github-mirrors/vortex
1
0
Fork 0
mirror of https://github.com/vortexgpgpu/vortex.git synced 2025-04-23 13:27:29 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

cache specialization for in-order DRAM reponses

Browse source
This commit is contained in:
Blaise Tine 2021-02-13 20:23:29 -08:00
parent 4aaaebab6e
commit 3c37db877a
7 changed files with 84 additions and 64 deletions
Download patch file Download diff file Expand all files Collapse all files

4
hw/rtl/VX_config.vh
View file

@ -244,7 +244,7 @@
// Size of LSU Request Queue
`ifndef LSUQ_SIZE
`define LSUQ_SIZE (`NUM_WARPS * `NUM_THREADS)
`define LSUQ_SIZE 8
`endif
// Size of FPU Request Queue
@ -313,7 +313,7 @@
// Miss Handling Register Size
`ifndef DMSHR_SIZE
`define DMSHR_SIZE (`LSUQ_SIZE / 2)
`define DMSHR_SIZE `LSUQ_SIZE
`endif
// DRAM Request Queue Size

2
hw/rtl/VX_platform.vh
View file

@ -46,7 +46,7 @@
if (!(cond)) $error msg; \
endgenerate
`define SASSERT(cond, msg) \
`define RUNTIME_ASSERT(cond, msg) \
always @(posedge clk) \
assert(cond) else $error msg; \

