shorter patch; new performance data for zgemm

Author: ajocksch

cscs-checks/libraries/magma/magma_checks.py

@@ -40,6 +40,7 @@ def __init__(self, subtest):
                 },
             }
         elif subtest == 'zgemm':
+            self.executable_opts = ['--range 1088:3136:1024']
             self.perf_patterns = {
                 'magma': sn.extractsingle(
                              r'MAGMA GFlops: (?P<magma_gflops>\S+)',
@@ -51,12 +52,12 @@ def __init__(self, subtest):
             }
             self.reference = {
                 'daint:gpu': {
-                    'magma':  (3749, -0.05, None, 'Gflop/s'),
-                    'cublas': (4050, -0.05, None, 'Gflop/s'),
+                    'magma':  (3692.65, -0.05, None, 'Gflop/s'),
+                    'cublas': (4269.31, -0.0847, None, 'Gflop/s'),
                 },
                 'dom:gpu': {
-                    'magma':  (3749, -0.05, None, 'Gflop/s'),
-                    'cublas': (4050, -0.05, None, 'Gflop/s'),
+                    'magma':  (3692.65, -0.05, None, 'Gflop/s'),
+                    'cublas': (4269.31, -0.0847, None, 'Gflop/s'),
                 },
             }
         elif subtest == 'zsymmetrize':

cscs-checks/libraries/magma/src/patch.txt

@@ -67,16 +67,6 @@ diff -Naur src/testing_zgemm.cpp src_patched/testing_zgemm.cpp
 
      // Allow 3*eps; complex needs 2*sqrt(2) factor; see Higham, 2002, sec. 3.6.
      double eps = lapackf77_dlamch("E");
-@@ -87,7 +88,8 @@
-                 g_platform_str, g_platform_str );
-     #endif
-     printf("%%========================================================================================================\n");
--    for( int itest = 0; itest < opts.ntest; ++itest ) {
-+//ajajajaj    for( int itest = 0; itest < opts.ntest; ++itest ) {
-+    for( int itest = 0; itest < 3; itest++ ) {
-         for( int iter = 0; iter < opts.niter; ++iter ) {
-             M = opts.msize[itest];
-             N = opts.nsize[itest];
 @@ -220,6 +222,9 @@
                             cpu_perf,    1000.*cpu_time,
                             magma_error, dev_error,