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

cache pipeline optimization

Browse Source
This commit is contained in:
Blaise Tine
2021-01-17 17:19:52 -08:00
parent ed216ab39d
commit a046bd7a73
11 changed files with 129 additions and 166 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
hw/rtl/VX_config.vh
View File

@@ -264,7 +264,7 @@
// Size of cache in bytes // Size of cache in bytes
`ifndef ICACHE_SIZE `ifndef ICACHE_SIZE
`define ICACHE_SIZE 8192 `define ICACHE_SIZE 16384
`endif `endif
// Core Request Queue Size // Core Request Queue Size
@@ -296,7 +296,7 @@
// Size of cache in bytes // Size of cache in bytes
`ifndef DCACHE_SIZE `ifndef DCACHE_SIZE
`define DCACHE_SIZE 8192 `define DCACHE_SIZE 16384
`endif `endif
// Number of banks // Number of banks
@@ -360,7 +360,7 @@
// Size of cache in bytes // Size of cache in bytes
`ifndef L2CACHE_SIZE `ifndef L2CACHE_SIZE
`define L2CACHE_SIZE 131072 `define L2CACHE_SIZE 65536
`endif `endif
// Number of banks // Number of banks

1
hw/rtl/VX_ibuffer.v
View File

@@ -43,7 +43,6 @@ module VX_ibuffer #(
VX_fifo_queue #( VX_fifo_queue #(
.DATAW (DATAW), .DATAW (DATAW),
.SIZE (SIZE), .SIZE (SIZE),
.BUFFERED (1),
.FASTRAM (1) .FASTRAM (1)
) queue ( ) queue (
.clk (clk), .clk (clk),

2
hw/rtl/Vortex.v
View File

@@ -165,7 +165,7 @@ module Vortex (
VX_cache #( VX_cache #(
.CACHE_ID (`L3CACHE_ID), .CACHE_ID (`L3CACHE_ID),
.CACHE_SIZE (`L3CACHE_SIZE), .CACHE_SIZE (`L3CACHE_SIZE),
.CACHE_LINE_SIZE (`L3CACHE_LINE_SIZE), .CACHE_LINE_SIZE (`L3CACHE_LINE_SIZE),
.NUM_BANKS (`L3NUM_BANKS), .NUM_BANKS (`L3NUM_BANKS),
.WORD_SIZE (`L3WORD_SIZE), .WORD_SIZE (`L3WORD_SIZE),
.NUM_REQS (`NUM_CLUSTERS), .NUM_REQS (`NUM_CLUSTERS),

