| 1 |
16 |
csantifort |
//////////////////////////////////////////////////////////////////
|
| 2 |
|
|
// //
|
| 3 |
|
|
// L1 Instruction Cache for Amber 25 Core //
|
| 4 |
|
|
// //
|
| 5 |
|
|
// This file is part of the Amber project //
|
| 6 |
|
|
// http://www.opencores.org/project,amber //
|
| 7 |
|
|
// //
|
| 8 |
|
|
// Description //
|
| 9 |
|
|
// Synthesizable L1 Unified Data and Instruction Cache //
|
| 10 |
|
|
// Cache is 4-way, 256 line and 16 bytes per line for //
|
| 11 |
|
|
// a total of 16KB. The cache policy is write-through and //
|
| 12 |
|
|
// read allocate. For swap instructions (SWP and SWPB) the //
|
| 13 |
|
|
// location is evicted from the cache and read from main //
|
| 14 |
|
|
// memory. //
|
| 15 |
|
|
// //
|
| 16 |
|
|
// Author(s): //
|
| 17 |
|
|
// - Conor Santifort, csantifort.amber@gmail.com //
|
| 18 |
|
|
// //
|
| 19 |
|
|
//////////////////////////////////////////////////////////////////
|
| 20 |
|
|
// //
|
| 21 |
|
|
// Copyright (C) 2011 Authors and OPENCORES.ORG //
|
| 22 |
|
|
// //
|
| 23 |
|
|
// This source file may be used and distributed without //
|
| 24 |
|
|
// restriction provided that this copyright statement is not //
|
| 25 |
|
|
// removed from the file and that any derivative work contains //
|
| 26 |
|
|
// the original copyright notice and the associated disclaimer. //
|
| 27 |
|
|
// //
|
| 28 |
|
|
// This source file is free software; you can redistribute it //
|
| 29 |
|
|
// and/or modify it under the terms of the GNU Lesser General //
|
| 30 |
|
|
// Public License as published by the Free Software Foundation; //
|
| 31 |
|
|
// either version 2.1 of the License, or (at your option) any //
|
| 32 |
|
|
// later version. //
|
| 33 |
|
|
// //
|
| 34 |
|
|
// This source is distributed in the hope that it will be //
|
| 35 |
|
|
// useful, but WITHOUT ANY WARRANTY; without even the implied //
|
| 36 |
|
|
// warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR //
|
| 37 |
|
|
// PURPOSE. See the GNU Lesser General Public License for more //
|
| 38 |
|
|
// details. //
|
| 39 |
|
|
// //
|
| 40 |
|
|
// You should have received a copy of the GNU Lesser General //
|
| 41 |
|
|
// Public License along with this source; if not, download it //
|
| 42 |
|
|
// from http://www.opencores.org/lgpl.shtml //
|
| 43 |
|
|
// //
|
| 44 |
|
|
//////////////////////////////////////////////////////////////////
|
| 45 |
|
|
|
| 46 |
|
|
`include "a25_config_defines.v"
|
| 47 |
|
|
|
| 48 |
|
|
module a25_icache
|
| 49 |
|
|
#(
|
| 50 |
|
|
|
| 51 |
|
|
// ---------------------------------------------------------
|
| 52 |
|
|
// Cache Configuration
|
| 53 |
|
|
|
| 54 |
|
|
// Limited to Linux 4k page sizes -> 256 lines
|
| 55 |
|
|
parameter CACHE_LINES = 256,
|
| 56 |
|
|
|
| 57 |
|
|
// This cannot be changed without some major surgeory on
|
| 58 |
|
|
// this module
|
| 59 |
|
|
parameter CACHE_WORDS_PER_LINE = 4,
|
| 60 |
|
|
|
| 61 |
|
|
// Changing this parameter is the recommended
|
| 62 |
|
|
// way to change the overall cache size; 2, 4 and 8 ways are supported.
|
| 63 |
|
|
// 2 ways -> 8KB cache
|
| 64 |
|
|
// 4 ways -> 16KB cache
|
| 65 |
|
|
// 8 ways -> 32KB cache
|
| 66 |
|
|
parameter WAYS = `A25_ICACHE_WAYS ,
|
| 67 |
|
|
|
| 68 |
|
|
// derived configuration parameters
|
| 69 |
|
|
parameter CACHE_ADDR_WIDTH = log2 ( CACHE_LINES ), // = 8
|
| 70 |
|
|
parameter WORD_SEL_WIDTH = log2 ( CACHE_WORDS_PER_LINE ), // = 2
|
| 71 |
|
|
parameter TAG_ADDR_WIDTH = 32 - CACHE_ADDR_WIDTH - WORD_SEL_WIDTH - 2, // = 20
|
| 72 |
|
|
parameter TAG_WIDTH = TAG_ADDR_WIDTH + 1, // = 21, including Valid flag
|
| 73 |
|
|
parameter CACHE_LINE_WIDTH = CACHE_WORDS_PER_LINE * 32, // = 128
|
| 74 |
|
|
parameter TAG_ADDR32_LSB = CACHE_ADDR_WIDTH + WORD_SEL_WIDTH + 2, // = 12
|
| 75 |
|
|
parameter CACHE_ADDR32_MSB = CACHE_ADDR_WIDTH + WORD_SEL_WIDTH + 2 - 1, // = 11
|
| 76 |
|
|
parameter CACHE_ADDR32_LSB = WORD_SEL_WIDTH + 2 , // = 4
|
| 77 |
|
|
parameter WORD_SEL_MSB = WORD_SEL_WIDTH + 2 - 1, // = 3
|
| 78 |
|
|
parameter WORD_SEL_LSB = 2 // = 2
|
| 79 |
|
|
// ---------------------------------------------------------
|
| 80 |
|
|
)
|
| 81 |
|
|
|
| 82 |
|
|
|
| 83 |
|
|
(
|
| 84 |
|
|
input i_clk,
|
| 85 |
|
|
input i_core_stall,
|
| 86 |
|
|
output o_stall,
|
| 87 |
|
|
|
| 88 |
|
|
// Read / Write requests from core
|
| 89 |
|
|
input i_select,
|
| 90 |
|
|
input [31:0] i_address, // registered address from execute
