AM5728: How can I improve PCIe performance?

DaeWon Lim

Part Number: AM5728

Tool/software:

We are using DDR3 memory via the EMIF of the AM5728, with 16MB of the memory starting from address 0xBD000000 set as an inbound region for PCIe. This setup is being used on a PC running Windows. The PCIe configuration is currently set to PCIe 2.0 with a single lane, which should theoretically provide a transfer speed of up to 500MB/s. However, the observed performance is significantly lower, at only 2.5MB/s.

The source code being used is from the pcie.c file located in the am5728 device directory of PDK version 1.0.19.

over 2 years ago

0 Takuma Fujiwara over 2 years ago

TI__Guru 62173 points

Hi DaeWon,

Can you check if the PCIe driver is using DMA for memory transfers? For a significant dip in performance as you mentioned, lack of DMA usage is what comes to mind.

Regards,

Takuma

0 DaeWon Lim over 2 years ago in reply to Takuma Fujiwara

Prodigy 160 points

Thank you for your response.

Is there any reference example within the PDK that I can refer to? In the system we are using, the specified region is being used like shared memory, which is causing Cache Coherency Issues. We are currently managing this by using writeback and invalidate operations. If we use DMA, can these issues be handled appropriately?

Thank you.

+1 Takuma Fujiwara over 2 years ago in reply to DaeWon Lim

TI__Guru 62173 points

Hi DaeWon,

DMA will require stricter cache management since DMA does not have cache coherency with other cores initiating transactions. What DMA would do is allow offloading of data transactions to DMA instead of the ARM cores, so if the ARM cores were the bottleneck this should improve performance.

As for reference examples, we have an old PCIe example here: https://git.ti.com/cgit/keystone-rtos/common-csl-ip/tree/example/pcie/write_loopback/rc/rc_write_loopback_app_main.c?h=legacy_master.

RTOS PCIe has been descoped for newer platforms, and this old PCIe example is unmaintained so support will be limited from us. However, I think this is the best example online. Otherwise, there are EDMA examples that are generic and non PCIe specific that can be referenced.

Regards,

Takuma

Processors

Processors forum

AM5728: How can I improve PCIe performance?