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

cache fixes and opyimization - fmax moved from 162 mhz to 220 mhz!!!

Browse Source
This commit is contained in:
Blaise Tine
2020-11-05 03:49:50 -08:00
parent 4c6a74fa87
commit af2bb3b789
8 changed files with 334 additions and 212 deletions
Download Patch File Download Diff File Expand all files Collapse all files

143
hw/rtl/cache/VX_bank.v vendored
View File

@@ -197,14 +197,14 @@ module VX_bank #(
wire reqq_pop; wire reqq_pop;
wire reqq_empty; wire reqq_empty;
wire reqq_full; wire reqq_full;
wire [`REQS_BITS-1:0] reqq_req_tid_st0; wire [`REQS_BITS-1:0] reqq_tid_st0;
wire reqq_req_rw_st0; wire reqq_rw_st0;
wire [WORD_SIZE-1:0] reqq_req_byteen_st0; wire [WORD_SIZE-1:0] reqq_byteen_st0;
`IGNORE_WARNINGS_BEGIN `IGNORE_WARNINGS_BEGIN
wire [`WORD_ADDR_WIDTH-1:0] reqq_req_addr_st0; wire [`WORD_ADDR_WIDTH-1:0] reqq_addr_st0;
`IGNORE_WARNINGS_END `IGNORE_WARNINGS_END
wire [`WORD_WIDTH-1:0] reqq_req_writeword_st0; wire [`WORD_WIDTH-1:0] reqq_writeword_st0;
wire [CORE_TAG_WIDTH-1:0] reqq_req_tag_st0; wire [CORE_TAG_WIDTH-1:0] reqq_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 = !reqq_full;
@@ -229,12 +229,12 @@ module VX_bank #(
// Dequeue // Dequeue
.reqq_pop (reqq_pop), .reqq_pop (reqq_pop),
.reqq_req_tid_st0 (reqq_req_tid_st0), .reqq_tid_st0 (reqq_tid_st0),
.reqq_req_rw_st0 (reqq_req_rw_st0), .reqq_rw_st0 (reqq_rw_st0),
.reqq_req_byteen_st0 (reqq_req_byteen_st0), .reqq_byteen_st0 (reqq_byteen_st0),
.reqq_req_addr_st0 (reqq_req_addr_st0), .reqq_addr_st0 (reqq_addr_st0),
.reqq_req_writedata_st0(reqq_req_writeword_st0), .reqq_writedata_st0(reqq_writeword_st0),
.reqq_req_tag_st0 (reqq_req_tag_st0), .reqq_tag_st0 (reqq_tag_st0),
.reqq_empty (reqq_empty), .reqq_empty (reqq_empty),
.reqq_full (reqq_full) .reqq_full (reqq_full)
); );
@@ -309,12 +309,12 @@ module VX_bank #(
assign addr_st0 = msrq_pop_unqual ? msrq_addr_st0 : assign addr_st0 = msrq_pop_unqual ? msrq_addr_st0 :
dfpq_pop_unqual ? dfpq_addr_st0 : dfpq_pop_unqual ? dfpq_addr_st0 :
reqq_pop_unqual ? reqq_req_addr_st0[`LINE_SELECT_ADDR_RNG] : reqq_pop_unqual ? reqq_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_req_addr_st0[`WORD_SELECT_WIDTH-1:0] : assign wsel_st0 = reqq_pop_unqual ? reqq_addr_st0[`WORD_SELECT_WIDTH-1:0] :
msrq_pop_unqual ? msrq_wsel_st0 : msrq_pop_unqual ? msrq_wsel_st0 :
0; 0;
end else begin end else begin
@@ -325,7 +325,7 @@ module VX_bank #(
assign writedata_st0 = dfpq_filldata_st0; assign writedata_st0 = dfpq_filldata_st0;
assign inst_meta_st0 = msrq_pop_unqual ? {`REQ_TAG_WIDTH'(msrq_tag_st0) , msrq_rw_st0, msrq_byteen_st0, msrq_tid_st0} : assign inst_meta_st0 = msrq_pop_unqual ? {`REQ_TAG_WIDTH'(msrq_tag_st0) , msrq_rw_st0, msrq_byteen_st0, msrq_tid_st0} :
reqq_pop_unqual ? {`REQ_TAG_WIDTH'(reqq_req_tag_st0), reqq_req_rw_st0, reqq_req_byteen_st0, reqq_req_tid_st0} : reqq_pop_unqual ? {`REQ_TAG_WIDTH'(reqq_tag_st0), reqq_rw_st0, reqq_byteen_st0, reqq_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;
@@ -338,7 +338,7 @@ module VX_bank #(
0; 0;
assign writeword_st0 = msrq_pop_unqual ? msrq_writeword_st0 : assign writeword_st0 = msrq_pop_unqual ? msrq_writeword_st0 :
reqq_pop_unqual ? reqq_req_writeword_st0 : reqq_pop_unqual ? reqq_writeword_st0 :
0; 0;
