[Bug 2060619] [NEW] gcc-14 offload support for amd gfx103x and gfx110x gpus is broken

[oscarbg]

Public bug reported:

Hi,

for context bug was posted here first:
https://github.com/gcc-mirror/gcc/commit/1bf18629c54adf4893c8db5227a36e1952ee69a3#commitcomment-140648051

gcc Andrew Stubbs author mentions:
"Yes, it looks like the libgomp device-side library is missing. Probably not listed in the packaging script?"

using gcc-14:
gcc-14 --version
gcc-14 (Ubuntu 14-20240330-1ubuntu2) 14.0.1 20240330 (experimental) [master r14-9728-g6fc84f680d0]

using the code posted below as sum.c

testing building with offload support on gfx1036:

gcc-14 -fopenmp -foffload=amdgcn-amdhsa -foffload-options=-march=gfx1036
-fcf-protection=none sum.c -o sumamd1030.exe

fails with:
ld: error: incompatible mach: /usr/lib/gcc/x86_64-linux-gnu/14/accel/amdgcn-amdhsa/libgomp.a(parallel.o)
collect2: error: ld returned 1 exit status
gcn mkoffload: fatal error: x86_64-linux-gnu-accel-amdgcn-amdhsa-gcc-14 returned 1 exit status
compilation terminated.
lto-wrapper: fatal error: /usr/libexec/gcc/x86_64-linux-gnu/14//accel/amdgcn-amdhsa/mkoffload returned 1 exit status
compilation terminated.
/usr/bin/ld: error: lto-wrapper failed
collect2: error: ld returned 1 exit status


instead selecting gfx900 works:
gcc-14 -fopenmp -foffload=amdgcn-amdhsa -foffload-options=-march=gfx900 -fcf-protection=none sum.c -o sumamd900.exe

note gfx103x and gfx11xx targets are new to GCC-14..

inspecting files seems the error is due to package 
gcc-14-offload-amdgcn_14-20240330-1ubuntu2_amd64.deb
missing libgomp.so files in
\usr\lib\gcc\x86_64-linux-gnu\14\accel\amdgcn-amdhsa\gfx1036 folder
but having this file in gfx9xx folders for ex:
\usr\lib\gcc\x86_64-linux-gnu\14\accel\amdgcn-amdhsa\gfx906 folder


code used:
sum.c:
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <omp.h>

int main(int argc, char *argv[]) {
	long long int j, largeN;
	double sum = 0.0;
	clock_t tstart, tend;
	int num_dev;

        num_dev = omp_get_num_devices();

        printf("%d\n", num_dev);

        largeN = (long long int) atoll(argv[1]);

        printf("Large Number = %lld\n", largeN);

	tstart = clock();
#pragma omp parallel for private(j) reduction(+:sum)
	for(j=1; j<largeN; j++) {
		sum += (double) (1.0/j);
	}
	tend = clock();
	printf("Sum [1, ... , %lld] = %.6f\n", largeN, sum);
	printf("\nCompleted in %.2f seconds%45s\n\n", ( (double) (tend - tstart) ) / CLOCKS_PER_SEC, "");

	sum = 0.0;
	tstart = clock();
#pragma omp target map(to:largeN) map(tofrom:sum)
#pragma omp parallel for simd private(j) reduction(+:sum)
	for (j=1; j<largeN; j++) {
		sum += (double) (1.0/j);
	}
	tend = clock();
	printf("Sum [1, ... , %lld] = %.6f\n", largeN, sum);
	printf("\nCompleted in %.2f seconds%45s\n\n", ( (double) (tend - tstart) ) / CLOCKS_PER_SEC, "");

	return(0);
}

** Affects: gcc-14 (Ubuntu)
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Ubuntu
Foundations Bugs, which is subscribed to gcc-14 in Ubuntu.
https://bugs.launchpad.net/bugs/2060619

