rtl refactoring

This commit is contained in:
Blaise Tine
2020-05-04 20:12:05 -04:00
parent 69f607b73e
commit f142afac80
39 changed files with 31067 additions and 31607 deletions

117
hw/rtl/cache/VX_bank.v vendored
View File

@@ -12,9 +12,7 @@ module VX_bank #(
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
@@ -38,11 +36,23 @@ module VX_bank #(
// Fill Forward SNP Queue
parameter FFSQ_SIZE = 8,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1
// Enable cache writeable
parameter WRITE_ENABLE = 1,
// Enable dram update
parameter DRAM_ENABLE = 1,
// Enable snoop forwarding
parameter SNOOP_FORWARDING_ENABLE = 0,
// core request tag size
parameter CORE_TAG_WIDTH = 1,
// size of tag id in core request tag
parameter CORE_TAG_ID_BITS = 0
) (
input wire clk,
input wire reset,
@@ -50,11 +60,11 @@ module VX_bank #(
// Core Request
input wire core_req_ready,
input wire [NUM_REQUESTS-1:0] core_req_valids,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_read,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_write,
input wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] core_req_read,
input wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] core_req_write,
input wire [NUM_REQUESTS-1:0][31:0] core_req_addr,
input wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_req_data,
input wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_req_tag,
input wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] core_req_tag,
output wire core_req_full,
// Core Response
@@ -98,7 +108,7 @@ module VX_bank #(
if (reset) begin
snoop_state <= 0;
end else begin
snoop_state <= (snoop_state | snp_req_valid) && ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID));
snoop_state <= (snoop_state | snp_req_valid) && SNOOP_FORWARDING_ENABLE;
end
end
@@ -156,12 +166,12 @@ module VX_bank #(
`IGNORE_WARNINGS_END
wire [`WORD_WIDTH-1:0] reqq_req_writeword_st0;
wire [CORE_TAG_WIDTH-1:0] reqq_req_tag_st0;
wire [`WORD_SEL_BITS-1:0] reqq_req_mem_read_st0;
wire [`WORD_SEL_BITS-1:0] reqq_req_mem_write_st0;
wire [`BYTE_EN_BITS-1:0] reqq_req_mem_read_st0;
wire [`BYTE_EN_BITS-1:0] reqq_req_mem_write_st0;
assign reqq_push = core_req_ready && (|core_req_valids);
VX_cache_req_queue #(
VX_cache_req_queue #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
@@ -177,7 +187,8 @@ module VX_bank #(
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS)
) req_queue (
.clk (clk),
.reset (reset),
@@ -212,17 +223,17 @@ module VX_bank #(
wire [`BASE_ADDR_BITS-1:0] mrvq_wsel_st0;
wire [`WORD_WIDTH-1:0] mrvq_writeword_st0;
wire [CORE_TAG_WIDTH-1:0] mrvq_tag_st0;
wire [`WORD_SEL_BITS-1:0] mrvq_mem_read_st0;
wire [`WORD_SEL_BITS-1:0] mrvq_mem_write_st0;
wire [`BYTE_EN_BITS-1:0] mrvq_mem_read_st0;
wire [`BYTE_EN_BITS-1:0] mrvq_mem_write_st0;
wire miss_add;
wire[`LINE_ADDR_WIDTH-1:0] miss_add_addr;
wire[`BASE_ADDR_BITS-1:0] miss_add_wsel;
wire[`WORD_WIDTH-1:0] miss_add_data;
wire[`LOG2UP(NUM_REQUESTS)-1:0] miss_add_tid;
wire[CORE_TAG_WIDTH-1:0] miss_add_tag;
wire[`WORD_SEL_BITS-1:0] miss_add_mem_read;
wire[`WORD_SEL_BITS-1:0] miss_add_mem_write;
wire[`BYTE_EN_BITS-1:0] miss_add_mem_read;
wire[`BYTE_EN_BITS-1:0] miss_add_mem_write;
wire[`LINE_ADDR_WIDTH-1:0] addr_st2;
wire is_fill_st2;
@@ -294,8 +305,8 @@ module VX_bank #(
0;
assign qual_going_to_write_st0 = dfpq_pop ? 1 :
(mrvq_pop && (mrvq_mem_write_st0 != `WORD_SEL_NO)) ? 1 :
(reqq_pop && (reqq_req_mem_write_st0 != `WORD_SEL_NO)) ? 1 :
(mrvq_pop && (mrvq_mem_write_st0 != `BYTE_EN_NO)) ? 1 :
(reqq_pop && (reqq_req_mem_write_st0 != `BYTE_EN_NO)) ? 1 :
(snrq_pop) ? 1 :
0;
@@ -339,8 +350,8 @@ module VX_bank #(
wire [CORE_TAG_WIDTH-1:0] tag_st1e;
wire [`LOG2UP(NUM_REQUESTS)-1:0] tid_st1e;
`DEBUG_END
wire [`WORD_SEL_BITS-1:0] mem_read_st1e;
wire [`WORD_SEL_BITS-1:0] mem_write_st1e;
wire [`BYTE_EN_BITS-1:0] mem_read_st1e;
wire [`BYTE_EN_BITS-1:0] mem_write_st1e;
wire fill_saw_dirty_st1e;
wire is_snp_st1e;
@@ -354,7 +365,6 @@ module VX_bank #(
.WORD_SIZE (WORD_SIZE),
.NUM_REQUESTS (NUM_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
@@ -363,7 +373,9 @@ module VX_bank #(
.DWBQ_SIZE (DWBQ_SIZE),
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE)
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.DRAM_ENABLE (DRAM_ENABLE),
