Merge pull request #2129 from devitocodes/drop-collapse1-nvc

compiler: Avoid generating collapse(1)

devito/passes/iet/languages/openacc.py

@@ -26,13 +26,13 @@ def _make_construct(cls, **kwargs):
         return 'acc parallel loop'
 
     @classmethod
-    def _make_clauses(cls, ncollapsed=None, reduction=None, tile=None, **kwargs):
+    def _make_clauses(cls, ncollapsed=0, reduction=None, tile=None, **kwargs):
         clauses = []
 
         if tile:
             clauses.append('tile(%s)' % ','.join(str(i) for i in tile))
-        elif ncollapsed:
-            clauses.append('collapse(%d)' % (ncollapsed or 1))
+        elif ncollapsed > 1:
+            clauses.append('collapse(%d)' % ncollapsed)
 
         if reduction:
             clauses.append(cls._make_clause_reduction_from_imask(reduction))

devito/passes/iet/languages/openmp.py

@@ -40,11 +40,12 @@ def _make_construct(cls, parallel=False, **kwargs):
             return 'omp for'
 
     @classmethod
-    def _make_clauses(cls, ncollapsed=None, chunk_size=None, nthreads=None,
+    def _make_clauses(cls, ncollapsed=0, chunk_size=None, nthreads=None,
                       reduction=None, schedule=None, **kwargs):
         clauses = []
 
-        clauses.append('collapse(%d)' % (ncollapsed or 1))
+        if ncollapsed > 1:
+            clauses.append('collapse(%d)' % ncollapsed)
 
         if chunk_size is not False:
             clauses.append('schedule(%s,%s)' % (schedule or 'dynamic',

examples/performance/00_overview.ipynb

@@ -716,7 +716,7 @@
       "    const int tid = omp_get_thread_num();\n",
       "    float (*restrict r0)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr0[tid];\n",
       "\n",
-      "    #pragma omp for collapse(1) schedule(dynamic,1)\n",
+      "    #pragma omp for schedule(dynamic,1)\n",
       "    for (int x = x_m; x <= x_M; x += 1)\n",
       "    {\n",
       "      for (int y = y_m - 2; y <= y_M + 2; y += 1)\n",
@@ -855,7 +855,7 @@
       "    const int tid = omp_get_thread_num();\n",
       "    float (*restrict r1)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr1[tid];\n",
       "\n",
-      "    #pragma omp for collapse(1) schedule(dynamic,1)\n",
+      "    #pragma omp for schedule(dynamic,1)\n",
       "    for (int x = x_m; x <= x_M; x += 1)\n",
       "    {\n",
       "      for (int y = y_m - 2; y <= y_M + 2; y += 1)\n",
@@ -991,7 +991,7 @@
       "    const int tid = omp_get_thread_num();\n",
       "    float (*restrict r0)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr0[tid];\n",
       "\n",
-      "    #pragma omp for collapse(1) schedule(dynamic,1)\n",
+      "    #pragma omp for schedule(dynamic,1)\n",
       "    for (int x = x_m; x <= x_M; x += 1)\n",
       "    {\n",
       "      for (int y = y_m - 2; y <= y_M + 2; y += 1)\n",
@@ -1557,7 +1557,7 @@
       "      const int tid = omp_get_thread_num();\n",
       "      float (*restrict r2)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr2[tid];\n",
       "\n",
-      "      #pragma omp for collapse(1) schedule(dynamic,1)\n",
+      "      #pragma omp for schedule(dynamic,1)\n",
       "      for (int x = x_m; x <= x_M; x += 1)\n",
       "      {\n",
       "        for (int y = y_m - 2; y <= y_M + 2; y += 1)\n",

tests/test_dle.py

@@ -816,7 +816,8 @@ def test_incs_no_atomic(self):
         # Now only `x` is parallelized
         op1 = Operator([Eq(v[t, x, 0, 0], v[t, x, 0, 0] + 1), Inc(uf, 1)],
                        opt=('advanced', {'openmp': True, 'par-collapse-ncores': 1}))
-        assert 'collapse(1)' in str(op1)
+        assert 'omp for' in str(op1)
+        assert 'collapse' not in str(op1)
         assert 'atomic' not in str(op1)
 
