From 0bf07d32cbd460220c67d726900772cf3692746d Mon Sep 17 00:00:00 2001 From: marha Date: Tue, 1 Feb 2011 10:02:38 +0000 Subject: libX11 libXinerama mesa git update 1 Feb 2011 --- mesalib/src/mesa/program/ir_to_mesa.cpp | 42 +- mesalib/src/mesa/program/register_allocate.c | 913 +++++++++++++-------------- mesalib/src/mesa/program/sampler.cpp | 278 ++++---- 3 files changed, 614 insertions(+), 619 deletions(-) (limited to 'mesalib/src/mesa/program') diff --git a/mesalib/src/mesa/program/ir_to_mesa.cpp b/mesalib/src/mesa/program/ir_to_mesa.cpp index 404b6c646..3794c0de0 100644 --- a/mesalib/src/mesa/program/ir_to_mesa.cpp +++ b/mesalib/src/mesa/program/ir_to_mesa.cpp @@ -105,13 +105,13 @@ extern ir_to_mesa_src_reg ir_to_mesa_undef; class ir_to_mesa_instruction : public exec_node { public: - /* Callers of this talloc-based new need not call delete. It's - * easier to just talloc_free 'ctx' (or any of its ancestors). */ + /* Callers of this ralloc-based new need not call delete. It's + * easier to just ralloc_free 'ctx' (or any of its ancestors). */ static void* operator new(size_t size, void *ctx) { void *node; - node = talloc_zero_size(ctx, size); + node = rzalloc_size(ctx, size); assert(node != NULL); return node; @@ -318,7 +318,7 @@ fail_link(struct gl_shader_program *prog, const char *fmt, ...) { va_list args; va_start(args, fmt); - prog->InfoLog = talloc_vasprintf_append(prog->InfoLog, fmt, args); + ralloc_vasprintf_append(&prog->InfoLog, fmt, args); va_end(args); prog->LinkStatus = GL_FALSE; @@ -1570,7 +1570,7 @@ ir_to_mesa_visitor::visit(ir_dereference_array *ir) this->result, src_reg_for_float(element_size)); } - src_reg.reladdr = talloc(mem_ctx, ir_to_mesa_src_reg); + src_reg.reladdr = ralloc(mem_ctx, ir_to_mesa_src_reg); memcpy(src_reg.reladdr, &index_reg, sizeof(index_reg)); } @@ -1927,7 +1927,7 @@ ir_to_mesa_visitor::get_function_signature(ir_function_signature *sig) return entry; } - entry = talloc(mem_ctx, function_entry); + entry = ralloc(mem_ctx, function_entry); entry->sig = sig; entry->sig_id = this->next_signature_id++; entry->bgn_inst = NULL; @@ -2264,12 +2264,12 @@ ir_to_mesa_visitor::ir_to_mesa_visitor() next_temp = 1; next_signature_id = 1; current_function = NULL; - mem_ctx = talloc_new(NULL); + mem_ctx = ralloc_context(NULL); } ir_to_mesa_visitor::~ir_to_mesa_visitor() { - talloc_free(mem_ctx); + ralloc_free(mem_ctx); } static struct prog_src_register @@ -2318,8 +2318,8 @@ set_branchtargets(ir_to_mesa_visitor *v, } } - if_stack = talloc_zero_array(v->mem_ctx, int, if_count); - loop_stack = talloc_zero_array(v->mem_ctx, int, loop_count); + if_stack = rzalloc_array(v->mem_ctx, int, if_count); + loop_stack = rzalloc_array(v->mem_ctx, int, loop_count); for (i = 0; i < num_instructions; i++) { switch (mesa_instructions[i].Opcode) { @@ -2462,7 +2462,7 @@ add_uniforms_to_parameters_list(struct gl_shader_program *shader_program, unsigned int next_sampler = 0, num_uniforms = 0; struct uniform_sort *sorted_uniforms; - sorted_uniforms = talloc_array(NULL, struct uniform_sort, + sorted_uniforms = ralloc_array(NULL, struct uniform_sort, shader_program->Uniforms->NumUniforms); for (i = 0; i < shader_program->Uniforms->NumUniforms; i++) { @@ -2541,7 +2541,7 @@ add_uniforms_to_parameters_list(struct gl_shader_program *shader_program, } } - talloc_free(sorted_uniforms); + ralloc_free(sorted_uniforms); } static void @@ -2557,7 +2557,7 @@ set_uniform_initializer(struct gl_context *ctx, void *mem_ctx, for (unsigned int i = 0; i < type->length; i++) { const glsl_type *field_type = type->fields.structure[i].type; - const char *field_name = talloc_asprintf(mem_ctx, "%s.%s", name, + const char *field_name = ralloc_asprintf(mem_ctx, "%s.