Re: [PATCH] i386: using __bf16 for AVX512BF16 intrinsics
On Fri, Oct 28, 2022 at 2:20 PM Kong, Lingling via Gcc-patches
wrote:
>
> Hi,
>
> Previously we use unsigned short to represent bf16. It's not a good
> expression, and at the time the front end didn't support bf16 type.
> Now we introduced __bf16 to X86 psABI. So we can switch intrinsics to the new
> type.
>
> Ok for trunk ?
LGTM, but please don't commit it until next week to leave some time
for others to take a look.
Also please update GCC13 doc for it.
https://gcc.gnu.org/gcc-13/changes.html.
>
> Thanks,
> Lingling
>
> gcc/ChangeLog:
>
> * config/i386/avx512bf16intrin.h (__attribute__): Change short to
> bf16.
> (_mm_cvtsbh_ss): Ditto.
> (_mm512_cvtne2ps_pbh): Ditto.
> (_mm512_mask_cvtne2ps_pbh): Ditto.
> (_mm512_maskz_cvtne2ps_pbh): Ditto.
> * config/i386/avx512bf16vlintrin.h (__attribute__): Ditto.
> (_mm256_cvtne2ps_pbh): Ditto.
> (_mm256_mask_cvtne2ps_pbh): Ditto.
> (_mm256_maskz_cvtne2ps_pbh): Ditto.
> (_mm_cvtne2ps_pbh): Ditto.
> (_mm_mask_cvtne2ps_pbh): Ditto.
> (_mm_maskz_cvtne2ps_pbh): Ditto.
> (_mm_cvtness_sbh): Ditto.
> * config/i386/i386-builtin-types.def (V8BF): Add new
> DEF_VECTOR_TYPE for BFmode.
> (V16BF): Ditto.
> (V32BF): Ditto.
> * config/i386/i386-builtin.def (BDESC): Fixed builtins.
> * config/i386/i386-expand.cc (ix86_expand_args_builtin): Changed
> avx512bf16 ix86_builtin_func_type included HI to BF.
> * config/i386/immintrin.h: Add SSE2 depend for avx512bf16.
> * config/i386/sse.md (TARGET_AVX512VL): Changed HI vector to BF
> vector.
> (avx512f_cvtneps2bf16_v4sf): New define_expand.
> (*avx512f_cvtneps2bf16_v4sf): New define_insn.
> (avx512f_cvtneps2bf16_v4sf_maskz):Ditto.
> (avx512f_cvtneps2bf16_v4sf_mask): Ditto.
> (avx512f_cvtneps2bf16_v4sf_mask_1): Ditto.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/avx512bf16-cvtsbh2ss-1.c: Add fpmath option.
> * gcc.target/i386/avx512bf16-vdpbf16ps-2.c: Fixed
> scan-assembler.
> * gcc.target/i386/avx512bf16vl-cvtness2sbh-1.c: Add x/y suffix
> for vcvtneps2bf16.
> * gcc.target/i386/avx512bf16vl-vcvtneps2bf16-1.c: Ditto.
