Restore default hw_decoder_load to 0.65 in imgcodec decoder

[jantonguirao]

Summary

• Change the default value of the hw_decoder_load argument in the nvImageCodec-based image decoder (dali/operators/imgcodec/decoder_schema.cc) from 0.90f back to 0.65f, matching the legacy decoder's default.

Motivation

The legacy decoder (dali/operators/decoder/image_decoder.cc) defaults hw_decoder_load to 0.65. The new nvImageCodec-based decoder ships with 0.90, which causes a meaningful behavior change for users who relied on the default value.

Nsys profiling of EfficientNet + AutoAugment ImageNet training (single-GPU, batch 256) on the two defaults shows:

	legacy default (0.65)	new default (0.90)
Measured HW path share	64.8 %	90.2 %
Hybrid (CPU+GPU) share	34.6 %	9.8 %
Decoder p50	15.6 ms	17.7 ms
Decoder p99	24.5 ms	34.3 ms
Decoder max	30.8 ms	72.7 ms
End-to-end iters/s	18.2	18.6 (+1.8 %)

Users who want the higher HW-routing fraction can still opt in explicitly via the operator argument.

Test plan

• Existing decoder tests pass
• internal_tools/hw_decoder_bench.py shows no regression at default settings

[Copilot]

Pull request overview

Restores the nvImageCodec-based image decoder’s default hardware JPEG routing (hw_decoder_load) to match the historical/legacy behavior, reducing unintended behavior changes for users relying on defaults.

Changes:

• Change default hw_decoder_load from 0.90f to 0.65f in the imgcodec decoder schema.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[greptile-apps]

Greptile Summary

Restores the hw_decoder_load default in the nvImageCodec-based decoder (decoder_schema.cc) from 0.90f back to 0.65f, matching the legacy decoder's long-standing default. The motivation is well-documented: profiling shows that 0.90 increases worst-case decode latency significantly (30.8 ms → 72.7 ms max) due to synchronous cuMemFree_v2 calls under high HW-engine pressure, while end-to-end throughput is essentially unchanged.

• One-line change in decoder_schema.cc; the copyright header (2023-2026) and argument docstring are already correct and require no update.
• No documentation, changelog, or test files reference the 0.90 value, so no secondary cleanup is needed.

Confidence Score: 5/5

Safe to merge — restores a well-understood default to align with the legacy decoder, backed by profiling data showing measurably better worst-case latency.

The change is a single-constant restoration with strong empirical justification. The legacy decoder, all tests, and all documentation are already consistent with 0.65; no secondary files need updating. No correctness, memory, or API contract concerns arise from the change.

No files require special attention.

Important Files Changed

Filename	Overview
dali/operators/imgcodec/decoder_schema.cc	Single-line change restoring hw_decoder_load default from 0.90f to 0.65f, matching the legacy decoder; copyright year already includes 2026.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Image Batch] --> B{hw_decoder_load routing}
    B -->|"≤ 0.65 fraction (default)"| C[HW JPEG Decoder\nNVIDIA A100 engine]
    B -->|"> 0.65 fraction"| D[Hybrid CPU+GPU Decoder\nnvJPEG software path]
    C --> E[Decoded Output]
    D --> E
    style B fill:#f0f4ff,stroke:#3366cc
    style C fill:#d4edda,stroke:#28a745
    style D fill:#fff3cd,stroke:#ffc107

Loading

_{Reviews (1): Last reviewed commit: "Restore default hw_decoder_load to 0.65 ..." | Re-trigger Greptile}

[dali-automaton]

CI MESSAGE: [52351570]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [52351570]: BUILD PASSED