[Buildroot] Analysis of build results

[Peter Seiderer <ps.report@gmx.net>]

Hello Romain,

On Sat, 29 Feb 2020 22:15:10 +0100, Romain Naour <romain.naour@gmail.com> wrote:

> Hi Peter,
> 
> Le 28/02/2020 ? 16:02, Peter Seiderer a ?crit?:
> > Hello Romain,
> > 
> > On Fri, 28 Feb 2020 08:32:18 +0100, Romain Naour <romain.naour@gmail.com> wrote:
> >   
> >>>  
> >>>>     arm      |         mesa3d-19.3.4          | NOK |  
> >>> http://autobuild.buildroot.net/results/6387b0a99e1a0922811919623d9a10b0943988df
> >>> |
> >>>
> >>> {standard input}: Assembler messages:
> >>> {standard input}:334: Error: selected processor does not support `vldm
> >>> r4,{q0,q1,q2,q3}' in ARM mode
> >>> {standard input}:335: Error: selected processor does not support `vst1.8
> >>> d0,[r3],r2' in ARM mode
> >>> {standard input}:336: Error: selected processor does not support `vst1.8
> >>> d1,[r3],r2' in ARM mode
> >>> {standard input}:337: Error: selected processor does not support `vst1.8
> >>> d2,[r3],r2' in ARM mode
> >>>
> >>> Bad stuff happens when building the vc4 driver. It probably needs some
> >>> minimal ARM architecture variant.
> >>>  
> >>
> >> I believe we should re-add the dependency on neon for vc4 driver. There is
> >> -mfpu=neon added by the build system.  
> > 
> > The BR2_ARM_CPU_HAS_NEON dependency was dropped with commit [1] because mesa3d
> > commit [2] added a compile time detected neon support, in the meantime this
> > was changed to a runtime detection [3] and finally enabled for the meson
> > cross compile build with commit [4] (assuming toolchain support for the
> > neon assembler instructions)...  
> 
> Thank you for the detailed explanation.
> 
> I saw the commit [4] and conclude that upstream mesa3d now require neon support
> when vc4 driver is enabled.
> 
> > 
> > Maybe the compile error could be fixed by re-adding partly the compile time
> > check for the neon support (tested):
> > 
> > --- mesa3d-19.3.4/src/broadcom/common/v3d_cpu_tiling.h_orig	2020-02-28 14:47:44.232634657 +0100
> > +++ mesa3d-19.3.4/src/broadcom/common/v3d_cpu_tiling.h	2020-02-28 14:49:21.518425222 +0100
> > @@ -31,7 +31,7 @@
> >  v3d_load_utile(void *cpu, uint32_t cpu_stride,
> >                 void *gpu, uint32_t gpu_stride)
> >  {
> > -#if defined(V3D_BUILD_NEON) && defined(PIPE_ARCH_ARM)
> > +#if defined(V3D_BUILD_NEON) && defined(__ARM_ARCH) && __ARM_ARCH >= 7
> >          if (gpu_stride == 8) {
> >                  __asm__ volatile (
> >                          /* Load from the GPU in one shot, no interleave, to
> > @@ -141,7 +141,7 @@
> >  v3d_store_utile(void *gpu, uint32_t gpu_stride,
> >                  void *cpu, uint32_t cpu_stride)
> >  {
> > -#if defined(V3D_BUILD_NEON) && defined(PIPE_ARCH_ARM)
> > +#if defined(V3D_BUILD_NEON) && defined(__ARM_ARCH) && __ARM_ARCH >= 7
> >          if (gpu_stride == 8) {
> >                  __asm__ volatile (
> >                          /* Load each 8-byte line from cpu-side source,
> >   
> 
> But -mfpu=neon is still present on the command line.

But without problems, as all the neon stuff in handcrafted via
inline assambler ;-)

> 
> Do you think if the neon support can be handled/forced by a new meson option?

The following patch fixes the problem for the vc4 driver (but not for the
v3d one which needs neon unconditionally - or fixed by the previous patch):

diff --git a/meson.build b/meson.build
index 7e2295b..a3134b1 100644
--- a/meson.build
+++ b/meson.build
@@ -1112,8 +1112,10 @@ elif host_machine.cpu_family() == 'x86_64'
   endif
 elif host_machine.cpu_family() == 'arm'
   if system_has_kms_drm
-    with_asm_arch = 'arm'
-    pre_args += ['-DUSE_ARM_ASM']
+    if not (host_machine.cpu().contains('4') or host_machine.cpu().contains('5') or host_machine.cpu().contains('6'))
+      with_asm_arch = 'arm'
+      pre_args += ['-DUSE_ARM_ASM']
+    endif
   endif
 elif host_machine.cpu_family() == 'aarch64'
   if system_has_kms_drm
diff --git a/src/gallium/drivers/vc4/meson.build b/src/gallium/drivers/vc4/meson.build
index 5ce5af5..2bb5d06 100644
--- a/src/gallium/drivers/vc4/meson.build
+++ b/src/gallium/drivers/vc4/meson.build
@@ -84,7 +84,7 @@ files_libvc4 = files(
 vc4_c_args = []
 
 libvc4_neon = []
-if host_machine.cpu_family() == 'arm'
+if host_machine.cpu_family() == 'arm' and not (host_machine.cpu().contains('4') or host_machine.cpu().contains('5') or host_machine.cpu().contains('6'))
   libvc4_neon = static_library(
     'vc4_neon',
     'vc4_tiling_lt_neon.c',

An alternative to the check on ARMvX >=7 a new meson option should be possible...

Regards,
Peter

> 
> Best regards,
> Romain
> 
> > Regards,
> > Peter
> > 
> > [1] https://git.buildroot.net/buildroot/commit/package/mesa3d?id=350cb0d32ece533b9723a5f3ca6fbf7e6f071c90
> > [2] https://cgit.freedesktop.org/mesa/mesa/commit/?h=17.1&id=4d30024238efa829cabc72c1601beeee18c3dbf2
> > [3] https://cgit.freedesktop.org/mesa/mesa/commit/?h=staging/19.3&id=ece06defe77a77d2db40abeddee5a2e0e45654ce
> > [4] https://cgit.freedesktop.org/mesa/mesa/commit/?id=932ed9c00b99e6ec92146ec9e820f546cf3e6551
> >   
> >>
> >> Romain  
> >   
>