cntr7[0],cntr7[7:2]}];
-    wire [OUT_WIDTH-1:0] java_data_dtt_rot1 = jav_dtt_rot[{n7[2:0], cntr7[1],cntr7[0],cntr7[6:2],1'b1}];  //java_dtt_rot0[{cntr7[1],cntr7[0],cntr7[7:2]}];
-    initial begin
-        while (RST) @(negedge CLK);
-        for (n7 = 0; n7 < 6; n7 = n7+1) begin
-            while (!FIRST_OUT) begin
-                @(negedge CLK);
-            end
-            for (cntr7 = 0; cntr7 < 128; cntr7 = cntr7 + 1) begin
-                #1;
-                diff70 = dout0 - java_data_dtt_rot0;
-                diff71 = dout1 - java_data_dtt_rot1;
-                @(negedge CLK);
-            end
-        end
+    always @ (posedge CLK) begin
+        diff7r <= dv_r? (dout_r - java_dout_r) : 'bz;
+        diff7b <= dv_b? (dout_b - java_dout_b) : 'bz;
+        diff7g <= dv_g? (dout_g - java_dout_g) : 'bz;
     end
 
     reg FIRST_OUTa;
@@ -641,7 +651,7 @@ module  mclt_test_05 ();
     reg                         page3; // 1/2-nd bayer tile
     reg                         pre_run;
     reg                   [1:0] pre_run_cntr;
-    wire                  [2:0] color_page = pre_run_cntr + 3 * page3;
+    wire                  [2:0] color_page = pre_run_cntr + 3 * page3; // SuppressThisWarning VEditor - VDT bug (used as index)
     always @ (posedge CLK) begin
         if (START)                  page3 <= (SUB_PAGE > 2);
         
@@ -673,7 +683,8 @@ module  mclt_test_05 ();
         .DSP_B_WIDTH     (DSP_B_WIDTH),
         .DSP_A_WIDTH     (DSP_A_WIDTH),
         .DSP_P_WIDTH     (DSP_P_WIDTH),
-        .DEAD_CYCLES     (DEAD_CYCLES)
+        .DEAD_CYCLES     (DEAD_CYCLES),
+        .OUTS_AT_ONCE    (OUTS_AT_ONCE)
     ) mclt16x16_bayer3_i (
         .clk            (CLK),                        // input
         .rst            (RST),                        // input

dsp/phase_rotator.v

@@ -53,7 +53,10 @@ module  phase_rotator#(
     input                            start,         //!< single-cycle start pulse that goes 1 cycle before first data
     input signed   [SHIFT_WIDTH-1:0] shift_h,       //!< subpixel shift horizontal
     input signed   [SHIFT_WIDTH-1:0] shift_v,       //!< subpixel shift vertical
-    input                            inv_checker,   //!< negate 2-nd and fourth samples (for handling inverted checkerboard)
+//    input                            inv_checker,   //!< negate 2-nd and fourth samples (for handling inverted checkerboard)
+// fitst sample is never negated
+    input                      [2:0] inv,           //!< bit 0 - invert 2-nd sample, 1 - third, 2 - fourth (for green: 5)
+    
     // input data CC,CS,SC,SS in column scan order (matching DTT)
     input signed      [FD_WIDTH-1:0] fd_din,        //!< frequency domain data in, LATENCY=3 from start
     output reg signed [FD_WIDTH-1:0] fd_out,        //!< frequency domain data in
@@ -95,8 +98,11 @@ module  phase_rotator#(
     reg         [SHIFT_WIDTH-1:0] shift_v0;
     reg         [SHIFT_WIDTH-1:0] shift_vr;
     reg         [SHIFT_WIDTH-1:0] shift_hv; // combined horizonta and vertical shifts to match cntr_mux;
-    reg                           inv_checker_r;
-    reg                           inv_checker_r2;
+//    reg                           inv_checker_r;
+//    reg                           inv_checker_r2;
+    reg                     [2:0] inv_r;
+    reg                     [2:0] inv_r4;
+    reg                     [2:0] inv_r5;
     reg                     [4:0] sign_cs;  // sign for cos / sin, feed to DSP
     wire                          sign_cs_d; // sign_cs delayed by 3 clocks
     reg                     [1:0] sign_cs_r; // sign_cs delayed by 5 clocks      
@@ -115,10 +121,12 @@ module  phase_rotator#(
         
