diff mbox series

rs6000: Add optimizations for _mm_sad_epu8

Message ID 20211022172849.499625-1-pc@us.ibm.com
State New
Headers show
Series rs6000: Add optimizations for _mm_sad_epu8 | expand

Commit Message

Paul A. Clarke Oct. 22, 2021, 5:28 p.m. UTC 
Power9 ISA added `vabsdub` instruction which is realized in the
`vec_absd` instrinsic.

Use `vec_absd` for `_mm_sad_epu8` compatibility intrinsic, when
`_ARCH_PWR9`.

Also, the realization of `vec_sum2s` on little-endian includes
two shifts in order to position the input and output to match
the semantics of `vec_sum2s`:
- Shift the second input vector left 12 bytes. In the current usage,
  that vector is `{0}`, so this shift is unnecessary, but is currently
  not eliminated under optimization.
- Shift the vector produced by the `vsum2sws` instruction left 4 bytes.
  The two words within each doubleword of this (shifted) result must then
  be explicitly swapped to match the semantics of `_mm_sad_epu8`,
  effectively reversing this shift.  So, this shift (and a susequent swap)
  are unnecessary, but not currently removed under optimization.

Using `__builtin_altivec_vsum2sws` retains both shifts, so is not an
option for removing the shifts.

For little-endian, use the `vsum2sws` instruction directly, and
eliminate the explicit shift (swap).

2021-10-22  Paul A. Clarke  <pc@us.ibm.com>

gcc
	* config/rs6000/emmintrin.h (_mm_sad_epu8): Use vec_absd
	when _ARCH_PWR9, optimize vec_sum2s when LE.
---
Tested on powerpc64le-linux on Power9, with and without `-mcpu=power9`,
and on powerpc/powerpc64-linux on Power8.

OK for trunk?

 gcc/config/rs6000/emmintrin.h | 24 +++++++++++++++++-------
 1 file changed, 17 insertions(+), 7 deletions(-)

Comments

Paul A. Clarke Nov. 8, 2021, 5:43 p.m. UTC | #1 
Gentle ping...

