OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Merge commit 'origin/x86' into loongson3a

tags/v0.1alpha1
Xianyi Zhang 14 years ago
parent
6eb02bbb9c ce78abe37e
commit
88cbfcc5b5
11 changed files with 168 additions and 113 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +5
    -0
      Changelog.txt
  2. +9
    -1
      Makefile
  3. +3
    -0
      Makefile.rule
  4. +8
    -8
      exports/Makefile
  5. +4
    -1
      exports/gensymbol
  6. +15
    -6
      interface/Makefile
  7. +3
    -0
      kernel/Makefile.LA
  8. +99
    -97
      kernel/x86/zdot_sse2.S
  9. +1
    -0
      utest/common_utest.h
  10. +2
    -0
      utest/main.c
  11. +19
    -0
      utest/test_dotu.c

+ 5
- 0
Changelog.txt View File

@@ -4,6 +4,8 @@ Version 0.1 (in development)
26-Feb-2011

common:
* Support "make NO_LAPACK=1" to build the library without
LAPACK functions.
* Fixed randomly SEGFAULT when nodemask==NULL with above Linux 2.6.34.
Thank Mr.Ei-ji Nakama providing this patch. (Refs issue #12 on github)
* Added DEBUG=1 rule in Makefile.rule to build debug version.
@@ -13,6 +15,9 @@ common:
* Imported GotoBLAS2 1.13 BSD version

x86/x86 64:
* On x86 32bits, gcc 4.4.3 generated wrong codes (movsd) from movlps
in zdot_sse2.S line 191. This would casue zdotu & zdotc failures.
Instead,Walk around it. (Refs issue #8 #9 on github)
* Modified ?axpy functions to return same netlib BLAS results
when incx==0 or incy==0 (Refs issue #7 on github)
* Modified ?swap functions to return same netlib BLAS results


+ 9
- 1
Makefile View File

@@ -15,7 +15,10 @@ ifdef SANITY_CHECK
BLASDIRS += reference
endif

SUBDIRS = $(BLASDIRS) lapack
SUBDIRS = $(BLASDIRS)
ifneq ($(NO_LAPACK), 1)
SUBDIRS += lapack
endif

SUBDIRS_ALL = $(SUBDIRS) test ctest utest exports benchmark ../laswp ../bench

@@ -149,10 +152,15 @@ hpl_p :
fi; \
done

ifeq ($(NO_LAPACK), 1)
netlib :

else
netlib : lapack-3.1.1 patch.for_lapack-3.1.1 lapack-3.1.1/make.inc
ifndef NOFORTRAN
-@$(MAKE) -C lapack-3.1.1 lapacklib
endif
endif

prof_lapack : lapack-3.1.1 lapack-3.1.1/make.inc
-@$(MAKE) -C lapack-3.1.1 lapack_prof


+ 3
- 0
Makefile.rule View File

@@ -43,6 +43,9 @@ VERSION = 0.1
# If you don't need CBLAS interface, please comment it in.
# NO_CBLAS = 1

# If you don't need LAPACK, please comment it in.
# NO_LAPACK = 1

# If you want to use legacy threaded Level 3 implementation.
# USE_SIMPLE_THREADED_LEVEL3 = 1



+ 8
- 8
exports/Makefile View File

@@ -75,13 +75,13 @@ libgoto2_shared.dll : ../$(LIBNAME) libgoto2_shared.def
-Wl,--out-implib,libgoto2_shared.lib $(FEXTRALIB)

libgoto2.def : gensymbol
perl ./gensymbol win2k $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) > $(@F)
perl ./gensymbol win2k $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > $(@F)

libgoto2_shared.def : gensymbol
perl ./gensymbol win2k $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) > $(@F)
perl ./gensymbol win2k $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > $(@F)

libgoto_hpl.def : gensymbol
perl ./gensymbol win2khpl $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) > $(@F)
perl ./gensymbol win2khpl $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > $(@F)

