Function holoscan::utils::transmit_data_per_model(gxf_context_t&, const HoloInfer::MultiMappings&, HoloInfer::DataMap&, OutputContext&, std::vector&, HoloInfer::DimType&, bool, bool, const nvidia::gxf::Handle&, const std::string&, const cudaStream_t&, TensorTransmitCache&)

Transmits multiple buffers via GXF Transmitters with persistent message caching.

Unlike the other overloads, this variant allocates the output GXF entity and inserts tensor components only once (on the first call). On subsequent calls the pre-allocated tensor buffers are reused and only the data is updated via memcpy / cudaMemcpyAsync. A tensor is reshaped (and its backing memory reallocated) only when the incoming buffer size exceeds the current allocation capacity. The persistent entity is emitted by reference-counted copy each frame rather than being moved, so the cache remains valid for the next call.

Parameters

• cont – GXF context for transmission

• model_to_tensor_map – Map of model name as key, mapped to a vector of tensor names

• input_data_map – Map of tensor name as key, mapped to the data buffer

• op_output – Output context. Assumes the output port name is “transmitter”.

• out_tensors – Output tensor names

• tensor_out_dims_map – Map with model name as key mapped to output tensor dimensions

• cuda_buffer_in – Whether input buffers reside on the GPU

• cuda_buffer_out – Whether the output message should reside on the GPU

• allocator_ – GXF Memory allocator

• module – Module name used in error reporting

• cstream – CUDA stream to use for async memory copies

• cache – Persistent cache holding the reusable output entity and per-tensor allocation info

Returns

GXF result code