This commit is contained in:
Blaise Tine 2020-12-31 22:40:34 -05:00
commit 138db29310
17 changed files with 376 additions and 396 deletions

View file

@ -116,7 +116,7 @@ case $APP in
APP_PATH=$VORTEX_HOME/benchmarks/opencl/sgemm APP_PATH=$VORTEX_HOME/benchmarks/opencl/sgemm
;; ;;
vecadd) vecadd)
APP_PATH=$VORTEX_HOME/benchmarks/opencl/vacadd APP_PATH=$VORTEX_HOME/benchmarks/opencl/vecadd
;; ;;
basic) basic)
APP_PATH=$VORTEX_HOME/driver/tests/basic APP_PATH=$VORTEX_HOME/driver/tests/basic

View file

@ -292,7 +292,7 @@ extern int vx_dump_perf(vx_device_h device, FILE* stream) {
int avg_dram_lat = (int)(double(dram_lat_per_core) / double(dram_reads_per_core)); int avg_dram_lat = (int)(double(dram_lat_per_core) / double(dram_reads_per_core));
int dram_utilization = (int)((1.0 - (double(dram_reads_per_core + dram_writes_per_core) / double(dram_reads_per_core + dram_writes_per_core + dram_stalls_per_core))) * 100); int dram_utilization = (int)((1.0 - (double(dram_reads_per_core + dram_writes_per_core) / double(dram_reads_per_core + dram_writes_per_core + dram_stalls_per_core))) * 100);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dram requests=%ld (reads=%ld, writes=%ld)\n", core_id, (dram_reads_per_core + dram_writes_per_core), dram_reads_per_core, dram_writes_per_core); if (num_cores > 1) fprintf(stream, "PERF: core%d: dram requests=%ld (reads=%ld, writes=%ld)\n", core_id, (dram_reads_per_core + dram_writes_per_core), dram_reads_per_core, dram_writes_per_core);
if (num_cores > 1) fprintf(stream, "PERF: core%d: dram stalls=%d (utilization=%d%%)\n", core_id, dram_stalls_per_core, dram_utilization); if (num_cores > 1) fprintf(stream, "PERF: core%d: dram stalls=%ld (utilization=%d%%)\n", core_id, dram_stalls_per_core, dram_utilization);
if (num_cores > 1) fprintf(stream, "PERF: core%d: average dram latency=%d cycles\n", core_id, avg_dram_lat); if (num_cores > 1) fprintf(stream, "PERF: core%d: average dram latency=%d cycles\n", core_id, avg_dram_lat);
dram_reads += dram_reads_per_core; dram_reads += dram_reads_per_core;
dram_writes += dram_writes_per_core; dram_writes += dram_writes_per_core;

View file

@ -204,10 +204,7 @@ module VX_cluster #(
.dram_rsp_valid (dram_rsp_valid), .dram_rsp_valid (dram_rsp_valid),
.dram_rsp_tag (dram_rsp_tag), .dram_rsp_tag (dram_rsp_tag),
.dram_rsp_data (dram_rsp_data), .dram_rsp_data (dram_rsp_data),
.dram_rsp_ready (dram_rsp_ready), .dram_rsp_ready (dram_rsp_ready)
// Miss status
`UNUSED_PIN (miss_vec)
); );
end else begin end else begin

View file

@ -234,8 +234,8 @@
/////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////
`ifdef DBG_CACHE_REQ_INFO // pc, rd, wid `ifdef DBG_CACHE_REQ_INFO // wid PC
`define DBG_CACHE_REQ_MDATAW (32 + `NR_BITS + `NW_BITS) `define DBG_CACHE_REQ_MDATAW (`NW_BITS + 32)
`else `else
`define DBG_CACHE_REQ_MDATAW 0 `define DBG_CACHE_REQ_MDATAW 0
`endif `endif

View file

@ -51,7 +51,7 @@ module VX_icache_stage #(
assign ifetch_req_if.ready = icache_req_if.ready; assign ifetch_req_if.ready = icache_req_if.ready;
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
assign icache_req_if.tag = {ifetch_req_if.PC, `NR_BITS'(0), ifetch_req_if.wid, req_tag}; assign icache_req_if.tag = {ifetch_req_if.PC, ifetch_req_if.wid, req_tag};
`else `else
assign icache_req_if.tag = req_tag; assign icache_req_if.tag = req_tag;
`endif `endif

View file

@ -123,7 +123,7 @@ module VX_issue #(
`SCOPE_ASSIGN (writeback_pc, writeback_if.PC); `SCOPE_ASSIGN (writeback_pc, writeback_if.PC);
`SCOPE_ASSIGN (writeback_rd, writeback_if.rd); `SCOPE_ASSIGN (writeback_rd, writeback_if.rd);
`SCOPE_ASSIGN (writeback_data, writeback_if.data); `SCOPE_ASSIGN (writeback_data, writeback_if.data);
`SCOPE_ASSIGN (writeback_eop, writeback_if.eof); `SCOPE_ASSIGN (writeback_eop, writeback_if.eop);
`ifdef PERF_ENABLE `ifdef PERF_ENABLE
reg [63:0] perf_ibf_stalls; reg [63:0] perf_ibf_stalls;

View file

@ -172,7 +172,7 @@ module VX_lsu_unit #(
assign dcache_req_if.data = req_data; assign dcache_req_if.data = req_data;
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
assign dcache_req_if.tag = {`NUM_THREADS{{req_pc, req_rd, req_wid, req_tag}}}; assign dcache_req_if.tag = {`NUM_THREADS{{req_pc, req_wid, req_tag}}};
`else `else
assign dcache_req_if.tag = {`NUM_THREADS{req_tag}}; assign dcache_req_if.tag = {`NUM_THREADS{req_tag}};
`endif `endif

View file

@ -136,10 +136,7 @@ module VX_mem_unit # (
.dram_rsp_valid (icache_dram_rsp_if.valid), .dram_rsp_valid (icache_dram_rsp_if.valid),
.dram_rsp_data (icache_dram_rsp_if.data), .dram_rsp_data (icache_dram_rsp_if.data),
.dram_rsp_tag (icache_dram_rsp_if.tag), .dram_rsp_tag (icache_dram_rsp_if.tag),
.dram_rsp_ready (icache_dram_rsp_if.ready), .dram_rsp_ready (icache_dram_rsp_if.ready)
// Miss status
`UNUSED_PIN (miss_vec)
); );
VX_cache #( VX_cache #(
@ -197,10 +194,7 @@ module VX_mem_unit # (
.dram_rsp_valid (dcache_dram_rsp_if.valid), .dram_rsp_valid (dcache_dram_rsp_if.valid),
.dram_rsp_data (dcache_dram_rsp_if.data), .dram_rsp_data (dcache_dram_rsp_if.data),
.dram_rsp_tag (dcache_dram_rsp_if.tag), .dram_rsp_tag (dcache_dram_rsp_if.tag),
.dram_rsp_ready (dcache_dram_rsp_if.ready), .dram_rsp_ready (dcache_dram_rsp_if.ready)
// Miss status
`UNUSED_PIN (miss_vec)
); );
if (`SM_ENABLE) begin if (`SM_ENABLE) begin
@ -260,10 +254,7 @@ module VX_mem_unit # (
.dram_rsp_valid (0), .dram_rsp_valid (0),
.dram_rsp_data (0), .dram_rsp_data (0),
.dram_rsp_tag (0), .dram_rsp_tag (0),
`UNUSED_PIN (dram_rsp_ready), `UNUSED_PIN (dram_rsp_ready)
// Miss status
`UNUSED_PIN (miss_vec)
); );
end end

View file

@ -136,7 +136,7 @@ task print_ex_op;
default: $write("?"); default: $write("?");
endcase endcase
end end
default:; default: $write("?");
endcase endcase
end end
endtask endtask

View file

@ -206,10 +206,7 @@ module Vortex (
.dram_rsp_valid (dram_rsp_valid), .dram_rsp_valid (dram_rsp_valid),
.dram_rsp_data (dram_rsp_data), .dram_rsp_data (dram_rsp_data),
.dram_rsp_tag (dram_rsp_tag), .dram_rsp_tag (dram_rsp_tag),
.dram_rsp_ready (dram_rsp_ready), .dram_rsp_ready (dram_rsp_ready)
// Miss status
`UNUSED_PIN (miss_vec)
); );
end else begin end else begin

