summaryrefslogtreecommitdiffstats
path: root/src/glsl/lower_instructions.cpp
blob: 49316d0020e6a19820d1d04a941d38909a3a9b6b (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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
/*
 * 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
 * - INT_DIV_TO_MUL_RCP
 * - EXP_TO_EXP2
 * - POW_TO_EXP2
 * - LOG_TO_LOG2
 * - MOD_TO_FRACT
 * - LDEXP_TO_ARITH
 * - BITFIELD_INSERT_TO_BFM_BFI
 *
 * 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 and INT_DIV_TO_MUL_RCP:
 * --------------------------------------
 * Breaks an ir_binop_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.
 *
 * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
 * handles the integer case, converting to and from floating point so that
 * RCP is possible.
 *
 * 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.
 *
 * POW_TO_EXP2:
 * -----------
 * Many older GPUs don't have an x**y instruction.  For these GPUs, convert
 * x**y to 2**(y * log2(x)).
 *
 * MOD_TO_FRACT:
 * -------------
 * Breaks an ir_binop_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.
 *
 * LDEXP_TO_ARITH:
 * -------------
 * Converts ir_binop_ldexp to arithmetic and bit operations.
 *
 * BITFIELD_INSERT_TO_BFM_BFI:
 * ---------------------------
 * Breaks ir_quadop_bitfield_insert into ir_binop_bfm (bitfield mask) and
 * ir_triop_bfi (bitfield insert).
 *
 * Many GPUs implement the bitfieldInsert() built-in from ARB_gpu_shader_5
 * with a pair of instructions.
 *
 */

#include "main/core.h" /* for M_LOG2E */
#include "glsl_types.h"
#include "ir.h"
#include "ir_builder.h"
#include "ir_optimization.h"

using namespace ir_builder;

namespace {

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 int_div_to_mul_rcp(ir_expression *);
   void mod_to_fract(ir_expression *);
   void exp_to_exp2(ir_expression *);
   void pow_to_exp2(ir_expression *);
   void log_to_log2(ir_expression *);
   void bitfield_insert_to_bfm_bfi(ir_expression *);
   void ldexp_to_arith(ir_expression *);
};

} /* anonymous namespace */

/**
 * 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)
{
   assert(ir->operands[1]->type->is_float());

   /* 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]);

   /* op0 / op1 -> op0 * (1.0 / op1) */
   ir->operation = ir_binop_mul;
   ir->operands[1] = expr;

   this->progress = true;
}

void
lower_instructions_visitor::int_div_to_mul_rcp(ir_expression *ir)
{
   assert(ir->operands[1]->type->is_integer());

   /* 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);

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

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

   if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) {
      ir->operation = ir_unop_f2i;
      ir->operands[0] = op0;
   } else {
      ir->operation = ir_unop_i2u;
      ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, 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(float(M_LOG2E));

   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::pow_to_exp2(ir_expression *ir)
{
   ir_expression *const log2_x =
      new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
			    ir->operands[0]);

   ir->operation = ir_unop_exp2;
   ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type,
					   ir->operands[1], log2_x);
   ir->operands[1] = NULL;
   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(float(1.0 / M_LOG2E));
   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;
}

void
lower_instructions_visitor::bitfield_insert_to_bfm_bfi(ir_expression *ir)
{
   /* Translates
    *    ir_quadop_bitfield_insert base insert offset bits
    * into
    *    ir_triop_bfi (ir_binop_bfm bits offset) insert base
    */

   ir_rvalue *base_expr = ir->operands[0];

   ir->operation = ir_triop_bfi;
   ir->operands[0] = new(ir) ir_expression(ir_binop_bfm,
                                           ir->type->get_base_type(),
                                           ir->operands[3],
                                           ir->operands[2]);
   /* ir->operands[1] is still the value to insert. */
   ir->operands[2] = base_expr;
   ir->operands[3] = NULL;

   this->progress = true;
}

void
lower_instructions_visitor::ldexp_to_arith(ir_expression *ir)
{
   /* Translates
    *    ir_binop_ldexp x exp
    * into
    *
    *    extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
    *    resulting_biased_exp = extracted_biased_exp + exp;
    *
    *    if (resulting_biased_exp < 1) {
    *       return copysign(0.0, x);
    *    }
    *
    *    return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
    *                       lshift(i2u(resulting_biased_exp), exp_shift));
    *
    * which we can't actually implement as such, since the GLSL IR doesn't
    * have vectorized if-statements. We actually implement it without branches
    * using conditional-select:
    *
    *    extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
    *    resulting_biased_exp = extracted_biased_exp + exp;
    *
    *    is_not_zero_or_underflow = gequal(resulting_biased_exp, 1);
    *    x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
    *    resulting_biased_exp = csel(is_not_zero_or_underflow,
    *                                resulting_biased_exp, 0);
    *
    *    return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
    *                       lshift(i2u(resulting_biased_exp), exp_shift));
    */

   const unsigned vec_elem = ir->type->vector_elements;