On Fri, Oct 22, 2021 at 12:28:49PM -0500, Paul A. Clarke via Gcc-patches wrote:
> Power9 ISA added `vabsdub` instruction which is realized in the
> `vec_absd` instrinsic.
> 
> Use `vec_absd` for `_mm_sad_epu8` compatibility intrinsic, when
> `_ARCH_PWR9`.
> 
> Also, the realization of `vec_sum2s` on little-endian includes
> two shifts in order to position the input and output to match
> the semantics of `vec_sum2s`:
> - Shift the second input vector left 12 bytes. In the current usage,
>   that vector is `{0}`, so this shift is unnecessary, but is currently
>   not eliminated under optimization.
> - Shift the vector produced by the `vsum2sws` instruction left 4 bytes.
>   The two words within each doubleword of this (shifted) result must then
>   be explicitly swapped to match the semantics of `_mm_sad_epu8`,
>   effectively reversing this shift.  So, this shift (and a susequent swap)
>   are unnecessary, but not currently removed under optimization.
> 
> Using `__builtin_altivec_vsum2sws` retains both shifts, so is not an
> option for removing the shifts.
> 
> For little-endian, use the `vsum2sws` instruction directly, and
> eliminate the explicit shift (swap).
> 
> 2021-10-22  Paul A. Clarke  <pc@us.ibm.com>
> 
> gcc
> 	* config/rs6000/emmintrin.h (_mm_sad_epu8): Use vec_absd
> 	when _ARCH_PWR9, optimize vec_sum2s when LE.
> ---
> Tested on powerpc64le-linux on Power9, with and without `-mcpu=power9`,
> and on powerpc/powerpc64-linux on Power8.
> 
> OK for trunk?
> 
>  gcc/config/rs6000/emmintrin.h | 24 +++++++++++++++++-------
>  1 file changed, 17 insertions(+), 7 deletions(-)
> 
> diff --git a/gcc/config/rs6000/emmintrin.h b/gcc/config/rs6000/emmintrin.h
> index ab16c13c379e..c4758be0e777 100644
> --- a/gcc/config/rs6000/emmintrin.h
> +++ b/gcc/config/rs6000/emmintrin.h
> @@ -2197,27 +2197,37 @@ extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __arti
>  _mm_sad_epu8 (__m128i __A, __m128i __B)
>  {
>    __v16qu a, b;
> -  __v16qu vmin, vmax, vabsdiff;
> +  __v16qu vabsdiff;
>    __v4si vsum;
>    const __v4su zero = { 0, 0, 0, 0 };
>    __v4si result;
>  
>    a = (__v16qu) __A;
>    b = (__v16qu) __B;
> -  vmin = vec_min (a, b);
> -  vmax = vec_max (a, b);
> +#ifndef _ARCH_PWR9
> +  __v16qu vmin = vec_min (a, b);
> +  __v16qu vmax = vec_max (a, b);
>    vabsdiff = vec_sub (vmax, vmin);
> +#else
> +  vabsdiff = vec_absd (a, b);
> +#endif
>    /* Sum four groups of bytes into integers.  */
>    vsum = (__vector signed int) vec_sum4s (vabsdiff, zero);
> +#ifdef __LITTLE_ENDIAN__
> +  /* Sum across four integers with two integer results.  */
> +  asm ("vsum2sws %0,%1,%2" : "=v" (result) : "v" (vsum), "v" (zero));
> +  /* Note: vec_sum2s could be used here, but on little-endian, vector
> +     shifts are added that are not needed for this use-case.
> +     A vector shift to correctly position the 32-bit integer results
> +     (currently at [0] and [2]) to [1] and [3] would then need to be
> +     swapped back again since the desired results are two 64-bit
> +     integers ([1]|[0] and [3]|[2]).  Thus, no shift is performed.  */
> +#else
>    /* Sum across four integers with two integer results.  */
>    result = vec_sum2s (vsum, (__vector signed int) zero);
>    /* Rotate the sums into the correct position.  */
> -#ifdef __LITTLE_ENDIAN__
> -  result = vec_sld (result, result, 4);
> -#else
>    result = vec_sld (result, result, 6);
>  #endif
> -  /* Rotate the sums into the correct position.  */
>    return (__m128i) result;
>  }
>  
> -- 
> 2.27.0
>
Paul A. Clarke Nov. 19, 2021, 2:26 a.m. UTC | #2 
On Mon, Nov 08, 2021 at 11:43:26AM -0600, Paul A. Clarke via Gcc-patches wrote:
> Gentle ping...

Gentle re-ping.