161
hw/rtl/cache/VX_bank.v vendored
View File

@@ -117,7 +117,6 @@ module VX_bank #(
.data_out_next ({drsq_addr_next, drsq_filldata_next, drsq_flush_next}), .data_out_next ({drsq_addr_next, drsq_filldata_next, drsq_flush_next}),
.empty_next (drsq_empty_next), .empty_next (drsq_empty_next),
.full (drsq_full), .full (drsq_full),
`UNUSED_PIN (almost_full),
`UNUSED_PIN (size) `UNUSED_PIN (size)
); );
@@ -164,12 +163,14 @@ module VX_bank #(
.data_out_next({creq_tag_next, creq_tid_next, creq_rw_next, creq_byteen_next, creq_addr_next_unqual, creq_writeword_next}), .data_out_next({creq_tag_next, creq_tid_next, creq_rw_next, creq_byteen_next, creq_addr_next_unqual, creq_writeword_next}),
`UNUSED_PIN (empty_next), `UNUSED_PIN (empty_next),
.full (creq_full), .full (creq_full),
`UNUSED_PIN (almost_full),
`UNUSED_PIN (size) `UNUSED_PIN (size)
); );
wire mshr_pop; wire crsq_alm_full;
wire mshr_almost_full; wire dreq_alm_full;
wire mshr_alm_full;
wire mshr_pop;
wire mshr_pending_unqual_st0; wire mshr_pending_unqual_st0;
wire mshr_valid; wire mshr_valid;
wire mshr_valid_next; wire mshr_valid_next;
@@ -181,14 +182,11 @@ module VX_bank #(
wire mshr_rw_next; wire mshr_rw_next;
wire [WORD_SIZE-1:0] mshr_byteen_next; wire [WORD_SIZE-1:0] mshr_byteen_next;
wire dreq_almost_full;
wire [`LINE_ADDR_WIDTH-1:0] addr_st0, addr_st1; wire [`LINE_ADDR_WIDTH-1:0] addr_st0, addr_st1;
wire [`UP(`WORD_SELECT_BITS)-1:0] wsel_st0, wsel_st1; wire [`UP(`WORD_SELECT_BITS)-1:0] wsel_st0, wsel_st1;
wire mem_rw_st0, mem_rw_st1; wire mem_rw_st0, mem_rw_st1;
wire [WORD_SIZE-1:0] byteen_st0, byteen_st1; wire [WORD_SIZE-1:0] byteen_st0, byteen_st1;
wire [`WORD_WIDTH-1:0] writeword_st0, writeword_st1; wire [`CACHE_LINE_WIDTH-1:0] data_st0, data_st1;
wire [`CACHE_LINE_WIDTH-1:0] filldata_st0, filldata_st1;
wire [`REQS_BITS-1:0] req_tid_st0, req_tid_st1; wire [`REQS_BITS-1:0] req_tid_st0, req_tid_st1;
wire [`REQ_TAG_WIDTH-1:0] tag_st0, tag_st1; wire [`REQ_TAG_WIDTH-1:0] tag_st0, tag_st1;
wire valid_st0, valid_st1; wire valid_st0, valid_st1;
@@ -199,35 +197,35 @@ module VX_bank #(
wire force_miss_st0, force_miss_st1; wire force_miss_st0, force_miss_st1;
wire do_writeback_st0, do_writeback_st1; wire do_writeback_st0, do_writeback_st1;
wire writeen_unqual_st0, writeen_unqual_st1; wire writeen_unqual_st0, writeen_unqual_st1;
wire mshr_push_unqual_st0, mshr_push_unqual_st1;
wire dreq_push_unqual_st0, dreq_push_unqual_st1; wire dreq_push_unqual_st0, dreq_push_unqual_st1;
wire writeen_st1; wire writeen_st1;
wire core_req_hit_st1; wire core_req_hit_st1;
wire is_flush_st0; wire is_flush_st0;
wire mshr_push_stall;
wire crsq_push_stall;
wire dreq_push_stall;
wire pipeline_stall;
wire is_mshr_miss_st1 = valid_st1 && is_mshr_st1 && (miss_st1 || force_miss_st1); wire is_mshr_miss_st1 = valid_st1 && is_mshr_st1 && (miss_st1 || force_miss_st1);
// determine which queue to pop next in piority order // determine which queue to pop next in piority order
wire mshr_pop_unqual = mshr_valid; wire mshr_pop_unqual = mshr_valid;
wire drsq_pop_unqual = !mshr_pop_unqual && !drsq_empty; wire drsq_pop_unqual = !mshr_pop_unqual && !drsq_empty;
wire creq_pop_unqual = !mshr_pop_unqual && !drsq_pop_unqual && !creq_empty && !mshr_almost_full && !dreq_almost_full; wire creq_pop_unqual = !mshr_pop_unqual && !drsq_pop_unqual && !creq_empty;
assign mshr_pop = mshr_pop_unqual && !pipeline_stall assign mshr_pop = mshr_pop_unqual
&& !is_mshr_miss_st1; // do not schedule another mshr request when the previous one missed && !crsq_alm_full // ensure core response ready
assign drsq_pop = drsq_pop_unqual && !pipeline_stall; && !is_mshr_miss_st1; // do not schedule another mshr request when the previous one missed
assign creq_pop = creq_pop_unqual && !pipeline_stall;
assign drsq_pop = drsq_pop_unqual;
assign creq_pop = creq_pop_unqual
&& !crsq_alm_full // ensure core response ready
&& !dreq_alm_full // ensure dram request ready
&& !mshr_alm_full; // ensure mshr enqueue ready
assign valid_st0 = mshr_pop || drsq_pop || creq_pop; assign valid_st0 = mshr_pop || drsq_pop || creq_pop;
assign is_mshr_st0 = mshr_pop_unqual; assign is_mshr_st0 = mshr_pop_unqual;
