Error compiling PAPI with intel_gpu

[Jordi Alcaraz]

I'm trying to install PAPI with the component "intel_gpu" but I'm unable to, the followinf error appears:
g++ -g -fpic -g -diag-disable 188,869,271  -DPAPI_NUM_COMP=4  -g -DDEBUG  -Icomponents/intel_gpu/internal -Icomponents/intel_gpu/internal/inc -D_GLIBCXX_USE_CXX11_ABI=1 -I/hsa/include -I/hsa/include/hsa -I/rocm_smi/include -I/rocm_smi/include/rocm_smi -I/include  -Icomponents/intel_gpu/internal/inc -I/usr/include/level_zero/ -o GPUMetricInterface.o -c components/intel_gpu/internal/src/GPUMetricInterface.cpp
g++: error: 188,869,271: No such file or directory

In the attached files you can check the outputs of configure and make

My steps to configure and compile PAPI are:

• source /opt/intel/oneapi/setvars.sh
• export INTEL_L0_HEADERS=/usr/include/level_zero/
• cd papi-7.0.0/src
• make clean
• ./configure --with-components="intel_gpu"
• make
  
  What could be the problem?
  

[Servat, Harald]

Hola Jordi,

 

  I think that the problem comes from the fact that -diag-disable is a flag for the classic Intel compilers (icc and icpc, not the new icx and icpx). Can you compile PAPI using these compilers? If not, can you tweak the PAPI Makefiles or configure scripts in order not to use “-diag-disable 188,869,271” ?

 

Salut

--
You received this message because you are subscribed to the Google Groups "ptools-perfapi" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ptools-perfap...@icl.utk.edu.
To view this discussion on the web visit https://groups.google.com/a/icl.utk.edu/d/msgid/ptools-perfapi/e634315f-bb77-4b54-b8eb-6a3e89307f40n%40icl.utk.edu.

-----------------------------------------------------------
Intel Corporation Iberia, S.A.
Registered Office: Torre Picasso, 25th Floor,
Plaza Pablo Ruiz Picasso, no. 1, 28020  Madrid

Este mensaje se dirige exclusivamente a su destinatario y puede
contener informacion privilegiada o confidencial. Si no es vd.
el destinatario indicado, queda notificado de que la lectura,
utilizacion, divulgacion y,o copia sin autorizacion esta prohibida
en virtud de la legislacion vigente. Si ha recibido este mensaje por
error, le rogamos que nos lo communique inmediatamente por
esta misma via y proceda a su destruccion.

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.

[Jordi Alcaraz]

Hola Harald,

I removed that part from the configure and tried to compile again (I may have not understood what you wrote as it is using icc and not icx).

With this change, a lot lines similar to the following appeared:
g++ -g  -DPAPI_NUM_COMP=4 -DDEBUG  -I../../../testlib -o gpu_query_gemm gpu_query_gemm.o gpu_common_utils.o -L/usr/lib -lze_loader ../../../libpapi.a -lsupc++ -lstdc++ -ldl -lstdc++
/usr/bin/ld: ../../../libpapi.a(papi.o): relocation R_X86_64_32 against `.rodata.str1.4' can not be used when making a PIE object; recompile with -fPIE
/usr/bin/ld: ../../../libpapi.a(papi_internal.o): relocation R_X86_64_32 against `.rodata.str1.4' can not be used when making a PIE object; recompile with -fPIE
/usr/bin/ld: ../../../libpapi.a(papi_hl.o): relocation R_X86_64_32 against `.rodata.str1.4' can not be used when making a PIE object; recompile with -fPIE
/usr/bin/ld: ../../../libpapi.a(extras.o): relocation R_X86_64_32 against `.rodata.str1.4' can not be used when making a PIE object; recompile with -fPIE

Followed by:
usr/bin/ld: gpu_common_utils.o: in function `parseInputParam':
/storage/users/jalcaraz/papi7/papi-7.0.0/src/components/intel_gpu/tests/gpu_common_utils.c:103: undefined reference to `__intel_sse2_strdup'
/usr/bin/ld: gpu_common_utils.o: in function `parseMetricList':
/storage/users/jalcaraz/papi7/papi-7.0.0/src/components/intel_gpu/tests/gpu_common_utils.c:57: undefined reference to `__intel_sse2_strtok'
/usr/bin/ld: /storage/users/jalcaraz/papi7/papi-7.0.0/src/components/intel_gpu/tests/gpu_common_utils.c:65: undefined reference to `__intel_sse2_strtok'

In the case of trying to use icpx:

• export CC=icx
• export CPP=icpx

• make clean
•  ./configure --with-components="intel_gpu" 

Outputs:

• configure: error: C preprocessor "icpx" fails sanity check

[Servat, Harald]

Hola,

 

  Looks like now you’re in a moment in which you have some object files compiled with Intel classic compilers (icc/icpc) and some others with GNU. I’m not sure if that would work without any further change (especially given the __intel_sse2_* references).