> On Fri, Oct 22, 2021 at 12:28:49PM -0500, Paul A. Clarke via Gcc-patches wrote:
> > Power9 ISA added `vabsdub` instruction which is realized in the
> > `vec_absd` instrinsic.
> > 
> > Use `vec_absd` for `_mm_sad_epu8` compatibility intrinsic, when
> > `_ARCH_PWR9`.
> > 
> > Also, the realization of `vec_sum2s` on little-endian includes
> > two shifts in order to position the input and output to match
> > the semantics of `vec_sum2s`:
> > - Shift the second input vector left 12 bytes. In the current usage,
> >   that vector is `{0}`, so this shift is unnecessary, but is currently
> >   not eliminated under optimization.
> > - Shift the vector produced by the `vsum2sws` instruction left 4 bytes.
> >   The two words within each doubleword of this (shifted) result must then
> >   be explicitly swapped to match the semantics of `_mm_sad_epu8`,
> >   effectively reversing this shift.  So, this shift (and a susequent swap)
> >   are unnecessary, but not currently removed under optimization.
> > 
> > Using `__builtin_altivec_vsum2sws` retains both shifts, so is not an
> > option for removing the shifts.
> > 
> > For little-endian, use the `vsum2sws` instruction directly, and
> > eliminate the explicit shift (swap).
> > 
> > 2021-10-22  Paul A. Clarke  <pc@us.ibm.com>
> > 
> > gcc
> > 	* config/rs6000/emmintrin.h (_mm_sad_epu8): Use vec_absd
> > 	when _ARCH_PWR9, optimize vec_sum2s when LE.
> > ---
> > Tested on powerpc64le-linux on Power9, with and without `-mcpu=power9`,
> > and on powerpc/powerpc64-linux on Power8.
> > 
> > OK for trunk?
> > 
> >  gcc/config/rs6000/emmintrin.h | 24 +++++++++++++++++-------
> >  1 file changed, 17 insertions(+), 7 deletions(-)
> > 
> > diff --git a/gcc/config/rs6000/emmintrin.h b/gcc/config/rs6000/emmintrin.h
> > index ab16c13c379e..c4758be0e777 100644
> > --- a/gcc/config/rs6000/emmintrin.h
> > +++ b/gcc/config/rs6000/emmintrin.h
> > @@ -2197,27 +2197,37 @@ extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __arti
> >  _mm_sad_epu8 (__m128i __A, __m128i __B)
> >  {
> >    __v16qu a, b;
> > -  __v16qu vmin, vmax, vabsdiff;
> > +  __v16qu vabsdiff;
> >    __v4si vsum;
> >    const __v4su zero = { 0, 0, 0, 0 };
> >    __v4si result;
> >  
> >    a = (__v16qu) __A;
> >    b = (__v16qu) __B;
> > -  vmin = vec_min (a, b);
> > -  vmax = vec_max (a, b);
> > +#ifndef _ARCH_PWR9
> > +  __v16qu vmin = vec_min (a, b);
> > +  __v16qu vmax = vec_max (a, b);
> >    vabsdiff = vec_sub (vmax, vmin);
> > +#else
> > +  vabsdiff = vec_absd (a, b);
> > +#endif
> >    /* Sum four groups of bytes into integers.  */
> >    vsum = (__vector signed int) vec_sum4s (vabsdiff, zero);
> > +#ifdef __LITTLE_ENDIAN__
> > +  /* Sum across four integers with two integer results.  */
> > +  asm ("vsum2sws %0,%1,%2" : "=v" (result) : "v" (vsum), "v" (zero));
> > +  /* Note: vec_sum2s could be used here, but on little-endian, vector
> > +     shifts are added that are not needed for this use-case.
> > +     A vector shift to correctly position the 32-bit integer results
> > +     (currently at [0] and [2]) to [1] and [3] would then need to be
> > +     swapped back again since the desired results are two 64-bit
> > +     integers ([1]|[0] and [3]|[2]).  Thus, no shift is performed.  */
> > +#else
> >    /* Sum across four integers with two integer results.  */
> >    result = vec_sum2s (vsum, (__vector signed int) zero);
> >    /* Rotate the sums into the correct position.  */
> > -#ifdef __LITTLE_ENDIAN__
> > -  result = vec_sld (result, result, 4);
> > -#else
> >    result = vec_sld (result, result, 6);
> >  #endif
> > -  /* Rotate the sums into the correct position.  */
> >    return (__m128i) result;
> >  }
> >  
> > -- 
> > 2.27.0
> >
Segher Boessenkool Nov. 19, 2021, 6:09 p.m. UTC | #3 
Hi!

On Fri, Oct 22, 2021 at 12:28:49PM -0500, Paul A. Clarke wrote:
> Power9 ISA added `vabsdub` instruction which is realized in the
> `vec_absd` instrinsic.
> 
> Use `vec_absd` for `_mm_sad_epu8` compatibility intrinsic, when
> `_ARCH_PWR9`.
> 
> Also, the realization of `vec_sum2s` on little-endian includes
> two shifts in order to position the input and output to match
> the semantics of `vec_sum2s`:
> - Shift the second input vector left 12 bytes. In the current usage,
>   that vector is `{0}`, so this shift is unnecessary, but is currently
>   not eliminated under optimization.

The vsum2sws implementation uses an unspec, so there is almost no chance
of anything with it being optimised :-(

It rotates it right by 4 bytes btw, it's not a shift.

