Skip to content

Fix profiler defects (Resolves #3544) - #3546

Open
DRJ-YDV wants to merge 1 commit into
NVIDIA:mainfrom
DRJ-YDV:fix-profiler-defects-3544
Open

Fix profiler defects (Resolves #3544)#3546
DRJ-YDV wants to merge 1 commit into
NVIDIA:mainfrom
DRJ-YDV:fix-profiler-defects-3544

Conversation

@DRJ-YDV

Copy link
Copy Markdown

Fixes#3544.
Knocked out those three minor profiler bugs mentioned in the issue:

  1. Device ID validation: Fixed the off-by-one bug in options.cu so it properly errors out on device >= num_devices instead of letting it slip through and crashing later.
  2. NCHW layout typo: In device_allocation.cu, kTensorNCHW was accidentally returning TensorNHWC due to a copy-paste error. Swapped it out to use the correct TensorNCHW layout.
  3. Sparse-meta type mismatches: Tweaked the host and device BlockFillRandomSparseMeta initialization functions so they both handle signed (kS16, kS32) and unsigned (kU16, kU32) type IDs. This stops the host initialization from silently falling through to all-zeros when the types don't line up.

Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Small profiler defects: device-ID validation off by one, kTensorNCHW mapped to TensorNHWC, sparse-meta host/device type mismatch

1 participant

@DRJ-YDV