Title:
  gcc-14 offload support for amd gfx103x and gfx110x gpus  is broken

Status in gcc-14 package in Ubuntu:
  New

Bug description:
  Hi,

  for context bug was posted here first:
  https://github.com/gcc-mirror/gcc/commit/1bf18629c54adf4893c8db5227a36e1952ee69a3#commitcomment-140648051

  gcc Andrew Stubbs author mentions:
  "Yes, it looks like the libgomp device-side library is missing. Probably not listed in the packaging script?"

  using gcc-14:
  gcc-14 --version
  gcc-14 (Ubuntu 14-20240330-1ubuntu2) 14.0.1 20240330 (experimental) [master r14-9728-g6fc84f680d0]

  using the code posted below as sum.c

  testing building with offload support on gfx1036:

  gcc-14 -fopenmp -foffload=amdgcn-amdhsa -foffload-
  options=-march=gfx1036 -fcf-protection=none sum.c -o sumamd1030.exe

  fails with:
  ld: error: incompatible mach: /usr/lib/gcc/x86_64-linux-gnu/14/accel/amdgcn-amdhsa/libgomp.a(parallel.o)
  collect2: error: ld returned 1 exit status
  gcn mkoffload: fatal error: x86_64-linux-gnu-accel-amdgcn-amdhsa-gcc-14 returned 1 exit status
  compilation terminated.
  lto-wrapper: fatal error: /usr/libexec/gcc/x86_64-linux-gnu/14//accel/amdgcn-amdhsa/mkoffload returned 1 exit status
  compilation terminated.
  /usr/bin/ld: error: lto-wrapper failed
  collect2: error: ld returned 1 exit status

  
  instead selecting gfx900 works:
  gcc-14 -fopenmp -foffload=amdgcn-amdhsa -foffload-options=-march=gfx900 -fcf-protection=none sum.c -o sumamd900.exe

  note gfx103x and gfx11xx targets are new to GCC-14..

  inspecting files seems the error is due to package 
  gcc-14-offload-amdgcn_14-20240330-1ubuntu2_amd64.deb
  missing libgomp.so files in
  \usr\lib\gcc\x86_64-linux-gnu\14\accel\amdgcn-amdhsa\gfx1036 folder
  but having this file in gfx9xx folders for ex:
  \usr\lib\gcc\x86_64-linux-gnu\14\accel\amdgcn-amdhsa\gfx906 folder

  
  code used:
  sum.c:
  #include <stdio.h>
  #include <stdlib.h>
  #include <time.h>
  #include <omp.h>

  int main(int argc, char *argv[]) {
  	long long int j, largeN;
  	double sum = 0.0;
  	clock_t tstart, tend;
  	int num_dev;

          num_dev = omp_get_num_devices();

          printf("%d\n", num_dev);

          largeN = (long long int) atoll(argv[1]);

          printf("Large Number = %lld\n", largeN);

  	tstart = clock();
  #pragma omp parallel for private(j) reduction(+:sum)
  	for(j=1; j<largeN; j++) {
  		sum += (double) (1.0/j);
  	}
  	tend = clock();
  	printf("Sum [1, ... , %lld] = %.6f\n", largeN, sum);
  	printf("\nCompleted in %.2f seconds%45s\n\n", ( (double) (tend - tstart) ) / CLOCKS_PER_SEC, "");

  	sum = 0.0;
  	tstart = clock();
  #pragma omp target map(to:largeN) map(tofrom:sum)
  #pragma omp parallel for simd private(j) reduction(+:sum)
  	for (j=1; j<largeN; j++) {
  		sum += (double) (1.0/j);
  	}
  	tend = clock();
  	printf("Sum [1, ... , %lld] = %.6f\n", largeN, sum);
  	printf("\nCompleted in %.2f seconds%45s\n\n", ( (double) (tend - tstart) ) / CLOCKS_PER_SEC, "");

  	return(0);
  }

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/gcc-14/+bug/2060619/+subscriptions