[gentoo-commits] proj/sci:master commit in: sci-libs/magma/, sci-libs/magma/files/

[Sebastien Fabbro]

commit:     19e8945bd388537423c2b8290a889f4453edfe08
Author:     Sebastien Fabbro <sfabbro <AT> uvic <DOT> ca>
AuthorDate: Tue Jul  3 19:31:08 2012 +0000
Commit:     Sebastien Fabbro <bicatali <AT> gentoo <DOT> org>
CommitDate: Tue Jul  3 19:31:08 2012 +0000
URL:        http://git.overlays.gentoo.org/gitweb/?p=proj/sci.git;a=commit;h=19e8945b

sci-libs/magma: Version bump. Updated patches

(Portage version: 2.2.01.20430-prefix/git/Linux x86_64, unsigned Manifest commit)

---
 sci-libs/magma/ChangeLog                           |    9 +
 sci-libs/magma/files/magma-1.2.0-cblas-dotc.patch  |   71 --
 .../files/magma-1.2.0-duplicate-symbols.patch      | 1227 --------------------
 .../files/magma-1.2.1-duplicate-symbols.patch      |  612 ++++++++++
 .../magma/files/magma-1.2.1-no-cuda-driver.patch   |   38 +
 .../{magma-1.2.0.ebuild => magma-1.2.1.ebuild}     |   18 +-
 6 files changed, 667 insertions(+), 1308 deletions(-)

diff --git a/sci-libs/magma/ChangeLog b/sci-libs/magma/ChangeLog
index 05bcaf9..460bde8 100644
--- a/sci-libs/magma/ChangeLog
+++ b/sci-libs/magma/ChangeLog
@@ -2,6 +2,15 @@
 # Copyright 1999-2012 Gentoo Foundation; Distributed under the GPL v2
 # $Header: $
 
+*magma-1.2.1 (03 Jul 2012)
+
+  03 Jul 2012; Sébastien Fabbro <bicatali@g.o>
+  +files/magma-1.2.1-duplicate-symbols.patch,
+  +files/magma-1.2.1-no-cuda-driver.patch, +magma-1.2.1.ebuild,
+  -files/magma-1.2.0-cblas-dotc.patch,
+  -files/magma-1.2.0-duplicate-symbols.patch, -magma-1.2.0.ebuild:
+  sci-libs/magma: Version bump. Updated patches
+
 *magma-1.2.0 (25 Jun 2012)
 
   25 Jun 2012; Sébastien Fabbro <bicatali@g.o>

diff --git a/sci-libs/magma/files/magma-1.2.0-cblas-dotc.patch b/sci-libs/magma/files/magma-1.2.0-cblas-dotc.patch
deleted file mode 100644
index e2d16a4..0000000
--- a/sci-libs/magma/files/magma-1.2.0-cblas-dotc.patch
+++ /dev/null
@@ -1,71 +0,0 @@
-Description: Use cblas instead of calls to non existing codtc and zdotc C routines
-Author: Sebastien Fabbro <bicatali@g.o>
-
-diff -Nur magma-1.2.0.orig/src/clatrd2.cpp magma-1.2.0/src/clatrd2.cpp
---- magma-1.2.0.orig/src/clatrd2.cpp	2012-06-06 17:41:50.000000000 +0100
-+++ magma-1.2.0/src/clatrd2.cpp	2012-06-06 17:45:23.000000000 +0100
-@@ -12,6 +12,7 @@
- 
- */
- #include "common_magma.h"
-+#include <cblas.h>
- 
- #define PRECISION_c
- 
-@@ -270,7 +271,9 @@
-           blasf77_cscal(&i, &tau[i - 1], W(0, iw), &ione);
- 
- #if defined(PRECISION_z) || defined(PRECISION_c)
--          blasf77_cdotc(&value, &i, W(0, iw), &ione, A(0, i), &ione);
-+          cblas_cdotc_sub(i, W(0, iw), ione, A(0, i), ione, &value);
-+
-+//          blasf77_cdotc(&value, &i, W(0, iw), &ione, A(0, i), &ione);
-           alpha = tau[i - 1] * -.5f * value;
- #else
-           alpha = tau[i - 1] * -.5f * blasf77_cdotc(&i, W(0, iw), &ione, A(0, i), &ione);
-@@ -349,7 +352,10 @@
-                             W(0, i), &ione, &c_one, W(i+1, i), &ione);
-               blasf77_cscal(&i_n, &tau[i], W(i+1,i), &ione);
-               #if defined(PRECISION_z) || defined(PRECISION_c)
--                    blasf77_cdotc(&value, &i_n, W(i+1,i), &ione, A(i+1, i), &ione);
-+                        cblas_cdotc_sub(i_n, W(i +1, i), ione,
-+                                        A(i +1, i), ione, &value);
-+                    
-+                  //blasf77_cdotc(&value, &i_n, W(i+1,i), &ione, A(i+1, i), &ione);
-                   alpha = tau[i]* -.5f * value;
-               #else
-                   alpha = tau[i]* -.5f* blasf77_cdotc(&i_n, W(i+1,i), &ione, A(i+1, i), &ione);
-diff -Nur magma-1.2.0.orig/src/zlatrd2.cpp magma-1.2.0/src/zlatrd2.cpp
---- magma-1.2.0.orig/src/zlatrd2.cpp	2012-06-06 17:41:50.000000000 +0100
-+++ magma-1.2.0/src/zlatrd2.cpp	2012-06-06 17:46:33.000000000 +0100
-@@ -12,6 +12,7 @@
- 
- */
- #include "common_magma.h"
-+#include <cblas.h> 
- 
- #define PRECISION_z
- 
-@@ -270,7 +271,9 @@
-           blasf77_zscal(&i, &tau[i - 1], W(0, iw), &ione);
- 
- #if defined(PRECISION_z) || defined(PRECISION_c)
--          blasf77_zdotc(&value, &i, W(0, iw), &ione, A(0, i), &ione);
-+          cblas_zdotc_sub(i, W(0, iw), ione, A(0, i), ione, &value);
-+
-+//          blasf77_zdotc(&value, &i, W(0, iw), &ione, A(0, i), &ione);
-           alpha = tau[i - 1] * -.5f * value;
- #else
-           alpha = tau[i - 1] * -.5f * blasf77_zdotc(&i, W(0, iw), &ione, A(0, i), &ione);
-@@ -349,7 +352,10 @@
-                             W(0, i), &ione, &c_one, W(i+1, i), &ione);
-               blasf77_zscal(&i_n, &tau[i], W(i+1,i), &ione);
-               #if defined(PRECISION_z) || defined(PRECISION_c)
--                    blasf77_zdotc(&value, &i_n, W(i+1,i), &ione, A(i+1, i), &ione);
-+                        cblas_zdotc_sub(i_n, W(i +1, i), ione,
-+                                        A(i +1, i), ione, &value);
-+                    
-+                  //blasf77_zdotc(&value, &i_n, W(i+1,i), &ione, A(i+1, i), &ione);
-                   alpha = tau[i]* -.5f * value;
-               #else
-                   alpha = tau[i]* -.5f* blasf77_zdotc(&i_n, W(i+1,i), &ione, A(i+1, i), &ione);

