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 * - INT_DIV_TO_MUL_RCP
36 * - EXP_TO_EXP2
37 * - POW_TO_EXP2
38 * - LOG_TO_LOG2
39 * - MOD_TO_FRACT
40 *
41 * SUB_TO_ADD_NEG:
42 * ---------------
43 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
44 *
45 * This simplifies expression reassociation, and for many backends
46 * there is no subtract operation separate from adding the negation.
47 * For backends with native subtract operations, they will probably
48 * want to recognize add(op0, neg(op1)) or the other way around to
49 * produce a subtract anyway.
50 *
51 * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP:
52 * --------------------------------------
53 * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
54 *
55 * Many GPUs don't have a divide instruction (945 and 965 included),
56 * but they do have an RCP instruction to compute an approximate
57 * reciprocal. By breaking the operation down, constant reciprocals
58 * can get constant folded.
59 *
60 * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
61 * handles the integer case, converting to and from floating point so that
62 * RCP is possible.
63 *
64 * EXP_TO_EXP2 and LOG_TO_LOG2:
65 * ----------------------------
66 * Many GPUs don't have a base e log or exponent instruction, but they
67 * do have base 2 versions, so this pass converts exp and log to exp2
68 * and log2 operations.
69 *
70 * POW_TO_EXP2:
71 * -----------
72 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
73 * x**y to 2**(y * log2(x)).
74 *
75 * MOD_TO_FRACT:
76 * -------------
77 * Breaks an ir_binop_mod expression down to (op1 * fract(op0 / op1))
78 *
79 * Many GPUs don't have a MOD instruction (945 and 965 included), and
80 * if we have to break it down like this anyway, it gives an
81 * opportunity to do things like constant fold the (1.0 / op1) easily.
82 */
83
84 #include "main/core.h" /* for M_LOG2E */
85 #include "glsl_types.h"
86 #include "ir.h"
87 #include "ir_optimization.h"
88
89 class lower_instructions_visitor : public ir_hierarchical_visitor {
90 public:
lower_instructions_visitor(unsigned lower)91 lower_instructions_visitor(unsigned lower)
92 : progress(false), lower(lower) { }
93
94 ir_visitor_status visit_leave(ir_expression *);
95
96 bool progress;
97
98 private:
99 unsigned lower; /** Bitfield of which operations to lower */
100
101 void sub_to_add_neg(ir_expression *);
102 void div_to_mul_rcp(ir_expression *);
103 void int_div_to_mul_rcp(ir_expression *);
104 void mod_to_fract(ir_expression *);
105 void exp_to_exp2(ir_expression *);
106 void pow_to_exp2(ir_expression *);
107 void log_to_log2(ir_expression *);
108 };
109
110 /**
111 * Determine if a particular type of lowering should occur
112 */
113 #define lowering(x) (this->lower & x)
114
115 bool
lower_instructions(exec_list * instructions,unsigned what_to_lower)116 lower_instructions(exec_list *instructions, unsigned what_to_lower)
117 {
118 lower_instructions_visitor v(what_to_lower);
119
120 visit_list_elements(&v, instructions);
121 return v.progress;
122 }
123
124 void
sub_to_add_neg(ir_expression * ir)125 lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
126 {
127 ir->operation = ir_binop_add;
128 ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type,
129 ir->operands[1], NULL);
130 this->progress = true;
131 }
132
133 void
div_to_mul_rcp(ir_expression * ir)134 lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
135 {
136 assert(ir->operands[1]->type->is_float());
137
138 /* New expression for the 1.0 / op1 */
139 ir_rvalue *expr;
140 expr = new(ir) ir_expression(ir_unop_rcp,
141 ir->operands[1]->type,
142 ir->operands[1]);
143
144 /* op0 / op1 -> op0 * (1.0 / op1) */
145 ir->operation = ir_binop_mul;
146 ir->operands[1] = expr;
147
148 this->progress = true;
149 }
150
151 void
int_div_to_mul_rcp(ir_expression * ir)152 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression *ir)
153 {
154 assert(ir->operands[1]->type->is_integer());
155
156 /* Be careful with integer division -- we need to do it as a
157 * float and re-truncate, since rcp(n > 1) of an integer would
158 * just be 0.
