Files
b210-k7-fpga/lib/rfnoc/blocks/rfnoc_block_fft/fft_reorder.sv
T
Wade Fife 6b19ec030c fpga: rfnoc: fft: Support multiple samples per cycle
This adds the NIPC parameter, which configures support for processing
multiple items or samples per clock cycle. With this enabled, the FFT
block can process at rates higher than 250 MSPS, such as 500 MSPS and
beyond.


Original-commit: fc76aa940e121fe1f85a3513f6d90df4667338cf
2025-03-07 12:39:34 -06:00

623 lines
22 KiB
Systemverilog

//
// Copyright 2024 Ettus Research, a National Instruments Brand
//
// SPDX-License-Identifier: LGPL-3.0-or-later
//
// Module: fft_reorder
//
// Description:
//
// This module optionally rearranges the order of FFT bins to put them in the
// desired order. It also supports cyclic prefix insertion.
//
// The input order that this module receives is a parameter that must be
// chosen at compile time. The following input orders are supported:
//
// NATURAL: Positive frequencies are input first, starting with 0 Hz,
// followed by negative frequencies. Frequencies are input in
// ascending order.
// BIT_REVERSE: Like natural, but the bits of the indices are in reverse
// order. For example, for a size 16 FFT, bin 0000 is input
// first, followed by bin 1000, 0100, 1100, 0010, etc.
//
// The output order can be chosen at run time. The following output orders
// are supported:
//
// NORMAL: Negative frequencies first, then positive frequencies. 0 Hz
// is in the center. Frequencies are output in ascending order.
// REVERSE: Reverse order of NORMAL. Positive frequencies first, then
// negative frequencies. 0 Hz in the center. Frequencies are
// output in descending order.
// NATURAL: Positive frequencies are first, starting with 0 Hz,
// followed by negative frequencies. Frequencies are output in
// ascending order.
// BIT_REVERSE: Like natural, but the bits of the indices are in reverse
// order. For example, for a size 16 FFT, bin 0000 is output
// first, followed by bin 1000, 0100, 1100, 0010, etc.
//
// Typically the FFT IP feeding this module will output data in BIT_REVERSE
// order. The FFT IP may have an option to rearrange the data into NATURAL
// order but enabling this feature causes a large memory to be added to the
// IP to do the reordering. Since we want to also be able to provide NORMAL
// order, and we don't want to add a second memory for that reordering, we do
// all the reordering here in one memory.
//
// If the FFT core is outputting in the order you want, then this module
// should probably be removed to save RAM and logic.
//
// The TLAST input/output corresponds to when the FFT input/output ends for a
// single FFT-sized sequence of data. i_tlast must be asserted during the
// last transfer of the input FFT to reset things for the next FFT input.
//
// For cyclic prefix insertion, the EN_CP_INSERTION parameter must be true
// and i_tuser contains the cyclic prefix size to insert. It must be valid
// during the first transfer of the packet. It can be any size from 0 to
// 2**MAX_FFT_LEN_LOG2-1.
//
// Parameters:
//
// IN_FIFO_LOG2 : Log base-2 of the input FIFO size. Set to -1 to remove
// the input FIFO. This FIFO is intended as a pipeline
// stage to cut the timing path on the input.
// OUT_FIFO_LOG2 : Log base-2 of the output FIFO size. This must be set to
// at least 3.
// INPUT_ORDER : BIT_REVERSE or NATURAL. See fft_reorder_pkg for values.
// MAX_FFT_LEN_LOG2 : Ceiling of log base-2 of the maximum FFT size to be
// supported.
// DATA_W : Data width. Typically 32 for sc16 data type.
// EN_CP_INSERTION : Controls whether or not the CP insertion logic is
// included.
//
// Signals:
//
// i_t* : AXI-Stream data input. Each packet is one FFT to be processed. The
// length of the packet must match the FFT size. i_tuser contains the
// cyclic prefix size to insert for this packet and must be valid
// during the first transfer of the packet.
