[Mesa-dev] [PATCH 05/12] glsl: Combine nop-swizzle optimization with swizzle-swizzle optimization
Thomas Helland
thomashelland90 at gmail.com
Thu Nov 2 22:53:25 UTC 2017
Changes all look really good, and some performance numbers
I have for a similar patch I've written shows a marginal benefit
in compiler runtime performance (perf-stat -> cycles executed)
For patches 4 and 5.
Reviewed-by: <thomashelland90 at gmail.com>
2. nov. 2017 21.26 skrev "Ian Romanick" <idr at freedesktop.org>:
From: Ian Romanick <ian.d.romanick at intel.com>
Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
---
src/compiler/Makefile.sources | 3 +-
src/compiler/glsl/glsl_parser_extras.cpp | 3 +-
src/compiler/glsl/ir_optimization.h | 3 +-
src/compiler/glsl/meson.build | 3 +-
.../glsl/{opt_noop_swizzle.cpp => opt_swizzle.cpp} | 56 ++++++++++---
src/compiler/glsl/opt_swizzle_swizzle.cpp | 96
----------------------
src/compiler/glsl/test_optpass.cpp | 6 +-
7 files changed, 52 insertions(+), 118 deletions(-)
rename src/compiler/glsl/{opt_noop_swizzle.cpp => opt_swizzle.cpp} (56%)
delete mode 100644 src/compiler/glsl/opt_swizzle_swizzle.cpp
diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index b80468c..60814e2 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -130,11 +130,10 @@ LIBGLSL_FILES = \
glsl/opt_function_inlining.cpp \
glsl/opt_if_simplification.cpp \
glsl/opt_minmax.cpp \
- glsl/opt_noop_swizzle.cpp \
glsl/opt_rebalance_tree.cpp \
glsl/opt_redundant_jumps.cpp \
glsl/opt_structure_splitting.cpp \
- glsl/opt_swizzle_swizzle.cpp \
+ glsl/opt_swizzle.cpp \
glsl/opt_tree_grafting.cpp \
glsl/opt_vectorize.cpp \
glsl/program.h \
diff --git a/src/compiler/glsl/glsl_parser_extras.cpp
b/src/compiler/glsl/glsl_parser_extras.cpp
index 822301a..5982173 100644
--- a/src/compiler/glsl/glsl_parser_extras.cpp
+++ b/src/compiler/glsl/glsl_parser_extras.cpp
@@ -2226,8 +2226,7 @@ do_common_optimization(exec_list *ir, bool linked,
options->EmitNoCont, options->EmitNoLoops);
OPT(do_vec_index_to_swizzle, ir);
OPT(lower_vector_insert, ir, false);
- OPT(do_swizzle_swizzle, ir);
- OPT(do_noop_swizzle, ir);
+ OPT(optimize_swizzles, ir);
OPT(optimize_split_arrays, ir, linked);
OPT(optimize_redundant_jumps, ir);
diff --git a/src/compiler/glsl/ir_optimization.h b/src/compiler/glsl/ir_
optimization.h
index f44ddcb..2b8c195 100644
--- a/src/compiler/glsl/ir_optimization.h
+++ b/src/compiler/glsl/ir_optimization.h
@@ -123,9 +123,8 @@ bool lower_if_to_cond_assign(gl_shader_stage stage,
exec_list *instructions,
unsigned max_depth = 0, unsigned
min_branch_cost = 0);
bool do_mat_op_to_vec(exec_list *instructions);
bool do_minmax_prune(exec_list *instructions);
-bool do_noop_swizzle(exec_list *instructions);
bool do_structure_splitting(exec_list *instructions);
-bool do_swizzle_swizzle(exec_list *instructions);
+bool optimize_swizzles(exec_list *instructions);
bool do_vectorize(exec_list *instructions);
bool do_tree_grafting(exec_list *instructions);
bool do_vec_index_to_cond_assign(exec_list *instructions);
diff --git a/src/compiler/glsl/meson.build b/src/compiler/glsl/meson.build
index 76fcafb..6284d0c 100644
--- a/src/compiler/glsl/meson.build
+++ b/src/compiler/glsl/meson.build
