[6] | 1 | `timescale 1ns / 1ps |
---|
| 2 | ////////////////////////////////////////////////////////////////////////////////// |
---|
| 3 | // Company: (C) Athree, 2009 |
---|
| 4 | // Engineer: Dmitry Rozhdestvenskiy |
---|
| 5 | // Email [email protected] [email protected] [email protected] |
---|
| 6 | // |
---|
| 7 | // Design Name: Bridge from SPARC Core to Wishbone Master |
---|
| 8 | // Module Name: os2wb |
---|
| 9 | // Project Name: SPARC SoC single-core |
---|
| 10 | // |
---|
| 11 | // LICENSE: |
---|
| 12 | // This is a Free Hardware Design; you can redistribute it and/or |
---|
| 13 | // modify it under the terms of the GNU General Public License |
---|
| 14 | // version 2 as published by the Free Software Foundation. |
---|
| 15 | // The above named program is distributed in the hope that it will |
---|
| 16 | // be useful, but WITHOUT ANY WARRANTY; without even the implied |
---|
| 17 | // warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. |
---|
| 18 | // See the GNU General Public License for more details. |
---|
| 19 | // |
---|
| 20 | ////////////////////////////////////////////////////////////////////////////////// |
---|
| 21 | module os2wb_dual( |
---|
| 22 | input clk, |
---|
| 23 | input rstn, |
---|
| 24 | |
---|
| 25 | // Core interface |
---|
| 26 | input [ 4:0] pcx_req, |
---|
| 27 | input pcx_atom, |
---|
| 28 | input [123:0] pcx_data, |
---|
| 29 | output reg [ 4:0] pcx_grant, |
---|
| 30 | output reg cpx_ready, |
---|
| 31 | output reg [144:0] cpx_packet, |
---|
| 32 | |
---|
| 33 | // Core 2nd interface |
---|
| 34 | input [ 4:0] pcx1_req, |
---|
| 35 | input pcx1_atom, |
---|
| 36 | input [123:0] pcx1_data, |
---|
| 37 | output reg [ 4:0] pcx1_grant, |
---|
| 38 | output reg cpx1_ready, |
---|
| 39 | output reg [144:0] cpx1_packet, |
---|
| 40 | |
---|
| 41 | // Wishbone master interface |
---|
| 42 | input [ 63:0] wb_data_i, |
---|
| 43 | input wb_ack, |
---|
| 44 | output reg wb_cycle, |
---|
| 45 | output reg wb_strobe, |
---|
| 46 | output reg wb_we, |
---|
| 47 | output reg [ 7:0] wb_sel, |
---|
| 48 | output reg [ 63:0] wb_addr, |
---|
| 49 | output reg [ 63:0] wb_data_o, |
---|
| 50 | |
---|
| 51 | // FPU interface |
---|
| 52 | output reg [123:0] fp_pcx, |
---|
| 53 | output reg fp_req, |
---|
| 54 | input [144:0] fp_cpx, |
---|
| 55 | input fp_rdy, |
---|
| 56 | |
---|
| 57 | // Ethernet interrupt, sensed on posedge, mapped to vector 'd29 |
---|
| 58 | input eth_int |
---|
| 59 | ); |
---|
| 60 | |
---|
| 61 | reg [123:0] pcx_packet_d; // Latched incoming PCX packet |
---|
| 62 | reg [123:0] pcx_packet_2nd; // Second packet for atomic (CAS) |
---|
| 63 | reg [ 4:0] pcx_req_d; // Latched request |
---|
| 64 | reg pcx_atom_d; // Latched atomic flasg |
---|
| 65 | reg [ 4:0] state; // FSM state |
---|
| 66 | reg [144:0] cpx_packet_1; // First CPX packet |
---|
| 67 | reg [144:0] cpx_packet_2; // Second CPX packet (for atomics and cached IFILLs) |
---|
| 68 | reg cpx_two_packet; // CPX answer is two-packet (!=atomic, SWAP has atomic==0 and answer is two-packet) |
---|
| 69 | |
---|
| 70 | wire [111:0] inval_vect0; // Invalidate, instr/data, way |
---|
| 71 | wire [111:0] inval_vect1; // IFill may cause two D lines invalidation at a time |
---|
| 72 | |
---|
| 73 | wire [1:0] othercachehit; |
---|
| 74 | wire [1:0] othercpuhit; |
---|
| 75 | wire [1:0] wayval0; |
---|
| 76 | wire [1:0] wayval1; |
---|
| 77 | |
---|
| 78 | `define TEST_DRAM_1 5'b00000 |
---|
| 79 | `define TEST_DRAM_2 5'b00001 |
---|
| 80 | `define TEST_DRAM_3 5'b00010 |
---|
| 81 | `define TEST_DRAM_4 5'b00011 |
---|
| 82 | `define INIT_DRAM_1 5'b00100 |
---|
| 83 | `define INIT_DRAM_2 5'b00101 |
---|
| 84 | `define WAKEUP 5'b00110 |
---|
| 85 | `define PCX_IDLE 5'b00111 |
---|
| 86 | `define GOT_PCX_REQ 5'b01000 |
---|
| 87 | `define PCX_REQ_2ND 5'b01001 |
---|
| 88 | `define PCX_REQ_STEP1 5'b01010 |
---|
| 89 | `define PCX_REQ_STEP1_1 5'b01011 |
---|
| 90 | `define PCX_REQ_STEP2 5'b01100 |
---|
| 91 | `define PCX_REQ_STEP2_1 5'b01101 |
---|
| 92 | `define PCX_REQ_STEP3 5'b01110 |
---|
| 93 | `define PCX_REQ_STEP3_1 5'b01111 |
---|
| 94 | `define PCX_REQ_STEP4 5'b10000 |
---|
| 95 | `define PCX_REQ_STEP4_1 5'b10001 |
---|
| 96 | `define PCX_BIS 5'b10010 |
---|
| 97 | `define PCX_BIS_1 5'b10011 |
---|
| 98 | `define PCX_BIS_2 5'b10100 |
---|
| 99 | `define CPX_READY_1 5'b10101 |
---|
| 100 | `define CPX_READY_2 5'b10110 |
---|
| 101 | `define PCX_REQ_STEP1_2 5'b10111 |
---|
| 102 | `define PCX_UNKNOWN 5'b11000 |
---|
| 103 | `define PCX_FP_1 5'b11001 |
---|
| 104 | `define PCX_FP_2 5'b11010 |
---|
| 105 | `define FP_WAIT 5'b11011 |
---|
| 106 | `define CPX_FP 5'b11100 |
---|
| 107 | `define CPX_SEND_ETH_IRQ 5'b11101 |
---|
| 108 | `define CPX_INT_VEC_DIS 5'b11110 |
---|
| 109 | `define PCX_REQ_CAS_COMPARE 5'b11111 |
---|
| 110 | |
---|
| 111 | `define MEM_SIZE 64'h00000000_10000000 |
---|
| 112 | |
---|
| 113 | `define TEST_DRAM 1 |
---|
| 114 | `define DEBUGGING 1 |
---|
| 115 | |
---|
| 116 | reg cache_init; |
---|
| 117 | wire [3:0] dcache0_hit; |
---|
| 118 | wire [3:0] dcache1_hit; |
---|
| 119 | wire [3:0] icache_hit; |
---|
| 120 | reg multi_hit; |
---|
| 121 | reg multi_hit1; |
---|
| 122 | reg eth_int_d; |
---|
| 123 | reg eth_int_send; |
---|
| 124 | reg eth_int_sent; |
---|
| 125 | reg [3:0] cnt; |
---|
| 126 | |
---|
| 127 | // PCX channel FIFO |
---|
| 128 | wire [129:0] pcx_data_fifo; |
---|
| 129 | wire pcx_fifo_empty; |
---|
| 130 | reg [ 4:0] pcx_req_1; |
---|
| 131 | reg [ 4:0] pcx_req_2; |
---|
| 132 | reg pcx_atom_1; |
---|
| 133 | reg pcx_atom_2; |
---|
| 134 | reg pcx_data_123_d; |
---|
| 135 | |
---|
