RTL code refactoring
This commit is contained in:
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
`include "VX_define.vh"
|
||||
module VX_bank
|
||||
#(
|
||||
module VX_bank #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -17,8 +16,7 @@ module VX_bank
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 0,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
// Miss Reserv Queue Knob
|
||||
@@ -28,7 +26,7 @@ module VX_bank
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -43,12 +41,9 @@ module VX_bank
|
||||
// Fill Invalidator Size {Fill invalidator must be active}
|
||||
parameter FILL_INVALIDAOR_SIZE = 16,
|
||||
|
||||
// Dram knobs
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
|
||||
@@ -77,30 +72,29 @@ module VX_bank
|
||||
output wire [31:0] bank_wb_address,
|
||||
|
||||
// Dram Fill Requests
|
||||
output wire dram_fill_req,
|
||||
output wire dram_fill_req_valid,
|
||||
output wire[31:0] dram_fill_req_addr,
|
||||
output wire dram_because_of_snp,
|
||||
output wire dram_snp_full,
|
||||
output wire dram_fill_req_is_snp,
|
||||
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_WORDS-1:0][`WORD_SIZE-1:0] dram_fill_rsp_data,
|
||||
output wire dram_fill_accept,
|
||||
input wire dram_rsp_valid,
|
||||
input wire [31:0] dram_rsp_addr,
|
||||
input wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] dram_rsp_data,
|
||||
output wire dram_rsp_ready,
|
||||
|
||||
// Dram WB Requests
|
||||
input wire dram_wb_queue_pop,
|
||||
output wire dram_wb_req,
|
||||
output wire dram_wb_req_valid,
|
||||
output wire[31:0] dram_wb_req_addr,
|
||||
output wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] dram_wb_req_data,
|
||||
|
||||
// Snp Request
|
||||
input wire snp_req,
|
||||
input wire snp_req_valid,
|
||||
input wire[31:0] snp_req_addr,
|
||||
output wire snrq_full,
|
||||
output wire snp_req_full,
|
||||
|
||||
output wire snp_fwd,
|
||||
output wire snp_fwd_valid,
|
||||
output wire[31:0] snp_fwd_addr,
|
||||
input wire snp_fwd_pop
|
||||
);
|
||||
@@ -111,7 +105,7 @@ module VX_bank
|
||||
if (reset) begin
|
||||
snoop_state <= 0;
|
||||
end else begin
|
||||
snoop_state <= (snoop_state | snp_req) && ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID));
|
||||
snoop_state <= (snoop_state | snp_req_valid) && ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID));
|
||||
end
|
||||
end
|
||||
|
||||
@@ -123,16 +117,20 @@ module VX_bank
|
||||
wire[31:0] snrq_addr_st0;
|
||||
|
||||
assign snrq_valid_st0 = !snrq_empty;
|
||||
VX_generic_queue_ll #(.DATAW(32), .SIZE(SNRQ_SIZE)) snr_queue(
|
||||
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW(32),
|
||||
.SIZE(SNRQ_SIZE)
|
||||
) snr_queue (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.push (snp_req),
|
||||
.push (snp_req_valid),
|
||||
.in_data (snp_req_addr),
|
||||
.pop (snrq_pop),
|
||||
.out_data(snrq_addr_st0),
|
||||
.empty (snrq_empty),
|
||||
.full (snrq_full)
|
||||
);
|
||||
.full (snp_req_full)
|
||||
);
|
||||
|
||||
wire dfpq_pop;
|
||||
wire dfpq_empty;
|
||||
@@ -140,13 +138,16 @@ module VX_bank
|
||||
wire[31:0] dfpq_addr_st0;
|
||||
wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] dfpq_filldata_st0;
|
||||
|
||||
assign dram_fill_accept = !dfpq_full;
|
||||
assign dram_rsp_ready = !dfpq_full;
|
||||
|
||||
VX_generic_queue_ll #(.DATAW(32+(`BANK_LINE_WORDS*`WORD_SIZE)), .SIZE(DFPQ_SIZE)) dfp_queue(
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW(32+(`BANK_LINE_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}),
|
||||
.push (dram_rsp_valid),
|
||||
.in_data ({dram_rsp_addr, dram_rsp_data}),
|
||||
.pop (dfpq_pop),
|
||||
.out_data({dfpq_addr_st0, dfpq_filldata_st0}),
|
||||
.empty (dfpq_empty),
|
||||
@@ -186,9 +187,7 @@ module VX_bank
|
||||
.LLVQ_SIZE (LLVQ_SIZE),
|
||||
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
|
||||
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
|
||||
)
|
||||
req_queue
|
||||
(
|
||||
) req_queue (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
// Enqueue
|
||||
@@ -217,7 +216,7 @@ module VX_bank
|
||||
.reqq_req_pc_st0 (reqq_req_pc_st0),
|
||||
.reqq_empty (reqq_empty),
|
||||
.reqq_full (reqq_full)
|
||||
);
|
||||
);
|
||||
|
||||
wire mrvq_pop;
|
||||
wire mrvq_full;
|
||||
@@ -265,9 +264,7 @@ module VX_bank
|
||||
.LLVQ_SIZE (LLVQ_SIZE),
|
||||
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
|
||||
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
|
||||
)
|
||||
mrvq_queue
|
||||
(
|
||||
) mrvq_queue (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
// Enqueue
|
||||
@@ -300,35 +297,39 @@ module VX_bank
|
||||
.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];
|
||||
/* verilator lint_off UNUSED */
|
||||
wire going_to_write_st1[STAGE_1_CYCLES-1:0];
|
||||
/* verilator lint_on UNUSED */
|
||||
wire [31:0] addr_st1 [STAGE_1_CYCLES-1:0];
|
||||
|
||||
integer 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;
|
||||
if (is_fill_st1[p_stage]) begin
|
||||
is_fill_in_pipe = 1;
|
||||
end
|
||||
end
|
||||
|
||||
if (is_fill_st2) is_fill_in_pipe = 1;
|
||||
if (is_fill_st2) begin
|
||||
is_fill_in_pipe = 1;
|
||||
end
|
||||
end
|
||||
|
||||
// assign is_fill_in_pipe = (|is_fill_st1) || is_fill_st2;
|
||||
// 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;
|
||||
@@ -384,13 +385,15 @@ module VX_bank
|
||||
reqq_pop ? reqq_req_writeword_st0 :
|
||||
0;
|
||||
|
||||
VX_generic_register #(.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_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]})
|
||||
VX_generic_register #(
|
||||
.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_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;
|
||||
@@ -414,14 +417,14 @@ module VX_bank
|
||||
wire miss_st1e;
|
||||
wire dirty_st1e;
|
||||
wire[31:0] pc_st1e;
|
||||
|
||||
|
||||
/* verilator lint_off UNUSED */
|
||||
wire [4:0] rd_st1e;
|
||||
wire [1:0] wb_st1e;
|
||||
wire [`NW_BITS-1:0] warp_num_st1e;
|
||||
wire [2:0] mem_read_st1e;
|
||||
wire [2:0] mem_write_st1e;
|
||||
wire [`LOG2UP(NUM_REQUESTS)-1:0] tid_st1e;
|
||||
/* verilator lint_on UNUSED */
|
||||
wire [2:0] mem_read_st1e;
|
||||
wire [2:0] mem_write_st1e;
|
||||
wire fill_saw_dirty_st1e;
|
||||
wire is_snp_st1e;
|
||||
|
||||
@@ -429,7 +432,6 @@ module VX_bank
|
||||
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),
|
||||
@@ -448,9 +450,7 @@ module VX_bank
|
||||
.LLVQ_SIZE (LLVQ_SIZE),
|
||||
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
|
||||
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
|
||||
)
|
||||
VX_tag_data_access
|
||||
(
|
||||
) vx_tag_data_access (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.stall (stall_bank_pipe),
|
||||
@@ -494,16 +494,16 @@ module VX_bank
|
||||
wire is_snp_st2;
|
||||
wire [31:0] pc_st2;
|
||||
|
||||
|
||||
VX_generic_register #(.N( 1+1+1+1+32+`WORD_SIZE+`WORD_SIZE+(`BANK_LINE_WORDS * `WORD_SIZE) + `REQ_INST_META_SIZE + `TAG_SELECT_NUM_BITS + 32 + 2)) st_1e_2 (
