AMDMIGraphX
AMDMIGraphX copied to clipboard
ROCmSoftwarePlatform
Add support for GridSample operator
Add support for GridSample onnx operator. Currently the following feature set is supported:
- inputs: 4-D inputs
- align_corners: both
0and1 - mode:
linearandnearest - padding_mode:
zeros,borderandreflectionFixes: https://github.com/ROCm/AMDMIGraphX/issues/2923
Codecov Report
All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 91.90%. Comparing base (
bceef13) to head (a1cf519).
:exclamation: Current head a1cf519 differs from pull request most recent head 8bb3d24
Please upload reports for the commit 8bb3d24 to get more accurate results.
Additional details and impacted files
@@ Coverage Diff @@
## develop #2909 +/- ##
===========================================
+ Coverage 91.82% 91.90% +0.08%
===========================================
Files 486 487 +1
Lines 18991 19192 +201
===========================================
+ Hits 17438 17639 +201
Misses 1553 1553
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
![codecov[bot] avatar](https://avatars.githubusercontent.com/in/254?v=4 "Published by a pub.dev verified publisher"){kind=small}
| Test | Batch | Rate new b6172d |
Rate old 5fcf86 |
Diff | Compare |
|---|---|---|---|---|---|
| torchvision-resnet50 | 64 | 1,751.55 | 1,751.74 | -0.01% | :white_check_mark: |
| torchvision-resnet50_fp16 | 64 | 4,084.47 | 4,084.34 | 0.00% | :white_check_mark: |
| torchvision-densenet121 | 32 | 1,467.78 | 1,467.39 | 0.03% | :white_check_mark: |
| torchvision-densenet121_fp16 | 32 | 2,529.08 | 2,525.45 | 0.14% | :white_check_mark: |
| torchvision-inceptionv3 | 32 | 889.72 | 889.64 | 0.01% | :white_check_mark: |
| torchvision-inceptionv3_fp16 | 32 | 1,483.73 | 1,483.57 | 0.01% | :white_check_mark: |
| cadene-inceptionv4 | 16 | 412.38 | 412.40 | -0.01% | :white_check_mark: |
| cadene-resnext64x4 | 16 | 419.71 | 419.50 | 0.05% | :white_check_mark: |
| slim-mobilenet | 64 | 4,007.09 | 4,006.71 | 0.01% | :white_check_mark: |
| slim-nasnetalarge | 64 | 101.02 | 101.01 | 0.01% | :white_check_mark: |
| slim-resnet50v2 | 64 | 1,680.79 | 1,680.60 | 0.01% | :white_check_mark: |
| bert-mrpc-onnx | 8 | 615.83 | 618.22 | -0.39% | :white_check_mark: |
| bert-mrpc-tf | 1 | 277.98 | 279.81 | -0.65% | :white_check_mark: |
| pytorch-examples-wlang-gru | 1 | 320.54 | 319.57 | 0.30% | :white_check_mark: |
| pytorch-examples-wlang-lstm | 1 | 288.76 | 289.36 | -0.21% | :white_check_mark: |
| torchvision-resnet50_1 | 1 | 471.26 | 471.89 | -0.13% | :white_check_mark: |
| cadene-dpn92_1 | 1 | 247.18 | 247.09 | 0.04% | :white_check_mark: |
| cadene-resnext101_1 | 1 | 203.92 | 204.23 | -0.15% | :white_check_mark: |
| onnx-taau-downsample | 1 | 206.48 | 206.24 | 0.11% | :white_check_mark: |
| dlrm-criteoterabyte | 1 | 22.91 | 22.90 | 0.04% | :white_check_mark: |
| dlrm-criteoterabyte_fp16 | 1 | 42.72 | 42.73 | -0.02% | :white_check_mark: |
| agentmodel | 1 | 6,369.83 | 6,323.67 | 0.73% | :white_check_mark: |
| unet_fp16 | 2 | 34.18 | 34.21 | -0.07% | :white_check_mark: |
| resnet50v1_fp16 | 1 | 574.81 | 589.36 | -2.47% | :white_check_mark: |
| resnet50v1_int8 | 1 | 578.26 | 573.50 | 0.83% | :white_check_mark: |
| bert_base_cased_fp16 | 64 | 646.28 | 646.33 | -0.01% | :white_check_mark: |
| bert_large_uncased_fp16 | 32 | 199.02 | 198.99 | 0.02% | :white_check_mark: |
| bert_large_fp16 | 1 | 117.21 | 117.54 | -0.27% | :white_check_mark: |
| distilgpt2_fp16 | 16 | 1,212.43 | 1,211.40 | 0.09% | :white_check_mark: |
| yolov5s | 1 | 301.12 | 301.25 | -0.04% | :white_check_mark: |
| tinyllama | 1 | 23.33 | 23.34 | -0.03% | :white_check_mark: |
| vicuna-fastchat | 1 | 133.85 | 133.65 | 0.15% | :white_check_mark: |
| whisper-tiny-encoder | 1 | 244.42 | 244.39 | 0.01% | :white_check_mark: |
| whisper-tiny-decoder | 1 | 256.29 | 256.66 | -0.14% | :white_check_mark: |
This build is OK for merge :white_check_mark:
:red_circle:bert_large_uncased_fp16: FAILED: MIGraphX is not within tolerance - check verbose output
@causten The PR is ready for review
~~This sould not be marged yet, because I've found an accuracy issue with fuse_pointwise pass with the GPU target when computing GridSample results (with linear inetrploation).
Collected my findings on that issue in https://github.com/ROCm/AMDMIGraphX/issues/2923~~
I had to exclude logical_and and where operators from pointwise fusion, so the GPU tests are passing now.
https://github.com/ROCm/AMDMIGraphX/issues/2923 is fixed by that.
@TedThemistokleous @pfultz2 I've adressed the majority of you commets except the Look into alternative to avoid O(n^4) updates and ops. part.
@TedThemistokleous @pfultz2 I've managed to move out the op creation and updates outside the O(n^4) part, so the heavy lifting is not in that part anymore. The remaining code there is only for generating literals for gathernd indices.
@pfultz2 I've removed the fuse pointwise related changes from the PR in favor of https://github.com/ROCm/AMDMIGraphX/pull/3054
Codecov Report
All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 92.05%. Comparing base (
5fcf86e) to head (b6172d3). Report is 154 commits behind head on develop.
Additional details and impacted files
@@ Coverage Diff @@
## develop #2909 +/- ##
===========================================
+ Coverage 91.97% 92.05% +0.08%
===========================================
Files 489 490 +1
Lines 19398 19599 +201
===========================================
+ Hits 17841 18042 +201
Misses 1557 1557
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
@gyulaz-htec you don't need hit the "update branch" button. Once the review are complete I'll handle it
I would like to see additional documentation and possible refactoring since the parse_gridsample.cpp is quite large. Not blocking for merge.