diff --git a/sci-libs/magma/files/magma-1.2.0-duplicate-symbols.patch b/sci-libs/magma/files/magma-1.2.0-duplicate-symbols.patch
deleted file mode 100644
index 0627f7f..0000000
--- a/sci-libs/magma/files/magma-1.2.0-duplicate-symbols.patch
+++ /dev/null
@@ -1,1227 +0,0 @@
-diff -Nur src.orig/cgeqrf_mgpu-trace.cpp src/cgeqrf_mgpu-trace.cpp
---- src.orig/cgeqrf_mgpu-trace.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/cgeqrf_mgpu-trace.cpp	2012-06-25 17:10:21.000000000 +0100
-@@ -16,108 +16,42 @@
- #include <sys/time.h>
- #include <assert.h>
- 
--float get_current_cpu_time(void)
--{
--  struct timeval  time_val;
--
--  gettimeofday(&time_val, NULL);
--
--  return (float)(time_val.tv_sec) + (float)(time_val.tv_usec) / 1000000.0;
--}
-+extern float get_current_cpu_time_float(void);
- 
- #define MAX_THREADS 5
- 
- //#define MAX_EVENTS 163840
- #define MAX_EVENTS 1048576
- 
--int    event_num        [MAX_THREADS]               __attribute__ ((aligned (128)));
--float event_start_time [MAX_THREADS]               __attribute__ ((aligned (128)));
--float event_end_time   [MAX_THREADS]               __attribute__ ((aligned (128)));
--float event_log        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
--int log_events = 1;
-+extern int    event_num        [MAX_THREADS]               __attribute__ ((aligned (128)));
-+extern float event_start_time_float [MAX_THREADS]               __attribute__ ((aligned (128)));
-+extern float event_end_time_float   [MAX_THREADS]               __attribute__ ((aligned (128)));
-+extern float event_log_float        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
-+extern int log_events;
- 
- #define core_cpu_event_start(my_core_id)                      \
--  event_start_time[my_core_id] = get_current_cpu_time();      \
-+  event_start_time_float[my_core_id] = get_current_cpu_time_float();      \
- 
- #define core_cpu_event_end(my_core_id)                        \
--  event_end_time[my_core_id] = get_current_cpu_time();        \
-+  event_end_time_float[my_core_id] = get_current_cpu_time_float();        \
- 
- #define core_gpu_event_start(my_core_id, e1, e2)              \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_start_time[my_core_id] = ctime/1000.+dtime;           \
-+  event_start_time_float[my_core_id] = ctime/1000.+dtime;           \
- 
- #define core_gpu_event_end(my_core_id, e1, e2)                \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_end_time[my_core_id] = ctime/1000.+dtime;             \
-+  event_end_time_float[my_core_id] = ctime/1000.+dtime;             \
- 
- #define core_log_event(event, my_core_id)             \
--  event_log[my_core_id][event_num[my_core_id]+0] = my_core_id;\
--  event_log[my_core_id][event_num[my_core_id]+1] = event_start_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+2] = event_end_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+3] = (event);\
-+  event_log_float[my_core_id][event_num[my_core_id]+0] = my_core_id;\
-+  event_log_float[my_core_id][event_num[my_core_id]+1] = event_start_time_float[my_core_id];\
-+  event_log_float[my_core_id][event_num[my_core_id]+2] = event_end_time_float[my_core_id];\
-+  event_log_float[my_core_id][event_num[my_core_id]+3] = (event);\
-   event_num[my_core_id] += (log_events << 2);         \
-   event_num[my_core_id] &= (MAX_EVENTS-1);
- 
--void dump_trace(int cores_num)
--{
--  char trace_file_name[32];
--  FILE *trace_file;
--  int event;
--  int core;
--
--  float scale = 100000.0;
--  float large = 100.0;
--
--  sprintf(trace_file_name, "trace.svg");
--  trace_file = fopen(trace_file_name, "w");
--  assert(trace_file != NULL);
--
--  fprintf(trace_file,
--          "<?xml version=\"1.0\" standalone=\"no\"?>"
--          "<svg version=\"1.1\" baseProfile=\"full\" xmlns=\"http://www.w3.org/2000/svg\" "
--          "xmlns:xlink=\"http://www.w3.org/1999/xlink\" xmlns:ev=\"http://www.w3.org/2001/xml-events\"  "
--          ">\n"
--          "  <g font-size=\"20\">\n");   
--  
--  for (core = 0; core < cores_num; core++)
--    for (event = 0; event < event_num[core]; event += 4)
--      {
--        int    thread = event_log[core][event+0];
--        float start  = event_log[core][event+1];
--        float end    = event_log[core][event+2];
--        int    color  = event_log[core][event+3];
--
--        start -= event_log[core][2];
--        end   -= event_log[core][2];
--        /*
--            fprintf(trace_file,
--                "    "
--                "<rect x=\"%.2lf\" y=\"%.0lf\" width=\"%.2lf\" height=\"%.0lf\" "
--                "fill=\"#%06x\" stroke=\"#000000\" stroke-width=\"1\"/>\n",
--                start * scale,
--                thread * 100.0,
--                (end - start) * scale,
--                90.0,
--                color);
--        */
--        fprintf(trace_file,
--            "    "
--            "<rect x=\"%.2lf\" y=\"%.0lf\" width=\"%.2lf\" height=\"%.0lf\" "
--                //            "fill=\"#%06x\" />\n",
--                "fill=\"#%06x\" stroke=\"#000000\" stroke-width=\"1\"/>\n",
--                start * scale,
--                thread * (large+20.0),
--                (end - start) * scale,
--                large,
--                color);
--      }
--
--  fprintf(trace_file,
--          "  </g>\n"
--          "</svg>\n");
--
--  fclose(trace_file);
--}
-+extern void dump_trace_float(int cores_num);
- 
- //===========================================================================
- 
-@@ -279,7 +213,7 @@
-     core_cpu_event_end(num_gpus);
-     core_log_event(0x666666, num_gpus);
-     
--    dtime = get_current_cpu_time();
-+    dtime = get_current_cpu_time_float();
- 
-     for(j=0; j<num_gpus; j++){
-       magma_setdevice(j);
-@@ -536,7 +470,7 @@
-     }
- 
-     magma_setdevice(cdevice);
--    dump_trace(num_gpus+1);
-+    dump_trace_float(num_gpus+1);
- 
-     return *info;
- } /* magma_cgeqrf2_mgpu */
-diff -Nur src.orig/cheevr_gpu.cpp src/cheevr_gpu.cpp
---- src.orig/cheevr_gpu.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/cheevr_gpu.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -13,34 +13,6 @@
- #include "common_magma.h"
- 
- /* These interfaces are used for TAU profiling */
--extern "C" {
--    void Mylapackf77_cstemr(const char *jobz, const char *range, magma_int_t *n, float *d, float *e,
--                            float *vl, float *vu, magma_int_t *il, magma_int_t *iu, 
--                            magma_int_t *m, float *w, cuFloatComplex *z, magma_int_t *ldz,
--                            magma_int_t *nzc, magma_int_t *isuppz, magma_int_t *tryrac,
--                            float *work, magma_int_t *lwork, magma_int_t *iwork, 
--                            magma_int_t *liwork, magma_int_t *info)
--    {
--        lapackf77_cstemr(jobz, range, n, d, e, vl, vu, il, iu, m, w, z, ldz, nzc, 
--                         isuppz, tryrac, work, lwork, iwork, liwork, info);
--    }
--
--    void Mylapackf77_cstein(int *n, float *d, float *e, int *m, float *w, int *iblock,
--                            int *isplit, cuFloatComplex *z, int *ldz, float *work, 
--                            int *iwork, int *ifail, int *info)
--    {
--        lapackf77_cstein(n, d, e, m, w, iblock, isplit, z, ldz, work, iwork, ifail, info);
--    }
--  
--    void Mylapackf77_sstebz(const char *range, const char *order, int *n, float *vl, 
--                            float *vu, int *il, int *iu, float *abstol,
--                            float *d, float *e, int *m, int *nsplit, float *w, 
--                            int *iblock, int *isplit, float *work, int *iwork, int *info)
--    {
--        lapackf77_sstebz(range, order, n, vl, vu, il, iu, abstol, d, e, m,
--                         nsplit, w, iblock, isplit, work, iwork,info);
--    }
--}
- 
- extern "C" {
- #ifdef ADD_
-@@ -497,7 +469,7 @@
-     else
-       tryrac=0;
-     
--    Mylapackf77_cstemr(jobz_, range_, &n, &rwork[indrdd], &rwork[indree], &vl, &vu, &il, 
-+    lapackf77_cstemr(jobz_, range_, &n, &rwork[indrdd], &rwork[indree], &vl, &vu, &il, 
-                      &iu, m, &w[1], wz, &ldwz, &n, &isuppz[1], &tryrac, &rwork[indrwk],
-                      &llrwork, &iwork[1], &liwork, info);
-     
-@@ -514,10 +486,10 @@
-     printf("B/I\n");
-     *info = 0;
-     
--    Mylapackf77_sstebz(range_, "B", &n, &vl, &vu, &il, &iu, &abstol, &rwork[indrd], &rwork[indre], m,
-+    lapackf77_sstebz(range_, "B", &n, &vl, &vu, &il, &iu, &abstol, &rwork[indrd], &rwork[indre], m,
-                      &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwo], info);
-     
--    Mylapackf77_cstein(&n, &rwork[indrd], &rwork[indre], m, &w[1], &iwork[indibl], &iwork[indisp],
-+    lapackf77_cstein(&n, &rwork[indrd], &rwork[indre], m, &w[1], &iwork[indibl], &iwork[indisp],
-                        wz, &ldwz, &rwork[indrwk], &iwork[indiwo], &iwork[indifl], info);
-       
-       /*        Apply unitary matrix used in reduction to tridiagonal   
-diff -Nur src.orig/cheevx_gpu.cpp src/cheevx_gpu.cpp
---- src.orig/cheevx_gpu.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/cheevx_gpu.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,27 +12,6 @@
-  */
- #include "common_magma.h"
- 
--/* These interfaces are used for TAU profiling */
--extern"C"{
--    void Mylapackf77_cstein(magma_int_t *n, float *d, float *e, magma_int_t *m, 
--                            float *w, magma_int_t *iblock, magma_int_t *isplit,
--                            cuFloatComplex *z, magma_int_t *ldz, float *work, 
--                            magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info)
--    {
--        lapackf77_cstein(n, d, e, m, w, iblock, isplit, z, ldz, work, iwork, ifail, info);
--    }
--
--    void Mylapackf77_sstebz(char *range, char *order, magma_int_t *n, float *vl,
--                            float *vu, magma_int_t *il, magma_int_t *iu, float *abstol,
--                            float *d, float *e, magma_int_t *m, magma_int_t *nsplit, 
--                            float *w, magma_int_t *iblock, magma_int_t *isplit, 
--                            float *work, magma_int_t *iwork, magma_int_t *info)
--    {
--        lapackf77_sstebz(range, order, n, vl, vu, il, iu, abstol, 
--                         d, e, m, nsplit, w, iblock, isplit, work, iwork,info);
--    }
--}
--
- extern "C" magma_int_t 
- magma_cheevx_gpu(char jobz, char range, char uplo, magma_int_t n, 
-                  cuFloatComplex *da, magma_int_t ldda, float vl, float vu, 
-@@ -404,12 +383,12 @@
-     indisp = indibl + n;
-     indiwk = indisp + n;
- 
--    Mylapackf77_sstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
-+    lapackf77_sstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
-                      &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwk], info);
-     
-     if (wantz) {
-       
--      Mylapackf77_cstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
-+      lapackf77_cstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
-                        wz, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info);
-       
-       magma_csetmatrix( n, *m, wz, ldwz, dz, lddz );
-diff -Nur src.orig/chegvd.cpp src/chegvd.cpp
---- src.orig/chegvd.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/chegvd.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -13,26 +13,6 @@
- */
- #include "common_magma.h"
- 
--/* This ctrmm interface is used for TAU profiling */
--void Mymagma_ctrmm(char side, char uplo, char trans, char unit,
--                   magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--  magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--  magma_device_sync();
--}
--
