[RFC,2/2] aarch64: add vector sin, cos, log and pow abi symbols

Message ID e8de3b3c-52d6-ac60-2ef1-65e0903b7027@arm.com
State New
Headers show
Series
  • [RFC,1/2] Aarch64: Add simd exp/expf ABI symbols
Related show

Commit Message

Szabolcs Nagy July 4, 2019, 11:09 a.m.
Add simple assembly implementations that fall back to scalar code,
similar to the vector exp code.

These are the symbols we expect to optimize.

(TODO: deal with sincos)

2019-07-04  Szabolcs Nagy  <szabolcs.nagy@arm.com>

	* sysdeps/aarch64/fpu/Makefile: Add functions.
	* sysdeps/aarch64/fpu/Versions: Add symbols.
	* sysdeps/aarch64/fpu/libmvec_double_vlen2_cos.S: New file.
	* sysdeps/aarch64/fpu/libmvec_double_vlen2_log.S: New file.
	* sysdeps/aarch64/fpu/libmvec_double_vlen2_pow.S: New file.
	* sysdeps/aarch64/fpu/libmvec_double_vlen2_sin.S: New file.
	* sysdeps/aarch64/fpu/libmvec_float_vlen4_cosf.S: New file.
	* sysdeps/aarch64/fpu/libmvec_float_vlen4_logf.S: New file.
	* sysdeps/aarch64/fpu/libmvec_float_vlen4_powf.S: New file.
	* sysdeps/aarch64/fpu/libmvec_float_vlen4_sinf.S: New file.
	* sysdeps/aarch64/fpu/test-double-vlen2-wrappers.c: Add wrappers.
	* sysdeps/aarch64/fpu/test-float-vlen4-wrappers.c: Add wrappers.
	* sysdeps/aarch64/libm-test-ulps: Update.
	* sysdeps/unix/sysv/linux/aarch64/libmvec.abilist: Update.
---
 sysdeps/aarch64/fpu/Makefile                  | 12 +++-
 sysdeps/aarch64/fpu/Versions                  |  4 ++
 .../aarch64/fpu/libmvec_double_vlen2_cos.S    | 21 ++++++
 .../aarch64/fpu/libmvec_double_vlen2_log.S    | 21 ++++++
 .../aarch64/fpu/libmvec_double_vlen2_pow.S    | 62 ++++++++++++++++
 .../aarch64/fpu/libmvec_double_vlen2_sin.S    | 21 ++++++
 .../aarch64/fpu/libmvec_float_vlen4_cosf.S    | 21 ++++++
 .../aarch64/fpu/libmvec_float_vlen4_logf.S    | 21 ++++++
 .../aarch64/fpu/libmvec_float_vlen4_powf.S    | 70 +++++++++++++++++++
 .../aarch64/fpu/libmvec_float_vlen4_sinf.S    | 21 ++++++
 .../aarch64/fpu/test-double-vlen2-wrappers.c  | 12 ++++
 .../aarch64/fpu/test-float-vlen4-wrappers.c   | 12 ++++
 sysdeps/aarch64/libm-test-ulps                | 18 +++++
 .../unix/sysv/linux/aarch64/libmvec.abilist   |  8 +++
 14 files changed, 322 insertions(+), 2 deletions(-)
 create mode 100644 sysdeps/aarch64/fpu/libmvec_double_vlen2_cos.S
 create mode 100644 sysdeps/aarch64/fpu/libmvec_double_vlen2_log.S
 create mode 100644 sysdeps/aarch64/fpu/libmvec_double_vlen2_pow.S
 create mode 100644 sysdeps/aarch64/fpu/libmvec_double_vlen2_sin.S
 create mode 100644 sysdeps/aarch64/fpu/libmvec_float_vlen4_cosf.S
 create mode 100644 sysdeps/aarch64/fpu/libmvec_float_vlen4_logf.S
 create mode 100644 sysdeps/aarch64/fpu/libmvec_float_vlen4_powf.S
 create mode 100644 sysdeps/aarch64/fpu/libmvec_float_vlen4_sinf.S

