Observed issues › nested/nvhpc-c-pointee-before-parent-in-one-map-list-copy-fault

Appears to behave incorrectlyRun-time error · C · OpenMP · Open

NVHPC: pointee listed before its parent in one map list faults in separate memory

Legality: contested. An independent reading finds the result not required: r->nx[0:1] is mapped before r[0:1] on the same construct. The attached-pointer rule applies when a corresponding pointer for the base pointer r->nx exists upon entry to the region. Here r->nx lives in r[0:1], which is created by this same construct, so the rule does not clearly guarantee attachment. The device r->nx may hold the host address, so acceptance is not clearly required and the observed illegal address is not clearly a bug. How this is checked

In separate memory NVHPC faults when a map list names a pointee before the structure holding the pointer.

Affected variants

Compiler	MI300A gfx942 unified memory (XNACK 1)	MI300A gfx942 copy (XNACK 0)	MI250X gfx90a copy (XNACK 0)	MI300X gfx942 copy (XNACK 0)	MI350X gfx950 copy (XNACK 0)	A100 sm_80 separate memory	A100 sm_80 managed memory
AFAR 24.3.0	correct	correct	unknown not run	correct	correct	n/a	n/a
CCE 19.0.0	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
CCE 20.0.2	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
CCE 21.0.3	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
CCE 21.0.4	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
NVHPC 25.11	n/a	n/a	n/a	n/a	n/a	behaves incorrectly	correct
NVHPC 26.9	n/a	n/a	n/a	n/a	n/a	behaves incorrectly	correct
GCC 16.2.0	correct	correct	correct	correct	correct	correct	n/a

"behaves incorrectly" / "not supported": this issue, confirmed in up to four environments (another build directory, another node, a repeat). "fails another way": a different failure, not counted as this issue. "unknown": not run there, or the runs disagree.

Messages

Result	Compiler and GPU	Message
behaves incorrectly	NVHPC 25.11, A100 separate NVHPC 26.9, A100 separate	show Failing in Thread:1 Accelerator Fatal Error: call to cuMemcpyDtoHAsync returned error 700 (CUDA_ERROR_ILLEGAL_ADDRESS): Illegal address during kernel execution File: <dir> Function: main:31 Line: 44

Reproduce

NVHPC 25.11, A100 separate

export LD_LIBRARY_PATH=$NVHPC_25_11/compilers/lib
nvc -mp=gpu -gpu=cc80 -gpu=mem:separate -O2 repro.c -o repro
OMP_TARGET_OFFLOAD=mandatory ./repro

NVHPC 26.9, A100 separate

export LD_LIBRARY_PATH=$NVHPC_26_9/compilers/lib
nvc -mp=gpu -gpu=cc80 -gpu=mem:separate -O2 repro.c -o repro
OMP_TARGET_OFFLOAD=mandatory ./repro

AFAR 24.3.0, MI300A unified (runs correctly)

$AFAR/bin/amdclang -fopenmp --offload-arch=gfx942 -O2 repro.c -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro

Program (48 lines)

