aarch64: Optimize SVE cos & cosf
Checks
| Context |
Check |
Description |
| redhat-pt-bot/TryBot-apply_patch |
success
|
Patch applied to master at the time it was sent
|
| redhat-pt-bot/TryBot-32bit |
success
|
Build for i686
|
| linaro-tcwg-bot/tcwg_glibc_build--master-aarch64 |
success
|
Testing passed
|
| linaro-tcwg-bot/tcwg_glibc_check--master-aarch64 |
success
|
Testing passed
|
| linaro-tcwg-bot/tcwg_glibc_build--master-arm |
success
|
Testing passed
|
| linaro-tcwg-bot/tcwg_glibc_check--master-arm |
success
|
Testing passed
|
Commit Message
Saves a mov by ensuring return value does not need to be moved out of
the way before special-case branch. Also change to use overloaded
intrinsics.
---
Thanks,
Joe
sysdeps/aarch64/fpu/cos_sve.c | 53 +++++++++++++++++-----------------
sysdeps/aarch64/fpu/cosf_sve.c | 47 ++++++++++++++----------------
2 files changed, 47 insertions(+), 53 deletions(-)
Comments
The 10/04/2023 10:37, Joe Ramsay wrote:
> Saves a mov by ensuring return value does not need to be moved out of
> the way before special-case branch. Also change to use overloaded
> intrinsics.
looks good. committed.
> ---
> Thanks,
> Joe
> sysdeps/aarch64/fpu/cos_sve.c | 53 +++++++++++++++++-----------------
> sysdeps/aarch64/fpu/cosf_sve.c | 47 ++++++++++++++----------------
> 2 files changed, 47 insertions(+), 53 deletions(-)
>
> diff --git a/sysdeps/aarch64/fpu/cos_sve.c b/sysdeps/aarch64/fpu/cos_sve.c
> index d7a9b134da..c804e52fc6 100644
> --- a/sysdeps/aarch64/fpu/cos_sve.c
> +++ b/sysdeps/aarch64/fpu/cos_sve.c
> @@ -37,9 +37,9 @@ static const struct data
> #define RangeVal 0x4160000000000000 /* asuint64 (0x1p23). */
>
> static svfloat64_t NOINLINE
> -special_case (svfloat64_t x, svfloat64_t y, svbool_t out_of_bounds)
> +special_case (svfloat64_t x, svfloat64_t y, svbool_t oob)
> {
> - return sv_call_f64 (cos, x, y, out_of_bounds);
> + return sv_call_f64 (cos, x, y, oob);
> }
>
> /* A fast SVE implementation of cos based on trigonometric
> @@ -51,42 +51,41 @@ svfloat64_t SV_NAME_D1 (cos) (svfloat64_t x, const svbool_t pg)
> {
> const struct data *d = ptr_barrier (&data);
>
> - svfloat64_t r = svabs_f64_x (pg, x);
> - svbool_t out_of_bounds
> - = svcmpge_n_u64 (pg, svreinterpret_u64_f64 (r), RangeVal);
> + svfloat64_t r = svabs_x (pg, x);
> + svbool_t oob = svcmpge (pg, svreinterpret_u64 (r), RangeVal);
>
> /* Load some constants in quad-word chunks to minimise memory access. */
> svbool_t ptrue = svptrue_b64 ();
> - svfloat64_t invpio2_and_pio2_1 = svld1rq_f64 (ptrue, &d->inv_pio2);
> - svfloat64_t pio2_23 = svld1rq_f64 (ptrue, &d->pio2_2);