Patch

diff --git a/sysdeps/aarch64/fpu/Makefile b/sysdeps/aarch64/fpu/Makefile
index 220b664323..fe72a74aec 100644
--- a/sysdeps/aarch64/fpu/Makefile
+++ b/sysdeps/aarch64/fpu/Makefile
@@ -15,15 +15,23 @@  endif
 
 ifeq ($(subdir),mathvec)
 libmvec-support += \
+  libmvec_double_vlen2_cos \
   libmvec_double_vlen2_exp \
+  libmvec_double_vlen2_log \
+  libmvec_double_vlen2_pow \
+  libmvec_double_vlen2_sin \
+  libmvec_float_vlen4_cosf \
   libmvec_float_vlen4_expf \
+  libmvec_float_vlen4_logf \
+  libmvec_float_vlen4_powf \
+  libmvec_float_vlen4_sinf \
 
 endif
 
 ifeq ($(subdir),math)
 ifeq ($(build-mathvec),yes)
-double-vlen2-funcs = exp
-float-vlen4-funcs = exp
+double-vlen2-funcs = cos exp log pow sin
+float-vlen4-funcs = cos exp log pow sin
 ifeq ($(test-mathvec),yes)
 libmvec-tests += double-vlen2 float-vlen4
 endif
diff --git a/sysdeps/aarch64/fpu/Versions b/sysdeps/aarch64/fpu/Versions
index da36f3c495..94ffaeee6d 100644
--- a/sysdeps/aarch64/fpu/Versions
+++ b/sysdeps/aarch64/fpu/Versions
@@ -1,5 +1,9 @@ 
 libmvec {
   GLIBC_2.30 {
+    _ZGVnN2v_cos; _ZGVnN4v_cosf;
     _ZGVnN2v_exp; _ZGVnN4v_expf;
+    _ZGVnN2v_log; _ZGVnN4v_logf;
+    _ZGVnN2vv_pow; _ZGVnN4vv_powf;
+    _ZGVnN2v_sin; _ZGVnN4v_sinf;
   }
 }
diff --git a/sysdeps/aarch64/fpu/libmvec_double_vlen2_cos.S b/sysdeps/aarch64/fpu/libmvec_double_vlen2_cos.S
new file mode 100644
index 0000000000..f4ad3c75f4
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_double_vlen2_cos.S
@@ -0,0 +1,21 @@ 
+/* Double-precision 2 element vector cos function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define SCALAR_FUNCTION cos
+#define VECTOR_FUNCTION _ZGVnN2v_cos
+#include "libmvec_double_vlen2.h"
diff --git a/sysdeps/aarch64/fpu/libmvec_double_vlen2_log.S b/sysdeps/aarch64/fpu/libmvec_double_vlen2_log.S
new file mode 100644
index 0000000000..b802a2608a
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_double_vlen2_log.S
@@ -0,0 +1,21 @@ 
+/* Double-precision 2 element vector log function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define SCALAR_FUNCTION log
+#define VECTOR_FUNCTION _ZGVnN2v_log
+#include "libmvec_double_vlen2.h"
diff --git a/sysdeps/aarch64/fpu/libmvec_double_vlen2_pow.S b/sysdeps/aarch64/fpu/libmvec_double_vlen2_pow.S
new file mode 100644
index 0000000000..85151482bf
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_double_vlen2_pow.S
@@ -0,0 +1,62 @@ 
+/* Double-precision 2 element vector x^y function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+ENTRY (_ZGVnN2vv_pow)
+	stp	x29, x30, [sp, -304]!
+	cfi_adjust_cfa_offset (304)
+	cfi_rel_offset (x29, 0)
+	cfi_rel_offset (x30, 8)
+	mov	x29, sp
+	stp	 q8,  q9, [sp, 16]
+	stp	q10, q11, [sp, 48]
+	stp	q12, q13, [sp, 80]
+	stp	q14, q15, [sp, 112]
+	stp	q16, q17, [sp, 144]
+	stp	q18, q19, [sp, 176]
+	stp	q20, q21, [sp, 208]
+	stp	q22, q23, [sp, 240]
+
+	// Use per lane load/store to avoid endianness issues.
+	str	q0, [sp, 272]
+	str	q1, [sp, 288]
+	ldr	d0, [sp, 272]
+	ldr	d1, [sp, 288]
+	bl pow
+	str	d0, [sp, 272]
+	ldr	d0, [sp, 280]
+	ldr	d1, [sp, 296]
+	bl pow
+	str	d0, [sp, 280]
+	ldr	q0, [sp, 272]
+
+	ldp	q8, q9, [sp, 16]
+	ldp	q10, q11, [sp, 48]
+	ldp	q12, q13, [sp, 80]
+	ldp	q14, q15, [sp, 112]
+	ldp	q16, q17, [sp, 144]
+	ldp	q18, q19, [sp, 176]
+	ldp	q20, q21, [sp, 208]
+	ldp	q22, q23, [sp, 240]
+	ldp	x29, x30, [sp], 304
+	cfi_adjust_cfa_offset (304)
+	cfi_restore (x29)
+	cfi_restore (x30)
+	ret
+END (_ZGVnN2vv_pow)
diff --git a/sysdeps/aarch64/fpu/libmvec_double_vlen2_sin.S b/sysdeps/aarch64/fpu/libmvec_double_vlen2_sin.S
new file mode 100644
