project directories restructuring

This commit is contained in:
Blaise Tine
2020-04-14 08:24:49 -04:00
parent fc155e1223
commit 3bbaea1332
46 changed files with 0 additions and 5199 deletions

View File

@@ -0,0 +1,637 @@
`include "VX_cache_config.v"
`include "VX_define.v"
module VX_bank
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Forward SNP Queue
parameter FFSQ_SIZE = 8,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
// Input Core Request
input wire delay_req,
input wire [NUMBER_REQUESTS-1:0] bank_valids,
input wire [NUMBER_REQUESTS-1:0][31:0] bank_addr,
input wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] bank_writedata,
input wire [4:0] bank_rd,
input wire [NUMBER_REQUESTS-1:0][1:0] bank_wb,
input wire [31:0] bank_pc,
input wire [`NW_M1:0] bank_warp_num,
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_read,
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_write,
output wire reqq_full,
// Output Core WB
input wire bank_wb_pop,
output wire bank_wb_valid,
output wire [`vx_clog2(NUMBER_REQUESTS)-1:0] bank_wb_tid,
output wire [4:0] bank_wb_rd,
output wire [1:0] bank_wb_wb,
output wire [`NW_M1:0] bank_wb_warp_num,
output wire [`WORD_SIZE_RNG] bank_wb_data,
output wire [31:0] bank_wb_pc,
output wire [31:0] bank_wb_address,
// Dram Fill Requests
output wire dram_fill_req,
output wire[31:0] dram_fill_req_addr,
output wire dram_because_of_snp,
output wire dram_snp_full,
input wire dram_fill_req_queue_full,
// Dram Fill Response
input wire dram_fill_rsp,
input wire [31:0] dram_fill_addr,
input wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dram_fill_rsp_data,
output wire dram_fill_accept,
// Dram WB Requests
input wire dram_wb_queue_pop,
output wire dram_wb_req,
output wire[31:0] dram_wb_req_addr,
output wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dram_wb_req_data,
// Snp Request
input wire snp_req,
input wire[31:0] snp_req_addr,
output wire snrq_full,
output wire snp_fwd,
output wire[31:0] snp_fwd_addr,
input wire snp_fwd_pop
);
reg snoop_state = 0;
always @(posedge clk) begin
if (reset) begin
snoop_state <= 0;
end else begin
snoop_state <= (snoop_state | snp_req) && ((FUNC_ID == `LLFUNC_ID) || (FUNC_ID == `L3FUNC_ID));
end
end
wire snrq_pop;
wire snrq_empty;
wire snrq_valid_st0;
wire[31:0] snrq_addr_st0;
assign snrq_valid_st0 = !snrq_empty;
VX_generic_queue_ll #(.DATAW(32), .SIZE(SNRQ_SIZE)) snr_queue(
.clk (clk),
.reset (reset),
.push (snp_req),
.in_data (snp_req_addr),
.pop (snrq_pop),
.out_data(snrq_addr_st0),
.empty (snrq_empty),
.full (snrq_full)
);
wire dfpq_pop;
wire dfpq_empty;
wire dfpq_full;
wire[31:0] dfpq_addr_st0;
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dfpq_filldata_st0;
assign dram_fill_accept = !dfpq_full;
VX_generic_queue_ll #(.DATAW(32+(`BANK_LINE_SIZE_WORDS*`WORD_SIZE)), .SIZE(DFPQ_SIZE)) dfp_queue(
.clk (clk),
.reset (reset),
.push (dram_fill_rsp),
.in_data ({dram_fill_addr, dram_fill_rsp_data}),
.pop (dfpq_pop),
.out_data({dfpq_addr_st0, dfpq_filldata_st0}),
.empty (dfpq_empty),
.full (dfpq_full)
);
wire reqq_pop;
wire reqq_push;
wire reqq_empty;
wire reqq_req_st0;
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] reqq_req_tid_st0;
wire [31:0] reqq_req_addr_st0;
wire [`WORD_SIZE_RNG] reqq_req_writeword_st0;
wire [4:0] reqq_req_rd_st0;
wire [1:0] reqq_req_wb_st0;
wire [`NW_M1:0] reqq_req_warp_num_st0;
wire [2:0] reqq_req_mem_read_st0;
wire [2:0] reqq_req_mem_write_st0;
wire [31:0] reqq_req_pc_st0;
assign reqq_push = !delay_req && (|bank_valids);
VX_cache_req_queue #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
req_queue
(
.clk (clk),
.reset (reset),
// Enqueue
.reqq_push (reqq_push),
.bank_valids (bank_valids),
.bank_addr (bank_addr),
.bank_writedata (bank_writedata),
.bank_rd (bank_rd),
.bank_pc (bank_pc),
.bank_wb (bank_wb),
.bank_warp_num (bank_warp_num),
.bank_mem_read (bank_mem_read),
.bank_mem_write (bank_mem_write),
// Dequeue
.reqq_pop (reqq_pop),
.reqq_req_st0 (reqq_req_st0),
.reqq_req_tid_st0 (reqq_req_tid_st0),
.reqq_req_addr_st0 (reqq_req_addr_st0),
.reqq_req_writedata_st0(reqq_req_writeword_st0),
.reqq_req_rd_st0 (reqq_req_rd_st0),
.reqq_req_wb_st0 (reqq_req_wb_st0),
.reqq_req_warp_num_st0 (reqq_req_warp_num_st0),
.reqq_req_mem_read_st0 (reqq_req_mem_read_st0),
.reqq_req_mem_write_st0(reqq_req_mem_write_st0),
.reqq_req_pc_st0 (reqq_req_pc_st0),
.reqq_empty (reqq_empty),
.reqq_full (reqq_full)
);
wire mrvq_pop;
wire mrvq_full;
wire mrvq_stop;
wire mrvq_valid_st0;
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] mrvq_tid_st0;
wire [31:0] mrvq_addr_st0;
wire [`WORD_SIZE_RNG] mrvq_writeword_st0;
wire [4:0] mrvq_rd_st0;
wire [1:0] mrvq_wb_st0;
wire [31:0] miss_resrv_pc_st0;
wire [`NW_M1:0] mrvq_warp_num_st0;
wire [2:0] mrvq_mem_read_st0;
wire [2:0] mrvq_mem_write_st0;
wire miss_add;
wire[31:0] miss_add_addr;
wire[`WORD_SIZE_RNG] miss_add_data;
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] miss_add_tid;
wire[4:0] miss_add_rd;
wire[1:0] miss_add_wb;
wire[`NW_M1:0] miss_add_warp_num;
wire[2:0] miss_add_mem_read;
wire[2:0] miss_add_mem_write;
wire[31:0] miss_add_pc;
wire[31:0] addr_st2;
wire is_fill_st2;
VX_cache_miss_resrv #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
mrvq_queue
(
.clk (clk),
.reset (reset),
// Enqueue
.miss_add (miss_add), // Need to do all
.miss_add_addr (miss_add_addr),
.miss_add_data (miss_add_data),
.miss_add_tid (miss_add_tid),
.miss_add_rd (miss_add_rd),
.miss_add_wb (miss_add_wb),
.miss_add_warp_num (miss_add_warp_num),
.miss_add_mem_read (miss_add_mem_read),
.miss_add_mem_write (miss_add_mem_write),
.miss_add_pc (miss_add_pc),
.miss_resrv_full (mrvq_full),
.miss_resrv_stop (mrvq_stop),
// Broadcast
.is_fill_st1 (is_fill_st2),
.fill_addr_st1 (addr_st2),
// Dequeue
.miss_resrv_pop (mrvq_pop),
.miss_resrv_valid_st0 (mrvq_valid_st0),
.miss_resrv_addr_st0 (mrvq_addr_st0),
.miss_resrv_data_st0 (mrvq_writeword_st0),
.miss_resrv_tid_st0 (mrvq_tid_st0),
.miss_resrv_rd_st0 (mrvq_rd_st0),
.miss_resrv_wb_st0 (mrvq_wb_st0),
.miss_resrv_pc_st0 (miss_resrv_pc_st0),
.miss_resrv_warp_num_st0 (mrvq_warp_num_st0),
.miss_resrv_mem_read_st0 (mrvq_mem_read_st0),
.miss_resrv_mem_write_st0(mrvq_mem_write_st0)
);
wire stall_bank_pipe;
reg is_fill_in_pipe;
wire valid_st1 [STAGE_1_CYCLES-1:0];
wire is_fill_st1 [STAGE_1_CYCLES-1:0];
wire going_to_write_st1[STAGE_1_CYCLES-1:0];
wire [31:0] addr_st1 [STAGE_1_CYCLES-1:0];
reg[16:0] p_stage;
always @(*) begin
is_fill_in_pipe = 0;
for (p_stage = 0; p_stage < STAGE_1_CYCLES; p_stage=p_stage+1) begin
if (is_fill_st1[p_stage]) is_fill_in_pipe = 1;
end
if (is_fill_st2) is_fill_in_pipe = 1;
end
// assign is_fill_in_pipe = (|is_fill_st1) || is_fill_st2;
assign mrvq_pop = mrvq_valid_st0 && !stall_bank_pipe;
assign dfpq_pop = !mrvq_pop && !dfpq_empty && !stall_bank_pipe;
assign reqq_pop = !mrvq_stop && !mrvq_pop && !dfpq_pop && !reqq_empty && reqq_req_st0 && !stall_bank_pipe && !is_fill_st1[0] && !is_fill_in_pipe;
assign snrq_pop = !reqq_pop && !reqq_pop && !mrvq_pop && !dfpq_pop && snrq_valid_st0 && !stall_bank_pipe;
integer st1_cycle;
wire qual_is_fill_st0;
wire qual_valid_st0;
wire [31:0] qual_addr_st0;
wire [`WORD_SIZE_RNG] qual_writeword_st0;
wire [`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] qual_writedata_st0;
wire [`REQ_INST_META_SIZE-1:0] qual_inst_meta_st0;
wire qual_going_to_write_st0;
wire qual_is_snp;
wire [31:0] qual_pc_st0;
wire [`WORD_SIZE_RNG] writeword_st1 [STAGE_1_CYCLES-1:0];
wire [`REQ_INST_META_SIZE-1:0] inst_meta_st1 [STAGE_1_CYCLES-1:0];
wire [`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] writedata_st1[STAGE_1_CYCLES-1:0];
wire is_snp_st1 [STAGE_1_CYCLES-1:0];
wire [31:0] pc_st1 [STAGE_1_CYCLES-1:0];
assign qual_is_fill_st0 = dfpq_pop;
// always @(*) begin
// if (qual_is_fill_st0 && (FUNC_ID == 3)) begin
// $display("WHAT THE FUCK FUNC_ID: %x", FUNC_ID);
// end
// end
assign qual_valid_st0 = dfpq_pop || mrvq_pop || reqq_pop || snrq_pop;
assign qual_addr_st0 = dfpq_pop ? dfpq_addr_st0 :
mrvq_pop ? mrvq_addr_st0 :
reqq_pop ? reqq_req_addr_st0 :
snrq_pop ? snrq_addr_st0 :
0;
assign qual_writedata_st0 = dfpq_pop ? dfpq_filldata_st0 : 57;
assign qual_inst_meta_st0 = mrvq_pop ? {mrvq_rd_st0 , mrvq_wb_st0 , mrvq_warp_num_st0 , mrvq_mem_read_st0 , mrvq_mem_write_st0 , mrvq_tid_st0 } :
reqq_pop ? {reqq_req_rd_st0, reqq_req_wb_st0, reqq_req_warp_num_st0, reqq_req_mem_read_st0, reqq_req_mem_write_st0, reqq_req_tid_st0} :
0;