--/* This ctrsm interface is used for TAU profiling */
--void Mymagma_ctrsm(char side, char uplo, char trans, char unit,
--                   magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--  magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--  magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_chegvd(magma_int_t itype, char jobz, char uplo, magma_int_t n,
-              cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb,
-@@ -324,7 +304,7 @@
-                 *(unsigned char *)trans = MagmaNoTrans;
-             }
- 
--            Mymagma_ctrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-+            magma_ctrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-                           n, n, c_one, db, lddb, da, ldda);
- 
-         } else if (itype == 3)
-@@ -337,7 +317,7 @@
-                 *(unsigned char *)trans = MagmaConjTrans;
-             }
- 
--            Mymagma_ctrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-+            magma_ctrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-                           n, n, c_one, db, lddb, da, ldda);
-         }
- 
-diff -Nur src.orig/chegvdx.cpp src/chegvdx.cpp
---- src.orig/chegvdx.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/chegvdx.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,22 +12,6 @@
- */
- #include "common_magma.h"
- 
--void Mymagma_ctrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--    magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
--void Mymagma_ctrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--    magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_chegvdx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
-               cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb, 
-@@ -363,7 +347,7 @@
-                 *(unsigned char *)trans = MagmaNoTrans;
-             }
- 
--            Mymagma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
-+            magma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
- 
-         } else if (itype == 3) {
- 
-@@ -375,7 +359,7 @@
-                 *(unsigned char *)trans = MagmaConjTrans;
-             }
- 
--            Mymagma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
-+            magma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
- 
-         }
- 
-diff -Nur src.orig/chegvr.cpp src/chegvr.cpp
---- src.orig/chegvr.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/chegvr.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,24 +12,6 @@
-  */
- #include "common_magma.h"
- 
--void Mymagma_ctrmm(char side, char uplo, char trans, char unit, 
--                   magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--    magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
--void Mymagma_ctrsm(char side, char uplo, char trans, char unit, 
--                   magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--    magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_chegvr(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
-              cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb, 
-@@ -408,7 +390,7 @@
-                 *(unsigned char *)trans = MagmaNoTrans;
-             }
-             
--            Mymagma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-+            magma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-                           db, lddb, dz, lddz);
-       
-         } else if (itype == 3) {
-@@ -421,7 +403,7 @@
-                 *(unsigned char *)trans = MagmaConjTrans;
-             }
-             
--            Mymagma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-+            magma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-                           db, lddb, dz, lddz);
-         }
-         
-diff -Nur src.orig/chegvx.cpp src/chegvx.cpp
---- src.orig/chegvx.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/chegvx.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,22 +12,6 @@
-  */
- #include "common_magma.h"
- 
--void Mymagma_ctrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--    magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
--void Mymagma_ctrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
--                   cuFloatComplex *dz, magma_int_t lddz)
--{
--    magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_chegvx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
-              cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb, 
-@@ -330,7 +314,7 @@
-         *(unsigned char *)trans = MagmaNoTrans;
-       }
-       
--      Mymagma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-+      magma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-       
-     } else if (itype == 3) {
-       
-@@ -343,7 +327,7 @@
-         *(unsigned char *)trans = MagmaConjTrans;
-       }
-       
--      Mymagma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-+      magma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-       
-     }
-     
-diff -Nur src.orig/cstedx_m.cpp src/cstedx_m.cpp
---- src.orig/cstedx_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/cstedx_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -16,10 +16,7 @@
-                                float* work, magma_int_t lwork, magma_int_t* iwork, magma_int_t liwork,
-                                magma_int_t* info);
- 
--    magma_int_t get_cstedx_smlsize()
--    {
--        return 25;
--    }
-+   magma_int_t get_cstedx_smlsize();
- }
- 
- extern "C" magma_int_t
-diff -Nur src.orig/dgeqrf_mgpu-trace.cpp src/dgeqrf_mgpu-trace.cpp
---- src.orig/dgeqrf_mgpu-trace.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/dgeqrf_mgpu-trace.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -1,4 +1,4 @@
--/*
-+/*5A5A5A
-     -- MAGMA (version 1.2.0) --
-        Univ. of Tennessee, Knoxville
-        Univ. of California, Berkeley
-@@ -16,7 +16,7 @@
- #include <sys/time.h>
- #include <assert.h>
- 
--double get_current_cpu_time(void)
-+double get_current_cpu_time_double(void)
- {
-   struct timeval  time_val;
- 
-@@ -30,35 +30,35 @@
- //#define MAX_EVENTS 163840
- #define MAX_EVENTS 1048576
- 
--int    event_num        [MAX_THREADS]               __attribute__ ((aligned (128)));
--double event_start_time [MAX_THREADS]               __attribute__ ((aligned (128)));
--double event_end_time   [MAX_THREADS]               __attribute__ ((aligned (128)));
--double event_log        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
--int log_events = 1;
-+extern int    event_num        [MAX_THREADS]               __attribute__ ((aligned (128)));
-+double event_start_time_double [MAX_THREADS]               __attribute__ ((aligned (128)));
-+double event_end_time_double   [MAX_THREADS]               __attribute__ ((aligned (128)));
-+double event_log_double        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
-+extern int log_events;
- 
- #define core_cpu_event_start(my_core_id)                      \
--  event_start_time[my_core_id] = get_current_cpu_time();      \
-+  event_start_time_double[my_core_id] = get_current_cpu_time_double();      \
- 
- #define core_cpu_event_end(my_core_id)                        \
--  event_end_time[my_core_id] = get_current_cpu_time();        \
-+  event_end_time_double[my_core_id] = get_current_cpu_time_double();        \
- 
- #define core_gpu_event_start(my_core_id, e1, e2)              \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_start_time[my_core_id] = ctime/1000.+dtime;           \
-+  event_start_time_double[my_core_id] = ctime/1000.+dtime;           \
- 
- #define core_gpu_event_end(my_core_id, e1, e2)                \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_end_time[my_core_id] = ctime/1000.+dtime;             \
-+  event_end_time_double[my_core_id] = ctime/1000.+dtime;             \
- 
- #define core_log_event(event, my_core_id)             \
--  event_log[my_core_id][event_num[my_core_id]+0] = my_core_id;\
--  event_log[my_core_id][event_num[my_core_id]+1] = event_start_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+2] = event_end_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+3] = (event);\
-+  event_log_double[my_core_id][event_num[my_core_id]+0] = my_core_id;\
-+  event_log_double[my_core_id][event_num[my_core_id]+1] = event_start_time_double[my_core_id];\
-+  event_log_double[my_core_id][event_num[my_core_id]+2] = event_end_time_double[my_core_id];\
-+  event_log_double[my_core_id][event_num[my_core_id]+3] = (event);\
-   event_num[my_core_id] += (log_events << 2);         \
-   event_num[my_core_id] &= (MAX_EVENTS-1);
- 
--void dump_trace(int cores_num)
-+void dump_trace_double(int cores_num)
- {
-   char trace_file_name[32];
-   FILE *trace_file;
-@@ -82,13 +82,13 @@
-   for (core = 0; core < cores_num; core++)
-     for (event = 0; event < event_num[core]; event += 4)
-       {
--        int    thread = event_log[core][event+0];
--        double start  = event_log[core][event+1];
--        double end    = event_log[core][event+2];
--        int    color  = event_log[core][event+3];
-+        int    thread = event_log_double[core][event+0];
-+        double start  = event_log_double[core][event+1];
-+        double end    = event_log_double[core][event+2];
-+        int    color  = event_log_double[core][event+3];
- 
--        start -= event_log[core][2];
--        end   -= event_log[core][2];
-+        start -= event_log_double[core][2];
-+        end   -= event_log_double[core][2];
-         /*
-             fprintf(trace_file,
-                 "    "
-@@ -279,7 +279,7 @@
-     core_cpu_event_end(num_gpus);
-     core_log_event(0x666666, num_gpus);
-     
--    dtime = get_current_cpu_time();
-+    dtime = get_current_cpu_time_double();
- 
-     for(j=0; j<num_gpus; j++){
-       magma_setdevice(j);
-@@ -536,7 +536,7 @@
-     }
- 
-     magma_setdevice(cdevice);
--    dump_trace(num_gpus+1);
-+    dump_trace_double(num_gpus+1);
- 
-     return *info;
- } /* magma_dgeqrf2_mgpu */
-diff -Nur src.orig/dlaex0_m.cpp src/dlaex0_m.cpp
---- src.orig/dlaex0_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/dlaex0_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -23,10 +23,7 @@
- 
-     int magma_get_dlaex3_m_nb();
- 
--    magma_int_t get_dlaex0_smlsize()
--    {
--        return 25;
--    }
-+    magma_int_t get_dlaex0_smlsize();
- }
- 
- extern "C" magma_int_t
-diff -Nur src.orig/dlaex3_m.cpp src/dlaex3_m.cpp
---- src.orig/dlaex3_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/dlaex3_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -36,7 +36,7 @@
- }
- 
- extern"C"{
--    void dvrange(magma_int_t k, double *d, magma_int_t *il, magma_int_t *iu, double vl, double vu)
-+    void ddvrange(magma_int_t k, double *d, magma_int_t *il, magma_int_t *iu, double vl, double vu)
-     {
-         magma_int_t i;
- 
-@@ -53,7 +53,7 @@
-         return;
-     }
- 
--    void dirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
-+    void ddirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
-     {
-         magma_int_t i;
- 
-@@ -364,9 +364,9 @@
- 
-     //compute the lower and upper bound of the non-deflated eigenvectors
-     if (valeig)
--        dvrange(k, d, &iil, &iiu, vl, vu);
-+        ddvrange(k, d, &iil, &iiu, vl, vu);
-     else if (indeig)
--        dirange(k, indxq, &iil, &iiu, il, iu);
-+        ddirange(k, indxq, &iil, &iiu, il, iu);
-     else {
-         iil = 1;
-         iiu = k;
-diff -Nur src.orig/dstedx_m.cpp src/dstedx_m.cpp
---- src.orig/dstedx_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/dstedx_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -26,10 +26,7 @@
- 
-     double lapackf77_dlanst( char* norm, magma_int_t* n, double* d, double* e);