> - Shift the vector produced by the `vsum2sws` instruction left 4 bytes.
>   The two words within each doubleword of this (shifted) result must then
>   be explicitly swapped to match the semantics of `_mm_sad_epu8`,
>   effectively reversing this shift.  So, this shift (and a susequent swap)
>   are unnecessary, but not currently removed under optimization.

Rotate left by 4 -- same thing once you consider word 0 and 2 are set
to zeroes by the sum2sws.

Not sure why it is not optimised, what do the dump files say?  -dap and
I'd start looking at the combine dump.

> Using `__builtin_altivec_vsum2sws` retains both shifts, so is not an
> option for removing the shifts.
> 
> For little-endian, use the `vsum2sws` instruction directly, and
> eliminate the explicit shift (swap).
> 
> 2021-10-22  Paul A. Clarke  <pc@us.ibm.com>
> 
> gcc
> 	* config/rs6000/emmintrin.h (_mm_sad_epu8): Use vec_absd
> 	when _ARCH_PWR9, optimize vec_sum2s when LE.

Please don't break changelog lines early.

> -  vmin = vec_min (a, b);
> -  vmax = vec_max (a, b);
> +#ifndef _ARCH_PWR9
> +  __v16qu vmin = vec_min (a, b);
> +  __v16qu vmax = vec_max (a, b);
>    vabsdiff = vec_sub (vmax, vmin);
> +#else
> +  vabsdiff = vec_absd (a, b);
> +#endif

So hrm, maybe we should have the vec_absd macro (or the builtin) always,
just expanding to three insns if necessary.

Okay for trunk with approproate changelog and commit message changes.
Thanks!


Segher
diff mbox series

Patch

diff --git a/gcc/config/rs6000/emmintrin.h b/gcc/config/rs6000/emmintrin.h
index ab16c13c379e..c4758be0e777 100644
--- a/gcc/config/rs6000/emmintrin.h
+++ b/gcc/config/rs6000/emmintrin.h
@@ -2197,27 +2197,37 @@  extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __arti
 _mm_sad_epu8 (__m128i __A, __m128i __B)
 {
   __v16qu a, b;
-  __v16qu vmin, vmax, vabsdiff;
+  __v16qu vabsdiff;
   __v4si vsum;
   const __v4su zero = { 0, 0, 0, 0 };
   __v4si result;
 
   a = (__v16qu) __A;
   b = (__v16qu) __B;
-  vmin = vec_min (a, b);
-  vmax = vec_max (a, b);
+#ifndef _ARCH_PWR9
+  __v16qu vmin = vec_min (a, b);
+  __v16qu vmax = vec_max (a, b);
   vabsdiff = vec_sub (vmax, vmin);
+#else
+  vabsdiff = vec_absd (a, b);
+#endif
   /* Sum four groups of bytes into integers.  */
   vsum = (__vector signed int) vec_sum4s (vabsdiff, zero);
+#ifdef __LITTLE_ENDIAN__
+  /* Sum across four integers with two integer results.  */
+  asm ("vsum2sws %0,%1,%2" : "=v" (result) : "v" (vsum), "v" (zero));
+  /* Note: vec_sum2s could be used here, but on little-endian, vector
+     shifts are added that are not needed for this use-case.
+     A vector shift to correctly position the 32-bit integer results
+     (currently at [0] and [2]) to [1] and [3] would then need to be
+     swapped back again since the desired results are two 64-bit
+     integers ([1]|[0] and [3]|[2]).  Thus, no shift is performed.  */
+#else
   /* Sum across four integers with two integer results.  */
   result = vec_sum2s (vsum, (__vector signed int) zero);
   /* Rotate the sums into the correct position.  */
-#ifdef __LITTLE_ENDIAN__
-  result = vec_sld (result, result, 4);
-#else
   result = vec_sld (result, result, 6);
 #endif
-  /* Rotate the sums into the correct position.  */
   return (__m128i) result;
 }