Optimized cache writeback path by 1) VX_fair_arbiter and 2) Added a wb register between LSU and WB arbiter

hw/rtl/VX_back_end.v

@@ -76,7 +76,7 @@ module VX_back_end #(
         .clk            (clk),
         .reset          (reset),
         .lsu_req_if     (lsu_req_if),
-        .mem_wb_if      (mem_wb_if),
+        .mem_wb_if_p1   (mem_wb_if),
         .dcache_req_if  (dcache_req_if),
         .dcache_rsp_if  (dcache_rsp_if),
         .delay          (mem_delay),

hw/rtl/VX_lsu_unit.v

@@ -12,7 +12,7 @@ module VX_lsu_unit #(
     VX_lsu_req_if           lsu_req_if,
 
     // Write back to GPR
-    VX_wb_if                mem_wb_if,
+    VX_wb_if                mem_wb_if_p1,
 
    // Dcache interface
     VX_cache_core_req_if    dcache_req_if,
@@ -20,6 +20,9 @@ module VX_lsu_unit #(
 
     output wire             delay
 );
+
+    VX_wb_if                mem_wb_if;
+
     wire[`NUM_THREADS-1:0][31:0]    use_address;
     wire[`NUM_THREADS-1:0][31:0]    use_store_data;
     wire[`NUM_THREADS-1:0]          use_valid;
@@ -156,7 +159,20 @@ module VX_lsu_unit #(
     assign mem_wb_if.data  = core_rsp_data;
 
     // Can't accept new response
-    assign dcache_rsp_if.core_rsp_ready = !no_slot_mem;    
+    assign dcache_rsp_if.core_rsp_ready = !no_slot_mem & (|mem_wb_if_p1.valid);
+
+
+
+    // From LSU to WB
+    localparam WB_REQ_SIZE = (`NUM_THREADS) + (`NUM_THREADS * 32) + (`NW_BITS) + (5) + (2) + 32;
+    VX_generic_register #(.N(WB_REQ_SIZE)) lsu_to_wb(
+        .clk   (clk),
+        .reset (reset),
+        .stall (no_slot_mem),
+        .flush (1'b0),
+        .in    ({mem_wb_if.valid   , mem_wb_if.data   , mem_wb_if.warp_num   , mem_wb_if.rd   , mem_wb_if.wb   , mem_wb_if.curr_PC   }),
+        .out   ({mem_wb_if_p1.valid, mem_wb_if_p1.data, mem_wb_if_p1.warp_num, mem_wb_if_p1.rd, mem_wb_if_p1.wb, mem_wb_if_p1.curr_PC})
+    );
 
     `SCOPE_ASSIGN(scope_dcache_req_valid, dcache_req_if.core_req_valid);    
     `SCOPE_ASSIGN(scope_dcache_req_warp_num, use_warp_num);

hw/rtl/cache/VX_cache_core_rsp_merge.v

@@ -35,7 +35,7 @@ module VX_cache_core_rsp_merge #(
 
     wire [`BANK_BITS-1:0] main_bank_index;
 
-    VX_fixed_arbiter #(
+    VX_fair_arbiter #(
         .N(NUM_BANKS)
     ) sel_bank (
         .clk         (clk),

hw/rtl/libs/VX_fair_arbiter.v

@@ -0,0 +1,68 @@
+`include "VX_define.vh"
+
+module VX_fair_arbiter #(
+    parameter N = 0
+) (
+    input  wire                  clk,
+    input  wire                  reset,
+    input  wire [N-1:0]          requests,           
+    output wire [`LOG2UP(N)-1:0] grant_index,
+    output wire [N-1:0]          grant_onehot,   
+    output wire                  grant_valid
+  );
+
+
+    if (N == 1)  begin        
+        
+        `UNUSED_VAR (clk)
+        `UNUSED_VAR (reset)
+        assign grant_index  = 0;
+        assign grant_onehot = requests;
+        assign grant_valid  = requests[0];
+
+    end else begin
+    
+
+       reg  [N-1:0] requests_use;
+       wire [N-1:0] update_value;
+
+       wire         refill;
+       wire [N-1:0] refill_value;
+       reg  [N-1:0] refill_original;
+
+       always @(posedge clk) begin
+            if (reset) begin
+                requests_use    <= 0;
+                refill_original <= 0;
+            end else if (refill) begin
+                requests_use    <= refill_value;
+                refill_original <= refill_value;
+            end else begin
+                requests_use <= update_value;
+            end
+       end
+
+       assign refill       = (requests_use == 0);
+       assign refill_value = requests;
+        
+        reg [N-1:0] grant_onehot_r; 
+
+        VX_priority_encoder # (
+            .N(N)
+        ) priority_encoder (
+            .data_in   (requests_use),
+            .data_out  (grant_index ),
+            .valid_out (grant_valid )
+        );
+
+        always @(*) begin
+            grant_onehot_r = N'(0);
+            grant_onehot_r[grant_index] = 1;
+        end
+        assign grant_onehot = grant_onehot_r;    
+       
+        assign update_value = (requests_use & ~grant_onehot_r) | ((refill_original ^ requests) & ~refill_original);
+
+    end
+    
+endmodule