Problem with vectorization for Intel Compiler 11

[gbncn]

I have a simple test function as

int test(float* U2,float* U3){
for (int k=0; k<100; k++){
float *u2=U2+k*1000, *u3=U3+k*1000;

#pragma ivdep
#pragma vector aligned
#pragma vector always
for(int i=0; i<1000; i++) u3 = u2;
}
return 0;
}

Compile option as " -vec-report3 -xP -O3 "

Intel compiler 9.* vecotrized above codes. But for Intel compiler 11, it gave message like:

" loop was not vectorized: dereference too complex "

Can anyone point out what I did wrong here.

Thanks,

[mecej4]

You did not state the full version number of the compiler and your OS. On Suse-X64 11.3, with 11.1.073, I find:

[cpp] icc -vec-report3 -xHost -O3 -c vec.cpp 
vec.cpp(2): (col. 4) remark: loop was not vectorized: not inner loop.
vec.cpp(8): (col. 7) remark: LOOP WAS VECTORIZED.
[/cpp]

I do not see any problems with this.

[gbncn]

Thank you.

The compiler is 11.1, with 64 bit installation

The OS is redhat Linux version 2.6.9-89.0.20.ELsmp

When I run icpc -vec-report3 -c -o -v, the output is like:

Version 11.1
/apps/compilers/intel/Compiler/11.1/056/bin/intel64/mcpcom -_g -mP3OPT_inline_alloca -D__HONOR_STD -D__ICC=1110 -D__INTEL_COMP ILER=1110 -D_MT "-_Asystem(unix)" -D__ELF__ "-_Acpu(x86_64)" "-_Amachine(x86_64)" -D__INTEL_COMPILER_BUILD_DATE=20090827 -D__PTRD IFF_TYPE__=long "-D__SIZE_TYPE__=unsigned long" -D__WCHAR_TYPE__=int "-D__WINT_TYPE__=unsigned int" "-D__INTMAX_TYPE__=long int" "-D__UINTMAX_TYPE__=long unsigned int" -D__LONG_MAX__=9223372036854775807L -D__QMSPP_ -D__OPTIMIZE__ -D__NO_MATH_INLINES -D__NO_S TRING_INLINES -D__GNUG__=3 -D__GNUC__=3 -D__GNUC_MINOR__=4 -D__GNUC_PATCHLEVEL__=6 -D__NO_INLINE__ -D__i686 -D__i686__ -D__pentiu mpro -D__pentiumpro__ -D__pentium4 -D__pentium4__ -D__SSE2__ -D__tune_pentium4__ -D__MMX__ -D__SSE__ -D__LP64__ -D_LP64 -D_GNU_SO URCE=1 -D__DEPRECATED=1 -D__GXX_WEAK__=1 -D__GXX_ABI_VERSION=1002 "-D__USER_LABEL_PREFIX__= " -D__REGISTER_PREFIX__= -D__INTEL_RT TI__ -D__EXCEPTIONS=1 -D__unix__ -D__unix -D__linux__ -D__linux -D__gnu_linux__ -B -Dunix -Dlinux -D__x86_64 -D__x86_64__ -_k -_8 -_l -_a -_b --gnu_version=346 -_W5 --gcc-extern-inline -p --bool -tused -mGLOB_eh_linux -x --mspp --multibyte_chars --bool -mP1O PT_version=11.1-intel64 -mGLOB_diag_enable_disable=E:vec -mGLOB_diag_file=tune.diag -mP1OPT_print_version=FALSE -mP3OPT_use_mspp_ call_convention -mCG_use_gas_got_workaround=F -mP2OPT_align_option_used=TRUE "-mGLOB_options_string=-vec-report3 -c -o tune.o -v" -mGLOB_cxx_limited_range=FALSE -mGLOB_as_output_backup_file_name=/tmp/icpcb07sjvas_.s -mIPOPT_activate -mIPOPT_lite -mGLOB_machi ne_model=GLOB_MACHINE_MODEL_EFI2 -mGLOB_extended_instructions=0x8 -mIPOPT_args_in_regs=0 -mPGOPTI_value_profile_use=T -mP2OPT_hlo _level=2 -mP2OPT_hlo -mP2OPT_vec_verbose=3 -mP3OPT_debug_linenum_only -mGLOB_debug_format=GLOB_DEBUG_FORMAT_DWARF20 -mIPOPT_obj_o utput_file_name=tune.o "-mGLOB_linker_version=2.15.92.0.2 20040927" -mP3OPT_asm_target=P3OPT_ASM_TARGET_GAS -mGLOB_obj_output_fil e=tune.o -mGLOB_source_dialect=GLOB_SOURCE_DIALECT_C_PLUS_PLUS -mP1OPT_source_file_name=../tune.cc ../tune.cc
#include "..." search starts here:
#include <...> search starts here:
/apps/compilers/intel/Compiler/11.1/056/include/intel64
/usr/include/c++/3.4.6
/usr/include/c++/3.4.6/x86_64-redhat-linux
/usr/include/c++/3.4.6/backward
/apps/compilers/intel/Compiler/11.1/056/include/intel64
/apps/compilers/intel/Compiler/11.1/056/include
/usr/local/include
/usr/include
/usr/lib/gcc/x86_64-redhat-linux/3.4.6/include

End of search list.
../tune.cc(5): (col. 4) remark: loop was not vectorized: not inner loop.
../tune.cc(11): (col. 7) remark: loop was not vectorized: not inner loop.
../tune.cc(11): (col. 34) remark: loop was not vectorized: dereference too complex.

[mecej4 · Accepted solution]

You are running 11.1.056. If you are eligible for the update released in August (11.1.073), installing it may be the simplest solution.

[TimP]

Perhaps your pragmas aren't recognized unless they begin in column 1. ivdep is almost certainly required to vectorize if you don't declare the pointers with appropriate restrict syntax.

[gbncn]

Thank you, all. It is unlikely that I can convince system admin to install latest version yet. I also tried this code on Intel compiler 10.*, vectorization doesn't work either.

I am sure "#pragma" is recogized by the compiler since it would give me dependency message if it did not recognize ivdep.

[TimP]

Yes, I see that it does report dependency when ivdep is omitted, and restrict is not added. However, it still produces an identical call to __intel_fast_memcpy() even while reporting inability to vectorize.
I note that the current default Red Hat gcc-4.4 will vectorize this without pragmas, with the addition of restrict.

[gbncn]

Sys Op finally give the 073 a try, and it works. Thanks, everyone.