wu-arch/vortex
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

using mshr pending request size

Browse Source
This commit is contained in:
Blaise Tine
2020-12-01 00:54:25 -08:00
parent 97739e9dcf
commit f68af3bb84
9 changed files with 147 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
hw/rtl/VX_cluster.v
View File

@@ -256,11 +256,6 @@ module VX_cluster #(
.csr_io_req_data_in (csr_io_req_data), .csr_io_req_data_in (csr_io_req_data),
.csr_io_req_ready_in (csr_io_req_ready), .csr_io_req_ready_in (csr_io_req_ready),
// input responses
.csr_io_rsp_valid_in (per_core_csr_io_rsp_valid),
.csr_io_rsp_data_in (per_core_csr_io_rsp_data),
.csr_io_rsp_ready_in (per_core_csr_io_rsp_ready),
// output request // output request
.csr_io_req_valid_out (per_core_csr_io_req_valid), .csr_io_req_valid_out (per_core_csr_io_req_valid),
.csr_io_req_addr_out (per_core_csr_io_req_addr), .csr_io_req_addr_out (per_core_csr_io_req_addr),
@@ -268,6 +263,11 @@ module VX_cluster #(
.csr_io_req_data_out (per_core_csr_io_req_data), .csr_io_req_data_out (per_core_csr_io_req_data),
.csr_io_req_ready_out (per_core_csr_io_req_ready), .csr_io_req_ready_out (per_core_csr_io_req_ready),
// input responses
.csr_io_rsp_valid_in (per_core_csr_io_rsp_valid),
.csr_io_rsp_data_in (per_core_csr_io_rsp_data),
.csr_io_rsp_ready_in (per_core_csr_io_rsp_ready),
// output response // output response
.csr_io_rsp_valid_out (csr_io_rsp_valid), .csr_io_rsp_valid_out (csr_io_rsp_valid),
.csr_io_rsp_data_out (csr_io_rsp_data), .csr_io_rsp_data_out (csr_io_rsp_data),

4
hw/rtl/VX_config.vh
View File

@@ -209,7 +209,7 @@
// Miss Handling Register Size // Miss Handling Register Size
`ifndef DMSHR_SIZE `ifndef DMSHR_SIZE
`define DMSHR_SIZE `MAX(`LSUQ_SIZE, 4) `define DMSHR_SIZE `LSUQ_SIZE
`endif `endif
// DRAM Request Queue Size // DRAM Request Queue Size
@@ -251,7 +251,7 @@
// Miss Handling Register Size // Miss Handling Register Size
`ifndef IMSHR_SIZE `ifndef IMSHR_SIZE
`define IMSHR_SIZE `MAX(`NUM_WARPS, 4) `define IMSHR_SIZE `NUM_WARPS
`endif `endif
// DRAM Request Queue Size // DRAM Request Queue Size

10
hw/rtl/VX_csr_io_arb.v
View File

@@ -16,11 +16,6 @@ module VX_csr_io_arb #(
input wire [31:0] csr_io_req_data_in, input wire [31:0] csr_io_req_data_in,
output wire csr_io_req_ready_in, output wire csr_io_req_ready_in,
// input response
input wire [NUM_REQUESTS-1:0] csr_io_rsp_valid_in,
input wire [NUM_REQUESTS-1:0][31:0] csr_io_rsp_data_in,
output wire [NUM_REQUESTS-1:0] csr_io_rsp_ready_in,
// output request // output request
output wire [NUM_REQUESTS-1:0] csr_io_req_valid_out, output wire [NUM_REQUESTS-1:0] csr_io_req_valid_out,
output wire [NUM_REQUESTS-1:0][11:0] csr_io_req_addr_out, output wire [NUM_REQUESTS-1:0][11:0] csr_io_req_addr_out,
@@ -28,6 +23,11 @@ module VX_csr_io_arb #(
output wire [NUM_REQUESTS-1:0][31:0] csr_io_req_data_out, output wire [NUM_REQUESTS-1:0][31:0] csr_io_req_data_out,
input wire [NUM_REQUESTS-1:0] csr_io_req_ready_out, input wire [NUM_REQUESTS-1:0] csr_io_req_ready_out,
// input response
input wire [NUM_REQUESTS-1:0] csr_io_rsp_valid_in,
input wire [NUM_REQUESTS-1:0][31:0] csr_io_rsp_data_in,
output wire [NUM_REQUESTS-1:0] csr_io_rsp_ready_in,
// output response // output response
output wire csr_io_rsp_valid_out, output wire csr_io_rsp_valid_out,
output wire [31:0] csr_io_rsp_data_out, output wire [31:0] csr_io_rsp_data_out,