.WRITE_ENABLE (WRITE_ENABLE)
) tag_data_access (
.clk (clk),
.reset (reset),
@@ -429,7 +441,17 @@ module VX_bank #(
wire invalidate_fill;
// Enqueue to miss reserv if it's a valid miss
assign miss_add = valid_st2 && !is_snp_st2 && miss_st2 && !mrvq_full && !(should_flush && dwbq_push) && !((is_snp_st2 && valid_st2 && ffsq_full) ||((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
assign miss_add = valid_st2
&& !is_snp_st2
&& miss_st2
&& !mrvq_full
&& !(should_flush && dwbq_push)
&& !((is_snp_st2 && valid_st2 && ffsq_full)
|| ((valid_st2 && !miss_st2) && cwbq_full)
|| (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full)
|| (valid_st2 && miss_st2 && mrvq_full)
|| (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
assign miss_add_addr = addr_st2;
assign miss_add_wsel = wsel_st2;
assign miss_add_data = writeword_st2;
@@ -484,7 +506,14 @@ module VX_bank #(
);
// Enqueue to CWB Queue
wire cwbq_push = (valid_st2 && !miss_st2) && !cwbq_full && !((FUNC_ID == `L2FUNC_ID) && (miss_add_mem_write == `WORD_SEL_NO)) && !((is_snp_st2 && valid_st2 && ffsq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
wire cwbq_push = (valid_st2 && !miss_st2)
&& !cwbq_full
&& !(SNOOP_FORWARDING_ENABLE && (miss_add_mem_write == `BYTE_EN_NO))
&& !((is_snp_st2 && valid_st2 && ffsq_full)
|| (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full)
|| (valid_st2 && miss_st2 && mrvq_full)
|| (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
wire [`WORD_WIDTH-1:0] cwbq_data = readword_st2;
wire [`LOG2UP(NUM_REQUESTS)-1:0] cwbq_tid = miss_add_tid;
wire [CORE_TAG_WIDTH-1:0] cwbq_tag = miss_add_tag;
@@ -507,18 +536,27 @@ module VX_bank #(
.full (cwbq_full)
);
assign should_flush = snoop_state && valid_st2 && (miss_add_mem_write != `WORD_SEL_NO) && !is_snp_st2 && !is_fill_st2;
assign should_flush = snoop_state
&& valid_st2
&& (miss_add_mem_write != `BYTE_EN_NO)
&& !is_snp_st2 && !is_fill_st2;
// Enqueue to DWB Queue
assign dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2 || should_flush) && !dwbq_full && !((is_snp_st2 && valid_st2 && ffsq_full) ||((valid_st2 && !miss_st2) && cwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
assign dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2 || should_flush)
&& !dwbq_full
&& !((is_snp_st2 && valid_st2 && ffsq_full)
|| ((valid_st2 && !miss_st2) && cwbq_full)
|| (valid_st2 && miss_st2 && mrvq_full)
|| (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
wire[`LINE_ADDR_WIDTH-1:0] dwbq_req_addr;
wire dwbq_empty;
wire[`BANK_LINE_WORDS-1:0][`WORD_WIDTH-1:0] dwbq_req_data;
if ((FUNC_ID == `L2FUNC_ID) || (FUNC_ID == `L3FUNC_ID)) begin
if (SNOOP_FORWARDING_ENABLE) begin
assign dwbq_req_data = (should_flush && dwbq_push) ? writeword_st2 : readdata_st2;
assign dwbq_req_addr = (should_flush && dwbq_push) ? (addr_st2) : {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
assign dwbq_req_addr = (should_flush && dwbq_push) ? addr_st2 : {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
end else begin
assign dwbq_req_data = readdata_st2;
assign dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
@@ -579,7 +617,14 @@ module VX_bank #(
wire snp_fwd_push;
wire ffsq_empty;
assign snp_fwd_push = is_snp_st2 && valid_st2 && !ffsq_full && !(((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
assign snp_fwd_push = is_snp_st2
&& valid_st2
&& !ffsq_full
&& !(((valid_st2 && !miss_st2) && cwbq_full)
|| (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full)
|| (valid_st2 && miss_st2 && mrvq_full)
|| (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full));
assign snp_fwd_valid = !ffsq_empty;
VX_generic_queue #(
@@ -596,6 +641,10 @@ module VX_bank #(
.full (ffsq_full)
);
assign stall_bank_pipe = (is_snp_st2 && valid_st2 && ffsq_full) || ((valid_st2 && !miss_st2) && cwbq_full) || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) || (valid_st2 && miss_st2 && mrvq_full) || (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full);
assign stall_bank_pipe = (is_snp_st2 && valid_st2 && ffsq_full)
|| ((valid_st2 && !miss_st2) && cwbq_full)
|| (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full)
|| (valid_st2 && miss_st2 && mrvq_full)
|| (valid_st2 && miss_st2 && !invalidate_fill && dram_fill_req_full);
endmodule : VX_bank

