This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH 9/N v2] [x86_64] Vectorized math functions
- From: Andrew Senkevich <andrew dot n dot senkevich at gmail dot com>
- To: Joseph Myers <joseph at codesourcery dot com>
- Cc: libc-alpha <libc-alpha at sourceware dot org>
- Date: Mon, 8 Dec 2014 22:37:11 +0400
- Subject: Re: [PATCH 9/N v2] [x86_64] Vectorized math functions
- Authentication-results: sourceware.org; auth=none
- References: <CAMXFM3uqG=KhwofMe7w-ToPHL0ur+apMT0sh4W82KYHyXnEEKg at mail dot gmail dot com> <alpine dot DEB dot 2 dot 10 dot 1412031814590 dot 20515 at digraph dot polyomino dot org dot uk>
Here is updated patch with tests.
* math/Makefile: Added rules for tests.
* sysdeps/x86_64/fpu/Makefile: Likewise.
* math/test-double-vlen2.h: New file.
* math/test-double-vlen4.h: New file.
* math/test-vec-loop.h: New file.
* sysdeps/x86_64/fpu/test-double-vlen2.c: New file.
* sysdeps/x86_64/fpu/test-double-vlen4-avx2.c: New file.
* sysdeps/x86_64/fpu/test-double-vlen4.c: New file.
* sysdeps/x86_64/fpu/math-tests-arch.h: AVX2 availability
runtime check set up.
* sysdeps/x86_64/fpu/libm-test-ulps: Regenarated.
diff --git a/math/Makefile b/math/Makefile
index e6285e9..bddd6eb 100644
--- a/math/Makefile
+++ b/math/Makefile
@@ -114,8 +114,9 @@ tests-static = test-fpucw-static test-fpucw-ieee-static
test-longdouble-yes = test-ldouble test-ildoubl
ifneq (no,$(PERL))
+libm-vec-tests = $(addprefix test-,$(libmvec-tests))
libm-tests = test-float test-double $(test-longdouble-$(long-double-fcts)) \
- test-ifloat test-idouble
+ test-ifloat test-idouble $(libm-vec-tests)
libm-tests.o = $(addsuffix .o,$(libm-tests))
tests += $(libm-tests)
@@ -142,8 +143,21 @@ $(objpfx)test-double.o: $(objpfx)libm-test.stmp
$(objpfx)test-idouble.o: $(objpfx)libm-test.stmp
$(objpfx)test-ldouble.o: $(objpfx)libm-test.stmp
$(objpfx)test-ildoubl.o: $(objpfx)libm-test.stmp
+
+$(objpfx)test-double-vlen2.o: $(objpfx)libm-test.stmp
+$(objpfx)test-double-vlen4.o: $(objpfx)libm-test.stmp
+
+$(objpfx)test-double-vlen2: $(libmvec) $(objpfx)init-arch.o
+$(objpfx)test-double-vlen4: $(libmvec) $(objpfx)init-arch.o
endif
+libm-test-fast-math-cflags = -fno-builtin -D__FAST_MATH__ -DTEST_FAST_MATH
+libm-test-vec-cflags = $(libm-test-fast-math-cflags) -fno-inline \
+ -ffloat-store -D_OPENMP=201307 -Wno-unknown-pragmas
+
+CFLAGS-test-double-vlen2.c = $(libm-test-vec-cflags) $(vlen2-arch-ext-cflags)
+CFLAGS-test-double-vlen4.c = $(libm-test-vec-cflags) $(vlen4-arch-ext-cflags)
+
CFLAGS-test-float.c = -fno-inline -ffloat-store -fno-builtin
CFLAGS-test-double.c = -fno-inline -ffloat-store -fno-builtin
CFLAGS-test-ldouble.c = -fno-inline -ffloat-store -fno-builtin
@@ -154,12 +168,13 @@ CFLAGS-test-tgmath.c = -fno-builtin
CFLAGS-test-tgmath2.c = -fno-builtin
CFLAGS-test-tgmath-ret.c = -fno-builtin
CFLAGS-test-powl.c = -fno-builtin
-CPPFLAGS-test-ifloat.c = -U__LIBC_INTERNAL_MATH_INLINES -D__FAST_MATH__ \
- -DTEST_FAST_MATH -fno-builtin
-CPPFLAGS-test-idouble.c = -U__LIBC_INTERNAL_MATH_INLINES -D__FAST_MATH__ \
- -DTEST_FAST_MATH -fno-builtin
-CPPFLAGS-test-ildoubl.c = -U__LIBC_INTERNAL_MATH_INLINES -D__FAST_MATH__ \
- -DTEST_FAST_MATH -fno-builtin
+
+CPPFLAGS-test-ifloat.c = -U__LIBC_INTERNAL_MATH_INLINES \
+ $(libm-test-fast-math-cflags)
+CPPFLAGS-test-idouble.c = -U__LIBC_INTERNAL_MATH_INLINES \
+ $(libm-test-fast-math-cflags)
+CPPFLAGS-test-ildoubl.c = -U__LIBC_INTERNAL_MATH_INLINES \
+ $(libm-test-fast-math-cflags)