|
| 91 |
|
|
input [31:0] i_address_nxt, // un-registered version of address from execute stage
|
| 92 |
|
|
input i_cache_enable, // from co-processor 15 configuration register
|
| 93 |
|
|
input i_cache_flush, // from co-processor 15 register
|
| 94 |
|
|
|
| 95 |
|
|
output [31:0] o_read_data,
|
| 96 |
|
|
|
| 97 |
|
|
// WB Read Request
|
| 98 |
|
|
output o_wb_req, // Read Request
|
| 99 |
|
|
input [31:0] i_wb_read_data,
|
| 100 |
|
|
input i_wb_ready
|
| 101 |
|
|
);
|
| 102 |
|
|
|
| 103 |
|
|
`include "a25_localparams.v"
|
| 104 |
|
|
`include "a25_functions.v"
|
| 105 |
|
|
|
| 106 |
|
|
// One-hot encoded
|
| 107 |
|
|
localparam C_INIT = 0,
|
| 108 |
|
|
C_CORE = 1,
|
| 109 |
|
|
C_FILL = 2,
|
| 110 |
|
|
C_INVA = 3,
|
| 111 |
|
|
C_STATES = 4;
|
| 112 |
|
|
|
| 113 |
|
|
localparam [3:0] CS_INIT = 4'd0,
|
| 114 |
|
|
CS_IDLE = 4'd1,
|
| 115 |
|
|
CS_FILL0 = 4'd2,
|
| 116 |
|
|
CS_FILL1 = 4'd3,
|
| 117 |
|
|
CS_FILL2 = 4'd4,
|
| 118 |
|
|
CS_FILL3 = 4'd5,
|
| 119 |
|
|
CS_FILL4 = 4'd6,
|
| 120 |
|
|
CS_FILL_COMPLETE = 4'd7,
|
| 121 |
|
|
CS_TURN_AROUND = 4'd8,
|
| 122 |
|
|
CS_WRITE_HIT1 = 4'd9,
|
| 123 |
|
|
CS_EX_DELETE = 4'd10;
|
| 124 |
|
|
|
| 125 |
|
|
|
| 126 |
|
|
reg [3:0] c_state = CS_IDLE;
|
| 127 |
|
|
reg [C_STATES-1:0] source_sel = 1'd1 << C_CORE;
|
| 128 |
|
|
reg [CACHE_ADDR_WIDTH:0] init_count = 'd0;
|
| 129 |
|
|
|
| 130 |
|
|
wire [TAG_WIDTH-1:0] tag_rdata_way [WAYS-1:0];
|
| 131 |
|
|
wire [CACHE_LINE_WIDTH-1:0] data_rdata_way[WAYS-1:0];
|
| 132 |
|
|
wire [WAYS-1:0] data_wenable_way;
|
| 133 |
|
|
wire [WAYS-1:0] data_hit_way;
|
| 134 |
|
|
wire [WAYS-1:0] tag_wenable_way;
|
| 135 |
|
|
reg [WAYS-1:0] select_way = 'd0;
|
| 136 |
|
|
wire [WAYS-1:0] next_way;
|
| 137 |
|
|
reg [WAYS-1:0] valid_bits_r = 'd0;
|
| 138 |
|
|
|
| 139 |
|
|
reg [3:0] random_num = 4'hf;
|
| 140 |
|
|
|
| 141 |
|
|
wire [CACHE_ADDR_WIDTH-1:0] tag_address;
|
| 142 |
|
|
wire [TAG_WIDTH-1:0] tag_wdata;
|
| 143 |
|
|
wire tag_wenable;
|
| 144 |
|
|
|
| 145 |
|
|
wire [CACHE_LINE_WIDTH-1:0] data_wdata;
|
| 146 |
|
|
wire [CACHE_ADDR_WIDTH-1:0] data_address;
|
| 147 |
|
|
wire [31:0] write_data_word;
|
| 148 |
|
|
|
| 149 |
|
|
wire idle_hit;
|
| 150 |
|
|
wire read_miss;
|
| 151 |
|
|
wire read_miss_fill;
|
| 152 |
|
|
wire invalid_read;
|
| 153 |
|
|
wire fill_state;
|
| 154 |
|
|
|
| 155 |
|
|
reg [31:0] miss_address = 'd0;
|
| 156 |
|
|
wire [CACHE_LINE_WIDTH-1:0] hit_rdata;
|
| 157 |
|
|
|
| 158 |
|
|
wire cache_busy_stall;
|
| 159 |
|
|
wire read_stall;
|
| 160 |
|
|
|
| 161 |
|
|
wire enable;
|
| 162 |
|
|
wire [CACHE_ADDR_WIDTH-1:0] address;
|
| 163 |
|
|
wire [31:0] address_c;
|
| 164 |
|
|
reg [31:0] address_r = 'd0;
|
| 165 |
|
|
|
| 166 |
|
|
reg [CACHE_LINE_WIDTH-1:0] wb_rdata_burst_r = 'd0;
|
| 167 |
|
|
wire [CACHE_LINE_WIDTH-1:0] wb_rdata_burst;
|
| 168 |
|
|
|
| 169 |
|
|
reg [31:0] wb_address = 'd0;
|
| 170 |
|
|
wire rbuf_hit = 'd0;
|
| 171 |
|
|
wire wb_hit;
|
| 172 |
|
|
genvar i;
|
| 173 |
|
|
|
| 174 |
|
|
// ======================================
|
| 175 |
|
|
// Address to use for cache access
|
| 176 |
|
|
// ======================================
|
| 177 |
|
|
// If currently stalled then the address for the next
|
| 178 |
|
|
// cycle will be the same as it is in the current cycle
|
| 179 |
|
|
//
|
| 180 |
|
|
assign address_c = i_core_stall ? i_address : //[CACHE_ADDR32_MSB:CACHE_ADDR32_LSB] :
|
| 181 |
|
|
i_address_nxt; //[CACHE_ADDR32_MSB:CACHE_ADDR32_LSB] ;
|
| 182 |
|
|
|
| 183 |
|
|
assign address = address_c[CACHE_ADDR32_MSB:CACHE_ADDR32_LSB];
|
| 184 |
|
|
|
| 185 |
|
|
// ======================================
|
| 186 |
|
|
// Outputs
|
| 187 |
|
|
// ======================================
|
| 188 |
|
|
assign o_read_data = wb_hit ? i_wb_read_data :
|
| 189 |
|
|
i_address[WORD_SEL_MSB:WORD_SEL_LSB] == 2'd0 ? hit_rdata [31:0] :
|
| 190 |
|
|
i_address[WORD_SEL_MSB:WORD_SEL_LSB] == 2'd1 ? hit_rdata [63:32] :
|
| 191 |
|
|
i_address[WORD_SEL_MSB:WORD_SEL_LSB] == 2'd2 ? hit_rdata [95:64] :
|
| 192 |
|
|
hit_rdata [127:96] ;
|
| 193 |
|
|
|
| 194 |
|
|
// Don't allow the cache to stall the wb i/f for an exclusive access
|
| 195 |
|
|
// The cache needs a couple of cycles to flush a potential copy of the exclusive
|
| 196 |
|
|
// address, but the wb can do the access in parallel. So there is no
|
| 197 |
|
|
// stall in the state CS_EX_DELETE, even though the cache is out of action.
