External Memory Interface Handbook Volume 3: Reference Material: For UniPHY-based Device Families

ID 683841
Date 3/06/2023
Public
Document Table of Contents

4.6.2. MPFE Multi-Port Arbitration

The HPS SDRAM controller multi-port front end (MPFE) contains a programmable arbiter. The arbiter decides which MPFE port gains access to the single-port memory controller.

The SDRAM transaction size that is arbitrated is a burst of two beats. This burst size ensures that the arbiter does not favor one port over another when the incoming transaction size is a large burst.

The arbiter makes decisions based on two criteria: priority and weight. The priority is an absolute arbitration priority where the higher priority ports always win arbitration over the lower priority ports. Because multiple ports can be set to the same priority, the weight value refines the port choice by implementing a round-robin arbitration among ports set to the same priority. This programmable weight allows you to assign a higher arbitration value to a port in comparison to others such that the highest weighted port receives more transaction bandwidth than the lower weighted ports of the same priority.

Before arbitration is performed, the MPFE buffers are checked for any incoming transactions. The priority of each port that has buffered transactions is compared and the highest priority wins. If multiple ports are of the same highest priority value, the port weight is applied to determine which port wins. Because the arbiter only allows SDRAM-sized bursts into the single-port memory controller, large transactions may need to be serviced multiple times before the read or write command is fully accepted to the single-port memory controller. The MPFE supports dynamic tuning of the priority and weight settings for each port, with changes committed into the SDRAM controller at fixed intervals of time.

Arbitration settings are applied to each port of the MPFE. The memory controller supports a mix of Avalon-MM and AXI protocols. As defined in the "Port Mappings" section, the Avalon-MM ports consume a single command port while the AXI ports consume a pair of command ports to support simultaneous read and write transactions. In total, there are ten command ports for the MPFE to arbitrate. The following table illustrates the command port mapping within the HPS as well as the ports exposed to the FPGA fabric.

Table 29.  HPS SDRAM MPFE Command Port Mapping

Command Port

Allowed Functions

Data Size

0, 2, 4

FPGA fabric AXI read command ports

FPGA fabric Avalon-MM read or write command ports

32-bit to 256-bit data

1, 3, 5

FPGA fabric AXI write command ports

FPGA fabric Avalon-MM read or write command ports

6

L3 AXI read command port

32-bit data

7

MPU AXI read command port

64-bit data

8

L3 AXI write command port

32-bit data

9

MPU AXI write command port

64-bit data

When the FPGA ports are configured for AXI, the command ports are always assigned in groups of two starting with even number ports 0, 2, or 4 assigned to the read command channel. For example, if you configure the first FPGA-to-SDRAM port as AXI and the second port as Avalon-MM, you can expect the following mapping:

  • Command port 0 = AXI read
  • Command port 1 = AXI write
  • Command port 2 = Avalon-MM read and write

Setting the MPFE Priority

The priority of each of the ten command ports is configured through the userpriority field of the mppriority register. This 30-bit register uses 3 bits per port to configure the priority. The lowest priority is 0x0 and the highest priority is 0x7. The bits are mapped in ascending order with bits [2:0] assigned to command port 0 and bits [29:27] assigned to command port 9.

Setting the MPFE Static Weights

The static weight settings used in the round-robin command port priority scheme are programmed in a 128-bit field distributed among four 32-bit registers:

  • mpweight_0_4
  • mpweight_1_4
  • mpweight_2_4
  • mpweight_3_4

Each port is assigned a 5-bit value within the 128-bit field, such that port 0 is assigned to bits [4:0] of the mpweight_0_4 register, port 1 is assigned to bits [9:5] of the mpweight_0_4 register up to port 9, which is assigned to bits[49:45] contained in the mpweight_1_4 register. The valid weight range for each port is 0x0 to 0x1F, with larger static weights representing a larger arbitration share.