View File

@@ -12,9 +12,7 @@ module VX_cache #(
// Number of Word requests per cycle {1, 2, 4, 8, ...}
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 3,
parameter STAGE_1_CYCLES = 2,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
@@ -40,14 +38,28 @@ module VX_cache #(
parameter FFSQ_SIZE = 8,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
parameter FILL_INVALIDAOR_SIZE = 16,
// Enable cache writeable
parameter WRITE_ENABLE = 1,
// Enable dram update
parameter DRAM_ENABLE = 1,
// Enable snoop forwarding
parameter SNOOP_FORWARDING_ENABLE = 0,
// Prefetcher
parameter PRFQ_SIZE = 64,
parameter PRFQ_STRIDE = 0,
// caceh requests tag size
// core request tag size
parameter CORE_TAG_WIDTH = 1,
// size of tag id in core request tag
parameter CORE_TAG_ID_BITS = 0,
// dram request tag size
parameter DRAM_TAG_WIDTH = 1
) (
input wire clk,
@@ -55,17 +67,17 @@ module VX_cache #(
// Core request
input wire [NUM_REQUESTS-1:0] core_req_valid,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_read,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] core_req_write,
input wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] core_req_read,
input wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] core_req_write,
input wire [NUM_REQUESTS-1:0][31:0] core_req_addr,
input wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_req_data,
input wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_req_tag,
input wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] core_req_tag,
output wire core_req_ready,
// Core response
output wire [NUM_REQUESTS-1:0] core_rsp_valid,
output wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_rsp_data,
output wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
output wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
input wire core_rsp_ready,
// DRAM request
@@ -127,7 +139,7 @@ module VX_cache #(
assign snp_req_ready = ~(|per_bank_snp_req_full);
assign dram_rsp_ready = (|per_bank_dram_fill_rsp_ready);
VX_cache_core_req_bank_sel #(
VX_cache_core_req_bank_sel #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
@@ -154,10 +166,10 @@ module VX_cache #(
for (i = 0; i < NUM_BANKS; i = i + 1) begin
wire [NUM_REQUESTS-1:0] curr_bank_core_req_valids;
wire [NUM_REQUESTS-1:0][31:0] curr_bank_core_req_addr;
wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] curr_bank_core_req_tag;
wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] curr_bank_core_req_tag;
wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] curr_bank_core_req_data;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] curr_bank_core_req_read;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] curr_bank_core_req_write;
wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] curr_bank_core_req_read;
wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] curr_bank_core_req_write;
wire curr_bank_core_rsp_pop;
wire curr_bank_core_rsp_valid;
@@ -241,7 +253,6 @@ module VX_cache #(
.WORD_SIZE (WORD_SIZE),
.NUM_REQUESTS (NUM_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
@@ -252,7 +263,11 @@ module VX_cache #(
.LLVQ_SIZE (LLVQ_SIZE),
.FFSQ_SIZE (FFSQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
.DRAM_ENABLE (DRAM_ENABLE),
.WRITE_ENABLE (WRITE_ENABLE),
.SNOOP_FORWARDING_ENABLE(SNOOP_FORWARDING_ENABLE),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS)
) bank (
.clk (clk),
.reset (reset),
@@ -304,14 +319,13 @@ module VX_cache #(
end
endgenerate
VX_cache_core_rsp_merge #(
VX_cache_core_rsp_merge #(
.CACHE_SIZE (CACHE_SIZE),
.BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE),
.NUM_REQUESTS (NUM_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
@@ -321,7 +335,8 @@ module VX_cache #(
.DFQQ_SIZE (DFQQ_SIZE),
.LLVQ_SIZE (LLVQ_SIZE),
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
.CORE_TAG_WIDTH (CORE_TAG_WIDTH),
.CORE_TAG_ID_BITS (CORE_TAG_ID_BITS)
) cache_core_rsp_merge (
.per_bank_core_rsp_tid (per_bank_core_rsp_tid),
.per_bank_core_rsp_valid (per_bank_core_rsp_valid),

