Stratix10 P-Tile AVMM IP for PCIe does multiple read requests for a single s/w 64-byte read request

[rkv]

The software is able to do 64-byte single read transaction to on-chip ram by making the BAR region as WC region using the instruction below.

__m256i a = _mm256_stream_load_si256 ((__m256i*)mem_base + 0);
__m256i b = _mm256_stream_load_si256 ((__m256i*)mem_base + 1);
_mm_mfence();

But when we tried single 64 Byte read to a custom component it repeats the transaction of the same kind three times. Please note that the custom component asserts read data valid after some time. when we checked in signal tap, read is asserted again, immediately after read data valid is asserted. screenshot of the same is attached.

64byte single write transaction is happening properly to the custom component. 

[SengKok_L_Intel]

Hi,

Just to clarify, please do let me know if my understanding is correct:

1. 64 Bytes single read from Host to the FPGA endpoint (On-chip memory)

• A single read was observed, where the readdatavalid and *_m_read_o only asserted once

1. 64 Bytes single read from Host to FPGA endpoint (Customer component)

• Three single reads were observed, where the  readdatavalid and *_m_read_o asserted three-time
  • All three read are read out at the same address with the same data

To further understand the problem, could you please also capture the Avalon-ST interface, this can further help to identify whether the host is sending three memory read request continuously?

You should be able to add the tx_st* and rx_st* signal into the signal tap.

https://www.intel.com/content/dam/www/programmable/us/en/pdfs/literature/ug/ug_ptile_pcie_avst.pdf

Regards -SK 

[rkv]

Your understanding is correct. We are using AVMM interface and the screenshot of the same is attached already. In that screenshot you can see that immediately after read_data_valid is asserted from the FPGA custom component, read is again asserted by PCIe P-Tile IP.

Also please note that the FPGA custom component asserts read_data_valid after some time. So does this has any impact? We also tried with 32B Read transaction and didn't face any issue.

[SengKok_L_Intel]

Did you observe completion Timeout status is asserted from the Advanced uncorrectable error status register?

Does the latency of “read_data_valid” for 64bytes larger than 32bytes? How long it requires to wait for *_valid signal to assert?

Regards -SK  

[SengKok_L_Intel]

If further support is needed in this thread, please post a response within 15 days. After 15 days, this thread will be transitioned to community support. The community users will be able to help you with your follow-up questions.