Compiling error: expected "#pragma acc" clause


I’m compiling an openacc code with gcc, but i get the following error:

/home/cesar/Documents/dso/src/main_dso_pangolin.cpp:418:23: error: expected ‘#pragma acc’ clause before ‘loop’
#pragma acc parallel loop

But, as you can see, i do have that clause before loop. I have the code like this:

#pragma acc parallel loop


Why is this generating an error?

I’m compiling with cmake and i pass the following flags:

“${SSE_FLAGS} -O3 -g -std=c++0x -march=native -fopenacc -fopt-info-note-omp”


Hi Caap,

I don’t use g++ with OpenACC very much, but I just tried using GNU 7.3 on a simple OpenACC parallel loop and it worked fine.

Can you please let me know what GNU version you’re using and give a complete reproducing example of the error?

Though, I do see the following similar error on Bugzilla. Perhaps the GNU folks fix the problem already?


Hello Mat,

I actually dont have this error anymore. I followed this tutorial to install GCC with offloading support:

And i created a simple script to test it, looking like this:

#pragma acc parallel loop
		for (int j = 0; j < 10; j++) {
		  x[j] = j;
		  y[j] = -j;

I can compile and execute withou error. Although, i don’t think GPU is being used at all, i dont get parallelization messages and i even ran the profiler of pgi to check and indeed no GPU is being used.

Do you have any idea why?

Also, in that same link check the last comment on comment section, that is me and has more information.


I’m not sure. I just tried one of my codes and GNU was able to compile and run it on my P100, and I was able to get a GPU profile.

I see that you posted a question on Krister’s blog. Hopefully they can help.

If not, please post a full reproducing example as well as the compile and run commands that you’re using. Also, please let me know what kind of NVIDIA device you’re using and which CUDA version is installed.

Here’s my run using SPEC ACCEL’s 303.ostencil benchmark:

% gcc --version
gcc (GCC) 7.3.0
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO

% sh -x make.out
+ gcc -c -o main.o -DSPEC -DSPEC_ACCEL -DNDEBUG -I./ -I./pbcommon_sources -O3 -fno-fast-math -foffload=-lm -lm -fopenacc main.c
+ gcc -c -o file.o -DSPEC -DSPEC_ACCEL -DNDEBUG -I./ -I./pbcommon_sources -O3 -fno-fast-math -foffload=-lm -lm -fopenacc file.c
+ gcc -c -o kernels.o -DSPEC -DSPEC_ACCEL -DNDEBUG -I./ -I./pbcommon_sources -O3 -fno-fast-math -foffload=-lm -lm -fopenacc kernels.c
+ gcc -c -o pbcommon_sources/parboil.o -DSPEC -DSPEC_ACCEL -DNDEBUG -I./ -I./pbcommon_sources -O3 -fno-fast-math -foffload=-lm -lm -fopenacc pbcommon_sources/parboil.c
+ gcc -O3 -fno-fast-math -foffload=-lm -lm -fopenacc main.o file.o kernels.o pbcommon_sources/parboil.o -lm -o ostencil_exe

% pgprof ./ostencil_exe -o 512x512x98.out -- 512 512 98 200
CPU-based 7 points stencil codes****
Original version by Li-Wen Chang <> and I-Jui Sung<>
This version maintained by Chris Rodrigues  ***********
==48574== PGPROF is profiling process 48574, command: ./ostencil_exe -o 512x512x98.out -- 512 512 98 200
IO        : 1.146225
Compute   : 2.969159
Timer Wall Time: 4.115386
==48574== Profiling application: ./ostencil_exe -o 512x512x98.out -- 512 512 98 200
==48574== Profiling result:
Time(%)      Time     Calls       Avg       Min       Max  Name
 94.01%  839.47ms       200  4.1973ms  4.1162ms  4.4123ms  cpu_stencil$_omp_fn$0
  3.75%  33.443ms      1002  33.375us     831ns  16.419ms  [CUDA memcpy HtoD]
  2.12%  18.934ms         2  9.4671ms  9.2627ms  9.6714ms  [CUDA memcpy DtoH]
  0.13%  1.1472ms       200  5.7350us  4.9890us  11.198us  [CUDA memcpy HtoH]



And what did you do in order to install gcc with offloading capability? Did you do something similar with those steps on the blog?

My GPU is a NVIDIA GTX 950M.

nvcc --version gives me:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Sep__1_21:08:03_CDT_2017
Cuda compilation tools, release 9.0, V9.0.176