// # Finding (nested-data): NVHPC, separate memory: a target construct that lists the pointee of a pointer member before the
// # structure that holds the pointer fails with CUDA_ERROR_ILLEGAL_ADDRESS (the other order is correct)
//
// - toolchains and modes that fail: nvhpc (NVHPC 25.11), nvhpc269 (NVHPC 26.9); modes: copy; -O2
// - build: nvc -mp=gpu -gpu=cc80 -gpu=mem:separate -O2
// - failure: run_error
// - observed: `exit_status: 1; without mandatory: exit_status: 1`; identical signature on 2 of 2 failing cells
// - expected: compiles and prints `tot= 1638` with `check=pass`
// - rule: attachment is decided on the storage of the whole construct: OpenMP 6.0 §7.9.6 p.284 l.8-10: "for which a
//   corresponding pointer exists in the device data environment after all mappable storage blocks are mapped" (the
//   OpenMP 5.2 text at §5.8.3 p.152 l.5-7 is worded for the state on entry and is weaker; the finding rests on the 6.0
//   wording), so the order of the clauses of one construct does not change the result
// - same-compiler control: controls/v2-ctl-parent-first.c (one change from this file; correct on nvhpc (NVHPC 25.11), nvhpc269
//   (NVHPC 26.9); rows results/nested/red/red13.jsonl, results/nested/red/tuo12.jsonl, results/nested/red/wgs.jsonl,
//   results/nested/red/wnv.jsonl); controls/v4-ctl-embedded-member.c (one change from this file; correct on nvhpc (NVHPC 25.11),
//   nvhpc269 (NVHPC 26.9); rows results/nested/red/red13.jsonl, results/nested/red/tuo12.jsonl, results/nested/red/wgs.jsonl,
//   results/nested/red/wnv.jsonl)
// - other families: afar (AFAR 24.3.0), cce19 (CCE 19.0.0), cce20 (CCE 20.0.2), cce21 (CCE 21.0.3), cce21_4 (CCE 21.0.4), gcc_g90
//   (GCC 16.2.0), gcc_sm80 (GCC 16.2.0) run the file correctly (rows results/nested/red/red13.jsonl,
//   results/nested/red/tuo12.jsonl, results/nested/red/wgs.jsonl)
// - gate and confirmer: 2 of 2 cells reproduce in all four environments (a-d; (c) node not recorded for every cell; rows
//   results/nested/red/wgate_abd.jsonl, results/nested/red/wgate_c.jsonl, results/nested/red/wnv.jsonl)
// - legality: the rules (offload.engine.rules) accept it; AFAR and CCE compile and run it correctly (front ends of two other
//   families)
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
struct s1 { int a[8]; };
struct s0 { int v; struct s1 *nx; };
int main(void)
{
  int hf = 1, i, tot = 0, ok;
  #pragma omp target map(from: hf)
  hf = omp_is_initial_device() ? 1 : 0;
  struct s0 *r = malloc(sizeof *r);
  r->v = 10;
  r->nx = malloc(sizeof *r->nx);
  for (i = 0; i < 8; i++) r->nx->a[i] = 200 + i;
  #pragma omp target map(tofrom: tot) map(to: r->nx[0:1]) map(to: r[0:1])
  {
    tot += r->v;
    for (int q = 0; q < 8; q++) tot += r->nx->a[q];
  }
  ok = tot == 1638;
  printf("tot= %d\n", tot);
  printf("RESULT ran_on_host=%d check=%s\n", hf, ok ? "pass" : "fail");
  return 0;
}

Source in the repository

Something to add? Dispute its legality · Report it fixed in a newer release · Report a mistake on this page

Expected and observed

Observed where it behaves incorrectly

Failing in Thread:1
Accelerator Fatal Error: call to cuMemcpyDtoHAsync returned error 700 (CUDA_ERROR_ILLEGAL_ADDRESS): Illegal address during kernel execution
 File: <dir>
 Function: main:31
 Line: 44

What the specification requires

The specification text this program relies on and the compiler is expected to support.

OpenMP 6.0 §7.9.6 p.284 l.8-10

for which a corresponding pointer exists in the device data environment after all mappable storage blocks are mapped

OpenMP 5.2 §5.8.3 p.152 l.5-7

Evidence

Same compiler, one change

one change from the finding; correct on nvhpc, nvhpc269: not run yet

Control program

// control: the same two list items, the parent first
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
struct s1 { int a[8]; };
struct s0 { int v; struct s1 *nx; };
int main(void)
{
  int hf = 1, i, tot = 0, ok;
  #pragma omp target map(from: hf)
  hf = omp_is_initial_device() ? 1 : 0;
  struct s0 *r = malloc(sizeof *r);
  r->v = 10;
  r->nx = malloc(sizeof *r->nx);
  for (i = 0; i < 8; i++) r->nx->a[i] = 200 + i;
  #pragma omp target map(tofrom: tot) map(to: r[0:1]) map(to: r->nx[0:1])
  {
    tot += r->v;
    for (int q = 0; q < 8; q++) tot += r->nx->a[q];
  }
  ok = tot == 1638;
  printf("tot= %d\n", tot);
  printf("RESULT ran_on_host=%d check=%s\n", hf, ok ? "pass" : "fail");
  return 0;
}

one change from the finding; correct on nvhpc, nvhpc269: not run yet

Control program

// control: the member is embedded, not a pointer
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
struct s1 { int a[8]; };
struct s0 { int v; struct s1 nx; };
int main(void)
{
  int hf = 1, i, tot = 0, ok;
  #pragma omp target map(from: hf)
  hf = omp_is_initial_device() ? 1 : 0;
  struct s0 *r = malloc(sizeof *r);
  r->v = 10;
  for (i = 0; i < 8; i++) r->nx.a[i] = 200 + i;
  #pragma omp target map(tofrom: tot) map(to: r[0:1])
  {
    tot += r->v;
    for (int q = 0; q < 8; q++) tot += r->nx.a[q];
  }
  ok = tot == 1638;
  printf("tot= %d\n", tot);
  printf("RESULT ran_on_host=%d check=%s\n", hf, ok ? "pass" : "fail");
  return 0;
}

Other compilers run it correctly

AFAR 24.3.0, CCE 19.0.0, CCE 20.0.2, CCE 21.0.3, CCE 21.0.4, GCC 16.2.0