// we have a miss in msrq or in stage 3 for the current address // we have a miss in msrq or in stage 3 for the current address
@@ -548,24 +548,24 @@ module VX_bank #(
// 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] miss_add_tid; wire[`REQS_BITS-1:0] req_tid_st3;
wire[`REQ_TAG_WIDTH-1:0] miss_add_tag; wire[`REQ_TAG_WIDTH-1:0] req_tag_st3;
wire miss_add_rw; wire req_rw_st3;
wire[WORD_SIZE-1:0] miss_add_byteen; wire[WORD_SIZE-1:0] req_byteen_st3;
wire miss_add_unqual = miss_st3 || force_miss_st3; wire msrq_push_unqual = miss_st3 || force_miss_st3;
assign msrq_push_stall = (miss_st3 || force_miss_st3) && msrq_full; assign msrq_push_stall = (miss_st3 || force_miss_st3) && msrq_full;
wire miss_add = miss_add_unqual wire msrq_push = msrq_push_unqual
&& !msrq_full && !msrq_full
&& !cwbq_push_stall && !cwbq_push_stall
&& !dwbq_push_stall && !dwbq_push_stall
&& !snpq_push_stall; && !snpq_push_stall;
assign {miss_add_tag, miss_add_rw, miss_add_byteen, miss_add_tid} = inst_meta_st3; assign {req_tag_st3, req_rw_st3, req_byteen_st3, req_tid_st3} = inst_meta_st3;
// a matching fill request is comming // check if a matching fill request is comming
wire incoming_st0_fill_st3 = 0 /*is_fill_st0 && (addr_st3 == dfpq_addr_st0)*/; // clock delay issue wire incoming_st0_fill_st3 = is_fill_st0 && (addr_st3 == dfpq_addr_st0);
wire incoming_st1_fill_st3 = is_fill_st1 && (addr_st3 == addr_st1); wire incoming_st1_fill_st3 = is_fill_st1 && (addr_st3 == addr_st1);
wire incoming_st2_fill_st3 = is_fill_st2 && (addr_st3 == addr_st2); wire incoming_st2_fill_st3 = is_fill_st2 && (addr_st3 == addr_st2);
wire incoming_fill = incoming_st2_fill_st3 wire incoming_fill = incoming_st2_fill_st3
@@ -573,15 +573,15 @@ module VX_bank #(
|| incoming_st0_fill_st3; || incoming_st0_fill_st3;
if (DRAM_ENABLE) begin if (DRAM_ENABLE) begin
wire msrq_dequeue_st3 = valid_st3 && is_msrq_st3 && !miss_add_unqual && !pipeline_stall; wire msrq_dequeue_st3 = valid_st3 && is_msrq_st3 && !msrq_push_unqual && !pipeline_stall;
// mark msrq entry that match DRAM fill as 'ready' // mark msrq entry that match DRAM fill as 'ready'
wire update_ready_st0 = dfpq_pop; wire update_ready_st0 = dfpq_pop;
// push missed requests as 'ready' // push missed requests as 'ready' is this was a forced missed
// if it didn't actually missed but had to abort because of pending requets in msrq // this request will be queued behind prior requests so will only pop when the fill arrives.
// or if a matching fill request is coming wire msrq_init_ready_state_st3 = !miss_st3
wire msrq_init_ready_state_st3 = !miss_st3 || incoming_fill; || incoming_fill;
VX_cache_miss_resrv #( VX_cache_miss_resrv #(
.BANK_ID (BANK_ID), .BANK_ID (BANK_ID),
@@ -610,14 +610,14 @@ module VX_bank #(
`endif `endif
// enqueue // enqueue
.enqueue_st3 (miss_add), .enqueue_st3 (msrq_push),
.enqueue_addr_st3 (addr_st3), .enqueue_addr_st3 (addr_st3),
.enqueue_wsel_st3 (wsel_st3), .enqueue_wsel_st3 (wsel_st3),
.enqueue_data_st3 (writeword_st3), .enqueue_data_st3 (writeword_st3),
.enqueue_tid_st3 (miss_add_tid), .enqueue_tid_st3 (req_tid_st3),
.enqueue_tag_st3 (miss_add_tag), .enqueue_tag_st3 (req_tag_st3),
.enqueue_rw_st3 (miss_add_rw), .enqueue_rw_st3 (req_rw_st3),
.enqueue_byteen_st3 (miss_add_byteen), .enqueue_byteen_st3 (req_byteen_st3),