%s", name, type->fields.structure[i].name); set_uniform_initializer(ctx, mem_ctx, shader_program, field_name, field_type, field_constant); @@ -2588,7 +2588,7 @@ set_uniform_initializer(struct gl_context *ctx, void *mem_ctx, void *values; if (element_type->base_type == GLSL_TYPE_BOOL) { - int *conv = talloc_array(mem_ctx, int, element_type->components()); + int *conv = ralloc_array(mem_ctx, int, element_type->components()); for (unsigned int j = 0; j < element_type->components(); j++) { conv[j] = element->value.b[j]; } @@ -2634,14 +2634,14 @@ set_uniform_initializers(struct gl_context *ctx, continue; if (!mem_ctx) - mem_ctx = talloc_new(NULL); + mem_ctx = ralloc_context(NULL); set_uniform_initializer(ctx, mem_ctx, shader_program, var->name, var->type, var->constant_value); } } - talloc_free(mem_ctx); + ralloc_free(mem_ctx); } /* @@ -2667,7 +2667,7 @@ set_uniform_initializers(struct gl_context *ctx, void ir_to_mesa_visitor::copy_propagate(void) { - ir_to_mesa_instruction **acp = talloc_zero_array(mem_ctx, + ir_to_mesa_instruction **acp = rzalloc_array(mem_ctx, ir_to_mesa_instruction *, this->next_temp * 4); @@ -2771,7 +2771,7 @@ ir_to_mesa_visitor::copy_propagate(void) } } - talloc_free(acp); + ralloc_free(acp); } @@ -2870,7 +2870,7 @@ get_mesa_program(struct gl_context *ctx, mesa_instructions = (struct prog_instruction *)calloc(num_instructions, sizeof(*mesa_instructions)); - mesa_instruction_annotation = talloc_array(v.mem_ctx, ir_instruction *, + mesa_instruction_annotation = ralloc_array(v.mem_ctx, ir_instruction *, num_instructions); v.copy_propagate(); @@ -3127,7 +3127,7 @@ _mesa_glsl_compile_shader(struct gl_context *ctx, struct gl_shader *shader) _mesa_glsl_lexer_dtor(state); } - talloc_free(shader->ir); + ralloc_free(shader->ir); shader->ir = new(shader) exec_list; if (!state->error && !state->translation_unit.is_empty()) _mesa_ast_to_hir(shader->ir, state); @@ -3174,7 +3174,7 @@ _mesa_glsl_compile_shader(struct gl_context *ctx, struct gl_shader *shader) /* Retain any live IR, but trash the rest. */ reparent_ir(shader->ir, shader->ir); - talloc_free(state); + ralloc_free(state); if (shader->CompileStatus) { if (!ctx->Driver.CompileShader(ctx, shader)) diff --git a/mesalib/src/mesa/program/register_allocate.c b/mesalib/src/mesa/program/register_allocate.c index 8992b91b6..95a9bde40 100644 --- a/mesalib/src/mesa/program/register_allocate.c +++ b/mesalib/src/mesa/program/register_allocate.c @@ -1,458 +1,455 @@ -/* - * 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. - * - * Authors: - * Eric Anholt - * - */ - -/** @file register_allocate.c - * - * Graph-coloring register allocator. - */ - -#include - -#include "main/imports.h" -#include "main/macros.h" -#include "main/mtypes.h" -#include "register_allocate.h" - -struct ra_reg { - GLboolean *conflicts; - unsigned int *conflict_list; - unsigned int conflict_list_size; - unsigned int num_conflicts; -}; - -struct ra_regs { - struct ra_reg *regs; - unsigned int count; - - struct ra_class **classes; - unsigned int class_count; -}; - -struct ra_class { - GLboolean *regs; - - /** - * p_B in Runeson/Nyström paper. - * - * This is "how many regs are in the set." - */ - unsigned int p; - - /** - * q_B,C in Runeson/Nyström paper. - */ - unsigned int *q; -}; - -struct ra_node { - GLboolean *adjacency; - unsigned int *adjacency_list; - unsigned int class; - unsigned int adjacency_count; - unsigned int reg; - GLboolean in_stack; - float spill_cost; -}; - -struct ra_graph { - struct ra_regs *regs; - /** - * the variables that need register allocation. - */ - struct ra_node *nodes; - unsigned int count; /**< count of nodes. */ - - unsigned int *stack; - unsigned int stack_count; -}; - -struct ra_regs * -ra_alloc_reg_set(unsigned int count) -{ - unsigned int i; - struct ra_regs *regs; - - regs = talloc_zero(NULL, struct ra_regs); - regs->count = count; - regs->regs = talloc_zero_array(regs, struct ra_reg, count); - - for (i = 0; i < count; i++) { - regs->regs[i].conflicts = talloc_zero_array(regs->regs, GLboolean, count); - regs->regs[i].conflicts[i] = GL_TRUE; - - regs->regs[i].conflict_list = talloc_array(regs->regs, unsigned int, 4); - regs->regs[i].conflict_list_size = 4; - regs->regs[i].conflict_list[0] = i; - regs->regs[i].num_conflicts = 1; - } - - return regs; -} - -static void -ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2) -{ - struct ra_reg *reg1 = ®s->regs[r1]; - - if (reg1->conflict_list_size == reg1->num_conflicts) { - reg1->conflict_list_size *= 2; - reg1->conflict_list = talloc_realloc(regs->regs, - reg1->conflict_list, - unsigned int, - reg1->conflict_list_size); - } - reg1->conflict_list[reg1->num_conflicts++] = r2; - reg1->conflicts[r2] = GL_TRUE; -} - -void -ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2) -{ - if (!regs->regs[r1].conflicts[r2]) { - ra_add_conflict_list(regs, r1, r2); - ra_add_conflict_list(regs, r2, r1); - } -} - -unsigned int -ra_alloc_reg_class(struct ra_regs *regs) -{ - struct ra_class *class; - - regs->classes = talloc_realloc(regs->regs, regs->classes, - struct ra_class *, - regs->class_count + 1); - - class = talloc_zero(regs, struct ra_class); - regs->classes[regs->class_count] = class; - - class->regs = talloc_zero_array(class, GLboolean, regs->count); - - return regs->class_count++; -} - -void -ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r) -{ - struct ra_class *class = regs->classes[c]; - - class->regs[r] = GL_TRUE; - class->p++; -} - -/** - * Must be called after all conflicts and register classes have been - * set up and before the register set is used for allocation. - */ -void -ra_set_finalize(struct ra_regs *regs) -{ - unsigned int b, c; - - for (b = 0; b < regs->class_count; b++) { - regs->classes[b]->q = talloc_array(regs, unsigned int, regs->class_count); - } - - /* Compute, for each class B and C, how many regs of B an - * allocation to C could conflict with. - */ - for (b = 0; b < regs->class_count; b++) { - for (c = 0; c < regs->class_count; c++) { - unsigned int rc; - int max_conflicts = 0; - - for (rc = 0; rc < regs->count; rc++) { - int conflicts = 0; - int i; - - if (!regs->classes[c]->regs[rc]) - continue; - - for (i = 0; i < regs->regs[rc].num_conflicts; i++) { - unsigned int rb = regs->regs[rc].conflict_list[i]; - if (regs->classes[b]->regs[rb]) - conflicts++; - } - max_conflicts = MAX2(max_conflicts, conflicts); - } - regs->classes[b]->q[c] = max_conflicts; - } - } -} - -static void -ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2) -{ - g->nodes[n1].adjacency[n2] = GL_TRUE; - g->nodes[n1].adjacency_list[g->nodes[n1].adjacency_count] = n2; - g->nodes[n1].adjacency_count++; -} - -struct ra_graph * -ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count) -{ - struct ra_graph *g; - unsigned int i; - - g = talloc_zero(regs, struct ra_graph); - g->regs = regs; - g->nodes = talloc_zero_array(g, struct ra_node, count); - g->count = count; - - g->stack = talloc_zero_array(g, unsigned int, count); - - for (i = 0; i < count; i++) { - g->nodes[i].adjacency = talloc_zero_array(g, GLboolean, count); - g->nodes[i].adjacency_list = talloc_array(g, unsigned int, count); - g->nodes[i].adjacency_count = 0; - ra_add_node_adjacency(g, i, i); - g->nodes[i].reg = ~0; - } - - return g; -} - -void -ra_set_node_class(struct ra_graph *g, - unsigned int n, unsigned int class) -{ - g->nodes[n].class = class; -} - -void -ra_add_node_interference(struct ra_graph *g, - unsigned int n1, unsigned int n2) -{ - if (!g->nodes[n1].adjacency[n2]) { - ra_add_node_adjacency(g, n1, n2); - ra_add_node_adjacency(g, n2, n1); - } -} - -static GLboolean pq_test(struct ra_graph *g, unsigned int n) -{ - unsigned int j; - unsigned int q = 0; - int n_class = g->nodes[n].class; - - for (j = 0; j < g->nodes[n].adjacency_count; j++) { - unsigned int n2 = g->nodes[n].adjacency_list[j]; - unsigned int n2_class = g->nodes[n2].class; - - if (n != n2 && !g->nodes[n2].in_stack) { - q += g->regs->classes[n_class]->q[n2_class]; - } - } - - return q < g->regs->classes[n_class]->p; -} - -/** - * Simplifies the interference graph by pushing all - * trivially-colorable nodes into a stack of nodes to be colored, - * removing them from the graph, and rinsing and