| 136 | // PCX 2nf channel FIFO |
---|
| 137 | wire [129:0] pcx1_data_fifo; |
---|
| 138 | wire pcx1_fifo_empty; |
---|
| 139 | reg [ 4:0] pcx1_req_1; |
---|
| 140 | reg [ 4:0] pcx1_req_2; |
---|
| 141 | reg pcx1_atom_1; |
---|
| 142 | reg pcx1_atom_2; |
---|
| 143 | reg pcx1_data_123_d; |
---|
| 144 | |
---|
| 145 | reg fifo_rd; |
---|
| 146 | reg fifo_rd1; |
---|
| 147 | |
---|
| 148 | always @(posedge clk) |
---|
| 149 | begin |
---|
| 150 | pcx_req_1<=pcx_req; |
---|
| 151 | pcx_atom_1<=pcx_atom; |
---|
| 152 | pcx_atom_2<=pcx_atom_1; |
---|
| 153 | pcx_req_2<=pcx_atom_1 ? pcx_req_1:5'b0; |
---|
| 154 | pcx_grant<=(pcx_req_1 | pcx_req_2); |
---|
| 155 | pcx_data_123_d<=pcx_data[123]; |
---|
| 156 | |
---|
| 157 | pcx1_req_1<=pcx1_req; |
---|
| 158 | pcx1_atom_1<=pcx1_atom; |
---|
| 159 | pcx1_atom_2<=pcx1_atom_1; |
---|
| 160 | pcx1_req_2<=pcx1_atom_1 ? pcx1_req_1:5'b0; |
---|
| 161 | pcx1_grant<=(pcx1_req_1 | pcx1_req_2); |
---|
| 162 | pcx1_data_123_d<=pcx1_data[123]; |
---|
| 163 | end |
---|
| 164 | |
---|
| 165 | /*pcx_fifo pcx_fifo_inst( |
---|
| 166 | // FIFO should be first word fall-through |
---|
| 167 | // It has no full flag as the core will send only limited number of requests, |
---|
| 168 | // in original design we used it 32 words deep |
---|
| 169 | // Just make it deeper if you experience overflow - |
---|
| 170 | // you can't just send no grant on full because the core expects immediate |
---|
| 171 | // grant for at least two requests for each zone |
---|
| 172 | .aclr(!rstn), |
---|
| 173 | .clock(clk), |
---|
| 174 | .data({pcx_atom_1,pcx_req_1,pcx_data}), |
---|
| 175 | .rdreq(fifo_rd), |
---|
| 176 | .wrreq((pcx_req_1!=5'b00000 && pcx_data[123]) || (pcx_atom_2 && pcx_data_123_d)), |
---|
| 177 | // Second atomic packet for FPU may be invalid, but should be sent to FPU |
---|
| 178 | // so if the first atomic packet is valid we latch both |
---|
| 179 | .empty(pcx_fifo_empty), |
---|
| 180 | .q(pcx_data_fifo) |
---|
| 181 | ); |
---|
| 182 | */ |
---|
| 183 | pcx_fifo pcx_fifo_inst( |
---|
| 184 | .clk(clk), |
---|
| 185 | .rst(!rstn), |
---|
| 186 | .din({pcx_atom_1,pcx_req_1,pcx_data}), |
---|
| 187 | .rd_en(fifo_rd), |
---|
| 188 | .wr_en((pcx_req_1!=5'b00000 && pcx_data[123]) || (pcx_atom_2 && pcx_data_123_d)), |
---|
| 189 | .empty(pcx_fifo_empty), |
---|
| 190 | .dout(pcx_data_fifo) |
---|
| 191 | ); |
---|
| 192 | |
---|
| 193 | |
---|
| 194 | pcx_fifo pcx_fifo_inst1( |
---|
| 195 | .clk(clk), |
---|
| 196 | .rst(!rstn), |
---|
| 197 | .din({pcx1_atom_1,pcx1_req_1,pcx1_data}), |
---|
| 198 | .rd_en(fifo_rd1), |
---|
| 199 | .wr_en((pcx1_req_1!=5'b00000 && pcx1_data[123]) || (pcx1_atom_2 && pcx1_data_123_d)), |
---|
| 200 | .empty(pcx1_fifo_empty), |
---|
| 201 | .dout(pcx1_data_fifo) |
---|
| 202 | ); |
---|
| 203 | // -------------------------- |
---|
| 204 | |
---|
| 205 | reg wb_ack_d; |
---|
| 206 | |
---|
| 207 | always @(posedge clk or negedge rstn) |
---|
| 208 | begin |
---|
| 209 | if(!rstn) |
---|
| 210 | eth_int_send<=0; |
---|
| 211 | else |
---|
| 212 | begin |
---|
| 213 | wb_ack_d<=wb_ack; |
---|
| 214 | eth_int_d<=eth_int; |
---|
| 215 | if(eth_int && !eth_int_d) |
---|
| 216 | eth_int_send<=1; |
---|
| 217 | else |
---|
| 218 | if(eth_int_sent) |
---|
| 219 | eth_int_send<=0; |
---|
| 220 | end |
---|
| 221 | end |
---|
| 222 | wire [123:0] pcx_packet; |
---|
| 223 | reg cpu; |
---|
| 224 | assign pcx_packet=cpu ? pcx1_data_fifo[123:0]:pcx_data_fifo[123:0]; |
---|
| 225 | reg cpu2; |
---|
| 226 | |
---|
| 227 | always @(posedge clk or negedge rstn) |
---|
| 228 | if(rstn==0) |
---|
| 229 | begin |
---|
| 230 | if(`TEST_DRAM) |
---|
| 231 | state<=`TEST_DRAM_1; |
---|
| 232 | else |
---|
| 233 | state<=`INIT_DRAM_1; // DRAM initialization is mandatory! |
---|
| 234 | cpx_ready<=0; |
---|
| 235 | fifo_rd<=0; |
---|
| 236 | cpx_packet<=145'b0; |
---|
| 237 | wb_cycle<=0; |
---|
| 238 | wb_strobe<=0; |
---|
| 239 | wb_we<=0; |
---|
| 240 | wb_sel<=0; |
---|
| 241 | wb_addr<=64'b0; |
---|
| 242 | wb_data_o<=64'b0; |
---|
| 243 | pcx_packet_d<=124'b0; |
---|
| 244 | fp_pcx<=124'b0; |
---|
| 245 | fp_req<=0; |
---|
| 246 | end |
---|
| 247 | else |
---|
| 248 | case(state) |
---|
| 249 | `TEST_DRAM_1: |
---|
| 250 | begin |
---|
| 251 | wb_cycle<=1; |
---|
| 252 | wb_strobe<=1; |
---|
| 253 | wb_sel<=8'hFF; |
---|
| 254 | wb_we<=1; |
---|
| 255 | state<=`TEST_DRAM_2; |
---|
| 256 | end |
---|
| 257 | `TEST_DRAM_2: |
---|
| 258 | if(wb_ack) |
---|
| 259 | begin |
---|
| 260 | wb_strobe<=0; |
---|
| 261 | if(wb_addr<`MEM_SIZE-8) |
---|
| 262 | begin |
---|
| 263 | wb_addr[31:0]<=wb_addr[31:0]+8; |
---|
| 264 | wb_data_o<={wb_addr[31:0]+8,wb_addr[31:0]+8}; |
---|
| 265 | state<=`TEST_DRAM_1; |
---|
| 266 | end |
---|
| 267 | else |
---|
| 268 | begin |
---|
| 269 | state<=`TEST_DRAM_3; |
---|
| 270 | wb_cycle<=0; |
---|
| 271 | wb_sel<=0; |
---|
| 272 | wb_we<=0; |
---|
| 273 | wb_data_o<=64'b0; |
---|
| 274 | wb_addr<=64'b0; |
---|
| 275 | end |
---|
| 276 | end |
---|
| 277 | `TEST_DRAM_3: |
---|
| 278 | begin |
---|
| 279 | wb_cycle<=1; |
---|
| 280 | wb_strobe<=1; |
---|
| 281 | wb_sel<=8'hFF; |
---|
| 282 | state<=`TEST_DRAM_4; |
---|
| 283 | end |
---|
| 284 | `TEST_DRAM_4: |
---|
| 285 | if(wb_ack) |
---|
| 286 | begin |
---|
| 287 | wb_strobe<=0; |
---|
| 288 | if(wb_addr<`MEM_SIZE-8) |
---|
| 289 | begin |
---|
| 290 | if(wb_data_i=={wb_addr[31:0],wb_addr[31:0]}) |
---|
| 291 | begin |
---|
| 292 | wb_addr[31:0]<=wb_addr[31:0]+8; |
---|
| 293 | state<=`TEST_DRAM_3; |
---|
| 294 | end |
---|
| 295 | end |
---|
| 296 | else |
---|
| 297 | begin |
---|
| 298 | state<=`INIT_DRAM_1; |
---|
| 299 | wb_cycle<=0; |