.enqueue_is_snp_st3 (is_snp_st3), .enqueue_is_snp_st3 (is_snp_st3),
.enqueue_snp_inv_st3 (snp_invalidate_st3), .enqueue_snp_inv_st3 (snp_invalidate_st3),
.enqueue_msrq_st3 (is_msrq_st3), .enqueue_msrq_st3 (is_msrq_st3),
@@ -645,11 +645,11 @@ module VX_bank #(
.dequeue_st3 (msrq_dequeue_st3) .dequeue_st3 (msrq_dequeue_st3)
); );
end else begin end else begin
`UNUSED_VAR (miss_add) `UNUSED_VAR (msrq_push)
`UNUSED_VAR (wsel_st3) `UNUSED_VAR (wsel_st3)
`UNUSED_VAR (writeword_st3) `UNUSED_VAR (writeword_st3)
`UNUSED_VAR (snp_invalidate_st3) `UNUSED_VAR (snp_invalidate_st3)
`UNUSED_VAR (miss_add_byteen) `UNUSED_VAR (req_byteen_st3)
assign msrq_pending_hazard_unqual_st0 = 0; assign msrq_pending_hazard_unqual_st0 = 0;
assign msrq_full = 0; assign msrq_full = 0;
assign msrq_almfull = 0; assign msrq_almfull = 0;
@@ -669,7 +669,7 @@ module VX_bank #(
wire cwbq_empty, cwbq_full; wire cwbq_empty, cwbq_full;
wire cwbq_push_unqual = valid_st3 && !is_fill_st3 && !is_snp_st3 && !miss_st3 && !force_miss_st3 && !miss_add_rw; wire cwbq_push_unqual = valid_st3 && !is_fill_st3 && !is_snp_st3 && !miss_st3 && !force_miss_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
@@ -680,9 +680,9 @@ module VX_bank #(
wire cwbq_pop = core_rsp_valid && core_rsp_ready; wire cwbq_pop = core_rsp_valid && core_rsp_ready;
wire [`REQS_BITS-1:0] cwbq_tid = miss_add_tid; wire [`REQS_BITS-1:0] cwbq_tid_st3 = req_tid_st3;
wire [CORE_TAG_WIDTH-1:0] cwbq_tag = CORE_TAG_WIDTH'(miss_add_tag); wire [CORE_TAG_WIDTH-1:0] cwbq_tag_st3 = CORE_TAG_WIDTH'(req_tag_st3);
wire [`WORD_WIDTH-1:0] cwbq_data = readword_st3; wire [`WORD_WIDTH-1:0] cwbq_data_st3 = readword_st3;
VX_generic_queue #( VX_generic_queue #(
.DATAW(`REQS_BITS + CORE_TAG_WIDTH + `WORD_WIDTH), .DATAW(`REQS_BITS + CORE_TAG_WIDTH + `WORD_WIDTH),
@@ -692,7 +692,7 @@ module VX_bank #(
.reset (reset), .reset (reset),
.push (cwbq_push), .push (cwbq_push),
.pop (cwbq_pop), .pop (cwbq_pop),
.data_in ({cwbq_tid, cwbq_tag, cwbq_data}), .data_in ({cwbq_tid_st3, cwbq_tag_st3, cwbq_data_st3}),
.data_out({core_rsp_tid, core_rsp_tag, core_rsp_data}), .data_out({core_rsp_tid, core_rsp_tag, core_rsp_data}),
.empty (cwbq_empty), .empty (cwbq_empty),
.full (cwbq_full), .full (cwbq_full),
@@ -705,13 +705,14 @@ module VX_bank #(
wire dwbq_empty, dwbq_full; wire dwbq_empty, dwbq_full;
wire dwbq_is_dfl_in = valid_st3 && miss_st3 && !incoming_fill && (!force_miss_st3 || is_msrq_st3); wire dwbq_is_dfl_in = valid_st3 && miss_st3 && (!force_miss_st3 || is_msrq_st3);
wire dwbq_is_dwb_in = valid_st3 && dirty_st3 && !force_miss_st3 && (is_fill_st3 || is_snp_st3); wire dwbq_is_dwb_in = valid_st3 && dirty_st3 && (is_fill_st3 || (!force_miss_st3 && is_snp_st3));
wire dwbq_push_unqual = dwbq_is_dfl_in || dwbq_is_dwb_in; wire dwbq_push_unqual = dwbq_is_dfl_in || dwbq_is_dwb_in;
assign dwbq_push_stall = dwbq_push_unqual && dwbq_full; assign dwbq_push_stall = dwbq_push_unqual && dwbq_full;
wire dwbq_push = dwbq_push_unqual wire dwbq_push = dwbq_push_unqual
&& !(dwbq_is_dfl_in && incoming_fill) // not in 'dwbq_push_stall' to reduce clock delay
&& !dwbq_full && !dwbq_full
&& !msrq_push_stall && !msrq_push_stall
&& !cwbq_push_stall && !cwbq_push_stall
@@ -719,11 +720,10 @@ module VX_bank #(
