add openmp flag and by pass workaround flag

pyop2/codegen/rep2loopy.py

@@ -455,6 +455,15 @@ def generate(builder, wrapper_name=None):
                                 name=wrapper_name,
                                 # TODO, should these really be silenced?
                                 silenced_warnings=["write_race*"])
+    from pyop2.configuration import configuration
+    if configuration["time"]:
+        batch_size = configuration["simd_width"]
+        if builder.extruded:
+            start, end = parameters.layer_start, parameters.layer_end
+        else:
+            start, end = "start", "end"
+        wrapper = loopy.assume(wrapper, "{0} mod {1} = 0".format(end, batch_size))
+        wrapper = loopy.assume(wrapper, "exists zz: zz > 0 and {0} = {1}*zz + {2}".format(end, configuration["simd_width"], start))
 
     # prioritize loops
     for indices in context.index_ordering:

pyop2/compilation.py

@@ -211,7 +211,8 @@ def workaround_cflags(self):
             if version.StrictVersion("7.3") <= ver < version.StrictVersion("7.5"):
                 # GCC bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90055
                 # See also https://github.com/firedrakeproject/firedrake/issues/1442
-                return ["-fno-tree-loop-vectorize"]
+                pass  # enable vectorization for paper
+                # return ["-fno-tree-loop-vectorize"]
         return []
 
     @collective
@@ -385,7 +386,7 @@ class LinuxCompiler(Compiler):
     :kwarg comm: Optional communicator to compile the code on (only
     rank 0 compiles code) (defaults to COMM_WORLD)."""
     def __init__(self, cppargs=[], ldargs=[], cpp=False, comm=None):
-        opt_flags = ['-march=native', '-O3', '-ffast-math']
+        opt_flags = ['-O3', '-ffast-math', '-fopenmp']
         if configuration['debug']:
             opt_flags = ['-O0', '-g']
         cc = "mpicc"

pyop2/sequential.py

@@ -83,15 +83,18 @@ def vectorise(wrapper, iname, batch_size):
     kernel = kernel.copy(temporary_variables=tmps)
 
     # split iname and vectorize the inner loop
+    slabs = (1, 1)
+    if configuration["time"]:
+        slabs = (0, 0)
     inner_iname = iname + "_batch"
 
     if configuration["vectorization_strategy"] == "ve":
         # vectorize using vector extenstions
-        kernel = loopy.split_iname(kernel, iname, batch_size, slabs=(0, 1), inner_tag="c_vec", inner_iname=inner_iname)
+        kernel = loopy.split_iname(kernel, iname, batch_size, slabs=slabs, inner_tag="c_vec", inner_iname=inner_iname)
     else:
         # vectoriza using omp pragma simd
         assert configuration["vectorization_strategy"] == "omp"
-        kernel = loopy.split_iname(kernel, iname, batch_size, slabs=(0, 1), inner_tag="omp_simd", inner_iname=inner_iname)
+        kernel = loopy.split_iname(kernel, iname, batch_size, slabs=slabs, inner_tag="omp_simd", inner_iname=inner_iname)
 
     alignment = configuration["alignment"]
     tmps = dict((name, tv.copy(alignment=alignment)) for name, tv in kernel.temporary_variables.items())
@@ -242,7 +245,7 @@ def set_nbytes(self, args):
         seen = set()
         for arg in args:
             if arg.access is INC:
-                nbytes += arg.data.nbytes
+                nbytes += arg.data.nbytes * 2
             else:
                 nbytes += arg.data.nbytes
             for map_ in arg.map_tuple: