@@ -40,80 +40,82 @@ def __init__(self, subtest):
4040 },
4141 }
4242 elif subtest == 'zgemm' :
43+ self .executable_opts = ['--range 1088:3136:1024' ]
4344 self .perf_patterns = {
44- 'magma' : sn .extractsingle (r'MAGMA GFlops: (?P<magma_gflops>\S+)' ,
45- self .stdout , 'magma_gflops' , float ),
45+ 'magma' : sn .extractsingle (
46+ r'MAGMA GFlops: (?P<magma_gflops>\S+)' ,
47+ self .stdout , 'magma_gflops' , float , 2
48+ ),
4649 'cublas' : sn .extractsingle (
4750 r'cuBLAS GFlops: (?P<cublas_gflops>\S+)' , self .stdout ,
48- 'cublas_gflops' , float ),
49- 'cpu' : sn .extractsingle (r'CPU GFlops: (?P<cpu_gflops>\S+)' ,
50- self .stdout , 'cpu_gflops' , float )
51+ 'cublas_gflops' , float , 2 )
5152 }
5253 self .reference = {
5354 'daint:gpu' : {
54- 'magma' : (3344 , - 0.05 , None , 'Gflop/s' ),
55- 'cublas' : (3709 , - 0.05 , None , 'Gflop/s' ),
56- 'cpu' : (42.8 , - 0.27 , None , 'Gflop/s' ),
55+ 'magma' : (3692.65 , - 0.05 , None , 'Gflop/s' ),
56+ 'cublas' : (4269.31 , - 0.09 , None , 'Gflop/s' ),
5757 },
5858 'dom:gpu' : {
59- 'magma' : (3344 , - 0.05 , None , 'Gflop/s' ),
60- 'cublas' : (3709 , - 0.05 , None , 'Gflop/s' ),
61- 'cpu' : (42.8 , - 0.27 , None , 'Gflop/s' ),
59+ 'magma' : (3692.65 , - 0.05 , None , 'Gflop/s' ),
60+ 'cublas' : (4269.31 , - 0.09 , None , 'Gflop/s' ),
61+ },
62+ '*' : {
63+ 'magma' : (0 , None , None , 'Gflop/s' ),
64+ 'cublas' : (0 , None , None , 'Gflop/s' ),
6265 },
6366 }
6467 elif subtest == 'zsymmetrize' :
6568 self .perf_patterns = {
66- 'cpu_perf' : sn .extractsingle (r'CPU performance: (\S+)' ,
67- self .stdout , 1 , float ),
6869 'gpu_perf' : sn .extractsingle (r'GPU performance: (\S+)' ,
6970 self .stdout , 1 , float ),
7071 }
7172 self .reference = {
7273 'daint:gpu' : {
73- 'cpu_perf' : (0.91 , - 0.05 , None , 'GB/s' ),
7474 'gpu_perf' : (158.3 , - 0.05 , None , 'GB/s' ),
7575 },
7676 'dom:gpu' : {
77- 'cpu_perf' : (0.91 , - 0.05 , None , 'GB/s' ),
7877 'gpu_perf' : (158.3 , - 0.05 , None , 'GB/s' ),
7978 },
79+ '*' : {
80+ 'gpu_perf' : (0 , None , None , 'GB/s' ),
81+ },
8082 }
8183 elif subtest == 'ztranspose' :
8284 self .perf_patterns = {
83- 'cpu_perf' :
84- sn .extractsingle (r'CPU performance: (?P<cpu_performance>\S+)' ,
85- self .stdout , 'cpu_performance' , float ),
8685 'gpu_perf' :
87- sn .extractsingle (r'GPU performance: (?P<gpu_performance>\S+)' ,
88- self .stdout , 'gpu_performance' , float )
86+ sn .extractsingle (
87+ r'GPU performance: (?P<gpu_performance>\S+)' ,
88+ self .stdout , 'gpu_performance' , float
89+ )
8990 }
9091 self .reference = {
9192 'daint:gpu' : {
92- 'cpu_perf' : (1.51 , - 0.05 , None , 'GB/s' ),
9393 'gpu_perf' : (498.2 , - 0.05 , None , 'GB/s' ),
9494 },
9595 'dom:gpu' : {
96- 'cpu_perf' : (1.51 , - 0.05 , None , 'GB/s' ),
9796 'gpu_perf' : (498.2 , - 0.05 , None , 'GB/s' ),
9897 },
98+ '*' : {
99+ 'gpu_perf' : (0 , None , None , 'GB/s' ),
100+ },
99101 }
100102 elif subtest == 'zunmbr' :
101103 # This test fails to compile with Magma 2.4
102104 self .perf_patterns = {
103- 'cpu_perf' :
104- sn .extractsingle (r'CPU performance: (?P<cpu_performance>\S+)' ,
105- self .stdout , 'cpu_performance' , float ),
106105 'gpu_perf' :
107- sn .extractsingle (r'GPU performance: (?P<gpu_performance>\S+)' ,
108- self .stdout , 'gpu_performance' , float )
106+ sn .extractsingle (
107+ r'GPU performance: (?P<gpu_performance>\S+)' ,
108+ self .stdout , 'gpu_performance' , float
109+ )
109110 }
110111 self .reference = {
111112 'daint:gpu' : {
112- 'cpu_perf' : (36.6 , - 0.05 , None , 'Gflop/s' ),
113113 'gpu_perf' : (254.7 , - 0.05 , None , 'Gflop/s' ),
114114 },
115115 'dom:gpu' : {
116- 'cpu_perf' : (36.6 , - 0.05 , None , 'Gflop/s' ),
117116 'gpu_perf' : (254.7 , - 0.05 , None , 'Gflop/s' ),
118117 },
118+ '*' : {
119+ 'gpu_perf' : (0 , None , None , 'Gflop/s' ),
120+ },
119121 }
0 commit comments