wire dwbq_pop = dram_req_valid && dram_req_ready; wire dwbq_pop = dram_req_valid && dram_req_ready;
if (DRAM_ENABLE) begin
wire [`LINE_ADDR_WIDTH-1:0] dwbq_req_addr = dwbq_is_dwb_in ? {readtag_st3, addr_st3[`LINE_SELECT_BITS-1:0]} : wire [`LINE_ADDR_WIDTH-1:0] dwbq_req_addr = dwbq_is_dwb_in ? {readtag_st3, addr_st3[`LINE_SELECT_BITS-1:0]} :
addr_st3; addr_st3;
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)
@@ -744,13 +744,14 @@ module VX_bank #(
`UNUSED_VAR (readtag_st3) `UNUSED_VAR (readtag_st3)
`UNUSED_VAR (dirtyb_st3) `UNUSED_VAR (dirtyb_st3)
`UNUSED_VAR (readdata_st3) `UNUSED_VAR (readdata_st3)
`UNUSED_VAR (dwbq_req_addr)
`UNUSED_VAR (dram_req_ready)
assign dwbq_empty = 1; assign dwbq_empty = 1;
assign dwbq_full = 0; assign dwbq_full = 0;
assign dram_req_rw = 0; assign dram_req_rw = 0;
assign dram_req_byteen = 0; assign dram_req_byteen = 0;
assign dram_req_addr = 0; assign dram_req_addr = 0;
assign dram_req_data = 0; assign dram_req_data = 0;
`UNUSED_VAR (dram_req_ready)
end end
assign dram_req_valid = !dwbq_empty; assign dram_req_valid = !dwbq_empty;
@@ -771,10 +772,9 @@ module VX_bank #(
wire snpq_pop = snp_rsp_valid && snp_rsp_ready; wire snpq_pop = snp_rsp_valid && snp_rsp_ready;
wire [SNP_REQ_TAG_WIDTH-1:0] snpq_tag_st3 = SNP_REQ_TAG_WIDTH'(miss_add_tag); wire [SNP_REQ_TAG_WIDTH-1:0] snpq_tag_st3 = SNP_REQ_TAG_WIDTH'(req_tag_st3);
if (FLUSH_ENABLE) begin if (FLUSH_ENABLE) begin
VX_generic_queue #( VX_generic_queue #(
.DATAW(SNP_REQ_TAG_WIDTH), .DATAW(SNP_REQ_TAG_WIDTH),
.SIZE(SNPQ_SIZE) .SIZE(SNPQ_SIZE)
@@ -809,7 +809,7 @@ module VX_bank #(
|| snpq_push_stall; || snpq_push_stall;
`SCOPE_ASSIGN (valid_st0, valid_st0); `SCOPE_ASSIGN (valid_st0, valid_st0);
`SCOPE_ASSIGN (valid_st, 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 (valid_st3, valid_st3);
@@ -817,7 +817,7 @@ module VX_bank #(
`SCOPE_ASSIGN (miss_st1, miss_st1); `SCOPE_ASSIGN (miss_st1, miss_st1);
`SCOPE_ASSIGN (dirty_st1, dirty_st1); `SCOPE_ASSIGN (dirty_st1, dirty_st1);
`SCOPE_ASSIGN (force_miss_st1, force_miss_st1); `SCOPE_ASSIGN (force_miss_st1, force_miss_st1);
`SCOPE_ASSIGN (stall_pipe, pipeline_stall); `SCOPE_ASSIGN (pipeline_stall, pipeline_stall);
`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));
@@ -826,48 +826,33 @@ module VX_bank #(
`ifdef DBG_PRINT_CACHE_BANK `ifdef DBG_PRINT_CACHE_BANK
always @(posedge clk) begin always @(posedge clk) begin
if (miss_st3 && (incoming_st0_fill_st3 || incoming_st2_fill_st3)) begin if (miss_st3 && (incoming_st0_fill_st3 || incoming_st1_fill_st3 || incoming_st2_fill_st3)) begin
$display("%t: incoming fill - addr=%0h, st0=%b, st1=%b", $time, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), incoming_st0_fill_st3, incoming_st2_fill_st3); $display("%t: incoming fill - addr=%0h, st0=%b, st1=%b, st2=%b", $time, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), incoming_st0_fill_st3, incoming_st1_fill_st3, incoming_st2_fill_st3);
assert(!is_msrq_st3); assert(!is_msrq_st3);
end end
if ((|core_req_valid) && core_req_ready) begin if (pipeline_stall) begin
$display("%t: cache%0d:%0d core-req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(core_req_addr[0], BANK_ID), core_req_tag); $display("%t: cache%0d:%0d pipeline-stall: msrq=%b, cwbq=%b, dwbq=%b, snpq=%b", $time, CACHE_ID, BANK_ID, msrq_push_stall, cwbq_push_stall, dwbq_push_stall, snpq_push_stall);
