bf16mul and execute_if fix

Author: NikhilRout

ci/regression.sh.in

@@ -482,7 +482,7 @@ tensor()
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_DPI" make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=2 --no-infinities --no-nans
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_DSP" LATENCY=42 make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=2 --no-infinities --no-nans
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_BHF" LATENCY=13 make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=2 --no-infinities --no-nans
-    #(failed)make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_DRL" make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=2 --no-zeros --no-subnormals --no-infinities --no-nans
+    make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_DRL" make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=2 --no-zeros --no-subnormals --no-infinities --no-nans
 
     # test fp8_e4m3
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_DPI" make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=3 --no-infinities --no-nans

hw/rtl/interfaces/VX_execute_if.sv

@@ -18,7 +18,6 @@ interface VX_execute_if import VX_gpu_pkg::*; #(
 );
     logic  valid;
     data_t data;
-    `UNUSED_VAR(data);
     logic  ready;
 
     modport master (

hw/rtl/tcu/VX_tcu_core.sv

@@ -66,7 +66,7 @@ module VX_tcu_core import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
     wire [3:0] fmt_s = execute_if.data.op_args.tcu.fmt_s;
     wire [3:0] fmt_d = execute_if.data.op_args.tcu.fmt_d;
 
-    `UNUSED_VAR ({step_m, step_n, fmt_s, fmt_d});
+    `UNUSED_VAR ({step_m, step_n, fmt_s, fmt_d, execute_if.data});
 
     wire mdata_queue_full;
 

hw/rtl/tcu/drl/VX_tcu_drl_bf16mul.sv

@@ -36,16 +36,19 @@ module VX_tcu_drl_bf16mul (
     assign sign_y = sign_a ^ sign_b;
 
     //Raw result exponent calculation
-    wire [7:0] neg_bias = -8'd126;    //-127+1
+    wire [9:0] neg_bias = 10'b1110000010; //-127+1
+    wire [9:0] raw_exp_y_signed;
+    `UNUSED_VAR (raw_exp_y_signed);
     VX_csa_tree #(
         .N(3),
-        .W(8),
-        .S(8)
+        .W(10),
+        .S(10)
     ) biasexp_bf16(
-        .operands({exp_a, exp_b, neg_bias}),
-        .sum     (raw_exp_y),
+        .operands({{2'd0, exp_a}, {2'd0, exp_b}, neg_bias}),
+        .sum     (raw_exp_y_signed),
         `UNUSED_PIN (cout)
     );
+    assign raw_exp_y = raw_exp_y_signed[9] ? -raw_exp_y_signed[7:0] : raw_exp_y_signed[7:0];
 
     //Mantissa Calculation
     wire [7:0] full_mant_a = {1'b1, frac_a};