xref: /aosp_15_r20/external/mesa3d/src/compiler/glsl/opt_algebraic.cpp (revision 6104692788411f58d303aa86923a9ff6ecaded22)
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 opt_algebraic.cpp
26  *
27  * Takes advantage of association, commutivity, and other algebraic
28  * properties to simplify expressions.
29  */
30 
31 #include "ir.h"
32 #include "ir_visitor.h"
33 #include "ir_rvalue_visitor.h"
34 #include "ir_optimization.h"
35 #include "ir_builder.h"
36 #include "compiler/glsl_types.h"
37 #include "main/consts_exts.h"
38 
39 using namespace ir_builder;
40 
41 namespace {
42 
43 /**
44  * Visitor class for replacing expressions with ir_constant values.
45  */
46 
47 class ir_algebraic_visitor : public ir_rvalue_visitor {
48 public:
ir_algebraic_visitor(bool native_integers,const struct gl_shader_compiler_options * options)49    ir_algebraic_visitor(bool native_integers,
50                         const struct gl_shader_compiler_options *options)
51       : options(options)
52    {
53       this->progress = false;
54       this->mem_ctx = NULL;
55       this->native_integers = native_integers;
56    }
57 
~ir_algebraic_visitor()58    virtual ~ir_algebraic_visitor()
59    {
60    }
61 
62    virtual ir_visitor_status visit_enter(ir_assignment *ir);
63 
64    ir_rvalue *handle_expression(ir_expression *ir);
65    void handle_rvalue(ir_rvalue **rvalue);
66    bool reassociate_constant(ir_expression *ir1,
67 			     int const_index,
68 			     ir_constant *constant,
69 			     ir_expression *ir2);
70    void reassociate_operands(ir_expression *ir1,
71 			     int op1,
72 			     ir_expression *ir2,
73 			     int op2);
74    ir_rvalue *swizzle_if_required(ir_expression *expr,
75 				  ir_rvalue *operand);
76 
77    const struct gl_shader_compiler_options *options;
78    void *mem_ctx;
79 
80    bool native_integers;
81    bool progress;
82 };
83 
84 } /* unnamed namespace */
85 
86 ir_visitor_status
visit_enter(ir_assignment * ir)87 ir_algebraic_visitor::visit_enter(ir_assignment *ir)
88 {
89    ir_variable *var = ir->lhs->variable_referenced();
90    if (var->data.invariant || var->data.precise) {
91       /* If we're assigning to an invariant or precise variable, just bail.
92        * Most of the algebraic optimizations aren't precision-safe.
93        *
94        * FINISHME: Find out which optimizations are precision-safe and enable
95        * then only for invariant or precise trees.
96        */
97       return visit_continue_with_parent;
98    } else {
99       return visit_continue;
100    }
101 }
102 
103 static inline bool
is_valid_vec_const(ir_constant * ir)104 is_valid_vec_const(ir_constant *ir)
105 {
106    if (ir == NULL)
107       return false;
108 
109    if (!glsl_type_is_scalar(ir->type) && !glsl_type_is_vector(ir->type))
110       return false;
111 
112    return true;
113 }
114 
115 static inline bool
is_less_than_one(ir_constant * ir)116 is_less_than_one(ir_constant *ir)
117 {
118    assert(glsl_type_is_float(ir->type));
119 
120    if (!is_valid_vec_const(ir))
121       return false;
122 
123    unsigned component = 0;
124    for (int c = 0; c < ir->type->vector_elements; c++) {
125       if (ir->get_float_component(c) < 1.0f)
126          component++;
127    }
128 
129    return (component == ir->type->vector_elements);
130 }
131 
132 static inline bool
is_greater_than_zero(ir_constant * ir)133 is_greater_than_zero(ir_constant *ir)
134 {
135    assert(glsl_type_is_float(ir->type));
136 
137    if (!is_valid_vec_const(ir))
138       return false;
139 
140    unsigned component = 0;
141    for (int c = 0; c < ir->type->vector_elements; c++) {
142       if (ir->get_float_component(c) > 0.0f)
143          component++;
144    }
145 
146    return (component == ir->type->vector_elements);
147 }
148 
149 static void
update_type(ir_expression * ir)150 update_type(ir_expression *ir)
151 {
152    if (glsl_type_is_vector(ir->operands[0]->type))
153       ir->type = ir->operands[0]->type;
154    else
155       ir->type = ir->operands[1]->type;
156 }
157 
158 void
reassociate_operands(ir_expression * ir1,int op1,ir_expression * ir2,int op2)159 ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
160 					   int op1,
161 					   ir_expression *ir2,
162 					   int op2)
163 {
164    ir_rvalue *temp = ir2->operands[op2];
165    ir2->operands[op2] = ir1->operands[op1];
166    ir1->operands[op1] = temp;
167 
168    /* Update the type of ir2.  The type of ir1 won't have changed --
169     * base types matched, and at least one of the operands of the 2
170     * binops is still a vector if any of them were.