// o_t* : AXI-Stream data output. Each packet is one FFT with optional cyclic
// prefix.
//
`default_nettype none
module fft_reorder
import fft_reorder_pkg::*;
#(
parameter int IN_FIFO_LOG2 = 1,
parameter int OUT_FIFO_LOG2 = 3,
parameter fft_order_t INPUT_ORDER = BIT_REVERSE,
parameter int MAX_FFT_LEN_LOG2 = 12,
parameter int DATA_W = 32,
parameter bit EN_CP_INSERTION = 1,
localparam int FFT_LEN_LOG2_W = $clog2(MAX_FFT_LEN_LOG2+1),
localparam int CP_LEN_W = MAX_FFT_LEN_LOG2
) (
input wire clk,
input wire rst,
input wire fft_cfg_wr,
input wire [FFT_LEN_LOG2_W-1:0] fft_len_log2,
input fft_order_t fft_out_order,
// Data Input
input wire [ DATA_W-1:0] i_tdata,
input wire [CP_LEN_W-1:0] i_tuser,
input wire i_tlast,
input wire i_tvalid,
output wire i_tready,
// Data Output
output wire [DATA_W-1:0] o_tdata,
output wire o_tlast,
output wire o_tvalid,
input wire o_tready
);
// These registers track if the current read/write buffers are OK to use
logic ok_to_write = 1'b1; // Current write buffer is free for writes
logic ok_to_read = 1'b0; // Current read buffer has data to read
//---------------------------------------------------------------------------
// Optional Data Input Pipeline
//---------------------------------------------------------------------------
logic [ DATA_W-1:0] in_fifo_o_tdata;
logic [CP_LEN_W-1:0] in_fifo_o_tuser;
logic in_fifo_o_tvalid;
logic in_fifo_o_tready;
logic in_fifo_o_tlast;
if (IN_FIFO_LOG2 >= 0) begin : gen_in_fifo
axi_fifo #(
.WIDTH(1 + CP_LEN_W + DATA_W),
.SIZE (IN_FIFO_LOG2)
) axi_fifo_in (
.clk (clk),
.reset (rst),
.clear ('0),
.i_tdata ({i_tlast, i_tuser, i_tdata}),
.i_tvalid(i_tvalid),
.i_tready(i_tready),
.o_tdata ({in_fifo_o_tlast, in_fifo_o_tuser, in_fifo_o_tdata}),
.o_tvalid(in_fifo_o_tvalid),
.o_tready(in_fifo_o_tready),
.space (),
.occupied()
);
end else begin : gen_no_in_fifo
assign in_fifo_o_tdata = i_tdata;
assign in_fifo_o_tuser = i_tuser;
assign in_fifo_o_tlast = i_tlast;
assign in_fifo_o_tvalid = i_tvalid;
assign i_tready = in_fifo_o_tready;
end
//---------------------------------------------------------------------------
// Optional Data Output Pipeline
//---------------------------------------------------------------------------
if (OUT_FIFO_LOG2 < 3) begin
OUT_FIFO_LOG2_must_be_at_least_3();
end
logic [DATA_W-1:0] out_fifo_i_tdata;
logic out_fifo_i_tvalid;
logic out_fifo_i_tlast;
// We use out_fifo_space instead of out_fifo_i_tready to allow extra space
// for the RAM output read delay.
logic [15:0] out_fifo_space;
axi_fifo #(
.WIDTH(DATA_W+1),
.SIZE (OUT_FIFO_LOG2)
) axi_fifo_out (
.clk (clk),
.reset (rst),
.clear ('0),
.i_tdata ({out_fifo_i_tlast, out_fifo_i_tdata}),
.i_tvalid(out_fifo_i_tvalid),
.i_tready(),
.o_tdata ({o_tlast, o_tdata}),
.o_tvalid(o_tvalid),
.o_tready(o_tready),
.space (out_fifo_space),
.occupied()
);
//---------------------------------------------------------------------------
// Configuration Registers
//---------------------------------------------------------------------------
//
// Store relevant FFT configuration values in registers for use elsewhere. We
// assume that the configuration is set in advance of any operation and is
// only changed when the FFT is idle, so we ignore the latency here.
