[webgpu] update perf test #6438
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
haoyunfeix
force-pushed
the
webgpu_update_perf_test
branch
2 times, most recently
from
05264a8
to
2d2c688
Compare
xhcao
reviewed
Sep 30, 2022
| const m = tf.matMul(tensorsWarmUp.tensorA, tensorsWarmUp.tensorB); | ||
| result.dispose(); | ||
| m.dispose(); | ||
| } |
There was a problem hiding this comment.
In tf.profile,
- I think we mainly profile
const result = tf.matMul(tensors[i].tensorA, tensors[i].tensorB), would includingwarmup modelandwarmup gpuparts influence the correctness? - Why we still need to execute
tf.matMul(tensorsWarmUp.tensorA, tensorsWarmUp.tensorB)in the 3rd for-loop, I think the cost time of it may be larger thantf.matMul(tensors[i].tensorA, tensors[i].tensorB), which is the real main workload. - How to make sure that all matMul commands are executed one by one? I think they may be executed synchronously here?
There was a problem hiding this comment.
- Yes,
warmup modelis for the first run of each matmul shape(costing large shader generate and buffer creating time ... which leads to a significant deviation from the average value) andwarmup gpuis to make every shape of inputs could be measured in the same frequency. - The idea to insert a large warmup tensor between each main workload is also to keep high frequency, the real main workload will pull down the frequency if they execute together, please see below pictures:
with: keep frequency at almost 1200M
without: Visible jitter, break the measurement accurate (only happens on WebGPU)
- You're right, serial execution would be considered in another PR.
| for (let i = 0; i < inputs.length; i++) { | ||
| let dimAOuter = parseInt(inputs[i].split(',')[0]); | ||
| let dimInner = parseInt(inputs[i].split(',')[1]); | ||
| let dimBOuter = parseInt(inputs[i].split(',')[2]); |
There was a problem hiding this comment.
Why we wrap the dimAOuter | dimInner | dimBOuter into a string here, and should encode and decode the string? Why not wrap this value into an array or a structure?
There was a problem hiding this comment.
Each row has a rerun button to help rerun a single case for double check. And this button catches shape infos(e.g. [16, 512, 16]) from HTML element, which value must be a string. So here I used the raw data what I get from the page.
haoyunfeix
force-pushed
the
webgpu_update_perf_test
branch
from
07bc3bc
to
b035bec
Compare
qjia7
approved these changes
Oct 12, 2022
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is