$(LIBDYNNAME) : ../$(LIBNAME) osx.def
$(PREFIX)gcc $(CFLAGS) -all_load -dynamiclib -o $(LIBDYNNAME) $< -Wl,-exported_symbols_list,osx.def $(FEXTRALIB)
@@ -162,23 +162,23 @@ static : ../$(LIBNAME)
rm -f goto.$(SUFFIX)

linux.def : gensymbol ../Makefile.system ../getarch.c
perl ./gensymbol linux $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) > $(@F)
perl ./gensymbol linux $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > $(@F)

osx.def : gensymbol ../Makefile.system ../getarch.c
perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) > $(@F)
perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > $(@F)

aix.def : gensymbol ../Makefile.system ../getarch.c
perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) > $(@F)
perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > $(@F)

symbol.S : gensymbol
perl ./gensymbol win2kasm noarch dummy $(EXPRECISION) $(NO_CBLAS) > symbol.S
perl ./gensymbol win2kasm noarch dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > symbol.S

test : linktest.c
$(CC) $(CFLAGS) -w -o linktest linktest.c ../$(LIBSONAME) -lm && echo OK.
rm -f linktest

linktest.c : gensymbol ../Makefile.system ../getarch.c
perl ./gensymbol linktest $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) > linktest.c
perl ./gensymbol linktest $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) > linktest.c

clean ::
@rm -f *.def *.dylib __.SYMDEF*


+ 4
- 1
exports/gensymbol View File

@@ -298,7 +298,10 @@
zcgesv, zlag2c, clag2z,
);

if (-d "../lapack-3.1.1") {
if ($ARGV[5] == 1) {
#NO_LAPACK=1
@objs = (@blasobjs);
} elsif (-d "../lapack-3.1.1") {
@objs = (@blasobjs, @lapackobjs, @lapackobjs2);
} else {
@objs = (@blasobjs, @lapackobjs);


+ 15
- 6
interface/Makefile View File

@@ -342,36 +342,45 @@ CBLASOBJS = $(CBLAS1OBJS) $(CBLAS2OBJS) $(CBLAS3OBJS)
ZBLASOBJS = $(ZBLAS1OBJS) $(ZBLAS2OBJS) $(ZBLAS3OBJS)
XBLASOBJS = $(XBLAS1OBJS) $(XBLAS2OBJS) $(XBLAS3OBJS)

SBLASOBJS += \
SLAPACKOBJS = \
sgetf2.$(SUFFIX) sgetrf.$(SUFFIX) slauu2.$(SUFFIX) slauum.$(SUFFIX) \
spotf2.$(SUFFIX) spotrf.$(SUFFIX) strti2.$(SUFFIX) strtri.$(SUFFIX) \
slaswp.$(SUFFIX) sgetrs.$(SUFFIX) sgesv.$(SUFFIX) spotri.$(SUFFIX) \

DBLASOBJS += \
DLAPACKOBJS = \
dgetf2.$(SUFFIX) dgetrf.$(SUFFIX) dlauu2.$(SUFFIX) dlauum.$(SUFFIX) \
dpotf2.$(SUFFIX) dpotrf.$(SUFFIX) dtrti2.$(SUFFIX) dtrtri.$(SUFFIX) \
dlaswp.$(SUFFIX) dgetrs.$(SUFFIX) dgesv.$(SUFFIX) dpotri.$(SUFFIX) \

QBLASOBJS += \
QLAPACKOBJS = \
qgetf2.$(SUFFIX) qgetrf.$(SUFFIX) qlauu2.$(SUFFIX) qlauum.$(SUFFIX) \
qpotf2.$(SUFFIX) qpotrf.$(SUFFIX) qtrti2.$(SUFFIX) qtrtri.$(SUFFIX) \
qlaswp.$(SUFFIX) qgetrs.$(SUFFIX) qgesv.$(SUFFIX) qpotri.$(SUFFIX) \