View File

@@ -3,19 +3,19 @@
`include "VX_define.vh"
`define WORD_SEL_NO 3'h7
`define WORD_SEL_LB 3'h0
`define WORD_SEL_LH 3'h1
`define WORD_SEL_LW 3'h2
`define WORD_SEL_HB 3'h4
`define WORD_SEL_HH 3'h5
`define WORD_SEL_BITS 3
`define BYTE_EN_NO 3'h7
`define BYTE_EN_LB 3'h0
`define BYTE_EN_LH 3'h1
`define BYTE_EN_LW 3'h2
`define BYTE_EN_HB 3'h4
`define BYTE_EN_HH 3'h5
`define BYTE_EN_BITS 3
// data tid tag read write base addr
`define MRVQ_METADATA_WIDTH (`WORD_WIDTH + `LOG2UP(NUM_REQUESTS) + CORE_TAG_WIDTH + `WORD_SEL_BITS + `WORD_SEL_BITS + `BASE_ADDR_BITS)
`define MRVQ_METADATA_WIDTH (`WORD_WIDTH + `LOG2UP(NUM_REQUESTS) + CORE_TAG_WIDTH + `BYTE_EN_BITS + `BYTE_EN_BITS + `BASE_ADDR_BITS)
// tag read write reqs
`define REQ_INST_META_WIDTH (CORE_TAG_WIDTH + `WORD_SEL_BITS + `WORD_SEL_BITS + `LOG2UP(NUM_REQUESTS))
`define REQ_INST_META_WIDTH (CORE_TAG_WIDTH + `BYTE_EN_BITS + `BYTE_EN_BITS + `LOG2UP(NUM_REQUESTS))
`define WORD_WIDTH (8 * WORD_SIZE)
`define BYTE_WIDTH (`WORD_WIDTH / 4)
@@ -66,13 +66,8 @@
///////////////////////////////////////////////////////////////////////////////
// Core request tag width pc, wb, rd, warp_num
`define CORE_REQ_TAG_WIDTH (32 + 2 + 5 + `NW_BITS)
`define CORE_REQ_TAG_COUNT ((CORE_TAG_ID_BITS != 0) ? 1 : NUM_REQUESTS)
// Core request tag info rd + warp_num
`define CORE_REQ_TAG_WARP(x) x[(5 + `NW_BITS)-1:0]
// DRAM response tag bank info
`define DRAM_ADDR_BANK(x) x[`BANK_SELECT_BITS-1:0]
`define DRAM_TO_LINE_ADDR(x) x[`DRAM_ADDR_WIDTH-1:`BANK_SELECT_BITS]

