lower_instructions.cpp revision ed92b912120394f3b19958effaa819d29bc6d059
1/* 2 * Copyright © 2010 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 21 * DEALINGS IN THE SOFTWARE. 22 */ 23 24/** 25 * \file lower_instructions.cpp 26 * 27 * Many GPUs lack native instructions for certain expression operations, and 28 * must replace them with some other expression tree. This pass lowers some 29 * of the most common cases, allowing the lowering code to be implemented once 30 * rather than in each driver backend. 31 * 32 * Currently supported transformations: 33 * - SUB_TO_ADD_NEG 34 * - DIV_TO_MUL_RCP 35 * - EXP_TO_EXP2 36 * - POW_TO_EXP2 37 * - LOG_TO_LOG2 38 * - MOD_TO_FRACT 39 * 40 * SUB_TO_ADD_NEG: 41 * --------------- 42 * Breaks an ir_binop_sub expression down to add(op0, neg(op1)) 43 * 44 * This simplifies expression reassociation, and for many backends 45 * there is no subtract operation separate from adding the negation. 46 * For backends with native subtract operations, they will probably 47 * want to recognize add(op0, neg(op1)) or the other way around to 48 * produce a subtract anyway. 49 * 50 * DIV_TO_MUL_RCP: 51 * --------------- 52 * Breaks an ir_unop_div expression down to op0 * (rcp(op1)). 53 * 54 * Many GPUs don't have a divide instruction (945 and 965 included), 55 * but they do have an RCP instruction to compute an approximate 56 * reciprocal. By breaking the operation down, constant reciprocals 57 * can get constant folded. 58 * 59 * EXP_TO_EXP2 and LOG_TO_LOG2: 60 * ---------------------------- 61 * Many GPUs don't have a base e log or exponent instruction, but they 62 * do have base 2 versions, so this pass converts exp and log to exp2 63 * and log2 operations. 64 * 65 * POW_TO_EXP2: 66 * ----------- 67 * Many older GPUs don't have an x**y instruction. For these GPUs, convert 68 * x**y to 2**(y * log2(x)). 69 * 70 * MOD_TO_FRACT: 71 * ------------- 72 * Breaks an ir_unop_mod expression down to (op1 * fract(op0 / op1)) 73 * 74 * Many GPUs don't have a MOD instruction (945 and 965 included), and 75 * if we have to break it down like this anyway, it gives an 76 * opportunity to do things like constant fold the (1.0 / op1) easily. 77 */ 78 79#include "main/core.h" /* for M_LOG2E */ 80#include "glsl_types.h" 81#include "ir.h" 82#include "ir_optimization.h" 83 84class lower_instructions_visitor : public ir_hierarchical_visitor { 85public: 86 lower_instructions_visitor(unsigned lower) 87 : progress(false), lower(lower) { } 88 89 ir_visitor_status visit_leave(ir_expression *); 90 91 bool progress; 92 93private: 94 unsigned lower; /** Bitfield of which operations to lower */ 95 96 void sub_to_add_neg(ir_expression *); 97 void div_to_mul_rcp(ir_expression *); 98 void mod_to_fract(ir_expression *); 99 void exp_to_exp2(ir_expression *); 100 void pow_to_exp2(ir_expression *); 101 void log_to_log2(ir_expression *); 102}; 103 104/** 105 * Determine if a particular type of lowering should occur 106 */ 107#define lowering(x) (this->lower & x) 108 109bool 110lower_instructions(exec_list *instructions, unsigned what_to_lower) 111{ 112 lower_instructions_visitor v(what_to_lower); 113 114 visit_list_elements(&v, instructions); 115 return v.progress; 116} 117 118void 119lower_instructions_visitor::sub_to_add_neg(ir_expression *ir) 120{ 121 ir->operation = ir_binop_add; 122 ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type, 123 ir->operands[1], NULL); 124 this->progress = true; 125} 126 127void 128lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir) 129{ 130 if (!ir->operands[1]->type->is_integer()) { 131 /* New expression for the 1.0 / op1 */ 132 ir_rvalue *expr; 133 expr = new(ir) ir_expression(ir_unop_rcp, 134 ir->operands[1]->type, 135 ir->operands[1], 136 NULL); 137 138 /* op0 / op1 -> op0 * (1.0 / op1) */ 139 ir->operation = ir_binop_mul; 140 ir->operands[1] = expr; 141 } else { 142 /* Be careful with integer division -- we need to do it as a 143 * float and re-truncate, since rcp(n > 1) of an integer would 144 * just be 0. 