 

  Since it looks that you have icc and icpc, is there any way you can force the module to be compiled with these compilers rather than using GNU compiler?

 

Salut,

[Jordi Alcaraz]

It fails the sanity check with icc/icpc, you can find attached the config.log where the compilations of the sanity check fail.
The problems seems to be after line 152, where it literally says "choke me" when trying to compile with ifort and later it also fails with icpc.

[Servat, Harald]

Hello,

 

  @Jordi Alcaraz – just checked on my end . The following worked for me. You may want to start with a clean package – or use make distclean before doing this in order to clean your object files.

 

  FC=gfortran CC=gcc CXX=g++ ./configure --with-components="intel_gpu"

 

  @PAPI team – please note that the following configure line fails because the configure below does not keep track of CXX and some of the intel_gpu files are written in C++ and since CXX is not captured, g++ (rather than icpc) is used. That leads to flag collisions between icc/icpc and gnu compiler flags.

 

  FC=ifort CC=icc CXX=icpc ./configure --with-components="intel_gpu" --prefix=$PWD/install

 

  I was looking if PAPI could support it. Basically, configure.in would need an AC_PROG_CXX statement similar to AC_PROG_CC and then pass the CXX compiler to the necessary Makefiles. However, I don’t see how components/intel_gpu/Rules.intel_gpu can use that CXX definition.

 

Best,

To view this discussion on the web visit https://groups.google.com/a/icl.utk.edu/d/msgid/ptools-perfapi/91c70782-679c-4228-930b-7c83ed0ff529n%40icl.utk.edu.

[Jordi Alcaraz]

Hello Harald,

Now it compiles and installs without problems.
However, there seems to be another problem, as I can execute code in the GPU:

• cd papi-7.0.0/src/components/intel_gpu/tests
• ./gpu_gemm

Output:
Level Zero Matrix Multiplication (matrix size: 1024 x 1024, repeats 4 times)
Target device: Intel(R) Iris(R) Xe MAX Graphics [0x4905]
Matrix multiplication time: 0.0161849 sec
Matrix multiplication time: 0.0154192 sec
Matrix multiplication time: 0.0154332 sec
Matrix multiplication time: 0.0154433 sec
Total execution time: 0.191503 sec




But after I  do as the readme.txt says:

• export ZET_ENABLE_METRICS=1
• export ZET_ENABLE_API_TRACING_EXP=1
• cat /proc/sys/kernel/perf_event_paranoid
  -1
• cat  /proc/sys/dev/i915/perf_stream_paranoid
  0

And execute the example again:
./gpu_gemm
gpu_gemm: gpu_query_gemm.cc:357: int main(int, char**): Assertion `status == ZE_RESULT_SUCCESS' failed.
Aborted (core dumped)

The problem comes from enabling ZET_ENABLE_METRICS


Thank you,
Jordi Alcaraz

[Peinan Zhang]

Hi, Jordi,

Can you unset ZET_ENABLE_API_TRACING_EXP to run "gpu_gemm"?  You can have ZET_ENABLE_METRICS=1 to enable metrics collection through a time-based approach.

ZET_ENABLE_API_TRACING_EXP setting is to enable metrics query through API tracing, which may not well support in some early HWs and some early kernel versions.

Thanks,

--Peinan

To view this discussion on the web visit https://groups.google.com/a/icl.utk.edu/d/msgid/ptools-perfapi/0020e847-de0e-440e-82b4-8d963088e4can%40icl.utk.edu.

