src/glsl/lower_instructions.cpp

   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  * - LDEXP_TO_ARITH
  41  * - BITFIELD_INSERT_TO_BFM_BFI
  42  *
  43  * SUB_TO_ADD_NEG:
  44  * ---------------
  45  * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
  46  *
  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.
  52  *
  53  * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP:
  54  * --------------------------------------
  55  * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
  56  *
  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.
  61  *
  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
  64  * RCP is possible.
  65  *
  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.
  71  *
  72  * POW_TO_EXP2:
  73  * -----------
  74  * Many older GPUs don't have an x**y instruction.  For these GPUs, convert
  75  * x**y to 2**(y * log2(x)).
  76  *
  77  * MOD_TO_FRACT:
  78  * -------------
  79  * Breaks an ir_binop_mod expression down to (op1 * fract(op0 / op1))
  80  *
  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.
  84  *
  85  * LDEXP_TO_ARITH:
  86  * -------------
  87  * Converts ir_binop_ldexp to arithmetic and bit operations.
  88  *
  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).
  93  *
  94  * Many GPUs implement the bitfieldInsert() built-in from ARB_gpu_shader_5
  95  * with a pair of instructions.
  96  *
  97  */
  98
  99 #include "main/core.h" /* for M_LOG2E */
 100 #include "glsl_types.h"
 101 #include "ir.h"
 102 #include "ir_builder.h"
 103 #include "ir_optimization.h"
 104
 105 using namespace ir_builder;
 106
 107 namespace {
 108
 109 class lower_instructions_visitor : public ir_hierarchical_visitor {
 110 public:
 111    lower_instructions_visitor(unsigned lower)
 112       : progress(false), lower(lower) { }
 113
 114    ir_visitor_status visit_leave(ir_expression *);
 115
 116    bool progress;
 117
 118 private:
 119    unsigned lower; /** Bitfield of which operations to lower */
 120
 121    void sub_to_add_neg(ir_expression *);
 122    void div_to_mul_rcp(ir_expression *);
 123    void int_div_to_mul_rcp(ir_expression *);
 124    void mod_to_fract(ir_expression *);
 125    void exp_to_exp2(ir_expression *);
 126    void pow_to_exp2(ir_expression *);
 127    void log_to_log2(ir_expression *);
 128    void bitfield_insert_to_bfm_bfi(ir_expression *);
 129    void ldexp_to_arith(ir_expression *);
 130 };
 131
 132 } /* anonymous namespace */
 133
 134 /**
 135  * Determine if a particular type of lowering should occur
 136  */
 137 #define lowering(x) (this->lower & x)
 138
 139 bool
 140 lower_instructions(exec_list *instructions, unsigned what_to_lower)
 141 {
 142    lower_instructions_visitor v(what_to_lower);
 143
 144    visit_list_elements(&v, instructions);
 145    return v.progress;
 146 }
 147
 148 void
 149 lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
 150 {
 151    ir->operation = ir_binop_add;
 152    ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type,
 153                                            ir->operands[1], NULL);
 154    this->progress = true;
 155 }
 156
 157 void
 158 lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
 159 {
 160    assert(ir->operands[1]->type->is_float());
 161
 162    /* New expression for the 1.0 / op1 */
 163    ir_rvalue *expr;
 164    expr = new(ir) ir_expression(ir_unop_rcp,
 165                                 ir->operands[1]->type,
 166                                 ir->operands[1]);
 167
 168    /* op0 / op1 -> op0 * (1.0 / op1) */
 169    ir->operation = ir_binop_mul;
 170    ir->operands[1] = expr;
 171
 172    this->progress = true;
 173 }
 174
 175 void
 176 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression *ir)
 177 {
 178    assert(ir->operands[1]->type->is_integer());
 179
 180    /* Be careful with integer division -- we need to do it as a
 181     * float and re-truncate, since rcp(n > 1) of an integer would
 182     * just be 0.
