Re: [PATCH v2 0/9] riscv: implement accelerated crc using zbc

[David Marchand <david.marchand@redhat.com>]

On Fri, Jul 12, 2024 at 5:47 PM Daniel Gregory
<daniel.gregory@bytedance.com> wrote:
>
> The RISC-V Zbc extension adds instructions for carry-less multiplication
> we can use to implement CRC in hardware. This patch set contains two new
> implementations:
>
> - one in lib/hash/rte_crc_riscv64.h that uses a Barrett reduction to
>   implement the four rte_hash_crc_* functions
> - one in lib/net/net_crc_zbc.c that uses repeated single-folds to reduce
>   the buffer until it is small enough for a Barrett reduction to
>   implement rte_crc16_ccitt_zbc_handler and rte_crc32_eth_zbc_handler
>
> My approach is largely based on the Intel's "Fast CRC Computation Using
> PCLMULQDQ Instruction" white paper
> https://www.researchgate.net/publication/263424619_Fast_CRC_computation
> and a post about "Optimizing CRC32 for small payload sizes on x86"
> https://mary.rs/lab/crc32/
>
> Whether these new implementations are enabled is controlled by new
> build-time and run-time detection of the RISC-V extensions present in
> the compiler and on the target system.
>
> I have carried out some performance comparisons between the generic
> table implementations and the new hardware implementations. Listed below
> is the number of cycles it takes to compute the CRC hash for buffers of
> various sizes (as reported by rte_get_timer_cycles()). These results
> were collected on a Kendryte K230 and averaged over 20 samples:
>
> |Buffer    | CRC32-ETH (lib/net) | CRC32C (lib/hash)   |
> |Size (MB) | Table    | Hardware | Table    | Hardware |
> |----------|----------|----------|----------|----------|
> |        1 |   155168 |    11610 |    73026 |    18385 |
> |        2 |   311203 |    22998 |   145586 |    35886 |
> |        3 |   466744 |    34370 |   218536 |    53939 |
> |        4 |   621843 |    45536 |   291574 |    71944 |
> |        5 |   777908 |    56989 |   364152 |    89706 |
> |        6 |   932736 |    68023 |   437016 |   107726 |
> |        7 |  1088756 |    79236 |   510197 |   125426 |
> |        8 |  1243794 |    90467 |   583231 |   143614 |
>
> These results suggest a speed-up of lib/net by thirteen times, and of
> lib/hash by four times.
>
> I have also run the hash_functions_autotest benchmark in dpdk_test,
> which measures the performance of the lib/hash implementation on small
> buffers, getting the following times:
>
> | Key Length | Time (ticks/op)     |
> | (bytes)    | Table    | Hardware |
> |------------|----------|----------|
> |          1 |     0.47 |     0.85 |
> |          2 |     0.57 |     0.87 |
> |          4 |     0.99 |     0.88 |
> |          8 |     1.35 |     0.88 |
> |          9 |     1.20 |     1.09 |
> |         13 |     1.76 |     1.35 |
> |         16 |     1.87 |     1.02 |
> |         32 |     2.96 |     0.98 |
> |         37 |     3.35 |     1.45 |
> |         40 |     3.49 |     1.12 |
> |         48 |     4.02 |     1.25 |
> |         64 |     5.08 |     1.54 |

Thanks for the submission.
This series comes late for v24.07 and there was no review, it is
deferred to v24.11.

Cc: Sachin for info.


-- 
David Marchand