[Mesa-dev] [PATCH 06/12] glsl: Add lowering pass for ir_triop_vector_insert
Kenneth Graunke
kenneth at whitecape.org
Mon May 6 23:52:50 PDT 2013
On 05/03/2013 04:07 PM, Ian Romanick wrote:
> From: Ian Romanick <ian.d.romanick at intel.com>
>
> This will eventually replace do_vec_index_to_cond_assign. This lowering
> pass is called in all the places where do_vec_index_to_cond_assign or
> do_vec_index_to_swizzle is called.
>
> v2: Use WRITEMASK_* instead of integer literals. Use a more concise
> method of generating broadcast_index. Both suggested by Eric.
>
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> Reviewed-by: Eric Anholt <eric at anholt.net>
> ---
> src/glsl/Makefile.sources | 1 +
> src/glsl/glsl_parser_extras.cpp | 1 +
> src/glsl/ir_optimization.h | 1 +
> src/glsl/lower_vector_insert.cpp | 160 +++++++++++++++++++++++++++++
> src/mesa/drivers/dri/i965/brw_shader.cpp | 1 +
> src/mesa/program/ir_to_mesa.cpp | 1 +
> src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 1 +
> 7 files changed, 166 insertions(+)
> create mode 100644 src/glsl/lower_vector_insert.cpp
>
> diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
> index 674a05f..8e2dc1b 100644
> --- a/src/glsl/Makefile.sources
> +++ b/src/glsl/Makefile.sources
> @@ -69,6 +69,7 @@ LIBGLSL_FILES = \
> $(GLSL_SRCDIR)/lower_vec_index_to_cond_assign.cpp \
> $(GLSL_SRCDIR)/lower_vec_index_to_swizzle.cpp \
> $(GLSL_SRCDIR)/lower_vector.cpp \
> + $(GLSL_SRCDIR)/lower_vector_insert.cpp \
> $(GLSL_SRCDIR)/lower_output_reads.cpp \
> $(GLSL_SRCDIR)/lower_ubo_reference.cpp \
> $(GLSL_SRCDIR)/opt_algebraic.cpp \
> diff --git a/src/glsl/glsl_parser_extras.cpp b/src/glsl/glsl_parser_extras.cpp
> index 0992294..d38e967 100644
> --- a/src/glsl/glsl_parser_extras.cpp
> +++ b/src/glsl/glsl_parser_extras.cpp
> @@ -1236,6 +1236,7 @@ do_common_optimization(exec_list *ir, bool linked,
> progress = do_algebraic(ir) || progress;
> progress = do_lower_jumps(ir) || progress;
> progress = do_vec_index_to_swizzle(ir) || progress;
> + progress = lower_vector_insert(ir, false) || progress;
> progress = do_swizzle_swizzle(ir) || progress;
> progress = do_noop_swizzle(ir) || progress;
>
> diff --git a/src/glsl/ir_optimization.h b/src/glsl/ir_optimization.h
> index a8885d7..0216e46 100644
> --- a/src/glsl/ir_optimization.h
> +++ b/src/glsl/ir_optimization.h
> @@ -106,6 +106,7 @@ void lower_ubo_reference(struct gl_shader *shader, exec_list *instructions);
> void lower_packed_varyings(void *mem_ctx, unsigned location_base,
> unsigned locations_used, ir_variable_mode mode,
> gl_shader *shader);
> +bool lower_vector_insert(exec_list *instructions, bool lower_nonconstant_index);
> bool optimize_redundant_jumps(exec_list *instructions);
> bool optimize_split_arrays(exec_list *instructions, bool linked);
>
> diff --git a/src/glsl/lower_vector_insert.cpp b/src/glsl/lower_vector_insert.cpp
> new file mode 100644
> index 0000000..3dbc263
> --- /dev/null
> +++ b/src/glsl/lower_vector_insert.cpp
> @@ -0,0 +1,160 @@
> +/*
> + * Copyright © 2013 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 "ir.h"
> +#include "ir_builder.h"
> +#include "ir_rvalue_visitor.h"
> +#include "ir_optimization.h"
> +
> +using namespace ir_builder;
> +
> +class vector_insert_visitor : public ir_rvalue_visitor {
> +public:
> + vector_insert_visitor(bool lower_nonconstant_index)
> + : progress(false), lower_nonconstant_index(lower_nonconstant_index)
> + {
> + factory.instructions = &factory_instructions;
> + }
> +
> + virtual ~vector_insert_visitor()
> + {
> + assert(factory_instructions.is_empty());
> + }
> +
> + virtual void handle_rvalue(ir_rvalue **rv);
> +
> + ir_factory factory;
> + exec_list factory_instructions;
> + bool progress;
> + bool lower_nonconstant_index;
> +};
> +
> +
> +void
> +vector_insert_visitor::handle_rvalue(ir_rvalue **rv)
> +{
> + if (*rv == NULL || (*rv)->ir_type != ir_type_expression)
> + return;
> +
> + ir_expression *const expr = (ir_expression *) *rv;
> +
> + if (likely(expr->operation != ir_triop_vector_insert))
> + return;
> +
> + factory.mem_ctx = ralloc_parent(expr);
> +
> + ir_constant *const idx = expr->operands[2]->constant_expression_value();
> + if (idx != NULL) {
> + /* Replace (vector_insert (vec) (scalar) (index)) with a dereference of
> + * a new temporary. The new temporary gets assigned as
> + *
> + * t = vec
> + * t.mask = scalar
> + *
> + * where mask is the component selected by index.