View File

@@ -13,8 +13,6 @@ module VX_cache_core_rsp_merge #(
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
@@ -39,8 +37,12 @@ module VX_cache_core_rsp_merge #(
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1,
// core request tag size
parameter CORE_TAG_WIDTH = 1,
// size of tag id in core request tag
parameter CORE_TAG_ID_BITS = 0,
// dram request tag size
parameter DRAM_TAG_WIDTH = 1
) (
// Per Bank WB
@@ -53,7 +55,7 @@ module VX_cache_core_rsp_merge #(
// Core Writeback
output reg [NUM_REQUESTS-1:0] core_rsp_valid,
output reg [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] core_rsp_data,
output reg [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
output reg [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] core_rsp_tag,
input wire core_rsp_ready
);
@@ -73,43 +75,48 @@ module VX_cache_core_rsp_merge #(
);
integer i;
generate
if (CORE_TAG_ID_BITS != 0) begin
assign core_rsp_tag = per_bank_core_rsp_tag[main_bank_index];
always @(*) begin
core_rsp_valid = 0;
core_rsp_data = 0;
core_rsp_tag = 0;
for (i = 0; i < NUM_BANKS; i = i + 1) begin
if ((FUNC_ID == `L2FUNC_ID)
|| (FUNC_ID == `L3FUNC_ID)) begin
if (found_bank
&& per_bank_core_rsp_valid[i]
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_core_rsp_tid[i] != per_bank_core_rsp_tid[main_bank_index]))) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_tag[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_tag[i];
per_bank_core_rsp_pop_unqual[i] = 1;
end else begin
per_bank_core_rsp_pop_unqual[i] = 0;
end
for (i = 0; i < NUM_BANKS; i = i + 1) begin
if (found_bank
&& per_bank_core_rsp_valid[i]
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_core_rsp_tid[i] != per_bank_core_rsp_tid[main_bank_index]))
&& (per_bank_core_rsp_tag[i][CORE_TAG_ID_BITS-1:0] == per_bank_core_rsp_tag[main_bank_index][CORE_TAG_ID_BITS-1:0])) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
per_bank_core_rsp_pop_unqual[i] = 1;
end else begin
if (found_bank
&& per_bank_core_rsp_valid[i]
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_core_rsp_tid[i] != per_bank_core_rsp_tid[main_bank_index]))
&& (`CORE_REQ_TAG_WARP(per_bank_core_rsp_tag[i]) == `CORE_REQ_TAG_WARP(per_bank_core_rsp_tag[main_bank_index]))) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_tag[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_tag[i];
per_bank_core_rsp_pop_unqual[i] = 1;
end else begin
per_bank_core_rsp_pop_unqual[i] = 0;
end
per_bank_core_rsp_pop_unqual[i] = 0;
end
end
end
end
endgenerate
end else begin
always @(*) begin
core_rsp_valid = 0;
core_rsp_data = 0;
core_rsp_tag = 0;
for (i = 0; i < NUM_BANKS; i = i + 1) begin
if (found_bank
&& per_bank_core_rsp_valid[i]
&& !core_rsp_valid[per_bank_core_rsp_tid[i]]
&& ((main_bank_index == `LOG2UP(NUM_BANKS)'(i))
|| (per_bank_core_rsp_tid[i] != per_bank_core_rsp_tid[main_bank_index]))) begin
core_rsp_valid[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];
core_rsp_tag[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_tag[i];
per_bank_core_rsp_pop_unqual[i] = 1;
end else begin
per_bank_core_rsp_pop_unqual[i] = 0;
end
end
end
end
endmodule

View File

@@ -87,7 +87,7 @@ module VX_cache_dfq_queue #(
assign qual_bank_dram_fill_req_addr = use_empty ? out_per_bank_dram_fill_req_addr : use_per_bank_dram_fill_req_addr;
wire[`LOG2UP(NUM_BANKS)-1:0] qual_request_index;
wire qual_has_request;
wire qual_has_request;
VX_generic_priority_encoder #(
.N(NUM_BANKS)
@@ -105,12 +105,12 @@ module VX_cache_dfq_queue #(
always @(posedge clk) begin
if (reset) begin
use_per_bank_dram_fill_req_valid <= 0;
use_per_bank_dram_fill_req_addr <= 0;
use_per_bank_dram_fill_req_valid <= 0;
use_per_bank_dram_fill_req_addr <= 0;
end else begin
if (dfqq_pop && qual_has_request) begin
use_per_bank_dram_fill_req_valid <= updated_bank_dram_fill_req_valid;
use_per_bank_dram_fill_req_addr <= qual_bank_dram_fill_req_addr;
use_per_bank_dram_fill_req_addr <= qual_bank_dram_fill_req_addr;
end
end
end

View File

@@ -79,7 +79,7 @@ module VX_cache_dram_req_arb #(
.PRFQ_STRIDE (PRFQ_STRIDE),
.BANK_LINE_SIZE(BANK_LINE_SIZE),
.WORD_SIZE (WORD_SIZE)
) prfqq (
) prfqq (
.clk (clk),
.reset (reset),

View File

@@ -1,4 +1,3 @@
`include "VX_cache_config.vh"
module VX_cache_miss_resrv #(
@@ -51,8 +50,8 @@ module VX_cache_miss_resrv #(
input wire[`WORD_WIDTH-1:0] miss_add_data,
input wire[`LOG2UP(NUM_REQUESTS)-1:0] miss_add_tid,
input wire[CORE_TAG_WIDTH-1:0] miss_add_tag,
input wire[`WORD_SEL_BITS-1:0] miss_add_mem_read,
input wire[`WORD_SEL_BITS-1:0] miss_add_mem_write,
input wire[`BYTE_EN_BITS-1:0] miss_add_mem_read,
input wire[`BYTE_EN_BITS-1:0] miss_add_mem_write,
output wire miss_resrv_full,
output wire miss_resrv_stop,
@@ -72,8 +71,8 @@ module VX_cache_miss_resrv #(
output wire[`WORD_WIDTH-1:0] miss_resrv_data_st0,
output wire[`LOG2UP(NUM_REQUESTS)-1:0] miss_resrv_tid_st0,
output wire[CORE_TAG_WIDTH-1:0] miss_resrv_tag_st0,
output wire[`WORD_SEL_BITS-1:0] miss_resrv_mem_read_st0,
output wire[`WORD_SEL_BITS-1:0] miss_resrv_mem_write_st0
output wire[`BYTE_EN_BITS-1:0] miss_resrv_mem_read_st0,
output wire[`BYTE_EN_BITS-1:0] miss_resrv_mem_write_st0
);
reg [`MRVQ_METADATA_WIDTH-1:0] metadata_table[MRVQ_SIZE-1:0];
reg [MRVQ_SIZE-1:0][`LINE_ADDR_WIDTH-1:0] addr_table;

