https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116685
Bug ID: 116685
Summary: RISC-V: missed optimization on vector dot products
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: other
Assignee: unassigned at gcc dot gnu.org
Reporter: ewlu at rivosinc dot com
Target Milestone: ---
Godbolt: https://godbolt.org/z/arM88Pa81
Vector definitions and expanded routines extracted from povray
https://github.com/POV-Ray/povray/blob/master/source/core/math/vector.h and
related files
I expanded some of the templates and added a couple basic vector routines which
may be used in a workload.
Gcc tends to avoid vectorizing dot products or fails to vectorize appropriately
2 element vector dot product (failure to vectorize):
dotDbl_2d(GenericVector2d<double> const&, GenericVector2d<double> const&):
fld fa5,8(a0) <-- could reduce number of loads with vsetivli
zero,2
fld fa3,8(a1)
fld fa0,0(a0)
fld fa4,0(a1)
fmul.d fa5,fa5,fa3
fmadd.d fa0,fa0,fa4,fa5
ret
3 element vector dot product (failure to vectorize):
dotDbl_3d(GenericVector3d<double> const&, GenericVector3d<double> const&):
fld fa3,8(a1) <-- could use with vsetivli zero,2 or vsetivli zero,3
fld fa4,8(a0)
fld fa5,0(a0)
fld fa2,0(a1)
fmul.d fa4,fa4,fa3
fld fa0,16(a0)
fld fa3,16(a1)
fmadd.d fa5,fa5,fa2,fa4
fmadd.d fa0,fa0,fa3,fa5
ret
4 element vector dot product (improper vectorization):
V4D_Dot(double&, double const*, double const*):
vsetivli zero,2,e64,m1,ta,ma <-- should be vsetivli zero,4
addi a4,a1,16
addi a5,a2,16
vle64.v v3,0(a5)
vle64.v v2,0(a4)
vle64.v v1,0(a1)
vle64.v v4,0(a2)
vfmul.vv v2,v2,v3
vmv.s.x v3,zero
vfmadd.vv v1,v4,v2
vfredusum.vs v1,v1,v3
vfmv.f.s fa5,v1
fsd fa5,0(a0)
ret
For odd element vectorization like the 3 element vector case, clang currently
uses vsetivli zero,2 and performs the last part of the dot product using scalar
code which gcc could potentially do as well.
While the current scalar code does use 1 fewer instruction compared to
vectorized code, it issues around twice as many loads than vectorized which may
impact performance.
