Method launch

cake-core/src/backends/cuda/ops.rs:197–234

(
            a: &CudaSlice<T>,
            l_a: &Layout,
            b: &CudaSlice<T>,
            l_b: &Layout,
            c: &CudaSlice<T>,
            l_c: &Layout,
            dev: &CudaDevice,

        #[allow(clippy::too_many_arguments)]
        fn launch<T: DeviceRepr + WithDType>(
            a: &CudaSlice<T>,
            l_a: &Layout,
            b: &CudaSlice<T>,
            l_b: &Layout,
            c: &CudaSlice<T>,
            l_c: &Layout,
            dev: &CudaDevice,
        ) -> Result<CudaSlice<T>> {
            let a = match l_a.contiguous_offsets() {
                Some((o1, o2)) => a.slice(o1..o2),
                None => candle_core::bail!("add3: a must be contiguous"),
            };
            let b = match l_b.contiguous_offsets() {
                Some((o1, o2)) => b.slice(o1..o2),
                None => candle_core::bail!("add3: b must be contiguous"),
            };
            let c = match l_c.contiguous_offsets() {
                Some((o1, o2)) => c.slice(o1..o2),
                None => candle_core::bail!("add3: c must be contiguous"),
            };
            let el = l_a.shape().elem_count();
            let cfg = LaunchConfig::for_num_elems(el as u32);
            let func = dev.get_or_load_custom_func(
                &kernel_name::<T>("add3"),
                "cake_fused_ops",
                FUSED_OPS_PTX,
            )?;
            let out = unsafe { dev.alloc::<T>(el)? };
            let mut builder = func.builder();
            builder.arg(&el);
            builder.arg(&a);
            builder.arg(&b);
            builder.arg(&c);
            builder.arg(&out);
            unsafe { builder.launch(cfg) }.w()?;
            Ok(out)
        }

        use candle_core::backend::BackendStorage;
        use candle_core::cuda_backend::CudaStorageSlice as SS;