assign qual_going_to_write_st0 = dfpq_pop ? 1 :
(mrvq_pop && (mrvq_mem_write_st0 != `NO_MEM_WRITE)) ? 1 :
(reqq_pop && (reqq_req_mem_write_st0 != `NO_MEM_WRITE)) ? 1 :
(snrq_pop) ? 1 :
0;
assign qual_pc_st0 = (reqq_pop) ? reqq_req_pc_st0 :
(mrvq_pop) ? miss_resrv_pc_st0 :
(dfpq_pop) ? 32'hdeadbeef :
(snrq_pop) ? 32'hb00b0000 :
32'h0;
assign qual_is_snp = snrq_pop ? 1 : 0;
assign qual_writeword_st0 = mrvq_pop ? mrvq_writeword_st0 :
reqq_pop ? reqq_req_writeword_st0 :
0;
VX_generic_register #(.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_SIZE_WORDS*`WORD_SIZE) + 1 + 32)) s0_1_c0 (
.clk (clk),
.reset(reset),
.stall(stall_bank_pipe),
.flush(0),
.in ({qual_is_snp , qual_going_to_write_st0, qual_valid_st0, qual_addr_st0, qual_writeword_st0, qual_inst_meta_st0, qual_is_fill_st0, qual_writedata_st0, qual_pc_st0 }),
.out ({is_snp_st1[0], going_to_write_st1[0] , valid_st1[0] , addr_st1[0] , writeword_st1[0] , inst_meta_st1[0] , is_fill_st1[0] , writedata_st1[0] , pc_st1[0]})
);
genvar curr_stage;
generate
for (curr_stage = 1; curr_stage < STAGE_1_CYCLES; curr_stage = curr_stage + 1) begin
VX_generic_register #(.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_SIZE_WORDS*`WORD_SIZE) + 1 + 32)) s0_1_cc (
.clk (clk),
.reset(reset),
.stall(stall_bank_pipe),
.flush(0),
.in ({is_snp_st1[curr_stage-1], going_to_write_st1[curr_stage-1], valid_st1[curr_stage-1], addr_st1[curr_stage-1], writeword_st1[curr_stage-1], inst_meta_st1[curr_stage-1], is_fill_st1[curr_stage-1] , writedata_st1[curr_stage-1], pc_st1[curr_stage-1]}),
.out ({is_snp_st1[curr_stage] , going_to_write_st1[curr_stage] , valid_st1[curr_stage] , addr_st1[curr_stage] , writeword_st1[curr_stage] , inst_meta_st1[curr_stage] , is_fill_st1[curr_stage] , writedata_st1[curr_stage] , pc_st1[curr_stage]})
);
end
endgenerate
wire[`WORD_SIZE_RNG] readword_st1e;
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] readdata_st1e;
wire[`TAG_SELECT_SIZE_RNG] readtag_st1e;
wire miss_st1e;
wire dirty_st1e;
wire[31:0] pc_st1e;
wire [4:0] rd_st1e;
wire [1:0] wb_st1e;
wire [`NW_M1:0] warp_num_st1e;
wire [2:0] mem_read_st1e;
wire [2:0] mem_write_st1e;
wire [`vx_clog2(NUMBER_REQUESTS)-1:0] tid_st1e;
wire fill_saw_dirty_st1e;
wire is_snp_st1e;
assign is_snp_st1e = is_snp_st1[STAGE_1_CYCLES-1];
assign pc_st1e = pc_st1[STAGE_1_CYCLES-1];
assign {rd_st1e, wb_st1e, warp_num_st1e, mem_read_st1e, mem_write_st1e, tid_st1e} = inst_meta_st1[STAGE_1_CYCLES-1];
VX_tag_data_access #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
VX_tag_data_access
(
.clk (clk),
.reset (reset),
.stall (stall_bank_pipe),
.stall_bank_pipe(stall_bank_pipe),
// Initial Read
.readaddr_st10 (addr_st1[0]),
// Actual Read/Write
.valid_req_st1e(valid_st1[STAGE_1_CYCLES-1]),
.writefill_st1e(is_fill_st1[STAGE_1_CYCLES-1]),
.writeaddr_st1e(addr_st1[STAGE_1_CYCLES-1]),
.writeword_st1e(writeword_st1[STAGE_1_CYCLES-1]),
.writedata_st1e(writedata_st1[STAGE_1_CYCLES-1]),
.mem_write_st1e(mem_write_st1e),
.mem_read_st1e (mem_read_st1e),
.is_snp_st1e (is_snp_st1e),
// Read Data
.readword_st1e (readword_st1e),
.readdata_st1e (readdata_st1e),
.readtag_st1e (readtag_st1e),
.miss_st1e (miss_st1e),
.dirty_st1e (dirty_st1e),
.fill_saw_dirty_st1e(fill_saw_dirty_st1e)
);
wire qual_valid_st1e_2 = valid_st1[STAGE_1_CYCLES-1] && !is_fill_st1[STAGE_1_CYCLES-1];
wire valid_st2;
wire[`WORD_SIZE_RNG] writeword_st2;
wire[`WORD_SIZE_RNG] readword_st2;
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] readdata_st2;
wire miss_st2;
wire dirty_st2;
wire[`REQ_INST_META_SIZE-1:0] inst_meta_st2;
wire[`TAG_SELECT_SIZE_RNG] readtag_st2;
wire fill_saw_dirty_st2;
wire is_snp_st2;
wire [31:0] pc_st2;
VX_generic_register #(.N( 1+1+1+1+32+`WORD_SIZE+`WORD_SIZE+(`BANK_LINE_SIZE_WORDS * `WORD_SIZE) + `REQ_INST_META_SIZE + `TAG_SELECT_NUM_BITS + 32 + 2)) st_1e_2 (
.clk (clk),
.reset(reset),
.stall(stall_bank_pipe),
.flush(0),
.in ({is_snp_st1e, fill_saw_dirty_st1e, is_fill_st1[STAGE_1_CYCLES-1] , qual_valid_st1e_2, addr_st1[STAGE_1_CYCLES-1], writeword_st1[STAGE_1_CYCLES-1], readword_st1e, readdata_st1e, readtag_st1e, miss_st1e, dirty_st1e, pc_st1e, inst_meta_st1[STAGE_1_CYCLES-1]}),
.out ({is_snp_st2 , fill_saw_dirty_st2 , is_fill_st2 , valid_st2 , addr_st2 , writeword_st2 , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , pc_st2 , inst_meta_st2 })
);
wire should_flush;
wire dwbq_push;
wire cwbq_full;
wire dwbq_full;
wire ffsq_full;
wire invalidate_fill;
// Enqueue to miss reserv if it's a valid miss
assign miss_add = valid_st2 && !is_snp_st2 && miss_st2 && !mrvq_full && !(should_flush && dwbq_push) && !((is_snp_st2 && valid_st2 && ffsq_full) ||((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full));
assign miss_add_pc = pc_st2;
assign miss_add_addr = addr_st2;
assign miss_add_data = writeword_st2;
assign {miss_add_rd, miss_add_wb, miss_add_warp_num, miss_add_mem_read, miss_add_mem_write, miss_add_tid} = inst_meta_st2;
// Enqueue to CWB Queue
wire cwbq_push = (valid_st2 && !miss_st2) && !cwbq_full && !((FUNC_ID == `LLFUNC_ID) && (miss_add_wb == 0)) && !((is_snp_st2 && valid_st2 && ffsq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full));
wire [`WORD_SIZE_RNG] cwbq_data = readword_st2;
wire [`vx_clog2(NUMBER_REQUESTS)-1:0] cwbq_tid = miss_add_tid;
wire [4:0] cwbq_rd = miss_add_rd;
wire [1:0] cwbq_wb = miss_add_wb;
wire [`NW_M1:0] cwbq_warp_num = miss_add_warp_num;
wire [31:0] cwbq_pc = pc_st2;
wire cwbq_empty;
assign bank_wb_valid = !cwbq_empty;
VX_generic_queue_ll #(.DATAW( `vx_clog2(NUMBER_REQUESTS) + 5 + 2 + (`NW_M1+1) + `WORD_SIZE + 32 + 32), .SIZE(CWBQ_SIZE)) cwb_queue(
.clk (clk),
.reset (reset),
.push (cwbq_push),
.in_data ({cwbq_tid, cwbq_rd, cwbq_wb, cwbq_warp_num, cwbq_data, cwbq_pc, addr_st2}),
.pop (bank_wb_pop),
.out_data({bank_wb_tid, bank_wb_rd, bank_wb_wb, bank_wb_warp_num, bank_wb_data, bank_wb_pc, bank_wb_address}),
.empty (cwbq_empty),
.full (cwbq_full)
);
assign should_flush = snoop_state && valid_st2 && (miss_add_mem_write != `NO_MEM_WRITE) && !is_snp_st2 && !is_fill_st2;
// Enqueue to DWB Queue
assign dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2 || should_flush) && !dwbq_full && !((is_snp_st2 && valid_st2 && ffsq_full) ||((valid_st2 && !miss_st2) && cwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full));
wire[31:0] dwbq_req_addr;
wire dwbq_empty;
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dwbq_req_data;
if ((FUNC_ID == `LLFUNC_ID) || (FUNC_ID == `L3FUNC_ID)) begin
assign dwbq_req_data = (should_flush && dwbq_push) ? writeword_st2 : readdata_st2;
assign dwbq_req_addr = (should_flush && dwbq_push) ? (addr_st2) : ({readtag_st2, addr_st2[`LINE_SELECT_ADDR_END:0]} & `BASE_ADDR_MASK);
end else begin
assign dwbq_req_data = readdata_st2;
assign dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_ADDR_END:0]} & `BASE_ADDR_MASK;
end
wire possible_fill = valid_st2 && miss_st2 && !dram_fill_req_queue_full && !is_snp_st2;
wire[31:0] fill_invalidator_addr = addr_st2 & `BASE_ADDR_MASK;
VX_fill_invalidator #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
VX_fill_invalidator
(
.clk (clk),
.reset (reset),
.possible_fill (possible_fill),
.success_fill (is_fill_st2),
.fill_addr (fill_invalidator_addr),
.invalidate_fill (invalidate_fill)
);
// Enqueu in dram_fill_req
assign dram_fill_req = possible_fill && !invalidate_fill;
assign dram_because_of_snp = is_snp_st2 && valid_st2 && miss_st2;
assign dram_snp_full = snrq_full && snp_req;
assign dram_fill_req_addr = addr_st2 & `BASE_ADDR_MASK;
assign dram_wb_req = !dwbq_empty;
VX_generic_queue_ll #(.DATAW( 32 + (`BANK_LINE_SIZE_WORDS * `WORD_SIZE)), .SIZE(DWBQ_SIZE)) dwb_queue(
.clk (clk),
.reset (reset),
.push (dwbq_push),
.in_data ({dwbq_req_addr, dwbq_req_data}),
.pop (dram_wb_queue_pop),
.out_data({dram_wb_req_addr, dram_wb_req_data}),
.empty (dwbq_empty),
.full (dwbq_full)
);
wire snp_fwd_push;
wire ffsq_empty;
assign snp_fwd_push = is_snp_st2 && valid_st2 && !ffsq_full && !(((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full));
assign snp_fwd = !ffsq_empty;
VX_generic_queue_ll #(.DATAW(32), .SIZE(FFSQ_SIZE)) ffs_queue(
.clk (clk),
.reset (reset),
.push (snp_fwd_push),
.in_data ({addr_st2}),
.pop (snp_fwd_pop),
.out_data({snp_fwd_addr}),
.empty (ffsq_empty),
.full (ffsq_full)
);
assign stall_bank_pipe = (is_snp_st2 && valid_st2 && ffsq_full) || ((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_queue_full);
endmodule : VX_bank

