Method suggested_launch_configuration

crates/cust/src/function.rs:380–405

Returns a reasonable block and grid size to achieve the maximum capacity for the launch (the max number of active warps with the fewest blocks per multiprocessor). # Params `dynamic_smem_size` is the amount of dynamic shared memory required by this function. We currently do not expose a way of determining this dynamically based on block size due to safety concerns. `block_size_limit` is the max

(
        &self,
        dynamic_smem_size: usize,
        block_size_limit: BlockSize,
    )

    ///
    /// Note: all panics by `dynamic_smem_size` will be ignored and the function will instead use `0`.
    pub fn suggested_launch_configuration(
        &self,
        dynamic_smem_size: usize,
        block_size_limit: BlockSize,
    ) -> CudaResult<(u32, u32)> {
        let mut min_grid_size = MaybeUninit::uninit();
        let mut block_size = MaybeUninit::uninit();

        let total_block_size_limit = block_size_limit.x * block_size_limit.y * block_size_limit.z;

        unsafe {
            cuda::cuOccupancyMaxPotentialBlockSize(
                min_grid_size.as_mut_ptr(),
                block_size.as_mut_ptr(),
                self.to_raw(),
                None,
                dynamic_smem_size,
                total_block_size_limit as i32,
            )
            .to_result()?;
            Ok((
                min_grid_size.assume_init() as u32,
                block_size.assume_init() as u32,
            ))
        }
    }
}

/// Launch a kernel function asynchronously.