Context Navigation

source: XOpenSparcT1/trunk/os2wb/os2wb.v @ 22

Revision 22, 59.0 KB checked in by pntsvt00, 14 years ago (diff)
checkpoint: la DDR effettua l'init

Rev	Line
[10]	1	`timescale 1ns / 1ps
	2	//////////////////////////////////////////////////////////////////////////////////
	3	// Company: (C) Athree, 2009
	4	// Engineer: Dmitry Rozhdestvenskiy
	5	// Email dmitry.rozhdestvenskiy@srisc.com dmitryr@a3.spb.ru divx4log@narod.ru
	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(
	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	// Wishbone master interface
	34	input [ 63:0] wb_data_i,
	35	input wb_ack,
	36	output reg wb_cycle,
	37	output reg wb_strobe,
	38	output reg wb_we,
	39	output reg [ 7:0] wb_sel,
	40	output reg [ 63:0] wb_addr,
	41	output reg [ 63:0] wb_data_o,
	42
	43	// FPU interface
	44	output reg [123:0] fp_pcx,
	45	output reg fp_req,
	46	input [144:0] fp_cpx,
	47	input fp_rdy,
	48
	49	// Ethernet interrupt, sensed on posedge, mapped to vector 'd29
	50	input eth_int
	51	);
	52
	53	reg [123:0] pcx_packet_d; // Latched incoming PCX packet
	54	reg [123:0] pcx_packet_2nd; // Second packet for atomic (CAS)
	55	reg [ 4:0] pcx_req_d; // Latched request
	56	reg pcx_atom_d; // Latched atomic flasg
	57	reg [ 4:0] state; // FSM state
	58	reg [144:0] cpx_packet_1; // First CPX packet
	59	reg [144:0] cpx_packet_2; // Second CPX packet (for atomics and cached IFILLs)
	60	reg cpx_two_packet; // CPX answer is two-packet (!=atomic, SWAP has atomic==0 and answer is two-packet)
	61
	62	reg [ 3:0] inval_vect0; // Invalidate, instr/data, way
	63	reg [ 3:0] inval_vect1; // IFill may cause two D lines invalidation at a time
	64
	65	wire [111:0] store_inv_vec; // Store invalidation vector
	66
	67	assign store_inv_vec[111:91]=0;
	68	assign store_inv_vec[90:88]=((pcx_packet_d[64+5:64+4]==2'b11) && inval_vect0[3:2]==2'b11) ? {inval_vect0[1:0],1'b1}:3'b000;
	69	assign store_inv_vec[87:60]=0;
	70	assign store_inv_vec[59:56]=((pcx_packet_d[64+5:64+4]==2'b10) && inval_vect0[3:2]==2'b11) \|\| ((pcx_packet_d[64+5]==1'b1) && inval_vect0[3:2]==2'b10) ? {inval_vect0[1:0],!inval_vect0[2],inval_vect0[2]}:4'b0000;
	71	assign store_inv_vec[55:35]=0;
	72	assign store_inv_vec[34:32]=((pcx_packet_d[64+5:64+4]==2'b01) && inval_vect0[3:2]==2'b11) ? {inval_vect0[1:0],1'b1}:3'b000;
	73	assign store_inv_vec[31:4]=0;
	74	assign store_inv_vec[3:0]=((pcx_packet_d[64+5:64+4]==2'b00) && inval_vect0[3:2]==2'b11) \|\| ((pcx_packet_d[64+5]==1'b0) && inval_vect0[3:2]==2'b10) ? {inval_vect0[1:0],!inval_vect0[2],inval_vect0[2]}:4'b0000;
	75
	76	wire [28:0] dcache0_do0;
	77	wire [28:0] dcache0_do1;
	78	wire [28:0] dcache1_do0;
	79	wire [28:0] dcache1_do1;
	80	wire [28:0] dcache2_do0;
	81	wire [28:0] dcache2_do1;
	82	wire [28:0] dcache3_do0;
	83	wire [28:0] dcache3_do1;
	84	wire [28:0] icache0_do;
	85	wire [28:0] icache1_do;
	86	wire [28:0] icache2_do;
	87	wire [28:0] icache3_do;
	88
	89	`define TEST_DRAM_1 5'b00000
	90	`define TEST_DRAM_2 5'b00001
	91	`define TEST_DRAM_3 5'b00010
	92	`define TEST_DRAM_4 5'b00011
	93	`define INIT_DRAM_1 5'b00100
	94	`define INIT_DRAM_2 5'b00101
	95	`define WAKEUP 5'b00110
	96	`define PCX_IDLE 5'b00111
	97	`define GOT_PCX_REQ 5'b01000
	98	`define PCX_REQ_2ND 5'b01001
	99	`define PCX_REQ_STEP1 5'b01010
	100	`define PCX_REQ_STEP1_1 5'b01011
	101	`define PCX_REQ_STEP2 5'b01100
	102	`define PCX_REQ_STEP2_1 5'b01101
	103	`define PCX_REQ_STEP3 5'b01110
	104	`define PCX_REQ_STEP3_1 5'b01111
	105	`define PCX_REQ_STEP4 5'b10000
	106	`define PCX_REQ_STEP4_1 5'b10001
	107	`define PCX_BIS 5'b10010
	108	`define PCX_BIS_1 5'b10011
	109	`define PCX_BIS_2 5'b10100
	110	`define CPX_READY_1 5'b10101
	111	`define CPX_READY_2 5'b10110
	112	`define PCX_UNKNOWN 5'b11000
	113	`define PCX_FP_1 5'b11001
	114	`define PCX_FP_2 5'b11010
	115	`define FP_WAIT 5'b11011
	116	`define CPX_FP 5'b11100
	117	`define CPX_SEND_ETH_IRQ 5'b11101
	118	`define CPX_INT_VEC_DIS 5'b11110
	119	`define PCX_REQ_CAS_COMPARE 5'b11111
	120
[22]	121	//`define MEM_SIZE 64'h00000000_10000000 //256 MB
	122	//`define MEM_SIZE 64'h00000000_00100000 //1MB
	123	`define MEM_SIZE 64'h00000000_00001000 //256KB
[10]	124
[22]	125	// sal: escludo test della DRAM `define TEST_DRAM 1
[10]	126	`define TEST_DRAM 1
	127	`define DEBUGGING 1
	128
	129	reg cache_init;
	130	wire [3:0] dcache0_hit;
	131	wire [3:0] dcache1_hit;
	132	wire [3:0] icache_hit;
	133	reg multi_hit;
	134	reg multi_hit1;
	135	reg eth_int_d;
	136	reg eth_int_send;
	137	reg eth_int_sent;
	138	reg [3:0] cnt;
	139
	140	// PCX channel FIFO
	141	wire [129:0] pcx_data_fifo;
	142	wire pcx_fifo_empty;
	143	reg [ 4:0] pcx_req_1;
	144	reg [ 4:0] pcx_req_2;
	145	reg pcx_atom_1;
	146	reg pcx_atom_2;
	147	reg pcx_data_123_d;
	148
	149	always @(posedge clk)
	150	begin
	151	pcx_req_1<=pcx_req;
	152	pcx_atom_1<=pcx_atom;
	153	pcx_atom_2<=pcx_atom_1;
	154	pcx_req_2<=pcx_atom_1 ? pcx_req_1:5'b0;
	155	pcx_grant<=(pcx_req_1 \| pcx_req_2);
	156	pcx_data_123_d<=pcx_data[123];
	157	end
	158
	159	/*pcx_fifo pcx_fifo_inst(
	160	// FIFO should be first word fall-through
	161	// It has no full flag as the core will send only limited number of requests,
	162	// in original design we used it 32 words deep
