perf: conv1d_and_silu consume/produce [B,S,C] directly, drop transposes - Vectorize over the channel dimension (innermost, contiguous) so each tap is an aligned 64-lane vmul+vadd instead of a per-tap vgather. - Read/write the token-major [B,S,C] layout directly, removing the two [B,S,C]<->[B,C,S] Transpose passes around the fused conv in the uniform prefill path. - Use bulk multi-burst GM<->UB DMA (dstGap) for tile load/store. - Restore cross-pipe (MTE2/V/MTE3/S) event-flag fences for UB data ordering. - Update XliteOpConv1dAndSiLU uniform validation to [B,S,C]. Signed-off-by: Igor Prusov <prusov.igor@huawei.com> Co-authored-by: DeepSeek V4 Pro
xlite: renormalize line endings Git index had some CRLF files which are checkout on Unix as-is. This commit runs git add --renormalize . command and unifies line endings as LF. Current line endings in repo and in working space can be seen via git ls-files --eol command. Having all files in LF format in index allows both Unix and Windows work correctly with default git settings. Signed-off-by: Igor Prusov <prusov.igor@huawei.com>
