pluto_hdl_adi/library/data_offload
David Winter 6a5d2f76d5 data_offload: Fix oversized TX input transactions
Signed-off-by: David Winter <david.winter@analog.com>
2021-10-27 14:26:22 +03:00
..
docs data_offload: Update README and generic block design 2021-08-06 11:55:24 +03:00
Makefile Makefile: Fix misc makefiles from projects and library 2021-10-05 14:24:48 +03:00
README.md Update README.md 2021-10-05 11:59:51 +03:00
data_offload.v Revert "data_offload: Fix timing violation" 2021-10-08 11:42:05 +03:00
data_offload_constr.ttcl data_offload: Improve external synchronization 2021-09-08 11:58:01 +03:00
data_offload_control.v data_offload: Initial commit 2021-08-06 11:55:24 +03:00
data_offload_fsm.v data_offload: Fix oversized TX input transactions 2021-10-27 14:26:22 +03:00
data_offload_ip.tcl data_offload: Improve external synchronization 2021-09-08 11:58:01 +03:00
data_offload_regmap.v data_offload: Fix support for > 4 GiB of storage 2021-08-06 11:55:24 +03:00
data_offload_sv.ttcl data_offload: Initial commit 2021-08-06 11:55:24 +03:00

README.md

Data offload IP core

Description, general use cases

Data offload module for high-speed converters:

NOTE: This IP will always have a storage unit (internal or external to the FPGA) and is designed to handle high data rates. If your data paths will run in a lower data rate, and your intention is just to transfer the data to another clock domain or to adjust the bus width of the data path, you may want to check out the util_axis_fifo or util_axis_fifo_asym IPs.

The initialization and data transfer looks as follows:

  • in case of DAC, the DMA initialize the storage unit, after that the controller will push the data to the DAC interface in one-shot or cyclic way

  • in case of ADC, the DMA request a transfer, the controller will save the data into the storage unit, after that will push it to the DMA

  • BYPASS mode: simple streaming FIFO in case of clock rate or data width differences between source and sink interfaces (data rate MUST match in order to work); the BYPASS mode is used when an initially high rate path is downgraded to lower rates.

Table of content

Generic arhitecture

The main role of our data paths, is to stream data from point A to point B in a particular system. There are always a SOURCE and a DESTINATION point, which can be a device (ADC or DAC), a DMA (for system memory) or any other data processing IP.

In the context of Data Offload IP, we don't need to know who is the source and who is the destination. Both interface is a AXI4 Stream interface, which can be supported in both Xilinx's an Intel's architecture, and can be connected to any device core or DMA.

The storage unit is connected to the Data Offload controller via two FIFO interface. This way the same controller can be used for various storage solutions. (BRAM, URAM, external memory etc.)

Block diagram

Generic Block Diagram

Parameters

NAME TYPE DEFAULT DESCRIPTION
ID integer 0 Instance ID number
MEM_TYPE [ 0:0] 0 Define the used storage type: FPGA RAM - 0; external DDR - 1
MEM_SIZE [31:0] 1024 Define the size of the storage element
MEMC_UIF_DATA_WIDTH [ 0:0] 512 The valid data depends on the DDRx memory controller IP.
TX_OR_RXN_PATH [ 0:0] 1 If set TX path enabled, otherwise RX
SRC_DATA_WIDTH integer 64 The data width of the source interface
SRC_RAW_DATA_EN [ 0:0] 0 Enable if the data path does extend samples to 16 bits
SRC_ADDR_WIDTH integer 8 The address width of the source interface, should be defined relative to the MEM_SIZE (MEM_SIZE/SRC_DATA_WIDTH/8)
DST_ADDR_WIDTH integer 7 The address width of the source interface, should be defined relative to the MEM_SIZE (MEM_SIZE/DST_DATA_WIDTH/8)
DST_DATA_WIDTH integer 64 The data width of the destination interface
DST_RAW_DATA_EN [ 0:0] 0 Enable if the data path does extend samples to 16 bits
DST_CYCLIC_EN [ 0:0] 0 Enables CYCLIC mode for destinations like DAC
AUTO_BRINUP [ 0:0] 0 If enabled the IP runs automatically after bootup

Interfaces

AXI4 Lite Memory Mapped Slave (S_AXI4_LITE)

This interface is used to access the register map.

