rtl cache refactory

This commit is contained in:
Blaise Tine
2020-04-30 17:12:18 -04:00
parent 814ac50d12
commit a1dc90b951
67 changed files with 51076 additions and 51059 deletions

492
hw/rtl/cache/VX_bank.v vendored
View File

@@ -2,13 +2,13 @@
`include "VX_define.vh"
module VX_bank #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -39,63 +39,56 @@ module VX_bank #(
parameter FFSQ_SIZE = 8,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1
) (
input wire clk,
input wire reset,
// Input Core Request
// Core Request
input wire core_req_ready,
input wire [NUM_REQUESTS-1:0] core_req_valids,
input wire [NUM_REQUESTS-1:0][2:0] core_req_read,
input wire [NUM_REQUESTS-1:0][2:0] core_req_write,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_read,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_write,
input wire [NUM_REQUESTS-1:0][31:0] core_req_addr,
input wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] core_req_data,
input wire [4:0] core_req_rd,
input wire [NUM_REQUESTS-1:0][1:0] core_req_wb,
input wire [31:0] core_req_pc,
input wire [`NW_BITS-1:0] core_req_warp_num,
input wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_req_data,
input wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_req_tag,
output wire core_req_full,
// Output Core WB
// Core Response
output wire core_rsp_valid,
output wire [`LOG2UP(NUM_REQUESTS)-1:0] core_rsp_tid,
output wire [4:0] core_rsp_rd,
output wire [1:0] core_rsp_wb,
output wire [`NW_BITS-1:0] core_rsp_warp_num,
output wire [`WORD_SIZE_RNG] core_rsp_data,
output wire [31:0] core_rsp_pc,
output wire [31:0] core_rsp_addr,
output wire [`WORD_WIDTH-1:0] core_rsp_data,
output wire [CORE_TAG_WIDTH-1:0] core_rsp_tag,
input wire core_rsp_pop,
// Dram Fill Requests
output wire dram_fill_req_valid,
output wire[31:0] dram_fill_req_addr,
output wire[`LINE_ADDR_WIDTH-1:0] dram_fill_req_addr,
output wire dram_fill_req_is_snp,
input wire dram_fill_req_full,
// Dram Fill Response
input wire dram_fill_rsp_valid,
input wire [31:0] dram_fill_rsp_addr,
input wire [`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] dram_fill_rsp_data,
input wire dram_fill_rsp_valid,
input wire [`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] dram_fill_rsp_data,
input wire [`LINE_ADDR_WIDTH-1:0] dram_fill_rsp_addr,
output wire dram_fill_rsp_ready,
// Dram WB Requests
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,
output wire [`LINE_ADDR_WIDTH-1:0] dram_wb_req_addr,
output wire [`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] dram_wb_req_data,
input wire dram_wb_req_pop,
// Snp Request
input wire snp_req_valid,
input wire [31:0] snp_req_addr,
input wire [`LINE_ADDR_WIDTH-1:0] snp_req_addr,
output wire snp_req_full,
output wire snp_fwd_valid,
output wire [31:0] snp_fwd_addr,
output wire [`LINE_ADDR_WIDTH-1:0] snp_fwd_addr,
input wire snp_fwd_pop
);
@@ -109,17 +102,16 @@ module VX_bank #(
end
end
wire snrq_pop;
wire snrq_empty;
wire snrq_valid_st0;
wire[31:0] snrq_addr_st0;
wire[`LINE_ADDR_WIDTH-1:0] snrq_addr_st0;
assign snrq_valid_st0 = !snrq_empty;
VX_generic_queue #(
.DATAW(32),
.DATAW($bits(snp_req_addr)),
.SIZE(SNRQ_SIZE)
) snr_queue (
.clk (clk),
@@ -132,16 +124,16 @@ module VX_bank #(
.full (snp_req_full)
);
wire dfpq_pop;
wire dfpq_empty;
wire dfpq_full;
wire[31:0] dfpq_addr_st0;
wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] dfpq_filldata_st0;
wire dfpq_pop;
wire dfpq_empty;
wire dfpq_full;
wire [`LINE_ADDR_WIDTH-1:0] dfpq_addr_st0;
wire [`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] dfpq_filldata_st0;
assign dram_fill_rsp_ready = !dfpq_full;
VX_generic_queue #(
.DATAW(32 + (`BANK_LINE_WORDS*`WORD_SIZE)),
.DATAW($bits(dram_fill_rsp_addr) + $bits(dram_fill_rsp_data)),
.SIZE(DFPQ_SIZE)
) dfp_queue (
.clk (clk),
@@ -159,34 +151,33 @@ module VX_bank #(
wire reqq_empty;
wire reqq_req_st0;
wire[`LOG2UP(NUM_REQUESTS)-1:0] reqq_req_tid_st0;
`IGNORE_WARNINGS_BEGIN
wire [31:0] reqq_req_addr_st0;
wire [`WORD_SIZE_RNG] reqq_req_writeword_st0;
wire [4:0] reqq_req_rd_st0;
wire [1:0] reqq_req_wb_st0;
wire [`NW_BITS-1:0] reqq_req_warp_num_st0;
wire [2:0] reqq_req_mem_read_st0;
wire [2:0] reqq_req_mem_write_st0;
wire [31:0] reqq_req_pc_st0;
`IGNORE_WARNINGS_END
wire [`WORD_WIDTH-1:0] reqq_req_writeword_st0;
wire [CORE_TAG_WIDTH-1:0] reqq_req_tag_st0;
wire [`WORD_SEL_BITS-1:0] reqq_req_mem_read_st0;
wire [`WORD_SEL_BITS-1:0] reqq_req_mem_write_st0;
assign reqq_push = core_req_ready && (|core_req_valids);
VX_cache_req_queue #(
.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)
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
) req_queue (
.clk (clk),
.reset (reset),
@@ -195,10 +186,7 @@ module VX_bank #(
.bank_valids (core_req_valids),
.bank_addr (core_req_addr),
.bank_writedata (core_req_data),
.bank_rd (core_req_rd),
.bank_pc (core_req_pc),
.bank_wb (core_req_wb),
.bank_warp_num (core_req_warp_num),
.bank_tag (core_req_tag),
.bank_mem_read (core_req_read),
.bank_mem_write (core_req_write),
@@ -208,12 +196,9 @@ module VX_bank #(
.reqq_req_tid_st0 (reqq_req_tid_st0),
.reqq_req_addr_st0 (reqq_req_addr_st0),
.reqq_req_writedata_st0(reqq_req_writeword_st0),
.reqq_req_rd_st0 (reqq_req_rd_st0),
.reqq_req_wb_st0 (reqq_req_wb_st0),
.reqq_req_warp_num_st0 (reqq_req_warp_num_st0),
.reqq_req_tag_st0 (reqq_req_tag_st0),
.reqq_req_mem_read_st0 (reqq_req_mem_read_st0),
.reqq_req_mem_write_st0(reqq_req_mem_write_st0),
.reqq_req_pc_st0 (reqq_req_pc_st0),
.reqq_empty (reqq_empty),
.reqq_full (core_req_full)
);
@@ -223,92 +208,33 @@ module VX_bank #(
wire mrvq_stop;
wire mrvq_valid_st0;
wire[`LOG2UP(NUM_REQUESTS)-1:0] mrvq_tid_st0;
wire [31:0] mrvq_addr_st0;
wire [`WORD_SIZE_RNG] mrvq_writeword_st0;
wire [4:0] mrvq_rd_st0;
wire [1:0] mrvq_wb_st0;
wire [31:0] miss_resrv_pc_st0;
wire [`NW_BITS-1:0] mrvq_warp_num_st0;
wire [2:0] mrvq_mem_read_st0;
wire [2:0] mrvq_mem_write_st0;
wire [`LINE_ADDR_WIDTH-1:0] mrvq_addr_st0;
wire [`BASE_ADDR_BITS-1:0] mrvq_wsel_st0;
wire [`WORD_WIDTH-1:0] mrvq_writeword_st0;
wire [CORE_TAG_WIDTH-1:0] mrvq_tag_st0;
wire [`WORD_SEL_BITS-1:0] mrvq_mem_read_st0;
wire [`WORD_SEL_BITS-1:0] mrvq_mem_write_st0;
wire miss_add;
wire[31:0] miss_add_addr;
wire[`WORD_SIZE_RNG] miss_add_data;
wire[`LINE_ADDR_WIDTH-1:0] miss_add_addr;
wire[`BASE_ADDR_BITS-1:0] miss_add_wsel;
wire[`WORD_WIDTH-1:0] miss_add_data;
wire[`LOG2UP(NUM_REQUESTS)-1:0] miss_add_tid;
wire[4:0] miss_add_rd;
wire[1:0] miss_add_wb;
wire[`NW_BITS-1:0] miss_add_warp_num;
wire[2:0] miss_add_mem_read;
wire[2:0] miss_add_mem_write;
wire[CORE_TAG_WIDTH-1:0] miss_add_tag;
wire[`WORD_SEL_BITS-1:0] miss_add_mem_read;
wire[`WORD_SEL_BITS-1:0] miss_add_mem_write;
wire[31:0] miss_add_pc;
wire[31:0] addr_st2;
wire[`LINE_ADDR_WIDTH-1:0] addr_st2;
wire is_fill_st2;
VX_cache_miss_resrv #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.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)
) mrvq_queue (
.clk (clk),
.reset (reset),
// Enqueue
.miss_add (miss_add), // Need to do all
.miss_add_addr (miss_add_addr),
.miss_add_data (miss_add_data),
.miss_add_tid (miss_add_tid),
.miss_add_rd (miss_add_rd),
.miss_add_wb (miss_add_wb),
.miss_add_warp_num (miss_add_warp_num),
.miss_add_mem_read (miss_add_mem_read),
.miss_add_mem_write (miss_add_mem_write),
.miss_add_pc (miss_add_pc),
.miss_resrv_full (mrvq_full),
.miss_resrv_stop (mrvq_stop),
// Broadcast
.is_fill_st1 (is_fill_st2),
.fill_addr_st1 (addr_st2),
// Dequeue
.miss_resrv_pop (mrvq_pop),
.miss_resrv_valid_st0 (mrvq_valid_st0),
.miss_resrv_addr_st0 (mrvq_addr_st0),
.miss_resrv_data_st0 (mrvq_writeword_st0),
.miss_resrv_tid_st0 (mrvq_tid_st0),
.miss_resrv_rd_st0 (mrvq_rd_st0),
.miss_resrv_wb_st0 (mrvq_wb_st0),
.miss_resrv_pc_st0 (miss_resrv_pc_st0),
.miss_resrv_warp_num_st0 (mrvq_warp_num_st0),
.miss_resrv_mem_read_st0 (mrvq_mem_read_st0),
.miss_resrv_mem_write_st0(mrvq_mem_write_st0)
);
wire stall_bank_pipe;
reg is_fill_in_pipe;
wire valid_st1 [STAGE_1_CYCLES-1:0];
wire is_fill_st1 [STAGE_1_CYCLES-1:0];
`DEBUG_BEGIN
wire going_to_write_st1[STAGE_1_CYCLES-1:0];
`DEBUG_END
wire [31:0] addr_st1 [STAGE_1_CYCLES-1:0];
integer i;
always @(*) begin
is_fill_in_pipe = 0;
@@ -323,8 +249,6 @@ module VX_bank #(
end
end
// assign is_fill_in_pipe = (|is_fill_st1) || is_fill_st2;
assign mrvq_pop = mrvq_valid_st0 && !stall_bank_pipe;
assign dfpq_pop = !mrvq_pop && !dfpq_empty && !stall_bank_pipe;
assign reqq_pop = !mrvq_stop && !mrvq_pop && !dfpq_pop && !reqq_empty && reqq_req_st0 && !stall_bank_pipe && !is_fill_st1[0] && !is_fill_in_pipe;
@@ -332,123 +256,114 @@ module VX_bank #(
wire qual_is_fill_st0;
wire qual_valid_st0;
wire [31:0] qual_addr_st0;
wire [`WORD_SIZE_RNG] qual_writeword_st0;
wire [`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] qual_writedata_st0;
wire [`REQ_INST_META_SIZE-1:0] qual_inst_meta_st0;
wire [`LINE_ADDR_WIDTH-1:0] qual_addr_st0;
wire [`WORD_SELECT_ADDR_END:0] qual_wsel_st0;
wire [`WORD_WIDTH-1:0] qual_writeword_st0;
wire [`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] qual_writedata_st0;
wire [`REQ_INST_META_WIDTH-1:0] qual_inst_meta_st0;
wire qual_going_to_write_st0;
wire qual_is_snp;
wire [31:0] qual_pc_st0;
wire [`WORD_SIZE_RNG] writeword_st1 [STAGE_1_CYCLES-1:0];
wire [`REQ_INST_META_SIZE-1:0] inst_meta_st1 [STAGE_1_CYCLES-1:0];
wire [`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] writedata_st1[STAGE_1_CYCLES-1:0];
wire valid_st1 [STAGE_1_CYCLES-1:0];
wire [`LINE_ADDR_WIDTH-1:0] addr_st1 [STAGE_1_CYCLES-1:0];
wire [`WORD_SELECT_ADDR_END:0] wsel_st1 [STAGE_1_CYCLES-1:0];
wire [`WORD_WIDTH-1:0] writeword_st1 [STAGE_1_CYCLES-1:0];
wire [`REQ_INST_META_WIDTH-1:0] inst_meta_st1 [STAGE_1_CYCLES-1:0];
wire [`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] writedata_st1[STAGE_1_CYCLES-1:0];
wire is_snp_st1 [STAGE_1_CYCLES-1:0];
wire [31:0] pc_st1 [STAGE_1_CYCLES-1:0];
assign qual_is_fill_st0 = dfpq_pop;
// always @(*) begin
// if (qual_is_fill_st0 && (FUNC_ID == 3)) begin
// $display("WHAT THE FUCK FUNC_ID: %x", FUNC_ID);
// end
// end
assign qual_valid_st0 = dfpq_pop || mrvq_pop || reqq_pop || snrq_pop;
assign qual_addr_st0 = dfpq_pop ? dfpq_addr_st0 :
mrvq_pop ? mrvq_addr_st0 :
reqq_pop ? reqq_req_addr_st0 :
reqq_pop ? reqq_req_addr_st0[31:`LINE_SELECT_ADDR_START] :
snrq_pop ? snrq_addr_st0 :
0;
0;
assign qual_wsel_st0 = reqq_pop ? reqq_req_addr_st0[`BASE_ADDR_BITS-1:0] :
mrvq_pop ? mrvq_wsel_st0 :
0;
assign qual_writedata_st0 = dfpq_pop ? dfpq_filldata_st0 : 57;
assign qual_inst_meta_st0 = mrvq_pop ? {mrvq_rd_st0 , mrvq_wb_st0 , mrvq_warp_num_st0 , mrvq_mem_read_st0 , mrvq_mem_write_st0 , mrvq_tid_st0 } :
reqq_pop ? {reqq_req_rd_st0, reqq_req_wb_st0, reqq_req_warp_num_st0, reqq_req_mem_read_st0, reqq_req_mem_write_st0, reqq_req_tid_st0} :
0;
assign qual_inst_meta_st0 = mrvq_pop ? {mrvq_tag_st0 , mrvq_mem_read_st0, mrvq_mem_write_st0, mrvq_tid_st0} :
reqq_pop ? {reqq_req_tag_st0, reqq_req_mem_read_st0, reqq_req_mem_write_st0, reqq_req_tid_st0} :
0;
assign qual_going_to_write_st0 = dfpq_pop ? 1 :