|
| 198 |
|
|
// This works fine as long as the wb is stalling the core
|
| 199 |
|
|
assign o_stall = read_stall || cache_busy_stall;
|
| 200 |
|
|
|
| 201 |
|
|
assign o_wb_req = read_miss && c_state == CS_IDLE;
|
| 202 |
|
|
|
| 203 |
|
|
|
| 204 |
|
|
// ======================================
|
| 205 |
|
|
// Cache State Machine
|
| 206 |
|
|
// ======================================
|
| 207 |
|
|
|
| 208 |
|
|
// Little State Machine to Flush Tag RAMS
|
| 209 |
|
|
always @ ( posedge i_clk )
|
| 210 |
|
|
if ( i_cache_flush )
|
| 211 |
|
|
begin
|
| 212 |
|
|
c_state <= CS_INIT;
|
| 213 |
|
|
source_sel <= 1'd1 << C_INIT;
|
| 214 |
|
|
init_count <= 'd0;
|
| 215 |
|
|
`ifdef A25_CACHE_DEBUG
|
| 216 |
|
|
`TB_DEBUG_MESSAGE
|
| 217 |
|
|
$display("Cache Flush");
|
| 218 |
|
|
`endif
|
| 219 |
|
|
end
|
| 220 |
|
|
else
|
| 221 |
|
|
case ( c_state )
|
| 222 |
|
|
CS_INIT :
|
| 223 |
|
|
if ( init_count < CACHE_LINES [CACHE_ADDR_WIDTH:0] )
|
| 224 |
|
|
begin
|
| 225 |
|
|
init_count <= init_count + 1'd1;
|
| 226 |
|
|
source_sel <= 1'd1 << C_INIT;
|
| 227 |
|
|
end
|
| 228 |
|
|
else
|
| 229 |
|
|
begin
|
| 230 |
|
|
source_sel <= 1'd1 << C_CORE;
|
| 231 |
|
|
c_state <= CS_TURN_AROUND;
|
| 232 |
|
|
end
|
| 233 |
|
|
|
| 234 |
|
|
CS_IDLE :
|
| 235 |
|
|
begin
|
| 236 |
|
|
source_sel <= 1'd1 << C_CORE;
|
| 237 |
|
|
|
| 238 |
|
|
if ( read_miss )
|
| 239 |
|
|
c_state <= CS_FILL0;
|
| 240 |
|
|
end
|
| 241 |
|
|
|
| 242 |
|
|
CS_FILL0 :
|
| 243 |
|
|
begin
|
| 244 |
|
|
// wb read request asserted, wait for ack
|
| 245 |
|
|
if ( i_wb_ready )
|
| 246 |
|
|
c_state <= CS_FILL1;
|
| 247 |
|
|
end
|
| 248 |
|
|
|
| 249 |
|
|
CS_FILL1 :
|
| 250 |
|
|
begin
|
| 251 |
|
|
// wb read request asserted, wait for ack
|
| 252 |
|
|
if ( i_wb_ready )
|
| 253 |
|
|
c_state <= CS_FILL2;
|
| 254 |
|
|
end
|
| 255 |
|
|
|
| 256 |
|
|
|
| 257 |
|
|
CS_FILL2 :
|
| 258 |
|
|
// first read of burst of 4
|
| 259 |
|
|
// wb read request asserted, wait for ack
|
| 260 |
|
|
if ( i_wb_ready )
|
| 261 |
|
|
c_state <= CS_FILL3;
|
| 262 |
|
|
|
| 263 |
|
|
|
| 264 |
|
|
CS_FILL3 :
|
| 265 |
|
|
begin
|
| 266 |
|
|
select_way <= next_way;
|
| 267 |
|
|
random_num <= {random_num[2], random_num[1], random_num[0],
|
| 268 |
|
|
random_num[3]^random_num[2]};
|
| 269 |
|
|
|
| 270 |
|
|
// third read of burst of 4
|
| 271 |
|
|
// wb read request asserted, wait for ack
|
| 272 |
|
|
if ( i_wb_ready )
|
| 273 |
|
|
begin
|
| 274 |
|
|
c_state <= CS_FILL_COMPLETE;
|
| 275 |
|
|
|
| 276 |
|
|
// Pick a way to write the cache update into
|
| 277 |
|
|
// Either pick one of the invalid caches, or if all are valid, then pick
|
| 278 |
|
|
// one randomly
|
| 279 |
|
|
|
| 280 |
|
|
end
|
| 281 |
|
|
end
|
| 282 |
|
|
|
| 283 |
|
|
|
| 284 |
|
|
// Write the read fetch data in this cycle
|
| 285 |
|
|
CS_FILL_COMPLETE :
|
| 286 |
|
|
begin
|
| 287 |
|
|
// Back to normal cache operations, but
|
| 288 |
|
|
// use physical address for first read as
|
| 289 |
|
|
// address moved before the stall was asserted for the read_miss
|
| 290 |
|
|
// However don't use it if its a non-cached address!
