On Tue, 15 Sep 2026, Tamar Christina wrote:
> > -----Original Message-----
> > From: Richard Biener <[email protected]>
> > Sent: 15 September 2026 10:16
> > To: Tamar Christina <[email protected]>
> > Cc: Ezra Sitorus <[email protected]>; [email protected];
> > [email protected]
> > Subject: RE: [RFC PATCH 3/5] gcc: Add new COMPLEX_DOT_PROD_ROT0
> > IFN/optabs
> >
> > On Tue, 15 Sep 2026, Tamar Christina wrote:
> >
> > > > -----Original Message-----
> > > > From: Richard Biener <[email protected]>
> > > > Sent: 15 September 2026 09:23
> > > > To: Tamar Christina <[email protected]>
> > > > Cc: Ezra Sitorus <[email protected]>; [email protected];
> > > > [email protected]
> > > > Subject: RE: [RFC PATCH 3/5] gcc: Add new COMPLEX_DOT_PROD_ROT0
> > > > IFN/optabs
> > > >
> > > > On Tue, 15 Sep 2026, Tamar Christina wrote:
> > > >
> > > > >
> > > > >
> > > > > > -----Original Message-----
> > > > > > From: [email protected] <[email protected]>
> > > > > > Sent: 09 September 2026 19:12
> > > > > > To: [email protected]
> > > > > > Cc: Tamar Christina <[email protected]>; [email protected];
> > > > > > [email protected]; Ezra Sitorus
> > <[email protected]>
> > > > > > Subject: [RFC PATCH 3/5] gcc: Add new COMPLEX_DOT_PROD_ROT0
> > > > > > IFN/optabs
> > > > > >
> > > > > > From: Ezra Sitorus <[email protected]>
> > > > > >
> > > > > > Add new internal function COMPLEX_DOT_PROD_ROT0 and optab
> > > > > > cdot_prod_rot0_optab
> > > > > > to represent complex integer dot product.
> > > > > >
> > > > > > COMPLEX_DOT_PROD_ROT0 (A, B, ACC) is a lane-reducing signed
> > integer
> > > > > > complex
> > > > > > dot product. The 2 input vectors contain interleaved W-bit real and
> > > > imaginary
> > > > > > components in even/odd numbered lanes respectively. Each 4W-bit
> > result
> > > > > > element
> > > > > > accumulates the 0-rotation products of the 2 complex input pairs in
> > > > > > the
> > > > > > corresponding group of 4 input lanes:
> > > > > >
> > > > > > result[j] = acc[j]
> > > > > > + a[k].re * b[k].re
> > > > > > - a[k].im * b[k].im
> > > > > > + a[k + 1].re * b[k + 1].re
> > > > > > - a[k + 1].im * b[k + 1].im
> > > > > >
> > > > > > where 2*j = k.
> > > > > >
> > > > > > 0-rotation means that the imaginary components in the second source
> > > > vector
> > > > > > are
> > > > > > negated, and so the overall operation is accumulating the real part
> > > > > > of
> > the
> > > > > > complex dot product. Other rotations can be added in future patches.
> > > >
> > > > I other places we use _conj, not rot0:
> > > >
> > > > OPTAB_D (cmul_conj_optab, "cmul_conj$a3")
> > > >
> > > > so why not use that for consistency?
> > >
> > > That's a different operation. complex dotprod is numerically
> >
> > Sure, I meant why cdot_prod_rot0 rather than cdot_prod_conj,
> > complex conjugate is exactly negating the imaginary part?
>
> Because the existing cmul/cadd/cdot *_conj optabs are intended to
> represent the full mathematical complex operation which create both
> the real and imaginary results.
>
> Yes, cdot #0 conjugates the second operand in the sense that it uses
> ar * br - ai * bi. But it only computes the real component of the
> complex dot product. The imaginary component is not computed.
Oh, but then the name cdot_* is even more confusing? Why even bother
with naming it "complex" then? It's not even a dot product?
res += a[i].real * b[i].real
- a[i].imag * b[i].imag;
that suggests cdot0 computes
res + x * y - z * w
or is it, for vectors, even more complicates as
res + x[2*n] * y[2*n] - x[2*n+1] * y[2*n+1]
? Aka it's an alternating plus/minus non-"complex" dot-product
(res + x[2*n] * y[2*n]) - x[2*n+1] * y[2*n+1]
as if negating every other lane in x[] and then using a regular
(non-complex) dot product? And it's also a reduction, so
reduc_dot_<?>_scal? with <?> standing for whatever we invent
for the alternating negation (we have vec_addsub and vec_fmaddsub
and vec_fmsubadd for example, but that's not reductions).
