Bug 11996

Summary: opencv3 SIGILL on Cortex-A5 with VFPv4-D16
Product: buildroot Reporter: James Cowgill <jcowgill+busybox>
Component: OtherAssignee: unassigned
Status: RESOLVED FIXED    
Severity: normal CC: buildroot
Priority: P5    
Version: 2019.02.3   
Target Milestone: ---   
Hardware: All   
OS: Linux   
Host: Target:
Build:

Description James Cowgill 2019-07-01 17:17:56 UTC
I recently upgraded from 2018.02 to 2019.02 and discovered that opencv3 has broken on my board (it worked with 2018.02). The board has a Cortex-A5 with VFPv4-D16 enabled but without NEON and VFPv4 (so it only has 16 double floating point registers).

These are the target options I have set:
BR2_arm=y
BR2_BINFMT_ELF=y
BR2_cortex_a5=y
BR2_ARM_ENABLE_NEON=n
BR2_ARM_ENABLE_VFP=y
BR2_ARM_EABIHF=y
BR2_ARM_FPU_VFPV4D16=y
BR2_ARM_INSTRUCTIONS_THUMB2=y

I can see that my application gets a SIGILL because it tries to load a value into the d20 register which doesn't exist in VFPv4-D16.

0xb661d3d2 in cv::interpolateLanczos4 (coeffs=0xbef4602c, x=0.03125)
    at /home/jcowgill/workspace/bsp/buildroot/output/build/opencv3-3.4.3/modules/imgproc/src/imgwarp.cpp:176
176     /home/jcowgill/workspace/bsp/buildroot/output/build/opencv3-3.4.3/modules/imgproc/src/imgwarp.cpp: No such file or directory.
(gdb) disassemble
...
   0xb661d3cc <+796>:   ldr     r3, [pc, #240]  ; (0xb661d4c0 <cv::initInterTab2D(int, bool)+1040>)
   0xb661d3ce <+798>:   vmov    s14, r8
=> 0xb661d3d2 <+802>:   vldr    d20, [r10]
   0xb661d3d6 <+806>:   vldr    d19, [r9]
   0xb661d3da <+810>:   add     r3, pc
   0xb661d3dc <+812>:   mov     r2, r7

========

I think this code in opencv3.mk is wrong (or maybe the ARM fpu config options are wrong):

ifeq ($(BR2_ARCH_IS_64):$(BR2_ARM_CPU_HAS_VFPV3),:y)
OPENCV3_CONF_OPTS += -DENABLE_VFPV3=ON
else
OPENCV3_CONF_OPTS += -DENABLE_VFPV3=OFF
endif

Apparently I have BR2_ARM_CPU_HAS_VFPV3 set (even though I only have the -D16 version), but when you pass -DENABLE_VFPV3=ON to OpenCV, it passes -mfpu=vfpv3 to the compiler and this is what's causing the SIGILL.
Comment 1 Arnout Vandecappelle 2019-07-01 21:47:09 UTC
> I think this code in opencv3.mk is wrong (or maybe the ARM fpu config options
> are wrong):
>
> ifeq ($(BR2_ARCH_IS_64):$(BR2_ARM_CPU_HAS_VFPV3),:y)
> OPENCV3_CONF_OPTS += -DENABLE_VFPV3=ON
> else
> OPENCV3_CONF_OPTS += -DENABLE_VFPV3=OFF
> endif
> 
> Apparently I have BR2_ARM_CPU_HAS_VFPV3 set (even though I only have the 
> -D16 version), but when you pass -DENABLE_VFPV3=ON to OpenCV, it passes 
> -mfpu=vfpv3 to the compiler and this is what's causing the SIGILL.

If the only thing that ENABLE_VFPV3 does is to set the -mfpu=... option, then we should just always set it to OFF, because we already pass that option.

Same for NEON, by the way.
Comment 2 Thomas Petazzoni 2019-11-27 22:20:51 UTC
So, I did a bit more research on this, but couldn't come to a useful conclusion. The handling of CPU optimizations in opencv3 is complicated. ENABLE_VFPV3 and ENABLE_NEON seem to be obsolete options, but the replacement CPU_BASELINE is not very clear.

With a VFPv3-D16 case, I get:

--   CPU/HW features:
--     Baseline:
--       requested:                 DETECT
--       disabled:                  VFPV3 NEON

which is quite expected.

When NEON is enabled as the FPU, I get:

--   CPU/HW features:
--     Baseline:                    NEON
--       requested:                 DETECT
--       disabled:                  VFPV3 NEON

So NEON seems to be detected, but it's also listed in the "disabled" features... which doesn't make a lot of sense.

Then, if I use VFPv3 as the FPU, I get:

--   CPU/HW features:
--     Baseline:
--       requested:                 DETECT
--       disabled:                  VFPV3 NEON

I.e, it doesn't detect that I have VFPv3.

(Of course the tests above are after removing ENABLE_VFPV3/ENABLE_NEON).