- 
--    magma_int_t get_dstedx_smlsize()
--    {
--        return 25;
--    }
-+  magma_int_t get_dstedx_smlsize();
- }
- 
- extern "C" magma_int_t
-diff -Nur src.orig/sgeqrf_mgpu-trace.cpp src/sgeqrf_mgpu-trace.cpp
---- src.orig/sgeqrf_mgpu-trace.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/sgeqrf_mgpu-trace.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -16,7 +16,7 @@
- #include <sys/time.h>
- #include <assert.h>
- 
--float get_current_cpu_time(void)
-+float get_current_cpu_time_float(void)
- {
-   struct timeval  time_val;
- 
-@@ -31,34 +31,34 @@
- #define MAX_EVENTS 1048576
- 
- int    event_num        [MAX_THREADS]               __attribute__ ((aligned (128)));
--float event_start_time [MAX_THREADS]               __attribute__ ((aligned (128)));
--float event_end_time   [MAX_THREADS]               __attribute__ ((aligned (128)));
--float event_log        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
-+float event_start_time_float [MAX_THREADS]               __attribute__ ((aligned (128)));
-+float event_end_time_float   [MAX_THREADS]               __attribute__ ((aligned (128)));
-+float event_log_float        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
- int log_events = 1;
- 
- #define core_cpu_event_start(my_core_id)                      \
--  event_start_time[my_core_id] = get_current_cpu_time();      \
-+  event_start_time_float[my_core_id] = get_current_cpu_time_float();      \
- 
- #define core_cpu_event_end(my_core_id)                        \
--  event_end_time[my_core_id] = get_current_cpu_time();        \
-+  event_end_time_float[my_core_id] = get_current_cpu_time_float();        \
- 
- #define core_gpu_event_start(my_core_id, e1, e2)              \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_start_time[my_core_id] = ctime/1000.+dtime;           \
-+  event_start_time_float[my_core_id] = ctime/1000.+dtime;           \
- 
- #define core_gpu_event_end(my_core_id, e1, e2)                \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_end_time[my_core_id] = ctime/1000.+dtime;             \
-+  event_end_time_float[my_core_id] = ctime/1000.+dtime;             \
- 
- #define core_log_event(event, my_core_id)             \
--  event_log[my_core_id][event_num[my_core_id]+0] = my_core_id;\
--  event_log[my_core_id][event_num[my_core_id]+1] = event_start_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+2] = event_end_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+3] = (event);\
-+  event_log_float[my_core_id][event_num[my_core_id]+0] = my_core_id;\
-+  event_log_float[my_core_id][event_num[my_core_id]+1] = event_start_time_float[my_core_id];\
-+  event_log_float[my_core_id][event_num[my_core_id]+2] = event_end_time_float[my_core_id];\
-+  event_log_float[my_core_id][event_num[my_core_id]+3] = (event);\
-   event_num[my_core_id] += (log_events << 2);         \
-   event_num[my_core_id] &= (MAX_EVENTS-1);
- 
--void dump_trace(int cores_num)
-+void dump_trace_float(int cores_num)
- {
-   char trace_file_name[32];
-   FILE *trace_file;
-@@ -82,13 +82,13 @@
-   for (core = 0; core < cores_num; core++)
-     for (event = 0; event < event_num[core]; event += 4)
-       {
--        int    thread = event_log[core][event+0];
--        float start  = event_log[core][event+1];
--        float end    = event_log[core][event+2];
--        int    color  = event_log[core][event+3];
-+        int    thread = event_log_float[core][event+0];
-+        float start  = event_log_float[core][event+1];
-+        float end    = event_log_float[core][event+2];
-+        int    color  = event_log_float[core][event+3];
- 
--        start -= event_log[core][2];
--        end   -= event_log[core][2];
-+        start -= event_log_float[core][2];
-+        end   -= event_log_float[core][2];
-         /*
-             fprintf(trace_file,
-                 "    "
-@@ -279,7 +279,7 @@
-     core_cpu_event_end(num_gpus);
-     core_log_event(0x666666, num_gpus);
-     
--    dtime = get_current_cpu_time();
-+    dtime = get_current_cpu_time_float();
- 
-     for(j=0; j<num_gpus; j++){
-       magma_setdevice(j);
-@@ -536,7 +536,7 @@
-     }
- 
-     magma_setdevice(cdevice);
--    dump_trace(num_gpus+1);
-+    dump_trace_float(num_gpus+1);
- 
-     return *info;
- } /* magma_sgeqrf2_mgpu */
-diff -Nur src.orig/slaex0_m.cpp src/slaex0_m.cpp
---- src.orig/slaex0_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/slaex0_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -23,10 +23,7 @@
- 
-     int magma_get_slaex3_m_nb();
- 
--    magma_int_t get_slaex0_smlsize()
--    {
--        return 25;
--    }
-+    magma_int_t get_slaex0_smlsize();
- }
- 
- extern "C" magma_int_t
-diff -Nur src.orig/slaex3_m.cpp src/slaex3_m.cpp
---- src.orig/slaex3_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/slaex3_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -36,7 +36,7 @@
- }
- 
- extern"C"{
--    void dvrange(magma_int_t k, float *d, magma_int_t *il, magma_int_t *iu, float vl, float vu)
-+    void sdvrange(magma_int_t k, float *d, magma_int_t *il, magma_int_t *iu, float vl, float vu)
-     {
-         magma_int_t i;
- 
-@@ -53,7 +53,7 @@
-         return;
-     }
- 
--    void dirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
-+    void sdirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
-     {
-         magma_int_t i;
- 
-@@ -364,9 +364,9 @@
- 
-     //compute the lower and upper bound of the non-deflated eigenvectors
-     if (valeig)
--        dvrange(k, d, &iil, &iiu, vl, vu);
-+        sdvrange(k, d, &iil, &iiu, vl, vu);
-     else if (indeig)
--        dirange(k, indxq, &iil, &iiu, il, iu);
-+        sdirange(k, indxq, &iil, &iiu, il, iu);
-     else {
-         iil = 1;
-         iiu = k;
-diff -Nur src.orig/sstedx_m.cpp src/sstedx_m.cpp
---- src.orig/sstedx_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/sstedx_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -26,10 +26,7 @@
- 
-     float lapackf77_slanst( char* norm, magma_int_t* n, float* d, float* e);
- 
--    magma_int_t get_sstedx_smlsize()
--    {
--        return 25;
--    }
-+    magma_int_t get_sstedx_smlsize();
- }
- 
- extern "C" magma_int_t
-diff -Nur src.orig/zgeqrf_mgpu-trace.cpp src/zgeqrf_mgpu-trace.cpp
---- src.orig/zgeqrf_mgpu-trace.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zgeqrf_mgpu-trace.cpp	2012-06-25 17:20:18.000000000 +0100
-@@ -16,108 +16,42 @@
- #include <sys/time.h>
- #include <assert.h>
- 
--double get_current_cpu_time(void)
--{
--  struct timeval  time_val;
--
--  gettimeofday(&time_val, NULL);
--
--  return (double)(time_val.tv_sec) + (double)(time_val.tv_usec) / 1000000.0;
--}
-+extern double get_current_cpu_time_double(void);
- 
- #define MAX_THREADS 5
- 
- //#define MAX_EVENTS 163840
- #define MAX_EVENTS 1048576
- 
--int    event_num        [MAX_THREADS]               __attribute__ ((aligned (128)));
--double event_start_time [MAX_THREADS]               __attribute__ ((aligned (128)));
--double event_end_time   [MAX_THREADS]               __attribute__ ((aligned (128)));
--double event_log        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
--int log_events = 1;
-+extern int    event_num        [MAX_THREADS]               __attribute__ ((aligned (128)));
-+extern double event_start_time_double [MAX_THREADS]               __attribute__ ((aligned (128)));
-+extern double event_end_time_double   [MAX_THREADS]               __attribute__ ((aligned (128)));
-+extern double event_log_double        [MAX_THREADS][MAX_EVENTS]   __attribute__ ((aligned (128)));
-+extern int log_events;
- 
- #define core_cpu_event_start(my_core_id)                      \
--  event_start_time[my_core_id] = get_current_cpu_time();      \
-+  event_start_time_double[my_core_id] = get_current_cpu_time_double();      \
- 
- #define core_cpu_event_end(my_core_id)                        \
--  event_end_time[my_core_id] = get_current_cpu_time();        \
-+  event_end_time_double[my_core_id] = get_current_cpu_time_double();        \
- 
- #define core_gpu_event_start(my_core_id, e1, e2)              \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_start_time[my_core_id] = ctime/1000.+dtime;           \
-+  event_start_time_double[my_core_id] = ctime/1000.+dtime;           \
- 
- #define core_gpu_event_end(my_core_id, e1, e2)                \
-   cudaEventElapsedTime(&ctime, e1, e2);                       \
--  event_end_time[my_core_id] = ctime/1000.+dtime;             \
-+  event_end_time_double[my_core_id] = ctime/1000.+dtime;             \
- 
- #define core_log_event(event, my_core_id)             \
--  event_log[my_core_id][event_num[my_core_id]+0] = my_core_id;\
--  event_log[my_core_id][event_num[my_core_id]+1] = event_start_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+2] = event_end_time[my_core_id];\
--  event_log[my_core_id][event_num[my_core_id]+3] = (event);\
-+  event_log_double[my_core_id][event_num[my_core_id]+0] = my_core_id;\
-+  event_log_double[my_core_id][event_num[my_core_id]+1] = event_start_time_double[my_core_id];\
-+  event_log_double[my_core_id][event_num[my_core_id]+2] = event_end_time_double[my_core_id];\
-+  event_log_double[my_core_id][event_num[my_core_id]+3] = (event);\
-   event_num[my_core_id] += (log_events << 2);         \
-   event_num[my_core_id] &= (MAX_EVENTS-1);
- 
--void dump_trace(int cores_num)
--{
--  char trace_file_name[32];
--  FILE *trace_file;
--  int event;
--  int core;
--
--  double scale = 100000.0;
--  double large = 100.0;
--
--  sprintf(trace_file_name, "trace.svg");
--  trace_file = fopen(trace_file_name, "w");
--  assert(trace_file != NULL);
--
--  fprintf(trace_file,
--          "<?xml version=\"1.0\" standalone=\"no\"?>"
--          "<svg version=\"1.1\" baseProfile=\"full\" xmlns=\"http://www.w3.org/2000/svg\" "
--          "xmlns:xlink=\"http://www.w3.org/1999/xlink\" xmlns:ev=\"http://www.w3.org/2001/xml-events\"  "
--          ">\n"
--          "  <g font-size=\"20\">\n");   
--  
--  for (core = 0; core < cores_num; core++)
--    for (event = 0; event < event_num[core]; event += 4)
--      {
--        int    thread = event_log[core][event+0];
--        double start  = event_log[core][event+1];
--        double end    = event_log[core][event+2];
--        int    color  = event_log[core][event+3];
--
--        start -= event_log[core][2];
--        end   -= event_log[core][2];
--        /*
--            fprintf(trace_file,
--                "    "
--                "<rect x=\"%.2lf\" y=\"%.0lf\" width=\"%.2lf\" height=\"%.0lf\" "
--                "fill=\"#%06x\" stroke=\"#000000\" stroke-width=\"1\"/>\n",
--                start * scale,
--                thread * 100.0,
--                (end - start) * scale,
--                90.0,
--                color);
--        */
--        fprintf(trace_file,
--            "    "
--            "<rect x=\"%.2lf\" y=\"%.0lf\" width=\"%.2lf\" height=\"%.0lf\" "
--                //            "fill=\"#%06x\" />\n",
--                "fill=\"#%06x\" stroke=\"#000000\" stroke-width=\"1\"/>\n",
--                start * scale,
--                thread * (large+20.0),
--                (end - start) * scale,
--                large,
--                color);
--      }
--
--  fprintf(trace_file,
--          "  </g>\n"
--          "</svg>\n");
--
--  fclose(trace_file);
--}
-+extern void dump_trace_double(int cores_num);
- 
- //===========================================================================
- 
-@@ -279,7 +213,7 @@
-     core_cpu_event_end(num_gpus);
-     core_log_event(0x666666, num_gpus);
-     
--    dtime = get_current_cpu_time();
-+    dtime = get_current_cpu_time_double();
- 
-     for(j=0; j<num_gpus; j++){