@@ -169,11 +169,10 @@ files_libglsl = files(
'opt_function_inlining.cpp',
'opt_if_simplification.cpp',
'opt_minmax.cpp',
- 'opt_noop_swizzle.cpp',
'opt_rebalance_tree.cpp',
'opt_redundant_jumps.cpp',
'opt_structure_splitting.cpp',
- 'opt_swizzle_swizzle.cpp',
+ 'opt_swizzle.cpp',
'opt_tree_grafting.cpp',
'opt_vectorize.cpp',
'program.h',
diff --git a/src/compiler/glsl/opt_noop_swizzle.cpp b/src/compiler/glsl/opt_
swizzle.cpp
similarity index 56%
rename from src/compiler/glsl/opt_noop_swizzle.cpp
rename to src/compiler/glsl/opt_swizzle.cpp
index 41890ab..2fbe362 100644
--- a/src/compiler/glsl/opt_noop_swizzle.cpp
+++ b/src/compiler/glsl/opt_swizzle.cpp
@@ -22,11 +22,14 @@
*/
/**
- * \file opt_noop_swizzle.cpp
+ * \file opt_swizzle.cpp
+ * Optimize swizzle operations.
*
- * If a swizzle doesn't change the order or count of components, then
- * remove the swizzle so that other optimization passes see the value
- * behind it.
+ * First, compact a sequence of swizzled swizzles into a single swizzle.
+ *
+ * If the final resulting swizzle doesn't change the order or count of
+ * components, then remove the swizzle so that other optimization passes
see
+ * the value behind it.
*/
#include "ir.h"
@@ -36,9 +39,9 @@
namespace {
-class ir_noop_swizzle_visitor : public ir_rvalue_visitor {
+class ir_opt_swizzle_visitor : public ir_rvalue_visitor {
public:
- ir_noop_swizzle_visitor()
+ ir_opt_swizzle_visitor()
{
this->progress = false;
}
@@ -50,13 +53,46 @@ public:
} /* unnamed namespace */
void
-ir_noop_swizzle_visitor::handle_rvalue(ir_rvalue **rvalue)
+ir_opt_swizzle_visitor::handle_rvalue(ir_rvalue **rvalue)
{
if (!*rvalue)
return;
ir_swizzle *swiz = (*rvalue)->as_swizzle();
- if (!swiz || swiz->type != swiz->val->type)
+
+ if (!swiz)
+ return;
+
+ ir_swizzle *swiz2;
+
+ while ((swiz2 = swiz->val->as_swizzle()) != NULL) {
+ int mask2[4];
+
+ memset(&mask2, 0, sizeof(mask2));
+ if (swiz2->mask.num_components >= 1)
+ mask2[0] = swiz2->mask.x;
+ if (swiz2->mask.num_components >= 2)
+ mask2[1] = swiz2->mask.y;
+ if (swiz2->mask.num_components >= 3)
+ mask2[2] = swiz2->mask.z;
+ if (swiz2->mask.num_components >= 4)
+ mask2[3] = swiz2->mask.w;
+
+ if (swiz->mask.num_components >= 1)
+ swiz->mask.x = mask2[swiz->mask.x];
+ if (swiz->mask.num_components >= 2)
+ swiz->mask.y = mask2[swiz->mask.y];
+ if (swiz->mask.num_components >= 3)
+ swiz->mask.z = mask2[swiz->mask.z];
+ if (swiz->mask.num_components >= 4)
+ swiz->mask.w = mask2[swiz->mask.w];
+
+ swiz->val = swiz2->val;
+
+ this->progress = true;
+ }
+
+ if (swiz->type != swiz->val->type)
return;
int elems = swiz->val->type->vector_elements;
@@ -74,9 +110,9 @@ ir_noop_swizzle_visitor::handle_rvalue(ir_rvalue
**rvalue)
}
bool
-do_noop_swizzle(exec_list *instructions)
+optimize_swizzles(exec_list *instructions)
{
- ir_noop_swizzle_visitor v;
+ ir_opt_swizzle_visitor v;
visit_list_elements(&v, instructions);
return v.progress;
diff --git a/src/compiler/glsl/opt_swizzle_swizzle.cpp
b/src/compiler/glsl/opt_swizzle_swizzle.cpp
deleted file mode 100644
index 40ce268..0000000
--- a/src/compiler/glsl/opt_swizzle_swizzle.cpp
+++ /dev/null
@@ -1,96 +0,0 @@
-/*
- * Copyright © 2010 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.
