src/mesa/drivers/dri/i965/brw_fs_channel_expressions.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 brw_wm_channel_expressions.cpp
  26  *
  27  * Breaks vector operations down into operations on each component.
  28  *
  29  * The 965 fragment shader receives 8 or 16 pixels at a time, so each
  30  * channel of a vector is laid out as 1 or 2 8-float registers.  Each
  31  * ALU operation operates on one of those channel registers.  As a
  32  * result, there is no value to the 965 fragment shader in tracking
  33  * "vector" expressions in the sense of GLSL fragment shaders, when
  34  * doing a channel at a time may help in constant folding, algebraic
  35  * simplification, and reducing the liveness of channel registers.
  36  *
  37  * The exception to the desire to break everything down to floats is
  38  * texturing.  The texture sampler returns a writemasked masked
  39  * 4/8-register sequence containing the texture values.  We don't want
  40  * to dispatch to the sampler separately for each channel we need, so
  41  * we do retain the vector types in that case.
  42  */
  43
  44 extern "C" {
  45 #include "main/core.h"
  46 #include "brw_wm.h"
  47 }
  48 #include "glsl/ir.h"
  49 #include "glsl/ir_expression_flattening.h"
  50 #include "glsl/glsl_types.h"
  51
  52 class ir_channel_expressions_visitor : public ir_hierarchical_visitor {
  53 public:
  54    ir_channel_expressions_visitor()
  55    {
  56       this->progress = false;
  57       this->mem_ctx = NULL;
  58    }
  59
  60    ir_visitor_status visit_leave(ir_assignment *);
  61
  62    ir_rvalue *get_element(ir_variable *var, unsigned int element);
  63    void assign(ir_assignment *ir, int elem, ir_rvalue *val);
  64
  65    bool progress;
  66    void *mem_ctx;
  67 };
  68
  69 static bool
  70 channel_expressions_predicate(ir_instruction *ir)
  71 {
  72    ir_expression *expr = ir->as_expression();
  73    unsigned int i;
  74
  75    if (!expr)
  76       return false;
  77
  78    switch (expr->operation) {
  79       /* these opcodes need to act on the whole vector,
  80        * just like texturing.
  81        */
  82       case ir_unop_interpolate_at_centroid:
  83       case ir_binop_interpolate_at_offset:
  84       case ir_binop_interpolate_at_sample:
  85          return false;
  86       default:
  87          break;
  88    }
  89
  90    for (i = 0; i < expr->get_num_operands(); i++) {
  91       if (expr->operands[i]->type->is_vector())
  92          return true;
  93    }
  94
  95    return false;
  96 }
  97
  98 bool
  99 brw_do_channel_expressions(exec_list *instructions)
 100 {
 101    ir_channel_expressions_visitor v;
 102
 103    /* Pull out any matrix expression to a separate assignment to a
 104     * temp.  This will make our handling of the breakdown to
 105     * operations on the matrix's vector components much easier.
 106     */
 107    do_expression_flattening(instructions, channel_expressions_predicate);
 108
 109    visit_list_elements(&v, instructions);
 110
 111    return v.progress;
 112 }
 113
 114 ir_rvalue *
 115 ir_channel_expressions_visitor::get_element(ir_variable *var, unsigned int elem)
 116 {
 117    ir_dereference *deref;
 118
 119    if (var->type->is_scalar())
 120       return new(mem_ctx) ir_dereference_variable(var);
 121
 122    assert(elem < var->type->components());
 123    deref = new(mem_ctx) ir_dereference_variable(var);
 124    return new(mem_ctx) ir_swizzle(deref, elem, 0, 0, 0, 1);
 125 }
 126
 127 void
 128 ir_channel_expressions_visitor::assign(ir_assignment *ir, int elem, ir_rvalue *val)
 129 {
 130    ir_dereference *lhs = ir->lhs->clone(mem_ctx, NULL);
 131    ir_assignment *assign;
 132
 133    /* This assign-of-expression should have been generated by the
 134     * expression flattening visitor (since we never short circit to
 135     * not flatten, even for plain assignments of variables), so the
 136     * writemask is always full.