# The -lieee module sets the _LIB_VERSION_ switch to IEEE mode
# for error handling in the -lm functions.
diff --git a/math/test-double-vlen2.h b/math/test-double-vlen2.h
new file mode 100644
index 0000000..5d94260
--- /dev/null
+++ b/math/test-double-vlen2.h
@@ -0,0 +1,42 @@
+/* Copyright (C) 2014 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 FLOAT double
+#define FUNC(function) function
+#define TEST_MSG "testing double vector math (without inline functions)\n"
+#define MATHCONST(x) x
+#define CHOOSE(Clongdouble,Cdouble,Cfloat,Cinlinelongdouble,Cinlinedouble,Cinlinefloat)
Cdouble
+#define PRINTF_EXPR "e"
+#define PRINTF_XEXPR "a"
+#define PRINTF_NEXPR "f"
+#define TEST_DOUBLE 1
+#define TEST_MATHVEC 1
+
+#ifndef __NO_MATH_INLINES
+# define __NO_MATH_INLINES
+#endif
+
+#define EXCEPTION_TESTS_double 0
+#define ROUNDING_TESTS_double(MODE) ((MODE) == FE_TONEAREST)
+
+#define VEC_SUFF _vlen2
+
+#define CONCAT(a, b) __CONCAT (a, b)
+
+#define WRAPPER_NAME(function) CONCAT (function, VEC_SUFF)
+
+#define FUNC_TEST(function) function ## _VEC_SUFF
diff --git a/math/test-double-vlen4.h b/math/test-double-vlen4.h
new file mode 100644
index 0000000..e112de1
--- /dev/null
+++ b/math/test-double-vlen4.h
@@ -0,0 +1,40 @@
+/* Copyright (C) 2014 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 FLOAT double
+#define FUNC(function) function
+#define TEST_MSG "testing double vector math (without inline functions)\n"
+#define MATHCONST(x) x
+#define CHOOSE(Clongdouble,Cdouble,Cfloat,Cinlinelongdouble,Cinlinedouble,Cinlinefloat)
Cdouble
+#define PRINTF_EXPR "e"
+#define PRINTF_XEXPR "a"
+#define PRINTF_NEXPR "f"
+#define TEST_DOUBLE 1
+#define TEST_MATHVEC 1
+
+#ifndef __NO_MATH_INLINES
+# define __NO_MATH_INLINES
+#endif
+
+#define EXCEPTION_TESTS_double 0
+#define ROUNDING_TESTS_double(MODE) ((MODE) == FE_TONEAREST)
+
+#define CONCAT(a, b) __CONCAT (a, b)
+
+#define WRAPPER_NAME(function) CONCAT (function, VEC_SUFF)
+
+#define FUNC_TEST(function) function ## _VEC_SUFF
diff --git a/math/test-vec-loop.h b/math/test-vec-loop.h
new file mode 100644
index 0000000..f0c5cf3
--- /dev/null
+++ b/math/test-vec-loop.h
@@ -0,0 +1,31 @@
+/* Copyright (C) 2014 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/>. */
+
+/* This macros is used in VECTOR_WRAPPER macros for vector tests. */
+#define TEST_VEC_LOOP(len) \
+ do \
+ { \
+ for (i = 1; i < len; i++) \
+ { \
+ if (((FLOAT *) &mr)[0] != ((FLOAT *) &mr)[i]) \
+ { \
+ return ((FLOAT *) &mr)[0] + 0.1; \
+ } \
+ } \
+ return ((FLOAT *) &mr)[0]; \
+ } \
+ while (0)
diff --git a/sysdeps/x86_64/fpu/Makefile b/sysdeps/x86_64/fpu/Makefile
index 25fe0d4..973b9a6 100644
--- a/sysdeps/x86_64/fpu/Makefile
+++ b/sysdeps/x86_64/fpu/Makefile
@@ -2,3 +2,19 @@ ifeq ($(subdir),mathvec)
libmvec-support += svml_d_cos2_core svml_d_cos4_core_avx \
svml_d_cos4_core_avx2 svml_d_cos_data
endif
+
+# Rules for libmvec tests.
+ifeq ($(subdir),math)
+ifeq ($(build-mathvec),yes)
+libmvec-tests += double-vlen2 double-vlen4 double-vlen4-avx2
+
+vlen2-arch-ext-cflags = -msse4
+vlen4-arch-ext-cflags = -mavx
+
+$(objpfx)test-double-vlen4-avx2.o: $(objpfx)libm-test.stmp
+
+$(objpfx)test-double-vlen4-avx2: $(libmvec) $(objpfx)init-arch.o
+
+CFLAGS-test-double-vlen4-avx2.c = $(libm-test-vec-cflags) -mavx2
+endif
+endif
diff --git a/sysdeps/x86_64/fpu/libm-test-ulps
b/sysdeps/x86_64/fpu/libm-test-ulps
index 36e1b76..e4de5b4 100644
--- a/sysdeps/x86_64/fpu/libm-test-ulps
+++ b/sysdeps/x86_64/fpu/libm-test-ulps
@@ -905,6 +905,15 @@ idouble: 1
ildouble: 2
ldouble: 2
+Function: "cos_vlen2":
+double: 1
+
+Function: "cos_vlen4_avx":
+double: 1
+
+Function: "cos_vlen4_avx2":
+double: 1
+
Function: "cosh":
double: 1
float: 1
diff --git a/sysdeps/x86_64/fpu/math-tests-arch.h
b/sysdeps/x86_64/fpu/math-tests-arch.h
new file mode 100644
index 0000000..4a357b3
--- /dev/null
+++ b/sysdeps/x86_64/fpu/math-tests-arch.h
@@ -0,0 +1,43 @@
+/* Runtime architecture check for math tests. x86_64 version.
