Re: [RFC] How to add vector math functions to Glibc

[Andrew Senkevich <andrew dot n dot senkevich at gmail dot com>]

Hi, Joseph,

here is the patch with test suite changes with fixes based on your
previous comments.

>> @@ -6258,7 +6274,11 @@ static const struct test_f_f_data cos_test_data[] =
>>  static void
>>  cos_test (void)
>>  {
>> +#ifndef TEST_MATHVEC
>>    ALL_RM_TEST (cos, 0, cos_test_data, RUN_TEST_LOOP_f_f, END);
>> +#else
>> +  TN_RM_TEST (vector_cos, 0, cos_test_data, RUN_TEST_LOOP_f_f, END);
>> +#endif
>>  }
>
> And I don't think we want conditionals like this for every function -
> indeed, the tests shouldn't need to know which functions have vector
> versions at all.

Do you mean to use the same *_test function for testing vector
(through wrapper)?
I have such scheme now but it requires to add macros named as standard
function and it also caused changes in START macros.

>> +CFLAGS-test-vec-double.c = -fno-inline -ffloat-store -fno-builtin
>> -frounding-math -mavx2 -Wno-unused-function
>
> And since you can't determine at configure
> time what host the tests might run on, instruction set features such as
> AVX need testing for at runtime (this means building separate source files
> for the test with separate options so that you know the compiler won't
> generate AVX code before you've tested for AVX availability).

Because of vector tests grouped by ISA we have different test driver
names containing vector length (test-double-vlen4.c for AVX2).
Scalar wrappers (called from test driver) will be in separate files
(test-double-vlen4-wrapper.c) and will be built with
architecture-specific options specified in sysdeps Makefile.
For runtime check we need to insert condition before wrapper start so
with help of new macros added in *_test function that condition could
be defined in test driver.
Now I have built test-double-vlen4 manually and if this way is ok I
will prepare that sysdeps Makefile.

diff --git a/math/libm-test.inc b/math/libm-test.inc
index f86a4fa..9ddb77e 100644
--- a/math/libm-test.inc
+++ b/math/libm-test.inc
@@ -684,7 +684,7 @@ static void
 test_single_errno (const char *test_name, int errno_value,
    int expected_value, const char *expected_name)
 {
-#ifndef TEST_INLINE
+#if !defined TEST_INLINE && !defined TEST_MATHVEC
   if (errno_value == expected_value)
     {
       if (print_screen (1))
@@ -1691,8 +1691,9 @@ struct test_fFF_11_data
   ROUND_RESTORE_ ## ROUNDING_MODE

 /* Start and end the tests for a given function.  */
-#define START(FUNC, EXACT) \
-  const char *this_func = #FUNC; \
+#define STR_CON(x,y) __STRING(x##y)
+#define START(FUNC, SUFF, EXACT) \
+  const char *this_func = STR_CON (FUNC, SUFF); \
   init_max_error (this_func, EXACT)
 #define END \
   print_max_error (this_func)
@@ -1705,28 +1706,28 @@ struct test_fFF_11_data
     { \
       do \
  { \
-  START (FUNC, EXACT); \
+  START (FUNC, , EXACT); \
   LOOP_MACRO (FUNC, ARRAY, , ## __VA_ARGS__); \
   END_MACRO; \
  } \
       while (0); \
       do \
  { \
-  START (FUNC ## _downward, EXACT); \
+  START (FUNC, _downward, EXACT); \
   LOOP_MACRO (FUNC, ARRAY, FE_DOWNWARD, ## __VA_ARGS__); \
   END_MACRO; \
  } \
       while (0); \
       do \
  { \
-  START (FUNC ## _towardzero, EXACT); \
+  START (FUNC, _towardzero, EXACT); \
   LOOP_MACRO (FUNC, ARRAY, FE_TOWARDZERO, ## __VA_ARGS__); \
   END_MACRO; \
  } \
       while (0); \
       do \
  { \
-  START (FUNC ## _upward, EXACT); \
+  START (FUNC, _upward, EXACT); \
   LOOP_MACRO (FUNC, ARRAY, FE_UPWARD, ## __VA_ARGS__); \
   END_MACRO; \
  } \
@@ -6034,7 +6035,7 @@ static const struct test_c_c_data cexp_test_data[] =
 static void
 cexp_test (void)
 {
-  START (cexp, 0);
+  START (cexp, , 0);
   RUN_TEST_LOOP_c_c (cexp, cexp_test_data, );
   END_COMPLEX;
 }
@@ -6247,7 +6248,7 @@ copysign_test (void)