---|
| 300 | wb_sel<=0; |
---|
| 301 | wb_we<=0; |
---|
| 302 | wb_data_o<=64'b0; |
---|
| 303 | wb_addr<=64'b0; |
---|
| 304 | end |
---|
| 305 | end |
---|
| 306 | `INIT_DRAM_1: |
---|
| 307 | begin |
---|
| 308 | wb_cycle<=1; |
---|
| 309 | wb_strobe<=1; |
---|
| 310 | wb_sel<=8'hFF; |
---|
| 311 | wb_we<=1; |
---|
| 312 | cache_init<=1; // We also init cache directories here |
---|
| 313 | state<=`INIT_DRAM_2; |
---|
| 314 | end |
---|
| 315 | `INIT_DRAM_2: |
---|
| 316 | if(wb_ack) |
---|
| 317 | begin |
---|
| 318 | wb_strobe<=0; |
---|
| 319 | if(wb_addr<`MEM_SIZE-8) |
---|
| 320 | begin |
---|
| 321 | wb_addr[31:0]<=wb_addr[31:0]+8; |
---|
| 322 | pcx_packet_d[64+11:64+4]<=pcx_packet_d[64+11:64+4]+1; // Address for cachedir init |
---|
| 323 | state<=`INIT_DRAM_1; |
---|
| 324 | end |
---|
| 325 | else |
---|
| 326 | begin |
---|
| 327 | state<=`WAKEUP; |
---|
| 328 | wb_cycle<=0; |
---|
| 329 | wb_sel<=0; |
---|
| 330 | wb_we<=0; |
---|
| 331 | cache_init<=0; |
---|
| 332 | wb_addr<=64'b0; |
---|
| 333 | end |
---|
| 334 | end |
---|
| 335 | `WAKEUP: |
---|
| 336 | begin |
---|
| 337 | cpx_packet<=145'h1700000000000000000000000000000010001; |
---|
| 338 | cpx_ready<=1; |
---|
| 339 | state<=`PCX_IDLE; |
---|
| 340 | end |
---|
| 341 | `PCX_IDLE: |
---|
| 342 | begin |
---|
| 343 | cnt<=0; |
---|
| 344 | cpx_packet<=145'b0; |
---|
| 345 | cpx_ready<=0; |
---|
| 346 | cpx1_packet<=145'b0; |
---|
| 347 | cpx1_ready<=0; |
---|
| 348 | cpx_two_packet<=0; |
---|
| 349 | multi_hit<=0; |
---|
| 350 | multi_hit1<=0; |
---|
| 351 | if(eth_int_send) |
---|
| 352 | begin |
---|
| 353 | state<=`CPX_SEND_ETH_IRQ; |
---|
| 354 | eth_int_sent<=1; |
---|
| 355 | end |
---|
| 356 | else |
---|
| 357 | if(!pcx_fifo_empty) |
---|
| 358 | begin |
---|
| 359 | pcx_req_d<=pcx_data_fifo[128:124]; |
---|
| 360 | pcx_atom_d<=pcx_data_fifo[129]; |
---|
| 361 | fifo_rd<=1; |
---|
| 362 | state<=`GOT_PCX_REQ; |
---|
| 363 | cpu<=0; |
---|
| 364 | cpu2<=0; |
---|
| 365 | end |
---|
| 366 | else |
---|
| 367 | if(!pcx1_fifo_empty) |
---|
| 368 | begin |
---|
| 369 | pcx_req_d<=pcx1_data_fifo[128:124]; |
---|
| 370 | pcx_atom_d<=pcx1_data_fifo[129]; |
---|
| 371 | fifo_rd1<=1; |
---|
| 372 | state<=`GOT_PCX_REQ; |
---|
| 373 | cpu<=1; |
---|
| 374 | cpu2<=1; |
---|
| 375 | end |
---|
| 376 | end |
---|
| 377 | `GOT_PCX_REQ: |
---|
| 378 | begin |
---|
| 379 | pcx_packet_d<=pcx_packet; |
---|
| 380 | if(`DEBUGGING) |
---|
| 381 | begin |
---|
| 382 | wb_sel[1:0]<=pcx_packet[113:112]; |
---|
| 383 | wb_sel[2]<=1; |
---|
| 384 | end |
---|
| 385 | if(pcx_packet[103:64]==40'h9800000800 && pcx_packet[122:118]==5'b00001) |
---|
| 386 | begin |
---|
| 387 | state<=`CPX_INT_VEC_DIS; |
---|
| 388 | fifo_rd<=0; |
---|
| 389 | fifo_rd1<=0; |
---|
| 390 | end |
---|
| 391 | else |
---|
| 392 | if(pcx_atom_d==0) |
---|
| 393 | begin |
---|
| 394 | fifo_rd<=0; |
---|
| 395 | fifo_rd1<=0; |
---|
| 396 | if(pcx_packet[122:118]==5'b01010) // FP req |
---|
| 397 | begin |
---|
| 398 | state<=`PCX_FP_1; |
---|
| 399 | pcx_packet_2nd[123]<=0; |
---|
| 400 | end |
---|
| 401 | else |
---|
| 402 | state<=`PCX_REQ_STEP1; |
---|
| 403 | end |
---|
| 404 | else |
---|
| 405 | state<=`PCX_REQ_2ND; |
---|
| 406 | end |
---|
| 407 | `PCX_REQ_2ND: |
---|
| 408 | begin |
---|
| 409 | pcx_packet_2nd<=pcx_packet; //Latch second packet for atomics |
---|
| 410 | if(`DEBUGGING) |
---|
| 411 | if(pcx_fifo_empty) |
---|
| 412 | wb_sel<=8'h67; |
---|
| 413 | fifo_rd<=0; |
---|
| 414 | fifo_rd1<=0; |
---|
| 415 | if(pcx_packet_d[122:118]==5'b01010) // FP req |
---|
| 416 | state<=`PCX_FP_1; |
---|
| 417 | else |
---|
| 418 | state<=`PCX_REQ_STEP1; |
---|
| 419 | end |
---|
| 420 | `PCX_REQ_STEP1: |
---|
| 421 | begin |
---|
| 422 | if(pcx_packet_d[111]==1'b1) // Invalidate request |
---|
| 423 | begin |
---|
| 424 | cpx_packet_1[144]<=1; // Valid |
---|
| 425 | cpx_packet_1[143:140]<=4'b0100; // Invalidate reply is Store ACK |
---|
| 426 | cpx_packet_1[139]<=1; // L2 miss |
---|
| 427 | cpx_packet_1[138:137]<=0; // Error |
---|
| 428 | cpx_packet_1[136]<=pcx_packet_d[117]; // Non-cacheble |
---|
| 429 | cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID |
---|
| 430 | cpx_packet_1[133:131]<=0; // Way valid |
---|
| 431 | cpx_packet_1[130]<=((pcx_packet_d[122:118]==5'b10000) && (pcx_req_d==5'b10000)) ? 1:0; // Four byte fill |
---|
| 432 | cpx_packet_1[129]<=pcx_atom_d; |
---|
| 433 | cpx_packet_1[128]<=pcx_packet_d[110]; // Prefetch |
---|
| 434 | cpx_packet_1[127:0]<={2'b0,pcx_packet_d[109]/*BIS*/,pcx_packet_d[122:118]==5'b00000 ? 2'b01:2'b10,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],112'b0}; |
---|
| 435 | state<=`CPX_READY_1; |
---|
| 436 | end |
---|
| 437 | else |
---|
| 438 | if(pcx_packet_d[122:118]!=5'b01001) // Not INT |
---|
| 439 | begin |
---|
| 440 | wb_cycle<=1'b1; |
---|
| 441 | wb_strobe<=1'b1; |
---|
| 442 | if((pcx_packet_d[122:118]==5'b00000 && !pcx_req_d[4]) || pcx_packet_d[122:118]==5'b00010 || pcx_packet_d[122:118]==5'b00100 || pcx_packet_d[122:118]==5'b00110) |
---|
| 443 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+4],4'b0000}; //DRAM load/streamload, CAS and SWAP always use DRAM and load first |
---|
| 444 | else |
---|
| 445 | if(pcx_packet_d[122:118]==5'b10000 && !pcx_req_d[4]) |
---|
| 446 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b00000}; //DRAM ifill |
---|
| 447 | else |
---|
| 448 | if(pcx_packet_d[64+39:64+28]==12'hFFF && pcx_packet_d[64+27:64+24]!=4'b0) // flash remap FFF1->FFF8 |
---|
| 449 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3]+37'h0000E00000,3'b000}; |
---|
| 450 | else |
---|