|
||||
VX_generic_register #(
|
||||
.N( 1+1+1+1+32+`WORD_SIZE+`WORD_SIZE+(`BANK_LINE_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;
|
||||
@@ -520,7 +520,6 @@ module VX_bank
|
||||
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 == `L2FUNC_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;
|
||||
@@ -532,7 +531,10 @@ module VX_bank
|
||||
|
||||
wire cwbq_empty;
|
||||
assign bank_wb_valid = !cwbq_empty;
|
||||
VX_generic_queue_ll #(.DATAW( `LOG2UP(NUM_REQUESTS) + 5 + 2 + (`NW_BITS-1+1) + `WORD_SIZE + 32 + 32), .SIZE(CWBQ_SIZE)) cwb_queue(
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW( `LOG2UP(NUM_REQUESTS) + 5 + 2 + (`NW_BITS-1+1) + `WORD_SIZE + 32 + 32),
|
||||
.SIZE(CWBQ_SIZE)
|
||||
) cwb_queue(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -543,13 +545,13 @@ module VX_bank
|
||||
.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;
|
||||
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_WORDS-1:0][`WORD_SIZE-1:0] dwbq_req_data;
|
||||
if ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID)) begin
|
||||
@@ -560,10 +562,9 @@ module VX_bank
|
||||
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),
|
||||
@@ -581,9 +582,7 @@ module VX_bank
|
||||
.LLVQ_SIZE (LLVQ_SIZE),
|
||||
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
|
||||
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
|
||||
)
|
||||
VX_fill_invalidator
|
||||
(
|
||||
) vx_fill_invalidator (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.possible_fill (possible_fill),
|
||||
@@ -591,16 +590,19 @@ module VX_bank
|
||||
.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;
|
||||
// Enqueue in dram_fill_req
|
||||
assign dram_fill_req_valid = possible_fill && !invalidate_fill;
|
||||
assign dram_fill_req_is_snp = is_snp_st2 && valid_st2 && miss_st2;
|
||||
assign dram_fill_req_addr = addr_st2 & `BASE_ADDR_MASK;
|
||||
|
||||
assign dram_wb_req = !dwbq_empty;
|
||||
VX_generic_queue_ll #(.DATAW( 32 + (`BANK_LINE_WORDS * `WORD_SIZE)), .SIZE(DWBQ_SIZE)) dwb_queue(
|
||||
assign dram_wb_req_valid = !dwbq_empty;
|
||||
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW( 32 + (`BANK_LINE_WORDS * `WORD_SIZE)),
|
||||
.SIZE(DWBQ_SIZE)
|
||||
) dwb_queue (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -611,14 +613,18 @@ module VX_bank
|
||||
.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(
|
||||
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_valid = !ffsq_empty;
|
||||
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW(32),
|
||||
.SIZE(FFSQ_SIZE)
|
||||
) ffs_queue (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.push (snp_fwd_push),
|
||||
@@ -627,7 +633,7 @@ module VX_bank
|
||||
.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);
|
||||
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_cache
|
||||
#(
|
||||
module VX_cache #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -17,7 +16,7 @@ module VX_cache
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 3,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
@@ -28,7 +27,7 @@ module VX_cache
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -47,21 +46,18 @@ module VX_cache
|
||||
parameter PRFQ_SIZE = 64,
|
||||
parameter PRFQ_STRIDE = 0,
|
||||
|
||||
// Dram knobs
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
|
||||
// Req Info
|
||||
// Req Info
|
||||
input wire [NUM_REQUESTS-1:0] core_req_valid,
|
||||
input wire [NUM_REQUESTS-1:0][31:0] core_req_addr,
|
||||
input wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] core_req_writedata,
|
||||
input wire[NUM_REQUESTS-1:0][2:0] core_req_mem_read,
|
||||
input wire[NUM_REQUESTS-1:0][2:0] core_req_mem_write,
|
||||
input wire [NUM_REQUESTS-1:0][2:0] core_req_mem_read,
|
||||
input wire [NUM_REQUESTS-1:0][2:0] core_req_mem_write,
|
||||
|
||||
// Req meta
|
||||
input wire [4:0] core_req_rd,
|
||||
@@ -80,39 +76,31 @@ module VX_cache
|
||||
output wire [NUM_REQUESTS-1:0][31:0] core_wb_pc,
|
||||
output wire [NUM_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_WORDS-1:0][31:0] dram_fill_rsp_data,
|
||||
output wire dram_fill_accept,
|
||||
input wire dram_rsp_valid,
|
||||
input wire [31:0] dram_rsp_addr,
|
||||
input wire [`IBANK_LINE_WORDS-1:0][31:0] dram_rsp_data,
|
||||
output wire dram_rsp_ready,
|
||||
|
||||
// Dram request
|
||||
output wire dram_req,
|
||||
output wire dram_req_write,
|
||||
output wire dram_req_read,
|
||||
output wire dram_req_write,
|
||||
output wire [31:0] dram_req_addr,
|
||||
output wire [31:0] dram_req_size,
|
||||
output wire [`IBANK_LINE_WORDS-1:0][31:0] dram_req_data,
|
||||
output wire dram_req_because_of_wb,
|
||||
input wire dram_req_delay,
|
||||
|
||||
output wire dram_snp_full,
|
||||
|
||||
input wire dram_req_full,
|
||||
|
||||
|
||||
// Snoop Req
|
||||
input wire snp_req,
|
||||
input wire[31:0] snp_req_addr,
|
||||
output wire snp_req_delay,
|
||||
input wire snp_req_valid,
|
||||
input wire [31:0] snp_req_addr,
|
||||
output wire snp_req_full,
|
||||
|
||||
// Snoop Forward
|
||||
output wire snp_fwd,
|
||||
output wire[31:0] snp_fwd_addr,
|
||||
input wire snp_fwd_delay
|
||||
|
||||
output wire snp_fwd_valid,
|
||||
output wire [31:0] snp_fwd_addr,
|
||||
input wire snp_fwd_full
|
||||
);
|
||||
|
||||
|
||||
wire [NUM_BANKS-1:0][NUM_REQUESTS-1:0] per_bank_valids;
|
||||
wire [NUM_BANKS-1:0] per_bank_wb_pop;
|
||||
wire [NUM_BANKS-1:0] per_bank_wb_valid;
|
||||
@@ -124,104 +112,90 @@ module VX_cache
|
||||
wire [NUM_BANKS-1:0][31:0] per_bank_wb_pc;
|
||||
wire [NUM_BANKS-1:0][31:0] per_bank_wb_address;
|
||||
|
||||
|
||||
wire dfqq_full;
|
||||
wire[NUM_BANKS-1:0] per_bank_dram_fill_req;
|
||||
wire[NUM_BANKS-1:0][31:0] per_bank_dram_fill_req_addr;
|
||||
wire[NUM_BANKS-1:0] per_bank_dram_fill_accept;
|
||||
wire [NUM_BANKS-1:0] per_bank_dram_fill_req_valid;
|
||||
wire [NUM_BANKS-1:0][31:0] per_bank_dram_fill_req_addr;
|
||||
/* verilator lint_off UNUSED */
|
||||
wire [NUM_BANKS-1:0] per_bank_dram_fill_req_is_snp;
|
||||
/* verilator lint_on UNUSED */
|
||||
wire [NUM_BANKS-1:0] per_bank_dram_rsp_ready;
|
||||
|
||||
wire[NUM_BANKS-1:0] per_bank_dram_wb_queue_pop;
|
||||
wire[NUM_BANKS-1:0] per_bank_dram_wb_req;
|
||||
wire[NUM_BANKS-1:0] per_bank_dram_because_of_snp;
|
||||
wire[NUM_BANKS-1:0][31:0] per_bank_dram_wb_req_addr;
|
||||
wire[NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] per_bank_dram_wb_req_data;
|
||||
wire [NUM_BANKS-1:0] per_bank_dram_wb_queue_pop;
|
||||
wire [NUM_BANKS-1:0] per_bank_dram_wb_req_valid;
|
||||
wire [NUM_BANKS-1:0][31:0] per_bank_dram_wb_req_addr;
|
||||
wire [NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] per_bank_dram_wb_req_data;
|
||||
|
||||
wire[NUM_BANKS-1:0] per_bank_reqq_full;
|
||||
|
||||
wire[NUM_BANKS-1:0] per_bank_snrq_full;
|
||||
|
||||
wire[NUM_BANKS-1:0] per_bank_snp_fwd;
|
||||