CBLASOBJS += \
CLAPACKOBJS = \
cgetf2.$(SUFFIX) cgetrf.$(SUFFIX) clauu2.$(SUFFIX) clauum.$(SUFFIX) \
cpotf2.$(SUFFIX) cpotrf.$(SUFFIX) ctrti2.$(SUFFIX) ctrtri.$(SUFFIX) \
claswp.$(SUFFIX) cgetrs.$(SUFFIX) cgesv.$(SUFFIX) cpotri.$(SUFFIX) \

ZBLASOBJS += \
ZLAPACKOBJS = \
zgetf2.$(SUFFIX) zgetrf.$(SUFFIX) zlauu2.$(SUFFIX) zlauum.$(SUFFIX) \
zpotf2.$(SUFFIX) zpotrf.$(SUFFIX) ztrti2.$(SUFFIX) ztrtri.$(SUFFIX) \
zlaswp.$(SUFFIX) zgetrs.$(SUFFIX) zgesv.$(SUFFIX) zpotri.$(SUFFIX) \

XBLASOBJS += \
XLAPACKOBJS = \
xgetf2.$(SUFFIX) xgetrf.$(SUFFIX) xlauu2.$(SUFFIX) xlauum.$(SUFFIX) \
xpotf2.$(SUFFIX) xpotrf.$(SUFFIX) xtrti2.$(SUFFIX) xtrtri.$(SUFFIX) \
xlaswp.$(SUFFIX) xgetrs.$(SUFFIX) xgesv.$(SUFFIX) xpotri.$(SUFFIX) \

ifneq ($(NO_LAPACK), 1)
SBLASOBJS += $(SLAPACKOBJS)
DBLASOBJS += $(DLAPACKOBJS)
QBLASOBJS += $(QLAPACKOBJS)
CBLASOBJS += $(CLAPACKOBJS)
ZBLASOBJS += $(ZLAPACKOBJS)
XBLASOBJS += $(XLAPACKOBJS)

endif

FUNCOBJS = $(SBLASOBJS) $(DBLASOBJS) $(CBLASOBJS) $(ZBLASOBJS)



+ 3
- 0
kernel/Makefile.LA View File

@@ -1,3 +1,4 @@
ifneq ($(NO_LAPACK), 1)
SBLASOBJS += sneg_tcopy$(TSUFFIX).$(SUFFIX) slaswp_ncopy$(TSUFFIX).$(SUFFIX)

DBLASOBJS += dneg_tcopy$(TSUFFIX).$(SUFFIX) dlaswp_ncopy$(TSUFFIX).$(SUFFIX)
@@ -10,6 +11,8 @@ ZBLASOBJS += zneg_tcopy$(TSUFFIX).$(SUFFIX) zlaswp_ncopy$(TSUFFIX).$(SUFFIX)

XBLASOBJS += xneg_tcopy$(TSUFFIX).$(SUFFIX) xlaswp_ncopy$(TSUFFIX).$(SUFFIX)

endif

$(KDIR)sneg_tcopy$(TSUFFIX).$(SUFFIX) $(KDIR)sneg_tcopy$(TSUFFIX).$(PSUFFIX) : generic/neg_tcopy_$(SGEMM_UNROLL_M).c
$(CC) -c $(CFLAGS) $< -o $@



+ 99
- 97
kernel/x86/zdot_sse2.S View File

@@ -60,7 +60,9 @@
#undef movsd

#ifndef OPTERON
#define movlps movsd
#define MOVLPS movsd
#else
#define MOVLPS movlps
#endif

PROLOGUE
@@ -351,11 +353,11 @@
sarl $3, %eax
jle .L25

movlps -16 * SIZE(X), %xmm4
MOVLPS -16 * SIZE(X), %xmm4
movhps -15 * SIZE(X), %xmm4
movaps -16 * SIZE(Y), %xmm6

movlps -14 * SIZE(X), %xmm5
MOVLPS -14 * SIZE(X), %xmm5
movhps -13 * SIZE(X), %xmm5
movaps -14 * SIZE(Y), %xmm7