(mrvq_pop && (mrvq_mem_write_st0 != `NO_MEM_WRITE)) ? 1 :
(reqq_pop && (reqq_req_mem_write_st0 != `NO_MEM_WRITE)) ? 1 :
(mrvq_pop && (mrvq_mem_write_st0 != `WORD_SEL_NO)) ? 1 :
(reqq_pop && (reqq_req_mem_write_st0 != `WORD_SEL_NO)) ? 1 :
(snrq_pop) ? 1 :
0;
assign qual_pc_st0 = (reqq_pop) ? reqq_req_pc_st0 :
(mrvq_pop) ? miss_resrv_pc_st0 :
(dfpq_pop) ? 32'hdeadbeef :
(snrq_pop) ? 32'hb00b0000 :
32'h0;
assign qual_is_snp = snrq_pop ? 1 : 0;
assign qual_is_snp = snrq_pop ? 1 : 0;
assign qual_writeword_st0 = mrvq_pop ? mrvq_writeword_st0 :
reqq_pop ? reqq_req_writeword_st0 :
0;
VX_generic_register #(
.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_WORDS*`WORD_SIZE) + 1 + 32)
.N(1 + 1 + 1 + `LINE_ADDR_WIDTH + `BASE_ADDR_BITS + `WORD_WIDTH + `REQ_INST_META_WIDTH + 1 + (`BANK_LINE_WORDS*`WORD_WIDTH))
) 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]})
.in ({qual_is_snp, qual_going_to_write_st0, qual_valid_st0, qual_addr_st0, qual_wsel_st0, qual_writeword_st0, qual_inst_meta_st0, qual_is_fill_st0, qual_writedata_st0}),
.out ({is_snp_st1[0], going_to_write_st1[0], valid_st1[0], addr_st1[0], wsel_st1[0], writeword_st1[0], inst_meta_st1[0], is_fill_st1[0], writedata_st1[0]})
);
genvar curr_stage;
generate
for (curr_stage = 1; curr_stage < STAGE_1_CYCLES; curr_stage = curr_stage + 1) begin
VX_generic_register #(.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_WORDS*`WORD_SIZE) + 1 + 32)) s0_1_cc (
genvar stage;
for (stage = 1; stage < STAGE_1_CYCLES; stage = stage + 1) begin
VX_generic_register #(
.N(1 + 1 + 1 + `LINE_ADDR_WIDTH + `BASE_ADDR_BITS + `WORD_WIDTH + `REQ_INST_META_WIDTH + 1 + (`BANK_LINE_WORDS*`WORD_WIDTH))
) s0_1_cc (
.clk (clk),
.reset(reset),
.stall(stall_bank_pipe),
.flush(0),
.in ({is_snp_st1[curr_stage-1], going_to_write_st1[curr_stage-1], valid_st1[curr_stage-1], addr_st1[curr_stage-1], writeword_st1[curr_stage-1], inst_meta_st1[curr_stage-1], is_fill_st1[curr_stage-1] , writedata_st1[curr_stage-1], pc_st1[curr_stage-1]}),
.out ({is_snp_st1[curr_stage] , going_to_write_st1[curr_stage] , valid_st1[curr_stage] , addr_st1[curr_stage] , writeword_st1[curr_stage] , inst_meta_st1[curr_stage] , is_fill_st1[curr_stage] , writedata_st1[curr_stage] , pc_st1[curr_stage]})
);
end
endgenerate
.in ({is_snp_st1[stage-1], going_to_write_st1[stage-1], valid_st1[stage-1], addr_st1[stage-1], wsel_st1[stage-1], writeword_st1[stage-1], inst_meta_st1[stage-1], is_fill_st1[stage-1], writedata_st1[stage-1]}),
.out ({is_snp_st1[stage], going_to_write_st1[stage], valid_st1[stage], addr_st1[stage], wsel_st1[stage], writeword_st1[stage], inst_meta_st1[stage], is_fill_st1[stage], writedata_st1[stage]})
);
end
wire[`WORD_SIZE_RNG] readword_st1e;
wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] readdata_st1e;
wire[`WORD_WIDTH-1:0] readword_st1e;
wire[`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] readdata_st1e;
wire[`TAG_SELECT_BITS-1:0] readtag_st1e;
wire miss_st1e;
wire dirty_st1e;
wire[31:0] pc_st1e;
`DEBUG_BEGIN
wire [4:0] rd_st1e;
wire [1:0] wb_st1e;
wire [`NW_BITS-1:0] warp_num_st1e;
wire [CORE_TAG_WIDTH-1:0] tag_st1e;
wire [`LOG2UP(NUM_REQUESTS)-1:0] tid_st1e;
`DEBUG_END
wire [2:0] mem_read_st1e;
wire [2:0] mem_write_st1e;
wire [`WORD_SEL_BITS-1:0] mem_read_st1e;
wire [`WORD_SEL_BITS-1:0] mem_write_st1e;
wire fill_saw_dirty_st1e;
wire is_snp_st1e;
assign is_snp_st1e = is_snp_st1[STAGE_1_CYCLES-1];
assign pc_st1e = pc_st1[STAGE_1_CYCLES-1];
assign {rd_st1e, wb_st1e, warp_num_st1e, mem_read_st1e, mem_write_st1e, tid_st1e} = inst_meta_st1[STAGE_1_CYCLES-1];
assign {tag_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),
.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),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
VX_tag_data_access #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE)
) tag_data_access (
.clk (clk),
.reset (reset),
@@ -456,12 +371,13 @@ module VX_bank #(
.stall_bank_pipe(stall_bank_pipe),
// Initial Read
.readaddr_st10 (addr_st1[0]),
.readaddr_st10 (addr_st1[0][`LINE_SELECT_BITS-1:0]),
// Actual Read/Write
.valid_req_st1e(valid_st1[STAGE_1_CYCLES-1]),
.writefill_st1e(is_fill_st1[STAGE_1_CYCLES-1]),
.writeaddr_st1e(addr_st1[STAGE_1_CYCLES-1]),
.writewsel_st1e(wsel_st1[STAGE_1_CYCLES-1]),
.writeword_st1e(writeword_st1[STAGE_1_CYCLES-1]),
.writedata_st1e(writedata_st1[STAGE_1_CYCLES-1]),
@@ -482,26 +398,26 @@ module VX_bank #(
wire qual_valid_st1e_2 = valid_st1[STAGE_1_CYCLES-1] && !is_fill_st1[STAGE_1_CYCLES-1];
wire valid_st2;
wire[`WORD_SIZE_RNG] writeword_st2;
wire[`WORD_SIZE_RNG] readword_st2;
wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] readdata_st2;
wire [`BASE_ADDR_BITS-1:0] wsel_st2;
wire [`WORD_WIDTH-1:0] writeword_st2;
wire [`WORD_WIDTH-1:0] readword_st2;
wire [`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] readdata_st2;
wire miss_st2;
wire dirty_st2;
wire[`REQ_INST_META_SIZE-1:0] inst_meta_st2;
wire[`TAG_SELECT_BITS-1:0] readtag_st2;
wire [`REQ_INST_META_WIDTH-1:0] inst_meta_st2;
wire [`TAG_SELECT_BITS-1:0] readtag_st2;
wire fill_saw_dirty_st2;
wire is_snp_st2;
wire [31:0] pc_st2;
VX_generic_register #(
.N( 1+1+1+1+32+`WORD_SIZE+`WORD_SIZE+(`BANK_LINE_WORDS * `WORD_SIZE) + `REQ_INST_META_SIZE + `TAG_SELECT_BITS + 32 + 2)
.N(1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `BASE_ADDR_BITS + `WORD_WIDTH + `WORD_WIDTH + (`BANK_LINE_WORDS * `WORD_WIDTH) + `TAG_SELECT_BITS + 1 + 1 + `REQ_INST_META_WIDTH)
) 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 })
.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], wsel_st1[STAGE_1_CYCLES-1], writeword_st1[STAGE_1_CYCLES-1], readword_st1e, readdata_st1e, readtag_st1e, miss_st1e, dirty_st1e, inst_meta_st1[STAGE_1_CYCLES-1]}),
.out ({is_snp_st2 , fill_saw_dirty_st2 , is_fill_st2 , valid_st2 , addr_st2 , wsel_st2, writeword_st2 , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , inst_meta_st2 })
);
wire should_flush;
@@ -514,73 +430,119 @@ module VX_bank #(
// Enqueue to miss reserv if it's a valid miss
assign miss_add = valid_st2 && !is_snp_st2 && miss_st2 && !mrvq_full && !(should_flush && dwbq_push) && !((is_snp_st2 && valid_st2 && ffsq_full) ||((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
assign miss_add_pc = pc_st2;
assign miss_add_addr = addr_st2;
assign miss_add_wsel = wsel_st2;
assign miss_add_data = writeword_st2;
assign {miss_add_rd, miss_add_wb, miss_add_warp_num, miss_add_mem_read, miss_add_mem_write, miss_add_tid} = inst_meta_st2;
assign {miss_add_tag, miss_add_mem_read, miss_add_mem_write, miss_add_tid} = inst_meta_st2;
VX_cache_miss_resrv #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
) cache_miss_resrv (
.clk (clk),
.reset (reset),
// Enqueue
.miss_add (miss_add), // Need to do all
.miss_add_addr (miss_add_addr),
.miss_add_wsel (miss_add_wsel),
.miss_add_data (miss_add_data),
.miss_add_tid (miss_add_tid),
.miss_add_tag (miss_add_tag),
.miss_add_mem_read (miss_add_mem_read),
.miss_add_mem_write (miss_add_mem_write),
.miss_resrv_full (mrvq_full),
.miss_resrv_stop (mrvq_stop),
// Broadcast
.is_fill_st1 (is_fill_st2),
.fill_addr_st1 (addr_st2),
// Dequeue
.miss_resrv_pop (mrvq_pop),
.miss_resrv_valid_st0 (mrvq_valid_st0),
.miss_resrv_addr_st0 (mrvq_addr_st0),
.miss_resrv_wsel_st0 (mrvq_wsel_st0),
.miss_resrv_data_st0 (mrvq_writeword_st0),
.miss_resrv_tid_st0 (mrvq_tid_st0),
.miss_resrv_tag_st0 (mrvq_tag_st0),
.miss_resrv_mem_read_st0 (mrvq_mem_read_st0),
.miss_resrv_mem_write_st0(mrvq_mem_write_st0)
);
// 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_full));
wire [`WORD_SIZE_RNG] cwbq_data = readword_st2;
wire [`LOG2UP(NUM_REQUESTS)-1:0] cwbq_tid = miss_add_tid;
wire [4:0] cwbq_rd = miss_add_rd;
wire [1:0] cwbq_wb = miss_add_wb;
wire [`NW_BITS-1:0] cwbq_warp_num = miss_add_warp_num;
wire [31:0] cwbq_pc = pc_st2;
wire cwbq_push = (valid_st2 && !miss_st2) && !cwbq_full && !((FUNC_ID == `L2FUNC_ID) && (miss_add_mem_write == `WORD_SEL_NO)) && !((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_full));
wire [`WORD_WIDTH-1:0] cwbq_data = readword_st2;
wire [`LOG2UP(NUM_REQUESTS)-1:0] cwbq_tid = miss_add_tid;
wire [CORE_TAG_WIDTH-1:0] cwbq_tag = miss_add_tag;
wire cwbq_empty;
wire cwbq_empty;
assign core_rsp_valid = !cwbq_empty;
VX_generic_queue #(
.DATAW(`LOG2UP(NUM_REQUESTS) + 5 + 2 + (`NW_BITS-1+1) + `WORD_SIZE + 32 + 32),
.DATAW(`LOG2UP(NUM_REQUESTS) + CORE_TAG_WIDTH + `WORD_WIDTH),
.SIZE(CWBQ_SIZE)
) cwb_queue(
) cwb_queue (
.clk (clk),
.reset (reset),
.push (cwbq_push),
.data_in ({cwbq_tid, cwbq_rd, cwbq_wb, cwbq_warp_num, cwbq_data, cwbq_pc, addr_st2}),
.data_in ({cwbq_tid, cwbq_tag, cwbq_data}),
.pop (core_rsp_pop),
.data_out({core_rsp_tid, core_rsp_rd, core_rsp_wb, core_rsp_warp_num, core_rsp_data, core_rsp_pc, core_rsp_addr}),
.data_out({core_rsp_tid, core_rsp_tag, core_rsp_data}),
.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;
assign should_flush = snoop_state && valid_st2 && (miss_add_mem_write != `WORD_SEL_NO) && !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_full));
wire[31:0] dwbq_req_addr;
wire[`LINE_ADDR_WIDTH-1:0] dwbq_req_addr;
wire dwbq_empty;
wire[`BANK_LINE_WORDS-1:0][`WORD_SIZE-1:0] dwbq_req_data;
wire[`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] dwbq_req_data;
if ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID)) begin
assign dwbq_req_data = (should_flush && dwbq_push) ? writeword_st2 : readdata_st2;
assign dwbq_req_addr = (should_flush && dwbq_push) ? (addr_st2) : ({readtag_st2, addr_st2[`LINE_SELECT_ADDR_END:0]} & `BASE_ADDR_MASK);
assign dwbq_req_addr = (should_flush && dwbq_push) ? (addr_st2) : {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
end else begin
assign dwbq_req_data = readdata_st2;
assign dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_ADDR_END:0]} & `BASE_ADDR_MASK;
assign dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
end
wire possible_fill = valid_st2 && miss_st2 && !dram_fill_req_full && !is_snp_st2;
wire[31:0] fill_invalidator_addr = addr_st2 & `BASE_ADDR_MASK;
wire [`LINE_ADDR_WIDTH-1:0] fill_invalidator_addr = addr_st2;
VX_fill_invalidator #(
.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_fill_invalidator #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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)
) fill_invalidator (
.clk (clk),
.reset (reset),
@@ -594,12 +556,12 @@ module VX_bank #(
// 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_fill_req_addr = addr_st2;
assign dram_wb_req_valid = !dwbq_empty;
VX_generic_queue #(
.DATAW(32 + (`BANK_LINE_WORDS * `WORD_SIZE)),
.DATAW(`LINE_ADDR_WIDTH + (`BANK_LINE_WORDS * `WORD_WIDTH)),
.SIZE(DWBQ_SIZE)
) dwb_queue (
.clk (clk),
@@ -621,15 +583,15 @@ module VX_bank #(
assign snp_fwd_valid = !ffsq_empty;
VX_generic_queue #(
.DATAW(32),
.DATAW(`LINE_ADDR_WIDTH),
.SIZE(FFSQ_SIZE)
) ffs_queue (
.clk (clk),
.reset (reset),
.push (snp_fwd_push),
.data_in ({addr_st2}),
.data_in (addr_st2),
.pop (snp_fwd_pop),
.data_out({snp_fwd_addr}),
.data_out(snp_fwd_addr),
.empty (ffsq_empty),
.full (ffsq_full)
);