assign is_fill_st0 = drsq_pop_unqual; assign is_fill_st0 = drsq_pop_unqual;
VX_pipe_register #( VX_pipe_register #(
.DATAW (`LINE_ADDR_WIDTH + `UP(`WORD_SELECT_BITS) + 1 + WORD_SIZE + `WORD_WIDTH + `REQS_BITS + `REQ_TAG_WIDTH + `CACHE_LINE_WIDTH + 1), .DATAW (`LINE_ADDR_WIDTH + `UP(`WORD_SELECT_BITS) + 1 + WORD_SIZE + `CACHE_LINE_WIDTH + `REQS_BITS + `REQ_TAG_WIDTH + 1),
.RESETW (0) .RESETW (0)
) pipe_reg0 ( ) pipe_reg0 (
.clk (clk), .clk (clk),
@@ -238,13 +236,12 @@ module VX_bank #(
mshr_valid_next ? mshr_wsel_next : creq_wsel_next, mshr_valid_next ? mshr_wsel_next : creq_wsel_next,
mshr_valid_next ? mshr_rw_next : creq_rw_next, mshr_valid_next ? mshr_rw_next : creq_rw_next,
mshr_valid_next ? mshr_byteen_next : creq_byteen_next, mshr_valid_next ? mshr_byteen_next : creq_byteen_next,
mshr_valid_next ? mshr_writeword_next : creq_writeword_next, mshr_valid_next ? {`WORDS_PER_LINE{mshr_writeword_next}} : (!drsq_empty_next ? drsq_filldata_next : {`WORDS_PER_LINE{creq_writeword_next}}),
mshr_valid_next ? mshr_tid_next : creq_tid_next, mshr_valid_next ? mshr_tid_next : creq_tid_next,
mshr_valid_next ? `REQ_TAG_WIDTH'(mshr_tag_next) : `REQ_TAG_WIDTH'(creq_tag_next), mshr_valid_next ? `REQ_TAG_WIDTH'(mshr_tag_next) : `REQ_TAG_WIDTH'(creq_tag_next),
drsq_filldata_next,
drsq_flush_next drsq_flush_next
}), }),
.data_out ({addr_st0, wsel_st0, mem_rw_st0, byteen_st0, writeword_st0, req_tid_st0, tag_st0, filldata_st0, is_flush_st0}) .data_out ({addr_st0, wsel_st0, mem_rw_st0, byteen_st0, data_st0, req_tid_st0, tag_st0, is_flush_st0})
); );
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
@@ -300,17 +297,15 @@ module VX_bank #(
assign dreq_push_unqual_st0 = send_fill_req_st0 || do_writeback_st0; assign dreq_push_unqual_st0 = send_fill_req_st0 || do_writeback_st0;
assign mshr_push_unqual_st0 = !is_fill_st0 && !mem_rw_st0;
VX_pipe_register #( VX_pipe_register #(
.DATAW (1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_BITS) + `WORD_WIDTH + `CACHE_LINE_WIDTH + 1 + WORD_SIZE + `REQS_BITS + `REQ_TAG_WIDTH), .DATAW (1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + `LINE_ADDR_WIDTH + `UP(`WORD_SELECT_BITS) + `CACHE_LINE_WIDTH + 1 + WORD_SIZE + `REQS_BITS + `REQ_TAG_WIDTH),
.RESETW (1) .RESETW (1)
) pipe_reg1 ( ) pipe_reg1 (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.enable (!pipeline_stall), .enable (1'b1),
.data_in ({valid_st0, is_mshr_st0, is_fill_st0, writeen_unqual_st0, mshr_push_unqual_st0, dreq_push_unqual_st0, do_writeback_st0, miss_st0, force_miss_st0, addr_st0, wsel_st0, writeword_st0, filldata_st0, mem_rw_st0, byteen_st0, req_tid_st0, tag_st0}), .data_in ({valid_st0, is_mshr_st0, is_fill_st0, writeen_unqual_st0, dreq_push_unqual_st0, do_writeback_st0, miss_st0, force_miss_st0, addr_st0, wsel_st0, data_st0, mem_rw_st0, byteen_st0, req_tid_st0, tag_st0}),
.data_out ({valid_st1, is_mshr_st1, is_fill_st1, writeen_unqual_st1, mshr_push_unqual_st1, dreq_push_unqual_st1, do_writeback_st1, miss_st1, force_miss_st1, addr_st1, wsel_st1, writeword_st1, filldata_st1, mem_rw_st1, byteen_st1, req_tid_st1, tag_st1}) .data_out ({valid_st1, is_mshr_st1, is_fill_st1, writeen_unqual_st1, dreq_push_unqual_st1, do_writeback_st1, miss_st1, force_miss_st1, addr_st1, wsel_st1, data_st1, mem_rw_st1, byteen_st1, req_tid_st1, tag_st1})
); );
assign 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;
@@ -319,7 +314,7 @@ module VX_bank #(
wire dreq_push_st1 = dreq_push_unqual_st1 && (do_writeback_st1 || !force_miss_st1); wire dreq_push_st1 = dreq_push_unqual_st1 && (do_writeback_st1 || !force_miss_st1);
wire mshr_push_st1 = mshr_push_unqual_st1 && (miss_st1 || force_miss_st1); wire mshr_push_st1 = !is_fill_st1 && !mem_rw_st1 && (miss_st1 || force_miss_st1);
wire crsq_push_st1 = core_req_hit_st1 && !mem_rw_st1; wire crsq_push_st1 = core_req_hit_st1 && !mem_rw_st1;
@@ -333,27 +328,26 @@ module VX_bank #(
.WORD_SIZE (WORD_SIZE), .WORD_SIZE (WORD_SIZE),
.WRITE_ENABLE (WRITE_ENABLE) .WRITE_ENABLE (WRITE_ENABLE)
) data_access ( ) data_access (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
.debug_pc (debug_pc_st1), .debug_pc (debug_pc_st1),
