[01/10] optabs: Make all `*dot_prod_optab's modeled as conversions

Message ID 20240710140602.1707875-2-victor.donascimento@arm.com
State New
Headers
Series Make `dot_prod' a convert-type optab |

Commit Message

Victor Do Nascimento July 10, 2024, 2:05 p.m. UTC
  Given the specification in the GCC internals manual defines the
{u|s}dot_prod<m> standard name as taking "two signed elements of the
same mode, adding them to a third operand of wider mode", there is
currently ambiguity in the relationship between the mode of the first
two arguments and that of the third.

This vagueness means that, in theory, different modes may be
supportable in the third argument.  This flexibility would allow for a
given backend to add to the accumulator a different number of
vectorized products, e.g. A backend may provide instructions for both:

  accum += a[0] * b[0] + a[1] * b[1] + a[2] * b[2] + a[3] * b[3]

and

  accum += a[0] * b[0] + a[1] * b[1],

as is now seen in the SVE2.1 extension to AArch64.  In spite of the
aforementioned flexibility, modeling the dot-product operation as a
direct optab means that we have no way to encode both input and the
accumulator data modes into the backend pattern name, which prevents
us from harnessing this flexibility.

We therefore make all dot_prod optabs conversions, allowing, for
example, for the encoding of both 2-way and 4-way dot product backend
patterns.

gcc/ChangeLog:

	* optabs.def (sdot_prod_optab): Convert from OPTAB_D to
	OPTAB_CD.
	(udot_prod_optab): Likewise.
	(usdot_prod_optab): Likewise.
	* doc/md.texi (Standard Names): update entries for u,s and us
	dot_prod names.
---
 gcc/doc/md.texi | 18 +++++++++---------
 gcc/optabs.def  |  6 +++---
 2 files changed, 12 insertions(+), 12 deletions(-)
  

Comments

Richard Sandiford July 11, 2024, 2:16 p.m. UTC | #1
Victor Do Nascimento <victor.donascimento@arm.com> writes:
> Given the specification in the GCC internals manual defines the
> {u|s}dot_prod<m> standard name as taking "two signed elements of the
> same mode, adding them to a third operand of wider mode", there is
> currently ambiguity in the relationship between the mode of the first
> two arguments and that of the third.
>
> This vagueness means that, in theory, different modes may be
> supportable in the third argument.  This flexibility would allow for a
> given backend to add to the accumulator a different number of
> vectorized products, e.g. A backend may provide instructions for both:
>
>   accum += a[0] * b[0] + a[1] * b[1] + a[2] * b[2] + a[3] * b[3]
>
> and
>
>   accum += a[0] * b[0] + a[1] * b[1],
>
> as is now seen in the SVE2.1 extension to AArch64.  In spite of the
> aforementioned flexibility, modeling the dot-product operation as a
> direct optab means that we have no way to encode both input and the
> accumulator data modes into the backend pattern name, which prevents
> us from harnessing this flexibility.
>
> We therefore make all dot_prod optabs conversions, allowing, for
> example, for the encoding of both 2-way and 4-way dot product backend
> patterns.
>
> gcc/ChangeLog:
>
> 	* optabs.def (sdot_prod_optab): Convert from OPTAB_D to
> 	OPTAB_CD.
> 	(udot_prod_optab): Likewise.
> 	(usdot_prod_optab): Likewise.
> 	* doc/md.texi (Standard Names): update entries for u,s and us
> 	dot_prod names.
> ---
>  gcc/doc/md.texi | 18 +++++++++---------
>  gcc/optabs.def  |  6 +++---
>  2 files changed, 12 insertions(+), 12 deletions(-)
>
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index 7f4335e0aac..2a74e473f05 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -5748,15 +5748,15 @@ for (i = 0; i < LEN + BIAS; i++)
>      operand0 += operand2[i];
>  @end smallexample
>  
> -@cindex @code{sdot_prod@var{m}} instruction pattern
> -@item @samp{sdot_prod@var{m}}
> +@cindex @code{sdot_prod@var{m}@var{n}} instruction pattern
> +@item @samp{sdot_prod@var{m}@var{n}}
>  
>  Compute the sum of the products of two signed elements.
>  Operand 1 and operand 2 are of the same mode. Their
>  product, which is of a wider mode, is computed and added to operand 3.
>  Operand 3 is of a mode equal or wider than the mode of the product. The
>  result is placed in operand 0, which is of the same mode as operand 3.
> -@var{m} is the mode of operand 1 and operand 2.
> +@var{m} is the mode of operands 0 and 3 and @var{n} the mode of operands 1 and 2.