// interface clock -- system clock -- 100 MHz
input                   s_axi_aclk
// interface resetn -- synchronous reset active low
input                   s_axi_aresetn

/* write address channel */

// validates the address on the bus
input                   s_axi_awvalid
// write address
input       [15:0]      s_axi_awaddr
// protection type -- not used in the core
input       [ 2:0]      s_axi_awprot
// write ready, indicates that the slave can accept the address
output                  s_axi_awready

/* write data channel */

// validate the data on the bus
input                   s_axi_wvalid
// write data
input       [31:0]      s_axi_wdata
// write strobe, indicates which byte lanes to update
input       [ 3:0]      s_axi_wstrb
// write ready, indicates that the slave can accept the data
output                  s_axi_wready

/* write response channel */

// validates the write response of the slave
output                  s_axi_bvalid
// write response, indicate the status of the transfer
output      [ 1:0]      s_axi_bresp
// response ready, indicates that the master can accept the data
input                   s_axi_bready

/* read address channel */

// validates the address on the bus
input                   s_axi_arvalid
// read address
input       [15:0]      s_axi_araddr
// protection type -- not used in the core
input       [ 2:0]      s_axi_arprot
// read ready, indicates that the slave can accept the address
output                  s_axi_arready

/* read data channel */

// validate the data on the bus
output                  s_axi_rvalid
// read response, indicate the status of the transfer
output      [ 1:0]      s_axi_rresp
// read data drivers by the slave
output      [31:0]      s_axi_rdata
// read ready, indicates that the master can accept the data
input                   s_axi_rready

Supported data interfaces

NOTE: To simplify the design both the source and destination data interface is an AXI4 streaming interface. A FIFO write (ADC) interface can be treated as AXI4 stream where only the master controles the data rate (s_axis_ready is always asserted), and a FIFO read (DAC) interface can be treated as an AXI4 stream where only the slave controles the data rate. (m_axis_valid is always asserted).

AXI4 Stream interface (S_AXIS | M_AXIS)

  • The AXI Stream Slave (S_AXIS) interface is used to receive AXI stream from the transmit DMA or ADC device.

  • The AXI Stream Master (M_AXIS) interface is used to transmit AXI stream to receive DMA or DAC device

// NOTE: this reference is a master interface

// interface clock -- can be device/core clock or DMA clock
input                        m_axis_aclk
// interface resetn -- synchronous reset with the system clock
input                        m_axis_resetn
// indicates that the slave can accept a transfer in the current cycle (in case of an ADC core, this will control the stream)
input                        m_axis_ready
// indicates that the master is driving a valid transfer
output                       m_axis_valid
// primary payload
output [DATA_WIDTH-1:0]      m_axis_data
// indicates the boundary of a packet
output                       m_axis_last
// byte qualifier, we need this so we can have different DMA and device data widths
output [(DATA_WIDTH/8)-1:0]  m_axis_tkeep

NOTE: A packet will always be a full buffer. All the data beats going to be full beats (all the bytes of the bus are valid), except the last one. axis_last and axis_tkeep will be used to indicate a partial last beat. This information should be transferred from the source domain to the sink domain, so we can read back the data from memory correctly.

FIFO source and destination interface to the storage unit

This is non-blocking (no back-pressure) interface for the storage unit, having an address bus too, so an ad_mem module can be connected directly to controller IP.

// This is a FIFO source interface - it's clocked on the source clock (s_axis_aclk)
// Reset signal
output                         fifo_src_resetn
// write enable signal
output                         fifo_src_wen
// address bus for internal memory
output   [SRC_ADDR_WIDTH-1:0]  fifo_src_waddr
// source data bus
output   [SRC_DATA_WIDTH-1:0]  fifo_src_wdata
// write last, indicates the last valid transfer
output                         fifo_src_wlast
// This is a FIFO destination interface - it's clocked on the destination clock (m_axis_aclk)
// Reset signal
output                         fifo_dst_resetn
// read enable signal
output                         fifo_dst_ren
// indicate if the storage is ready to accept read requests
output                         fifo_dst_ready,
// address bus for internal memory
output   [DST_ADDR_WIDTH-1:0]  fifo_dst_raddr
// destination data bus
output   [DST_DATA_WIDTH-1:0]  fifo_dst_rdata
// This is a Slave FIFO Read interface
// device digital interface clock, or core clock
input                     fifo_rd_clk
// enables the channel --  in our case this is redundant -- maybe we do neet to use it at all
input                     fifo_rd_enable
// validates the data on the bus, it's driven by the device indicates when the core latches the data
input                     fifo_rd_valid
// primary payload, its data width is equal with the channel's data width
output  [DATA_WIDTH-1:0]  fifo_rd_data
// indicates an underflow, the source (offload FIFO in this case) can not produce the data fast enough
output                    fifo_rd_unf

Initialization request interface

Define a simple request/acknowledge interface to initialize the memory:

  • The request will comes from the system and will put the data offload FSM into a standby/ready state.