159 */
160 ir_rvalue *op0, *op1;
161 const struct glsl_type *vec_type;
162
163 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
164 ir->operands[1]->type->vector_elements,
165 ir->operands[1]->type->matrix_columns);
166
167 if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
168 op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
169 else
170 op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
171
172 op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
173
174 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
175 ir->operands[0]->type->vector_elements,
176 ir->operands[0]->type->matrix_columns);
177
178 if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
179 op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
180 else
181 op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
182
183 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
184 ir->type->vector_elements,
185 ir->type->matrix_columns);
186
187 op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
188
189 if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) {
190 ir->operation = ir_unop_f2i;
191 ir->operands[0] = op0;
192 } else {
193 ir->operation = ir_unop_i2u;
194 ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, op0);
195 }
196 ir->operands[1] = NULL;
197
198 this->progress = true;
199 }
200
201 void
exp_to_exp2(ir_expression * ir)202 lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
203 {
204 ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E));
205
206 ir->operation = ir_unop_exp2;
207 ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
208 ir->operands[0], log2_e);
209 this->progress = true;
210 }
211
212 void
pow_to_exp2(ir_expression * ir)213 lower_instructions_visitor::pow_to_exp2(ir_expression *ir)
214 {
215 ir_expression *const log2_x =
216 new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
217 ir->operands[0]);
218
219 ir->operation = ir_unop_exp2;
220 ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type,
221 ir->operands[1], log2_x);
222 ir->operands[1] = NULL;
223 this->progress = true;
224 }
225
226 void
log_to_log2(ir_expression * ir)227 lower_instructions_visitor::log_to_log2(ir_expression *ir)
228 {
229 ir->operation = ir_binop_mul;
230 ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
231 ir->operands[0], NULL);
232 ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E));
233 this->progress = true;
234 }
235
236 void
mod_to_fract(ir_expression * ir)237 lower_instructions_visitor::mod_to_fract(ir_expression *ir)
238 {
239 ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b",
240 ir_var_temporary);
241 this->base_ir->insert_before(temp);
242
243 ir_assignment *const assign =
244 new(ir) ir_assignment(new(ir) ir_dereference_variable(temp),
245 ir->operands[1], NULL);
246
247 this->base_ir->insert_before(assign);
248
249 ir_expression *const div_expr =
250 new(ir) ir_expression(ir_binop_div, ir->operands[0]->type,
251 ir->operands[0],
252 new(ir) ir_dereference_variable(temp));
253
254 /* Don't generate new IR that would need to be lowered in an additional
255 * pass.
256 */
257 if (lowering(DIV_TO_MUL_RCP))
258 div_to_mul_rcp(div_expr);
259
260 ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract,
261 ir->operands[0]->type,
262 div_expr,
263 NULL);
264
265 ir->operation = ir_binop_mul;
266 ir->operands[0] = new(ir) ir_dereference_variable(temp);
267 ir->operands[1] = expr;
268 this->progress = true;
269 }
270
271 ir_visitor_status
visit_leave(ir_expression * ir)272 lower_instructions_visitor::visit_leave(ir_expression *ir)
273 {
274 switch (ir->operation) {
275 case ir_binop_sub:
276 if (lowering(SUB_TO_ADD_NEG))
277 sub_to_add_neg(ir);
278 break;
279
280 case ir_binop_div:
281 if (ir->operands[1]->type->is_integer() && lowering(INT_DIV_TO_MUL_RCP))
282 int_div_to_mul_rcp(ir);
283 else if (ir->operands[1]->type->is_float() && lowering(DIV_TO_MUL_RCP))
284 div_to_mul_rcp(ir);
285 break;
286
287 case ir_unop_exp:
288 if (lowering(EXP_TO_EXP2))
289 exp_to_exp2(ir);
290 break;
291
292 case ir_unop_log:
293 if (lowering(LOG_TO_LOG2))
294 log_to_log2(ir);
295 break;
296
297 case ir_binop_mod:
298 if (lowering(MOD_TO_FRACT) && ir->type->is_float())
299 mod_to_fract(ir);
300 break;
301
302 case ir_binop_pow:
303 if (lowering(POW_TO_EXP2))
304 pow_to_exp2(ir);
305 break;
306
307 default:
308 return visit_continue;
309 }
310
311 return visit_continue;
312 }
313