Add VSX support #195

shibatch · 2018-05-27T15:16:34Z

This patch adds support for POWER VSX.
clang-5.0 and later is supported at this time.
gcc-8 should be able to compile the VSX code, but I haven't checked.

@avmgithub Please check if this works on your environment.

avmgithub · 2018-05-27T16:04:19Z

@shibatch ok , will do, do you which unit test in pytorch exercises sleef ?

shibatch · 2018-05-27T16:09:33Z

@avmgithub I don't fully understand you mean, but please check if ctest in sleef passes.
I only checked with binfmt_misc + qemu, so I'm not confident if this works on other environments.
For PyTorch, I haven't done any test. Do you know what the standard set of tests is in a case like this?

avmgithub · 2018-05-28T12:10:59Z

@shibatch I cloned and checkedout commit ca35e1a . I hope that is the correct one. Here were my steps:
cd sleef
git checkout ca35e1a
mkdir build
cd build
cmake ..
make VERBOSE=1 <<<< this failed stating I need to add -std=c99 . I had to add them into 2 places .
first in line 193 ; set(FLAGS_FASTMATH "-std=c99 -ffast-math") and the other in 236 ; set(SLEEF_C_FLAGS "-std=c99 " "${FLAGS_WALL} ${FLAGS_STRICTMATH} ${FLAGS_NO_ERRNO}") of Configure.cmake.

here is the error:
/home/freddie/projects/sleef/src/libm/mkmasked_gnuabi.c:58:3: error: ‘for’ loop initial declarations are only allowed in C99 mode
for(int i=0;funcList[i].name != NULL;i++) {
^
/home/freddie/projects/sleef/src/libm/mkmasked_gnuabi.c:58:3: note: use option -std=c99 or -std=gnu99 to compile your code

Once I corrected the error above. All built fine. I then ran "make test"
It gave me errors in the last 8 tests (see below)
......
18/26 Test #18: roundtriptest2ddp_4_4 ............***Failed 10.44 sec
Start 19: roundtriptest2ddp_8_8
19/26 Test #19: roundtriptest2ddp_8_8 ............***Failed 0.53 sec
Start 20: roundtriptest2ddp_10_10
20/26 Test #20: roundtriptest2ddp_10_10 ..........***Failed 0.21 sec
Start 21: roundtriptest2ddp_5_15
21/26 Test #21: roundtriptest2ddp_5_15 ...........***Failed 0.30 sec
Start 22: roundtriptest2dsp_2_2
22/26 Test #22: roundtriptest2dsp_2_2 ............ Passed 0.16 sec
Start 23: roundtriptest2dsp_4_4
23/26 Test #23: roundtriptest2dsp_4_4 ............***Failed 8.49 sec
Start 24: roundtriptest2dsp_8_8
24/26 Test #24: roundtriptest2dsp_8_8 ............***Failed 0.50 sec
Start 25: roundtriptest2dsp_10_10
25/26 Test #25: roundtriptest2dsp_10_10 ..........***Failed 0.20 sec
Start 26: roundtriptest2dsp_5_15
26/26 Test #26: roundtriptest2dsp_5_15 ...........***Failed 0.30 sec

69% tests passed, 8 tests failed out of 26

Total Test time (real) = 29.56 sec

The following tests FAILED:
18 - roundtriptest2ddp_4_4 (Failed)
19 - roundtriptest2ddp_8_8 (Failed)
20 - roundtriptest2ddp_10_10 (Failed)
21 - roundtriptest2ddp_5_15 (Failed)
23 - roundtriptest2dsp_4_4 (Failed)
24 - roundtriptest2dsp_8_8 (Failed)
25 - roundtriptest2dsp_10_10 (Failed)
26 - roundtriptest2dsp_5_15 (Failed)
Errors while running CTest
make: *** [test] Error 8

shibatch · 2018-05-28T12:22:11Z

Can I see the output from cmake?