I see you are following the aarch64 SVE2 'CDOT' instruction
notation with its #rotate specification. I'm a bit unsure
how to translate that appropriately also given the unfortunate(?)
choice to compute a "packed" real part from two-component inputs.
> For example, cmul_conj today lowers to both:
>
> fcmla #180
> fcmla #270
>
> because both operations are needed to compute the full complex multiply
> by a conjugated operand.
>
> Similarly, computing both components of a complex dot product would need
> Two instructions and rotations:
>
> cdot #0
> cdot #90
>
> So naming the single cdot #0 operation cdot_prod_conj would be
> misleading: it suggests the optab computes the full conjugated complex
> dot product, whereas it only computes one rotated component.
>
> This is the same distinction as for complex multiply rotations, discussed
> in PR121925, where we settled on the "rot" suffix for the lane-rotation forms.
>
> This isn't dissimilar to complex multiply and rotate which we discussed here
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121925 and settled on the "rot"
> suffix for them.
>
> Thanks,
> Tamar
> >
> > > #include <stdint.h>
> > >
> > > typedef struct {
> > > int8_t real;
> > > int8_t imag;
> > > } cint16_t;
> > >
> > > typedef struct {
> > > int32_t real;
> > > int32_t imag;
> > > } cint32_t;
> > >
> > > cint32_t cdot(cint16_t *a, cint16_t *b, uint32_t N)
> > > {
> > > cint32_t res = {0, 0};
> > >
> > > for (uint32_t i = 0; i < N; i++) {
> > > res.real += a[i].real * b[i].real
> > > - a[i].imag * b[i].imag;
> > >
> > > res.imag += a[i].real * b[i].imag
> > > + a[i].imag * b[i].real;
> > > }
> > >
> > > return res;
> > > }
> > >
> > > Which would be what we'd call cdot.
> > >
> > > cdot_conj would be
> > >
> > > #include <stdint.h>
> > >
> > > typedef struct {
> > > int8_t real;
> > > int8_t imag;
> > > } cint16_t;
> > >
> > > typedef struct {
> > > int32_t real;
> > > int32_t imag;
> > > } cint32_t;
> > >
> > > cint32_t cdot_conj(const cint16_t *a,
> > > const cint16_t *b,
> > > uint32_t N)
> > > {
> > > cint32_t res = {0, 0};
> > >
> > > for (uint32_t i = 0; i < N; i++) {
> > > res.real += a[i].real * b[i].real
> > > + a[i].imag * b[i].imag;
> > >
> > > res.imag += a[i].real * b[i].imag
> > > - a[i].imag * b[i].real;
> > > }
> > >
> > > return res;
> > > }
> > >
> > > i.e. normal dual lane operations.
> > >
> > > however this patch recognizes the single lane reduction,
> > >
> > > #include <stdint.h>
> > >
> > > typedef struct {
> > > int8_t real;
> > > int8_t imag;
> > > } cint16_t;
> > >
> > > int32_t cdot0(const cint16_t *a, const cint16_t *b, uint32_t N)
> > > {
> > > int32_t res = 0;
> > >
> > > for (uint32_t i = 0; i < N; i++)
> > > res += a[i].real * b[i].real
> > > - a[i].imag * b[i].imag;
> > >
> > > return res;
> > > }
> > >
> > > Which is not numerically a complex dot product. Note that the end goal
> > > is to recognize all of these.
> > >
> > > Thanks,
> > > Tamar
> > > >
> > > > > > In internal-fn.cc: complex_dot_prod_direct is defined to { -1, 1,
> > > > > > false }
> > > > > > because it uses the return type as the accumulator/result type and
> > > > > > the
> > > > input
> > > > > > type is either of the narrower source type.
> > > > > >
> > > > > > In AArch64, this corresponds to the CDOT instruction. This patch
> > > > > > also
> > > > teaches
> > > > > > GCC to convert the optab into the correct define_insn pattern.
> > > > > >
> > > > > > gcc/ChangeLog:
> > > > > > * config/aarch64/aarch64-sve2.md
> > > > > > (cdot_prod_rot0<mode><vsi2qi>): New
> > > > > > define_expand.