10
hw/rtl/Vortex.v
View File

@@ -297,11 +297,6 @@ module Vortex (
.csr_io_req_data_in (csr_io_req_data), .csr_io_req_data_in (csr_io_req_data),
.csr_io_req_ready_in (csr_io_req_ready), .csr_io_req_ready_in (csr_io_req_ready),
// input responses
.csr_io_rsp_valid_in (per_cluster_csr_io_rsp_valid),
.csr_io_rsp_data_in (per_cluster_csr_io_rsp_data),
.csr_io_rsp_ready_in (per_cluster_csr_io_rsp_ready),
// output request // output request
.csr_io_req_valid_out (per_cluster_csr_io_req_valid), .csr_io_req_valid_out (per_cluster_csr_io_req_valid),
.csr_io_req_addr_out (per_cluster_csr_io_req_addr), .csr_io_req_addr_out (per_cluster_csr_io_req_addr),
@@ -309,6 +304,11 @@ module Vortex (
.csr_io_req_data_out (per_cluster_csr_io_req_data), .csr_io_req_data_out (per_cluster_csr_io_req_data),
.csr_io_req_ready_out (per_cluster_csr_io_req_ready), .csr_io_req_ready_out (per_cluster_csr_io_req_ready),
// input responses
.csr_io_rsp_valid_in (per_cluster_csr_io_rsp_valid),
.csr_io_rsp_data_in (per_cluster_csr_io_rsp_data),
.csr_io_rsp_ready_in (per_cluster_csr_io_rsp_ready),
// output response // output response
.csr_io_rsp_valid_out (csr_io_rsp_valid), .csr_io_rsp_valid_out (csr_io_rsp_valid),
.csr_io_rsp_data_out (csr_io_rsp_data), .csr_io_rsp_data_out (csr_io_rsp_data),