repeating. - * - * Returns GL_TRUE if all nodes were removed from the graph. GL_FALSE - * means that either spilling will be required, or optimistic coloring - * should be applied. - */ -GLboolean -ra_simplify(struct ra_graph *g) -{ - GLboolean progress = GL_TRUE; - int i; - - while (progress) { - progress = GL_FALSE; - - for (i = g->count - 1; i >= 0; i--) { - if (g->nodes[i].in_stack) - continue; - - if (pq_test(g, i)) { - g->stack[g->stack_count] = i; - g->stack_count++; - g->nodes[i].in_stack = GL_TRUE; - progress = GL_TRUE; - } - } - } - - for (i = 0; i < g->count; i++) { - if (!g->nodes[i].in_stack) - return GL_FALSE; - } - - return GL_TRUE; -} - -/** - * Pops nodes from the stack back into the graph, coloring them with - * registers as they go. - * - * If all nodes were trivially colorable, then this must succeed. If - * not (optimistic coloring), then it may return GL_FALSE; - */ -GLboolean -ra_select(struct ra_graph *g) -{ - int i; - - while (g->stack_count != 0) { - unsigned int r; - int n = g->stack[g->stack_count - 1]; - struct ra_class *c = g->regs->classes[g->nodes[n].class]; - - /* Find the lowest-numbered reg which is not used by a member - * of the graph adjacent to us. - */ - for (r = 0; r < g->regs->count; r++) { - if (!c->regs[r]) - continue; - - /* Check if any of our neighbors conflict with this register choice. */ - for (i = 0; i < g->nodes[n].adjacency_count; i++) { - unsigned int n2 = g->nodes[n].adjacency_list[i]; - - if (!g->nodes[n2].in_stack && - g->regs->regs[r].conflicts[g->nodes[n2].reg]) { - break; - } - } - if (i == g->nodes[n].adjacency_count) - break; - } - if (r == g->regs->count) - return GL_FALSE; - - g->nodes[n].reg = r; - g->nodes[n].in_stack = GL_FALSE; - g->stack_count--; - } - - return GL_TRUE; -} - -/** - * Optimistic register coloring: Just push the remaining nodes - * on the stack. They'll be colored first in ra_select(), and - * if they succeed then the locally-colorable nodes are still - * locally-colorable and the rest of the register allocation - * will succeed. - */ -void -ra_optimistic_color(struct ra_graph *g) -{ - unsigned int i; - - for (i = 0; i < g->count; i++) { - if (g->nodes[i].in_stack) - continue; - - g->stack[g->stack_count] = i; - g->stack_count++; - g->nodes[i].in_stack = GL_TRUE; - } -} - -GLboolean -ra_allocate_no_spills(struct ra_graph *g) -{ - if (!ra_simplify(g)) { - ra_optimistic_color(g); - } - return ra_select(g); -} - -unsigned int -ra_get_node_reg(struct ra_graph *g, unsigned int n) -{ - return g->nodes[n].reg; -} - -static float -ra_get_spill_benefit(struct ra_graph *g, unsigned int n) -{ - int j; - float benefit = 0; - int n_class = g->nodes[n].class; - - /* Define the benefit of eliminating an interference between n, n2 - * through spilling as q(C, B) / p(C). This is similar to the - * "count number of edges" approach of traditional graph coloring, - * but takes classes into account. - */ - for (j = 0; j < g->nodes[n].adjacency_count; j++) { - unsigned int n2 = g->nodes[n].adjacency_list[j]; - if (n != n2) { - unsigned int n2_class = g->nodes[n2].class; - benefit += ((float)g->regs->classes[n_class]->q[n2_class] / - g->regs->classes[n_class]->p); - } - } - - return benefit; -} - -/** - * Returns a node number to be spilled according to the cost/benefit using - * the pq test, or -1 if there are no spillable nodes. - */ -int -ra_get_best_spill_node(struct ra_graph *g) -{ - unsigned int best_node = -1; - unsigned int best_benefit = 0.0; - unsigned int n; - - for (n = 0; n < g->count; n++) { - float cost = g->nodes[n].spill_cost; - float benefit; - - if (cost <= 0.0) - continue; - - benefit = ra_get_spill_benefit(g, n); - - if (benefit / cost > best_benefit) { - best_benefit = benefit / cost; - best_node = n; - } - } - - return best_node; -} - -/** - * Only nodes with a spill cost set (cost != 0.0) will be considered - * for register spilling. - */ -void -ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost) -{ - g->nodes[n].spill_cost = cost; -} +/* + * 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. + * + * Authors: + * Eric Anholt + * + */ + +/** @file register_allocate.c + * + * Graph-coloring register allocator. + */ + +#include + +#include "main/imports.h" +#include "main/macros.h" +#include "main/mtypes.h" +#include "register_allocate.h" + +struct ra_reg { + GLboolean *conflicts; + unsigned int *conflict_list; + unsigned int conflict_list_size; + unsigned int num_conflicts; +}; + +struct ra_regs { + struct ra_reg *regs; + unsigned int count; + + struct ra_class **classes; + unsigned int class_count; +}; + +struct ra_class { + GLboolean *regs; + + /** + * p_B in Runeson/Nyström paper. + * + * This is "how many regs are in the set." + */ + unsigned int p; + + /** + * q_B,C in Runeson/Nyström paper. + */ + unsigned int *q; +}; + +struct ra_node { + GLboolean *adjacency; + unsigned int *adjacency_list; + unsigned int class; + unsigned int adjacency_count; + unsigned int reg; + GLboolean in_stack; + float spill_cost; +}; + +struct ra_graph { + struct ra_regs *regs; + /** + * the variables that need register allocation. + */ + struct ra_node *nodes; + unsigned int count; /**< count of nodes. */ + + unsigned int *stack; + unsigned int stack_count; +}; + +struct ra_regs * +ra_alloc_reg_set(unsigned int count) +{ + unsigned int i; + struct ra_regs *regs; + + regs = rzalloc(NULL, struct ra_regs); + regs->count = count; + regs->regs = rzalloc_array(regs, struct ra_reg, count); + + for (i = 0; i < count; i++) { + regs->regs[i].conflicts = rzalloc_array(regs->regs, GLboolean, count); + regs->regs[i].conflicts[i] = GL_TRUE; + + regs->regs[i].conflict_list = ralloc_array(regs->regs, unsigned int, 4); + regs->regs[i].conflict_list_size = 4; + regs->regs[i].conflict_list[0] = i; + regs->regs[i].num_conflicts = 1; + } + + return regs; +} + +static void +ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2) +{ + struct ra_reg *reg1 = ®s->regs[r1]; + + if (reg1->conflict_list_size == reg1->num_conflicts) { + reg1->conflict_list_size *= 2; + reg1->conflict_list = reralloc(regs->regs, reg1->conflict_list, + unsigned int, reg1->conflict_list_size); + } + reg1->conflict_list[reg1->num_conflicts++] = r2; + reg1->conflicts[r2] = GL_TRUE; +} + +void +ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2) +{ + if (!regs->regs[r1].conflicts[r2]) { + ra_add_conflict_list(regs, r1, r2); + ra_add_conflict_list(regs, r2, r1); + } +} + +unsigned int +ra_alloc_reg_class(struct ra_regs *regs) +{ + struct ra_class *class; + + regs->classes = reralloc(regs->regs, regs->classes, struct ra_class *, + regs->class_count + 1); + + class = rzalloc(regs, struct ra_class); + regs->classes[regs->class_count] = class; + + class->regs = rzalloc_array(class, GLboolean, regs->count); + + return regs->class_count++; +} + +void +ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r) +{ + struct ra_class *class = regs->classes[c]; + + class->regs[r] = GL_TRUE; + class->p++; +} + +/** + * Must be called after all conflicts and register classes have been + * set up and before the register set is used for allocation. + */ +void +ra_set_finalize(struct ra_regs *regs) +{ + unsigned int b, c; + + for (b = 0; b < regs->class_count; b++) { + regs->classes[b]->q = ralloc_array(regs, unsigned int, regs->class_count); + } + + /* Compute, for each class B and C, how many regs of B an + * allocation to C could conflict with. + */ + for (b = 0; b < regs->class_count; b++) { + for (c = 0; c < regs->class_count; c++) { + unsigned int rc; + int max_conflicts = 0; + + for (rc = 0; rc < regs->count; rc++) { + int conflicts = 0; + int i; + + if (!regs->classes[c]->regs[rc]) + continue; + + for (i = 0; i < regs->regs[rc].num_conflicts; i++) { + unsigned int rb = regs->regs[rc].conflict_list[i]; + if (regs->classes[b]->regs[rb]) + conflicts++; + } + max_conflicts = MAX2(max_conflicts, conflicts); + } + regs->classes[b]->q[c] = max_conflicts; + } + } +} + +static void +ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2) +{ + g->nodes[n1].adjacency[n2] = GL_TRUE; + g->nodes[n1].adjacency_list[g->nodes[n1].adjacency_count] = n2; + g->nodes[n1].adjacency_count++; +} + +struct ra_graph * +ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count) +{ + struct ra_graph *g; + unsigned int i; + + g = rzalloc(regs, struct ra_graph); + g->regs = regs; + g->nodes = rzalloc_array(g, struct ra_node, count); + g->count = count; + + g->stack = rzalloc_array(g, unsigned int, count); + + for (i = 0; i < count; i++) { + g->nodes[i].adjacency = rzalloc_array(g, GLboolean, count); + g->nodes[i].adjacency_list = ralloc_array(g, unsigned int, count); + g->nodes[i].adjacency_count = 0; + ra_add_node_adjacency(g, i, i); + g->nodes[i].reg = ~0; + } + + return g; +} + +void +ra_set_node_class(struct ra_graph *g, + unsigned int n, unsigned int class) +{ + g->nodes[n].class = class; +} + +void +ra_add_node_interference(struct ra_graph *g, + unsigned int n1, unsigned int n2) +{ + if (!g->nodes[n1].adjacency[n2]) { + ra_add_node_adjacency(g, n1, n2); + ra_add_node_adjacency(g, n2, n1); + } +} + +static GLboolean pq_test(struct ra_graph *g, unsigned int n) +{ + unsigned int j; + unsigned int q = 0; + int n_class = g->nodes[n].class; + + for (j = 0; j < g->nodes[n].adjacency_count; j++) { + unsigned int n2 = g->nodes[n].adjacency_list[j]; + unsigned int n2_class = g->nodes[n2].class; + + if (n != n2 && !g->nodes[n2].in_stack) { + q += g->regs->classes[n_class]->q[n2_class]; + } + } + + return q < g->regs->classes[n_class]->p; +} + +/** + * Simplifies the interference graph by pushing all + * trivially-colorable nodes into a stack of nodes to be colored, + * removing them from the graph, and rinsing and repeating. + * + * Returns GL_TRUE if all nodes were removed from the graph. GL_FALSE + * means that either spilling will be required, or optimistic coloring + * should be applied. + */ +GLboolean +ra_simplify(struct ra_graph *g) +{ + GLboolean progress = GL_TRUE; + int i; + + while (progress) { + progress = GL_FALSE; + + for (i = g->count - 1; i >= 0; i--) { + if (g->nodes[i].in_stack) + continue; + + if (pq_test(g, i)) { + g->stack[g->stack_count] = i; + g->stack_count++; + g->nodes[i].in_stack = GL_TRUE; + progress = GL_TRUE; + } + } + } + + for (i = 0; i < g->count; i++) { + if (!g->nodes[i].in_stack) + return GL_FALSE; + } + + return GL_TRUE; +} + +/** + * Pops nodes from the stack back into the graph, coloring them with + * registers as they go. + * + * If all nodes were trivially colorable, then this must succeed. If + * not (optimistic coloring), then it may return GL_FALSE; + */ +GLboolean +ra_select(struct ra_graph *g) +{ + int i; + + while (g->stack_count != 0) { + unsigned int r; + int n = g->stack[g->stack_count - 1]; + struct ra_class *c = g->regs->classes[g->nodes[n].class]; + + /* Find the lowest-numbered reg which is not used by a member + * of the graph adjacent to us. + */ + for (r = 0; r < g->regs->count; r++) { + if (!c->regs[r]) + continue; + + /* Check if any of our neighbors conflict with this register choice. */ + for (i = 0; i < g->nodes[n].adjacency_count; i++) { + unsigned int n2 = g->nodes[n].adjacency_list[i]; + + if (!g->nodes[n2].in_stack && + g->regs->regs[r].conflicts[g->nodes[n2].reg]) { + break; + } + } + if (i == g->nodes[n].adjacency_count) + break; + } + if (r == g->regs->count) + return GL_FALSE; + + g->nodes[n].reg = r; + g->nodes[n].in_stack = GL_FALSE; + g->stack_count--; + } + + return GL_TRUE; +} + +/** + * Optimistic register coloring: Just push the remaining nodes + * on the stack. They'll be colored first in ra_select(), and + * if they succeed then the locally-colorable nodes are still + * locally-colorable and the rest of the register allocation + * will succeed. + */ +void +ra_optimistic_color(struct ra_graph *g) +{ + unsigned int i; + + for (i = 0; i < g->count; i++) { + if (g->nodes[i].in_stack) + continue; + + g->stack[g->stack_count] = i; + g->stack_count++; + g->nodes[i].in_stack = GL_TRUE; + } +} + +GLboolean +ra_allocate_no_spills(struct ra_graph *g) +{ + if (!ra_simplify(g)) { + ra_optimistic_color(g); + } + return ra_select(g); +} + +unsigned int +ra_get_node_reg(struct ra_graph *g, unsigned int n) +{ + return g->nodes[n].reg; +} + +static float +ra_get_spill_benefit(struct ra_graph *g, unsigned int n) +{ + int j; + float benefit = 0; + int n_class = g->nodes[n].class; + + /* Define the benefit of eliminating an interference between n, n2 + * through spilling as q(C, B) / p(C). This is similar to the + * "count number of edges" approach of traditional graph coloring, + * but takes classes into account. + */ + for (j = 0; j < g->nodes[n].adjacency_count; j++) { + unsigned int n2 = g->nodes[n].adjacency_list[j]; + if (n != n2) { + unsigned int n2_class = g->nodes[n2].class; + benefit += ((float)g->regs->classes[n_class]->q[n2_class] / + g->regs->classes[n_class]->p); + } + } + + return benefit; +} + +/** + * Returns a node number to be spilled according to the cost/benefit using + * the pq test, or -1 if there are no spillable nodes. + */ +int +ra_get_best_spill_node(struct ra_graph *g) +{ + unsigned int best_node = -1; + unsigned int best_benefit = 0.0; + unsigned int n; + + for (n = 0; n < g->count; n++) { + float cost = g->nodes[n].spill_cost; + float benefit; + + if (cost <= 0.0) + continue; + + benefit = ra_get_spill_benefit(g, n); + + if (benefit / cost > best_benefit) { + best_benefit = benefit / cost; + best_node = n; + } + } + + return best_node; +} + +/** + * Only nodes with a spill cost set (cost != 0.0) will be considered + * for register spilling. + */ +void +ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost) +{ + g->nodes[n].spill_cost = cost; +} diff --git a/mesalib/src/mesa/program/sampler.cpp b/mesalib/src/mesa/program/sampler.cpp index cd44467a3..12c4a40a2 100644 --- a/mesalib/src/mesa/program/sampler.cpp +++ b/mesalib/src/mesa/program/sampler.cpp @@ -1,140 +1,138 @@ -/* - * Copyright (C) 2005-2007 Brian Paul All Rights Reserved. - * Copyright (C) 2008 VMware, Inc. All Rights Reserved. - * 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. - */ - -#include -#include "ir.h" -#include "glsl_types.h" -#include "ir_visitor.h" - -extern "C" { -#include "main/compiler.h" -#include "main/mtypes.h" -#include "program/prog_parameter.h" -} - -static void fail_link(struct gl_shader_program *prog, const char *fmt, ...) PRINTFLIKE(2, 3); - -static void fail_link(struct gl_shader_program *prog, const char *fmt, ...) -{ - va_list args; - va_start(args, fmt); - prog->InfoLog = talloc_vasprintf_append(prog->InfoLog, fmt, args); - va_end(args); - - prog->LinkStatus = GL_FALSE; -} - -class get_sampler_name : public ir_hierarchical_visitor -{ -public: - get_sampler_name(ir_dereference *last, - struct gl_shader_program *shader_program) - { - this->mem_ctx = talloc_new(NULL); - this->shader_program = shader_program; - this->name = NULL; - this->offset = 0; - this->last = last; - } - - ~get_sampler_name() - { - talloc_free(this->mem_ctx); - } - - virtual ir_visitor_status visit(ir_dereference_variable *ir) - { - this->name = ir->var->name; - return visit_continue; - } - - virtual ir_visitor_status visit_leave(ir_dereference_record *ir) - { - this->name = talloc_asprintf(mem_ctx, "%s.%s", name, ir->field); - return visit_continue; - } - - virtual ir_visitor_status visit_leave(ir_dereference_array *ir) - { - ir_constant *index = ir->array_index->as_constant(); - int i; - - if (index) { - i = index->value.i[0]; - } else { - /* GLSL 1.10 and 1.20 allowed variable sampler array indices, - * while GLSL 1.30 requires that the array indices be - * constant integer expressions. We don't expect any driver - * to actually work with a really variable array index, so - * all that would work would be an unrolled loop counter that ends - * up being constant above. - */ - shader_program->InfoLog = - talloc_asprintf_append(shader_program->InfoLog, - "warning: Variable sampler array index " - "unsupported.