|
| 291 |
|
|
source_sel <= 1'd1 << C_CORE;
|
| 292 |
|
|
c_state <= CS_TURN_AROUND;
|
| 293 |
|
|
end
|
| 294 |
|
|
|
| 295 |
|
|
|
| 296 |
|
|
// Ignore the tag read data in this cycle
|
| 297 |
|
|
// Wait 1 cycle to pre-read the cache and return to normal operation
|
| 298 |
|
|
CS_TURN_AROUND :
|
| 299 |
|
|
begin
|
| 300 |
|
|
c_state <= CS_IDLE;
|
| 301 |
|
|
end
|
| 302 |
|
|
|
| 303 |
|
|
endcase
|
| 304 |
|
|
|
| 305 |
|
|
|
| 306 |
|
|
// ======================================
|
| 307 |
|
|
// Capture WB Block Read - burst of 4 words
|
| 308 |
|
|
// ======================================
|
| 309 |
|
|
assign wb_rdata_burst = {i_wb_read_data, wb_rdata_burst_r[127:32]};
|
| 310 |
|
|
always @ ( posedge i_clk )
|
| 311 |
|
|
if ( i_wb_ready )
|
| 312 |
|
|
wb_rdata_burst_r <= wb_rdata_burst;
|
| 313 |
|
|
|
| 314 |
|
|
|
| 315 |
|
|
// ======================================
|
| 316 |
|
|
// Miss Address
|
| 317 |
|
|
// ======================================
|
| 318 |
|
|
always @ ( posedge i_clk )
|
| 319 |
|
|
if ( c_state == CS_IDLE )
|
| 320 |
|
|
miss_address <= i_address;
|
| 321 |
|
|
|
| 322 |
|
|
|
| 323 |
|
|
always @ ( posedge i_clk )
|
| 324 |
|
|
address_r <= address_c;
|
| 325 |
|
|
|
| 326 |
|
|
assign invalid_read = address_r != i_address;
|
| 327 |
|
|
|
| 328 |
|
|
|
| 329 |
|
|
always @(posedge i_clk)
|
| 330 |
|
|
if ( o_wb_req )
|
| 331 |
|
|
wb_address <= i_address;
|
| 332 |
|
|
else if ( i_wb_ready && fill_state )
|
| 333 |
|
|
wb_address <= {wb_address[31:4], wb_address[3:2] + 1'd1, 2'd0};
|
| 334 |
|
|
|
| 335 |
|
|
assign fill_state = c_state == CS_FILL0 || c_state == CS_FILL1 || c_state == CS_FILL2 || c_state == CS_FILL3 ;
|
| 336 |
|
|
assign wb_hit = i_address == wb_address && i_wb_ready && fill_state;
|
| 337 |
|
|
|
| 338 |
|
|
assign tag_address = read_miss_fill ? miss_address [CACHE_ADDR32_MSB:CACHE_ADDR32_LSB] :
|
| 339 |
|
|
source_sel[C_INIT] ? init_count[CACHE_ADDR_WIDTH-1:0] :
|
| 340 |
|
|
address ;
|
| 341 |
|
|
|
| 342 |
|
|
|
| 343 |
|
|
assign data_address = read_miss_fill ? miss_address[CACHE_ADDR32_MSB:CACHE_ADDR32_LSB] :
|
| 344 |
|
|
address ;
|
| 345 |
|
|
|
| 346 |
|
|
|
| 347 |
|
|
assign tag_wdata = read_miss_fill ? {1'd1, miss_address[31:TAG_ADDR32_LSB]} :
|
| 348 |
|
|
{TAG_WIDTH{1'd0}} ;
|
| 349 |
|
|
|
| 350 |
|
|
|
| 351 |
|
|
// Data comes in off the WB bus in wrap4 with the missed data word first
|
| 352 |
|
|
assign data_wdata = miss_address[3:2] == 2'd0 ? { wb_rdata_burst[127:0] }:
|
| 353 |
|
|
miss_address[3:2] == 2'd1 ? { wb_rdata_burst[95:0], wb_rdata_burst[127:96] }:
|
| 354 |
|
|
miss_address[3:2] == 2'd2 ? { wb_rdata_burst[63:0], wb_rdata_burst[127:64] }:
|
| 355 |
|
|
{ wb_rdata_burst[31:0], wb_rdata_burst[127:32] };
|
| 356 |
|
|
|
| 357 |
|
|
|
| 358 |
|
|
assign read_miss_fill = c_state == CS_FILL3 && i_wb_ready;
|
| 359 |
|
|
|
| 360 |
|
|
|
| 361 |
|
|
|
| 362 |
|
|
assign tag_wenable = read_miss_fill ? 1'd1 :
|
| 363 |
|
|
source_sel[C_INVA] ? 1'd1 :
|
| 364 |
|
|
source_sel[C_FILL] ? 1'd1 :
|
| 365 |
|
|
source_sel[C_INIT] ? 1'd1 :
|
| 366 |
|
|
source_sel[C_CORE] ? 1'd0 :
|
| 367 |
|
|
1'd0 ;
|
| 368 |
|
|
|
| 369 |
|
|
|
| 370 |
|
|
assign enable = i_select && i_cache_enable;
|
| 371 |
|
|
|
| 372 |
|
|
assign idle_hit = |data_hit_way;
|
| 373 |
|
|
|
| 374 |
|
|
assign read_miss = enable && !idle_hit && !invalid_read;
|
| 375 |
|
|
|
| 376 |
|
|
assign read_stall = enable && !idle_hit && !rbuf_hit && !wb_hit;
|
| 377 |
|
|
|
| 378 |
|
|
assign cache_busy_stall = (c_state == CS_TURN_AROUND && enable) || c_state == CS_INIT;
|
| 379 |
|
|
|
| 380 |
|
|
|
| 381 |
|
|
// ======================================
|
| 382 |
|
|
// Instantiate RAMS