[Jordi Alcaraz]

Hi Peinan,

The problem I see is that with ZET_ENABLE_METRICS=1, the output is

./gpu_gemm
gpu_gemm: gpu_query_gemm.cc:357: int main(int, char**): Assertion `status == ZE_RESULT_SUCCESS' failed.
Aborted (core dumped)

In the same way, the output of papi_component_avail is:


InitMetricDevices failed, device does not exist or cannot find dependent libraries, abort
Available components and hardware information.
--------------------------------------------------------------------------------
PAPI version             : 7.0.0.0
Operating system         : Linux 5.4.48-xe-max
Vendor string and code   : GenuineIntel (1, 0x1)
Model string and code    : Intel(R) Xeon(R) Gold 6348H CPU @ 2.30GHz (85, 0x55)
CPU revision             : 11.000000
CPUID                    : Family/Model/Stepping 6/85/11, 0x06/0x55/0x0b
CPU Max MHz              : 4200
CPU Min MHz              : 1000
Total cores              : 192
SMT threads per core     : 2
Cores per socket         : 24
Sockets                  : 4
Cores per NUMA region    : 48
NUMA regions             : 4
Running in a VM          : no
Number Hardware Counters : 10
Max Multiplex Counters   : 384
Fast counter read (rdpmc): yes
--------------------------------------------------------------------------------

Compiled-in components:
Name:   perf_event              Linux perf_event CPU counters
Name:   perf_event_uncore       Linux perf_event CPU uncore and northbridge
Name:   intel_gpu               Intel GPU performance metrics
   \-> Disabled: The intel_gpu component does not detect metrics device.
Name:   sysdetect               System info detection component

Active components:
Name:   perf_event              Linux perf_event CPU counters
                                Native: 175, Preset: 59, Counters: 10
                                PMUs supported: ix86arch, perf, perf_raw, clx

Name:   perf_event_uncore       Linux perf_event CPU uncore and northbridge
                                Native: 3350, Preset: 0, Counters: 188
                                PMUs supported: rapl, skx_unc_cha0, skx_unc_cha1, skx_unc_cha2, skx_unc_cha3, skx_unc_cha4
                                                skx_unc_cha5, skx_unc_cha6, skx_unc_cha7, skx_unc_cha8, skx_unc_cha9
                                                skx_unc_cha10, skx_unc_cha11, skx_unc_cha12, skx_unc_cha13, skx_unc_cha14
                                                skx_unc_cha15, skx_unc_cha16, skx_unc_cha17, skx_unc_cha18, skx_unc_cha19
                                                skx_unc_cha20, skx_unc_cha21, skx_unc_cha22, skx_unc_cha23, skx_unc_iio0
                                                skx_unc_iio1, skx_unc_iio2, skx_unc_iio3, skx_unc_iio4, skx_unc_iio5, skx_unc_imc0
                                                skx_unc_imc1, skx_unc_imc2, skx_unc_imc3, skx_unc_imc4, skx_unc_imc5, skx_unc_m2m0
                                                skx_unc_m2m1, skx_unc_m3upi0, skx_unc_m3upi1, skx_unc_pcu, skx_unc_ubo
                                                skx_unc_upi0, skx_unc_upi1, skx_unc_upi2

Name:   sysdetect               System info detection component
                                Native: 0, Preset: 0, Counters: 0


--------------------------------------------------------------------------------

Could it be possible that in the same way you say in some early versions the tracing is not supported, the metrics may not be supported?

Best regards,
Jordi Alcaraz

[Servat, Harald]

Hello,

 

    Please note the following message from your output.

 

Target LEVEL0 RTL -->   error: Double type is not supported on this platform.
Target LEVEL0 RTL -->   in kernel: 'compute_target(double*, double*, double*, int, int, int)'
Target LEVEL0 RTL -->   error: backend compiler failed build.

 

    DG1 does not support double precision natively. It supports double precision through emulation, though. So my suggestion is to follow one of the following:

• Convert the test to use single-precision rather than double-precision
• Try a different test
• Try emulation (although the code may run much slower) – you can get some directions from the following link
• https://community.intel.com/t5/GPU-Compute-Software/Iris-Xe-Graphics-can-t-compute-double-precision-floating-point/m-p/1281976

 

Best,

[Jordi Alcaraz]

Hi Harald,

The problem I have is that after setting ZET_ENABLE_METRICS=1, the code fails with:

InitMetricDevices failed, device does not exist or cannot find dependent libraries, abort

Could it be possible that in the same way you say in some early versions the tracing is not supported, the metrics may not be supported?

Thank you,
Jordi Alcaraz

[Servat, Harald]

Hello,

 

  I’m sorry to say that I’m not sure about the support for metrics in Xe Max. Peinan may know better.

 

  Note that Peinan also pointed out that you may could try with a more recent kernel.