.debug_wid (debug_wid_st1), .debug_wid (debug_wid_st1),
`endif `endif
.addr (addr_st1), .addr (addr_st1),
// reading // reading
.readen (valid_st1 && !mem_rw_st1 && !is_fill_st1 && ~pipeline_stall), .readen (valid_st1 && !mem_rw_st1 && !is_fill_st1),
.readdata (readdata_st1), .rddata (readdata_st1),
// writing // writing
.writeen (valid_st1 && writeen_st1 && ~pipeline_stall), .writeen (valid_st1 && writeen_st1),
.is_fill (is_fill_st1), .is_fill (is_fill_st1),
.wsel (wsel_st1), .wsel (wsel_st1),
.byteen (byteen_st1), .byteen (byteen_st1),
.writeword (writeword_st1), .wrdata (data_st1)
.filldata (filldata_st1)
); );
`ifdef DBG_CACHE_REQ_INFO `ifdef DBG_CACHE_REQ_INFO
@@ -364,20 +358,14 @@ module VX_bank #(
end end
`endif `endif
wire mshr_push_unqual = valid_st1 && mshr_push_st1;
assign mshr_push_stall = 0;
wire mshr_push = mshr_push_unqual
&& !crsq_push_stall
&& !dreq_push_stall;
wire incoming_fill_st1 = valid_st0 && is_fill_st0 && (addr_st1 == addr_st0); wire incoming_fill_st1 = valid_st0 && is_fill_st0 && (addr_st1 == addr_st0);
wire mshr_dequeue_st1 = valid_st1 && is_mshr_st1 && !mshr_push_unqual && !pipeline_stall; wire mshr_push = valid_st1 && mshr_push_st1;
wire mshr_dequeue = valid_st1 && is_mshr_st1 && !mshr_push_st1;
// push a missed request as 'ready' if it was a forced miss that actually had a hit // 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 // or the fill request for this block is comming
wire mshr_init_ready_state_st1 = !miss_st1 || incoming_fill_st1; wire mshr_init_ready_state = !miss_st1 || incoming_fill_st1;
VX_miss_resrv #( VX_miss_resrv #(
.BANK_ID (BANK_ID), .BANK_ID (BANK_ID),
@@ -404,10 +392,10 @@ module VX_bank #(
// enqueue // enqueue
.enqueue (mshr_push), .enqueue (mshr_push),
.enqueue_addr (addr_st1), .enqueue_addr (addr_st1),
.enqueue_data ({writeword_st1, req_tid_st1, tag_st1, mem_rw_st1, byteen_st1, wsel_st1}), .enqueue_data ({data_st1[`WORD_WIDTH-1:0], req_tid_st1, tag_st1, mem_rw_st1, byteen_st1, wsel_st1}),
.enqueue_is_mshr (is_mshr_st1), .enqueue_is_mshr (is_mshr_st1),
.enqueue_as_ready (mshr_init_ready_state_st1), .enqueue_as_ready (mshr_init_ready_state),
.enqueue_almfull (mshr_almost_full), .enqueue_almfull (mshr_alm_full),
// lookup // lookup
.lookup_ready (drsq_pop && !is_flush_st0), .lookup_ready (drsq_pop && !is_flush_st0),
@@ -424,22 +412,15 @@ module VX_bank #(
.schedule_data_next ({mshr_writeword_next, mshr_tid_next, mshr_tag_next, mshr_rw_next, mshr_byteen_next, mshr_wsel_next}), .schedule_data_next ({mshr_writeword_next, mshr_tid_next, mshr_tag_next, mshr_rw_next, mshr_byteen_next, mshr_wsel_next}),
// dequeue // dequeue
.dequeue (mshr_dequeue_st1) .dequeue (mshr_dequeue)
); );
// Enqueue core response // Enqueue core response
wire crsq_empty, crsq_full; wire crsq_empty;
wire crsq_push_unqual = valid_st1 && crsq_push_st1; wire crsq_push = valid_st1 && crsq_push_st1;
assign crsq_push_stall = crsq_push_unqual && crsq_full; wire crsq_pop = core_rsp_valid && core_rsp_ready;
wire crsq_push = crsq_push_unqual
&& !crsq_full
&& !mshr_push_stall
&& !dreq_push_stall;
wire crsq_pop = core_rsp_valid && core_rsp_ready;
wire [`REQS_BITS-1:0] crsq_tid_st1 = req_tid_st1; wire [`REQS_BITS-1:0] crsq_tid_st1 = req_tid_st1;
wire [CORE_TAG_WIDTH-1:0] crsq_tag_st1 = CORE_TAG_WIDTH'(tag_st1); wire [CORE_TAG_WIDTH-1:0] crsq_tag_st1 = CORE_TAG_WIDTH'(tag_st1);
@@ -459,7 +440,8 @@ module VX_bank #(
VX_fifo_queue #( VX_fifo_queue #(
.DATAW (`REQS_BITS + CORE_TAG_WIDTH + `WORD_WIDTH), .DATAW (`REQS_BITS + CORE_TAG_WIDTH + `WORD_WIDTH),
.SIZE (CRSQ_SIZE), .SIZE (CRSQ_SIZE),
.BUFFERED (NUM_BANKS == 1), .ALM_FULL (CRSQ_SIZE-1),
.BUFFERED (1),
.FASTRAM (1) .FASTRAM (1)
) core_rsp_queue ( ) core_rsp_queue (
.clk (clk), .clk (clk),
@@ -469,9 +451,9 @@ module VX_bank #(
.data_in ({crsq_tid_st1, crsq_tag_st1, crsq_data_st1}), .data_in ({crsq_tid_st1, crsq_tag_st1, crsq_data_st1}),
.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), .alm_full(crsq_alm_full),
`UNUSED_PIN (alm_empty), `UNUSED_PIN (full),
`UNUSED_PIN (alm_full), `UNUSED_PIN (alm_empty),
`UNUSED_PIN (size) `UNUSED_PIN (size)
); );
@@ -480,14 +462,9 @@ module VX_bank #(