> > > > > > * internal-fn.cc (complex_dot_prod_direct): New static
> > > > > > initializers.
> > > > > > * internal-fn.def (COMPLEX_DOT_PROD_ROT0) New internal
> > > > > > function.
> > > > > > * optabs.def (cdot_prod_rot0_optab): New optab.
> > > > > > ---
> > > > > > gcc/config/aarch64/aarch64-sve2.md | 10 ++++++++++
> > > > > > gcc/internal-fn.cc | 5 +++++
> > > > > > gcc/internal-fn.def | 1 +
> > > > > > gcc/optabs.def | 1 +
> > > > >
> > > > > The new optabs and IFN need to be documented in ifn.texi and md.texi
> > > > > and split the aarch64 implementation to it's own patch.
> > > > >
> > > > > But why not add the other rotations at the same time?
> > > > >
> > > > > We'd likely also later (in a different patch) need cdot_prod and
> > > > cdot_prod_conj,
> > > > > because then both lanes are visible the SLP tree will look much
> > > > > different.
> > i.e.
> > > > >
> > > > > #include <stdint.h>
> > > > >
> > > > > typedef _Complex int8_t cint16_t;
> > > > > typedef _Complex int32_t cint32_t;
> > > > >
> > > > > cint32_t cdot(cint16_t *a, cint16_t *b, uint32_t N)
> > > > > {
> > > > > cint32_t res = 0;
> > > > >
> > > > > for (uint32_t i = 0; i < N; i++)
> > > > > res += a[i] * b[i];
> > > > >
> > > > > return res;
> > > > > }
> > > > >
> > > > > > 4 files changed, 17 insertions(+)
> > > > > >
> > > > > > diff --git a/gcc/config/aarch64/aarch64-sve2.md
> > > > > > b/gcc/config/aarch64/aarch64-sve2.md
> > > > > > index c33df283cc2a..ad4febb60bf8 100644
> > > > > > --- a/gcc/config/aarch64/aarch64-sve2.md
> > > > > > +++ b/gcc/config/aarch64/aarch64-sve2.md
> > > > > > @@ -3578,6 +3578,16 @@
> > > > > > ;; - CDOT
> > > > > > ;;
> > > > > > -------------------------------------------------------------------------
> > > > > >
> > > > > > +(define_expand "cdot_prod_rot0<mode><vsi2qi>"
> > > > > > + [(set (match_operand:SVE_FULL_SDI 0 "register_operand")
> > > > > > + (unspec:SVE_FULL_SDI
> > > > > > + [(match_operand:SVE_FULL_SDI 3 "register_operand")
> > > > > > + (match_operand:<VSI2QI> 1 "register_operand")
> > > > > > + (match_operand:<VSI2QI> 2 "register_operand")]
> > > > > > + UNSPEC_CDOT))]
> > > > > > + "TARGET_SVE2"
> > > > > > +)
> > > > > > +
> > > > > > (define_insn "@aarch64_sve_<optab><mode>"
> > > > > > [(set (match_operand:SVE_FULL_SDI 0 "register_operand")
> > > > > > (unspec:SVE_FULL_SDI
> > > > > > diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
> > > > > > index c57b594ac591..cc968109261b 100644
> > > > > > --- a/gcc/internal-fn.cc
> > > > > > +++ b/gcc/internal-fn.cc
> > > > > > @@ -180,6 +180,7 @@ init_internal_fns ()
> > > > > > #define unary_convert_direct { -1, 0, true }
> > > > > > #define binary_direct { 0, 0, true }
> > > > > > #define ternary_direct { 0, 0, true }
> > > > > > +#define complex_dot_prod_direct { -1, 1, false }
> > > > > > #define cond_unary_direct { 1, 1, true }
> > > > > > #define cond_binary_direct { 1, 1, true }
> > > > > > #define cond_ternary_direct { 1, 1, true }
> > > > > > @@ -4315,6 +4316,9 @@ expand_reduc_sbool_optab_fn (internal_fn
> > fn,
> > > > > > gcall *stmt, direct_optab optab)
> > > > > > #define expand_select_vl_optab_fn(FN, STMT, OPTAB) \