wire[NUM_BANKS-1:0][31:0] per_bank_snp_fwd_addr;
|
||||
wire[NUM_BANKS-1:0] per_bank_snp_fwd_pop;
|
||||
wire [NUM_BANKS-1:0] per_bank_reqq_full;
|
||||
wire [NUM_BANKS-1:0] per_bank_snrq_full;
|
||||
|
||||
wire [NUM_BANKS-1:0] per_bank_snp_fwd;
|
||||
wire [NUM_BANKS-1:0][31:0] per_bank_snp_fwd_addr;
|
||||
wire [NUM_BANKS-1:0] per_bank_snp_fwd_pop;
|
||||
|
||||
assign delay_req = (|per_bank_reqq_full);
|
||||
assign snp_req_full = (|per_bank_snrq_full);
|
||||
|
||||
|
||||
assign snp_req_delay = (|per_bank_snrq_full);
|
||||
|
||||
|
||||
// assign dram_fill_accept = (NUM_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);
|
||||
// assign dram_rsp_ready = (NUM_BANKS == 1) ? per_bank_dram_rsp_ready[0] : per_bank_dram_rsp_ready[dram_rsp_addr[`BANK_SELECT_ADDR_RNG]];
|
||||
assign dram_rsp_ready = (|per_bank_dram_rsp_ready);
|
||||
|
||||
VX_cache_dram_req_arb #(
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUM_REQUESTS (NUM_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)
|
||||
);
|
||||
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUM_REQUESTS (NUM_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_valid(per_bank_dram_fill_req_valid),
|
||||
.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_valid (per_bank_dram_wb_req_valid),
|
||||
.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_read (dram_req_read),
|
||||
.dram_req_write (dram_req_write),
|
||||
.dram_req_addr (dram_req_addr),
|
||||
.dram_req_data (dram_req_data),
|
||||
.dram_req_full (dram_req_full)
|
||||
);
|
||||
|
||||
VX_cache_core_req_bank_sel #(
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUM_REQUESTS (NUM_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)
|
||||
);
|
||||
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUM_REQUESTS (NUM_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),
|
||||
@@ -241,9 +215,7 @@ module VX_cache
|
||||
.LLVQ_SIZE (LLVQ_SIZE),
|
||||
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
|
||||
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
|
||||
)
|
||||
VX_cache_core_wb_sel_merge
|
||||
(
|
||||
) 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),
|
||||
@@ -262,28 +234,27 @@ module VX_cache
|
||||
.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 #(.NUM_BANKS(NUM_BANKS)) VX_snp_fwd_arb(
|
||||
VX_snp_fwd_arb #(
|
||||
.NUM_BANKS(NUM_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_valid (snp_fwd_valid),
|
||||
.snp_fwd_addr (snp_fwd_addr),
|
||||
.snp_fwd_delay (snp_fwd_delay)
|
||||
);
|
||||
.snp_fwd_full (snp_fwd_full)
|
||||
);
|
||||
|
||||
// Snoop Forward Logic
|
||||
|
||||
genvar curr_bank;
|
||||
generate
|
||||
for (curr_bank = 0; curr_bank < NUM_BANKS; curr_bank=curr_bank+1) begin
|
||||
wire [NUM_REQUESTS-1:0] curr_bank_valids;
|
||||
wire [NUM_REQUESTS-1:0][31:0] curr_bank_addr;
|
||||
wire [NUM_REQUESTS-1:0] curr_bank_valids;
|
||||
wire [NUM_REQUESTS-1:0][31:0] curr_bank_addr;
|
||||
wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] curr_bank_writedata;
|
||||
wire [4:0] curr_bank_rd;
|
||||
wire [NUM_REQUESTS-1:0][1:0] curr_bank_wb;
|
||||
@@ -294,7 +265,7 @@ module VX_cache
|
||||
|
||||
wire curr_bank_wb_pop;
|
||||
wire curr_bank_wb_valid;
|
||||
wire [`LOG2UP(NUM_REQUESTS)-1:0] curr_bank_wb_tid;
|
||||
wire [`LOG2UP(NUM_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;
|
||||
@@ -302,19 +273,18 @@ module VX_cache
|
||||
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_WORDS-1:0][`WORD_SIZE-1:0] curr_bank_dram_fill_rsp_data;
|
||||
wire curr_bank_dram_fill_accept;
|
||||
wire curr_bank_dram_rsp_valid;
|
||||
wire [31:0] curr_bank_dram_rsp_addr;
|
||||
wire [`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] curr_bank_dram_rsp_data;
|
||||
wire curr_bank_dram_rsp_ready;
|
||||
|
||||
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 curr_bank_dram_fill_req_valid;
|
||||
wire curr_bank_dram_fill_req_is_snp;
|
||||
wire[31:0] curr_bank_dram_fill_req_addr;
|
||||
|
||||
wire curr_bank_dram_wb_queue_pop;
|
||||
wire curr_bank_dram_wb_req;
|
||||
wire curr_bank_dram_wb_req_valid;
|
||||
wire[31:0] curr_bank_dram_wb_req_addr;
|
||||
wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] curr_bank_dram_wb_req_data;
|
||||
|
||||
@@ -326,9 +296,7 @@ module VX_cache
|
||||
wire curr_bank_snp_fwd;
|
||||
wire[31:0] curr_bank_snp_fwd_addr;
|
||||
wire curr_bank_snp_fwd_pop;
|
||||
wire curr_bank_snrq_full;
|
||||
|
||||
|
||||
wire curr_bank_snp_req_full;
|
||||
|
||||
// Core Req
|
||||
assign curr_bank_valids = per_bank_valids[curr_bank];
|
||||
@@ -354,56 +322,53 @@ module VX_cache
|
||||
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;
|
||||
assign curr_bank_dfqq_full = dfqq_full;
|
||||
assign per_bank_dram_fill_req_valid[curr_bank] = curr_bank_dram_fill_req_valid;
|
||||
assign per_bank_dram_fill_req_addr[curr_bank] = curr_bank_dram_fill_req_addr;
|
||||
assign per_bank_dram_fill_req_is_snp[curr_bank] = curr_bank_dram_fill_req_is_snp;
|
||||
|
||||
// Dram fill response
|
||||
assign curr_bank_dram_fill_rsp = (NUM_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;
|
||||
assign curr_bank_dram_rsp_valid = (NUM_BANKS == 1) || (dram_rsp_valid && (curr_bank_dram_rsp_addr[`BANK_SELECT_ADDR_RNG] == curr_bank));
|
||||
assign curr_bank_dram_rsp_addr = dram_rsp_addr;
|
||||
assign curr_bank_dram_rsp_data = dram_rsp_data;
|
||||
assign per_bank_dram_rsp_ready[curr_bank] = curr_bank_dram_rsp_ready;
|
||||
|
||||
// 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_valid[curr_bank] = curr_bank_dram_wb_req_valid;
|
||||
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;
|
||||
assign curr_bank_snp_req = snp_req_valid && (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_snp_req_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),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUM_REQUESTS (NUM_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
|
||||
(
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
|
||||
.NUM_BANKS (NUM_BANKS),
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUM_REQUESTS (NUM_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
|
||||
@@ -431,39 +396,35 @@ module VX_cache
|
||||
.bank_wb_address (curr_bank_wb_address),
|
||||
|
||||
// Dram fill req
|
||||
.dram_fill_req (curr_bank_dram_fill_req),
|
||||
.dram_fill_req_valid (curr_bank_dram_fill_req_valid),
|
||||
.dram_fill_req_addr (curr_bank_dram_fill_req_addr),
|
||||
.dram_fill_req_is_snp (curr_bank_dram_fill_req_is_snp),
|
||||
.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_rsp_valid (curr_bank_dram_rsp_valid),
|
||||
.dram_rsp_addr (curr_bank_dram_rsp_addr),
|
||||
.dram_rsp_data (curr_bank_dram_rsp_data),
|
||||
.dram_rsp_ready (curr_bank_dram_rsp_ready),
|
||||
|
||||
// Dram writeback
|
||||
.dram_wb_queue_pop (curr_bank_dram_wb_queue_pop),
|
||||
.dram_wb_req (curr_bank_dram_wb_req),
|
||||
.dram_wb_req_valid (curr_bank_dram_wb_req_valid),
|
||||
