Cao E

Fix reshape and empty input convolution issues for channels last memory format

1

* Fixes #78611 Reshape tensors witch are channels_last will get unexpected stride. * Fixes empty input convolution issue : when input is empty e.g. shape of (0, 3, 3, 4)...

Add Mx2, Mx4, 2xN, and 4xN avx512 transpose

6

Add Mx2, Mx4, 2xN, and 4xN specific transposes on avx512 to improve the transpose performance of shapes of Mx2, Mx4, 2xN, and 4xN. * When the shape is Mx2 or...

Add inference, amp and channel_last supports for began.

Make q, k, and v contiguous to get better performance for normalize. After the rearrange operations for q, k, and v, normalizations on the last dim for q and k...

Use 32x32 block size for i16 transpose

1

Use 32x32 block size for i16 transpose.

add cl3d support for conv deconv and add arg is_channels_last

1

Fixes #ISSUE_NUMBER This PR depends on https://github.com/intel/ideep/pull/165

Set simdlen based on ATEN_CPU_CAPABILITY

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #123514 It is part of https://github.com/pytorch/pytorch/issues/123224. Set simdlen based on the environment ATEN_CPU_CAPABILITY to control CPU vec ISA like eager. cc @voznesenskym...

Add oneDNN BRGEMM support on CPU

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #133194 * #131879 * __->__ #131878 cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @sanchitintel @ashokei @jingxu10 @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @snadampal

integrate oneDNN brgemm into cpp gemm template for Half

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #133194 * #131879 * #131878 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225...

Use brgemm with pack for Half/BFloat16 flash attention forward kernel

2

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #133194 * __->__ #131879 * #131878 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10