Re: [PATCH i-g-t] tests/xe_exec_system_allocator: Don't use NaN in an input buffer

[Matthew Brost <matthew.brost@intel.com>]

On Mon, May 04, 2026 at 05:28:21PM +0200, Zbigniew Kempczyński wrote:
> Setting input buffer memory to random byte (in this particular case
> 0xff) and then use it as a float causes doing compute square on NaN
> numbers. It's nothing wrong about this but comparing NaN in between
> returns false according to IEEE-754 rules. Buffer compare function
> does assertion if output is not equal to expected output so we hit
> this randomly in 'eu-fault*' tests.
> 
> Replace memset to loop which sets all input floats to real numbers
> thus avoiding NaN arithmetic and comparison.
> 
> Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
> Cc: Francois Dugast <francois.dugast@intel.com>
> ---
>  tests/intel/xe_exec_system_allocator.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/tests/intel/xe_exec_system_allocator.c b/tests/intel/xe_exec_system_allocator.c
> index 5580099f78..0f1d6ee9c7 100644
> --- a/tests/intel/xe_exec_system_allocator.c
> +++ b/tests/intel/xe_exec_system_allocator.c
> @@ -2339,7 +2339,8 @@ test_compute(int fd, struct drm_xe_engine_class_instance *eci, size_t size,
>  		igt_assert(compute_input);
>  		env.input_addr = to_user_pointer(compute_input);
>  
> -		memset(compute_input, rand() % 255 + 1, size);
> +		for (int j = 0; j < env.array_size; j++)
> +			compute_input[j] = rand() / (float)RAND_MAX;

This logic can still get 0xff or 0x0...

If you want to avoid 0xff or 0x0 I think this logic would work:

memset(compute_input, rand() % 254 + 1, size);

>  
>  		xe_run_intel_compute_kernel_on_engine(fd, eci, &env, EXECENV_PREF_SYSTEM);
>  
> -- 
> 2.43.0
>