View File

@@ -0,0 +1,469 @@
`include "VX_cache_config.v"
module VX_cache
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 16,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 3,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Forward SNP Queue
parameter FFSQ_SIZE = 8,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Prefetcher
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 0,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
// Req Info
input wire [NUMBER_REQUESTS-1:0] core_req_valid,
input wire [NUMBER_REQUESTS-1:0][31:0] core_req_addr,
input wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] core_req_writedata,
input wire[NUMBER_REQUESTS-1:0][2:0] core_req_mem_read,
input wire[NUMBER_REQUESTS-1:0][2:0] core_req_mem_write,
// Req meta
input wire [4:0] core_req_rd,
input wire [NUMBER_REQUESTS-1:0][1:0] core_req_wb,
input wire [`NW_M1:0] core_req_warp_num,
input wire [31:0] core_req_pc,
output wire delay_req,
// Core Writeback
input wire core_no_wb_slot,
output wire [NUMBER_REQUESTS-1:0] core_wb_valid,
output wire [4:0] core_wb_req_rd,
output wire [1:0] core_wb_req_wb,
output wire [`NW_M1:0] core_wb_warp_num,
output wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] core_wb_readdata,
output wire [NUMBER_REQUESTS-1:0][31:0] core_wb_pc,
output wire [NUMBER_REQUESTS-1:0][31:0] core_wb_address,
// Dram Fill Response
input wire dram_fill_rsp,
input wire [31:0] dram_fill_rsp_addr,
input wire [`IBANK_LINE_SIZE_RNG][31:0] dram_fill_rsp_data,
output wire dram_fill_accept,
// Dram request
output wire dram_req,
output wire dram_req_write,
output wire dram_req_read,
output wire [31:0] dram_req_addr,
output wire [31:0] dram_req_size,
output wire [`IBANK_LINE_SIZE_RNG][31:0] dram_req_data,
output wire dram_req_because_of_wb,
input wire dram_req_delay,
output wire dram_snp_full,
// Snoop Req
input wire snp_req,
input wire[31:0] snp_req_addr,
output wire snp_req_delay,
// Snoop Forward
output wire snp_fwd,
output wire[31:0] snp_fwd_addr,
input wire snp_fwd_delay
);
wire [NUMBER_BANKS-1:0][NUMBER_REQUESTS-1:0] per_bank_valids;
wire [NUMBER_BANKS-1:0] per_bank_wb_pop;
wire [NUMBER_BANKS-1:0] per_bank_wb_valid;
wire [NUMBER_BANKS-1:0][`vx_clog2(NUMBER_REQUESTS)-1:0] per_bank_wb_tid;
wire [NUMBER_BANKS-1:0][4:0] per_bank_wb_rd;
wire [NUMBER_BANKS-1:0][1:0] per_bank_wb_wb;
wire [NUMBER_BANKS-1:0][`NW_M1:0] per_bank_wb_warp_num;
wire [NUMBER_BANKS-1:0][`WORD_SIZE_RNG] per_bank_wb_data;
wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_pc;
wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_address;
wire dfqq_full;
wire[NUMBER_BANKS-1:0] per_bank_dram_fill_req;
wire[NUMBER_BANKS-1:0][31:0] per_bank_dram_fill_req_addr;
wire[NUMBER_BANKS-1:0] per_bank_dram_fill_accept;
wire[NUMBER_BANKS-1:0] per_bank_dram_wb_queue_pop;
wire[NUMBER_BANKS-1:0] per_bank_dram_wb_req;
wire[NUMBER_BANKS-1:0] per_bank_dram_because_of_snp;
wire[NUMBER_BANKS-1:0][31:0] per_bank_dram_wb_req_addr;
wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] per_bank_dram_wb_req_data;
wire[NUMBER_BANKS-1:0] per_bank_reqq_full;
wire[NUMBER_BANKS-1:0] per_bank_snrq_full;
wire[NUMBER_BANKS-1:0] per_bank_snp_fwd;
wire[NUMBER_BANKS-1:0][31:0] per_bank_snp_fwd_addr;
wire[NUMBER_BANKS-1:0] per_bank_snp_fwd_pop;
assign delay_req = (|per_bank_reqq_full);
assign snp_req_delay = (|per_bank_snrq_full);
// assign dram_fill_accept = (NUMBER_BANKS == 1) ? per_bank_dram_fill_accept[0] : per_bank_dram_fill_accept[dram_fill_rsp_addr[`BANK_SELECT_ADDR_RNG]];
assign dram_fill_accept = (|per_bank_dram_fill_accept);
VX_cache_dram_req_arb #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.PRFQ_SIZE (PRFQ_SIZE),
.PRFQ_STRIDE (PRFQ_STRIDE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
VX_cache_dram_req_arb
(
.clk (clk),
.reset (reset),
.dfqq_full (dfqq_full),
.per_bank_dram_fill_req (per_bank_dram_fill_req),
.per_bank_dram_fill_req_addr(per_bank_dram_fill_req_addr),
.per_bank_dram_wb_queue_pop (per_bank_dram_wb_queue_pop),
.per_bank_dram_wb_req (per_bank_dram_wb_req),
.per_bank_dram_because_of_snp(per_bank_dram_because_of_snp),
.per_bank_dram_wb_req_addr (per_bank_dram_wb_req_addr),
.per_bank_dram_wb_req_data (per_bank_dram_wb_req_data),
.dram_req (dram_req),
.dram_req_write (dram_req_write),
.dram_req_read (dram_req_read),
.dram_req_addr (dram_req_addr),
.dram_req_size (dram_req_size),
.dram_req_data (dram_req_data),
.dram_req_because_of_wb (dram_req_because_of_wb),
.dram_req_delay (dram_req_delay)
);
VX_cache_core_req_bank_sel #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
VX_cache_core_req_bank_sell
(
.core_req_valid (core_req_valid),
.core_req_addr (core_req_addr),
.per_bank_valids(per_bank_valids)
);
VX_cache_wb_sel_merge #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
VX_cache_core_wb_sel_merge
(
.per_bank_wb_valid (per_bank_wb_valid),
.per_bank_wb_tid (per_bank_wb_tid),
.per_bank_wb_rd (per_bank_wb_rd),
.per_bank_wb_pc (per_bank_wb_pc),
.per_bank_wb_wb (per_bank_wb_wb),
.per_bank_wb_warp_num(per_bank_wb_warp_num),
.per_bank_wb_data (per_bank_wb_data),
.per_bank_wb_pop (per_bank_wb_pop),
.per_bank_wb_address (per_bank_wb_address),
.core_no_wb_slot (core_no_wb_slot),
.core_wb_valid (core_wb_valid),
.core_wb_req_rd (core_wb_req_rd),
.core_wb_req_wb (core_wb_req_wb),
.core_wb_warp_num (core_wb_warp_num),
.core_wb_readdata (core_wb_readdata),
.core_wb_address (core_wb_address),
.core_wb_pc (core_wb_pc)
);
// Snoop Forward Logic
VX_snp_fwd_arb #(.NUMBER_BANKS(NUMBER_BANKS)) VX_snp_fwd_arb(
.per_bank_snp_fwd (per_bank_snp_fwd),
.per_bank_snp_fwd_addr(per_bank_snp_fwd_addr),
.per_bank_snp_fwd_pop (per_bank_snp_fwd_pop),
.snp_fwd (snp_fwd),
.snp_fwd_addr (snp_fwd_addr),
.snp_fwd_delay (snp_fwd_delay)
);
// Snoop Forward Logic
genvar curr_bank;
generate
for (curr_bank = 0; curr_bank < NUMBER_BANKS; curr_bank=curr_bank+1) begin
wire [NUMBER_REQUESTS-1:0] curr_bank_valids;
wire [NUMBER_REQUESTS-1:0][31:0] curr_bank_addr;
wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] curr_bank_writedata;
wire [4:0] curr_bank_rd;
wire [NUMBER_REQUESTS-1:0][1:0] curr_bank_wb;
wire [`NW_M1:0] curr_bank_warp_num;
wire [NUMBER_REQUESTS-1:0][2:0] curr_bank_mem_read;
wire [NUMBER_REQUESTS-1:0][2:0] curr_bank_mem_write;
wire [31:0] curr_bank_pc;
wire curr_bank_wb_pop;
wire curr_bank_wb_valid;
wire [`vx_clog2(NUMBER_REQUESTS)-1:0] curr_bank_wb_tid;
wire [31:0] curr_bank_wb_pc;
wire [4:0] curr_bank_wb_rd;
wire [1:0] curr_bank_wb_wb;
wire [`NW_M1:0] curr_bank_wb_warp_num;
wire [`WORD_SIZE_RNG] curr_bank_wb_data;
wire [31:0] curr_bank_wb_address;
wire curr_bank_dram_fill_rsp;
wire [31:0] curr_bank_dram_fill_rsp_addr;
wire [`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] curr_bank_dram_fill_rsp_data;
wire curr_bank_dram_fill_accept;
wire curr_bank_dfqq_full;
wire curr_bank_dram_fill_req;
wire curr_bank_dram_because_of_snp;
wire curr_bank_dram_snp_full;
wire[31:0] curr_bank_dram_fill_req_addr;
wire curr_bank_dram_wb_queue_pop;
wire curr_bank_dram_wb_req;
wire[31:0] curr_bank_dram_wb_req_addr;
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] curr_bank_dram_wb_req_data;
wire curr_bank_snp_req;
wire[31:0] curr_bank_snp_req_addr;
wire curr_bank_reqq_full;
wire curr_bank_snp_fwd;
wire[31:0] curr_bank_snp_fwd_addr;
wire curr_bank_snp_fwd_pop;
wire curr_bank_snrq_full;
// Core Req
assign curr_bank_valids = per_bank_valids[curr_bank];
assign curr_bank_addr = core_req_addr;
assign curr_bank_writedata = core_req_writedata;
assign curr_bank_rd = core_req_rd;
assign curr_bank_wb = core_req_wb;
assign curr_bank_pc = core_req_pc;
assign curr_bank_warp_num = core_req_warp_num;
assign curr_bank_mem_read = core_req_mem_read;
assign curr_bank_mem_write = core_req_mem_write;
assign per_bank_reqq_full[curr_bank] = curr_bank_reqq_full;
// Core WB
assign curr_bank_wb_pop = per_bank_wb_pop[curr_bank];
assign per_bank_wb_valid [curr_bank] = curr_bank_wb_valid;
assign per_bank_wb_tid [curr_bank] = curr_bank_wb_tid;
assign per_bank_wb_rd [curr_bank] = curr_bank_wb_rd;
assign per_bank_wb_wb [curr_bank] = curr_bank_wb_wb;
assign per_bank_wb_warp_num[curr_bank] = curr_bank_wb_warp_num;
assign per_bank_wb_data [curr_bank] = curr_bank_wb_data;
assign per_bank_wb_pc [curr_bank] = curr_bank_wb_pc;
assign per_bank_wb_address [curr_bank] = curr_bank_wb_address;
// Dram fill request
assign curr_bank_dfqq_full = dfqq_full;
assign per_bank_dram_fill_req[curr_bank] = curr_bank_dram_fill_req;
assign per_bank_dram_fill_req_addr[curr_bank] = curr_bank_dram_fill_req_addr;
// Dram fill response
assign curr_bank_dram_fill_rsp = (NUMBER_BANKS == 1) || (dram_fill_rsp && (curr_bank_dram_fill_rsp_addr[`BANK_SELECT_ADDR_RNG] == curr_bank));
assign curr_bank_dram_fill_rsp_addr = dram_fill_rsp_addr;
assign curr_bank_dram_fill_rsp_data = dram_fill_rsp_data;
assign per_bank_dram_fill_accept[curr_bank] = curr_bank_dram_fill_accept;
// Dram writeback request
assign curr_bank_dram_wb_queue_pop = per_bank_dram_wb_queue_pop[curr_bank];
assign per_bank_dram_wb_req[curr_bank] = curr_bank_dram_wb_req;
assign per_bank_dram_because_of_snp[curr_bank] = curr_bank_dram_because_of_snp;
assign per_bank_dram_wb_req_addr[curr_bank] = curr_bank_dram_wb_req_addr;
assign per_bank_dram_wb_req_data[curr_bank] = curr_bank_dram_wb_req_data;
// Snoop Request
assign curr_bank_snp_req = snp_req && (snp_req_addr[`BANK_SELECT_ADDR_RNG] == curr_bank);
assign curr_bank_snp_req_addr = snp_req_addr;
assign per_bank_snrq_full[curr_bank] = curr_bank_snrq_full;
// Snoop Fwd
assign curr_bank_snp_fwd_pop = per_bank_snp_fwd_pop[curr_bank];
assign per_bank_snp_fwd[curr_bank] = curr_bank_snp_fwd;
assign per_bank_snp_fwd_addr[curr_bank] = curr_bank_snp_fwd_addr;
VX_bank #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FFSQ_SIZE (FFSQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
bank
(
.clk (clk),
.reset (reset),
// Core req
.delay_req (delay_req),
.bank_valids (curr_bank_valids),
.bank_addr (curr_bank_addr),
.bank_writedata (curr_bank_writedata),
.bank_rd (curr_bank_rd),
.bank_wb (curr_bank_wb),
.bank_pc (curr_bank_pc),
.bank_warp_num (curr_bank_warp_num),
.bank_mem_read (curr_bank_mem_read),
.bank_mem_write (curr_bank_mem_write),
.reqq_full (curr_bank_reqq_full),
// Output core wb
.bank_wb_pop (curr_bank_wb_pop),
.bank_wb_valid (curr_bank_wb_valid),
.bank_wb_tid (curr_bank_wb_tid),
.bank_wb_rd (curr_bank_wb_rd),
.bank_wb_wb (curr_bank_wb_wb),
.bank_wb_warp_num (curr_bank_wb_warp_num),
.bank_wb_data (curr_bank_wb_data),
.bank_wb_pc (curr_bank_wb_pc),
.bank_wb_address (curr_bank_wb_address),
// Dram fill req
.dram_fill_req (curr_bank_dram_fill_req),
.dram_fill_req_addr (curr_bank_dram_fill_req_addr),
.dram_fill_req_queue_full(curr_bank_dfqq_full),
// Dram fill rsp
.dram_fill_rsp (curr_bank_dram_fill_rsp),
.dram_fill_addr (curr_bank_dram_fill_rsp_addr),
.dram_fill_rsp_data (curr_bank_dram_fill_rsp_data),
.dram_fill_accept (curr_bank_dram_fill_accept),
// Dram writeback
.dram_wb_queue_pop (curr_bank_dram_wb_queue_pop),
.dram_wb_req (curr_bank_dram_wb_req),
.dram_wb_req_addr (curr_bank_dram_wb_req_addr),
.dram_wb_req_data (curr_bank_dram_wb_req_data),
.dram_because_of_snp (curr_bank_dram_because_of_snp),
.dram_snp_full (curr_bank_dram_snp_full),
// Snoop Request
.snp_req (curr_bank_snp_req),
.snp_req_addr (curr_bank_snp_req_addr),
.snrq_full (curr_bank_snrq_full),
// Snoop Fwd
.snp_fwd (curr_bank_snp_fwd),
.snp_fwd_addr (curr_bank_snp_fwd_addr),
.snp_fwd_pop (curr_bank_snp_fwd_pop)
);
end
endgenerate
endmodule

