cache bank area optimization + multi-porting fix for l2/l3 caches
This commit is contained in:
122
hw/rtl/cache/VX_nc_bypass.v
vendored
122
hw/rtl/cache/VX_nc_bypass.v
vendored
@@ -1,6 +1,7 @@
|
||||
`include "VX_cache_define.vh"
|
||||
|
||||
module VX_nc_bypass #(
|
||||
parameter NUM_PORTS = 1,
|
||||
parameter NUM_REQS = 1,
|
||||
parameter NUM_RSP_TAGS = 0,
|
||||
parameter NC_TAG_BIT = 0,
|
||||
@@ -10,13 +11,14 @@ module VX_nc_bypass #(
|
||||
parameter CORE_TAG_IN_WIDTH = 1,
|
||||
|
||||
parameter MEM_ADDR_WIDTH = 1,
|
||||
parameter MEM_DATA_SIZE = 1,
|
||||
parameter MEM_DATA_SIZE = 1,
|
||||
parameter MEM_TAG_IN_WIDTH = 1,
|
||||
parameter MEM_TAG_OUT_WIDTH = 1,
|
||||
|
||||
localparam CORE_DATA_WIDTH = CORE_DATA_SIZE * 8,
|
||||
localparam MEM_DATA_WIDTH = MEM_DATA_SIZE * 8,
|
||||
localparam CORE_TAG_OUT_WIDTH = CORE_TAG_IN_WIDTH - 1
|
||||
localparam CORE_DATA_WIDTH = CORE_DATA_SIZE * 8,
|
||||
localparam MEM_DATA_WIDTH = MEM_DATA_SIZE * 8,
|
||||
localparam CORE_TAG_OUT_WIDTH = CORE_TAG_IN_WIDTH - 1,
|
||||
localparam MEM_SELECT_BITS = `UP(`CLOG2(MEM_DATA_SIZE / CORE_DATA_SIZE))
|
||||
) (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
@@ -57,8 +59,9 @@ module VX_nc_bypass #(
|
||||
input wire mem_req_valid_in,
|
||||
input wire mem_req_rw_in,
|
||||
input wire [MEM_ADDR_WIDTH-1:0] mem_req_addr_in,
|
||||
input wire [MEM_DATA_SIZE-1:0] mem_req_byteen_in,
|
||||
input wire [MEM_DATA_WIDTH-1:0] mem_req_data_in,
|
||||
input wire [NUM_PORTS-1:0][CORE_DATA_SIZE-1:0] mem_req_byteen_in,
|
||||
input wire [NUM_PORTS-1:0][MEM_SELECT_BITS-1:0] mem_req_wsel_in,
|
||||
input wire [NUM_PORTS-1:0][CORE_DATA_WIDTH-1:0] mem_req_data_in,
|
||||
input wire [MEM_TAG_IN_WIDTH-1:0] mem_req_tag_in,
|
||||
output wire mem_req_ready_in,
|
||||
|
||||
@@ -66,8 +69,9 @@ module VX_nc_bypass #(
|
||||
output wire mem_req_valid_out,
|
||||
output wire mem_req_rw_out,
|
||||
output wire [MEM_ADDR_WIDTH-1:0] mem_req_addr_out,
|
||||
output wire [MEM_DATA_SIZE-1:0] mem_req_byteen_out,
|
||||
output wire [MEM_DATA_WIDTH-1:0] mem_req_data_out,
|
||||
output wire [NUM_PORTS-1:0][CORE_DATA_SIZE-1:0] mem_req_byteen_out,
|
||||