Bits[113:50] in the mpweight_1_4, mpweight_2_4 and mpweight_3_4 registers, hold the sum of weights for each priority. This 64-bit field is divided into eight fields of 8-bits, each representing the sum of static weights. Bits[113:50] are mapped in ascending order with bits [57:50] holding the sum of static weights for all ports with priority setting 0x0, and bits [113:106] holding the sum of static weights for all ports with priority setting 0x7.

Example Using MPFE Priority and Weights

In this example, the following settings apply:

  • FPGA MPFE ports 0 is assigned to AXI read commands and port 1 is assigned to AXI write commands.
  • FPGA MPFE port 2 is assigned to Avalon-MM read and write commands.
  • L3 master ports (ports 6 and 8) and the MPU ports (port 7 and 9) are given the lowest priority but the MPU ports are configured with more arbitration static weight than the L3 master ports.
  • The FPGA MPFE command ports are given the highest priority; however, AXI ports 0 and 1 are given a larger static weight because they carry the highest priority traffic in the entire system. Assigning a high priority and larger static weight ensures ports 0 and 1 receives the highest quality-of-service (QoS).

The table below details the port weights and sum of weights.

Table 30.  SDRAM MPFE Port Priority, Weights and Sum of Weights

Priority

Weights

Sum of Weights

-

Port 0

Port 1

Port 2

Port 3

Port 4

Port 5

Port 6

Port 7

Port 8

Port 9

-

1

10

10

5

0

0

0

0

0

0

0

25

0

0

0

0

0

0

0

1

4

1

4

10

If the FPGA-to-SDRAM ports are configured according to the table and if both ports are accessed concurrently, you can expect the AXI port to receive 80% of the total service. This value is determined by taking the sum of port 0 and 1 weights divided by the total weight for all ports of priority 1. The remaining 20% of bandwidth is allocated to the Avalon-MM port. With these port settings, any FPGA transaction buffered by the MPFE for either slave port blocks the MPU and L3 masters from having their buffered transactions serviced. To avoid transaction starvation, you should assign ports the same priority level and adjust the bandwidth allocated to each port by adjusting the static weights.

MPFE Weight Calculation

The MPFE applies a deficit round-robin arbitration scheme to determine which port is serviced. The larger the port weight, the more often it is serviced. Ports are serviced only when they have buffered transactions and are set to the highest priority of all the ports that also have buffered transactions. The arbiter determines which port to service by examining the running weight of all the same ports at the same priority level and the largest running weight is serviced.

Each time a port is drained of all transactions, its running weight is set to 0x80. Each time a port is serviced, the static weight is added and the sum of weights is subtracted from the running weight for that port. Each time a port is not serviced (same priority as another port but has a lower running weight), the static weight for the port is added to the running weight of the port for that particular priority level. The running weight additions and subtractions are only applied to one priority level, so any time ports of a different priority level are being serviced, the running weight of a lower priority port is not modified.

MPFE Multi-Port Arbitration Considerations for Use

When using MPFE multi-port arbitration, the following considerations apply:

  • To ensure that the dynamic weight value does not roll over when a port is serviced, the following equation should always be true:
    (sum of weights - static weight) < 128
    If the running weight remains less than 128, arbitration for that port remains functional.
  • The memory controller commits the priority and weight registers into the MPFE arbiter once every 10 SDRAM clock cycles. As a result, when the mppriority register and mpweight_*_4 registers are configured at run time, the update interval can occur while the registers are still being written and ports can have different priority or weights than intended for a brief period. Because the mppriority and mpweight_*_4 registers can be updated in a single 32-bit transaction, Intel recommends updating first to ensure that transactions that need to be serviced have the appropriate priority after the next update. Because the static weights are divided among four 32-bit registers
  • In addition to the mppriority register and mpweight_*_* registers, the remappriority register adds another level of priority to the port scheduling. By programming bit N in the priorityremap field of the remappriority register, any port with an absolute priority N is sent to the front of the single port command queue and is serviced before any other transaction. Please refer to the remappriority register for more details.

The scheduler is work‑conserving. Write operations can only be scheduled when enough data for the SDRAM burst has been received. Read operations can only be scheduled when sufficient internal memory is free and the port is not occupying too much of the read buffer.