|
| 383 |
|
|
// ======================================
|
| 384 |
|
|
|
| 385 |
|
|
generate
|
| 386 |
|
|
for ( i=0; i<WAYS;i=i+1 ) begin : rams
|
| 387 |
|
|
|
| 388 |
|
|
// Tag RAMs
|
| 389 |
|
|
`ifdef XILINX_SPARTAN6_FPGA
|
| 390 |
|
|
xs6_sram_256x21_line_en
|
| 391 |
|
|
`endif
|
| 392 |
|
|
|
| 393 |
|
|
`ifdef XILINX_VIRTEX6_FPGA
|
| 394 |
|
|
xv6_sram_256x21_line_en
|
| 395 |
|
|
`endif
|
| 396 |
|
|
|
| 397 |
|
|
`ifndef XILINX_FPGA
|
| 398 |
|
|
generic_sram_line_en
|
| 399 |
|
|
`endif
|
| 400 |
|
|
|
| 401 |
|
|
#(
|
| 402 |
|
|
.DATA_WIDTH ( TAG_WIDTH ),
|
| 403 |
|
|
.INITIALIZE_TO_ZERO ( 1 ),
|
| 404 |
|
|
.ADDRESS_WIDTH ( CACHE_ADDR_WIDTH ))
|
| 405 |
|
|
u_tag (
|
| 406 |
|
|
.i_clk ( i_clk ),
|
| 407 |
|
|
.i_write_data ( tag_wdata ),
|
| 408 |
|
|
.i_write_enable ( tag_wenable_way[i] ),
|
| 409 |
|
|
.i_address ( tag_address ),
|
| 410 |
|
|
|
| 411 |
|
|
.o_read_data ( tag_rdata_way[i] )
|
| 412 |
|
|
);
|
| 413 |
|
|
|
| 414 |
|
|
// Data RAMs
|
| 415 |
|
|
`ifdef XILINX_SPARTAN6_FPGA
|
| 416 |
|
|
xs6_sram_256x128_byte_en
|
| 417 |
|
|
`endif
|
| 418 |
|
|
|
| 419 |
|
|
`ifdef XILINX_VIRTEX6_FPGA
|
| 420 |
|
|
xv6_sram_256x128_byte_en
|
| 421 |
|
|
`endif
|
| 422 |
|
|
|
| 423 |
|
|
`ifndef XILINX_FPGA
|
| 424 |
|
|
generic_sram_byte_en
|
| 425 |
|
|
`endif
|
| 426 |
|
|
|
| 427 |
|
|
#(
|
| 428 |
|
|
.DATA_WIDTH ( CACHE_LINE_WIDTH) ,
|
| 429 |
|
|
.ADDRESS_WIDTH ( CACHE_ADDR_WIDTH) )
|
| 430 |
|
|
u_data (
|
| 431 |
|
|
.i_clk ( i_clk ),
|
| 432 |
|
|
.i_write_data ( data_wdata ),
|
| 433 |
|
|
.i_write_enable ( data_wenable_way[i] ),
|
| 434 |
|
|
.i_address ( data_address ),
|
| 435 |
|
|
.i_byte_enable ( {CACHE_LINE_WIDTH/8{1'd1}} ),
|
| 436 |
|
|
.o_read_data ( data_rdata_way[i] )
|
| 437 |
|
|
);
|
| 438 |
|
|
|
| 439 |
|
|
|
| 440 |
|
|
// Per tag-ram write-enable
|
| 441 |
|
|
assign tag_wenable_way[i] = tag_wenable && ( select_way[i] || source_sel[C_INIT] );
|
| 442 |
|
|
|
| 443 |
|
|
// Per data-ram write-enable
|
| 444 |
|
|
assign data_wenable_way[i] = ( source_sel[C_FILL] || read_miss_fill ) && select_way[i];
|
| 445 |
|
|
|
| 446 |
|
|
// Per data-ram idle_hit flag
|
| 447 |
|
|
assign data_hit_way[i] = tag_rdata_way[i][TAG_WIDTH-1] &&
|
| 448 |
|
|
tag_rdata_way[i][TAG_ADDR_WIDTH-1:0] == i_address[31:TAG_ADDR32_LSB] &&
|
| 449 |
|
|
c_state == CS_IDLE;
|
| 450 |
|
|
end
|
| 451 |
|
|
endgenerate
|
| 452 |
|
|
|
| 453 |
|
|
|
| 454 |
|
|
// ======================================
|
| 455 |
|
|
// Register Valid Bits
|
| 456 |
|
|
// ======================================
|
| 457 |
|
|
generate
|
| 458 |
|
|
if ( WAYS == 2 ) begin : valid_bits_2ways
|
| 459 |
|
|
|
| 460 |
|
|
always @ ( posedge i_clk )
|
| 461 |
|
|
if ( c_state == CS_IDLE )
|
| 462 |
|
|
valid_bits_r <= {tag_rdata_way[1][TAG_WIDTH-1],
|
| 463 |
|
|
tag_rdata_way[0][TAG_WIDTH-1]};
|
| 464 |
|
|
|
| 465 |
|
|
end
|
| 466 |
|
|
else if ( WAYS == 3 ) begin : valid_bits_3ways
|
| 467 |
|
|
|
| 468 |
|
|
always @ ( posedge i_clk )
|
| 469 |
|
|
if ( c_state == CS_IDLE )
|
| 470 |
|
|
valid_bits_r <= {tag_rdata_way[2][TAG_WIDTH-1],
|
| 471 |
|
|
tag_rdata_way[1][TAG_WIDTH-1],
|
| 472 |
|
|
tag_rdata_way[0][TAG_WIDTH-1]};
|
| 473 |
|
|
|
| 474 |
|
|
end
|
| 475 |
|
|
else if ( WAYS == 4 ) begin : valid_bits_4ways
|
| 476 |
|
|
|
| 477 |
|
|
always @ ( posedge i_clk )
|
| 478 |
|
|
if ( c_state == CS_IDLE )
|
| 479 |
|
|
valid_bits_r <= {tag_rdata_way[3][TAG_WIDTH-1],
|
| 480 |
|
|
tag_rdata_way[2][TAG_WIDTH-1],
|
| 481 |
|
|
tag_rdata_way[1][TAG_WIDTH-1],
|
| 482 |
|
|
tag_rdata_way[0][TAG_WIDTH-1]};
|
| 483 |
|
|
|
| 484 |
|
|
end
|
| 485 |
|
|
else begin : valid_bits_8ways
|
| 486 |
|
|
|
| 487 |
|
|
always @ ( posedge i_clk )
