bibek

In CK for (conv + bias + activation) we currently have activation as RELU. It would be great to have other activations. We can start with the ones that take...

List of GEMM+epilogues fusion needed for Transformer Engine

1

The ones we need for Transformer Engine are the following: 1) CUBLASLT_EPILOGUE_GELU_AUX step 1 : matrix multiplication step 2 : apply gelu step 3 : store the result to seperate...

Currently the [conv + bias + activ kernel is only for type f16](https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp). It would be great if we had similar kernel for F32 and other data types.

Integrate conv bilinear ck solver lwp 633

1

Integrate CK's 3D bilinear solvers. This PR is combination of https://github.com/ROCm/MIOpen/pull/2817 and https://github.com/ROCm/MIOpen/pull/2912

address issues raised in PR #2154

5

* [use struct to return data ](https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2154/files#r1211864942) * [see if passing of unique_ptr is really necesary](https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2154/files#r1199790458) * [remove item added to target_link_library that was part of workaround](https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2177) This issue is...

measurement tool useful for debugging of performance regressions

5

* This PR finds the performance hipMemcpy, hipMemset and SetTensor The performance results for tensor {4, 4, 1115, 1115} in micro sec were : hipMemcpy = 29717 hipMemset = 1...

CK's fp8 col major solver integration into MIOpen

How to tune using API?

5

Can we do tuning using API? Lets say if I want to tune fused solver or regular solver. I can do: MIOPEN_FIND_ENFORCE=3 MIOPEN_FIND_MODE=1 ./bin/MIOpenDriver CBAInfer -F 4 -n 128 -c...

adding mha as static lib

2

* This PR is to generate the mha static lib from generate.py

- Enable NCHW/NHWC layout from driver command for batch norm. - Moved `GpumemTensor` to `driver/driver.hpp` - Stopped using old and slow `miopenBNFwdTrainPerActivationRunHost`, `miopenBNFwdTrainSpatialRunHost`, `miopenBNFwdInferPerActivationRunHost`, `miopenBNFwdInferSpatialRunHost`, `miopenBNBwdPerActivationRunHost` and `miopenBNBwdSpatialRunHost` since they...