avmgithub · 2018-05-28T12:25:56Z

sure , here it is
$ cmake ..
-- The C compiler identification is GNU 4.8.5
-- Check for working C compiler: /bin/cc
-- Check for working C compiler: /bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Setting build type to 'Release' (required for full support).
-- Looking for sys/types.h
-- Looking for sys/types.h - found
-- Looking for stdint.h
-- Looking for stdint.h - found
-- Looking for stddef.h
-- Looking for stddef.h - found
-- Check size of long double
-- Check size of long double - done
-- Performing Test COMPILER_SUPPORTS_LONG_DOUBLE
-- Performing Test COMPILER_SUPPORTS_LONG_DOUBLE - Success
-- Performing Test COMPILER_SUPPORTS_FLOAT128
-- Performing Test COMPILER_SUPPORTS_FLOAT128 - Failed
-- Performing Test COMPILER_SUPPORTS_SSE2
-- Performing Test COMPILER_SUPPORTS_SSE2 - Failed
-- Performing Test COMPILER_SUPPORTS_SSE4
-- Performing Test COMPILER_SUPPORTS_SSE4 - Failed
-- Performing Test COMPILER_SUPPORTS_AVX
-- Performing Test COMPILER_SUPPORTS_AVX - Failed
-- Performing Test COMPILER_SUPPORTS_FMA4
-- Performing Test COMPILER_SUPPORTS_FMA4 - Failed
-- Performing Test COMPILER_SUPPORTS_AVX2
-- Performing Test COMPILER_SUPPORTS_AVX2 - Failed
-- Performing Test COMPILER_SUPPORTS_SVE
-- Performing Test COMPILER_SUPPORTS_SVE - Failed
-- Performing Test COMPILER_SUPPORTS_AVX512F
-- Performing Test COMPILER_SUPPORTS_AVX512F - Failed
-- Performing Test COMPILER_SUPPORTS_VSX
-- Performing Test COMPILER_SUPPORTS_VSX - Failed
-- Try OpenMP C flag = [-fopenmp]
-- Performing Test OpenMP_FLAG_DETECTED
-- Performing Test OpenMP_FLAG_DETECTED - Success
-- Found OpenMP: -fopenmp
-- Performing Test COMPILER_SUPPORTS_OPENMP
-- Performing Test COMPILER_SUPPORTS_OPENMP - Success
-- Performing Test COMPILER_SUPPORTS_WEAK_ALIASES
-- Performing Test COMPILER_SUPPORTS_WEAK_ALIASES - Success
-- Performing Test COMPILER_SUPPORTS_BUILTIN_MATH
-- Performing Test COMPILER_SUPPORTS_BUILTIN_MATH - Success
-- Unroll target for DP : unroll_0_vecextdp.c;unroll_2_vecextdp.c
-- Unroll target for SP : unroll_0_vecextsp.c;unroll_2_vecextsp.c
-- Configuring build for SLEEF-v3.2
Target system: Linux-3.10.0-693.11.6.el7.ppc64le
Target processor: ppc64le
Host system: Linux-3.10.0-693.11.6.el7.ppc64le
Host processor: ppc64le
Detected C compiler: GNU @ /bin/cc
-- Using option -Wall -Wno-unused -Wno-attributes -Wno-unused-result -Wno-psabi -ffp-contract=off -fno-math-errno -fno-trapping-math to compile libsleef
-- Building shared libs : ON
-- MPFR : LIB_MPFR-NOTFOUND
-- GMP : LIBGMP-NOTFOUND
-- RUNNING_ON_TRAVIS : 0
-- COMPILER_SUPPORTS_OPENMP : 1
-- Configuring done
-- Generating done
-- Build files have been written to: /home/freddie/projects/sleef/build

shibatch · 2018-05-28T12:27:09Z

Your compiler is too old. Try using clang-5 or later, or gcc-8.
VSX intrinsics are fully supported only by recent compilers.

avmgithub · 2018-05-28T12:39:16Z

I'll see what I can do, unfortunately most of ppc64le users use the OS distro gcc. Right now we're supporting RH 7.4 which comes with gcc 4.8. Even if I switch to the latest Advanced Toolkit for ppc64le it only goes to gcc 7.3.1 and that becomes a problem with pytorch as pytorch only seem to support gcc 6.