 183     */
 184    ir_rvalue *op0, *op1;
 185    const struct glsl_type *vec_type;
 186
 187    vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
 188                                       ir->operands[1]->type->vector_elements,
 189                                       ir->operands[1]->type->matrix_columns);
 190
 191    if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
 192       op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
 193    else
 194       op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
 195
 196    op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
 197
 198    vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
 199                                       ir->operands[0]->type->vector_elements,
 200                                       ir->operands[0]->type->matrix_columns);
 201
 202    if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
 203       op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
 204    else
 205       op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
 206
 207    vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
 208                                       ir->type->vector_elements,
 209                                       ir->type->matrix_columns);
 210
 211    op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
 212
 213    if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) {
 214       ir->operation = ir_unop_f2i;
 215       ir->operands[0] = op0;
 216    } else {
 217       ir->operation = ir_unop_i2u;
 218       ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, op0);
 219    }
 220    ir->operands[1] = NULL;
 221
 222    this->progress = true;
 223 }
 224
 225 void
 226 lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
 227 {
 228    ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E));
 229
 230    ir->operation = ir_unop_exp2;
 231    ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
 232                                            ir->operands[0], log2_e);
 233    this->progress = true;
 234 }
 235
 236 void
 237 lower_instructions_visitor::pow_to_exp2(ir_expression *ir)
 238 {
 239    ir_expression *const log2_x =
 240       new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
 241                             ir->operands[0]);
 242
 243    ir->operation = ir_unop_exp2;
 244    ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type,
 245                                            ir->operands[1], log2_x);
 246    ir->operands[1] = NULL;
 247    this->progress = true;
 248 }
 249
 250 void
 251 lower_instructions_visitor::log_to_log2(ir_expression *ir)
 252 {
 253    ir->operation = ir_binop_mul;
 254    ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
 255                                            ir->operands[0], NULL);
 256    ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E));
 257    this->progress = true;
 258 }
 259
 260 void
 261 lower_instructions_visitor::mod_to_fract(ir_expression *ir)
 262 {
 263    ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b",
 264                                            ir_var_temporary);
 265    this->base_ir->insert_before(temp);
 266
 267    ir_assignment *const assign =
 268       new(ir) ir_assignment(new(ir) ir_dereference_variable(temp),
 269                             ir->operands[1], NULL);
 270
 271    this->base_ir->insert_before(assign);
 272
 273    ir_expression *const div_expr =
 274       new(ir) ir_expression(ir_binop_div, ir->operands[0]->type,
 275                             ir->operands[0],
 276                             new(ir) ir_dereference_variable(temp));
 277
 278    /* Don't generate new IR that would need to be lowered in an additional
 279     * pass.
 280     */
 281    if (lowering(DIV_TO_MUL_RCP))
 282       div_to_mul_rcp(div_expr);
 283
 284    ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract,
 285                                            ir->operands[0]->type,
 286                                            div_expr,
 287                                            NULL);
 288
 289    ir->operation = ir_binop_mul;
 290    ir->operands[0] = new(ir) ir_dereference_variable(temp);
 291    ir->operands[1] = expr;
 292    this->progress = true;
 293 }
 294
 295 void
 296 lower_instructions_visitor::bitfield_insert_to_bfm_bfi(ir_expression *ir)
 297 {
 298    /* Translates
 299     *    ir_quadop_bitfield_insert base insert offset bits
 300     * into
 301     *    ir_triop_bfi (ir_binop_bfm bits offset) insert base
 302     */
 303
 304    ir_rvalue *base_expr = ir->operands[0];
 305
 306    ir->operation = ir_triop_bfi;
 307    ir->operands[0] = new(ir) ir_expression(ir_binop_bfm,
 308                                            ir->type->get_base_type(),
 309                                            ir->operands[3],
 310                                            ir->operands[2]);
 311    /* ir->operands[1] is still the value to insert. */
 312    ir->operands[2] = base_expr;
 313    ir->operands[3] = NULL;
 314
 315    this->progress = true;
 316 }
 317
 318 void
 319 lower_instructions_visitor::ldexp_to_arith(ir_expression *ir)
 320 {
 321    /* Translates
 322     *    ir_binop_ldexp x exp
 323     * into
 324     *
 325     *    extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
 326     *    resulting_biased_exp = extracted_biased_exp + exp;
 327     *
 328     *    if (resulting_biased_exp < 1) {
 329     *       return copysign(0.0, x);
 330     *    }
 331     *
 332     *    return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
 333     *                       lshift(i2u(resulting_biased_exp), exp_shift));
 334     *
 335     * which we can't actually implement as such, since the GLSL IR doesn't
 336     * have vectorized if-statements. We actually implement it without branches
 337     * using conditional-select:
 338     *
 339     *    extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
