On 01/16/2015 08:15 AM, Matt Turner wrote:
> From: Ian Romanick <ian.d.roman...@intel.com>
>
> Some shaders end up with code that looks something like:
>
> if (some_condition)
> result = 1.0;
> else
> result = 0.0;
>
> This pass converts those if-statements to
>
> result = float(some_condition);
>
> total instructions in shared programs: 5934568 -> 5921818 (-0.21%)
> instructions in affected programs: 723433 -> 710683 (-1.76%)
> GAINED: 26
> LOST: 7
I was getting really different results here... on my more recent version
(that should produce the same IR), I was getting basically the reverse
lost / gained. The other results were about the same.
I updated both Mesa and shader-db, and I'm not able to do a full
shader-db run now. At semi-arbitrary points in the run I get
Exception: Only shader 0 found.
shaders/closed/unreal4/sun-temple/180.shader_test
If I run that shader_test by hand, I get multiple shader emitted. I'm
using the older run.py... is it possible that's broken?
Anyway... I pushed an updated branch as bool-optimizations-v2. I was
trying to get results on 33619be and bab3b1d. I know the latter
duplicates one of the NIR optimizations, but with the former patch it
gave some nice gains.
> 12 vertex shaders contain loops that we were now able to unroll.
> Since their instruction count deltas are not representative of this
> change, I've removed them from the stats.
>
> Reviewed-by: Matt Turner <matts...@gmail.com>
> ---
> idr: Want to put an S-o-b on this?
>
> src/glsl/Makefile.sources | 1 +
> src/glsl/glsl_parser_extras.cpp | 1 +
> src/glsl/ir_optimization.h | 1 +
> src/glsl/opt_if_to_bool_cast.cpp | 178
> +++++++++++++++++++++++++++++++++++++++
> 4 files changed, 181 insertions(+)
> create mode 100644 src/glsl/opt_if_to_bool_cast.cpp
>
> diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
> index 006e947..61de09b 100644
> --- a/src/glsl/Makefile.sources
> +++ b/src/glsl/Makefile.sources
> @@ -162,6 +162,7 @@ LIBGLSL_FILES = \
> $(GLSL_SRCDIR)/opt_flip_matrices.cpp \
> $(GLSL_SRCDIR)/opt_function_inlining.cpp \
> $(GLSL_SRCDIR)/opt_if_simplification.cpp \
> + $(GLSL_SRCDIR)/opt_if_to_bool_cast.cpp \
> $(GLSL_SRCDIR)/opt_minmax.cpp \
> $(GLSL_SRCDIR)/opt_noop_swizzle.cpp \
> $(GLSL_SRCDIR)/opt_rebalance_tree.cpp \
> diff --git a/src/glsl/glsl_parser_extras.cpp b/src/glsl/glsl_parser_extras.cpp
> index 6d18cd4..c96b6c9 100644
> --- a/src/glsl/glsl_parser_extras.cpp
> +++ b/src/glsl/glsl_parser_extras.cpp
> @@ -1603,6 +1603,7 @@ do_common_optimization(exec_list *ir, bool linked,
> progress = do_structure_splitting(ir) || progress;
> }
> progress = do_if_simplification(ir) || progress;
> + progress = optimize_if_to_bool_cast(ir) || progress;
> progress = opt_flatten_nested_if_blocks(ir) || progress;
> progress = do_copy_propagation(ir) || progress;
> progress = do_copy_propagation_elements(ir) || progress;
> diff --git a/src/glsl/ir_optimization.h b/src/glsl/ir_optimization.h
> index 34e0b4b..4dbbd7c 100644
> --- a/src/glsl/ir_optimization.h
> +++ b/src/glsl/ir_optimization.h
> @@ -131,6 +131,7 @@ void optimize_dead_builtin_variables(exec_list
> *instructions,
> enum ir_variable_mode other);
>
> bool lower_vertex_id(gl_shader *shader);
> +bool optimize_if_to_bool_cast(exec_list *instructions);
>
> ir_rvalue *
> compare_index_block(exec_list *instructions, ir_variable *index,
> diff --git a/src/glsl/opt_if_to_bool_cast.cpp
> b/src/glsl/opt_if_to_bool_cast.cpp
> new file mode 100644
> index 0000000..32c1f99
> --- /dev/null
> +++ b/src/glsl/opt_if_to_bool_cast.cpp
> @@ -0,0 +1,178 @@
> +/*
> + * Copyright © 2014 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_if_to_bool_cast.cpp
> + * Optimize certain if-statements to just casts from the condition
> + *
> + * Some shaders end up with code that looks something like:
> + *
> + * if (some_condition)
> + * result = 1.0;
> + * else
> + * result = 0.0;
> + *
> + * This pass converts those if-statements to
> + *
> + * result = float(some_condition);
> + */
> +
> +#include "main/imports.h"
> +#include "ir.h"
> +#include "ir_builder.h"
> +#include "program/prog_instruction.h"
> +
> +using namespace ir_builder;
> +
> +namespace {
> +
> +class ir_if_to_bool_cast_visitor : public ir_hierarchical_visitor {
> +public:
> + ir_if_to_bool_cast_visitor()
> + : progress(false)
> + {
> + /* empty */
> + }
> +
> + ir_visitor_status visit_leave(ir_if *);
> +
> + bool progress;
> +};
> +
> +} /* unnamed namespace */
> +
> +bool
> +optimize_if_to_bool_cast(exec_list *instructions)
> +{
> + ir_if_to_bool_cast_visitor v;
> +
> + v.run(instructions);
> + return v.progress;
> +}
> +
> +ir_visitor_status
> +ir_if_to_bool_cast_visitor::visit_leave(ir_if *ir)
> +{
> + /* Both the then-block and the else-block must contain a single
> instruction
> + * that is an assignment.