| 451 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3],3'b000}; |
---|
| 452 | wb_data_o<=pcx_packet_d[63:0]; |
---|
| 453 | state<=`PCX_REQ_STEP1_1; |
---|
| 454 | end |
---|
| 455 | else |
---|
| 456 | //if((pcx_packet_d[12:10]!=3'b000) && !pcx_packet_d[117]) // Not FLUSH int and not this core |
---|
| 457 | // state<=`PCX_IDLE; |
---|
| 458 | //else |
---|
| 459 | state<=`CPX_READY_1; |
---|
| 460 | case(pcx_packet_d[122:118]) // Packet type |
---|
| 461 | 5'b00000://Load |
---|
| 462 | begin |
---|
| 463 | wb_we<=0; |
---|
| 464 | if(!pcx_req_d[4]) |
---|
| 465 | wb_sel<=8'b11111111; // DRAM requests are always 128 bit |
---|
| 466 | else |
---|
| 467 | case(pcx_packet_d[106:104]) //Size |
---|
| 468 | 3'b000://Byte |
---|
| 469 | case(pcx_packet_d[64+2:64]) |
---|
| 470 | 3'b000:wb_sel<=8'b10000000; |
---|
| 471 | 3'b001:wb_sel<=8'b01000000; |
---|
| 472 | 3'b010:wb_sel<=8'b00100000; |
---|
| 473 | 3'b011:wb_sel<=8'b00010000; |
---|
| 474 | 3'b100:wb_sel<=8'b00001000; |
---|
| 475 | 3'b101:wb_sel<=8'b00000100; |
---|
| 476 | 3'b110:wb_sel<=8'b00000010; |
---|
| 477 | 3'b111:wb_sel<=8'b00000001; |
---|
| 478 | endcase |
---|
| 479 | 3'b001://Halfword |
---|
| 480 | case(pcx_packet_d[64+2:64+1]) |
---|
| 481 | 2'b00:wb_sel<=8'b11000000; |
---|
| 482 | 2'b01:wb_sel<=8'b00110000; |
---|
| 483 | 2'b10:wb_sel<=8'b00001100; |
---|
| 484 | 2'b11:wb_sel<=8'b00000011; |
---|
| 485 | endcase |
---|
| 486 | 3'b010://Word |
---|
| 487 | wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; |
---|
| 488 | 3'b011://Doubleword |
---|
| 489 | wb_sel<=8'b11111111; |
---|
| 490 | 3'b100://Quadword |
---|
| 491 | wb_sel<=8'b11111111; |
---|
| 492 | 3'b111://Cacheline |
---|
| 493 | wb_sel<=8'b11111111; |
---|
| 494 | default: |
---|
| 495 | wb_sel<=8'b01011010; // Unreal eye-catching value for debug |
---|
| 496 | endcase |
---|
| 497 | end |
---|
| 498 | 5'b00001://Store |
---|
| 499 | begin |
---|
| 500 | wb_we<=1; |
---|
| 501 | if(pcx_packet_d[110:109]!=2'b00) //Block (or init) store |
---|
| 502 | wb_sel<=8'b11111111; // Blocks are always 64 bit |
---|
| 503 | else |
---|
| 504 | case(pcx_packet_d[106:104]) //Size |
---|
| 505 | 3'b000://Byte |
---|
| 506 | case(pcx_packet_d[64+2:64]) |
---|
| 507 | 3'b000:wb_sel<=8'b10000000; |
---|
| 508 | 3'b001:wb_sel<=8'b01000000; |
---|
| 509 | 3'b010:wb_sel<=8'b00100000; |
---|
| 510 | 3'b011:wb_sel<=8'b00010000; |
---|
| 511 | 3'b100:wb_sel<=8'b00001000; |
---|
| 512 | 3'b101:wb_sel<=8'b00000100; |
---|
| 513 | 3'b110:wb_sel<=8'b00000010; |
---|
| 514 | 3'b111:wb_sel<=8'b00000001; |
---|
| 515 | endcase |
---|
| 516 | 3'b001://Halfword |
---|
| 517 | case(pcx_packet_d[64+2:64+1]) |
---|
| 518 | 2'b00:wb_sel<=8'b11000000; |
---|
| 519 | 2'b01:wb_sel<=8'b00110000; |
---|
| 520 | 2'b10:wb_sel<=8'b00001100; |
---|
| 521 | 2'b11:wb_sel<=8'b00000011; |
---|
| 522 | endcase |
---|
| 523 | 3'b010://Word |
---|
| 524 | wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; |
---|
| 525 | 3'b011://Doubleword |
---|
| 526 | wb_sel<=8'b11111111; |
---|
| 527 | default: |
---|
| 528 | if(`DEBUGGING) |
---|
| 529 | wb_sel<=8'b01011010; // Unreal eye-catching value for debug |
---|
| 530 | endcase |
---|
| 531 | end |
---|
| 532 | 5'b00010://CAS |
---|
| 533 | begin |
---|
| 534 | wb_we<=0; //Load first |
---|
| 535 | wb_sel<=8'b11111111; // CAS loads are as cacheline |
---|
| 536 | end |
---|
| 537 | 5'b00100://STRLOAD |
---|
| 538 | begin |
---|
| 539 | wb_we<=0; |
---|
| 540 | wb_sel<=8'b11111111; // Stream loads are always 128 bit |
---|
| 541 | end |
---|
| 542 | 5'b00101://STRSTORE |
---|
| 543 | begin |
---|
| 544 | wb_we<=1; |
---|
| 545 | case(pcx_packet_d[106:104]) //Size |
---|
| 546 | 3'b000://Byte |
---|
| 547 | case(pcx_packet_d[64+2:64]) |
---|
| 548 | 3'b000:wb_sel<=8'b10000000; |
---|
| 549 | 3'b001:wb_sel<=8'b01000000; |
---|
| 550 | 3'b010:wb_sel<=8'b00100000; |
---|
| 551 | 3'b011:wb_sel<=8'b00010000; |
---|
| 552 | 3'b100:wb_sel<=8'b00001000; |
---|
| 553 | 3'b101:wb_sel<=8'b00000100; |
---|
| 554 | 3'b110:wb_sel<=8'b00000010; |
---|
| 555 | 3'b111:wb_sel<=8'b00000001; |
---|
| 556 | endcase |
---|
| 557 | 3'b001://Halfword |
---|
| 558 | case(pcx_packet_d[64+2:64+1]) |
---|
| 559 | 2'b00:wb_sel<=8'b11000000; |
---|
| 560 | 2'b01:wb_sel<=8'b00110000; |
---|
| 561 | 2'b10:wb_sel<=8'b00001100; |
---|
| 562 | 2'b11:wb_sel<=8'b00000011; |
---|
| 563 | endcase |
---|
| 564 | 3'b010://Word |
---|
| 565 | wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; |
---|
| 566 | 3'b011://Doubleword |
---|
| 567 | wb_sel<=8'b11111111; |
---|
| 568 | 3'b100://Quadword |
---|
| 569 | wb_sel<=8'b11111111; |
---|
| 570 | 3'b111://Cacheline |
---|
| 571 | wb_sel<=8'b11111111; |
---|
| 572 | default: |
---|
| 573 | wb_sel<=8'b01011010; // Unreal eye-catching value for debug |
---|
| 574 | endcase |
---|
| 575 | end |
---|
| 576 | 5'b00110://SWAP/LDSTUB |
---|
| 577 | begin |
---|
| 578 | wb_we<=0; // Load first, as CAS |
---|
| 579 | wb_sel<=8'b11111111; // SWAP/LDSTUB loads are as cacheline |
---|
| 580 | end |
---|
| 581 | 5'b01001://INT |
---|
| 582 | if(pcx_packet_d[117]) // Flush |
---|
| 583 | begin |
---|
| 584 | cpx_packet_1<={9'h171,pcx_packet_d[113:112],11'h0,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],30'h0,pcx_packet_d[17:0],46'b0,pcx_packet_d[17:0]}; //FLUSH instruction answer |
---|
| 585 | //cpx_packet_2<={9'h171,pcx_packet_d[113:112],11'h0,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],30'h0,pcx_packet_d[17:0],46'b0,pcx_packet_d[17:0]}; //FLUSH instruction answer |
---|
| 586 | //cpx_two_packet<=1; |
---|
| 587 | //cpu2<=!cpu; // Flush should be sent to both cores |
---|
| 588 | end |
---|
| 589 | else // Tread-to-thread interrupt |