188
hw/rtl/cache/VX_bank.v vendored
View File

@@ -214,20 +214,20 @@ module VX_bank #(
assign dram_rsp_ready = 0; assign dram_rsp_ready = 0;
end end
wire reqq_pop; wire creq_pop;
wire reqq_empty; wire creq_empty;
wire reqq_full; wire creq_full;
wire [`REQS_BITS-1:0] reqq_tid_st0; wire [`REQS_BITS-1:0] creq_tid_st0;
wire reqq_rw_st0; wire creq_rw_st0;
wire [WORD_SIZE-1:0] reqq_byteen_st0; wire [WORD_SIZE-1:0] creq_byteen_st0;
`IGNORE_WARNINGS_BEGIN `IGNORE_WARNINGS_BEGIN
wire [`WORD_ADDR_WIDTH-1:0] reqq_addr_st0; wire [`WORD_ADDR_WIDTH-1:0] creq_addr_st0;
`IGNORE_WARNINGS_END `IGNORE_WARNINGS_END
wire [`WORD_WIDTH-1:0] reqq_writeword_st0; wire [`WORD_WIDTH-1:0] creq_writeword_st0;
wire [CORE_TAG_WIDTH-1:0] reqq_tag_st0; wire [CORE_TAG_WIDTH-1:0] creq_tag_st0;
wire core_req_fire = (| core_req_valid) && core_req_ready; wire core_req_fire = (| core_req_valid) && core_req_ready;
assign core_req_ready = !reqq_full; assign core_req_ready = !creq_full;
VX_bank_core_req_arb #( VX_bank_core_req_arb #(
.WORD_SIZE (WORD_SIZE), .WORD_SIZE (WORD_SIZE),
@@ -249,22 +249,21 @@ module VX_bank #(
.writedata_in (core_req_data), .writedata_in (core_req_data),
// Dequeue // Dequeue
.pop (reqq_pop), .pop (creq_pop),
.tag_out (reqq_tag_st0), .tag_out (creq_tag_st0),
.tid_out (reqq_tid_st0), .tid_out (creq_tid_st0),
.rw_out (reqq_rw_st0), .rw_out (creq_rw_st0),
.byteen_out (reqq_byteen_st0), .byteen_out (creq_byteen_st0),
.addr_out (reqq_addr_st0), .addr_out (creq_addr_st0),
.writedata_out (reqq_writeword_st0), .writedata_out (creq_writeword_st0),
// States // States
.empty (reqq_empty), .empty (creq_empty),
.full (reqq_full) .full (creq_full)
); );
reg [$clog2(MSHR_SIZE+1)-1:0] mshr_pending_size;
wire mshr_pop; wire mshr_pop;
wire mshr_full;
wire mshr_almfull;
wire mshr_valid_st0; wire mshr_valid_st0;
wire[`REQS_BITS-1:0] mshr_tid_st0; wire[`REQS_BITS-1:0] mshr_tid_st0;
wire [`LINE_ADDR_WIDTH-1:0] mshr_addr_st0; wire [`LINE_ADDR_WIDTH-1:0] mshr_addr_st0;
@@ -275,26 +274,6 @@ module VX_bank #(
wire [WORD_SIZE-1:0] mshr_byteen_st0; wire [WORD_SIZE-1:0] mshr_byteen_st0;
wire mshr_is_snp_st0; wire mshr_is_snp_st0;
wire mshr_snp_inv_st0; wire mshr_snp_inv_st0;
wire is_mshr_miss_st2;
wire is_mshr_miss_st3;
wire mshr_push_stall;
wire cwbq_push_stall;
wire dwbq_push_stall;
wire snpq_push_stall;
wire pipeline_stall;
// determine which queue to pop next in piority order
wire mshr_pop_unqual = mshr_valid_st0;
wire dfpq_pop_unqual = !mshr_pop_unqual && !dfpq_empty;
wire reqq_pop_unqual = !mshr_pop_unqual && !dfpq_pop_unqual && !reqq_empty && !mshr_almfull;
wire snrq_pop_unqual = !mshr_pop_unqual && !dfpq_pop_unqual && !reqq_pop_unqual && !snrq_empty && !mshr_almfull;
assign mshr_pop = mshr_pop_unqual && !pipeline_stall
&& !(is_mshr_miss_st2 || is_mshr_miss_st3); // stop if previous request was a miss
assign dfpq_pop = dfpq_pop_unqual && !pipeline_stall;
assign reqq_pop = reqq_pop_unqual && !pipeline_stall;
assign snrq_pop = snrq_pop_unqual && !pipeline_stall;
wire is_fill_st0; wire is_fill_st0;
wire is_mshr_st0; wire is_mshr_st0;
@@ -350,24 +329,64 @@ module VX_bank #(
wire force_miss_st2; wire force_miss_st2;
wire[`LINE_ADDR_WIDTH-1:0] addr_st2; wire[`LINE_ADDR_WIDTH-1:0] addr_st2;
wire writeen_st2; wire writeen_st2;
wire core_req_hit_st2;
wire valid_st3;
wire is_mshr_st3;
wire miss_st3; wire miss_st3;
wire force_miss_st3; wire force_miss_st3;
wire [`LINE_ADDR_WIDTH-1:0] addr_st3; wire [`LINE_ADDR_WIDTH-1:0] addr_st3;
wire core_req_hit_st1;
wire mshr_push_stall;
wire cwbq_push_stall;
wire dwbq_push_stall;
wire snpq_push_stall;
wire pipeline_stall;
wire is_mshr_miss_st2 = valid_st2 && is_mshr_st2 && (miss_st2 || force_miss_st2);
wire is_mshr_miss_st3 = valid_st3 && is_mshr_st3 && (miss_st3 || force_miss_st3);
wire creq_commit = valid_st1 && core_req_hit_st1 && !pipeline_stall;
wire mshr_going_full = (mshr_pending_size == MSHR_SIZE);
// determine which queue to pop next in piority order
wire mshr_pop_unqual = mshr_valid_st0;
wire dfpq_pop_unqual = !mshr_pop_unqual && !dfpq_empty;
wire creq_pop_unqual = !mshr_pop_unqual && !dfpq_pop_unqual && !creq_empty && !mshr_going_full;
wire snrq_pop_unqual = !mshr_pop_unqual && !dfpq_pop_unqual && !creq_pop_unqual && !snrq_empty && !mshr_going_full;
assign mshr_pop = mshr_pop_unqual && !pipeline_stall
&& !(is_mshr_miss_st2 || is_mshr_miss_st3); // stop if previous request was a miss
assign dfpq_pop = dfpq_pop_unqual && !pipeline_stall;
assign creq_pop = creq_pop_unqual && !pipeline_stall;
assign snrq_pop = snrq_pop_unqual && !pipeline_stall;
// MSHR pending size
always @(posedge clk) begin
if (reset) begin
mshr_pending_size <= 0;
end else begin
mshr_pending_size <= mshr_pending_size +
((creq_pop && !creq_commit) ? 1 : ((creq_commit && !creq_pop) ? -1 : 0));
end
end
assign is_mshr_st0 = mshr_pop_unqual; assign is_mshr_st0 = mshr_pop_unqual;
assign is_fill_st0 = dfpq_pop_unqual; assign is_fill_st0 = dfpq_pop_unqual;
assign valid_st0 = dfpq_pop || mshr_pop || reqq_pop || snrq_pop; assign valid_st0 = dfpq_pop || mshr_pop || creq_pop || snrq_pop;
assign addr_st0 = mshr_pop_unqual ? mshr_addr_st0 : assign addr_st0 = mshr_pop_unqual ? mshr_addr_st0 :
dfpq_pop_unqual ? dfpq_addr_st0 : dfpq_pop_unqual ? dfpq_addr_st0 :
