Refs #205. Merge boegel's codes about downloading LAPACK.

diff --git a/Makefile b/Makefile
index 9992d0b..2668ab3 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -154,6 +154,9 @@ ifeq ($(DYNAMIC_ARCH), 1)
        done
        echo DYNAMIC_ARCH=1 >> Makefile.conf_last
 endif
+ifdef USE_THREAD
+       echo USE_THREAD=$(USE_THREAD) >>  Makefile.conf_last
+endif
        touch lib.grd
 
 prof : prof_blas prof_lapack
@@ -268,7 +271,7 @@ else
 ifeq ($(OSNAME), FreeBSD)
        fetch $(LAPACK_URL);
 else
-       wget $(LAPACK_URL);
+       wget -O $@ $(LAPACK_URL);
 endif
 endif
        fi

diff --git a/Makefile.rule b/Makefile.rule
index 4e23857..9c921d3 100644 (file)
--- a/Makefile.rule
+++ b/Makefile.rule
@@ -81,6 +81,9 @@ VERSION = 0.2.6
 # and OS. However, the performance is low.
 # NO_AVX = 1
 
+# Don't use parallel make.
+# NO_PARALLEL_MAKE = 1
+
 # If you would like to know minute performance report of GotoBLAS.
 # FUNCTION_PROFILE = 1
 
@@ -104,8 +107,8 @@ VERSION = 0.2.6
 
 # If any gemm arguement m, n or k is less or equal this threshold, gemm will be execute 
 # with single thread. You can use this flag to avoid the overhead of multi-threading 
-# in small matrix sizes. The default value is 50.
-# GEMM_MULTITHREAD_THRESHOLD = 50
+# in small matrix sizes. The default value is 4.
+# GEMM_MULTITHREAD_THRESHOLD = 4
 
 # If you need santy check by comparing reference BLAS. It'll be very
 # slow (Not implemented yet).

diff --git a/Makefile.system b/Makefile.system
index eac61e9..b5e64c0 100644 (file)
--- a/Makefile.system
+++ b/Makefile.system
@@ -53,7 +53,7 @@ GETARCH_FLAGS += -DUSE64BITINT
 endif
 
 ifndef GEMM_MULTITHREAD_THRESHOLD
-GEMM_MULTITHREAD_THRESHOLD=50
+GEMM_MULTITHREAD_THRESHOLD=4
 endif
 GETARCH_FLAGS  += -DGEMM_MULTITHREAD_THRESHOLD=$(GEMM_MULTITHREAD_THRESHOLD) 
 
@@ -65,6 +65,10 @@ ifeq ($(DEBUG), 1)
 GETARCH_FLAGS  += -g
 endif
 
+ifndef NO_PARALLEL_MAKE
+NO_PARALLEL_MAKE=0
+endif
+GETARCH_FLAGS  += -DNO_PARALLEL_MAKE=$(NO_PARALLEL_MAKE)
 # This operation is expensive, so execution should be once.
 ifndef GOTOBLAS_MAKEFILE
 export GOTOBLAS_MAKEFILE = 1

diff --git a/README.md b/README.md
index f932b51..2ecea32 100644 (file)
--- a/README.md
+++ b/README.md
@@ -91,7 +91,8 @@ If you compile this lib with USE_OPENMP=1, you should use the above functions, t
 Please add a issue in https://github.com/xianyi/OpenBLAS/issues
 
 ## Contact
-OpenBLAS users mailing list: http://list.rdcps.ac.cn/mailman/listinfo/openblas
+* OpenBLAS users mailing list: https://groups.google.com/forum/#!forum/openblas-users
+* OpenBLAS developers mailing list: https://groups.google.com/forum/#!forum/openblas-dev
 
 ## ChangeLog
 Please see Changelog.txt to obtain the differences between GotoBLAS2 1.13 BSD version.

diff --git a/c_check b/c_check
index 4d82237..5663114 100644 (file)
--- a/c_check
+++ b/c_check
@@ -241,13 +241,13 @@ print CONFFILE "#define FUNDERSCORE\t$need_fu\n" if $need_fu ne "";
 
 if ($os eq "LINUX") {
     
-    @pthread = split(/\s+/, `nm /lib/libpthread.so* | grep _pthread_create`);
+#    @pthread = split(/\s+/, `nm /lib/libpthread.so* | grep _pthread_create`);
     