.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),
|
||||
.dram_wb_req_data (curr_bank_dram_wb_req_data),
|
||||
|
||||
// Snoop Request
|
||||
.snp_req (curr_bank_snp_req),
|
||||
.snp_req_valid (curr_bank_snp_req),
|
||||
.snp_req_addr (curr_bank_snp_req_addr),
|
||||
.snrq_full (curr_bank_snrq_full),
|
||||
.snp_req_full (curr_bank_snp_req_full),
|
||||
|
||||
// Snoop Fwd
|
||||
.snp_fwd (curr_bank_snp_fwd),
|
||||
.snp_fwd_valid (curr_bank_snp_fwd),
|
||||
.snp_fwd_addr (curr_bank_snp_fwd_addr),
|
||||
.snp_fwd_pop (curr_bank_snp_fwd_pop)
|
||||
|
||||
);
|
||||
|
||||
);
|
||||
end
|
||||
|
||||
endgenerate
|
||||
|
||||
|
||||
|
||||
|
||||
endmodule
|
||||
@@ -9,24 +9,24 @@
|
||||
// 5 + 2 + 4 + 3 + 3 + 1
|
||||
`define REQ_INST_META_SIZE (5 + 2 + (`NW_BITS-1+1) + 3 + 3 + `LOG2UP(NUM_REQUESTS))
|
||||
|
||||
`define WORD_SIZE (8*WORD_SIZE_BYTES)
|
||||
`define WORD_SIZE (8 * WORD_SIZE_BYTES)
|
||||
`define WORD_SIZE_RNG (`WORD_SIZE)-1:0
|
||||
|
||||
// 128
|
||||
`define BANK_SIZE_BYTES CACHE_SIZE_BYTES/NUM_BANKS
|
||||
`define BANK_SIZE_BYTES (CACHE_SIZE_BYTES / NUM_BANKS)
|
||||
|
||||
// 8
|
||||
`define BANK_LINE_COUNT (`BANK_SIZE_BYTES/BANK_LINE_SIZE_BYTES)
|
||||
`define BANK_LINE_COUNT (`BANK_SIZE_BYTES / BANK_LINE_SIZE_BYTES)
|
||||
// 4
|
||||
`define BANK_LINE_WORDS (BANK_LINE_SIZE_BYTES / WORD_SIZE_BYTES)
|
||||
|
||||
// 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
|
||||
`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 (`LOG2UP(`BANK_LINE_WORDS))
|
||||
@@ -55,17 +55,14 @@
|
||||
|
||||
// 3
|
||||
`define LINE_SELECT_NUM_BITS (`LOG2UP(`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)
|
||||
`define LINE_SELECT_ADDR_END (`LINE_SELECT_NUM_BITS+`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
|
||||
|
||||
`define LINE_SELECT_SIZE_RNG `LINE_SELECT_NUM_BITS-1:0
|
||||
|
||||
// 10
|
||||
`define TAG_SELECT_ADDR_START (1+`LINE_SELECT_ADDR_END)
|
||||
@@ -76,9 +73,10 @@
|
||||
// 22
|
||||
`define TAG_SELECT_SIZE_END (`TAG_SELECT_NUM_BITS)
|
||||
// 21:0
|
||||
`define TAG_SELECT_SIZE_RNG `TAG_SELECT_SIZE_END-1:0
|
||||
`define TAG_SELECT_SIZE_RNG `TAG_SELECT_NUM_BITS-1:0
|
||||
|
||||
`define TAG_LINE_SELECT_BITS (`TAG_SELECT_NUM_BITS+`LINE_SELECT_NUM_BITS)
|
||||
|
||||
`define BASE_ADDR_MASK (~((1<<(`WORD_SELECT_ADDR_END+1))-1))
|
||||
|
||||
`endif
|
||||
|
||||
|
||||
@@ -54,8 +54,6 @@ module VX_cache_core_req_bank_sel
|
||||
output reg [NUM_BANKS-1:0][NUM_REQUESTS-1:0] per_bank_valids
|
||||
);
|
||||
|
||||
wire[31:0] req_address;
|
||||
|
||||
generate
|
||||
integer curr_req;
|
||||
always @(*) begin
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_cache_dfq_queue
|
||||
#(
|
||||
module VX_cache_dfq_queue #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -15,8 +14,7 @@ module VX_cache_dfq_queue
|
||||
// Number of cycles to complete stage 1 (read from memory)
|
||||
parameter STAGE_1_CYCLES = 2,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
// Miss Reserv Queue Knob
|
||||
@@ -26,7 +24,7 @@ module VX_cache_dfq_queue
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -39,16 +37,13 @@ module VX_cache_dfq_queue
|
||||
// Fill Invalidator Size {Fill invalidator must be active}
|
||||
parameter FILL_INVALIDAOR_SIZE = 16,
|
||||
|
||||
// Dram knobs
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
input wire dfqq_push,
|
||||
input wire[NUM_BANKS-1:0] per_bank_dram_fill_req,
|
||||
input wire[NUM_BANKS-1:0] per_bank_dram_fill_req_valid,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_dram_fill_req_addr,
|
||||
|
||||
input wire dfqq_pop,
|
||||
@@ -61,17 +56,14 @@ module VX_cache_dfq_queue
|
||||
wire[NUM_BANKS-1:0] out_per_bank_dram_fill_req;
|
||||
wire[NUM_BANKS-1:0][31:0] out_per_bank_dram_fill_req_addr;
|
||||
|
||||
|
||||
reg [NUM_BANKS-1:0] use_per_bank_dram_fill_req;
|
||||
reg [NUM_BANKS-1:0][31:0] use_per_bank_dram_fill_req_addr;
|
||||
|
||||
|
||||
wire[NUM_BANKS-1:0] qual_bank_dram_fill_req;
|
||||
wire[NUM_BANKS-1:0][31:0] qual_bank_dram_fill_req_addr;
|
||||
|
||||
wire[NUM_BANKS-1:0] updated_bank_dram_fill_req;
|
||||
|
||||
|
||||
wire o_empty;
|
||||
|
||||
wire use_empty = !(|use_per_bank_dram_fill_req);
|
||||
@@ -79,27 +71,34 @@ module VX_cache_dfq_queue
|
||||
|
||||
wire push_qual = dfqq_push && !dfqq_full;
|
||||
wire pop_qual = dfqq_pop && use_empty && !out_empty;
|
||||
VX_generic_queue_ll #(.DATAW(NUM_BANKS * (1+32)), .SIZE(DFQQ_SIZE)) dfqq_queue(
|
||||
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW(NUM_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}),
|
||||
.in_data ({per_bank_dram_fill_req_valid, 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 & {NUM_BANKS{!o_empty}}) : (use_per_bank_dram_fill_req & {NUM_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[`LOG2UP(NUM_BANKS)-1:0] qual_request_index;
|
||||
wire qual_has_request;
|
||||
VX_generic_priority_encoder #(.N(NUM_BANKS)) VX_sel_bank(
|
||||
wire qual_has_request;
|
||||
|
||||
VX_generic_priority_encoder #(
|
||||
.N(NUM_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];
|
||||
@@ -119,5 +118,4 @@ module VX_cache_dfq_queue
|
||||
end
|
||||
end
|
||||
|
||||
|
||||
endmodule
|
||||
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_cache_dram_req_arb
|
||||
#(
|
||||
module VX_cache_dram_req_arb #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -15,7 +14,7 @@ module VX_cache_dram_req_arb
|
||||
// Number of cycles to complete stage 1 (read from memory)
|
||||
parameter STAGE_1_CYCLES = 2,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
@@ -26,7 +25,7 @@ module VX_cache_dram_req_arb
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -45,39 +44,29 @@ module VX_cache_dram_req_arb
|
||||
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
|
||||
|
||||
// Fill Request
|
||||
output wire dfqq_full,
|
||||
input wire[NUM_BANKS-1:0] per_bank_dram_fill_req,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_dram_fill_req_addr,
|
||||
|
||||
// DFQ Request
|
||||
output wire[NUM_BANKS-1:0] per_bank_dram_wb_queue_pop,
|
||||
input wire[NUM_BANKS-1:0] per_bank_dram_wb_req,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_dram_wb_req_addr,
|
||||
input wire[NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] per_bank_dram_wb_req_data,
|
||||
input wire[NUM_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_WORDS-1:0][31:0] dram_req_data,
|
||||
output wire dram_req_because_of_wb,
|
||||
|
||||
input wire dram_req_delay
|
||||
output wire dfqq_full,
|
||||
input wire[NUM_BANKS-1:0] per_bank_dram_fill_req_valid,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_dram_fill_req_addr,
|
||||
|
||||
);
|
||||
// DFQ Request
|
||||
output wire[NUM_BANKS-1:0] per_bank_dram_wb_queue_pop,
|
||||
input wire[NUM_BANKS-1:0] per_bank_dram_wb_req_valid,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_dram_wb_req_addr,
|
||||
input wire[NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] per_bank_dram_wb_req_data,