-       magma_setdevice(j);
-@@ -536,7 +470,7 @@
-     }
- 
-     magma_setdevice(cdevice);
--    dump_trace(num_gpus+1);
-+    dump_trace_double(num_gpus+1);
- 
-     return *info;
- } /* magma_zgeqrf2_mgpu */
-diff -Nur src.orig/zheevr_gpu.cpp src/zheevr_gpu.cpp
---- src.orig/zheevr_gpu.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zheevr_gpu.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,36 +12,6 @@
- */
- #include "common_magma.h"
- 
--/* These interfaces are used for TAU profiling */
--extern "C" {
--    void Mylapackf77_zstemr(const char *jobz, const char *range, magma_int_t *n, double *d, double *e,
--                            double *vl, double *vu, magma_int_t *il, magma_int_t *iu, 
--                            magma_int_t *m, double *w, cuDoubleComplex *z, magma_int_t *ldz,
--                            magma_int_t *nzc, magma_int_t *isuppz, magma_int_t *tryrac,
--                            double *work, magma_int_t *lwork, magma_int_t *iwork, 
--                            magma_int_t *liwork, magma_int_t *info)
--    {
--        lapackf77_zstemr(jobz, range, n, d, e, vl, vu, il, iu, m, w, z, ldz, nzc, 
--                         isuppz, tryrac, work, lwork, iwork, liwork, info);
--    }
--
--    void Mylapackf77_zstein(int *n, double *d, double *e, int *m, double *w, int *iblock,
--                            int *isplit, cuDoubleComplex *z, int *ldz, double *work, 
--                            int *iwork, int *ifail, int *info)
--    {
--        lapackf77_zstein(n, d, e, m, w, iblock, isplit, z, ldz, work, iwork, ifail, info);
--    }
--  
--    void Mylapackf77_dstebz(const char *range, const char *order, int *n, double *vl, 
--                            double *vu, int *il, int *iu, double *abstol,
--                            double *d, double *e, int *m, int *nsplit, double *w, 
--                            int *iblock, int *isplit, double *work, int *iwork, int *info)
--    {
--        lapackf77_dstebz(range, order, n, vl, vu, il, iu, abstol, d, e, m,
--                         nsplit, w, iblock, isplit, work, iwork,info);
--    }
--}
--
- extern "C" {
- #ifdef ADD_
- #    define lapackf77_ieeeck   ieeeck_
-@@ -497,7 +467,7 @@
-     else
-       tryrac=0;
-     
--    Mylapackf77_zstemr(jobz_, range_, &n, &rwork[indrdd], &rwork[indree], &vl, &vu, &il, 
-+    lapackf77_zstemr(jobz_, range_, &n, &rwork[indrdd], &rwork[indree], &vl, &vu, &il, 
-                      &iu, m, &w[1], wz, &ldwz, &n, &isuppz[1], &tryrac, &rwork[indrwk],
-                      &llrwork, &iwork[1], &liwork, info);
-     
-@@ -514,10 +484,10 @@
-     printf("B/I\n");
-     *info = 0;
-     
--    Mylapackf77_dstebz(range_, "B", &n, &vl, &vu, &il, &iu, &abstol, &rwork[indrd], &rwork[indre], m,
-+    lapackf77_dstebz(range_, "B", &n, &vl, &vu, &il, &iu, &abstol, &rwork[indrd], &rwork[indre], m,
-                      &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwo], info);
-     
--    Mylapackf77_zstein(&n, &rwork[indrd], &rwork[indre], m, &w[1], &iwork[indibl], &iwork[indisp],
-+    lapackf77_zstein(&n, &rwork[indrd], &rwork[indre], m, &w[1], &iwork[indibl], &iwork[indisp],
-                        wz, &ldwz, &rwork[indrwk], &iwork[indiwo], &iwork[indifl], info);
-       
-       /*        Apply unitary matrix used in reduction to tridiagonal   
-diff -Nur src.orig/zheevx_gpu.cpp src/zheevx_gpu.cpp
---- src.orig/zheevx_gpu.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zheevx_gpu.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,27 +12,6 @@
-  */
- #include "common_magma.h"
- 
--/* These interfaces are used for TAU profiling */
--extern"C"{
--    void Mylapackf77_zstein(magma_int_t *n, double *d, double *e, magma_int_t *m, 
--                            double *w, magma_int_t *iblock, magma_int_t *isplit,
--                            cuDoubleComplex *z, magma_int_t *ldz, double *work, 
--                            magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info)
--    {
--        lapackf77_zstein(n, d, e, m, w, iblock, isplit, z, ldz, work, iwork, ifail, info);
--    }
--
--    void Mylapackf77_dstebz(char *range, char *order, magma_int_t *n, double *vl,
--                            double *vu, magma_int_t *il, magma_int_t *iu, double *abstol,
--                            double *d, double *e, magma_int_t *m, magma_int_t *nsplit, 
--                            double *w, magma_int_t *iblock, magma_int_t *isplit, 
--                            double *work, magma_int_t *iwork, magma_int_t *info)
--    {
--        lapackf77_dstebz(range, order, n, vl, vu, il, iu, abstol, 
--                         d, e, m, nsplit, w, iblock, isplit, work, iwork,info);
--    }
--}
--
- extern "C" magma_int_t 
- magma_zheevx_gpu(char jobz, char range, char uplo, magma_int_t n, 
-                  cuDoubleComplex *da, magma_int_t ldda, double vl, double vu, 
-@@ -404,12 +383,12 @@
-     indisp = indibl + n;
-     indiwk = indisp + n;
- 
--    Mylapackf77_dstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
-+    lapackf77_dstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
-                      &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwk], info);
-     
-     if (wantz) {
-       
--      Mylapackf77_zstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
-+      lapackf77_zstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
-                        wz, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info);
-       
-       magma_zsetmatrix( n, *m, wz, ldwz, dz, lddz );
-diff -Nur src.orig/zhegvd.cpp src/zhegvd.cpp
---- src.orig/zhegvd.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zhegvd.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -13,26 +13,6 @@
- */
- #include "common_magma.h"
- 
--/* This ztrmm interface is used for TAU profiling */
--void Mymagma_ztrmm(char side, char uplo, char trans, char unit,
--                   magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--  magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--  magma_device_sync();
--}
--
--/* This ztrsm interface is used for TAU profiling */
--void Mymagma_ztrsm(char side, char uplo, char trans, char unit,
--                   magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--  magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--  magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_zhegvd(magma_int_t itype, char jobz, char uplo, magma_int_t n,
-              cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb,
-@@ -324,7 +304,7 @@
-                 *(unsigned char *)trans = MagmaNoTrans;
-             }
- 
--            Mymagma_ztrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-+            magma_ztrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-                           n, n, c_one, db, lddb, da, ldda);
- 
-         } else if (itype == 3)
-@@ -337,7 +317,7 @@
-                 *(unsigned char *)trans = MagmaConjTrans;
-             }
- 
--            Mymagma_ztrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-+            magma_ztrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
-                           n, n, c_one, db, lddb, da, ldda);
-         }
- 
-diff -Nur src.orig/zhegvdx.cpp src/zhegvdx.cpp
---- src.orig/zhegvdx.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zhegvdx.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,22 +12,6 @@
- */
- #include "common_magma.h"
- 
--void Mymagma_ztrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--    magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
--void Mymagma_ztrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--    magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_zhegvdx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
-               cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb, 
-@@ -363,7 +347,7 @@
-                 *(unsigned char *)trans = MagmaNoTrans;
-             }
- 
--            Mymagma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
-+            magma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
- 
-         } else if (itype == 3) {
- 
-@@ -375,7 +359,7 @@
-                 *(unsigned char *)trans = MagmaConjTrans;
-             }
- 
--            Mymagma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
-+            magma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
- 
-         }
- 
-diff -Nur src.orig/zhegvr.cpp src/zhegvr.cpp
---- src.orig/zhegvr.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zhegvr.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,24 +12,6 @@
-  */
- #include "common_magma.h"
- 
--void Mymagma_ztrmm(char side, char uplo, char trans, char unit, 
--                   magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--    magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
--void Mymagma_ztrsm(char side, char uplo, char trans, char unit, 
--                   magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--    magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_zhegvr(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
-              cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb, 
-@@ -408,7 +390,7 @@
-                 *(unsigned char *)trans = MagmaNoTrans;
-             }
-             
--            Mymagma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-+            magma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-                           db, lddb, dz, lddz);
-       
-         } else if (itype == 3) {
-@@ -421,7 +403,7 @@
-                 *(unsigned char *)trans = MagmaConjTrans;
-             }
-             
--            Mymagma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-+            magma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
-                           db, lddb, dz, lddz);
-         }
-         
-diff -Nur src.orig/zhegvx.cpp src/zhegvx.cpp
---- src.orig/zhegvx.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zhegvx.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -12,22 +12,6 @@
-  */
- #include "common_magma.h"
- 
--void Mymagma_ztrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--    magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
--void Mymagma_ztrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
--                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
--                   cuDoubleComplex *dz, magma_int_t lddz)
--{
--    magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
--    magma_device_sync();
--}
--
- extern "C" magma_int_t
- magma_zhegvx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
-              cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb, 
-@@ -330,7 +314,7 @@
-         *(unsigned char *)trans = MagmaNoTrans;
-       }
-       
--      Mymagma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-+      magma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-       
-     } else if (itype == 3) {
-       
-@@ -343,7 +327,7 @@
-         *(unsigned char *)trans = MagmaConjTrans;
-       }
-       
--      Mymagma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-+      magma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
-       
-     }
-     
-diff -Nur src.orig/zstedx_m.cpp src/zstedx_m.cpp
---- src.orig/zstedx_m.cpp	2012-06-23 21:52:09.000000000 +0100
-+++ src/zstedx_m.cpp	2012-06-25 06:59:15.000000000 +0100
-@@ -16,10 +16,7 @@
-                                double* work, magma_int_t lwork, magma_int_t* iwork, magma_int_t liwork,
-                                magma_int_t* info);
- 
--    magma_int_t get_zstedx_smlsize()
--    {
--        return 25;
--    }
-+    magma_int_t get_zstedx_smlsize();
- }
- 
- extern "C" magma_int_t

