search

JvanKatwijk / qt-dab

Qt-DAB, a general software DAB (DAB+) decoder with a (slight) focus on showing the signal
http://www.sdr-j.tk
GNU General Public License v2.0
296 stars 61 forks source link

Raspberry Pi CMake build broken: emmintrin.h: No such file or directory #153

Closed sjlongland closed 4 years ago

sjlongland commented 4 years ago 
[ 69%] Building CXX object CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-jan/viterbi-handler.cpp.o
[ 69%] Building CXX object CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-spiral/viterbi-spiral.cpp.o
[ 70%] Building C object CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-spiral/spiral-sse.c.o
/tmp/qt-dab/src/support/viterbi-spiral/spiral-sse.c:23:10: fatal error: emmintrin.h: No such file or directory
 #include <emmintrin.h>
          ^~~~~~~~~~~~~
compilation terminated.
make[2]: *** [CMakeFiles/qt-dab-3.21-Beta.dir/build.make:1275: CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-spiral/spiral-sse.c.o] Error 1

It would appear it tries to drag in the x86 SSE implementation of Viterbi, when it should be pulling in a ARM NEON version instead… digging around, I found the following diff gets it compiling on the Pi (but will break x86).

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 460e805..8673db7 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -14,7 +14,7 @@ endif()
 if(MINGW)
     add_definitions ( -municode)
 else (MINGW)
-       add_definitions (-DSSE_AVAILABLE)
+       add_definitions (-DNEON_AVAILABLE)
 endif()

 IF(EXISTS (".git") AND IS_DIRECTORY (".git"))
@@ -267,7 +267,7 @@ add_definitions (-DPRESET_NAME)
             ./includes/support/ensemble-printer.h
             ./includes/support/viterbi-jan/viterbi-handler.h
             ./includes/support/viterbi-spiral/viterbi-spiral.h
-              ./src/support/viterbi-spiral/spiral-sse.h
+            ./src/support/viterbi-spiral/spiral-neon.h
             ./includes/support/preset-handler.h
             ./includes/support/presetcombobox.h
             ./includes/support/smallqlistview.h
@@ -358,7 +358,7 @@ add_definitions (-DPRESET_NAME)
             ./src/support/history-handler.cpp
             ./src/support/viterbi-jan/viterbi-handler.cpp
             ./src/support/viterbi-spiral/viterbi-spiral.cpp
-            ./src/support/viterbi-spiral/spiral-sse.c
+            ./src/support/viterbi-spiral/spiral-neon.c
             ./src/scopes-qwt6/iqdisplay.cpp 
             ./spectrum-viewer/spectrum-viewer.cpp 
             ./correlation-viewer/correlation-viewer.cpp

I think the real fix is to define the implementation of the Viterbi algorithm being used someplace else based on processor architecture, and feed that in here. I'll slot it into the back-log somewhere, but just documenting what I've seen in case others hit the same issue. :-)

JvanKatwijk commented 4 years ago

For x64 linux system uncomment SSE

For any other system comment SSE out and uncomment NO_SSE

CONFIG += SSE

CONFIG += NO_SSE

Use the NO_SSE, SSE is just for x64

Op zo 19 jan. 2020 om 09:32 schreef Stuart Longland < notifications@github.com>:

[ 69%] Building CXX object CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-jan/viterbi-handler.cpp.o

[ 69%] Building CXX object CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-spiral/viterbi-spiral.cpp.o

[ 70%] Building C object CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-spiral/spiral-sse.c.o

/tmp/qt-dab/src/support/viterbi-spiral/spiral-sse.c:23:10: fatal error: emmintrin.h: No such file or directory

include 

      ^~~~~~~~~~~~~

compilation terminated.

make[2]: *** [CMakeFiles/qt-dab-3.21-Beta.dir/build.make:1275: CMakeFiles/qt-dab-3.21-Beta.dir/src/support/viterbi-spiral/spiral-sse.c.o] Error 1

It would appear it tries to drag in the x86 SSE implementation of Viterbi, when it should be pulling in a ARM NEON version instead… digging around, I found the following diff gets it compiling on the Pi (but will break x86).

diff --git a/CMakeLists.txt b/CMakeLists.txt

index 460e805..e0fca6e 100755

--- a/CMakeLists.txt

+++ b/CMakeLists.txt

@@ -267,7 +267,7 @@ add_definitions (-DPRESET_NAME)

         ./includes/support/ensemble-printer.h

         ./includes/support/viterbi-jan/viterbi-handler.h

         ./includes/support/viterbi-spiral/viterbi-spiral.h

  • ./src/support/viterbi-spiral/spiral-sse.h

  • ./src/support/viterbi-spiral/spiral-neon.h

         ./includes/support/preset-handler.h

     ./includes/support/presetcombobox.h

     ./includes/support/smallqlistview.h

@@ -358,7 +358,7 @@ add_definitions (-DPRESET_NAME)

         ./src/support/history-handler.cpp

         ./src/support/viterbi-jan/viterbi-handler.cpp

         ./src/support/viterbi-spiral/viterbi-spiral.cpp

  • ./src/support/viterbi-spiral/spiral-sse.c

  • ./src/support/viterbi-spiral/spiral-neon.c

         ./src/scopes-qwt6/iqdisplay.cpp

     ./spectrum-viewer/spectrum-viewer.cpp

     ./correlation-viewer/correlation-viewer.cpp

I think the real fix is to define the implementation of the Viterbi algorithm being used someplace else based on processor architecture, and feed that in here. I'll slot it into the back-log somewhere, but just documenting what I've seen in case others hit the same issue. :-)

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/JvanKatwijk/qt-dab/issues/153?email_source=notifications&email_token=ACCPHQC4GCQST3ZX6276IR3Q6QFY7A5CNFSM4KIX4LF2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4IHFDXDQ, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACCPHQC37NXXW5U44LKSAW3Q6QFY7ANCNFSM4KIX4LFQ .

