https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120598
Bug ID: 120598
Summary: Compiler is unable to vectorise scalar code
Product: gcc
Version: 15.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libgcc
Assignee: unassigned at gcc dot gnu.org
Reporter: Manjul.Mohan at ibm dot com
Target Milestone: ---
Created attachment 61605
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=61605&action=edit
As we can see, the scalar function is not getting vectorised in case of MariaDB
In MariaDB
There is a function dot_product, which is not getting vectorised, even with the
latest gcc (15.1)
The function signature is as follows:
float dot_product (const int16_t *v1, const int16_t *v2, size_t len);
Currently we are manually optimised code in mariadb by the use of vector
intrinsics (https://github.com/MariaDB/server/pull/3850/files)
We’ve observed that while functions in the PGVector library benefit from both
loop unrolling and auto-vectorization (even with earlier versions of GCC, like
13.3 and 11.5), the same does not hold true for the dot_product function in the
MariaDB library — even when compiled with GCC 15.1
One potential reason could be that, unlike PGVector where both the input
parameters and the return value from the function is of type float:
VECTOR_TARGET_CLONES static float VectorInnerProduct(int dim, float *ax, float
*bx);
Whereas in case of MariaDB implementation requires promotion from int16_t to
float:
float dot_product (const int16_t *v1, const int16_t *v2, size_t len);
This implicit type promotion may be preventing effective vectorization.