	163	// Just make it deeper if you experience overflow -
	164	// you can't just send no grant on full because the core expects immediate
	165	// grant for at least two requests for each zone
	166	.aclr(!rstn),
	167	.clock(clk),
	168	.data({pcx_atom_1,pcx_req_1,pcx_data}),
	169	.rdreq(fifo_rd),
	170	.wrreq((pcx_req_1!=5'b00000 && pcx_data[123]) \|\| (pcx_atom_2 && pcx_data_123_d)),
	171	// Second atomic packet for FPU may be invalid, but should be sent to FPU
	172	// so if the first atomic packet is valid we latch both
	173	.empty(pcx_fifo_empty),
	174	.q(pcx_data_fifo)
	175	);
	176	*/
[14]	177	reg fifo_rd;
	178	wire [123:0] pcx_packet;
	179
	180
[10]	181	pcx_fifo pcx_fifo_inst(
	182	.clk(clk),
[22]	183	.rst(!rstn),
[10]	184	.din({pcx_atom_1,pcx_req_1,pcx_data}),
	185	.rd_en(fifo_rd),
	186	.wr_en((pcx_req_1!=5'b00000 && pcx_data[123]) \|\| (pcx_atom_2 && pcx_data_123_d)),
	187	.empty(pcx_fifo_empty),
[22]	188	.full(),
[10]	189	.dout(pcx_data_fifo)
	190	);
	191
	192
	193
	194	// --------------------------
	195
	196	always @(posedge clk or negedge rstn)
	197	if(!rstn)
	198	eth_int_send<=0;
	199	else
	200	begin
	201	eth_int_d<=eth_int;
	202	if(eth_int && !eth_int_d)
	203	eth_int_send<=1;
	204	else
	205	if(eth_int_sent)
	206	eth_int_send<=0;
	207	end
	208
[14]	209
[10]	210	assign pcx_packet=pcx_data_fifo[123:0];
	211
	212	always @(posedge clk or negedge rstn)
	213	if(rstn==0)
	214	begin
[22]	215	//$display("INFO: OS2WB: RST_DRAM at %t",$time);
[10]	216	if(`TEST_DRAM)
	217	state<=`TEST_DRAM_1;
	218	else
[22]	219	state<=`INIT_DRAM_1; // DRAM initialization is mandatory!
[10]	220	cpx_ready<=0;
	221	fifo_rd<=0;
	222	cpx_packet<=145'b0;
	223	wb_cycle<=0;
	224	wb_strobe<=0;
	225	wb_we<=0;
	226	wb_sel<=0;
	227	wb_addr<=64'b0;
	228	wb_data_o<=64'b0;
	229	pcx_packet_d<=124'b0;
	230	fp_pcx<=124'b0;
	231	fp_req<=0;
	232	end
	233	else
	234	case(state)
	235	`TEST_DRAM_1:
	236	begin
[22]	237	$display("INFO: OS2WB: TEST_DRAM_1");
[10]	238	wb_cycle<=1;
	239	wb_strobe<=1;
	240	wb_sel<=8'hFF;
	241	wb_we<=1;
	242	state<=`TEST_DRAM_2;
	243	end
	244	`TEST_DRAM_2:
	245	if(wb_ack)
	246	begin
[22]	247	$display("INFO: OS2WB: TEST_DRAM_2 at time %d with wb_addr=%x",$time,wb_addr);
[10]	248	wb_strobe<=0;
	249	if(wb_addr<`MEM_SIZE-8)
	250	begin
	251	wb_addr[31:0]<=wb_addr[31:0]+8;
	252	wb_data_o<={wb_addr[31:0]+8,wb_addr[31:0]+8};
	253	state<=`TEST_DRAM_1;
	254	end
	255	else
	256	begin
	257	state<=`TEST_DRAM_3;
	258	wb_cycle<=0;
	259	wb_sel<=0;
	260	wb_we<=0;
	261	wb_data_o<=64'b0;
	262	wb_addr<=64'b0;
	263	end
	264	end
	265	`TEST_DRAM_3:
	266	begin
[22]	267	$display("INFO: OS2WB: TEST_DRAM_3");
[10]	268	wb_cycle<=1;
	269	wb_strobe<=1;
	270	wb_sel<=8'hFF;
	271	state<=`TEST_DRAM_4;
	272	end
	273	`TEST_DRAM_4:
	274	if(wb_ack)
	275	begin
[22]	276	$display("INFO: OS2WB: TEST_DRAM_4 at %t",$time);
[10]	277	wb_strobe<=0;
	278	if(wb_addr<`MEM_SIZE-8)
	279	begin
	280	if(wb_data_i=={wb_addr[31:0],wb_addr[31:0]})
	281	begin
	282	wb_addr[31:0]<=wb_addr[31:0]+8;
	283	state<=`TEST_DRAM_3;
	284	end
[22]	285	else
	286	$display("INFO: OS2WB: TEST_DRAM_4 error in read addres %x at %t",wb_addr,$time);
	287
[10]	288	end
	289	else
	290	begin
[22]	291	$display("INFO: OS2WB: INIT_DRAM at %t",$time);
[10]	292	state<=`INIT_DRAM_1;
	293	wb_cycle<=0;
	294	wb_sel<=0;
	295	wb_we<=0;
	296	wb_data_o<=64'b0;
	297	wb_addr<=64'b0;
	298	end
	299	end
	300	`INIT_DRAM_1:
	301	begin
	302	wb_cycle<=1;
	303	wb_strobe<=1;
	304	wb_sel<=8'hFF;
	305	wb_we<=1;
	306	cache_init<=1; // We also init cache directories here
	307	state<=`INIT_DRAM_2;
	308	end
	309	`INIT_DRAM_2:
	310	if(wb_ack)
	311	begin
	312	wb_strobe<=0;
	313	if(wb_addr<`MEM_SIZE-8)
	314	begin
[22]	315	//for debug
	316	// if (wb_addr[10:3]==8'b0)
	317	// $display("INFO: OS2WB: INIT_DRAM_2 at time %d with wb_addr=%x",$time,wb_addr);
	318	//
[10]	319	wb_addr[31:0]<=wb_addr[31:0]+8;
	320	pcx_packet_d[64+11:64+4]<=pcx_packet_d[64+11:64+4]+1; // Address for cachedir init
	321	state<=`INIT_DRAM_1;
	322	end
	323	else
	324	begin
[22]	325	$display("INFO: OS2WB: WAKEUP_DRAM at %t",$time);
[10]	326	state<=`WAKEUP;
	327	wb_cycle<=0;
	328	wb_sel<=0;
	329	wb_we<=0;
	330	cache_init<=0;
	331	wb_addr<=64'b0;
	332	end
	333	end
	334	`WAKEUP:
	335	begin
	336	cpx_packet<=145'h1700000000000000000000000000000010001;
	337	cpx_ready<=1;
	338	state<=`PCX_IDLE;
	339	end
	340	`PCX_IDLE:
	341	begin
	342	cnt<=0;
	343	cpx_packet<=145'b0;
	344	cpx_ready<=0;
	345	cpx_two_packet<=0;
	346	inval_vect0[3]<=0;
	347	inval_vect1[3]<=0;
	348	multi_hit<=0;
	349	multi_hit1<=0;
	350	if(eth_int_send)
	351	begin
	352	state<=`CPX_SEND_ETH_IRQ;
	353	eth_int_sent<=1;
	354	end
	355	else
	356	if(!pcx_fifo_empty)
	357	begin
	358	pcx_req_d<=pcx_data_fifo[128:124];
	359	pcx_atom_d<=pcx_data_fifo[129];
	360	fifo_rd<=1;
	361	state<=`GOT_PCX_REQ;
	362	end
	363	end
	364	`GOT_PCX_REQ:
	365	begin
	366	pcx_packet_d<=pcx_packet;
	367	if(`DEBUGGING)
	368	begin
[22]	369	$display("INFO: OS2WB: GOT_PCX_REQ");
[10]	370	wb_sel[1:0]<=pcx_packet[113:112];
	371	wb_sel[2]<=1;
	372	end
	373	if(pcx_packet[103:64]==40'h9800000800 && pcx_packet[122:118]==5'b00001)
	374	begin
	375	state<=`CPX_INT_VEC_DIS;
	376	fifo_rd<=0;
	377	end
	378	else
	379	if(pcx_atom_d==0)
	380	begin
	381	fifo_rd<=0;
