mor1kx-bemicrocv/ip/altera/ddr3/altera_merlin_arbitrator.sv

273 lines
9.1 KiB
Systemverilog
Raw Normal View History

2016-08-04 19:22:38 +02:00
// (C) 2001-2015 Altera Corporation. All rights reserved.
// Your use of Altera Corporation's design tools, logic functions and other
// software and tools, and its AMPP partner logic functions, and any output
// files any of the foregoing (including device programming or simulation
// files), and any associated documentation or information are expressly subject
// to the terms and conditions of the Altera Program License Subscription
// Agreement, Altera MegaCore Function License Agreement, or other applicable
// license agreement, including, without limitation, that your use is for the
// sole purpose of programming logic devices manufactured by Altera and sold by
// Altera or its authorized distributors. Please refer to the applicable
// agreement for further details.
// (C) 2001-2010 Altera Corporation. All rights reserved.
// Your use of Altera Corporation's design tools, logic functions and other
// software and tools, and its AMPP partner logic functions, and any output
// files any of the foregoing (including device programming or simulation
// files), and any associated documentation or information are expressly subject
// to the terms and conditions of the Altera Program License Subscription
// Agreement, Altera MegaCore Function License Agreement, or other applicable
// license agreement, including, without limitation, that your use is for the
// sole purpose of programming logic devices manufactured by Altera and sold by
// Altera or its authorized distributors. Please refer to the applicable
// agreement for further details.
// $Id: //acds/main/ip/merlin/altera_merlin_std_arbitrator/altera_merlin_std_arbitrator_core.sv#3 $
// $Revision: #3 $
// $Date: 2010/07/07 $
// $Author: jyeap $
/* -----------------------------------------------------------------------
Round-robin/fixed arbitration implementation.
Q: how do you find the least-significant set-bit in an n-bit binary number, X?
A: M = X & (~X + 1)
Example: X = 101000100
101000100 &
010111011 + 1 =
101000100 &
010111100 =
-----------
000000100
The method can be generalized to find the first set-bit
at a bit index no lower than bit-index N, simply by adding
2**N rather than 1.
Q: how does this relate to round-robin arbitration?
A:
Let X be the concatenation of all request signals.
Let the number to be added to X (hereafter called the
top_priority) initialize to 1, and be assigned from the
concatenation of the previous saved-grant, left-rotated
by one position, each time arbitration occurs. The
concatenation of grants is then M.
Problem: consider this case:
top_priority = 010000
request = 001001
~request + top_priority = 000110
next_grant = 000000 <- no one is granted!
There was no "set bit at a bit index no lower than bit-index 4", so
the result was 0.
We need to propagate the carry out from (~request + top_priority) to the LSB, so
that the sum becomes 000111, and next_grant is 000001. This operation could be
called a "circular add".
A bit of experimentation on the circular add reveals a significant amount of
delay in exiting and re-entering the carry chain - this will vary with device
family. Quartus also reports a combinational loop warning. Finally,
Modelsim 6.3g has trouble with the expression, evaluating it to 'X'. But
Modelsim _doesn't_ report a combinational loop!)
An alternate solution: concatenate the request vector with itself, and OR
corresponding bits from the top and bottom halves to determine next_grant.
Example:
top_priority = 010000
{request, request} = 001001 001001
{~request, ~request} + top_priority = 110111 000110
result of & operation = 000001 000000
next_grant = 000001
Notice that if request = 0, the sum operation will overflow, but we can ignore
this; the next_grant result is 0 (no one granted), as you might expect.
In the implementation, the last-granted value must be maintained as
a non-zero value - best probably simply not to update it when no requests
occur.
