Bartłomiej Kocot

Results 11 comments of Bartłomiej Kocot

I have good idea for this problem: - Add struct SCALAR : public BaseTensorLayout - Pass Tuple as layouts for Ds - Pass only one value with shape ={1}, strides={0}...

Hi @iq136boy, due to error message: `max err: 0.0001745224, number of errors: 1760, 0.02872243% wrong values` I suppose it is catastrophic cancellation. CPU can produce better results due to internal...

> May I ask when CK will deprecate the support of 2d convolution instances? This has a significant impact on MIOpen side because there are several solvers in MIOpen based...

Hi @hclearner , I prepared some example of creating layout using tensor descriptor and some draft of Layout wrapper: #1054 Can You look at this? Which version is more preferred?

@hclearner Yes, I prepared this example to make sure that such API: https://github.com/ROCmSoftwarePlatform/composable_kernel/pull/1054/files#diff-cd0bbbc81c77205e3a568fe45046b2e20f8cc7fe5a2d3c991a59add9f2c63d61 is needed. So I will move this from the examples to the exact library and develop it...

Hi @hclearner , thanks. Maybe I'm wrong but I don't see CuTe usage in https://github.com/facebookresearch/xformers/tree/main/xformers/csrc/attention/cuda/fmha. Or do you need this just because it would be easier for You to use...

Hi @hclearner , after rapid development we prepared basic version of requested API. You can refer to: - [Documentation](https://rocm.docs.amd.com/projects/composable_kernel/en/develop/wrapper.html) - [GEMM Tutorial](https://github.com/ROCm/composable_kernel/blob/develop/client_example/25_wrapper/README.md) - [Image to column example](https://github.com/ROCm/composable_kernel/blob/develop/client_example/25_wrapper/wrapper_img2col.cpp) - [ Basic...

Hi @chengl-amd , try[ this branch ](https://github.com/ROCm/composable_kernel/tree/barkocot/relu-bias-int8) , I changed AccDataType and CShuffleDataType to `int32_t` (which is supported with `int8_t`).

Hi @ozturkosu why this example is called as complex? I see that you use real data types, not complex.

Hi @ozturkosu , does this PR contain changes from this PR #1061 ?