---|
| 590 | begin |
---|
| 591 | cpx_packet_1<={9'h170,pcx_packet_d[113:112],52'h0,pcx_packet_d[17:0],46'h0,pcx_packet_d[17:0]}; |
---|
| 592 | cpu<=pcx_packet_d[10]; |
---|
| 593 | end |
---|
| 594 | //5'b01010: FP1 - processed by separate state |
---|
| 595 | //5'b01011: FP2 - processed by separate state |
---|
| 596 | //5'b01101: FWDREQ - not implemented |
---|
| 597 | //5'b01110: FWDREPL - not implemented |
---|
| 598 | 5'b10000://IFILL |
---|
| 599 | begin |
---|
| 600 | wb_we<=0; |
---|
| 601 | if(pcx_req_d[4]) // I/O access |
---|
| 602 | wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; |
---|
| 603 | else |
---|
| 604 | wb_sel<=8'b11111111; |
---|
| 605 | end |
---|
| 606 | default: |
---|
| 607 | begin |
---|
| 608 | wb_we<=0; |
---|
| 609 | wb_sel<=8'b10101010; // Unreal eye-catching value for debug |
---|
| 610 | end |
---|
| 611 | endcase |
---|
| 612 | end |
---|
| 613 | `PCX_REQ_STEP1_1: |
---|
| 614 | state<=`PCX_REQ_STEP1_2; // Delay for L1 directory |
---|
| 615 | `PCX_REQ_STEP1_2: |
---|
| 616 | begin |
---|
| 617 | if(wb_ack || wb_ack_d) |
---|
| 618 | begin |
---|
| 619 | cpx_packet_1[144]<=1; // Valid |
---|
| 620 | cpx_packet_1[139]<=(pcx_packet_d[122:118]==5'b00000) || (pcx_packet_d[122:118]==5'b10000) ? 1:0; // L2 always miss on load and ifill |
---|
| 621 | cpx_packet_1[138:137]<=0; // Error |
---|
| 622 | cpx_packet_1[136]<=pcx_packet_d[117] || (pcx_packet_d[122:118]==5'b00001) ? 1:0; // Non-cacheble is set on store too |
---|
| 623 | cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID |
---|
| 624 | if((pcx_packet_d[122:118]==5'b00000 && !pcx_packet_d[117] && !pcx_packet_d[110]) || (pcx_packet_d[122:118]==5'b10000)) // Cacheble Load or IFill |
---|
| 625 | cpx_packet_1[133:131]<={othercachehit[0],wayval0}; |
---|
| 626 | else |
---|
| 627 | cpx_packet_1[133:131]<=3'b000; // Way valid |
---|
| 628 | if(pcx_packet_d[122:118]==5'b00100) // Strload |
---|
| 629 | cpx_packet_1[130]<=pcx_packet_d[106]; // A |
---|
| 630 | else |
---|
| 631 | if(pcx_packet_d[122:118]==5'b00101) // Stream store |
---|
| 632 | cpx_packet_1[130]<=pcx_packet_d[108]; // A |
---|
| 633 | else |
---|
| 634 | cpx_packet_1[130]<=((pcx_packet_d[122:118]==5'b10000) && pcx_req_d[4]) ? 1:0; // Four byte fill |
---|
| 635 | if(pcx_packet_d[122:118]==5'b00100) // Strload |
---|
| 636 | cpx_packet_1[129]<=pcx_packet_d[105]; // B |
---|
| 637 | else |
---|
| 638 | cpx_packet_1[129]<=pcx_atom_d || (pcx_packet_d[122:118]==5'b00110); // SWAP is single-packet but needs atom in CPX |
---|
| 639 | cpx_packet_1[128]<=pcx_packet_d[110] && pcx_packet_d[122:118]==5'b00000; // Prefetch |
---|
| 640 | cpx_packet_2[144]<=1; // Valid |
---|
| 641 | cpx_packet_2[139]<=0; // L2 miss |
---|
| 642 | cpx_packet_2[138:137]<=0; // Error |
---|
| 643 | cpx_packet_2[136]<=pcx_packet_d[117] || (pcx_packet_d[122:118]==5'b00001) ? 1:0; // Non-cacheble is set on store too |
---|
| 644 | cpx_packet_2[135:134]<=pcx_packet_d[113:112]; // Thread ID |
---|
| 645 | if(pcx_packet_d[122:118]==5'b10000) // IFill |
---|
| 646 | cpx_packet_2[133:131]<={othercachehit[1],wayval1}; |
---|
| 647 | else |
---|
| 648 | cpx_packet_2[133:131]<=3'b000; // Way valid |
---|
| 649 | cpx_packet_2[130]<=0; // Four byte fill |
---|
| 650 | cpx_packet_2[129]<=pcx_atom_d || (pcx_packet_d[122:118]==5'b00110) || ((pcx_packet_d[122:118]==5'b10000) && !pcx_req_d[4]); |
---|
| 651 | cpx_packet_2[128]<=0; // Prefetch |
---|
| 652 | wb_strobe<=0; |
---|
| 653 | wb_sel<=8'b0; |
---|
| 654 | wb_addr<=64'b0; |
---|
| 655 | wb_data_o<=64'b0; |
---|
| 656 | wb_we<=0; |
---|
| 657 | case(pcx_packet_d[122:118]) // Packet type |
---|
| 658 | 5'b00000://Load |
---|
| 659 | begin |
---|
| 660 | cpx_packet_1[143:140]<=4'b0000; // Type |
---|
| 661 | if(!pcx_req_d[4]) |
---|
| 662 | begin |
---|
| 663 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 664 | state<=`PCX_REQ_STEP2; |
---|
| 665 | end |
---|
| 666 | else |
---|
| 667 | case(pcx_packet_d[106:104]) //Size |
---|
| 668 | 3'b000://Byte |
---|
| 669 | begin |
---|
| 670 | case(pcx_packet_d[64+2:64]) |
---|
| 671 | 3'b000:cpx_packet_1[127:0]<={wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56]}; |
---|
| 672 | 3'b001:cpx_packet_1[127:0]<={wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48]}; |
---|
| 673 | 3'b010:cpx_packet_1[127:0]<={wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40]}; |
---|
| 674 | 3'b011:cpx_packet_1[127:0]<={wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32]}; |
---|
| 675 | 3'b100:cpx_packet_1[127:0]<={wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24]}; |
---|
| 676 | 3'b101:cpx_packet_1[127:0]<={wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16]}; |
---|
| 677 | 3'b110:cpx_packet_1[127:0]<={wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8]}; |
---|
| 678 | 3'b111:cpx_packet_1[127:0]<={wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0]}; |
---|
| 679 | endcase |
---|
| 680 | wb_cycle<=0; |
---|
| 681 | state<=`CPX_READY_1; |
---|
| 682 | end |
---|
| 683 | 3'b001://Halfword |
---|
| 684 | begin |
---|
| 685 | case(pcx_packet_d[64+2:64+1]) |
---|
| 686 | 2'b00:cpx_packet_1[127:0]<={wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48]}; |
---|
| 687 | 2'b01:cpx_packet_1[127:0]<={wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32]}; |
---|
| 688 | 2'b10:cpx_packet_1[127:0]<={wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16]}; |
---|
| 689 | 2'b11:cpx_packet_1[127:0]<={wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0]}; |
---|
| 690 | endcase |
---|
| 691 | wb_cycle<=0; |
---|
| 692 | state<=`CPX_READY_1; |
---|
| 693 | end |
---|