View File

@@ -0,0 +1,130 @@
`ifndef VX_CACHE_CONFIG
`define VX_CACHE_CONFIG
`include "../VX_define.v"
// data tid rd wb warp_num read write
`define vx_clog2(value) ((value == 1) ? 1 : $clog2(value))
`define MRVQ_METADATA_SIZE (`WORD_SIZE + `vx_clog2(NUMBER_REQUESTS) + 5 + 2 + (`NW_M1 + 1) + 3 + 3)
// 5 + 2 + 4 + 3 + 3 + 1
`define REQ_INST_META_SIZE (5 + 2 + (`NW_M1+1) + 3 + 3 + `vx_clog2(NUMBER_REQUESTS))
// `define vx_clog2_h(value, x) (value == (1 << x)) ? (x)
// `define vx_clog2(value) (value == 0 ) ? 0 : \
// (value == 1 ) ? 1 : \
// `vx_clog2_h(value, 2 ) : \
// `vx_clog2_h(value, 3 ) : \
// `vx_clog2_h(value, 4 ) : \
// `vx_clog2_h(value, 5 ) : \
// `vx_clog2_h(value, 6 ) : \
// `vx_clog2_h(value, 7 ) : \
// `vx_clog2_h(value, 8 ) : \
// `vx_clog2_h(value, 9 ) : \
// `vx_clog2_h(value, 10) : \
// `vx_clog2_h(value, 11) : \
// `vx_clog2_h(value, 12) : \
// `vx_clog2_h(value, 13) : \
// `vx_clog2_h(value, 14) : \
// `vx_clog2_h(value, 15) : \
// `vx_clog2_h(value, 16) : \
// `vx_clog2_h(value, 17) : \
// `vx_clog2_h(value, 18) : \
// `vx_clog2_h(value, 19) : \
// `vx_clog2_h(value, 20) : \
// `vx_clog2_h(value, 21) : \
// `vx_clog2_h(value, 22) : \
// `vx_clog2_h(value, 23) : \
// `vx_clog2_h(value, 24) : \
// `vx_clog2_h(value, 25) : \
// `vx_clog2_h(value, 26) : \
// `vx_clog2_h(value, 27) : \
// `vx_clog2_h(value, 28) : \
// `vx_clog2_h(value, 29) : \
// `vx_clog2_h(value, 30) : \
// `vx_clog2_h(value, 31) : \
// 0
`define WORD_SIZE (8*WORD_SIZE_BYTES)
`define WORD_SIZE_RNG (`WORD_SIZE)-1:0
// 128
`define BANK_SIZE_BYTES CACHE_SIZE_BYTES/NUMBER_BANKS
// 8
`define BANK_LINE_COUNT (`BANK_SIZE_BYTES/BANK_LINE_SIZE_BYTES)
// 4
`define BANK_LINE_SIZE_WORDS (BANK_LINE_SIZE_BYTES / WORD_SIZE_BYTES)
// 3:0
`define BANK_LINE_SIZE_RNG `BANK_LINE_SIZE_WORDS-1:0
// Offset is fixed
`define OFFSET_ADDR_NUM_BITS 2
`define OFFSET_SIZE_END 1
`define OFFSET_ADDR_START 0
`define OFFSET_ADDR_END 1
`define OFFSET_ADDR_RNG `OFFSET_ADDR_END:`OFFSET_ADDR_START
`define OFFSET_SIZE_RNG `OFFSET_SIZE_END:0
// 2
`define WORD_SELECT_NUM_BITS (`vx_clog2(`BANK_LINE_SIZE_WORDS))
// 2
`define WORD_SELECT_SIZE_END (`WORD_SELECT_NUM_BITS)
// 2
`define WORD_SELECT_ADDR_START (1+`OFFSET_ADDR_END)
// 3
`define WORD_SELECT_ADDR_END (`WORD_SELECT_SIZE_END+`OFFSET_ADDR_END)
// 3:2
`define WORD_SELECT_ADDR_RNG `WORD_SELECT_ADDR_END:`WORD_SELECT_ADDR_START
`define WORD_SELECT_SIZE_RNG `WORD_SELECT_SIZE_END-1:0
// 3
`define BANK_SELECT_NUM_BITS (`vx_clog2(NUMBER_BANKS))
// 3
`define BANK_SELECT_SIZE_END (`BANK_SELECT_NUM_BITS)
// 4
`define BANK_SELECT_ADDR_START (1+`WORD_SELECT_ADDR_END)
// 6
`define BANK_SELECT_ADDR_END (`BANK_SELECT_SIZE_END+`BANK_SELECT_ADDR_START-1)
// 6:4
`define BANK_SELECT_ADDR_RNG `BANK_SELECT_ADDR_END:`BANK_SELECT_ADDR_START
// 2:0
`define BANK_SELECT_SIZE_RNG `BANK_SELECT_SIZE_END-1:0
// 3
`define LINE_SELECT_NUM_BITS (`vx_clog2(`BANK_LINE_COUNT))
// 3
`define LINE_SELECT_SIZE_END (`LINE_SELECT_NUM_BITS)
// 7
`define LINE_SELECT_ADDR_START (1+`BANK_SELECT_ADDR_END)
// 9
`define LINE_SELECT_ADDR_END (`LINE_SELECT_SIZE_END+`LINE_SELECT_ADDR_START-1)
// 9:7
`define LINE_SELECT_ADDR_RNG `LINE_SELECT_ADDR_END:`LINE_SELECT_ADDR_START
// 2:0
`define LINE_SELECT_SIZE_RNG `LINE_SELECT_SIZE_END-1:0
// 10
`define TAG_SELECT_ADDR_START (1+`LINE_SELECT_ADDR_END)
// 31:10
`define TAG_SELECT_ADDR_RNG 31:`TAG_SELECT_ADDR_START
// 22
`define TAG_SELECT_NUM_BITS (32-`TAG_SELECT_ADDR_START)
// 22
`define TAG_SELECT_SIZE_END (`TAG_SELECT_NUM_BITS)
// 21:0
`define TAG_SELECT_SIZE_RNG `TAG_SELECT_SIZE_END-1:0
`define BASE_ADDR_MASK (~((1<<(`WORD_SELECT_ADDR_END+1))-1))
`endif

View File

@@ -0,0 +1,74 @@
`include "VX_cache_config.v"
module VX_cache_core_req_bank_sel
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire [NUMBER_REQUESTS-1:0] core_req_valid,
input wire [NUMBER_REQUESTS-1:0][31:0] core_req_addr,
output reg [NUMBER_BANKS-1:0][NUMBER_REQUESTS-1:0] per_bank_valids
);
wire[31:0] req_address;
generate
integer curr_req;
always @(*) begin
per_bank_valids = 0;
for (curr_req = 0; curr_req < NUMBER_REQUESTS; curr_req = curr_req + 1) begin
if (NUMBER_BANKS == 1) begin
// If there is only one bank, then only map requests to that bank
per_bank_valids[0][curr_req] = core_req_valid[curr_req];
end else begin
per_bank_valids[core_req_addr[curr_req][`BANK_SELECT_ADDR_RNG]][curr_req] = core_req_valid[curr_req];
end
end
end
endgenerate
endmodule

View File

@@ -0,0 +1,124 @@
`include "VX_cache_config.v"
module VX_cache_dfq_queue
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
input wire dfqq_push,
input wire[NUMBER_BANKS-1:0] per_bank_dram_fill_req,
input wire[NUMBER_BANKS-1:0][31:0] per_bank_dram_fill_req_addr,
input wire dfqq_pop,
output wire dfqq_req,
output wire[31:0] dfqq_req_addr,
output wire dfqq_empty,
output wire dfqq_full
);
wire[NUMBER_BANKS-1:0] out_per_bank_dram_fill_req;
wire[NUMBER_BANKS-1:0][31:0] out_per_bank_dram_fill_req_addr;
reg [NUMBER_BANKS-1:0] use_per_bank_dram_fill_req;
reg [NUMBER_BANKS-1:0][31:0] use_per_bank_dram_fill_req_addr;
wire[NUMBER_BANKS-1:0] qual_bank_dram_fill_req;
wire[NUMBER_BANKS-1:0][31:0] qual_bank_dram_fill_req_addr;
wire[NUMBER_BANKS-1:0] updated_bank_dram_fill_req;
wire o_empty;
wire use_empty = !(|use_per_bank_dram_fill_req);
wire out_empty = !(|out_per_bank_dram_fill_req) || o_empty;
wire push_qual = dfqq_push && !dfqq_full;
wire pop_qual = dfqq_pop && use_empty && !out_empty;
VX_generic_queue_ll #(.DATAW(NUMBER_BANKS * (1+32)), .SIZE(DFQQ_SIZE)) dfqq_queue(
.clk (clk),
.reset (reset),
.push (push_qual),
.in_data ({per_bank_dram_fill_req, per_bank_dram_fill_req_addr}),
.pop (pop_qual),
.out_data({out_per_bank_dram_fill_req, out_per_bank_dram_fill_req_addr}),
.empty (o_empty),
.full (dfqq_full)
);
assign qual_bank_dram_fill_req = use_empty ? (out_per_bank_dram_fill_req & {NUMBER_BANKS{!o_empty}}) : (use_per_bank_dram_fill_req & {NUMBER_BANKS{!use_empty}});
assign qual_bank_dram_fill_req_addr = use_empty ? out_per_bank_dram_fill_req_addr : use_per_bank_dram_fill_req_addr;
wire[`vx_clog2(NUMBER_BANKS)-1:0] qual_request_index;
wire qual_has_request;
VX_generic_priority_encoder #(.N(NUMBER_BANKS)) VX_sel_bank(
.valids(qual_bank_dram_fill_req),
.index (qual_request_index),
.found (qual_has_request)
);
assign dfqq_empty = !qual_has_request;
assign dfqq_req = qual_bank_dram_fill_req [qual_request_index];
assign dfqq_req_addr = qual_bank_dram_fill_req_addr[qual_request_index];
assign updated_bank_dram_fill_req = qual_bank_dram_fill_req & (~(1 << qual_request_index));
always @(posedge clk) begin
if (reset) begin
use_per_bank_dram_fill_req <= 0;
use_per_bank_dram_fill_req_addr <= 0;
end else begin
if (dfqq_pop && qual_has_request) begin
use_per_bank_dram_fill_req <= updated_bank_dram_fill_req;
use_per_bank_dram_fill_req_addr <= qual_bank_dram_fill_req_addr;
end
end
end
endmodule