diff --git a/sci-libs/magma/files/magma-1.2.1-duplicate-symbols.patch b/sci-libs/magma/files/magma-1.2.1-duplicate-symbols.patch
new file mode 100644
index 0000000..8fd2d0d
--- /dev/null
+++ b/sci-libs/magma/files/magma-1.2.1-duplicate-symbols.patch
@@ -0,0 +1,612 @@
+diff -Nur magma-1.2.1.orig/src/cheevx_gpu.cpp magma-1.2.1/src/cheevx_gpu.cpp
+--- magma-1.2.1.orig/src/cheevx_gpu.cpp	2012-07-03 11:59:28.923659921 -0700
++++ magma-1.2.1/src/cheevx_gpu.cpp	2012-07-03 11:59:45.428746349 -0700
+@@ -12,27 +12,6 @@
+  */
+ #include "common_magma.h"
+ 
+-/* These interfaces are used for TAU profiling */
+-extern"C"{
+-    void Mylapackf77_cstein(magma_int_t *n, float *d, float *e, magma_int_t *m, 
+-                            float *w, magma_int_t *iblock, magma_int_t *isplit,
+-                            cuFloatComplex *z, magma_int_t *ldz, float *work, 
+-                            magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info)
+-    {
+-        lapackf77_cstein(n, d, e, m, w, iblock, isplit, z, ldz, work, iwork, ifail, info);
+-    }
+-
+-    void Mylapackf77_sstebz(char *range, char *order, magma_int_t *n, float *vl,
+-                            float *vu, magma_int_t *il, magma_int_t *iu, float *abstol,
+-                            float *d, float *e, magma_int_t *m, magma_int_t *nsplit, 
+-                            float *w, magma_int_t *iblock, magma_int_t *isplit, 
+-                            float *work, magma_int_t *iwork, magma_int_t *info)
+-    {
+-        lapackf77_sstebz(range, order, n, vl, vu, il, iu, abstol, 
+-                         d, e, m, nsplit, w, iblock, isplit, work, iwork,info);
+-    }
+-}
+-
+ extern "C" magma_int_t 
+ magma_cheevx_gpu(char jobz, char range, char uplo, magma_int_t n, 
+                  cuFloatComplex *da, magma_int_t ldda, float vl, float vu, 
+@@ -404,12 +383,12 @@
+     indisp = indibl + n;
+     indiwk = indisp + n;
+ 
+-    Mylapackf77_sstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
++    lapackf77_sstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
+                      &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwk], info);
+     
+     if (wantz) {
+       
+-      Mylapackf77_cstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
++      lapackf77_cstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
+                        wz, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info);
+       
+       magma_csetmatrix( n, *m, wz, ldwz, dz, lddz );
+diff -Nur magma-1.2.1.orig/src/chegvd.cpp magma-1.2.1/src/chegvd.cpp
+--- magma-1.2.1.orig/src/chegvd.cpp	2012-07-03 11:59:28.932659966 -0700
++++ magma-1.2.1/src/chegvd.cpp	2012-07-03 11:59:45.428746349 -0700
+@@ -13,26 +13,6 @@
+ */
+ #include "common_magma.h"
+ 
+-/* This ctrmm interface is used for TAU profiling */
+-void Mymagma_ctrmm(char side, char uplo, char trans, char unit,
+-                   magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-  magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-  magma_device_sync();
+-}
+-
+-/* This ctrsm interface is used for TAU profiling */
+-void Mymagma_ctrsm(char side, char uplo, char trans, char unit,
+-                   magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-  magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-  magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_chegvd(magma_int_t itype, char jobz, char uplo, magma_int_t n,
+              cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb,
+@@ -324,7 +304,7 @@
+                 *(unsigned char *)trans = MagmaNoTrans;
+             }
+ 
+-            Mymagma_ctrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
++            magma_ctrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
+                           n, n, c_one, db, lddb, da, ldda);
+ 
+         } else if (itype == 3)
+@@ -337,7 +317,7 @@
+                 *(unsigned char *)trans = MagmaConjTrans;
+             }
+ 
+-            Mymagma_ctrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
++            magma_ctrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
+                           n, n, c_one, db, lddb, da, ldda);
+         }
+ 
+diff -Nur magma-1.2.1.orig/src/chegvdx.cpp magma-1.2.1/src/chegvdx.cpp
+--- magma-1.2.1.orig/src/chegvdx.cpp	2012-07-03 11:59:28.921659909 -0700
++++ magma-1.2.1/src/chegvdx.cpp	2012-07-03 11:59:45.428746349 -0700
+@@ -12,22 +12,6 @@
+ */
+ #include "common_magma.h"
+ 
+-void Mymagma_ctrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-    magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+-void Mymagma_ctrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-    magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_chegvdx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
+               cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb, 
+@@ -363,7 +347,7 @@
+                 *(unsigned char *)trans = MagmaNoTrans;
+             }
+ 
+-            Mymagma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
++            magma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
+ 
+         } else if (itype == 3) {
+ 
+@@ -375,7 +359,7 @@
+                 *(unsigned char *)trans = MagmaConjTrans;
+             }
+ 
+-            Mymagma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
++            magma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
+ 
+         }
+ 
+diff -Nur magma-1.2.1.orig/src/chegvr.cpp magma-1.2.1/src/chegvr.cpp
+--- magma-1.2.1.orig/src/chegvr.cpp	2012-07-03 11:59:28.932659966 -0700
++++ magma-1.2.1/src/chegvr.cpp	2012-07-03 11:59:45.429746354 -0700
+@@ -12,24 +12,6 @@
+  */
+ #include "common_magma.h"
+ 
+-void Mymagma_ctrmm(char side, char uplo, char trans, char unit, 
+-                   magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-    magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+-void Mymagma_ctrsm(char side, char uplo, char trans, char unit, 
+-                   magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-    magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_chegvr(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
+              cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb, 
+@@ -408,7 +390,7 @@
+                 *(unsigned char *)trans = MagmaNoTrans;
+             }
+             
+-            Mymagma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
++            magma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
+                           db, lddb, dz, lddz);
+       
+         } else if (itype == 3) {
+@@ -421,7 +403,7 @@
+                 *(unsigned char *)trans = MagmaConjTrans;
+             }
+             
+-            Mymagma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
++            magma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
+                           db, lddb, dz, lddz);
+         }
+         
+diff -Nur magma-1.2.1.orig/src/chegvx.cpp magma-1.2.1/src/chegvx.cpp
+--- magma-1.2.1.orig/src/chegvx.cpp	2012-07-03 11:59:28.923659921 -0700
++++ magma-1.2.1/src/chegvx.cpp	2012-07-03 11:59:45.429746354 -0700
+@@ -12,22 +12,6 @@
+  */
+ #include "common_magma.h"
+ 
+-void Mymagma_ctrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb,
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-    magma_ctrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+-void Mymagma_ctrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuFloatComplex alpha, cuFloatComplex *db, magma_int_t lddb, 
+-                   cuFloatComplex *dz, magma_int_t lddz)
+-{
+-    magma_ctrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_chegvx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
+              cuFloatComplex *a, magma_int_t lda, cuFloatComplex *b, magma_int_t ldb, 
+@@ -330,7 +314,7 @@
+         *(unsigned char *)trans = MagmaNoTrans;
+       }
+       
+-      Mymagma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
++      magma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
+       
+     } else if (itype == 3) {
+       
+@@ -343,7 +327,7 @@
+         *(unsigned char *)trans = MagmaConjTrans;
+       }
+       
+-      Mymagma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
++      magma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
+       
+     }
+     
+diff -Nur magma-1.2.1.orig/src/cstedx_m.cpp magma-1.2.1/src/cstedx_m.cpp
+--- magma-1.2.1.orig/src/cstedx_m.cpp	2012-07-03 11:59:28.921659909 -0700
++++ magma-1.2.1/src/cstedx_m.cpp	2012-07-03 11:59:45.429746354 -0700
+@@ -16,10 +16,7 @@
+                                float* work, magma_int_t lwork, magma_int_t* iwork, magma_int_t liwork,
+                                magma_int_t* info);
+ 
+-    magma_int_t get_cstedx_smlsize()
+-    {
+-        return 25;
+-    }
++   magma_int_t get_cstedx_smlsize();
+ }
+ 
+ extern "C" magma_int_t
+diff -Nur magma-1.2.1.orig/src/dlaex0_m.cpp magma-1.2.1/src/dlaex0_m.cpp
+--- magma-1.2.1.orig/src/dlaex0_m.cpp	2012-07-03 11:59:28.933659971 -0700
++++ magma-1.2.1/src/dlaex0_m.cpp	2012-07-03 11:59:45.430746359 -0700
+@@ -22,10 +22,7 @@
+ 
+     int magma_get_dlaex3_m_nb();
+ 
+-    magma_int_t get_dlaex0_smlsize()
+-    {
+-        return 25;
+-    }
++    magma_int_t get_dlaex0_smlsize();
+ }
+ 
+ extern "C" magma_int_t
+diff -Nur magma-1.2.1.orig/src/dlaex3_m.cpp magma-1.2.1/src/dlaex3_m.cpp
+--- magma-1.2.1.orig/src/dlaex3_m.cpp	2012-07-03 11:59:28.929659951 -0700
++++ magma-1.2.1/src/dlaex3_m.cpp	2012-07-03 11:59:45.430746359 -0700
+@@ -34,7 +34,7 @@
+ }
+ 
+ extern"C"{
+-    void dvrange(magma_int_t k, double *d, magma_int_t *il, magma_int_t *iu, double vl, double vu)
++    void ddvrange(magma_int_t k, double *d, magma_int_t *il, magma_int_t *iu, double vl, double vu)
+     {
+         magma_int_t i;
+ 
+@@ -51,7 +51,7 @@
+         return;
+     }
+ 
+-    void dirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
++    void ddirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
+     {
+         magma_int_t i;
+ 
+@@ -362,9 +362,9 @@
+ 
+     //compute the lower and upper bound of the non-deflated eigenvectors
+     if (valeig)
+-        dvrange(k, d, &iil, &iiu, vl, vu);
++        ddvrange(k, d, &iil, &iiu, vl, vu);
+     else if (indeig)
+-        dirange(k, indxq, &iil, &iiu, il, iu);
++        ddirange(k, indxq, &iil, &iiu, il, iu);
+     else {
+         iil = 1;
+         iiu = k;
+diff -Nur magma-1.2.1.orig/src/dstedx_m.cpp magma-1.2.1/src/dstedx_m.cpp
+--- magma-1.2.1.orig/src/dstedx_m.cpp	2012-07-03 11:59:28.920659903 -0700
++++ magma-1.2.1/src/dstedx_m.cpp	2012-07-03 11:59:45.430746359 -0700
+@@ -26,10 +26,7 @@
+ 
+     double lapackf77_dlanst( char* norm, magma_int_t* n, double* d, double* e);
+ 
+-    magma_int_t get_dstedx_smlsize()
+-    {
+-        return 25;
+-    }
++  magma_int_t get_dstedx_smlsize();
+ }
+ 
+ extern "C" magma_int_t
+diff -Nur magma-1.2.1.orig/src/slaex0_m.cpp magma-1.2.1/src/slaex0_m.cpp
+--- magma-1.2.1.orig/src/slaex0_m.cpp	2012-07-03 11:59:28.931659961 -0700
++++ magma-1.2.1/src/slaex0_m.cpp	2012-07-03 11:59:45.431746364 -0700
+@@ -22,10 +22,7 @@
+ 
+     int magma_get_slaex3_m_nb();
+ 
+-    magma_int_t get_slaex0_smlsize()
+-    {
+-        return 25;
+-    }
++    magma_int_t get_slaex0_smlsize();
+ }
+ 
+ extern "C" magma_int_t
+diff -Nur magma-1.2.1.orig/src/slaex3_m.cpp magma-1.2.1/src/slaex3_m.cpp
+--- magma-1.2.1.orig/src/slaex3_m.cpp	2012-07-03 11:59:28.920659903 -0700
++++ magma-1.2.1/src/slaex3_m.cpp	2012-07-03 11:59:45.431746364 -0700
+@@ -34,7 +34,7 @@
+ }
+ 
+ extern"C"{
+-    void dvrange(magma_int_t k, float *d, magma_int_t *il, magma_int_t *iu, float vl, float vu)
++    void sdvrange(magma_int_t k, float *d, magma_int_t *il, magma_int_t *iu, float vl, float vu)
+     {
+         magma_int_t i;
+ 
+@@ -51,7 +51,7 @@
+         return;
+     }
+ 
+-    void dirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
++    void sdirange(magma_int_t k, magma_int_t* indxq, magma_int_t *iil, magma_int_t *iiu, magma_int_t il, magma_int_t iu)
+     {
+         magma_int_t i;
+ 
+@@ -362,9 +362,9 @@
+ 
+     //compute the lower and upper bound of the non-deflated eigenvectors
+     if (valeig)
+-        dvrange(k, d, &iil, &iiu, vl, vu);
++        sdvrange(k, d, &iil, &iiu, vl, vu);
+     else if (indeig)
+-        dirange(k, indxq, &iil, &iiu, il, iu);
++        sdirange(k, indxq, &iil, &iiu, il, iu);
+     else {
+         iil = 1;
+         iiu = k;
+diff -Nur magma-1.2.1.orig/src/sstedx_m.cpp magma-1.2.1/src/sstedx_m.cpp
+--- magma-1.2.1.orig/src/sstedx_m.cpp	2012-07-03 11:59:28.932659966 -0700
++++ magma-1.2.1/src/sstedx_m.cpp	2012-07-03 11:59:45.431746364 -0700
+@@ -26,10 +26,7 @@
+ 
+     float lapackf77_slanst( char* norm, magma_int_t* n, float* d, float* e);
+ 
+-    magma_int_t get_sstedx_smlsize()
+-    {
+-        return 25;
+-    }
++    magma_int_t get_sstedx_smlsize();
+ }
+ 
+ extern "C" magma_int_t
+diff -Nur magma-1.2.1.orig/src/zheevx_gpu.cpp magma-1.2.1/src/zheevx_gpu.cpp
+--- magma-1.2.1.orig/src/zheevx_gpu.cpp	2012-07-03 11:59:28.934659976 -0700
++++ magma-1.2.1/src/zheevx_gpu.cpp	2012-07-03 11:59:45.431746364 -0700
+@@ -12,27 +12,6 @@
+  */
+ #include "common_magma.h"
+ 
+-/* These interfaces are used for TAU profiling */
+-extern"C"{
+-    void Mylapackf77_zstein(magma_int_t *n, double *d, double *e, magma_int_t *m, 
+-                            double *w, magma_int_t *iblock, magma_int_t *isplit,
+-                            cuDoubleComplex *z, magma_int_t *ldz, double *work, 
+-                            magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info)
+-    {
+-        lapackf77_zstein(n, d, e, m, w, iblock, isplit, z, ldz, work, iwork, ifail, info);
+-    }
+-
+-    void Mylapackf77_dstebz(char *range, char *order, magma_int_t *n, double *vl,
+-                            double *vu, magma_int_t *il, magma_int_t *iu, double *abstol,
+-                            double *d, double *e, magma_int_t *m, magma_int_t *nsplit, 
+-                            double *w, magma_int_t *iblock, magma_int_t *isplit, 
+-                            double *work, magma_int_t *iwork, magma_int_t *info)
+-    {
+-        lapackf77_dstebz(range, order, n, vl, vu, il, iu, abstol, 
+-                         d, e, m, nsplit, w, iblock, isplit, work, iwork,info);
+-    }
+-}
+-
+ extern "C" magma_int_t 
+ magma_zheevx_gpu(char jobz, char range, char uplo, magma_int_t n, 
+                  cuDoubleComplex *da, magma_int_t ldda, double vl, double vu, 
+@@ -404,12 +383,12 @@
+     indisp = indibl + n;
+     indiwk = indisp + n;
+ 
+-    Mylapackf77_dstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
++    lapackf77_dstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m,
+                      &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwk], info);
+     
+     if (wantz) {
+       
+-      Mylapackf77_zstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
++      lapackf77_zstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp],
+                        wz, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info);
+       
+       magma_zsetmatrix( n, *m, wz, ldwz, dz, lddz );
+diff -Nur magma-1.2.1.orig/src/zhegvd.cpp magma-1.2.1/src/zhegvd.cpp
+--- magma-1.2.1.orig/src/zhegvd.cpp	2012-07-03 11:59:28.920659903 -0700
++++ magma-1.2.1/src/zhegvd.cpp	2012-07-03 11:59:45.432746370 -0700
+@@ -13,26 +13,6 @@
+ */
+ #include "common_magma.h"
+ 
+-/* This ztrmm interface is used for TAU profiling */
+-void Mymagma_ztrmm(char side, char uplo, char trans, char unit,
+-                   magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-  magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-  magma_device_sync();
+-}
+-
+-/* This ztrsm interface is used for TAU profiling */
+-void Mymagma_ztrsm(char side, char uplo, char trans, char unit,
+-                   magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-  magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-  magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_zhegvd(magma_int_t itype, char jobz, char uplo, magma_int_t n,
+              cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb,
+@@ -324,7 +304,7 @@
+                 *(unsigned char *)trans = MagmaNoTrans;
+             }
+ 
+-            Mymagma_ztrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
++            magma_ztrsm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
+                           n, n, c_one, db, lddb, da, ldda);
+ 
+         } else if (itype == 3)
+@@ -337,7 +317,7 @@
+                 *(unsigned char *)trans = MagmaConjTrans;
+             }
+ 
+-            Mymagma_ztrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
++            magma_ztrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit,
+                           n, n, c_one, db, lddb, da, ldda);
+         }
+ 
+diff -Nur magma-1.2.1.orig/src/zhegvdx.cpp magma-1.2.1/src/zhegvdx.cpp
+--- magma-1.2.1.orig/src/zhegvdx.cpp	2012-07-03 11:59:28.922659915 -0700
++++ magma-1.2.1/src/zhegvdx.cpp	2012-07-03 11:59:45.432746370 -0700
+@@ -12,22 +12,6 @@
+ */
+ #include "common_magma.h"
+ 
+-void Mymagma_ztrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-    magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+-void Mymagma_ztrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-    magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_zhegvdx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
+               cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb, 
+@@ -363,7 +347,7 @@
+                 *(unsigned char *)trans = MagmaNoTrans;
+             }
+ 
+-            Mymagma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
++            magma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
+ 
+         } else if (itype == 3) {
+ 
+@@ -375,7 +359,7 @@
+                 *(unsigned char *)trans = MagmaConjTrans;
+             }
+ 
+-            Mymagma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
++            magma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, da, ldda);
+ 
+         }
+ 
+diff -Nur magma-1.2.1.orig/src/zhegvr.cpp magma-1.2.1/src/zhegvr.cpp
+--- magma-1.2.1.orig/src/zhegvr.cpp	2012-07-03 11:59:28.930659956 -0700
++++ magma-1.2.1/src/zhegvr.cpp	2012-07-03 11:59:45.432746370 -0700
+@@ -12,24 +12,6 @@
+  */
+ #include "common_magma.h"
+ 
+-void Mymagma_ztrmm(char side, char uplo, char trans, char unit, 
+-                   magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-    magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+-void Mymagma_ztrsm(char side, char uplo, char trans, char unit, 
+-                   magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-    magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_zhegvr(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
+              cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb, 
+@@ -408,7 +390,7 @@
+                 *(unsigned char *)trans = MagmaNoTrans;
+             }
+             
+-            Mymagma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
++            magma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
+                           db, lddb, dz, lddz);
+       
+         } else if (itype == 3) {
+@@ -421,7 +403,7 @@
+                 *(unsigned char *)trans = MagmaConjTrans;
+             }
+             
+-            Mymagma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
++            magma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, 
+                           db, lddb, dz, lddz);
+         }
+         
+diff -Nur magma-1.2.1.orig/src/zhegvx.cpp magma-1.2.1/src/zhegvx.cpp
+--- magma-1.2.1.orig/src/zhegvx.cpp	2012-07-03 11:59:28.914659873 -0700
++++ magma-1.2.1/src/zhegvx.cpp	2012-07-03 11:59:45.433746376 -0700
+@@ -12,22 +12,6 @@
+  */
+ #include "common_magma.h"
+ 
+-void Mymagma_ztrmm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb,
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-    magma_ztrmm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+-void Mymagma_ztrsm(char side, char uplo, char trans, char unit, magma_int_t n, magma_int_t m,
+-                   cuDoubleComplex alpha, cuDoubleComplex *db, magma_int_t lddb, 
+-                   cuDoubleComplex *dz, magma_int_t lddz)
+-{
+-    magma_ztrsm(side, uplo, trans, unit, n, m, alpha, db, lddb, dz, lddz);
+-    magma_device_sync();
+-}
+-
+ extern "C" magma_int_t
+ magma_zhegvx(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
+              cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *b, magma_int_t ldb, 
+@@ -330,7 +314,7 @@
+         *(unsigned char *)trans = MagmaNoTrans;
+       }
+       
+-      Mymagma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
++      magma_ztrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
+       
+     } else if (itype == 3) {
+       
+@@ -343,7 +327,7 @@
+         *(unsigned char *)trans = MagmaConjTrans;
+       }
+       
+-      Mymagma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
++      magma_ztrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz);
+       
+     }
+     
+diff -Nur magma-1.2.1.orig/src/zstedx_m.cpp magma-1.2.1/src/zstedx_m.cpp
+--- magma-1.2.1.orig/src/zstedx_m.cpp	2012-07-03 11:59:28.921659909 -0700
++++ magma-1.2.1/src/zstedx_m.cpp	2012-07-03 11:59:45.433746376 -0700
+@@ -16,10 +16,7 @@
+                                double* work, magma_int_t lwork, magma_int_t* iwork, magma_int_t liwork,
+                                magma_int_t* info);
+ 
+-    magma_int_t get_zstedx_smlsize()
+-    {
+-        return 25;
+-    }
++    magma_int_t get_zstedx_smlsize();
+ }
+ 
+ extern "C" magma_int_t