145 */ 146 ir_rvalue *op0, *op1; 147 const struct glsl_type *vec_type; 148 149 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT, 150 ir->operands[1]->type->vector_elements, 151 ir->operands[1]->type->matrix_columns); 152 153 if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) 154 op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL); 155 else 156 op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL); 157 158 op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL); 159 160 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT, 161 ir->operands[0]->type->vector_elements, 162 ir->operands[0]->type->matrix_columns); 163 164 if (ir->operands[0]->type->base_type == GLSL_TYPE_INT) 165 op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL); 166 else 167 op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL); 168 169 op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1); 170 171 if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) { 172 ir->operation = ir_unop_f2i; 173 ir->operands[0] = op0; 174 } else { 175 ir->operation = ir_unop_i2u; 176 ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, op0); 177 } 178 ir->operands[1] = NULL; 179 } 180 181 this->progress = true; 182} 183 184void 185lower_instructions_visitor::exp_to_exp2(ir_expression *ir) 186{ 187 ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E)); 188 189 ir->operation = ir_unop_exp2; 190 ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type, 191 ir->operands[0], log2_e); 192 this->progress = true; 193} 194 195void 196lower_instructions_visitor::pow_to_exp2(ir_expression *ir) 197{ 198 ir_expression *const log2_x = 199 new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type, 200 ir->operands[0]); 201 202 ir->operation = ir_unop_exp2; 203 ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type, 204 ir->operands[1], log2_x); 205 ir->operands[1] = NULL; 206 this->progress = true; 207} 208 209void 210lower_instructions_visitor::log_to_log2(ir_expression *ir) 211{ 212 ir->operation = ir_binop_mul; 213 ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type, 214 ir->operands[0], NULL); 215 ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E)); 216 this->progress = true; 217} 218 219void 220lower_instructions_visitor::mod_to_fract(ir_expression *ir) 221{ 222 ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b", 223 ir_var_temporary); 224 this->base_ir->insert_before(temp); 225 226 ir_assignment *const assign = 227 new(ir) ir_assignment(new(ir) ir_dereference_variable(temp), 228 ir->operands[1], NULL); 229 230 this->base_ir->insert_before(assign); 231 232 ir_expression *const div_expr = 233 new(ir) ir_expression(ir_binop_div, ir->operands[0]->type, 234 ir->operands[0], 235 new(ir) ir_dereference_variable(temp)); 236 237 /* Don't generate new IR that would need to be lowered in an additional 238 * pass. 239 */ 240 if (lowering(DIV_TO_MUL_RCP)) 241 div_to_mul_rcp(div_expr); 242 243 ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract, 244 ir->operands[0]->type, 245 div_expr, 246 NULL); 247 248 ir->operation = ir_binop_mul; 249 ir->operands[0] = new(ir) ir_dereference_variable(temp); 250 ir->operands[1] = expr; 251 this->progress = true; 252} 253 254ir_visitor_status 255lower_instructions_visitor::visit_leave(ir_expression *ir) 256{ 257 switch (ir->operation) { 258 case ir_binop_sub: 259 if (lowering(SUB_TO_ADD_NEG)) 260 sub_to_add_neg(ir); 261 break; 262 263 case ir_binop_div: 264 if (lowering(DIV_TO_MUL_RCP)) 265 div_to_mul_rcp(ir); 266 break; 267 268 case ir_unop_exp: 269 if (lowering(EXP_TO_EXP2)) 270 exp_to_exp2(ir); 271 break; 272 273 case ir_unop_log: 274 if (lowering(LOG_TO_LOG2)) 275 log_to_log2(ir); 276 break; 277 278 case ir_binop_mod: 279 if (lowering(MOD_TO_FRACT)) 280 mod_to_fract(ir); 281 break; 282 283 case ir_binop_pow: 284 if (lowering(POW_TO_EXP2)) 285 pow_to_exp2(ir); 286 break; 287 288 default: 289 return visit_continue; 290 } 291 292 return visit_continue; 293} 294