// Enqueue DRAM request // Enqueue DRAM request
wire dreq_empty; wire dreq_empty;
wire dreq_push_unqual = valid_st1 && dreq_push_st1; wire dreq_push = valid_st1 && dreq_push_st1
assign dreq_push_stall = 0; && (do_writeback_st1 || !incoming_fill_st1);
wire dreq_push = dreq_push_unqual
&& (do_writeback_st1 || !incoming_fill_st1)
&& !mshr_push_stall
&& !crsq_push_stall;
wire dreq_pop = dram_req_valid && dram_req_ready; wire dreq_pop = dram_req_valid && dram_req_ready;
@@ -501,12 +478,11 @@ module VX_bank #(
if (`WORD_SELECT_BITS != 0) begin if (`WORD_SELECT_BITS != 0) begin
for (genvar i = 0; i < `WORDS_PER_LINE; i++) begin for (genvar i = 0; i < `WORDS_PER_LINE; i++) begin
assign dreq_byteen_unqual[i * WORD_SIZE +: WORD_SIZE] = (wsel_st1 == `WORD_SELECT_BITS'(i)) ? byteen_st1 : {WORD_SIZE{1'b0}}; assign dreq_byteen_unqual[i * WORD_SIZE +: WORD_SIZE] = (wsel_st1 == `WORD_SELECT_BITS'(i)) ? byteen_st1 : {WORD_SIZE{1'b0}};
assign dreq_data[i * `WORD_WIDTH +: `WORD_WIDTH] = writeword_st1;
end end
end else begin end else begin
assign dreq_byteen_unqual = byteen_st1; assign dreq_byteen_unqual = byteen_st1;
assign dreq_data = writeword_st1;
end end
assign dreq_data = {`WORDS_PER_LINE{data_st1[`WORD_WIDTH-1:0]}};
assign dreq_byteen = writeback ? dreq_byteen_unqual : {CACHE_LINE_SIZE{1'b1}}; assign dreq_byteen = writeback ? dreq_byteen_unqual : {CACHE_LINE_SIZE{1'b1}};
@@ -524,16 +500,13 @@ module VX_bank #(
.data_in ({writeback, dreq_byteen, dreq_addr, dreq_data}), .data_in ({writeback, dreq_byteen, dreq_addr, dreq_data}),
.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),
.alm_full(dreq_almost_full), .alm_full(dreq_alm_full),
`UNUSED_PIN (full), `UNUSED_PIN (full),
`UNUSED_PIN (alm_empty), `UNUSED_PIN (alm_empty),
`UNUSED_PIN (size) `UNUSED_PIN (size)
); );
assign dram_req_valid = !dreq_empty; assign dram_req_valid = !dreq_empty;
// bank pipeline stall
assign pipeline_stall = crsq_push_stall;
`SCOPE_ASSIGN (valid_st0, valid_st0); `SCOPE_ASSIGN (valid_st0, valid_st0);
`SCOPE_ASSIGN (valid_st1, valid_st1); `SCOPE_ASSIGN (valid_st1, valid_st1);
@@ -542,15 +515,17 @@ module VX_bank #(
`SCOPE_ASSIGN (miss_st0, miss_st0); `SCOPE_ASSIGN (miss_st0, miss_st0);
`SCOPE_ASSIGN (force_miss_st0, force_miss_st0); `SCOPE_ASSIGN (force_miss_st0, force_miss_st0);
`SCOPE_ASSIGN (mshr_push, mshr_push); `SCOPE_ASSIGN (mshr_push, mshr_push);
`SCOPE_ASSIGN (pipeline_stall, pipeline_stall); `SCOPE_ASSIGN (crsq_alm_full, crsq_alm_full);
`SCOPE_ASSIGN (dreq_alm_full, dreq_alm_full);
`SCOPE_ASSIGN (mshr_alm_full, mshr_alm_full);
`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));
`ifdef PERF_ENABLE `ifdef PERF_ENABLE
assign perf_read_misses = valid_st1 && !pipeline_stall && !is_fill_st1 && !is_mshr_st1 && miss_st1 && !mem_rw_st1; assign perf_read_misses = valid_st1 && !is_fill_st1 && !is_mshr_st1 && miss_st1 && !mem_rw_st1;
assign perf_write_misses = valid_st1 && !pipeline_stall && !is_fill_st1 && !is_mshr_st1 && miss_st1 && mem_rw_st1; assign perf_write_misses = valid_st1 && !is_fill_st1 && !is_mshr_st1 && miss_st1 && mem_rw_st1;
assign perf_pipe_stalls = pipeline_stall || mshr_almost_full || dreq_almost_full; assign perf_pipe_stalls = crsq_alm_full || dreq_alm_full || mshr_alm_full;
assign perf_mshr_stalls = mshr_almost_full; assign perf_mshr_stalls = mshr_alm_full;
`endif `endif
`ifdef DBG_PRINT_CACHE_BANK `ifdef DBG_PRINT_CACHE_BANK
@@ -559,8 +534,8 @@ module VX_bank #(
$display("%t: miss with incoming fill - addr=%0h", $time, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID)); $display("%t: miss with incoming fill - addr=%0h", $time, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID));
assert(!is_mshr_st1); assert(!is_mshr_st1);
end end
if (crsq_push_stall || mshr_almost_full || dreq_almost_full) begin if (crsq_alm_full || dreq_alm_full || mshr_alm_full) begin
$display("%t: cache%0d:%0d pipeline-stall: mshr=%b, cwbq=%b, dwbq=%b", $time, CACHE_ID, BANK_ID, mshr_almost_full, crsq_push_stall, dreq_almost_full); $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 end
if (drsq_pop) begin if (drsq_pop) begin
if (is_flush_st0) if (is_flush_st0)