+ Copyright (C) 2013-2014 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/>. */
+
+#ifdef REQUIRE_AVX2
+# include <init-arch.h>
+
+ /* Set to 1 if AVX2 supported. */
+ static int avx2_usable;
+
+# define INIT_ARCH_EXT \
+ do \
+ { \
+ __init_cpu_features (); \
+ avx2_usable = __cpu_features.feature[index_AVX2_Usable] \
+ & bit_AVX2_Usable; \
+ } \
+ while (0)
+
+# define CHECK_ARCH_EXT \
+ do \
+ { \
+ if (!avx2_usable) return; \
+ } \
+ while (0)
+
+#else
+# include <sysdeps/generic/math-tests-arch.h>
+#endif
diff --git a/sysdeps/x86_64/fpu/test-double-vlen2.c
b/sysdeps/x86_64/fpu/test-double-vlen2.c
new file mode 100644
index 0000000..5810795
--- /dev/null
+++ b/sysdeps/x86_64/fpu/test-double-vlen2.c
@@ -0,0 +1,38 @@
+/* Tests for SSE4 ISA versions of vector math functions.
+ Copyright (C) 2014 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 "test-double-vlen2.h"
+#include "test-vec-loop.h"
+#include <immintrin.h>
+
+// Wrapper from scalar to vector function implemented in SSE4.
+#define VECTOR_WRAPPER(scalar_func, vector_func) \
+extern __m128d vector_func (__m128d); \
+FLOAT scalar_func (FLOAT x) \
+{ \
+ int i; \
+ __m128d mx = _mm_set1_pd (x); \
+ __m128d mr = vector_func (mx); \
+ TEST_VEC_LOOP (2); \
+}
+
+VECTOR_WRAPPER (WRAPPER_NAME (cos), _ZGVbN2v_cos)
+
+#define TEST_VECTOR_cos 1
+
+#include "libm-test.c"
diff --git a/sysdeps/x86_64/fpu/test-double-vlen4-avx2.c
b/sysdeps/x86_64/fpu/test-double-vlen4-avx2.c
new file mode 100644
index 0000000..e052c31
--- /dev/null
+++ b/sysdeps/x86_64/fpu/test-double-vlen4-avx2.c
@@ -0,0 +1,42 @@
+/* Tests for AVX2 ISA versions of vector math functions.
+ Copyright (C) 2014 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 "test-double-vlen4.h"
+#include "test-vec-loop.h"
+#include <immintrin.h>
+
+// Wrapper from scalar to vector function implemented in AVX2.
+#define VECTOR_WRAPPER(scalar_func, vector_func) \
+extern __m256d vector_func (__m256d); \
+FLOAT scalar_func (FLOAT x) \
+{ \
+ int i; \
+ __m256d mx = _mm256_set1_pd (x); \
+ __m256d mr = vector_func (mx); \
+ TEST_VEC_LOOP (4); \
+}
+
+#define VEC_SUFF _vlen4_avx2
+
+VECTOR_WRAPPER (WRAPPER_NAME (cos), _ZGVdN4v_cos)
+
+#define TEST_VECTOR_cos 1
+
+#define REQUIRE_AVX2
+
+#include "libm-test.c"
diff --git a/sysdeps/x86_64/fpu/test-double-vlen4.c
b/sysdeps/x86_64/fpu/test-double-vlen4.c
new file mode 100644
index 0000000..7b5462f
--- /dev/null
+++ b/sysdeps/x86_64/fpu/test-double-vlen4.c
@@ -0,0 +1,40 @@
+/* Tests for AVX ISA versions of vector math functions.
+ Copyright (C) 2014 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 "test-double-vlen4.h"
+#include "test-vec-loop.h"
+#include <immintrin.h>
+
+// Wrapper from scalar to vector function implemented in AVX.
+#define VECTOR_WRAPPER(scalar_func, vector_func) \
+extern __m256d vector_func (__m256d); \
+FLOAT scalar_func (FLOAT x) \
+{ \
+ int i; \
+ __m256d mx = _mm256_set1_pd (x); \
+ __m256d mr = vector_func (mx); \
+ TEST_VEC_LOOP (4); \
+}
+
+#define VEC_SUFF _vlen4_avx
+
+VECTOR_WRAPPER (WRAPPER_NAME (cos), _ZGVcN4v_cos)
+
+#define TEST_VECTOR_cos 1
+
+#include "libm-test.c"
--
WBR,
Andrew