---
src/compiler/Makefile.sources | 1 +
src/compiler/nir/nir.h | 2 +
src/compiler/nir/nir_opt_loop_unroll.c | 394 +++++++++++++++++++++++++++++++++
3 files changed, 397 insertions(+)
create mode 100644 src/compiler/nir/nir_opt_loop_unroll.c
diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index 79de484..a9f104d 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -231,6 +231,7 @@ NIR_FILES = \
nir/nir_opt_dead_cf.c \
nir/nir_opt_gcm.c \
nir/nir_opt_global_to_local.c \
+ nir/nir_opt_loop_unroll.c \
nir/nir_opt_peephole_select.c \
nir/nir_opt_remove_phis.c \
nir/nir_opt_undef.c \
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 9083bd0..81d9dfc 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -2676,6 +2676,8 @@ bool nir_opt_dead_cf(nir_shader *shader);
void nir_opt_gcm(nir_shader *shader);
+bool nir_opt_loop_unroll(nir_shader *shader);
+
bool nir_opt_peephole_select(nir_shader *shader);
bool nir_opt_remove_phis(nir_shader *shader);
diff --git a/src/compiler/nir/nir_opt_loop_unroll.c
b/src/compiler/nir/nir_opt_loop_unroll.c
new file mode 100644
index 0000000..22530c9
--- /dev/null
+++ b/src/compiler/nir/nir_opt_loop_unroll.c
@@ -0,0 +1,394 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#include "nir.h"
+#include "nir_builder.h"
+#include "nir_control_flow.h"
+
+typedef struct {
+ /* Array of loops */
+ nir_loop *loops;
+
+ /* Array of unroll factors */
+ int *factors;
+} unroll_vector;
+
+typedef struct {
+ /* Array of loop infos for the loop nest */
+ nir_loop_info *li;
+
+ /* List of unroll vectors */
+ struct list_head unroll_vectors;
+
+ nir_shader_compiler_options *options;
+} loop_unroll_state;
+
+static void
+extract_loop_body(nir_cf_list *extracted, nir_cf_node *node, nir_shader *ns)
+{
+ nir_cf_node *end = node;
+ while (!nir_cf_node_is_last(end))
+ end = nir_cf_node_next(end);
+
+ nir_cf_loop_list_extract(extracted, node, end);
+}
+
+static void
+clone_list(nir_shader *ns, nir_loop *loop, nir_cf_list *src_cf_list,
+ nir_cf_list *cloned_cf_list, struct hash_table *remap_table,
+ struct hash_table *phi_remap)
+{
+ /* Dest list needs to at least have one block */
+ nir_block *nblk = nir_block_create(ns);
+ nblk->cf_node.parent = loop->cf_node.parent;
+ exec_list_push_tail(&cloned_cf_list->list, &nblk->cf_node.node);
+
+ nir_clone_loop_list(&cloned_cf_list->list, &src_cf_list->list,
+ remap_table, phi_remap, ns);
+}
+
+static void
+remove_unrolled_loop(nir_cf_node *loop, nir_block *block_before_loop,
+ struct hash_table *remap_table,
+ struct hash_table *phi_remap,
+ nir_function_impl *impl)
+{
+ /* Fixup LCSSA-phi srcs */
+ nir_block *prev_block = nir_cf_node_as_block(nir_cf_node_prev(loop));
+ nir_cf_node *cf_node = nir_cf_node_next(loop);
+ assert(cf_node->type == nir_cf_node_block);
+
+ nir_block *block = nir_cf_node_as_block(cf_node);
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type == nir_instr_type_phi) {
+ nir_phi_instr *phi = nir_instr_as_phi(instr);
+ assert(phi->is_lcssa_phi);
+
+ if (nir_cf_node_as_loop(loop)->info->trip_count != 0) {
+ nir_foreach_phi_src_safe(src, phi) {
+ /* Update predecessor */
+ src->pred = prev_block;
+
+ /* Update src */
+ struct hash_entry *hte =
+ _mesa_hash_table_search(phi_remap, src->src.ssa);
+ if (hte) {
+ nir_src new_src = nir_src_for_ssa((nir_ssa_def *) hte->data);
+ nir_instr_rewrite_src(instr, &src->src, new_src);
+ } else {
+ /* If the LCSSA src was not a phi fallback to following the
+ * the remappings in the remap table.
+ */
+ nir_ssa_def *ssa_def = src->src.ssa;
+ while ((hte = _mesa_hash_table_search(remap_table,
+ ssa_def))) {
+ ssa_def = (nir_ssa_def *) hte->data;
+ }
+
+ if (ssa_def != src->src.ssa){
+ nir_src new_src = nir_src_for_ssa(ssa_def);
+ nir_instr_rewrite_src(instr, &src->src, new_src);
+ }
+ }
+ }
+ }
+ nir_opt_remove_phi(phi);
+ } else {
+ /* There should be no more LCSSA-phis */
+ break;
+ }
+ }
+
+ /* After cloning the loop and fixing up the LCSSA-phis we need to remove
+ * any obsolete phis at the beginning of the new unrolled block. To avoid
+ * failing validation.