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

@@ -4,7 +4,7 @@ module VX_cache #(
parameter CACHE_ID = 0, parameter CACHE_ID = 0,
// Size of cache in bytes // Size of cache in bytes
parameter CACHE_SIZE = 1048576, parameter CACHE_SIZE = 16384,
// Size of line inside a bank in bytes // Size of line inside a bank in bytes
parameter CACHE_LINE_SIZE = 64, parameter CACHE_LINE_SIZE = 64,
// Number of banks // Number of banks

17
hw/rtl/cache/VX_data_access.v vendored
View File

@@ -34,20 +34,18 @@ module VX_data_access #(
// reading // reading
input wire readen, input wire readen,
output wire [`CACHE_LINE_WIDTH-1:0] readdata, output wire [`CACHE_LINE_WIDTH-1:0] rddata,
// writing // writing
input wire writeen, input wire writeen,
input wire [`UP(`WORD_SELECT_BITS)-1:0] wsel,
input wire [WORD_SIZE-1:0] byteen,
input wire is_fill, input wire is_fill,
input wire [`WORD_WIDTH-1:0] writeword, input wire [`UP(`WORD_SELECT_BITS)-1:0] wsel,
input wire [`CACHE_LINE_WIDTH-1:0] filldata input wire [WORD_SIZE-1:0] byteen,
input wire [`CACHE_LINE_WIDTH-1:0] wrdata
); );
`UNUSED_VAR (reset) `UNUSED_VAR (reset)
wire [CACHE_LINE_SIZE-1:0] byte_enable; wire [CACHE_LINE_SIZE-1:0] byte_enable;
wire [`CACHE_LINE_WIDTH-1:0] write_data;
wire [`LINE_SELECT_BITS-1:0] line_addr = addr[`LINE_SELECT_BITS-1:0]; wire [`LINE_SELECT_BITS-1:0] line_addr = addr[`LINE_SELECT_BITS-1:0];
@@ -62,8 +60,8 @@ module VX_data_access #(
.wren(writeen), .wren(writeen),
.byteen(byte_enable), .byteen(byte_enable),
.rden(1'b1), .rden(1'b1),
.din(write_data), .din(wrdata),
.dout(readdata) .dout(rddata)
); );
wire [`WORDS_PER_LINE-1:0][WORD_SIZE-1:0] byteen_qual; wire [`WORDS_PER_LINE-1:0][WORD_SIZE-1:0] byteen_qual;
@@ -78,7 +76,6 @@ module VX_data_access #(
end end
assign byte_enable = is_fill ? {CACHE_LINE_SIZE{1'b1}} : byteen_qual; assign byte_enable = is_fill ? {CACHE_LINE_SIZE{1'b1}} : byteen_qual;
assign write_data = is_fill ? filldata : {`WORDS_PER_LINE{writeword}};
`UNUSED_VAR (readen) `UNUSED_VAR (readen)

