Loopy unable to generate the device code for a kernel with conditional branching

[nipunayf]

Description

When attempting to convert and generate the device code for a C kernel with a if condition, it generates the following error:

/home/hp/mambaforge/envs/libnomp/bin/python /home/hp/projects/nomp/libnomp/python/loopy_api.py 
Traceback (most recent call last):
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/loopy/tools.py", line 933, in wrapper
    result = transform_cache[cache_key]
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/pytools/persistent_dict.py", line 583, in __getitem__
    return self.fetch(key, _stacklevel=1)
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/pytools/persistent_dict.py", line 702, in fetch
    raise NoSuchEntryError(key)
pytools.persistent_dict.NoSuchEntryError: ('preprocess_program', 'preprocess_program', (TranslationUnit(func_id_to_in_knl_callable_mappers=[], entrypoints=frozenset({'foo'}), callables_table=pmap({'foo': CallableKernel(arg_id_to_descr=None, arg_id_to_dtype=None, name='foo', subkernel=LoopKernel(domains=[BasicSet("[N] -> { [i] : 0 <= i <= N }")], instructions=[Assignment(conflicts_with_groups=frozenset(), assignee=Subscript(Variable('a'), Variable('i')), tags=frozenset(), depends_on=frozenset(), no_sync_with=frozenset(), within_inames=frozenset({'i'}), groups=frozenset(), id='_nomp_insn', within_inames_is_final=False, expression=Product((Subscript(Variable('a'), Variable('i')), Sum((Variable('i'), 1)))), predicates=frozenset({Comparison(Subscript(Variable('a'), Variable('i')), '>', 0)}), temp_var_type=Optional(), atomicity=(), depends_on_is_final=True, priority=0)], args=[<a: ArrayArg, type: np:dtype('float64'), shape: unknown in/out aspace: global>, <N: ValueArg, type: np:dtype('int32')>], assumptions=BasicSet("[N] -> {  :  }"), temporary_variables={}, inames={'i': Iname(name='i', tags=frozenset())}, substitutions={}, options=Options(disable_global_barriers=False, enforce_variable_access_ordered=True, return_dict=False, trace_assignments=False, insert_gbarriers=False, check_dep_resolution=True, no_numpy=False, annotate_inames=False, cl_exec_manage_array_events=True, edit_code=False, write_code=False, allow_fp_reordering=True, enforce_array_accesses_within_bounds=True, trace_assignment_values=False, skip_arg_checks=False, allow_terminal_colors=True, build_options=[], write_wrapper=False), target=<loopy.target.opencl.OpenCLTarget object at 0x7fc1fcd3b670>, tags=frozenset(), state=<KernelState.INITIAL: 0>, name='foo', preambles=(), preamble_generators=(), symbol_manglers=(), linearization=None, iname_slab_increments=immutables.Map({}), loop_priority=frozenset(), applied_iname_rewrites=(), index_dtype=np:dtype('int32'), silenced_warnings=[], overridden_get_grid_sizes_for_insn_ids=None))}), target=<loopy.target.opencl.OpenCLTarget object at 0x7fc1fcd3b670>),), {})

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/hp/projects/nomp/libnomp/python/loopy_api.py", line 624, in <module>
    print(lp.generate_code_v2(lp_knl).device_code())
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/loopy/codegen/__init__.py", line 597, in generate_code_v2
    program = preprocess_program(program)
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/loopy/tools.py", line 949, in wrapper
    result = func(*args, **kwargs)
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/loopy/preprocess.py", line 843, in preprocess_program
    new_subkernel = _preprocess_single_kernel(
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/loopy/preprocess.py", line 761, in _preprocess_single_kernel
    check_for_writes_to_predicates(kernel)
  File "/home/hp/mambaforge/envs/libnomp/lib/python3.10/site-packages/loopy/preprocess.py", line 67, in check_for_writes_to_predicates
    raise LoopyError("In instruction '%s': may not write to "
loopy.diagnostic.LoopyError: In instruction '_nomp_insn': may not write to variable(s) 'a' involved in the instruction's predicates

Process finished with exit code 1

Reproduce Steps

1. Use the loopy_api.py to convert the following C kernel.

void foo(double *a, int N) {
  for (int i = 0; i < N + 1; i++)
    if (a[i] > 0)
      a[i] = a[i] * (i + 1);
}

2. Attempt to generate the device code with,

print(lp.generate_code_v2(lp_knl).device_code())

[thilinarmtb]

Thanks for reporting this !

[nipunayf]

Additional note to this issue,

The issue appears when we access an array element in the condition, in this case a[i]. If we rewrite the kernel as below, the issue will be resolved.

void foo(double *a, int N) {
  for (int i = 0; i < N + 1; i++)
    int temp = a[i];
    if (temp > 0)
      a[i] = a[i] * (i + 1);
}

[thilinarmtb]

Additional note to this issue,

The issue appears when we access an array element in the condition, in this case a[i]. If we rewrite the kernel as below, the issue will be resolved.

void foo(double *a, int N) {
  for (int i = 0; i < N + 1; i++)
    int temp = a[i];
    if (temp > 0)
      a[i] = a[i] * (i + 1);
}

Yes, we need to make sure that if condition is a dependency for the code in then and else branch.