 137     */
 138    assert(ir->write_mask == (1 << ir->lhs->type->components()) - 1);
 139
 140    assign = new(mem_ctx) ir_assignment(lhs, val, NULL, (1 << elem));
 141    ir->insert_before(assign);
 142 }
 143
 144 ir_visitor_status
 145 ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
 146 {
 147    ir_expression *expr = ir->rhs->as_expression();
 148    bool found_vector = false;
 149    unsigned int i, vector_elements = 1;
 150    ir_variable *op_var[3];
 151
 152    if (!expr)
 153       return visit_continue;
 154
 155    if (!this->mem_ctx)
 156       this->mem_ctx = ralloc_parent(ir);
 157
 158    for (i = 0; i < expr->get_num_operands(); i++) {
 159       if (expr->operands[i]->type->is_vector()) {
 160          found_vector = true;
 161          vector_elements = expr->operands[i]->type->vector_elements;
 162          break;
 163       }
 164    }
 165    if (!found_vector)
 166       return visit_continue;
 167
 168    switch (expr->operation) {
 169       case ir_unop_interpolate_at_centroid:
 170       case ir_binop_interpolate_at_offset:
 171       case ir_binop_interpolate_at_sample:
 172          return visit_continue;
 173
 174       default:
 175          break;
 176    }
 177
 178    /* Store the expression operands in temps so we can use them
 179     * multiple times.
 180     */
 181    for (i = 0; i < expr->get_num_operands(); i++) {
 182       ir_assignment *assign;
 183       ir_dereference *deref;
 184
 185       assert(!expr->operands[i]->type->is_matrix());
 186
 187       op_var[i] = new(mem_ctx) ir_variable(expr->operands[i]->type,
 188                                            "channel_expressions",
 189                                            ir_var_temporary);
 190       ir->insert_before(op_var[i]);
 191
 192       deref = new(mem_ctx) ir_dereference_variable(op_var[i]);
 193       assign = new(mem_ctx) ir_assignment(deref,
 194                                           expr->operands[i],
 195                                           NULL);
 196       ir->insert_before(assign);
 197    }
 198
 199    const glsl_type *element_type = glsl_type::get_instance(ir->lhs->type->base_type,
 200                                                            1, 1);
 201
 202    /* OK, time to break down this vector operation. */
 203    switch (expr->operation) {
 204    case ir_unop_bit_not:
 205    case ir_unop_logic_not:
 206    case ir_unop_neg:
 207    case ir_unop_abs:
 208    case ir_unop_sign:
 209    case ir_unop_rcp:
 210    case ir_unop_rsq:
 211    case ir_unop_sqrt:
 212    case ir_unop_exp:
 213    case ir_unop_log:
 214    case ir_unop_exp2:
 215    case ir_unop_log2:
 216    case ir_unop_bitcast_i2f:
 217    case ir_unop_bitcast_f2i:
 218    case ir_unop_bitcast_f2u:
 219    case ir_unop_bitcast_u2f:
 220    case ir_unop_i2u:
 221    case ir_unop_u2i:
 222    case ir_unop_f2i:
 223    case ir_unop_f2u:
 224    case ir_unop_i2f:
 225    case ir_unop_f2b:
 226    case ir_unop_b2f:
 227    case ir_unop_i2b:
 228    case ir_unop_b2i:
 229    case ir_unop_u2f:
 230    case ir_unop_trunc:
 231    case ir_unop_ceil:
 232    case ir_unop_floor:
 233    case ir_unop_fract:
 234    case ir_unop_round_even:
 235    case ir_unop_sin:
 236    case ir_unop_cos:
 237    case ir_unop_sin_reduced:
 238    case ir_unop_cos_reduced:
 239    case ir_unop_dFdx:
 240    case ir_unop_dFdx_coarse:
 241    case ir_unop_dFdx_fine:
 242    case ir_unop_dFdy:
 243    case ir_unop_dFdy_coarse:
 244    case ir_unop_dFdy_fine:
 245    case ir_unop_bitfield_reverse:
 246    case ir_unop_bit_count:
 247    case ir_unop_find_msb:
 248    case ir_unop_find_lsb:
 249       for (i = 0; i < vector_elements; i++) {
 250          ir_rvalue *op0 = get_element(op_var[0], i);
 251
 252          assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
 253                                                   element_type,
 254                                                   op0,
 255                                                   NULL));
 256       }
 257       break;
 258
 259    case ir_binop_add:
 260    case ir_binop_sub:
 261    case ir_binop_mul:
 262    case ir_binop_imul_high:
 263    case ir_binop_div:
 264    case ir_binop_carry:
 265    case ir_binop_borrow:
 266    case ir_binop_mod:
 267    case ir_binop_min:
 268    case ir_binop_max:
 269    case ir_binop_pow:
 270    case ir_binop_lshift:
 271    case ir_binop_rshift:
 272    case ir_binop_bit_and:
 273    case ir_binop_bit_xor:
 274    case ir_binop_bit_or:
 275    case ir_binop_less:
 276    case ir_binop_greater:
 277    case ir_binop_lequal:
 278    case ir_binop_gequal:
 279    case ir_binop_equal:
 280    case ir_binop_nequal:
 281       for (i = 0; i < vector_elements; i++) {
 282          ir_rvalue *op0 = get_element(op_var[0], i);
 283          ir_rvalue *op1 = get_element(op_var[1], i);
 284
 285          assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
 286                                                   element_type,
 287                                                   op0,
 288                                                   op1));
 289       }
 290       break;
 291
 292    case ir_unop_any: {
 293       ir_expression *temp;
 294       temp = new(mem_ctx) ir_expression(ir_binop_logic_or,
 295                                         element_type,
 296                                         get_element(op_var[0], 0),
 297                                         get_element(op_var[0], 1));
 298
 299       for (i = 2; i < vector_elements; i++) {
 300          temp = new(mem_ctx) ir_expression(ir_binop_logic_or,
 301                                            element_type,
 302                                            get_element(op_var[0], i),
 303                                            temp);
 304       }
 305       assign(ir, 0, temp);
 306       break;
 307    }
 308
 309    case ir_binop_dot: {
 310       ir_expression *last = NULL;
 311       for (i = 0; i < vector_elements; i++) {
 312          ir_rvalue *op0 = get_element(op_var[0], i);
 313          ir_rvalue *op1 = get_element(op_var[1], i);
 314          ir_expression *temp;
 315
 316          temp = new(mem_ctx) ir_expression(ir_binop_mul,
 317                                            element_type,
 318                                            op0,
 319                                            op1);
 320          if (last) {
 321             last = new(mem_ctx) ir_expression(ir_binop_add,
 322                                               element_type,
 323                                               temp,
 324                                               last);
 325          } else {
 326             last = temp;
 327          }
 328       }
 329       assign(ir, 0, last);
 330       break;
 331    }
 332
 333    case ir_binop_logic_and:
 334    case ir_binop_logic_xor:
 335    case ir_binop_logic_or:
 336       ir->fprint(stderr);
 337       fprintf(stderr, "\n");
 338       unreachable("not reached: expression operates on scalars only");
 339    case ir_binop_all_equal:
 340    case ir_binop_any_nequal: {
 341       ir_expression *last = NULL;
 342       for (i = 0; i < vector_elements; i++) {
 343          ir_rvalue *op0 = get_element(op_var[0], i);
 344          ir_rvalue *op1 = get_element(op_var[1], i);
 345          ir_expression *temp;
 346          ir_expression_operation join;
 347
 348          if (expr->operation == ir_binop_all_equal)
 349             join = ir_binop_logic_and;
 350          else
 351             join = ir_binop_logic_or;
 352
 353          temp = new(mem_ctx) ir_expression(expr->operation,
 354                                            element_type,
 355                                            op0,
 356                                            op1);
 357          if (last) {
 358             last = new(mem_ctx) ir_expression(join,
 359                                               element_type,
 360                                               temp,
 361                                               last);
 362          } else {
 363             last = temp;
 364          }
 365       }
 366       assign(ir, 0, last);
 367       break;
 368    }
 369    case ir_unop_noise:
 370       unreachable("noise should have been broken down to function call");
 371
 372    case ir_binop_bfm: {
 373       /* Does not need to be scalarized, since its result will be identical
 374        * for all channels.