reqq_pop_unqual ? reqq_addr_st0[`LINE_SELECT_ADDR_RNG] : creq_pop_unqual ? creq_addr_st0[`LINE_SELECT_ADDR_RNG] :
snrq_pop_unqual ? snrq_addr_st0 : snrq_pop_unqual ? snrq_addr_st0 :
0; 0;
if (`WORD_SELECT_WIDTH != 0) begin if (`WORD_SELECT_WIDTH != 0) begin
assign wsel_st0 = reqq_pop_unqual ? reqq_addr_st0[`WORD_SELECT_WIDTH-1:0] : assign wsel_st0 = creq_pop_unqual ? creq_addr_st0[`WORD_SELECT_WIDTH-1:0] :
mshr_pop_unqual ? mshr_wsel_st0 : mshr_pop_unqual ? mshr_wsel_st0 :
0; 0;
end else begin end else begin
@@ -378,7 +397,7 @@ module VX_bank #(
assign writedata_st0 = dfpq_filldata_st0; assign writedata_st0 = dfpq_filldata_st0;
assign inst_meta_st0 = mshr_pop_unqual ? {`REQ_TAG_WIDTH'(mshr_tag_st0), mshr_rw_st0, mshr_byteen_st0, mshr_tid_st0} : assign inst_meta_st0 = mshr_pop_unqual ? {`REQ_TAG_WIDTH'(mshr_tag_st0), mshr_rw_st0, mshr_byteen_st0, mshr_tid_st0} :
reqq_pop_unqual ? {`REQ_TAG_WIDTH'(reqq_tag_st0), reqq_rw_st0, reqq_byteen_st0, reqq_tid_st0} : creq_pop_unqual ? {`REQ_TAG_WIDTH'(creq_tag_st0), creq_rw_st0, creq_byteen_st0, creq_tid_st0} :
snrq_pop_unqual ? {`REQ_TAG_WIDTH'(snrq_tag_st0), 1'b0, WORD_SIZE'(0), `REQS_BITS'(0)} : snrq_pop_unqual ? {`REQ_TAG_WIDTH'(snrq_tag_st0), 1'b0, WORD_SIZE'(0), `REQS_BITS'(0)} :
0; 0;
@@ -391,7 +410,7 @@ module VX_bank #(
0; 0;
assign writeword_st0 = mshr_pop_unqual ? mshr_writeword_st0 : assign writeword_st0 = mshr_pop_unqual ? mshr_writeword_st0 :
reqq_pop_unqual ? reqq_writeword_st0 : creq_pop_unqual ? creq_writeword_st0 :
0; 0;
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
@@ -432,12 +451,10 @@ if (DRAM_ENABLE) begin
assign {tag_st1, mem_rw_st1, mem_byteen_st1, tid_st1} = inst_meta_st1; assign {tag_st1, mem_rw_st1, mem_byteen_st1, tid_st1} = inst_meta_st1;
// we have a matching previous request that missed alreedy
wire st2_pending_hazard_st1 = valid_st2 && (miss_st2 || force_miss_st2) && (addr_st2 == addr_st1);
wire st3_pending_hazard_st1 = valid_st3 && (miss_st3 || force_miss_st3) && (addr_st3 == addr_st1);
// force miss to ensure commit order when a new request has pending previous requests to same block // force miss to ensure commit order when a new request has pending previous requests to same block
// also force a miss for msrq requests when previous requests got a miss // also force a miss for msrq requests when previous requests got a miss
wire st2_pending_hazard_st1 = valid_st2 && (miss_st2 || force_miss_st2) && (addr_st2 == addr_st1);
wire st3_pending_hazard_st1 = valid_st3 && (miss_st3 || force_miss_st3) && (addr_st3 == addr_st1);
assign force_miss_st1 = (valid_st1 && !is_mshr_st1 && !is_fill_st1 assign force_miss_st1 = (valid_st1 && !is_mshr_st1 && !is_fill_st1
&& (mshr_pending_hazard_st1 || st2_pending_hazard_st1 || st3_pending_hazard_st1)) && (mshr_pending_hazard_st1 || st2_pending_hazard_st1 || st3_pending_hazard_st1))
|| (valid_st1 && is_mshr_st1 && is_mshr_miss_st2); || (valid_st1 && is_mshr_st1 && is_mshr_miss_st2);
@@ -481,23 +498,26 @@ if (DRAM_ENABLE) begin
.writeen_out (writeen_st1) .writeen_out (writeen_st1)
); );
assign core_req_hit_st1 = !is_fill_st1 && !is_snp_st1 && !miss_st1 && !force_miss_st1;
assign misses = miss_st1; assign misses = miss_st1;
VX_generic_register #( VX_generic_register #(
.N(1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + `TAG_SELECT_BITS + 1 + 1 + `BANK_LINE_WIDTH + WORD_SIZE + `REQ_INST_META_WIDTH), .N(1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + `TAG_SELECT_BITS + 1 + `BANK_LINE_WIDTH + WORD_SIZE + `REQ_INST_META_WIDTH),
.R(1) .R(1)
) pipe_reg1 ( ) pipe_reg1 (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.stall (pipeline_stall), .stall (pipeline_stall),
.flush (1'b0), .flush (1'b0),
.in ({valid_st1, is_mshr_st1, writeen_st1, force_miss_st1, is_snp_st1, snp_inv_st1, is_fill_st1, addr_st1, wsel_st1, writeword_st1, readtag_st1, miss_st1, dirty_st1, writedata_st1, mem_byteen_st1, inst_meta_st1}), .in ({valid_st1, core_req_hit_st1, is_mshr_st1, writeen_st1, force_miss_st1, dirty_st1, is_snp_st1, snp_inv_st1, is_fill_st1, addr_st1, wsel_st1, writeword_st1, readtag_st1, miss_st1, writedata_st1, mem_byteen_st1, inst_meta_st1}),
.out ({valid_st2, is_mshr_st2, writeen_st2, force_miss_st2, is_snp_st2, snp_inv_st2, is_fill_st2, addr_st2, wsel_st2, writeword_st2, readtag_st2, miss_st2, dirty_st2, writedata_st2, mem_byteen_st2, inst_meta_st2}) .out ({valid_st2, core_req_hit_st2, is_mshr_st2, writeen_st2, force_miss_st2, dirty_st2, is_snp_st2, snp_inv_st2, is_fill_st2, addr_st2, wsel_st2, writeword_st2, readtag_st2, miss_st2, writedata_st2, mem_byteen_st2, inst_meta_st2})
); );
end else begin end else begin
`UNUSED_VAR (mshr_pending_hazard_unqual_st0) `UNUSED_VAR (mshr_pending_hazard_unqual_st0)
`UNUSED_VAR (dram_rsp_fire)
`UNUSED_VAR (addr_st0) `UNUSED_VAR (addr_st0)
assign {tag_st1, mem_rw_st1, mem_byteen_st1, tid_st1} = inst_meta_st1; assign {tag_st1, mem_rw_st1, mem_byteen_st1, tid_st1} = inst_meta_st1;
@@ -511,7 +531,7 @@ end else begin
assign writedata_st1= writedata_st0; assign writedata_st1= writedata_st0;
assign inst_meta_st1= inst_meta_st0; assign inst_meta_st1= inst_meta_st0;
assign snp_inv_st1 = snp_inv_st0; assign snp_inv_st1 = snp_inv_st0;
assign addr_st1 = reqq_addr_st0[`LINE_SELECT_ADDR_RNG]; assign addr_st1 = creq_addr_st0[`LINE_SELECT_ADDR_RNG];