-- Jan van Katwijk

sjlongland commented 4 years ago

Where's the CONFIG line in CMakeLists.txt?

JvanKatwijk commented 4 years ago

good question, I am talking about the ".pro" file.

Op zo 19 jan. 2020 om 10:07 schreef Stuart Longland < notifications@github.com>:

Where's the CONFIG line in CMakeLists.txt?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/JvanKatwijk/qt-dab/issues/153?email_source=notifications&email_token=ACCPHQGT7KEEGHGACSVF2SDQ6QJ3VA5CNFSM4KIX4LF2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEJKMYVI#issuecomment-575982677, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACCPHQAABLUZGGX667E2VEDQ6QJ3VANCNFSM4KIX4LFQ .

-- Jan van Katwijk

sjlongland commented 4 years ago

For what it's worth cmake upstream haven't got a "standard" way to do this yet: https://gitlab.kitware.com/cmake/cmake/issues/18445

A "good-enough" work-around on Unix-like platforms might be to look at uname -m … if it's x86_64, there's a fair bet it'll support SSE (which was introduced with the Pentium III and the AMD Athlon XP/Duron)… if it's armhf, then probably try building with NEON.

After making those earlier changes to CMakeLists.txt I got some compiler errors regarding NEON instructions…

/usr/lib/gcc/arm-linux-gnueabihf/8/include/arm_neon.h:6733:1: error: inlining failed in call to always_inline ‘vdupq_n_s32’: target specific option mismatch
 vdupq_n_s32 (int32_t __a)
 ^~~~~~~~~~~
In file included from /tmp/qt-dab/src/support/viterbi-spiral/spiral-neon.c:27:
/tmp/qt-dab/src/support/viterbi-spiral/SSE2NEON.h:312:33: note: called from here

But that's progress… looks very similar to https://github.com/kcat/openal-soft/issues/54 so I'll try seeing if a few C compiler flags will crack this nut and get the CMake build system working again.

sjlongland commented 4 years ago

Above patch + build with options:

cmake -DRTLSDR=ON -DRTLTCP=ON -DHACKRF=ON -DCMAKE_C_FLAGS="-mfpu=neon" .

got things to compile. I have Qt-DAB 3.21 beta working on the Pi3 now. I think CMake should have enough intelligence to at least tell us what OS and CPU architecture we're compiling for, and thus we can make some sane assumptions. (i.e. pretty sure every x86-based Apple machine ever made supports SSE, most Windows machines running current versions would too, as would anything AMD64.)

I'll do some reading up on the CMake docs and see what can be figured out.

JvanKatwijk commented 4 years ago

I added as options to the CMakeLists.txt file VITERBI_SSE VITERBI_NEON When e.g. adding -DVITERBI_NEON=ON to the command line, the right files for inclusion are selected and the proper define is set

sjlongland commented 4 years ago

On 21/1/20 6:36 am, JvanKatwijk wrote:

I added as options to the CMakeLists.txt file VITERBI_SSE VITERBI_NEON When e.g. adding -DVITERBI_NEON=ON to the command line, the right files for inclusion are selected and the proper define is set

Fantastic. I guess that reduces the urgency on the autodetection aspect: we can just document "on ARM compile with -DVITERBI_NEON" in the meantime.

Later on, we should be able to figure out which is needed and define the appropriate parameter automagically.

Is there a non-optimised (i.e. non-SIMD) Viterbi implementation that is selected if neither parameter is specified?

JvanKatwijk commented 4 years ago

If nothing is specified a non-sse implementation is chosen. Note that these implementations are for the FIC blocks, not for the audio decoding. The "optimized" versions are generated for a limited size of the input, the audio segments often extend this length

Op ma 20 jan. 2020 om 21:49 schreef Stuart Longland < notifications@github.com>:

On 21/1/20 6:36 am, JvanKatwijk wrote:

I added as options to the CMakeLists.txt file VITERBI_SSE VITERBI_NEON When e.g. adding -DVITERBI_NEON=ON to the command line, the right files for inclusion are selected and the proper define is set

Fantastic. I guess that reduces the urgency on the autodetection aspect: we can just document "on ARM compile with -DVITERBI_NEON" in the meantime.

Later on, we should be able to figure out which is needed and define the appropriate parameter automagically.

Is there a non-optimised (i.e. non-SIMD) Viterbi implementation that is selected if neither parameter is specified?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/JvanKatwijk/qt-dab/issues/153?email_source=notifications&email_token=ACCPHQDQBMDKEATOBFT5ICDQ6YE6ZA5CNFSM4KIX4LF2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEJNZBUY#issuecomment-576426195, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACCPHQG6MU5T5NUIIVXK2Q3Q6YE6ZANCNFSM4KIX4LFQ .

-- Jan van Katwijk