View File

@@ -0,0 +1,151 @@
`include "VX_cache_config.v"
module VX_cache_dram_req_arb
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Prefetcher
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 2,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
// Fill Request
output wire dfqq_full,
input wire[NUMBER_BANKS-1:0] per_bank_dram_fill_req,
input wire[NUMBER_BANKS-1:0][31:0] per_bank_dram_fill_req_addr,
// DFQ Request
output wire[NUMBER_BANKS-1:0] per_bank_dram_wb_queue_pop,
input wire[NUMBER_BANKS-1:0] per_bank_dram_wb_req,
input wire[NUMBER_BANKS-1:0][31:0] per_bank_dram_wb_req_addr,
input wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] per_bank_dram_wb_req_data,
input wire[NUMBER_BANKS-1:0] per_bank_dram_because_of_snp,
// real Dram request
output wire dram_req,
output wire dram_req_write,
output wire dram_req_read,
output wire [31:0] dram_req_addr,
output wire [31:0] dram_req_size,
output wire [`IBANK_LINE_SIZE_RNG][31:0] dram_req_data,
output wire dram_req_because_of_wb,
input wire dram_req_delay
);
wire pref_pop;
wire pref_valid;
wire[31:0] pref_addr;
wire dwb_valid;
wire dfqq_req;
assign pref_pop = !dwb_valid && !dfqq_req && !dram_req_delay && pref_valid;
VX_prefetcher #(
.PRFQ_SIZE (PRFQ_SIZE),
.PRFQ_STRIDE (PRFQ_STRIDE),
.BANK_LINE_SIZE_BYTES(BANK_LINE_SIZE_BYTES),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES)
)
prfqq
(
.clk (clk),
.reset (reset),
.dram_req (dram_req && dram_req_read),
.dram_req_addr(dram_req_addr),
.pref_pop (pref_pop),
.pref_valid (pref_valid),
.pref_addr (pref_addr)
);
wire[31:0] dfqq_req_addr;
wire dfqq_empty;
wire dfqq_pop = !dwb_valid && dfqq_req && !dram_req_delay; // If no dwb, and dfqq has valids, then pop
wire dfqq_push = (|per_bank_dram_fill_req);
VX_cache_dfq_queue VX_cache_dfq_queue(
.clk (clk),
.reset (reset),
.dfqq_push (dfqq_push),
.per_bank_dram_fill_req (per_bank_dram_fill_req),
.per_bank_dram_fill_req_addr(per_bank_dram_fill_req_addr),
.dfqq_pop (dfqq_pop),
.dfqq_req (dfqq_req),
.dfqq_req_addr (dfqq_req_addr),
.dfqq_empty (dfqq_empty),
.dfqq_full (dfqq_full)
);
wire[`vx_clog2(NUMBER_BANKS)-1:0] dwb_bank;
// wire[NUMBER_BANKS-1:0] use_wb_valid = per_bank_dram_wb_req | per_bank_dram_because_of_snp;
wire[NUMBER_BANKS-1:0] use_wb_valid = per_bank_dram_wb_req;
VX_generic_priority_encoder #(.N(NUMBER_BANKS)) VX_sel_dwb(
.valids(use_wb_valid),
.index (dwb_bank),
.found (dwb_valid)
);
assign per_bank_dram_wb_queue_pop = dram_req_delay ? 0 : use_wb_valid & ((1 << dwb_bank));
assign dram_req = dwb_valid || dfqq_req || pref_pop;
assign dram_req_write = dwb_valid && dram_req;
assign dram_req_read = ((dfqq_req && !dwb_valid) || pref_pop) && dram_req;
assign dram_req_addr = (dwb_valid ? per_bank_dram_wb_req_addr[dwb_bank] : (dfqq_req ? dfqq_req_addr : pref_addr)) & `BASE_ADDR_MASK;
assign dram_req_size = BANK_LINE_SIZE_BYTES;
assign {dram_req_data} = dwb_valid ? {per_bank_dram_wb_req_data[dwb_bank] }: 0;
// assign dram_req_because_of_wb = dwb_valid ? per_bank_dram_because_of_snp[dwb_bank] : 0;
assign dram_req_because_of_wb = 0;
endmodule

View File

@@ -0,0 +1,176 @@
`include "VX_cache_config.v"
module VX_cache_miss_resrv
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
// Miss enqueue
input wire miss_add,
input wire[31:0] miss_add_addr,
input wire[`WORD_SIZE_RNG] miss_add_data,
input wire[`vx_clog2(NUMBER_REQUESTS)-1:0] miss_add_tid,
input wire[4:0] miss_add_rd,
input wire[1:0] miss_add_wb,
input wire[`NW_M1:0] miss_add_warp_num,
input wire[2:0] miss_add_mem_read,
input wire[2:0] miss_add_mem_write,
input wire[31:0] miss_add_pc,
output wire miss_resrv_full,
output wire miss_resrv_stop,
// Broadcast Fill
input wire is_fill_st1,
input wire[31:0] fill_addr_st1,
// Miss dequeue
input wire miss_resrv_pop,
output wire miss_resrv_valid_st0,
output wire[31:0] miss_resrv_addr_st0,
output wire[`WORD_SIZE_RNG] miss_resrv_data_st0,
output wire[`vx_clog2(NUMBER_REQUESTS)-1:0] miss_resrv_tid_st0,
output wire[4:0] miss_resrv_rd_st0,
output wire[1:0] miss_resrv_wb_st0,
output wire[`NW_M1:0] miss_resrv_warp_num_st0,
output wire[2:0] miss_resrv_mem_read_st0,
output wire[31:0] miss_resrv_pc_st0,
output wire[2:0] miss_resrv_mem_write_st0
);
// Size of metadata = 32 + `vx_clog2(NUMBER_REQUESTS) + 5 + 2 + (`NW_M1 + 1)
reg[`MRVQ_METADATA_SIZE-1:0] metadata_table[MRVQ_SIZE-1:0];
reg[MRVQ_SIZE-1:0][31:0] addr_table;
reg[MRVQ_SIZE-1:0][31:0] pc_table;
reg[MRVQ_SIZE-1:0] valid_table;
reg[MRVQ_SIZE-1:0] ready_table;
reg[`vx_clog2(MRVQ_SIZE)-1:0] head_ptr;
reg[`vx_clog2(MRVQ_SIZE)-1:0] tail_ptr;
reg[31:0] size;
// assign miss_resrv_full = (MRVQ_SIZE != 2) && (tail_ptr+1) == head_ptr;
assign miss_resrv_full = (MRVQ_SIZE != 2) && (size == MRVQ_SIZE );
assign miss_resrv_stop = (MRVQ_SIZE != 2) && (size > (MRVQ_SIZE-5));
wire enqueue_possible = !miss_resrv_full;
wire[`vx_clog2(MRVQ_SIZE)-1:0] enqueue_index = tail_ptr;
reg[MRVQ_SIZE-1:0] make_ready;
genvar curr_e;
generate
for (curr_e = 0; curr_e < MRVQ_SIZE; curr_e=curr_e+1) begin
assign make_ready[curr_e] = is_fill_st1 && valid_table[curr_e]
&& addr_table[curr_e][31:`LINE_SELECT_ADDR_START] == fill_addr_st1[31:`LINE_SELECT_ADDR_START];
end
endgenerate
wire dequeue_possible = valid_table[head_ptr] && ready_table[head_ptr];
wire[`vx_clog2(MRVQ_SIZE)-1:0] dequeue_index = head_ptr;
assign miss_resrv_valid_st0 = (MRVQ_SIZE != 2) && dequeue_possible;
assign miss_resrv_pc_st0 = pc_table[dequeue_index];
assign miss_resrv_addr_st0 = addr_table[dequeue_index];
assign {miss_resrv_data_st0, miss_resrv_tid_st0, miss_resrv_rd_st0, miss_resrv_wb_st0, miss_resrv_warp_num_st0, miss_resrv_mem_read_st0, miss_resrv_mem_write_st0} = metadata_table[dequeue_index];
wire mrvq_push = miss_add && enqueue_possible && (MRVQ_SIZE != 2);
wire mrvq_pop = miss_resrv_pop && dequeue_possible;
wire update_ready = (|make_ready);
integer i;
always @(posedge clk) begin
if (reset) begin
for (i = 0; i < MRVQ_SIZE; i=i+1) metadata_table[i] <= 0;
valid_table <= 0;
ready_table <= 0;
addr_table <= 0;
pc_table <= 0;
size <= 0;
head_ptr <= 0;
tail_ptr <= 0;
end else begin
if (mrvq_push) begin
valid_table[enqueue_index] <= 1;
ready_table[enqueue_index] <= 0;
pc_table[enqueue_index] <= miss_add_pc;
addr_table[enqueue_index] <= miss_add_addr;
metadata_table[enqueue_index] <= {miss_add_data, miss_add_tid, miss_add_rd, miss_add_wb, miss_add_warp_num, miss_add_mem_read, miss_add_mem_write};
tail_ptr <= tail_ptr + 1;
end
if (update_ready) begin
ready_table <= ready_table | make_ready;
end
if (mrvq_pop) begin
valid_table[dequeue_index] <= 0;
ready_table[dequeue_index] <= 0;
addr_table[dequeue_index] <= 0;
metadata_table[dequeue_index] <= 0;
pc_table[dequeue_index] <= 0;
head_ptr <= head_ptr + 1;
end
if (!(mrvq_push && mrvq_pop)) begin
if (mrvq_push) begin
size <= size + 1;
end
if (mrvq_pop) begin
size <= size - 1;
end
end
end
end
endmodule