> ---
> gcc/config/i386/avx512bf16intrin.h| 12 +--
> gcc/config/i386/avx512bf16vlintrin.h | 29 ++---
> gcc/config/i386/i386-builtin-types.def| 51 -
> gcc/config/i386/i386-builtin.def | 54 +-
> gcc/config/i386/i386-expand.cc| 48 -
> gcc/config/i386/immintrin.h | 2 +
> gcc/config/i386/sse.md| 101 ++
> .../gcc.target/i386/avx512bf16-cvtsbh2ss-1.c | 2 +-
> .../gcc.target/i386/avx512bf16-vdpbf16ps-2.c | 2 +-
> .../i386/avx512bf16vl-cvtness2sbh-1.c | 2 +-
> .../i386/avx512bf16vl-vcvtneps2bf16-1.c | 12 +--
> 11 files changed, 189 insertions(+), 126 deletions(-)
>
> diff --git a/gcc/config/i386/avx512bf16intrin.h
> b/gcc/config/i386/avx512bf16intrin.h
> index b6e9ddad157..ea1d0125b3f 100644
> --- a/gcc/config/i386/avx512bf16intrin.h
> +++ b/gcc/config/i386/avx512bf16intrin.h
> @@ -35,16 +35,16 @@
> #endif /* __AVX512BF16__ */
>
> /* Internal data types for implementing the intrinsics. */
> -typedef short __v32bh __attribute__ ((__vector_size__ (64)));
> +typedef __bf16 __v32bf __attribute__ ((__vector_size__ (64)));
>
> /* The Intel API is flexible enough that we must allow aliasing with other
> vector types, and their scalar components. */
> -typedef short __m512bh __attribute__ ((__vector_size__ (64), __may_alias__));
> +typedef __bf16 __m512bh __attribute__ ((__vector_size__ (64),
> __may_alias__));
>
> /* Convert One BF16 Data to One Single Float Data. */
> extern __inline float
> __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> -_mm_cvtsbh_ss (__bfloat16 __A)
> +_mm_cvtsbh_ss (__bf16 __A)
> {
>union{ float a; unsigned int b;} __tmp;
>__tmp.b = ((unsigned int)(__A)) << 16;
> @@ -57,21 +57,21 @@ extern __inline __m512bh
> __attribute__((__gnu_inline__, __always_inline__, __artificial__))
> _mm512_cvtne2ps_pbh (__m512 __A, __m512 __B)
> {
> - return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi(__A, __B);
> + return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf(__A, __B);
> }
>
> extern __inline __m512bh
> __attribute__((__gnu_inline__, __always_inline__, __artificial__))
> _mm512_mask_cvtne2ps_pbh (__m512bh __A, __mmask32 __B, __m512 __C, __m512
> __D)
> {
> - return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi_mask(__C, __D, __A,
> __B);
> + return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf_mask(__C, __D, __A,
> __B);
> }
>
> extern __inline __m512bh
> __attribute__((__gnu_inline__, __always_inline__, __artificial__))
> _mm512_maskz_cvtne2ps_pbh (__mmask32 __A, __m512 __B, __m512 __C)
> {
> - return
[PATCH] i386: using __bf16 for AVX512BF16 intrinsics
Hi,
Previously we use unsigned short to represent bf16. It's not a good expression,
and at the time the front end didn't support bf16 type.
Now we introduced __bf16 to X86 psABI. So we can switch intrinsics to the new
type.
Ok for trunk ?
Thanks,
Lingling
gcc/ChangeLog:
* config/i386/avx512bf16intrin.h (__attribute__): Change short to bf16.
(_mm_cvtsbh_ss): Ditto.
(_mm512_cvtne2ps_pbh): Ditto.
(_mm512_mask_cvtne2ps_pbh): Ditto.
(_mm512_maskz_cvtne2ps_pbh): Ditto.
* config/i386/avx512bf16vlintrin.h (__attribute__): Ditto.
(_mm256_cvtne2ps_pbh): Ditto.
(_mm256_mask_cvtne2ps_pbh): Ditto.
(_mm256_maskz_cvtne2ps_pbh): Ditto.
(_mm_cvtne2ps_pbh): Ditto.
(_mm_mask_cvtne2ps_pbh): Ditto.
(_mm_maskz_cvtne2ps_pbh): Ditto.
(_mm_cvtness_sbh): Ditto.
* config/i386/i386-builtin-types.def (V8BF): Add new
DEF_VECTOR_TYPE for BFmode.
(V16BF): Ditto.
(V32BF): Ditto.
* config/i386/i386-builtin.def (BDESC): Fixed builtins.
* config/i386/i386-expand.cc (ix86_expand_args_builtin): Changed
avx512bf16 ix86_builtin_func_type included HI to BF.
* config/i386/immintrin.h: Add SSE2 depend for avx512bf16.
* config/i386/sse.md (TARGET_AVX512VL): Changed HI vector to BF
vector.
(avx512f_cvtneps2bf16_v4sf): New define_expand.
(*avx512f_cvtneps2bf16_v4sf): New define_insn.