> + svfloat64_t invpio2_and_pio2_1 = svld1rq (ptrue, &d->inv_pio2);
> + svfloat64_t pio2_23 = svld1rq (ptrue, &d->pio2_2);
>
> /* n = rint(|x|/(pi/2)). */
> - svfloat64_t q = svmla_lane_f64 (sv_f64 (d->shift), r, invpio2_and_pio2_1, 0);
> - svfloat64_t n = svsub_n_f64_x (pg, q, d->shift);
> + svfloat64_t q = svmla_lane (sv_f64 (d->shift), r, invpio2_and_pio2_1, 0);
> + svfloat64_t n = svsub_x (pg, q, d->shift);
>
> /* r = |x| - n*(pi/2) (range reduction into -pi/4 .. pi/4). */
> - r = svmls_lane_f64 (r, n, invpio2_and_pio2_1, 1);
> - r = svmls_lane_f64 (r, n, pio2_23, 0);
> - r = svmls_lane_f64 (r, n, pio2_23, 1);
> + r = svmls_lane (r, n, invpio2_and_pio2_1, 1);
> + r = svmls_lane (r, n, pio2_23, 0);
> + r = svmls_lane (r, n, pio2_23, 1);
>
> /* cos(r) poly approx. */
> - svfloat64_t r2 = svtsmul_f64 (r, svreinterpret_u64_f64 (q));
> + svfloat64_t r2 = svtsmul (r, svreinterpret_u64 (q));
> svfloat64_t y = sv_f64 (0.0);
> - y = svtmad_f64 (y, r2, 7);
> - y = svtmad_f64 (y, r2, 6);
> - y = svtmad_f64 (y, r2, 5);
> - y = svtmad_f64 (y, r2, 4);
> - y = svtmad_f64 (y, r2, 3);
> - y = svtmad_f64 (y, r2, 2);
> - y = svtmad_f64 (y, r2, 1);
> - y = svtmad_f64 (y, r2, 0);
> + y = svtmad (y, r2, 7);
> + y = svtmad (y, r2, 6);
> + y = svtmad (y, r2, 5);
> + y = svtmad (y, r2, 4);
> + y = svtmad (y, r2, 3);
> + y = svtmad (y, r2, 2);
> + y = svtmad (y, r2, 1);
> + y = svtmad (y, r2, 0);
>
> /* Final multiplicative factor: 1.0 or x depending on bit #0 of q. */
> - svfloat64_t f = svtssel_f64 (r, svreinterpret_u64_f64 (q));
> - /* Apply factor. */
> - y = svmul_f64_x (pg, f, y);
> + svfloat64_t f = svtssel (r, svreinterpret_u64 (q));
> +
> + if (__glibc_unlikely (svptest_any (pg, oob)))
> + return special_case (x, svmul_x (svnot_z (pg, oob), y, f), oob);
>
> - if (__glibc_unlikely (svptest_any (pg, out_of_bounds)))
> - return special_case (x, y, out_of_bounds);
> - return y;
> + /* Apply factor. */
> + return svmul_x (pg, f, y);
> }
> diff --git a/sysdeps/aarch64/fpu/cosf_sve.c b/sysdeps/aarch64/fpu/cosf_sve.c
> index 577cbd864e..a0be56ec7e 100644
> --- a/sysdeps/aarch64/fpu/cosf_sve.c
> +++ b/sysdeps/aarch64/fpu/cosf_sve.c
> @@ -37,9 +37,9 @@ static const struct data
> #define RangeVal 0x49800000 /* asuint32(0x1p20f). */
>
> static svfloat32_t NOINLINE
> -special_case (svfloat32_t x, svfloat32_t y, svbool_t out_of_bounds)
> +special_case (svfloat32_t x, svfloat32_t y, svbool_t oob)
> {
> - return sv_call_f32 (cosf, x, y, out_of_bounds);
> + return sv_call_f32 (cosf, x, y, oob);
> }
>
> /* A fast SVE implementation of cosf based on trigonometric
