Documentation review #1451
Documentation review #1451
Conversation
Includes screenshots and minor edits to the documentation, including a new index page. Signed-off-by: Melissa Weber Mendonça <melissawm@gmail.com>
Signed-off-by: Pavithra Eswaramoorthy <pavithraes@outlook.com>
Signed-off-by: Pavithra Eswaramoorthy <pavithraes@outlook.com>
Signed-off-by: Melissa Weber Mendonça <melissawm@gmail.com>
Signed-off-by: Melissa Weber Mendonça <melissawm@gmail.com>
docs/hlo_op_stats.md
Outdated
|
|
||
| For GPUs, HLO ops have an N:M relationship with the kernels that actually get | ||
| executed. For statistics at the kernel level, see the GPU Kernel Stats tool. | ||
| executed. For statistics at the kernel level, see the _GPU Kernel Stats_ tool. |
There was a problem hiding this comment.
I assume the GPU Kernel Stats tool is this one: https://www.tensorflow.org/guide/profiler#gpu_kernel_stats
We may want to create a page from it with similar (but updated) content.
| @@ -60,18 +60,20 @@ The HLO Op Stats tool has the following key components: | |||
|
|
|||
| ### HLO Operation Statistics Table Details | |||
|
|
|||
|  | |||
There was a problem hiding this comment.
Some of the columns listed below do not show up at all for me (any columns past the Avg. time are missing). I wonder if this has to do with the fact that I can't seem to change the size of the output on colab, and maybe the columns are just overflowing the size of the screen. Unfortunately I tried different browsers and sizes and still I could not see them.
| @@ -8,11 +8,11 @@ utilizes hardware resources, identify performance bottlenecks, and optimize your | |||
| model for faster execution. The Trace Viewer UI is based on the one used in | |||
| `chrome://tracing` and therefore requires that you use the Chrome browser. | |||
There was a problem hiding this comment.
This worked fine for me on Firefox, so maybe this is not a requirement anymore?
| * Clicking on an XLA op provides additional information in the details pane. | ||
| For example, it links to the op in the Graph Viewer tool. It may also | ||
| provide pointers to source code and/or the Python stack trace, the framework | ||
| op that caused this XLA op to get generated, etc. (if present in the | ||
| profile). It may also show FLOPS (number of floating point operations | ||
| executed by the op) and bytes accessed by the op; this information is | ||
| statically acquired from XLA during compilation, rather than runtime | ||
| information from the profile. |
There was a problem hiding this comment.
I reorganized this content to be included in the details pane description as I felt it would be more self-contained there.
docs/graph_viewer.md
Outdated
There was a problem hiding this comment.
I notice a formatting issue with this page -- I'll push a commit to update it.
|
I have just added a skeleton of a page for the GPU Kernel Stats tool. I'm not sure if this is something we want to keep or not, please let me know! |
| * For TPUs only, time per HLO by replica group: A drop down lets you pick | ||
| from the different collective operations executed during the profiling | ||
| session. Different instances of that collective op may have been | ||
| executed among different replica groups (e.g., | ||
| [AllGather](https://openxla.org/xla/operation_semantics#allgather)); a | ||
| pie chart shows the distribution of time between these different | ||
| instances. |
There was a problem hiding this comment.
I do not see this component, not sure if something went wrong on my side.
| * A chart titled **Time spent on outside compilation**. Outside | ||
| compilation is a TensorFlow feature that enables certain ops within an | ||
| XLA computation to transparently run on the host CPU rather than the | ||
| accelerator device (e.g., `tf.summary` or `tf.print` that requires I/O | ||
| access that the device does not possess). |
There was a problem hiding this comment.
This chart shows up empty for me. Not sure if this is the correct wording or if it should be reviewed.
This PR includes a number of improvements to documentation, including a few content changes and some screenshots of the XProf Tensorboard Plugin.
I have a few questions I will ask as comments below. This PR is co-authored by @pavithraes , and we are happy to address any feedback.
We also found a couple of other documents containing similar content:
Should we include some of that content here, link out to those pages, or update the pages themselves to have up-to-date content?
Thank you!