72
hw/rtl/cache/VX_bank.v vendored
View file

@ -40,7 +40,10 @@ module VX_bank #(
parameter CORE_TAG_ID_BITS = 0,
// bank offset from beginning of index range
parameter BANK_ADDR_OFFSET = 0
parameter BANK_ADDR_OFFSET = 0,
// in-order DRAN
parameter IN_ORDER_DRAM = 0
) (
`SCOPE_IO_VX_bank
@ -85,8 +88,11 @@ module VX_bank #(
input wire dram_rsp_valid,
input wire [`LINE_ADDR_WIDTH-1:0] dram_rsp_addr,
input wire [`CACHE_LINE_WIDTH-1:0] dram_rsp_data,
input wire dram_rsp_flush,
output wire dram_rsp_ready
output wire dram_rsp_ready,
// flush
input wire flush_enable,
input wire [`LINE_SELECT_BITS-1:0] flush_addr
);
`ifdef DBG_CACHE_REQ_INFO
@ -97,7 +103,8 @@ module VX_bank #(
`endif
wire creq_pop;
wire creq_full, creq_empty;
wire creq_full;
wire creq_empty;
wire [NUM_PORTS-1:0] creq_pmask;
wire [NUM_PORTS-1:0][`UP(`WORD_SELECT_BITS)-1:0] creq_wsel;
wire [NUM_PORTS-1:0][WORD_SIZE-1:0] creq_byteen;
@ -181,14 +188,13 @@ module VX_bank #(
wire mshr_pop_unqual = mshr_valid
&& !dreq_alm_full; // ensure DRAM request queue not full (deadlock prevention)
wire drsq_pop_unqual = !mshr_pop_unqual && dram_rsp_valid;
wire creq_pop_unqual = !mshr_pop_unqual && !drsq_pop_unqual && !creq_empty;
wire creq_pop_unqual = !mshr_pop_unqual && !drsq_pop_unqual && !creq_empty && !flush_enable;
wire is_miss_st1 = valid_st1 && !is_fill_st1 && (miss_st1 || force_miss_st1);
wire is_miss_st1 = valid_st1 && (miss_st1 || force_miss_st1);
assign mshr_pop = mshr_pop_unqual
&& !crsq_alm_full // ensure core response ready
&& !(is_miss_st1 && is_mshr_st1); // do not schedule another mshr request if the previous one missed
&& !crsq_alm_full // ensure core response ready
&& !(!IN_ORDER_DRAM && is_miss_st1 && is_mshr_st1); // do not schedule another mshr request if the previous one missed
assign drsq_pop = drsq_pop_unqual;
assign creq_pop = creq_pop_unqual
@ -231,6 +237,14 @@ module VX_bank #(
assign creq_line_data = creq_data;
end
wire [`LINE_ADDR_WIDTH-1:0] dram_rsp_addr_qual;
if (IN_ORDER_DRAM) begin
`UNUSED_VAR (dram_rsp_addr)
assign dram_rsp_addr_qual = mshr_addr;
end else begin
assign dram_rsp_addr_qual = dram_rsp_addr;
end
VX_pipe_register #(
.DATAW (1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `CACHE_LINE_WIDTH + (`UP(`WORD_SELECT_BITS) + WORD_SIZE + `REQS_BITS + 1) * NUM_PORTS + CORE_TAG_WIDTH + 1 + 1),
.RESETW (1)
@ -239,21 +253,21 @@ module VX_bank #(
.reset (reset),
.enable (1'b1),
.data_in ({
mshr_pop || drsq_pop || creq_pop,
flush_enable || mshr_pop || drsq_pop || creq_pop,
flush_enable,
mshr_pop_unqual,
drsq_pop_unqual,
drsq_pop_unqual || flush_enable,
mshr_pop_unqual ? 1'b0 : creq_rw,
mshr_pop_unqual ? mshr_addr : (dram_rsp_valid ? dram_rsp_addr : creq_addr),
flush_enable ? (`LINE_ADDR_WIDTH'(flush_addr)) : (mshr_pop_unqual ? mshr_addr : (dram_rsp_valid ? dram_rsp_addr_qual : creq_addr)),
dram_rsp_valid ? dram_rsp_data : creq_line_data,
mshr_pop_unqual ? mshr_wsel : creq_wsel,
mshr_pop_unqual ? mshr_byteen : creq_byteen,
mshr_pop_unqual ? mshr_tid : creq_tid,
mshr_pop_unqual ? mshr_pmask : creq_pmask,
mshr_pop_unqual ? mshr_tag : creq_tag,
mshr_pending_sel,
dram_rsp_flush
mshr_pending_sel
}),
.data_out ({valid_st0, is_mshr_st0, is_fill_st0, mem_rw_st0, addr_st0, data_st0, wsel_st0, byteen_st0, req_tid_st0, pmask_st0, tag_st0, mshr_pending_st0, is_flush_st0})
.data_out ({valid_st0, is_flush_st0, is_mshr_st0, is_fill_st0, mem_rw_st0, addr_st0, data_st0, wsel_st0, byteen_st0, req_tid_st0, pmask_st0, tag_st0, mshr_pending_st0})
);
`ifdef DBG_CACHE_REQ_INFO
@ -263,6 +277,8 @@ module VX_bank #(
assign {debug_pc_st0, debug_wid_st0} = 0;
end
`endif
wire tag_match;
VX_tag_access #(
.BANK_ID (BANK_ID),
@ -286,24 +302,26 @@ module VX_bank #(
.addr (addr_st0),
.fill (valid_st0 && is_fill_st0),