> @@ -51,40 +51,35 @@ svfloat32_t SV_NAME_F1 (cos) (svfloat32_t x, const svbool_t pg)
> {
> const struct data *d = ptr_barrier (&data);
>
> - svfloat32_t r = svabs_f32_x (pg, x);
> - svbool_t out_of_bounds
> - = svcmpge_n_u32 (pg, svreinterpret_u32_f32 (r), RangeVal);
> + svfloat32_t r = svabs_x (pg, x);
> + svbool_t oob = svcmpge (pg, svreinterpret_u32 (r), RangeVal);
>
> /* Load some constants in quad-word chunks to minimise memory access. */
> - svfloat32_t negpio2_and_invpio2
> - = svld1rq_f32 (svptrue_b32 (), &d->neg_pio2_1);
> + svfloat32_t negpio2_and_invpio2 = svld1rq (svptrue_b32 (), &d->neg_pio2_1);
>
> /* n = rint(|x|/(pi/2)). */
> - svfloat32_t q
> - = svmla_lane_f32 (sv_f32 (d->shift), r, negpio2_and_invpio2, 3);
> - svfloat32_t n = svsub_n_f32_x (pg, q, d->shift);
> + svfloat32_t q = svmla_lane (sv_f32 (d->shift), r, negpio2_and_invpio2, 3);
> + svfloat32_t n = svsub_x (pg, q, d->shift);
>
> /* r = |x| - n*(pi/2) (range reduction into -pi/4 .. pi/4). */
> - r = svmla_lane_f32 (r, n, negpio2_and_invpio2, 0);
> - r = svmla_lane_f32 (r, n, negpio2_and_invpio2, 1);
> - r = svmla_lane_f32 (r, n, negpio2_and_invpio2, 2);
> + r = svmla_lane (r, n, negpio2_and_invpio2, 0);
> + r = svmla_lane (r, n, negpio2_and_invpio2, 1);
> + r = svmla_lane (r, n, negpio2_and_invpio2, 2);
>
> /* Final multiplicative factor: 1.0 or x depending on bit #0 of q. */
> - svfloat32_t f = svtssel_f32 (r, svreinterpret_u32_f32 (q));
> + svfloat32_t f = svtssel (r, svreinterpret_u32 (q));
>
> /* cos(r) poly approx. */
> - svfloat32_t r2 = svtsmul_f32 (r, svreinterpret_u32_f32 (q));
> + svfloat32_t r2 = svtsmul (r, svreinterpret_u32 (q));
> svfloat32_t y = sv_f32 (0.0f);
> - y = svtmad_f32 (y, r2, 4);
> - y = svtmad_f32 (y, r2, 3);
> - y = svtmad_f32 (y, r2, 2);
> - y = svtmad_f32 (y, r2, 1);
> - y = svtmad_f32 (y, r2, 0);
> -
> + y = svtmad (y, r2, 4);
> + y = svtmad (y, r2, 3);
> + y = svtmad (y, r2, 2);
> + y = svtmad (y, r2, 1);
> + y = svtmad (y, r2, 0);
> +
> + if (__glibc_unlikely (svptest_any (pg, oob)))
> + return special_case (x, svmul_x (svnot_z (pg, oob), f, y), oob);
> /* Apply factor. */
> - y = svmul_f32_x (pg, f, y);
> -
> - if (__glibc_unlikely (svptest_any (pg, out_of_bounds)))
> - return special_case (x, y, out_of_bounds);
> - return y;
> + return svmul_x (pg, f, y);
> }
> --
> 2.27.0
>
@@ -37,9 +37,9 @@ static const struct data
#define RangeVal 0x4160000000000000 /* asuint64 (0x1p23). */
static svfloat64_t NOINLINE
-special_case (svfloat64_t x, svfloat64_t y, svbool_t out_of_bounds)
+special_case (svfloat64_t x, svfloat64_t y, svbool_t oob)
{
