Identifying Parallelization for an FPGA Application - 2023.1 English

Vitis Tutorials: Hardware Acceleration (XD099)

Document ID
Release Date
2023.1 English

In Software, the flow will look similar to the following figure. missing image

  • Murmurhash2 functions are calculated for all the words up front, and output flags are set in the local memory. Each of the loops in the hash compute functions are run sequentially.

  • After all hashes have computed, only then can another loop be called for all the documents to calculate the compute score.

When you run the application on the FPGA, an additional delay is added when the data is transferred from the host to device memory and read back. You can split the entire application time and create a budgeted-based run based on following requirements:

  1. Transferring document data of size 1400 MB from the host application to the device DDR using PCIe. Using a PCIe Write BW of approximately 9 GBps, the approximate time for transfer = 1400 MB/9.5 GBps = ~147 ms.

  2. Compute the hashes on the FPGA.

  3. Transferring flags data of size 350 MB from Device to Host using PCIe. Using a PCIe Read BW of approximately 12 GBps, the approximate time for transfer = 350 MB/12G = ~30 ms.

  4. Calculate the compute score after all the flags are available from the FPGA. This takes about a total of 380 ms on the CPU.

For achieving application target of 380 ms, adding 1470 ms + 30 ms + 380 ms is clearly above your goal of 380 ms for 100,000 documents computation, and the FPGA time for computing the hashes has not yet been accounted for. However unlike the CPU, parallelization can be achieved while working with the FPGA. You should also architect an accelerator for “Compute hashes on FPGA” for the best possible performance to ensure the FPGA compute is not the bottleneck.

If steps 1 through 4 are carried out sequentially like the CPU, you cannot achieve your performance goal. You will need to take advantage of concurrent processing and overlapping of the FPGA.

  1. Parallelism between the host to device data transfer and Compute on FPGA. Split the 100,000 documents into multiple buffers, and send the buffers to the device so the kernel does not need to wait until the whole buffer is transferred.

  2. Parallelism between Compute on FPGA and Compute Profile score on CPU.

  3. Increase the number of words to be processed in parallel to increase the concurrent hash compute processing. The hash compute in the CPU is performed on a 32-bit word. Because the hash compute can be completed independently on several words, you can explore computing four, eight, or even 16 words in parallel.

Based on your earlier analysis, the Murmurhash2 function is a good candidate for computation on the FPGA, and the Compute Profile Score calculation can be carried out on the host. For a hardware function on the FPGA, based on the established goal in previous lab, you should process hashes as fast as possible by processing multiples of these words every clock cycle.

The application conceptually can similar to the following figure.

missing image

Transferring words from the host to CPU, Compute on FPGA, transferring words from the FPGA to CPU all can be executed in parallel. The CPU starts to calculate the profile score as soon as the flags are received;essentially, the profile score on the CPU can also start calculations in parallel. With the pipelining as shown above, the latency of the Compute on FPGA will become invisible.