@@ -373,7 +375,7 @@
addpd %xmm6, %xmm0
movaps -12 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps -12 * SIZE(X), %xmm4
MOVLPS -12 * SIZE(X), %xmm4
movhps -11 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -382,7 +384,7 @@
addpd %xmm7, %xmm0
movaps -10 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps -10 * SIZE(X), %xmm5
MOVLPS -10 * SIZE(X), %xmm5
movhps -9 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -395,7 +397,7 @@
addpd %xmm6, %xmm0
movaps -8 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps -8 * SIZE(X), %xmm4
MOVLPS -8 * SIZE(X), %xmm4
movhps -7 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -404,7 +406,7 @@
addpd %xmm7, %xmm0
movaps -6 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps -6 * SIZE(X), %xmm5
MOVLPS -6 * SIZE(X), %xmm5
movhps -5 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -417,7 +419,7 @@
addpd %xmm6, %xmm0
movaps -4 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps -4 * SIZE(X), %xmm4
MOVLPS -4 * SIZE(X), %xmm4
movhps -3 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -426,7 +428,7 @@
addpd %xmm7, %xmm0
movaps -2 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps -2 * SIZE(X), %xmm5
MOVLPS -2 * SIZE(X), %xmm5
movhps -1 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -439,7 +441,7 @@
addpd %xmm6, %xmm0
movaps 0 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -448,7 +450,7 @@
addpd %xmm7, %xmm0
movaps 2 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps 2 * SIZE(X), %xmm5
MOVLPS 2 * SIZE(X), %xmm5
movhps 3 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -465,7 +467,7 @@
addpd %xmm6, %xmm0
movaps -12 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps -12 * SIZE(X), %xmm4
MOVLPS -12 * SIZE(X), %xmm4
movhps -11 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -474,7 +476,7 @@
addpd %xmm7, %xmm0
movaps -10 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps -10 * SIZE(X), %xmm5
MOVLPS -10 * SIZE(X), %xmm5
movhps -9 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -483,7 +485,7 @@
addpd %xmm6, %xmm0
movaps -8 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps -8 * SIZE(X), %xmm4
MOVLPS -8 * SIZE(X), %xmm4
movhps -7 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -492,7 +494,7 @@
addpd %xmm7, %xmm0
movaps -6 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps -6 * SIZE(X), %xmm5
MOVLPS -6 * SIZE(X), %xmm5
movhps -5 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -501,7 +503,7 @@
addpd %xmm6, %xmm0
movaps -4 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps -4 * SIZE(X), %xmm4
MOVLPS -4 * SIZE(X), %xmm4
movhps -3 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -510,7 +512,7 @@
addpd %xmm7, %xmm0
movaps -2 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps -2 * SIZE(X), %xmm5
MOVLPS -2 * SIZE(X), %xmm5
movhps -1 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -534,11 +536,11 @@
testl $4, N
jle .L26

movlps -16 * SIZE(X), %xmm4
MOVLPS -16 * SIZE(X), %xmm4
movhps -15 * SIZE(X), %xmm4
movaps -16 * SIZE(Y), %xmm6

movlps -14 * SIZE(X), %xmm5
MOVLPS -14 * SIZE(X), %xmm5
movhps -13 * SIZE(X), %xmm5
movaps -14 * SIZE(Y), %xmm7

@@ -547,7 +549,7 @@
addpd %xmm6, %xmm0
movaps -12 * SIZE(Y), %xmm6
mulpd %xmm4, %xmm3
movlps -12 * SIZE(X), %xmm4
MOVLPS -12 * SIZE(X), %xmm4
movhps -11 * SIZE(X), %xmm4
addpd %xmm3, %xmm1

