Analog Devices EE206 Application Notes

Engineer To Engineer Note EE-206

a

Technical Notes on using Analog Devices' DSP components and development tools

Contact our technical support by phone: (800) ANALOG-D or e-mail: [e-mail]
Or visit our on-line resources http://www.analog.com/dsp and http://www.analog.com/dsp/EZAnswers

ADSP-BF535 Blackfin® Processor PCI Interface Performance

Contributed by Jorge Manguane September 3, 2003

Introduction

This Engineer to Engineer Note will briefly
discuss the performance of the ADSP-BF535
Blackfin® Processor PCI interface. In general,
maximum PCI performance is achieved during
burst transfers, where a single address phase
(one clock cycle at the beginning of the
transaction where the address and transfer type
are output on the AD bus and on the C/BE lines,
respectively) is followed by multiple data
phases. It is therefore easy to see how single
data transfers may yield lower throughput. We
will look at the ramifications of each of these
transfer types, namely, burst and single word
transfers.

To validate this analysis the Eagle-35 and the
Hawk-35 boards were used as a system host and
PCI device, respectively. Both boards have the
ADSP-BF535 Blackfin® Processor as the main
processor and are sold and supported by
Momentum Data Systems, MDS. Also, a
VMETRO PCI bus analyzer/exerciser was used
to both monitor the PCI traffic along with its
metrics and to exercise the bus.

It should be noted that a PCI-to-PCI bridge is
used on the Hawk-35 board to interface the
BF535 PCI core interface to the PCI bus signals.
The main purpose of this bridge is to perform
automatic voltage detection and translation in
order to allow the board to be plugged into either
a 3.3 Volt or 5 Volt system.

data directly to/from internal (L2 memory) or
external memory spaces without involving the
processor core.

Burst transfers

When the BF535 is the bus master, burst
transfers are accomplished through Direct
Memory Access (DMA), in particular through
the processor’s Memory DMA (MemDMA)
engine. For PCI transfers, MemDMA allows
memory to memory DMA transfers between PCI
memory space and either internal L2 or SDRAM
memory.

The MemDMA engine’s burst size is 8 words,
which means that after every 8 words transferred
the PCI interface issues the next PCI address.
This situation is illustrated in the PCI trace
shown in Figure 1, where every address phase
(Start) is followed by eight data words. A new
Start (address phase) transfer follows, and the
process goes on until all data has been
transferred. Not every transaction (address
phase, data phase) will be the same (i.e., 8 data
words for each address phase) because at some
point the target may issue ”retries.” These retries
will force the PCI bus master to reissue a new
address. Figure 2 shows the PCI bus statistics
including the Data Transfers per Transaction ,
which indicate an average number of data words
transferred during each burst access.

On the BF535, a dedicated bus is available on
chip to allow an external bus master to transfer

Copyright 2003, Analog Devices, Inc. All rights reserved. Analog Devices assumes no responsibility for customer product design or the use or application of
customers’ products or for any infringements of patents or rights of others which may result from Analog Devices assistance. All trademarks and logos are property
of their respective holders. Information furnished by Analog Devices Applications and Development Tools Engineers is believed to be accurate and reliable, however
no responsibility is assumed by Analog Devices regarding technical accuracy and topicality of the content provided in Analog Devices’ Engineer-to-Engineer Notes.

Figure 1. Eagle-35 Burst write trace

slave). The target memory accessed was the on­chip ADSP-BF535 L2 memory. This will be the
case for the remainder of the traces shown in this
application note. Accesses to SDRAM yielded a
slight (insignificant) decrease in throughput –
refer to the metrics in T . As it can be seen

Figure 2. Eagle-35 Burst Write statistics

from , a
high throughput of 89.12 MB/s was achieved, as
well as an efficiency of 72.83%. Here the
efficiency is measured as a ratio of the
percentage of data transfers against the
percentage of total transactions.