View File

@@ -2,13 +2,13 @@
module VX_cache #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 16,
parameter WORD_SIZE = 16,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -46,119 +46,313 @@ module VX_cache #(
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 0,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1,
parameter DRAM_TAG_WIDTH = 1
) (
input wire clk,
input wire reset,
// Core request
input wire [NUM_REQUESTS-1:0] core_req_valid,
input wire [NUM_REQUESTS-1:0][2:0] core_req_read,
input wire [NUM_REQUESTS-1:0][2:0] core_req_write,
input wire [NUM_REQUESTS-1:0][31:0] core_req_addr,
input wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] core_req_data,
output wire core_req_ready,
// Core request meta data
input wire [4:0] core_req_rd,
input wire [NUM_REQUESTS-1:0][1:0] core_req_wb,
input wire [`NW_BITS-1:0] core_req_warp_num,
input wire [31:0] core_req_pc,
input wire [NUM_REQUESTS-1:0] core_req_valid,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_read,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_write,
input wire [NUM_REQUESTS-1:0][31:0] core_req_addr,
input wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_req_data,
input wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_req_tag,
output wire core_req_ready,
// Core response
output wire [NUM_REQUESTS-1:0] core_rsp_valid,
output wire [4:0] core_rsp_read,
output wire [1:0] core_rsp_write,
output wire [NUM_REQUESTS-1:0][31:0] core_rsp_addr,
output wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] core_rsp_data,
input wire core_rsp_ready,
// Core response meta data
output wire [`NW_BITS-1:0] core_rsp_warp_num,
output wire [NUM_REQUESTS-1:0][31:0] core_rsp_pc,
output wire [NUM_REQUESTS-1:0] core_rsp_valid,
output wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_rsp_data,
output wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
input wire core_rsp_ready,
// 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_ready,
output wire dram_req_read,
output wire dram_req_write,
output wire [`DRAM_ADDR_WIDTH-1:0] dram_req_addr,
output wire [`BANK_LINE_WIDTH-1:0] dram_req_data,
output wire [DRAM_TAG_WIDTH-1:0] dram_req_tag,
input wire dram_req_ready,
// DRAM response
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,
input wire dram_rsp_valid,
input wire [`BANK_LINE_WIDTH-1:0] dram_rsp_data,
input wire [DRAM_TAG_WIDTH-1:0] dram_rsp_tag,
output wire dram_rsp_ready,
// Snoop Req
input wire snp_req_valid,
input wire [31:0] snp_req_addr,
output wire snp_req_ready,
input wire snp_req_valid,
input wire [`DRAM_ADDR_WIDTH-1:0] snp_req_addr,
output wire snp_req_ready,
// Snoop Forward
output wire snp_fwd_valid,
output wire [31:0] snp_fwd_addr,
input wire snp_fwd_ready
output wire snp_fwd_valid,
output wire [`DRAM_ADDR_WIDTH-1:0] snp_fwd_addr,
input wire snp_fwd_ready
);
wire [NUM_BANKS-1:0][NUM_REQUESTS-1:0] per_bank_valids;
wire [NUM_BANKS-1:0] per_bank_core_rsp_pop;
wire [NUM_BANKS-1:0] per_bank_core_rsp_valid;
wire [NUM_BANKS-1:0][`LOG2UP(NUM_REQUESTS)-1:0] per_bank_core_rsp_tid;
wire [NUM_BANKS-1:0][4:0] per_bank_core_rsp_rd;
wire [NUM_BANKS-1:0][1:0] per_bank_core_rsp_wb;
wire [NUM_BANKS-1:0][`NW_BITS-1:0] per_bank_core_rsp_warp_num;
wire [NUM_BANKS-1:0][`WORD_SIZE_RNG] per_bank_core_rsp_data;
wire [NUM_BANKS-1:0][31:0] per_bank_core_rsp_pc;
wire [NUM_BANKS-1:0][31:0] per_bank_core_rsp_addr;
wire [NUM_BANKS-1:0][`LOG2UP(NUM_REQUESTS)-1:0] per_bank_core_rsp_tid;
wire [NUM_BANKS-1:0][`WORD_WIDTH-1:0] per_bank_core_rsp_data;
wire [NUM_BANKS-1:0][CORE_TAG_WIDTH-1:0] per_bank_core_rsp_tag;
wire dfqq_full;
wire [NUM_BANKS-1:0] per_bank_dram_fill_req_valid;
wire [NUM_BANKS-1:0][31:0] per_bank_dram_fill_req_addr;
`DEBUG_BEGIN
wire [NUM_BANKS-1:0] per_bank_dram_fill_req_is_snp;
`DEBUG_END
wire [NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] per_bank_dram_fill_req_addr;
wire [NUM_BANKS-1:0] per_bank_dram_fill_rsp_ready;
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][`DRAM_ADDR_WIDTH-1:0] per_bank_dram_wb_req_addr;
wire [NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] per_bank_dram_wb_req_data;
wire [NUM_BANKS-1:0] per_bank_reqq_full;
wire [NUM_BANKS-1:0] per_bank_snp_req_full;
wire [NUM_BANKS-1:0] per_bank_snp_fwd_valid;
wire [NUM_BANKS-1:0][31:0] per_bank_snp_fwd_addr;
wire [NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] per_bank_snp_fwd_addr;
wire [NUM_BANKS-1:0] per_bank_snp_fwd_pop;
assign core_req_ready = ~(|per_bank_reqq_full);
assign snp_req_ready = ~(|per_bank_snp_req_full);
`DEBUG_BEGIN
wire [NUM_BANKS-1:0] per_bank_dram_fill_req_is_snp;
`DEBUG_END
// assign dram_rsp_ready = (NUM_BANKS == 1) ? per_bank_dram_fill_rsp_ready[0] : per_bank_dram_fill_rsp_ready[dram_rsp_addr[`BANK_SELECT_ADDR_RNG]];
assign dram_rsp_ready = (|per_bank_dram_fill_rsp_ready);
assign dram_req_tag = dram_req_addr;
assign core_req_ready = ~(|per_bank_reqq_full);
assign snp_req_ready = ~(|per_bank_snp_req_full);
assign dram_rsp_ready = (|per_bank_dram_fill_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_core_req_bank_sel #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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)
) cache_core_req_bank_sell (
.core_req_valid (core_req_valid),
.core_req_addr (core_req_addr),
.per_bank_valids (per_bank_valids)
);
genvar i;
generate
for (i = 0; i < NUM_BANKS; i = i + 1) begin
wire [NUM_REQUESTS-1:0] curr_bank_core_req_valids;
wire [NUM_REQUESTS-1:0][31:0] curr_bank_core_req_addr;
wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] curr_bank_core_req_tag;
wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] curr_bank_core_req_data;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] curr_bank_core_req_read;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] curr_bank_core_req_write;
wire curr_bank_core_rsp_pop;
wire curr_bank_core_rsp_valid;
wire [`LOG2UP(NUM_REQUESTS)-1:0] curr_bank_core_rsp_tid;
wire [`WORD_WIDTH-1:0] curr_bank_core_rsp_data;
wire [CORE_TAG_WIDTH-1:0] curr_bank_core_rsp_tag;
wire curr_bank_dram_fill_rsp_valid;
wire [`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] curr_bank_dram_fill_rsp_data;
wire [`LINE_ADDR_WIDTH-1:0] curr_bank_dram_fill_rsp_addr;
wire curr_bank_dram_fill_rsp_ready;
wire curr_bank_dram_fill_req_full;
wire curr_bank_dram_fill_req_valid;
wire curr_bank_dram_fill_req_is_snp;
wire [`LINE_ADDR_WIDTH-1:0] curr_bank_dram_fill_req_addr;
wire curr_bank_dram_wb_req_pop;
wire curr_bank_dram_wb_req_valid;
wire [`LINE_ADDR_WIDTH-1:0] curr_bank_dram_wb_req_addr;
wire[`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] curr_bank_dram_wb_req_data;
wire curr_bank_snp_req_valid;
wire [`LINE_ADDR_WIDTH-1:0] curr_bank_snp_req_addr;
wire curr_bank_snp_req_full;
wire curr_bank_snp_fwd_valid;
wire [`LINE_ADDR_WIDTH-1:0] curr_bank_snp_fwd_addr;
wire curr_bank_snp_fwd_pop;
wire curr_bank_reqq_full;
// Core Req
assign curr_bank_core_req_valids = per_bank_valids[i];
assign curr_bank_core_req_addr = core_req_addr;
assign curr_bank_core_req_data = core_req_data;
assign curr_bank_core_req_tag = core_req_tag;
assign curr_bank_core_req_read = core_req_read;
assign curr_bank_core_req_write = core_req_write;
assign per_bank_reqq_full[i] = curr_bank_reqq_full;
// Core WB
assign curr_bank_core_rsp_pop = per_bank_core_rsp_pop[i];
assign per_bank_core_rsp_valid [i] = curr_bank_core_rsp_valid;
assign per_bank_core_rsp_tid [i] = curr_bank_core_rsp_tid;
assign per_bank_core_rsp_tag [i] = curr_bank_core_rsp_tag;
assign per_bank_core_rsp_data [i] = curr_bank_core_rsp_data;
// Dram fill request
assign curr_bank_dram_fill_req_full = dfqq_full;
assign per_bank_dram_fill_req_valid[i] = curr_bank_dram_fill_req_valid;
assign per_bank_dram_fill_req_addr[i] = `LINE_TO_DRAM_ADDR(curr_bank_dram_fill_req_addr, i);
assign per_bank_dram_fill_req_is_snp[i] = curr_bank_dram_fill_req_is_snp;
// Dram fill response
assign curr_bank_dram_fill_rsp_valid = dram_rsp_valid && (`DRAM_ADDR_BANK(dram_rsp_tag) == i);
assign curr_bank_dram_fill_rsp_addr = `DRAM_TO_LINE_ADDR(dram_rsp_tag);
assign curr_bank_dram_fill_rsp_data = dram_rsp_data;
assign per_bank_dram_fill_rsp_ready[i] = curr_bank_dram_fill_rsp_ready;
// Dram writeback request
assign curr_bank_dram_wb_req_pop = per_bank_dram_wb_queue_pop[i];
assign per_bank_dram_wb_req_valid[i] = curr_bank_dram_wb_req_valid;
assign per_bank_dram_wb_req_addr[i] = `LINE_TO_DRAM_ADDR(curr_bank_dram_wb_req_addr, i);
assign per_bank_dram_wb_req_data[i] = curr_bank_dram_wb_req_data;
// Snoop Request
assign curr_bank_snp_req_valid = snp_req_valid && (`DRAM_ADDR_BANK(snp_req_addr) == i);
assign curr_bank_snp_req_addr = `DRAM_TO_LINE_ADDR(snp_req_addr);
assign per_bank_snp_req_full[i] = curr_bank_snp_req_full;
// Snoop Fwd
assign per_bank_snp_fwd_valid[i] = curr_bank_snp_fwd_valid;
assign per_bank_snp_fwd_addr[i] = `LINE_TO_DRAM_ADDR(curr_bank_snp_fwd_addr, i);
assign curr_bank_snp_fwd_pop = per_bank_snp_fwd_pop[i];
VX_bank #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
) bank (
.clk (clk),
.reset (reset),
// Core request
.core_req_valids (curr_bank_core_req_valids),
.core_req_read (curr_bank_core_req_read),
.core_req_write (curr_bank_core_req_write),
.core_req_addr (curr_bank_core_req_addr),
.core_req_data (curr_bank_core_req_data),
.core_req_tag (curr_bank_core_req_tag),
.core_req_full (curr_bank_reqq_full),
.core_req_ready (core_req_ready),
// Core response
.core_rsp_valid (curr_bank_core_rsp_valid),
.core_rsp_tid (curr_bank_core_rsp_tid),
.core_rsp_data (curr_bank_core_rsp_data),
.core_rsp_tag (curr_bank_core_rsp_tag),
.core_rsp_pop (curr_bank_core_rsp_pop),
// Dram fill request
.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_full (curr_bank_dram_fill_req_full),
// Dram fill response
.dram_fill_rsp_valid (curr_bank_dram_fill_rsp_valid),
.dram_fill_rsp_data (curr_bank_dram_fill_rsp_data),
.dram_fill_rsp_addr (curr_bank_dram_fill_rsp_addr),
.dram_fill_rsp_ready (curr_bank_dram_fill_rsp_ready),
// Dram writeback request
.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_wb_req_pop (curr_bank_dram_wb_req_pop),
// Snoop request
.snp_req_valid (curr_bank_snp_req_valid),
.snp_req_addr (curr_bank_snp_req_addr),
.snp_req_full (curr_bank_snp_req_full),
// Snoop forwarding
.snp_fwd_valid (curr_bank_snp_fwd_valid),
.snp_fwd_addr (curr_bank_snp_fwd_addr),
.snp_fwd_pop (curr_bank_snp_fwd_pop)
);
end
endgenerate
VX_cache_core_rsp_merge #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
) cache_core_rsp_merge (
.per_bank_core_rsp_tid (per_bank_core_rsp_tid),
.per_bank_core_rsp_valid (per_bank_core_rsp_valid),
.per_bank_core_rsp_data (per_bank_core_rsp_data),
.per_bank_core_rsp_tag (per_bank_core_rsp_tag),
.per_bank_core_rsp_pop (per_bank_core_rsp_pop),
.core_rsp_valid (core_rsp_valid),
.core_rsp_data (core_rsp_data),
.core_rsp_tag (core_rsp_tag),
.core_rsp_ready (core_rsp_ready)
);
VX_cache_dram_req_arb #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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)
) cache_dram_req_arb (
.clk (clk),
.reset (reset),
@@ -174,73 +368,11 @@ module VX_cache #(
.dram_req_addr (dram_req_addr),
.dram_req_data (dram_req_data),
.dram_req_ready (dram_req_ready)
);
);
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)
) cache_core_req_bank_sell (
.core_req_valid (core_req_valid),
.core_req_addr (core_req_addr),
.per_bank_valids (per_bank_valids)
);
VX_cache_wb_sel_merge #(
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
.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),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
) cache_core_rsp_sel_merge (
.per_bank_wb_valid (per_bank_core_rsp_valid),
.per_bank_wb_tid (per_bank_core_rsp_tid),
.per_bank_wb_rd (per_bank_core_rsp_rd),
.per_bank_wb_pc (per_bank_core_rsp_pc),
.per_bank_wb_wb (per_bank_core_rsp_wb),
.per_bank_wb_warp_num(per_bank_core_rsp_warp_num),
.per_bank_wb_data (per_bank_core_rsp_data),
.per_bank_wb_pop (per_bank_core_rsp_pop),
.per_bank_wb_addr (per_bank_core_rsp_addr),
.core_rsp_ready (core_rsp_ready),
.core_rsp_valid (core_rsp_valid),
.core_rsp_read (core_rsp_read),
.core_rsp_write (core_rsp_write),
.core_rsp_warp_num (core_rsp_warp_num),
.core_rsp_data (core_rsp_data),
.core_rsp_addr (core_rsp_addr),
.core_rsp_pc (core_rsp_pc)
);
// Snoop Forward Logic
VX_snp_fwd_arb #(
.NUM_BANKS(NUM_BANKS)
.NUM_BANKS(NUM_BANKS),
.BANK_LINE_SIZE(BANK_LINE_SIZE)
) snp_fwd_arb (
.per_bank_snp_fwd_valid (per_bank_snp_fwd_valid),
.per_bank_snp_fwd_addr (per_bank_snp_fwd_addr),
@@ -249,184 +381,5 @@ module VX_cache #(
.snp_fwd_addr (snp_fwd_addr),
.snp_fwd_ready (snp_fwd_ready)
);
// 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_core_req_valids;
wire [NUM_REQUESTS-1:0][31:0] curr_bank_core_req_addr;
wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] curr_bank_core_req_data;
wire [4:0] curr_bank_core_req_rd;
wire [NUM_REQUESTS-1:0][1:0] curr_bank_core_req_wb;
wire [`NW_BITS-1:0] curr_bank_core_warp_num;
wire [NUM_REQUESTS-1:0][2:0] curr_bank_core_req_read;
wire [NUM_REQUESTS-1:0][2:0] curr_bank_core_req_write;
wire [31:0] curr_bank_core_req_pc;
wire curr_bank_core_rsp_pop;
wire curr_bank_core_rsp_valid;
wire [`LOG2UP(NUM_REQUESTS)-1:0] curr_bank_core_rsp_tid;
wire [31:0] curr_bank_core_rsp_pc;
wire [4:0] curr_bank_core_rsp_rd;
wire [1:0] curr_bank_core_rsp_wb;
wire [`NW_BITS-1:0] curr_bank_core_rsp_warp_num;
wire [`WORD_SIZE_RNG] curr_bank_core_rsp_data;
wire [31:0] curr_bank_core_rsp_addr;
wire curr_bank_dram_fill_rsp_valid;
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_rsp_ready;
wire curr_bank_dram_fill_req_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_req_pop;
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;
wire curr_bank_snp_req;
wire[31:0] curr_bank_snp_req_addr;
wire curr_bank_reqq_full;
wire curr_bank_snp_fwd_valid;
wire[31:0] curr_bank_snp_fwd_addr;
wire curr_bank_snp_fwd_pop;
wire curr_bank_snp_req_full;
// Core Req
assign curr_bank_core_req_valids = per_bank_valids[curr_bank];
assign curr_bank_core_req_addr = core_req_addr;
assign curr_bank_core_req_data = core_req_data;
assign curr_bank_core_req_rd = core_req_rd;
assign curr_bank_core_req_wb = core_req_wb;
assign curr_bank_core_req_pc = core_req_pc;
assign curr_bank_core_warp_num = core_req_warp_num;
assign curr_bank_core_req_read = core_req_read;
assign curr_bank_core_req_write = core_req_write;
assign per_bank_reqq_full[curr_bank] = curr_bank_reqq_full;
// Core WB
assign curr_bank_core_rsp_pop = per_bank_core_rsp_pop[curr_bank];
assign per_bank_core_rsp_valid [curr_bank] = curr_bank_core_rsp_valid;
assign per_bank_core_rsp_tid [curr_bank] = curr_bank_core_rsp_tid;
assign per_bank_core_rsp_rd [curr_bank] = curr_bank_core_rsp_rd;
assign per_bank_core_rsp_wb [curr_bank] = curr_bank_core_rsp_wb;
assign per_bank_core_rsp_warp_num[curr_bank] = curr_bank_core_rsp_warp_num;
assign per_bank_core_rsp_data [curr_bank] = curr_bank_core_rsp_data;
assign per_bank_core_rsp_pc [curr_bank] = curr_bank_core_rsp_pc;
assign per_bank_core_rsp_addr [curr_bank] = curr_bank_core_rsp_addr;
// Dram fill request
assign curr_bank_dram_fill_req_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_valid = (NUM_BANKS == 1) || (dram_rsp_valid && (curr_bank_dram_fill_rsp_addr[`BANK_SELECT_ADDR_RNG] == curr_bank));
assign curr_bank_dram_fill_rsp_addr = dram_rsp_addr;
assign curr_bank_dram_fill_rsp_data = dram_rsp_data;
assign per_bank_dram_fill_rsp_ready[curr_bank] = curr_bank_dram_fill_rsp_ready;
// Dram writeback request
assign curr_bank_dram_wb_req_pop = per_bank_dram_wb_queue_pop[curr_bank];
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_valid && (snp_req_addr[`BANK_SELECT_ADDR_RNG] == curr_bank);
assign curr_bank_snp_req_addr = snp_req_addr;
assign per_bank_snp_req_full[curr_bank] = curr_bank_snp_req_full;
// Snoop Fwd
assign per_bank_snp_fwd_valid[curr_bank] = curr_bank_snp_fwd_valid;
assign per_bank_snp_fwd_addr[curr_bank] = curr_bank_snp_fwd_addr;
assign curr_bank_snp_fwd_pop = per_bank_snp_fwd_pop[curr_bank];
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 (
.clk (clk),
.reset (reset),
// Core request
.core_req_valids (curr_bank_core_req_valids),
.core_req_read (curr_bank_core_req_read),
.core_req_write (curr_bank_core_req_write),
.core_req_addr (curr_bank_core_req_addr),
.core_req_data (curr_bank_core_req_data),
.core_req_rd (curr_bank_core_req_rd),
.core_req_wb (curr_bank_core_req_wb),
.core_req_pc (curr_bank_core_req_pc),
.core_req_warp_num (curr_bank_core_warp_num),
.core_req_full (curr_bank_reqq_full),
.core_req_ready (core_req_ready),
// Core response
.core_rsp_valid (curr_bank_core_rsp_valid),
.core_rsp_tid (curr_bank_core_rsp_tid),
.core_rsp_rd (curr_bank_core_rsp_rd),
.core_rsp_wb (curr_bank_core_rsp_wb),
.core_rsp_warp_num (curr_bank_core_rsp_warp_num),
.core_rsp_data (curr_bank_core_rsp_data),
.core_rsp_pc (curr_bank_core_rsp_pc),
.core_rsp_addr (curr_bank_core_rsp_addr),
.core_rsp_pop (curr_bank_core_rsp_pop),
// Dram fill request
.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_full (curr_bank_dram_fill_req_full),
// Dram fill response
.dram_fill_rsp_valid (curr_bank_dram_fill_rsp_valid),
.dram_fill_rsp_addr (curr_bank_dram_fill_rsp_addr),
.dram_fill_rsp_data (curr_bank_dram_fill_rsp_data),
.dram_fill_rsp_ready (curr_bank_dram_fill_rsp_ready),
// Dram writeback request
.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_wb_req_pop (curr_bank_dram_wb_req_pop),
// Snoop request
.snp_req_valid (curr_bank_snp_req),
.snp_req_addr (curr_bank_snp_req_addr),
.snp_req_full (curr_bank_snp_req_full),
// Snoop forwarding
.snp_fwd_valid (curr_bank_snp_fwd_valid),
.snp_fwd_addr (curr_bank_snp_fwd_addr),
.snp_fwd_pop (curr_bank_snp_fwd_pop)
);
end
endgenerate
endmodule