> > > > > > expand_convert_optab_fn (FN, STMT, OPTAB, 3)
> > > > > >
> > > > > > +#define expand_complex_dot_prod_optab_fn(FN, STMT, OPTAB) \
> > > > > > + expand_convert_optab_fn (FN, STMT, OPTAB, 3)
> > > > > > +
> > > > > > /* Expanders for optabs that can use expand_convert_optab_fn. */
> > > > > >
> > > > > > #define expand_unary_convert_optab_fn(FN, STMT, OPTAB) \
> > > > > > @@ -4398,6 +4402,7 @@ multi_vector_optab_supported_p
> > > > (convert_optab
> > > > > > optab, tree_pair types,
> > > > > > #define direct_unary_convert_optab_supported_p
> > > > > > convert_optab_supported_p
> > > > > > #define direct_binary_optab_supported_p direct_optab_supported_p
> > > > > > #define direct_ternary_optab_supported_p direct_optab_supported_p
> > > > > > +#define direct_complex_dot_prod_optab_supported_p
> > > > > > convert_optab_supported_p
> > > > > > #define direct_cond_unary_optab_supported_p
> > > > direct_optab_supported_p
> > > > > > #define direct_cond_binary_optab_supported_p
> > > > direct_optab_supported_p
> > > > > > #define direct_cond_ternary_optab_supported_p
> > > > direct_optab_supported_p
> > > > > > diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> > > > > > index af9f92950c73..08bdedc30045 100644
> > > > > > --- a/gcc/internal-fn.def
> > > > > > +++ b/gcc/internal-fn.def
> > > > > > @@ -453,6 +453,7 @@ DEF_INTERNAL_OPTAB_FN
> > > > (COMPLEX_ADD_ROT90,
> > > > > > ECF_CONST, cadd90, binary)
> > > > > > DEF_INTERNAL_OPTAB_FN (COMPLEX_ADD_ROT270, ECF_CONST,
> > > > cadd270,
> > > > > > binary)
> > > > > > DEF_INTERNAL_OPTAB_FN (COMPLEX_MUL, ECF_CONST, cmul,
> > binary)
> > > > > > DEF_INTERNAL_OPTAB_FN (COMPLEX_MUL_CONJ, ECF_CONST,
> > > > cmul_conj,
> > > > > > binary)
> > > > > > +DEF_INTERNAL_OPTAB_FN (COMPLEX_DOT_PROD_ROT0,
> > ECF_CONST,
> > > > > > cdot_prod_rot0, complex_dot_prod)
> > > > > > DEF_INTERNAL_OPTAB_FN (VEC_ADDSUB, ECF_CONST, vec_addsub,
> > > > binary)
> > > > > > DEF_INTERNAL_OPTAB_FN (VEC_TRUNC_ADD_HIGH, ECF_CONST |
> > > > > > ECF_NOTHROW,
> > > > > > vec_trunc_add_high, binary)
> > > > > > diff --git a/gcc/optabs.def b/gcc/optabs.def
> > > > > > index 9d95f047cb50..b30e2bec6b2b 100644
> > > > > > --- a/gcc/optabs.def
> > > > > > +++ b/gcc/optabs.def
> > > > > > @@ -114,6 +114,7 @@ OPTAB_CD(vec_init_optab, "vec_init$a$b")
> > > > > > OPTAB_CD (sdot_prod_optab, "sdot_prod$I$a$b")
> > > > > > OPTAB_CD (udot_prod_optab, "udot_prod$I$a$b")
> > > > > > OPTAB_CD (usdot_prod_optab, "usdot_prod$I$a$b")
> > > > > > +OPTAB_CD (cdot_prod_rot0_optab, "cdot_prod_rot0$I$a$b")
> > > > > >
> > > > > > OPTAB_CD (while_ult_optab, "while_ult$a$b")
> > > > > > OPTAB_CD (select_vl_optab, "select_vl$a$b")
> > > > > > --
> > > > > > 2.53.0
> > > > >
> > > > >
> > > >
> > > > --
> > > > Richard Biener <[email protected]>
> > > > SUSE Software Solutions Germany GmbH,
> > > > Frankenstrasse 146, 90461 Nuernberg, Germany;
> > > > GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)
> > >
> >
> > --
> > Richard Biener <[email protected]>
> > SUSE Software Solutions Germany GmbH,
> > Frankenstrasse 146, 90461 Nuernberg, Germany;
> > GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)