- */
-
-/**
- * \file opt_swizzle_swizzle.cpp
- * Compact a sequence of swizzled swizzles into a single swizzle.
- */
-
-#include "ir.h"
-#include "ir_visitor.h"
-#include "ir_optimization.h"
-#include "compiler/glsl_types.h"
-
-namespace {
-
-class ir_swizzle_swizzle_visitor : public ir_hierarchical_visitor {
-public:
- ir_swizzle_swizzle_visitor()
- {
- progress = false;
- }
-
- virtual ir_visitor_status visit_enter(ir_swizzle *);
-
- bool progress;
-};
-
-} /* unnamed namespace */
-
-ir_visitor_status
-ir_swizzle_swizzle_visitor::visit_enter(ir_swizzle *ir)
-{
- ir_swizzle *swiz2;
-
- while ((swiz2 = ir->val->as_swizzle()) != NULL) {
- int mask2[4];
-
- memset(&mask2, 0, sizeof(mask2));
- if (swiz2->mask.num_components >= 1)
- mask2[0] = swiz2->mask.x;
- if (swiz2->mask.num_components >= 2)
- mask2[1] = swiz2->mask.y;
- if (swiz2->mask.num_components >= 3)
- mask2[2] = swiz2->mask.z;
- if (swiz2->mask.num_components >= 4)
- mask2[3] = swiz2->mask.w;
-
- if (ir->mask.num_components >= 1)
- ir->mask.x = mask2[ir->mask.x];
- if (ir->mask.num_components >= 2)
- ir->mask.y = mask2[ir->mask.y];
- if (ir->mask.num_components >= 3)
- ir->mask.z = mask2[ir->mask.z];
- if (ir->mask.num_components >= 4)
- ir->mask.w = mask2[ir->mask.w];
-
- ir->val = swiz2->val;
-
- this->progress = true;
- }
-
- return visit_continue;
-}
-
-/**
- * Does a copy propagation pass on the code present in the instruction
stream.
- */
-bool
-do_swizzle_swizzle(exec_list *instructions)
-{
- ir_swizzle_swizzle_visitor v;
-
- v.run(instructions);
-
- return v.progress;
-}
diff --git a/src/compiler/glsl/test_optpass.cpp b/src/compiler/glsl/test_
optpass.cpp
index c6e9788..5fd1b2b 100644
--- a/src/compiler/glsl/test_optpass.cpp
+++ b/src/compiler/glsl/test_optpass.cpp
@@ -102,12 +102,10 @@ do_optimization(struct exec_list *ir, const char
*optimization,
return lower_if_to_cond_assign(MESA_SHADER_VERTEX, ir, int_0);
} else if (strcmp(optimization, "do_mat_op_to_vec") == 0) {
return do_mat_op_to_vec(ir);
- } else if (strcmp(optimization, "do_noop_swizzle") == 0) {
- return do_noop_swizzle(ir);
+ } else if (strcmp(optimization, "optimize_swizzles") == 0) {
+ return optimize_swizzles(ir);
} else if (strcmp(optimization, "do_structure_splitting") == 0) {
return do_structure_splitting(ir);
- } else if (strcmp(optimization, "do_swizzle_swizzle") == 0) {
- return do_swizzle_swizzle(ir);
} else if (strcmp(optimization, "do_tree_grafting") == 0) {
return do_tree_grafting(ir);
} else if (strcmp(optimization, "do_vec_index_to_cond_assign") == 0) {
--
2.9.5
_______________________________________________
mesa-dev mailing list
mesa-dev at lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20171102/cb3e5bdc/attachment-0001.html>
More information about the mesa-dev
mailing list