Re: [ARM] Enable FP16 vector arithmetic operations.

[Kyrill Tkachov <kyrylo dot tkachov at foss dot arm dot com>]

Hi Tamar,

On 02/05/17 15:46, Tamar Christina wrote:
> Hi All,
>
> I'm taking this one over from Matthew, I think it slipped through the cracks before.
>
> Since it still applies cleanly on trunk I'm just pinging it.
>
> Ok for trunk?

Sorry for missing this.
For the record you are referring to the patch at:
https://gcc.gnu.org/ml/gcc-patches/2016-09/msg01700.html

This is ok and in line with what we do for the f32 intrinsics.
My only concern was that we can do this only if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
is defined from the architecture/fpu level, but these intrinsics are already
gated on that in arm_neon.h.

This is ok for trunk if a bootstrap and test run on arm-none-linux-gnueabihf with
current trunk shows no issues.

Thanks,
Kyrill

> Tamar
> ________________________________________
> From: gcc-patches-owner@gcc.gnu.org <gcc-patches-owner@gcc.gnu.org> on behalf of Matthew Wahab <matthew.wahab@foss.arm.com>
> Sent: Friday, September 23, 2016 4:02 PM
> To: gcc-patches
> Subject: [ARM] Enable FP16 vector arithmetic operations.
>
> Hello,
>
> Support for the ARMv8.2-A FP16 NEON arithmetic instructions was added
> using non-standard names for the instruction patterns. This was needed
> because the NEON floating point semantics meant that their use by the
> compiler for HFmode arithmetic operations needed to be restricted. This
> follows the implementation for 32-bit NEON intructions.
>
> As with the 32-bit instructions, the restriction on the HFmode
> operation can be lifted when -funsafe-math-optimizations is
> enabled. This patch does that, defining the standard pattern names
> addhf3, subhf3, mulhf3 and fmahf3.
>
> This patch also updates the NEON intrinsics to use the arithmetic
> operations when -ffast-math is enabled. This is to make keep the 16-bit
> support consistent with the 32-bit supportd. It is needed so that code
> using the f16 intrinsics are subject to the same optimizations as code
> using the f32 intrinsics would be.
>
> Tested for arm-none-linux-gnueabihf with native bootstrap and make check
> on ARMv8-A and for arm-none-eabi and armeb-none-eabi with cross-compiled
> make check on an ARMv8.2-A emulator.
>
> Ok for trunk?
> Matthew
>
> gcc/
> 2016-09-23  Matthew Wahab  <matthew.wahab@arm.com>
>
>         * config/arm/arm_neon.h (vadd_f16): Use standard arithmetic
>         operations in fast-math mode.
>         (vaddq_f16): Likewise.
>         (vmul_f16): Likewise.
>         (vmulq_f16): Likewise.
>         (vsub_f16): Likewise.
>         (vsubq_f16): Likewise.
>         * config/arm/neon.md (add<mode>3): New.
>         (sub<mode>3): New.
>         (fma:<VH:mode>3): New.  Also remove outdated comment.
>         (mul<mode>3): New.
>
> testsuite/
> 2016-09-23  Matthew Wahab  <matthew.wahab@arm.com>
>
>         * gcc.target/arm/armv8_2-fp16-arith-1.c: Expand comment.  Update
>         expected output of vadd, vsub and vmul instructions.
>         * gcc.target/arm/armv8_2-fp16-arith-2.c: New.
>         * gcc.target/arm/armv8_2-fp16-neon-2.c: New.
>         * gcc.target/arm/armv8_2-fp16-neon-3.c: New.