+ */
+ nir_foreach_instr_safe(instr, block_before_loop) {
+ if (instr->type != nir_instr_type_phi)
+ continue;
+
+ /* Replace phi */
+ nir_phi_instr *phi = nir_instr_as_phi(instr);
+ nir_opt_remove_phi(phi);
+ }
+
+ /* Remove the loop */
+ nir_loop *nir_loop = nir_cf_node_as_loop(loop);
+ foreach_list_typed(nir_cf_node, child, node, &nir_loop->body)
+ nir_cleanup_cf_node(child, impl, false);
+
+ exec_node_remove(&loop->node);
+
+ stitch_blocks(prev_block, block);
+}
+
+/**
+ * Unroll a loop which does not contain any jumps. For example, if the input
+ * is:
+ *
+ * (loop (...) ...instrs...)
+ *
+ * And the iteration count is 3, the output will be:
+ *
+ * ...instrs... ...instrs... ...instrs...
+ */
+static void
+simple_unroll(nir_function *fn, nir_loop *loop, nir_builder *b)
+{
+ nir_shader *ns = fn->shader;
+ nir_block *prev_block =
+ nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
+
+ struct exec_node *loop_node = exec_list_get_head(&loop->body);
+ nir_cf_node *lp_header_cf_node = exec_node_data(nir_cf_node, loop_node,
+ node);
+
+ struct hash_table *remap_table =
+ _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ struct hash_table *phi_remap =
+ _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ /* Get the loop header this contains a bunch of phis and the loops
+ * conditional.
+ */
+ assert(lp_header_cf_node->type == nir_cf_node_block);
+ nir_block *loop_header_blk = nir_cf_node_as_block(lp_header_cf_node);
+
+ /* Remove any phi preds that are from the loop itself nir_opt_remove_phi()
+ * will clean this up further for us.
+ */
+ nir_foreach_instr_safe(instr, loop_header_blk) {
+ if (instr->type != nir_instr_type_phi)
+ break;
+
+ nir_phi_instr *phi = nir_instr_as_phi(instr);
+ nir_foreach_phi_src_safe(src, phi) {
+ /* Is the pred from the block itself? */
+ if (src->pred->index > phi->instr.block->index &&
+ src->pred->cf_node.parent == &loop->cf_node) {
+
+ /* Store the phi dest and src to be used for remaping */
+ _mesa_hash_table_insert(phi_remap, &phi->dest.ssa, src->src.ssa);
+
+ list_del(&src->src.use_link);
+ exec_node_remove(&src->node);
+ }
+ }
+ }
+
+ /* Skip over if node (loop condition) and get the real loop body we need
+ * to do this now because we are about to start messing with the control
+ * flow.
+ */
+ nir_cf_node *if_node = nir_cf_node_next(lp_header_cf_node);
+ nir_cf_node *cf_node = nir_cf_node_next(if_node);
+ assert(cf_node->type == nir_cf_node_block);
+
+ /* Pluck out the loop header */
+ nir_cf_list lp_header;
+ nir_cf_extract(&lp_header, nir_before_cf_node(lp_header_cf_node),
+ nir_after_cf_node(lp_header_cf_node));
+
+ loop_node = exec_list_get_head(&loop->body);
+ lp_header_cf_node = exec_node_data(nir_cf_node, loop_node, node);
+
+ /* Extract phis from loop header and place before the loop */
+ nir_cf_list lp_header_phis;
+ nir_cf_loop_list_extract(&lp_header_phis, lp_header_cf_node,
+ lp_header_cf_node);
+ b->cursor = nir_before_cf_node(&loop->cf_node);
+ nir_cf_reinsert(&lp_header_phis, b->cursor);
+
+ /* Pluck out the loop body */
+ nir_cf_list loop_body;
+ extract_loop_body(&loop_body, cf_node, ns);
+
+ /* Initialise remap table with phi remap table */
+ struct hash_entry *hte;
+ hash_table_foreach(phi_remap, hte) {
+ _mesa_hash_table_insert(remap_table, hte->key, hte->data);
+ }
+
+ /* Clone loop header and append to the loop body */
+ if (loop->info->trip_count > 0) {
+ nir_cf_list cloned_header;
+ exec_list_make_empty(&cloned_header.list);
+ cloned_header.impl = loop_body.impl;
+
+ clone_list(ns, loop, &lp_header, &cloned_header, remap_table, phi_remap);
+
+ /* Append loop body with header instructions */
+ struct exec_node *lp_body_node = exec_list_get_tail(&loop_body.list);
+ nir_cf_node *end = exec_node_data(nir_cf_node, lp_body_node, node);
+ nir_block *lp_body_blk = nir_cf_node_as_block(end);
+
+ struct exec_node *lp_header_node =
+ exec_list_get_head(&cloned_header.list);
+ nir_block *lp_header_blk =
