Re: [dpdk-dev] [PATCH v6 2/2] net/hns3: refactor SVE code compile method

[fengchengwen <fengchengwen@huawei.com>]

Fix in v7, thanks

On 2021/5/24 18:03, Ruifeng Wang wrote:
>> -----Original Message-----
>> From: fengchengwen <fengchengwen@huawei.com>
>> Sent: Monday, May 24, 2021 4:44 PM
>> To: Ruifeng Wang <Ruifeng.Wang@arm.com>; thomas@monjalon.net;
>> ferruh.yigit@intel.com
>> Cc: dev@dpdk.org; jerinj@marvell.com; viktorin@rehivetech.com;
>> bruce.richardson@intel.com; Honnappa Nagarahalli
>> <Honnappa.Nagarahalli@arm.com>; jerinjacobk@gmail.com;
>> juraj.linkes@pantheon.tech; nd <nd@arm.com>
>> Subject: Re: [PATCH v6 2/2] net/hns3: refactor SVE code compile method
>>
>>
>>
>> On 2021/5/24 13:38, Ruifeng Wang wrote:
>>>> -----Original Message-----
>>>> From: fengchengwen <fengchengwen@huawei.com>
>>>> Sent: Friday, May 21, 2021 2:53 PM
>>>> To: Ruifeng Wang <Ruifeng.Wang@arm.com>; thomas@monjalon.net;
>>>> ferruh.yigit@intel.com
>>>> Cc: dev@dpdk.org; jerinj@marvell.com; viktorin@rehivetech.com;
>>>> bruce.richardson@intel.com; Honnappa Nagarahalli
>>>> <Honnappa.Nagarahalli@arm.com>; jerinjacobk@gmail.com;
>>>> juraj.linkes@pantheon.tech; nd <nd@arm.com>
>>>> Subject: Re: [PATCH v6 2/2] net/hns3: refactor SVE code compile
>>>> method
>>>>
>>>>
>>>>
>>>> On 2021/5/21 13:21, Ruifeng Wang wrote:
>>>>>> -----Original Message-----
>>>>>> From: fengchengwen <fengchengwen@huawei.com>
>>>>>> Sent: Thursday, May 20, 2021 6:55 PM
>>>>>> To: Ruifeng Wang <Ruifeng.Wang@arm.com>; thomas@monjalon.net;
>>>>>> ferruh.yigit@intel.com
>>>>>> Cc: dev@dpdk.org; jerinj@marvell.com; viktorin@rehivetech.com;
>>>>>> bruce.richardson@intel.com; Honnappa Nagarahalli
>>>>>> <Honnappa.Nagarahalli@arm.com>; jerinjacobk@gmail.com;
>>>>>> juraj.linkes@pantheon.tech; nd <nd@arm.com>
>>>>>> Subject: Re: [PATCH v6 2/2] net/hns3: refactor SVE code compile
>>>>>> method
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 2021/5/20 16:24, Ruifeng Wang wrote:
>>>>>>>> -----Original Message-----
>>>>>>>> From: Chengwen Feng <fengchengwen@huawei.com>
>>>>>>>> Sent: Wednesday, May 19, 2021 9:26 PM
>>>>>>>> To: thomas@monjalon.net; ferruh.yigit@intel.com
>>>>>>>> Cc: dev@dpdk.org; jerinj@marvell.com; Ruifeng Wang
>>>>>>>> <Ruifeng.Wang@arm.com>; viktorin@rehivetech.com;
>>>>>>>> bruce.richardson@intel.com; Honnappa Nagarahalli
>>>>>>>> <Honnappa.Nagarahalli@arm.com>; jerinjacobk@gmail.com;
>>>>>>>> juraj.linkes@pantheon.tech; nd <nd@arm.com>
>>>>>>>> Subject: [PATCH v6 2/2] net/hns3: refactor SVE code compile
>>>>>>>> method
>>>>>>>>
>>>>>>>> Currently, the SVE code is compiled only when -march supports SVE
>>>>>>>> (e.g. '- march=armv8.2a+sve'), there maybe some problem[1] with
>>>>>>>> this
>>>>>> approach.
>>>>>>>>
>>>>>>>> The solution:
>>>>>>>> a. If the minimum instruction set support SVE then compiles it.
>>>>>>>> b. Else if the compiler support SVE then compiles it.
>>>>>>>> c. Otherwise don't compile it.
>>>>>>>>
>>>>>>>> [1] https://mails.dpdk.org/archives/dev/2021-April/208189.html
>>>>>>>>
>>>>>>>> Fixes: 8c25b02b082a ("net/hns3: fix enabling SVE Rx/Tx")
>>>>>>>> Fixes: 952ebacce4f2 ("net/hns3: support SVE Rx")
>>>>>>>> Cc: stable@dpdk.org
>>>>>>>>
>>>>>>>> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
>>>>>>>> ---
>>>>>>>>  drivers/net/hns3/hns3_rxtx.c |  2 +-
>>>>>>>> drivers/net/hns3/meson.build
>>>>>>>> |
>>>>>>>> 21 ++++++++++++++++++++-
>>>>>>>>  2 files changed, 21 insertions(+), 2 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/net/hns3/hns3_rxtx.c
>>>>>>>> b/drivers/net/hns3/hns3_rxtx.c index 1d7a769..4ef20c6 100644
>>>>>>>> --- a/drivers/net/hns3/hns3_rxtx.c
>>>>>>>> +++ b/drivers/net/hns3/hns3_rxtx.c
>>>>>>>> @@ -2808,7 +2808,7 @@ hns3_get_default_vec_support(void)
>>>>>>>>  static bool
>>>>>>>>  hns3_get_sve_support(void)
>>>>>>>>  {
>>>>>>>> -#if defined(RTE_ARCH_ARM64) && defined(__ARM_FEATURE_SVE)
>>>>>>>> +#if defined(CC_SVE_SUPPORT)
>>>>>>>>  	if (rte_vect_get_max_simd_bitwidth() <
>> RTE_VECT_SIMD_256)
>>>>>>>>  		return false;
>>>>>>>>  	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SVE))
>>>>>>>> diff --git a/drivers/net/hns3/meson.build
>>>>>>>> b/drivers/net/hns3/meson.build index 53c7df7..5f9af9b 100644
>>>>>>>> --- a/drivers/net/hns3/meson.build
>>>>>>>> +++ b/drivers/net/hns3/meson.build
>>>>>>>> @@ -35,7 +35,26 @@ deps += ['hash']
>>>>>>>>
>>>>>>>>  if arch_subdir == 'arm' and dpdk_conf.get('RTE_ARCH_64')
>>>>>>>>      sources += files('hns3_rxtx_vec.c')
>>>>>>>> -    if cc.get_define('__ARM_FEATURE_SVE', args: machine_args) != ''
>>>>>>>> +
>>>>>>>> +    # compile SVE when:
>>>>>>>> +    # a. support SVE in minimum instruction set baseline
>>>>>>>> +    # b. it's not minimum instruction set, but compiler support
>>>>>>>> +    if cc.get_define('__ARM_FEATURE_SVE', args: machine_args) !=
>> ''
>>>>>>>> + and
>>>>>>>> cc.check_header('arm_sve.h')
>>>>>>>> +        cflags += ['-DCC_SVE_SUPPORT']
>>>>>>> With SVE build fix patch [1], CC_SVE_ACLE_SUPPORT will be defined.
>>>>>>> Here we can use CC_SVE_ACLE_SUPPORT and not to add a new one.
>>>>>>>
>>>>>>
>>>>>> The CC_SVE_ACLE_SUPPORT was defined under default machine_args
>>>> which
>>>>>> support SVE, it can't deals with the situation: the default
>>>>>> machine_args don't support SVE but compiler support SVE.
>>>>>> So the CC_SVE_SUPPORT marco is necessary.
>>>>> Agree that macro for SVE is also needed here. And we can also use '-
>>>> DCC_SVE_ACLE_SUPPORT' here right?
>>>>> I think there is no difference between CC_SVE_ACLE_SUPPORT and
>>>> CC_SVE_SUPPORT when they are used in source code.
>>>>> IMO the same macro name can be used, and it removes redundancy and
>>>> confusion.
>>>>>
>>>>
>>>> You are right, no difference between CC_SVE_ACLE_SUPPORT and
>>>> CC_SVE_SUPPORT But the hns3 sve already support 20.11, and
>>>> CC_SVE_ACLE_SUPPORT was newly defined, there maybe some
>> problems when
>>>> backporting.
>>> 20.11 release has no machine enabled SVE extension.
>>>
>>>>
>>>> Or we could redefine CC_SVE_ACLE_SUPPORT under default
>> machine_args:
>>>>     if cc.get_define('__ARM_FEATURE_SVE', args: machine_args) != ''
>>>> and
>>>> cc.check_header('arm_sve.h')
>>>>         cflags += ['-DCC_SVE_ACLE_SUPPORT']
>>> 'if dpdk_conf.get(CC_SVE_ACLE_SUPPORT)' should be fine?
>>> Stable branch has no SVE enabled in machine_args.
>>>
>>
>> But 20.11 use could use hns3 SVE path when compile with gcc10.
> 20.11 user will be able to use hns3 SVE path. 
> Implementation 'a' should be fine for both 20.11 and 21.08.
> 
>>
>> If we reuse the CC_SVE_ACLE_SUPPORT macro, there maybe problem when
>> backporting:
>> a. In 21.08 we could depend on CC_SVE_ACLE_SUPPORT, so it will be:
>>     if dpdk_conf.get('CC_SVE_ACLE_SUPPORT')
>>         sources += files('hns3_rxtx_vec_sve.c')
>>     elif cc.has_argument('-march=armv8.2-a+sve') and
>> cc.check_header('arm_sve.h')
> gcc10 user will go into this branch, and SVE path will be included.
> It is identical in 'a' and 'b'.
> 
>>         sve_cflags = ['-DCC_SVE_ACLE_SUPPORT']
>>         ...
>> b. But for backport to 20.11, we should use another impl:
>>     if cc.get_define('__ARM_FEATURE_SVE', args: machine_args) != '' and
>> cc.check_header('arm_sve.h')
> 'if' clause in implementation 'a' will have the same behavior as this one in 20.11.
>  Both of the checks will be false. 
> 'CC_SVE_ACLE_SUPPORT' is not defined in 20.11 -> result in false.
> No machine_args have sve enabled in 20.11 -> result in false.
> So I think there is a chance we can use a single macro.
> 
>>         cflags += ['-DCC_SVE_ACLE_SUPPORT']
>>         sources += files('hns3_rxtx_vec_sve.c')
>>     elif cc.has_argument('-march=armv8.2-a+sve') and
>> cc.check_header('arm_sve.h')
>>         sve_cflags = ['-DCC_SVE_ACLE_SUPPORT']
>>         ...
>> As you see, the above two are not unified.
>>
>> So here I think use the CC_SVE_SUPPORT is appropriate.
>>
>> @Ferruh  what's your opinion ?
>>
>>>>         sources += files('hns3_rxtx_vec_sve.c')
>>>>     elif cc.has_argument('-march=armv8.2-a+sve') and
>>>> cc.check_header('arm_sve.h')
>>>>         sve_cflags = ['-DCC_SVE_ACLE_SUPPORT']
>>> This is fine. Macro name is consistent.
>>>
>>>>         foreach flag: cflags
>>>>             # filterout -march -mcpu -mtune
>>>>             if not (flag.startswith('-march=') or
>>>> flag.startswith('-mcpu=') or
>>>> flag.startswith('-mtune='))
>>>>                 sve_cflags += flag
>>>>             endif
>>>>         endforeach
>>>> but this way may introduce coupling, I think.
>>>>
>>>>>>
>>>>>>> [1]
>>>>>>> http://patches.dpdk.org/project/dpdk/patch/1621495007-28387-1-
>> git-
>>>>>>> se
>>>>>>> nd
>>>>>>> -email-fengchengwen@huawei.com/
>>>>>>>
>>>>>>>>          sources += files('hns3_rxtx_vec_sve.c')
>>>>>>>> +    elif cc.has_argument('-march=armv8.2-a+sve') and
>>>>>>>> cc.check_header('arm_sve.h')
>>>>>>>> +        sve_cflags = ['-DCC_SVE_SUPPORT']
>>>>>>>> +        foreach flag: cflags
>>>>>>>> +            # filterout -march -mcpu -mtune
>>>>>>>> +            if not (flag.startswith('-march=') or
>>>>>>>> + flag.startswith('-mcpu=') or
>>>>>>>> flag.startswith('-mtune='))
>>>>>>>> +                sve_cflags += flag
>>>>>>>> +            endif
>>>>>>>> +        endforeach
>>>>>>>> +        hns3_sve_lib = static_library('hns3_sve_lib',
>>>>>>>> +                        'hns3_rxtx_vec_sve.c',
>>>>>>>> +                        dependencies: [static_rte_ethdev],
>>>>>>>> +                        include_directories: includes,
>>>>>>>> +                        c_args: [sve_cflags, '-march=armv8.2-a+sve'])
>>>>>>>> +        objs +=
>>>>>>>> + hns3_sve_lib.extract_objects('hns3_rxtx_vec_sve.c')
>>>>>>>>      endif
>>>>>>>>  endif
>>>>>>>> --
>>>>>>>> 2.8.1
>>>>>>>
>>>>>>>
>>>>>>> .
>>>>>>>
>>>>>
>>>
>