 375        */
 376       ir_rvalue *op0 = get_element(op_var[0], 0);
 377       ir_rvalue *op1 = get_element(op_var[1], 0);
 378
 379       assign(ir, 0, new(mem_ctx) ir_expression(expr->operation,
 380                                                element_type,
 381                                                op0,
 382                                                op1));
 383       break;
 384    }
 385
 386    case ir_binop_ubo_load:
 387       unreachable("not yet supported");
 388
 389    case ir_triop_fma:
 390    case ir_triop_lrp:
 391    case ir_triop_csel:
 392    case ir_triop_bitfield_extract:
 393       for (i = 0; i < vector_elements; i++) {
 394          ir_rvalue *op0 = get_element(op_var[0], i);
 395          ir_rvalue *op1 = get_element(op_var[1], i);
 396          ir_rvalue *op2 = get_element(op_var[2], i);
 397
 398          assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
 399                                                   element_type,
 400                                                   op0,
 401                                                   op1,
 402                                                   op2));
 403       }
 404       break;
 405
 406    case ir_triop_bfi: {
 407       /* Only a single BFM is needed for multiple BFIs. */
 408       ir_rvalue *op0 = get_element(op_var[0], 0);
 409
 410       for (i = 0; i < vector_elements; i++) {
 411          ir_rvalue *op1 = get_element(op_var[1], i);
 412          ir_rvalue *op2 = get_element(op_var[2], i);
 413
 414          assign(ir, i, new(mem_ctx) ir_expression(expr->operation,
 415                                                   element_type,
 416                                                   op0->clone(mem_ctx, NULL),
 417                                                   op1,
 418                                                   op2));
 419       }
 420       break;
 421    }
 422
 423    case ir_unop_pack_snorm_2x16:
 424    case ir_unop_pack_snorm_4x8:
 425    case ir_unop_pack_unorm_2x16:
 426    case ir_unop_pack_unorm_4x8:
 427    case ir_unop_pack_half_2x16:
 428    case ir_unop_unpack_snorm_2x16:
 429    case ir_unop_unpack_snorm_4x8:
 430    case ir_unop_unpack_unorm_2x16:
 431    case ir_unop_unpack_unorm_4x8:
 432    case ir_unop_unpack_half_2x16:
 433    case ir_binop_ldexp:
 434    case ir_binop_vector_extract:
 435    case ir_triop_vector_insert:
 436    case ir_quadop_bitfield_insert:
 437    case ir_quadop_vector:
 438       unreachable("should have been lowered");
 439
 440    case ir_unop_unpack_half_2x16_split_x:
 441    case ir_unop_unpack_half_2x16_split_y:
 442    case ir_binop_pack_half_2x16_split:
 443    case ir_unop_interpolate_at_centroid:
 444    case ir_binop_interpolate_at_offset:
 445    case ir_binop_interpolate_at_sample:
 446       unreachable("not reached: expression operates on scalars only");
 447    }
 448
 449    ir->remove();
 450    this->progress = true;
 451
 452    return visit_continue;
 453 }