|
||||
|
||||
// real Dram request
|
||||
output wire dram_req_read,
|
||||
output wire dram_req_write,
|
||||
output wire [31:0] dram_req_addr,
|
||||
output wire [`IBANK_LINE_WORDS-1:0][31:0] dram_req_data,
|
||||
|
||||
input wire dram_req_full
|
||||
);
|
||||
|
||||
wire pref_pop;
|
||||
wire pref_valid;
|
||||
@@ -86,66 +75,62 @@ module VX_cache_dram_req_arb
|
||||
wire dwb_valid;
|
||||
wire dfqq_req;
|
||||
|
||||
assign pref_pop = !dwb_valid && !dfqq_req && !dram_req_delay && pref_valid;
|
||||
assign pref_pop = !dwb_valid && !dfqq_req && !dram_req_full && 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
|
||||
(
|
||||
) prfqq (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
.dram_req (dram_req && dram_req_read),
|
||||
.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;
|
||||
/* verilator lint_off UNUSED */
|
||||
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);
|
||||
/* verilator lint_on UNUSED */
|
||||
wire dfqq_pop = !dwb_valid && dfqq_req && !dram_req_full; // If no dwb, and dfqq has valids, then pop
|
||||
wire dfqq_push = (|per_bank_dram_fill_req_valid);
|
||||
|
||||
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)
|
||||
);
|
||||
VX_cache_dfq_queue vx_cache_dfq_queue(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.dfqq_push (dfqq_push),
|
||||
.per_bank_dram_fill_req_valid (per_bank_dram_fill_req_valid),
|
||||
.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[`LOG2UP(NUM_BANKS)-1:0] dwb_bank;
|
||||
// wire[NUM_BANKS-1:0] use_wb_valid = per_bank_dram_wb_req | per_bank_dram_because_of_snp;
|
||||
wire[NUM_BANKS-1:0] use_wb_valid = per_bank_dram_wb_req;
|
||||
VX_generic_priority_encoder #(.N(NUM_BANKS)) VX_sel_dwb(
|
||||
|
||||
wire[NUM_BANKS-1:0] use_wb_valid = per_bank_dram_wb_req_valid;
|
||||
|
||||
VX_generic_priority_encoder #(
|
||||
.N(NUM_BANKS)
|
||||
) vx_sel_dwb (
|
||||
.valids(use_wb_valid),
|
||||
.index (dwb_bank),
|
||||
.found (dwb_valid)
|
||||
);
|
||||
);
|
||||
|
||||
assign per_bank_dram_wb_queue_pop = dram_req_full ? 0 : use_wb_valid & ((1 << dwb_bank));
|
||||
|
||||
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;
|
||||
wire dram_req = dwb_valid || dfqq_req || pref_pop;
|
||||
assign dram_req_read = ((dfqq_req && !dwb_valid) || pref_pop) && dram_req;
|
||||
assign dram_req_write = dwb_valid && 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_data} = dwb_valid ? {per_bank_dram_wb_req_data[dwb_bank] }: 0;
|
||||
|
||||
endmodule
|
||||
@@ -1,8 +1,7 @@
|
||||
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_cache_miss_resrv
|
||||
#(
|
||||
module VX_cache_miss_resrv #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -16,8 +15,7 @@ module VX_cache_miss_resrv
|
||||
// Number of cycles to complete stage 1 (read from memory)
|
||||
parameter STAGE_1_CYCLES = 2,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
// Miss Reserv Queue Knob
|
||||
@@ -27,7 +25,7 @@ module VX_cache_miss_resrv
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -40,12 +38,9 @@ module VX_cache_miss_resrv
|
||||
// Fill Invalidator Size {Fill invalidator must be active}
|
||||
parameter FILL_INVALIDAOR_SIZE = 16,
|
||||
|
||||
// Dram knobs
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
|
||||
@@ -65,7 +60,11 @@ module VX_cache_miss_resrv
|
||||
|
||||
// Broadcast Fill
|
||||
input wire is_fill_st1,
|
||||
|
||||
/* verilator lint_off UNUSED */
|
||||
// TODO: should fix this
|
||||
input wire[31:0] fill_addr_st1,
|
||||
/* verilator lint_on UNUSED */
|
||||
|
||||
// Miss dequeue
|
||||
input wire miss_resrv_pop,
|
||||
@@ -81,96 +80,91 @@ module VX_cache_miss_resrv
|
||||
output wire[2:0] miss_resrv_mem_write_st0
|
||||
|
||||
);
|
||||
// Size of metadata = 32 + `LOG2UP(NUM_REQUESTS) + 5 + 2 + (`NW_BITS-1 + 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 [`LOG2UP(MRVQ_SIZE)-1:0] head_ptr;
|
||||
reg [`LOG2UP(MRVQ_SIZE)-1:0] tail_ptr;
|
||||
|
||||
// Size of metadata = 32 + `LOG2UP(NUM_REQUESTS) + 5 + 2 + (`NW_BITS-1 + 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[`LOG2UP(MRVQ_SIZE)-1:0] head_ptr;
|
||||
reg[`LOG2UP(MRVQ_SIZE)-1:0] tail_ptr;
|
||||
reg [31:0] size;
|
||||
|
||||
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 [`LOG2UP(MRVQ_SIZE)-1:0] enqueue_index = tail_ptr;
|
||||
|
||||
// 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));
|
||||
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 enqueue_possible = !miss_resrv_full;
|
||||
wire[`LOG2UP(MRVQ_SIZE)-1:0] enqueue_index = tail_ptr;
|
||||
wire dequeue_possible = valid_table[head_ptr] && ready_table[head_ptr];
|
||||
wire [`LOG2UP(MRVQ_SIZE)-1:0] dequeue_index = head_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];
|
||||
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) begin
|
||||
metadata_table[i] <= 0;
|
||||
end
|
||||
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
|
||||
endgenerate
|
||||
|
||||
if (update_ready) begin
|
||||
ready_table <= ready_table | make_ready;
|
||||
end
|
||||
|
||||
wire dequeue_possible = valid_table[head_ptr] && ready_table[head_ptr];
|
||||
wire[`LOG2UP(MRVQ_SIZE)-1:0] dequeue_index = head_ptr;
|
||||
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
|
||||
|
||||
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 && mrvq_pop)) 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;
|
||||
size <= size + 1;
|
||||
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;
|
||||
size <= size - 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
|
||||
|
||||
|
||||
end
|
||||
|
||||
endmodule
|
||||
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_cache_req_queue
|
||||
#(
|
||||
module VX_cache_req_queue #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -15,8 +14,7 @@ module VX_cache_req_queue
|
||||
// Number of cycles to complete stage 1 (read from memory)
|
||||
parameter STAGE_1_CYCLES = 2,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
// Miss Reserv Queue Knob
|
||||
@@ -26,7 +24,7 @@ module VX_cache_req_queue
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -39,12 +37,9 @@ module VX_cache_req_queue
|
||||
// Fill Invalidator Size {Fill invalidator must be active}
|
||||
parameter FILL_INVALIDAOR_SIZE = 16,
|
||||
|
||||
// Dram knobs
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
|
||||
@@ -88,7 +83,6 @@ module VX_cache_req_queue
|
||||
wire [NUM_REQUESTS-1:0][2:0] out_per_mem_write;
|
||||
wire [31:0] out_per_pc;
|
||||
|
||||
|
||||
reg [NUM_REQUESTS-1:0] use_per_valids;
|
||||
reg [NUM_REQUESTS-1:0][31:0] use_per_addr;
|
||||
reg [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] use_per_writedata;
|
||||
@@ -99,7 +93,6 @@ module VX_cache_req_queue
|
||||
reg [NUM_REQUESTS-1:0][2:0] use_per_mem_read;
|
||||
reg [NUM_REQUESTS-1:0][2:0] use_per_mem_write;
|
||||
|
||||
|
||||
wire [NUM_REQUESTS-1:0] qual_valids;
|
||||
wire [NUM_REQUESTS-1:0][31:0] qual_addr;
