[dpdk-dev] [PATCH] sched: enable neon optimizations

[Sekhar, Ashwin]

On Friday 28 April 2017 09:20 AM, Jianbo Liu wrote:
> On 27 April 2017 at 21:00, Ashwin Sekhar T K
> <ashwin.sekhar at caviumnetworks.com> wrote:
>> * Enabled CONFIG_RTE_SCHED_VECTOR for arm64
>> * Verified the changes with sched_autotest unit test case
>>
>> Signed-off-by: Ashwin Sekhar T K <ashwin.sekhar at caviumnetworks.com>
>> ---
>>  config/defconfig_arm64-armv8a-linuxapp-gcc |  2 +-
>>  lib/librte_sched/rte_sched.c               | 22 ++++++++++++++++++++++
>>  2 files changed, 23 insertions(+), 1 deletion(-)
>>
>> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc
>> index 65888ce..021044a 100644
>> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc
>> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc
>> @@ -48,4 +48,4 @@ CONFIG_RTE_LIBRTE_FM10K_PMD=n
>>  CONFIG_RTE_LIBRTE_SFC_EFX_PMD=n
>>  CONFIG_RTE_LIBRTE_AVP_PMD=n
>>
>> -CONFIG_RTE_SCHED_VECTOR=n
>> +CONFIG_RTE_SCHED_VECTOR=y
>
> It's enough to remove this line only, I don't think you must enable it
> explicitly in the armv8a common config.
>
Tried removing this line from armv8a config. But in that case 
RTE_SCHED_VECTOR doesn't get defined.
./config/common_base has "CONFIG_RTE_SCHED_VECTOR=n" as the default 
setting. So enabling explicitly is required.

- Ashwin

>> diff --git a/lib/librte_sched/rte_sched.c b/lib/librte_sched/rte_sched.c
>> index 614705d..4ba476a 100644
>> --- a/lib/librte_sched/rte_sched.c
>> +++ b/lib/librte_sched/rte_sched.c
>> @@ -58,6 +58,8 @@
>>
>>  #if defined(__SSE4__)
>>  #define SCHED_VECTOR_SSE4
>> +#elif defined(RTE_MACHINE_CPUFLAG_NEON)
>> +#define SCHED_VECTOR_NEON
>>  #endif
>>
>>  #endif
>> @@ -1732,6 +1734,26 @@ grinder_pipe_exists(struct rte_sched_port *port, uint32_t base_pipe)
>>         return 1;
>>  }
>>
>> +#elif defined(SCHED_VECTOR_NEON)
>> +
>> +static inline int
>> +grinder_pipe_exists(struct rte_sched_port *port, uint32_t base_pipe)
>> +{
>> +       uint32x4_t index, pipes;
>> +       uint32_t *pos = (uint32_t *)port->grinder_base_bmp_pos;
>> +
>> +       index = vmovq_n_u32(base_pipe);
>> +       pipes = vld1q_u32(pos);
>> +       if (!vminvq_u32(veorq_u32(pipes, index)))
>> +               return 1;
>> +
>> +       pipes = vld1q_u32(pos + 4);
>> +       if (!vminvq_u32(veorq_u32(pipes, index)))
>> +               return 1;
>> +
>> +       return 0;
>> +}
>> +
>>  #else
>>
>>  static inline int
>> --
>> 2.7.4
>>
>