shibatch · 2018-05-28T12:54:55Z

I see, but I think it's basically impossible to support VSX with gcc 4.8.
We need to wait until newer compilers are supported.

fpetrogalli · 2018-05-29T12:24:52Z

Configure.cmake

@@ -192,7 +208,7 @@ if(CMAKE_C_COMPILER_ID MATCHES "(GNU|Clang)")
    # src/arch/helpervecext.h:88
    string(CONCAT FLAGS_WALL ${FLAGS_WALL} " -Wno-psabi")
    set(FLAGS_ENABLE_NEON32 "-mfpu=neon")
-  endif(CMAKE_C_COMPILER_ID MATCHES "GNU")
+  endif()


It is good practice to enclose the if condition in the endif statement. We don't have a coding format, but i think we should keep it. Please revert this change.

fpetrogalli · 2018-05-29T12:53:17Z

src/dft-tester/fftwtest1d.c

 #include <time.h>
+
+#if defined(UNDEF_USE_EXTERN_INLINES)


This is a workaround specific to this patch. Could you please add a comment as you did in the CMakeLists.txt file?

fpetrogalli · 2018-05-29T12:53:24Z

src/dft-tester/fftwtest2d.c

 #include <time.h>
+
+#if defined(UNDEF_USE_EXTERN_INLINES)


This is a workaround specific to this patch. Could you please add a comment as you did in the CMakeLists.txt file?

fpetrogalli · 2018-05-29T12:53:37Z

src/dft-tester/measuredft.c

 #include <time.h>
 #include <unistd.h>
 #include <sys/time.h>

+#if defined(UNDEF_USE_EXTERN_INLINES)


This is a workaround specific to this patch. Could you please add a comment as you did in the CMakeLists.txt file?

fpetrogalli · 2018-05-29T12:53:44Z

src/dft-tester/naivetest.c

@@ -7,9 +7,17 @@
 #include <stdlib.h>
 #include <stdint.h>
 #include <assert.h>
+#include <time.h>
+
+#if defined(UNDEF_USE_EXTERN_INLINES)


This is a workaround specific to this patch. Could you please add a comment as you did in the CMakeLists.txt file?

fpetrogalli · 2018-05-29T13:00:13Z

src/libm/sleefsimddp.c

@@ -268,7 +277,7 @@ EXPORT CONST vdouble xsin(vdouble d) {
  d = vmla_vd_vd_vd_vd(dqh, vcast_vd_d(-PI_C), d);
  d = vmla_vd_vd_vd_vd(dql, vcast_vd_d(-PI_C), d);
  d = vmla_vd_vd_vd_vd(vadd_vd_vd_vd(dqh, dql), vcast_vd_d(-PI_D), d);
-  
+


Please don't change this line, we don't have a coding standard that enforces to remove lines with spaces.

fpetrogalli · 2018-05-29T13:00:19Z

src/libm/sleefsimddp.c

@@ -289,7 +298,7 @@ EXPORT CONST vdouble xsin(vdouble d) {
 					vor_vo_vo_vo(visnegzero_vo_vd(r),
 						     vgt_vo_vd_vd(vabs_vd_vd(r), vcast_vd_d(TRIGRANGEMAX)))),
 		       vcast_vd_d(-0.0), u);
-  
+


Please don't change this line, we don't have a coding standard that enforces to remove lines with spaces.

fpetrogalli · 2018-05-29T13:08:42Z

src/libm/mkrename.c

-    char *vfloatname = argv[4];
-    char *vintname = argv[5];
-    char *vint2name = argv[6];
+    char *vdoublename = argv[3], *vdoublename_escspace = escapeSpace(vdoublename);


Can you do it the other way around, so that you don't have to change the variable used in the printf calls? We have a patch downstream that adds the AArch64 Vector PCS attribute [1] to the renaming functions (of course, we will upstream it), your changes will you be easier to merge if you leave the original name in the printf calls.