//
//---------------------------------------------------------------------------
// Number of bits needed to represent the maximum FFT size
localparam FFT_LEN_W = MAX_FFT_LEN_LOG2+1;
logic fft_cfg_wr_stb = 1'b0;
fft_order_t fft_out_order_reg = NORMAL;
logic [FFT_LEN_LOG2_W-1:0] fft_len_log2_reg = MAX_FFT_LEN_LOG2;
logic [FFT_LEN_W-1:0] fft_len = 1 << MAX_FFT_LEN_LOG2;
logic [FFT_LEN_W-1:0] fft_len_m1 = (1 << MAX_FFT_LEN_LOG2)-1;
always_ff @(posedge clk) begin
if(rst) begin
fft_cfg_wr_stb <= 1'b0;
fft_out_order_reg <= NORMAL;
fft_len_log2_reg <= MAX_FFT_LEN_LOG2;
fft_len <= 1 << MAX_FFT_LEN_LOG2;
fft_len_m1 <= (1 << MAX_FFT_LEN_LOG2)-1;
end else begin
fft_cfg_wr_stb <= 1'b0;
if (fft_cfg_wr) begin
fft_cfg_wr_stb <= 1'b1;
fft_out_order_reg <= fft_out_order;
fft_len_log2_reg <= fft_len_log2;
fft_len <= (1 << fft_len_log2);
fft_len_m1 <= (1 << fft_len_log2)-1;
end
end
end
//---------------------------------------------------------------------------
// RAM Buffer
//---------------------------------------------------------------------------
//
// This RAM stores the data that's being input, writing it the order needed
// such that when read out sequentially, it will be in the correct order.
//
// The RAM is divided into two halves, which we'll call buffers. Each buffer
// is used exclusively for read or write, until they switch.
//
//---------------------------------------------------------------------------
// Address width for each buffer. Must be big enough to store the maximum
// length FFT.
localparam ADDR_W = MAX_FFT_LEN_LOG2;
// RAM read latency
localparam READ_LATENCY = 2;
logic ram_rd_buffer; // Indicates which buffer is currently used for reads
logic ram_wr_buffer; // Indicates which buffer is currently used for writes
logic ram_wr_en;
logic ram_wr_en_0; // One RAM read enable for each buffer
logic ram_wr_en_1;
logic [ADDR_W-1:0] ram_wr_addr;
logic [DATA_W-1:0] ram_wr_data;
logic ram_rd_en;
logic [ADDR_W-1:0] ram_rd_addr;
logic [DATA_W-1:0] ram_rd_data_raw_0; // One RAM read output for each buffer
logic [DATA_W-1:0] ram_rd_data_raw_1;
ram_2port #(
.DWIDTH (DATA_W),
.AWIDTH (ADDR_W), // Make the RAM two buffers big
.OUT_REG(1)
) ram_2port_0 (
.clka (clk),
.ena ('1),
.wea (ram_wr_en_0),
.addra(ram_wr_addr),
.dia (ram_wr_data),
.doa (),
.clkb (clk),
.enb ('1),
.web ('0),
.addrb(ram_rd_addr),
.dib ('0),
.dob (ram_rd_data_raw_0)
);
ram_2port #(
.DWIDTH (DATA_W),
.AWIDTH (ADDR_W), // Make the RAM two buffers big
.OUT_REG(1)
) ram_2port_1 (
.clka (clk),
.ena ('1),
.wea (ram_wr_en_1),
.addra(ram_wr_addr),
.dia (ram_wr_data),
.doa (),
.clkb (clk),
.enb ('1),
.web ('0),
.addrb(ram_rd_addr),
.dib ('0),
.dob (ram_rd_data_raw_1)
);
//---------------------------------------------------------------------------
// Write Logic
//---------------------------------------------------------------------------
//
// Here we write the data into the memory in a carefully controlled order
// such that we can read it out in sequential or bit-reversed order to get
// the order we want.
