Module: vkd3d Branch: master Commit: b940d79aa8a0783a48bf28d566cce5c4e6259f08 URL: https://gitlab.winehq.org/wine/vkd3d/-/commit/b940d79aa8a0783a48bf28d566cce5...
Author: Giovanni Mascellani gmascellani@codeweavers.com Date: Fri Feb 2 23:35:21 2024 +0100
vkd3d-shader/ir: Keep loops contiguous in the topological sort.
---
libs/vkd3d-shader/ir.c | 94 ++++++++++++++++++++++++++++++++++++++++++++++---- 1 file changed, 88 insertions(+), 6 deletions(-)
diff --git a/libs/vkd3d-shader/ir.c b/libs/vkd3d-shader/ir.c index 894324c5..3c862f33 100644 --- a/libs/vkd3d-shader/ir.c +++ b/libs/vkd3d-shader/ir.c @@ -3121,6 +3121,11 @@ static void vsir_block_list_sort(struct vsir_block_list *list) qsort(list->blocks, list->count, sizeof(*list->blocks), block_compare); }
+static bool vsir_block_list_search(struct vsir_block_list *list, struct vsir_block *block) +{ + return !!bsearch(&block, list->blocks, list->count, sizeof(*list->blocks), block_compare); +} + struct vsir_cfg { struct vkd3d_shader_message_context *message_context; @@ -3487,16 +3492,37 @@ static enum vkd3d_result vsir_cfg_compute_loops(struct vsir_cfg *cfg) struct vsir_cfg_node_sorter { struct vsir_cfg *cfg; + struct vsir_cfg_node_sorter_stack_item + { + struct vsir_block_list *loop; + unsigned int seen_count; + } *stack; + size_t stack_count, stack_capacity; struct vsir_block_list available_blocks; };
static enum vkd3d_result vsir_cfg_node_sorter_make_node_available(struct vsir_cfg_node_sorter *sorter, struct vsir_block *block) { + struct vsir_block_list *loop = NULL; + struct vsir_cfg_node_sorter_stack_item *item; enum vkd3d_result ret;
+ if (sorter->cfg->loops_by_header[block->label - 1] != SIZE_MAX) + loop = &sorter->cfg->loops[sorter->cfg->loops_by_header[block->label - 1]]; + if ((ret = vsir_block_list_add_checked(&sorter->available_blocks, block)) < 0) return ret;
+ if (!loop) + return VKD3D_OK; + + if (!vkd3d_array_reserve((void **)&sorter->stack, &sorter->stack_capacity, sorter->stack_count + 1, sizeof(*sorter->stack))) + return VKD3D_ERROR_OUT_OF_MEMORY; + + item = &sorter->stack[sorter->stack_count++]; + item->loop = loop; + item->seen_count = 0; + return VKD3D_OK; }
@@ -3507,10 +3533,23 @@ static enum vkd3d_result vsir_cfg_node_sorter_make_node_available(struct vsir_cf * available list of all the blocks whose incoming degree has reached * zero. At each step we pick a block from the available list and * strip it away from the graph, updating the incoming degrees and - * available list. We prefer picking the most recently added block - * (i.e., from the end of the available list) because it will keep - * together related blocks in the order, which allow us to generate - * fewer control flow primitives. */ + * available list. + * + * In principle at each step we can pick whatever node we want from + * the available list, and will get a topological sort + * anyway. However, we use these two criteria to give to the computed + * order additional properties: + * + * 1. we keep track of which loops we're into, and pick blocks + * belonging to the current innermost loop, so that loops are kept + * contiguous in the order; this can always be done when the input + * CFG is reducible; + * + * 2. subject to the requirement above, we always pick the most + * recently added block to the available list, because this tends + * to keep related blocks and require fewer control flow + * primitives. + */ static enum vkd3d_result vsir_cfg_sort_nodes(struct vsir_cfg *cfg) { struct vsir_cfg_node_sorter sorter = { .cfg = cfg }; @@ -3566,12 +3605,51 @@ static enum vkd3d_result vsir_cfg_sort_nodes(struct vsir_cfg *cfg)
while (sorter.available_blocks.count != 0) { - struct vsir_block *block = sorter.available_blocks.blocks[sorter.available_blocks.count - 1]; + struct vsir_cfg_node_sorter_stack_item *inner_stack_item = NULL; + struct vsir_block *block; + size_t new_seen_count; + + if (sorter.stack_count != 0) + inner_stack_item = &sorter.stack[sorter.stack_count - 1]; + + for (i = sorter.available_blocks.count - 1; ; --i) + { + if (i == SIZE_MAX) + { + ERR("Couldn't find any viable next block, is the input CFG reducible?\n"); + ret = VKD3D_ERROR_INVALID_SHADER; + goto fail; + } + + block = sorter.available_blocks.blocks[i]; + + if (!inner_stack_item || vsir_block_list_search(inner_stack_item->loop, block)) + break; + }
- vsir_block_list_remove_index(&sorter.available_blocks, sorter.available_blocks.count - 1); + vsir_block_list_remove_index(&sorter.available_blocks, i); if ((ret = vsir_block_list_add_checked(&cfg->order, block)) < 0) goto fail;
+ /* Close loops: since each loop is a strict subset of any + * outer loop, we just need to track how many blocks we've + * seen; when I close a loop I mark the same number of seen + * blocks for the next outer loop. */ + new_seen_count = 1; + while (sorter.stack_count != 0) + { + inner_stack_item = &sorter.stack[sorter.stack_count - 1]; + + inner_stack_item->seen_count += new_seen_count; + + assert(inner_stack_item->seen_count <= inner_stack_item->loop->count); + if (inner_stack_item->seen_count != inner_stack_item->loop->count) + break; + + new_seen_count = inner_stack_item->loop->count; + --sorter.stack_count; + } + /* Remove (forward) edges and make new nodes available. */ for (i = 0; i < block->successors.count; ++i) { @@ -3601,7 +3679,10 @@ static enum vkd3d_result vsir_cfg_sort_nodes(struct vsir_cfg *cfg) goto fail; }
+ assert(sorter.stack_count == 0); + vkd3d_free(in_degrees); + vkd3d_free(sorter.stack); vsir_block_list_cleanup(&sorter.available_blocks);
if (TRACE_ON()) @@ -3619,6 +3700,7 @@ static enum vkd3d_result vsir_cfg_sort_nodes(struct vsir_cfg *cfg)
fail: vkd3d_free(in_degrees); + vkd3d_free(sorter.stack); vsir_block_list_cleanup(&sorter.available_blocks);
return ret;