Some of the comments on the BFMMLA/BFMLA[LT] patch apply here too.
Delia Burduv <[email protected]> writes:
> This patch adds the Armv8.6-a ACLE intrinsics for bfmmla, bfmlalb and
> bfmlalt as part of the BFloat16 extension.
That's the other patch :-)
> [...]
> diff --git a/gcc/config/aarch64/aarch64-simd.md
> b/gcc/config/aarch64/aarch64-simd.md
> index
> 55660ae248f4fa75d35ba2949cd4b9d5139ff5f5..ff7a1f5f34a19b05eba48dba96c736dfdfdf7bac
> 100644
> --- a/gcc/config/aarch64/aarch64-simd.md
> +++ b/gcc/config/aarch64/aarch64-simd.md
> @@ -7027,3 +7027,32 @@
> "xtn\t%0.<Vntype>, %1.<Vtype>"
> [(set_attr "type" "neon_shift_imm_narrow_q")]
> )
> +
> +;; bfcvtn
> +(define_insn "aarch64_bfcvtn<q><mode>"
> + [(set (match_operand:V4SF_TO_BF 0 "register_operand" "=w")
> + (unspec:V4SF_TO_BF [(match_operand:V4SF 1 "register_operand" "w")]
> + UNSPEC_BFCVTN))]
> + "TARGET_BF16_SIMD"
> + "bfcvtn\\t%0.4h, %1.4s"
> + [(set_attr "type" "f_cvt")]
> +)
> +
If I've understood the naming convention correctly, the closest type
seems to be "neon_fp_cvt_narrow_s_q".
> +(define_insn "aarch64_bfcvtn2v8bf"
> + [(set (match_operand:V8BF 0 "register_operand" "=w")
> + (unspec:V8BF [(match_operand:V8BF 1 "register_operand" "w")
> + (match_operand:V4SF 2 "register_operand" "w")]
> + UNSPEC_BFCVTN2))]
> + "TARGET_BF16_SIMD"
> + "bfcvtn2\\t%0.8h, %2.4s"
> + [(set_attr "type" "f_cvt")]
> +)
Same here.
The constraint on operand 1 needs to be "0", otherwise operands 1 and 0
could end up in different registers. You could test for this using
something like:
bfloat16x8_t test_bfcvtnq2_untied (bfloat16x8_t unused, bfloat16x8_t inactive,
float32x4_t a)
{
return vcvtq_high_bf16_f32 (inactive, a);
}
which when compiled at -O should produce something like:
/*
**test_bfcvtnq2_untied:
** mov v0\.8h, v1\.8h
** bfcvtn2 v0\.8h, v2\.4s
** ret
*/
(Completely untested, the code above is probably wrong.)
> +
> +(define_insn "aarch64_bfcvtbf"
> + [(set (match_operand:BF 0 "register_operand" "=w")
> + (unspec:BF [(match_operand:SF 1 "register_operand" "w")]
> + UNSPEC_BFCVT))]
> + "TARGET_BF16_SIMD"
I think this just needs the scalar macro rather than *_SIMD.
> + "bfcvt\\t%h0, %s1"
> + [(set_attr "type" "f_cvt")]
> +)
> diff --git a/gcc/config/aarch64/arm_bf16.h b/gcc/config/aarch64/arm_bf16.h
> index
> aedb0972735ce549fac1870bacd1ef3101e8fd26..1b9ab3690d35e153cd4f24b9e3bbb5b4cc4b4f4d
> 100644
> --- a/gcc/config/aarch64/arm_bf16.h
> +++ b/gcc/config/aarch64/arm_bf16.h
> @@ -34,7 +34,15 @@
> #ifdef __ARM_FEATURE_BF16_SCALAR_ARITHMETIC
>
> typedef __bf16 bfloat16_t;
> -
> +typedef float float32_t;
> +
> +__extension__ extern __inline bfloat16_t
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vcvth_bf16_f32 \
> + (float32_t __a)
No need for the line break here.
> +{
> + return __builtin_aarch64_bfcvtbf (__a);
> +}
>
> #endif
> #pragma GCC pop_options
> diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h
> index
> 6cdbf381f0156ed993f03b847228b36ebbdd14f8..120f4b7d8827aee51834e75aeaa6ab8f8451980e
> 100644
> --- a/gcc/config/aarch64/arm_neon.h
> +++ b/gcc/config/aarch64/arm_neon.h
> @@ -34610,6 +34610,35 @@ vrnd64xq_f64 (float64x2_t __a)
>
> #include "arm_bf16.h"
>
> +#pragma GCC push_options
> +#pragma GCC target ("arch=armv8.2-a+bf16")
> +#ifdef __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
> +
> +__extension__ extern __inline bfloat16x4_t
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vcvt_bf16_f32 (float32x4_t __a)
> +{
> + return __builtin_aarch64_bfcvtnv4bf (__a);
> +
> +}
Nit: extra blank line.
