Struct ChromaConvert 

struct ChromaConvert {
    module: *mut c_void,
    kernel: *mut c_void,
    kernel_tex: *mut c_void,
    layout: ConvertLayout,
    surface: CUdeviceptr,
    pitch: usize,
    registered: NV_ENC_REGISTERED_PTR,
    mapped: NV_ENC_INPUT_PTR,
}

Fields

module: *mut c_void

kernel: *mut c_void

kernel_tex: *mut c_void

layout: ConvertLayout

surface: CUdeviceptr

pitch: usize

registered: NV_ENC_REGISTERED_PTR

mapped: NV_ENC_INPUT_PTR

Implementations

impl ChromaConvert

unsafe fn new( cuda: &CudaFunctions, funcs: &NV_ENCODE_API_FUNCTION_LIST, session: *mut c_void, width: u32, height: u32, layout: ConvertLayout, ) -> Option<Self>

JIT the module, allocate the widthxheight surface of layout, and register it with the session. None where any step refuses: the session then encodes the packed RGB itself, which sites chroma at the left of each block but converts with the matrix the session declares all the same.

fn input(&self) -> (NV_ENC_INPUT_PTR, NV_ENC_BUFFER_FORMAT)

The registered surface and its format, as NVENC is handed them.

unsafe fn run( &self, cuda: &CudaFunctions, src: CUdeviceptr, src_pitch: usize, width: u32, height: u32, swap_rb: bool, ) -> Result<(), String>

Convert the widthxheight packed surface at src/src_pitch into the surface, on the default stream so it is ordered behind the upload and ahead of the encode. swap_rb marks an RGBA byte order rather than BGRA.

unsafe fn run_texture( &self, cuda: &CudaFunctions, tex: CUtexObject, width: u32, height: u32, swap_rb: bool, ) -> Result<(), String>

The same convert reading an array-typed import through tex, which leaves its RGB where the compositor put it instead of copying it into linear memory first.

unsafe fn launch( &self, cuda: &CudaFunctions, kernel: *mut c_void, params: &mut [*mut c_void], width: u32, height: u32, ) -> Result<(), String>

One chroma sample per thread, on the default stream so the convert is ordered behind whatever produced the source and ahead of the encode.

unsafe fn texture_for(cuda: &CudaFunctions, array: CUarray) -> CUtexObject

A point-sampled, byte-valued texture over an array-typed import, or zero where the driver refuses one — the frame then stays on NVENC’s own conversion.

unsafe fn release( &self, cuda: &CudaFunctions, funcs: &NV_ENCODE_API_FUNCTION_LIST, session: *mut c_void, )

Inner handle before outer, as the rest of the teardown does.