         if (start)      shift_hr <= shift_h;
         if (start)      shift_v0 <= shift_v;
-        if (start)      inv_checker_r <= inv_checker; 
+//        if (start)      inv_checker_r <= inv_checker; 
+        if (start)      inv_r <= inv; 
         if (start_d[3]) shift_vr <= shift_v0;
-        if (start_d[4]) inv_checker_r2 <= inv_checker_r;
-        
+//        if (start_d[4]) inv_checker_r2 <= inv_checker_r;
+        if (start_d[3]) inv_r4 <= inv_r;
+        inv_r5 <= inv_r4;
         if   (rst)                                               run_h <= 0;
         else if (start)                                          run_h <= 1;
 //        else if (&cntr_h_consec[6:0] && (cntr_h[7] || DECIMATE)) run_h <= 0;
@@ -231,10 +239,16 @@ module  phase_rotator#(
         negm_2 <= ((ph[5] & ~sign_cs[3]) | (ph[6] & sign_cs[4])) ^
                   (inv_checker_r2 & (ph[5] | ph[7]));  // invert negation when using Bayer patterns
         */
-        negm_1 <= (ph[4] & ~sign_cs[2]) | (ph[5] & sign_cs[3]); 
-        negm_2 <= ((ph[5] & ~sign_cs[3]) | (ph[6] & sign_cs[4])) ^ inv_checker_r2;
-//                  (inv_checker_r2 & (|ph[7:4]));  // invert negation when using Bayer patterns
+//        negm_1 <= (ph[4] & ~sign_cs[2]) | (ph[5] & sign_cs[3]); 
+//        negm_2 <= ((ph[5] & ~sign_cs[3]) | (ph[6] & sign_cs[4])) ^ inv_checker_r2;
         
+        negm_1 <= ((ph[4] & ~sign_cs[2]) | (ph[5] & sign_cs[3])) ^
+                  (inv_r4[1] & (ph[4] | ph[6]));  // invert negation when using Bayer patterns
+        negm_2 <= ((ph[5] & ~sign_cs[3]) | (ph[6] & sign_cs[4])) ^
+                  ((inv_r5[0] & (ph[4] | ph[6])) |
+                   (inv_r5[2] & (ph[5] | ph[7])));  // invert negation when using Bayer patterns
+
+
         
         accum_1 <= ph[4] | ph[6]; accum_2 <= ph[5] | ph[7];
     // vertical shift DSPs

dsp/phase_rotator_rgb.v

@@ -55,7 +55,7 @@ module  phase_rotator_rgb#(
     input signed   [SHIFT_WIDTH-1:0] shift_h,       //!< subpixel shift horizontal
     input signed   [SHIFT_WIDTH-1:0] shift_v,       //!< subpixel shift vertical
     input                            inv_checker,   //!< negate 2-nd and fourth samples (for handling inverted checkerboard)
-    input                            inv_rows,      //!< 0 : use first row, 1 - second row (when GREEN=0) 
+    input                            odd_rows,      //!< when not GEEN (R or B) 0: even (first) rows non-zero, 1: odd (second)  
     // input data CC,CS,SC,SS in column scan order (matching DTT)
     output             [GREEN + 6:0] in_addr,       //!< input buffer address
     output                     [1:0] in_re,         //!< input buffer re/regen      
@@ -69,10 +69,8 @@ module  phase_rotator_rgb#(
 