//
//---------------------------------------------------------------------------
logic [FFT_LEN_W-1:0] fft_addr_mask;
logic [FFT_LEN_W-1:0] wr_count;
logic ram_wr_last;
assign ram_wr_data = in_fifo_o_tdata;
assign ram_wr_en = in_fifo_o_tvalid && in_fifo_o_tready;
assign ram_wr_en_0 = ram_wr_en && (ram_wr_buffer == 1'b0);
assign ram_wr_en_1 = ram_wr_en && (ram_wr_buffer == 1'b1);
assign in_fifo_o_tready = ok_to_write;
assign ram_wr_last = in_fifo_o_tlast;
always_ff @(posedge clk) begin
if (fft_cfg_wr_stb || (ram_wr_en && ram_wr_last)) begin
if (fft_out_order_reg == NATURAL) begin
// Natural to natural. No mask needed to affect the order.
fft_addr_mask <= '0;
ram_wr_addr <= '0;
end else if (fft_out_order_reg == REVERSE) begin
// Natural to reverse. Invert all bits except the MSB. Inverting the
// lower bits reverses the order. Leaving the MSB unchanged ensures we
// output positive frequencies first, then negative frequencies.
fft_addr_mask <= fft_len_m1 >> 1; // e.g., 8'b0111_1111
ram_wr_addr <= fft_len_m1 >> 1;
end else if (fft_out_order_reg == NORMAL) begin
// Natural to normal. Invert the MSB, so that we output negative
// frequencies first, then positive frequencies.
fft_addr_mask <= fft_len >> 1; // e.g., 8'b1000_0000
ram_wr_addr <= fft_len >> 1;
end else begin // (fft_order_t == BIT_REVERSE)
// Natural to bit-reverse. For this we also use natural order, and we
// enable/disable the bit-reversal on the read side as needed.
fft_addr_mask <= '0;
ram_wr_addr <= '0;
end
end
if (ram_wr_en) begin
wr_count <= wr_count+1;
if (ram_wr_last) begin
// Switch to the other buffer
ram_wr_buffer <= ~ram_wr_buffer;
wr_count <= '0;
end else begin
// Calculate the the next write address
if (
(INPUT_ORDER == BIT_REVERSE && fft_out_order_reg != BIT_REVERSE) ||
(INPUT_ORDER == NATURAL && fft_out_order_reg == BIT_REVERSE)
) begin : bit_reversed
// If the input is bit-reversed and we're not outputting
// bit-reversed, then we bit reverse the RAM address to convert from
// bit-reversed to natural order. Then apply the mask to that to
// convert from natural to the desired output order.
ram_wr_addr <= bit_reverse(wr_count+1, fft_len_log2_reg) ^ fft_addr_mask;
end else begin : natural
// Apply the mask to convert from natural to to the desired output
// order.
ram_wr_addr <= (wr_count+1) ^ fft_addr_mask;
end
end
end
if (rst) begin
ram_wr_buffer <= '0;
ram_wr_addr <= '0;
wr_count <= '0;
end
end
//---------------------------------------------------------------------------
// CP Insertion Length FIFO
//---------------------------------------------------------------------------
// Cyclic prefix logic interface signals
logic cp_valid; // Indicates the CP FIFO has an output
logic cp_non_zero; // Indicates the CP value is > 0
logic [ADDR_W-1:0] cp_start_addr; // Indicates the CP RAM start address
logic cp_consume; // Control to indicate we've captured the CP length output
if (EN_CP_INSERTION) begin: gen_cp_ins_fifo
logic [CP_LEN_W-1:0] cp_len_tdata;
logic cp_len_tvalid;
logic tmp_i_tvalid;
logic in_fifo_o_tfirst = '1; // First transfer of packet
// Create a register that indicates when the next transfer is the start of
// a new packet.