464
hw/rtl/cache/VX_bank.v vendored
View file

@ -47,6 +47,13 @@ module VX_bank #(
input wire clk, input wire clk,
input wire reset, input wire reset,
`ifdef PERF_ENABLE
output wire perf_read_misses,
output wire perf_write_misses,
output wire perf_mshr_stalls,
output wire perf_pipe_stalls,
`endif
// Core Request // Core Request
input wire core_req_valid, input wire core_req_valid,
input wire [`REQS_BITS-1:0] core_req_tid, input wire [`REQS_BITS-1:0] core_req_tid,
@ -76,39 +83,21 @@ module VX_bank #(
input wire dram_rsp_valid, input wire dram_rsp_valid,
input wire [`LINE_ADDR_WIDTH-1:0] dram_rsp_addr, input wire [`LINE_ADDR_WIDTH-1:0] dram_rsp_addr,
input wire [`BANK_LINE_WIDTH-1:0] dram_rsp_data, input wire [`BANK_LINE_WIDTH-1:0] dram_rsp_data,
output wire dram_rsp_ready, output wire dram_rsp_ready
`ifdef PERF_ENABLE
output wire perf_read_misses,
output wire perf_write_misses,
output wire perf_mshr_stalls,
output wire perf_pipe_stalls,
`endif
// Misses
output wire misses
); );
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
/* verilator lint_off UNUSED */ /* verilator lint_off UNUSED */
wire [31:0] debug_pc_st0; wire [31:0] debug_pc_st0;
wire [`NR_BITS-1:0] debug_rd_st0;
wire [`NW_BITS-1:0] debug_wid_st0; wire [`NW_BITS-1:0] debug_wid_st0;
wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st0;
wire [31:0] debug_pc_st1; wire [31:0] debug_pc_st1;
wire [`NR_BITS-1:0] debug_rd_st1;
wire [`NW_BITS-1:0] debug_wid_st1; wire [`NW_BITS-1:0] debug_wid_st1;
wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st1;
wire [31:0] debug_pc_st12;
wire [`NW_BITS-1:0] debug_wid_st12;
wire [31:0] debug_pc_st2; wire [31:0] debug_pc_st2;
wire [`NR_BITS-1:0] debug_rd_st2;
wire [`NW_BITS-1:0] debug_wid_st2; wire [`NW_BITS-1:0] debug_wid_st2;
wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st2;
wire [31:0] debug_pc_st3;
wire [`NR_BITS-1:0] debug_rd_st3;
wire [`NW_BITS-1:0] debug_wid_st3;
wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st3;
/* verilator lint_on UNUSED */ /* verilator lint_on UNUSED */
`endif `endif
@ -121,7 +110,6 @@ module VX_bank #(
wire drsq_push = dram_rsp_valid && dram_rsp_ready; wire drsq_push = dram_rsp_valid && dram_rsp_ready;
if (DRAM_ENABLE) begin if (DRAM_ENABLE) begin
wire drsq_full; wire drsq_full;
assign dram_rsp_ready = !drsq_full; assign dram_rsp_ready = !drsq_full;
@ -183,10 +171,11 @@ module VX_bank #(
`UNUSED_PIN (size) `UNUSED_PIN (size)
); );
wire mshr_pop;
reg [$clog2(MSHR_SIZE+1)-1:0] mshr_pending_size; reg [$clog2(MSHR_SIZE+1)-1:0] mshr_pending_size;
wire [$clog2(MSHR_SIZE+1)-1:0] mshr_pending_size_n; wire [$clog2(MSHR_SIZE+1)-1:0] mshr_pending_size_n;
reg mshr_going_full; reg mshr_going_full;
wire mshr_pop;
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;
@ -213,7 +202,10 @@ module VX_bank #(
wire is_mshr_st1; wire is_mshr_st1;
wire valid_st1; wire valid_st1;
wire [`LINE_ADDR_WIDTH-1:0] addr_st1; wire [`LINE_ADDR_WIDTH-1:0] addr_st1;
wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st1; wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st1;
wire [`WORD_WIDTH-1:0] readword_st1;
wire [`BANK_LINE_WIDTH-1:0] readdata_st1;
wire [BANK_LINE_SIZE-1:0] dirtyb_st1;
wire [`WORD_WIDTH-1:0] writeword_st1; wire [`WORD_WIDTH-1:0] writeword_st1;
wire [`BANK_LINE_WIDTH-1:0] writedata_st1; wire [`BANK_LINE_WIDTH-1:0] writedata_st1;
wire [`TAG_SELECT_BITS-1:0] readtag_st1; wire [`TAG_SELECT_BITS-1:0] readtag_st1;
@ -224,7 +216,21 @@ module VX_bank #(
wire [`REQ_TAG_WIDTH-1:0] tag_st1; wire [`REQ_TAG_WIDTH-1:0] tag_st1;
wire mem_rw_st1; wire mem_rw_st1;
wire [WORD_SIZE-1:0] byteen_st1; wire [WORD_SIZE-1:0] byteen_st1;
wire [`REQS_BITS-1:0] req_tid_st1; wire [`REQS_BITS-1:0] req_tid_st1;
wire core_req_hit_st1;
wire incoming_fill_st1;
wire do_writeback_st1;
wire mshr_push_st1;
wire crsq_push_st1;
wire dreq_push_st1;
wire valid_st12;
wire writeen_st12;
wire [`LINE_ADDR_WIDTH-1:0] addr_st12;
wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st12;
wire [WORD_SIZE-1:0] byteen_st12;
wire [`WORD_WIDTH-1:0] writeword_st12;
wire [`REQ_TAG_WIDTH-1:0] tag_st12;
wire valid_st2; wire valid_st2;
wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st2; wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st2;
@ -232,7 +238,6 @@ module VX_bank #(
wire [`WORD_WIDTH-1:0] writeword_st2; wire [`WORD_WIDTH-1:0] writeword_st2;
wire [`BANK_LINE_WIDTH-1:0] readdata_st2; wire [`BANK_LINE_WIDTH-1:0] readdata_st2;
wire [`BANK_LINE_WIDTH-1:0] writedata_st2; wire [`BANK_LINE_WIDTH-1:0] writedata_st2;
wire dirty_st2;
wire [BANK_LINE_SIZE-1:0] dirtyb_st2; wire [BANK_LINE_SIZE-1:0] dirtyb_st2;
wire [`TAG_SELECT_BITS-1:0] readtag_st2; wire [`TAG_SELECT_BITS-1:0] readtag_st2;
wire is_fill_st2; wire is_fill_st2;
@ -240,23 +245,17 @@ module VX_bank #(
wire miss_st2; wire miss_st2;
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 incoming_fill_st2;
wire [`REQ_TAG_WIDTH-1:0] tag_st2; wire [`REQ_TAG_WIDTH-1:0] tag_st2;
wire mem_rw_st2; wire mem_rw_st2;
wire [WORD_SIZE-1:0] byteen_st2; wire [WORD_SIZE-1:0] byteen_st2;
wire [`REQS_BITS-1:0] req_tid_st2; wire [`REQS_BITS-1:0] req_tid_st2;
wire core_req_hit_st2;
wire valid_st3; wire incoming_fill_st2;
wire is_mshr_st3; wire do_writeback_st2;
wire miss_st3; wire mshr_push_st2;
wire force_miss_st3; wire crsq_push_st2;
wire [`LINE_ADDR_WIDTH-1:0] addr_st3; wire dreq_push_st2;
wire [`REQ_TAG_WIDTH-1:0] tag_st3;
wire mem_rw_st3;
wire [WORD_SIZE-1:0] byteen_st3;
wire [`REQS_BITS-1:0] req_tid_st3;
wire mshr_push_stall; wire mshr_push_stall;
wire crsq_push_stall; wire crsq_push_stall;
@ -264,7 +263,6 @@ module VX_bank #(
wire pipeline_stall; wire pipeline_stall;
wire is_mshr_miss_st2 = valid_st2 && is_mshr_st2 && (miss_st2 || force_miss_st2); 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_st2 wire creq_commit = valid_st2
&& (core_req_hit_st2 || (WRITE_THROUGH && mem_rw_st2)) && (core_req_hit_st2 || (WRITE_THROUGH && mem_rw_st2))
@ -276,7 +274,7 @@ module VX_bank #(
wire creq_pop_unqual = !mshr_pop_unqual && !drsq_pop_unqual && !creq_empty && !mshr_going_full; wire creq_pop_unqual = !mshr_pop_unqual && !drsq_pop_unqual && !creq_empty && !mshr_going_full;
assign mshr_pop = mshr_pop_unqual && !pipeline_stall assign mshr_pop = mshr_pop_unqual && !pipeline_stall
&& !(is_mshr_miss_st2 || is_mshr_miss_st3); // stop if previous request was a miss && !is_mshr_miss_st2; // stop if previous request was a miss
assign drsq_pop = drsq_pop_unqual && !pipeline_stall; assign drsq_pop = drsq_pop_unqual && !pipeline_stall;
assign creq_pop = creq_pop_unqual && !pipeline_stall; assign creq_pop = creq_pop_unqual && !pipeline_stall;
@ -336,9 +334,9 @@ module VX_bank #(
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin
assign {debug_pc_st0, debug_rd_st0, debug_wid_st0, debug_tagid_st0} = tag_st0; assign {debug_pc_st0, debug_wid_st0} = tag_st0[CORE_TAG_WIDTH-1:CORE_TAG_ID_BITS];
end else begin end else begin
assign {debug_pc_st0, debug_rd_st0, debug_wid_st0, debug_tagid_st0} = 0; assign {debug_pc_st0, debug_wid_st0} = 0;
end end
`endif `endif
@ -346,14 +344,14 @@ if (DRAM_ENABLE) begin
wire mshr_pending_hazard_st1; wire mshr_pending_hazard_st1;
// we have a miss in msrq or in stage 3 for the current address // we have a miss in mshr or in stage 3 for the current address
wire mshr_pending_hazard_st0 = mshr_pending_hazard_unqual_st0 wire mshr_pending_hazard_st0 = mshr_pending_hazard_unqual_st0
|| (valid_st3 && (miss_st3 || force_miss_st3) && (addr_st3 == addr_st0)); || (valid_st2 && (miss_st2 || force_miss_st2) && (addr_st2 == addr_st0));
VX_pipe_register #( VX_pipe_register #(