@@ -556,7 +558,7 @@
addpd %xmm7, %xmm0
movaps -10 * SIZE(Y), %xmm7
mulpd %xmm5, %xmm3
movlps -10 * SIZE(X), %xmm5
MOVLPS -10 * SIZE(X), %xmm5
movhps -9 * SIZE(X), %xmm5
addpd %xmm3, %xmm1

@@ -580,7 +582,7 @@
testl $2, N
jle .L27

movlps -16 * SIZE(X), %xmm4
MOVLPS -16 * SIZE(X), %xmm4
movhps -15 * SIZE(X), %xmm4
movaps -16 * SIZE(Y), %xmm6

@@ -590,7 +592,7 @@
mulpd %xmm4, %xmm3
addpd %xmm3, %xmm1

movlps -14 * SIZE(X), %xmm5
MOVLPS -14 * SIZE(X), %xmm5
movhps -13 * SIZE(X), %xmm5
movaps -14 * SIZE(Y), %xmm7

@@ -608,7 +610,7 @@
testl $1, N
jle .L98

movlps -16 * SIZE(X), %xmm4
MOVLPS -16 * SIZE(X), %xmm4
movhps -15 * SIZE(X), %xmm4
movaps -16 * SIZE(Y), %xmm6

@@ -628,11 +630,11 @@
sarl $3, %eax
jle .L35

movlps -16 * SIZE(Y), %xmm4
MOVLPS -16 * SIZE(Y), %xmm4
movhps -15 * SIZE(Y), %xmm4
movaps -16 * SIZE(X), %xmm6

movlps -14 * SIZE(Y), %xmm5
MOVLPS -14 * SIZE(Y), %xmm5
movhps -13 * SIZE(Y), %xmm5
movaps -14 * SIZE(X), %xmm7

@@ -650,7 +652,7 @@
addpd %xmm6, %xmm0
movaps -12 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps -12 * SIZE(Y), %xmm4
MOVLPS -12 * SIZE(Y), %xmm4
movhps -11 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -659,7 +661,7 @@
addpd %xmm7, %xmm0
movaps -10 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps -10 * SIZE(Y), %xmm5
MOVLPS -10 * SIZE(Y), %xmm5
movhps -9 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -671,7 +673,7 @@
addpd %xmm6, %xmm0
movaps -8 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps -8 * SIZE(Y), %xmm4
MOVLPS -8 * SIZE(Y), %xmm4
movhps -7 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -680,7 +682,7 @@
addpd %xmm7, %xmm0
movaps -6 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps -6 * SIZE(Y), %xmm5
MOVLPS -6 * SIZE(Y), %xmm5
movhps -5 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -693,7 +695,7 @@
addpd %xmm6, %xmm0
movaps -4 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps -4 * SIZE(Y), %xmm4
MOVLPS -4 * SIZE(Y), %xmm4
movhps -3 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -702,7 +704,7 @@
addpd %xmm7, %xmm0
movaps -2 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps -2 * SIZE(Y), %xmm5
MOVLPS -2 * SIZE(Y), %xmm5
movhps -1 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -715,7 +717,7 @@
addpd %xmm6, %xmm0
movaps 0 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps 0 * SIZE(Y), %xmm4
MOVLPS 0 * SIZE(Y), %xmm4
movhps 1 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -724,7 +726,7 @@
addpd %xmm7, %xmm0
movaps 2 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps 2 * SIZE(Y), %xmm5
MOVLPS 2 * SIZE(Y), %xmm5
movhps 3 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -741,7 +743,7 @@
addpd %xmm6, %xmm0
movaps -12 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps -12 * SIZE(Y), %xmm4
MOVLPS -12 * SIZE(Y), %xmm4
movhps -11 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -750,7 +752,7 @@
addpd %xmm7, %xmm0
movaps -10 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps -10 * SIZE(Y), %xmm5
MOVLPS -10 * SIZE(Y), %xmm5
movhps -9 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -759,7 +761,7 @@
addpd %xmm6, %xmm0
movaps -8 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps -8 * SIZE(Y), %xmm4
MOVLPS -8 * SIZE(Y), %xmm4
movhps -7 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -768,7 +770,7 @@
addpd %xmm7, %xmm0
movaps -6 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps -6 * SIZE(Y), %xmm5
MOVLPS -6 * SIZE(Y), %xmm5
movhps -5 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -777,7 +779,7 @@
addpd %xmm6, %xmm0
movaps -4 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps -4 * SIZE(Y), %xmm4
MOVLPS -4 * SIZE(Y), %xmm4
movhps -3 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -786,7 +788,7 @@
addpd %xmm7, %xmm0
movaps -2 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps -2 * SIZE(Y), %xmm5
MOVLPS -2 * SIZE(Y), %xmm5
movhps -1 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -810,11 +812,11 @@
testl $4, N
jle .L36