171     */
172    update_type(ir2);
173 
174    this->progress = true;
175 }
176 
177 /**
178  * Reassociates a constant down a tree of adds or multiplies.
179  *
180  * Consider (2 * (a * (b * 0.5))).  We want to end up with a * b.
181  */
182 bool
reassociate_constant(ir_expression * ir1,int const_index,ir_constant * constant,ir_expression * ir2)183 ir_algebraic_visitor::reassociate_constant(ir_expression *ir1, int const_index,
184 					   ir_constant *constant,
185 					   ir_expression *ir2)
186 {
187    if (!ir2 || ir1->operation != ir2->operation)
188       return false;
189 
190    /* Don't want to even think about matrices. */
191    if (glsl_type_is_matrix(ir1->operands[0]->type) ||
192        glsl_type_is_matrix(ir1->operands[1]->type) ||
193        glsl_type_is_matrix(ir2->operands[0]->type) ||
194        glsl_type_is_matrix(ir2->operands[1]->type))
195       return false;
196 
197    void *mem_ctx = ralloc_parent(ir2);
198 
199    ir_constant *ir2_const[2];
200    ir2_const[0] = ir2->operands[0]->constant_expression_value(mem_ctx);
201    ir2_const[1] = ir2->operands[1]->constant_expression_value(mem_ctx);
202 
203    if (ir2_const[0] && ir2_const[1])
204       return false;
205 
206    if (ir2_const[0]) {
207       reassociate_operands(ir1, const_index, ir2, 1);
208       return true;
209    } else if (ir2_const[1]) {
210       reassociate_operands(ir1, const_index, ir2, 0);
211       return true;
212    }
213 
214    if (reassociate_constant(ir1, const_index, constant,
215 			    ir2->operands[0]->as_expression())) {
216       update_type(ir2);
217       return true;
218    }
219 
220    if (reassociate_constant(ir1, const_index, constant,
221 			    ir2->operands[1]->as_expression())) {
222       update_type(ir2);
223       return true;
224    }
225 
226    return false;
227 }
228 
229 /* When eliminating an expression and just returning one of its operands,
230  * we may need to swizzle that operand out to a vector if the expression was
231  * vector type.
232  */
233 ir_rvalue *
swizzle_if_required(ir_expression * expr,ir_rvalue * operand)234 ir_algebraic_visitor::swizzle_if_required(ir_expression *expr,
235 					  ir_rvalue *operand)
236 {
237    if (glsl_type_is_vector(expr->type) && glsl_type_is_scalar(operand->type)) {
238       return new(mem_ctx) ir_swizzle(operand, 0, 0, 0, 0,
239 				     expr->type->vector_elements);
240    } else
241       return operand;
242 }
243 
244 ir_rvalue *
handle_expression(ir_expression * ir)245 ir_algebraic_visitor::handle_expression(ir_expression *ir)
246 {
247    ir_constant *op_const[4] = {NULL, NULL, NULL, NULL};
248    ir_expression *op_expr[4] = {NULL, NULL, NULL, NULL};
249 
250    if (ir->operation == ir_binop_mul &&
251        glsl_type_is_matrix(ir->operands[0]->type) &&
252        glsl_type_is_vector(ir->operands[1]->type)) {
253       ir_expression *matrix_mul = ir->operands[0]->as_expression();
254 
255       if (matrix_mul && matrix_mul->operation == ir_binop_mul &&
256          glsl_type_is_matrix(matrix_mul->operands[0]->type) &&
257          glsl_type_is_matrix(matrix_mul->operands[1]->type)) {
258 
259          return mul(matrix_mul->operands[0],
260                     mul(matrix_mul->operands[1], ir->operands[1]));
261       }
262    }
263 
264    assert(ir->num_operands <= 4);
265    for (unsigned i = 0; i < ir->num_operands; i++) {
266       if (glsl_type_is_matrix(ir->operands[i]->type))
267 	 return ir;
268 
269       op_const[i] =
270          ir->operands[i]->constant_expression_value(ralloc_parent(ir));
271       op_expr[i] = ir->operands[i]->as_expression();
272    }
273 
274    if (this->mem_ctx == NULL)
275       this->mem_ctx = ralloc_parent(ir);
276 
277    switch (ir->operation) {
278    case ir_binop_add:
279       /* Reassociate addition of constants so that we can do constant
280        * folding.
281        */
282       if (op_const[0] && !op_const[1])
283 	 reassociate_constant(ir, 0, op_const[0], op_expr[1]);
284       if (op_const[1] && !op_const[0])
285 	 reassociate_constant(ir, 1, op_const[1], op_expr[0]);
286 
287       break;
288 
289    case ir_binop_mul:
290       /* Reassociate multiplication of constants so that we can do
291        * constant folding.