	382	if(pcx_packet[122:118]==5'b01010) // FP req
	383	begin
	384	state<=`PCX_FP_1;
	385	pcx_packet_2nd[123]<=0;
	386	end
	387	else
	388	state<=`PCX_REQ_STEP1;
	389	end
	390	else
	391	state<=`PCX_REQ_2ND;
	392	end
	393	`PCX_REQ_2ND:
	394	begin
	395	pcx_packet_2nd<=pcx_packet; //Latch second packet for atomics
	396	if(`DEBUGGING)
	397	if(pcx_fifo_empty)
	398	wb_sel<=8'h67;
	399	fifo_rd<=0;
	400	if(pcx_packet_d[122:118]==5'b01010) // FP req
	401	state<=`PCX_FP_1;
	402	else
	403	state<=`PCX_REQ_STEP1;
	404	end
	405	`PCX_REQ_STEP1:
	406	begin
	407	if(pcx_packet_d[111]==1'b1) // Invalidate request
	408	begin
	409	cpx_packet_1[144]<=1; // Valid
	410	cpx_packet_1[143:140]<=4'b0100; // Invalidate reply is Store ACK
	411	cpx_packet_1[139]<=1; // L2 miss
	412	cpx_packet_1[138:137]<=0; // Error
	413	cpx_packet_1[136]<=pcx_packet_d[117]; // Non-cacheble
	414	cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID
	415	cpx_packet_1[133:131]<=0; // Way valid
	416	cpx_packet_1[130]<=((pcx_packet_d[122:118]==5'b10000) && (pcx_req_d==5'b10000)) ? 1:0; // Four byte fill
	417	cpx_packet_1[129]<=pcx_atom_d;
	418	cpx_packet_1[128]<=pcx_packet_d[110]; // Prefetch
	419	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],3'b0,pcx_packet_d[64+11:64+6],112'b0};
	420	state<=`CPX_READY_1;
	421	end
	422	else
	423	if(pcx_packet_d[122:118]!=5'b01001) // Not INT
	424	begin
	425	wb_cycle<=1'b1;
	426	wb_strobe<=1'b1;
	427	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)
	428	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
	429	else
	430	if(pcx_packet_d[122:118]==5'b10000 && !pcx_req_d[4])
	431	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b00000}; //DRAM ifill
	432	else
	433	if(pcx_packet_d[64+39:64+28]==12'hFFF && pcx_packet_d[64+27:64+24]!=4'b0) // flash remap FFF1->FFF8
	434	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3]+37'h0000E00000,3'b000};
	435	else
	436	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3],3'b000};
	437	wb_data_o<=pcx_packet_d[63:0];
	438	state<=`PCX_REQ_STEP1_1;
	439	end
	440	else
	441	if((pcx_packet_d[12:10]!=3'b000) && !pcx_packet_d[117]) // Not FLUSH int and not this core
	442	state<=`PCX_IDLE;
	443	else
	444	state<=`CPX_READY_1;
	445	case(pcx_packet_d[122:118]) // Packet type
	446	5'b00000://Load
	447	begin
	448	wb_we<=0;
	449	if(!pcx_packet_d[110] && !pcx_packet_d[117])
	450	case(icache_hit)
	451	4'b0000:;
	452	4'b0001:inval_vect0<=4'b1_0_00;
	453	4'b0010:inval_vect0<=4'b1_0_01;
	454	4'b0100:inval_vect0<=4'b1_0_10;
	455	4'b1000:inval_vect0<=4'b1_0_11;
	456	default:multi_hit<=1;
	457	endcase
	458	if(!pcx_req_d[4])
	459	wb_sel<=8'b11111111; // DRAM requests are always 128 bit
	460	else
	461	case(pcx_packet_d[106:104]) //Size
	462	3'b000://Byte
	463	case(pcx_packet_d[64+2:64])
	464	3'b000:wb_sel<=8'b10000000;
	465	3'b001:wb_sel<=8'b01000000;
	466	3'b010:wb_sel<=8'b00100000;
	467	3'b011:wb_sel<=8'b00010000;
	468	3'b100:wb_sel<=8'b00001000;
	469	3'b101:wb_sel<=8'b00000100;
	470	3'b110:wb_sel<=8'b00000010;
	471	3'b111:wb_sel<=8'b00000001;
	472	endcase
	473	3'b001://Halfword
	474	case(pcx_packet_d[64+2:64+1])
	475	2'b00:wb_sel<=8'b11000000;
	476	2'b01:wb_sel<=8'b00110000;
	477	2'b10:wb_sel<=8'b00001100;
	478	2'b11:wb_sel<=8'b00000011;
	479	endcase
	480	3'b010://Word
	481	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	482	3'b011://Doubleword
	483	wb_sel<=8'b11111111;
	484	3'b100://Quadword
	485	wb_sel<=8'b11111111;
	486	3'b111://Cacheline
	487	wb_sel<=8'b11111111;
	488	default:
	489	wb_sel<=8'b01011010; // Unreal eye-catching value for debug
	490	endcase
	491	end
	492	5'b00001://Store
	493	begin
	494	wb_we<=1;
	495	case({icache_hit,dcache0_hit})
	496	8'b00000000:;
	497	8'b00000001:inval_vect0<=4'b1_1_00;
	498	8'b00000010:inval_vect0<=4'b1_1_01;
	499	8'b00000100:inval_vect0<=4'b1_1_10;
	500	8'b00001000:inval_vect0<=4'b1_1_11;
	501	8'b00010000:inval_vect0<=4'b1_0_00;
	502	8'b00100000:inval_vect0<=4'b1_0_01;
	503	8'b01000000:inval_vect0<=4'b1_0_10;
	504	8'b10000000:inval_vect0<=4'b1_0_11;
	505	default:multi_hit<=1;
	506	endcase
	507	if(pcx_packet_d[110:109]!=2'b00) //Block (or init) store
	508	wb_sel<=8'b11111111; // Blocks are always 64 bit
	509	else
	510	case(pcx_packet_d[106:104]) //Size
	511	3'b000://Byte
	512	case(pcx_packet_d[64+2:64])
	513	3'b000:wb_sel<=8'b10000000;
	514	3'b001:wb_sel<=8'b01000000;
	515	3'b010:wb_sel<=8'b00100000;
	516	3'b011:wb_sel<=8'b00010000;
	517	3'b100:wb_sel<=8'b00001000;
	518	3'b101:wb_sel<=8'b00000100;
	519	3'b110:wb_sel<=8'b00000010;
	520	3'b111:wb_sel<=8'b00000001;
	521	endcase
	522	3'b001://Halfword
	523	case(pcx_packet_d[64+2:64+1])
	524	2'b00:wb_sel<=8'b11000000;
	525	2'b01:wb_sel<=8'b00110000;
	526	2'b10:wb_sel<=8'b00001100;
	527	2'b11:wb_sel<=8'b00000011;
	528	endcase
	529	3'b010://Word
	530	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	531	3'b011://Doubleword
	532	wb_sel<=8'b11111111;
	533	default:
	534	if(`DEBUGGING)
	535	wb_sel<=8'b01011010; // Unreal eye-catching value for debug
	536	endcase
	537	end
	538	5'b00010://CAS
	539	begin
	540	wb_we<=0; //Load first
	541	case({icache_hit,dcache0_hit})
	542	8'b00000000:;
	543	8'b00000001:inval_vect0<=4'b1_1_00;
	544	8'b00000010:inval_vect0<=4'b1_1_01;
	545	8'b00000100:inval_vect0<=4'b1_1_10;