| 694 | 3'b010://Word |
---|
| 695 | begin |
---|
| 696 | if(pcx_packet_d[64+2]==0) |
---|
| 697 | cpx_packet_1[127:0]<={wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32]}; |
---|
| 698 | else |
---|
| 699 | cpx_packet_1[127:0]<={wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0]}; |
---|
| 700 | wb_cycle<=0; |
---|
| 701 | state<=`CPX_READY_1; |
---|
| 702 | end |
---|
| 703 | 3'b011://Doubleword |
---|
| 704 | begin |
---|
| 705 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 706 | wb_cycle<=0; |
---|
| 707 | state<=`CPX_READY_1; |
---|
| 708 | end |
---|
| 709 | 3'b100://Quadword |
---|
| 710 | begin |
---|
| 711 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 712 | wb_cycle<=0; |
---|
| 713 | state<=`CPX_READY_1; // 16 byte access to PROM should just duplicate the data |
---|
| 714 | end |
---|
| 715 | 3'b111://Cacheline |
---|
| 716 | begin |
---|
| 717 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 718 | wb_cycle<=0; |
---|
| 719 | state<=`CPX_READY_1; // 16 byte access to PROM should just duplicate the data |
---|
| 720 | end |
---|
| 721 | default: |
---|
| 722 | begin |
---|
| 723 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 724 | wb_cycle<=0; |
---|
| 725 | state<=`PCX_UNKNOWN; |
---|
| 726 | end |
---|
| 727 | endcase |
---|
| 728 | end |
---|
| 729 | 5'b00001://Store |
---|
| 730 | begin |
---|
| 731 | cpx_packet_1[143:140]<=4'b0100; // Type |
---|
| 732 | cpx_packet_1[127:0]<={2'b0,pcx_packet_d[109]/*BIS*/,2'b0,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],inval_vect0}; |
---|
| 733 | // if((pcx_packet_d[110:109]==2'b01) && (pcx_packet_d[64+5:64]==0) && !inval_vect0[3] && !inval_vect1[3]) // Block init store |
---|
| 734 | // state<=`PCX_BIS; |
---|
| 735 | // else |
---|
| 736 | // begin |
---|
| 737 | wb_cycle<=0; |
---|
| 738 | state<=`CPX_READY_1; |
---|
| 739 | // end |
---|
| 740 | end |
---|
| 741 | 5'b00010://CAS |
---|
| 742 | begin |
---|
| 743 | cpx_packet_1[143:140]<=4'b0000; // Load return for first packet |
---|
| 744 | cpx_packet_2[143:140]<=4'b0100; // Store ACK for second packet |
---|
| 745 | cpx_packet_2[127:0]<={5'b0,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],inval_vect0}; |
---|
| 746 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 747 | state<=`PCX_REQ_STEP2; |
---|
| 748 | end |
---|
| 749 | 5'b00100://STRLOAD |
---|
| 750 | begin |
---|
| 751 | cpx_packet_1[143:140]<=4'b0010; // Type |
---|
| 752 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 753 | state<=`PCX_REQ_STEP2; |
---|
| 754 | end |
---|
| 755 | 5'b00101://STRSTORE |
---|
| 756 | begin |
---|
| 757 | cpx_packet_1[143:140]<=4'b0110; // Type |
---|
| 758 | cpx_packet_1[127:0]<={5'b0,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],inval_vect0}; |
---|
| 759 | wb_cycle<=0; |
---|
| 760 | state<=`CPX_READY_1; |
---|
| 761 | end |
---|
| 762 | 5'b00110://SWAP/LDSTUB |
---|
| 763 | begin |
---|
| 764 | cpx_packet_1[143:140]<=4'b0000; // Load return for first packet |
---|
| 765 | cpx_packet_2[143:140]<=4'b0100; // Store ACK for second packet |
---|
| 766 | cpx_packet_2[127:0]<={5'b0,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],inval_vect0}; |
---|
| 767 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 768 | state<=`PCX_REQ_STEP2; |
---|
| 769 | end |
---|
| 770 | 5'b10000://IFILL |
---|
| 771 | begin |
---|
| 772 | cpx_packet_1[143:140]<=4'b0001; // Type |
---|
| 773 | cpx_packet_2[143:140]<=4'b0001; // Type |
---|
| 774 | if(pcx_req_d[4]) // I/O access |
---|
| 775 | begin |
---|
| 776 | if(pcx_packet_d[64+2]==0) |
---|
| 777 | cpx_packet_1[127:0]<={wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32]}; |
---|
| 778 | else |
---|
| 779 | cpx_packet_1[127:0]<={wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0]}; |
---|
| 780 | state<=`CPX_READY_1; |
---|
| 781 | wb_cycle<=0; |
---|
| 782 | end |
---|
| 783 | else |
---|
| 784 | begin |
---|
| 785 | cpx_packet_1[127:0]<={wb_data_i,wb_data_i}; |
---|
| 786 | state<=`PCX_REQ_STEP2; |
---|
| 787 | end |
---|
| 788 | end |
---|
| 789 | default: |
---|
| 790 | begin |
---|
| 791 | wb_cycle<=0; |
---|
| 792 | state<=`PCX_UNKNOWN; |
---|
| 793 | end |
---|
| 794 | endcase |
---|
| 795 | end |
---|
| 796 | end |
---|
| 797 | `PCX_REQ_STEP2: // IFill, Load/strload, CAS, SWAP, LDSTUB - alwas load |
---|
| 798 | begin |
---|
| 799 | wb_strobe<=1'b1; |
---|
| 800 | if(pcx_packet_d[122:118]==5'b10000) |
---|
| 801 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b01000}; |
---|
| 802 | else |
---|
| 803 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+4],4'b1000}; |
---|
| 804 | wb_sel<=8'b11111111; // It is always full width for subsequent IFill and load accesses |
---|
| 805 | state<=`PCX_REQ_STEP2_1; |
---|
| 806 | end |
---|
| 807 | `PCX_REQ_STEP2_1: |
---|
| 808 | if(wb_ack==1) |
---|
| 809 | begin |
---|
| 810 | wb_strobe<=0; |
---|
| 811 | wb_sel<=8'b0; |
---|
| 812 | wb_addr<=64'b0; |
---|
| 813 | wb_data_o<=64'b0; |
---|
| 814 | wb_we<=0; |
---|
| 815 | cpx_packet_1[63:0]<=wb_data_i; |
---|
| 816 | if((pcx_packet_d[122:118]!=5'b00000) && (pcx_packet_d[122:118]!=5'b00100)) |
---|
| 817 | if(pcx_packet_d[122:118]!=5'b00010) // IFill, SWAP |
---|
| 818 | state<=`PCX_REQ_STEP3; |
---|
| 819 | else |
---|
| 820 | state<=`PCX_REQ_CAS_COMPARE; // CAS |
---|
| 821 | else |
---|
| 822 | begin |
---|
| 823 | wb_cycle<=0; |
---|
| 824 | state<=`CPX_READY_1; |
---|
| 825 | end |
---|
| 826 | end |
---|
| 827 | `PCX_REQ_CAS_COMPARE: |
---|
| 828 | begin |
---|
| 829 | cpx_two_packet<=1; |
---|
| 830 | if(pcx_packet_d[106:104]==3'b010) // 32-bit |
---|
| 831 | case(pcx_packet_d[64+3:64+2]) |
---|
