Struct KernelLaunch 

pub struct KernelLaunch<'a> { /* private fields */ }

Builder for type-safe kernel launches across CUDA + Metal.

Accumulates grid dimensions, block dimensions, and typed kernel arguments. launch() packages the args as &[KernelArg] and calls GpuBackend::launch_typed.

Implementations

impl<'a> KernelLaunch<'a>

pub fn new(gpu: &'a dyn GpuBackend, kernel: KernelHandle) -> Self

pub fn grid(self, grid: [u32; 3]) -> Self

pub fn block(self, block: [u32; 3]) -> Self

pub fn shared_mem(self, bytes: u32) -> Self

pub fn arg_ptr(self, p: DevicePtr) -> Self

Add a DevicePtr (u64) argument.

pub fn arg_tensormap(self, map: &[u8; 128]) -> Self

Add a 128-byte CUtensorMap by value, for a kernel parameter declared __grid_constant__ const CUtensorMap.

TMA descriptors are the one argument on this path that is not pointer-or-scalar sized: the driver copies all 128 bytes into the parameter buffer, so the bytes must land in ceil(128/8) = 16 CONSECUTIVE slots contributing ONE param entry. See gpu::pack_kernel_args, and a_128_byte_arg_is_not_truncated.

pub fn arg_u32(self, v: u32) -> Self

Add a u32 argument.

pub fn arg_u64(self, v: u64) -> Self

Add a u64 argument.

pub fn arg_i32(self, v: i32) -> Self

Add an i32 argument.

pub fn arg_f32(self, v: f32) -> Self

Add an f32 argument.

pub fn launch(self, stream: u64) -> Result<()>

Execute the kernel launch via GpuBackend::launch_typed.

Builds a typed KernelArg slice from the recorded storage + kinds. The cuda backend’s default launch_typed flattens this back into the legacy void** shape; the metal backend overrides launch_typed to use setBuffer: / setBytes: per arg. The storage vec is not reallocated between building the args and launching, so all byte slices remain valid.