(avx512f_cvtneps2bf16_v4sf_maskz):Ditto.
(avx512f_cvtneps2bf16_v4sf_mask): Ditto.
(avx512f_cvtneps2bf16_v4sf_mask_1): Ditto.
gcc/testsuite/ChangeLog:
* gcc.target/i386/avx512bf16-cvtsbh2ss-1.c: Add fpmath option.
* gcc.target/i386/avx512bf16-vdpbf16ps-2.c: Fixed
scan-assembler.
* gcc.target/i386/avx512bf16vl-cvtness2sbh-1.c: Add x/y suffix
for vcvtneps2bf16.
* gcc.target/i386/avx512bf16vl-vcvtneps2bf16-1.c: Ditto.
---
gcc/config/i386/avx512bf16intrin.h| 12 +--
gcc/config/i386/avx512bf16vlintrin.h | 29 ++---
gcc/config/i386/i386-builtin-types.def| 51 -
gcc/config/i386/i386-builtin.def | 54 +-
gcc/config/i386/i386-expand.cc| 48 -
gcc/config/i386/immintrin.h | 2 +
gcc/config/i386/sse.md| 101 ++
.../gcc.target/i386/avx512bf16-cvtsbh2ss-1.c | 2 +-
.../gcc.target/i386/avx512bf16-vdpbf16ps-2.c | 2 +-
.../i386/avx512bf16vl-cvtness2sbh-1.c | 2 +-
.../i386/avx512bf16vl-vcvtneps2bf16-1.c | 12 +--
11 files changed, 189 insertions(+), 126 deletions(-)
diff --git a/gcc/config/i386/avx512bf16intrin.h
b/gcc/config/i386/avx512bf16intrin.h
index b6e9ddad157..ea1d0125b3f 100644
--- a/gcc/config/i386/avx512bf16intrin.h
+++ b/gcc/config/i386/avx512bf16intrin.h
@@ -35,16 +35,16 @@
#endif /* __AVX512BF16__ */
/* Internal data types for implementing the intrinsics. */
-typedef short __v32bh __attribute__ ((__vector_size__ (64)));
+typedef __bf16 __v32bf __attribute__ ((__vector_size__ (64)));
/* The Intel API is flexible enough that we must allow aliasing with other
vector types, and their scalar components. */
-typedef short __m512bh __attribute__ ((__vector_size__ (64), __may_alias__));
+typedef __bf16 __m512bh __attribute__ ((__vector_size__ (64), __may_alias__));
/* Convert One BF16 Data to One Single Float Data. */
extern __inline float
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
-_mm_cvtsbh_ss (__bfloat16 __A)
+_mm_cvtsbh_ss (__bf16 __A)
{
union{ float a; unsigned int b;} __tmp;
__tmp.b = ((unsigned int)(__A)) << 16;
@@ -57,21 +57,21 @@ extern __inline __m512bh
__attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm512_cvtne2ps_pbh (__m512 __A, __m512 __B)
{
- return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi(__A, __B);
+ return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf(__A, __B);
}
extern __inline __m512bh
__attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm512_mask_cvtne2ps_pbh (__m512bh __A, __mmask32 __B, __m512 __C, __m512 __D)
{
- return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi_mask(__C, __D, __A, __B);
+ return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf_mask(__C, __D, __A, __B);
}
extern __inline __m512bh
__attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm512_maskz_cvtne2ps_pbh (__mmask32 __A, __m512 __B, __m512 __C)
{
- return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi_maskz(__B, __C, __A);
+ return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf_maskz(__B, __C, __A);
}
/* vcvtneps2bf16 */
diff --git a/gcc/config/i386/avx512bf16vlintrin.h
b/gcc/config/i386/avx512bf16vlintrin.h
index 969335ff358..56c28f14cf6 100644
--- a/gcc/config/i386/avx512bf16vlintrin.h
+++ b/gcc/config/i386/avx512bf16vlintrin.h
@@ -35,57 +35,58 @@
#endif /* __AVX512BF16__ */
/* Internal data types for