assign dirty_st1 = 0; assign dirty_st1 = 0;
assign readtag_st1 = 0; assign readtag_st1 = 0;
assign miss_st1 = 0; assign miss_st1 = 0;
@@ -528,13 +548,19 @@ end else begin
assign inst_meta_st2= inst_meta_st1; assign inst_meta_st2= inst_meta_st1;
assign snp_inv_st2 = snp_inv_st1; assign snp_inv_st2 = snp_inv_st1;
assign addr_st2 = addr_st1; assign addr_st2 = addr_st1;
assign mem_byteen_st2 = mem_byteen_st1;
assign dirty_st2 = dirty_st1; assign dirty_st2 = dirty_st1;
assign mem_byteen_st2 = mem_byteen_st1;
assign readtag_st2 = readtag_st1; assign readtag_st2 = readtag_st1;
assign miss_st2 = miss_st1; assign miss_st2 = miss_st1;
assign writeen_st2 = writeen_st1; assign writeen_st2 = writeen_st1;
assign force_miss_st2 = force_miss_st1; assign force_miss_st2 = force_miss_st1;
assign core_req_hit_st1 = 0;
assign core_req_hit_st2 = 0;
assign send_dwb_req_st2 = 0;
assign do_writeback_st2 = 0;
assign incoming_fill_st2 = 0;
assign misses = 0; assign misses = 0;
end end
@@ -547,8 +573,6 @@ end
end end
`endif `endif
assign is_mshr_miss_st2 = valid_st2 && is_mshr_st2 && (miss_st2 || force_miss_st2);
VX_data_access #( VX_data_access #(
.BANK_ID (BANK_ID), .BANK_ID (BANK_ID),
.CACHE_ID (CACHE_ID), .CACHE_ID (CACHE_ID),
@@ -587,7 +611,6 @@ end
.dirtyb_out (dirtyb_st2) .dirtyb_out (dirtyb_st2)
); );
wire valid_st3;
wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st3; wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st3;
wire [`WORD_WIDTH-1:0] writeword_st3; wire [`WORD_WIDTH-1:0] writeword_st3;
wire [`WORD_WIDTH-1:0] readword_st3; wire [`WORD_WIDTH-1:0] readword_st3;
@@ -597,15 +620,11 @@ end
wire [`TAG_SELECT_BITS-1:0] readtag_st3; wire [`TAG_SELECT_BITS-1:0] readtag_st3;
wire is_snp_st3; wire is_snp_st3;
wire snp_inv_st3; wire snp_inv_st3;
wire is_mshr_st3; wire core_req_hit_st3;
wire send_core_rsp_st3;
wire send_dwb_req_st3; wire send_dwb_req_st3;
wire do_writeback_st3; wire do_writeback_st3;
wire send_snp_rsp_st3;
wire incoming_fill_st3; wire incoming_fill_st3;
wire send_core_rsp_st2 = !is_fill_st2 && !is_snp_st2 && !miss_st2 && !force_miss_st2;
// check if a matching fill request is comming // check if a matching fill request is comming
wire incoming_fill_dfp_st2 = dram_rsp_fire && (addr_st2 == dram_rsp_addr); wire incoming_fill_dfp_st2 = dram_rsp_fire && (addr_st2 == dram_rsp_addr);
wire incoming_fill_st0_st2 = !dfpq_empty && (addr_st2 == dfpq_addr_st0); wire incoming_fill_st0_st2 = !dfpq_empty && (addr_st2 == dfpq_addr_st0);
@@ -625,18 +644,16 @@ end
wire send_dwb_req_st2 = send_fill_req_st2 || do_writeback_st2; wire send_dwb_req_st2 = send_fill_req_st2 || do_writeback_st2;
wire send_snp_rsp_st2 = is_snp_st2 && !force_miss_st2;
VX_generic_register #( VX_generic_register #(
.N(1 + 1+ 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + `WORD_WIDTH + `BANK_LINE_WIDTH + `TAG_SELECT_BITS + 1 + 1 + BANK_LINE_SIZE + `REQ_INST_META_WIDTH), .N(1 + 1+ 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + `WORD_WIDTH + `BANK_LINE_WIDTH + `TAG_SELECT_BITS + 1 + 1 + BANK_LINE_SIZE + `REQ_INST_META_WIDTH),
.R(1) .R(1)
) pipe_reg2 ( ) pipe_reg2 (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.stall (pipeline_stall), .stall (pipeline_stall),
.flush (1'b0), .flush (1'b0),
.in ({valid_st2, send_core_rsp_st2, send_dwb_req_st2, do_writeback_st2, send_snp_rsp_st2, incoming_fill_st2, force_miss_st2, is_mshr_st2, is_snp_st2, snp_inv_st2, addr_st2, wsel_st2, writeword_st2, readword_st2, readdata_st2, readtag_st2, miss_st2, dirtyb_st2, inst_meta_st2}), .in ({valid_st2, core_req_hit_st2, send_dwb_req_st2, do_writeback_st2, incoming_fill_st2, force_miss_st2, is_mshr_st2, is_snp_st2, snp_inv_st2, addr_st2, wsel_st2, writeword_st2, readword_st2, readdata_st2, readtag_st2, miss_st2, dirtyb_st2, inst_meta_st2}),
.out ({valid_st3, send_core_rsp_st3, send_dwb_req_st3, do_writeback_st3, send_snp_rsp_st3, incoming_fill_st3, force_miss_st3, is_mshr_st3, is_snp_st3, snp_inv_st3, addr_st3, wsel_st3, writeword_st3, readword_st3, readdata_st3, readtag_st3, miss_st3, dirtyb_st3, inst_meta_st3}) .out ({valid_st3, core_req_hit_st3, send_dwb_req_st3, do_writeback_st3, incoming_fill_st3, force_miss_st3, is_mshr_st3, is_snp_st3, snp_inv_st3, addr_st3, wsel_st3, writeword_st3, readword_st3, readdata_st3, readtag_st3, miss_st3, dirtyb_st3, inst_meta_st3})
); );
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
@@ -647,8 +664,6 @@ end
end end
`endif `endif
assign is_mshr_miss_st3 = valid_st3 && is_mshr_st3 && (miss_st3 || force_miss_st3);
// Enqueue to miss reserv if it's a valid miss // Enqueue to miss reserv if it's a valid miss
wire[`REQS_BITS-1:0] req_tid_st3; wire[`REQS_BITS-1:0] req_tid_st3;
@@ -657,17 +672,22 @@ end
wire[WORD_SIZE-1:0] req_byteen_st3; wire[WORD_SIZE-1:0] req_byteen_st3;
wire mshr_push_unqual = valid_st3 && (miss_st3 || force_miss_st3); wire mshr_push_unqual = valid_st3 && (miss_st3 || force_miss_st3);
assign mshr_push_stall = mshr_push_unqual && mshr_full; assign mshr_push_stall = 0;
wire mshr_push = mshr_push_unqual wire mshr_push = mshr_push_unqual
&& !mshr_full
&& !cwbq_push_stall && !cwbq_push_stall
&& !dwbq_push_stall && !dwbq_push_stall
&& !snpq_push_stall; && !snpq_push_stall;