end
if (core_rsp_valid && core_rsp_ready) begin
$display("%t: cache%0d:%0d core-rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
end
if (dram_req_valid && dram_req_ready) begin
$display("%t: cache%0d:%0d dram-req: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_req_addr, BANK_ID), dram_req_data);
end
if (dram_rsp_valid && dram_rsp_ready) begin
$display("%t: cache%0d:%0d dram-rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_rsp_addr, BANK_ID), dram_rsp_data);
end
if (snp_req_valid && snp_req_ready) begin
$display("%t: cache%0d:%0d snp-req: addr=%0h, tag=%0h, invalidate=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(snp_req_addr, BANK_ID), snp_req_tag, snp_req_invalidate);
end
if (snp_rsp_valid && snp_rsp_ready) begin
$display("%t: cache%0d:%0d snp-rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
end
if (msrq_pop) begin
$display("%t: cache%0d:%0d msrq-pop: addr=%0h wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), debug_wid_st0, debug_pc_st0);
end end
if (dfpq_pop) begin if (dfpq_pop) begin
$display("%t: cache%0d:%0d dfpq-pop: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID)); $display("%t: cache%0d:%0d dram-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 (reqq_pop) begin
$display("%t: cache%0d:%0d reqq-pop: addr=%0h wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), debug_wid_st0, debug_pc_st0); $display("%t: cache%0d:%0d core-req: addr=%0h, tag=%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), reqq_tag_st0, debug_wid_st0, debug_pc_st0);
end end
if (snrq_pop) begin if (snrq_pop) begin
$display("%t: cache%0d:%0d snrq-pop: addr=%0h tag=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), snrq_tag_st0); $display("%t: cache%0d:%0d snp-req: addr=%0h, tag=%0d, invalidate=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st0, BANK_ID), snrq_tag_st0, snrq_invalidate_st0);
end end
if (cwbq_push) begin if (cwbq_push) begin
$display("%t: cache%0d:%0d cwbq-push: addr=%0h wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), debug_wid_st3, debug_pc_st3); $display("%t: cache%0d:%0d core-rsp: addr=%0h, tag=%0d, data=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), cwbq_tag_st3, cwbq_data_st3, debug_wid_st3, debug_pc_st3);
end end
if (dwbq_push) begin if (dwbq_push) begin
$display("%t: cache%0d:%0d dwbq-push: addr=%0h wid=%0d, PC=%0h, rw=%b", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), debug_wid_st3, debug_pc_st3, dwbq_is_dwb_in); if (dwbq_is_dwb_in)
$display("%t: cache%0d:%0d dram-wb: addr=%0h, data=%0h, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dwbq_req_addr, BANK_ID), readdata_st3, dirtyb_st3, debug_wid_st3, debug_pc_st3);
else
$display("%t: cache%0d:%0d dram-fill: addr=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dwbq_req_addr, BANK_ID), debug_wid_st3, debug_pc_st3);
end end
if (snpq_push) begin if (snpq_push) begin
$display("%t: cache%0d:%0d snpq-push: addr=%0h tag=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), snpq_tag_st3); $display("%t: cache%0d:%0d snp-rsp: addr=%0h, tag=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st3, BANK_ID), snpq_tag_st3);
end end
end end
`endif `endif

