[PATCH] Enable GCC support for AMX
Hongyu Wang
wwwhhhyyy333@gmail.com
Fri Sep 11 17:00:36 GMT 2020
Hi
Thanks for your review, and sorry for the late reply. It took a while
to finish the runtime test.
> > diff --git a/gcc/config.gcc b/gcc/config.gcc
> > index 797f0ad5edd..d0e59e86a5c 100644
> > --- a/gcc/config.gcc
> > +++ b/gcc/config.gcc
> > @@ -412,7 +412,7 @@ i[34567]86-*-*)
> > waitpkgintrin.h cldemoteintrin.h avx512bf16vlintrin.h
> > avx512bf16intrin.h enqcmdintrin.h serializeintrin.h
> > avx512vp2intersectintrin.h avx512vp2intersectvlintrin.h
> > - tsxldtrkintrin.h"
> > + tsxldtrkintrin.h amxtileintrin.h amxint8intrin.h amxbf16intrin.h"
>
> Line more than 80 chars.
>
> > ;;
> > x86_64-*-*)
> > cpu_type=i386
> > @@ -447,7 +447,7 @@ x86_64-*-*)
> > waitpkgintrin.h cldemoteintrin.h avx512bf16vlintrin.h
> > avx512bf16intrin.h enqcmdintrin.h serializeintrin.h
> > avx512vp2intersectintrin.h avx512vp2intersectvlintrin.h
> > - tsxldtrkintrin.h"
> > + tsxldtrkintrin.h amxtileintrin.h amxint8intrin.h amxbf16intrin.h"
>
> Ditto.
Changed.
>
> > diff --git a/gcc/config/i386/amxbf16intrin.h b/gcc/config/i386/amxbf16intrin.h
> > new file mode 100644
> > index 00000000000..df0e2262d50
> > --- /dev/null
> > +++ b/gcc/config/i386/amxbf16intrin.h
> > @@ -0,0 +1,25 @@
> > +#if !defined _IMMINTRIN_H_INCLUDED
> > +#error "Never use <amxbf16intrin.h> directly; include <immintrin.h> instead."
> > +#endif
> > +
> > +#ifndef _AMXBF16INTRIN_H_INCLUDED
> > +#define _AMXBF16INTRIN_H_INCLUDED
> > +
> > +#if !defined(__AMX_BF16__)
> > +#pragma GCC push_options
> > +#pragma GCC target("amx-bf16")
> > +#define __DISABLE_AMX_BF16__
> > +#endif /* __AMX_BF16__ */
> > +
> > +#if defined(__x86_64__) && defined(__AMX_BF16__)
> > +#define _tile_dpbf16ps(dst,src1,src2) \
> > + __asm__ volatile\
> > + ("{tdpbf16ps\t%%tmm"#src2", %%tmm"#src1", %%tmm"#dst"|tdpbf16ps\t%%tmm"#dst", %%tmm"#src1", %%tmm"#src2"}" ::)
> > +#endif
>
> I hope in future we'll replace it with unspecs at least...
Currently we think it is redundant to add builtins with just const int
parameters, which are supposed to be replaced in the future.
>
> > diff --git a/gcc/config/i386/i386.opt b/gcc/config/i386/i386.opt
> > index c9f7195d423..9389dc24948 100644
> > diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
> > index bca8c856dc8..a46e31f5862 100644
> > --- a/gcc/doc/invoke.texi
> > +++ b/gcc/doc/invoke.texi
> > @@ -1357,6 +1357,7 @@ See RS/6000 and PowerPC Options.
> > -mvpclmulqdq -mavx512bitalg -mmovdiri -mmovdir64b -mavx512vpopcntdq @gol
> > -mavx5124fmaps -mavx512vnni -mavx5124vnniw -mprfchw -mrdpid @gol
> > -mrdseed -msgx -mavx512vp2intersect -mserialize -mtsxldtrk@gol
> > +-mamx-tile -mamx-int8 -mamx-bf16@gol
>
> Add space please.
Changed.
>
> > diff --git a/gcc/testsuite/gcc.target/i386/amxbf16-asmintel-2.c b/gcc/testsuite/gcc.target/i386/amxbf16-asmintel-2.c
> > new file mode 100644
> > index 00000000000..605a44df3f8
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/amxbf16-asmintel-2.c
> > @@ -0,0 +1,4 @@
> > +/* { dg-do assemble { target { ! ia32 } } } */
> > +/* { dg-options "-O2 -mamx-bf16 -masm=intel" } */
> > +/* { dg-require-effective-target amx_bf16 } */
> > +#include"amxbf16-asmintel-1.c"
>
> I didn't get it. We ususally use second tescase to actually execute
> it and (well, a little) verify that semantics is ok. E.g. that
> operands order is correct. Could you please do that?
