Multi-stage GPU compute pipeline orchestration.

Pipelines chain multiple GPU operations (kernels, buffer copies, etc.) into a single executable graph via EXCubeCL's native pipeline API.

Example: Image processing pipeline

pipeline = Dala.Gpu.Compute.pipeline()
pipeline
|> Dala.Gpu.Compute.pipeline_add(%{
  op: :run_kernel,
  kernel: :blur,
  inputs: [input_buf],
  output: blurred_buf,
  params: %{radius: 3, sigma: 1.5}
})
|> Dala.Gpu.Compute.pipeline_add(%{
  op: :run_kernel,
  kernel: :sharpen,
  inputs: [blurred_buf],
  output: sharpened_buf,
  params: %{amount: 0.5}
})
|> Dala.Gpu.Compute.pipeline_add(%{
  op: :run_kernel,
  kernel: :grayscale,
  inputs: [sharpened_buf],
  output: output_buf,
  params: %{}
})
Dala.Gpu.Compute.pipeline_run(pipeline)

Example: AI inference pipeline

pipeline = Dala.Gpu.Compute.pipeline()
pipeline
|> Dala.Gpu.Compute.pipeline_add(%{
  op: :run_kernel,
  kernel: :preprocess,
  inputs: [camera_buf],
  output: preprocessed_buf,
  params: %{normalize: true, size: {224, 224}}
})
|> Dala.Gpu.Compute.pipeline_add(%{
  op: :run_kernel,
  kernel: :mobilenet_v2,
  inputs: [preprocessed_buf],
  output: logits_buf,
  params: %{}
})
|> Dala.Gpu.Compute.pipeline_add(%{
  op: :run_kernel,
  kernel: :softmax,
  inputs: [logits_buf],
  output: probs_buf,
  params: %{}
})
Dala.Gpu.Compute.pipeline_run(pipeline)

Stage Specs

Each stage is a map with:

• :op — operation type (:run_kernel, :copy_buffer, :barrier)
• :kernel — kernel atom (for :run_kernel ops)
• :inputs — list of input Buffer structs
• :output — output Buffer struct
• :params — map of kernel-specific parameters

EXCubeCL Backend

Under the hood, stages are compiled into an EXCubeCL native pipeline (ExCubecl.pipeline/0 + ExCubecl.pipeline_add/5 + ExCubecl.pipeline_run/1) for efficient batch submission. The pipeline is freed after execution.