.is_flush (is_flush_st0),
.missed (miss_st0)
.tag_match (tag_match)
);
// redundant fills
wire is_redundant_fill = is_fill_st0 && !miss_st0;
wire is_redundant_fill = !IN_ORDER_DRAM && is_fill_st0 && tag_match;
// we had a miss with prior request for the current address
assign prev_miss_dep_st0 = is_miss_st1 && (addr_st0 == addr_st1);
assign miss_st0 = !is_fill_st0 && !tag_match;
// force miss to ensure commit order when a new request has pending previous requests to same block
// also force a miss for mshr requests when previous requests got a miss
assign force_miss_st0 = (!is_fill_st0 && !is_mshr_st0 && (mshr_pending_st0 || prev_miss_dep_st0))
|| (is_mshr_st0 && is_miss_st1 && is_mshr_st1);
assign writeen_unqual_st0 = (!is_fill_st0 && !miss_st0 && mem_rw_st0)
assign writeen_unqual_st0 = (!is_fill_st0 && tag_match && mem_rw_st0)
|| (is_fill_st0 && !is_redundant_fill);
assign incoming_fill_st0 = dram_rsp_valid && (addr_st0 == dram_rsp_addr);
assign incoming_fill_st0 = dram_rsp_valid && (addr_st0 == dram_rsp_addr_qual);
VX_pipe_register #(
.DATAW (1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `CACHE_LINE_WIDTH + (`UP(`WORD_SELECT_BITS) + WORD_SIZE + `REQS_BITS + 1) * NUM_PORTS + CORE_TAG_WIDTH),
@ -330,7 +348,7 @@ module VX_bank #(
wire mshr_push_st1 = !is_fill_st1 && !mem_rw_st1 && (miss_st1 || force_miss_st1);
wire incoming_fill_qual_st1 = (dram_rsp_valid && (addr_st1 == dram_rsp_addr))
wire incoming_fill_qual_st1 = (dram_rsp_valid && (addr_st1 == dram_rsp_addr_qual))
|| incoming_fill_st1;
wire send_fill_req_st1 = !is_fill_st1 && !mem_rw_st1 && miss_st1
@ -394,13 +412,15 @@ module VX_bank #(
assign mshr_push = valid_st1 && mshr_push_st1;
wire mshr_dequeue = valid_st1 && is_mshr_st1 && !mshr_push_st1;
wire mshr_restore = !IN_ORDER_DRAM && is_mshr_st1;
`RUNTIME_ASSERT(!IN_ORDER_DRAM || !(mshr_push && mshr_restore), ("Oops!"))
// push a missed request as 'ready' if it was a forced miss that actually had a hit
// or the fill request for this block is comming
wire mshr_init_ready_state = !miss_st1 || incoming_fill_qual_st1;
// use dram rsp or core req address to lookup the mshr
wire [`LINE_ADDR_WIDTH-1:0] lookup_addr = dram_rsp_valid ? dram_rsp_addr : creq_addr;
wire [`LINE_ADDR_WIDTH-1:0] lookup_addr = dram_rsp_valid ? dram_rsp_addr_qual : creq_addr;
VX_miss_resrv #(
.BANK_ID (BANK_ID),
@ -429,7 +449,7 @@ module VX_bank #(
.enqueue (mshr_push),
.enqueue_addr (addr_st1),
.enqueue_data ({wsel_st1, byteen_st1, tag_st1, req_tid_st1, pmask_st1}),
.enqueue_is_mshr (is_mshr_st1),
.enqueue_is_mshr (mshr_restore),
.enqueue_as_ready (mshr_init_ready_state),
`UNUSED_PIN (enqueue_almfull),
`UNUSED_PIN (enqueue_full),
@ -562,11 +582,11 @@ module VX_bank #(
if (crsq_alm_full || dreq_alm_full || mshr_alm_full) begin
$display("%t: cache%0d:%0d pipeline-stall: cwbq=%b, dwbq=%b, mshr=%b", $time, CACHE_ID, BANK_ID, crsq_alm_full, dreq_alm_full, mshr_alm_full);
end
if (flush_enable) begin
$display("%t: cache%0d:%0d flush: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(flush_addr, BANK_ID));
end
if (drsq_pop) begin
if (dram_rsp_flush)
$display("%t: cache%0d:%0d flush: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_rsp_addr, BANK_ID));
else
$display("%t: cache%0d:%0d fill-rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_rsp_addr, BANK_ID), dram_rsp_data);
$display("%t: cache%0d:%0d fill-rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_rsp_addr_qual, BANK_ID), dram_rsp_data);
end
if (mshr_pop) begin
$display("%t: cache%0d:%0d mshr-rd-req: addr=%0h, tag=%0h, pmask=%0b, tid=%0d, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mshr_addr, BANK_ID), mshr_tag, mshr_pmask, mshr_tid, mshr_byteen, debug_wid_sel, debug_pc_sel);