[1] https://developer.arm.com/products/software-development-tools/hpc/arm-compiler-for-hpc/vector-function-abi

Not all the arguments of printf calls are space-escaped.
I don't think of a good way of fulfilling your request.

Let me put it this way: how is it possible that argv[3] has spaces in it? Wouldn't a space break the argument in two arguments?

In altivec or VSX intrinsics, vector data types are like "vector float".

https://gcc.gnu.org/onlinedocs/gcc/PowerPC-AltiVec_002fVSX-Built-in-Functions.html
http://openpowerfoundation.org/wp-content/uploads/resources/leabi/content/dbdoclet.50655245_56497.html

Got it, thank you. Please add your useful answer as a comment to the definition of the function escapeSpace.

h
Hi @shibatch , sorry for bothering you again with this. I had a second (or maybe third!!) thought, and if my understanding is correctly, the following variables don't have space in their names, unless when targeting VSX:

vdoublename vfloatname vintname vint2name

Given that the function escapeSpace do not modify the values of those variables when not targeting VSX, I'd rather have you redefine them as follows:

char *vdoublename = escapeSpace(argv[3]); char *vfloatname = escapeSpace(argv[4]); char *vintname = escapeSpace(argv[5]); char *vint2name = escapeSpace(argv[6]);

This should keep the functionality of the current patch, without the need to modify the variables name in the printf calls.

fpetrogalli · 2018-05-29T13:12:59Z

Configure.cmake

+
+# When cross compiling for ppc64, this bug-workaround is needed
+if(CMAKE_CROSSCOMPILING AND CMAKE_SYSTEM_PROCESSOR MATCHES "^(powerpc|ppc)64")
+  set(COMMON_TARGET_DEFINITIONS ${COMMON_TARGET_DEFINITIONS} UNDEF_USE_EXTERN_INLINES=1)


I have commented all the places where you use this define, asking to add a comment saying that the use of UNDEF_USE_EXTERN_INLINES is platform specific.

I was thinking, instead of adding the comments, how about making the variable names specific to PPC64, so there is not doubt it target specific? May I ask you to change it to POWER64_UNDEF_USE_EXTERN_INLINES? With this name change, I am happy for you to avoid adding the comments all around.

fpetrogalli · 2018-05-29T13:16:05Z

CHANGELOG.md

@@ -15,6 +15,8 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 - 3.5-ULP versions of sinh, cosh, tanh, sinhf, coshf, tanhf, and the
  corresponding testing functionalities are added.
  https://github.com/shibatch/sleef/pull/192
+- Power VSX target support is added to libm.


Love this! :)

UPDATE: I just realized the changelog should say libsleef, not libm.

fpetrogalli · 2018-05-30T14:24:52Z

src/libm/mkrename.c

-    char *vfloatname = argv[4];
-    char *vintname = argv[5];
-    char *vint2name = argv[6];
+    char *vdoublename = argv[3], *vdoublename_escspace = escapeSpace(vdoublename);


Got it, thank you. Please add your useful answer as a comment to the definition of the function escapeSpace.

fpetrogalli · 2018-05-30T18:51:12Z

src/libm/mkrename.c

-    char *vfloatname = argv[4];
-    char *vintname = argv[5];
-    char *vint2name = argv[6];
+    char *vdoublename = argv[3], *vdoublename_escspace = escapeSpace(vdoublename);


h
Hi @shibatch , sorry for bothering you again with this. I had a second (or maybe third!!) thought, and if my understanding is correctly, the following variables don't have space in their names, unless when targeting VSX:

vdoublename vfloatname vintname vint2name

Given that the function escapeSpace do not modify the values of those variables when not targeting VSX, I'd rather have you redefine them as follows:

char *vdoublename = escapeSpace(argv[3]); char *vfloatname = escapeSpace(argv[4]); char *vintname = escapeSpace(argv[5]); char *vint2name = escapeSpace(argv[6]);

This should keep the functionality of the current patch, without the need to modify the variables name in the printf calls.

fpetrogalli · 2018-05-30T18:57:19Z

src/libm/mkrename.c

@@ -10,6 +10,13 @@

 #include "funcproto.h"

+char *escapeSpace(char *str) {
+  char *ret = malloc(strlen(str) + 10);


This memory allocation is never released. you should modify the content of str in place, and definitely not return ret.