View File

@@ -37,8 +37,11 @@ module VX_cache_req_queue #(
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16,
// caceh requests tag size
parameter CORE_TAG_WIDTH = 1
// core request tag size
parameter CORE_TAG_WIDTH = 1,
// size of tag id in core request tag
parameter CORE_TAG_ID_BITS = 0
) (
input wire clk,
input wire reset,
@@ -46,18 +49,18 @@ module VX_cache_req_queue #(
// Enqueue Data
input wire reqq_push,
input wire [NUM_REQUESTS-1:0] bank_valids,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] bank_mem_read,
input wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] bank_mem_write,
input wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] bank_mem_read,
input wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] bank_mem_write,
input wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] bank_writedata,
input wire [NUM_REQUESTS-1:0][31:0] bank_addr,
input wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] bank_tag,
input wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] bank_tag,
// Dequeue Data
input wire reqq_pop,
output wire reqq_req_st0,
output wire [`LOG2UP(NUM_REQUESTS)-1:0] reqq_req_tid_st0,
output wire [`WORD_SEL_BITS-1:0] reqq_req_mem_read_st0,
output wire [`WORD_SEL_BITS-1:0] reqq_req_mem_write_st0,
output wire [`BYTE_EN_BITS-1:0] reqq_req_mem_read_st0,
output wire [`BYTE_EN_BITS-1:0] reqq_req_mem_write_st0,
output wire [`WORD_WIDTH-1:0] reqq_req_writedata_st0,
output wire [31:0] reqq_req_addr_st0,
output wire [CORE_TAG_WIDTH-1:0] reqq_req_tag_st0,
@@ -70,23 +73,23 @@ module VX_cache_req_queue #(
wire [NUM_REQUESTS-1:0] out_per_valids;
wire [NUM_REQUESTS-1:0][31:0] out_per_addr;
wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] out_per_writedata;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] out_per_mem_read;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] out_per_mem_write;
wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] out_per_tag;
wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] out_per_mem_read;
wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] out_per_mem_write;
wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] out_per_tag;
reg [NUM_REQUESTS-1:0] use_per_valids;
reg [NUM_REQUESTS-1:0][31:0] use_per_addr;
reg [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] use_per_writedata;
reg [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] use_per_mem_read;
reg [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] use_per_mem_write;
reg [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] use_per_tag;
reg [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] use_per_mem_read;
reg [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] use_per_mem_write;
reg [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] use_per_tag;
wire [NUM_REQUESTS-1:0] qual_valids;
wire [NUM_REQUESTS-1:0][31:0] qual_addr;
wire [NUM_REQUESTS-1:0][`WORD_WIDTH-1:0] qual_writedata;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] qual_mem_read;
wire [NUM_REQUESTS-1:0][`WORD_SEL_BITS-1:0] qual_mem_write;
wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] qual_tag;
wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] qual_mem_read;
wire [NUM_REQUESTS-1:0][`BYTE_EN_BITS-1:0] qual_mem_write;
wire [`CORE_REQ_TAG_COUNT-1:0][CORE_TAG_WIDTH-1:0] qual_tag;
`DEBUG_BEGIN
reg [NUM_REQUESTS-1:0] updated_valids;
@@ -123,8 +126,8 @@ module VX_cache_req_queue #(
assign qual_mem_read = use_per_mem_read;
assign qual_mem_write = use_per_mem_write;
wire[`LOG2UP(NUM_REQUESTS)-1:0]qual_request_index;
wire qual_has_request;
wire[`LOG2UP(NUM_REQUESTS)-1:0] qual_request_index;
wire qual_has_request;
VX_generic_priority_encoder #(
.N(NUM_REQUESTS)
@@ -139,9 +142,15 @@ module VX_cache_req_queue #(
assign reqq_req_tid_st0 = qual_request_index;
assign reqq_req_addr_st0 = qual_addr[qual_request_index];
assign reqq_req_writedata_st0 = qual_writedata[qual_request_index];
assign reqq_req_tag_st0 = qual_tag[qual_request_index];
assign reqq_req_mem_read_st0 = qual_mem_read [qual_request_index];
assign reqq_req_mem_write_st0 = qual_mem_write[qual_request_index];
if (CORE_TAG_ID_BITS != 0) begin
assign reqq_req_tag_st0 = qual_tag;
end else begin
assign reqq_req_tag_st0 = qual_tag[qual_request_index];
end
assign reqq_req_mem_read_st0 = qual_mem_read [qual_request_index];
assign reqq_req_mem_write_st0 = qual_mem_write[qual_request_index];
always @(*) begin
updated_valids = qual_valids;