|
| 488 |
|
|
if ( c_state == CS_IDLE )
|
| 489 |
|
|
valid_bits_r <= {tag_rdata_way[7][TAG_WIDTH-1],
|
| 490 |
|
|
tag_rdata_way[6][TAG_WIDTH-1],
|
| 491 |
|
|
tag_rdata_way[5][TAG_WIDTH-1],
|
| 492 |
|
|
tag_rdata_way[4][TAG_WIDTH-1],
|
| 493 |
|
|
tag_rdata_way[3][TAG_WIDTH-1],
|
| 494 |
|
|
tag_rdata_way[2][TAG_WIDTH-1],
|
| 495 |
|
|
tag_rdata_way[1][TAG_WIDTH-1],
|
| 496 |
|
|
tag_rdata_way[0][TAG_WIDTH-1]};
|
| 497 |
|
|
|
| 498 |
|
|
end
|
| 499 |
|
|
endgenerate
|
| 500 |
|
|
|
| 501 |
|
|
|
| 502 |
|
|
// ======================================
|
| 503 |
|
|
// Select read hit data
|
| 504 |
|
|
// ======================================
|
| 505 |
|
|
generate
|
| 506 |
|
|
if ( WAYS == 2 ) begin : read_data_2ways
|
| 507 |
|
|
|
| 508 |
|
|
assign hit_rdata = data_hit_way[0] ? data_rdata_way[0] :
|
| 509 |
|
|
data_hit_way[1] ? data_rdata_way[1] :
|
| 510 |
|
|
{CACHE_LINE_WIDTH{1'd1}} ; // all 1's for debug
|
| 511 |
|
|
|
| 512 |
|
|
end
|
| 513 |
|
|
else if ( WAYS == 3 ) begin : read_data_3ways
|
| 514 |
|
|
|
| 515 |
|
|
assign hit_rdata = data_hit_way[0] ? data_rdata_way[0] :
|
| 516 |
|
|
data_hit_way[1] ? data_rdata_way[1] :
|
| 517 |
|
|
data_hit_way[2] ? data_rdata_way[2] :
|
| 518 |
|
|
{CACHE_LINE_WIDTH{1'd1}} ; // all 1's for debug
|
| 519 |
|
|
|
| 520 |
|
|
end
|
| 521 |
|
|
else if ( WAYS == 4 ) begin : read_data_4ways
|
| 522 |
|
|
|
| 523 |
|
|
assign hit_rdata = data_hit_way[0] ? data_rdata_way[0] :
|
| 524 |
|
|
data_hit_way[1] ? data_rdata_way[1] :
|
| 525 |
|
|
data_hit_way[2] ? data_rdata_way[2] :
|
| 526 |
|
|
data_hit_way[3] ? data_rdata_way[3] :
|
| 527 |
|
|
{CACHE_LINE_WIDTH{1'd1}} ; // all 1's for debug
|
| 528 |
|
|
|
| 529 |
|
|
end
|
| 530 |
|
|
else begin : read_data_8ways
|
| 531 |
|
|
|
| 532 |
|
|
assign hit_rdata = data_hit_way[0] ? data_rdata_way[0] :
|
| 533 |
|
|
data_hit_way[1] ? data_rdata_way[1] :
|
| 534 |
|
|
data_hit_way[2] ? data_rdata_way[2] :
|
| 535 |
|
|
data_hit_way[3] ? data_rdata_way[3] :
|
| 536 |
|
|
data_hit_way[4] ? data_rdata_way[4] :
|
| 537 |
|
|
data_hit_way[5] ? data_rdata_way[5] :
|
| 538 |
|
|
data_hit_way[6] ? data_rdata_way[6] :
|
| 539 |
|
|
data_hit_way[7] ? data_rdata_way[7] :
|
| 540 |
|
|
{CACHE_LINE_WIDTH{1'd1}} ; // all 1's for debug
|
| 541 |
|
|
|
| 542 |
|
|
end
|
| 543 |
|
|
endgenerate
|
| 544 |
|
|
|
| 545 |
|
|
|
| 546 |
|
|
// ======================================
|
| 547 |
|
|
// Function to select the way to use
|
| 548 |
|
|
// for fills
|
| 549 |
|
|
// ======================================
|
| 550 |
|
|
generate
|
| 551 |
|
|
if ( WAYS == 2 ) begin : pick_way_2ways
|
| 552 |
|
|
|
| 553 |
|
|
assign next_way = pick_way ( valid_bits_r, random_num );
|
| 554 |
|
|
|
| 555 |
|
|
function [WAYS-1:0] pick_way;
|
| 556 |
|
|
input [WAYS-1:0] valid_bits;
|
| 557 |
|
|
input [3:0] random_num;
|
| 558 |
|
|
begin
|
| 559 |
|
|
if ( valid_bits[0] == 1'd0 )
|
| 560 |
|
|
// way 0 not occupied so use it
|
| 561 |
|
|
pick_way = 2'b01;
|
| 562 |
|
|
else if ( valid_bits[1] == 1'd0 )
|
| 563 |
|
|
// way 1 not occupied so use it
|
| 564 |
|
|
pick_way = 2'b10;
|
| 565 |
|
|
else
|
| 566 |
|
|
begin
|
| 567 |
|
|
// All ways occupied so pick one randomly
|
| 568 |
|
|
case (random_num[3:1])
|
| 569 |
|
|
3'd0, 3'd3,
|
| 570 |
|
|
3'd5, 3'd6: pick_way = 2'b10;
|
| 571 |
|
|
default: pick_way = 2'b01;
|
| 572 |
|
|
endcase
|
| 573 |
|
|
end
|
| 574 |
|
|
end
|
| 575 |
|
|
endfunction
|
| 576 |
|
|
|
| 577 |
|
|
end
|
| 578 |
|
|
else if ( WAYS == 3 ) begin : pick_way_3ways
|
| 579 |
|
|
|
| 580 |
|
|
assign next_way = pick_way ( valid_bits_r, random_num );
|
| 581 |
|
|
|
| 582 |
|
|
function [WAYS-1:0] pick_way;
|
| 583 |
|
|