| 832 | 2'b00:state<=cpx_packet_1[127:96]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1; |
---|
| 833 | 2'b01:state<=cpx_packet_1[95:64]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1; |
---|
| 834 | 2'b10:state<=cpx_packet_1[63:32]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1; |
---|
| 835 | 2'b11:state<=cpx_packet_1[31:0]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1; |
---|
| 836 | endcase |
---|
| 837 | else |
---|
| 838 | if(pcx_packet_d[64+3]==0) |
---|
| 839 | state<=cpx_packet_1[127:64]==pcx_packet_d[63:0] ? `PCX_REQ_STEP3:`CPX_READY_1; |
---|
| 840 | else |
---|
| 841 | state<=cpx_packet_1[63:0]==pcx_packet_d[63:0] ? `PCX_REQ_STEP3:`CPX_READY_1; |
---|
| 842 | end |
---|
| 843 | `PCX_REQ_STEP3: // 256-bit IFILL; CAS, SWAP and LDSTUB store |
---|
| 844 | begin |
---|
| 845 | if(pcx_packet_d[122:118]==5'b10000) |
---|
| 846 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b10000}; |
---|
| 847 | else |
---|
| 848 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3],3'b000}; // CAS or SWAP save |
---|
| 849 | cpx_two_packet<=1; |
---|
| 850 | if(pcx_packet_d[122:118]==5'b10000) |
---|
| 851 | wb_we<=0; |
---|
| 852 | else |
---|
| 853 | wb_we<=1; |
---|
| 854 | wb_strobe<=1'b1; |
---|
| 855 | if(pcx_packet_d[122:118]==5'b00010) // CAS |
---|
| 856 | if(pcx_packet_d[106:104]==3'b010) |
---|
| 857 | wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; |
---|
| 858 | else |
---|
| 859 | wb_sel<=8'b11111111; //CASX |
---|
| 860 | else |
---|
| 861 | if(pcx_packet_d[122:118]==5'b00110) //SWAP or LDSTUB |
---|
| 862 | if(pcx_packet_d[106:104]==3'b000) //LDSTUB |
---|
| 863 | case(pcx_packet_d[64+2:64]) |
---|
| 864 | 3'b000:wb_sel<=8'b10000000; |
---|
| 865 | 3'b001:wb_sel<=8'b01000000; |
---|
| 866 | 3'b010:wb_sel<=8'b00100000; |
---|
| 867 | 3'b011:wb_sel<=8'b00010000; |
---|
| 868 | 3'b100:wb_sel<=8'b00001000; |
---|
| 869 | 3'b101:wb_sel<=8'b00000100; |
---|
| 870 | 3'b110:wb_sel<=8'b00000010; |
---|
| 871 | 3'b111:wb_sel<=8'b00000001; |
---|
| 872 | endcase |
---|
| 873 | else |
---|
| 874 | wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; ///SWAP is always 32-bit |
---|
| 875 | else |
---|
| 876 | wb_sel<=8'b11111111; // It is always full width for subsequent IFill accesses |
---|
| 877 | if(pcx_packet_d[122:118]==5'b00110) //SWAP or LDSTUB |
---|
| 878 | wb_data_o<={pcx_packet_d[63:32],pcx_packet_d[63:32]}; |
---|
| 879 | // wb_data_o<=pcx_packet_d[63:0]; |
---|
| 880 | else |
---|
| 881 | wb_data_o<=pcx_packet_2nd[63:0]; // CAS store second packet data |
---|
| 882 | // if(pcx_packet_d[106:104]==3'b010) |
---|
| 883 | // wb_data_o<={pcx_packet_2nd[63:32],pcx_packet_2nd[63:32]}; // CAS store second packet data |
---|
| 884 | // else |
---|
| 885 | // wb_data_o<=pcx_packet_2nd[63:0]; |
---|
| 886 | state<=`PCX_REQ_STEP3_1; |
---|
| 887 | end |
---|
| 888 | `PCX_REQ_STEP3_1: |
---|
| 889 | if(wb_ack==1) |
---|
| 890 | begin |
---|
| 891 | wb_strobe<=0; |
---|
| 892 | wb_sel<=8'b0; |
---|
| 893 | wb_addr<=64'b0; |
---|
| 894 | wb_we<=0; |
---|
| 895 | wb_data_o<=64'b0; |
---|
| 896 | if(pcx_packet_d[122:118]==5'b10000) // IFill |
---|
| 897 | begin |
---|
| 898 | cpx_packet_2[127:64]<=wb_data_i; |
---|
| 899 | state<=`PCX_REQ_STEP4; |
---|
| 900 | end |
---|
| 901 | else |
---|
| 902 | begin |
---|
| 903 | wb_cycle<=0; |
---|
| 904 | state<=`CPX_READY_1; |
---|
| 905 | end |
---|
| 906 | end |
---|
| 907 | `PCX_REQ_STEP4: // 256-bit IFILL only |
---|
| 908 | begin |
---|
| 909 | wb_strobe<=1'b1; |
---|
| 910 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b11000}; |
---|
| 911 | wb_sel<=8'b11111111; // It is always full width for subsequent accesses |
---|
| 912 | state<=`PCX_REQ_STEP4_1; |
---|
| 913 | end |
---|
| 914 | `PCX_REQ_STEP4_1: |
---|
| 915 | if(wb_ack==1) |
---|
| 916 | begin |
---|
| 917 | wb_cycle<=0; |
---|
| 918 | wb_strobe<=0; |
---|
| 919 | wb_sel<=8'b0; |
---|
| 920 | wb_addr<=64'b0; |
---|
| 921 | wb_we<=0; |
---|
| 922 | cpx_packet_2[63:0]<=wb_data_i; |
---|
| 923 | state<=`CPX_READY_1; |
---|
| 924 | end |
---|
| 925 | `PCX_BIS: // Block init store |
---|
| 926 | begin |
---|
| 927 | wb_strobe<=1'b1; |
---|
| 928 | wb_we<=1; |
---|
| 929 | wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+6],6'b001000}; |
---|
| 930 | wb_sel<=8'b11111111; |
---|
| 931 | wb_data_o<=64'b0; |
---|
| 932 | state<=`PCX_BIS_1; |
---|
| 933 | end |
---|
| 934 | `PCX_BIS_1: |
---|
| 935 | if(wb_ack) |
---|
| 936 | begin |
---|
| 937 | wb_strobe<=0; |
---|
| 938 | if(wb_addr[39:0]<(pcx_packet_d[64+39:64]+8*7)) |
---|
| 939 | state<=`PCX_BIS_2; |
---|
| 940 | else |
---|
| 941 | begin |
---|
| 942 | wb_cycle<=0; |
---|
| 943 | wb_sel<=0; |
---|
| 944 | wb_we<=0; |
---|
| 945 | wb_addr<=64'b0; |
---|
| 946 | state<=`CPX_READY_1; |
---|
| 947 | end |
---|
| 948 | end |
---|
| 949 | `PCX_BIS_2: |
---|
| 950 | begin |
---|
| 951 | wb_strobe<=1'b1; |
---|
| 952 | wb_addr[5:0]<=wb_addr[5:0]+8; |
---|
| 953 | state<=`PCX_BIS_1; |
---|
| 954 | end |
---|
| 955 | `PCX_FP_1: |
---|
| 956 | begin |
---|
| 957 | fp_pcx<=pcx_packet_d; |
---|
| 958 | fp_req<=1; |
---|
| 959 | state<=`PCX_FP_2; |
---|
| 960 | if(`DEBUGGING) |
---|
| 961 | begin |
---|
| 962 | wb_addr<=pcx_packet_d[103:64]; |
---|
| 963 | wb_data_o<=pcx_packet_d[63:0]; |
---|
| 964 | wb_sel<=8'h22; |
---|
| 965 | end |
---|
| 966 | end |
---|
| 967 | `PCX_FP_2: |
---|
| 968 | begin |
---|
| 969 | fp_pcx<=pcx_packet_2nd; |
---|
| 970 | state<=`FP_WAIT; |
---|
| 971 | if(`DEBUGGING) |
---|
| 972 | begin |
---|
| 973 | wb_addr<=pcx_packet_2nd[103:64]; |
---|
| 974 | wb_data_o<=pcx_packet_d[63:0]; |
---|
| 975 | wb_sel<=8'h23; |
---|
| 976 | end |
---|
| 977 | end |