View File

@@ -0,0 +1,208 @@
`include "VX_cache_config.v"
module VX_cache_req_queue
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
// Enqueue Data
input wire reqq_push,
input wire [NUMBER_REQUESTS-1:0] bank_valids,
input wire [NUMBER_REQUESTS-1:0][31:0] bank_addr,
input wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] bank_writedata,
input wire [4:0] bank_rd,
input wire [NUMBER_REQUESTS-1:0][1:0] bank_wb,
input wire [`NW_M1:0] bank_warp_num,
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_read,
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_write,
input wire [31:0] bank_pc,
// Dequeue Data
input wire reqq_pop,
output wire reqq_req_st0,
output wire [`vx_clog2(NUMBER_REQUESTS)-1:0] reqq_req_tid_st0,
output wire [31:0] reqq_req_addr_st0,
output wire [`WORD_SIZE_RNG] reqq_req_writedata_st0,
output wire [4:0] reqq_req_rd_st0,
output wire [1:0] reqq_req_wb_st0,
output wire [`NW_M1:0] reqq_req_warp_num_st0,
output wire [2:0] reqq_req_mem_read_st0,
output wire [2:0] reqq_req_mem_write_st0,
output wire [31:0] reqq_req_pc_st0,
// State Data
output wire reqq_empty,
output wire reqq_full
);
wire [NUMBER_REQUESTS-1:0] out_per_valids;
wire [NUMBER_REQUESTS-1:0][31:0] out_per_addr;
wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] out_per_writedata;
wire [4:0] out_per_rd;
wire [NUMBER_REQUESTS-1:0][1:0] out_per_wb;
wire [`NW_M1:0] out_per_warp_num;
wire [NUMBER_REQUESTS-1:0][2:0] out_per_mem_read;
wire [NUMBER_REQUESTS-1:0][2:0] out_per_mem_write;
wire [31:0] out_per_pc;
reg [NUMBER_REQUESTS-1:0] use_per_valids;
reg [NUMBER_REQUESTS-1:0][31:0] use_per_addr;
reg [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] use_per_writedata;
reg [4:0] use_per_rd;
reg [NUMBER_REQUESTS-1:0][1:0] use_per_wb;
reg [31:0] use_per_pc;
reg [`NW_M1:0] use_per_warp_num;
reg [NUMBER_REQUESTS-1:0][2:0] use_per_mem_read;
reg [NUMBER_REQUESTS-1:0][2:0] use_per_mem_write;
wire [NUMBER_REQUESTS-1:0] qual_valids;
wire [NUMBER_REQUESTS-1:0][31:0] qual_addr;
wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] qual_writedata;
wire [4:0] qual_rd;
wire [NUMBER_REQUESTS-1:0][1:0] qual_wb;
wire [`NW_M1:0] qual_warp_num;
wire [NUMBER_REQUESTS-1:0][2:0] qual_mem_read;
wire [NUMBER_REQUESTS-1:0][2:0] qual_mem_write;
wire [31:0] qual_pc;
reg [NUMBER_REQUESTS-1:0] updated_valids;
wire o_empty;
wire use_empty = !(|use_per_valids);
wire out_empty = !(|out_per_valids) || o_empty;
wire push_qual = reqq_push && !reqq_full;
wire pop_qual = !out_empty && use_empty;
VX_generic_queue_ll #(.DATAW( (NUMBER_REQUESTS * (1+32+`WORD_SIZE)) + 5 + (NUMBER_REQUESTS*2) + (`NW_M1+1) + (NUMBER_REQUESTS * (3 + 3)) + 32 ), .SIZE(REQQ_SIZE)) reqq_queue(
.clk (clk),
.reset (reset),
.push (push_qual),
.in_data ({bank_valids , bank_addr , bank_writedata , bank_rd , bank_wb , bank_warp_num , bank_mem_read , bank_mem_write , bank_pc}),
.pop (pop_qual),
.out_data({out_per_valids, out_per_addr, out_per_writedata, out_per_rd, out_per_wb, out_per_warp_num, out_per_mem_read, out_per_mem_write, out_per_pc}),
.empty (o_empty),
.full (reqq_full)
);
wire[NUMBER_REQUESTS-1:0] real_out_per_valids = out_per_valids & {NUMBER_REQUESTS{~out_empty}};
assign qual_valids = use_per_valids;
assign qual_addr = use_per_addr;
assign qual_writedata = use_per_writedata;
assign qual_rd = use_per_rd;
assign qual_wb = use_per_wb;
assign qual_warp_num = use_per_warp_num;
assign qual_mem_read = use_per_mem_read;
assign qual_mem_write = use_per_mem_write;
assign qual_pc = use_per_pc;
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] qual_request_index;
wire qual_has_request;
VX_generic_priority_encoder #(.N(NUMBER_REQUESTS)) VX_sel_bank(
.valids(qual_valids),
.index (qual_request_index),
.found (qual_has_request)
);
assign reqq_empty = !qual_has_request;
assign reqq_req_st0 = qual_has_request;
assign reqq_req_tid_st0 = qual_request_index;
assign reqq_req_addr_st0 = qual_addr [qual_request_index];
assign reqq_req_writedata_st0 = qual_writedata[qual_request_index];
assign reqq_req_rd_st0 = qual_rd;
assign reqq_req_wb_st0 = qual_wb[qual_request_index];
assign reqq_req_warp_num_st0 = qual_warp_num;
assign reqq_req_mem_read_st0 = qual_mem_read [qual_request_index];
assign reqq_req_mem_write_st0 = qual_mem_write[qual_request_index];
assign reqq_req_pc_st0 = qual_pc;
always @(*) begin
updated_valids = qual_valids;
if (qual_has_request) begin
updated_valids[qual_request_index] = 0;
end
end
always @(posedge clk) begin
if (reset) begin
use_per_valids <= 0;
use_per_addr <= 0;
use_per_writedata <= 0;
use_per_rd <= 0;
use_per_wb <= 0;
use_per_warp_num <= 0;
use_per_mem_read <= 0;
use_per_mem_write <= 0;
use_per_pc <= 0;
end else begin
if (pop_qual) begin
use_per_valids <= real_out_per_valids;
use_per_addr <= out_per_addr;
use_per_writedata <= out_per_writedata;
use_per_rd <= out_per_rd;
use_per_wb <= out_per_wb;
use_per_warp_num <= out_per_warp_num;
use_per_mem_read <= out_per_mem_read;
use_per_mem_write <= out_per_mem_write;
use_per_pc <= out_per_pc;
end else if (reqq_pop) begin
use_per_valids[qual_request_index] <= 0;
end
// else if (reqq_pop) begin
// use_per_valids[qual_request_index] <= updated_valids;
// end
end
end
endmodule

View File

@@ -0,0 +1,140 @@
`include "VX_cache_config.v"
module VX_cache_wb_sel_merge
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
// Per Bank WB
input wire [NUMBER_BANKS-1:0] per_bank_wb_valid,
input wire [NUMBER_BANKS-1:0][`vx_clog2(NUMBER_REQUESTS)-1:0] per_bank_wb_tid,
input wire [NUMBER_BANKS-1:0][4:0] per_bank_wb_rd,
input wire [NUMBER_BANKS-1:0][1:0] per_bank_wb_wb,
input wire [NUMBER_BANKS-1:0][`NW_M1:0] per_bank_wb_warp_num,
input wire [NUMBER_BANKS-1:0][`WORD_SIZE_RNG] per_bank_wb_data,
input wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_pc,
input wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_address,
output wire [NUMBER_BANKS-1:0] per_bank_wb_pop,
// Core Writeback
input wire core_no_wb_slot,
output reg [NUMBER_REQUESTS-1:0] core_wb_valid,
output reg [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] core_wb_readdata,
output reg [NUMBER_REQUESTS-1:0][31:0] core_wb_pc,
output wire [4:0] core_wb_req_rd,
output wire [1:0] core_wb_req_wb,
output wire [`NW_M1:0] core_wb_warp_num,
output reg [NUMBER_REQUESTS-1:0][31:0] core_wb_address
);
reg [NUMBER_BANKS-1:0] per_bank_wb_pop_unqual;
assign per_bank_wb_pop = per_bank_wb_pop_unqual & {NUMBER_BANKS{~core_no_wb_slot}};
// wire[NUMBER_BANKS-1:0] bank_wants_wb;
// genvar curr_bank;
// generate
// for (curr_bank = 0; curr_bank < NUMBER_BANKS; curr_bank=curr_bank+1) begin
// assign bank_wants_wb[curr_bank] = (|per_bank_wb_valid[curr_bank]);
// end
// endgenerate
wire [(`vx_clog2(NUMBER_BANKS))-1:0] main_bank_index;
wire found_bank;
VX_generic_priority_encoder #(.N(NUMBER_BANKS)) VX_sel_bank(
.valids(per_bank_wb_valid),
.index (main_bank_index),
.found (found_bank)
);
assign core_wb_req_rd = per_bank_wb_rd [main_bank_index];
assign core_wb_req_wb = per_bank_wb_wb [main_bank_index];
assign core_wb_warp_num = per_bank_wb_warp_num[main_bank_index];
integer this_bank;
generate
always @(*) begin
core_wb_valid = 0;
core_wb_readdata = 0;
core_wb_pc = 0;
core_wb_address = 0;
for (this_bank = 0; this_bank < NUMBER_BANKS; this_bank = this_bank + 1) begin
if ((FUNC_ID == `LLFUNC_ID) || (FUNC_ID == `L3FUNC_ID)) begin
if (found_bank && !core_wb_valid[per_bank_wb_tid[this_bank]] && per_bank_wb_valid[this_bank] && ((this_bank == main_bank_index) || (per_bank_wb_tid[this_bank] != per_bank_wb_tid[main_bank_index]))) begin
core_wb_valid[per_bank_wb_tid[this_bank]] = 1;
core_wb_readdata[per_bank_wb_tid[this_bank]] = per_bank_wb_data[this_bank];
core_wb_pc[per_bank_wb_tid[this_bank]] = per_bank_wb_pc[this_bank];
core_wb_address[per_bank_wb_tid[this_bank]] = per_bank_wb_address[this_bank];
per_bank_wb_pop_unqual[this_bank] = 1;
end else begin
per_bank_wb_pop_unqual[this_bank] = 0;
end
end else begin
if (((this_bank == main_bank_index) || (per_bank_wb_tid[this_bank] != per_bank_wb_tid[main_bank_index])) && found_bank && !core_wb_valid[per_bank_wb_tid[this_bank]] && (per_bank_wb_valid[this_bank]) && (per_bank_wb_rd[this_bank] == per_bank_wb_rd[main_bank_index]) && (per_bank_wb_warp_num[this_bank] == per_bank_wb_warp_num[main_bank_index])) begin
core_wb_valid[per_bank_wb_tid[this_bank]] = 1;
core_wb_readdata[per_bank_wb_tid[this_bank]] = per_bank_wb_data[this_bank];
core_wb_pc[per_bank_wb_tid[this_bank]] = per_bank_wb_pc[this_bank];
core_wb_address[per_bank_wb_tid[this_bank]] = per_bank_wb_address[this_bank];
per_bank_wb_pop_unqual[this_bank] = 1;
end else begin
per_bank_wb_pop_unqual[this_bank] = 0;
end
end
end
end
endgenerate
endmodule

View File

@@ -0,0 +1,85 @@
`include "VX_cache_config.v"
module VX_dcache_llv_resp_bank_sel
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
output reg [NUMBER_BANKS-1:0] per_bank_llvq_pop,
input wire[NUMBER_BANKS-1:0] per_bank_llvq_valid,
input wire[NUMBER_BANKS-1:0][31:0] per_bank_llvq_res_addr,
input wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][31:0] per_bank_llvq_res_data,
input wire[NUMBER_BANKS-1:0][`vx_clog2(NUMBER_REQUESTS)-1:0] per_bank_llvq_res_tid,
input wire llvq_pop,
output reg[NUMBER_REQUESTS-1:0] llvq_valid,
output reg[NUMBER_REQUESTS-1:0][31:0] llvq_res_addr,
output reg[NUMBER_REQUESTS-1:0][`BANK_LINE_SIZE_RNG][31:0] llvq_res_data
);
wire [(`vx_clog2(NUMBER_BANKS))-1:0] main_bank_index;
wire found_bank;
VX_generic_priority_encoder #(.N(NUMBER_BANKS)) VX_sel_bank(
.valids(per_bank_llvq_valid),
.index (main_bank_index),
.found (found_bank)
);
always @(*) begin
llvq_valid = 0;
llvq_res_addr = 0;
llvq_res_data = 0;
per_bank_llvq_pop = 0;
if (found_bank && llvq_pop) begin
llvq_valid [per_bank_llvq_res_tid[main_bank_index]] = 1'b1;
llvq_res_addr[per_bank_llvq_res_tid[main_bank_index]] = per_bank_llvq_res_addr[main_bank_index];
llvq_res_data[per_bank_llvq_res_tid[main_bank_index]] = per_bank_llvq_res_data[main_bank_index];
per_bank_llvq_pop[main_bank_index] = 1'b1;
end
end
endmodule

