Re: [PATCH i-g-t] tests/intel/xe_exec_system_allocator: Enhance SVM compute test

[Francois Dugast <francois.dugast@intel.com>]

Hi,

On Mon, Mar 09, 2026 at 04:34:00AM +0000, nishit.sharma@intel.com wrote:
> From: Nishit Sharma <nishit.sharma@intel.com>
> 
> Updated test to validate output values against expected results after
> workload execution. These enhancements improve test coverage, reliability,
> and debugging for SVM compute workloads. Updated test to validate output
> values against expected results after kernel execution.

Using a malloc'ed array for output and validating output in this test both
revert what was done in this commit, for good reasons:
373a6dbfb tests/intel/xe_exec_system_allocator: Cleanup compute test

This is because the test is used to finely trigger page fault floods from
compute kernel context (EUs) and we want to avoid interference with 1. extra
GPU page faults and 2. CPU page faults, so that we can measure the exact
impact of KMD optimizations on page fault servicing latency. Please see:
https://patchwork.freedesktop.org/series/160784/#rev2

This change would bring us back to the known previous state, which prevented
those measurements. Can you elaborate on why we need to revert?

Thanks,
Francois

> 
> Signed-off-by: Nishit Sharma <nishit.sharma@intel.com>
> ---
>  tests/intel/xe_exec_system_allocator.c | 20 ++++++++++++++++----
>  1 file changed, 16 insertions(+), 4 deletions(-)
> 
> diff --git a/tests/intel/xe_exec_system_allocator.c b/tests/intel/xe_exec_system_allocator.c
> index 14f5e358f..e670ee007 100644
> --- a/tests/intel/xe_exec_system_allocator.c
> +++ b/tests/intel/xe_exec_system_allocator.c
> @@ -2257,7 +2257,7 @@ test_compute(int fd, size_t size, unsigned int flags)
>  		 */
>  		.array_size = size / sizeof(float),
>  	};
> -	float *compute_input;
> +	float *compute_input, *compute_output;
>  	int i;
>  
>  	vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_LR_MODE | DRM_XE_VM_CREATE_FLAG_FAULT_MODE, 0);
> @@ -2267,19 +2267,31 @@ test_compute(int fd, size_t size, unsigned int flags)
>  	xe_wait_ufence(fd, &bo_sync->sync, USER_FENCE_VALUE, 0, FIVE_SEC);
>  
>  	compute_input = aligned_alloc(SZ_2M, size);
> -	igt_assert(compute_input);
> +	compute_output = aligned_alloc(SZ_2M, size);
> +	igt_assert(compute_input && compute_output);
>  
> -	env.loop_count = (flags & TOUCH_ONCE) ? 1 : env.array_size;
> +	env.loop_count = (flags & TOUCH_ONCE) ? 1 : env.array_size / 2;
>  	env.skip_results_check = !(flags & ACCESS_DEVICE_HOST);
>  	env.input_addr = to_user_pointer(compute_input);
> +	env.output_addr = to_user_pointer(compute_output);
>  	env.vm = vm;
>  
>  	for (i = 0; i < env.loop_count; i++)
> -		compute_input[i] = rand() / (float)RAND_MAX;
> +		compute_input[i] = (float)i + 2.0f;
>  
>  	run_intel_compute_kernel(fd, &env, EXECENV_PREF_SYSTEM);
>  
> +	for (i = 0; i < env.loop_count; i++) {
> +		float expected_output = compute_input[i] * compute_input[i];
> +
> +		if (compute_output[i] != expected_output || compute_output[i] == 0.0f)
> +			igt_debug("[%4d] input:%f output:%f expected_output:%f\n",
> +				  i, compute_input[i], compute_output[i], expected_output);
> +		igt_assert_eq_double(compute_output[i], expected_output);
> +	}
> +
>  	free(compute_input);
> +	free(compute_output);
>  	unbind_system_allocator();
>  	xe_vm_destroy(fd, vm);
>  }
> -- 
> 2.43.0
>