     reg signed   [SHIFT_WIDTH-1:0] shift_h_r;
     reg signed   [SHIFT_WIDTH-1:0] shift_v_r;
-    reg                            inv_checker_r;
-    reg                            inv_rows_r;
     reg                            wpage_r;
-    wire                           negate = inv_checker_r ^ inv_rows_r; // FIXME: put real
+    reg                      [2:0] inv;
     reg                      [1:0] dtt_start_out;
     reg                      [7:0] dtt_dly_cntr;
     reg                      [4:0] dtt_rd_regen_dv;
@@ -81,14 +79,16 @@ module  phase_rotator_rgb#(
     reg                      [8:0] out_addr_r;
     assign  in_addr = in_addr_r[GREEN + 6:0];
     assign  in_re = dtt_rd_regen_dv[2:1];
-    assign  fd_wa = {out_addr_r[8], out_addr_r[0],out_addr_r[1],out_addr_r[4:2],out_addr_r[7:5]};
+//    assign  fd_wa = {out_addr_r[8], out_addr_r[0],out_addr_r[1],out_addr_r[4:2],out_addr_r[7:5]};
+    assign  fd_wa = {out_addr_r[8], out_addr_r[1],out_addr_r[0],out_addr_r[4:2],out_addr_r[7:5]};
     
     always @ (posedge clk) begin
         if (start) begin
             shift_h_r <=     shift_h;
             shift_v_r <=     shift_v;
-            inv_checker_r <= inv_checker;
-            inv_rows_r <=    inv_rows;
+            inv <= inv_checker ?
+                   (( GREEN || odd_rows) ? 5 : 6):
+                   ((!GREEN && odd_rows) ? 3 : 0);
             wpage_r <=       wpage;
         end
         
@@ -110,11 +110,11 @@ module  phase_rotator_rgb#(
         else if (dtt_rd_regen_dv[0])    dtt_rd_cntr_pre <= dtt_rd_cntr_pre + 1;
         
         if (GREEN) in_addr_r <= {dtt_rd_cntr_pre[8],
-                                 dtt_rd_cntr_pre[0] ^ dtt_rd_cntr_pre[0],
+                                 dtt_rd_cntr_pre[0] ^ dtt_rd_cntr_pre[1],
                                  dtt_rd_cntr_pre[0] ? (~dtt_rd_cntr_pre[7:2]) : dtt_rd_cntr_pre[7:2]};
         else       in_addr_r <= {1'b0, 
                                  dtt_rd_cntr_pre[8],
-//                               dtt_rd_cntr_pre[0] ^ dtt_rd_cntr_pre[0],
+//                               dtt_rd_cntr_pre[0] ^ dtt_rd_cntr_pre[1],
                                  dtt_rd_cntr_pre[1] ?
                                  (dtt_rd_cntr_pre[0] ? (~dtt_rd_cntr_pre[7:2]) : {~dtt_rd_cntr_pre[7:5],dtt_rd_cntr_pre[4:2]}):
                                  (dtt_rd_cntr_pre[0] ? {dtt_rd_cntr_pre[7:5],~dtt_rd_cntr_pre[4:2]} : dtt_rd_cntr_pre[7:2])};
@@ -139,10 +139,10 @@ module  phase_rotator_rgb#(
         .clk           (clk),            // input
         .rst           (rst),            // input
         .start         (dtt_start_out[1]), // input
-        // are these shift OK? Will need to be valis only @ dtt_start_out
+        // are these shift OK? Will need to be valid only @ dtt_start_out
         .shift_h       (shift_h_r),      // input[6:0] signed 
         .shift_v       (shift_v_r),      // input[6:0] signed
-        .inv_checker   (negate),         // input only used for Bayer mosaic data 
+        .inv           (inv),            // input [2:0] 
         .fd_din        (fd_din),         // input[24:0] signed. Expected latency = 3 from start  
         .fd_out        (fd_out),         // output[24:0] reg signed 
         .pre_first_out (pre_first_out),  // output reg