> + */
> + ir_variable *const temp =
> + factory.make_temp(expr->operands[0]->type, "vec_tmp");
> +
> + const int mask = 1 << idx->value.i[0];
> +
> + factory.emit(assign(temp, expr->operands[0]));
> + factory.emit(assign(temp, expr->operands[1], mask));
> +
> + this->progress = true;
> + *rv = new(factory.mem_ctx) ir_dereference_variable(temp);
> + } else if (this->lower_nonconstant_index) {
> + /* Replace (vector_insert (vec) (scalar) (index)) with a dereference of
> + * a new temporary. The new temporary gets assigned as
> + *
> + * t = vec
> + * if (index == 0)
> + * t.x = scalar
> + * if (index == 1)
> + * t.y = scalar
> + * if (index == 2)
> + * t.z = scalar
> + * if (index == 3)
> + * t.w = scalar
lower_vec_index_to_cond_assign generates conditional moves here, rather
than single-statement if-trees. That seems preferable to me. Tiny
if-trees like this lead to choppy basic-block structure, which cuts some
of the i965's backend optimization opportunity.
Yes, we should fix that (and Eric and I have patches to do so), but I
still think conditional moves would be better.
Feel free to object if you disagree, though.
Otherwise, this pass seems fine, so as is:
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
> + */
> + ir_variable *const temp =
> + factory.make_temp(expr->operands[0]->type, "vec_tmp");
> +
> + const unsigned components = expr->type->vector_elements;
> +
> + ir_constant_data cmp_indices_data;
> + memset(&cmp_indices_data, 0, sizeof(cmp_indices_data));
> + cmp_indices_data.i[0] = 0;
> + cmp_indices_data.i[1] = 1;
> + cmp_indices_data.i[2] = 2;
> + cmp_indices_data.i[3] = 3;
> +
> + ir_constant *const cmp_indices =
> + new(factory.mem_ctx) ir_constant(glsl_type::get_instance(GLSL_TYPE_INT,
> + components,
> + 1),
> + &cmp_indices_data);
> +
> + ir_variable *const cmp_result =
> + factory.make_temp(glsl_type::get_instance(GLSL_TYPE_BOOL,
> + components,
> + 1),
> + "index_condition");
> +
> + ir_variable *const src_temp =
> + factory.make_temp(expr->operands[1]->type, "src_temp");
> +
> + ir_rvalue *broadcast_index =
> + swizzle_for_size(swizzle_xxxx(expr->operands[2]), components);
> +
> + factory.emit(assign(temp, expr->operands[0]));
> + factory.emit(assign(src_temp, expr->operands[1]));
> + factory.emit(assign(cmp_result, equal(broadcast_index, cmp_indices)));
> + factory.emit(if_tree(swizzle_x(cmp_result),
> + assign(temp, src_temp, WRITEMASK_X)));
> + factory.emit(if_tree(swizzle_y(cmp_result),
> + assign(temp, src_temp, WRITEMASK_Y)));
> +
> + if (components > 2)
> + factory.emit(if_tree(swizzle_z(cmp_result),
> + assign(temp, src_temp, WRITEMASK_Z)));
> +
> + if (components > 3)
> + factory.emit(if_tree(swizzle_w(cmp_result),
> + assign(temp, src_temp, WRITEMASK_W)));
> +
> + this->progress = true;
> + *rv = new(factory.mem_ctx) ir_dereference_variable(temp);
> + }
> +
> + base_ir->insert_before(factory.instructions);
> +}
> +
> +bool
> +lower_vector_insert(exec_list *instructions, bool lower_nonconstant_index)
> +{
> + vector_insert_visitor v(lower_nonconstant_index);
> +
> + visit_list_elements(&v, instructions);
> +
> + return v.progress;
> +}
> diff --git a/src/mesa/drivers/dri/i965/brw_shader.cpp b/src/mesa/drivers/dri/i965/brw_shader.cpp
> index 75468b1..6dd1419 100644
> --- a/src/mesa/drivers/dri/i965/brw_shader.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_shader.cpp
> @@ -170,6 +170,7 @@ brw_link_shader(struct gl_context *ctx, struct gl_shader_program *shProg)
> do_lower_texture_projection(shader->ir);
> brw_lower_texture_gradients(intel, shader->ir);
> do_vec_index_to_cond_assign(shader->ir);
> + lower_vector_insert(shader->ir, true);
> brw_do_cubemap_normalize(shader->ir);
> lower_noise(shader->ir);
> lower_quadop_vector(shader->ir, false);
> diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp
> index eb64347..bcc8af6 100644
> --- a/src/mesa/program/ir_to_mesa.cpp
> +++ b/src/mesa/program/ir_to_mesa.cpp
> @@ -3041,6 +3041,7 @@ _mesa_ir_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
> || progress;
>
> progress = do_vec_index_to_cond_assign(ir) || progress;
> + progress = lower_vector_insert(ir, true) || progress;
> } while (progress);
>
> validate_ir_tree(ir);
> diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> index f2eb3e7..5ec9c41 100644
> --- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> +++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> @@ -5233,6 +5233,7 @@ st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
>
> lower_ubo_reference(prog->_LinkedShaders[i], ir);
> do_vec_index_to_cond_assign(ir);
> + lower_vector_insert(ir, true);
> lower_quadop_vector(ir, false);
> lower_noise(ir);
> if (options->MaxIfDepth == 0) {
>
More information about the mesa-dev
mailing list