 340     *    resulting_biased_exp = extracted_biased_exp + exp;
 341     *
 342     *    is_not_zero_or_underflow = gequal(resulting_biased_exp, 1);
 343     *    x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
 344     *    resulting_biased_exp = csel(is_not_zero_or_underflow,
 345     *                                resulting_biased_exp, 0);
 346     *
 347     *    return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
 348     *                       lshift(i2u(resulting_biased_exp), exp_shift));
 349     */
 350
 351    const unsigned vec_elem = ir->type->vector_elements;
 352
 353    /* Types */
 354    const glsl_type *ivec = glsl_type::get_instance(GLSL_TYPE_INT, vec_elem, 1);
 355    const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);
 356
 357    /* Constants */
 358    ir_constant *zeroi = ir_constant::zero(ir, ivec);
 359
 360    ir_constant *sign_mask = new(ir) ir_constant(0x80000000u, vec_elem);
 361
 362    ir_constant *exp_shift = new(ir) ir_constant(23u, vec_elem);
 363    ir_constant *exp_width = new(ir) ir_constant(8u, vec_elem);
 364
 365    /* Temporary variables */
 366    ir_variable *x = new(ir) ir_variable(ir->type, "x", ir_var_temporary);
 367    ir_variable *exp = new(ir) ir_variable(ivec, "exp", ir_var_temporary);
 368
 369    ir_variable *zero_sign_x = new(ir) ir_variable(ir->type, "zero_sign_x",
 370                                                   ir_var_temporary);
 371
 372    ir_variable *extracted_biased_exp =
 373       new(ir) ir_variable(ivec, "extracted_biased_exp", ir_var_temporary);
 374    ir_variable *resulting_biased_exp =
 375       new(ir) ir_variable(ivec, "resulting_biased_exp", ir_var_temporary);
 376
 377    ir_variable *is_not_zero_or_underflow =
 378       new(ir) ir_variable(bvec, "is_not_zero_or_underflow", ir_var_temporary);
 379
 380    ir_instruction &i = *base_ir;
 381
 382    /* Copy <x> and <exp> arguments. */
 383    i.insert_before(x);
 384    i.insert_before(assign(x, ir->operands[0]));
 385    i.insert_before(exp);
 386    i.insert_before(assign(exp, ir->operands[1]));
 387
 388    /* Extract the biased exponent from <x>. */
 389    i.insert_before(extracted_biased_exp);
 390    i.insert_before(assign(extracted_biased_exp,
 391                           rshift(bitcast_f2i(abs(x)), exp_shift)));
 392
 393    i.insert_before(resulting_biased_exp);
 394    i.insert_before(assign(resulting_biased_exp,
 395                           add(extracted_biased_exp, exp)));
 396
 397    /* Test if result is ±0.0, subnormal, or underflow by checking if the
 398     * resulting biased exponent would be less than 0x1. If so, the result is
 399     * 0.0 with the sign of x. (Actually, invert the conditions so that
 400     * immediate values are the second arguments, which is better for i965)
 401     */
 402    i.insert_before(zero_sign_x);
 403    i.insert_before(assign(zero_sign_x,
 404                           bitcast_u2f(bit_and(bitcast_f2u(x), sign_mask))));
 405
 406    i.insert_before(is_not_zero_or_underflow);
 407    i.insert_before(assign(is_not_zero_or_underflow,
 408                           gequal(resulting_biased_exp,
 409                                   new(ir) ir_constant(0x1, vec_elem))));
 410    i.insert_before(assign(x, csel(is_not_zero_or_underflow,
 411                                   x, zero_sign_x)));
 412    i.insert_before(assign(resulting_biased_exp,
 413                           csel(is_not_zero_or_underflow,
 414                                resulting_biased_exp, zeroi)));
 415
 416    /* We could test for overflows by checking if the resulting biased exponent
 417     * would be greater than 0xFE. Turns out we don't need to because the GLSL
 418     * spec says:
 419     *
 420     *    "If this product is too large to be represented in the
 421     *     floating-point type, the result is undefined."
 422     */
 423
 424    ir_constant *exp_shift_clone = exp_shift->clone(ir, NULL);
 425    ir->operation = ir_unop_bitcast_i2f;
 426    ir->operands[0] = bitfield_insert(bitcast_f2i(x), resulting_biased_exp,
 427                                      exp_shift_clone, exp_width);
 428    ir->operands[1] = NULL;
 429
 430    /* Don't generate new IR that would need to be lowered in an additional
 431     * pass.
 432     */
 433    if (lowering(BITFIELD_INSERT_TO_BFM_BFI))
 434       bitfield_insert_to_bfm_bfi(ir->operands[0]->as_expression());
 435
 436    this->progress = true;
 437 }
 438
 439 ir_visitor_status
 440 lower_instructions_visitor::visit_leave(ir_expression *ir)
 441 {
 442    switch (ir->operation) {
 443    case ir_binop_sub:
 444       if (lowering(SUB_TO_ADD_NEG))
 445          sub_to_add_neg(ir);
 446       break;
 447
 448    case ir_binop_div:
 449       if (ir->operands[1]->type->is_integer() && lowering(INT_DIV_TO_MUL_RCP))
 450          int_div_to_mul_rcp(ir);
 451       else if (ir->operands[1]->type->is_float() && lowering(DIV_TO_MUL_RCP))
 452          div_to_mul_rcp(ir);
 453       break;
 454
 455    case ir_unop_exp:
 456       if (lowering(EXP_TO_EXP2))
 457          exp_to_exp2(ir);
 458       break;
 459
 460    case ir_unop_log:
 461       if (lowering(LOG_TO_LOG2))
 462          log_to_log2(ir);
 463       break;
 464
 465    case ir_binop_mod:
 466       if (lowering(MOD_TO_FRACT) && ir->type->is_float())
 467          mod_to_fract(ir);
 468       break;
 469
 470    case ir_binop_pow:
 471       if (lowering(POW_TO_EXP2))
 472          pow_to_exp2(ir);
 473       break;
 474
 475    case ir_quadop_bitfield_insert:
 476       if (lowering(BITFIELD_INSERT_TO_BFM_BFI))
 477          bitfield_insert_to_bfm_bfi(ir);
 478       break;
 479
 480    case ir_binop_ldexp:
 481       if (lowering(LDEXP_TO_ARITH))
 482          ldexp_to_arith(ir);
 483       break;
 484
 485    default:
 486       return visit_continue;
 487    }
 488
 489    return visit_continue;
 490 }