292        */
293       if (op_const[0] && !op_const[1])
294 	 reassociate_constant(ir, 0, op_const[0], op_expr[1]);
295       if (op_const[1] && !op_const[0])
296 	 reassociate_constant(ir, 1, op_const[1], op_expr[0]);
297       break;
298 
299    case ir_binop_min:
300    case ir_binop_max:
301       if (!glsl_type_is_float(ir->type))
302          break;
303 
304       /* Replace min(max) operations and its commutative combinations with
305        * a saturate operation
306        */
307       for (int op = 0; op < 2; op++) {
308          ir_expression *inner_expr = op_expr[op];
309          ir_constant *outer_const = op_const[1 - op];
310          ir_expression_operation op_cond = (ir->operation == ir_binop_max) ?
311             ir_binop_min : ir_binop_max;
312 
313          if (!inner_expr || !outer_const || (inner_expr->operation != op_cond))
314             continue;
315 
316          /* One of these has to be a constant */
317          if (!inner_expr->operands[0]->as_constant() &&
318              !inner_expr->operands[1]->as_constant())
319             break;
320 
321          /* Found a min(max) combination. Now try to see if its operands
322           * meet our conditions that we can do just a single saturate operation
323           */
324          for (int minmax_op = 0; minmax_op < 2; minmax_op++) {
325             ir_rvalue *x = inner_expr->operands[minmax_op];
326             ir_rvalue *y = inner_expr->operands[1 - minmax_op];
327 
328             ir_constant *inner_const = y->as_constant();
329             if (!inner_const)
330                continue;
331 
332             /* min(max(x, 0.0), 1.0) is sat(x) */
333             if (ir->operation == ir_binop_min &&
334                 inner_const->is_zero() &&
335                 outer_const->is_one())
336                return saturate(x);
337 
338             /* max(min(x, 1.0), 0.0) is sat(x) */
339             if (ir->operation == ir_binop_max &&
340                 inner_const->is_one() &&
341                 outer_const->is_zero())
342                return saturate(x);
343 
344             /* min(max(x, 0.0), b) where b < 1.0 is sat(min(x, b)) */
345             if (ir->operation == ir_binop_min &&
346                 inner_const->is_zero() &&
347                 is_less_than_one(outer_const))
348                return saturate(expr(ir_binop_min, x, outer_const));
349 
350             /* max(min(x, b), 0.0) where b < 1.0 is sat(min(x, b)) */
351             if (ir->operation == ir_binop_max &&
352                 is_less_than_one(inner_const) &&
353                 outer_const->is_zero())
354                return saturate(expr(ir_binop_min, x, inner_const));
355 
356             /* max(min(x, 1.0), b) where b > 0.0 is sat(max(x, b)) */
357             if (ir->operation == ir_binop_max &&
358                 inner_const->is_one() &&
359                 is_greater_than_zero(outer_const))
360                return saturate(expr(ir_binop_max, x, outer_const));
361 
362             /* min(max(x, b), 1.0) where b > 0.0 is sat(max(x, b)) */
363             if (ir->operation == ir_binop_min &&
364                 is_greater_than_zero(inner_const) &&
365                 outer_const->is_one())
366                return saturate(expr(ir_binop_max, x, inner_const));
367          }
368       }
369 
370       break;
371 
372    /* Remove interpolateAt* instructions for demoted inputs. They are
373     * assigned a constant expression to facilitate this.
374     */
375    case ir_unop_interpolate_at_centroid:
376    case ir_binop_interpolate_at_offset:
377    case ir_binop_interpolate_at_sample:
378       if (op_const[0])
379          return ir->operands[0];
380       break;
381 
382    default:
383       break;
384    }
385 
386    return ir;
387 }
388 
389 void
handle_rvalue(ir_rvalue ** rvalue)390 ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
391 {
392    if (!*rvalue)
393       return;
394 
395    ir_expression *expr = (*rvalue)->as_expression();
396    if (!expr || expr->operation == ir_quadop_vector)
397       return;
398 
399    ir_rvalue *new_rvalue = handle_expression(expr);
400    if (new_rvalue == *rvalue)
401       return;
402 
403    /* If the expr used to be some vec OP scalar returning a vector, and the
404     * optimization gave us back a scalar, we still need to turn it into a
405     * vector.
406     */
407    *rvalue = swizzle_if_required(expr, new_rvalue);
408 
409    this->progress = true;
410 }
411 
412 bool
do_algebraic(exec_list * instructions,bool native_integers,const struct gl_shader_compiler_options * options)413 do_algebraic(exec_list *instructions, bool native_integers,
414              const struct gl_shader_compiler_options *options)
415 {
416    ir_algebraic_visitor v(native_integers, options);
417 
418    visit_list_elements(&v, instructions);
419 
420    return v.progress;
421 }
422