- return sv_call_f64 (cos, x, y, out_of_bounds);
+ return sv_call_f64 (cos, x, y, oob);
}
/* A fast SVE implementation of cos based on trigonometric
@@ -51,42 +51,41 @@ svfloat64_t SV_NAME_D1 (cos) (svfloat64_t x, const svbool_t pg)
{
const struct data *d = ptr_barrier (&data);
- svfloat64_t r = svabs_f64_x (pg, x);
- svbool_t out_of_bounds
- = svcmpge_n_u64 (pg, svreinterpret_u64_f64 (r), RangeVal);
+ svfloat64_t r = svabs_x (pg, x);
+ svbool_t oob = svcmpge (pg, svreinterpret_u64 (r), RangeVal);
/* Load some constants in quad-word chunks to minimise memory access. */
svbool_t ptrue = svptrue_b64 ();
- svfloat64_t invpio2_and_pio2_1 = svld1rq_f64 (ptrue, &d->inv_pio2);
- svfloat64_t pio2_23 = svld1rq_f64 (ptrue, &d->pio2_2);
+ svfloat64_t invpio2_and_pio2_1 = svld1rq (ptrue, &d->inv_pio2);
+ svfloat64_t pio2_23 = svld1rq (ptrue, &d->pio2_2);
/* n = rint(|x|/(pi/2)). */
- svfloat64_t q = svmla_lane_f64 (sv_f64 (d->shift), r, invpio2_and_pio2_1, 0);
- svfloat64_t n = svsub_n_f64_x (pg, q, d->shift);
+ svfloat64_t q = svmla_lane (sv_f64 (d->shift), r, invpio2_and_pio2_1, 0);
+ svfloat64_t n = svsub_x (pg, q, d->shift);
/* r = |x| - n*(pi/2) (range reduction into -pi/4 .. pi/4). */
- r = svmls_lane_f64 (r, n, invpio2_and_pio2_1, 1);
- r = svmls_lane_f64 (r, n, pio2_23, 0);
- r = svmls_lane_f64 (r, n, pio2_23, 1);
+ r = svmls_lane (r, n, invpio2_and_pio2_1, 1);
+ r = svmls_lane (r, n, pio2_23, 0);
+ r = svmls_lane (r, n, pio2_23, 1);
/* cos(r) poly approx. */
- svfloat64_t r2 = svtsmul_f64 (r, svreinterpret_u64_f64 (q));
+ svfloat64_t r2 = svtsmul (r, svreinterpret_u64 (q));
svfloat64_t y = sv_f64 (0.0);
- y = svtmad_f64 (y, r2, 7);
- y = svtmad_f64 (y, r2, 6);
- y = svtmad_f64 (y, r2, 5);
- y = svtmad_f64 (y, r2, 4);
- y = svtmad_f64 (y, r2, 3);
- y = svtmad_f64 (y, r2, 2);
- y = svtmad_f64 (y, r2, 1);
- y = svtmad_f64 (y, r2, 0);
+ y = svtmad (y, r2, 7);
+ y = svtmad (y, r2, 6);
+ y = svtmad (y, r2, 5);
+ y = svtmad (y, r2, 4);
+ y = svtmad (y, r2, 3);
+ y = svtmad (y, r2, 2);
+ y = svtmad (y, r2, 1);
+ y = svtmad (y, r2, 0);
/* Final multiplicative factor: 1.0 or x depending on bit #0 of q. */
- svfloat64_t f = svtssel_f64 (r, svreinterpret_u64_f64 (q));
- /* Apply factor. */
- y = svmul_f64_x (pg, f, y);
+ svfloat64_t f = svtssel (r, svreinterpret_u64 (q));
+
+ if (__glibc_unlikely (svptest_any (pg, oob)))
+ return special_case (x, svmul_x (svnot_z (pg, oob), y, f), oob);
- if (__glibc_unlikely (svptest_any (pg, out_of_bounds)))
- return special_case (x, y, out_of_bounds);
- return y;
+ /* Apply factor. */
+ return svmul_x (pg, f, y);
}