+ nir_cf_node_as_block(exec_node_data(nir_cf_node, lp_header_node,
+ node));
+
+ foreach_list_typed(nir_instr, instr, node, &lp_header_blk->instr_list) {
+ instr->block = lp_body_blk;
+ }
+
+ exec_list_append(&lp_body_blk->instr_list, &lp_header_blk->instr_list);
+ }
+
+ /* Reinsert original header before the loop */
+ nir_cf_reinsert(&lp_header, b->cursor);
+
+ nir_cf_list *prev_lp = &loop_body;
+
+ /* Create temp array to store the cloned loop body as we unroll */
+ nir_cf_list unrolled_lp_body[2];
+ exec_list_make_empty(&unrolled_lp_body[0].list);
+ exec_list_make_empty(&unrolled_lp_body[1].list);
+ unrolled_lp_body[0].impl = loop_body.impl;
+ unrolled_lp_body[1].impl = loop_body.impl;
+
+ for (unsigned i = 1; i < loop->info->trip_count; i++) {
+
+ /* Clone the loop body */
+ if (exec_list_is_empty(&unrolled_lp_body[0].list)){
+ clone_list(ns, loop, prev_lp, &unrolled_lp_body[0], remap_table,
+ phi_remap);
+
+ /* Insert unrolled loop body before the loop. */
+ b->cursor = nir_before_cf_node(&loop->cf_node);
+ nir_cf_reinsert(prev_lp, b->cursor);
+
+ prev_lp = &unrolled_lp_body[0];
+ } else {
+ clone_list(ns, loop, prev_lp, &unrolled_lp_body[1], remap_table,
+ phi_remap);
+
+ /* Insert unrolled loop body before the loop. */
+ b->cursor = nir_before_cf_node(&loop->cf_node);
+ nir_cf_reinsert(prev_lp, b->cursor);
+
+ prev_lp = &unrolled_lp_body[1];
+ }
+ }
+
+ b->cursor = nir_before_cf_node(&loop->cf_node);
+ if (loop->info->trip_count != 0) {
+ nir_cf_reinsert(prev_lp, b->cursor);
+ } else {
+ /* Its possible to end up with a trip count of zero. For example:
+ *
+ * int 0;
+ * do res = res.yzwx; while (++i < 1);
+ *
+ * In this case we simple delete the loop and fix up the LCSSA phis.
+ */
+ nir_cf_delete(prev_lp);
+ }
+
+ /* The loop has been unrolled so remove it. */
+ remove_unrolled_loop(&loop->cf_node, prev_block, remap_table, phi_remap,
+ fn->impl);
+
+ _mesa_hash_table_destroy(remap_table, NULL);
+ _mesa_hash_table_destroy(phi_remap, NULL);
+}
+
+static bool
+process_loops(nir_cf_node *cf_node, nir_builder *b, bool *innermost_loop)
+{
+ bool progress = false;
+ nir_loop *loop;
+
+ switch (cf_node->type) {
+ case nir_cf_node_block:
+ return progress;
+ case nir_cf_node_if: {
+ nir_if *if_stmt = nir_cf_node_as_if(cf_node);
+ foreach_list_typed_safe(nir_cf_node, nested_node, node,
&if_stmt->then_list)
+ progress |= process_loops(nested_node, b, innermost_loop);
+ foreach_list_typed_safe(nir_cf_node, nested_node, node,
&if_stmt->else_list)
+ progress |= process_loops(nested_node, b, innermost_loop);
+ return progress;
+ }
+ case nir_cf_node_loop: {
+ loop = nir_cf_node_as_loop(cf_node);
+ foreach_list_typed_safe(nir_cf_node, nested_node, node, &loop->body)
+ progress |= process_loops(nested_node, b, innermost_loop);
+ break;
+ }
+ default:
+ unreachable("unknown cf node type");
+ }
+
+ if (*innermost_loop) {
+ nir_function *fn = nir_cf_node_get_function(&loop->cf_node)->function;
+ if (is_simple_for_loop(fn->shader, loop->info)) {
+ simple_unroll(fn, loop, b);
+ progress = true;
+ /* Don't unroll outer loops loops until after loop analysis have run
+ * again.
+ */
+ *innermost_loop = false;
+ } else {
+ /* We couldn't unroll the innermost loop so don't try to unroll the
+ * outerloop.
+ */
+ *innermost_loop = false;
+ }
+ }
+
+ return progress;
+}
+
+static bool
+nir_opt_loop_unroll_impl(nir_function_impl *impl)
+{
+ bool progress = false;
+ nir_metadata_require(impl, nir_metadata_loop_analysis);
+
+ nir_builder b;
+ nir_builder_init(&b, impl);
+
+ foreach_list_typed_safe(nir_cf_node, node, node, &impl->body) {
+ bool innermost_loop = true;
+ progress |= process_loops(node, &b, &innermost_loop);
+ }
+
+ return progress;
+}
+
+bool
+nir_opt_loop_unroll(nir_shader *shader)
+{
+ bool progress = false;
+
+ nir_foreach_function(function, shader) {
+ if (function->impl) {
+ progress |= nir_opt_loop_unroll_impl(function->impl);
+ }
+ }
+ return false;
+}
--
2.7.4
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