always_ff @(posedge clk) begin
if (rst) begin
in_fifo_o_tfirst <= '1;
end else begin
if (in_fifo_o_tvalid && in_fifo_o_tready) begin
in_fifo_o_tfirst <= in_fifo_o_tlast;
end
end
end
// Write the first tuser word of the packet into the CP length FIFO
assign tmp_i_tvalid = in_fifo_o_tvalid && in_fifo_o_tready && in_fifo_o_tfirst;
// The dual RAM buffer can only hold two FFTs at a time, so we can
// guarantee this FIFO has sufficient room and will always be ready by
// setting its size appropriately.
axi_fifo #(
.WIDTH(CP_LEN_W),
.SIZE (1)
) axi_fifo_cp_length (
.clk (clk),
.reset (rst),
.clear ('0),
.i_tdata (in_fifo_o_tuser),
.i_tvalid(tmp_i_tvalid),
.i_tready(),
.o_tdata (cp_len_tdata),
.o_tvalid(cp_len_tvalid),
.o_tready(cp_consume),
.space (),
.occupied()
);
// Add a register to calculate the cyclic prefix start read address and
// figure out if we need to do a cyclic prefix insertion. The latency of
// this register will be much less than the FFT write time.
always_ff @(posedge clk) begin
cp_valid <= cp_len_tvalid;
cp_non_zero <= (cp_len_tdata != 0);
cp_start_addr <= fft_len - cp_len_tdata;
end
end else begin : gen_no_cp_ins_fifo
assign cp_valid = '0;
assign cp_non_zero = '0;
assign cp_start_addr = '0;
end
//---------------------------------------------------------------------------
// Read Logic
//---------------------------------------------------------------------------
typedef enum logic [1:0] { READ_CHECK, READ_CP, READ_FFT} read_state_t;
read_state_t read_state = EN_CP_INSERTION ? READ_CHECK : READ_FFT;
read_state_t read_state_nx;
logic [ADDR_W-1:0] ram_rd_addr_nx;
logic ram_rd_buffer_nx;
logic ram_rd_last; // Indicates when ram_rd_en asserts for the last sample
logic out_fifo_avail;
// Delayed versions of read signals to align with read output timing
logic [READ_LATENCY-1:0] ram_rd_buffer_del;
logic [READ_LATENCY-1:0] ram_rd_en_del;
logic [READ_LATENCY-1:0] ram_rd_last_del;
logic [DATA_W-1:0] ram_rd_data;
logic ram_rd_data_valid; // Indicates ram_rd_data has data
logic ram_rd_data_last; // Indicates ram_rd_data is the last of the FFT
assign out_fifo_i_tdata = ram_rd_data;
assign out_fifo_i_tvalid = ram_rd_data_valid;
assign out_fifo_i_tlast = ram_rd_data_last;
always_ff @(posedge clk) begin : read_fsm_reg
if (rst) begin
read_state <= EN_CP_INSERTION ? READ_CHECK : READ_FFT;
ram_rd_buffer <= '0;
ram_rd_addr <= '0;
ram_rd_buffer_del <= '0;
ram_rd_en_del <= '0;
ram_rd_last_del <= '0;
ram_rd_data <= 'X;
ram_rd_data_valid <= '0;
ram_rd_data_last <= '0;
out_fifo_avail <= '0;
end else begin
read_state <= read_state_nx;
ram_rd_buffer <= ram_rd_buffer_nx;
ram_rd_addr <= ram_rd_addr_nx;
// Pipeline the buffer selection, enable, and last to align with RAM output
ram_rd_buffer_del <= (ram_rd_buffer_del << 1) | ram_rd_buffer;
ram_rd_en_del <= (ram_rd_en_del << 1) | ram_rd_en;
ram_rd_last_del <= (ram_rd_last_del << 1) | ram_rd_last;
// Select the RAM output that was used for the read
ram_rd_data <= ram_rd_buffer_del[READ_LATENCY-1] ?