.DATAW (1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + 1 + `BANK_LINE_WIDTH + 1 + WORD_SIZE + `REQS_BITS + `REQ_TAG_WIDTH), .DATAW (1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + 1 + `BANK_LINE_WIDTH + 1 + WORD_SIZE + `REQS_BITS + `REQ_TAG_WIDTH),
.RESETW (1) .RESETW (1)
) pipe_reg0 ( ) pipe_reg1 (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.enable (!pipeline_stall), .enable (!pipeline_stall),
@ -363,18 +361,17 @@ if (DRAM_ENABLE) begin
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin
assign {debug_pc_st1, debug_rd_st1, debug_wid_st1, debug_tagid_st1} = tag_st1; assign {debug_pc_st1, debug_wid_st1} = tag_st1[CORE_TAG_WIDTH-1:CORE_TAG_ID_BITS];
end else begin end else begin
assign {debug_pc_st1, debug_rd_st1, debug_wid_st1, debug_tagid_st1} = 0; assign {debug_pc_st1, debug_wid_st1} = 0;
end end
`endif `endif
// 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 mshr requests when previous requests got a miss
wire st2_pending_hazard_st1 = valid_st2 && (miss_st2 || force_miss_st2) && (addr_st2 == addr_st1); 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))
|| (valid_st1 && is_mshr_st1 && is_mshr_miss_st2); || (valid_st1 && is_mshr_st1 && is_mshr_miss_st2);
VX_tag_access #( VX_tag_access #(
@ -392,9 +389,7 @@ if (DRAM_ENABLE) begin
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
.debug_pc (debug_pc_st1), .debug_pc (debug_pc_st1),
.debug_rd (debug_rd_st1),
.debug_wid (debug_wid_st1), .debug_wid (debug_wid_st1),
.debug_tagid (debug_tagid_st1),
`endif `endif
.stall (pipeline_stall), .stall (pipeline_stall),
@ -413,76 +408,118 @@ if (DRAM_ENABLE) begin
.writeen_out (writeen_st1) .writeen_out (writeen_st1)
); );
assign misses = miss_st1; assign valid_st12 = valid_st2;
assign writeen_st12 = writeen_st2;
assign addr_st12 = addr_st2;
assign wsel_st12 = wsel_st2;
assign byteen_st12 = byteen_st2;
assign writeword_st12 = writeword_st2;
assign tag_st12 = tag_st2;
wire core_req_hit_st1 = !is_fill_st1 && !miss_st1 && !force_miss_st1; assign core_req_hit_st1 = !is_fill_st1 && !miss_st1 && !force_miss_st1;
wire incoming_fill_st1 = !drsq_empty && (addr_st1 == drsq_addr_st0); assign incoming_fill_st1 = !drsq_empty && (addr_st1 == drsq_addr_st0);
VX_pipe_register #( wire do_fill_req_st1 = miss_st1
.DATAW (1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + `TAG_SELECT_BITS + 1 + `BANK_LINE_WIDTH + 1 + WORD_SIZE + `REQS_BITS + `REQ_TAG_WIDTH), && !(WRITE_THROUGH && mem_rw_st1)
.RESETW (1) && (!force_miss_st1
) pipe_reg1 ( || (is_mshr_st1 && addr_st1 != addr_st2))
.clk (clk), && !incoming_fill_st1;
.reset (reset),
.enable (!pipeline_stall), assign do_writeback_st1 = (WRITE_THROUGH && mem_rw_st1)
.data_in ({valid_st1, incoming_fill_st1, core_req_hit_st1, is_mshr_st1, writeen_st1, force_miss_st1, dirty_st1, is_fill_st1, addr_st1, wsel_st1, writeword_st1, readtag_st1, miss_st1, writedata_st1, mem_rw_st1, byteen_st1, req_tid_st1, tag_st1}), || (!WRITE_THROUGH && dirty_st1 && is_fill_st1);
.data_out ({valid_st2, incoming_fill_st2, core_req_hit_st2, is_mshr_st2, writeen_st2, force_miss_st2, dirty_st2, is_fill_st2, addr_st2, wsel_st2, writeword_st2, readtag_st2, miss_st2, writedata_st2, mem_rw_st2, byteen_st2, req_tid_st2, tag_st2})
); assign dreq_push_st1 = do_fill_req_st1 || do_writeback_st1;
assign mshr_push_st1 = (miss_st1 || force_miss_st1)
&& !(WRITE_THROUGH && mem_rw_st1);
assign crsq_push_st1 = core_req_hit_st1 && !mem_rw_st1;
end else begin end else begin
`UNUSED_VAR (mshr_pending_hazard_unqual_st0) `UNUSED_VAR (mshr_pending_hazard_unqual_st0)
`UNUSED_VAR (drsq_push) `UNUSED_VAR (drsq_push)
`UNUSED_VAR (addr_st0) `UNUSED_VAR (dirty_st1)
`UNUSED_VAR (writeen_st2)
`ifdef DBG_CACHE_REQ_INFO
assign debug_pc_st1 = debug_pc_st0;
assign debug_wid_st1 = debug_wid_st0;
`endif
assign is_fill_st1 = is_fill_st0; assign is_fill_st1 = is_fill_st0;
assign is_mshr_st1 = is_mshr_st0; assign is_mshr_st1 = is_mshr_st0;
assign valid_st1 = valid_st0; assign valid_st1 = valid_st0;
assign wsel_st1 = wsel_st0; assign wsel_st1 = wsel_st0;
assign writeword_st1= writeword_st0; assign writeword_st1 = writeword_st0;
assign writedata_st1= writedata_st0; assign writedata_st1 = writedata_st0;
assign addr_st1 = creq_addr_st0[`LINE_SELECT_ADDR_RNG]; assign addr_st1 = creq_addr_st0[`LINE_SELECT_ADDR_RNG];
assign dirty_st1 = 0; assign tag_st1 = tag_st0;
assign readtag_st1 = 0; assign mem_rw_st1 = mem_rw_st0;
assign miss_st1 = 0; assign byteen_st1 = byteen_st0;
assign writeen_st1 = mem_rw_st1; assign req_tid_st1 = req_tid_st0;
assign force_miss_st1 = 0; assign dirty_st1 = 0;
assign tag_st1 = tag_st0; assign readtag_st1 = 0;
assign mem_rw_st1 = mem_rw_st0; assign miss_st1 = 0;
assign byteen_st1 = byteen_st0; assign writeen_st1 = mem_rw_st0;
assign req_tid_st1 = req_tid_st0; assign force_miss_st1 = 0;
assign is_fill_st2 = is_fill_st1;
assign is_mshr_st2 = is_mshr_st1;
assign valid_st2 = valid_st1;
assign wsel_st2 = wsel_st1;
assign writeword_st2= writeword_st1;
assign writedata_st2= writedata_st1;
assign addr_st2 = addr_st1;
assign dirty_st2 = dirty_st1;
assign readtag_st2 = readtag_st1;
assign miss_st2 = miss_st1;
assign writeen_st2 = writeen_st1;
assign force_miss_st2 = force_miss_st1;
assign tag_st2 = tag_st1;
assign mem_rw_st2 = mem_rw_st1;
assign byteen_st2 = byteen_st1;
assign req_tid_st2 = req_tid_st1;
assign core_req_hit_st2 = 1; assign valid_st12 = valid_st0;
assign incoming_fill_st2 = 0; assign writeen_st12 = mem_rw_st0;
assign addr_st12 = addr_st0;
assign wsel_st12 = wsel_st0;
assign byteen_st12 = byteen_st0;
assign writeword_st12 = writeword_st0;
assign tag_st12 = tag_st0;
assign incoming_fill_st1= 0;
assign core_req_hit_st1 = 1;
assign do_writeback_st1 = 0;
assign mshr_push_st1 = 0;
assign crsq_push_st1 = !mem_rw_st0;
assign dreq_push_st1 = 0;
end
assign misses = 0; VX_pipe_register #(
end .DATAW (1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + `WORD_WIDTH + `TAG_SELECT_BITS + 1 + `BANK_LINE_WIDTH + 1 + WORD_SIZE + `REQS_BITS + `REQ_TAG_WIDTH),
.RESETW (1)
) pipe_reg2 (
.clk (clk),
.reset (reset),
.enable (!pipeline_stall),
.data_in ({valid_st1, mshr_push_st1, crsq_push_st1, dreq_push_st1, do_writeback_st1, incoming_fill_st1, core_req_hit_st1, is_mshr_st1, writeen_st1, force_miss_st1, is_fill_st1, addr_st1, wsel_st1, readword_st1, writeword_st1, readtag_st1, miss_st1, writedata_st1, mem_rw_st1, byteen_st1, req_tid_st1, tag_st1}),
.data_out ({valid_st2, mshr_push_st2, crsq_push_st2, dreq_push_st2, do_writeback_st2, incoming_fill_st2, core_req_hit_st2, is_mshr_st2, writeen_st2, force_miss_st2, is_fill_st2, addr_st2, wsel_st2, readword_st2, writeword_st2, readtag_st2, miss_st2, writedata_st2, mem_rw_st2, byteen_st2, req_tid_st2, tag_st2})
);
if (WRITE_THROUGH) begin
assign dirtyb_st2 = dirtyb_st1;
assign readdata_st2 = readdata_st1;
end else begin
VX_pipe_register #(
.DATAW (BANK_LINE_SIZE + `BANK_LINE_WIDTH),
.RESETW (0)
) pipe_reg2b (
.clk (clk),
.reset (reset),
.enable (!pipeline_stall),
.data_in ({dirtyb_st1, readdata_st1}),
.data_out ({dirtyb_st2, readdata_st2})
);
end
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin
assign {debug_pc_st2, debug_rd_st2, debug_wid_st2, debug_tagid_st2} = tag_st2; assign {debug_pc_st12, debug_wid_st12} = tag_st12[CORE_TAG_WIDTH-1:CORE_TAG_ID_BITS];
end else begin end else begin
assign {debug_pc_st2, debug_rd_st2, debug_wid_st2, debug_tagid_st2} = 0; assign {debug_pc_st12, debug_wid_st12} = 0;
end end
`endif `endif
`UNUSED_VAR (tag_st12)
VX_data_access #( VX_data_access #(
.BANK_ID (BANK_ID), .BANK_ID (BANK_ID),
@ -491,6 +528,7 @@ end
