From: Martin Kronbichler Date: Thu, 7 May 2020 11:03:15 +0000 (+0200) Subject: Add a comment X-Git-Tag: v9.2.0-rc1~79^2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=a443eaad306598a73fd50a297201305b8483eee1;p=dealii.git Add a comment --- diff --git a/source/matrix_free/task_info.cc b/source/matrix_free/task_info.cc index 953041fc53..c45ce9100f 100644 --- a/source/matrix_free/task_info.cc +++ b/source/matrix_free/task_info.cc @@ -777,6 +777,27 @@ namespace internal std::vector & renumbering, std::vector & incompletely_filled_vectorization) { + // This function is decomposed into several steps to determine a good + // ordering that satisfies the following constraints: + // a. Only cells belonging to the same category (or next higher if the + // cell_vectorization_categories_strict is false) can be grouped into + // the same SIMD batch + // b. hp adaptive computations must form contiguous ranges for the same + // degree (category) in cell_partition_data + // c. We want to group the cells with the same parent in the same SIMD + // lane if possible + // d. The cell order should be similar to the initial one + // e. Form sets without MPI communication and those with to overlap + // communication with computation + // + // These constraints are satisfied by first grouping by the categories + // and, within the groups, to distinguish between cells with a parent + // and those without. All of this is set up with batches of cells (with + // padding if the size does not match). Then we define a vector of + // arrays where we define sorting criteria for the cell batches to + // satisfy the items b and d together, split by different parts to + // satisfy item e. + // Give the compiler a chance to detect that vectorization_length is a // power of two, which allows it to replace integer divisions by shifts unsigned int vectorization_length_bits = 0;