input [WAYS-1:0] valid_bits;
|
| 584 |
|
|
input [3:0] random_num;
|
| 585 |
|
|
begin
|
| 586 |
|
|
if ( valid_bits[0] == 1'd0 )
|
| 587 |
|
|
// way 0 not occupied so use it
|
| 588 |
|
|
pick_way = 3'b001;
|
| 589 |
|
|
else if ( valid_bits[1] == 1'd0 )
|
| 590 |
|
|
// way 1 not occupied so use it
|
| 591 |
|
|
pick_way = 3'b010;
|
| 592 |
|
|
else if ( valid_bits[2] == 1'd0 )
|
| 593 |
|
|
// way 2 not occupied so use it
|
| 594 |
|
|
pick_way = 3'b100;
|
| 595 |
|
|
else
|
| 596 |
|
|
begin
|
| 597 |
|
|
// All ways occupied so pick one randomly
|
| 598 |
|
|
case (random_num[3:1])
|
| 599 |
|
|
3'd0, 3'd1, 3'd2: pick_way = 3'b010;
|
| 600 |
|
|
3'd2, 3'd3, 3'd4: pick_way = 3'b100;
|
| 601 |
|
|
default: pick_way = 3'b001;
|
| 602 |
|
|
endcase
|
| 603 |
|
|
end
|
| 604 |
|
|
end
|
| 605 |
|
|
endfunction
|
| 606 |
|
|
|
| 607 |
|
|
end
|
| 608 |
|
|
else if ( WAYS == 4 ) begin : pick_way_4ways
|
| 609 |
|
|
|
| 610 |
|
|
assign next_way = pick_way ( valid_bits_r, random_num );
|
| 611 |
|
|
|
| 612 |
|
|
function [WAYS-1:0] pick_way;
|
| 613 |
|
|
input [WAYS-1:0] valid_bits;
|
| 614 |
|
|
input [3:0] random_num;
|
| 615 |
|
|
begin
|
| 616 |
|
|
if ( valid_bits[0] == 1'd0 )
|
| 617 |
|
|
// way 0 not occupied so use it
|
| 618 |
|
|
pick_way = 4'b0001;
|
| 619 |
|
|
else if ( valid_bits[1] == 1'd0 )
|
| 620 |
|
|
// way 1 not occupied so use it
|
| 621 |
|
|
pick_way = 4'b0010;
|
| 622 |
|
|
else if ( valid_bits[2] == 1'd0 )
|
| 623 |
|
|
// way 2 not occupied so use it
|
| 624 |
|
|
pick_way = 4'b0100;
|
| 625 |
|
|
else if ( valid_bits[3] == 1'd0 )
|
| 626 |
|
|
// way 3 not occupied so use it
|
| 627 |
|
|
pick_way = 4'b1000;
|
| 628 |
|
|
else
|
| 629 |
|
|
begin
|
| 630 |
|
|
// All ways occupied so pick one randomly
|
| 631 |
|
|
case (random_num[3:1])
|
| 632 |
|
|
3'd0, 3'd1: pick_way = 4'b0100;
|
| 633 |
|
|
3'd2, 3'd3: pick_way = 4'b1000;
|
| 634 |
|
|
3'd4, 3'd5: pick_way = 4'b0001;
|
| 635 |
|
|
default: pick_way = 4'b0010;
|
| 636 |
|
|
endcase
|
| 637 |
|
|
end
|
| 638 |
|
|
end
|
| 639 |
|
|
endfunction
|
| 640 |
|
|
|
| 641 |
|
|
end
|
| 642 |
|
|
else begin : pick_way_8ways
|
| 643 |
|
|
|
| 644 |
|
|
assign next_way = pick_way ( valid_bits_r, random_num );
|
| 645 |
|
|
|
| 646 |
|
|
function [WAYS-1:0] pick_way;
|
| 647 |
|
|
input [WAYS-1:0] valid_bits;
|
| 648 |
|
|
input [3:0] random_num;
|
| 649 |
|
|
begin
|
| 650 |
|
|
if ( valid_bits[0] == 1'd0 )
|
| 651 |
|
|
// way 0 not occupied so use it
|
| 652 |
|
|
pick_way = 8'b00000001;
|
| 653 |
|
|
else if ( valid_bits[1] == 1'd0 )
|
| 654 |
|
|
// way 1 not occupied so use it
|
| 655 |
|
|
pick_way = 8'b00000010;
|
| 656 |
|
|
else if ( valid_bits[2] == 1'd0 )
|
| 657 |
|
|
// way 2 not occupied so use it
|
| 658 |
|
|
pick_way = 8'b00000100;
|
| 659 |
|
|
else if ( valid_bits[3] == 1'd0 )
|
| 660 |
|
|
// way 3 not occupied so use it
|
| 661 |
|
|
pick_way = 8'b00001000;
|
| 662 |
|
|
else if ( valid_bits[4] == 1'd0 )
|
| 663 |
|
|
// way 3 not occupied so use it
|
| 664 |
|
|
pick_way = 8'b00010000;
|
| 665 |
|
|
else if ( valid_bits[5] == 1'd0 )
|
| 666 |
|
|
// way 3 not occupied so use it
|
| 667 |
|
|
pick_way = 8'b00100000;
|
| 668 |
|
|
else if ( valid_bits[6] == 1'd0 )
|
| 669 |
|
|
// way 3 not occupied so use it
|
| 670 |
|
|
pick_way = 8'b01000000;
|
| 671 |
|
|
else if ( valid_bits[7] == 1'd0 )
|
| 672 |
|
|
// way 3 not occupied so use it
|
| 673 |
|
|
pick_way = 8'b10000000;
|
| 674 |
|
|
else
|
| 675 |
|
|
begin
|
| 676 |
|
|
// All ways occupied so pick one randomly
|
| 677 |
|
|
case (random_num[3:1])
|
| 678 |
|
|
3'd0: pick_way = 8'b00010000;
|
| 679 |
|
|
3'd1: pick_way = 8'b00100000;
|
| 680 |
|
|
3'd2: pick_way = 8'b01000000;
|