.CACHE_SIZE (CACHE_SIZE), .CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE), .BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS), .NUM_BANKS (NUM_BANKS),
.DRAM_ENABLE (DRAM_ENABLE),
.WORD_SIZE (WORD_SIZE), .WORD_SIZE (WORD_SIZE),
.WRITE_ENABLE (WRITE_ENABLE), .WRITE_ENABLE (WRITE_ENABLE),
.WRITE_THROUGH (WRITE_THROUGH) .WRITE_THROUGH (WRITE_THROUGH)
@ -499,99 +537,59 @@ end
.reset (reset), .reset (reset),
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
.debug_pc (debug_pc_st2), .rdebug_pc (debug_pc_st1),
.debug_rd (debug_rd_st2), .rdebug_wid (debug_wid_st1),
.debug_wid (debug_wid_st2), .wdebug_pc (debug_pc_st12),
.debug_tagid (debug_tagid_st2), .wdebug_wid (debug_wid_st12),
`endif `endif
.stall (pipeline_stall), .stall (pipeline_stall),
// Inputs // reading
.valid_in (valid_st2), .readen_in (~writeen_st1 && valid_st1),
.addr_in (addr_st2), .raddr_in (addr_st1),
.writeen_in (writeen_st2), .rwsel_in (wsel_st1),
.is_fill_in (is_fill_st2), .rbyteen_in (byteen_st1),
.wordsel_in (wsel_st2), .readword_out (readword_st1),
.byteen_in (byteen_st2), .readdata_out (readdata_st1),
.writeword_in (writeword_st2), .dirtyb_out (dirtyb_st1),
.writedata_in (writedata_st2),
// Outputs // writing
.readword_out (readword_st2), .writeen_in (writeen_st12 && valid_st12),
.readdata_out (readdata_st2), .waddr_in (addr_st12),
.dirtyb_out (dirtyb_st2) .wfill_in (is_fill_st2),
); .wwsel_in (wsel_st12),
.wbyteen_in (byteen_st12),
wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st3; .writeword_in (writeword_st12),
wire [`WORD_WIDTH-1:0] writeword_st3; .writedata_in (writedata_st2)
wire [`WORD_WIDTH-1:0] readword_st3; );
wire [`BANK_LINE_WIDTH-1:0] readdata_st3;
wire [BANK_LINE_SIZE-1:0] dirtyb_st3;
wire [`TAG_SELECT_BITS-1:0] readtag_st3;
wire do_writeback_st3;
wire incoming_fill_st3;
wire mshr_push_st3;
wire crsq_push_st3;
wire dreq_push_st3;
wire incoming_fill_qual_st2 = (!drsq_empty && (addr_st2 == drsq_addr_st0)) || incoming_fill_st2;
wire do_fill_req_st2 = miss_st2
&& !(WRITE_THROUGH && mem_rw_st2)
&& (!force_miss_st2
|| (is_mshr_st2 && addr_st2 != addr_st3))
&& !incoming_fill_qual_st2;
wire do_writeback_st2 = (WRITE_THROUGH && mem_rw_st2)
|| (!WRITE_THROUGH && dirty_st2 && is_fill_st2);
wire dreq_push_st2 = do_fill_req_st2 || do_writeback_st2;
wire mshr_push_st2 = (miss_st2 || force_miss_st2)
&& !(WRITE_THROUGH && mem_rw_st2);
wire crsq_push_st2 = core_req_hit_st2 && !mem_rw_st2;
VX_pipe_register #(
.DATAW (1 + 1+ 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_WIDTH) + `WORD_WIDTH + `TAG_SELECT_BITS + BANK_LINE_SIZE + 1 + WORD_SIZE + `WORD_WIDTH + `BANK_LINE_WIDTH + `REQS_BITS + `REQ_TAG_WIDTH),
.RESETW (1)
) pipe_reg2 (
.clk (clk),
.reset (reset),
.enable (!pipeline_stall),
.data_in ({valid_st2, mshr_push_st2, crsq_push_st2, dreq_push_st2, do_writeback_st2, incoming_fill_qual_st2, force_miss_st2, is_mshr_st2, addr_st2, wsel_st2, writeword_st2, readtag_st2, miss_st2, dirtyb_st2, mem_rw_st2, byteen_st2, readword_st2, readdata_st2, req_tid_st2, tag_st2}),
.data_out ({valid_st3, mshr_push_st3, crsq_push_st3, dreq_push_st3, do_writeback_st3, incoming_fill_st3, force_miss_st3, is_mshr_st3, addr_st3, wsel_st3, writeword_st3, readtag_st3, miss_st3, dirtyb_st3, mem_rw_st3, byteen_st3, readword_st3, readdata_st3, req_tid_st3, tag_st3})
);
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin if (CORE_TAG_WIDTH != CORE_TAG_ID_BITS && CORE_TAG_ID_BITS != 0) begin
assign {debug_pc_st3, debug_rd_st3, debug_wid_st3, debug_tagid_st3} = tag_st3; assign {debug_pc_st2, debug_wid_st2} = tag_st2[CORE_TAG_WIDTH-1:CORE_TAG_ID_BITS];
end else begin end else begin
assign {debug_pc_st3, debug_rd_st3, debug_wid_st3, debug_tagid_st3} = 0; assign {debug_pc_st2, debug_wid_st2} = 0;
end end
`endif `endif
// Enqueue to miss reserv if it's a valid miss wire mshr_push_unqual = valid_st2 && mshr_push_st2;
wire mshr_push_unqual = valid_st3 && mshr_push_st3;
assign mshr_push_stall = 0; assign mshr_push_stall = 0;
wire mshr_push = mshr_push_unqual wire mshr_push = mshr_push_unqual
&& !crsq_push_stall && !crsq_push_stall
&& !dreq_push_stall; && !dreq_push_stall;
wire incoming_fill_qual_st3 = (!drsq_empty && (addr_st3 == drsq_addr_st0)) || incoming_fill_st3; wire incoming_fill_qual_st2 = (!drsq_empty && (addr_st2 == drsq_addr_st0)) || incoming_fill_st2;
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_st2 = valid_st2 && is_mshr_st2 && !mshr_push_unqual && !pipeline_stall;
// mark msrq entry that match DRAM fill as 'ready' // mark mshr entry that match DRAM fill as 'ready'
wire update_ready_st0 = drsq_pop; wire update_ready_st0 = drsq_pop;
// push missed requests as 'ready' if it was a forced miss but actually had a hit // push missed requests as 'ready' if it was a forced miss but actually had a hit
// or the fill request is comming for the missed block // or the fill request is comming for this block
wire mshr_init_ready_state_st3 = valid_st3 && (!miss_st3 || incoming_fill_qual_st3); wire mshr_init_ready_state_st2 = valid_st2 && (!miss_st2 || incoming_fill_qual_st2);
VX_miss_resrv #( VX_miss_resrv #(
.BANK_ID (BANK_ID), .BANK_ID (BANK_ID),
@ -608,44 +606,42 @@ end
.reset (reset), .reset (reset),
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
.debug_pc_st0 (debug_pc_st0), .deq_debug_pc (debug_pc_st0),
.debug_rd_st0 (debug_rd_st0), .deq_debug_wid (debug_wid_st0),
.debug_wid_st0 (debug_wid_st0), .enq_debug_pc (debug_pc_st2),
.debug_tagid_st0 (debug_tagid_st0), .enq_debug_wid (debug_wid_st2),
.debug_pc_st3 (debug_pc_st3),
.debug_rd_st3 (debug_rd_st3),
.debug_wid_st3 (debug_wid_st3),
.debug_tagid_st3 (debug_tagid_st3),
`endif `endif
// enqueue // enqueue
.enqueue_st3 (mshr_push), .enqueue (mshr_push),
.enqueue_addr_st3 (addr_st3), .enqueue_addr (addr_st2),
.enqueue_data_st3 ({writeword_st3, req_tid_st3, tag_st3, mem_rw_st3, byteen_st3, wsel_st3}), .enqueue_data ({writeword_st2, req_tid_st2, tag_st2, mem_rw_st2, byteen_st2, wsel_st2}),
.enqueue_is_mshr_st3(is_mshr_st3), .enqueue_is_mshr (is_mshr_st2),
.enqueue_ready_st3 (mshr_init_ready_state_st3), .enqueue_ready (mshr_init_ready_state_st2),
`UNUSED_PIN (enqueue_full), `UNUSED_PIN (enqueue_full),
// fill // lookup
.update_ready_st0 (update_ready_st0), .lookup_ready (update_ready_st0),
.addr_st0 (addr_st0), .lookup_addr (addr_st0),
.pending_hazard_st0 (mshr_pending_hazard_unqual_st0), .lookup_match (mshr_pending_hazard_unqual_st0),
// schedule
.schedule (mshr_pop),
.schedule_valid (mshr_valid_st0),
.schedule_addr (mshr_addr_st0),
.schedule_data ({mshr_writeword_st0, mshr_tid_st0, mshr_tag_st0, mshr_rw_st0, mshr_byteen_st0, mshr_wsel_st0}),
// dequeue // dequeue
.schedule_st0 (mshr_pop), .dequeue (mshr_dequeue_st2)
.dequeue_valid_st0 (mshr_valid_st0),
.dequeue_addr_st0 (mshr_addr_st0),
.dequeue_data_st0 ({mshr_writeword_st0, mshr_tid_st0, mshr_tag_st0, mshr_rw_st0, mshr_byteen_st0, mshr_wsel_st0}),
.dequeue_st3 (mshr_dequeue_st3)
); );
end else begin end else begin
`UNUSED_VAR (valid_st3) `UNUSED_VAR (valid_st2)
`UNUSED_VAR (mshr_push) `UNUSED_VAR (mshr_push)
`UNUSED_VAR (wsel_st3) `UNUSED_VAR (wsel_st2)
`UNUSED_VAR (writeword_st3) `UNUSED_VAR (writeword_st2)
`UNUSED_VAR (mem_rw_st3) `UNUSED_VAR (mem_rw_st2)
`UNUSED_VAR (byteen_st3) `UNUSED_VAR (byteen_st2)
`UNUSED_VAR (incoming_fill_st3) `UNUSED_VAR (incoming_fill_st2)
assign mshr_pending_hazard_unqual_st0 = 0; assign mshr_pending_hazard_unqual_st0 = 0;
assign mshr_valid_st0 = 0; assign mshr_valid_st0 = 0;
assign mshr_addr_st0 = 0; assign mshr_addr_st0 = 0;
@ -661,7 +657,7 @@ end
wire crsq_empty, crsq_full; wire crsq_empty, crsq_full;
wire crsq_push_unqual = valid_st3 && crsq_push_st3; wire crsq_push_unqual = valid_st2 && crsq_push_st2;