	546	8'b00001000:inval_vect0<=4'b1_1_11;
	547	8'b00010000:inval_vect0<=4'b1_0_00;
	548	8'b00100000:inval_vect0<=4'b1_0_01;
	549	8'b01000000:inval_vect0<=4'b1_0_10;
	550	8'b10000000:inval_vect0<=4'b1_0_11;
	551	default:multi_hit<=1;
	552	endcase
	553	wb_sel<=8'b11111111; // CAS loads are as cacheline
	554	end
	555	5'b00100://STRLOAD
	556	begin
	557	wb_we<=0;
	558	wb_sel<=8'b11111111; // Stream loads are always 128 bit
	559	end
	560	5'b00101://STRSTORE
	561	begin
	562	wb_we<=1;
	563	case({icache_hit,dcache0_hit})
	564	8'b00000000:;
	565	8'b00000001:inval_vect0<=4'b1_1_00;
	566	8'b00000010:inval_vect0<=4'b1_1_01;
	567	8'b00000100:inval_vect0<=4'b1_1_10;
	568	8'b00001000:inval_vect0<=4'b1_1_11;
	569	8'b00010000:inval_vect0<=4'b1_0_00;
	570	8'b00100000:inval_vect0<=4'b1_0_01;
	571	8'b01000000:inval_vect0<=4'b1_0_10;
	572	8'b10000000:inval_vect0<=4'b1_0_11;
	573	default:multi_hit<=1;
	574	endcase
	575	case(pcx_packet_d[106:104]) //Size
	576	3'b000://Byte
	577	case(pcx_packet_d[64+2:64])
	578	3'b000:wb_sel<=8'b10000000;
	579	3'b001:wb_sel<=8'b01000000;
	580	3'b010:wb_sel<=8'b00100000;
	581	3'b011:wb_sel<=8'b00010000;
	582	3'b100:wb_sel<=8'b00001000;
	583	3'b101:wb_sel<=8'b00000100;
	584	3'b110:wb_sel<=8'b00000010;
	585	3'b111:wb_sel<=8'b00000001;
	586	endcase
	587	3'b001://Halfword
	588	case(pcx_packet_d[64+2:64+1])
	589	2'b00:wb_sel<=8'b11000000;
	590	2'b01:wb_sel<=8'b00110000;
	591	2'b10:wb_sel<=8'b00001100;
	592	2'b11:wb_sel<=8'b00000011;
	593	endcase
	594	3'b010://Word
	595	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	596	3'b011://Doubleword
	597	wb_sel<=8'b11111111;
	598	3'b100://Quadword
	599	wb_sel<=8'b11111111;
	600	3'b111://Cacheline
	601	wb_sel<=8'b11111111;
	602	default:
	603	wb_sel<=8'b01011010; // Unreal eye-catching value for debug
	604	endcase
	605	end
	606	5'b00110://SWAP/LDSTUB
	607	begin
	608	case({icache_hit,dcache0_hit})
	609	8'b00000000:;
	610	8'b00000001:inval_vect0<=4'b1_1_00;
	611	8'b00000010:inval_vect0<=4'b1_1_01;
	612	8'b00000100:inval_vect0<=4'b1_1_10;
	613	8'b00001000:inval_vect0<=4'b1_1_11;
	614	8'b00010000:inval_vect0<=4'b1_0_00;
	615	8'b00100000:inval_vect0<=4'b1_0_01;
	616	8'b01000000:inval_vect0<=4'b1_0_10;
	617	8'b10000000:inval_vect0<=4'b1_0_11;
	618	default:multi_hit<=1;
	619	endcase
	620	wb_we<=0; // Load first, as CAS
	621	wb_sel<=8'b11111111; // SWAP/LDSTUB loads are as cacheline
	622	end
	623	5'b01001://INT
	624	if(pcx_packet_d[117]) // Flush
	625	cpx_packet_1<={9'h171,pcx_packet_d[113:112],11'h0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],30'h0,pcx_packet_d[17:0],46'b0,pcx_packet_d[17:0]}; //FLUSH instruction answer
	626	else // Tread-to-thread interrupt
	627	cpx_packet_1<={9'h170,pcx_packet_d[113:112],52'h0,pcx_packet_d[17:0],46'h0,pcx_packet_d[17:0]};
	628	//5'b01010: FP1 - processed by separate state
	629	//5'b01011: FP2 - processed by separate state
	630	//5'b01101: FWDREQ - not implemented
	631	//5'b01110: FWDREPL - not implemented
	632	5'b10000://IFILL
	633	begin
	634	wb_we<=0;
	635	if(!pcx_req_d[4]) // not I/O access
	636	begin
	637	case(dcache0_hit)
	638	4'b0000:;
	639	4'b0001:inval_vect0<=4'b1_1_00;
	640	4'b0010:inval_vect0<=4'b1_1_01;
	641	4'b0100:inval_vect0<=4'b1_1_10;
	642	4'b1000:inval_vect0<=4'b1_1_11;
	643	default:multi_hit<=1;
	644	endcase
	645	case(dcache1_hit)
	646	4'b0000:;
	647	4'b0001:inval_vect1<=4'b1_1_00;
	648	4'b0010:inval_vect1<=4'b1_1_01;
	649	4'b0100:inval_vect1<=4'b1_1_10;
	650	4'b1000:inval_vect1<=4'b1_1_11;
	651	default:multi_hit1<=1;
	652	endcase
	653	end
	654	if(pcx_req_d[4]) // I/O access
	655	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	656	else
	657	wb_sel<=8'b11111111;
	658	end
	659	default:
	660	begin
	661	wb_we<=0;
	662	wb_sel<=8'b10101010; // Unreal eye-catching value for debug
	663	end
	664	endcase
	665	end
	666	`PCX_REQ_STEP1_1:
	667	begin
	668	if(wb_ack)
	669	begin
	670	cpx_packet_1[144]<=1; // Valid
	671	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
	672	cpx_packet_1[138:137]<=0; // Error
	673	cpx_packet_1[136]<=pcx_packet_d[117] \|\| (pcx_packet_d[122:118]==5'b00001) ? 1:0; // Non-cacheble is set on store too
	674	cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID
	675	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
	676	cpx_packet_1[133:131]<={inval_vect0[3],inval_vect0[1:0]};
	677	else
	678	cpx_packet_1[133:131]<=3'b000; // Way valid
	679	if(pcx_packet_d[122:118]==5'b00100) // Strload
	680	cpx_packet_1[130]<=pcx_packet_d[106]; // A
	681	else
	682	if(pcx_packet_d[122:118]==5'b00101) // Stream store
	683	cpx_packet_1[130]<=pcx_packet_d[108]; // A
	684	else
	685	cpx_packet_1[130]<=((pcx_packet_d[122:118]==5'b10000) && pcx_req_d[4]) ? 1:0; // Four byte fill
	686	if(pcx_packet_d[122:118]==5'b00100) // Strload
	687	cpx_packet_1[129]<=pcx_packet_d[105]; // B
	688	else
	689	cpx_packet_1[129]<=pcx_atom_d \|\| (pcx_packet_d[122:118]==5'b00110); // SWAP is single-packet but needs atom in CPX
	690	cpx_packet_1[128]<=pcx_packet_d[110] && pcx_packet_d[122:118]==5'b00000; // Prefetch
	691	cpx_packet_2[144]<=1; // Valid
	692	cpx_packet_2[139]<=0; // L2 miss
	693	cpx_packet_2[138:137]<=0; // Error
	694	cpx_packet_2[136]<=pcx_packet_d[117] \|\| (pcx_packet_d[122:118]==5'b00001) ? 1:0; // Non-cacheble is set on store too