movlps -16 * SIZE(Y), %xmm4
MOVLPS -16 * SIZE(Y), %xmm4
movhps -15 * SIZE(Y), %xmm4
movaps -16 * SIZE(X), %xmm6

movlps -14 * SIZE(Y), %xmm5
MOVLPS -14 * SIZE(Y), %xmm5
movhps -13 * SIZE(Y), %xmm5
movaps -14 * SIZE(X), %xmm7

@@ -823,7 +825,7 @@
addpd %xmm6, %xmm0
movaps -12 * SIZE(X), %xmm6
mulpd %xmm4, %xmm3
movlps -12 * SIZE(Y), %xmm4
MOVLPS -12 * SIZE(Y), %xmm4
movhps -11 * SIZE(Y), %xmm4
addpd %xmm3, %xmm1

@@ -832,7 +834,7 @@
addpd %xmm7, %xmm0
movaps -10 * SIZE(X), %xmm7
mulpd %xmm5, %xmm3
movlps -10 * SIZE(Y), %xmm5
MOVLPS -10 * SIZE(Y), %xmm5
movhps -9 * SIZE(Y), %xmm5
addpd %xmm3, %xmm1

@@ -856,7 +858,7 @@
testl $2, N
jle .L37

movlps -16 * SIZE(Y), %xmm4
MOVLPS -16 * SIZE(Y), %xmm4
movhps -15 * SIZE(Y), %xmm4
movaps -16 * SIZE(X), %xmm6

@@ -866,7 +868,7 @@
mulpd %xmm4, %xmm3
addpd %xmm3, %xmm1

movlps -14 * SIZE(Y), %xmm5
MOVLPS -14 * SIZE(Y), %xmm5
movhps -13 * SIZE(Y), %xmm5
movaps -14 * SIZE(X), %xmm7

@@ -887,7 +889,7 @@
testl $1, N
jle .L98

movlps -16 * SIZE(Y), %xmm4
MOVLPS -16 * SIZE(Y), %xmm4
movhps -15 * SIZE(Y), %xmm4
movaps -16 * SIZE(X), %xmm6

@@ -1188,8 +1190,8 @@
testl $1, N
jle .L48

movlps -16 * SIZE(X), %xmm4
movlps -16 * SIZE(Y), %xmm6
movlpd -16 * SIZE(X), %xmm4
movlpd -16 * SIZE(Y), %xmm6

pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
@@ -1211,17 +1213,17 @@
sarl $3, %eax
jle .L55

movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y

movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y

@@ -1233,11 +1235,11 @@
pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1245,11 +1247,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1257,11 +1259,11 @@
pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1269,11 +1271,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1281,11 +1283,11 @@
pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1293,11 +1295,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1305,11 +1307,11 @@
pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1317,11 +1319,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1334,11 +1336,11 @@
pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1346,11 +1348,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1358,11 +1360,11 @@
pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1370,11 +1372,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1382,11 +1384,11 @@
pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1394,11 +1396,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1420,28 +1422,28 @@
testl $4, N
jle .L56

movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y

movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y

pshufd $0x4e, %xmm6, %xmm3
mulpd %xmm4, %xmm6
addpd %xmm6, %xmm0
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y
mulpd %xmm4, %xmm3
movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
addpd %xmm3, %xmm1
@@ -1449,11 +1451,11 @@
pshufd $0x4e, %xmm7, %xmm3
mulpd %xmm5, %xmm7
addpd %xmm7, %xmm0
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y
mulpd %xmm5, %xmm3
movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
addpd %xmm3, %xmm1
@@ -1475,10 +1477,10 @@
testl $2, N
jle .L57

movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
addl INCX, X
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6
addl INCY, Y

@@ -1488,10 +1490,10 @@
mulpd %xmm4, %xmm3
addpd %xmm3, %xmm1

movlps 0 * SIZE(X), %xmm5
MOVLPS 0 * SIZE(X), %xmm5
movhps 1 * SIZE(X), %xmm5
addl INCX, X
movlps 0 * SIZE(Y), %xmm7
MOVLPS 0 * SIZE(Y), %xmm7
movhps 1 * SIZE(Y), %xmm7
addl INCY, Y

@@ -1506,9 +1508,9 @@
testl $1, N
jle .L98

movlps 0 * SIZE(X), %xmm4
MOVLPS 0 * SIZE(X), %xmm4
movhps 1 * SIZE(X), %xmm4
movlps 0 * SIZE(Y), %xmm6
MOVLPS 0 * SIZE(Y), %xmm6
movhps 1 * SIZE(Y), %xmm6

pshufd $0x4e, %xmm6, %xmm3
@@ -1533,8 +1535,8 @@
.L999:
movl RESULT, %eax

movlps %xmm0, 0 * SIZE(%eax)
movlps %xmm1, 1 * SIZE(%eax)
MOVLPS %xmm0, 0 * SIZE(%eax)
MOVLPS %xmm1, 1 * SIZE(%eax)

popl %ebx
popl %esi


+ 1
- 0
utest/common_utest.h View File

@@ -55,5 +55,6 @@ void test_saxpy_inc_0(void);
void test_caxpy_inc_0(void);

void test_zdotu_n_1(void);
void test_zdotu_offset_1(void);

#endif

+ 2
- 0
utest/main.c View File

@@ -53,6 +53,8 @@ CU_TestInfo test_level1[]={
{"Testing zaxpy with incx || incy == 0",test_zaxpy_inc_0},

{"Testing zdotu with n == 1",test_zdotu_n_1},
{"Testing zdotu with input x & y offset == 1",test_zdotu_offset_1},
CU_TEST_INFO_NULL,
};



+ 19
- 0
utest/test_dotu.c View File

@@ -53,4 +53,23 @@ void test_zdotu_n_1(void)
}

void test_zdotu_offset_1(void)
{
int N=1,incX=1,incY=1;
double x1[]={1.0,2.0,3.0,4.0};
double y1[]={5.0,6.0,7.0,8.0};
double x2[]={1.0,2.0,3.0,4.0};
double y2[]={5.0,6.0,7.0,8.0};
double _Complex result1=0.0;
double _Complex result2=0.0;
//OpenBLAS
result1=BLASFUNC(zdotu)(&N,x1+1,&incX,y1+1,&incY);
//reference
result2=BLASFUNC_REF(zdotu)(&N,x2+1,&incX,y2+1,&incY);

CU_ASSERT_DOUBLE_EQUAL(creal(result1), creal(result2), CHECK_EPS);
CU_ASSERT_DOUBLE_EQUAL(cimag(result1), cimag(result2), CHECK_EPS);
// printf("\%lf,%lf\n",creal(result1),cimag(result1));
}


Write Preview
Loading…
Cancel
Save