\nThis feature of the language " - "was removed in GLSL 1.20 and is unlikely " - "to be supported for 1.10 in Mesa.\n"); - i = 0; - } - if (ir != last) { - this->name = talloc_asprintf(mem_ctx, "%s[%d]", name, i); - } else { - offset = i; - } - return visit_continue; - } - - struct gl_shader_program *shader_program; - const char *name; - void *mem_ctx; - int offset; - ir_dereference *last; -}; - -extern "C" { -int -_mesa_get_sampler_uniform_value(class ir_dereference *sampler, - struct gl_shader_program *shader_program, - const struct gl_program *prog) -{ - get_sampler_name getname(sampler, shader_program); - - sampler->accept(&getname); - - GLint index = _mesa_lookup_parameter_index(prog->Parameters, -1, - getname.name); - - if (index < 0) { - fail_link(shader_program, - "failed to find sampler named %s.\n", getname.name); - return 0; - } - - index += getname.offset; - - return prog->Parameters->ParameterValues[index][0]; -} -} +/* + * Copyright (C) 2005-2007 Brian Paul All Rights Reserved. + * Copyright (C) 2008 VMware, Inc. All Rights Reserved. + * 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. + */ + +#include +#include "ir.h" +#include "glsl_types.h" +#include "ir_visitor.h" + +extern "C" { +#include "main/compiler.h" +#include "main/mtypes.h" +#include "program/prog_parameter.h" +} + +static void fail_link(struct gl_shader_program *prog, const char *fmt, ...) PRINTFLIKE(2, 3); + +static void fail_link(struct gl_shader_program *prog, const char *fmt, ...) +{ + va_list args; + va_start(args, fmt); + ralloc_vasprintf_append(&prog->InfoLog, fmt, args); + va_end(args); + + prog->LinkStatus = GL_FALSE; +} + +class get_sampler_name : public ir_hierarchical_visitor +{ +public: + get_sampler_name(ir_dereference *last, + struct gl_shader_program *shader_program) + { + this->mem_ctx = ralloc_context(NULL); + this->shader_program = shader_program; + this->name = NULL; + this->offset = 0; + this->last = last; + } + + ~get_sampler_name() + { + ralloc_free(this->mem_ctx); + } + + virtual ir_visitor_status visit(ir_dereference_variable *ir) + { + this->name = ir->var->name; + return visit_continue; + } + + virtual ir_visitor_status visit_leave(ir_dereference_record *ir) + { + this->name = ralloc_asprintf(mem_ctx, "%s.%s", name, ir->field); + return visit_continue; + } + + virtual ir_visitor_status visit_leave(ir_dereference_array *ir) + { + ir_constant *index = ir->array_index->as_constant(); + int i; + + if (index) { + i = index->value.i[0]; + } else { + /* GLSL 1.10 and 1.20 allowed variable sampler array indices, + * while GLSL 1.30 requires that the array indices be + * constant integer expressions. We don't expect any driver + * to actually work with a really variable array index, so + * all that would work would be an unrolled loop counter that ends + * up being constant above. + */ + ralloc_strcat(&shader_program->InfoLog, + "warning: Variable sampler array index unsupported.\n" + "This feature of the language was removed in GLSL 1.20 " + "and is unlikely to be supported for 1.10 in Mesa.\n"); + i = 0; + } + if (ir != last) { + this->name = ralloc_asprintf(mem_ctx, "%s[%d]", name, i); + } else { + offset = i; + } + return visit_continue; + } + + struct gl_shader_program *shader_program; + const char *name; + void *mem_ctx; + int offset; + ir_dereference *last; +}; + +extern "C" { +int +_mesa_get_sampler_uniform_value(class ir_dereference *sampler, + struct gl_shader_program *shader_program, + const struct gl_program *prog) +{ + get_sampler_name getname(sampler, shader_program); + + sampler->accept(&getname); + + GLint index = _mesa_lookup_parameter_index(prog->Parameters, -1, + getname.name); + + if (index < 0) { + fail_link(shader_program, + "failed to find sampler named %s.\n", getname.name); + return 0; + } + + index += getname.offset; + + return prog->Parameters->ParameterValues[index][0]; +} +} -- cgit v1.2.3