28
hw/rtl/cache/VX_bank_core_req_arb.v vendored
View File

@@ -26,12 +26,12 @@ module VX_bank_core_req_arb #(
// Dequeue Data // Dequeue Data
input wire reqq_pop, input wire reqq_pop,
output wire [`REQS_BITS-1:0] reqq_req_tid_st0, output wire [`REQS_BITS-1:0] reqq_tid_st0,
output wire reqq_req_rw_st0, output wire reqq_rw_st0,
output wire [WORD_SIZE-1:0] reqq_req_byteen_st0, output wire [WORD_SIZE-1:0] reqq_byteen_st0,
output wire [`WORD_ADDR_WIDTH-1:0] reqq_req_addr_st0, output wire [`WORD_ADDR_WIDTH-1:0] reqq_addr_st0,
output wire [`WORD_WIDTH-1:0] reqq_req_writedata_st0, output wire [`WORD_WIDTH-1:0] reqq_writedata_st0,
output wire [CORE_TAG_WIDTH-1:0] reqq_req_tag_st0, output wire [CORE_TAG_WIDTH-1:0] reqq_tag_st0,
// State Data // State Data
output wire reqq_empty, output wire reqq_empty,
@@ -106,17 +106,17 @@ module VX_bank_core_req_arb #(
); );
assign reqq_empty = !qual_has_request; assign reqq_empty = !qual_has_request;
assign reqq_req_tid_st0 = qual_request_index; assign reqq_tid_st0 = qual_request_index;
assign reqq_req_byteen_st0 = qual_byteen[qual_request_index]; assign reqq_byteen_st0 = qual_byteen[qual_request_index];
assign reqq_req_addr_st0 = qual_addr[qual_request_index]; assign reqq_addr_st0 = qual_addr[qual_request_index];
assign reqq_req_writedata_st0 = qual_writedata[qual_request_index]; assign reqq_writedata_st0 = qual_writedata[qual_request_index];
if (CORE_TAG_ID_BITS != 0) begin if (CORE_TAG_ID_BITS != 0) begin
assign reqq_req_tag_st0 = qual_tag; assign reqq_tag_st0 = qual_tag;
assign reqq_req_rw_st0 = qual_rw; assign reqq_rw_st0 = qual_rw;
end else begin end else begin
assign reqq_req_tag_st0 = qual_tag[qual_request_index]; assign reqq_tag_st0 = qual_tag[qual_request_index];
assign reqq_req_rw_st0 = qual_rw[qual_request_index]; assign reqq_rw_st0 = qual_rw[qual_request_index];
end end
`DEBUG_BLOCK( `DEBUG_BLOCK(

8
hw/rtl/cache/VX_cache_miss_resrv.v vendored
View File

@@ -181,14 +181,14 @@ module VX_cache_miss_resrv #(
if (enqueue_st3 || schedule_st0 || dequeue_st3) begin if (enqueue_st3 || schedule_st0 || dequeue_st3) begin
if (enqueue_st3) begin if (enqueue_st3) begin
if (enqueue_msrq_st3) if (enqueue_msrq_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_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_st3, BANK_ID), enqueue_ready_st3);
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_st3, BANK_ID), enqueue_ready_st3, debug_wid_st3, debug_pc_st3);
end end
if (schedule_st0) if (schedule_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(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(dequeue_addr_st0, BANK_ID), debug_wid_st0, debug_pc_st0);
if (dequeue_st3) if (dequeue_st3)
$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, debug_wid_st3, debug_pc_st3);
$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 < MRVQ_SIZE; j++) begin for (integer j = 0; j < MRVQ_SIZE; j++) begin
if (valid_table[j]) begin if (valid_table[j]) begin