|
||||
wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] qual_writedata;
|
||||
@@ -110,7 +103,9 @@ module VX_cache_req_queue
|
||||
wire [NUM_REQUESTS-1:0][2:0] qual_mem_write;
|
||||
wire [31:0] qual_pc;
|
||||
|
||||
/* verilator lint_off UNUSED */
|
||||
reg [NUM_REQUESTS-1:0] updated_valids;
|
||||
/* verilator lint_on UNUSED */
|
||||
|
||||
wire o_empty;
|
||||
|
||||
@@ -120,17 +115,19 @@ module VX_cache_req_queue
|
||||
wire push_qual = reqq_push && !reqq_full;
|
||||
wire pop_qual = !out_empty && use_empty;
|
||||
|
||||
VX_generic_queue_ll #(.DATAW( (NUM_REQUESTS * (1+32+`WORD_SIZE)) + 5 + (NUM_REQUESTS*2) + (`NW_BITS-1+1) + (NUM_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)
|
||||
);
|
||||
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW( (NUM_REQUESTS * (1+32+`WORD_SIZE)) + 5 + (NUM_REQUESTS*2) + (`NW_BITS-1+1) + (NUM_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[NUM_REQUESTS-1:0] real_out_per_valids = out_per_valids & {NUM_REQUESTS{~out_empty}};
|
||||
|
||||
@@ -146,11 +143,13 @@ module VX_cache_req_queue
|
||||
|
||||
wire[`LOG2UP(NUM_REQUESTS)-1:0] qual_request_index;
|
||||
wire qual_has_request;
|
||||
VX_generic_priority_encoder #(.N(NUM_REQUESTS)) VX_sel_bank(
|
||||
VX_generic_priority_encoder #(
|
||||
.N(NUM_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;
|
||||
@@ -164,7 +163,6 @@ module VX_cache_req_queue
|
||||
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
|
||||
@@ -172,7 +170,6 @@ module VX_cache_req_queue
|
||||
end
|
||||
end
|
||||
|
||||
|
||||
always @(posedge clk) begin
|
||||
if (reset) begin
|
||||
use_per_valids <= 0;
|
||||
@@ -204,5 +201,4 @@ module VX_cache_req_queue
|
||||
end
|
||||
end
|
||||
|
||||
|
||||
endmodule
|
||||
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_cache_wb_sel_merge
|
||||
#(
|
||||
module VX_cache_wb_sel_merge #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -17,8 +16,7 @@ module VX_cache_wb_sel_merge
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 0,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
// Miss Reserv Queue Knob
|
||||
@@ -28,7 +26,7 @@ module VX_cache_wb_sel_merge
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -41,35 +39,29 @@ module VX_cache_wb_sel_merge
|
||||
// Fill Invalidator Size {Fill invalidator must be active}
|
||||
parameter FILL_INVALIDAOR_SIZE = 16,
|
||||
|
||||
// Dram knobs
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
|
||||
) (
|
||||
// Per Bank WB
|
||||
input wire [NUM_BANKS-1:0] per_bank_wb_valid,
|
||||
input wire [NUM_BANKS-1:0][`LOG2UP(NUM_REQUESTS)-1:0] per_bank_wb_tid,
|
||||
input wire [NUM_BANKS-1:0][4:0] per_bank_wb_rd,
|
||||
input wire [NUM_BANKS-1:0][1:0] per_bank_wb_wb,
|
||||
input wire [NUM_BANKS-1:0][`NW_BITS-1:0] per_bank_wb_warp_num,
|
||||
input wire [NUM_BANKS-1:0][`WORD_SIZE_RNG] per_bank_wb_data,
|
||||
input wire [NUM_BANKS-1:0][31:0] per_bank_wb_pc,
|
||||
input wire [NUM_BANKS-1:0][31:0] per_bank_wb_address,
|
||||
output wire [NUM_BANKS-1:0] per_bank_wb_pop,
|
||||
|
||||
input wire [NUM_BANKS-1:0] per_bank_wb_valid,
|
||||
input wire [NUM_BANKS-1:0][`LOG2UP(NUM_REQUESTS)-1:0] per_bank_wb_tid,
|
||||
input wire [NUM_BANKS-1:0][4:0] per_bank_wb_rd,
|
||||
input wire [NUM_BANKS-1:0][1:0] per_bank_wb_wb,
|
||||
input wire [NUM_BANKS-1:0][`NW_BITS-1:0] per_bank_wb_warp_num,
|
||||
input wire [NUM_BANKS-1:0][`WORD_SIZE_RNG] per_bank_wb_data,
|
||||
input wire [NUM_BANKS-1:0][31:0] per_bank_wb_pc,
|
||||
input wire [NUM_BANKS-1:0][31:0] per_bank_wb_address,
|
||||
output wire [NUM_BANKS-1:0] per_bank_wb_pop,
|
||||
|
||||
// Core Writeback
|
||||
input wire core_no_wb_slot,
|
||||
output reg [NUM_REQUESTS-1:0] core_wb_valid,
|
||||
output reg [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] core_wb_readdata,
|
||||
output reg [NUM_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_BITS-1:0] core_wb_warp_num,
|
||||
output reg [NUM_REQUESTS-1:0][31:0] core_wb_address
|
||||
|
||||
input wire core_no_wb_slot,
|
||||
output reg [NUM_REQUESTS-1:0] core_wb_valid,
|
||||
output reg [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] core_wb_readdata,
|
||||
output reg [NUM_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_BITS-1:0] core_wb_warp_num,
|
||||
output reg [NUM_REQUESTS-1:0][31:0] core_wb_address
|
||||
);
|
||||
|
||||
reg [NUM_BANKS-1:0] per_bank_wb_pop_unqual;
|
||||
@@ -83,15 +75,16 @@ module VX_cache_wb_sel_merge
|
||||
// end
|
||||
// endgenerate
|
||||
|
||||
|
||||
wire [`LOG2UP(NUM_BANKS)-1:0] main_bank_index;
|
||||
wire found_bank;
|
||||
wire found_bank;
|
||||
|
||||
VX_generic_priority_encoder #(.N(NUM_BANKS)) VX_sel_bank(
|
||||
VX_generic_priority_encoder #(
|
||||
.N(NUM_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];
|
||||
@@ -106,42 +99,36 @@ module VX_cache_wb_sel_merge
|
||||
core_wb_address = 0;
|
||||
for (this_bank = 0; this_bank < NUM_BANKS; this_bank = this_bank + 1) begin
|
||||
if ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID)) begin
|
||||
|
||||
if (found_bank
|
||||
&& !core_wb_valid[per_bank_wb_tid[this_bank]]
|
||||
&& per_bank_wb_valid[this_bank]
|
||||
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(this_bank))
|
||||
|| (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
|
||||
|
||||
if (found_bank
|
||||
&& !core_wb_valid[per_bank_wb_tid[this_bank]]
|
||||
&& per_bank_wb_valid[this_bank]
|
||||
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(this_bank))
|
||||
|| (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 (((main_bank_index == `LOG2UP(NUM_BANKS)'(this_bank))
|
||||
|| (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
|
||||
|
||||
if (((main_bank_index == `LOG2UP(NUM_BANKS)'(this_bank))
|
||||
|| (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
|
||||
|
||||
@@ -47,14 +47,14 @@ module VX_dcache_llv_resp_bank_sel
|
||||
(
|
||||
output reg [NUM_BANKS-1:0] per_bank_llvq_pop,
|
||||
input wire[NUM_BANKS-1:0] per_bank_llvq_valid,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_llvq_res_addr,
|
||||
input wire[NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][31:0] per_bank_llvq_res_data,
|
||||
input wire[NUM_BANKS-1:0][`LOG2UP(NUM_REQUESTS)-1:0] per_bank_llvq_res_tid,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_llvq_rsp_addr,
|
||||
input wire[NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][31:0] per_bank_llvq_rsp_data,
|
||||
input wire[NUM_BANKS-1:0][`LOG2UP(NUM_REQUESTS)-1:0] per_bank_llvq_rsp_tid,
|
||||
|
||||
input wire llvq_pop,
|
||||
output reg[NUM_REQUESTS-1:0] llvq_valid,
|
||||
output reg[NUM_REQUESTS-1:0][31:0] llvq_res_addr,
|
||||
output reg[NUM_REQUESTS-1:0][`BANK_LINE_WORDS-1:0][31:0] llvq_res_data
|
||||
output reg[NUM_REQUESTS-1:0][31:0] llvq_rsp_addr,
|
||||
output reg[NUM_REQUESTS-1:0][`BANK_LINE_WORDS-1:0][31:0] llvq_rsp_data
|
||||
|
||||
|
||||
);
|
||||
@@ -62,7 +62,7 @@ module VX_dcache_llv_resp_bank_sel