----------------------------------------------------------------------- */
`timescale 1 ns / 1 ns
module altera_merlin_arbitrator
#(
parameter NUM_REQUESTERS = 8,
// --------------------------------------
// Implemented schemes
// "round-robin"
// "fixed-priority"
// "no-arb"
// --------------------------------------
parameter SCHEME = "round-robin",
parameter PIPELINE = 0
)
(
input clk,
input reset,
// --------------------------------------
// Requests
// --------------------------------------
input [NUM_REQUESTERS-1:0] request,
// --------------------------------------
// Grants
// --------------------------------------
output [NUM_REQUESTERS-1:0] grant,
// --------------------------------------
// Control Signals
// --------------------------------------
input increment_top_priority,
input save_top_priority
);
// --------------------------------------
// Signals
// --------------------------------------
wire [NUM_REQUESTERS-1:0] top_priority;
reg [NUM_REQUESTERS-1:0] top_priority_reg;
reg [NUM_REQUESTERS-1:0] last_grant;
wire [2*NUM_REQUESTERS-1:0] result;
// --------------------------------------
// Scheme Selection
// --------------------------------------
generate
if (SCHEME == "round-robin" && NUM_REQUESTERS > 1) begin
assign top_priority = top_priority_reg;
end
else begin
// Fixed arbitration (or single-requester corner case)
assign top_priority = 1'b1;
end
endgenerate
// --------------------------------------
// Decision Logic
// --------------------------------------
altera_merlin_arb_adder
#(
.WIDTH (2 * NUM_REQUESTERS)
)
adder
(
.a ({ ~request, ~request }),
.b ({{NUM_REQUESTERS{1'b0}}, top_priority}),
.sum (result)
);
generate if (SCHEME == "no-arb") begin
// --------------------------------------
// No arbitration: just wire request directly to grant
// --------------------------------------
assign grant = request;
end else begin
// Do the math in double-vector domain
wire [2*NUM_REQUESTERS-1:0] grant_double_vector;
assign grant_double_vector = {request, request} & result;
// --------------------------------------
// Extract grant from the top and bottom halves
// of the double vector.
// --------------------------------------
assign grant =
grant_double_vector[NUM_REQUESTERS - 1 : 0] |
grant_double_vector[2 * NUM_REQUESTERS - 1 : NUM_REQUESTERS];
end
endgenerate
// --------------------------------------
// Left-rotate the last grant vector to create top_priority.
// --------------------------------------
always @(posedge clk or posedge reset) begin
if (reset) begin
top_priority_reg <= 1'b1;
end
else begin
if (PIPELINE) begin
if (increment_top_priority) begin
top_priority_reg <= (|request) ? {grant[NUM_REQUESTERS-2:0],
grant[NUM_REQUESTERS-1]} : top_priority_reg;
end
end else begin
if (increment_top_priority) begin
if (|request)
top_priority_reg <= { grant[NUM_REQUESTERS-2:0],
grant[NUM_REQUESTERS-1] };
else
top_priority_reg <= { top_priority_reg[NUM_REQUESTERS-2:0], top_priority_reg[NUM_REQUESTERS-1] };
end
else if (save_top_priority) begin
top_priority_reg <= grant;
end
end
end
end
endmodule
// ----------------------------------------------
// Adder for the standard arbitrator
// ----------------------------------------------
module altera_merlin_arb_adder
#(
parameter WIDTH = 8
)
(
input [WIDTH-1:0] a,
input [WIDTH-1:0] b,
output [WIDTH-1:0] sum
);
wire [WIDTH:0] sum_lint;
// ----------------------------------------------
// Benchmarks indicate that for small widths, the full
// adder has higher fmax because synthesis can merge
// it with the mux, allowing partial decisions to be
// made early.
//
// The magic number is 4 requesters, which means an
// 8 bit adder.
// ----------------------------------------------
genvar i;
generate if (WIDTH <= 8) begin : full_adder
wire cout[WIDTH-1:0];
assign sum[0] = (a[0] ^ b[0]);
assign cout[0] = (a[0] & b[0]);
for (i = 1; i < WIDTH; i = i+1) begin : arb
assign sum[i] = (a[i] ^ b[i]) ^ cout[i-1];
assign cout[i] = (a[i] & b[i]) | (cout[i-1] & (a[i] ^ b[i]));
end
end else begin : carry_chain
assign sum_lint = a + b;
assign sum = sum_lint[WIDTH-1:0];
end
endgenerate
endmodule