	695	cpx_packet_2[135:134]<=pcx_packet_d[113:112]; // Thread ID
	696	if(pcx_packet_d[122:118]==5'b10000) // IFill
	697	cpx_packet_2[133:131]<={inval_vect1[3],inval_vect1[1:0]};
	698	else
	699	cpx_packet_2[133:131]<=3'b000; // Way valid
	700	cpx_packet_2[130]<=0; // Four byte fill
	701	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]);
	702	cpx_packet_2[128]<=0; // Prefetch
	703	wb_strobe<=0;
	704	wb_sel<=8'b0;
	705	wb_addr<=64'b0;
	706	wb_data_o<=64'b0;
	707	wb_we<=0;
	708	case(pcx_packet_d[122:118]) // Packet type
	709	5'b00000://Load
	710	begin
	711	cpx_packet_1[143:140]<=4'b0000; // Type
	712	if(!pcx_req_d[4])
	713	begin
	714	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	715	state<=`PCX_REQ_STEP2;
	716	end
	717	else
	718	case(pcx_packet_d[106:104]) //Size
	719	3'b000://Byte
	720	begin
	721	case(pcx_packet_d[64+2:64])
	722	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]};
	723	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]};
	724	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]};
	725	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]};
	726	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]};
	727	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]};
	728	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]};
	729	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]};
	730	endcase
	731	wb_cycle<=0;
	732	state<=`CPX_READY_1;
	733	end
	734	3'b001://Halfword
	735	begin
	736	case(pcx_packet_d[64+2:64+1])
	737	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]};
	738	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]};
	739	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]};
	740	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]};
	741	endcase
	742	wb_cycle<=0;
	743	state<=`CPX_READY_1;
	744	end
	745	3'b010://Word
	746	begin
	747	if(pcx_packet_d[64+2]==0)
	748	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]};
	749	else
	750	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]};
	751	wb_cycle<=0;
	752	state<=`CPX_READY_1;
	753	end
	754	3'b011://Doubleword
	755	begin
	756	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	757	wb_cycle<=0;
	758	state<=`CPX_READY_1;
	759	end
	760	3'b100://Quadword
	761	begin
	762	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	763	wb_cycle<=0;
	764	state<=`CPX_READY_1; // 16 byte access to PROM should just duplicate the data
	765	end
	766	3'b111://Cacheline
	767	begin
	768	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	769	wb_cycle<=0;
	770	state<=`CPX_READY_1; // 16 byte access to PROM should just duplicate the data
	771	end
	772	default:
	773	begin
	774	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	775	wb_cycle<=0;
	776	state<=`PCX_UNKNOWN;
	777	end
	778	endcase
	779	end
	780	5'b00001://Store
	781	begin
	782	cpx_packet_1[143:140]<=4'b0100; // Type
	783	cpx_packet_1[127:0]<={2'b0,pcx_packet_d[109]/BIS/,2'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	784	// if((pcx_packet_d[110:109]==2'b01) && (pcx_packet_d[64+5:64]==0) && !inval_vect0[3] && !inval_vect1[3]) // Block init store
	785	// state<=`PCX_BIS;
	786	// else
	787	// begin
	788	wb_cycle<=0;
	789	state<=`CPX_READY_1;
	790	// end
	791	end
	792	5'b00010://CAS
	793	begin
	794	cpx_packet_1[143:140]<=4'b0000; // Load return for first packet
	795	cpx_packet_2[143:140]<=4'b0100; // Store ACK for second packet
	796	cpx_packet_2[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	797	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	798	state<=`PCX_REQ_STEP2;
	799	end
	800	5'b00100://STRLOAD
	801	begin
	802	cpx_packet_1[143:140]<=4'b0010; // Type
	803	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	804	state<=`PCX_REQ_STEP2;
	805	end
	806	5'b00101://STRSTORE
	807	begin
	808	cpx_packet_1[143:140]<=4'b0110; // Type
	809	cpx_packet_1[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	810	wb_cycle<=0;
	811	state<=`CPX_READY_1;
	812	end
	813	5'b00110://SWAP/LDSTUB
	814	begin
	815	cpx_packet_1[143:140]<=4'b0000; // Load return for first packet
	816	cpx_packet_2[143:140]<=4'b0100; // Store ACK for second packet
	817	cpx_packet_2[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	818	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	819	state<=`PCX_REQ_STEP2;
	820	end
	821	5'b10000://IFILL
	822	begin
	823	cpx_packet_1[143:140]<=4'b0001; // Type
	824	cpx_packet_2[143:140]<=4'b0001; // Type
	825	if(pcx_req_d[4]) // I/O access
	826	begin
	827	if(pcx_packet_d[64+2]==0)
	828	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]};
	829	else
	830	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]};
	831	state<=`CPX_READY_1;
	832	wb_cycle<=0;
	833	end
	834	else
	835	begin
	836	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	837	state<=`PCX_REQ_STEP2;
	838	end
	839	end
	840	default:
	841	begin
	842	wb_cycle<=0;
	843	state<=`PCX_UNKNOWN;
	844	end
	845	endcase
	846	end
	847	end
	848	`PCX_REQ_STEP2: // IFill, Load/strload, CAS, SWAP, LDSTUB - alwas load
	849	begin
	850	wb_strobe<=1'b1;
	851	if(pcx_packet_d[122:118]==5'b10000)
	852	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b01000};
	853	else
	854	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+4],4'b1000};