  • Both RX and TX path should have a separate initialization request interface.

  • Acknowledge will be asserted by the data offload IP as the FSM is ready to receive data. (from TX_DMA or ADC)

  • In case of ADC: after the acknowledge samples will be stored into the memory using one of the SYNC modes.

  • In case of the DAC: after acknowledge data from the DMA will be stored into the memory. Acknowledge will stay asserted until one of the SYNC mode is used, after that the source interface of the IP will stay in busy state. (all the DMA transfers will be blocked)

Synchronization modes

  • AUTOMATIC

    • ADC: As the acknowledge of the initialization interface is asserted, the IP will start to fill up the buffer with samples.
    • DAC: As the DMA will send a valid last, the FSM will start to send the stored data to the device.
  • HARDWARE

    • ADC and DAC: An external signal will trigger the write or read into or from the memory.
    • NOTE: In case of DAC, if the DMA does not sent all the data into the buffer, before a hardware sync event, the unsent data will be ignored. It's the user/software responsibility to sync up these events accordingly.
  • SOFTWARE

    • The software write a RW1C register which will trigger the reads or writes into or from the memory.

Register Map

WORD BYTE BITS NAME TYPE CLOCK DOMAIN DESCRIPTION
0x0000 0x0000 VERSION RO SYS Version number
[31:16] MAJOR
[15: 8] MINOR
[ 7: 0] PATCH
0x0001 0x0004 PERIPHERAL_ID RO SYS Value of the IP configuration parameter
0x0002 0x0008 SCRATCH RW SYS Scratch register
0x0003 0x000C IDENTIFICATION RO SYS Peripheral identification. Default value: 0x44414F46 - ('D','A','O','F')
0x0004 0x0010 SYNTHESIS_CONFIG RO SYS Core configuration registers
[ 1: 1] TX_OR_RXN_PATH RX Path => 0, TX => 1
[ 0: 0] MEMORY_TYPE The used storage type (embedded => 0 or external => 1)
0x0005 0x0014 MEMORY_SIZE_LSB RO SYS 32bits LSB of the memory size register
0x0006 0x0018 MEMORY_SIZE_MSB RO SYS 2bits MSB of the memory size register
[ 1: 0] MEMORY_SIZE_MSB
0x0007 0x001C TRANSFER_LENGTH RW SRC Transfer length
0x0020 0x0080 MEM_PHY_STATE RO DDR Status bits of the memory controller IP
[ 0: 0] CALIB_COMPLETE Indicates that the memory initialization and calibration have completed successfully
0x0021 0x0084 RESETN_OFFLOAD RW DST/SRC Reset all the internal address registers and state machines
[ 0: 0] RESETN
0x0022 0x0088 CONTROL RW DST A global control register
[ 1: 1] ONESHOT_EN By default the TX path runs on CYCLIC mode, set this bit to switch it to ONE-SHOT mode
[ 0: 0] OFFLOAD_BYPASS Bypass the offload storage, the data path consist just of a CDC FIFO
0x0040 0x0100 SYNC_TRIGGER RW1C SRC Synchronization setup for RX and TX path
[ 0: 0] SYNC_TRIGGER Trigger the data capture
0x0041 0x0104 SYNC_CONFIG RW SRC Synchronization setup
[ 1: 0] SYNC_CONFIG Auto - '0'; hardware - '1'; software - '2'
0x0080 0x0200 FSM_DBG RO Debug register for the offload FSM
[ 5: 4] FSM_STATE_READ SRC The current state of the read-offload state machine
[ 1: 0] FSM_STATE_WRITE DST The current state of the write-offload state machine
0x0081 0x0204 SAMPLE_COUNT_LSB RO SRC Stored sample count (32 LSB)
0x0082 0x0208 SAMPLE_COUNT_MSB RO SRC Stored sample count (32 MSB)

Clock tree

In general there are at least two different clock in the data offload module:

  • DMA or system clock : on this clock will run all the front end interfaces
  • Memory Controller user clock : user interface clock of the DDRx controller (optional)
  • Device clock : the digital interface clock of the converter

Clocks

A general frequency relationship of the above clocks are:

  CLKdma <= CLKddr <= CLKconverter

The clock domain crossing should be handled by the util_axis_fifo module.