> This applies to all *-2.c cases.
> I've checked and looks like public SDE simulator supports AMX.
>
Added runtime test. Tested and passed under SDE.
Also, we adjust the intrinsic call to accept macro parameters.
Updated patch.
> --
> K
> Hello,
>
> On 03 сен 08:17, H.J. Lu wrote:
> > On Thu, Sep 3, 2020 at 8:08 AM Kirill Yukhin via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Hello,
> > >
> > > On 06 июл 09:58, Hongyu Wang via Gcc-patches wrote:
> > > > Hi:
> > > >
> > > > This patch is about to support Intel Advanced Matrix Extensions (AMX)
> > > > which will be enabled in GLC.
> > > >
> > > > AMX is a new 64-bit programming paradigm consisting of two
> > > > compo nents: a set of 2-dimensional registers (tiles) representing
> > > > sub-arrays from a larger 2-dimensional memory image,
> > > > and an accelerator able to operate on tiles
> > > >
> > > > Supported instructions are
> > > >
> > > > AMX-TILE:ldtilecfg/sttilecfg/tileloadd/tileloaddt1/tilezero/tilerelease
> > > > AMX-INT8:tdpbssd/tdpbsud/tdpbusd/tdpbuud
> > > > AMX-BF16:tdpbf16ps
> > > >
> > > > The intrinsics adopts constant tile register number as its input parameters.
> > > >
> > > > For detailed information, please refer to
> > > > https://software.intel.com/content/dam/develop/public/us/en/documents/architecture-instruction-set-extensions-programming-reference.pdf
> > > >
> > > > Bootstrap ok, regression test on i386/x86 backend is ok.
> > > >
> > > > OK for master?
> > >
> > > I was trying to apply your patch to recent master and got
> > > compilation error:
> > >
> > > g++ -std=gnu++11 -fno-PIE -c -g -O2 -DIN_GCC -fno-exceptions -fno-rtti -fasynchronous-unwind-tables -W -Wall -Wno-narrowi
> > > ng -Wwrite-strings -Wcast-qual -Wmissing-format-attribute -Woverloaded-virtual -pedantic -Wno-long-long -Wno-variadic-macros -Wn
> > > o-overlength-strings -fno-common -DHAVE_CONFIG_H -I. -I. -I/export/kyukhin/gcc/src/gcc -I/export/kyukhin/gcc/src/gcc/. -I/expor
> > > t/kyukhin/gcc/src/gcc/../include -I/export/kyukhin/gcc/src/gcc/../libcpp/include -I/export/kyukhin/gcc/src/gcc/../libdecnumber
> > > -I/export/kyukhin/gcc/src/gcc/../libdecnumber/bid -I../libdecnumber -I/export/kyukhin/gcc/src/gcc/../libbacktrace -o i386-opti
> > > ons.o -MT i386-options.o -MMD -MP -MF ./.deps/i386-options.TPo /export/kyukhin/gcc/src/gcc/config/i386/i386-options.c
> > > /export/kyukhin/gcc/src/gcc/config/i386/i386-options.c: In function ‘bool ix86_option_override_internal(bool, gcc_options*, gcc_
> > > options*)’:
> > > /export/kyukhin/gcc/src/gcc/config/i386/i386-options.c:2263:41: error: ‘PTA_AMX_TILE’ was not declared in this scope
> > > if (((processor_alias_table[i].flags & PTA_AMX_TILE) != 0)
> > > ^
> > > /export/kyukhin/gcc/src/gcc/config/i386/i386-options.c:2267:41: error: ‘PTA_AMX_INT8’ was not declared in this scope
> > > if (((processor_alias_table[i].flags & PTA_AMX_INT8) != 0)
> > > ^
> > > /export/kyukhin/gcc/src/gcc/config/i386/i386-options.c:2271:41: error: ‘PTA_AMX_BF16’ was not declared in this scope
> > > if (((processor_alias_table[i].flags & PTA_AMX_BF16) != 0)
> > >
> > > Could you please fix that?
> >
> > Here is the rebased patch against
> >
> > commit 3c219134152f645103f2fcd50735b177ccd76cde
> > Author: Jonathan Wakely <jwakely@redhat.com>
> > Date: Thu Sep 3 12:38:50 2020 +0100
> >
> > libstdc++: Optimise GCD algorithms
> >
> > Thanks.
> >
> > --
> > H.J.