	855	wb_sel<=8'b11111111; // It is always full width for subsequent IFill and load accesses
	856	state<=`PCX_REQ_STEP2_1;
	857	end
	858	`PCX_REQ_STEP2_1:
	859	if(wb_ack==1)
	860	begin
	861	wb_strobe<=0;
	862	wb_sel<=8'b0;
	863	wb_addr<=64'b0;
	864	wb_data_o<=64'b0;
	865	wb_we<=0;
	866	cpx_packet_1[63:0]<=wb_data_i;
	867	if((pcx_packet_d[122:118]!=5'b00000) && (pcx_packet_d[122:118]!=5'b00100))
	868	if(pcx_packet_d[122:118]!=5'b00010) // IFill, SWAP
	869	state<=`PCX_REQ_STEP3;
	870	else
	871	state<=`PCX_REQ_CAS_COMPARE; // CAS
	872	else
	873	begin
	874	wb_cycle<=0;
	875	state<=`CPX_READY_1;
	876	end
	877	end
	878	`PCX_REQ_CAS_COMPARE:
	879	begin
	880	cpx_two_packet<=1;
	881	if(pcx_packet_d[106:104]==3'b010) // 32-bit
	882	case(pcx_packet_d[64+3:64+2])
	883	2'b00:state<=cpx_packet_1[127:96]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	884	2'b01:state<=cpx_packet_1[95:64]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	885	2'b10:state<=cpx_packet_1[63:32]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	886	2'b11:state<=cpx_packet_1[31:0]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	887	endcase
	888	else
	889	if(pcx_packet_d[64+3]==0)
	890	state<=cpx_packet_1[127:64]==pcx_packet_d[63:0] ? `PCX_REQ_STEP3:`CPX_READY_1;
	891	else
	892	state<=cpx_packet_1[63:0]==pcx_packet_d[63:0] ? `PCX_REQ_STEP3:`CPX_READY_1;
	893	end
	894	`PCX_REQ_STEP3: // 256-bit IFILL; CAS, SWAP and LDSTUB store
	895	begin
	896	if(pcx_packet_d[122:118]==5'b10000)
	897	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b10000};
	898	else
	899	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3],3'b000}; // CAS or SWAP save
	900	cpx_two_packet<=1;
	901	if(pcx_packet_d[122:118]==5'b10000)
	902	wb_we<=0;
	903	else
	904	wb_we<=1;
	905	wb_strobe<=1'b1;
	906	if(pcx_packet_d[122:118]==5'b00010) // CAS
	907	if(pcx_packet_d[106:104]==3'b010)
	908	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	909	else
	910	wb_sel<=8'b11111111; //CASX
	911	else
	912	if(pcx_packet_d[122:118]==5'b00110) //SWAP or LDSTUB
	913	if(pcx_packet_d[106:104]==3'b000) //LDSTUB
	914	case(pcx_packet_d[64+2:64])
	915	3'b000:wb_sel<=8'b10000000;
	916	3'b001:wb_sel<=8'b01000000;
	917	3'b010:wb_sel<=8'b00100000;
	918	3'b011:wb_sel<=8'b00010000;
	919	3'b100:wb_sel<=8'b00001000;
	920	3'b101:wb_sel<=8'b00000100;
	921	3'b110:wb_sel<=8'b00000010;
	922	3'b111:wb_sel<=8'b00000001;
	923	endcase
	924	else
	925	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; ///SWAP is always 32-bit
	926	else
	927	wb_sel<=8'b11111111; // It is always full width for subsequent IFill accesses
	928	if(pcx_packet_d[122:118]==5'b00110) //SWAP or LDSTUB
	929	wb_data_o<={pcx_packet_d[63:32],pcx_packet_d[63:32]};
	930	// wb_data_o<=pcx_packet_d[63:0];
	931	else
	932	wb_data_o<=pcx_packet_2nd[63:0]; // CAS store second packet data
	933	// if(pcx_packet_d[106:104]==3'b010)
	934	// wb_data_o<={pcx_packet_2nd[63:32],pcx_packet_2nd[63:32]}; // CAS store second packet data
	935	// else
	936	// wb_data_o<=pcx_packet_2nd[63:0];
	937	state<=`PCX_REQ_STEP3_1;
	938	end
	939	`PCX_REQ_STEP3_1:
	940	if(wb_ack==1)
	941	begin
	942	wb_strobe<=0;
	943	wb_sel<=8'b0;
	944	wb_addr<=64'b0;
	945	wb_we<=0;
	946	wb_data_o<=64'b0;
	947	if(pcx_packet_d[122:118]==5'b10000) // IFill
	948	begin
	949	cpx_packet_2[127:64]<=wb_data_i;
	950	state<=`PCX_REQ_STEP4;
	951	end
	952	else
	953	begin
	954	wb_cycle<=0;
	955	state<=`CPX_READY_1;
	956	end
	957	end
	958	`PCX_REQ_STEP4: // 256-bit IFILL only
	959	begin
	960	wb_strobe<=1'b1;
	961	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b11000};
	962	wb_sel<=8'b11111111; // It is always full width for subsequent accesses
	963	state<=`PCX_REQ_STEP4_1;
	964	end
	965	`PCX_REQ_STEP4_1:
	966	if(wb_ack==1)
	967	begin
	968	wb_cycle<=0;
	969	wb_strobe<=0;
	970	wb_sel<=8'b0;
	971	wb_addr<=64'b0;
	972	wb_we<=0;
	973	cpx_packet_2[63:0]<=wb_data_i;
	974	state<=`CPX_READY_1;
	975	end
	976	`PCX_BIS: // Block init store
	977	begin
	978	wb_strobe<=1'b1;
	979	wb_we<=1;
	980	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+6],6'b001000};
	981	wb_sel<=8'b11111111;
	982	wb_data_o<=64'b0;
	983	state<=`PCX_BIS_1;
	984	end
	985	`PCX_BIS_1:
	986	if(wb_ack)
	987	begin
	988	wb_strobe<=0;
	989	if(wb_addr[39:0]<(pcx_packet_d[64+39:64]+8*7))
	990	state<=`PCX_BIS_2;
	991	else
	992	begin
	993	wb_cycle<=0;
	994	wb_sel<=0;
	995	wb_we<=0;
	996	wb_addr<=64'b0;
	997	state<=`CPX_READY_1;
	998	end
	999	end
	1000	`PCX_BIS_2:
	1001	begin
	1002	wb_strobe<=1'b1;
	1003	wb_addr[5:0]<=wb_addr[5:0]+8;
	1004	state<=`PCX_BIS_1;
	1005	end
	1006	`PCX_FP_1:
	1007	begin
	1008	fp_pcx<=pcx_packet_d;
	1009	fp_req<=1;
	1010	state<=`PCX_FP_2;
	1011	if(`DEBUGGING)
	1012	begin
	1013	wb_addr<=pcx_packet_d[103:64];
	1014	wb_data_o<=pcx_packet_d[63:0];
	1015	wb_sel<=8'h22;
	1016	end
	1017	end
	1018	`PCX_FP_2:
	1019	begin
	1020	fp_pcx<=pcx_packet_2nd;
	1021	state<=`FP_WAIT;
	1022	if(`DEBUGGING)
	1023	begin
	1024	wb_addr<=pcx_packet_2nd[103:64];
	1025	wb_data_o<=pcx_packet_d[63:0];
	1026	wb_sel<=8'h23;
	1027	end
	1028	end
	1029	`FP_WAIT:
	1030	begin
	1031	fp_pcx<=124'b0;
	1032	fp_req<=0;
	1033	if(fp_rdy)
	1034	state<=`CPX_FP;
	1035	if(`DEBUGGING)
	1036	wb_sel<=8'h24;
	1037	end
	1038	`CPX_FP:
	1039	if(fp_cpx[144]) // Packet valid
