AMDMIGraphX
AMDMIGraphX copied to clipboard
Onehot dyn and refactor
- Creates a ref onehot operator that handles dynamic shapes
- New simplify matcher that replaces the new ref onehot operator with other ops and scatter
- Refactored to use scatter rather than gather for this op
- Added new mode to scatter on ref and gpu to skip out of bounds indices
Test | Batch | Rate new 6b7482 |
Rate old a03942 |
Diff | Compare |
---|---|---|---|---|---|
torchvision-resnet50 | 64 | 1,751.51 | 1,713.64 | 2.21% | :white_check_mark: |
torchvision-resnet50_fp16 | 64 | 4,083.92 | 3,811.63 | 7.14% | :high_brightness: |
torchvision-densenet121 | 32 | 1,471.89 | 1,454.47 | 1.20% | :white_check_mark: |
torchvision-densenet121_fp16 | 32 | 2,525.19 | 2,432.15 | 3.83% | :high_brightness: |
torchvision-inceptionv3 | 32 | 889.71 | 882.93 | 0.77% | :white_check_mark: |
torchvision-inceptionv3_fp16 | 32 | 1,485.10 | 1,414.78 | 4.97% | :high_brightness: |
cadene-inceptionv4 | 16 | 412.62 | 406.93 | 1.40% | :white_check_mark: |
cadene-resnext64x4 | 16 | 419.75 | 413.66 | 1.47% | :white_check_mark: |
slim-mobilenet | 64 | 4,009.12 | 3,816.13 | 5.06% | :high_brightness: |
slim-nasnetalarge | 64 | 101.03 | 97.00 | 4.15% | :high_brightness: |
slim-resnet50v2 | 64 | 1,681.76 | 1,651.31 | 1.84% | :white_check_mark: |
bert-mrpc-onnx | 8 | 617.10 | 589.49 | 4.68% | :high_brightness: |
bert-mrpc-tf | 1 | 279.72 | 290.01 | -3.55% | :red_circle: |
pytorch-examples-wlang-gru | 1 | 321.65 | 331.98 | -3.11% | :red_circle: |
pytorch-examples-wlang-lstm | 1 | 292.32 | 301.70 | -3.11% | :red_circle: |
torchvision-resnet50_1 | 1 | 470.97 | 454.96 | 3.52% | :high_brightness: |
cadene-dpn92_1 | 1 | 246.31 | 244.72 | 0.65% | :white_check_mark: |
cadene-resnext101_1 | 1 | 199.40 | 189.06 | 5.47% | :high_brightness: |
onnx-taau-downsample | 1 | 206.53 | 204.08 | 1.20% | :white_check_mark: |
dlrm-criteoterabyte | 1 | 22.92 | 22.27 | 2.91% | :white_check_mark: |
dlrm-criteoterabyte_fp16 | 1 | 42.66 | 41.57 | 2.64% | :white_check_mark: |
agentmodel | 1 | 6,298.58 | 6,009.87 | 4.80% | :high_brightness: |
unet_fp16 | 2 | 34.19 | 33.46 | 2.17% | :white_check_mark: |
resnet50v1_fp16 | 1 | 586.24 | 553.42 | 5.93% | :high_brightness: |
resnet50v1_int8 | 1 | 579.50 | 463.87 | 24.93% | :high_brightness: |
bert_base_cased_fp16 | 64 | 646.45 | 620.51 | 4.18% | :high_brightness: |
bert_large_uncased_fp16 | 32 | 199.05 | 193.82 | 2.70% | :white_check_mark: |
bert_large_fp16 | 1 | 117.26 | 103.87 | 12.89% | :high_brightness: |
distilgpt2_fp16 | 16 | 1,212.37 | 1,155.49 | 4.92% | :high_brightness: |
yolov5s | 1 | 302.46 | 298.89 | 1.19% | :white_check_mark: |
tinyllama | 1 | 23.33 | 23.33 | 0.01% | :white_check_mark: |
vicuna-fastchat | 1 | 134.21 | 132.40 | 1.37% | :white_check_mark: |
whisper-tiny-encoder | 1 | 244.34 | 241.13 | 1.33% | :white_check_mark: |
whisper-tiny-decoder | 1 | 256.72 | 245.64 | 4.51% | :high_brightness: |
This build is not recommended to merge :red_circle:
:red_circle:bert_large_uncased_fp16: FAILED: MIGraphX is not within tolerance - check verbose output
Codecov Report
All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 91.97%. Comparing base (
206e0fd
) to head (6b74824
). Report is 143 commits behind head on develop.
Additional details and impacted files
@@ Coverage Diff @@
## develop #3011 +/- ##
===========================================
+ Coverage 91.94% 91.97% +0.02%
===========================================
Files 489 489
Lines 19333 19390 +57
===========================================
+ Hits 17776 17833 +57
Misses 1557 1557
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.