View File

@@ -3,71 +3,80 @@
`include "VX_define.vh"
// data tid rd wb warp_num read write
`define MRVQ_METADATA_SIZE (`WORD_SIZE + `LOG2UP(NUM_REQUESTS) + 5 + 2 + (`NW_BITS) + 3 + 3)
`define WORD_SEL_NO 3'h7
`define WORD_SEL_LB 3'h0
`define WORD_SEL_LH 3'h1
`define WORD_SEL_LW 3'h2
`define WORD_SEL_HB 3'h4
`define WORD_SEL_HH 3'h5
`define WORD_SEL_BITS 3
// rd wb warp_num read write + reqs
`define REQ_INST_META_SIZE (5 + 2 + (`NW_BITS) + 3 + 3 + `LOG2UP(NUM_REQUESTS))
// data tid tag read write base addr
`define MRVQ_METADATA_WIDTH (`WORD_WIDTH + `LOG2UP(NUM_REQUESTS) + CORE_TAG_WIDTH + `WORD_SEL_BITS + `WORD_SEL_BITS + `BASE_ADDR_BITS)
`define WORD_SIZE (8 * WORD_SIZE_BYTES)
`define WORD_SIZE_RNG (`WORD_SIZE)-1:0
// tag read write reqs
`define REQ_INST_META_WIDTH (CORE_TAG_WIDTH + `WORD_SEL_BITS + `WORD_SEL_BITS + `LOG2UP(NUM_REQUESTS))
// 128
`define BANK_SIZE_BYTES (CACHE_SIZE_BYTES / NUM_BANKS)
`define WORD_WIDTH (8 * WORD_SIZE)
`define BYTE_WIDTH (`WORD_WIDTH / 4)
// 8
`define BANK_LINE_COUNT (`BANK_SIZE_BYTES / BANK_LINE_SIZE_BYTES)
// 4
`define BANK_LINE_WORDS (BANK_LINE_SIZE_BYTES / WORD_SIZE_BYTES)
`define BANK_LINE_WIDTH (8 * BANK_LINE_SIZE)
// Offset is fixed
`define OFFSET_ADDR_BITS 2
`define OFFSET_SIZE_END 1
`define BANK_SIZE (CACHE_SIZE / NUM_BANKS)
`define BANK_LINE_COUNT (`BANK_SIZE / BANK_LINE_SIZE)
`define BANK_LINE_WORDS (BANK_LINE_SIZE / WORD_SIZE)
// Offset select
`define OFFSET_ADDR_BITS `CLOG2(WORD_SIZE)
`define OFFSET_ADDR_START 0
`define OFFSET_ADDR_END 1
`define OFFSET_ADDR_END (`OFFSET_ADDR_START+`OFFSET_ADDR_BITS-1)
`define OFFSET_ADDR_RNG `OFFSET_ADDR_END:`OFFSET_ADDR_START
`define OFFSET_SIZE_RNG `OFFSET_SIZE_END:0
// 2
`define WORD_SELECT_BITS (`LOG2UP(`BANK_LINE_WORDS))
// 2
`define WORD_SELECT_SIZE_END (`WORD_SELECT_BITS)
// 2
// Word select
`define WORD_SELECT_BITS `CLOG2(`BANK_LINE_WORDS)
`define WORD_SELECT_ADDR_START (1+`OFFSET_ADDR_END)
// 3
`define WORD_SELECT_ADDR_END (`WORD_SELECT_SIZE_END+`OFFSET_ADDR_END)
// 3:2
`define WORD_SELECT_ADDR_END (`WORD_SELECT_ADDR_START+`WORD_SELECT_BITS-1)
`define WORD_SELECT_ADDR_RNG `WORD_SELECT_ADDR_END:`WORD_SELECT_ADDR_START
// 3
`define BANK_SELECT_BITS (`LOG2UP(NUM_BANKS))
// 3
`define BANK_SELECT_SIZE_END (`BANK_SELECT_BITS)
// 4
// Bank select
`define BANK_SELECT_BITS `CLOG2(NUM_BANKS)
`define BANK_SELECT_ADDR_START (1+`WORD_SELECT_ADDR_END)
// 6
`define BANK_SELECT_ADDR_END (`BANK_SELECT_SIZE_END+`BANK_SELECT_ADDR_START-1)
// 6:4
`define BANK_SELECT_ADDR_END (`BANK_SELECT_ADDR_START+`BANK_SELECT_BITS-1)
`define BANK_SELECT_ADDR_RNG `BANK_SELECT_ADDR_END:`BANK_SELECT_ADDR_START
// 3
`define LINE_SELECT_BITS (`LOG2UP(`BANK_LINE_COUNT))
// 7
// Line select
`define LINE_SELECT_BITS `CLOG2(`BANK_LINE_COUNT)
`define LINE_SELECT_ADDR_START (1+`BANK_SELECT_ADDR_END)
// 9
`define LINE_SELECT_ADDR_END (`LINE_SELECT_BITS+`LINE_SELECT_ADDR_START-1)
// 9:7
`define LINE_SELECT_ADDR_END (`LINE_SELECT_ADDR_START+`LINE_SELECT_BITS-1)
`define LINE_SELECT_ADDR_RNG `LINE_SELECT_ADDR_END:`LINE_SELECT_ADDR_START
// 10
// Tag select
`define TAG_SELECT_BITS (31-`LINE_SELECT_ADDR_END)
`define TAG_SELECT_ADDR_START (1+`LINE_SELECT_ADDR_END)
// 31:10
`define TAG_SELECT_ADDR_RNG 31:`TAG_SELECT_ADDR_START
// 22
`define TAG_SELECT_BITS (32-`TAG_SELECT_ADDR_START)
`define TAG_SELECT_ADDR_END 31
`define TAG_SELECT_ADDR_RNG `TAG_SELECT_ADDR_END:`TAG_SELECT_ADDR_START
`define TAG_LINE_SELECT_BITS (`TAG_SELECT_BITS+`LINE_SELECT_BITS)
`define DRAM_ADDR_WIDTH (32-`CLOG2(BANK_LINE_SIZE))
`define BASE_ADDR_MASK (~((1<<(`WORD_SELECT_ADDR_END+1))-1))
`define LINE_ADDR_WIDTH (`DRAM_ADDR_WIDTH-`BANK_SELECT_BITS)
`define TAG_LINE_ADDR_RNG `LINE_ADDR_WIDTH-1:`LINE_SELECT_BITS
`define BASE_ADDR_BITS (`WORD_SELECT_BITS+`OFFSET_ADDR_BITS)
///////////////////////////////////////////////////////////////////////////////
// Core request tag width pc, wb, rd, warp_num
`define CORE_REQ_TAG_WIDTH (32 + 2 + 5 + `NW_BITS)
// Core request tag info rd + warp_num
`define CORE_REQ_TAG_WARP(x) x[(5 + `NW_BITS)-1:0]
// DRAM response tag bank info
`define DRAM_ADDR_BANK(x) x[`BANK_SELECT_BITS-1:0]
`define DRAM_TO_LINE_ADDR(x) x[`DRAM_ADDR_WIDTH-1:`BANK_SELECT_BITS]
`define LINE_TO_DRAM_ADDR(x, i) {x, (`BANK_SELECT_BITS)'(i)};
`endif