ram_rd_data_raw_1 : ram_rd_data_raw_0;
ram_rd_data_valid <= ram_rd_en_del[READ_LATENCY-1];
ram_rd_data_last <= ram_rd_last_del[READ_LATENCY-1];
// Ensure there's enough room in the output FIFO to account for the
// latency through the read logic.
out_fifo_avail <= out_fifo_space > 4;
end
end
always_comb begin : read_fsm_comb
ram_rd_en = '0;
ram_rd_last = '0;
ram_rd_buffer_nx = ram_rd_buffer;
ram_rd_addr_nx = ram_rd_addr;
read_state_nx = read_state;
cp_consume = '0;
case (read_state)
READ_CHECK : begin
// Wait until the next cyclic prefix is available and update the RAM
// read address appropriately.
if (cp_valid) begin
cp_consume = '1;
if (cp_non_zero) begin
read_state_nx = READ_CP;
ram_rd_addr_nx = cp_start_addr;
end else begin
read_state_nx = READ_FFT;
ram_rd_addr_nx = '0;
end
end
end
READ_CP : begin
// Read out the cyclic prefix
ram_rd_en = (ok_to_read && out_fifo_avail);
if (ram_rd_en) begin
if (ram_rd_addr == fft_len_m1) begin
ram_rd_addr_nx = '0;
read_state_nx = READ_FFT;
end else begin
ram_rd_addr_nx = ram_rd_addr + 1;
end
end
end
default : begin // READ_FFT
// Read out the whole FFT
ram_rd_en = (ok_to_read && out_fifo_avail);
if (ram_rd_en) begin
if (ram_rd_addr == fft_len_m1) begin
ram_rd_last = '1;
ram_rd_addr_nx = '0;
ram_rd_buffer_nx = ~ram_rd_buffer;
read_state_nx = EN_CP_INSERTION ? READ_CHECK : READ_FFT;
end else begin
ram_rd_addr_nx = ram_rd_addr + 1;
end
end
end
endcase
end
//---------------------------------------------------------------------------
// Read/Write Arbitration Logic
//---------------------------------------------------------------------------
//
// Here we ensure that we only write when the write buffer is free and that
// we only read when the read buffer has an FFT in it. Because we're reading
// and writing simultaneously, we swap between the lower and upper parts of
// the RAM as data gets written and read out.
//
//---------------------------------------------------------------------------
always_ff @(posedge clk) begin
if (ram_wr_en && ram_rd_en) begin
if (ram_wr_last && ram_rd_last) begin
// Both buffers are switching on the same cycle
ok_to_write <= 1'b1;
ok_to_read <= 1'b1;
end else if (ram_wr_last) begin
// Switching write buffer to the one being used for reads
ok_to_write <= 1'b0;
end else if (ram_rd_last) begin
// Switching read buffer to the one being used for writes
ok_to_read <= 1'b0;
end
end else if (ram_wr_en && ram_wr_last) begin
// Write buffer is switching
if (ram_wr_buffer == ram_rd_buffer) begin
// Write buffer is switching away from the current read buffer
ok_to_write <= 1'b1;
ok_to_read <= 1'b1;
end else begin
// Write buffer is switching to the current read buffer
ok_to_write <= 1'b0;
end
end else if (ram_rd_en && ram_rd_last) begin
// Read buffer is switching
if (ram_wr_buffer == ram_rd_buffer) begin
// Read buffer is switching away from the current write buffer
ok_to_write <= 1'b1;
ok_to_read <= 1'b1;
end else begin
// Read buffer is switching to the current write buffer
ok_to_read <= 1'b0;
end
end
//synthesis translate_off
if (ram_wr_en && ram_rd_en && (ram_wr_buffer == ram_rd_buffer)) begin
$error("Attempt to read and write the same buffer!");
end
//synthesis translate_on
if (rst) begin
ok_to_write <= 1'b1; // Buffers empty after reset
ok_to_read <= 1'b0; // Can't read until we fill the first buffer
end
end
endmodule
`default_nettype wire