-    if ($pthread[2] ne "") {
-       print CONFFILE "#define PTHREAD_CREATE_FUNC     $pthread[2]\n";
-    } else {
+#    if ($pthread[2] ne "") {
+#      print CONFFILE "#define PTHREAD_CREATE_FUNC     $pthread[2]\n";
+#    } else {
        print CONFFILE "#define PTHREAD_CREATE_FUNC     pthread_create\n";
-    }
+#    }
 } else {
     print CONFFILE "#define PTHREAD_CREATE_FUNC        pthread_create\n";
 }

diff --git a/cpuid_power.c b/cpuid_power.c
index 24ff78f..9fd9ec9 100644 (file)
--- a/cpuid_power.c
+++ b/cpuid_power.c
@@ -114,6 +114,7 @@ int detect(void){
   if (!strncasecmp(p, "PPC970", 6)) return CPUTYPE_PPC970;
   if (!strncasecmp(p, "POWER5", 6)) return CPUTYPE_POWER5;
   if (!strncasecmp(p, "POWER6", 6)) return CPUTYPE_POWER6;
+  if (!strncasecmp(p, "POWER7", 6)) return CPUTYPE_POWER6;
   if (!strncasecmp(p, "Cell",   4)) return CPUTYPE_CELL;
   if (!strncasecmp(p, "7447",   4)) return CPUTYPE_PPCG4;
 

diff --git a/exports/gensymbol b/exports/gensymbol
index 99609b3..d39f9b2 100644 (file)
--- a/exports/gensymbol
+++ b/exports/gensymbol
@@ -49,7 +49,7 @@
               cblas_zhemv, cblas_zher2, cblas_zher2k, cblas_zher, cblas_zherk, cblas_zhpmv, cblas_zhpr2,
               cblas_zhpr, cblas_zscal, cblas_zswap, cblas_zsymm, cblas_zsyr2k, cblas_zsyrk,
               cblas_ztbmv, cblas_ztbsv, cblas_ztpmv, cblas_ztpsv, cblas_ztrmm, cblas_ztrmv, cblas_ztrsm,
-              cblas_ztrsv);
+              cblas_ztrsv, cblas_cdotc_sub, cblas_cdotu_sub, cblas_zdotc_sub, cblas_zdotu_sub );
 
 @exblasobjs = (
               qamax,qamin,qasum,qaxpy,qcabs1,qcopy,qdot,qgbmv,qgemm,

diff --git a/getarch.c b/getarch.c
index ac10f1c..48d2827 100644 (file)
--- a/getarch.c
+++ b/getarch.c
@@ -804,9 +804,13 @@ int main(int argc, char *argv[]){
 #endif
 #endif
 
+#if NO_PARALLEL_MAKE==1
+    printf("MAKE += -j 1\n");
+#else
 #ifndef OS_WINDOWS
     printf("MAKE += -j %d\n", get_num_cores());
 #endif
+#endif
 
     break;
 

diff --git a/kernel/x86_64/KERNEL.BULLDOZER b/kernel/x86_64/KERNEL.BULLDOZER
index d596685..2ac035f 100644 (file)
--- a/kernel/x86_64/KERNEL.BULLDOZER
+++ b/kernel/x86_64/KERNEL.BULLDOZER
@@ -13,8 +13,8 @@ SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 DGEMMKERNEL    =  dgemm_kernel_4x4_bulldozer.S
 DGEMMINCOPY    =
 DGEMMITCOPY    =
-DGEMMONCOPY    =  gemm_ncopy_4_opteron.S
-DGEMMOTCOPY    =  gemm_tcopy_4_opteron.S
+DGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
 DGEMMINCOPYOBJ =
 DGEMMITCOPYOBJ =
 DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)

diff --git a/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S b/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S
index f8a316b..05bad59 100644 (file)
--- a/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S
+++ b/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S
@@ -36,6 +36,29 @@
 /* or implied, of The University of Texas at Austin.                 */\r
 /*********************************************************************/\r
 \r
+/*********************************************************************\r
+* Changelog:\r
+*\r
+* 2013/04/15 Saar\r
+* Prefetch for A and B\r
+* unroll of inner Loop\r
+* using generic versions for ncopy and tcopy\r
+* moved vmovddup        ALPHA, %xmm7 down\r
+* define A_PR1 192\r
+* define B_PR1 512\r
+**********************************************************************/\r
+\r
+/*********************************************************************\r
+* 2013/04/12   Saar\r
+* Performance:\r
+* 3584x3584    89 GFLOPS with 8 threads on 4 modules\r
+*              76 GFLOPS with 4 threads on 4 modules\r
+*              53 GFLOPS with 4 threads on 2 modules\r
+*              46 GFLOPS with 2 threads on 2 modules\r
+*              28 GFLOPS with 2 threads on 1 module\r
+*            23,6 GFLOPS with 1 thread  on 1 module\r
+*********************************************************************/\r
+\r
 #define ASSEMBLER\r
 #include "common.h"\r
  \r
@@ -87,134 +110,136 @@
 #define movapd movaps\r
 #define movupd movups\r
 \r
+#define A_PR1 192\r
+#define B_PR1 512\r
+\r
 #define KERNEL1(xx) \\r
-       vmovups         -16 * SIZE(AO, %rax, 4),%xmm0 ;\\r
-       vmovups         -14 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup        -16 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup        -15 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups         -14 * SIZE(AO, %rax, 4),%xmm2 ;\\r
+       vfmaddpd %xmm8,%xmm6,%xmm7,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2,%xmm7,%xmm12 ;\\r
        vmovddup        -14 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vfmaddpd %xmm8,%xmm0,%xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2,%xmm1,%xmm12 ;\\r
-       vmovddup        -13 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vmovups         -12 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups         -10 * SIZE(AO, %rax, 4),%xmm6 ;\\r
-       vfmaddpd %xmm9,%xmm0,%xmm3,%xmm9 ;\\r
+       vmovddup        -13 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm9,%xmm6,%xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2,%xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0,%xmm5,%xmm10 ;\\r
+       vmovddup         -8 * SIZE(BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6,%xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm11,%xmm6,%xmm4,%xmm11 ;\\r
+       vmovddup        -12 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0,%xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vmovups         -8 * SIZE(AO, %rax, 4),%xmm6 ;\\r
 \r
 #define KERNEL2(xx) \\r
-       vmovddup        -12 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vmovups         -12 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups         -10 * SIZE(AO, %rax, 4),%xmm0 ;\\r
        vmovddup        -11 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vfmaddpd %xmm8, %xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\\r
        vmovddup        -10 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         -9 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8, %xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1,%xmm12 ;\\r
-       vfmaddpd %xmm9, %xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13,%xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm9, %xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup         -9 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL3(xx) \\r
-       vmovups         -8 * SIZE(AO, %rax, 4),%xmm0 ;\\r
