> -----Original Message-----
> From: Richard Biener <richard.guent...@gmail.com>
> Sent: Tuesday, August 20, 2024 10:37 AM
> To: Victor Do Nascimento <victor.donascime...@arm.com>
> Cc: gcc-patches@gcc.gnu.org; Tamar Christina <tamar.christ...@arm.com>;
> claz...@gmail.com; hongtao....@intel.com; s...@gcc.gnu.org; bernds_cb1@t-
> online.de; al...@redhat.com; Victor Do Nascimento
> <vicdo...@e125768.arm.com>
> Subject: Re: [PATCH V3 10/10] autovectorizer: Test autovectorization of
> different
> dot-prod modes.
>
> On Thu, Aug 15, 2024 at 10:46 AM Victor Do Nascimento
> <victor.donascime...@arm.com> wrote:
> >
> > From: Victor Do Nascimento <vicdo...@e125768.arm.com>
> >
> > Given the novel treatment of the dot product optab as a conversion, we
> > are now able to targe different relationships between output modes and
> > input modes.
> >
> > This is made clearer by way of example. Previously, on AArch64, the
> > following loop was vectorizable:
> >
> > uint32_t udot4(int n, uint8_t* data) {
> > uint32_t sum = 0;
> > for (int i=0; i<n; i+=1)
> > sum += data[i] * data[i];
> > return sum;
> > }
> >
> > while the following was not:
> >
> > uint32_t udot2(int n, uint16_t* data) {
> > uint32_t sum = 0;
> > for (int i=0; i<n; i+=1)
> > sum += data[i] * data[i];
> > return sum;
> > }
> >
> > Under the new treatment of the dot product optab, they are both now
> > vectorizable.
> >
> > This adds the relevant target-agnostic check to ensure this behaviour
> > in the autovectorizer, gated behind the new check_effective_target
> > `vect_dotprod_twoway' as well a runtime check targetting aarch64.
>
> I think vect_dotprod_twoway is not clear - does aarch64 now support
> all of qi->hi, hi->si, si->di variants while formerly only "fourway"
> qi->si and hi->di?
>
> If it's just hi->si that's now supported please use vect_dotprod_hisi
> to be specific as to what is required/supported.
At the moment we support only H->S, but other variants are being considered.
So perhaps vect_dotprod_single_step or something to make it clearer what it
does but still be general?
Thanks,
Tamar
>
> Richard.
>
> > gcc/testsuite/ChangeLog:
> >
> > * lib/target-supports.exp
> > (check_effective_target_vect_dotprod_twoway):
> > New.
> > * gcc.dg/vect/vect-dotprod-twoway.c: Likewise.
> > * gcc.target/aarch64/vect-dotprod-twoway.c: Likewise.
> > ---
> > .../gcc.dg/vect/vect-dotprod-twoway.c | 39 +++++++++++
> > .../gcc.target/aarch64/vect-dotprod-twoway.c | 65 +++++++++++++++++++
> > gcc/testsuite/lib/target-supports.exp | 8 +++
> > 3 files changed, 112 insertions(+)
> > create mode 100644 gcc/testsuite/gcc.dg/vect/vect-dotprod-twoway.c
> > create mode 100644 gcc/testsuite/gcc.target/aarch64/vect-dotprod-twoway.c
> >
> > diff --git a/gcc/testsuite/gcc.dg/vect/vect-dotprod-twoway.c
> b/gcc/testsuite/gcc.dg/vect/vect-dotprod-twoway.c
> > new file mode 100644
> > index 00000000000..ff6a2559dee
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/vect/vect-dotprod-twoway.c
> > @@ -0,0 +1,39 @@
> > +/* { dg-do compile } */
> > +/* { dg-require-effective-target vect_dotprod_twoway } */
> > +/* Ensure both the two-way and four-way dot products are autovectorized.