	1040	begin
	1041	cpx_packet_1<=fp_cpx;
	1042	state<=`CPX_READY_1;
	1043	if(`DEBUGGING)
	1044	begin
	1045	wb_addr<=fp_cpx[63:0];
	1046	wb_data_o<=fp_cpx[127:64];
	1047	end
	1048	end
	1049	else
	1050	if(!fp_rdy)
	1051	state<=`FP_WAIT; // Else wait for another one if it is not here still
	1052	`CPX_SEND_ETH_IRQ:
	1053	begin
	1054	cpx_packet_1<=145'h1_7_000_000000000000001D_000000000000_001D;
	1055	eth_int_sent<=0;
	1056	state<=`CPX_READY_1;
	1057	end
	1058	`CPX_INT_VEC_DIS:
	1059	begin
	1060	if(pcx_packet_d[12:10]==3'b000)
	1061	cpx_two_packet<=1; // Send interrupt only if it is for this core
	1062	cpx_packet_1[144:140]<=5'b10100;
	1063	cpx_packet_1[139:137]<=0;
	1064	cpx_packet_1[136]<=1;
	1065	cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID
	1066	cpx_packet_1[133:130]<=0;
	1067	cpx_packet_1[129]<=pcx_atom_d;
	1068	cpx_packet_1[128]<=0;
	1069	cpx_packet_1[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],112'b0};
	1070	cpx_packet_2<={9'h170,54'h0,pcx_packet_d[17:0],46'h0,pcx_packet_d[17:0]};
	1071	state<=`CPX_READY_1;
	1072	end
	1073	`CPX_READY_1:
	1074	begin
	1075	cpx_ready<=1;
	1076	cpx_packet<=cpx_packet_1;
	1077	cnt<=cnt+1;
	1078	if(`DEBUGGING)
	1079	if(multi_hit \|\| multi_hit1)
	1080	wb_sel<=8'h11;
	1081	if(!cpx_two_packet)
	1082	state<=`PCX_IDLE;
	1083	else
	1084	//if(cnt==4'b1111 \|\| pcx_packet_d[103:64]!=40'h9800000800)
	1085	state<=`CPX_READY_2;
	1086	end
	1087	`CPX_READY_2:
	1088	begin
	1089	cpx_ready<=1;
	1090	cpx_packet<=cpx_packet_2;
	1091	state<=`PCX_IDLE;
	1092	end
	1093	`PCX_UNKNOWN:
	1094	begin
	1095	wb_sel<=8'b10100101; // Illegal eye-catching value for debugging
	1096	state<=`PCX_IDLE;
	1097	end
	1098	endcase
	1099
	1100	/* Cache directory checking:
	1101	Load: allocate D if cacheable, check I, invalidate&deallocate if found
	1102	Store: check I, invalidate&deallocate if found; check D, invalidate if found
	1103	IFill: allocate I if cacheable, check D, invalidate&deallocate if found
	1104	SWAP/LDSTUB: check I, invalidate&deallocate if found; check D, invalidate&deallocate if found
	1105	CAS: Like SWAP
	1106
	1107	Allocation and querying is made simultaneously at GOT_PCX_REQ
	1108	(memory read mode does not matter as long as allocation and invalidation
	1109	are never made to the same directory, so if memory is written its output will not be checked)
	1110	Invalidation vectors are built during PCX_REQ_STEP1, or Invalidate all ways issued
	1111	During PCX_REQ_STEP1_1 directory is deallocated if needed
	1112
	1113	*/
	1114
	1115	// Directory enable
	1116	assign dir_en=((state==`GOT_PCX_REQ) \|\| (state==`PCX_REQ_STEP1) \|\| cache_init \|\|
	1117	((state==`PCX_REQ_STEP1_1) && wb_ack)) ? 1:0;
	1118
	1119	// ICache deallocation flag
	1120	assign loadstore=((pcx_packet_d[122:118]==5'b00000) && !pcx_packet_d[117] && !pcx_packet_d[110]) \|\| // cacheable load, not prefetch
	1121	(pcx_packet_d[122:118]==5'b00001) \|\| (pcx_packet_d[122:118]==5'b00010) \|\| // Store, CAS
	1122	(pcx_packet_d[122:118]==5'b00110) \|\| (pcx_packet_d[122:118]==5'b00101); // SWAP/LDSTUB, StrStore
	1123
	1124	// DCache deallocation flag
	1125	assign ifillcas=(pcx_packet_d[122:118]==5'b00110) \|\| (pcx_packet_d[122:118]==5'b00010) \|\| //SWAP, CAS
	1126	(pcx_packet_d[122:118]==5'b10000) \|\| (pcx_packet_d[122:118]==5'b00101) \|\| // IFill, StrStore
	1127	((pcx_packet_d[122:118]==5'b00001) && pcx_packet_d[110:109]!=2'b00); // Block (or init) store
	1128
	1129	// DCache allocation flag
	1130	assign cacheload=(pcx_packet[122:118]==5'b00000) && !pcx_packet[110] && !pcx_packet[117] && !pcx_packet[111];
	1131
	1132	// ICache allocation flag
	1133	assign cacheifill=(pcx_packet[122:118]==5'b10000) && !pcx_packet[117] && !pcx_packet[111];
	1134
	1135	assign dcache0_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b00) && cacheload;
	1136	assign dcache0_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_00) && ifillcas;
	1137	assign dcache0_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_00) && ifillcas;
	1138
	1139	assign dcache1_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b01) && cacheload;
	1140	assign dcache1_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_01) && ifillcas;
	1141	assign dcache1_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_01) && ifillcas;
	1142
	1143	assign dcache2_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b10) && cacheload;
	1144	assign dcache2_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_10) && ifillcas;
	1145	assign dcache2_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_10) && ifillcas;
	1146
	1147	assign dcache3_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b11) && cacheload;
	1148	assign dcache3_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_11) && ifillcas;
	1149	assign dcache3_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_11) && ifillcas;
	1150
	1151	assign icache0_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b00) && cacheifill;
	1152	assign icache0_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_00) && loadstore;
	1153
	1154	assign icache1_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b01) && cacheifill;
	1155	assign icache1_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_01) && loadstore;