View File

@@ -3,13 +3,13 @@
module VX_cache_core_req_bank_sel #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -38,10 +38,7 @@ module VX_cache_core_req_bank_sel #(
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
parameter FILL_INVALIDAOR_SIZE = 16
) (
input wire [NUM_REQUESTS-1:0] core_req_valid,
input wire [NUM_REQUESTS-1:0][31:0] core_req_addr,

115
hw/rtl/cache/VX_cache_core_rsp_merge.v vendored Normal file
View File

@@ -0,0 +1,115 @@
`include "VX_cache_config.vh"
module VX_cache_core_rsp_merge #(
// Size of cache in bytes
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1,
parameter DRAM_TAG_WIDTH = 1
) (
// Per Bank WB
input wire [NUM_BANKS-1:0][`LOG2UP(NUM_REQUESTS)-1:0] per_bank_core_rsp_tid,
input wire [NUM_BANKS-1:0] per_bank_core_rsp_valid,
input wire [NUM_BANKS-1:0][`WORD_WIDTH-1:0] per_bank_core_rsp_data,
input wire [NUM_BANKS-1:0][CORE_TAG_WIDTH-1:0] per_bank_core_rsp_tag,
output wire [NUM_BANKS-1:0] per_bank_core_rsp_pop,
// Core Writeback
output reg [NUM_REQUESTS-1:0] core_rsp_valid,
output reg [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_rsp_data,
output reg [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
input wire core_rsp_ready
);
reg [NUM_BANKS-1:0] per_bank_core_rsp_pop_unqual;
assign per_bank_core_rsp_pop = per_bank_core_rsp_pop_unqual & {NUM_BANKS{core_rsp_ready}};
wire [`LOG2UP(NUM_BANKS)-1:0] main_bank_index;
wire found_bank;
VX_generic_priority_encoder #(
.N(NUM_BANKS)
) sel_bank (
.valids(per_bank_core_rsp_valid),
.index (main_bank_index),
.found (found_bank)
);
integer i;
generate
always @(*) begin
core_rsp_valid = 0;
core_rsp_data = 0;
core_rsp_tag = 0;
for (i = 0; i < NUM_BANKS; i = i + 1) begin
if ((FUNC_ID == `L2FUNC_ID)
|| (FUNC_ID == `L3FUNC_ID)) begin
if (found_bank
&& per_bank_core_rsp_valid[i]
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_core_rsp_tid[i] != per_bank_core_rsp_tid[main_bank_index]))) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_tag[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_tag[i];
per_bank_core_rsp_pop_unqual[i] = 1;
end else begin
per_bank_core_rsp_pop_unqual[i] = 0;
end
end else begin
if (found_bank
&& per_bank_core_rsp_valid[i]
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_core_rsp_tid[i] != per_bank_core_rsp_tid[main_bank_index]))
&& (`CORE_REQ_TAG_WARP(per_bank_core_rsp_tag[i]) == `CORE_REQ_TAG_WARP(per_bank_core_rsp_tag[main_bank_index]))) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_tag[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_tag[i];
per_bank_core_rsp_pop_unqual[i] = 1;
end else begin
per_bank_core_rsp_pop_unqual[i] = 0;
end
end
end
end
endgenerate
endmodule

View File

@@ -2,13 +2,13 @@
module VX_cache_dfq_queue #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -35,45 +35,42 @@ module VX_cache_dfq_queue #(
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
parameter FILL_INVALIDAOR_SIZE = 16
) (
input wire clk,
input wire reset,
input wire dfqq_push,
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 clk,
input wire reset,
input wire dfqq_push,
input wire[NUM_BANKS-1:0] per_bank_dram_fill_req_valid,
input wire[NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] per_bank_dram_fill_req_addr,
input wire dfqq_pop,
output wire dfqq_req,
output wire[31:0] dfqq_req_addr,
output wire dfqq_empty,
output wire dfqq_full
input wire dfqq_pop,
output wire dfqq_req,
output wire[`DRAM_ADDR_WIDTH-1:0] dfqq_req_addr,
output wire dfqq_empty,
output wire dfqq_full
);
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;
wire[NUM_BANKS-1:0] out_per_bank_dram_fill_req_valid;
wire[NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] 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;
reg [NUM_BANKS-1:0] use_per_bank_dram_fill_req_valid;
reg [NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] 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] use_per_bqual_bank_dram_fill_req_valid;
wire[NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] qual_bank_dram_fill_req_addr;
wire[NUM_BANKS-1:0] updated_bank_dram_fill_req;
wire[NUM_BANKS-1:0] updated_bank_dram_fill_req_valid;
wire o_empty;
wire use_empty = !(|use_per_bank_dram_fill_req);
wire out_empty = !(|out_per_bank_dram_fill_req) || o_empty;
wire use_empty = !(|use_per_bank_dram_fill_req_valid);
wire out_empty = !(|out_per_bank_dram_fill_req_valid) || o_empty;
wire push_qual = dfqq_push && !dfqq_full;
wire pop_qual = dfqq_pop && use_empty && !out_empty;
wire pop_qual = dfqq_pop && use_empty && !out_empty;
VX_generic_queue #(
.DATAW(NUM_BANKS * (1+32)),
.DATAW(NUM_BANKS * (1+`DRAM_ADDR_WIDTH)),
.SIZE(DFQQ_SIZE)
) dfqq_queue (
.clk (clk),
@@ -81,38 +78,38 @@ module VX_cache_dfq_queue #(
.push (push_qual),
.data_in ({per_bank_dram_fill_req_valid, per_bank_dram_fill_req_addr}),
.pop (pop_qual),
.data_out({out_per_bank_dram_fill_req, out_per_bank_dram_fill_req_addr}),
.data_out({out_per_bank_dram_fill_req_valid, 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 use_per_bqual_bank_dram_fill_req_valid = use_empty ? (out_per_bank_dram_fill_req_valid & {NUM_BANKS{!o_empty}}) : (use_per_bank_dram_fill_req_valid & {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;
wire qual_has_request;
VX_generic_priority_encoder #(
.N(NUM_BANKS)
) sel_bank (
.valids(qual_bank_dram_fill_req),
.valids(use_per_bqual_bank_dram_fill_req_valid),
.index (qual_request_index),
.found (qual_has_request)
);
assign dfqq_empty = !qual_has_request;
assign dfqq_req = qual_bank_dram_fill_req [qual_request_index];
assign dfqq_req = use_per_bqual_bank_dram_fill_req_valid [qual_request_index];
assign dfqq_req_addr = qual_bank_dram_fill_req_addr[qual_request_index];
assign updated_bank_dram_fill_req = qual_bank_dram_fill_req & (~(1 << qual_request_index));
assign updated_bank_dram_fill_req_valid = use_per_bqual_bank_dram_fill_req_valid & (~(1 << qual_request_index));
always @(posedge clk) begin
if (reset) begin
use_per_bank_dram_fill_req <= 0;
use_per_bank_dram_fill_req_valid <= 0;
use_per_bank_dram_fill_req_addr <= 0;
end else begin
if (dfqq_pop && qual_has_request) begin
use_per_bank_dram_fill_req <= updated_bank_dram_fill_req;
use_per_bank_dram_fill_req_valid <= updated_bank_dram_fill_req_valid;
use_per_bank_dram_fill_req_addr <= qual_bank_dram_fill_req_addr;
end
end

View File

@@ -2,13 +2,13 @@
module VX_cache_dram_req_arb #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -40,37 +40,34 @@ module VX_cache_dram_req_arb #(
// Prefetcher
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 2,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
parameter PRFQ_STRIDE = 2
) (
input wire clk,
input wire reset,
// Fill Request
output wire dfqq_full,
// Fill Request
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 [NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] per_bank_dram_fill_req_addr,
output wire dfqq_full,
// DFQ Request
output wire [NUM_BANKS-1:0] per_bank_dram_wb_queue_pop,
// Writeback Request
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,
input wire [NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] per_bank_dram_wb_req_addr,
input wire [NUM_BANKS-1:0][`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] per_bank_dram_wb_req_data,
output wire [NUM_BANKS-1:0] per_bank_dram_wb_queue_pop,
// real Dram request
// Merged 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,
output wire [`DRAM_ADDR_WIDTH-1:0] dram_req_addr,
output wire [`BANK_LINE_WIDTH-1:0] dram_req_data,
input wire dram_req_ready
);
wire pref_pop;
wire pref_valid;
wire[31:0] pref_addr;
wire[`DRAM_ADDR_WIDTH-1:0] pref_addr;
wire dwb_valid;
wire dfqq_req;
@@ -78,10 +75,10 @@ module VX_cache_dram_req_arb #(
assign pref_pop = !dwb_valid && !dfqq_req && dram_req_ready && 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)
.PRFQ_SIZE (PRFQ_SIZE),
.PRFQ_STRIDE (PRFQ_STRIDE),
.BANK_LINE_SIZE(BANK_LINE_SIZE),
.WORD_SIZE (WORD_SIZE)
) prfqq (
.clk (clk),
.reset (reset),
@@ -94,7 +91,7 @@ module VX_cache_dram_req_arb #(
.pref_addr (pref_addr)
);
wire[31:0] dfqq_req_addr;
wire[`DRAM_ADDR_WIDTH-1:0] dfqq_req_addr;
`DEBUG_BEGIN
wire dfqq_empty;
@@ -130,10 +127,10 @@ module VX_cache_dram_req_arb #(
assign per_bank_dram_wb_queue_pop = dram_req_ready ? (use_wb_valid & ((1 << dwb_bank))) : 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;
wire dram_req_valid = dwb_valid || dfqq_req || pref_pop;
assign dram_req_read = ((dfqq_req && !dwb_valid) || pref_pop) && dram_req_valid;
assign dram_req_write = dwb_valid && dram_req_valid;
assign dram_req_addr = dwb_valid ? per_bank_dram_wb_req_addr[dwb_bank] : (dfqq_req ? dfqq_req_addr : pref_addr);
assign {dram_req_data} = dwb_valid ? {per_bank_dram_wb_req_data[dwb_bank] }: 0;
endmodule

