Configurable Cache & Bigger Cache

sim/Makefile

@@ -52,7 +52,7 @@ FUNC_SOURCE = $(wildcard ../resources/tb.sv ../resources/func_test/*.v ../resour
 ####################
 .phony: lint verilate func_soft tlb_soft build coverage run clean
 
-default: func_run
+default: run
 
 lint:
 	$(VERILATOR) --lint-only $(VERILATOR_FLAGS) $(INCLUDE) $(SOURCE) -top mycpu_top

src/Gadgets/mux9.sv

@@ -0,0 +1,31 @@
+module mux9 #(
+    parameter WIDTH = 8
+) (
+    input logic [WIDTH-1:0] d0,
+    input logic [WIDTH-1:0] d1,
+    input logic [WIDTH-1:0] d2,
+    input logic [WIDTH-1:0] d3,
+    input logic [WIDTH-1:0] d4,
+    input logic [WIDTH-1:0] d5,
+    input logic [WIDTH-1:0] d6,
+    input logic [WIDTH-1:0] d7,
+    input logic [WIDTH-1:0] d8,
+
+    input  logic [      3:0] s,
+    output logic [WIDTH-1:0] q
+);
+
+  always_comb begin
+    case (s)
+      4'b0000:  q = d0;
+      4'b0001:  q = d1;
+      4'b0010:  q = d2;
+      4'b0011:  q = d3;
+      4'b0100:  q = d4;
+      4'b0101:  q = d5;
+      4'b0110:  q = d6;
+      4'b0111:  q = d7;
+      default:  q = d8;
+    endcase
+  end
+endmodule

src/MU/DCache.sv

@@ -15,7 +15,7 @@ module DCache (
   DCDataRAM_t DataRAM0/*verilator split_var*/, DataRAM1/*verilator split_var*/, DataRAM2/*verilator split_var*/, DataRAM3/*verilator split_var*/;
 
   (* RAM_STYLE="distributed" *)
-  logic [3:0] LRU[128];
+  logic [3:0] LRU[`DC_INDEX_DEPTH];
   logic [3:0] nowLRU;
   logic [3:0] nxtLRU;
 
@@ -89,7 +89,7 @@ module DCache (
                   | (victim[1] ? tag[1].tag : {(32-`DC_TAGL){1'b0}})
                   | (victim[2] ? tag[2].tag : {(32-`DC_TAGL){1'b0}})
                   | (victim[3] ? tag[3].tag : {(32-`DC_TAGL){1'b0}});
-  assign port.dirt_addr = {dirt_tag, port.index, 4'b0};
+  assign port.dirt_addr = {dirt_tag, port.index, `DC_INDEXL'b0};
   assign port.dirt_row = (victim[0] ? data[0] : {`DC_DATA_LENGTH{1'b0}})
                        | (victim[1] ? data[1] : {`DC_DATA_LENGTH{1'b0}})
                        | (victim[2] ? data[2] : {`DC_DATA_LENGTH{1'b0}})
@@ -112,7 +112,7 @@ module DCache (
   assign nxtLRU = (setLRU_valid ? nxtsetLRU : 4'b0000)
                 | (clrLRU_valid ? nxtclrLRU : 4'b0000);
 
-  for (genvar i = 0; i < 128; i++)
+  for (genvar i = 0; i < `DC_INDEX_DEPTH; i++)
     initial LRU[i] = 4'b0;
 
   always_ff @(posedge clk) begin

src/MU/ICache.sv

@@ -15,7 +15,7 @@ module ICache (
   ICDataRAM_t DataRAM0/*verilator split_var*/, DataRAM1/*verilator split_var*/, DataRAM2/*verilator split_var*/, DataRAM3/*verilator split_var*/;
 
   (* RAM_STYLE="distributed" *)
-  logic [3:0] LRU[64];
+  logic [3:0] LRU[`IC_INDEX_DEPTH];
   logic [3:0] nowLRU;
   logic [3:0] nxtLRU;
 
@@ -81,7 +81,7 @@ module ICache (
   assign nxtLRU = (setLRU_valid ? nxtsetLRU : 4'b0000)
                 | (clrLRU_valid ? nxtclrLRU : 4'b0000);
 
