2 * Copyright © 2010 Intel Corporation
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:
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
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.
25 * \file lower_instructions.cpp
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.
32 * Currently supported transformations:
42 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
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.
52 * Breaks an ir_unop_div expression down to op0 * (rcp(op1)).
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.
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.
67 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
68 * x**y to 2**(y * log2(x)).
72 * Breaks an ir_unop_mod expression down to (op1 * fract(op0 / op1))
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.
79 #include "main/core.h" /* for M_LOG2E */
80 #include "glsl_types.h"
82 #include "ir_optimization.h"
84 class lower_instructions_visitor : public ir_hierarchical_visitor {
86 lower_instructions_visitor(unsigned lower)
87 : progress(false), lower(lower) { }
89 ir_visitor_status visit_leave(ir_expression *);
94 unsigned lower; /** Bitfield of which operations to lower */
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 *);
105 * Determine if a particular type of lowering should occur
107 #define lowering(x) (this->lower & x)
110 lower_instructions(exec_list *instructions, unsigned what_to_lower)
112 lower_instructions_visitor v(what_to_lower);
114 visit_list_elements(&v, instructions);
119 lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
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;
128 lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
130 if (!ir->operands[1]->type->is_integer()) {
131 /* New expression for the 1.0 / op1 */
133 expr = new(ir) ir_expression(ir_unop_rcp,
134 ir->operands[1]->type,
138 /* op0 / op1 -> op0 * (1.0 / op1) */
139 ir->operation = ir_binop_mul;
140 ir->operands[1] = expr;
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
146 ir_rvalue *op0, *op1;
147 const struct glsl_type *vec_type;
149 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
150 ir->operands[1]->type->vector_elements,
151 ir->operands[1]->type->matrix_columns);
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);
156 op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
158 op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
160 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
161 ir->operands[0]->type->vector_elements,
162 ir->operands[0]->type->matrix_columns);
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);
167 op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
169 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
170 ir->type->vector_elements,
171 ir->type->matrix_columns);
173 op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
175 ir->operation = ir_unop_f2i;
176 ir->operands[0] = op0;
177 ir->operands[1] = NULL;
180 this->progress = true;
184 lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
186 ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E));
188 ir->operation = ir_unop_exp2;
189 ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
190 ir->operands[0], log2_e);
191 this->progress = true;
195 lower_instructions_visitor::pow_to_exp2(ir_expression *ir)
197 ir_expression *const log2_x =
198 new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
201 ir->operation = ir_unop_exp2;
202 ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type,
203 ir->operands[1], log2_x);
204 ir->operands[1] = NULL;
205 this->progress = true;
209 lower_instructions_visitor::log_to_log2(ir_expression *ir)
211 ir->operation = ir_binop_mul;
212 ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
213 ir->operands[0], NULL);
214 ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E));
215 this->progress = true;
219 lower_instructions_visitor::mod_to_fract(ir_expression *ir)
221 ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b",
223 this->base_ir->insert_before(temp);
225 ir_assignment *const assign =
226 new(ir) ir_assignment(new(ir) ir_dereference_variable(temp),
227 ir->operands[1], NULL);
229 this->base_ir->insert_before(assign);
231 ir_expression *const div_expr =
232 new(ir) ir_expression(ir_binop_div, ir->operands[0]->type,
234 new(ir) ir_dereference_variable(temp));
236 /* Don't generate new IR that would need to be lowered in an additional
239 if (lowering(DIV_TO_MUL_RCP))
240 div_to_mul_rcp(div_expr);
242 ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract,
243 ir->operands[0]->type,
247 ir->operation = ir_binop_mul;
248 ir->operands[0] = new(ir) ir_dereference_variable(temp);
249 ir->operands[1] = expr;
250 this->progress = true;
254 lower_instructions_visitor::visit_leave(ir_expression *ir)
256 switch (ir->operation) {
258 if (lowering(SUB_TO_ADD_NEG))
263 if (lowering(DIV_TO_MUL_RCP))
268 if (lowering(EXP_TO_EXP2))
273 if (lowering(LOG_TO_LOG2))
278 if (lowering(MOD_TO_FRACT))
283 if (lowering(POW_TO_EXP2))
288 return visit_continue;
291 return visit_continue;