Merge pull request #783 from ajocksch/checks/magma_revised

[test] Revise MAGMA library benchmark

Author: Vasileios Karakasis

cscs-checks/libraries/magma/magma_checks.py

@@ -40,80 +40,82 @@ 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+)',
-                                          self.stdout, 'magma_gflops', float),
+                'magma': sn.extractsingle(
+                    r'MAGMA GFlops: (?P<magma_gflops>\S+)',
+                    self.stdout, 'magma_gflops', float, 2
+                ),
                 'cublas': sn.extractsingle(
                     r'cuBLAS GFlops: (?P<cublas_gflops>\S+)', self.stdout,
-                    'cublas_gflops', float),
-                'cpu': sn.extractsingle(r'CPU GFlops: (?P<cpu_gflops>\S+)',
-                                        self.stdout, 'cpu_gflops', float)
+                    'cublas_gflops', float, 2)
             }
             self.reference = {
                 'daint:gpu': {
-                    'magma':  (3344, -0.05, None, 'Gflop/s'),
-                    'cublas': (3709, -0.05, None, 'Gflop/s'),
-                    'cpu':    (42.8, -0.27, None, 'Gflop/s'),
+                    'magma':  (3692.65, -0.05, None, 'Gflop/s'),
+                    'cublas': (4269.31, -0.09, None, 'Gflop/s'),
                 },
                 'dom:gpu': {
-                    'magma':  (3344, -0.05, None, 'Gflop/s'),
-                    'cublas': (3709, -0.05, None, 'Gflop/s'),
-                    'cpu':    (42.8, -0.27, None, 'Gflop/s'),
+                    'magma':  (3692.65, -0.05, None, 'Gflop/s'),
+                    'cublas': (4269.31, -0.09, None, 'Gflop/s'),
+                },
+                '*': {
+                    'magma':  (0, None, None, 'Gflop/s'),
+                    'cublas': (0, None, None, 'Gflop/s'),
                 },
             }
         elif subtest == 'zsymmetrize':
             self.perf_patterns = {
-                'cpu_perf': sn.extractsingle(r'CPU performance: (\S+)',
-                                             self.stdout, 1, float),
                 'gpu_perf': sn.extractsingle(r'GPU performance: (\S+)',
                                              self.stdout, 1, float),
             }
             self.reference = {
                 'daint:gpu': {
-                    'cpu_perf': (0.91, -0.05, None, 'GB/s'),
                     'gpu_perf': (158.3, -0.05, None, 'GB/s'),
                 },
                 'dom:gpu': {
-                    'cpu_perf': (0.91, -0.05, None, 'GB/s'),
                     'gpu_perf': (158.3, -0.05, None, 'GB/s'),
                 },
+                '*': {
+                    'gpu_perf': (0, None, None, 'GB/s'),
+                },
             }
         elif subtest == 'ztranspose':
             self.perf_patterns = {
-                'cpu_perf':
-                    sn.extractsingle(r'CPU performance: (?P<cpu_performance>\S+)',
-                                     self.stdout, 'cpu_performance', float),
                 'gpu_perf':
-                    sn.extractsingle(r'GPU performance: (?P<gpu_performance>\S+)',
-                                     self.stdout, 'gpu_performance', float)
+                    sn.extractsingle(
+                        r'GPU performance: (?P<gpu_performance>\S+)',
+                        self.stdout, 'gpu_performance', float
+                    )
             }
             self.reference = {
                 'daint:gpu': {
-                    'cpu_perf': (1.51, -0.05, None, 'GB/s'),
                     'gpu_perf': (498.2, -0.05, None, 'GB/s'),
                 },
                 'dom:gpu': {
-                    'cpu_perf': (1.51, -0.05, None, 'GB/s'),
                     'gpu_perf': (498.2, -0.05, None, 'GB/s'),
                 },
+                '*': {
+                    'gpu_perf': (0, None, None, 'GB/s'),
+                },
             }
         elif subtest == 'zunmbr':
             # This test fails to compile with Magma 2.4
             self.perf_patterns = {
-                'cpu_perf':
-                    sn.extractsingle(r'CPU performance: (?P<cpu_performance>\S+)',
-                                     self.stdout, 'cpu_performance', float),
                 'gpu_perf':
-                    sn.extractsingle(r'GPU performance: (?P<gpu_performance>\S+)',
-                                     self.stdout, 'gpu_performance', float)
+                    sn.extractsingle(
+                        r'GPU performance: (?P<gpu_performance>\S+)',
+                        self.stdout, 'gpu_performance', float
+                    )
             }
             self.reference = {
                 'daint:gpu': {
-                    'cpu_perf': (36.6, -0.05, None, 'Gflop/s'),
                     'gpu_perf': (254.7, -0.05, None, 'Gflop/s'),
                 },
                 'dom:gpu': {
-                    'cpu_perf': (36.6, -0.05, None, 'Gflop/s'),
                     'gpu_perf': (254.7, -0.05, None, 'Gflop/s'),
                 },
+                '*': {
+                    'gpu_perf': (0, None, None, 'Gflop/s'),
+                },
             }

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 < opts.ntest; itest+=100 ) {
-         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,