[DRAFT] Fix op tensor nhwc #3190
Conversation
… tensors. Improve error messages.
…ut is not set explicitly and strides are known. [API] Extend layouts with 1/2/3D defauls and "unknown", for internal use.
… for internal use. [tensor] Add support for non-default 1/2/3D layouts.
…icit "NCHWc4" and "NCHWc8", "CHWNc" - to "CHWNc4" and "CHWNc8". This is necessary to fully support conversion of strings to enums.
… non-default layouts
|
@CAHEK7 @averinevg For early preview |
|
Tensor operations are layout agnostic, the most important limitation is all the tensors must have the same layout. Another limitation from the current implementation - all the kernels, even the generic one, do not support strides in the last dimension, so even if the layout is not default, but the last dimension is tightly packed, we can use most of the kernels (in the other words, we must check that the last dimension stride is 1 and skip any layout checks). Technically we even can omit this check |
… & adapt to changes from PR ROCm#3213 # RESOLVED Conflicts: # src/include/miopen/conv/problem_description.hpp # src/include/miopen/graphapi/tensor.hpp # src/include/miopen/tensor.hpp # src/tensor.cpp
…t4D5D, GetLayoutEnum, GetLayout_str: add suport for n_dims = {1,2,3}. Remove the "4D5D" suffices.
…oes not resolve the issue; remove it. [tensor] Removed notion of "default layout".
|
🛠️ Detection of "mismatched tensor layouts" cannot help to resolve the problem. The PR will be reworked. |
WIP for #2860