Release 0.2.0

What is new in 0.2.0

Bug/Issue Fixes

• Fixed incorrect use of double constants in some operators
• Fixed crash when loading models that were saved on OCL devices
• Fixed default parameter of torch.ocl.synchronize
• Fixed failure of printing on Intel devices with missing fp64 support

New nets Validated

Visual transformers vit_transformets and vit_x_NN ets validated

New operators implemented:

• resize_, arange, mm, bmm, amin, amax, addmm, _native_multi_head_attention and transform_bias_rescale_qkv, round, maximum, minimum, prod, atan, dropout_native
• lt,le,gt,ge,eq,ne for tensors
• bitwise ^, |, &, ~
• upsample_2d : bilinear, nearest and nearest exact, forward and backward

Fixed operators

• Fixed softmax and log softmax support of dim that is not last dim
• Fixed view operator and set_ storage
• cat now supports mixed types
• Fix handling of empty tensors with non empty storage
• Very limited half tensor handling
• Fixed tensor >, < ==, != scalar ops

New features:

• Added support of profiling via torch.ocl.profile API
• Improved benchmark scripts

Performance improvements

• Intel Arc, UHD - enabled winograd convolution, support of OpenCL 3.0 floating point add atomics, enabled k-reduction for GEMM operators
• NVidia - added use of native atomic float add (via PTX assembly)
• GELU major improvements due to faulty use of double instead of float

Release 0.1.0

This release supports pytorch 2.4 and introduces better way to use OpenCL pytorch

This time I provided binary distributions of the backend

• Linux for python 3.8 till 3.12, for torch=2.4
• Windows for python 3.11 and 3.12 for torch=2.4

To install - install CPU version of pytorch in virtual evironment, download whl file from release make sure torch version, python version and architecture matches your environment.

For example python 3.10, torch 2.4 on Linux it is:

pip install pytorch_ocl-0.1.0+torch2.4-cp310-none-linux_x86_64.whl

To use import pytorch_ocl