24
hw/rtl/cache/VX_fifo_queue_xt.v vendored
View File

@@ -3,7 +3,6 @@
module VX_fifo_queue_xt #( module VX_fifo_queue_xt #(
parameter DATAW = 1, parameter DATAW = 1,
parameter SIZE = 2, parameter SIZE = 2,
parameter ALM_FULL = (SIZE - 1),
parameter ADDRW = $clog2(SIZE), parameter ADDRW = $clog2(SIZE),
parameter SIZEW = $clog2(SIZE+1), parameter SIZEW = $clog2(SIZE+1),
parameter FASTRAM = 0 parameter FASTRAM = 0
@@ -18,22 +17,20 @@ module VX_fifo_queue_xt #(
output wire [DATAW-1:0] data_out_next, output wire [DATAW-1:0] data_out_next,
output wire empty_next, output wire empty_next,
output wire full, output wire full,
output wire almost_full,
output wire [SIZEW-1:0] size output wire [SIZEW-1:0] size
); );
wire [DATAW-1:0] dout; wire [DATAW-1:0] dout;
reg [DATAW-1:0] dout_r, dout_n_r; reg [DATAW-1:0] dout_r, dout_n_r;
reg [ADDRW-1:0] wr_ptr_r; reg [ADDRW-1:0] wr_ptr_r;
reg [ADDRW-1:0] rd_ptr_r, rd_ptr_n_r; reg [ADDRW-1:0] rd_ptr_r, rd_ptr_n_r;
reg full_r, almost_full_r; reg full_r;
reg empty_r, empty_n_r; reg empty_r, empty_n_r;
reg [ADDRW-1:0] used_r; reg [ADDRW-1:0] used_r;
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset) begin
full_r <= 0; full_r <= 0;
almost_full_r <= 0; used_r <= 0;
used_r <= 0;
end else begin end else begin
assert(!push || !full); assert(!push || !full);
assert(!pop || !empty_r); assert(!pop || !empty_r);
@@ -41,12 +38,8 @@ module VX_fifo_queue_xt #(
if (!pop) begin if (!pop) begin
if (used_r == ADDRW'(SIZE-1)) if (used_r == ADDRW'(SIZE-1))
full_r <= 1; full_r <= 1;
if (used_r == ADDRW'(ALM_FULL-1))
almost_full_r <= 1;
end end
end else if (pop) begin end else if (pop) begin
if (used_r == ADDRW'(ALM_FULL))
almost_full_r <= 0;
full_r <= 0; full_r <= 0;
end end
@@ -75,11 +68,11 @@ module VX_fifo_queue_xt #(
end end
VX_dp_ram #( VX_dp_ram #(
.DATAW(DATAW), .DATAW (DATAW),
.SIZE(SIZE), .SIZE (SIZE),
.BUFFERED(0), .BUFFERED (0),
.RWCHECK(1), .RWCHECK (1),
.FASTRAM(FASTRAM) .FASTRAM (FASTRAM)
) dp_ram ( ) dp_ram (
.clk(clk), .clk(clk),
.waddr(wr_ptr_r), .waddr(wr_ptr_r),
@@ -127,7 +120,6 @@ module VX_fifo_queue_xt #(
assign empty = empty_r; assign empty = empty_r;
assign empty_next = empty_n_r; assign empty_next = empty_n_r;
assign full = full_r; assign full = full_r;
assign almost_full = almost_full_r;
assign size = {full_r, used_r}; assign size = {full_r, used_r};
endmodule endmodule

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

@@ -186,9 +186,7 @@ module VX_miss_resrv #(
if (reset) begin if (reset) begin
schedule_valid_n_r = 0; schedule_valid_n_r = 0;
end else begin end else begin
if (restore) begin if (lookup_ready) begin
schedule_valid_n_r = enqueue_as_ready;
end else if (lookup_ready) begin
schedule_valid_n_r = schedule_valid_r || (schedule_addr_r == lookup_addr); schedule_valid_n_r = schedule_valid_r || (schedule_addr_r == lookup_addr);
end else if (schedule) begin end else if (schedule) begin
schedule_valid_n_r = ready_table[schedule_n_ptr]; schedule_valid_n_r = ready_table[schedule_n_ptr];