  • TODO : Make sure that we support both AXIS and FIFO
  • TODO : Add support for asymmetric aspect ratio.

All the back end paths (device side) are time critical. The module must read or write from or into the storage at the speed of the device.

  DDR data rate >= Device data rate
  DDR data rate >= ADC data rate + DAC data rate

Data path

Data path

  • The data path should be designed to support any kind of difference between the source, memory and sink data width.

  • The data width adjustments will be made by the CDC_FIFO.

  • In both path (ADC and DAC) the data stream at the front-end side is packatized, meaning there is a valid TLAS/TKEEP in the stream. While in the back-end side the stream is continuous. (no TLAST/TKEEP)

    • The DAC path have to have a depacketizer to get rid of the last partial beat from the stream.
    • Because the ADC path already arrive in a packed form, and we always will fill up the whole storage, we don't need to treat special use-cases.

Used storage elements

ZC706 ZCU102 A10SOC
FPGA XC7Z045 FFG900 2 XCZU9EG-2FFVB1156 10AS066N3F40E2SG
External Memory Type DDR3 SODIMM DDR4 DDR4 HILO
External Memory Size 1 GB 512 MB 2 GB
Embedded Memory Type BRAM BRAM M20K
Embedded Memory Size 19.1 Mb 32.1 Mb 41 Mb

Data width manipulation

  • data width differences should be treated by the CDC FIFO

  • the smallest granularity should be 8 bits. This constraints mainly will generate additional logic just in the TX path, taking the fact that the data from the ADC will come packed.

  • the gearbox main role is to improve the DDR's bandwidth, strips the padding bits of each samples, so the raw data could be stored into the memory.

Xilinx's MIG vs. Intel's EMIF

  • Incrementing burst support for 1 to 256 beats, the length of the burst should be defined by the internal controller

  • Concurrent read/write access, the external memory to be shared between an ADC and DAC

  • Dynamic burst length tuning: an FSM reads and writes dummy data until both ADC's overflow and DAC's underflow lines are de-asserted. Pre-requisites : both device's interface should be up and running.

  • TODO: prefetch the next transfer if it's possible, by driving the address channels ahead (e.g. Overlapping read burst in case of AXI4)

  • Optional gearbox to congest the samples in order to increase the maximum data rate.

  • In general we packing all samples into 16 bits. This can add a significant overhand to the maximum real data rate on the memory interface. The gearbox main role is to pack and unpack the device's samples into the required data width. (in general 512 or 1024 bit)

Boards with FPGA side DDR3/4 SODIMMs/HILO: ZC706, ZCU102, A10SOC

ZC706 ZCU102 A10SOC
Max data throughputs (MT/s) 1600 2400 2133
DDRx reference clocks 200 MHz 300 MHz 133 MHz
DDRx Data bus width 64 16 64
Memory to FPGA clock ratio 4:1 4:1 4:1
UI type & burst length AXI4-256 AXI4-256 Avalon Memory Map
UI data width 512 128 512

Internal cyclic buffer support for the TX path

Data path with external storage

  • On the front end side if the TX path, a special buffer will handle the data width up/down conversions and run in cyclic mode if the length of the data set is smaller than 4/8 AXI/Avalon burst. This way we can avoid to overload the memory interface with small bursts.

  • On the back end side, because the smallest granularity can be 8 bytes, we need a dynamic 'depackatizer' or re-aligner, which will filter out the invalid data bytes from the data stream. (this module will use the tlast and tkeep signal of the AXI stream interface)

Control path - Offload FSM

RX control FSM for internal RAM mode

RX_control FMS for internal RAM mode

TX control FSM for internal RAM mode

TX_control FMS for internal RAM mode

TODO FSMs for the external DDR mode

References

AMBA AXI

Avalon

Xilinx

Intel

Supported FPGA boards