assign crsq_push_stall = crsq_push_unqual && crsq_full; assign crsq_push_stall = crsq_push_unqual && crsq_full;
wire crsq_push = crsq_push_unqual wire crsq_push = crsq_push_unqual
@ -671,9 +667,9 @@ end
wire crsq_pop = core_rsp_valid && core_rsp_ready; wire crsq_pop = core_rsp_valid && core_rsp_ready;
wire [`REQS_BITS-1:0] crsq_tid_st3 = req_tid_st3; wire [`REQS_BITS-1:0] crsq_tid_st2 = req_tid_st2;
wire [CORE_TAG_WIDTH-1:0] crsq_tag_st3 = CORE_TAG_WIDTH'(tag_st3); wire [CORE_TAG_WIDTH-1:0] crsq_tag_st2 = CORE_TAG_WIDTH'(tag_st2);
wire [`WORD_WIDTH-1:0] crsq_data_st3 = readword_st3; wire [`WORD_WIDTH-1:0] crsq_data_st2 = readword_st2;
VX_fifo_queue #( VX_fifo_queue #(
.DATAW (`REQS_BITS + CORE_TAG_WIDTH + `WORD_WIDTH), .DATAW (`REQS_BITS + CORE_TAG_WIDTH + `WORD_WIDTH),
@ -685,7 +681,7 @@ end
.reset (reset), .reset (reset),
.push (crsq_push), .push (crsq_push),
.pop (crsq_pop), .pop (crsq_pop),
.data_in ({crsq_tid_st3, crsq_tag_st3, crsq_data_st3}), .data_in ({crsq_tid_st2, crsq_tag_st2, crsq_data_st2}),
.data_out({core_rsp_tid, core_rsp_tag, core_rsp_data}), .data_out({core_rsp_tid, core_rsp_tag, core_rsp_data}),
.empty (crsq_empty), .empty (crsq_empty),
.full (crsq_full), .full (crsq_full),
@ -698,23 +694,23 @@ end
wire dreq_empty, dreq_full; wire dreq_empty, dreq_full;
wire dreq_push_unqual = valid_st3 && dreq_push_st3; wire dreq_push_unqual = valid_st2 && dreq_push_st2;
assign dreq_push_stall = dreq_push_unqual && dreq_full; assign dreq_push_stall = dreq_push_unqual && dreq_full;
wire dreq_push = dreq_push_unqual wire dreq_push = dreq_push_unqual
&& (do_writeback_st3 || !incoming_fill_qual_st3) && (do_writeback_st2 || !incoming_fill_qual_st2)
&& !dreq_full && !dreq_full
&& !mshr_push_stall && !mshr_push_stall
&& !crsq_push_stall; && !crsq_push_stall;
wire dreq_pop = dram_req_valid && dram_req_ready; wire dreq_pop = dram_req_valid && dram_req_ready;
wire writeback = WRITE_ENABLE && do_writeback_st3; wire writeback = WRITE_ENABLE && do_writeback_st2;
wire [`LINE_ADDR_WIDTH-1:0] dreq_addr = (WRITE_THROUGH || !writeback) ? addr_st3 : wire [`LINE_ADDR_WIDTH-1:0] dreq_addr = (WRITE_THROUGH || !writeback) ? addr_st2 :
{readtag_st3, addr_st3[`LINE_SELECT_BITS-1:0]}; {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
wire [BANK_LINE_SIZE-1:0] dreq_byteen = writeback ? dirtyb_st3 : {BANK_LINE_SIZE{1'b1}}; wire [BANK_LINE_SIZE-1:0] dreq_byteen = writeback ? dirtyb_st2 : {BANK_LINE_SIZE{1'b1}};
if (DRAM_ENABLE) begin if (DRAM_ENABLE) begin
VX_fifo_queue #( VX_fifo_queue #(
@ -727,7 +723,7 @@ end
.reset (reset), .reset (reset),
.push (dreq_push), .push (dreq_push),
.pop (dreq_pop), .pop (dreq_pop),
.data_in ({writeback, dreq_byteen, dreq_addr, readdata_st3}), .data_in ({writeback, dreq_byteen, dreq_addr, readdata_st2}),
.data_out({dram_req_rw, dram_req_byteen, dram_req_addr, dram_req_data}), .data_out({dram_req_rw, dram_req_byteen, dram_req_addr, dram_req_data}),
.empty (dreq_empty), .empty (dreq_empty),
.full (dreq_full), .full (dreq_full),
@ -738,9 +734,9 @@ end
`UNUSED_VAR (dreq_pop) `UNUSED_VAR (dreq_pop)
`UNUSED_VAR (dreq_addr) `UNUSED_VAR (dreq_addr)
`UNUSED_VAR (dreq_byteen) `UNUSED_VAR (dreq_byteen)
`UNUSED_VAR (readtag_st3) `UNUSED_VAR (readtag_st2)
`UNUSED_VAR (dirtyb_st3) `UNUSED_VAR (dirtyb_st2)
`UNUSED_VAR (readdata_st3) `UNUSED_VAR (readdata_st2)
`UNUSED_VAR (writeback) `UNUSED_VAR (writeback)
`UNUSED_VAR (dram_req_ready) `UNUSED_VAR (dram_req_ready)
assign dreq_empty = 1; assign dreq_empty = 1;
@ -761,7 +757,6 @@ end
`SCOPE_ASSIGN (valid_st0, valid_st0); `SCOPE_ASSIGN (valid_st0, valid_st0);
`SCOPE_ASSIGN (valid_st1, valid_st1); `SCOPE_ASSIGN (valid_st1, valid_st1);
`SCOPE_ASSIGN (valid_st2, valid_st2); `SCOPE_ASSIGN (valid_st2, valid_st2);
`SCOPE_ASSIGN (valid_st3, valid_st3);
`SCOPE_ASSIGN (is_fill_st0, is_fill_st0); `SCOPE_ASSIGN (is_fill_st0, is_fill_st0);
`SCOPE_ASSIGN (is_mshr_st0, is_mshr_st0); `SCOPE_ASSIGN (is_mshr_st0, is_mshr_st0);
`SCOPE_ASSIGN (miss_st1, miss_st1); `SCOPE_ASSIGN (miss_st1, miss_st1);
@ -772,7 +767,6 @@ end
`SCOPE_ASSIGN (addr_st0, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID)); `SCOPE_ASSIGN (addr_st0, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID));
`SCOPE_ASSIGN (addr_st1, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID)); `SCOPE_ASSIGN (addr_st1, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID));
`SCOPE_ASSIGN (addr_st2, `LINE_TO_BYTE_ADDR(addr_st2, BANK_ID)); `SCOPE_ASSIGN (addr_st2, `LINE_TO_BYTE_ADDR(addr_st2, BANK_ID));
`SCOPE_ASSIGN (addr_st3, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID));
`ifdef PERF_ENABLE `ifdef PERF_ENABLE
assign perf_read_misses = !pipeline_stall && miss_st2 && !is_mshr_st2 && !mem_rw_st2; assign perf_read_misses = !pipeline_stall && miss_st2 && !is_mshr_st2 && !mem_rw_st2;
@ -782,14 +776,14 @@ end
`endif `endif
`ifdef DBG_PRINT_CACHE_BANK `ifdef DBG_PRINT_CACHE_BANK
wire incoming_fill_dfp_st3 = drsq_push && (addr_st3 == dram_rsp_addr); wire incoming_fill_dfp_st2 = drsq_push && (addr_st2 == dram_rsp_addr);
always @(posedge clk) begin always @(posedge clk) begin
if (valid_st3 && miss_st3 && (incoming_fill_st3 || incoming_fill_dfp_st3)) begin if (valid_st2 && miss_st2 && (incoming_fill_st2 || incoming_fill_dfp_st2)) begin
$display("%t: incoming fill - addr=%0h, st3=%b, dfp=%b", $time, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), incoming_fill_st3, incoming_fill_dfp_st3); $display("%t: incoming fill - addr=%0h, st3=%b, dfp=%b", $time, `LINE_TO_BYTE_ADDR(addr_st2, BANK_ID), incoming_fill_st2, incoming_fill_dfp_st2);
assert(!is_mshr_st3); assert(!is_mshr_st2);
end end
if (pipeline_stall) begin if (pipeline_stall) begin
$display("%t: cache%0d:%0d pipeline-stall: msrq=%b, cwbq=%b, dwbq=%b", $time, CACHE_ID, BANK_ID, mshr_push_stall, crsq_push_stall, dreq_push_stall); $display("%t: cache%0d:%0d pipeline-stall: mshr=%b, cwbq=%b, dwbq=%b", $time, CACHE_ID, BANK_ID, mshr_push_stall, crsq_push_stall, dreq_push_stall);
end end
if (drsq_pop) begin if (drsq_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), drsq_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), drsq_filldata_st0);
@ -801,13 +795,13 @@ end
$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, 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, debug_wid_st0, debug_pc_st0);
end end
if (crsq_push) begin if (crsq_push) begin
$display("%t: cache%0d:%0d core-rsp: addr=%0h, tag=%0h, tid=%0d, data=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), crsq_tag_st3, crsq_tid_st3, crsq_data_st3, debug_wid_st3, debug_pc_st3); $display("%t: cache%0d:%0d core-rsp: addr=%0h, tag=%0h, tid=%0d, data=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st2, BANK_ID), crsq_tag_st2, crsq_tid_st2, crsq_data_st2, debug_wid_st2, debug_pc_st2);
end end
if (dreq_push) begin if (dreq_push) begin
if (do_writeback_st3) if (do_writeback_st2)
$display("%t: cache%0d:%0d writeback: addr=%0h, data=%0h, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dreq_addr, BANK_ID), readdata_st3, dirtyb_st3, debug_wid_st3, debug_pc_st3); $display("%t: cache%0d:%0d writeback: addr=%0h, data=%0h, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dreq_addr, BANK_ID), readdata_st2, dreq_byteen, debug_wid_st2, debug_pc_st2);
else else
$display("%t: cache%0d:%0d fill-req: addr=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dreq_addr, BANK_ID), debug_wid_st3, debug_pc_st3); $display("%t: cache%0d:%0d fill-req: addr=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dreq_addr, BANK_ID), debug_wid_st2, debug_pc_st2);
end end
end end
`endif `endif