View File

@@ -13,8 +13,6 @@ module VX_tag_data_access #(
parameter NUM_REQUESTS = 2,
// Number of cycles to complete stage 1 (read from memory)
parameter STAGE_1_CYCLES = 2,
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
parameter FUNC_ID = 0,
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
// Core Request Queue Size
@@ -37,7 +35,13 @@ module VX_tag_data_access #(
parameter LLVQ_SIZE = 16,
// Fill Invalidator Size {Fill invalidator must be active}
parameter FILL_INVALIDAOR_SIZE = 16
parameter FILL_INVALIDAOR_SIZE = 16,
// Enable cache writeable
parameter WRITE_ENABLE = 1,
// Enable dram update
parameter DRAM_ENABLE = 1
) (
input wire clk,
input wire reset,
@@ -55,8 +59,8 @@ module VX_tag_data_access #(
`IGNORE_WARNINGS_BEGIN
input wire[`WORD_SELECT_ADDR_END:0] writewsel_st1e,
input wire[`WORD_SEL_BITS-1:0] mem_write_st1e,
input wire[`WORD_SEL_BITS-1:0] mem_read_st1e,
input wire[`BYTE_EN_BITS-1:0] mem_write_st1e,
input wire[`BYTE_EN_BITS-1:0] mem_read_st1e,
`IGNORE_WARNINGS_END
output wire[`WORD_WIDTH-1:0] readword_st1e,
@@ -101,7 +105,6 @@ module VX_tag_data_access #(
.WORD_SIZE (WORD_SIZE),
.NUM_REQUESTS (NUM_REQUESTS),
.STAGE_1_CYCLES (STAGE_1_CYCLES),
.FUNC_ID (FUNC_ID),
.REQQ_SIZE (REQQ_SIZE),
.MRVQ_SIZE (MRVQ_SIZE),
.DFPQ_SIZE (DFPQ_SIZE),
@@ -139,8 +142,8 @@ module VX_tag_data_access #(
.reset(reset),
.stall(stall),
.flush(0),
.in ({qual_read_valid_st1, qual_read_dirty_st1, qual_read_tag_st1, qual_read_data_st1}),
.out ({read_valid_st1c[0], read_dirty_st1c[0], read_tag_st1c[0], read_data_st1c[0]})
.in({qual_read_valid_st1, qual_read_dirty_st1, qual_read_tag_st1, qual_read_data_st1}),
.out({read_valid_st1c[0], read_dirty_st1c[0], read_tag_st1c[0], read_data_st1c[0]})
);
genvar i;
@@ -152,14 +155,14 @@ module VX_tag_data_access #(
.reset(reset),
.stall(stall),
.flush(0),
.in ({read_valid_st1c[i-1], read_dirty_st1c[i-1], read_tag_st1c[i-1], read_data_st1c[i-1]}),
.out ({read_valid_st1c[i], read_dirty_st1c[i], read_tag_st1c[i], read_data_st1c[i]})
.in({read_valid_st1c[i-1], read_dirty_st1c[i-1], read_tag_st1c[i-1], read_data_st1c[i-1]}),
.out({read_valid_st1c[i], read_dirty_st1c[i], read_tag_st1c[i], read_data_st1c[i]})
);
end
assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || (FUNC_ID == `SFUNC_ID); // If shared memory, always valid
assign use_read_dirty_st1e = read_dirty_st1c[STAGE_1_CYCLES-1] && (FUNC_ID != `SFUNC_ID); // Dirty only applies in Dcache
assign use_read_tag_st1e = (FUNC_ID == `SFUNC_ID) ? writeaddr_st1e[`TAG_LINE_ADDR_RNG] : read_tag_st1c[STAGE_1_CYCLES-1]; // Tag is always the same in SM
assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || ~DRAM_ENABLE; // If shared memory, always valid
assign use_read_dirty_st1e = read_dirty_st1c[STAGE_1_CYCLES-1] && DRAM_ENABLE; // Dirty only applies in Dcache
assign use_read_tag_st1e = DRAM_ENABLE ? read_tag_st1c[STAGE_1_CYCLES-1] : writeaddr_st1e[`TAG_LINE_ADDR_RNG]; // Tag is always the same in SM
for (i = 0; i < `BANK_LINE_WORDS; i = i + 1) begin
assign use_read_data_st1e[i * `WORD_WIDTH +: `WORD_WIDTH] = read_data_st1c[STAGE_1_CYCLES-1][i * `WORD_WIDTH +: `WORD_WIDTH];