        vmovups         -6 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup         -8 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup         -7 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vfmaddpd %xmm8, %xmm6, %xmm7, %xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\\r
        vmovddup         -6 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         -5 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8, %xmm0, %xmm1, %xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\\r
-       vmovups          -4 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups          -2 * SIZE(AO, %rax, 4),%xmm6 ;\\r
-       vfmaddpd %xmm9, %xmm0, %xmm3,%xmm9 ;\\r
+       vmovddup         -5 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm9, %xmm6, %xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\\r
+       vmovddup                  (BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0, %xmm7, %xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vmovddup         -4 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vfmaddpd %xmm11,%xmm6, %xmm4, %xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
+       vmovups                   (AO, %rax, 4), %xmm6 ;\\r
 \r
 #define KERNEL4(xx) \\r
-       vmovddup         -4 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vmovups          -4 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups          -2 * SIZE(AO, %rax, 4),%xmm0 ;\\r
        vmovddup         -3 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1 ,%xmm12;\\r
        vmovddup         -2 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         -1 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1 ,%xmm12;\\r
-       vfmaddpd %xmm9,%xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13,%xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup         -1 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL5(xx) \\r
-       vmovups                   (AO, %rax, 4), %xmm0 ;\\r
        vmovups           2 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup                  (BO, %rax, 4), %xmm1 ;\\r
        vmovddup          1 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\\r
        vmovddup          2 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup          3 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm0, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\\r
-       vmovups         4 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups         6 * SIZE(AO, %rax, 4),%xmm6 ;\\r
-       vfmaddpd %xmm9,%xmm0, %xmm3,%xmm9 ;\\r
+       vmovddup          3 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\\r
+       vmovddup          8 * SIZE(BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vmovddup        4 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
+       vmovups           8 * SIZE(AO, %rax, 4), %xmm6 ;\\r
 \r
 #define KERNEL6(xx) \\r
-       vmovddup        4 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vmovups         4 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups         6 * SIZE(AO, %rax, 4),%xmm0 ;\\r
        vmovddup        5 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\\r
        vmovddup          6 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup          7 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1,%xmm12 ;\\r
-       vfmaddpd %xmm9,%xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13,%xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup          7 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL7(xx) \\r
-       vmovups           8 * SIZE(AO, %rax, 4), %xmm0 ;\\r
        vmovups          10 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup          8 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup          9 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\\r
        vmovddup         10 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         11 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm0, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\\r