wire mshr_full;
always @(posedge clk) begin
assert(!mshr_push || !mshr_full); // mmshr stall is detected before issuing new requests
end
assign {req_tag_st3, req_rw_st3, req_byteen_st3, req_tid_st3} = inst_meta_st3; assign {req_tag_st3, req_rw_st3, req_byteen_st3, req_tid_st3} = inst_meta_st3;
if (DRAM_ENABLE) begin if (DRAM_ENABLE) begin
wire mshr_dequeue_st3 = valid_st3 && is_mshr_st3 && !mshr_push_unqual && !pipeline_stall; wire mshr_dequeue_st3 = valid_st3 && is_mshr_st3 && !mshr_push_unqual && !pipeline_stall;
// mark msrq entry that match DRAM fill as 'ready' // mark msrq entry that match DRAM fill as 'ready'
@@ -677,7 +697,7 @@ end
// or the fill request is comming for the missed block // or the fill request is comming for the missed block
wire mshr_init_ready_state_st3 = valid_st3 && (!miss_st3 || incoming_fill_st3); wire mshr_init_ready_state_st3 = valid_st3 && (!miss_st3 || incoming_fill_st3);
VX_cache_miss_resrv #( VX_miss_resrv #(
.BANK_ID (BANK_ID), .BANK_ID (BANK_ID),
.CACHE_ID (CACHE_ID), .CACHE_ID (CACHE_ID),
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS), .CORE_TAG_ID_BITS (CORE_TAG_ID_BITS),
@@ -688,7 +708,7 @@ end
.MSHR_SIZE (MSHR_SIZE), .MSHR_SIZE (MSHR_SIZE),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH), .CORE_TAG_WIDTH (CORE_TAG_WIDTH),
.SNP_TAG_WIDTH (SNP_TAG_WIDTH) .SNP_TAG_WIDTH (SNP_TAG_WIDTH)
) cache_miss_resrv ( ) miss_resrv (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -717,7 +737,6 @@ end
.enqueue_mshr_st3 (is_mshr_st3), .enqueue_mshr_st3 (is_mshr_st3),
.enqueue_ready_st3 (mshr_init_ready_state_st3), .enqueue_ready_st3 (mshr_init_ready_state_st3),
.enqueue_full (mshr_full), .enqueue_full (mshr_full),
.enqueue_almfull (mshr_almfull),
// fill // fill
.update_ready_st0 (update_ready_st0), .update_ready_st0 (update_ready_st0),
@@ -749,7 +768,6 @@ end
`UNUSED_VAR (incoming_fill_st3) `UNUSED_VAR (incoming_fill_st3)
assign mshr_pending_hazard_unqual_st0 = 0; assign mshr_pending_hazard_unqual_st0 = 0;
assign mshr_full = 0; assign mshr_full = 0;
assign mshr_almfull = 0;
assign mshr_valid_st0 = 0; assign mshr_valid_st0 = 0;
assign mshr_addr_st0 = 0; assign mshr_addr_st0 = 0;
assign mshr_wsel_st0 = 0; assign mshr_wsel_st0 = 0;
@@ -766,7 +784,7 @@ end
wire cwbq_empty, cwbq_full; wire cwbq_empty, cwbq_full;
wire cwbq_push_unqual = valid_st3 && send_core_rsp_st3 && !req_rw_st3; wire cwbq_push_unqual = valid_st3 && core_req_hit_st3 && !req_rw_st3;
assign cwbq_push_stall = cwbq_push_unqual && cwbq_full; assign cwbq_push_stall = cwbq_push_unqual && cwbq_full;
wire cwbq_push = cwbq_push_unqual wire cwbq_push = cwbq_push_unqual
@@ -823,7 +841,8 @@ end
if (DRAM_ENABLE) begin if (DRAM_ENABLE) begin
VX_generic_queue #( VX_generic_queue #(
.DATAW(1 + BANK_LINE_SIZE + `LINE_ADDR_WIDTH + `BANK_LINE_WIDTH), .DATAW(1 + BANK_LINE_SIZE + `LINE_ADDR_WIDTH + `BANK_LINE_WIDTH),
.SIZE(DREQ_SIZE) .SIZE(DREQ_SIZE),
.BUFFERED(1)
) dwb_queue ( ) dwb_queue (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -858,7 +877,7 @@ end
wire snpq_empty, snpq_full; wire snpq_empty, snpq_full;
wire snpq_push_unqual = valid_st3 && send_snp_rsp_st3; wire snpq_push_unqual = valid_st3 && is_snp_st3 && !force_miss_st3;
assign snpq_push_stall = snpq_push_unqual && snpq_full; assign snpq_push_stall = snpq_push_unqual && snpq_full;
@@ -875,7 +894,8 @@ end
if (FLUSH_ENABLE) begin if (FLUSH_ENABLE) begin
VX_generic_queue #( VX_generic_queue #(
.DATAW (SNP_TAG_WIDTH), .DATAW (SNP_TAG_WIDTH),
.SIZE (SNPQ_SIZE) .SIZE (SNPQ_SIZE),
.BUFFERED(1)
) snp_rsp_queue ( ) snp_rsp_queue (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -936,11 +956,11 @@ end
if (dfpq_pop) begin if (dfpq_pop) begin
$display("%t: cache%0d:%0d fill-rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), dfpq_filldata_st0); $display("%t: cache%0d:%0d fill-rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), dfpq_filldata_st0);
end end
if (reqq_pop) begin if (creq_pop) begin
if (reqq_rw_st0) if (creq_rw_st0)
$display("%t: cache%0d:%0d core-wr-req: addr=%0h, tag=%0h, tid=%0d, byteen=%b, data=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), reqq_tag_st0, reqq_tid_st0, reqq_byteen_st0, reqq_writeword_st0, debug_wid_st0, debug_pc_st0); $display("%t: cache%0d:%0d core-wr-req: addr=%0h, tag=%0h, tid=%0d, byteen=%b, data=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), creq_tag_st0, creq_tid_st0, creq_byteen_st0, creq_writeword_st0, debug_wid_st0, debug_pc_st0);
else else
$display("%t: cache%0d:%0d core-rd-req: addr=%0h, tag=%0h, tid=%0d, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), reqq_tag_st0, reqq_tid_st0, reqq_byteen_st0, reqq_writeword_st0, debug_wid_st0, debug_pc_st0); $display("%t: cache%0d:%0d core-rd-req: addr=%0h, tag=%0h, tid=%0d, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), creq_tag_st0, creq_tid_st0, creq_byteen_st0, creq_writeword_st0, debug_wid_st0, debug_pc_st0);
end end
if (snrq_pop) begin if (snrq_pop) begin
$display("%t: cache%0d:%0d snp-req: addr=%0h, tag=%0h, invalidate=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), snrq_tag_st0, snrq_inv_st0); $display("%t: cache%0d:%0d snp-req: addr=%0h, tag=%0h, invalidate=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), snrq_tag_st0, snrq_inv_st0);