output wire [NUM_PORTS-1:0][MEM_SELECT_BITS-1:0] mem_req_wsel_out,
|
||||
output wire [NUM_PORTS-1:0][CORE_DATA_WIDTH-1:0] mem_req_data_out,
|
||||
output wire [MEM_TAG_OUT_WIDTH-1:0] mem_req_tag_out,
|
||||
input wire mem_req_ready_out,
|
||||
|
||||
@@ -148,7 +152,7 @@ module VX_nc_bypass #(
|
||||
assign mem_req_valid_out = mem_req_valid_in || core_req_nc_valid;
|
||||
assign mem_req_ready_in = mem_req_ready_out;
|
||||
|
||||
wire [(MEM_TAG_IN_WIDTH+1)-1:0] mem_req_tag_in_nc;
|
||||
wire [(MEM_TAG_IN_WIDTH+1)-1:0] mem_req_tag_in_c;
|
||||
|
||||
VX_bits_insert #(
|
||||
.N (MEM_TAG_IN_WIDTH),
|
||||
@@ -157,74 +161,66 @@ module VX_nc_bypass #(
|
||||
) mem_req_tag_insert (
|
||||
.data_in (mem_req_tag_in),
|
||||
.sel_in ('0),
|
||||
.data_out (mem_req_tag_in_nc)
|
||||
.data_out (mem_req_tag_in_c)
|
||||
);
|
||||
|
||||
wire [CORE_TAG_IN_WIDTH-1:0] core_req_tag_in_sel;
|
||||
wire [CORE_DATA_WIDTH-1:0] core_req_data_in_sel;
|
||||
wire [CORE_DATA_SIZE-1:0] core_req_byteen_in_sel;
|
||||
wire [CORE_ADDR_WIDTH-1:0] core_req_addr_in_sel;
|
||||
wire core_req_rw_in_sel;
|
||||
|
||||
if (NUM_REQS > 1) begin
|
||||
|
||||
wire [CORE_TAG_IN_WIDTH-1:0] core_req_tag_in_sel;
|
||||
wire [CORE_DATA_WIDTH-1:0] core_req_data_in_sel;
|
||||
wire [CORE_DATA_SIZE-1:0] core_req_byteen_in_sel;
|
||||
wire [CORE_ADDR_WIDTH-1:0] core_req_addr_in_sel;
|
||||
wire core_req_rw_in_sel;
|
||||
|
||||
wire [NUM_REQS-1:0][MUX_DATAW-1:0] core_req_nc_mux_in;
|
||||
for (genvar i = 0; i < NUM_REQS; ++i) begin
|
||||
assign core_req_nc_mux_in[i] = {core_req_tag_in[i], core_req_data_in[i], core_req_byteen_in[i], core_req_addr_in[i], core_req_rw_in[i]};
|
||||
end
|
||||
|
||||
assign {core_req_tag_in_sel, core_req_data_in_sel, core_req_byteen_in_sel, core_req_addr_in_sel, core_req_rw_in_sel} = core_req_nc_mux_in[core_req_nc_tid];
|
||||
|
||||
assign mem_req_rw_out = mem_req_valid_in ? mem_req_rw_in : core_req_rw_in_sel;
|
||||
assign mem_req_addr_out = mem_req_valid_in ? mem_req_addr_in : core_req_addr_in_sel[D +: MEM_ADDR_WIDTH];
|
||||
|
||||
for (genvar i = 0; i < P; ++i) begin
|
||||
assign mem_req_data_out[i * CORE_DATA_WIDTH +: CORE_DATA_WIDTH] = mem_req_valid_in ?