	1156
	1157	assign icache2_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b10) && cacheifill;
	1158	assign icache2_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_10) && loadstore;
	1159
	1160	assign icache3_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b11) && cacheifill;
	1161	assign icache3_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_11) && loadstore;
	1162
	1163	assign dcache_inval_all=(state==`PCX_REQ_STEP1) && pcx_packet_d[111] && pcx_packet_d[122:118]==5'b00000;
	1164	assign icache_inval_all=(state==`PCX_REQ_STEP1) && pcx_packet_d[111] && pcx_packet_d[122:118]==5'b10000;
	1165
	1166	`define INVAL_TAG 29'h10000000
	1167
	1168	// DCache least address bit for first bank
	1169	// it should be 0 for IFill (1 is hardcoded for second bank)
	1170	assign dcache_la=(state==`GOT_PCX_REQ) ? (pcx_packet[122:118]==5'b10000 ? 1'b0:pcx_packet[64+4]):
	1171	(pcx_packet_d[122:118]==5'b10000 ? 1'b0:pcx_packet_d[64+4]);
	1172
	1173	wire [ 6:0] dcache_index;
	1174	wire [28:0] dcache_data;
	1175	assign dcache_index=(state==`GOT_PCX_REQ) ? pcx_packet[64+10:64+5]:pcx_packet_d[64+10:64+5];
	1176	assign dcache_data=(state==`GOT_PCX_REQ) ? pcx_packet[64+39:64+11]:`INVAL_TAG;
	1177
	1178	cachedir dcache0 (
	1179	.clock(clk),
	1180	.enable(dir_en),
	1181	.wren_a(dcache0_alloc \|\| dcache0_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1182	.address_a({1'b0,dcache_index,dcache_la}),
	1183	.data_a(dcache_data),
	1184	.q_a(dcache0_do0),
	1185
	1186	.wren_b(dcache0_dealloc1),
	1187	.address_b({1'b0,dcache_index,1'b1}),
	1188	.data_b(`INVAL_TAG),
	1189	.q_b(dcache0_do1)
	1190	);
	1191
	1192	cachedir dcache1 (
	1193	.clock(clk),
	1194	.enable(dir_en),
	1195	.wren_a(dcache1_alloc \|\| dcache1_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1196	.address_a({1'b0,dcache_index,dcache_la}),
	1197	.data_a(dcache_data),
	1198	.q_a(dcache1_do0),
	1199
	1200	.wren_b(dcache1_dealloc1),
	1201	.address_b({1'b0,dcache_index,1'b1}),
	1202	.data_b(`INVAL_TAG),
	1203	.q_b(dcache1_do1)
	1204	);
	1205
	1206	cachedir dcache2 (
	1207	.clock(clk),
	1208	.enable(dir_en),
	1209	.wren_a(dcache2_alloc \|\| dcache2_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1210	.address_a({1'b0,dcache_index,dcache_la}),
	1211	.data_a(dcache_data),
	1212	.q_a(dcache2_do0),
	1213
	1214	.wren_b(dcache2_dealloc1),
	1215	.address_b({1'b0,dcache_index,1'b1}),
	1216	.data_b(`INVAL_TAG),
	1217	.q_b(dcache2_do1)
	1218	);
	1219
	1220	cachedir dcache3 (
	1221	.clock(clk),
	1222	.enable(dir_en),
	1223	.wren_a(dcache3_alloc \|\| dcache3_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1224	.address_a({1'b0,dcache_index,dcache_la}),
	1225	.data_a(dcache_data),
	1226	.q_a(dcache3_do0),
	1227
	1228	.wren_b(dcache3_dealloc1),
	1229	.address_b({1'b0,dcache_index,1'b1}),
	1230	.data_b(`INVAL_TAG),
	1231	.q_b(dcache3_do1)
	1232	);
	1233
	1234	assign dcache0_hit={dcache3_do0==pcx_packet_d[64+39:64+11],
	1235	dcache2_do0==pcx_packet_d[64+39:64+11],
	1236	dcache1_do0==pcx_packet_d[64+39:64+11],
	1237	dcache0_do0==pcx_packet_d[64+39:64+11]};
	1238	assign dcache1_hit={dcache3_do1==pcx_packet_d[64+39:64+11],
	1239	dcache2_do1==pcx_packet_d[64+39:64+11],
	1240	dcache1_do1==pcx_packet_d[64+39:64+11],
	1241	dcache0_do1==pcx_packet_d[64+39:64+11]};
	1242
	1243	wire [ 6:0] icache_index;
	1244	wire [28:0] icache_data;
	1245	assign icache_index=(state==`GOT_PCX_REQ) ? pcx_packet[64+11:64+5]:pcx_packet_d[64+11:64+5];
	1246	assign icache_data=(state==`GOT_PCX_REQ) ? {pcx_packet[64+39:64+12],1'b0}:`INVAL_TAG;
	1247
	1248	cachedir icache01 (
	1249	.clock(clk),
	1250	.enable(dir_en),
	1251	.wren_a(icache0_alloc \|\| icache0_dealloc \|\| icache_inval_all \|\| cache_init),
[17]	1252	.address_a({2'b0,icache_index}),
[10]	1253	.data_a(icache_data),
	1254	.q_a(icache0_do),
	1255
	1256	.wren_b(icache1_alloc \|\| icache1_dealloc \|\| icache_inval_all \|\| cache_init),
[22]	1257	.address_b({2'b0,icache_index}),
[10]	1258	.data_b(icache_data),
	1259	.q_b(icache1_do)
	1260	);
	1261
	1262	cachedir icache23 (
	1263	.clock(clk),
	1264	.enable(dir_en),
	1265	.wren_a(icache2_alloc \|\| icache2_dealloc \|\| icache_inval_all \|\| cache_init),
[22]	1266	.address_a({2'b0,icache_index}),
[10]	1267	.data_a(icache_data),
	1268	.q_a(icache2_do),
	1269
	1270	.wren_b(icache3_alloc \|\| icache3_dealloc \|\| icache_inval_all \|\| cache_init),
[22]	1271	.address_b({2'b0,icache_index}),
[10]	1272	.data_b(icache_data),
	1273	.q_b(icache3_do)
	1274	);
	1275
	1276	assign icache_hit={icache3_do[28:1]==pcx_packet_d[64+39:64+12],
	1277	icache2_do[28:1]==pcx_packet_d[64+39:64+12],
	1278	icache1_do[28:1]==pcx_packet_d[64+39:64+12],
	1279	icache0_do[28:1]==pcx_packet_d[64+39:64+12]};
	1280
	1281	/*
	1282	case(pcx_packet_d[122:118]) // Packet type
	1283	5'b00000://Load
	1284	5'b00001://Store
	1285	5'b00010://CAS
	1286	5'b00100://STRLOAD
	1287	5'b00101://STRSTORE
	1288	5'b00110://SWAP
	1289	5'b01001://INT
	1290	//5'b01010://FP1
	1291	//5'b01011://FP2
	1292	//5'b01101://FWDREQ
	1293	//5'b01110://FWDREPL
	1294	5'b10000://IFILL
	1295	endcase
	1296	*/
[14]	1297	endmodule

Note: See TracBrowser for help on using the repository browser.

Download in other formats: