Re: [dpdk-dev] [RFC PATCH v2] build: add platform meson option

["Juraj Linkeš" <juraj.linkes@pantheon.tech>]

> -----Original Message-----
> From: Bruce Richardson <bruce.richardson@intel.com>
> Sent: Wednesday, January 6, 2021 3:43 PM
> To: David Christensen <drc@linux.vnet.ibm.com>
> Cc: Juraj Linkeš <juraj.linkes@pantheon.tech>; thomas@monjalon.net;
> Honnappa.Nagarahalli@arm.com; dev@dpdk.org
> Subject: Re: [RFC PATCH v2] build: add platform meson option
> 
> On Tue, Jan 05, 2021 at 02:17:44PM -0800, David Christensen wrote:
> > > > The current meson option 'machine' should only specify the ISA,
> > > > which is not sufficient for Arm, where setting ISA implies other setting as
> well.
> > > > Add a new meson option, 'platform', which differentiates the type
> > > > of the build
> > > > (native/generic) and sets machine accordingly, unless the user
> > > > chooses to override it.
> > > > The 'machine' option also doesn't describe very well what it sets,
> > > > so introduce a new option 'cpu_instruction_set', but keep
> > > > 'machine' for backward compatibility.
> > > > These two new variables, taken together, achieve what 'machine'
> > > > was setting per architecture - setting the ISA in x86 build and
> > > > setting native/default 'build type' in aarch64 build - is now
> > > > properly being set for all architectures in a uniform manner.
> > > >
> > > > Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
> > > > ---
> > > >   config/arm/meson.build        |  4 +--
> > > >   config/meson.build            | 47 +++++++++++++++++++++++++----------
> > > >   devtools/test-meson-builds.sh |  9 ++++---
> > > >   meson_options.txt             |  8 ++++--
> > > >   4 files changed, 47 insertions(+), 21 deletions(-)
> > > >
> > > > diff --git a/config/arm/meson.build b/config/arm/meson.build index
> > > > 42b4e43c7..6b09a74a7 100644
> > > > --- a/config/arm/meson.build
> > > > +++ b/config/arm/meson.build
> > > > @@ -3,10 +3,10 @@
> > > >   # Copyright(c) 2017 Cavium, Inc
> > > >
> > > >   # for checking defines we need to use the correct compiler flags
> > > > -march_opt = '-
> > > > march=@0@'.format(machine)
> > > > +march_opt = '-march=@0@'.format(cpu_instruction_set)
> > > >
> > > >   arm_force_native_march = false
> > > > -arm_force_default_march = (machine == 'default')
> > > > +arm_force_default_march = (platform == 'generic')
> > > >
> > > >   flags_common_default = [
> > > >   	# Accelarate rte_memcpy. Be sure to run unit test
> > > > (memcpy_perf_autotest) diff --git a/config/meson.build
> > > > b/config/meson.build index a3154e29c..647116513 100644
> > > > --- a/config/meson.build
> > > > +++ b/config/meson.build
> > > > @@ -63,42 +63,63 @@ if not is_windows
> > > >   			pmd_subdir_opt)
> > > >   endif
> > > >
> > > > -# set the machine type and cflags for it
> > > > +platform = get_option('platform')
> > > > +
> > > > +# set the cpu_instruction_set type and cflags for it
> > > >   if meson.is_cross_build()
> > > > -	machine = host_machine.cpu()
> > > > +	cpu_instruction_set = host_machine.cpu()
> > > >   else
> > > > -	machine = get_option('machine')
> > > > +	cpu_instruction_set = get_option('cpu_instruction_set')
> > > > +	if get_option('machine') != 'auto'
> > > > +		warning('The "machine" option is deprecated. ' +
> > > > +		        'Please use "cpu_instruction_set" instead.')
> > > > +		if cpu_instruction_set != 'auto'
> > > > +			error('Setting both "machine" and ' +
> > > > +			      '"cpu_instruction_set" is unsupported.')
> > > > +		endif
> > > > +		cpu_instruction_set = get_option('machine')
> > > > +	endif
> > > > +endif
> > > > +
> > > > +if platform == 'native'
> > > > +	if cpu_instruction_set == 'auto'
> > > > +		cpu_instruction_set = 'native'
> > > > +	endif
> > > > +elif platform == 'generic'
> > > > +	if cpu_instruction_set == 'auto'
> > > > +		cpu_instruction_set = 'default'
> > > > +	endif
> > > >   endif
> > > >
> > > > -# machine type 'default' is special, it defaults to the per arch
> > > > agreed common
> > > > +if cpu_instruction_set == 'default'
> > > > +# cpu_instruction_set type 'default' is special, it defaults to
> > > > +the per arch agreed common
> > > >   # minimal baseline needed for DPDK.
> > > >   # That might not be the most optimized, but the most portable
> > > > version while  # still being able to support the CPU features required for
> DPDK.
> > > >   # This can be bumped up by the DPDK project, but it can never be
> > > > an  # invariant like 'native'
> > > > -if machine == 'default'
> > > >   	if host_machine.cpu_family().startswith('x86')
> > > >   		# matches the old pre-meson build systems default
> > > > -		machine = 'corei7'
> > > > +		cpu_instruction_set = 'corei7'
> > > >   	elif host_machine.cpu_family().startswith('arm')
> > > > -		machine = 'armv7-a'
> > > > +		cpu_instruction_set = 'armv7-a'
> > > >   	elif host_machine.cpu_family().startswith('aarch')
> > > >   		# arm64 manages defaults in config/arm/meson.build
> > > > -		machine = 'default'
> > > > +		cpu_instruction_set = 'default'
> > > >   	elif host_machine.cpu_family().startswith('ppc')
> > > > -		machine = 'power8'
> > > > +		cpu_instruction_set = 'power8'
> > > >   	endif
> > > >   endif
> >
> > This change forces the build on a P9 system to use the P8 instruction set.
> > Prior to this change the "native" machine type was used which resulted
> > in P9 instructions when built on a P9 system.  How do I force the
> > build to use the
> > power9 instruction set in this case?
> >
> > Dave
> 
> From looking at the patch, setting the "platform" to "native", or the
> instruction_set to "native" should do this.
> While I consider generic builds a good thing, I wonder if there is an expectation
> that "native" is always the default build type for DPDK builds?
> 
> /Bruce