25
hw/rtl/cache/VX_tag_access.v vendored
View File

@@ -77,8 +77,6 @@ module VX_tag_access #(
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.stall (stall),
.read_addr (addrline_st1), .read_addr (addrline_st1),
.read_valid (qual_read_valid_st1), .read_valid (qual_read_valid_st1),
.read_dirty (qual_read_dirty_st1), .read_dirty (qual_read_dirty_st1),
@@ -109,14 +107,17 @@ module VX_tag_access #(
wire fill_write = valid_req_st1 && writefill_st1 wire fill_write = valid_req_st1 && writefill_st1
&& !tags_match; // discard redundant fills because the block could be dirty && !tags_match; // discard redundant fills because the block could be dirty
assign use_write_enable = normal_write || fill_write; assign use_write_enable = (normal_write || fill_write)
&& !stall;
assign use_invalidate = valid_req_st1 && is_snp_st1 && tags_match assign use_invalidate = valid_req_st1 && is_snp_st1
&& (use_read_dirty_st1 || snp_invalidate_st1) // block is dirty or need to force invalidation && tags_match
&& !force_miss_st1; && (use_read_dirty_st1 || snp_invalidate_st1) // block is dirty or should invalidate
&& !force_miss_st1
&& !stall;
wire core_req_miss = valid_req_st1 && !is_snp_st1 && !writefill_st1 // is core request wire core_req_miss = valid_req_st1 && !is_snp_st1 && !writefill_st1
&& (!use_read_valid_st1 || !tags_match); // block missing or has wrong tag && !tags_match;
assign miss_st1 = core_req_miss; assign miss_st1 = core_req_miss;
assign dirty_st1 = valid_req_st1 && use_read_valid_st1 && use_read_dirty_st1; assign dirty_st1 = valid_req_st1 && use_read_valid_st1 && use_read_dirty_st1;
@@ -130,15 +131,15 @@ module VX_tag_access #(
$display("%t: warning: redundant fill - addr=%0h", $time, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID)); $display("%t: warning: redundant fill - addr=%0h", $time, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID));
end end
if (miss_st1) begin if (miss_st1) begin
$display("%t: cache%0d:%0d data-miss: addr=%0h, wid=%0d, PC=%0h, valid=%b, tagmatch=%b, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), debug_wid_st1, debug_pc_st1, use_read_dirty_st1, tags_match, addrline_st1, addrtag_st1); $display("%t: cache%0d:%0d tag-miss: addr=%0h, wid=%0d, PC=%0h, valid=%b, blk_tag_id=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), debug_wid_st1, debug_pc_st1, use_read_dirty_st1, qual_read_tag_st1, addrline_st1, addrtag_st1);
end else if ((| use_write_enable)) begin end else if ((| use_write_enable)) begin
if (writefill_st1) begin if (writefill_st1) begin
$display("%t: cache%0d:%0d data-fill: addr=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), addrline_st1, addrtag_st1); $display("%t: cache%0d:%0d tag-fill: addr=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), addrline_st1, addrtag_st1);
end else begin end else begin
$display("%t: cache%0d:%0d data-write: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), debug_wid_st1, debug_pc_st1, addrline_st1, addrtag_st1); $display("%t: cache%0d:%0d tag-write: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), debug_wid_st1, debug_pc_st1, addrline_st1, addrtag_st1);
end end
end else begin end else begin
$display("%t: cache%0d:%0d data-read: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), debug_wid_st1, debug_pc_st1, addrline_st1, qual_read_tag_st1); $display("%t: cache%0d:%0d tag-read: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), debug_wid_st1, debug_pc_st1, addrline_st1, qual_read_tag_st1);
end end
end end
end end

4
hw/rtl/cache/VX_tag_store.v vendored
View File

@@ -13,8 +13,6 @@ module VX_tag_store #(
input wire clk, input wire clk,
input wire reset, input wire reset,
input wire stall,
input wire write_enable, input wire write_enable,
input wire write_fill, input wire write_fill,
input wire[`LINE_SELECT_BITS-1:0] write_addr, input wire[`LINE_SELECT_BITS-1:0] write_addr,
@@ -35,7 +33,7 @@ module VX_tag_store #(
valid[i] <= 0; valid[i] <= 0;
dirty[i] <= 0; dirty[i] <= 0;
end end
end else if(!stall) begin end else begin
if (write_enable) begin if (write_enable) begin
assert(!invalidate); assert(!invalidate);
dirty[write_addr] <= !write_fill; dirty[write_addr] <= !write_fill;

