summaryrefslogtreecommitdiff
path: root/src/glsl/lower_instructions.cpp
blob: d460ba1a97abd2d3fe00f368c2ef5eba4f0e7d55 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
/*
 * 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.
 */

/**
 * \file lower_instructions.cpp
 *
 * Many GPUs lack native instructions for certain expression operations, and
 * must replace them with some other expression tree.  This pass lowers some
 * of the most common cases, allowing the lowering code to be implemented once
 * rather than in each driver backend.
 *
 * Currently supported transformations:
 * - SUB_TO_ADD_NEG
 * - DIV_TO_MUL_RCP
 * - EXP_TO_EXP2
 * - LOG_TO_LOG2
 * - MOD_TO_FRACT
 *
 * SUB_TO_ADD_NEG:
 * ---------------
 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
 *
 * This simplifies expression reassociation, and for many backends
 * there is no subtract operation separate from adding the negation.
 * For backends with native subtract operations, they will probably
 * want to recognize add(op0, neg(op1)) or the other way around to
 * produce a subtract anyway.
 *
 * DIV_TO_MUL_RCP:
 * ---------------
 * Breaks an ir_unop_div expression down to op0 * (rcp(op1)).
 *
 * Many GPUs don't have a divide instruction (945 and 965 included),
 * but they do have an RCP instruction to compute an approximate
 * reciprocal.  By breaking the operation down, constant reciprocals
 * can get constant folded.
 *
 * EXP_TO_EXP2 and LOG_TO_LOG2:
 * ----------------------------
 * Many GPUs don't have a base e log or exponent instruction, but they
 * do have base 2 versions, so this pass converts exp and log to exp2
 * and log2 operations.
 *
 * MOD_TO_FRACT:
 * -------------
 * Breaks an ir_unop_mod expression down to (op1 * fract(op0 / op1))
 *
 * Many GPUs don't have a MOD instruction (945 and 965 included), and
 * if we have to break it down like this anyway, it gives an
 * opportunity to do things like constant fold the (1.0 / op1) easily.
 */

#include "main/core.h" /* for M_E */
#include "glsl_types.h"
#include "ir.h"
#include "ir_optimization.h"

class lower_instructions_visitor : public ir_hierarchical_visitor {
public:
   lower_instructions_visitor(unsigned lower)
      : progress(false), lower(lower) { }

   ir_visitor_status visit_leave(ir_expression *);

   bool progress;

private:
   unsigned lower; /** Bitfield of which operations to lower */

   void sub_to_add_neg(ir_expression *);
   void div_to_mul_rcp(ir_expression *);
   void mod_to_fract(ir_expression *);
   void exp_to_exp2(ir_expression *);
   void log_to_log2(ir_expression *);
};

/**
 * Determine if a particular type of lowering should occur
 */
#define lowering(x) (this->lower & x)

bool
lower_instructions(exec_list *instructions, unsigned what_to_lower)
{
   lower_instructions_visitor v(what_to_lower);

   visit_list_elements(&v, instructions);
   return v.progress;
}

void
lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
{
   ir->operation = ir_binop_add;
   ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type,
					   ir->operands[1], NULL);
   this->progress = true;
}

void
lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
{
   if (!ir->operands[1]->type->is_integer()) {
      /* New expression for the 1.0 / op1 */
      ir_rvalue *expr;
      expr = new(ir) ir_expression(ir_unop_rcp,
				   ir->operands[1]->type,
				   ir->operands[1],
				   NULL);

      /* op0 / op1 -> op0 * (1.0 / op1) */
      ir->operation = ir_binop_mul;
      ir->operands[1] = expr;
   } else {
      /* Be careful with integer division -- we need to do it as a
       * float and re-truncate, since rcp(n > 1) of an integer would
       * just be 0.
       */
      ir_rvalue *op0, *op1;
      const struct glsl_type *vec_type;

      vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
					 ir->operands[1]->type->vector_elements,
					 ir->operands[1]->type->matrix_columns);

      if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
	 op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
      else
	 op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);

      op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);

      vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
					 ir->operands[0]->type->vector_elements,
					 ir->operands[0]->type->matrix_columns);

      if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
	 op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
      else
	 op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);

      op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);

      ir->operation = ir_unop_f2i;
      ir->operands[0] = op0;
      ir->operands[1] = NULL;
   }

   this->progress = true;
}

void
lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
{
   ir_constant *log2_e = new(ir) ir_constant(log2f(M_E));

   ir->operation = ir_unop_exp2;
   ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
					   ir->operands[0], log2_e);
   this->progress = true;
}

void
lower_instructions_visitor::log_to_log2(ir_expression *ir)
{
   ir->operation = ir_binop_mul;
   ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
					   ir->operands[0], NULL);
   ir->operands[1] = new(ir) ir_constant(1.0f / log2f(M_E));
   this->progress = true;
}

void
lower_instructions_visitor::mod_to_fract(ir_expression *ir)
{
   ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b",
					   ir_var_temporary);
   this->base_ir->insert_before(temp);

   ir_assignment *const assign =
      new(ir) ir_assignment(new(ir) ir_dereference_variable(temp),
			    ir->operands[1], NULL);

   this->base_ir->insert_before(assign);

   ir_expression *const div_expr =
      new(ir) ir_expression(ir_binop_div, ir->operands[0]->type,
			    ir->operands[0],
			    new(ir) ir_dereference_variable(temp));

   /* Don't generate new IR that would need to be lowered in an additional
    * pass.
    */
   if (lowering(DIV_TO_MUL_RCP))
      div_to_mul_rcp(div_expr);

   ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract,
					   ir->operands[0]->type,
					   div_expr,
					   NULL);

   ir->operation = ir_binop_mul;
   ir->operands[0] = new(ir) ir_dereference_variable(temp);
   ir->operands[1] = expr;
   this->progress = true;
}

ir_visitor_status
lower_instructions_visitor::visit_leave(ir_expression *ir)
{
   switch (ir->operation) {
   case ir_binop_sub:
      if (lowering(SUB_TO_ADD_NEG))
	 sub_to_add_neg(ir);
      break;

   case ir_binop_div:
      if (lowering(DIV_TO_MUL_RCP))
	 div_to_mul_rcp(ir);
      break;

   case ir_unop_exp:
      if (lowering(EXP_TO_EXP2))
	 exp_to_exp2(ir);
      break;

   case ir_unop_log:
      if (lowering(LOG_TO_LOG2))
	 log_to_log2(ir);
      break;

   case ir_binop_mod:
      if (lowering(MOD_TO_FRACT))
	 mod_to_fract(ir);
      break;

   default:
      return visit_continue;
   }

   return visit_continue;
}