oneDNN
oneDNN copied to clipboard
Fix simd uni instructions for sse41
Description
This PR is fixing issues with sse41 for uni_* operations.
Consider the following example:
uni_vsubps(xmm_y0, xmm_pred_ctr_y, xmm_y0);
, where xmm_y0 = xmm0, xmm_pred_ctr_y = xmm8
Previously this example was converted on sse41 in wrong logical instructions:
movups(xmm0, xmm8);
addps(xmm0, xmm0);
As you can see, last instruction is logically wrong addps(xmm0, xmm0);
openvino PR: https://github.com/openvinotoolkit/openvino/pull/13472
@ceciliapeng2011 @dmitry-gorokhov @maxnick Please, take a look or assign reviewers
@dmitry-gorokhov, could you please take a look?
@dmitry-gorokhov @nshchego @chenhu-wang Do you have some comments for this PR ?
@dmitry-gorokhov @nshchego @chenhu-wang Do you have additional comments ?