@@ -170,30 +173,49 @@ module VX_tag_data_access #(
wire [`BANK_LINE_WORDS-1:0][3:0] we;
wire [`BANK_LINE_WIDTH-1:0] data_write;
if (WORD_SIZE == 4) begin
if (WORD_SIZE == BANK_LINE_SIZE) begin
wire should_write = ((mem_write_st1e != `BYTE_EN_NO))
&& valid_req_st1e
&& use_read_valid_st1e
&& !miss_st1e
&& !is_snp_st1e;
for (i = 0; i < `BANK_LINE_WORDS; i = i + 1) begin
assign we[i] = (force_write || (should_write && !real_writefill)) ? 4'b1111 : 4'b0000;
end
assign readword_st1e = read_data_st1c[STAGE_1_CYCLES-1];
assign data_write = force_write ? writedata_st1e : writeword_st1e;
end else begin
wire[`OFFSET_ADDR_BITS-1:0] byte_select = writewsel_st1e[`OFFSET_ADDR_RNG];
wire[`WORD_SELECT_BITS-1:0] block_offset = writewsel_st1e[`WORD_SELECT_ADDR_RNG];
wire lb = valid_req_st1e && (mem_read_st1e == `WORD_SEL_LB);
wire lh = valid_req_st1e && (mem_read_st1e == `WORD_SEL_LH);
wire lbu = valid_req_st1e && (mem_read_st1e == `WORD_SEL_HB);
wire lhu = valid_req_st1e && (mem_read_st1e == `WORD_SEL_HH);
wire lw = valid_req_st1e && (mem_read_st1e == `WORD_SEL_LW);
wire lb = valid_req_st1e && (mem_read_st1e == `BYTE_EN_LB);
wire lh = valid_req_st1e && (mem_read_st1e == `BYTE_EN_LH);
wire lbu = valid_req_st1e && (mem_read_st1e == `BYTE_EN_HB);
wire lhu = valid_req_st1e && (mem_read_st1e == `BYTE_EN_HH);
wire lw = valid_req_st1e && (mem_read_st1e == `BYTE_EN_LW);
wire b0 = (byte_select == 0);
wire b1 = (byte_select == 1);
wire b2 = (byte_select == 2);
wire b3 = (byte_select == 3);
wire sb = valid_req_st1e && (mem_write_st1e == `WORD_SEL_LB);
wire sh = valid_req_st1e && (mem_write_st1e == `WORD_SEL_LH);
wire sw = valid_req_st1e && (mem_write_st1e == `WORD_SEL_LW);
wire sb = valid_req_st1e && (mem_write_st1e == `BYTE_EN_LB);
wire sh = valid_req_st1e && (mem_write_st1e == `BYTE_EN_LH);
wire sw = valid_req_st1e && (mem_write_st1e == `BYTE_EN_LW);
wire [3:0] sb_mask = (b0 ? 4'b0001 : (b1 ? 4'b0010 : (b2 ? 4'b0100 : 4'b1000)));
wire [3:0] sh_mask = (b0 ? 4'b0011 : 4'b1100);
wire should_write = (sw || sb || sh) && valid_req_st1e && use_read_valid_st1e && !miss_st1e && !is_snp_st1e;
wire should_write = (sw || sb || sh)
&& valid_req_st1e
&& use_read_valid_st1e
&& !miss_st1e
&& !is_snp_st1e;
wire[`WORD_WIDTH-1:0] data_unmod = read_data_st1c[STAGE_1_CYCLES-1][block_offset * 32 +: 32];
wire[`WORD_WIDTH-1:0] data_unQual = (b0 || lw) ? (data_unmod) :
@@ -234,13 +256,6 @@ module VX_tag_data_access #(
assign data_write[i * `WORD_WIDTH +: `WORD_WIDTH] = force_write ? writedata_st1e[i * `WORD_WIDTH +: `WORD_WIDTH] : use_write_dat;
end
end else begin
wire should_write = ((mem_write_st1e != `WORD_SEL_NO)) && valid_req_st1e && use_read_valid_st1e && !miss_st1e && !is_snp_st1e;
for (i = 0; i < `BANK_LINE_WORDS; i = i + 1) begin
assign we[i] = (force_write || (should_write && !real_writefill)) ? 4'b1111 : 4'b0000;
end
assign readword_st1e = read_data_st1c[STAGE_1_CYCLES-1];
assign data_write = force_write ? writedata_st1e : writeword_st1e;
end
assign use_write_enable = (writefill_st1e && !real_writefill) ? 0 : we;