|
||||
wire [(`LOG2UP(NUM_BANKS))-1:0] main_bank_index;
|
||||
wire found_bank;
|
||||
|
||||
VX_generic_priority_encoder #(.N(NUM_BANKS)) VX_sel_bank(
|
||||
VX_generic_priority_encoder #(.N(NUM_BANKS)) vx_sel_bank(
|
||||
.valids(per_bank_llvq_valid),
|
||||
.index (main_bank_index),
|
||||
.found (found_bank)
|
||||
@@ -71,13 +71,13 @@ module VX_dcache_llv_resp_bank_sel
|
||||
|
||||
always @(*) begin
|
||||
llvq_valid = 0;
|
||||
llvq_res_addr = 0;
|
||||
llvq_res_data = 0;
|
||||
llvq_rsp_addr = 0;
|
||||
llvq_rsp_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];
|
||||
llvq_valid [per_bank_llvq_rsp_tid[main_bank_index]] = 1'b1;
|
||||
llvq_rsp_addr[per_bank_llvq_rsp_tid[main_bank_index]] = per_bank_llvq_rsp_addr[main_bank_index];
|
||||
llvq_rsp_data[per_bank_llvq_rsp_tid[main_bank_index]] = per_bank_llvq_rsp_data[main_bank_index];
|
||||
per_bank_llvq_pop[main_bank_index] = 1'b1;
|
||||
end
|
||||
end
|
||||
|
||||
@@ -82,17 +82,18 @@ module VX_fill_invalidator
|
||||
|
||||
|
||||
wire [(`LOG2UP(FILL_INVALIDAOR_SIZE))-1:0] enqueue_index;
|
||||
wire enqueue_found;
|
||||
VX_generic_priority_encoder #(.N(FILL_INVALIDAOR_SIZE)) VX_sel_bank(
|
||||
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;
|
||||
@@ -109,7 +110,6 @@ module VX_fill_invalidator
|
||||
end
|
||||
end
|
||||
|
||||
|
||||
// reg success_found;
|
||||
// reg[(`LOG2UP(FILL_INVALIDAOR_SIZE))-1:0] success_index;
|
||||
|
||||
@@ -133,21 +133,15 @@ module VX_fill_invalidator
|
||||
// end
|
||||
// end
|
||||
|
||||
|
||||
|
||||
|
||||
// wire [(`LOG2UP(FILL_INVALIDAOR_SIZE))-1:0] enqueue_index;
|
||||
// wire enqueue_found;
|
||||
|
||||
// VX_generic_priority_encoder #(.N(FILL_INVALIDAOR_SIZE)) VX_sel_bank(
|
||||
// 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;
|
||||
@@ -165,8 +159,6 @@ module VX_fill_invalidator
|
||||
// end
|
||||
// end
|
||||
|
||||
|
||||
end
|
||||
|
||||
|
||||
endmodule
|
||||
@@ -1,15 +1,13 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
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
|
||||
)
|
||||
(
|
||||
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,
|
||||
|
||||
@@ -21,24 +19,23 @@ module VX_prefetcher
|
||||
output wire[31:0] pref_addr
|
||||
|
||||
);
|
||||
|
||||
|
||||
reg[`LOG2UP(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(
|
||||
VX_generic_queue_ll #(
|
||||
.DATAW(32),
|
||||
.SIZE(PRFQ_SIZE)
|
||||
) pfq_queue (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -50,14 +47,11 @@ module VX_prefetcher
|
||||
|
||||
.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;
|
||||
@@ -70,7 +64,6 @@ module VX_prefetcher
|
||||
use_valid <= use_valid - 1;
|
||||
use_addr <= use_addr + BANK_LINE_SIZE_BYTES;
|
||||
end
|
||||
|
||||
end
|
||||
end
|
||||
|
||||
|
||||
@@ -5,28 +5,30 @@ module VX_snp_fwd_arb
|
||||
parameter NUM_BANKS = 8
|
||||
)
|
||||
(
|
||||
input wire[NUM_BANKS-1:0] per_bank_snp_fwd,
|
||||
input wire[NUM_BANKS-1:0][31:0] per_bank_snp_fwd_addr,
|
||||
output reg[NUM_BANKS-1:0] per_bank_snp_fwd_pop,
|
||||
input wire [NUM_BANKS-1:0] per_bank_snp_fwd,
|
||||
input wire [NUM_BANKS-1:0][31:0] per_bank_snp_fwd_addr,
|
||||
output reg [NUM_BANKS-1:0] per_bank_snp_fwd_pop,
|
||||
|
||||
output wire snp_fwd,
|
||||
output wire[31:0] snp_fwd_addr,
|
||||
input wire snp_fwd_delay
|
||||
output wire snp_fwd_valid,
|
||||
output wire [31:0] snp_fwd_addr,
|
||||
input wire snp_fwd_full
|
||||
|
||||
);
|
||||
|
||||
wire[NUM_BANKS-1:0] qual_per_bank_snp_fwd = per_bank_snp_fwd & {NUM_BANKS{!snp_fwd_delay}};
|
||||
wire[NUM_BANKS-1:0] qual_per_bank_snp_fwd = per_bank_snp_fwd & {NUM_BANKS{!snp_fwd_full}};
|
||||
|
||||
wire[`LOG2UP(NUM_BANKS)-1:0] fsq_bank;
|
||||
wire fsq_valid;
|
||||
wire fsq_valid;
|
||||
|
||||
VX_generic_priority_encoder #(.N(NUM_BANKS)) VX_sel_ffsq(
|
||||
VX_generic_priority_encoder #(
|
||||
.N(NUM_BANKS)
|
||||
) vx_sel_ffsq(
|
||||
.valids(qual_per_bank_snp_fwd),
|
||||
.index (fsq_bank),
|
||||
.found (fsq_valid)
|
||||
);
|
||||
);
|
||||
|
||||
assign snp_fwd = fsq_valid;
|
||||
assign snp_fwd_valid = fsq_valid;
|
||||
assign snp_fwd_addr = per_bank_snp_fwd_addr[fsq_bank];
|
||||
|
||||
always @(*) begin
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_tag_data_access
|
||||
#(
|
||||
module VX_tag_data_access #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -17,8 +16,7 @@ module VX_tag_data_access
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 0,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
// Miss Reserv Queue Knob
|
||||
@@ -28,7 +26,7 @@ module VX_tag_data_access
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -41,24 +39,26 @@ module VX_tag_data_access
|
||||
// Fill Invalidator Size {Fill invalidator must be active}
|
||||
parameter FILL_INVALIDAOR_SIZE = 16,
|
||||
|
||||
// Dram knobs
|
||||
// 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
|
||||
/* verilator lint_off UNUSED */
|
||||
// TODO:
|
||||
input wire[31:0] readaddr_st10,
|
||||
|
||||
/* verilator lint_on UNUSED */
|
||||
// Write/Read Logic
|
||||
input wire valid_req_st1e,
|
||||
input wire writefill_st1e,
|
||||
/* verilator lint_off UNUSED */
|
||||
// TODO:
|
||||
input wire[31:0] writeaddr_st1e,
|
||||
/* verilator lint_on UNUSED */
|
||||
input wire[`WORD_SIZE_RNG] writeword_st1e,
|
||||
input wire[`DBANK_LINE_WORDS-1:0][31:0] writedata_st1e,
|
||||
input wire[2:0] mem_write_st1e,
|
||||
@@ -69,19 +69,14 @@ module VX_tag_data_access
|
||||
output wire[`TAG_SELECT_SIZE_RNG] readtag_st1e,
|
||||
output wire miss_st1e,
|
||||
output wire dirty_st1e,
|
||||
output wire fill_saw_dirty_st1e
|
||||
|
||||
output wire fill_saw_dirty_st1e
|
||||
);
|
||||
|
||||
|
||||
reg[`DBANK_LINE_WORDS-1:0][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_WORDS-1:0][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;
|
||||
@@ -98,9 +93,9 @@ module VX_tag_data_access
|
||||
|
||||
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),
|
||||
@@ -119,14 +114,12 @@ module VX_tag_data_access
|
||||
.LLVQ_SIZE (LLVQ_SIZE),
|
||||
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
|
||||
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
|
||||
)
|
||||
VX_tag_data_structure
|
||||
(
|
||||
) vx_tag_data_structure (
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.stall_bank_pipe(stall_bank_pipe),
|
||||
|
||||
.read_addr (readaddr_st10),
|
||||
.read_addr (readaddr_st10[`LINE_SELECT_ADDR_RNG]),
|
||||
.read_valid (qual_read_valid_st1),
|
||||
.read_dirty (qual_read_dirty_st1),
|
||||
.read_tag (qual_read_tag_st1),
|
||||
@@ -135,13 +128,17 @@ module VX_tag_data_access
|
||||
.invalidate (invalidate_line),
|
||||
.write_enable(use_write_enable),
|
||||
.write_fill (real_writefill),
|
||||
.write_addr (writeaddr_st1e),
|
||||
.write_addr (writeaddr_st1e[`LINE_SELECT_ADDR_RNG]),
|
||||
.tag_index (writeaddr_st1e[`TAG_SELECT_ADDR_RNG]),
|
||||
.write_data (use_write_data),