View File

@@ -0,0 +1,172 @@
`include "VX_cache_config.v"
module VX_fill_invalidator
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
input wire possible_fill,
input wire success_fill,
input wire[31:0] fill_addr,
output reg invalidate_fill
);
if (FILL_INVALIDAOR_SIZE == 0) begin
assign invalidate_fill = 0;
end else begin
reg[FILL_INVALIDAOR_SIZE-1:0] fills_active;
reg[FILL_INVALIDAOR_SIZE-1:0][31:0] fills_address;
reg[FILL_INVALIDAOR_SIZE-1:0] matched_fill;
wire matched;
integer fi;
always @(*) begin
for (fi = 0; fi < FILL_INVALIDAOR_SIZE; fi+=1) begin
matched_fill[fi] = fills_active[fi] && (fills_address[fi][31:`LINE_SELECT_ADDR_START] == fill_addr[31:`LINE_SELECT_ADDR_START]);
end
end
assign matched = (|(matched_fill));
wire [(`vx_clog2(FILL_INVALIDAOR_SIZE))-1:0] enqueue_index;
wire enqueue_found;
VX_generic_priority_encoder #(.N(FILL_INVALIDAOR_SIZE)) VX_sel_bank(
.valids(~fills_active),
.index (enqueue_index),
.found (enqueue_found)
);
assign invalidate_fill = possible_fill && matched;
always @(posedge clk) begin
if (reset) begin
fills_active <= 0;
fills_address <= 0;
end else begin
if (possible_fill && !matched && enqueue_found) begin
fills_active [enqueue_index] <= 1;
fills_address[enqueue_index] <= fill_addr;
end else if (success_fill && matched) begin
fills_active <= fills_active & (~matched_fill);
end
end
end
// reg success_found;
// reg[(`vx_clog2(FILL_INVALIDAOR_SIZE))-1:0] success_index;
// integer curr_fill;
// always @(*) begin
// invalidate_fill = 0;
// success_found = 0;
// success_index = 0;
// for (curr_fill = 0; curr_fill < FILL_INVALIDAOR_SIZE; curr_fill=curr_fill+1) begin
// if (fill_addr[31:`LINE_SELECT_ADDR_START] == fills_address[curr_fill][31:`LINE_SELECT_ADDR_START]) begin
// if (possible_fill && fills_active[curr_fill]) begin
// invalidate_fill = 1;
// end
// if (success_fill) begin
// success_found = 1;
// success_index = curr_fill;
// end
// end
// end
// end
// wire [(`vx_clog2(FILL_INVALIDAOR_SIZE))-1:0] enqueue_index;
// wire enqueue_found;
// VX_generic_priority_encoder #(.N(FILL_INVALIDAOR_SIZE)) VX_sel_bank(
// .valids(~fills_active),
// .index (enqueue_index),
// .found (enqueue_found)
// );
// always @(posedge clk) begin
// if (reset) begin
// fills_active <= 0;
// fills_address <= 0;
// end else begin
// if (possible_fill && !invalidate_fill) begin
// fills_active[enqueue_index] <= 1;
// fills_address[enqueue_index] <= fill_addr;
// end
// if (success_found) begin
// fills_active[success_index] <= 0;
// end
// end
// end
end
endmodule

View File

@@ -0,0 +1,122 @@
module VX_mrv_queue
#(
parameter DATAW = 4,
parameter SIZE = 277
)
(
input wire clk,
input wire reset,
input wire push,
input wire[DATAW-1:0] in_data,
input wire pop,
output wire[DATAW-1:0] out_data,
output wire empty,
output wire full
);
if (SIZE == 0) begin
assign empty = 1;
assign out_data = 0;
assign full = 0;
end else begin
reg[DATAW-1:0] data[SIZE-1:0], curr_r, head_r;
reg[$clog2(SIZE+1)-1:0] size_r;
reg[$clog2(SIZE)-1:0] wr_ctr_r;
reg[$clog2(SIZE)-1:0] rd_ptr_r, rd_next_ptr_r;
reg empty_r, full_r, bypass_r;
wire reading, writing;
assign reading = pop && !empty;
assign writing = push && !full;
if (SIZE == 1) begin
always @(posedge clk) begin
if (reset) begin
size_r <= 0;
end else begin
if (writing && !reading) begin
size_r <= 1;
end else if (reading && !writing) begin
size_r <= 0;
end
if (writing) begin
head_r <= in_data;
end
end
end
assign out_data = head_r;
assign empty = (size_r == 0);
assign full = (size_r != 0) && !pop;
end else begin
always @(posedge clk) begin
if (reset) begin
wr_ctr_r <= 0;
end else begin
if (writing)
wr_ctr_r <= wr_ctr_r + 1;
end
end
always @(posedge clk) begin
if (reset) begin
size_r <= 0;
empty_r <= 1;
full_r <= 0;
end else begin
if (writing && !reading) begin
size_r <= size_r + 1;
empty_r <= 0;
if (size_r == SIZE-1)
full_r <= 1;
end else if (reading && !writing) begin
size_r <= size_r - 1;
if (size_r == 1)
empty_r <= 1;
full_r <= 0;
end
end
end
always @(posedge clk) begin
if (writing) begin
data[wr_ctr_r] <= in_data;
end
end
always @(posedge clk) begin
if (reset) begin
rd_ptr_r <= 0;
rd_next_ptr_r <= 1;
bypass_r <= 0;
end else begin
if (reading) begin
if (SIZE == 2) begin
rd_ptr_r <= rd_next_ptr_r;
rd_next_ptr_r <= ~rd_next_ptr_r;
end else if (SIZE > 2) begin
rd_ptr_r <= rd_next_ptr_r;
rd_next_ptr_r <= rd_ptr_r + 2;
end
end
bypass_r <= writing && (empty_r || (1 == size_r) && reading);
curr_r <= in_data;
head_r <= data[reading ? rd_next_ptr_r : rd_ptr_r];
end
end
assign out_data = bypass_r ? curr_r : head_r;
assign empty = empty_r;
assign full = full_r;
end
end
endmodule

View File

@@ -0,0 +1,77 @@
`include "VX_cache_config.v"
module VX_prefetcher
#(
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 2,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4
)
(
input wire clk,
input wire reset,
input wire dram_req,
input wire[31:0] dram_req_addr,
input wire pref_pop,
output wire pref_valid,
output wire[31:0] pref_addr
);
reg[`vx_clog2(PRFQ_STRIDE):0] use_valid;
reg[31:0] use_addr;
wire current_valid;
wire[31:0] current_addr;
wire current_full;
wire current_empty;
assign current_valid = ~current_empty;
wire update_use = ((use_valid == 0) || ((use_valid-1) == 0)) && current_valid;
VX_generic_queue_ll #(.DATAW(32), .SIZE(PRFQ_SIZE)) pfq_queue(
.clk (clk),
.reset (reset),
.push (dram_req && !current_full && !pref_pop),
.in_data (dram_req_addr & `BASE_ADDR_MASK),
.pop (update_use),
.out_data(current_addr),
.empty (current_empty),
.full (current_full)
);
assign pref_valid = use_valid != 0;
assign pref_addr = use_addr;
always @(posedge clk) begin
if (reset) begin
use_valid <= 0;
use_addr <= 0;
end else begin
if (update_use) begin
use_valid <= PRFQ_STRIDE;
use_addr <= current_addr + BANK_LINE_SIZE_BYTES;
end else if (pref_valid && pref_pop) begin
use_valid <= use_valid - 1;
use_addr <= use_addr + BANK_LINE_SIZE_BYTES;
end
end
end
endmodule

View File

@@ -0,0 +1,40 @@
`include "VX_cache_config.v"
module VX_snp_fwd_arb
#(
parameter NUMBER_BANKS = 8
)
(
input wire[NUMBER_BANKS-1:0] per_bank_snp_fwd,
input wire[NUMBER_BANKS-1:0][31:0] per_bank_snp_fwd_addr,
output reg[NUMBER_BANKS-1:0] per_bank_snp_fwd_pop,
output wire snp_fwd,
output wire[31:0] snp_fwd_addr,
input wire snp_fwd_delay
);
wire[NUMBER_BANKS-1:0] qual_per_bank_snp_fwd = per_bank_snp_fwd & {NUMBER_BANKS{!snp_fwd_delay}};
wire[`vx_clog2(NUMBER_BANKS)-1:0] fsq_bank;
wire fsq_valid;
VX_generic_priority_encoder #(.N(NUMBER_BANKS)) VX_sel_ffsq(
.valids(qual_per_bank_snp_fwd),
.index (fsq_bank),
.found (fsq_valid)
);
assign snp_fwd = fsq_valid;
assign snp_fwd_addr = per_bank_snp_fwd_addr[fsq_bank];
always @(*) begin
per_bank_snp_fwd_pop = 0;
if (fsq_valid) begin
per_bank_snp_fwd_pop[fsq_bank] = 1;
end
end
endmodule

