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:
35 * - INT_DIV_TO_MUL_RCP
41 * - BITFIELD_INSERT_TO_BFM_BFI
45 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
47 * This simplifies expression reassociation, and for many backends
48 * there is no subtract operation separate from adding the negation.
49 * For backends with native subtract operations, they will probably
50 * want to recognize add(op0, neg(op1)) or the other way around to
51 * produce a subtract anyway.
53 * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP:
54 * --------------------------------------
55 * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
57 * Many GPUs don't have a divide instruction (945 and 965 included),
58 * but they do have an RCP instruction to compute an approximate
59 * reciprocal. By breaking the operation down, constant reciprocals
60 * can get constant folded.
62 * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
63 * handles the integer case, converting to and from floating point so that
66 * EXP_TO_EXP2 and LOG_TO_LOG2:
67 * ----------------------------
68 * Many GPUs don't have a base e log or exponent instruction, but they
69 * do have base 2 versions, so this pass converts exp and log to exp2
70 * and log2 operations.
74 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
75 * x**y to 2**(y * log2(x)).
79 * Breaks an ir_binop_mod expression down to (op1 * fract(op0 / op1))
81 * Many GPUs don't have a MOD instruction (945 and 965 included), and
82 * if we have to break it down like this anyway, it gives an
83 * opportunity to do things like constant fold the (1.0 / op1) easily.
87 * Converts ir_triop_lrp to (op0 * (1.0f - op2)) + (op1 * op2).
89 * BITFIELD_INSERT_TO_BFM_BFI:
90 * ---------------------------
91 * Breaks ir_quadop_bitfield_insert into ir_binop_bfm (bitfield mask) and
92 * ir_triop_bfi (bitfield insert).
94 * Many GPUs implement the bitfieldInsert() built-in from ARB_gpu_shader_5
95 * with a pair of instructions.
99 #include "main/core.h" /* for M_LOG2E */
100 #include "glsl_types.h"
102 #include "ir_builder.h"
103 #include "ir_optimization.h"
105 using namespace ir_builder
;
107 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
109 lower_instructions_visitor(unsigned lower
)
110 : progress(false), lower(lower
) { }
112 ir_visitor_status
visit_leave(ir_expression
*);
117 unsigned lower
; /** Bitfield of which operations to lower */
119 void sub_to_add_neg(ir_expression
*);
120 void div_to_mul_rcp(ir_expression
*);
121 void int_div_to_mul_rcp(ir_expression
*);
122 void mod_to_fract(ir_expression
*);
123 void exp_to_exp2(ir_expression
*);
124 void pow_to_exp2(ir_expression
*);
125 void log_to_log2(ir_expression
*);
126 void lrp_to_arith(ir_expression
*);
127 void bitfield_insert_to_bfm_bfi(ir_expression
*);
131 * Determine if a particular type of lowering should occur
133 #define lowering(x) (this->lower & x)
136 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
138 lower_instructions_visitor
v(what_to_lower
);
140 visit_list_elements(&v
, instructions
);
145 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
147 ir
->operation
= ir_binop_add
;
148 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
149 ir
->operands
[1], NULL
);
150 this->progress
= true;
154 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
156 assert(ir
->operands
[1]->type
->is_float());
158 /* New expression for the 1.0 / op1 */
160 expr
= new(ir
) ir_expression(ir_unop_rcp
,
161 ir
->operands
[1]->type
,
164 /* op0 / op1 -> op0 * (1.0 / op1) */
165 ir
->operation
= ir_binop_mul
;
166 ir
->operands
[1] = expr
;
168 this->progress
= true;
172 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
174 assert(ir
->operands
[1]->type
->is_integer());
176 /* Be careful with integer division -- we need to do it as a
177 * float and re-truncate, since rcp(n > 1) of an integer would
180 ir_rvalue
*op0
, *op1
;
181 const struct glsl_type
*vec_type
;
183 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
184 ir
->operands
[1]->type
->vector_elements
,
185 ir
->operands
[1]->type
->matrix_columns
);
187 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
188 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
190 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
192 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