-  for (genvar i = 0; i < 64; i++)
+  for (genvar i = 0; i < `IC_INDEX_DEPTH; i++)
     initial LRU[i] = 4'b0;
 
   always_ff @(posedge clk) begin

src/MU/MU.sv

@@ -120,16 +120,9 @@ module MU (
   // ============
   logic    if_source_select_direct; // only used when non-cached
   ICData_t if_source_data;
-  // TODO: use macro/generator to dynamic mux
-  mux5 #(2 * `XLEN) if_rdata_mux(
-    if_source_data[ 63:  0],
-    if_source_data[127: 64],
-    if_source_data[191:128],
-    if_source_data[255:192],
-    if_source_data[255:192], // uncached
-    {if_source_select_direct, stored_instfetch_addr[`IC_INDEXL-1:3]},
-    {instfetch.rdata1, instfetch.rdata0}
-  );
+  assign   {instfetch.rdata1, instfetch.rdata0} = if_source_select_direct
+                                                ? if_source_data[(`IC_ROW_LENGTH-1)*32+:64]
+                                                : if_source_data[{stored_instfetch_addr[`IC_INDEXL-1:3], 6'b0}+:64];
 
   // =====================================
   // == InstFetch Control State Machine ==
@@ -368,16 +361,9 @@ module MU (
 
   logic    mem_rdata_select_direct; // only used when non-cached
   DCData_t mem_rdata_source_data;
-  // TODO: use macro/generator to dynamic mux
-  mux5 #(`XLEN) mem_rdata_mux(
-    mem_rdata_source_data[ 31:  0],
-    mem_rdata_source_data[ 63: 32],
-    mem_rdata_source_data[ 95: 64],
-    mem_rdata_source_data[127: 96],
-    mem_rdata_source_data[127: 96], // uncached
-    {mem_rdata_select_direct, stored_memory_addr[`DC_INDEXL-1:2]},
-    memory.rdata
-  );
+  assign   memory.rdata = mem_rdata_select_direct
+                        ? mem_rdata_source_data[`DC_ROW_LENGTH*32+:32]
+                        : mem_rdata_source_data[{stored_memory_addr[`DC_INDEXL-1:2], 5'b0}+:32];
 
   logic    mem_wstrb_direct;
   DCData_t mem_wdata_source_data, mem_wdata_output;

src/include/Cache.svh

@@ -4,11 +4,12 @@
 `include "defines.svh"
 
 // IC for I-Cache
-`define IC_TAGL        11
-`define IC_INDEXL      5
+`define IC_TAGL        12
+`define IC_INDEXL      6
 `define IC_TAG_LENGTH  (`XLEN - `IC_IGAL + 1) // Tag + Valid
-`define IC_DATA_LENGTH (2 ** `IC_INDEXL * 8)  // 32Bytes
+`define IC_DATA_LENGTH (2 ** `IC_INDEXL * 8)  // 64Bytes
 `define IC_ROW_LENGTH  (`IC_DATA_LENGTH / `XLEN - 1)
+`define IC_INDEX_DEPTH (2 ** (`IC_TAGL - `IC_INDEXL))
 
 typedef logic     [`IC_DATA_LENGTH-1:0] ICData_t;
 typedef logic         [32-`IC_TAGL-1:0] ICTagL_t;
@@ -41,11 +42,12 @@ typedef struct packed {
 
 
 // DC for D-Cache
-`define DC_TAGL        11
-`define DC_INDEXL      4
+`define DC_TAGL        12
+`define DC_INDEXL      5
 `define DC_TAG_LENGTH  (`XLEN - `DC_TAGL + 1 + 1) // Tag + Valid + Dirty
-`define DC_DATA_LENGTH (2 ** `DC_INDEXL * 8)      // 16Bytes
+`define DC_DATA_LENGTH (2 ** `DC_INDEXL * 8)      // 8Bytes
 `define DC_ROW_LENGTH  (`DC_DATA_LENGTH / `XLEN - 1)
+`define DC_INDEX_DEPTH (2 ** (`DC_TAGL - `DC_INDEXL))
 
 typedef logic     [`DC_DATA_LENGTH-1:0] DCData_t;
 typedef logic         [32-`DC_TAGL-1:0] DCTagL_t;

src/include/defines.svh

@@ -1,12 +1,12 @@
 `ifndef DEFINES_SVH
 `define DEFINES_SVH
 
-`define ILA_DEBUG
+// `define ILA_DEBUG
 
-`define ENABLE_CACHEOP
-`define ENABLE_TLB
-`define ENABLE_CpU
-`define ENABLE_TRAP
+// `define ENABLE_CACHEOP
+// `define ENABLE_TLB
+// `define ENABLE_CpU
+// `define ENABLE_TRAP
 
 `ifdef SIMULATION_VERILATOR
   `undef ENABLE_CpU