> > */
> > +#include <stdint.h>
> > +
> > +uint32_t udot4(int n, uint8_t* data) {
> > + uint32_t sum = 0;
> > + for (int i=0; i<n; i+=1) {
> > + sum += data[i] * data[i];
> > + }
> > + return sum;
> > +}
> > +
> > +int32_t sdot4(int n, int8_t* data) {
> > + int32_t sum = 0;
> > + for (int i=0; i<n; i+=1) {
> > + sum += data[i] * data[i];
> > + }
> > + return sum;
> > +}
> > +
> > +uint32_t udot2(int n, uint16_t* data) {
> > + uint32_t sum = 0;
> > + for (int i=0; i<n; i+=1) {
> > + sum += data[i] * data[i];
> > + }
> > + return sum;
> > +}
> > +
> > +int32_t sdot2(int n, int16_t* data) {
> > + int32_t sum = 0;
> > + for (int i=0; i<n; i+=1) {
> > + sum += data[i] * data[i];
> > + }
> > + return sum;
> > +}
> > +
> > +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 4 "vect" } } */
> > +/* { dg-final { scan-tree-dump-times "vect_recog_dot_prod_pattern:
> > detected"
> 4 "vect" } } */
> > diff --git a/gcc/testsuite/gcc.target/aarch64/vect-dotprod-twoway.c
> b/gcc/testsuite/gcc.target/aarch64/vect-dotprod-twoway.c
> > new file mode 100644
> > index 00000000000..bac1e1846da
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/aarch64/vect-dotprod-twoway.c
> > @@ -0,0 +1,65 @@
> > +/* { dg-do run } */
> > +/* { dg-require-effective-target vect_dotprod_twoway } */
> > +/* { dg-options "-march=armv8-a+sme2 -static -O3 -ftree-vectorize -fdump-
> tree-vect-details -save-temps" } */
> > +/* Ensure runtime correctness in the autovectorized two-way dot product
> operations. */
> > +
> > +#include <stdint.h>
> > +#include <stdlib.h>
> > +
> > +uint32_t
> > +udot2 (int n, uint16_t* data) __arm_streaming
> > +{
> > + uint32_t sum = 0;
> > + for (int i=0; i<n; i+=1) {
> > + sum += data[i] * data[i];
> > + }
> > + return sum;
> > +}
> > +
> > +int32_t
> > +sdot2 (int n, int16_t* data) __arm_streaming
> > +{
> > + int32_t sum = 0;
> > + for (int i=0; i<n; i+=1) {
> > + sum += data[i] * data[i];
> > + }
> > + return sum;
> > +}
> > +
> > +int
> > +main ()
> > +{
> > +
> > + uint16_t u_input_nil[] = { [0 ... 3] = 0 };
> > + uint16_t u_input_min[] = { [0 ... 3] = 1 };
> > + uint16_t u_input_max[] = { [0 ... 3] = 32767};
> > +
> > + uint32_t u_nil_dotprod = udot2 (4, u_input_nil);
> > + uint32_t u_min_dotprod = udot2 (4, u_input_min);
> > + uint32_t u_max_dotprod = udot2 (4, u_input_max);
> > +
> > + if (u_nil_dotprod != 0
> > + || u_min_dotprod != 4
> > + || u_max_dotprod != 4294705156)
> > + abort ();
> > +
> > + int16_t s_input_nil[] = { [0 ... 3] = 0 };
> > + int16_t s_input_min[] = { [0 ... 3] = -23170 };
> > + int16_t s_input_max[] = { [0 ... 3] = 23170 };
> > +
> > + int32_t s_nil_dotprod = sdot2 (4, s_input_nil);
> > + int32_t s_min_dotprod = sdot2 (4, s_input_min);
> > + int32_t s_max_dotprod = sdot2 (4, s_input_max);
> > +
> > + if (s_nil_dotprod != 0
> > + || s_min_dotprod != 2147395600
> > + || s_max_dotprod != 2147395600)
> > + abort ();
> > +
> > + return 0;
> > +}
> > +
> > +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 2 "vect" } } */
> > +/* { dg-final { scan-tree-dump-times "vect_recog_dot_prod_pattern:
> > detected"
> 46 "vect" } } */
> > +/* { dg-final { scan-assembler "\[ \t\]udot\tz\[0-9\]+.s, z\[0-9\]+.h,
> > z\[0-
> 9\]+.h" } } */
> > +/* { dg-final { scan-assembler "\[ \t\]sdot\tz\[0-9\]+.s, z\[0-9\]+.h,
> > z\[0-9\]+.h"
> } } */
> > diff --git a/gcc/testsuite/lib/target-supports.exp
> > b/gcc/testsuite/lib/target-
> supports.exp
> > index 11ba77ca404..41618d399a3 100644
> > --- a/gcc/testsuite/lib/target-supports.exp
> > +++ b/gcc/testsuite/lib/target-supports.exp
> > @@ -4258,6 +4258,14 @@ proc check_effective_target_vect_int { } {
> > }}]
> > }
> >
> > +# Return 1 if the target supports two-way dot products, or 0 otherwise.
> > +
> > +proc check_effective_target_vect_dotprod_twoway { } {
> > + return [check_cached_effective_target_indexed aarch64_sme2 {
> > + expr { [check_effective_target_aarch64_sme2]
> > + }}]
> > +}
> > +
> > # Return 1 if the target supports vectorization of early breaks,
> > # 0 otherwise.
> > #
> > --
> > 2.34.1
> >