View File

@@ -3,13 +3,13 @@
module VX_cache_miss_resrv #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -36,25 +36,23 @@ module VX_cache_miss_resrv #(
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1
) (
input wire clk,
input wire reset,
// Miss enqueue
input wire miss_add,
input wire[31:0] miss_add_addr,
input wire[`WORD_SIZE_RNG] miss_add_data,
input wire[`LINE_ADDR_WIDTH-1:0] miss_add_addr,
input wire[`BASE_ADDR_BITS-1:0] miss_add_wsel,
input wire[`WORD_WIDTH-1:0] miss_add_data,
input wire[`LOG2UP(NUM_REQUESTS)-1:0] miss_add_tid,
input wire[4:0] miss_add_rd,
input wire[1:0] miss_add_wb,
input wire[`NW_BITS-1:0] miss_add_warp_num,
input wire[2:0] miss_add_mem_read,
input wire[2:0] miss_add_mem_write,
input wire[31:0] miss_add_pc,
input wire[CORE_TAG_WIDTH-1:0] miss_add_tag,
input wire[`WORD_SEL_BITS-1:0] miss_add_mem_read,
input wire[`WORD_SEL_BITS-1:0] miss_add_mem_write,
output wire miss_resrv_full,
output wire miss_resrv_stop,
@@ -63,37 +61,31 @@ module VX_cache_miss_resrv #(
`IGNORE_WARNINGS_BEGIN
// TODO: should fix this
input wire[31:0] fill_addr_st1,
input wire[`LINE_ADDR_WIDTH-1:0] fill_addr_st1,
`IGNORE_WARNINGS_END
// Miss dequeue
input wire miss_resrv_pop,
output wire miss_resrv_valid_st0,
output wire[31:0] miss_resrv_addr_st0,
output wire[`WORD_SIZE_RNG] miss_resrv_data_st0,
output wire[`LINE_ADDR_WIDTH-1:0] miss_resrv_addr_st0,
output wire[`BASE_ADDR_BITS-1:0] miss_resrv_wsel_st0,
output wire[`WORD_WIDTH-1:0] miss_resrv_data_st0,
output wire[`LOG2UP(NUM_REQUESTS)-1:0] miss_resrv_tid_st0,
output wire[4:0] miss_resrv_rd_st0,
output wire[1:0] miss_resrv_wb_st0,
output wire[`NW_BITS-1:0] miss_resrv_warp_num_st0,
output wire[2:0] miss_resrv_mem_read_st0,
output wire[31:0] miss_resrv_pc_st0,
output wire[2:0] miss_resrv_mem_write_st0
output wire[CORE_TAG_WIDTH-1:0] miss_resrv_tag_st0,
output wire[`WORD_SEL_BITS-1:0] miss_resrv_mem_read_st0,
output wire[`WORD_SEL_BITS-1:0] 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_METADATA_WIDTH-1:0] metadata_table[MRVQ_SIZE-1:0];
reg [MRVQ_SIZE-1:0][`LINE_ADDR_WIDTH-1:0] addr_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 [`LOG2UP(MRVQ_SIZE+1)-1:0] 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));
assign miss_resrv_full = (size == $bits(size)'(MRVQ_SIZE));
assign miss_resrv_stop = (size > $bits(size)'(MRVQ_SIZE-5));
wire enqueue_possible = !miss_resrv_full;
wire [`LOG2UP(MRVQ_SIZE)-1:0] enqueue_index = tail_ptr;
@@ -102,8 +94,7 @@ module VX_cache_miss_resrv #(
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 make_ready[curr_e] = is_fill_st1 && valid_table[curr_e] && (addr_table[curr_e] == fill_addr_st1);
end
endgenerate
@@ -111,24 +102,19 @@ module VX_cache_miss_resrv #(
wire [`LOG2UP(MRVQ_SIZE)-1:0] dequeue_index = head_ptr;
assign miss_resrv_valid_st0 = (MRVQ_SIZE != 2) && dequeue_possible;
assign miss_resrv_pc_st0 = pc_table[dequeue_index];
assign miss_resrv_addr_st0 = addr_table[dequeue_index];
assign {miss_resrv_data_st0, miss_resrv_tid_st0, miss_resrv_rd_st0, miss_resrv_wb_st0, miss_resrv_warp_num_st0, miss_resrv_mem_read_st0, miss_resrv_mem_write_st0} = metadata_table[dequeue_index];
assign {miss_resrv_data_st0, miss_resrv_tid_st0, miss_resrv_tag_st0, miss_resrv_mem_read_st0, miss_resrv_mem_write_st0, miss_resrv_wsel_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;
@@ -136,9 +122,8 @@ module VX_cache_miss_resrv #(
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};
metadata_table[enqueue_index] <= {miss_add_data, miss_add_tid, miss_add_tag, miss_add_mem_read, miss_add_mem_write, miss_add_wsel};
tail_ptr <= tail_ptr + 1;
end
@@ -151,7 +136,6 @@ module VX_cache_miss_resrv #(
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
@@ -159,7 +143,6 @@ module VX_cache_miss_resrv #(
if (mrvq_push) begin
size <= size + 1;
end
if (mrvq_pop) begin
size <= size - 1;
end

View File

@@ -2,13 +2,13 @@
module VX_cache_req_queue #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -34,77 +34,62 @@ module VX_cache_req_queue #(
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1
) (
input wire clk,
input wire reset,
input wire clk,
input wire reset,
// Enqueue Data
input wire reqq_push,
input wire [NUM_REQUESTS-1:0] bank_valids,
input wire [NUM_REQUESTS-1:0][31:0] bank_addr,
input wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] bank_writedata,
input wire [4:0] bank_rd,
input wire [NUM_REQUESTS-1:0][1:0] bank_wb,
input wire [`NW_BITS-1:0] bank_warp_num,
input wire [NUM_REQUESTS-1:0][2:0] bank_mem_read,
input wire [NUM_REQUESTS-1:0][2:0] bank_mem_write,
input wire [31:0] bank_pc,
input wire reqq_push,
input wire [NUM_REQUESTS-1:0] bank_valids,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] bank_mem_read,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] bank_mem_write,
input wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] bank_writedata,
input wire [NUM_REQUESTS-1:0][31:0] bank_addr,
input wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] bank_tag,
// Dequeue Data
input wire reqq_pop,
output wire reqq_req_st0,
output wire [`LOG2UP(NUM_REQUESTS)-1:0] reqq_req_tid_st0,
output wire [`LOG2UP(NUM_REQUESTS)-1:0] reqq_req_tid_st0,
output wire [`WORD_SEL_BITS-1:0] reqq_req_mem_read_st0,
output wire [`WORD_SEL_BITS-1:0] reqq_req_mem_write_st0,
output wire [`WORD_WIDTH-1:0] reqq_req_writedata_st0,
output wire [31:0] reqq_req_addr_st0,
output wire [`WORD_SIZE_RNG] reqq_req_writedata_st0,
output wire [4:0] reqq_req_rd_st0,
output wire [1:0] reqq_req_wb_st0,
output wire [`NW_BITS-1:0] reqq_req_warp_num_st0,
output wire [2:0] reqq_req_mem_read_st0,
output wire [2:0] reqq_req_mem_write_st0,
output wire [31:0] reqq_req_pc_st0,
output wire [CORE_TAG_WIDTH-1:0] reqq_req_tag_st0,
// State Data
output wire reqq_empty,
output wire reqq_full
);
wire [NUM_REQUESTS-1:0] out_per_valids;
wire [NUM_REQUESTS-1:0][31:0] out_per_addr;
wire [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] out_per_writedata;
wire [4:0] out_per_rd;
wire [NUM_REQUESTS-1:0][1:0] out_per_wb;
wire [`NW_BITS-1:0] out_per_warp_num;
wire [NUM_REQUESTS-1:0][2:0] out_per_mem_read;
wire [NUM_REQUESTS-1:0][2:0] out_per_mem_write;
wire [31:0] out_per_pc;
wire [NUM_REQUESTS-1:0] out_per_valids;
wire [NUM_REQUESTS-1:0][31:0] out_per_addr;
wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] out_per_writedata;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] out_per_mem_read;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] out_per_mem_write;
wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] out_per_tag;
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;
reg [4:0] use_per_rd;
reg [NUM_REQUESTS-1:0][1:0] use_per_wb;
reg [31:0] use_per_pc;
reg [`NW_BITS-1:0] use_per_warp_num;
reg [NUM_REQUESTS-1:0][2:0] use_per_mem_read;
reg [NUM_REQUESTS-1:0][2:0] use_per_mem_write;
reg [NUM_REQUESTS-1:0] use_per_valids;
reg [NUM_REQUESTS-1:0][31:0] use_per_addr;
reg [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] use_per_writedata;
reg [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] use_per_mem_read;
reg [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] use_per_mem_write;
reg [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] use_per_tag;
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;
wire [4:0] qual_rd;
wire [NUM_REQUESTS-1:0][1:0] qual_wb;
wire [`NW_BITS-1:0] qual_warp_num;
wire [NUM_REQUESTS-1:0][2:0] qual_mem_read;
wire [NUM_REQUESTS-1:0][2:0] qual_mem_write;
wire [31:0] qual_pc;
wire [NUM_REQUESTS-1:0] qual_valids;
wire [NUM_REQUESTS-1:0][31:0] qual_addr;
wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] qual_writedata;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] qual_mem_read;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] qual_mem_write;
wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] qual_tag;
`DEBUG_BEGIN
reg [NUM_REQUESTS-1:0] updated_valids;
reg [NUM_REQUESTS-1:0] updated_valids;
`DEBUG_END
wire o_empty;
@@ -116,15 +101,15 @@ module VX_cache_req_queue #(
wire pop_qual = !out_empty && use_empty;
VX_generic_queue #(
.DATAW( (NUM_REQUESTS * (1+32+`WORD_SIZE)) + 5 + (NUM_REQUESTS*2) + (`NW_BITS-1+1) + (NUM_REQUESTS * (3 + 3)) + 32 ),
.DATAW($bits(bank_valids) + $bits(bank_addr) + $bits(bank_writedata) + $bits(bank_tag) + $bits(bank_mem_read) + $bits(bank_mem_write)),
.SIZE(REQQ_SIZE)
) reqq_queue (
.clk (clk),
.reset (reset),
.push (push_qual),
.data_in ({bank_valids , bank_addr , bank_writedata , bank_rd , bank_wb , bank_warp_num , bank_mem_read , bank_mem_write , bank_pc}),
.data_in ({bank_valids, bank_addr, bank_writedata, bank_tag, bank_mem_read, bank_mem_write}),
.pop (pop_qual),
.data_out ({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}),
.data_out ({out_per_valids, out_per_addr, out_per_writedata, out_per_tag, out_per_mem_read, out_per_mem_write}),
.empty (o_empty),
.full (reqq_full)
);
@@ -134,15 +119,12 @@ module VX_cache_req_queue #(
assign qual_valids = use_per_valids;
assign qual_addr = use_per_addr;
assign qual_writedata = use_per_writedata;
assign qual_rd = use_per_rd;
assign qual_wb = use_per_wb;
assign qual_warp_num = use_per_warp_num;
assign qual_tag = use_per_tag;
assign qual_mem_read = use_per_mem_read;
assign qual_mem_write = use_per_mem_write;
assign qual_pc = use_per_pc;
wire[`LOG2UP(NUM_REQUESTS)-1:0] qual_request_index;
wire qual_has_request;
wire[`LOG2UP(NUM_REQUESTS)-1:0]qual_request_index;
wire qual_has_request;
VX_generic_priority_encoder #(
.N(NUM_REQUESTS)
@@ -157,12 +139,9 @@ module VX_cache_req_queue #(
assign reqq_req_tid_st0 = qual_request_index;
assign reqq_req_addr_st0 = qual_addr[qual_request_index];
assign reqq_req_writedata_st0 = qual_writedata[qual_request_index];
assign reqq_req_rd_st0 = qual_rd;
assign reqq_req_wb_st0 = qual_wb[qual_request_index];
assign reqq_req_warp_num_st0 = qual_warp_num;
assign reqq_req_tag_st0 = qual_tag[qual_request_index];
assign reqq_req_mem_read_st0 = qual_mem_read [qual_request_index];
assign reqq_req_mem_write_st0 = qual_mem_write[qual_request_index];
assign reqq_req_pc_st0 = qual_pc;
always @(*) begin
updated_valids = qual_valids;
@@ -173,32 +152,18 @@ module VX_cache_req_queue #(
always @(posedge clk) begin
if (reset) begin
use_per_valids <= 0;
use_per_addr <= 0;
use_per_writedata <= 0;
use_per_rd <= 0;
use_per_wb <= 0;
use_per_warp_num <= 0;
use_per_mem_read <= 0;
use_per_mem_write <= 0;
use_per_pc <= 0;
use_per_valids <= 0;
end else begin
if (pop_qual) begin
use_per_valids <= real_out_per_valids;
use_per_addr <= out_per_addr;
use_per_writedata <= out_per_writedata;
use_per_rd <= out_per_rd;
use_per_wb <= out_per_wb;
use_per_warp_num <= out_per_warp_num;
use_per_tag <= out_per_tag;
use_per_mem_read <= out_per_mem_read;
use_per_mem_write <= out_per_mem_write;
use_per_pc <= out_per_pc;
end else if (reqq_pop) begin
use_per_valids[qual_request_index] <= 0;
end
// else if (reqq_pop) begin
// use_per_valids[qual_request_index] <= updated_valids;
// end
end
end

View File

@@ -1,137 +0,0 @@
`include "VX_cache_config.vh"
module VX_cache_wb_sel_merge #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
parameter REQQ_SIZE = 8,
// Miss Reserv Queue Knob
parameter MRVQ_SIZE = 8,
// Dram Fill Rsp Queue Size
parameter DFPQ_SIZE = 2,
// Snoop Req Queue
parameter SNRQ_SIZE = 8,
// Queues for writebacks Knobs {1, 2, 4, 8, ...}
// Core Writeback Queue Size
parameter CWBQ_SIZE = 8,
// Dram Writeback Queue Size
parameter DWBQ_SIZE = 4,
// Dram Fill Req Queue Size
parameter DFQQ_SIZE = 8,
// Lower Level Cache Hit Queue Size
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
) (
// Per Bank WB
input wire [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_addr,
output wire [NUM_BANKS-1:0] per_bank_wb_pop,
// Core Writeback
input wire core_rsp_ready,
output reg [NUM_REQUESTS-1:0] core_rsp_valid,
output reg [NUM_REQUESTS-1:0][`WORD_SIZE_RNG] core_rsp_data,
output reg [NUM_REQUESTS-1:0][31:0] core_rsp_pc,
output wire [4:0] core_rsp_read,
output wire [1:0] core_rsp_write,
output wire [`NW_BITS-1:0] core_rsp_warp_num,
output reg [NUM_REQUESTS-1:0][31:0] core_rsp_addr
);
reg [NUM_BANKS-1:0] per_bank_wb_pop_unqual;
assign per_bank_wb_pop = per_bank_wb_pop_unqual & {NUM_BANKS{core_rsp_ready}};
// wire[NUM_BANKS-1:0] bank_wants_wb;
// genvar curr_bank;
// generate
// for (curr_bank = 0; curr_bank < NUM_BANKS; curr_bank=curr_bank+1) begin
// assign bank_wants_wb[curr_bank] = (|per_bank_wb_valid[curr_bank]);
// end
// endgenerate
wire [`LOG2UP(NUM_BANKS)-1:0] main_bank_index;
wire found_bank;
VX_generic_priority_encoder #(
.N(NUM_BANKS)
) sel_bank (
.valids(per_bank_wb_valid),
.index (main_bank_index),
.found (found_bank)
);
assign core_rsp_read = per_bank_wb_rd[main_bank_index];
assign core_rsp_write = per_bank_wb_wb[main_bank_index];
assign core_rsp_warp_num = per_bank_wb_warp_num[main_bank_index];
integer i;
generate
always @(*) begin
core_rsp_valid = 0;
core_rsp_data = 0;
core_rsp_pc = 0;
core_rsp_addr = 0;
for (i = 0; i < NUM_BANKS; i = i + 1) begin
if ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID)) begin
if (found_bank
&& !core_rsp_valid[per_bank_wb_tid[i]]
&& per_bank_wb_valid[i]
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_wb_tid[i] != per_bank_wb_tid[main_bank_index]))) begin
core_rsp_valid[per_bank_wb_tid[i]] = 1;
core_rsp_data[per_bank_wb_tid[i]] = per_bank_wb_data[i];
core_rsp_pc[per_bank_wb_tid[i]] = per_bank_wb_pc[i];
core_rsp_addr[per_bank_wb_tid[i]] = per_bank_wb_addr[i];
per_bank_wb_pop_unqual[i] = 1;
end else begin
per_bank_wb_pop_unqual[i] = 0;
end
end else begin
if (((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_wb_tid[i] != per_bank_wb_tid[main_bank_index]))
&& found_bank
&& !core_rsp_valid[per_bank_wb_tid[i]]
&& (per_bank_wb_valid[i])
&& (per_bank_wb_rd[i] == per_bank_wb_rd[main_bank_index])
&& (per_bank_wb_warp_num[i] == per_bank_wb_warp_num[main_bank_index])) begin
core_rsp_valid[per_bank_wb_tid[i]] = 1;
core_rsp_data[per_bank_wb_tid[i]] = per_bank_wb_data[i];
core_rsp_pc[per_bank_wb_tid[i]] = per_bank_wb_pc[i];
core_rsp_addr[per_bank_wb_tid[i]] = per_bank_wb_addr[i];
per_bank_wb_pop_unqual[i] = 1;
end else begin
per_bank_wb_pop_unqual[i] = 0;
end
end
end
end
endgenerate
endmodule