     @pytest.mark.parametrize('exprs,simd_level,expected', [
@@ -879,18 +880,22 @@ def test_edge_cases(self, exprs, simd_level, expected):
         for i, e in enumerate(list(exprs)):
             exprs[i] = eval(e)
 
-        op = Operator(exprs, opt=('advanced', {'openmp': True}))
+        op = Operator(exprs, opt=('advanced', {'openmp': True,
+                                               'par-collapse-ncores': 1}))
 
         iterations = FindNodes(Iteration).visit(op)
+        parallel = [i for i in iterations if i.is_Parallel]
         try:
-            assert 'omp for collapse' in iterations[0].pragmas[0].value
+            assert 'omp for' in iterations[0].pragmas[0].value
+            if len(parallel) > 1 and simd_level is not None and simd_level > 1:
+                assert 'collapse' in iterations[0].pragmas[0].value
             if simd_level:
                 assert 'omp simd' in iterations[simd_level].pragmas[0].value
         except:
             # E.g. gcc-5 doesn't support array reductions, so the compiler will
             # generate different legal code
             assert not Ompizer._support_array_reduction(configuration['compiler'])
-            assert any('omp for collapse' in i.pragmas[0].value
+            assert any('omp for' in i.pragmas[0].value
                        for i in iterations if i.pragmas)
 
         op.apply()
@@ -910,7 +915,7 @@ def test_simd_space_invariant(self):
         op = Operator(eq, opt=('advanced', {'openmp': True}))
         iterations = FindNodes(Iteration).visit(op)
 
-        assert 'omp for collapse(1) schedule(static,1)' in iterations[0].pragmas[0].value
+        assert 'omp for schedule(static,1)' in iterations[0].pragmas[0].value
         assert 'omp simd' in iterations[1].pragmas[0].value
         assert 'omp simd' in iterations[3].pragmas[0].value
 
@@ -979,8 +984,8 @@ def test_basic(self):
         bns, _ = assert_blocking(op, {'x0_blk0'})
 
         iterations = FindNodes(Iteration).visit(bns['x0_blk0'])
-        assert iterations[0].pragmas[0].value == 'omp for collapse(1) schedule(dynamic,1)'
-        assert iterations[2].pragmas[0].value == ('omp parallel for collapse(1) '
+        assert iterations[0].pragmas[0].value == 'omp for schedule(dynamic,1)'
+        assert iterations[2].pragmas[0].value == ('omp parallel for '
                                                   'schedule(dynamic,1) '
                                                   'num_threads(nthreads_nested)')
 
@@ -1073,11 +1078,11 @@ def test_multiple_subnests_v1(self):
             'omp for collapse(2) schedule(dynamic,1)'
         assert not trees[0][2].pragmas
         assert not trees[0][3].pragmas
-        assert trees[0][4].pragmas[0].value == ('omp parallel for collapse(1) '
+        assert trees[0][4].pragmas[0].value == ('omp parallel for '
                                                 'schedule(dynamic,1) '
                                                 'num_threads(nthreads_nested)')
         assert not trees[1][2].pragmas
-        assert trees[1][3].pragmas[0].value == ('omp parallel for collapse(1) '
+        assert trees[1][3].pragmas[0].value == ('omp parallel for '
                                                 'schedule(dynamic,1) '
                                                 'num_threads(nthreads_nested)')
 

tests/test_gpu_openacc.py

@@ -110,7 +110,7 @@ def test_tile_insteadof_collapse(self, par_tile):
             'acc parallel loop tile(32,4) present(u)'
         # Only the AFFINE Iterations are tiled
         assert trees[3][1].pragmas[0].value ==\
-            'acc parallel loop collapse(1) present(src,src_coords,u)'
+            'acc parallel loop present(src,src_coords,u)'
 
     @pytest.mark.parametrize('par_tile', [((32, 4, 4), (8, 8)), ((32, 4), (8, 8)),
                                           ((32, 4, 4), (8, 8, 8))])