[Mesa-dev] [PATCH 09/20] nir: add partial loop unrolling support
Timothy Arceri
tarceri at itsqueeze.com
Fri Dec 7 03:08:09 UTC 2018
This adds partial loop unrolling support and makes use of a
guessed trip count based on array access.
The code is written so that we could use partial unrolling
more generally, but for now it's only use when we have guessed
the trip count.
We use partial unrolling for this guessed trip count because its
possible any out of bounds array access doesn't otherwise affect
the shader e.g the stores/loads to/from the array are unused. So
we insert a copy of the loop in the innermost continue branch of
the unrolled loop. Later on its possible for nir_opt_dead_cf()
to then remove the loop in some cases.
A Renderdoc capture from the Rise of the Tomb Raider benchmark,
reports the following change in an affected compute shader:
GPU duration: 350 -> 325 microseconds
shader-db results radeonsi VEGA (NIR backend):
Totals from affected shaders:
SGPRS: 1120 -> 928 (-17.14 %)
VGPRS: 768 -> 516 (-32.81 %)
Spilled SGPRs: 666 -> 157 (-76.43 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 44072 -> 51880 (17.72 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 108 -> 147 (36.11 %)
Wait states: 0 -> 0 (0.00 %
shader-db results i965 SKL:
total instructions in shared programs: 13098265 -> 13103359 (0.04%)
instructions in affected programs: 5126 -> 10220 (99.38%)
helped: 0
HURT: 21
total cycles in shared programs: 332039949 -> 331985622 (-0.02%)
cycles in affected programs: 289252 -> 234925 (-18.78%)
helped: 12
HURT: 9
vkpipeline-db results VEGA:
Totals from affected shaders:
SGPRS: 184 -> 184 (0.00 %)
VGPRS: 448 -> 448 (0.00 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 26460 -> 25092 (-5.17 %) bytes
LDS: 6 -> 6 (0.00 %) blocks
Max Waves: 5 -> 5 (0.00 %)
Wait states: 0 -> 0 (0.00 %)
---
src/compiler/nir/nir_opt_loop_unroll.c | 206 ++++++++++++++++++++++++-
1 file changed, 198 insertions(+), 8 deletions(-)
diff --git a/src/compiler/nir/nir_opt_loop_unroll.c b/src/compiler/nir/nir_opt_loop_unroll.c
index 8406880204..d8df619b32 100644
--- a/src/compiler/nir/nir_opt_loop_unroll.c
+++ b/src/compiler/nir/nir_opt_loop_unroll.c
@@ -556,19 +556,200 @@ wrapper_unroll(nir_loop *loop)
return true;
}
+static bool
+is_access_out_of_bounds(nir_loop_terminator *term, nir_deref_instr *deref,
+ unsigned trip_count)
+{
+ for (nir_deref_instr *d = deref; d; d = nir_deref_instr_parent(d)) {
+ if (d->deref_type != nir_deref_type_array)
+ continue;
+
+ nir_alu_instr *alu = nir_instr_as_alu(term->conditional_instr);
+ nir_src src = term->induction_rhs ? alu->src[1].src : alu->src[0].src;
+ if (!nir_srcs_equal(d->arr.index, src))
+ continue;
+
+ nir_deref_instr *parent = nir_deref_instr_parent(d);
+ assert(glsl_type_is_array(parent->type) ||
+ glsl_type_is_matrix(parent->type));
+
+ /* We have already unrolled the loop and the new one will be imbedded in
+ * the innermost continue branch. So unless the array is greater than
+ * the trip count any iteration over the loop will be an out of bounds
+ * access of the array.
+ */
+ return glsl_get_length(parent->type) <= trip_count;
+ }
+
+ return false;
+}
+
+/* If we know an array access is going to be out of bounds remove or replace
+ * the access with an undef. This can later result in the entire loop being
+ * removed by nir_opt_dead_cf().