View File

@@ -0,0 +1,295 @@
`include "VX_cache_config.v"
module VX_tag_data_access
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
input wire stall,
input wire is_snp_st1e,
input wire stall_bank_pipe,
// Initial Reading
input wire[31:0] readaddr_st10,
// Write/Read Logic
input wire valid_req_st1e,
input wire writefill_st1e,
input wire[31:0] writeaddr_st1e,
input wire[`WORD_SIZE_RNG] writeword_st1e,
input wire[`DBANK_LINE_SIZE_RNG][31:0] writedata_st1e,
input wire[2:0] mem_write_st1e,
input wire[2:0] mem_read_st1e,
output wire[`WORD_SIZE_RNG] readword_st1e,
output wire[`DBANK_LINE_SIZE_RNG][31:0] readdata_st1e,
output wire[`TAG_SELECT_SIZE_RNG] readtag_st1e,
output wire miss_st1e,
output wire dirty_st1e,
output wire fill_saw_dirty_st1e
);
reg[`DBANK_LINE_SIZE_RNG][31:0] readdata_st[STAGE_1_CYCLES-1:0];
reg read_valid_st1c[STAGE_1_CYCLES-1:0];
reg read_dirty_st1c[STAGE_1_CYCLES-1:0];
reg[`TAG_SELECT_SIZE_RNG] read_tag_st1c [STAGE_1_CYCLES-1:0];
reg[`DBANK_LINE_SIZE_RNG][31:0] read_data_st1c [STAGE_1_CYCLES-1:0];
wire qual_read_valid_st1;
wire qual_read_dirty_st1;
wire[`TAG_SELECT_SIZE_RNG] qual_read_tag_st1;
wire[`DBANK_LINE_SIZE_RNG][31:0] qual_read_data_st1;
wire use_read_valid_st1e;
wire use_read_dirty_st1e;
wire[`TAG_SELECT_SIZE_RNG] use_read_tag_st1e;
wire[`DBANK_LINE_SIZE_RNG][31:0] use_read_data_st1e;
wire[`DBANK_LINE_SIZE_RNG][3:0] use_write_enable;
wire[`DBANK_LINE_SIZE_RNG][31:0] use_write_data;
wire sw, sb, sh;
wire real_writefill = writefill_st1e && ((valid_req_st1e && !use_read_valid_st1e) || (valid_req_st1e && use_read_valid_st1e && (writeaddr_st1e[`TAG_SELECT_ADDR_RNG] != use_read_tag_st1e)));
wire fill_sent;
wire invalidate_line;
VX_tag_data_structure #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.NUMBER_BANKS (NUMBER_BANKS),
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
.NUMBER_REQUESTS (NUMBER_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
.SNRQ_SIZE (SNRQ_SIZE),
.CWBQ_SIZE (CWBQ_SIZE),
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
)
VX_tag_data_structure
(
.clk (clk),
.reset (reset),
.stall_bank_pipe(stall_bank_pipe),
.read_addr (readaddr_st10),
.read_valid (qual_read_valid_st1),
.read_dirty (qual_read_dirty_st1),
.read_tag (qual_read_tag_st1),
.read_data (qual_read_data_st1),
.invalidate (invalidate_line),
.write_enable(use_write_enable),
.write_fill (real_writefill),
.write_addr (writeaddr_st1e),
.write_data (use_write_data),
.fill_sent (fill_sent)
);
// VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_SIZE_WORDS*32) )) s0_1_c0 (
VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_SIZE_WORDS*32) ), .Valid(0)) s0_1_c0 (
.clk (clk),
.reset(reset),
.stall(stall),
.flush(0),
.in ({qual_read_valid_st1, qual_read_dirty_st1, qual_read_tag_st1, qual_read_data_st1}),
.out ({read_valid_st1c[0] , read_dirty_st1c[0] , read_tag_st1c[0] , read_data_st1c[0]})
);
genvar curr_stage;
generate
for (curr_stage = 1; curr_stage < STAGE_1_CYCLES-1; curr_stage = curr_stage + 1) begin
VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_SIZE_WORDS*32) )) s0_1_cc (
.clk (clk),
.reset(reset),
.stall(stall),
.flush(0),
.in ({read_valid_st1c[curr_stage-1] , read_dirty_st1c[curr_stage-1] , read_tag_st1c[curr_stage-1] , read_data_st1c[curr_stage-1]}),
.out ({read_valid_st1c[curr_stage] , read_dirty_st1c[curr_stage] , read_tag_st1c[curr_stage] , read_data_st1c[curr_stage] })
);
end
endgenerate
assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || (FUNC_ID == `SFUNC_ID); // If shared memory, always valid
assign use_read_dirty_st1e = read_dirty_st1c[STAGE_1_CYCLES-1] && (FUNC_ID != `SFUNC_ID); // Dirty only applies in Dcache
assign use_read_tag_st1e = (FUNC_ID == `SFUNC_ID) ? writeaddr_st1e[`TAG_SELECT_ADDR_RNG] : read_tag_st1c [STAGE_1_CYCLES-1]; // Tag is always the same in SM
genvar curr_w;
for (curr_w = 0; curr_w < `DBANK_LINE_SIZE_WORDS; curr_w = curr_w+1) assign use_read_data_st1e[curr_w][31:0] = read_data_st1c[STAGE_1_CYCLES-1][curr_w][31:0];
// assign use_read_data_st1e = read_data_st1c [STAGE_1_CYCLES-1];
/////////////////////// LOAD LOGIC ///////////////////
wire[`OFFSET_SIZE_RNG] byte_select = writeaddr_st1e[`OFFSET_ADDR_RNG];
wire[`WORD_SELECT_SIZE_RNG] block_offset = writeaddr_st1e[`WORD_SELECT_ADDR_RNG];
wire lw = valid_req_st1e && (mem_read_st1e == `LW_MEM_READ);
wire lb = valid_req_st1e && (mem_read_st1e == `LB_MEM_READ);
wire lh = valid_req_st1e && (mem_read_st1e == `LH_MEM_READ);
wire lhu = valid_req_st1e && (mem_read_st1e == `LHU_MEM_READ);
wire lbu = valid_req_st1e && (mem_read_st1e == `LBU_MEM_READ);
wire b0 = (byte_select == 0);
wire b1 = (byte_select == 1);
wire b2 = (byte_select == 2);
wire b3 = (byte_select == 3);
wire[31:0] w0 = read_data_st1c[STAGE_1_CYCLES-1][0][31:0];
wire[31:0] w1 = read_data_st1c[STAGE_1_CYCLES-1][1][31:0];
wire[31:0] w2 = read_data_st1c[STAGE_1_CYCLES-1][2][31:0];
wire[31:0] w3 = read_data_st1c[STAGE_1_CYCLES-1][3][31:0];
wire[31:0] data_unmod = read_data_st1c[STAGE_1_CYCLES-1][block_offset][31:0];
wire[31:0] data_unQual = (b0 || lw) ? (data_unmod) :
b1 ? (data_unmod >> 8) :
b2 ? (data_unmod >> 16) :
(data_unmod >> 24);
wire[31:0] lb_data = (data_unQual[7] ) ? (data_unQual | 32'hFFFFFF00) : (data_unQual & 32'hFF);
wire[31:0] lh_data = (data_unQual[15]) ? (data_unQual | 32'hFFFF0000) : (data_unQual & 32'hFFFF);
wire[31:0] lbu_data = (data_unQual & 32'hFF);
wire[31:0] lhu_data = (data_unQual & 32'hFFFF);
wire[31:0] lw_data = (data_unQual);
wire[31:0] sw_data = writeword_st1e[31:0];
wire[31:0] sb_data = b1 ? {{16{1'b0}}, writeword_st1e[7:0], { 8{1'b0}}} :
b2 ? {{ 8{1'b0}}, writeword_st1e[7:0], {16{1'b0}}} :
b3 ? {{ 0{1'b0}}, writeword_st1e[7:0], {24{1'b0}}} :
writeword_st1e[31:0];
wire[31:0] sh_data = b2 ? {writeword_st1e[15:0], {16{1'b0}}} : writeword_st1e[31:0];
wire[31:0] use_write_dat = sb ? sb_data :
sh ? sh_data :
sw_data;
wire[31:0] data_Qual = lb ? lb_data :
lh ? lh_data :
lhu ? lhu_data :
lbu ? lbu_data :
lw_data;
/////////////////////// STORE LOGIC ///////////////////
assign sw = valid_req_st1e && (mem_write_st1e == `SW_MEM_WRITE);
assign sb = valid_req_st1e && (mem_write_st1e == `SB_MEM_WRITE);
assign sh = valid_req_st1e && (mem_write_st1e == `SH_MEM_WRITE);
wire[3:0] sb_mask = (b0 ? 4'b0001 : (b1 ? 4'b0010 : (b2 ? 4'b0100 : 4'b1000)));
wire[3:0] sh_mask = (b0 ? 4'b0011 : 4'b1100);
wire should_write = (sw || sb || sh) && valid_req_st1e && use_read_valid_st1e && !miss_st1e && !is_snp_st1e;
wire force_write = real_writefill;
wire[`DBANK_LINE_SIZE_RNG][3:0] we;
wire[`DBANK_LINE_SIZE_RNG][31:0] data_write;
genvar g;
generate
for (g = 0; g < `DBANK_LINE_SIZE_WORDS; g = g + 1) begin : write_enables
wire normal_write = (block_offset == g[`WORD_SELECT_SIZE_RNG]) && should_write && !real_writefill;
assign we[g] = (force_write) ? 4'b1111 :
(should_write && !real_writefill && (FUNC_ID == `LLFUNC_ID)) ? 4'b1111 :
(normal_write && sw) ? 4'b1111 :
(normal_write && sb) ? sb_mask :
(normal_write && sh) ? sh_mask :
4'b0000;
if (!(FUNC_ID == `LLFUNC_ID)) assign data_write[g] = force_write ? writedata_st1e[g] : use_write_dat;
end
if ((FUNC_ID == `LLFUNC_ID)) begin
assign data_write = force_write ? writedata_st1e : writeword_st1e;
end
endgenerate
assign use_write_enable = (writefill_st1e && !real_writefill) ? 0 : we;
assign use_write_data = data_write;
///////////////////////
if (FUNC_ID == `LLFUNC_ID) begin
assign readword_st1e = read_data_st1c[STAGE_1_CYCLES-1];
end else begin
assign readword_st1e = data_Qual;
end
wire[`TAG_SELECT_ADDR_RNG] writeaddr_tag = writeaddr_st1e[`TAG_SELECT_ADDR_RNG];
wire tags_mismatch = writeaddr_tag != use_read_tag_st1e;
wire tags_match = writeaddr_tag == use_read_tag_st1e;
wire snoop_hit = valid_req_st1e && is_snp_st1e && use_read_valid_st1e && tags_match && use_read_dirty_st1e;
wire req_invalid = valid_req_st1e && !is_snp_st1e && !use_read_valid_st1e && !writefill_st1e;
wire req_miss = valid_req_st1e && !is_snp_st1e && use_read_valid_st1e && !writefill_st1e && tags_mismatch;
assign miss_st1e = snoop_hit || req_invalid || req_miss;
assign dirty_st1e = valid_req_st1e && use_read_valid_st1e && use_read_dirty_st1e;
assign readdata_st1e = use_read_data_st1e;
assign readtag_st1e = use_read_tag_st1e;
assign fill_sent = miss_st1e;
assign fill_saw_dirty_st1e = real_writefill && dirty_st1e;
assign invalidate_line = snoop_hit;
endmodule

View File

@@ -0,0 +1,122 @@
`include "VX_cache_config.v"
module VX_tag_data_structure
#(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUMBER_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUMBER_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
)
(
input wire clk,
input wire reset,
input wire stall_bank_pipe,
input wire[31:0] read_addr,
output wire read_valid,
output wire read_dirty,
output wire[`TAG_SELECT_SIZE_RNG] read_tag,
output wire[`DBANK_LINE_SIZE_RNG][31:0] read_data,
input wire invalidate,
input wire[`DBANK_LINE_SIZE_RNG][3:0] write_enable,
input wire write_fill,
input wire[31:0] write_addr,
input wire[`DBANK_LINE_SIZE_RNG][31:0] write_data,
input wire fill_sent
);
reg[`DBANK_LINE_SIZE_RNG][3:0][7:0] data [`BANK_LINE_COUNT-1:0];
reg[`TAG_SELECT_SIZE_RNG] tag [`BANK_LINE_COUNT-1:0];
reg valid[`BANK_LINE_COUNT-1:0];
reg dirty[`BANK_LINE_COUNT-1:0];
wire[`TAG_SELECT_ADDR_RNG] curr_tag = write_addr[`TAG_SELECT_ADDR_RNG];
wire[`LINE_SELECT_ADDR_RNG] curr_inx = write_addr[`LINE_SELECT_ADDR_RNG];
assign read_valid = valid[read_addr[`LINE_SELECT_ADDR_RNG]];
assign read_dirty = dirty[read_addr[`LINE_SELECT_ADDR_RNG]];
assign read_tag = tag [read_addr[`LINE_SELECT_ADDR_RNG]];
assign read_data = data [read_addr[`LINE_SELECT_ADDR_RNG]];
wire going_to_write = (|write_enable);
integer f;
integer l;
always @(posedge clk) begin
if (reset) begin
for (l = 0; l < `BANK_LINE_COUNT; l=l+1) begin
valid[l] <= 0;
// tag [l] <= 0;
dirty[l] <= 0;
// data [l] <= 0;
end
end else if (!stall_bank_pipe) begin
if (going_to_write) begin
valid[write_addr[`LINE_SELECT_ADDR_RNG]] <= 1;
tag [write_addr[`LINE_SELECT_ADDR_RNG]] <= write_addr[`TAG_SELECT_ADDR_RNG];
if (write_fill) begin
dirty[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
end else begin
dirty[write_addr[`LINE_SELECT_ADDR_RNG]] <= 1;
end
end else if (fill_sent) begin
dirty[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
// valid[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
end
if (invalidate) begin
valid[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
end
for (f = 0; f < `DBANK_LINE_SIZE_WORDS; f = f + 1) begin
if (write_enable[f][0]) data[write_addr[`LINE_SELECT_ADDR_RNG]][f][0] <= write_data[f][7 :0 ];
if (write_enable[f][1]) data[write_addr[`LINE_SELECT_ADDR_RNG]][f][1] <= write_data[f][15:8 ];
if (write_enable[f][2]) data[write_addr[`LINE_SELECT_ADDR_RNG]][f][2] <= write_data[f][23:16];
if (write_enable[f][3]) data[write_addr[`LINE_SELECT_ADDR_RNG]][f][3] <= write_data[f][31:24];
end
end
end
endmodule