diff --git a/sumpy/e2e.py b/sumpy/e2e.py index 2d13ba5aa158e1980c47fbe7d292fa4668915887..7774b62aaed9f24b613c4b5d2c16fa4f470db90e 100644 --- a/sumpy/e2e.py +++ b/sumpy/e2e.py @@ -649,6 +649,8 @@ class M2LGenerateTranslationClassesDependentData(E2EBase): loopy_knl = knl.prepare_loopy_kernel(loopy_knl) loopy_knl = lp.tag_inames(loopy_knl, "idim*:unr") + loopy_knl = lp.set_options(loopy_knl, + enforce_variable_access_ordered="no_check") return loopy_knl @@ -983,6 +985,8 @@ class E2EFromParent(E2EBase): loopy_knl = knl.prepare_loopy_kernel(loopy_knl) loopy_knl = lp.tag_inames(loopy_knl, "idim*:unr") + loopy_knl = lp.set_options(loopy_knl, + enforce_variable_access_ordered="no_check") return loopy_knl diff --git a/sumpy/e2p.py b/sumpy/e2p.py index 2a4ff531c0522514bb43cbb8267bc411ea9e05fa..25e03242b4507f28c71cd3ca3a97d6b1b9af0889 100644 --- a/sumpy/e2p.py +++ b/sumpy/e2p.py @@ -202,6 +202,8 @@ class E2PFromSingleBox(E2PBase): knl = self.get_kernel() knl = lp.tag_inames(knl, dict(itgt_box="g.0")) knl = self._allow_redundant_execution_of_knl_scaling(knl) + knl = lp.set_options(knl, + enforce_variable_access_ordered="no_check") return knl @@ -314,6 +316,8 @@ class E2PFromCSR(E2PBase): knl = self.get_kernel() knl = lp.tag_inames(knl, dict(itgt_box="g.0")) knl = self._allow_redundant_execution_of_knl_scaling(knl) + knl = lp.set_options(knl, + enforce_variable_access_ordered="no_check") return knl def __call__(self, queue, **kwargs): diff --git a/sumpy/p2e.py b/sumpy/p2e.py index cb4b6616dca435db2c77350f5aaf5952dacdf6d3..177d0b586ba8c04be2c1e01955a79d7b53f042f1 100644 --- a/sumpy/p2e.py +++ b/sumpy/p2e.py @@ -131,6 +131,8 @@ class P2EBase(KernelComputation, KernelCacheWrapper): knl = lp.tag_array_axes(knl, "centers", "sep,C") knl = self._allow_redundant_execution_of_knl_scaling(knl) + knl = lp.set_options(knl, + enforce_variable_access_ordered="no_check") return knl def __call__(self, queue, **kwargs): diff --git a/sumpy/p2p.py b/sumpy/p2p.py index 4853a2b1de94601978238f1ff4a81906dace713c..21dde61efe9ca0553dfea2eb8e3c62b5f903065e 100644 --- a/sumpy/p2p.py +++ b/sumpy/p2p.py @@ -187,6 +187,8 @@ class P2PBase(KernelComputation, KernelCacheWrapper): knl = lp.split_iname(knl, "itgt", 1024, outer_tag="g.0") knl = self._allow_redundant_execution_of_knl_scaling(knl) + knl = lp.set_options(knl, + enforce_variable_access_ordered="no_check") return knl @@ -395,6 +397,8 @@ class P2PMatrixSubsetGenerator(P2PBase): knl = lp.split_iname(knl, "imat", 1024, outer_tag="g.0") knl = self._allow_redundant_execution_of_knl_scaling(knl) + knl = lp.set_options(knl, + enforce_variable_access_ordered="no_check") return knl def __call__(self, queue, targets, sources, tgtindices, srcindices, **kwargs): @@ -534,6 +538,8 @@ class P2PFromCSR(P2PBase): knl = lp.split_iname(knl, "itgt_box", 4, outer_tag="g.0") knl = self._allow_redundant_execution_of_knl_scaling(knl) + knl = lp.set_options(knl, + enforce_variable_access_ordered="no_check") return knl