View file

@ -82,9 +82,7 @@ module VX_cache #(
input wire dram_rsp_valid, input wire dram_rsp_valid,
input wire [`BANK_LINE_WIDTH-1:0] dram_rsp_data, input wire [`BANK_LINE_WIDTH-1:0] dram_rsp_data,
input wire [DRAM_TAG_WIDTH-1:0] dram_rsp_tag, input wire [DRAM_TAG_WIDTH-1:0] dram_rsp_tag,
output wire dram_rsp_ready, output wire dram_rsp_ready
output wire [NUM_BANKS-1:0] miss_vec
); );
`STATIC_ASSERT(NUM_BANKS <= NUM_REQS, ("invalid value")) `STATIC_ASSERT(NUM_BANKS <= NUM_REQS, ("invalid value"))
@ -113,9 +111,6 @@ module VX_cache #(
wire [NUM_BANKS-1:0] per_bank_dram_rsp_ready; wire [NUM_BANKS-1:0] per_bank_dram_rsp_ready;
wire [NUM_BANKS-1:0] per_bank_miss;
assign miss_vec = per_bank_miss;
`ifdef PERF_ENABLE `ifdef PERF_ENABLE
wire [NUM_BANKS-1:0] perf_read_miss_per_bank; wire [NUM_BANKS-1:0] perf_read_miss_per_bank;
wire [NUM_BANKS-1:0] perf_write_miss_per_bank; wire [NUM_BANKS-1:0] perf_write_miss_per_bank;
@ -189,8 +184,6 @@ module VX_cache #(
wire [`LINE_ADDR_WIDTH-1:0] curr_bank_dram_rsp_addr; wire [`LINE_ADDR_WIDTH-1:0] curr_bank_dram_rsp_addr;
wire curr_bank_dram_rsp_ready; wire curr_bank_dram_rsp_ready;
wire curr_bank_miss;
// Core Req // Core Req
assign curr_bank_core_req_valid = per_bank_core_req_valid[i]; assign curr_bank_core_req_valid = per_bank_core_req_valid[i];
assign curr_bank_core_req_tid = per_bank_core_req_tid[i]; assign curr_bank_core_req_tid = per_bank_core_req_tid[i];
@ -230,9 +223,6 @@ module VX_cache #(
end end
assign curr_bank_dram_rsp_data = dram_rsp_data; assign curr_bank_dram_rsp_data = dram_rsp_data;
assign per_bank_dram_rsp_ready[i] = curr_bank_dram_rsp_ready; assign per_bank_dram_rsp_ready[i] = curr_bank_dram_rsp_ready;
//Misses
assign per_bank_miss[i] = curr_bank_miss;
VX_bank #( VX_bank #(
.BANK_ID (i), .BANK_ID (i),
@ -257,6 +247,13 @@ module VX_cache #(
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
`ifdef PERF_ENABLE
.perf_read_misses (perf_read_miss_per_bank[i]),
.perf_write_misses (perf_write_miss_per_bank[i]),
.perf_mshr_stalls (perf_mshr_stall_per_bank[i]),
.perf_pipe_stalls (perf_pipe_stall_per_bank[i]),
`endif
// Core request // Core request
.core_req_valid (curr_bank_core_req_valid), .core_req_valid (curr_bank_core_req_valid),
@ -287,17 +284,7 @@ module VX_cache #(
.dram_rsp_valid (curr_bank_dram_rsp_valid), .dram_rsp_valid (curr_bank_dram_rsp_valid),
.dram_rsp_data (curr_bank_dram_rsp_data), .dram_rsp_data (curr_bank_dram_rsp_data),
.dram_rsp_addr (curr_bank_dram_rsp_addr), .dram_rsp_addr (curr_bank_dram_rsp_addr),
.dram_rsp_ready (curr_bank_dram_rsp_ready), .dram_rsp_ready (curr_bank_dram_rsp_ready)
`ifdef PERF_ENABLE
.perf_read_misses (perf_read_miss_per_bank[i]),
.perf_write_misses (perf_write_miss_per_bank[i]),
.perf_mshr_stalls (perf_mshr_stall_per_bank[i]),
.perf_pipe_stalls (perf_pipe_stall_per_bank[i]),
`endif
//Misses
.misses (curr_bank_miss)
); );
end end