   /* Types */
   const glsl_type *ivec = glsl_type::get_instance(GLSL_TYPE_INT, vec_elem, 1);
   const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);

   /* Constants */
   ir_constant *zeroi = ir_constant::zero(ir, ivec);

   ir_constant *sign_mask = new(ir) ir_constant(0x80000000u, vec_elem);

   ir_constant *exp_shift = new(ir) ir_constant(23);
   ir_constant *exp_width = new(ir) ir_constant(8);

   /* Temporary variables */
   ir_variable *x = new(ir) ir_variable(ir->type, "x", ir_var_temporary);
   ir_variable *exp = new(ir) ir_variable(ivec, "exp", ir_var_temporary);

   ir_variable *zero_sign_x = new(ir) ir_variable(ir->type, "zero_sign_x",
                                                  ir_var_temporary);

   ir_variable *extracted_biased_exp =
      new(ir) ir_variable(ivec, "extracted_biased_exp", ir_var_temporary);
   ir_variable *resulting_biased_exp =
      new(ir) ir_variable(ivec, "resulting_biased_exp", ir_var_temporary);

   ir_variable *is_not_zero_or_underflow =
      new(ir) ir_variable(bvec, "is_not_zero_or_underflow", ir_var_temporary);

   ir_instruction &i = *base_ir;

   /* Copy <x> and <exp> arguments. */
   i.insert_before(x);
   i.insert_before(assign(x, ir->operands[0]));
   i.insert_before(exp);
   i.insert_before(assign(exp, ir->operands[1]));

   /* Extract the biased exponent from <x>. */
   i.insert_before(extracted_biased_exp);
   i.insert_before(assign(extracted_biased_exp,
                          rshift(bitcast_f2i(abs(x)), exp_shift)));

   i.insert_before(resulting_biased_exp);
   i.insert_before(assign(resulting_biased_exp,
                          add(extracted_biased_exp, exp)));

   /* Test if result is ±0.0, subnormal, or underflow by checking if the
    * resulting biased exponent would be less than 0x1. If so, the result is
    * 0.0 with the sign of x. (Actually, invert the conditions so that
    * immediate values are the second arguments, which is better for i965)
    */
   i.insert_before(zero_sign_x);
   i.insert_before(assign(zero_sign_x,
                          bitcast_u2f(bit_and(bitcast_f2u(x), sign_mask))));

   i.insert_before(is_not_zero_or_underflow);
   i.insert_before(assign(is_not_zero_or_underflow,
                          gequal(resulting_biased_exp,
                                  new(ir) ir_constant(0x1, vec_elem))));
   i.insert_before(assign(x, csel(is_not_zero_or_underflow,
                                  x, zero_sign_x)));
   i.insert_before(assign(resulting_biased_exp,
                          csel(is_not_zero_or_underflow,
                               resulting_biased_exp, zeroi)));

   /* We could test for overflows by checking if the resulting biased exponent
    * would be greater than 0xFE. Turns out we don't need to because the GLSL
    * spec says:
    *
    *    "If this product is too large to be represented in the
    *     floating-point type, the result is undefined."
    */

   ir_constant *exp_shift_clone = exp_shift->clone(ir, NULL);
   ir->operation = ir_unop_bitcast_i2f;
   ir->operands[0] = bitfield_insert(bitcast_f2i(x), resulting_biased_exp,
                                     exp_shift_clone, exp_width);
   ir->operands[1] = NULL;

   /* Don't generate new IR that would need to be lowered in an additional
    * pass.
    */
   if (lowering(BITFIELD_INSERT_TO_BFM_BFI))
      bitfield_insert_to_bfm_bfi(ir->operands[0]->as_expression());

   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 (ir->operands[1]->type->is_integer() && lowering(INT_DIV_TO_MUL_RCP))
	 int_div_to_mul_rcp(ir);
      else if (ir->operands[1]->type->is_float() && 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) && ir->type->is_float())
	 mod_to_fract(ir);
      break;

   case ir_binop_pow:
      if (lowering(POW_TO_EXP2))
	 pow_to_exp2(ir);
      break;

   case ir_quadop_bitfield_insert:
      if (lowering(BITFIELD_INSERT_TO_BFM_BFI))
         bitfield_insert_to_bfm_bfi(ir);
      break;

   case ir_binop_ldexp:
      if (lowering(LDEXP_TO_ARITH))
         ldexp_to_arith(ir);
      break;

   default:
      return visit_continue;
   }

   return visit_continue;
}