On 2/19/21 1:11 AM, Juraj Linkeš wrote:


-----Original Message-----
From: Bruce Richardson <bruce.richard...@intel.com>
Sent: Wednesday, January 6, 2021 3:43 PM
To: David Christensen <d...@linux.vnet.ibm.com>
Cc: Juraj Linkeš <juraj.lin...@pantheon.tech>; tho...@monjalon.net;
honnappa.nagaraha...@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.lin...@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.

POWER8 code (Power ISA v2.07) runs on POWER9 CPUs (Power ISA v3.0), so setting cpu_instruction_set to GENERIC is a reasonable option. POWER10 CPUs (Power ISA v3.1) are around the corner so I want to make sure developers can tune the application for the platform as easily as possible. I'm fine with supporting GENERIC, just need clear instructions on how to build for a particular ISA.

Dave

Reply via email to