> + */
> + ir_instruction *const then_inst =
> + (ir_instruction *) ir->then_instructions.get_head();
> + if (then_inst == NULL
> + || !then_inst->next->is_tail_sentinel()
> + || then_inst->ir_type != ir_type_assignment)
> + return visit_continue;
> +
> + ir_instruction *const else_inst =
> + (ir_instruction *) ir->else_instructions.get_head();
> + if (else_inst == NULL
> + || !else_inst->next->is_tail_sentinel()
> + || else_inst->ir_type != ir_type_assignment)
> + return visit_continue;
> +
> + /* That assignment must have the same write-mask and have the same LHS.
> + */
> + ir_assignment *const then_assign = then_inst->as_assignment();
> + ir_assignment *const else_assign = else_inst->as_assignment();
> +
> + assert(then_assign != NULL);
> + assert(else_assign != NULL);
> +
> + if (then_assign->write_mask != else_assign->write_mask
> + || then_assign->condition != NULL
> + || else_assign->condition != NULL)
> + return visit_continue;
> +
> + if (!then_assign->lhs->equals(else_assign->lhs))
> + return visit_continue;
> +
> + /* The RHS of the then-assignment must be 1, and the RHS of the
> + * else-assignment must be 0. Or vice-versa.
> + */
> + ir_constant *const then_rhs = then_assign->rhs->as_constant();
> + ir_constant *const else_rhs = else_assign->rhs->as_constant();
> +
> + if (then_rhs == NULL || else_rhs == NULL)
> + return visit_continue;
> +
> + if (then_rhs->is_zero()) {
> + if (!else_rhs->is_one() && !else_rhs->is_negative_one())
> + return visit_continue;
> + } else if (else_rhs->is_zero()) {
> + if (!then_rhs->is_one() && !then_rhs->is_negative_one())
> + return visit_continue;
> + } else {
> + return visit_continue;
> + }
> +
> + /* Generate the RHS of the new assignment. Note that if the old code
> + * looked like
> + *
> + * x = (cond) ? 0.0 : 1.0;
> + *
> + * we have to generate
> + *
> + * x = float(!cond);
> + */
> + ir_rvalue *const cond = then_rhs->is_zero()
> + ? logic_not(ir->condition) : ir->condition;
> +
> + ir_rvalue *rhs = NULL;
> +
> + switch (then_rhs->type->base_type) {
> + case GLSL_TYPE_UINT:
> + rhs = i2u(b2i(cond));
> + break;
> + case GLSL_TYPE_INT:
> + rhs = b2i(cond);
> + break;
> + case GLSL_TYPE_FLOAT:
> + rhs = b2f(cond);
> + break;
> + case GLSL_TYPE_BOOL:
> + rhs = cond;
> + break;
> + default:
> + unreachable("invalid base type");
> + }
> +
> + /* If the non-zero value was -1, negate the result of the b2[uif]
> + * operation.
> + */
> + if (then_rhs->is_negative_one() || else_rhs->is_negative_one())
> + rhs = neg(rhs);
> +
> + const unsigned size = _mesa_bitcount(then_assign->write_mask);
> +
> + assert(size >= 1 && size <= 4);
> +
> + ir_assignment *const a = assign(then_assign->lhs,
> + swizzle(rhs, SWIZZLE_XXXX, size),
> + then_assign->write_mask);
> +
> + ir->insert_before(a);
> + ir->remove();
> + progress = true;
> +
> + return visit_continue;
> + }
>
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