diff --git a/sci-libs/magma/files/magma-1.2.1-no-cuda-driver.patch b/sci-libs/magma/files/magma-1.2.1-no-cuda-driver.patch
new file mode 100644
index 0000000..9dc99a8
--- /dev/null
+++ b/sci-libs/magma/files/magma-1.2.1-no-cuda-driver.patch
@@ -0,0 +1,38 @@
+--- magma-1.2.1.orig/testing/lin/Makefile	2012-07-03 11:59:28.875659669 -0700
++++ magma-1.2.1/testing/lin/Makefile	2012-07-03 12:01:05.546166252 -0700
+@@ -266,7 +266,7 @@
+ 	../matgen/$(TMGLIB) -o xlintsts \
+ 	../fortran.o \
+ 	$(LDOPTS) -L../../lib \
+-	-lcuda -lmagma -lmagmablas \
++	-lmagma -lmagmablas \
+ 	$(LIBDIR) $(LIB) 
+ 
+ xlintstc : $(ALINTST) $(CLINTST) $(SCLNTST) ../fortran.o
+@@ -274,7 +274,7 @@
+ 	../matgen/$(TMGLIB) -o xlintstc \
+ 	../fortran.o \
+ 	$(LDOPTS) -L../../lib \
+-	-lcuda -lmagma -lmagmablas \
++	-lmagma -lmagmablas \
+ 	$(LIBDIR) $(LIB) 
+  
+ xlintstd : $(ALINTST) $(DLINTST) $(DZLNTST) ../fortran.o
+@@ -282,7 +282,7 @@
+ 	../matgen/$(TMGLIB) -o xlintstd \
+ 	../fortran.o \
+ 	$(LDOPTS) -L../../lib \
+-	-lcuda -lmagma -lmagmablas \
++	-lmagma -lmagmablas \
+ 	$(LIBDIR) $(LIB) 
+  
+ xlintstz : $(ALINTST) $(ZLINTST) $(DZLNTST) ../fortran.o
+@@ -290,7 +290,7 @@
+ 	../matgen/$(TMGLIB) -o xlintstz \
+ 	../fortran.o \
+ 	$(LDOPTS) -L../../lib \
+-	-lcuda -lmagma -lmagmablas -lmagma \
++	-lmagma -lmagmablas -lmagma \
+ 	$(LIBDIR) $(LIB) 
+  
+ $(ALINTST): $(FRC)

