User manual

Table Of Contents
Zynq-7000 AP SoC Technical Reference Manual www.xilinx.com 133
UG585 (v1.11) September 27, 2016
Chapter 5: Interconnect
Interconnect Issuance Throttling
To optimize the latency or throughput of other masters in the system such as the CPUs, it might be
desirable to constrain the number of outstanding transactions that a high-performance port
requests to the system interconnect. Issuing capability is the maximum number of outstanding
commands that a HP can request at any one time.
Control of the read and write command issuing capability of the high performance AXI interface is
available as a primary input from the logic. This option can be enabled by means o the
axi_hp.AFI_{RD, WR}CHAN_CTRL [FabricOutCmdEn] register fields.
The logic signals, SAXIHP{0-3}RDISSUECAP1_EN and SAXIHP{0-3}WRISSUECAP1_EN allow you to
change the issuing capability of the AFI block to the PS dynamically between two levels.
Write FIFO Store and Forward
The write channels can be configured to store and forward write commands or allow them to pass
through with no storage. The following two registers control the mode of write, store, and forward:
axi_hp.AFI_WRCHAN_CTRL [WrCmdReleaseMode]
axi_hp.AFI_WRCHAN_CTRL [WrDataThreshold]
The mode register selects between a complete AXI burst store and forward, a partial AXI burst store
and forward, or a pass through (no store at all). If absolute minimum latency for write commands is
required, the pass-through mode could be selected. However, in cases where multiple masters are
competing for the system slaves, better system performance can likely be achieved using at least the
partial AXI burst store and forward mode. This is because once an AXI write is committed at each
point throughout the PS, the entire burst must be processed before any other write data from other
write commands can be processed.
For example, using pass-through mode, if one HP port with a slow clock rate issues a long burst, a
second port with a faster clock rate might need to wait until the entire slower write data burst has
been transferred, even if all of the write data on the fast clock is available. This is different from the
case of reads, where read data interleaving is permitted.
32-bit Interface Considerations
Each physical high performance PL AXI interface is programmable to be either a 32-bit or 64-bit
interface through the register field axi_hp.AFI_{RD, WR}CHAN_CTRL [32BitEn]. Note that the read and
write channels have separate enables and can therefore be configured differently.
Upsizing and Expansion
In 32-bit mode, some form of translation between the 32-bit port and the 64-bit port is required. For
write data, the 32-bit data (and write strobes) must be aligned correctly onto the appropriate lanes
in the 64-bit domain. For read data, the appropriate lanes of the 64-bit data must be aligned onto
the 32-bit data bus. This data alignment between different width interfaces are automatically dealt
with by the high performance AXI interface module.