-       vmovups          12 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups          14 * SIZE(AO, %rax, 4), %xmm6 ;\\r
-       vfmaddpd %xmm9,%xmm0, %xmm3,%xmm9 ;\\r
+       vmovddup         11 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\\r
+       vmovddup         16 * SIZE(BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vmovddup         12 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
+       vmovups          16 * SIZE(AO, %rax, 4),%xmm6 ;\\r
 \r
 #define KERNEL8(xx) \\r
-       vmovddup         12 * SIZE(BO, %rax, 4), %xmm1 ;\\r
+       vmovups          12 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups          14 * SIZE(AO, %rax, 4), %xmm0 ;\\r
        vmovddup         13 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\\r
        vmovddup         14 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         15 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1,%xmm12 ;\\r
-       vfmaddpd %xmm9,%xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13, %xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
-       addq    $8 * SIZE, %rax ;\\r
+       vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13, %xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup         15 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL_SUB1(xx) \\r
        vmovups         -16 * SIZE(AO),%xmm0 ;\\r
@@ -368,7 +393,7 @@
        movq    A, AO           # aoffset = a\r
 \r
        movq    K, %rax\r
-       salq    $BASE_SHIFT + 2, %rax\r
+       salq    $BASE_SHIFT + 2, %rax    # k << 5 # K * 32\r
        leaq     (B, %rax), BB\r
 \r
        movq    M,  I\r
@@ -389,6 +414,8 @@
        leaq    (B,  %rax, 4), BO\r
 #endif \r
 \r
+       \r
+\r
        vxorpd  %xmm8, %xmm8,%xmm8\r
        vxorpd  %xmm9, %xmm9,%xmm9\r
        vxorpd  %xmm10, %xmm10,%xmm10\r
@@ -398,10 +425,10 @@
        vxorpd  %xmm14, %xmm14,%xmm14\r
        vxorpd  %xmm15, %xmm15,%xmm15\r
 \r
-       prefetcht0       (CO1)\r
-       prefetcht0       (CO1,LDC)\r
-       prefetcht0       (CO2)\r
-       prefetcht0       (CO2,LDC)\r
+       // prefetchw       (CO1)\r
+       // prefetchw       (CO1,LDC)\r
+       // prefetchw       (CO2)\r
+       // prefetchw       (CO2,LDC)\r
 \r
 #ifndef TRMMKERNEL\r
        movq    K, %rax\r
@@ -424,33 +451,145 @@
        leaq    (AO, %rax, 4), AO\r
        leaq    (BO, %rax, 4), BO\r
        negq    %rax\r
-       NOBRANCH\r
        je      .L15\r
        // ALIGN_4\r
 \r
-       .align 16\r
+       vmovups         -16 * SIZE(AO, %rax, 4),%xmm6 \r
+       vmovddup        -16 * SIZE(BO, %rax, 4), %xmm7 \r
 \r
-#define PR1 16\r
-#define PR2 24\r
+       .align 16\r
 \r
 .L12:\r
-       prefetcht0 PR1*SIZE(AO,%rax,4)\r
-       prefetcht0 PR2*SIZE(AO,%rax,4)\r
-       prefetcht0 PR1*SIZE(BO,%rax,4)\r
-       prefetcht0 PR2*SIZE(BO,%rax,4)\r
-       KERNEL1(16 *  0)\r
-       KERNEL2(16 *  0)\r
-       KERNEL3(16 *  0)\r
-       KERNEL4(16 *  0)\r
-       KERNEL5(16 *  0)\r
-       KERNEL6(16 *  0)\r
-       KERNEL7(16 *  0)\r
-       KERNEL8(16 *  0)\r
-       jl      .L12\r
-       ALIGN_4\r
+\r
+#ifndef SMP \r
+\r
+       prefetcht0 A_PR1(AO,%rax,4)\r
+        prefetcht0 B_PR1(BO,%rax,4)\r
+        KERNEL1(16 *  0)\r
+        KERNEL2(16 *  0)\r
+        prefetcht0 A_PR1+64(AO,%rax,4)\r
+        prefetcht0 B_PR1+64(BO,%rax,4)\r
+        KERNEL3(16 *  0)\r
+        KERNEL4(16 *  0)\r
+        prefetcht0 A_PR1+128(AO,%rax,4)\r
+        prefetcht0 B_PR1+128(BO,%rax,4)\r
+        KERNEL5(16 *  0)\r
+        KERNEL6(16 *  0)\r
+        prefetcht0 A_PR1+192(AO,%rax,4)\r
+        prefetcht0 B_PR1+192(BO,%rax,4)\r
+        KERNEL7(16 *  0)\r
+        KERNEL8(16 *  0)\r
+\r
+       addq    $8 * SIZE, %rax \r
+       je      .L15\r
+\r
+       prefetcht0 A_PR1(AO,%rax,4)\r
+        prefetcht0 B_PR1(BO,%rax,4)\r
+        KERNEL1(16 *  0)\r
+        KERNEL2(16 *  0)\r
+        prefetcht0 A_PR1+64(AO,%rax,4)\r
+        prefetcht0 B_PR1+64(BO,%rax,4)\r
+        KERNEL3(16 *  0)\r
+        KERNEL4(16 *  0)\r
+        prefetcht0 A_PR1+128(AO,%rax,4)\r