+ */
+static void
+remove_out_of_bounds_induction_use(nir_shader *shader, nir_loop *loop,
+ nir_loop_terminator *term,
+ nir_cf_list *lp_header,
+ nir_cf_list *lp_body,
+ unsigned trip_count)
+{
+ if (!loop->info->guessed_trip_count)
+ return;
+
+ /* Temporarily recreate the original loop so we can alter it */
+ nir_cf_reinsert(lp_header, nir_after_block(nir_loop_last_block(loop)));
+ nir_cf_reinsert(lp_body, nir_after_block(nir_loop_last_block(loop)));
+
+ nir_builder b;
+ nir_builder_init(&b, nir_cf_node_get_function(&loop->cf_node));
+
+ nir_foreach_block_in_cf_node(block, &loop->cf_node) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ /* Check for arrays variably-indexed by a loop induction variable.
+ * If this access is out of bounds remove the instruction or replace
+ * its use with an undefined instruction.
+ * If the loop is no longer useful we leave if for the appropriate
+ * pass to clean it up for us.
+ */
+ if (intrin->intrinsic == nir_intrinsic_load_deref ||
+ intrin->intrinsic == nir_intrinsic_store_deref ||
+ intrin->intrinsic == nir_intrinsic_copy_deref) {
+
+ if (is_access_out_of_bounds(term, nir_src_as_deref(intrin->src[0]),
+ trip_count)) {
+ if (intrin->intrinsic == nir_intrinsic_load_deref) {
+ assert(intrin->src[0].is_ssa);
+ nir_ssa_def *a_ssa = intrin->src[0].ssa;
+ nir_ssa_def *undef =
+ nir_ssa_undef(&b, intrin->num_components,
+ a_ssa->bit_size);
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
+ nir_src_for_ssa(undef));
+ } else {
+ nir_instr_remove(instr);
+ }
+ }
+
+ if (intrin->intrinsic == nir_intrinsic_copy_deref &&
+ is_access_out_of_bounds(term, nir_src_as_deref(intrin->src[1]),
+ trip_count)) {
+ assert(intrin->src[1].is_ssa);
+ nir_ssa_def *a_ssa = intrin->src[1].ssa;
+ nir_ssa_def *undef =
+ nir_ssa_undef(&b, intrin->num_components, a_ssa->bit_size);
+
+ /* Replace the copy with a store of the undefined value */
+ b.cursor = nir_before_instr(instr);
+ nir_store_deref(&b, nir_src_as_deref(intrin->src[0]), undef, ~0);
+ nir_instr_remove(instr);
+ }
+ }
+ }
+ }
+
+ /* Now that we are done extract the loop header and body again */
+ nir_cf_extract(lp_header, nir_before_block(nir_loop_first_block(loop)),
+ nir_before_cf_node(&term->nif->cf_node));
+ nir_cf_extract(lp_body, nir_before_block(nir_loop_first_block(loop)),
+ nir_after_block(nir_loop_last_block(loop)));
+}
+
+/* Partially unrolls loops that don't have a known trip count.
+ */
+static void
+partial_unroll(nir_shader *shader, nir_loop *loop, unsigned trip_count)
+{
+ assert(list_length(&loop->info->loop_terminator_list) == 1);
+
+ nir_loop_terminator *terminator =
+ list_first_entry(&loop->info->loop_terminator_list,
+ nir_loop_terminator, loop_terminator_link);
+
+ assert(nir_is_trivial_loop_if(terminator->nif, terminator->break_block));
+
+ loop_prepare_for_unroll(loop);
+
+ /* Pluck out the loop header */
+ nir_cf_list lp_header;
+ nir_cf_extract(&lp_header, nir_before_block(nir_loop_first_block(loop)),
+ nir_before_cf_node(&terminator->nif->cf_node));
+
+ struct hash_table *remap_table =
+ _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ nir_cf_list lp_body;
+ nir_cf_node *unroll_loc =
+ complex_unroll_loop_body(loop, terminator, &lp_header, &lp_body,
+ remap_table, trip_count);
+
+ /* Attempt to remove out of bounds array access */
+ remove_out_of_bounds_induction_use(shader, loop, terminator, &lp_header,
+ &lp_body, trip_count);
+
+ nir_cursor cursor =
+ get_complex_unroll_insert_location(unroll_loc,
+ terminator->continue_from_then);
+
+ /* Reinsert the loop in the innermost nested continue branch of the unrolled
+ * loop.
