bug fixes

hw/dpi/float_dpi.cpp

@@ -81,7 +81,7 @@ inline bool is_nan_boxed(uint64_t value) {
 
 inline bool is_nan_boxed16(uint64_t value) {
 #ifdef XLEN_64
-  return (uint32_t(value >> 16) == 0xffffffffffff);
+  return (uint64_t(value >> 16) == 0xffffffffffff);
 #else
   return true;
 #endif

hw/rtl/core/VX_ibuffer.sv

@@ -70,7 +70,7 @@ module VX_ibuffer import VX_gpu_pkg::*; #(
             .ready_out(uop_sequencer_if.ready)
         );
     `ifndef L1_ENABLE
-        assign decode_if.ibuf_pop[w] = uop_sequencer_if[w].valid && uop_sequencer_if[w].ready;
+        assign decode_if.ibuf_pop[w] = uop_sequencer_if.valid && uop_sequencer_if.ready;
     `endif
 
         VX_uop_sequencer uop_sequencer (

hw/rtl/tcu/VX_tcu_fedp_fp.sv

@@ -24,10 +24,10 @@ module VX_tcu_fedp_fp #(
     input  wire[2:0] fmt_s,
     input  wire[2:0] fmt_d,
 
-    input  wire [N-1:0][DATAW-1:0] a_row,
-    input  wire [N-1:0][DATAW-1:0] b_col,
-    input  wire [DATAW-1:0] c_val,
-    output wire [DATAW-1:0] d_val
+    input  wire [N-1:0][`XLEN-1:0] a_row,
+    input  wire [N-1:0][`XLEN-1:0] b_col,
+    input  wire [`XLEN-1:0] c_val,
+    output wire [`XLEN-1:0] d_val
 );
     wire [DATAW-1:0] a_row_p [0:N-1];
     wire [DATAW-1:0] b_col_p [0:N-1];
@@ -35,6 +35,8 @@ module VX_tcu_fedp_fp #(
     wire [2:0] fmt_s_p [0:N-1];
     wire [2:0] fmt_d_p [0:N-1];
 
+    `UNUSED_VAR ({a_row, b_col, c_val});
+
     for (genvar i = 0; i < N; i++) begin: g_pipe
         VX_pipe_register #(
             .DATAW (DATAW),
@@ -43,7 +45,7 @@ module VX_tcu_fedp_fp #(
             .clk      (clk),
             .reset    (reset),
             .enable   (enable),
-            .data_in  (a_row[i]),
+            .data_in  (a_row[i][DATAW-1:0]),
             .data_out (a_row_p[i])
         );
         VX_pipe_register #(
@@ -53,7 +55,7 @@ module VX_tcu_fedp_fp #(
             .clk      (clk),
             .reset    (reset),
             .enable   (enable),
-            .data_in  (b_col[i]),
+            .data_in  (b_col[i][DATAW-1:0]),
             .data_out (b_col_p[i])
         );
 
@@ -71,7 +73,7 @@ module VX_tcu_fedp_fp #(
 
     wire [DATAW-1:0] fma_out [0:N];
 
-    assign fma_out[0] = c_val;
+    assign fma_out[0] = c_val[DATAW-1:0];
 
     for (genvar i = 0; i < N; i++) begin : g_fmas
     `ifdef TCU_DPI
@@ -106,6 +108,6 @@ module VX_tcu_fedp_fp #(
 
     end
 
-    assign d_val = fma_out[N];
+    assign d_val = `XLEN'(fma_out[N]);
 
 endmodule

hw/rtl/tcu/VX_tcu_fedp_int.sv

@@ -24,10 +24,10 @@ module VX_tcu_fedp_int #(
     input  wire[2:0] fmt_s,
     input  wire[2:0] fmt_d,
 
-    input  wire [N-1:0][DATAW-1:0] a_row,
-    input  wire [N-1:0][DATAW-1:0] b_col,
-    input  wire [DATAW-1:0] c_val,
-    output wire [DATAW-1:0] d_val
+    input  wire [N-1:0][`XLEN-1:0] a_row,
+    input  wire [N-1:0][`XLEN-1:0] b_col,
+    input  wire [`XLEN-1:0] c_val,
+    output wire [`XLEN-1:0] d_val
 );
     wire [DATAW-1:0] a_row_p [0:N-1];
     wire [DATAW-1:0] b_col_p [0:N-1];
@@ -35,6 +35,8 @@ module VX_tcu_fedp_int #(
     wire [2:0] fmt_s_p [0:N-1];
     wire [2:0] fmt_d_p [0:N-1];
 
+    `UNUSED_VAR ({a_row, b_col, c_val});
+
     for (genvar i = 0; i < N; i++) begin: g_pipe
         VX_pipe_register #(
             .DATAW (DATAW),
@@ -43,7 +45,7 @@ module VX_tcu_fedp_int #(
             .clk      (clk),
             .reset    (reset),
             .enable   (enable),
-            .data_in  (a_row[i]),
+            .data_in  (a_row[i][DATAW-1:0]),
             .data_out (a_row_p[i])
         );
         VX_pipe_register #(
@@ -53,7 +55,7 @@ module VX_tcu_fedp_int #(
             .clk      (clk),
             .reset    (reset),
             .enable   (enable),
-            .data_in  (b_col[i]),
+            .data_in  (b_col[i][DATAW-1:0]),
             .data_out (b_col_p[i])
         );
 
@@ -71,7 +73,7 @@ module VX_tcu_fedp_int #(
 
     wire [DATAW-1:0] fma_out [0:N];
 
-    assign fma_out[0] = c_val;
+    assign fma_out[0] = c_val[DATAW-1:0];
 
     for (genvar i = 0; i < N; i++) begin : g_fmas
         VX_tcu_fma_int #(
@@ -89,6 +91,6 @@ module VX_tcu_fedp_int #(
         );
     end
 
-    assign d_val = fma_out[N];
+    assign d_val = `XLEN'(fma_out[N]);
 
 endmodule

hw/rtl/tcu/VX_tcu_fp.sv

@@ -110,7 +110,7 @@ module VX_tcu_fp import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
             wire [`XLEN-1:0] c_val = execute_if.data.rs3_data[i * TCU_TC_N + j];
 
             VX_tcu_fedp_fp #(
-                .DATAW (`XLEN),
+                .DATAW (32),
                 .N     (TCU_TC_K)
             ) fedp (
                 .clk   (clk),

hw/rtl/tcu/VX_tcu_int.sv

@@ -110,7 +110,7 @@ module VX_tcu_int import VX_gpu_pkg::*, VX_tcu_pkg::*; #(
             wire [`XLEN-1:0] c_val = execute_if.data.rs3_data[i * TCU_TC_N + j];
 
             VX_tcu_fedp_int #(
-                .DATAW (`XLEN),
+                .DATAW (32),
                 .N     (TCU_TC_K)
             ) fedp (
                 .clk   (clk),