@@ -37,9 +37,9 @@ static const struct data
#define RangeVal 0x49800000 /* asuint32(0x1p20f). */
static svfloat32_t NOINLINE
-special_case (svfloat32_t x, svfloat32_t y, svbool_t out_of_bounds)
+special_case (svfloat32_t x, svfloat32_t y, svbool_t oob)
{
- return sv_call_f32 (cosf, x, y, out_of_bounds);
+ return sv_call_f32 (cosf, x, y, oob);
}
/* A fast SVE implementation of cosf based on trigonometric
@@ -51,40 +51,35 @@ svfloat32_t SV_NAME_F1 (cos) (svfloat32_t x, const svbool_t pg)
{
const struct data *d = ptr_barrier (&data);
- svfloat32_t r = svabs_f32_x (pg, x);
- svbool_t out_of_bounds
- = svcmpge_n_u32 (pg, svreinterpret_u32_f32 (r), RangeVal);
+ svfloat32_t r = svabs_x (pg, x);
+ svbool_t oob = svcmpge (pg, svreinterpret_u32 (r), RangeVal);
/* Load some constants in quad-word chunks to minimise memory access. */
- svfloat32_t negpio2_and_invpio2
- = svld1rq_f32 (svptrue_b32 (), &d->neg_pio2_1);
+ svfloat32_t negpio2_and_invpio2 = svld1rq (svptrue_b32 (), &d->neg_pio2_1);
/* n = rint(|x|/(pi/2)). */
- svfloat32_t q
- = svmla_lane_f32 (sv_f32 (d->shift), r, negpio2_and_invpio2, 3);
- svfloat32_t n = svsub_n_f32_x (pg, q, d->shift);
+ svfloat32_t q = svmla_lane (sv_f32 (d->shift), r, negpio2_and_invpio2, 3);
+ svfloat32_t n = svsub_x (pg, q, d->shift);
/* r = |x| - n*(pi/2) (range reduction into -pi/4 .. pi/4). */
- r = svmla_lane_f32 (r, n, negpio2_and_invpio2, 0);
- r = svmla_lane_f32 (r, n, negpio2_and_invpio2, 1);
- r = svmla_lane_f32 (r, n, negpio2_and_invpio2, 2);
+ r = svmla_lane (r, n, negpio2_and_invpio2, 0);
+ r = svmla_lane (r, n, negpio2_and_invpio2, 1);
+ r = svmla_lane (r, n, negpio2_and_invpio2, 2);
/* Final multiplicative factor: 1.0 or x depending on bit #0 of q. */
- svfloat32_t f = svtssel_f32 (r, svreinterpret_u32_f32 (q));
+ svfloat32_t f = svtssel (r, svreinterpret_u32 (q));
/* cos(r) poly approx. */
- svfloat32_t r2 = svtsmul_f32 (r, svreinterpret_u32_f32 (q));
+ svfloat32_t r2 = svtsmul (r, svreinterpret_u32 (q));
svfloat32_t y = sv_f32 (0.0f);
- y = svtmad_f32 (y, r2, 4);
- y = svtmad_f32 (y, r2, 3);
- y = svtmad_f32 (y, r2, 2);
- y = svtmad_f32 (y, r2, 1);
- y = svtmad_f32 (y, r2, 0);
-
+ y = svtmad (y, r2, 4);
+ y = svtmad (y, r2, 3);
+ y = svtmad (y, r2, 2);
+ y = svtmad (y, r2, 1);
+ y = svtmad (y, r2, 0);
+
+ if (__glibc_unlikely (svptest_any (pg, oob)))
+ return special_case (x, svmul_x (svnot_z (pg, oob), f, y), oob);
/* Apply factor. */
- y = svmul_f32_x (pg, f, y);
-
- if (__glibc_unlikely (svptest_any (pg, out_of_bounds)))
- return special_case (x, y, out_of_bounds);
- return y;
+ return svmul_x (pg, f, y);
}