+        prefetcht0 B_PR1+128(BO,%rax,4)\r
+        KERNEL5(16 *  0)\r
+        KERNEL6(16 *  0)\r
+        prefetcht0 A_PR1+192(AO,%rax,4)\r
+        prefetcht0 B_PR1+192(BO,%rax,4)\r
+        KERNEL7(16 *  0)\r
+        KERNEL8(16 *  0)\r
+\r
+       addq    $8 * SIZE, %rax \r
+       je      .L15\r
+\r
+\r
+       prefetcht0 A_PR1(AO,%rax,4)\r
+        prefetcht0 B_PR1(BO,%rax,4)\r
+        KERNEL1(16 *  0)\r
+        KERNEL2(16 *  0)\r
+        prefetcht0 A_PR1+64(AO,%rax,4)\r
+        prefetcht0 B_PR1+64(BO,%rax,4)\r
+        KERNEL3(16 *  0)\r
+        KERNEL4(16 *  0)\r
+        prefetcht0 A_PR1+128(AO,%rax,4)\r
+        prefetcht0 B_PR1+128(BO,%rax,4)\r
+        KERNEL5(16 *  0)\r
+        KERNEL6(16 *  0)\r
+        prefetcht0 A_PR1+192(AO,%rax,4)\r
+        prefetcht0 B_PR1+192(BO,%rax,4)\r
+        KERNEL7(16 *  0)\r
+        KERNEL8(16 *  0)\r
+\r
+       addq    $8 * SIZE, %rax \r
+       je      .L15\r
+\r
+       prefetcht0 A_PR1(AO,%rax,4)\r
+        prefetcht0 B_PR1(BO,%rax,4)\r
+        KERNEL1(16 *  0)\r
+        KERNEL2(16 *  0)\r
+        prefetcht0 A_PR1+64(AO,%rax,4)\r
+        prefetcht0 B_PR1+64(BO,%rax,4)\r
+        KERNEL3(16 *  0)\r
+        KERNEL4(16 *  0)\r
+        prefetcht0 A_PR1+128(AO,%rax,4)\r
+        prefetcht0 B_PR1+128(BO,%rax,4)\r
+        KERNEL5(16 *  0)\r
+        KERNEL6(16 *  0)\r
+        prefetcht0 A_PR1+192(AO,%rax,4)\r
+        prefetcht0 B_PR1+192(BO,%rax,4)\r
+        KERNEL7(16 *  0)\r
+        KERNEL8(16 *  0)\r
+\r
+       addq    $8 * SIZE, %rax \r
+       jnz     .L12\r
+\r
+       .align 16\r
+#else\r
+#ifdef OPTMODULE\r
+\r
+       prefetcht0 A_PR1(AO,%rax,4)\r
+        prefetcht0 B_PR1(BO,%rax,4)\r
+        KERNEL1(16 *  0)\r
+        KERNEL2(16 *  0)\r
+        prefetcht0 A_PR1+64(AO,%rax,4)\r
+        prefetcht0 B_PR1+64(BO,%rax,4)\r
+        KERNEL3(16 *  0)\r
+        KERNEL4(16 *  0)\r
+        prefetcht0 A_PR1+128(AO,%rax,4)\r
+        prefetcht0 B_PR1+128(BO,%rax,4)\r
+        KERNEL5(16 *  0)\r
+        KERNEL6(16 *  0)\r
+        prefetcht0 A_PR1+192(AO,%rax,4)\r
+        prefetcht0 B_PR1+192(BO,%rax,4)\r
+        KERNEL7(16 *  0)\r
+        KERNEL8(16 *  0)\r
+\r
+       addq    $8 * SIZE, %rax \r
+       je      .L15\r
+       jmp     .L12\r
+       .align 16\r
+\r
+#else\r
+        KERNEL1(16 *  0)\r
+        KERNEL2(16 *  0)\r
+        KERNEL3(16 *  0)\r
+        KERNEL4(16 *  0)\r
+        KERNEL5(16 *  0)\r
+        KERNEL6(16 *  0)\r
+        KERNEL7(16 *  0)\r
+        KERNEL8(16 *  0)\r
+\r
+       addq    $8 * SIZE, %rax \r
+       je      .L15\r
+       jmp     .L12\r
+       .align 16\r
+\r
+#endif\r
+#endif\r
+\r
 \r
 .L15:\r
-       vmovddup        ALPHA, %xmm7\r
 \r
 #ifndef TRMMKERNEL\r
        movq    K, %rax\r
@@ -476,6 +615,7 @@
 #else\r
        movq    KKK, %rax\r
 #endif\r
+       vmovddup        ALPHA, %xmm7\r
        andq    $3, %rax                # if (k & 1)\r
        je .L19\r
 \r
@@ -518,13 +658,10 @@
 \r
        vfmaddpd (CO1),%xmm7, %xmm8,%xmm8\r
        vfmaddpd 2 * SIZE(CO1),%xmm7, %xmm12,%xmm12\r
-       .align 2\r
        vfmaddpd (CO1, LDC),%xmm7, %xmm9,%xmm9\r
        vfmaddpd 2 * SIZE(CO1, LDC),%xmm7, %xmm13,%xmm13\r
-       .align 2\r
        vfmaddpd (CO2),%xmm7, %xmm10,%xmm10\r
        vfmaddpd 2 * SIZE(CO2),%xmm7, %xmm14,%xmm14\r
-       .align 2\r
        vfmaddpd (CO2, LDC),%xmm7, %xmm11,%xmm11\r
        vfmaddpd 2 * SIZE(CO2, LDC),%xmm7, %xmm15,%xmm15\r
 \r

diff --git a/openblas_config_template.h b/openblas_config_template.h
index cf2c037..ccf7bac 100644 (file)
--- a/openblas_config_template.h
+++ b/openblas_config_template.h
@@ -51,7 +51,9 @@ typedef int blasint;
 #if (defined(__STDC_IEC_559_COMPLEX__) || __STDC_VERSION__ >= 199901L || \
      (__GNUC__ >= 3 && !defined(__cplusplus)))
   #define OPENBLAS_COMPLEX_C99
+#ifndef __cplusplus
   #include <complex.h>
+#endif
   typedef float _Complex openblas_complex_float;
   typedef double _Complex openblas_complex_double;
   typedef xdouble _Complex openblas_complex_xdouble;