[libclc] Improving vector code generated from scalar code #140008
Conversation
The previous method splits vector data into two halves. shuffle_vector concatenates the two results into a vector data of original size. This PR eliminates the use of shuffle_vector.
|
@frasercrmck please help to review, thanks. |
There was a problem hiding this comment.
LGTM. I don't think there's any reason to recursively split like this when we only provide the scalar case.
IIRC downstream we have some "have-2" variants of these macros where this splitting might be helpful, as explicitly providing a vector2 form indicates it might be more optimal to use that as the base case. But that's not for here!
I'd also like to rename those macros at some point as I think they obfuscate the kind of vectorization (basically, scalarization) going on.
| #define s4 2.66422703033638609560e-02 /* 0x3F9B481C, 0x7E939961 */ | ||
| #define s5 1.84028451407337715652e-03 /* 0x3F5E26B6, 0x7368F239 */ | ||
| #define s6 3.19475326584100867617e-05 /* 0x3F00BFEC, 0xDD17E945 */ | ||
| #define s0_d -7.72156649015328655494e-02 /* 0xBFB3C467, 0xE37DB0C8 */ |
LGTM thanks @frasercrmck for the review. Please help to merge, thanks. |
The previous method splits vector data into two halves. shuffle_vector concatenates the two results into a vector data of original size. This PR eliminates the use of shuffle_vector.