I think this is harmless, but I added free for these variables.

shibatch · 2018-05-30T19:21:57Z

The contents of vdoublename and other variables are used for its own type name and part of typedef name. So, we need both the original one and the replaced one. They cannot be unified.

fpetrogalli · 2018-05-30T20:08:55Z

The contents of vdoublename and other variables are used for its own type name and part of typedef name. So, we need both the original one and the replaced one. They cannot be unified.

@shibatch Sigh - apologies, now I have seen it. Sorry for keeping you on this.

fpetrogalli · 2018-05-30T20:33:36Z

Thank you for adding this new target @shibatch! This shows all the goodness of SLEEF!

shibatch added 22 commits May 27, 2018 20:43

no message

f8ccc52

no message

08317d6

no message

c2ed0b3

no message

2fdca8e

no message

ac5e692

no message

ec480aa

no message

0cb37d5

no message

cb2d22b

no message

3da03cc

no message

30cd57c

no message

024d10f

no message

95a2362

no message

46f6ab1

no message

9cf98ac

no message

4fc7063

no message

b40081c

no message

d7c1457

no message

50294ce

no message

92b4b74

no message

3224a1d

no message

306473f

no message

f9928a5

shibatch requested a review from fpetrogalli May 27, 2018 15:17

shibatch added 2 commits May 28, 2018 07:05

no message

63ba8c8

no message

ca35e1a

shibatch added 3 commits May 28, 2018 22:27

no message

a5775c0

no message

f3c36ed

no message

d639a59

fpetrogalli requested changes May 29, 2018

View reviewed changes

shibatch added 3 commits May 30, 2018 18:46

no message

84e0389

no message

032e572

Updated CHANGELOG.

7b0e398

fpetrogalli requested changes May 30, 2018

View reviewed changes

shibatch added 2 commits May 30, 2018 23:41

Added comment to mkrename.c

13a6656

Fix detection code for VSX compiler support

12dc81f

fpetrogalli requested changes May 30, 2018

View reviewed changes

Free malloced memory

1d5b9ba

fpetrogalli approved these changes May 30, 2018

View reviewed changes

shibatch merged commit 41b62dd into master May 31, 2018

shibatch deleted the Add_VSX_Support branch May 31, 2018 00:33

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add VSX support #195

Add VSX support #195

shibatch commented May 27, 2018 •

edited

Loading

avmgithub commented May 27, 2018

shibatch commented May 27, 2018

avmgithub commented May 28, 2018

shibatch commented May 28, 2018

avmgithub commented May 28, 2018

shibatch commented May 28, 2018

avmgithub commented May 28, 2018

shibatch commented May 28, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018

shibatch May 30, 2018

fpetrogalli May 30, 2018

shibatch May 30, 2018

fpetrogalli May 30, 2018

fpetrogalli May 30, 2018

fpetrogalli May 29, 2018

fpetrogalli May 29, 2018 •

edited

Loading

fpetrogalli May 30, 2018

fpetrogalli May 30, 2018

fpetrogalli May 30, 2018

shibatch May 30, 2018

shibatch commented May 30, 2018

fpetrogalli commented May 30, 2018

fpetrogalli commented May 30, 2018

Add VSX support #195

Add VSX support #195

Conversation

shibatch commented May 27, 2018 • edited Loading

avmgithub commented May 27, 2018

shibatch commented May 27, 2018

avmgithub commented May 28, 2018

shibatch commented May 28, 2018

avmgithub commented May 28, 2018

shibatch commented May 28, 2018

avmgithub commented May 28, 2018

shibatch commented May 28, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

fpetrogalli May 29, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

shibatch commented May 30, 2018

fpetrogalli commented May 30, 2018

fpetrogalli commented May 30, 2018

shibatch commented May 27, 2018 •

edited

Loading

fpetrogalli May 29, 2018 •

edited

Loading