I left this patch alone so that people could chime in, but noone did, so let's try to find some agreeable solution.

My current thoughts are as such:
The platform parameter specifies a set of DPDK options that will be used. This is what arm uses for its builds, x86 and ppc don't use this.
The cpu_instruction_set sets just one configuration among the "platform" configuration set.
We want the build to work on most machines of the machine architecture. That implies platform=generic (as in use config options that will work on everything of that architecture) and cpu_instruction_set=generic (as in use ISA that will work on all cpus of the build machine architecture).
Setting cpu_instruction_set=generic changes the build without cmdline options for ppc. Thus, the expectation may be that cpu_instruction_set should be native by default.

For arm, cpu_instruction_set is ignored (and thus the value doen't matter), since we can't use that without other config options (e.g. DPDK config for an SoC (such as RTE_ARM_FEATURE_ATOMICS) used with an invalid cpu_instuction_set). That means the only relevant parameter for Arm is platform and if we want to have a build usable on most machines of the build type, we have to use platform=generic.

For x86 and ppc, there's no difference between native and generic platform (as it's a new argument, the purpose of which is to differentiate DPDK config across platforms, which doesn't exist for x86 and ppc - DPDK config is the same (correct me if I'm wrong)).

So it basically boils down to what should be the default value of cpu_instruction_set when platform=generic (for platform=native, it's obviously native):
1. cpu_instruction_set=native, this would preserve the current behavior, but we won't use the 'auto' default. I think we can fall back to this if we don't agree on anything better.
2. cpu_instruction_set=auto, the same as cpu_instruction_set=generic,
3. cpu_instruction_set=generic, this changes behavior for ppc builds, but we may be able to remedy this:
Similarly to arm (arm is using platform for this, but the idea is the same), if cpu_instruction_set is generic, we can do some discovery for pcc and set the ISA accordingly (either power8 or power9). If I understand it correctly, power8 is a different architecture from power9 (I could be wrong on this), so this is desirable. There's some logic in config/ppc/meson.build, but it doesn't seem sufficient as a discovery mechanism between power8/power9.

I like 3 if we can find a way to discover power8/power9 ppc (that way we would be able to use cpu_instruction_set=auto), if not, we should probably go with 1 (in which case we can't use cpu_instruction_set=auto).

Or maybe we're overthingking this and we really should go with 1. What's (or should be) the difference between cpu_instruction_set=generic and cpu_instruction_set=native for x86 and ppc?