This pass flips (matrix * vector) operations to (vector * matrixTranspose) for certain built-in matrices (currently gl_ModelViewProjectionMatrix and gl_TextureMatrix).
This is equivalent, but results in dot products rather than multiplies and adds. On some hardware, this is more efficient. This pass is conditionalized on ctx->mvp_with_dp4, the flag drivers set to indicate they prefer dot products. Improves performance in Lightsmark by 1.01131% +/- 0.162069% (n = 10) on a Haswell GT2 system. Passes Piglit on Ivybridge. Signed-off-by: Kenneth Graunke <[email protected]> --- src/glsl/Makefile.sources | 1 + src/glsl/glsl_parser_extras.cpp | 7 +- src/glsl/ir_optimization.h | 4 +- src/glsl/linker.cpp | 2 +- src/glsl/main.cpp | 2 +- src/glsl/opt_flip_matrices.cpp | 122 +++++++++++++++++++++++++++++ src/glsl/test_optpass.cpp | 2 +- src/mesa/drivers/dri/i965/brw_shader.cpp | 2 +- src/mesa/main/ff_fragment_shader.cpp | 2 +- src/mesa/program/ir_to_mesa.cpp | 5 +- src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 3 +- 11 files changed, 142 insertions(+), 10 deletions(-) create mode 100644 src/glsl/opt_flip_matrices.cpp diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources index c294aa4..df1e9d5 100644 --- a/src/glsl/Makefile.sources +++ b/src/glsl/Makefile.sources @@ -80,6 +80,7 @@ LIBGLSL_FILES = \ $(GLSL_SRCDIR)/opt_dead_code.cpp \ $(GLSL_SRCDIR)/opt_dead_code_local.cpp \ $(GLSL_SRCDIR)/opt_dead_functions.cpp \ + $(GLSL_SRCDIR)/opt_flip_matrices.cpp \ $(GLSL_SRCDIR)/opt_function_inlining.cpp \ $(GLSL_SRCDIR)/opt_if_simplification.cpp \ $(GLSL_SRCDIR)/opt_noop_swizzle.cpp \ diff --git a/src/glsl/glsl_parser_extras.cpp b/src/glsl/glsl_parser_extras.cpp index 9740903..e0387b8 100644 --- a/src/glsl/glsl_parser_extras.cpp +++ b/src/glsl/glsl_parser_extras.cpp @@ -1206,7 +1206,8 @@ ast_struct_specifier::ast_struct_specifier(const char *identifier, bool do_common_optimization(exec_list *ir, bool linked, bool uniform_locations_assigned, - unsigned max_unroll_iterations) + unsigned max_unroll_iterations, + bool prefer_dp4) { GLboolean progress = GL_FALSE; @@ -1220,6 +1221,10 @@ do_common_optimization(exec_list *ir, bool linked, progress = do_if_simplification(ir) || progress; progress = do_copy_propagation(ir) || progress; progress = do_copy_propagation_elements(ir) || progress; + + if (prefer_dp4 && !linked) + progress = opt_flip_matrices(ir) || progress; + if (linked) progress = do_dead_code(ir, uniform_locations_assigned) || progress; else diff --git a/src/glsl/ir_optimization.h b/src/glsl/ir_optimization.h index 2454bbe..b404256 100644 --- a/src/glsl/ir_optimization.h +++ b/src/glsl/ir_optimization.h @@ -65,7 +65,8 @@ enum lower_packing_builtins_op { bool do_common_optimization(exec_list *ir, bool linked, bool uniform_locations_assigned, - unsigned max_unroll_iterations); + unsigned max_unroll_iterations, + bool prefer_dp4); bool do_algebraic(exec_list *instructions); bool do_constant_folding(exec_list *instructions); @@ -78,6 +79,7 @@ bool do_dead_code(exec_list *instructions, bool uniform_locations_assigned); bool do_dead_code_local(exec_list *instructions); bool do_dead_code_unlinked(exec_list *instructions); bool do_dead_functions(exec_list *instructions); +bool opt_flip_matrices(exec_list *instructions); bool do_function_inlining(exec_list *instructions); bool do_lower_jumps(exec_list *instructions, bool pull_out_jumps = true, bool lower_sub_return = true, bool lower_main_return = false, bool lower_continue = false, bool lower_break = false); bool do_lower_texture_projection(exec_list *instructions); diff --git a/src/glsl/linker.cpp b/src/glsl/linker.cpp index 2b30d2b..67a6c5a 100644 --- a/src/glsl/linker.cpp +++ b/src/glsl/linker.cpp @@ -1767,7 +1767,7 @@ link_shaders(struct gl_context *ctx, struct gl_shader_program *prog) unsigned max_unroll = ctx->ShaderCompilerOptions[i].MaxUnrollIterations; - while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false, max_unroll)) + while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false, max_unroll, ctx->mvp_with_dp4)) ; } diff --git a/src/glsl/main.cpp b/src/glsl/main.cpp index ce084b4..13dfdd3 100644 --- a/src/glsl/main.cpp +++ b/src/glsl/main.cpp @@ -176,7 +176,7 @@ compile_shader(struct gl_context *ctx, struct gl_shader *shader) if (!state->error && !shader->ir->is_empty()) { bool progress; do { - progress = do_common_optimization(shader->ir, false, false, 32); + progress = do_common_optimization(shader->ir, false, false, 32, false); } while (progress); validate_ir_tree(shader->ir); diff --git a/src/glsl/opt_flip_matrices.cpp b/src/glsl/opt_flip_matrices.cpp new file mode 100644 index 0000000..497513f --- /dev/null +++ b/src/glsl/opt_flip_matrices.cpp @@ -0,0 +1,122 @@ +/* + * 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. + */ + +/** + * \file opt_flip_matrices.cpp + * + * Convert (matrix * vector) operations to (vector * matrixTranspose), + * which can be done using dot products rather than multiplies and adds. + * On some hardware, this is more efficient. + * + * This currently only does the conversion for built-in matrices which + * already have transposed equivalents. Namely, gl_ModelViewProjectionMatrix + * and gl_TextureMatrix. + */ +#include "ir.h" +#include "ir_optimization.h" +#include "main/macros.h" + +namespace { +class matrix_flipper : public ir_hierarchical_visitor { +public: + matrix_flipper(exec_list *instructions) + { + progress = false; + mvp_transpose = NULL; + texmat_transpose = NULL; + + foreach_list(n, instructions) { + ir_instruction *ir = (ir_instruction *) n; + ir_variable *var = ir->as_variable(); + if (!var) + continue; + if (strcmp(var->name, "gl_ModelViewProjectionMatrixTranspose") == 0) + mvp_transpose = var; + if (strcmp(var->name, "gl_TextureMatrixTranspose") == 0) + texmat_transpose = var; + } + } + + ir_visitor_status visit_enter(ir_expression *ir); + + bool progress; + +private: + ir_variable *mvp_transpose; + ir_variable *texmat_transpose; +}; +} + +ir_visitor_status +matrix_flipper::visit_enter(ir_expression *ir) +{ + if (ir->operation != ir_binop_mul || + !ir->operands[0]->type->is_matrix() || + !ir->operands[1]->type->is_vector()) + return visit_continue; + + ir_variable *mat_var = ir->operands[0]->variable_referenced(); + if (!mat_var) + return visit_continue; + + if (mvp_transpose && + strcmp(mat_var->name, "gl_ModelViewProjectionMatrix") == 0) { + ir_dereference_variable *deref = ir->operands[0]->as_dereference_variable(); + assert(deref && deref->var == mat_var); + + void *mem_ctx = ralloc_parent(ir); + + ir->operands[0] = ir->operands[1]; + ir->operands[1] = new(mem_ctx) ir_dereference_variable(mvp_transpose); + + progress = true; + } else if (texmat_transpose && + strcmp(mat_var->name, "gl_TextureMatrix") == 0) { + ir_dereference_array *array_ref = ir->operands[0]->as_dereference_array(); + assert(array_ref != NULL); + ir_dereference_variable *var_ref = array_ref->array->as_dereference_variable(); + assert(var_ref && var_ref->var == mat_var); + + ir->operands[0] = ir->operands[1]; + ir->operands[1] = array_ref; + + var_ref->var = texmat_transpose; + + texmat_transpose->max_array_access = + MAX2(texmat_transpose->max_array_access, mat_var->max_array_access); + + progress = true; + } + + return visit_continue; +} + +bool +opt_flip_matrices(struct exec_list *instructions) +{ + matrix_flipper v(instructions); + + visit_list_elements(&v, instructions); + + return v.progress; +} diff --git a/src/glsl/test_optpass.cpp b/src/glsl/test_optpass.cpp index 117b0b0..87c3705 100644 --- a/src/glsl/test_optpass.cpp +++ b/src/glsl/test_optpass.cpp @@ -64,7 +64,7 @@ do_optimization(struct exec_list *ir, const char *optimization) if (sscanf(optimization, "do_common_optimization ( %d , %d ) ", &int_0, &int_1) == 2) { - return do_common_optimization(ir, int_0 != 0, false, int_1); + return do_common_optimization(ir, int_0 != 0, false, int_1, false); } else if (strcmp(optimization, "do_algebraic") == 0) { return do_algebraic(ir); } else if (strcmp(optimization, "do_constant_folding") == 0) { diff --git a/src/mesa/drivers/dri/i965/brw_shader.cpp b/src/mesa/drivers/dri/i965/brw_shader.cpp index 1a52039..54537b1 100644 --- a/src/mesa/drivers/dri/i965/brw_shader.cpp +++ b/src/mesa/drivers/dri/i965/brw_shader.cpp @@ -207,7 +207,7 @@ brw_link_shader(struct gl_context *ctx, struct gl_shader_program *shProg) false /* loops */ ) || progress; - progress = do_common_optimization(shader->ir, true, true, 32) + progress = do_common_optimization(shader->ir, true, true, 32, true) || progress; } while (progress); diff --git a/src/mesa/main/ff_fragment_shader.cpp b/src/mesa/main/ff_fragment_shader.cpp index 01a4542..fb45578 100644 --- a/src/mesa/main/ff_fragment_shader.cpp +++ b/src/mesa/main/ff_fragment_shader.cpp @@ -1336,7 +1336,7 @@ create_new_program(struct gl_context *ctx, struct state_key *key) validate_ir_tree(p.shader->ir); - while (do_common_optimization(p.shader->ir, false, false, 32)) + while (do_common_optimization(p.shader->ir, false, false, 32, ctx->mvp_with_dp4)) ; reparent_ir(p.shader->ir, p.shader->ir); diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp index 14cf5ba..9456571 100644 --- a/src/mesa/program/ir_to_mesa.cpp +++ b/src/mesa/program/ir_to_mesa.cpp @@ -3012,7 +3012,8 @@ _mesa_ir_link_shader(struct gl_context *ctx, struct gl_shader_program *prog) progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress; progress = do_common_optimization(ir, true, true, - options->MaxUnrollIterations) + options->MaxUnrollIterations, + ctx->mvp_with_dp4) || progress; progress = lower_quadop_vector(ir, true) || progress; @@ -3122,7 +3123,7 @@ _mesa_glsl_compile_shader(struct gl_context *ctx, struct gl_shader *shader) /* Do some optimization at compile time to reduce shader IR size * and reduce later work if the same shader is linked multiple times */ - while (do_common_optimization(shader->ir, false, false, 32)) + while (do_common_optimization(shader->ir, false, false, 32, ctx->mvp_with_dp4)) ; validate_ir_tree(shader->ir); diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp index 338c652..a23c705 100644 --- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp +++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp @@ -5228,7 +5228,8 @@ st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog) progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress; progress = do_common_optimization(ir, true, true, - options->MaxUnrollIterations) + options->MaxUnrollIterations, + ctx->mvp_with_dp4) || progress; progress = lower_if_to_cond_assign(ir, options->MaxIfDepth) || progress; -- 1.8.2 _______________________________________________ mesa-dev mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/mesa-dev