View file

@ -12,6 +12,9 @@ module VX_data_access #(
parameter NUM_BANKS = 1, parameter NUM_BANKS = 1,
// Size of a word in bytes // Size of a word in bytes
parameter WORD_SIZE = 1, parameter WORD_SIZE = 1,
// Enable dram update
parameter DRAM_ENABLE = 1,
// Enable cache writeable // Enable cache writeable
parameter WRITE_ENABLE = 1, parameter WRITE_ENABLE = 1,
@ -27,41 +30,49 @@ module VX_data_access #(
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
`IGNORE_WARNINGS_BEGIN `IGNORE_WARNINGS_BEGIN
input wire[31:0] debug_pc, input wire[31:0] rdebug_pc,
input wire[`NR_BITS-1:0] debug_rd, input wire[`NW_BITS-1:0] rdebug_wid,
input wire[`NW_BITS-1:0] debug_wid, input wire[31:0] wdebug_pc,
input wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid, input wire[`NW_BITS-1:0] wdebug_wid,
`IGNORE_WARNINGS_END `IGNORE_WARNINGS_END
`endif `endif
input wire stall, input wire stall,
// Inputs // reading
input wire valid_in, input wire readen_in,
`IGNORE_WARNINGS_BEGIN `IGNORE_WARNINGS_BEGIN
input wire[`LINE_ADDR_WIDTH-1:0] addr_in, input wire[`LINE_ADDR_WIDTH-1:0] raddr_in,
`IGNORE_WARNINGS_END `IGNORE_WARNINGS_END
input wire writeen_in, input wire [`UP(`WORD_SELECT_WIDTH)-1:0] rwsel_in,
input wire is_fill_in, input wire [WORD_SIZE-1:0] rbyteen_in,
input wire [`WORD_WIDTH-1:0] writeword_in,
input wire [`BANK_LINE_WIDTH-1:0] writedata_in,
input wire [WORD_SIZE-1:0] byteen_in,
input wire [`UP(`WORD_SELECT_WIDTH)-1:0] wordsel_in,
// Outputs
output wire[`WORD_WIDTH-1:0] readword_out, output wire[`WORD_WIDTH-1:0] readword_out,
output wire [`BANK_LINE_WIDTH-1:0] readdata_out, output wire [`BANK_LINE_WIDTH-1:0] readdata_out,
output wire [BANK_LINE_SIZE-1:0] dirtyb_out output wire [BANK_LINE_SIZE-1:0] dirtyb_out,
// writing
input wire writeen_in,
`IGNORE_WARNINGS_BEGIN
input wire[`LINE_ADDR_WIDTH-1:0] waddr_in,
`IGNORE_WARNINGS_END
input wire [`UP(`WORD_SELECT_WIDTH)-1:0] wwsel_in,
input wire [WORD_SIZE-1:0] wbyteen_in,
input wire wfill_in,
input wire [`WORD_WIDTH-1:0] writeword_in,
input wire [`BANK_LINE_WIDTH-1:0] writedata_in
); );
wire [BANK_LINE_SIZE-1:0] read_dirtyb_out; wire [BANK_LINE_SIZE-1:0] read_dirtyb, dirtyb_qual;
wire [`BANK_LINE_WIDTH-1:0] read_data; wire [`BANK_LINE_WIDTH-1:0] read_data, readdata_qual;
wire [`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0] byte_enable; wire [BANK_LINE_SIZE-1:0] byte_enable;
wire [`BANK_LINE_WIDTH-1:0] write_data; wire [`BANK_LINE_WIDTH-1:0] write_data;
wire write_enable; wire write_enable;
wire [`LINE_SELECT_BITS-1:0] addrline = addr_in[`LINE_SELECT_BITS-1:0]; wire [`LINE_SELECT_BITS-1:0] raddr = raddr_in[`LINE_SELECT_BITS-1:0];
wire [`LINE_SELECT_BITS-1:0] waddr = waddr_in[`LINE_SELECT_BITS-1:0];
`UNUSED_VAR (readen_in)
VX_data_store #( VX_data_store #(
.CACHE_SIZE (CACHE_SIZE), .CACHE_SIZE (CACHE_SIZE),
@ -73,66 +84,75 @@ module VX_data_access #(
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.read_addr (addrline), .read_addr (raddr),
.read_dirtyb (read_dirtyb_out),
.read_data (read_data), .read_data (read_data),
.read_dirtyb (read_dirtyb),
.write_enable(write_enable), .write_enable(write_enable),
.write_fill (is_fill_in), .write_fill (wfill_in),
.byte_enable (byte_enable), .write_addr (waddr),
.write_addr (addrline), .byte_enable (byte_enable),
.write_data (write_data) .write_data (write_data)
); );
if (`WORD_SELECT_WIDTH != 0) begin wire [`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0] wbyteen_qual;
wire [`WORD_WIDTH-1:0] readword = read_data[wordsel_in * `WORD_WIDTH +: `WORD_WIDTH]; wire [`BANK_LINE_WIDTH-1:0] writeword_qual;
for (genvar i = 0; i < WORD_SIZE; i++) begin
assign readword_out[i * 8 +: 8] = readword[i * 8 +: 8] & {8{byteen_in[i]}};
end
end else begin
for (genvar i = 0; i < WORD_SIZE; i++) begin
assign readword_out[i * 8 +: 8] = read_data[i * 8 +: 8] & {8{byteen_in[i]}};
end
end
wire [`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0] byte_enable_w;
wire [`BANK_LINE_WIDTH-1:0] write_data_w;
if (`WORD_SELECT_WIDTH != 0) begin if (`WORD_SELECT_WIDTH != 0) begin
for (genvar i = 0; i < `BANK_LINE_WORDS; i++) begin for (genvar i = 0; i < `BANK_LINE_WORDS; i++) begin
assign byte_enable_w[i] = (wordsel_in == `WORD_SELECT_WIDTH'(i)) ? byteen_in : {WORD_SIZE{1'b0}}; assign wbyteen_qual[i] = (wwsel_in == `WORD_SELECT_WIDTH'(i)) ? wbyteen_in : {WORD_SIZE{1'b0}};
assign write_data_w[i * `WORD_WIDTH +: `WORD_WIDTH] = writeword_in; assign writeword_qual[i * `WORD_WIDTH +: `WORD_WIDTH] = writeword_in;
end end
end else begin end else begin
assign byte_enable_w = byteen_in; `UNUSED_VAR (wwsel_in)
assign write_data_w = writeword_in; assign wbyteen_qual = wbyteen_in;
assign writeword_qual = writeword_in;
end end
assign byte_enable = is_fill_in ? {BANK_LINE_SIZE{1'b1}} : byte_enable_w; assign byte_enable = wfill_in ? {BANK_LINE_SIZE{1'b1}} : wbyteen_qual;
assign write_data = is_fill_in ? writedata_in : write_data_w; assign write_data = wfill_in ? writedata_in : writeword_qual;
assign write_enable = valid_in && writeen_in && !stall; assign write_enable = writeen_in && !stall;
wire rw_hazard = DRAM_ENABLE && (raddr == waddr) && writeen_in;
for (genvar i = 0; i < BANK_LINE_SIZE; i++) begin
assign dirtyb_qual[i] = rw_hazard ? byte_enable[i] : read_dirtyb[i];
assign readdata_qual[i * 8 +: 8] = (rw_hazard && byte_enable[i]) ? write_data[i * 8 +: 8] : read_data[i * 8 +: 8];
end
if (WRITE_THROUGH) begin if (WRITE_THROUGH) begin
`UNUSED_VAR (read_dirtyb_out) `UNUSED_VAR (dirtyb_qual)
assign dirtyb_out = byte_enable_w; assign dirtyb_out = wbyteen_qual;
assign readdata_out = write_data_w; assign readdata_out = writeword_qual;
end else begin end else begin
assign dirtyb_out = read_dirtyb_out; assign dirtyb_out = dirtyb_qual;
assign readdata_out = read_data; assign readdata_out = readdata_qual;
end
if (`WORD_SELECT_WIDTH != 0) begin
wire [`WORD_WIDTH-1:0] readword = readdata_qual[rwsel_in * `WORD_WIDTH +: `WORD_WIDTH];
for (genvar i = 0; i < WORD_SIZE; i++) begin
assign readword_out[i * 8 +: 8] = readword[i * 8 +: 8] & {8{rbyteen_in[i]}};
end
end else begin
`UNUSED_VAR (rwsel_in)
for (genvar i = 0; i < WORD_SIZE; i++) begin
assign readword_out[i * 8 +: 8] = readdata_qual[i * 8 +: 8] & {8{rbyteen_in[i]}};
end
end end
`ifdef DBG_PRINT_CACHE_DATA `ifdef DBG_PRINT_CACHE_DATA
always @(posedge clk) begin always @(posedge clk) begin
if (valid_in && !stall) begin if (!stall) begin
if (write_enable) begin if (writeen_in) begin
if (is_fill_in) begin if (wfill_in) begin
$display("%t: cache%0d:%0d data-fill: addr=%0h, dirty=%b, blk_addr=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_in, BANK_ID), dirtyb_out, addrline, write_data); $display("%t: cache%0d:%0d data-fill: addr=%0h, dirty=%b, blk_addr=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(waddr_in, BANK_ID), dirtyb_out, waddr, write_data);
end else begin end else begin
$display("%t: cache%0d:%0d data-write: addr=%0h, wid=%0d, PC=%0h, byteen=%b, dirty=%b, blk_addr=%0d, wsel=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_in, BANK_ID), debug_wid, debug_pc, byte_enable, dirtyb_out, addrline, wordsel_in, writeword_in); $display("%t: cache%0d:%0d data-write: addr=%0h, wid=%0d, PC=%0h, byteen=%b, dirty=%b, blk_addr=%0d, wsel=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(waddr_in, BANK_ID), rdebug_wid, rdebug_pc, byte_enable, dirtyb_out, waddr, wwsel_in, writeword_in);
end end
end else begin end
$display("%t: cache%0d:%0d data-read: addr=%0h, wid=%0d, PC=%0h, dirty=%b, blk_addr=%0d, wsel=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_in, BANK_ID), debug_wid, debug_pc, dirtyb_out, addrline, wordsel_in, read_data); if (readen_in) begin
$display("%t: cache%0d:%0d data-read: addr=%0h, wid=%0d, PC=%0h, dirty=%b, blk_addr=%0d, wsel=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(raddr_in, BANK_ID), rdebug_wid, rdebug_pc, dirtyb_out, raddr, rwsel_in, read_data);
end end
end end
end end

View file

@ -18,7 +18,7 @@ module VX_data_store #(
input wire write_enable, input wire write_enable,
input wire write_fill, input wire write_fill,
input wire[`BANK_LINE_WORDS-1:0][WORD_SIZE-1:0] byte_enable, input wire[BANK_LINE_SIZE-1:0] byte_enable,
input wire[`LINE_SELECT_BITS-1:0] write_addr, input wire[`LINE_SELECT_BITS-1:0] write_addr,
input wire[`BANK_LINE_WIDTH-1:0] write_data, input wire[`BANK_LINE_WIDTH-1:0] write_data,