10
hw/rtl/cache/VX_cache_miss_resrv.v → hw/rtl/cache/VX_miss_resrv.v vendored
View File

@@ -1,6 +1,6 @@
`include "VX_cache_config.vh" `include "VX_cache_config.vh"
module VX_cache_miss_resrv #( module VX_miss_resrv #(
parameter CACHE_ID = 0, parameter CACHE_ID = 0,
parameter BANK_ID = 0, parameter BANK_ID = 0,
@@ -51,7 +51,6 @@ module VX_cache_miss_resrv #(
input wire enqueue_mshr_st3, input wire enqueue_mshr_st3,
input wire enqueue_ready_st3, input wire enqueue_ready_st3,
output wire enqueue_full, output wire enqueue_full,
output wire enqueue_almfull,
// fill // fill
input wire update_ready_st0, input wire update_ready_st0,
@@ -72,9 +71,8 @@ module VX_cache_miss_resrv #(
output wire dequeue_snp_inv_st0, output wire dequeue_snp_inv_st0,
input wire dequeue_st3 input wire dequeue_st3
); );
localparam FULL_DISTANCE = 3; // need 3 cycles window to prevent pipeline lock
wire [`MSHR_METADATA_WIDTH-1:0] metadata_table; wire [`MSHR_METADATA_WIDTH-1:0] metadata_table;
`NO_RW_RAM_CHECK reg [`LINE_ADDR_WIDTH-1:0] addr_table [MSHR_SIZE-1:0]; `NO_RW_RAM_CHECK reg [`LINE_ADDR_WIDTH-1:0] addr_table [MSHR_SIZE-1:0];
reg [MSHR_SIZE-1:0] valid_table; reg [MSHR_SIZE-1:0] valid_table;
@@ -82,13 +80,9 @@ module VX_cache_miss_resrv #(
reg [`LOG2UP(MSHR_SIZE)-1:0] schedule_ptr, restore_ptr; reg [`LOG2UP(MSHR_SIZE)-1:0] schedule_ptr, restore_ptr;
reg [`LOG2UP(MSHR_SIZE)-1:0] head_ptr; reg [`LOG2UP(MSHR_SIZE)-1:0] head_ptr;
reg [`LOG2UP(MSHR_SIZE)-1:0] tail_ptr; reg [`LOG2UP(MSHR_SIZE)-1:0] tail_ptr;
reg [`LOG2UP(MSHR_SIZE+1)-1:0] size; reg [`LOG2UP(MSHR_SIZE+1)-1:0] size;
`STATIC_ASSERT(MSHR_SIZE > FULL_DISTANCE, ("invalid size"))
assign enqueue_full = (size == $bits(size)'(MSHR_SIZE)); assign enqueue_full = (size == $bits(size)'(MSHR_SIZE));
assign enqueue_almfull = (size >= $bits(size)'(MSHR_SIZE-FULL_DISTANCE));
wire [MSHR_SIZE-1:0] valid_address_match; wire [MSHR_SIZE-1:0] valid_address_match;
for (genvar i = 0; i < MSHR_SIZE; i++) begin for (genvar i = 0; i < MSHR_SIZE; i++) begin