137
hw/rtl/libs/VX_dp_ram.v
View File

@@ -8,7 +8,8 @@ module VX_dp_ram #(
parameter RWCHECK = 1, parameter RWCHECK = 1,
parameter RWBYPASS = 0, parameter RWBYPASS = 0,
parameter ADDRW = $clog2(SIZE), parameter ADDRW = $clog2(SIZE),
parameter SIZEW = $clog2(SIZE+1) parameter SIZEW = $clog2(SIZE+1),
parameter FASTRAM = 0
) ( ) (
input wire clk, input wire clk,
input wire [ADDRW-1:0] waddr, input wire [ADDRW-1:0] waddr,
@@ -20,6 +21,139 @@ module VX_dp_ram #(
output wire [DATAW-1:0] dout output wire [DATAW-1:0] dout
); );
if (FASTRAM) begin
if (BUFFERED) begin
`USE_FAST_BRAM reg [DATAW-1:0] mem [SIZE-1:0];
reg [DATAW-1:0] dout_r;
if (BYTEENW > 1) begin
always @(posedge clk) begin
if (wren) begin
for (integer i = 0; i < BYTEENW; i++) begin
if (byteen[i])
mem[waddr][i * 8 +: 8] <= din[i * 8 +: 8];
end
end
end
end else begin
always @(posedge clk) begin
if (wren && byteen)
mem[waddr] <= din;
end
end
always @(posedge clk) begin
if (rden)
dout_r <= mem[raddr];
end
if (RWBYPASS) begin
reg [DATAW-1:0] din_r;
wire writing;
if (BYTEENW > 1) begin
always @(posedge clk) begin
if (wren) begin
for (integer i = 0; i < BYTEENW; i++) begin
din_r[i * 8 +: 8] <= byteen[i] ? din[i * 8 +: 8] : mem[waddr][i * 8 +: 8];
end
end
end
end else begin
always @(posedge clk) begin
din_r <= din;
end
end
reg bypass_r;
always @(posedge clk) begin
bypass_r <= wren && (raddr == waddr);
end
assign dout = bypass_r ? din_r : dout_r;
end else begin
assign dout = dout_r;
end
end else begin
`UNUSED_VAR (rden)
if (RWCHECK) begin
`USE_FAST_BRAM reg [DATAW-1:0] mem [SIZE-1:0];
if (BYTEENW > 1) begin
always @(posedge clk) begin
if (wren) begin
for (integer i = 0; i < BYTEENW; i++) begin
if (byteen[i])
mem[waddr][i * 8 +: 8] <= din[i * 8 +: 8];
end
end
end
end else begin
always @(posedge clk) begin
if (wren && byteen)
mem[waddr] <= din;
end
end
if (RWBYPASS) begin
reg [DATAW-1:0] din_r;
wire writing;
if (BYTEENW > 1) begin
always @(posedge clk) begin
if (wren) begin
for (integer i = 0; i < BYTEENW; i++) begin
din_r[i * 8 +: 8] <= byteen[i] ? din[i * 8 +: 8] : mem[waddr][i * 8 +: 8];
end
end
end
end else begin
always @(posedge clk) begin
din_r <= din;
end
end
reg bypass_r;
always @(posedge clk) begin
bypass_r <= writing && (raddr == waddr);
end
assign dout = bypass_r ? din_r : mem[raddr];
end else begin
assign dout = mem[raddr];
end
end else begin
`USE_FAST_BRAM `NO_RW_RAM_CHECK reg [DATAW-1:0] mem [SIZE-1:0];
if (BYTEENW > 1) begin
always @(posedge clk) begin
if (wren) begin
for (integer i = 0; i < BYTEENW; i++) begin
if (byteen[i])
mem[waddr][i * 8 +: 8] <= din[i * 8 +: 8];
end
end
end
end else begin
always @(posedge clk) begin
if (wren && byteen)
mem[waddr] <= din;
end
end
assign dout = mem[raddr];
end
end
end else begin
if (BUFFERED) begin if (BUFFERED) begin
reg [DATAW-1:0] mem [SIZE-1:0]; reg [DATAW-1:0] mem [SIZE-1:0];
@@ -148,5 +282,6 @@ module VX_dp_ram #(
assign dout = mem[raddr]; assign dout = mem[raddr];
end end
end end
end
endmodule endmodule

9
hw/rtl/libs/VX_generic_queue.v
View File

@@ -5,7 +5,8 @@ module VX_generic_queue #(
parameter SIZE = 2, parameter SIZE = 2,
parameter BUFFERED = 0, parameter BUFFERED = 0,
parameter ADDRW = $clog2(SIZE), parameter ADDRW = $clog2(SIZE),
parameter SIZEW = $clog2(SIZE+1) parameter SIZEW = $clog2(SIZE+1),
parameter FASTRAM = 1
) ( ) (
input wire clk, input wire clk,
input wire reset, input wire reset,
@@ -108,7 +109,8 @@ module VX_generic_queue #(
.DATAW(DATAW), .DATAW(DATAW),
.SIZE(SIZE), .SIZE(SIZE),
.BUFFERED(0), .BUFFERED(0),
.RWCHECK(1) .RWCHECK(1),
.FASTRAM(FASTRAM)
) dp_ram ( ) dp_ram (
.clk(clk), .clk(clk),
.waddr(wr_ptr_a), .waddr(wr_ptr_a),
@@ -161,7 +163,8 @@ module VX_generic_queue #(
.DATAW(DATAW), .DATAW(DATAW),
.SIZE(SIZE), .SIZE(SIZE),
.BUFFERED(1), .BUFFERED(1),
.RWCHECK(0) .RWCHECK(0),
.FASTRAM(FASTRAM)
) dp_ram ( ) dp_ram (
.clk(clk), .clk(clk),
.waddr(wr_ptr_r), .waddr(wr_ptr_r),

4
hw/scripts/scope.json
View File

@@ -208,11 +208,11 @@
"addr_st1": 32, "addr_st1": 32,
"addr_st2": 32, "addr_st2": 32,
"addr_st3": 32, "addr_st3": 32,
"is_mrvq_st1": 1, "is_msrq_st1": 1,
"miss_st1": 1, "miss_st1": 1,
"dirty_st1": 1, "dirty_st1": 1,
"!force_miss_st1": 1, "!force_miss_st1": 1,
"!stall_pipe": 1 "!pipeline_stall": 1
} }
} }
} }