diff --git a/sci-libs/magma/magma-1.2.0.ebuild b/sci-libs/magma/magma-1.2.1.ebuild
similarity index 87%
rename from sci-libs/magma/magma-1.2.0.ebuild
rename to sci-libs/magma/magma-1.2.1.ebuild
index 18e16f2..65b822c 100644
--- a/sci-libs/magma/magma-1.2.0.ebuild
+++ b/sci-libs/magma/magma-1.2.1.ebuild
@@ -3,25 +3,22 @@
 # $Header: $
 
 EAPI=4
-FORTRAN_STANDARD="77 90"
 
+FORTRAN_STANDARD="77 90"
 inherit eutils fortran-2 toolchain-funcs versionator
 
-MYP=${PN}_${PV}
-
 DESCRIPTION="Matrix Algebra on GPU and Multicore Architectures"
 HOMEPAGE="http://icl.cs.utk.edu/magma/"
-SRC_URI="http://icl.cs.utk.edu/projectsfiles/${PN}/${MYP}.tar.gz"
+SRC_URI="http://icl.cs.utk.edu/projectsfiles/${PN}/${P}.tar.gz"
 
 LICENSE="BSD"
 SLOT="0"
-KEYWORDS="~amd64 ~x86"
+KEYWORDS="~amd64 ~x86 ~amd64-linux ~x86-linux"
 IUSE="fermi static-libs tesla"
 
 RDEPEND="dev-util/nvidia-cuda-toolkit
 	virtual/cblas
 	virtual/lapack"
-
 DEPEND="${RDEPEND}
 	virtual/pkgconfig"
 
@@ -40,8 +37,8 @@ make_shared_lib() {
 
 src_prepare() {
 	epatch \
-		"${FILESDIR}"/${P}-cblas-dotc.patch \
-		"${FILESDIR}"/${P}-duplicate-symbols.patch
+		"${FILESDIR}"/${P}-duplicate-symbols.patch \
+		"${FILESDIR}"/${P}-no-cuda-driver.patch
 
 	# distributed pc file not so useful so replace it
 	cat <<-EOF > ${PN}.pc
@@ -53,7 +50,7 @@ src_prepare() {
 		Version: ${PV}
 		URL: ${HOMEPAGE}
 		Libs: -L\${libdir} -lmagma -lmagmablas
-		Libs.private: -lm -lpthread -ldl -lcublas -lcudart -lcuda
+		Libs.private: -lm -lpthread -ldl -lcublas -lcudart
 		Cflags: -I\${includedir}
 		Requires: cblas lapack
 	EOF
@@ -70,12 +67,13 @@ src_configure() {
 		INC = -I${EPREFIX}/opt/cuda/include -DADD_
 		OPTS = ${CFLAGS} -fPIC
 		FOPTS = ${FFLAGS} -fPIC -x f95-cpp-input
+		F77OPTS = ${FFLAGS} -fPIC
 		NVOPTS = -DADD_ --compiler-options '-fPIC ${CFLAGS}' -DUNIX
 		LOADER = $(tc-getFC)
 		LIBBLAS = $(pkg-config --libs cblas)
 		LIBLAPACK = $(pkg-config --libs lapack)
 		CUDADIR = ${EPREFIX}/opt/cuda
-		LIBCUDA = -L\$(CUDADIR)/$(get_libdir) -lcublas -lcudart -lcuda
+		LIBCUDA = -L\$(CUDADIR)/$(get_libdir) -lcublas -lcudart
 		LIB = -pthread -lm -ldl \$(LIBCUDA) \$(LIBBLAS) \$(LIBLAPACK) -lstdc++
 	EOF
 	if use fermi; then