index 0000000000..c01e4399cd
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_double_vlen2_sin.S
@@ -0,0 +1,21 @@ 
+/* Double-precision 2 element vector sin function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define SCALAR_FUNCTION sin
+#define VECTOR_FUNCTION _ZGVnN2v_sin
+#include "libmvec_double_vlen2.h"
diff --git a/sysdeps/aarch64/fpu/libmvec_float_vlen4_cosf.S b/sysdeps/aarch64/fpu/libmvec_float_vlen4_cosf.S
new file mode 100644
index 0000000000..2d9ea9fb36
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_float_vlen4_cosf.S
@@ -0,0 +1,21 @@ 
+/* Single-precision 4 element vector cos function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define SCALAR_FUNCTION cosf
+#define VECTOR_FUNCTION _ZGVnN4v_cosf
+#include "libmvec_float_vlen4.h"
diff --git a/sysdeps/aarch64/fpu/libmvec_float_vlen4_logf.S b/sysdeps/aarch64/fpu/libmvec_float_vlen4_logf.S
new file mode 100644
index 0000000000..df961eadba
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_float_vlen4_logf.S
@@ -0,0 +1,21 @@ 
+/* Single-precision 4 element vector log function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define SCALAR_FUNCTION logf
+#define VECTOR_FUNCTION _ZGVnN4v_logf
+#include "libmvec_float_vlen4.h"
diff --git a/sysdeps/aarch64/fpu/libmvec_float_vlen4_powf.S b/sysdeps/aarch64/fpu/libmvec_float_vlen4_powf.S
new file mode 100644
index 0000000000..95e593c151
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_float_vlen4_powf.S
@@ -0,0 +1,70 @@ 
+/* Single-precision 4 element vector x^y function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+ENTRY (_ZGVnN4vv_powf)
+	stp	x29, x30, [sp, -304]!
+	cfi_adjust_cfa_offset (304)
+	cfi_rel_offset (x29, 0)
+	cfi_rel_offset (x30, 8)
+	mov	x29, sp
+	stp	 q8,  q9, [sp, 16]
+	stp	q10, q11, [sp, 48]
+	stp	q12, q13, [sp, 80]
+	stp	q14, q15, [sp, 112]
+	stp	q16, q17, [sp, 144]
+	stp	q18, q19, [sp, 176]
+	stp	q20, q21, [sp, 208]
+	stp	q22, q23, [sp, 240]
+
+	// Use per lane load/store to avoid endianness issues.
+	str	q0, [sp, 272]
+	str	q1, [sp, 288]
+	ldr	s0, [sp, 272]
+	ldr	s1, [sp, 288]
+	bl powf
+	str	s0, [sp, 272]
+	ldr	s0, [sp, 276]
+	ldr	s1, [sp, 292]
+	bl powf
+	str	s0, [sp, 276]
+	ldr	s0, [sp, 280]
+	ldr	s1, [sp, 296]
+	bl powf
+	str	s0, [sp, 280]
+	ldr	s0, [sp, 284]
+	ldr	s1, [sp, 300]
+	bl powf
+	str	s0, [sp, 284]
+	ldr	q0, [sp, 272]
+
+	ldp	q8, q9, [sp, 16]
+	ldp	q10, q11, [sp, 48]
+	ldp	q12, q13, [sp, 80]
+	ldp	q14, q15, [sp, 112]
+	ldp	q16, q17, [sp, 144]
+	ldp	q18, q19, [sp, 176]
+	ldp	q20, q21, [sp, 208]
+	ldp	q22, q23, [sp, 240]
+	ldp	x29, x30, [sp], 304
+	cfi_adjust_cfa_offset (304)
+	cfi_restore (x29)
+	cfi_restore (x30)
+	ret
+END (_ZGVnN4vv_powf)
diff --git a/sysdeps/aarch64/fpu/libmvec_float_vlen4_sinf.S b/sysdeps/aarch64/fpu/libmvec_float_vlen4_sinf.S
new file mode 100644
index 0000000000..49b8e95a91
--- /dev/null
+++ b/sysdeps/aarch64/fpu/libmvec_float_vlen4_sinf.S
@@ -0,0 +1,21 @@ 
+/* Single-precision 4 element vector sin function.
+   Copyright (C) 2019 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define SCALAR_FUNCTION sinf
+#define VECTOR_FUNCTION _ZGVnN4v_sinf
+#include "libmvec_float_vlen4.h"
diff --git a/sysdeps/aarch64/fpu/test-double-vlen2-wrappers.c b/sysdeps/aarch64/fpu/test-double-vlen2-wrappers.c
index 6c6c44d6b5..00c5f5bd4b 100644
--- a/sysdeps/aarch64/fpu/test-double-vlen2-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-double-vlen2-wrappers.c
@@ -25,4 +25,16 @@ 
    placing it here happens to work, should be fixed in test-math-vector.h.  */
 __attribute__ ((aarch64_vector_pcs))
 