Now that we can put names to both modes, how about replacing the
description with something like this:

Multiply operand 1 by operand 2 without loss of precision, given that
both operands contain signed elements.  Add each product to the overlapping
element of operand 3 and store the result in operand 0.  Operands 0 and 3
have mode @var{m} and operands 1 and 2 have mode @var{n}, with @var{n}
having narrower elements than @var{m}.

This is all personal taste though, so it's just a suggestion.

Same idea for the others.

OK with that change from my POV, but happy to hear other suggestions.

Thanks,
Richard

>  Semantically the expressions perform the multiplication in the following signs
>  
> @@ -5766,15 +5766,15 @@ sdot<signed op0, signed op1, signed op2, signed op3> ==
>  @dots{}
>  @end smallexample
>  
> -@cindex @code{udot_prod@var{m}} instruction pattern
> -@item @samp{udot_prod@var{m}}
> +@cindex @code{udot_prod@var{m}@var{n}} instruction pattern
> +@item @samp{udot_prod@var{m}@var{n}}
>  
>  Compute the sum of the products of two unsigned elements.
>  Operand 1 and operand 2 are of the same mode. Their
>  product, which is of a wider mode, is computed and added to operand 3.
>  Operand 3 is of a mode equal or wider than the mode of the product. The
>  result is placed in operand 0, which is of the same mode as operand 3.
> -@var{m} is the mode of operand 1 and operand 2.
> +@var{m} is the mode of operands 0 and 3 and @var{n} the mode of operands 1 and 2.
>  
>  Semantically the expressions perform the multiplication in the following signs
>  
> @@ -5784,14 +5784,14 @@ udot<unsigned op0, unsigned op1, unsigned op2, unsigned op3> ==
>  @dots{}
>  @end smallexample
>  
> -@cindex @code{usdot_prod@var{m}} instruction pattern
> -@item @samp{usdot_prod@var{m}}
> +@cindex @code{usdot_prod@var{m}@var{n}} instruction pattern
> +@item @samp{usdot_prod@var{m}@var{n}}
>  Compute the sum of the products of elements of different signs.
>  Operand 1 must be unsigned and operand 2 signed. Their
>  product, which is of a wider mode, is computed and added to operand 3.
>  Operand 3 is of a mode equal or wider than the mode of the product. The
>  result is placed in operand 0, which is of the same mode as operand 3.
> -@var{m} is the mode of operand 1 and operand 2.
> +@var{m} is the mode of operands 0 and 3 and @var{n} the mode of operands 1 and 2.
>  
>  Semantically the expressions perform the multiplication in the following signs
>  
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 45e117a7f50..fce4b2d5b08 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -106,6 +106,9 @@ OPTAB_CD(mask_scatter_store_optab, "mask_scatter_store$a$b")
>  OPTAB_CD(mask_len_scatter_store_optab, "mask_len_scatter_store$a$b")
>  OPTAB_CD(vec_extract_optab, "vec_extract$a$b")
>  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$b")
>  
>  OPTAB_CD (while_ult_optab, "while_ult$a$b")
>  
> @@ -409,10 +412,7 @@ OPTAB_D (savg_floor_optab, "avg$a3_floor")
>  OPTAB_D (uavg_floor_optab, "uavg$a3_floor")
>  OPTAB_D (savg_ceil_optab, "avg$a3_ceil")
>  OPTAB_D (uavg_ceil_optab, "uavg$a3_ceil")
> -OPTAB_D (sdot_prod_optab, "sdot_prod$I$a")
>  OPTAB_D (ssum_widen_optab, "widen_ssum$I$a3")
> -OPTAB_D (udot_prod_optab, "udot_prod$I$a")
> -OPTAB_D (usdot_prod_optab, "usdot_prod$I$a")
>  OPTAB_D (usum_widen_optab, "widen_usum$I$a3")
>  OPTAB_D (usad_optab, "usad$I$a")
>  OPTAB_D (ssad_optab, "ssad$I$a")
  