>
> > diff --git a/gcc/config.gcc b/gcc/config.gcc
> > index 797f0ad5edd..d0e59e86a5c 100644
> > --- a/gcc/config.gcc
> > +++ b/gcc/config.gcc
> > @@ -412,7 +412,7 @@ i[34567]86-*-*)
> > waitpkgintrin.h cldemoteintrin.h avx512bf16vlintrin.h
> > avx512bf16intrin.h enqcmdintrin.h serializeintrin.h
> > avx512vp2intersectintrin.h avx512vp2intersectvlintrin.h
> > - tsxldtrkintrin.h"
> > + tsxldtrkintrin.h amxtileintrin.h amxint8intrin.h amxbf16intrin.h"
>
> Line more than 80 chars.
>
> > ;;
> > x86_64-*-*)
> > cpu_type=i386
> > @@ -447,7 +447,7 @@ x86_64-*-*)
> > waitpkgintrin.h cldemoteintrin.h avx512bf16vlintrin.h
> > avx512bf16intrin.h enqcmdintrin.h serializeintrin.h
> > avx512vp2intersectintrin.h avx512vp2intersectvlintrin.h
> > - tsxldtrkintrin.h"
> > + tsxldtrkintrin.h amxtileintrin.h amxint8intrin.h amxbf16intrin.h"
>
> Ditto.
>
> > diff --git a/gcc/config/i386/amxbf16intrin.h b/gcc/config/i386/amxbf16intrin.h
> > new file mode 100644
> > index 00000000000..df0e2262d50
> > --- /dev/null
> > +++ b/gcc/config/i386/amxbf16intrin.h
> > @@ -0,0 +1,25 @@
> > +#if !defined _IMMINTRIN_H_INCLUDED
> > +#error "Never use <amxbf16intrin.h> directly; include <immintrin.h> instead."
> > +#endif
> > +
> > +#ifndef _AMXBF16INTRIN_H_INCLUDED
> > +#define _AMXBF16INTRIN_H_INCLUDED
> > +
> > +#if !defined(__AMX_BF16__)
> > +#pragma GCC push_options
> > +#pragma GCC target("amx-bf16")
> > +#define __DISABLE_AMX_BF16__
> > +#endif /* __AMX_BF16__ */
> > +
> > +#if defined(__x86_64__) && defined(__AMX_BF16__)
> > +#define _tile_dpbf16ps(dst,src1,src2) \
> > + __asm__ volatile\
> > + ("{tdpbf16ps\t%%tmm"#src2", %%tmm"#src1", %%tmm"#dst"|tdpbf16ps\t%%tmm"#dst", %%tmm"#src1", %%tmm"#src2"}" ::)
> > +#endif
>
> I hope in future we'll replace it with unspecs at least...
>
> > diff --git a/gcc/config/i386/i386.opt b/gcc/config/i386/i386.opt
> > index c9f7195d423..9389dc24948 100644
> > diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
> > index bca8c856dc8..a46e31f5862 100644
> > --- a/gcc/doc/invoke.texi
> > +++ b/gcc/doc/invoke.texi
> > @@ -1357,6 +1357,7 @@ See RS/6000 and PowerPC Options.
> > -mvpclmulqdq -mavx512bitalg -mmovdiri -mmovdir64b -mavx512vpopcntdq @gol
> > -mavx5124fmaps -mavx512vnni -mavx5124vnniw -mprfchw -mrdpid @gol
> > -mrdseed -msgx -mavx512vp2intersect -mserialize -mtsxldtrk@gol
> > +-mamx-tile -mamx-int8 -mamx-bf16@gol
>
> Add space please.
>
> > diff --git a/gcc/testsuite/gcc.target/i386/amxbf16-asmintel-2.c b/gcc/testsuite/gcc.target/i386/amxbf16-asmintel-2.c
> > new file mode 100644
> > index 00000000000..605a44df3f8
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/amxbf16-asmintel-2.c
> > @@ -0,0 +1,4 @@
> > +/* { dg-do assemble { target { ! ia32 } } } */
> > +/* { dg-options "-O2 -mamx-bf16 -masm=intel" } */
> > +/* { dg-require-effective-target amx_bf16 } */
> > +#include"amxbf16-asmintel-1.c"
>
> I didn't get it. We ususally use second tescase to actually execute
> it and (well, a little) verify that semantics is ok. E.g. that
> operands order is correct. Could you please do that?
> This applies to all *-2.c cases.
> I've checked and looks like public SDE simulator supports AMX.
>
> --
> K
-------------- next part --------------
A non-text attachment was scrubbed...
Name: GCC_AMX_support_v2.patch
Type: application/octet-stream
Size: 64726 bytes
Desc: not available
URL: <https://gcc.gnu.org/pipermail/gcc-patches/attachments/20200912/bdaf53b0/attachment-0001.obj>
More information about the Gcc-patches
mailing list