View file

@ -24,36 +24,34 @@ module VX_miss_resrv #(
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
`IGNORE_WARNINGS_BEGIN `IGNORE_WARNINGS_BEGIN
input wire[31:0] debug_pc_st0, input wire[31:0] deq_debug_pc,
input wire[`NR_BITS-1:0] debug_rd_st0, input wire[`NW_BITS-1:0] deq_debug_wid,
input wire[`NW_BITS-1:0] debug_wid_st0, input wire[31:0] enq_debug_pc,
input wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st0, input wire[`NW_BITS-1:0] enq_debug_wid,
input wire[31:0] debug_pc_st3,
input wire[`NR_BITS-1:0] debug_rd_st3,
input wire[`NW_BITS-1:0] debug_wid_st3,
input wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid_st3,
`IGNORE_WARNINGS_END `IGNORE_WARNINGS_END
`endif `endif
// enqueue // enqueue
input wire enqueue_st3, input wire enqueue,
input wire [`LINE_ADDR_WIDTH-1:0] enqueue_addr_st3, input wire [`LINE_ADDR_WIDTH-1:0] enqueue_addr,
input wire [`MSHR_DATA_WIDTH-1:0] enqueue_data_st3, input wire [`MSHR_DATA_WIDTH-1:0] enqueue_data,
input wire enqueue_is_mshr_st3, input wire enqueue_is_mshr,
input wire enqueue_ready_st3, input wire enqueue_ready,
output wire enqueue_full, output wire enqueue_full,
// fill // lookup
input wire update_ready_st0, input wire lookup_ready,
input wire [`LINE_ADDR_WIDTH-1:0] addr_st0, input wire [`LINE_ADDR_WIDTH-1:0] lookup_addr,
output wire pending_hazard_st0, output wire lookup_match,
// schedule
input wire schedule,
output wire schedule_valid,
output wire [`LINE_ADDR_WIDTH-1:0] schedule_addr,
output wire [`MSHR_DATA_WIDTH-1:0] schedule_data,
// dequeue // dequeue
input wire schedule_st0, input wire dequeue
output wire dequeue_valid_st0,
output wire [`LINE_ADDR_WIDTH-1:0] dequeue_addr_st0,
output wire [`MSHR_DATA_WIDTH-1:0] dequeue_data_st0,
input wire dequeue_st3
); );
`USE_FAST_BRAM reg [`LINE_ADDR_WIDTH-1:0] addr_table [MSHR_SIZE-1:0]; `USE_FAST_BRAM reg [`LINE_ADDR_WIDTH-1:0] addr_table [MSHR_SIZE-1:0];
@ -67,17 +65,17 @@ module VX_miss_resrv #(
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
assign valid_address_match[i] = valid_table[i] && (addr_table[i] == addr_st0); assign valid_address_match[i] = valid_table[i] && (addr_table[i] == lookup_addr);
end end
assign pending_hazard_st0 = (| valid_address_match); assign lookup_match = (| valid_address_match);
wire dequeue_ready = ready_table[schedule_ptr]; wire dequeue_ready = ready_table[schedule_ptr];
assign dequeue_valid_st0 = dequeue_ready; assign schedule_valid = dequeue_ready;
assign dequeue_addr_st0 = addr_table[schedule_ptr]; assign schedule_addr = addr_table[schedule_ptr];
wire mshr_push = enqueue_st3 && !enqueue_is_mshr_st3; wire mshr_push = enqueue && !enqueue_is_mshr;
wire [`LOG2UP(MSHR_SIZE)-1:0] head_ptr_n = head_ptr + $bits(head_ptr)'(1); wire [`LOG2UP(MSHR_SIZE)-1:0] head_ptr_n = head_ptr + $bits(head_ptr)'(1);
@ -91,33 +89,33 @@ module VX_miss_resrv #(
tail_ptr <= 0; tail_ptr <= 0;
size <= 0; size <= 0;
end else begin end else begin
if (update_ready_st0) begin if (lookup_ready) begin
ready_table <= ready_table | valid_address_match; ready_table <= ready_table | valid_address_match;
end end
if (enqueue_st3) begin if (enqueue) begin
assert(!enqueue_full); assert(!enqueue_full);
if (enqueue_is_mshr_st3) begin if (enqueue_is_mshr) begin
// returning missed msrq entry, restore schedule // returning missed msrq entry, restore schedule
valid_table[restore_ptr] <= 1; valid_table[restore_ptr] <= 1;
ready_table[restore_ptr] <= enqueue_ready_st3; ready_table[restore_ptr] <= enqueue_ready;
restore_ptr <= restore_ptr + $bits(restore_ptr)'(1); restore_ptr <= restore_ptr + $bits(restore_ptr)'(1);
schedule_ptr <= head_ptr; schedule_ptr <= head_ptr;
end else begin end else begin
valid_table[tail_ptr] <= 1; valid_table[tail_ptr] <= 1;
ready_table[tail_ptr] <= enqueue_ready_st3; ready_table[tail_ptr] <= enqueue_ready;
tail_ptr <= tail_ptr + $bits(tail_ptr)'(1); tail_ptr <= tail_ptr + $bits(tail_ptr)'(1);
size <= size + $bits(size)'(1); size <= size + $bits(size)'(1);
end end
end else if (dequeue_st3) begin end else if (dequeue) begin
head_ptr <= head_ptr_n; head_ptr <= head_ptr_n;
restore_ptr <= head_ptr_n; restore_ptr <= head_ptr_n;
valid_table[head_ptr] <= 0; valid_table[head_ptr] <= 0;
size <= size - $bits(size)'(1); size <= size - $bits(size)'(1);
end end
if (schedule_st0) begin if (schedule) begin
assert(dequeue_valid_st0); assert(schedule_valid);
valid_table[schedule_ptr] <= 0; valid_table[schedule_ptr] <= 0;
ready_table[schedule_ptr] <= 0; ready_table[schedule_ptr] <= 0;
schedule_ptr <= schedule_ptr + $bits(schedule_ptr)'(1); schedule_ptr <= schedule_ptr + $bits(schedule_ptr)'(1);
@ -126,8 +124,8 @@ module VX_miss_resrv #(
end end
always @(posedge clk) begin always @(posedge clk) begin
if (enqueue_st3 && !enqueue_is_mshr_st3) begin if (enqueue && !enqueue_is_mshr) begin
addr_table[tail_ptr] <= enqueue_addr_st3; addr_table[tail_ptr] <= enqueue_addr;
end end
end end
@ -142,23 +140,23 @@ module VX_miss_resrv #(
.wren(mshr_push), .wren(mshr_push),
.byteen(1'b1), .byteen(1'b1),
.rden(1'b1), .rden(1'b1),
.din(enqueue_data_st3), .din(enqueue_data),
.dout(dequeue_data_st0) .dout(schedule_data)
); );
`ifdef DBG_PRINT_CACHE_MSHR `ifdef DBG_PRINT_CACHE_MSHR
always @(posedge clk) begin always @(posedge clk) begin
if (update_ready_st0 || schedule_st0 || enqueue_st3 || dequeue_st3) begin if (lookup_ready || schedule || enqueue || dequeue) begin
if (schedule_st0) if (schedule)
$display("%t: cache%0d:%0d msrq-schedule: addr%0d=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, schedule_ptr, `LINE_TO_BYTE_ADDR(dequeue_addr_st0, BANK_ID), debug_wid_st0, debug_pc_st0); $display("%t: cache%0d:%0d msrq-schedule: addr%0d=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, schedule_ptr, `LINE_TO_BYTE_ADDR(schedule_addr, BANK_ID), deq_debug_wid, deq_debug_pc);
if (enqueue_st3) begin if (enqueue) begin
if (enqueue_is_mshr_st3) if (enqueue_is_mshr)
$display("%t: cache%0d:%0d msrq-restore: addr%0d=%0h, ready=%b", $time, CACHE_ID, BANK_ID, restore_ptr, `LINE_TO_BYTE_ADDR(enqueue_addr_st3, BANK_ID), enqueue_ready_st3); $display("%t: cache%0d:%0d msrq-restore: addr%0d=%0h, ready=%b", $time, CACHE_ID, BANK_ID, restore_ptr, `LINE_TO_BYTE_ADDR(enqueue_addr, BANK_ID), enqueue_ready);
else else
$display("%t: cache%0d:%0d msrq-enq: addr%0d=%0h, ready=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, tail_ptr, `LINE_TO_BYTE_ADDR(enqueue_addr_st3, BANK_ID), enqueue_ready_st3, debug_wid_st3, debug_pc_st3); $display("%t: cache%0d:%0d msrq-enq: addr%0d=%0h, ready=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, tail_ptr, `LINE_TO_BYTE_ADDR(enqueue_addr, BANK_ID), enqueue_ready, enq_debug_wid, enq_debug_pc);
end end
if (dequeue_st3) if (dequeue)
$display("%t: cache%0d:%0d msrq-deq addr%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, head_ptr, debug_wid_st3, debug_pc_st3); $display("%t: cache%0d:%0d msrq-deq addr%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, head_ptr, enq_debug_wid, enq_debug_pc);
$write("%t: cache%0d:%0d msrq-table", $time, CACHE_ID, BANK_ID); $write("%t: cache%0d:%0d msrq-table", $time, CACHE_ID, BANK_ID);
for (integer j = 0; j < MSHR_SIZE; j++) begin for (integer j = 0; j < MSHR_SIZE; j++) begin
if (valid_table[j]) begin if (valid_table[j]) begin

View file

@ -25,9 +25,7 @@ module VX_tag_access #(
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
`IGNORE_WARNINGS_BEGIN `IGNORE_WARNINGS_BEGIN
input wire[31:0] debug_pc, input wire[31:0] debug_pc,
input wire[`NR_BITS-1:0] debug_rd,
input wire[`NW_BITS-1:0] debug_wid, input wire[`NW_BITS-1:0] debug_wid,
input wire[`UP(CORE_TAG_ID_BITS)-1:0] debug_tagid,
`IGNORE_WARNINGS_END `IGNORE_WARNINGS_END
`endif `endif

View file

@ -193,11 +193,9 @@
"?valid_st0": 1, "?valid_st0": 1,
"?valid_st1": 1, "?valid_st1": 1,
"?valid_st2": 1, "?valid_st2": 1,
"?valid_st3": 1,
"addr_st0": 32, "addr_st0": 32,
"addr_st1": 32, "addr_st1": 32,
"addr_st2": 32, "addr_st2": 32,
"addr_st3": 32,
"is_fill_st0": 1, "is_fill_st0": 1,
"is_mshr_st0": 1, "is_mshr_st0": 1,
"miss_st1": 1, "miss_st1": 1,