| 681 |
|
|
3'd3: pick_way = 8'b10000000;
|
| 682 |
|
|
3'd4: pick_way = 8'b00000001;
|
| 683 |
|
|
3'd5: pick_way = 8'b00000010;
|
| 684 |
|
|
3'd6: pick_way = 8'b00000100;
|
| 685 |
|
|
default: pick_way = 8'b00001000;
|
| 686 |
|
|
endcase
|
| 687 |
|
|
end
|
| 688 |
|
|
end
|
| 689 |
|
|
endfunction
|
| 690 |
|
|
|
| 691 |
|
|
end
|
| 692 |
|
|
endgenerate
|
| 693 |
|
|
|
| 694 |
|
|
|
| 695 |
|
|
// ========================================================
|
| 696 |
|
|
// Debug WB bus - not synthesizable
|
| 697 |
|
|
// ========================================================
|
| 698 |
|
|
//synopsys translate_off
|
| 699 |
|
|
wire [(6*8)-1:0] xSOURCE_SEL;
|
| 700 |
|
|
wire [(20*8)-1:0] xC_STATE;
|
| 701 |
|
|
|
| 702 |
|
|
assign xSOURCE_SEL = source_sel[C_CORE] ? "C_CORE" :
|
| 703 |
|
|
source_sel[C_INIT] ? "C_INIT" :
|
| 704 |
|
|
source_sel[C_FILL] ? "C_FILL" :
|
| 705 |
|
|
source_sel[C_INVA] ? "C_INVA" :
|
| 706 |
|
|
"UNKNON" ;
|
| 707 |
|
|
|
| 708 |
|
|
assign xC_STATE = c_state == CS_INIT ? "CS_INIT" :
|
| 709 |
|
|
c_state == CS_IDLE ? "CS_IDLE" :
|
| 710 |
|
|
c_state == CS_FILL0 ? "CS_FILL0" :
|
| 711 |
|
|
c_state == CS_FILL1 ? "CS_FILL1" :
|
| 712 |
|
|
c_state == CS_FILL2 ? "CS_FILL2" :
|
| 713 |
|
|
c_state == CS_FILL3 ? "CS_FILL3" :
|
| 714 |
|
|
c_state == CS_FILL4 ? "CS_FILL4" :
|
| 715 |
|
|
c_state == CS_FILL_COMPLETE ? "CS_FILL_COMPLETE" :
|
| 716 |
|
|
c_state == CS_EX_DELETE ? "CS_EX_DELETE" :
|
| 717 |
|
|
c_state == CS_TURN_AROUND ? "CS_TURN_AROUND" :
|
| 718 |
|
|
c_state == CS_WRITE_HIT1 ? "CS_WRITE_HIT1" :
|
| 719 |
|
|
"UNKNOWN" ;
|
| 720 |
|
|
|
| 721 |
|
|
|
| 722 |
|
|
generate
|
| 723 |
|
|
if ( WAYS == 2 ) begin : check_hit_2ways
|
| 724 |
|
|
|
| 725 |
|
|
always @( posedge i_clk )
|
| 726 |
|
|
if ( (data_hit_way[0] + data_hit_way[1] ) > 4'd1 )
|
| 727 |
|
|
begin
|
| 728 |
|
|
`TB_ERROR_MESSAGE
|
| 729 |
|
|
$display("Hit in more than one cache ways!");
|
| 730 |
|
|
end
|
| 731 |
|
|
|
| 732 |
|
|
end
|
| 733 |
|
|
else if ( WAYS == 3 ) begin : check_hit_3ways
|
| 734 |
|
|
|
| 735 |
|
|
always @( posedge i_clk )
|
| 736 |
|
|
if ( (data_hit_way[0] + data_hit_way[1] + data_hit_way[2] ) > 4'd1 )
|
| 737 |
|
|
begin
|
| 738 |
|
|
`TB_ERROR_MESSAGE
|
| 739 |
|
|
$display("Hit in more than one cache ways!");
|
| 740 |
|
|
end
|
| 741 |
|
|
|
| 742 |
|
|
end
|
| 743 |
|
|
else if ( WAYS == 4 ) begin : check_hit_4ways
|
| 744 |
|
|
|
| 745 |
|
|
always @( posedge i_clk )
|
| 746 |
|
|
if ( (data_hit_way[0] + data_hit_way[1] +
|
| 747 |
|
|
data_hit_way[2] + data_hit_way[3] ) > 4'd1 )
|
| 748 |
|
|
begin
|
| 749 |
|
|
`TB_ERROR_MESSAGE
|
| 750 |
|
|
$display("Hit in more than one cache ways!");
|
| 751 |
|
|
end
|
| 752 |
|
|
|
| 753 |
|
|
end
|
| 754 |
|
|
else if ( WAYS == 8 ) begin : check_hit_8ways
|
| 755 |
|
|
|
| 756 |
|
|
always @( posedge i_clk )
|
| 757 |
|
|
if ( (data_hit_way[0] + data_hit_way[1] +
|
| 758 |
|
|
data_hit_way[2] + data_hit_way[3] +
|
| 759 |
|
|
data_hit_way[4] + data_hit_way[5] +
|
| 760 |
|
|
data_hit_way[6] + data_hit_way[7] ) > 4'd1 )
|
| 761 |
|
|
begin
|
| 762 |
|
|
`TB_ERROR_MESSAGE
|
| 763 |
|
|
$display("Hit in more than one cache ways!");
|
| 764 |
|
|
end
|
| 765 |
|
|
|
| 766 |
|
|
end
|
| 767 |
|
|
else begin : check_hit_nways
|
| 768 |
|
|
|
| 769 |
|
|
initial
|
| 770 |
|
|
begin
|
| 771 |
|
|
`TB_ERROR_MESSAGE
|
| 772 |
|
|
$display("Unsupported number of ways %0d", WAYS);
|
| 773 |
|
|
$display("Set A25_ICACHE_WAYS in a25_config_defines.v to either 2,3,4 or 8");
|
| 774 |
|
|
end
|
| 775 |
|
|
|
| 776 |
|
|
end
|
| 777 |
|
|
endgenerate
|
| 778 |
|
|
|
| 779 |
|
|
//synopsys translate_on
|
| 780 |
|
|
|
| 781 |
|
|
endmodule
|
| 782 |
|
|
|