The processors’ system clocks, SCLK, on both
boards were set to 131 MHz, with the core clocks
running at 262 MHz. Since the maximum burst
length on the BF535 is 8, SCLK is a major factor
that influences PCI throughput. As an example,
when SCLK was set to 120 MHz, the observed
throughput was 80 MB/s. Because the
MemDMA engine operates in the SCLK domain,
it is apparent that a higher SCLK allows more
data transferred per unit time. Note, however,
that 133 MHz is the maximum frequency to
which SCLK can be set.

able 1

a

Figure 2. Eagle-35 Burst Write statistics

These metrics were obtained from a DMA write
transfer between the Eagle-35 board (as the bus
master) and the Hawk-35 board (as the bus

When the BF535 is the target (slave) of a burst
transfer, the initiator’s burst size will determine
the amount of data transferred per transaction. As
an example, the burst length of the VMETRO’s
PCI exerciser can be specified to an arbitrary
length. Burst length also influences throughput.
As the burst length increases, the number of
transactions needed to complete the transfer
decreases. Figure 3 shows the Hawk-35 being
accessed by the VMETRO’s PCI exerciser (read
access), and Figure 4 shows the corresponding
bus statistics. Here the burst length was set to
equal the number of bytes transferred.

ADSP-BF535 Blackfin® Processor PCI Interface Performance (EE-206) Page 2 of 5

Figure 3. Exerciser Burst Reads trace

Writes are posted as long as there’s space in both
the transaction and transmit data FIFOs ( 8 x 32­bit words deep). The posting of writes allows the
BF535 PCI interface to reach the theoretical
maximum throughput of 44 MB/s for single­word write accesses.

a

Figure 5. Single word writes trace. Eagle-35 (master)
to Hawk-35 (target)

On the other hand, reads are not posted, and for

Figure 4. PCI Exerciser Burst Read statistics

Single Word Accesses

For the case of single word accesses, each
transaction consists of one address phase
followed by one data word in the data phase.
Writes are posted, meaning that the transaction is
buffered at an intermediate agent (e.g., a bridge
from one bus to another -- transaction FIFOs on
the BF535) and completes at the source before
actually completing at the destination. The
BF535 transaction FIFO is 4 transactions deep.

ADSP-BF535 Blackfin® Processor PCI Interface Performance (EE-206) Page 3 of 5

every transaction there’s at least one extra cycle
added between the address and data phases – this
cycle is referred to as the “turnaround cycle,”
after which the target drives the data on the PCI
Address/Data bus. In addition, both the target
and the master may introduce wait states, or the
target may disconnect the transaction with no
data being transferred, which would greatly
reduce the PCI throughput for single read
accesses.

Figure 6. Single word reads trace. Eagle-35 (master)
to Hawk-35 (target)

The following tables summarizes the ADSP­BF535 PCI performance metrics:

a

Access Type Efficiency Throughput

Burst Reads 80.09 % 85.17 MB/s

Burst Writes 95.48 % 57.96 MB/s

Table 2. Transfer between the VMETRO’s PCI
exerciser (PCI master) and the Hawk-35 board (PCI
device)

Conclusion

This Engineer to Engineer note provides some
PCI benchmarks and performance considerations
that should be helpful in determining the usage
of the ADSP-BF535 PCI interface appropriate to
the architecture of a given PCI agent. For
instance, it may make sense to have the ADSP­BF535 PCI interface perform a write to a PCI
agent rather than have the PCI agent read the
ADSP-BF535 PCI memory. This may, however,
not be necessary if the reading agent has a rather
large burst length, as was the case above for the
VMETRO PCI exerciser.

Access
Type

Writes

Reads

Single
Writes

Single
Reads

Table 1. Transfer between Eagle-35 (PCI master)
and Hawk-35 (PCI device) both with an SCLK of 131
MHz.

Target
memory

L2 89.12 MB/s 72.83 % Burst

SDRAM 88.26 MB/s 72.44%

L2 26.04 MB/s 40 % Burst

SDRAM 25.97 MB/s 39.91%

L2 44.42 MB/s 33.33% Sequential

SDRAM 44.42 MB/s 33.33

L2 4.16 MB/s 9.37% Sequential

SDRAM 3.55 MB/s 8%

Throughput Efficiency

ADSP-BF535 Blackfin® Processor PCI Interface Performance (EE-206) Page 4 of 5

a

References

[1] ADSP-BF535 Blackfin® Processor Hardware Reference Revision 2.0, April 2003 Analog Devices,

Inc.

[2] PCI Hardware and Reference, 5

[3] PCI System Architecture, 4

[4] PCI Local Bus Specification Revision 2.2 December 18, 1998

th

Edition, Solari & Willse.

th

Edition, Mindshare, Inc

Document History

Version Description

September 3, 2003 by J Manguane Initial Final Release

August 4, 2003 by J Manguane Initial Draft Release

ADSP-BF535 Blackfin® Processor PCI Interface Performance (EE-206) Page 5 of 5