|
||||
mem_req_data_in[i * CORE_DATA_WIDTH +: CORE_DATA_WIDTH] : core_req_data_in_sel;
|
||||
end
|
||||
|
||||
if (D != 0) begin
|
||||
wire [D-1:0] req_addr_idx = core_req_addr_in_sel[D-1:0];
|
||||
reg [MEM_DATA_SIZE-1:0] mem_req_byteen_in_r;
|
||||
always @(*) begin
|
||||
mem_req_byteen_in_r = 0;
|
||||
mem_req_byteen_in_r[req_addr_idx * CORE_DATA_SIZE +: CORE_DATA_SIZE] = core_req_byteen_in_sel;
|
||||
end
|
||||
assign mem_req_byteen_out = mem_req_valid_in ? mem_req_byteen_in : mem_req_byteen_in_r;
|
||||
assign mem_req_tag_out = mem_req_valid_in ? MEM_TAG_OUT_WIDTH'(mem_req_tag_in_nc) : MEM_TAG_OUT_WIDTH'({core_req_nc_tid, req_addr_idx, core_req_tag_in_sel});
|
||||
end else begin
|
||||
assign mem_req_byteen_out = mem_req_valid_in ? mem_req_byteen_in : core_req_byteen_in_sel;
|
||||
assign mem_req_tag_out = mem_req_valid_in ? MEM_TAG_OUT_WIDTH'(mem_req_tag_in_nc) : MEM_TAG_OUT_WIDTH'({core_req_nc_tid, core_req_tag_in_sel});
|
||||
end
|
||||
end else begin
|
||||
`UNUSED_VAR (core_req_nc_tid)
|
||||
assign core_req_tag_in_sel = core_req_tag_in;
|
||||
assign core_req_data_in_sel = core_req_data_in;
|
||||
assign core_req_byteen_in_sel = core_req_byteen_in;
|
||||
assign core_req_addr_in_sel = core_req_addr_in;
|
||||
assign core_req_rw_in_sel = core_req_rw_in;
|
||||
end
|
||||
|
||||
assign mem_req_rw_out = mem_req_valid_in ? mem_req_rw_in : core_req_rw_in_sel;
|
||||
assign mem_req_addr_out = mem_req_valid_in ? mem_req_addr_in : core_req_addr_in_sel[D +: MEM_ADDR_WIDTH];
|
||||
|
||||
if (D != 0) begin
|
||||
reg [NUM_PORTS-1:0][CORE_DATA_SIZE-1:0] mem_req_byteen_in_r;
|
||||
reg [NUM_PORTS-1:0][MEM_SELECT_BITS-1:0] mem_req_wsel_in_r;
|
||||
reg [NUM_PORTS-1:0][CORE_DATA_WIDTH-1:0] mem_req_data_in_r;
|
||||
|
||||
wire [D-1:0] req_addr_idx = core_req_addr_in_sel[D-1:0];
|
||||
|
||||
assign mem_req_rw_out = mem_req_valid_in ? mem_req_rw_in : core_req_rw_in;
|
||||
assign mem_req_addr_out = mem_req_valid_in ? mem_req_addr_in : core_req_addr_in[0][D +: MEM_ADDR_WIDTH];
|
||||
always @(*) begin
|
||||
mem_req_byteen_in_r = 0;
|
||||
mem_req_byteen_in_r[0] = core_req_byteen_in_sel;
|
||||
|
||||
for (genvar i = 0; i < P; ++i) begin
|
||||
assign mem_req_data_out[i * CORE_DATA_WIDTH +: CORE_DATA_WIDTH] = mem_req_valid_in ?
|
||||
mem_req_data_in[i * CORE_DATA_WIDTH +: CORE_DATA_WIDTH] : core_req_data_in;
|
||||
mem_req_wsel_in_r = 'x;
|
||||
mem_req_wsel_in_r[0] = req_addr_idx;
|
||||
|
||||
mem_req_data_in_r = 'x;
|
||||
mem_req_data_in_r[0] = core_req_data_in_sel;
|
||||
end
|
||||
|
||||
if (D != 0) begin
|
||||
wire [D-1:0] req_addr_idx = core_req_addr_in[0][D-1:0];
|
||||
reg [MEM_DATA_SIZE-1:0] mem_req_byteen_in_r;
|
||||
always @(*) begin
|
||||
mem_req_byteen_in_r = 0;
|
||||
mem_req_byteen_in_r[req_addr_idx * CORE_DATA_SIZE +: CORE_DATA_SIZE] = core_req_byteen_in;
|
||||
end
|
||||
assign mem_req_byteen_out = mem_req_valid_in ? mem_req_byteen_in : mem_req_byteen_in_r;