|
||||
.fill_sent (fill_sent)
|
||||
);
|
||||
);
|
||||
|
||||
// VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_WORDS*32) )) s0_1_c0 (
|
||||
VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_WORDS*32) ), .Valid(0)) s0_1_c0 (
|
||||
VX_generic_register #(
|
||||
.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_WORDS*32) ),
|
||||
.PassThru(1)
|
||||
) s0_1_c0 (
|
||||
.clk (clk),
|
||||
.reset(reset),
|
||||
.stall(stall),
|
||||
@@ -153,7 +150,9 @@ module VX_tag_data_access
|
||||
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_WORDS*32) )) s0_1_cc (
|
||||
VX_generic_register #(
|
||||
.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_WORDS*32))
|
||||
) s0_1_cc (
|
||||
.clk (clk),
|
||||
.reset(reset),
|
||||
.stall(stall),
|
||||
@@ -164,7 +163,6 @@ module VX_tag_data_access
|
||||
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
|
||||
@@ -178,6 +176,7 @@ module VX_tag_data_access
|
||||
wire[`OFFSET_SIZE_RNG] byte_select = writeaddr_st1e[`OFFSET_ADDR_RNG];
|
||||
wire[`WORD_SELECT_SIZE_RNG] block_offset = writeaddr_st1e[`WORD_SELECT_ADDR_RNG];
|
||||
|
||||
/* verilator lint_off UNUSED */
|
||||
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);
|
||||
@@ -187,49 +186,15 @@ module VX_tag_data_access
|
||||
wire b0 = (byte_select == 0);
|
||||
wire b1 = (byte_select == 1);
|
||||
wire b2 = (byte_select == 2);
|
||||
wire b3 = (byte_select == 3);
|
||||
wire b3 = (byte_select == 3);
|
||||
/* verilator lint_on UNUSED */
|
||||
|
||||
/* verilator lint_off UNUSED */
|
||||
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;
|
||||
/* verilator lint_on UNUSED */
|
||||
|
||||
/////////////////////// STORE LOGIC ///////////////////
|
||||
|
||||
@@ -245,6 +210,7 @@ module VX_tag_data_access
|
||||
|
||||
wire[`DBANK_LINE_WORDS-1:0][3:0] we;
|
||||
wire[`DBANK_LINE_WORDS-1:0][31:0] data_write;
|
||||
|
||||
genvar g;
|
||||
generate
|
||||
for (g = 0; g < `DBANK_LINE_WORDS; g = g + 1) begin : write_enables
|
||||
@@ -257,9 +223,18 @@ module VX_tag_data_access
|
||||
(normal_write && sh) ? sh_mask :
|
||||
4'b0000;
|
||||
|
||||
if (!(FUNC_ID == `L2FUNC_ID)) assign data_write[g] = force_write ? writedata_st1e[g] : use_write_dat;
|
||||
if (FUNC_ID != `L2FUNC_ID) begin
|
||||
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] sw_data = 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;
|
||||
assign data_write[g] = force_write ? writedata_st1e[g] : use_write_dat;
|
||||
end
|
||||
end
|
||||
if ((FUNC_ID == `L2FUNC_ID)) begin
|
||||
if (FUNC_ID == `L2FUNC_ID) begin
|
||||
assign data_write = force_write ? writedata_st1e : writeword_st1e;
|
||||
end
|
||||
endgenerate
|
||||
@@ -268,13 +243,29 @@ module VX_tag_data_access
|
||||
assign use_write_data = data_write;
|
||||
|
||||
///////////////////////
|
||||
|
||||
if (FUNC_ID == `L2FUNC_ID) begin
|
||||
assign readword_st1e = read_data_st1c[STAGE_1_CYCLES-1];
|
||||
end else begin
|
||||
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] data_Qual = lb ? lb_data :
|
||||
lh ? lh_data :
|
||||
lhu ? lhu_data :
|
||||
lbu ? lbu_data :
|
||||
lw_data;
|
||||
|
||||
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;
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
`include "VX_cache_config.vh"
|
||||
|
||||
module VX_tag_data_structure
|
||||
#(
|
||||
module VX_tag_data_structure #(
|
||||
// Size of cache in bytes
|
||||
parameter CACHE_SIZE_BYTES = 1024,
|
||||
// Size of line inside a bank in bytes
|
||||
@@ -17,8 +16,7 @@ module VX_tag_data_structure
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 0,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Request Queue Size
|
||||
parameter REQQ_SIZE = 8,
|
||||
// Miss Reserv Queue Knob
|
||||
@@ -28,7 +26,7 @@ module VX_tag_data_structure
|
||||
// Snoop Req Queue
|
||||
parameter SNRQ_SIZE = 8,
|
||||
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
|
||||
// Core Writeback Queue Size
|
||||
parameter CWBQ_SIZE = 8,
|
||||
// Dram Writeback Queue Size
|
||||
@@ -41,44 +39,37 @@ module VX_tag_data_structure
|
||||
// Fill Invalidator Size {Fill invalidator must be active}
|
||||
parameter FILL_INVALIDAOR_SIZE = 16,
|
||||
|
||||
// Dram knobs
|
||||
// Dram knobs
|
||||
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
input wire stall_bank_pipe,
|
||||
|
||||
|
||||
)
|
||||
(
|
||||
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,
|
||||
input wire[`LINE_SELECT_SIZE_RNG] read_addr,
|
||||
output wire read_valid,
|
||||
output wire read_dirty,
|
||||
output wire[`TAG_SELECT_SIZE_RNG] read_tag,
|
||||
output wire[`DBANK_LINE_WORDS-1:0][31:0] read_data,
|
||||
|
||||
input wire invalidate,
|
||||
input wire[`DBANK_LINE_WORDS-1:0][3:0] write_enable,
|
||||
input wire write_fill,
|
||||
input wire[31:0] write_addr,
|
||||
input wire invalidate,
|
||||
input wire[`DBANK_LINE_WORDS-1:0][3:0] write_enable,
|
||||
input wire write_fill,
|
||||
input wire[`LINE_SELECT_SIZE_RNG] write_addr,
|
||||
input wire[`TAG_SELECT_SIZE_RNG] tag_index,
|
||||
input wire[`DBANK_LINE_WORDS-1:0][31:0] write_data,
|
||||
input wire fill_sent
|
||||
|
||||
input wire fill_sent
|
||||
);
|
||||
|
||||
reg[`DBANK_LINE_WORDS-1:0][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];
|
||||
reg [`DBANK_LINE_WORDS-1:0][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]];
|
||||
assign read_valid = valid [read_addr];
|
||||
assign read_dirty = dirty [read_addr];
|
||||
assign read_tag = tag [read_addr];
|
||||
assign read_data = data [read_addr];
|
||||
|
||||
wire going_to_write = (|write_enable);
|
||||
|
||||
@@ -94,27 +85,27 @@ module VX_tag_data_structure
|
||||
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];
|
||||
valid[write_addr] <= 1;
|
||||
tag [write_addr] <= tag_index;
|
||||
if (write_fill) begin
|
||||
dirty[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
|
||||
dirty[write_addr] <= 0;
|
||||
end else begin
|
||||
dirty[write_addr[`LINE_SELECT_ADDR_RNG]] <= 1;
|
||||
dirty[write_addr] <= 1;
|
||||
end
|
||||
end else if (fill_sent) begin
|
||||
dirty[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
|
||||
// valid[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
|
||||
dirty[write_addr] <= 0;
|
||||
// valid[write_addr] <= 0;
|
||||
end
|
||||
|
||||
if (invalidate) begin
|
||||
valid[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
|
||||
valid[write_addr] <= 0;
|
||||
end
|
||||
|
||||
for (f = 0; f < `DBANK_LINE_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];
|
||||
if (write_enable[f][0]) data[write_addr][f][0] <= write_data[f][7 :0 ];
|
||||
if (write_enable[f][1]) data[write_addr][f][1] <= write_data[f][15:8 ];
|
||||
if (write_enable[f][2]) data[write_addr][f][2] <= write_data[f][23:16];
|
||||
if (write_enable[f][3]) data[write_addr][f][3] <= write_data[f][31:24];
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
Reference in New Issue
Block a user