36
hw/rtl/cache/VX_cache.v vendored
View file

@ -42,7 +42,10 @@ module VX_cache #(
parameter DRAM_TAG_WIDTH = (32 - $clog2(CACHE_LINE_SIZE)),
// bank offset from beginning of index range
parameter BANK_ADDR_OFFSET = 0
parameter BANK_ADDR_OFFSET = 0,
// in-order DRAN
parameter IN_ORDER_DRAM = 0
) (
`SCOPE_IO_VX_cache
@ -117,7 +120,7 @@ module VX_cache #(
wire [`CACHE_LINE_WIDTH-1:0] dram_rsp_data_qual;
wire [DRAM_TAG_WIDTH-1:0] dram_rsp_tag_qual;
wire [`LINE_ADDR_WIDTH-1:0] flush_addr;
wire [`LINE_SELECT_BITS-1:0] flush_addr;
wire flush_enable;
`ifdef PERF_ENABLE
@ -151,13 +154,13 @@ module VX_cache #(
`UNUSED_PIN (alm_full),
`UNUSED_PIN (alm_empty),
`UNUSED_PIN (size)
);
);
if (NUM_BANKS == 1) begin
`UNUSED_VAR (dram_rsp_tag_qual)
assign drsq_pop = !drsq_empty && per_bank_dram_rsp_ready && !flush_enable;
assign drsq_pop = !drsq_empty && per_bank_dram_rsp_ready;
end else begin
assign drsq_pop = !drsq_empty && per_bank_dram_rsp_ready[`DRAM_ADDR_BANK(dram_rsp_tag_qual)] && !flush_enable;
assign drsq_pop = !drsq_empty && per_bank_dram_rsp_ready[`DRAM_ADDR_BANK(dram_rsp_tag_qual)];
end
///////////////////////////////////////////////////////////////////////////
@ -171,8 +174,7 @@ module VX_cache #(
.clk (clk),
.reset (reset),
.flush (flush),
.addr (flush_addr),
.ready_out ((& per_bank_dram_rsp_ready)),
.addr_out (flush_addr),
.valid_out (flush_enable)
);
@ -240,7 +242,6 @@ module VX_cache #(
wire curr_bank_dram_rsp_valid;
wire [`LINE_ADDR_WIDTH-1:0] curr_bank_dram_rsp_addr;
wire [`CACHE_LINE_WIDTH-1:0] curr_bank_dram_rsp_data;
wire curr_bank_dram_rsp_flush;
wire curr_bank_dram_rsp_ready;
// Core Req
@ -276,14 +277,13 @@ module VX_cache #(
// DRAM response
if (NUM_BANKS == 1) begin
assign curr_bank_dram_rsp_valid = !drsq_empty || flush_enable;
assign curr_bank_dram_rsp_addr = flush_enable ? flush_addr : dram_rsp_tag_qual;
assign curr_bank_dram_rsp_valid = !drsq_empty;
assign curr_bank_dram_rsp_addr = dram_rsp_tag_qual;
end else begin
assign curr_bank_dram_rsp_valid = (!drsq_empty && (`DRAM_ADDR_BANK(dram_rsp_tag_qual) == i)) || flush_enable;
assign curr_bank_dram_rsp_addr = flush_enable ? flush_addr : `DRAM_TO_LINE_ADDR(dram_rsp_tag_qual);
assign curr_bank_dram_rsp_valid = !drsq_empty && (`DRAM_ADDR_BANK(dram_rsp_tag_qual) == i);
assign curr_bank_dram_rsp_addr = `DRAM_TO_LINE_ADDR(dram_rsp_tag_qual);
end
assign curr_bank_dram_rsp_data = dram_rsp_data_qual;
assign curr_bank_dram_rsp_flush = flush_enable;
assign per_bank_dram_rsp_ready[i] = curr_bank_dram_rsp_ready;
VX_bank #(
@ -303,7 +303,8 @@ module VX_cache #(
.WRITE_ENABLE (WRITE_ENABLE),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS),
.BANK_ADDR_OFFSET (BANK_ADDR_OFFSET)
.BANK_ADDR_OFFSET (BANK_ADDR_OFFSET),
.IN_ORDER_DRAM (IN_ORDER_DRAM)
) bank (
`SCOPE_BIND_VX_cache_bank(i)
@ -348,8 +349,11 @@ module VX_cache #(
.dram_rsp_valid (curr_bank_dram_rsp_valid),
.dram_rsp_addr (curr_bank_dram_rsp_addr),
.dram_rsp_data (curr_bank_dram_rsp_data),
.dram_rsp_flush (curr_bank_dram_rsp_flush),
.dram_rsp_ready (curr_bank_dram_rsp_ready)
.dram_rsp_ready (curr_bank_dram_rsp_ready),
// flush
.flush_enable (flush_enable),
.flush_addr (flush_addr)
);
end

