Home Original page

Update nvImageCodec to 0.7.0 by jantonguirao · Pull Request #6105 · NVIDIA/DALI

Greptile Overview

Greptile Summary

Updates nvImageCodec dependency from 0.6.0 to 0.7.0, adapting to API changes and optimizing performance.

  • Removed deprecated buffer_size field from image_info structure (API change in 0.7.0)
  • Optimized resource reuse: sub code stream and image instances are now reused across iterations instead of being recreated, reducing allocation overhead
  • Simplified image descriptor creation by removing parallel initialization task and creating images inline in the main loop
  • Updated version requirements, tarball URLs, hashes, and git tags across cmake and conda configurations
  • Reordered stub generator function declarations to follow create-destroy pattern

Confidence Score: 4/5

  • This PR is a straightforward dependency update with well-contained API adaptations that follow established patterns in the codebase.
  • Score of 4 reflects: clean dependency version bump, logical API adaptation for removed buffer_size field, and reasonable optimization for instance reuse. The instance reuse pattern appears correct based on nvImageCodec's API semantics. Deducted 1 point because the image decoder changes involve subtle resource management that benefits from integration testing.
  • Pay attention to dali/operators/imgcodec/image_decoder.h - the instance reuse optimization should be verified with integration tests to ensure ROI handling works correctly.

Important Files Changed

File Analysis

Filename Score Overview
cmake/Dependencies.common.cmake 5/5 Version bump from 0.6.0 to 0.7.0 for nvImageCodec dependency. Updated min/max versions, tarball URL, hash, and git tag. Standard dependency update.
conda/third_party/dali_nvimagecodec/recipe/meta.yaml 5/5 Updated build_version and git_rev from 0.6.0 to 0.7.0 for the conda package recipe.
dali/operators/imgcodec/image_decoder.h 4/5 Major refactoring: removed deprecated buffer_size field, optimized sub code stream and image instance reuse across iterations, simplified image descriptor creation by removing parallel init task.
internal_tools/stub_generator/nvimgcodec.json 5/5 Reordered function declarations: moved nvimgcodecImageDestroy after nvimgcodecImageCreate to follow create-destroy pattern.

Sequence Diagram

sequenceDiagram
    participant TP as ThreadPool
    participant ID as ImageDecoder
    participant NIC as nvImageCodec
    participant State as SampleState

    Note over ID: RunImplImpl() called
    
    loop For each sample (parallel in blocks)
        ID->>State: ParseSample()
        alt ROI requested
            ID->>State: Check sub_encoded_stream exists
            alt Reuse existing stream
                ID->>NIC: nvimgcodecCodeStreamGetSubCodeStream(existing)
            else Create new stream
                ID->>NIC: NvImageCodecCodeStream::FromSubCodeStream()
            end
            NIC-->>State: Store sub_encoded_stream
        end
        ID->>State: PrepareOutput()
    end
    
    Note over ID: Allocate output memory
    
    loop For each sample (sequential)
        alt Not cached
            ID->>State: Set image_info.buffer
            alt Reuse existing image
                ID->>NIC: nvimgcodecImageCreate(existing)
            else Create new image
                ID->>NIC: NvImageCodecImage::Create()
            end
            ID->>ID: Add to batch_images
        end
    end
    
    ID->>NIC: nvimgcodecDecoderDecode(batch)
    NIC-->>ID: Future with results
    ID->>NIC: nvimgcodecFutureWaitForAll()
    
    opt Need processing
        loop For each decoded sample
            TP->>ID: ConvertGPU/ConvertCPU()
        end
    end
Loading