|
||||
assign mem_req_tag_out = mem_req_valid_in ? MEM_TAG_OUT_WIDTH'(mem_req_tag_in_nc) : MEM_TAG_OUT_WIDTH'({req_addr_idx, core_req_tag_in});
|
||||
end else begin
|
||||
assign mem_req_byteen_out = mem_req_valid_in ? mem_req_byteen_in : core_req_byteen_in;
|
||||
assign mem_req_tag_out = mem_req_valid_in ? MEM_TAG_OUT_WIDTH'(mem_req_tag_in_nc) : MEM_TAG_OUT_WIDTH'(core_req_tag_in);
|
||||
end
|
||||
assign mem_req_byteen_out = mem_req_valid_in ? mem_req_byteen_in : mem_req_byteen_in_r;
|
||||
assign mem_req_wsel_out = mem_req_valid_in ? mem_req_wsel_in : mem_req_wsel_in_r;
|
||||
assign mem_req_data_out = mem_req_valid_in ? mem_req_data_in : mem_req_data_in_r;
|
||||
assign mem_req_tag_out = mem_req_valid_in ? MEM_TAG_OUT_WIDTH'(mem_req_tag_in_c) : MEM_TAG_OUT_WIDTH'({core_req_nc_tid, req_addr_idx, core_req_tag_in_sel});
|
||||
end else begin
|
||||
`UNUSED_VAR (mem_req_wsel_in)
|
||||
assign mem_req_byteen_out = mem_req_valid_in ? mem_req_byteen_in : core_req_byteen_in_sel;
|
||||
assign mem_req_data_out = mem_req_valid_in ? mem_req_data_in : core_req_data_in_sel;
|
||||
assign mem_req_wsel_out = 0;
|
||||
assign mem_req_tag_out = mem_req_valid_in ? MEM_TAG_OUT_WIDTH'(mem_req_tag_in_c) : MEM_TAG_OUT_WIDTH'({core_req_nc_tid, core_req_tag_in_sel});
|
||||
end
|
||||
|
||||
// core response handling
|
||||
|
||||
wire [NUM_RSP_TAGS-1:0][CORE_TAG_IN_WIDTH-1:0] core_rsp_tag_out_unqual;
|
||||
wire [NUM_RSP_TAGS-1:0][CORE_TAG_IN_WIDTH-1:0] core_rsp_tag_out_c;
|
||||
|
||||
wire is_mem_rsp_nc = mem_rsp_valid_in && mem_rsp_tag_in[NC_TAG_BIT];
|
||||
|
||||
@@ -236,7 +232,7 @@ module VX_nc_bypass #(
|
||||
) core_rsp_tag_insert (
|
||||
.data_in (core_rsp_tag_in[i]),
|
||||
.sel_in ('0),
|
||||
.data_out (core_rsp_tag_out_unqual[i])
|
||||
.data_out (core_rsp_tag_out_c[i])
|
||||
);
|
||||
end
|
||||
|
||||
@@ -262,14 +258,14 @@ module VX_nc_bypass #(
|
||||
for (genvar i = 0; i < NUM_REQS; ++i) begin
|
||||
assign core_rsp_data_out[i] = core_rsp_valid_in[i] ? core_rsp_data_in[i] : mem_rsp_data_in;
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
for (genvar i = 0; i < NUM_REQS; ++i) begin
|
||||
assign core_rsp_tag_out[i] = core_rsp_valid_in[i] ? core_rsp_tag_out_unqual[i] : mem_rsp_tag_in[CORE_TAG_IN_WIDTH-1:0];
|
||||
assign core_rsp_tag_out[i] = core_rsp_valid_in[i] ? core_rsp_tag_out_c[i] : mem_rsp_tag_in[CORE_TAG_IN_WIDTH-1:0];
|
||||
end
|
||||
end else begin
|
||||
assign core_rsp_valid_out = core_rsp_valid_in || is_mem_rsp_nc;
|
||||
assign core_rsp_tag_out = core_rsp_valid_in ? core_rsp_tag_out_unqual : mem_rsp_tag_in[CORE_TAG_IN_WIDTH-1:0];
|
||||
assign core_rsp_tag_out = core_rsp_valid_in ? core_rsp_tag_out_c : mem_rsp_tag_in[CORE_TAG_IN_WIDTH-1:0];
|
||||
assign core_rsp_ready_in = core_rsp_ready_out;
|
||||
|
||||
if (NUM_REQS > 1) begin
|
||||
|
||||
Reference in New Issue
Block a user