ROCmSoftwarePlatform
Switch to c++23
This upgrades the version to c++23 for our source code and for kernels. We still cannot use newer std library features because sles uses gcc 7.
- Update
any_castof type erasure to handle casting to classes that dont match the interface(I am not sure why this worked before as it seems obvious why it would fail). In this case, theany_castmember now has the constraints so it cant be called with invalid class, and theany_castfunction are overloaded so ifany_castmember cant be called it will return null(or throw depending on which version). - Disable clang tidy checks for newer c++. We can look at trying to reenable this in the future, but most of them require a newer c++ library which we cannot use because of sles. Also, the fixits for the using
requirekeyword instead ofenable_ifis completely broken. - Disable the
-Wdeprecated-this-capturewarning, since it leads to too many false positives. If we fix the deprecated warning, then we usually get another warning about unused capture. We can try and address this at a later time. - Fix dangling references to temporary in parse_gridsample, most likely due to changes in the
std::accumulatealgorithm.
Since MSVC implements P0533R9 this breaks when compiling with hip. I opened issue here for the compiler team to fix this. For now, we need use c++17 on windows, but this may lead subpar workarounds until the issue is fixed.
This whole exercise should ideally be first verified on the Windows platform.
Msvc is not as up-2-date in its support matrix for c++23 as gcc & clang. https://en.cppreference.com/w/cpp/compiler_support/23.
| Test | Batch | Rate new ae7af2 |
Rate old 052e8f |
Diff | Compare |
|---|---|---|---|---|---|
| torchvision-resnet50 | 64 | 3,158.56 | 3,158.08 | 0.02% | :white_check_mark: |
| torchvision-resnet50_fp16 | 64 | 6,593.45 | 6,592.78 | 0.01% | :white_check_mark: |
| torchvision-densenet121 | 32 | 2,439.55 | 2,438.75 | 0.03% | :white_check_mark: |
| torchvision-densenet121_fp16 | 32 | 4,119.25 | 4,130.05 | -0.26% | :white_check_mark: |
| torchvision-inceptionv3 | 32 | 1,666.20 | 1,665.61 | 0.04% | :white_check_mark: |
| torchvision-inceptionv3_fp16 | 32 | 2,588.64 | 2,588.98 | -0.01% | :white_check_mark: |
| cadene-inceptionv4 | 16 | 794.47 | 795.18 | -0.09% | :white_check_mark: |
| cadene-resnext64x4 | 16 | 802.72 | 802.87 | -0.02% | :white_check_mark: |
| slim-mobilenet | 64 | 8,210.90 | 8,208.93 | 0.02% | :white_check_mark: |
| slim-nasnetalarge | 64 | 221.80 | 221.84 | -0.02% | :white_check_mark: |
| slim-resnet50v2 | 64 | 3,297.97 | 3,295.13 | 0.09% | :white_check_mark: |
| bert-mrpc-onnx | 8 | 1,133.71 | 1,135.52 | -0.16% | :white_check_mark: |
| bert-mrpc-tf | 1 | 488.74 | 488.63 | 0.02% | :white_check_mark: |
| pytorch-examples-wlang-gru | 1 | 315.46 | 309.76 | 1.84% | :white_check_mark: |
| pytorch-examples-wlang-lstm | 1 | 435.76 | 446.29 | -2.36% | :white_check_mark: |
| torchvision-resnet50_1 | 1 | 807.63 | 806.43 | 0.15% | :white_check_mark: |
| cadene-dpn92_1 | 1 | 433.48 | 425.59 | 1.85% | :white_check_mark: |
| cadene-resnext101_1 | 1 | 368.72 | 360.96 | 2.15% | :white_check_mark: |
| onnx-taau-downsample | 1 | 398.13 | 398.07 | 0.01% | :white_check_mark: |
| dlrm-criteoterabyte | 1 | 31.91 | 31.93 | -0.04% | :white_check_mark: |
| dlrm-criteoterabyte_fp16 | 1 | 50.98 | 51.01 | -0.05% | :white_check_mark: |
| agentmodel | 1 | 9,706.38 | 9,421.43 | 3.02% | :high_brightness: |
| unet_fp16 | 2 | 59.00 | 59.07 | -0.12% | :white_check_mark: |
| resnet50v1_fp16 | 1 | 999.50 | 993.23 | 0.63% | :white_check_mark: |
| resnet50v1_int8 | 1 | 992.98 | 999.58 | -0.66% | :white_check_mark: |
| bert_base_cased_fp16 | 64 | 1,099.51 | 1,099.48 | 0.00% | :white_check_mark: |
| bert_large_uncased_fp16 | 32 | 343.77 | 343.86 | -0.03% | :white_check_mark: |
| bert_large_fp16 | 1 | 198.50 | 197.89 | 0.31% | :white_check_mark: |
| distilgpt2_fp16 | 16 | 2,075.07 | 2,075.60 | -0.03% | :white_check_mark: |
| yolov5s | 1 | 589.40 | 588.33 | 0.18% | :white_check_mark: |
| tinyllama | 1 | 43.78 | 43.77 | 0.02% | :white_check_mark: |
| vicuna-fastchat | 1 | 45.06 | 44.80 | 0.58% | :white_check_mark: |
| whisper-tiny-encoder | 1 | 410.29 | 409.17 | 0.27% | :white_check_mark: |
| whisper-tiny-decoder | 1 | 414.26 | 413.25 | 0.25% | :white_check_mark: |
| llama2_7b | 1 | 19.12 | 19.11 | 0.05% | :white_check_mark: |
| qwen1.5-7b | 1 | 23.44 | 23.42 | 0.09% | :white_check_mark: |
| phi3-3.8b | 1 | 26.61 | 26.62 | -0.05% | :white_check_mark: |
| mask-rcnn | 1 | 11.96 | 12.20 | -1.99% | :white_check_mark: |
| llama3-8b | 1 | 21.68 | 21.68 | -0.03% | :white_check_mark: |
| whisper-large-encoder | 1 | 10.17 | 10.17 | 0.00% | :white_check_mark: |
| whisper-large-decoder | 1 | 99.95 | 99.50 | 0.45% | :white_check_mark: |
| mistral-7b | 1 | 23.66 | 23.66 | -0.03% | :white_check_mark: |
| FLUX.1-schnell | 1 | 727.86 | 727.11 | 0.10% | :white_check_mark: |
| nan | nan | nan | nan | nan% | :x: |
This build is not recommended to merge :red_circle:
:x:bert-mrpc-tf: ERROR - check error output
2025-09-10 10:38:02.763843: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
Traceback (most recent call last):
File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 359, in
main()
File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 306, in main
graph = load_tf_graph(model_name)
File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 300, in load_tf_graph
graph_def.ParseFromString(f.read())
File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/lib/io/file_io.py", line 116, in read
self._preread_check()
File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/lib/io/file_io.py", line 77, in _preread_check
self._read_buf = _pywrap_file_io.BufferedInputStream(
tensorflow.python.framework.errors_impl.UnimplementedError: File system scheme '[local]' not implemented (file: '/new-saved-models/tf-misc/bert_mrpc1.pb'):red_circle:bert_large_uncased_fp16: FAILED: MIGraphX is not within tolerance - check verbose output
:red_circle:mask-rcnn: FAILED: MIGraphX is not within tolerance - check verbose output