View File

@@ -2,13 +2,13 @@
module VX_fill_invalidator #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -35,20 +35,17 @@ module VX_fill_invalidator #(
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
parameter FILL_INVALIDAOR_SIZE = 16
) (
input wire clk,
input wire reset,
input wire clk,
input wire reset,
input wire possible_fill,
input wire success_fill,
input wire possible_fill,
input wire success_fill,
input wire[31:0] fill_addr,
input wire[`LINE_ADDR_WIDTH-1:0] fill_addr,
output reg invalidate_fill
output reg invalidate_fill
);
if (FILL_INVALIDAOR_SIZE == 0) begin
@@ -57,15 +54,16 @@ module VX_fill_invalidator #(
end else begin
reg [FILL_INVALIDAOR_SIZE-1:0] fills_active;
reg [FILL_INVALIDAOR_SIZE-1:0][31:0] fills_address;
reg [FILL_INVALIDAOR_SIZE-1:0] fills_active;
reg [FILL_INVALIDAOR_SIZE-1:0][`LINE_ADDR_WIDTH-1:0] fills_address;
reg [FILL_INVALIDAOR_SIZE-1:0] matched_fill;
wire matched;
integer fi;
always @(*) begin
for (fi = 0; fi < FILL_INVALIDAOR_SIZE; fi+=1) begin
matched_fill[fi] = fills_active[fi] && (fills_address[fi][31:`LINE_SELECT_ADDR_START] == fill_addr[31:`LINE_SELECT_ADDR_START]);
matched_fill[fi] = fills_active[fi] && (fills_address[fi] == fill_addr);
end
end
@@ -89,7 +87,6 @@ module VX_fill_invalidator #(
fills_active <= 0;
fills_address <= 0;
end else begin
if (possible_fill && !matched && enqueue_found) begin
fills_active [enqueue_index] <= 1;
fills_address[enqueue_index] <= fill_addr;
@@ -99,56 +96,6 @@ module VX_fill_invalidator #(
end
end
// reg success_found;
// reg[(`LOG2UP(FILL_INVALIDAOR_SIZE))-1:0] success_index;
// integer curr_fill;
// always @(*) begin
// invalidate_fill = 0;
// success_found = 0;
// success_index = 0;
// for (curr_fill = 0; curr_fill < FILL_INVALIDAOR_SIZE; curr_fill=curr_fill+1) begin
// if (fill_addr[31:`LINE_SELECT_ADDR_START] == fills_address[curr_fill][31:`LINE_SELECT_ADDR_START]) begin
// if (possible_fill && fills_active[curr_fill]) begin
// invalidate_fill = 1;
// end
// if (success_fill) begin
// success_found = 1;
// success_index = curr_fill;
// end
// end
// end
// end
// wire [(`LOG2UP(FILL_INVALIDAOR_SIZE))-1:0] enqueue_index;
// wire enqueue_found;
// VX_generic_priority_encoder #(.N(FILL_INVALIDAOR_SIZE)) sel_bank(
// .valids(~fills_active),
// .index (enqueue_index),
// .found (enqueue_found)
// );
// always @(posedge clk) begin
// if (reset) begin
// fills_active <= 0;
// fills_address <= 0;
// end else begin
// if (possible_fill && !invalidate_fill) begin
// fills_active[enqueue_index] <= 1;
// fills_address[enqueue_index] <= fill_addr;
// end
// if (success_found) begin
// fills_active[success_index] <= 0;
// end
// end
// end
end
endmodule

View File

@@ -1,29 +1,29 @@
`include "VX_cache_config.vh"
module VX_prefetcher #(
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 2,
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 2,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4
parameter WORD_SIZE = 4
) (
input wire clk,
input wire reset,
input wire clk,
input wire reset,
input wire dram_req,
input wire[31:0] dram_req_addr,
input wire dram_req,
input wire[`DRAM_ADDR_WIDTH-1:0] dram_req_addr,
input wire pref_pop,
output wire pref_valid,
output wire[31:0] pref_addr
input wire pref_pop,
output wire pref_valid,
output wire[`DRAM_ADDR_WIDTH-1:0] pref_addr
);
reg[`LOG2UP(PRFQ_STRIDE):0] use_valid;
reg[31:0] use_addr;
reg[`LOG2UP(PRFQ_STRIDE):0] use_valid;
reg[`DRAM_ADDR_WIDTH-1:0] use_addr;
wire current_valid;
wire[31:0] current_addr;
wire current_valid;
wire[`DRAM_ADDR_WIDTH-1:0] current_addr;
wire current_full;
wire current_empty;
@@ -33,14 +33,14 @@ module VX_prefetcher #(
wire update_use = ((use_valid == 0) || ((use_valid-1) == 0)) && current_valid;
VX_generic_queue #(
.DATAW(32),
.DATAW(`DRAM_ADDR_WIDTH),
.SIZE(PRFQ_SIZE)
) pfq_queue (
.clk (clk),
.reset (reset),
.push (dram_req && !current_full && !pref_pop),
.data_in (dram_req_addr & `BASE_ADDR_MASK),
.data_in (dram_req_addr),
.pop (update_use),
.data_out(current_addr),
@@ -49,7 +49,7 @@ module VX_prefetcher #(
.full (current_full)
);
assign pref_valid = use_valid != 0;
assign pref_valid = 0; // TODO use_valid != 0;
assign pref_addr = use_addr;
always @(posedge clk) begin
@@ -59,10 +59,10 @@ module VX_prefetcher #(
end else begin
if (update_use) begin
use_valid <= PRFQ_STRIDE;
use_addr <= current_addr + BANK_LINE_SIZE_BYTES;
use_addr <= current_addr + BANK_LINE_SIZE;
end else if (pref_valid && pref_pop) begin
use_valid <= use_valid - 1;
use_addr <= use_addr + BANK_LINE_SIZE_BYTES;
use_addr <= use_addr + BANK_LINE_SIZE;
end
end
end

View File

@@ -1,14 +1,15 @@
`include "VX_cache_config.vh"
module VX_snp_fwd_arb #(
parameter NUM_BANKS = 8
parameter NUM_BANKS = 1,
parameter BANK_LINE_SIZE = 1
) (
input wire [NUM_BANKS-1:0] per_bank_snp_fwd_valid,
input wire [NUM_BANKS-1:0][31:0] per_bank_snp_fwd_addr,
input wire [NUM_BANKS-1:0][`DRAM_ADDR_WIDTH-1:0] per_bank_snp_fwd_addr,
output reg [NUM_BANKS-1:0] per_bank_snp_fwd_pop,
output wire snp_fwd_valid,
output wire [31:0] snp_fwd_addr,
output wire [`DRAM_ADDR_WIDTH-1:0] snp_fwd_addr,
input wire snp_fwd_ready
);