38
hw/rtl/cache/VX_shared_mem.v vendored
View File

@@ -166,18 +166,18 @@ module VX_shared_mem #(
for (genvar i = 0; i < NUM_BANKS; i++) begin for (genvar i = 0; i < NUM_BANKS; i++) begin
VX_sp_ram #( VX_sp_ram #(
.DATAW(`WORD_WIDTH), .DATAW (`WORD_WIDTH),
.SIZE(`LINES_PER_BANK), .SIZE (`LINES_PER_BANK),
.BYTEENW(WORD_SIZE), .BYTEENW (WORD_SIZE),
.RWCHECK(1) .RWCHECK (1)
) data ( ) data (
.clk(clk), .clk (clk),
.addr(per_bank_core_req_addr[i]), .addr (per_bank_core_req_addr[i]),
.wren(per_bank_core_req_valid[i] && per_bank_core_req_rw[i] && ~crsq_full), .wren (per_bank_core_req_valid[i] && per_bank_core_req_rw[i] && ~crsq_full),
.byteen(per_bank_core_req_byteen[i]), .byteen (per_bank_core_req_byteen[i]),
.rden(1'b1), .rden (1'b1),
.din(per_bank_core_req_data[i]), .din (per_bank_core_req_data[i]),
.dout(per_bank_core_rsp_data[i]) .dout (per_bank_core_rsp_data[i])
); );
end end
@@ -220,14 +220,14 @@ module VX_shared_mem #(
.BUFFERED (1), .BUFFERED (1),
.FASTRAM (1) .FASTRAM (1)
) core_rsp_queue ( ) core_rsp_queue (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
.push (crsq_push), .push (crsq_push),
.pop (crsq_pop), .pop (crsq_pop),
.data_in ({core_rsp_valid_unqual, core_rsp_data_unqual, core_rsp_tag_unqual}), .data_in ({core_rsp_valid_unqual, core_rsp_data_unqual, core_rsp_tag_unqual}),
.data_out({core_rsp_valid_tmask, core_rsp_data, core_rsp_tag}), .data_out ({core_rsp_valid_tmask, core_rsp_data, core_rsp_tag}),
.empty (crsq_empty), .empty (crsq_empty),
.full (crsq_full), .full (crsq_full),
`UNUSED_PIN (alm_empty), `UNUSED_PIN (alm_empty),
`UNUSED_PIN (alm_full), `UNUSED_PIN (alm_full),
`UNUSED_PIN (size) `UNUSED_PIN (size)

36
hw/rtl/libs/VX_fifo_queue.v
View File

@@ -1,14 +1,14 @@
`include "VX_platform.vh" `include "VX_platform.vh"
module VX_fifo_queue #( module VX_fifo_queue #(
parameter DATAW = 1, parameter DATAW = 1,
parameter SIZE = 2, parameter SIZE = 2,
parameter ALM_FULL = (SIZE - 1), parameter ALM_FULL = (SIZE - 1),
parameter ALM_EMPTY= 1, parameter ALM_EMPTY = 1,
parameter ADDRW = $clog2(SIZE), parameter ADDRW = $clog2(SIZE),
parameter SIZEW = $clog2(SIZE+1), parameter SIZEW = $clog2(SIZE+1),
parameter BUFFERED = 0, parameter BUFFERED = 0,
parameter FASTRAM = 1 parameter FASTRAM = 1
) ( ) (
input wire clk, input wire clk,
input wire reset, input wire reset,
@@ -111,11 +111,11 @@ module VX_fifo_queue #(
end end
VX_dp_ram #( VX_dp_ram #(
.DATAW(DATAW), .DATAW (DATAW),
.SIZE(SIZE), .SIZE (SIZE),
.BUFFERED(0), .BUFFERED (0),
.RWCHECK(1), .RWCHECK (1),
.FASTRAM(FASTRAM) .FASTRAM (FASTRAM)
) dp_ram ( ) dp_ram (
.clk(clk), .clk(clk),
.waddr(wr_ptr_r), .waddr(wr_ptr_r),
@@ -156,11 +156,11 @@ module VX_fifo_queue #(
end end
VX_dp_ram #( VX_dp_ram #(
.DATAW(DATAW), .DATAW (DATAW),
.SIZE(SIZE), .SIZE (SIZE),
.BUFFERED(0), .BUFFERED (0),
.RWCHECK(1), .RWCHECK (1),
.FASTRAM(FASTRAM) .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

@@ -212,7 +212,9 @@
"miss_st0": 1, "miss_st0": 1,
"force_miss_st0": 1, "force_miss_st0": 1,
"mshr_push": 1, "mshr_push": 1,
"?pipeline_stall": 1 "?crsq_alm_full": 1,
"?dreq_alm_full": 1,
"?mshr_alm_full": 1
} }
} }
} }