I recompiled with the modified CPU detection, the SSE3 syntax error
and with the default configure options (after a distclean) and I get the
_incorrect_ result.
On 29/12/2019 22:15, Momtchil Momtchev wrote:
I get the same wrong results with and without Docker, with and
without the SSE/SSE3/AVX when using the configure script. I also tried
removing -O2. Is the SSE3 unpacking in GDAL or in libgeotiff?
On 29/12/2019 21:51, Even Rouault wrote:
Is it possible that this is a compiler issue? Anyone else that
can
test this on an AMD CPU? Mine are old Phenom II X4 965s.
Perhaps rather a runtime issue. I suspect this is related to a SSSE3
optimization that has been added in GDAL 2.2 that is used in this use
case
(unpacking of byte buffer with a 2 byte stride to packed buffer)
From what I found on the net, I think this CPU doesn't support the
SSSE3
instruction set (SSSE3 with three S, not to be confused with SSE3...)
Can you check the following ?
cat /proc/cpuinfo | grep ssse3
What is strange is that GDAL does have a runtime check to detect if
SSSE3 is
available, so I'm not sure why you would get that issue. Unless that
Docker
would wrongly expose SSSE3, but Docker is not a VM technology so it
shouldn't
mess up with CPU capability discoveries. This is weird.
Can you try the following ?
docker run --rm osgeo/gdal su -c "apt install -y cpuid; cpuid" | grep
SSSE3
In a DEBUG build of GDAL, you could disable at runtime the SSSE3
optimization
by defining the environment variable/configuration option
GDAL_USE_SSSE3 to
NO, but this will not work on the release builds available on Docker.
--
Momtchil Momtchev <[email protected]>
_______________________________________________
gdal-dev mailing list
[email protected]
https://lists.osgeo.org/mailman/listinfo/gdal-dev