194 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
195 ir
->operands
[0]->type
->vector_elements
,
196 ir
->operands
[0]->type
->matrix_columns
);
198 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
199 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
201 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
203 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
204 ir
->type
->vector_elements
,
205 ir
->type
->matrix_columns
);
207 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
209 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
210 ir
->operation
= ir_unop_f2i
;
211 ir
->operands
[0] = op0
;
213 ir
->operation
= ir_unop_i2u
;
214 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
216 ir
->operands
[1] = NULL
;
218 this->progress
= true;
222 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
224 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
226 ir
->operation
= ir_unop_exp2
;
227 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
228 ir
->operands
[0], log2_e
);
229 this->progress
= true;
233 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
235 ir_expression
*const log2_x
=
236 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
239 ir
->operation
= ir_unop_exp2
;
240 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
241 ir
->operands
[1], log2_x
);
242 ir
->operands
[1] = NULL
;
243 this->progress
= true;
247 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
249 ir
->operation
= ir_binop_mul
;
250 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
251 ir
->operands
[0], NULL
);
252 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
253 this->progress
= true;
257 lower_instructions_visitor::mod_to_fract(ir_expression
*ir
)
259 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_b",
261 this->base_ir
->insert_before(temp
);
263 ir_assignment
*const assign
=
264 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(temp
),
265 ir
->operands
[1], NULL
);
267 this->base_ir
->insert_before(assign
);
269 ir_expression
*const div_expr
=
270 new(ir
) ir_expression(ir_binop_div
, ir
->operands
[0]->type
,
272 new(ir
) ir_dereference_variable(temp
));
274 /* Don't generate new IR that would need to be lowered in an additional
277 if (lowering(DIV_TO_MUL_RCP
))
278 div_to_mul_rcp(div_expr
);
280 ir_rvalue
*expr
= new(ir
) ir_expression(ir_unop_fract
,
281 ir
->operands
[0]->type
,
285 ir
->operation
= ir_binop_mul
;
286 ir
->operands
[0] = new(ir
) ir_dereference_variable(temp
);
287 ir
->operands
[1] = expr
;
288 this->progress
= true;
292 lower_instructions_visitor::lrp_to_arith(ir_expression
*ir
)
294 /* (lrp x y a) -> x*(1-a) + y*a */
297 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[2]->type
, "lrp_factor",
299 this->base_ir
->insert_before(temp
);
300 this->base_ir
->insert_before(assign(temp
, ir
->operands
[2]));
302 ir_constant
*one
= new(ir
) ir_constant(1.0f
);
304 ir
->operation
= ir_binop_add
;
305 ir
->operands
[0] = mul(ir
->operands
[0], sub(one
, temp
));
306 ir
->operands
[1] = mul(ir
->operands
[1], temp
);
307 ir
->operands
[2] = NULL
;
309 this->progress
= true;
313 lower_instructions_visitor::bitfield_insert_to_bfm_bfi(ir_expression
*ir
)
316 * ir_quadop_bitfield_insert base insert offset bits
318 * ir_triop_bfi (ir_binop_bfm bits offset) insert base
321 ir_rvalue
*base_expr
= ir
->operands
[0];
323 ir
->operation
= ir_triop_bfi
;
324 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_bfm
,
325 ir
->type
->get_base_type(),
328 /* ir->operands[1] is still the value to insert. */
329 ir
->operands
[2] = base_expr
;
330 ir
->operands
[3] = NULL
;
332 this->progress
= true;
336 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
338 switch (ir
->operation
) {
340 if (lowering(SUB_TO_ADD_NEG
))
345 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
346 int_div_to_mul_rcp(ir
);
347 else if (ir
->operands
[1]->type
->is_float() && lowering(DIV_TO_MUL_RCP
))
352 if (lowering(EXP_TO_EXP2
))
357 if (lowering(LOG_TO_LOG2
))
362 if (lowering(MOD_TO_FRACT
) && ir
->type
->is_float())
367 if (lowering(POW_TO_EXP2
))
372 if (lowering(LRP_TO_ARITH
))
376 case ir_quadop_bitfield_insert
:
377 if (lowering(BITFIELD_INSERT_TO_BFM_BFI
))
378 bitfield_insert_to_bfm_bfi(ir
);
382 return visit_continue
;
385 return visit_continue
;