2
hw/rtl/cache/VX_tag_access.v vendored
View File

@@ -85,7 +85,7 @@ module VX_tag_access #(
); );
// use "case equality" to handle uninitialized tag when block entry is not valid // use "case equality" to handle uninitialized tag when block entry is not valid
wire tags_match = read_valid && (addrtag === read_tag); wire tags_match = read_valid && (addrtag == read_tag);
assign do_write = WRITE_ENABLE assign do_write = WRITE_ENABLE
&& valid_in && valid_in

1
hw/syn/quartus/project.tcl
View File

@@ -40,7 +40,6 @@ set_global_assignment -name VERILOG_MACRO NDEBUG
set_global_assignment -name MESSAGE_DISABLE 16818 set_global_assignment -name MESSAGE_DISABLE 16818
set_global_assignment -name VERILOG_MACRO FPU_FAST set_global_assignment -name VERILOG_MACRO FPU_FAST
set_global_assignment -name AUTO_SHIFT_REGISTER_RECOGNITION AUTO
set_global_assignment -name MIN_CORE_JUNCTION_TEMP 0 set_global_assignment -name MIN_CORE_JUNCTION_TEMP 0
set_global_assignment -name MAX_CORE_JUNCTION_TEMP 100 set_global_assignment -name MAX_CORE_JUNCTION_TEMP 100
set_global_assignment -name POWER_BOARD_THERMAL_MODEL "NONE (CONSERVATIVE)" set_global_assignment -name POWER_BOARD_THERMAL_MODEL "NONE (CONSERVATIVE)"