View File

@@ -2,13 +2,13 @@
module VX_tag_data_access #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -37,85 +37,86 @@ module VX_tag_data_access #(
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
parameter FILL_INVALIDAOR_SIZE = 16
) (
input wire clk,
input wire reset,
input wire stall,
input wire is_snp_st1e,
input wire stall_bank_pipe,
// Initial Reading
`IGNORE_WARNINGS_BEGIN
// TODO: should fix this
input wire[31:0] readaddr_st10,
input wire[31:0] writeaddr_st1e,
`IGNORE_WARNINGS_END
input wire valid_req_st1e,
input wire writefill_st1e,
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,
input wire[2:0] mem_read_st1e,
input wire clk,
input wire reset,
input wire stall,
input wire is_snp_st1e,
input wire stall_bank_pipe,
output wire[`WORD_SIZE_RNG] readword_st1e,
output wire[`DBANK_LINE_WORDS-1:0][31:0] readdata_st1e,
output wire[`TAG_SELECT_BITS-1:0] readtag_st1e,
output wire miss_st1e,
output wire dirty_st1e,
output wire fill_saw_dirty_st1e
input wire[`LINE_SELECT_BITS-1:0] readaddr_st10,
input wire[`LINE_ADDR_WIDTH-1:0] writeaddr_st1e,
input wire valid_req_st1e,
input wire writefill_st1e,
input wire[`WORD_WIDTH-1:0] writeword_st1e,
input wire[`BANK_LINE_WIDTH-1:0] writedata_st1e,
`IGNORE_WARNINGS_BEGIN
input wire[`WORD_SELECT_ADDR_END:0] writewsel_st1e,
input wire[`WORD_SEL_BITS-1:0] mem_write_st1e,
input wire[`WORD_SEL_BITS-1:0] mem_read_st1e,
`IGNORE_WARNINGS_END
output wire[`WORD_WIDTH-1:0] readword_st1e,
output wire[`BANK_LINE_WIDTH-1:0] readdata_st1e,
output wire[`TAG_SELECT_BITS-1:0] readtag_st1e,
output wire miss_st1e,
output wire dirty_st1e,
output wire fill_saw_dirty_st1e
);
reg read_valid_st1c[STAGE_1_CYCLES-1:0];
reg read_dirty_st1c[STAGE_1_CYCLES-1:0];
reg[`TAG_SELECT_BITS-1:0] read_tag_st1c [STAGE_1_CYCLES-1:0];
reg[`DBANK_LINE_WORDS-1:0][31:0] read_data_st1c [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_BITS-1:0] read_tag_st1c [STAGE_1_CYCLES-1:0];
reg[`BANK_LINE_WIDTH-1:0] read_data_st1c [STAGE_1_CYCLES-1:0];
wire qual_read_valid_st1;
wire qual_read_dirty_st1;
wire[`TAG_SELECT_BITS-1:0] qual_read_tag_st1;
wire[`DBANK_LINE_WORDS-1:0][31:0] qual_read_data_st1;
wire qual_read_valid_st1;
wire qual_read_dirty_st1;
wire[`TAG_SELECT_BITS-1:0] qual_read_tag_st1;
wire[`BANK_LINE_WIDTH-1:0] qual_read_data_st1;
wire use_read_valid_st1e;
wire use_read_dirty_st1e;
wire[`TAG_SELECT_BITS-1:0] use_read_tag_st1e;
wire[`DBANK_LINE_WORDS-1:0][31:0] use_read_data_st1e;
wire[`DBANK_LINE_WORDS-1:0][3:0] use_write_enable;
wire[`DBANK_LINE_WORDS-1:0][31:0] use_write_data;
wire sw, sb, sh;
wire real_writefill = writefill_st1e && ((valid_req_st1e && !use_read_valid_st1e) || (valid_req_st1e && use_read_valid_st1e && (writeaddr_st1e[`TAG_SELECT_ADDR_RNG] != use_read_tag_st1e)));
wire use_read_valid_st1e;
wire use_read_dirty_st1e;
wire[`TAG_SELECT_BITS-1:0] use_read_tag_st1e;
wire[`BANK_LINE_WIDTH-1:0] use_read_data_st1e;
wire[`BANK_LINE_WORDS-1:0][3:0] use_write_enable;
wire[`BANK_LINE_WIDTH-1:0] use_write_data;
wire fill_sent;
wire invalidate_line;
VX_tag_data_structure #(
.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),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
wire real_writefill = writefill_st1e
&& ((valid_req_st1e
&& !use_read_valid_st1e)
|| (valid_req_st1e
&& use_read_valid_st1e
&& (writeaddr_st1e[`TAG_LINE_ADDR_RNG] != use_read_tag_st1e)));
VX_tag_data_structure #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.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),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE)
) tag_data_structure (
.clk (clk),
.reset (reset),
.stall_bank_pipe(stall_bank_pipe),
.read_addr (readaddr_st10[`LINE_SELECT_ADDR_RNG]),
.read_addr (readaddr_st10),
.read_valid (qual_read_valid_st1),
.read_dirty (qual_read_dirty_st1),
.read_tag (qual_read_tag_st1),
@@ -124,15 +125,14 @@ module VX_tag_data_access #(
.invalidate (invalidate_line),
.write_enable(use_write_enable),
.write_fill (real_writefill),
.write_addr (writeaddr_st1e[`LINE_SELECT_ADDR_RNG]),
.tag_index (writeaddr_st1e[`TAG_SELECT_ADDR_RNG]),
.write_addr (writeaddr_st1e[`LINE_SELECT_BITS-1:0]),
.tag_index (writeaddr_st1e[`TAG_LINE_ADDR_RNG]),
.write_data (use_write_data),
.fill_sent (fill_sent)
);
// VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_BITS + (`DBANK_LINE_WORDS*32) )) s0_1_c0 (
VX_generic_register #(
.N( 1 + 1 + `TAG_SELECT_BITS + (`DBANK_LINE_WORDS*32) ),
.N(1 + 1 + `TAG_SELECT_BITS + `BANK_LINE_WIDTH),
.PassThru(1)
) s0_1_c0 (
.clk (clk),
@@ -140,134 +140,120 @@ module VX_tag_data_access #(
.stall(stall),
.flush(0),
.in ({qual_read_valid_st1, qual_read_dirty_st1, qual_read_tag_st1, qual_read_data_st1}),
.out ({read_valid_st1c[0] , read_dirty_st1c[0] , read_tag_st1c[0] , read_data_st1c[0]})
.out ({read_valid_st1c[0], read_dirty_st1c[0], read_tag_st1c[0], read_data_st1c[0]})
);
genvar curr_stage;
generate
for (curr_stage = 1; curr_stage < STAGE_1_CYCLES-1; curr_stage = curr_stage + 1) begin
VX_generic_register #(
.N( 1 + 1 + `TAG_SELECT_BITS + (`DBANK_LINE_WORDS*32))
) s0_1_cc (
.clk (clk),
.reset(reset),
.stall(stall),
.flush(0),
.in ({read_valid_st1c[curr_stage-1] , read_dirty_st1c[curr_stage-1] , read_tag_st1c[curr_stage-1] , read_data_st1c[curr_stage-1]}),
.out ({read_valid_st1c[curr_stage] , read_dirty_st1c[curr_stage] , read_tag_st1c[curr_stage] , read_data_st1c[curr_stage] })
);
end
endgenerate
genvar i;
for (i = 1; i < STAGE_1_CYCLES-1; i = i + 1) begin
VX_generic_register #(
.N( 1 + 1 + `TAG_SELECT_BITS + `BANK_LINE_WIDTH)
) s0_1_cc (
.clk (clk),
.reset(reset),
.stall(stall),
.flush(0),
.in ({read_valid_st1c[i-1], read_dirty_st1c[i-1], read_tag_st1c[i-1], read_data_st1c[i-1]}),
.out ({read_valid_st1c[i], read_dirty_st1c[i], read_tag_st1c[i], read_data_st1c[i]})
);
end
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
assign use_read_tag_st1e = (FUNC_ID == `SFUNC_ID) ? writeaddr_st1e[`TAG_LINE_ADDR_RNG] : read_tag_st1c[STAGE_1_CYCLES-1]; // Tag is always the same in SM
genvar curr_w;
for (curr_w = 0; curr_w < `DBANK_LINE_WORDS; curr_w = curr_w+1) assign use_read_data_st1e[curr_w][31:0] = read_data_st1c[STAGE_1_CYCLES-1][curr_w][31:0];
// assign use_read_data_st1e = read_data_st1c [STAGE_1_CYCLES-1];
for (i = 0; i < `BANK_LINE_WORDS; i = i + 1) begin
assign use_read_data_st1e[i * `WORD_WIDTH +: `WORD_WIDTH] = read_data_st1c[STAGE_1_CYCLES-1][i * `WORD_WIDTH +: `WORD_WIDTH];
end
/////////////////////// LOAD LOGIC ///////////////////
wire[`OFFSET_SIZE_RNG] byte_select = writeaddr_st1e[`OFFSET_ADDR_RNG];
wire[`WORD_SELECT_BITS-1:0] block_offset = writeaddr_st1e[`WORD_SELECT_ADDR_RNG];
`IGNORE_WARNINGS_BEGIN
wire lw = valid_req_st1e && (mem_read_st1e == `LW_MEM_READ);
wire lb = valid_req_st1e && (mem_read_st1e == `LB_MEM_READ);
wire lh = valid_req_st1e && (mem_read_st1e == `LH_MEM_READ);
wire lhu = valid_req_st1e && (mem_read_st1e == `LHU_MEM_READ);
wire lbu = valid_req_st1e && (mem_read_st1e == `LBU_MEM_READ);
wire b0 = (byte_select == 0);
wire b1 = (byte_select == 1);
wire b2 = (byte_select == 2);
wire b3 = (byte_select == 3);
`IGNORE_WARNINGS_END
`DEBUG_BEGIN
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];
`DEBUG_END
/////////////////////// STORE LOGIC ///////////////////
assign sw = valid_req_st1e && (mem_write_st1e == `SW_MEM_WRITE);
assign sb = valid_req_st1e && (mem_write_st1e == `SB_MEM_WRITE);
assign sh = valid_req_st1e && (mem_write_st1e == `SH_MEM_WRITE);
wire[3:0] sb_mask = (b0 ? 4'b0001 : (b1 ? 4'b0010 : (b2 ? 4'b0100 : 4'b1000)));
wire[3:0] sh_mask = (b0 ? 4'b0011 : 4'b1100);
wire should_write = (sw || sb || sh) && valid_req_st1e && use_read_valid_st1e && !miss_st1e && !is_snp_st1e;
wire force_write = real_writefill;
wire[`DBANK_LINE_WORDS-1:0][3:0] we;
wire[`DBANK_LINE_WORDS-1:0][31:0] data_write;
wire [`BANK_LINE_WORDS-1:0][3:0] we;
wire [`BANK_LINE_WIDTH-1:0] data_write;
genvar g;
generate
for (g = 0; g < `DBANK_LINE_WORDS; g = g + 1) begin : write_enables
wire normal_write = (block_offset == g[`WORD_SELECT_BITS-1:0]) && should_write && !real_writefill;
if (WORD_SIZE == 4) begin
assign we[g] = (force_write) ? 4'b1111 :
(should_write && !real_writefill && (FUNC_ID == `L2FUNC_ID)) ? 4'b1111 :
(normal_write && sw) ? 4'b1111 :
(normal_write && sb) ? sb_mask :
(normal_write && sh) ? sh_mask :
4'b0000;
wire[`OFFSET_ADDR_BITS-1:0] byte_select = writewsel_st1e[`OFFSET_ADDR_RNG];
wire[`WORD_SELECT_BITS-1:0] block_offset = writewsel_st1e[`WORD_SELECT_ADDR_RNG];
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
wire lb = valid_req_st1e && (mem_read_st1e == `WORD_SEL_LB);
wire lh = valid_req_st1e && (mem_read_st1e == `WORD_SEL_LH);
wire lbu = valid_req_st1e && (mem_read_st1e == `WORD_SEL_HB);
wire lhu = valid_req_st1e && (mem_read_st1e == `WORD_SEL_HH);
wire lw = valid_req_st1e && (mem_read_st1e == `WORD_SEL_LW);
wire b0 = (byte_select == 0);
wire b1 = (byte_select == 1);
wire b2 = (byte_select == 2);
wire b3 = (byte_select == 3);
wire sb = valid_req_st1e && (mem_write_st1e == `WORD_SEL_LB);
wire sh = valid_req_st1e && (mem_write_st1e == `WORD_SEL_LH);
wire sw = valid_req_st1e && (mem_write_st1e == `WORD_SEL_LW);
wire [3:0] sb_mask = (b0 ? 4'b0001 : (b1 ? 4'b0010 : (b2 ? 4'b0100 : 4'b1000)));
wire [3:0] sh_mask = (b0 ? 4'b0011 : 4'b1100);
wire should_write = (sw || sb || sh) && valid_req_st1e && use_read_valid_st1e && !miss_st1e && !is_snp_st1e;
wire[`WORD_WIDTH-1:0] data_unmod = read_data_st1c[STAGE_1_CYCLES-1][block_offset * 32 +: 32];
wire[`WORD_WIDTH-1:0] data_unQual = (b0 || lw) ? (data_unmod) :
b1 ? (data_unmod >> 8) :
b2 ? (data_unmod >> 16) :
(data_unmod >> 24);
wire[`WORD_WIDTH-1:0] lb_data = (data_unQual[7] ) ? (data_unQual | 32'hFFFFFF00) : (data_unQual & 32'hFF);
wire[`WORD_WIDTH-1:0] lh_data = (data_unQual[15]) ? (data_unQual | 32'hFFFF0000) : (data_unQual & 32'hFFFF);
wire[`WORD_WIDTH-1:0] lbu_data = (data_unQual & 32'hFF);
wire[`WORD_WIDTH-1:0] lhu_data = (data_unQual & 32'hFFFF);
wire[`WORD_WIDTH-1:0] lw_data = (data_unQual);
wire[`WORD_WIDTH-1:0] data_Qual = lb ? lb_data :
lh ? lh_data :
lhu ? lhu_data :
lbu ? lbu_data :
lw_data;
assign readword_st1e = data_Qual;
for (i = 0; i < `BANK_LINE_WORDS; i = i + 1) begin
wire normal_write = (block_offset == i[`WORD_SELECT_BITS-1:0]) && should_write && !real_writefill;
assign we[i] = (force_write) ? 4'b1111 :
(normal_write && sw) ? 4'b1111 :
(normal_write && sb) ? sb_mask :
(normal_write && sh) ? sh_mask :
4'b0000;
wire [`WORD_WIDTH-1:0] 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 [`WORD_WIDTH-1:0] sw_data = writeword_st1e[31:0];
wire [`WORD_WIDTH-1:0] sh_data = b2 ? {writeword_st1e[15:0], {16{1'b0}}} : writeword_st1e[31:0];
wire [`WORD_WIDTH-1:0] use_write_dat = sb ? sb_data : sh ? sh_data : sw_data;
assign data_write[i * `WORD_WIDTH +: `WORD_WIDTH] = force_write ? writedata_st1e[i * `WORD_WIDTH +: `WORD_WIDTH] : use_write_dat;
end
if (FUNC_ID == `L2FUNC_ID) begin
assign data_write = force_write ? writedata_st1e : writeword_st1e;
end
endgenerate
end else begin
wire should_write = ((mem_write_st1e != `WORD_SEL_NO)) && valid_req_st1e && use_read_valid_st1e && !miss_st1e && !is_snp_st1e;
for (i = 0; i < `BANK_LINE_WORDS; i = i + 1) begin
assign we[i] = (force_write || (should_write && !real_writefill)) ? 4'b1111 : 4'b0000;
end
assign readword_st1e = read_data_st1c[STAGE_1_CYCLES-1];
assign data_write = force_write ? writedata_st1e : writeword_st1e;
end
assign use_write_enable = (writefill_st1e && !real_writefill) ? 0 : we;
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[`TAG_SELECT_BITS-1:0] writeaddr_tag = writeaddr_st1e[`TAG_LINE_ADDR_RNG];
wire tags_mismatch = writeaddr_tag != use_read_tag_st1e;
wire tags_match = writeaddr_tag == use_read_tag_st1e;
wire snoop_hit = valid_req_st1e && is_snp_st1e && use_read_valid_st1e && tags_match && use_read_dirty_st1e;
wire req_invalid = valid_req_st1e && !is_snp_st1e && !use_read_valid_st1e && !writefill_st1e;
wire req_miss = valid_req_st1e && !is_snp_st1e && use_read_valid_st1e && !writefill_st1e && tags_mismatch;
wire snoop_hit = valid_req_st1e && is_snp_st1e && use_read_valid_st1e && tags_match && use_read_dirty_st1e;
wire req_invalid = valid_req_st1e && !is_snp_st1e && !use_read_valid_st1e && !writefill_st1e;
wire req_miss = valid_req_st1e && !is_snp_st1e && use_read_valid_st1e && !writefill_st1e && tags_mismatch;
assign miss_st1e = snoop_hit || req_invalid || req_miss;
assign dirty_st1e = valid_req_st1e && use_read_valid_st1e && use_read_dirty_st1e;

View File

@@ -2,13 +2,13 @@
module VX_tag_data_structure #(
// Size of cache in bytes
parameter CACHE_SIZE_BYTES = 1024,
parameter CACHE_SIZE = 1024,
// Size of line inside a bank in bytes
parameter BANK_LINE_SIZE_BYTES = 16,
parameter BANK_LINE_SIZE = 16,
// Number of banks {1, 2, 4, 8,...}
parameter NUM_BANKS = 8,
// Size of a word in bytes
parameter WORD_SIZE_BYTES = 4,
parameter WORD_SIZE = 4,
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
@@ -37,10 +37,7 @@ module VX_tag_data_structure #(
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Dram knobs
parameter SIMULATED_DRAM_LATENCY_CYCLES = 10
parameter FILL_INVALIDAOR_SIZE = 16
) (
input wire clk,
input wire reset,
@@ -50,21 +47,21 @@ module VX_tag_data_structure #(
output wire read_valid,
output wire read_dirty,
output wire[`TAG_SELECT_BITS-1:0] read_tag,
output wire[`DBANK_LINE_WORDS-1:0][31:0] read_data,
output wire[`BANK_LINE_WIDTH-1:0] read_data,
input wire invalidate,
input wire[`DBANK_LINE_WORDS-1:0][3:0] write_enable,
input wire[`BANK_LINE_WORDS-1:0][3:0] write_enable,
input wire write_fill,
input wire[`LINE_SELECT_BITS-1:0] write_addr,
input wire[`TAG_SELECT_BITS-1:0] tag_index,
input wire[`DBANK_LINE_WORDS-1:0][31:0] write_data,
input wire[`BANK_LINE_WIDTH-1:0] write_data,
input wire fill_sent
);
reg [`DBANK_LINE_WORDS-1:0][3:0][7:0] data [`BANK_LINE_COUNT-1:0];
reg [`TAG_SELECT_BITS-1:0] tag [`BANK_LINE_COUNT-1:0];
reg valid [`BANK_LINE_COUNT-1:0];
reg dirty [`BANK_LINE_COUNT-1:0];
reg [`BANK_LINE_WORDS-1:0][3:0][`BYTE_WIDTH-1:0] data [`BANK_LINE_COUNT-1:0];
reg [`TAG_SELECT_BITS-1:0] tag [`BANK_LINE_COUNT-1:0];
reg valid [`BANK_LINE_COUNT-1:0];
reg dirty [`BANK_LINE_COUNT-1:0];
assign read_valid = valid [read_addr];
assign read_dirty = dirty [read_addr];
@@ -73,15 +70,12 @@ module VX_tag_data_structure #(
wire going_to_write = (|write_enable);
integer f;
integer l;
integer i;
always @(posedge clk) begin
if (reset) begin
for (l = 0; l < `BANK_LINE_COUNT; l=l+1) begin
valid[l] <= 0;
// tag [l] <= 0;
dirty[l] <= 0;
// data [l] <= 0;
for (i = 0; i < `BANK_LINE_COUNT; i = i + 1) begin
valid[i] <= 0;
dirty[i] <= 0;
end
end else if (!stall_bank_pipe) begin
if (going_to_write) begin
@@ -94,18 +88,17 @@ module VX_tag_data_structure #(
end
end else if (fill_sent) begin
dirty[write_addr] <= 0;
// valid[write_addr] <= 0;
end
if (invalidate) begin
valid[write_addr] <= 0;
end
for (f = 0; f < `DBANK_LINE_WORDS; f = f + 1) begin
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];
for (i = 0; i < `BANK_LINE_WORDS; i = i + 1) begin
if (write_enable[i][0]) data[write_addr][i][0] <= write_data[i * `WORD_WIDTH + 0 * `BYTE_WIDTH +: `BYTE_WIDTH];
if (write_enable[i][1]) data[write_addr][i][1] <= write_data[i * `WORD_WIDTH + 1 * `BYTE_WIDTH +: `BYTE_WIDTH];
if (write_enable[i][2]) data[write_addr][i][2] <= write_data[i * `WORD_WIDTH + 2 * `BYTE_WIDTH +: `BYTE_WIDTH];
if (write_enable[i][3]) data[write_addr][i][3] <= write_data[i * `WORD_WIDTH + 3 * `BYTE_WIDTH +: `BYTE_WIDTH];
end
end
end