> +
> +__extension__ extern __inline bfloat16x8_t
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vcvtq_low_bf16_f32 (float32x4_t __a)
> +{
> + return __builtin_aarch64_bfcvtn_qv8bf (__a);
> +}
> +
> +__extension__ extern __inline bfloat16x8_t
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vcvtq_high_bf16_f32 (bfloat16x8_t __inactive, float32x4_t __a)
> +{
> + return __builtin_aarch64_bfcvtn2v8bf (__inactive, __a);
> +}
> +
> +#endif
> +#pragma GCC pop_options
> +
> #pragma GCC pop_options
>
> #undef __aarch64_vget_lane_any
> diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
> index
> 931166da5e47302afe810498eea9c8c2ab89b9de..f9f0bafb1eca4da42e564224fca1fd43d89f6ed1
> 100644
> --- a/gcc/config/aarch64/iterators.md
> +++ b/gcc/config/aarch64/iterators.md
> @@ -431,6 +431,9 @@
> ;; SVE predicate modes that control 16-bit, 32-bit or 64-bit elements.
> (define_mode_iterator PRED_HSD [VNx8BI VNx4BI VNx2BI])
>
> +;; Bfloat16 modes to which V4SF can be converted
> +(define_mode_iterator V4SF_TO_BF [V4BF V8BF])
> +
> ;; ------------------------------------------------------------------
> ;; Unspec enumerations for Advance SIMD. These could well go into
> ;; aarch64.md but for their use in int_iterators here.
> @@ -673,6 +676,9 @@
> UNSPEC_UMULHS ; Used in aarch64-sve2.md.
> UNSPEC_UMULHRS ; Used in aarch64-sve2.md.
> UNSPEC_ASRD ; Used in aarch64-sve.md.
> + UNSPEC_BFCVTN ; Used in aarch64-simd.md.
> + UNSPEC_BFCVTN2 ; Used in aarch64-simd.md.
> + UNSPEC_BFCVT ; Used in aarch64-simd.md.
> ])
>
> ;; ------------------------------------------------------------------
> diff --git a/gcc/config/arm/types.md b/gcc/config/arm/types.md
> index
> df39522f2ad63a52c910b1a6bcc7aa13aaf5d021..dbcb4d58798d7f51b1b8310cd446c58317d7b50d
> 100644
> --- a/gcc/config/arm/types.md
> +++ b/gcc/config/arm/types.md
> @@ -1097,7 +1097,8 @@
> crypto_sm4,\
> coproc,\
> tme,\
> - memtag"
> + memtag,\
> + bf_cvt"
This doesn't seem to be used.
> diff --git
> a/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/bfcvt-compile.c
> b/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/bfcvt-compile.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..ebe5b578c1fa82a6f2a166d55c7dc7e905b87135
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/bfcvt-compile.c
> @@ -0,0 +1,56 @@
> +/* { dg-do assemble { target { aarch64*-*-* } } } */
> +/* { dg-require-effective-target arm_v8_2a_bf16_neon_ok } */
> +/* { dg-add-options arm_v8_2a_bf16_neon } */
> +/* { dg-additional-options "-save-temps" } */
> +/* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
> +
> +#include <arm_neon.h>
> +
> +/*
> +**test_bfcvtn:
> +** ...
> +** bfcvtn\tv[0-9]+.4h, v[0-9]+.4s
> +** ...
> +*/
> +bfloat16x4_t test_bfcvtn (float32x4_t a)
> +{
> + return vcvt_bf16_f32 (a);
> +}
> +
> +/*
> +**test_bfcvtnq:
> +** ...
> +** bfcvtn v[0-9]+.4h, v[0-9]+.4s
> +** ...
> +*/
> +bfloat16x8_t test_bfcvtnq (float32x4_t a)
> +{
> + return vcvtq_low_bf16_f32 (a);
> +}
> +
> +/*
> +**test_bfcvtnq2:
> +** ...
> +** bfcvtn v[0-9]+.4h, v[0-9]+.4s
> +** ...
> +*/
> +bfloat16x8_t test_bfcvtnq2 (bfloat16x8_t inactive, float32x4_t a)
> +{
> + return vcvtq_high_bf16_f32 (inactive, a);
> +}
> +
> +/*
> +**test_bfcvt:
> +** ...
> +** bfcvt h[0-9]+, s[0-9]+
> +** ...
> +*/
> +bfloat16_t test_bfcvt (float32_t a)
> +{
> + return vcvth_bf16_f32 (a);
> +}
> +
> +/* { dg-final { scan-assembler {bfcvtn\tv[0-9]+.4h, v[0-9]+.4s} } } */
> +/* { dg-final { scan-assembler {bfcvtn\tv[0-9]+.4h, v[0-9]+.4s} } } */
> +/* { dg-final { scan-assembler {bfcvtn\tv[0-9]+.4h, v[0-9]+.4s} } } */
> +/* { dg-final { scan-assembler {bfcvt\th[0-9]+, s[0-9]+} } } */
Same comments as for the BFMMLA/BFMLA[BT] tests.
As well as testing all these combinations for the SIMD case,
it would be good to have a direct arm_bf16.h-only test for:
#pragma GCC target "arch=armv8.2-a+bf16+nosimd"
test_bfcvt should still work in that case.
It would also be good to have a test that test_bfcvt reports
an appropriate error if compiled after:
#pragma GCC target "arch=armv8.2-a+nobf16"
Thanks,
Richard
