A high-performance inference engine for AI models
use crate::backends::vulkan::{VkKernelError, VkPhysicalDevice};
pub fn check_work_group(
physical_device: &VkPhysicalDevice,
kernel: &'static str,
size: [u32; 3],
subgroup_size: Option<u32>,
) -> Result<(), VkKernelError> {
let limits = &physical_device.limits;
let work_group_error = || VkKernelError::WorkGroupSize {
size,
limit: limits.max_work_group_size,
invocations: limits.max_work_group_invocations,
};
if size.contains(&0) || size.iter().zip(limits.max_work_group_size).any(|(&size, limit)| size > limit) {
return Err(work_group_error());
}
let invocations = size.iter().product::<u32>();
if invocations > limits.max_work_group_invocations {
return Err(work_group_error());
}
if let Some(subgroup_size) = subgroup_size
&& (!size[0].is_multiple_of(subgroup_size) || invocations > subgroup_size * limits.max_work_group_subgroups)
{
return Err(VkKernelError::KernelPrecondition {
kernel,
condition: "a workgroup of whole subgroups of its required size, within maxComputeWorkgroupSubgroups",
});
}
Ok(())
}