+ */
+ nir_loop *new_loop = nir_loop_create(shader);
+ nir_cf_node_insert(cursor, &new_loop->cf_node);
+ new_loop->partially_unrolled = true;
+
+ /* Clone loop header and insert into new loop */
+ nir_cf_list_clone_and_reinsert(&lp_header, loop->cf_node.parent,
+ nir_after_cf_list(&new_loop->body),
+ remap_table);
+
+ /* Clone loop body and insert into new loop */
+ nir_cf_list_clone_and_reinsert(&lp_body, loop->cf_node.parent,
+ nir_after_cf_list(&new_loop->body),
+ remap_table);
+
+ /* Insert break back into terminator */
+ nir_jump_instr *brk = nir_jump_instr_create(shader, nir_jump_break);
+ nir_if *nif = nir_block_get_following_if(nir_loop_first_block(new_loop));
+ if (terminator->continue_from_then) {
+ nir_instr_insert_after_block(nir_if_last_else_block(nif), &brk->instr);
+ } else {
+ nir_instr_insert_after_block(nir_if_last_then_block(nif), &brk->instr);
+ }
+
+ /* Delete the original loop header and body */
+ nir_cf_delete(&lp_header);
+ nir_cf_delete(&lp_body);
+
+ /* The original loop has been replaced so remove it. */
+ nir_cf_node_remove(&loop->cf_node);
+
+ _mesa_hash_table_destroy(remap_table, NULL);
+}
+
static bool
is_loop_small_enough_to_unroll(nir_shader *shader, nir_loop_info *li)
{
unsigned max_iter = shader->options->max_unroll_iterations;
- if (li->max_trip_count > max_iter)
+ unsigned trip_count =
+ li->max_trip_count ? li->max_trip_count : li->guessed_trip_count;
+
+ if (trip_count > max_iter)
return false;
- if (li->force_unroll)
+ if (li->force_unroll && !li->guessed_trip_count)
return true;
bool loop_not_too_large =
- li->num_instructions * li->max_trip_count <= max_iter * LOOP_UNROLL_LIMIT;
+ li->num_instructions * trip_count <= max_iter * LOOP_UNROLL_LIMIT;
return loop_not_too_large;
}
@@ -620,15 +801,24 @@ process_loops(nir_shader *sh, nir_cf_node *cf_node, bool *has_nested_loop_out)
!loop->info->complex_loop) {
nir_block *last_loop_blk = nir_loop_last_block(loop);
- if (!nir_block_ends_in_break(last_loop_blk))
+ if (nir_block_ends_in_break(last_loop_blk)) {
+ progress = wrapper_unroll(loop);
goto exit;
+ }
- progress = wrapper_unroll(loop);
-
- goto exit;
+ /* If we were able to guess the loop iteration based on array access
+ * then do a partial unroll.
+ */
+ unsigned num_lt = list_length(&loop->info->loop_terminator_list);
+ if (!has_nested_loop && num_lt == 1 && !loop->partially_unrolled &&
+ loop->info->guessed_trip_count &&
+ is_loop_small_enough_to_unroll(sh, loop->info)) {
+ partial_unroll(sh, loop, loop->info->guessed_trip_count);
+ progress = true;
+ }
}
- if (has_nested_loop || loop->info->limiting_terminator == NULL)
+ if (has_nested_loop || !loop->info->limiting_terminator)
goto exit;
if (!is_loop_small_enough_to_unroll(sh, loop->info))
--
2.19.2
More information about the mesa-dev
mailing list