---|
| 978 | `FP_WAIT: |
---|
| 979 | begin |
---|
| 980 | fp_pcx<=124'b0; |
---|
| 981 | fp_req<=0; |
---|
| 982 | if(fp_rdy) |
---|
| 983 | state<=`CPX_FP; |
---|
| 984 | if(`DEBUGGING) |
---|
| 985 | wb_sel<=8'h24; |
---|
| 986 | end |
---|
| 987 | `CPX_FP: |
---|
| 988 | if(fp_cpx[144]) // Packet valid |
---|
| 989 | begin |
---|
| 990 | cpx_packet_1<=fp_cpx; |
---|
| 991 | state<=`CPX_READY_1; |
---|
| 992 | if(`DEBUGGING) |
---|
| 993 | begin |
---|
| 994 | wb_addr<=fp_cpx[63:0]; |
---|
| 995 | wb_data_o<=fp_cpx[127:64]; |
---|
| 996 | end |
---|
| 997 | end |
---|
| 998 | else |
---|
| 999 | if(!fp_rdy) |
---|
| 1000 | state<=`FP_WAIT; // Else wait for another one if it is not here still |
---|
| 1001 | `CPX_SEND_ETH_IRQ: |
---|
| 1002 | begin |
---|
| 1003 | cpx_packet_1<=145'h1_7_000_000000000000001D_000000000000_001D; |
---|
| 1004 | eth_int_sent<=0; |
---|
| 1005 | state<=`CPX_READY_1; |
---|
| 1006 | end |
---|
| 1007 | `CPX_INT_VEC_DIS: |
---|
| 1008 | begin |
---|
| 1009 | //if(pcx_packet_d[12:10]==3'b000) // Send interrupt only if it is for this core |
---|
| 1010 | cpx_two_packet<=1; |
---|
| 1011 | cpu2<=pcx_packet_d[10]; |
---|
| 1012 | cpx_packet_1[144:140]<=5'b10100; |
---|
| 1013 | cpx_packet_1[139:137]<=0; |
---|
| 1014 | cpx_packet_1[136]<=1; |
---|
| 1015 | cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID |
---|
| 1016 | cpx_packet_1[133:130]<=0; |
---|
| 1017 | cpx_packet_1[129]<=pcx_atom_d; |
---|
| 1018 | cpx_packet_1[128]<=0; |
---|
| 1019 | cpx_packet_1[127:0]<={5'b0,pcx_packet_d[64+5:64+4],2'b0,cpu,pcx_packet_d[64+11:64+6],112'b0}; |
---|
| 1020 | cpx_packet_2<={9'h170,54'h0,pcx_packet_d[17:0],46'h0,pcx_packet_d[17:0]}; |
---|
| 1021 | state<=`CPX_READY_1; |
---|
| 1022 | end |
---|
| 1023 | `CPX_READY_1: |
---|
| 1024 | begin |
---|
| 1025 | if(!cpu) |
---|
| 1026 | begin |
---|
| 1027 | cpx_ready<=1; |
---|
| 1028 | cpx_packet<=cpx_packet_1; |
---|
| 1029 | if(othercpuhit[0]) |
---|
| 1030 | begin |
---|
| 1031 | cpx1_ready<=1; |
---|
| 1032 | cpx1_packet<={1'b1,4'b0011,12'b0,5'b0,pcx_packet_d[64+5:64+4],3'b001,pcx_packet_d[64+11:64+6],inval_vect0}; |
---|
| 1033 | end |
---|
| 1034 | end |
---|
| 1035 | else |
---|
| 1036 | begin |
---|
| 1037 | cpx1_ready<=1; |
---|
| 1038 | cpx1_packet<=cpx_packet_1; |
---|
| 1039 | if(othercpuhit[0]) |
---|
| 1040 | begin |
---|
| 1041 | cpx_ready<=1; |
---|
| 1042 | cpx_packet<={1'b1,4'b0011,12'b0,5'b0,pcx_packet_d[64+5:64+4],3'b000,pcx_packet_d[64+11:64+6],inval_vect0}; |
---|
| 1043 | end |
---|
| 1044 | end |
---|
| 1045 | cnt<=cnt+1; |
---|
| 1046 | if(`DEBUGGING) |
---|
| 1047 | if(multi_hit || multi_hit1) |
---|
| 1048 | wb_sel<=8'h11; |
---|
| 1049 | state<=`CPX_READY_2; |
---|
| 1050 | end |
---|
| 1051 | `CPX_READY_2: |
---|
| 1052 | begin |
---|
| 1053 | if(cpx_two_packet && !cpu2) |
---|
| 1054 | begin |
---|
| 1055 | cpx_ready<=1; |
---|
| 1056 | cpx_packet<=cpx_packet_2; |
---|
| 1057 | end |
---|
| 1058 | else |
---|
| 1059 | if(cpu2 && othercpuhit[1]) |
---|
| 1060 | begin |
---|
| 1061 | cpx_ready<=1; |
---|
| 1062 | cpx_packet<={1'b1,4'b0011,12'b0,5'b0,pcx_packet_d[64+5],1'b1,3'b000,pcx_packet_d[64+11:64+6],inval_vect1}; |
---|
| 1063 | end |
---|
| 1064 | else |
---|
| 1065 | begin |
---|
| 1066 | cpx_ready<=0; |
---|
| 1067 | cpx_packet<=145'b0; |
---|
| 1068 | end |
---|
| 1069 | if(cpx_two_packet && cpu2) |
---|
| 1070 | begin |
---|
| 1071 | cpx1_ready<=1; |
---|
| 1072 | cpx1_packet<=cpx_packet_2; |
---|
| 1073 | end |
---|
| 1074 | else |
---|
| 1075 | if(!cpu2 && othercpuhit[1]) |
---|
| 1076 | begin |
---|
| 1077 | cpx1_ready<=1; |
---|
| 1078 | cpx1_packet<={1'b1,4'b0011,12'b0,5'b0,pcx_packet_d[64+5],1'b1,3'b001,pcx_packet_d[64+11:64+6],inval_vect1}; |
---|
| 1079 | end |
---|
| 1080 | else |
---|
| 1081 | begin |
---|
| 1082 | cpx1_ready<=0; |
---|
| 1083 | cpx1_packet<=145'b0; |
---|
| 1084 | end |
---|
| 1085 | state<=`PCX_IDLE; |
---|
| 1086 | end |
---|
| 1087 | `PCX_UNKNOWN: |
---|
| 1088 | begin |
---|
| 1089 | wb_sel<=8'b10100101; // Illegal eye-catching value for debugging |
---|
| 1090 | state<=`PCX_IDLE; |
---|
| 1091 | end |
---|
| 1092 | endcase |
---|
| 1093 | |
---|
| 1094 | l1dir l1dir_inst( |
---|
| 1095 | .clk(clk), |
---|
| 1096 | .reset(!rstn), |
---|
| 1097 | |
---|
| 1098 | .cpu(cpu), // Issuing CPU number |
---|
| 1099 | .strobe(state==`GOT_PCX_REQ), |
---|
| 1100 | .way(pcx_packet[108:107]), // Way to allocate for allocating loads |
---|
| 1101 | .address(pcx_packet[64+39:64]), |
---|
| 1102 | .load(pcx_packet[122:118]==5'b00000), |
---|
| 1103 | .ifill(pcx_packet[122:118]==5'b10000), |
---|
| 1104 | .store(pcx_packet[122:118]==5'b00001), |
---|
| 1105 | .cas(pcx_packet[122:118]==5'b00010), |
---|
| 1106 | .swap(pcx_packet[122:118]==5'b00110), |
---|
| 1107 | .strload(pcx_packet[122:118]==5'b00100), |
---|
| 1108 | .strstore(pcx_packet[122:118]==5'b00101), |
---|
| 1109 | .cacheable((!pcx_packet[117]) && (!pcx_req_d[4])), |
---|
| 1110 | .prefetch(pcx_packet[110]), |
---|
| 1111 | .invalidate(pcx_packet[111]), |
---|
| 1112 | .blockstore(pcx_packet[109] | pcx_packet[110]), |
---|
| 1113 | |
---|
| 1114 | .inval_vect0(inval_vect0), // Invalidation vector |
---|
| 1115 | .inval_vect1(inval_vect1), |
---|
| 1116 | .othercachehit(othercachehit), // Other cache hit in the same CPU, wayval0/wayval1 |
---|
| 1117 | .othercpuhit(othercpuhit), // Any cache hit in the other CPU, wayval0/wayval1 |
---|
| 1118 | .wayval0(wayval0), // Way valid |
---|
| 1119 | .wayval1(wayval1), // Second way valid for ifill |
---|
| 1120 | .ready(ready) // Directory init done |
---|
| 1121 | ); |
---|
| 1122 | |
---|
| 1123 | endmodule |
---|