 static const struct test_f_f_data cos_test_data[] =
-  {
+  {
     TEST_f_f (cos, plus_infty, qnan_value, INVALID_EXCEPTION|ERRNO_EDOM),
     TEST_f_f (cos, minus_infty, qnan_value, INVALID_EXCEPTION|ERRNO_EDOM),
     TEST_f_f (cos, qnan_value, qnan_value,
NO_INEXACT_EXCEPTION|ERRNO_UNCHANGED),
@@ -6255,9 +6256,14 @@ static const struct test_f_f_data cos_test_data[] =
     AUTO_TESTS_f_f (cos),
   };

+#ifndef CHECKARCH
+# define CHECKARCH
+#endif
+
 static void
 cos_test (void)
 {
+  CHECKARCH
   ALL_RM_TEST (cos, 0, cos_test_data, RUN_TEST_LOOP_f_f, END);
 }

@@ -7548,7 +7554,7 @@ static const struct test_if_f_data jn_test_data[] =
 static void
 jn_test (void)
 {
-  START (jn, 0);
+  START (jn, , 0);
   RUN_TEST_LOOP_if_f (jn, jn_test_data, );
   END;
 }
@@ -9374,7 +9380,7 @@ static const struct test_f_f_data tgamma_test_data[] =
 static void
 tgamma_test (void)
 {
-  START (tgamma, 0);
+  START (tgamma, , 0);
   RUN_TEST_LOOP_f_f (tgamma, tgamma_test_data, );
   END;
 }
@@ -9824,6 +9830,12 @@ main (int argc, char **argv)
   initialize ();
   printf (TEST_MSG);

+  /* Vector trigonometric functions:  */
+#ifdef TEST_MATHVEC
+
+  cos_test ();
+
+#else
   check_ulp ();

   /* Keep the tests a wee bit ordered (according to ISO C99).  */
@@ -9960,6 +9972,7 @@ main (int argc, char **argv)
   y0_test ();
   y1_test ();
   yn_test ();
+#endif

   if (output_ulps)
     fclose (ulps_file);

diff --git a/sysdeps/x86_64/fpu/libm-test-ulps
b/sysdeps/x86_64/fpu/libm-test-ulps
index 36e1b76..0e11cd5 100644
--- a/sysdeps/x86_64/fpu/libm-test-ulps
+++ b/sysdeps/x86_64/fpu/libm-test-ulps
@@ -905,6 +905,12 @@ idouble: 1
 ildouble: 2
 ldouble: 2

+
+Function: "vlen4_cos":
+double: 1
+
 Function: "cosh":
 double: 1
 float: 1

diff --git a/sysdeps/x86_64/fpu/test-double-vlen4-wrapper.c
b/sysdeps/x86_64/fpu/test-double-vlen4-wrapper.c
new file mode 100644
index 0000000..35e130e
--- /dev/null
+++ b/sysdeps/x86_64/fpu/test-double-vlen4-wrapper.c
@@ -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
+
+// 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);\
+  for(i=1;i<4;i++)\
+  {\
+    if (((FLOAT*)&mr)[0]!=((FLOAT*)&mr)[i])\
+    {\
+      return ((FLOAT*)&mr)[0]+0.1;\
+    }\
+  }\
+  return ((FLOAT*)&mr)[0];\
+}
+
+#include <immintrin.h>
+
+VECTOR_WRAPPER(vlen4_cos,_ZGVdN4v_cos)

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..ce40c04
--- /dev/null
+++ b/sysdeps/x86_64/fpu/test-double-vlen4.c
@@ -0,0 +1,44 @@
+/* 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 FUNC(function) function
+#define FLOAT double
+#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
+
+#ifndef __NO_MATH_INLINES
+# define __NO_MATH_INLINES
+#endif
+
+#define TEST_MATHVEC
+#define EXCEPTION_TESTS_double 0
+#define ROUNDING_TESTS_double(MODE) ((MODE) == FE_TONEAREST)
+
+#define cos vlen4_cos
+
+#include <init-arch.h>
+
+#define CHECKARCH \
+__init_cpu_features();\
+if (__cpu_features.feature[index_AVX2_Usable] & bit_AVX2_Usable)
+
+#include "libm-test.c"


--
WBR,
Andrew