Patch

diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
index 7f4335e0aac..2a74e473f05 100644
--- a/gcc/doc/md.texi
+++ b/gcc/doc/md.texi
@@ -5748,15 +5748,15 @@  for (i = 0; i < LEN + BIAS; i++)
     operand0 += operand2[i];
 @end smallexample
 
-@cindex @code{sdot_prod@var{m}} instruction pattern
-@item @samp{sdot_prod@var{m}}
+@cindex @code{sdot_prod@var{m}@var{n}} instruction pattern
+@item @samp{sdot_prod@var{m}@var{n}}
 
 Compute the sum of the products of two signed elements.
 Operand 1 and operand 2 are of the same mode. Their
 product, which is of a wider mode, is computed and added to operand 3.
 Operand 3 is of a mode equal or wider than the mode of the product. The
 result is placed in operand 0, which is of the same mode as operand 3.
-@var{m} is the mode of operand 1 and operand 2.
+@var{m} is the mode of operands 0 and 3 and @var{n} the mode of operands 1 and 2.
 
 Semantically the expressions perform the multiplication in the following signs
 
@@ -5766,15 +5766,15 @@  sdot<signed op0, signed op1, signed op2, signed op3> ==
 @dots{}
 @end smallexample
 
-@cindex @code{udot_prod@var{m}} instruction pattern
-@item @samp{udot_prod@var{m}}
+@cindex @code{udot_prod@var{m}@var{n}} instruction pattern
+@item @samp{udot_prod@var{m}@var{n}}
 
 Compute the sum of the products of two unsigned elements.
 Operand 1 and operand 2 are of the same mode. Their
 product, which is of a wider mode, is computed and added to operand 3.
 Operand 3 is of a mode equal or wider than the mode of the product. The
 result is placed in operand 0, which is of the same mode as operand 3.
-@var{m} is the mode of operand 1 and operand 2.
+@var{m} is the mode of operands 0 and 3 and @var{n} the mode of operands 1 and 2.
 
 Semantically the expressions perform the multiplication in the following signs
 
@@ -5784,14 +5784,14 @@  udot<unsigned op0, unsigned op1, unsigned op2, unsigned op3> ==
 @dots{}
 @end smallexample
 
-@cindex @code{usdot_prod@var{m}} instruction pattern
-@item @samp{usdot_prod@var{m}}
+@cindex @code{usdot_prod@var{m}@var{n}} instruction pattern
+@item @samp{usdot_prod@var{m}@var{n}}
 Compute the sum of the products of elements of different signs.
 Operand 1 must be unsigned and operand 2 signed. Their
 product, which is of a wider mode, is computed and added to operand 3.
 Operand 3 is of a mode equal or wider than the mode of the product. The
 result is placed in operand 0, which is of the same mode as operand 3.
-@var{m} is the mode of operand 1 and operand 2.
+@var{m} is the mode of operands 0 and 3 and @var{n} the mode of operands 1 and 2.
 
 Semantically the expressions perform the multiplication in the following signs
 
diff --git a/gcc/optabs.def b/gcc/optabs.def
index 45e117a7f50..fce4b2d5b08 100644
--- a/gcc/optabs.def
+++ b/gcc/optabs.def
@@ -106,6 +106,9 @@  OPTAB_CD(mask_scatter_store_optab, "mask_scatter_store$a$b")
 OPTAB_CD(mask_len_scatter_store_optab, "mask_len_scatter_store$a$b")
 OPTAB_CD(vec_extract_optab, "vec_extract$a$b")
 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 (while_ult_optab, "while_ult$a$b")
 
@@ -409,10 +412,7 @@  OPTAB_D (savg_floor_optab, "avg$a3_floor")
 OPTAB_D (uavg_floor_optab, "uavg$a3_floor")
 OPTAB_D (savg_ceil_optab, "avg$a3_ceil")
 OPTAB_D (uavg_ceil_optab, "uavg$a3_ceil")
-OPTAB_D (sdot_prod_optab, "sdot_prod$I$a")
 OPTAB_D (ssum_widen_optab, "widen_ssum$I$a3")
-OPTAB_D (udot_prod_optab, "udot_prod$I$a")
-OPTAB_D (usdot_prod_optab, "usdot_prod$I$a")
 OPTAB_D (usum_widen_optab, "widen_usum$I$a3")
 OPTAB_D (usad_optab, "usad$I$a")
 OPTAB_D (ssad_optab, "ssad$I$a")