9
hw/rtl/cache/VX_flush_ctrl.v vendored
View file

@ -2,7 +2,7 @@
module VX_flush_ctrl #(
// Size of cache in bytes
parameter CACHE_SIZE = 16384,
parameter CACHE_SIZE = 16384,
// Size of line inside a bank in bytes
parameter CACHE_LINE_SIZE = 1,
// Number of banks
@ -13,8 +13,7 @@ module VX_flush_ctrl #(
input wire clk,
input wire reset,
input wire flush,
output wire [`LINE_ADDR_WIDTH-1:0] addr,
input wire ready_out,
output wire [`LINE_SELECT_BITS-1:0] addr_out,
output wire valid_out
);
reg flush_enable;
@ -25,7 +24,7 @@ module VX_flush_ctrl #(
flush_enable <= 1;
flush_ctr <= 0;
end else begin
if (flush_enable && ready_out) begin
if (flush_enable) begin
if (flush_ctr == ((2 ** `LINE_SELECT_BITS)-1)) begin
flush_enable <= 0;
end
@ -34,7 +33,7 @@ module VX_flush_ctrl #(
end
end
assign addr = `LINE_ADDR_WIDTH'(flush_ctr);
assign addr_out = flush_ctr;
assign valid_out = flush_enable;
endmodule

2
hw/rtl/cache/VX_miss_resrv.v vendored
View file

@ -21,7 +21,7 @@ module VX_miss_resrv #(
// core request tag size
parameter CORE_TAG_WIDTH = 1,
// size of tag id in core request tag
parameter CORE_TAG_ID_BITS = 0
parameter CORE_TAG_ID_BITS = 0
) (
input wire clk,
input wire reset,

23
hw/rtl/cache/VX_tag_access.v vendored
View file

@ -29,9 +29,10 @@ module VX_tag_access #(
input wire[`LINE_ADDR_WIDTH-1:0] addr,
input wire fill,
input wire is_flush,
output wire missed
output wire tag_match
);
`UNUSED_VAR (reset)
`UNUSED_VAR (lookup)
wire read_valid;
wire [`TAG_SELECT_BITS-1:0] read_tag;
@ -54,25 +55,21 @@ module VX_tag_access #(
.dout({read_valid, read_tag})
);
wire tags_match = read_valid && (line_tag == read_tag);
assign missed = !tags_match;
wire do_lookup = lookup;
`UNUSED_VAR (do_lookup)
assign tag_match = read_valid && (line_tag == read_tag);
`ifdef DBG_PRINT_CACHE_TAG
always @(posedge clk) begin
if (fill) begin
if (is_flush)
if (is_flush) begin
$display("%t: cache%0d:%0d tag-flush: addr=%0h, blk_addr=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr);
else
end else begin
$display("%t: cache%0d:%0d tag-fill: addr=%0h, blk_addr=%0d, tag_id=%0h, old_tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr, line_tag, read_tag);
if (tags_match) begin
$display("%t: warning: redundant fill - addr=%0h", $time, `LINE_TO_BYTE_ADDR(addr, BANK_ID));
if (tag_match) begin
$display("%t: warning: redundant fill - addr=%0h", $time, `LINE_TO_BYTE_ADDR(addr, BANK_ID));
end
end
end else if (do_lookup) begin
if (tags_match) begin
end else if (lookup) begin
if (tag_match) begin
$display("%t: cache%0d:%0d tag-hit: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, line_tag);
end else begin
$display("%t: cache%0d:%0d tag-miss: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h, old_tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, line_tag, read_tag);