+VECTOR_WRAPPER (WRAPPER_NAME (cos), _ZGVnN2v_cos)
+
+__attribute__ ((aarch64_vector_pcs))
 VECTOR_WRAPPER (WRAPPER_NAME (exp), _ZGVnN2v_exp)
+
+__attribute__ ((aarch64_vector_pcs))
+VECTOR_WRAPPER (WRAPPER_NAME (log), _ZGVnN2v_log)
+
+__attribute__ ((aarch64_vector_pcs))
+VECTOR_WRAPPER_ff (WRAPPER_NAME (pow), _ZGVnN2vv_pow)
+
+__attribute__ ((aarch64_vector_pcs))
+VECTOR_WRAPPER (WRAPPER_NAME (sin), _ZGVnN2v_sin)
diff --git a/sysdeps/aarch64/fpu/test-float-vlen4-wrappers.c b/sysdeps/aarch64/fpu/test-float-vlen4-wrappers.c
index 5117633f1f..2b9cf6d31f 100644
--- a/sysdeps/aarch64/fpu/test-float-vlen4-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-float-vlen4-wrappers.c
@@ -25,4 +25,16 @@ 
    placing it here happens to work, should be fixed in test-math-vector.h.  */
 __attribute__ ((aarch64_vector_pcs))
 
+VECTOR_WRAPPER (WRAPPER_NAME (cosf), _ZGVnN4v_cosf)
+
+__attribute__ ((aarch64_vector_pcs))
 VECTOR_WRAPPER (WRAPPER_NAME (expf), _ZGVnN4v_expf)
+
+__attribute__ ((aarch64_vector_pcs))
+VECTOR_WRAPPER (WRAPPER_NAME (logf), _ZGVnN4v_logf)
+
+__attribute__ ((aarch64_vector_pcs))
+VECTOR_WRAPPER_ff (WRAPPER_NAME (powf), _ZGVnN4vv_powf)
+
+__attribute__ ((aarch64_vector_pcs))
+VECTOR_WRAPPER (WRAPPER_NAME (sinf), _ZGVnN4v_sinf)
diff --git a/sysdeps/aarch64/libm-test-ulps b/sysdeps/aarch64/libm-test-ulps
index 1ed4af9e55..f83213c48c 100644
--- a/sysdeps/aarch64/libm-test-ulps
+++ b/sysdeps/aarch64/libm-test-ulps
@@ -1043,6 +1043,12 @@  ifloat: 1
 ildouble: 2
 ldouble: 2
 
+Function: "cos_vlen2":
+double: 1
+
+Function: "cos_vlen4":
+float: 1
+
 Function: "cosh":
 double: 1
 float: 1
@@ -1977,6 +1983,12 @@  ifloat: 1
 ildouble: 2
 ldouble: 2
 
+Function: "pow_vlen2":
+double: 1
+
+Function: "pow_vlen4":
+float: 1
+
 Function: "sin":
 double: 1
 float: 1
@@ -2009,6 +2021,12 @@  ifloat: 1
 ildouble: 3
 ldouble: 3
 
+Function: "sin_vlen2":
+double: 1
+
+Function: "sin_vlen4":
+float: 1
+
 Function: "sincos":
 double: 1
 float: 1
diff --git a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
index 9e178253f7..20cc3dcd7f 100644
--- a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
+++ b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
@@ -1,2 +1,10 @@ 
+GLIBC_2.30 _ZGVnN2v_cos F
 GLIBC_2.30 _ZGVnN2v_exp F
+GLIBC_2.30 _ZGVnN2v_log F
+GLIBC_2.30 _ZGVnN2v_sin F
+GLIBC_2.30 _ZGVnN2vv_pow F
+GLIBC_2.30 _ZGVnN4v_cosf F
 GLIBC_2.30 _ZGVnN4v_expf F
+GLIBC_2.30 _ZGVnN4v_logf F
+GLIBC_2.30 _ZGVnN4v_sinf F
+GLIBC_2.30 _ZGVnN4vv_powf F