src/intel/compiler/brw_nir_opt_peephole_ffma.c

   1 /*
   2  * Copyright © 2014 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 DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Jason Ekstrand (jason@jlekstrand.net)
  25  *
  26  */
  27
  28 #include "brw_nir.h"
  29 #include "compiler/nir/nir_builder.h"
  30
  31 /*
  32  * Implements a small peephole optimization that looks for a multiply that
  33  * is only ever used in an add and replaces both with an fma.
  34  */
  35
  36 static inline bool
  37 are_all_uses_fadd(nir_ssa_def *def)
  38 {
  39    if (!list_empty(&def->if_uses))
  40       return false;
  41
  42    nir_foreach_use(use_src, def) {
  43       nir_instr *use_instr = use_src->parent_instr;
  44
  45       if (use_instr->type != nir_instr_type_alu)
  46          return false;
  47
  48       nir_alu_instr *use_alu = nir_instr_as_alu(use_instr);
  49       switch (use_alu->op) {
  50       case nir_op_fadd:
  51          break; /* This one's ok */
  52
  53       case nir_op_imov:
  54       case nir_op_fmov:
  55       case nir_op_fneg:
  56       case nir_op_fabs:
  57          assert(use_alu->dest.dest.is_ssa);
  58          if (!are_all_uses_fadd(&use_alu->dest.dest.ssa))
  59             return false;
  60          break;
  61
  62       default:
  63          return false;
  64       }
  65    }
  66
  67    return true;
  68 }
  69
  70 static nir_alu_instr *
  71 get_mul_for_src(nir_alu_src *src, unsigned num_components,
  72                 uint8_t swizzle[4], bool *negate, bool *abs)
  73 {
  74    uint8_t swizzle_tmp[4];
  75    assert(src->src.is_ssa && !src->abs && !src->negate);
  76
  77    nir_instr *instr = src->src.ssa->parent_instr;
  78    if (instr->type != nir_instr_type_alu)
  79       return NULL;
  80
  81    nir_alu_instr *alu = nir_instr_as_alu(instr);
  82
  83    /* We want to bail if any of the other ALU operations involved is labled
  84     * exact.  One reason for this is that, while the value that is changing is
  85     * actually the result of the add and not the multiply, the intention of
  86     * the user when they specify an exact multiply is that they want *that*
  87     * value and what they don't care about is the add.  Another reason is that
  88     * SPIR-V explicitly requires this behaviour.
  89     */
  90    if (alu->exact)
  91       return NULL;
  92
  93    switch (alu->op) {
  94    case nir_op_imov:
  95    case nir_op_fmov:
  96       alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components,
  97                             swizzle, negate, abs);
  98       break;
  99
 100    case nir_op_fneg:
 101       alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components,
 102                             swizzle, negate, abs);
 103       *negate = !*negate;
 104       break;
 105
 106    case nir_op_fabs:
 107       alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components,
 108                             swizzle, negate, abs);
 109       *negate = false;
 110       *abs = true;
 111       break;
 112
 113    case nir_op_fmul:
 114       /* Only absorb a fmul into a ffma if the fmul is only used in fadd
 115        * operations.  This prevents us from being too aggressive with our
 116        * fusing which can actually lead to more instructions.
 117        */
 118       if (!are_all_uses_fadd(&alu->dest.dest.ssa))
 119          return NULL;
 120       break;
 121
 122    default:
 123       return NULL;
 124    }
 125
 126    if (!alu)
 127       return NULL;
 128
 129    /* Copy swizzle data before overwriting it to avoid setting a wrong swizzle.
 130     *
 131     * Example:
 132     *   Former swizzle[] = xyzw
 133     *   src->swizzle[] = zyxx
 134     *
 135     *   Expected output swizzle = zyxx
 136     *   If we reuse swizzle in the loop, then output swizzle would be zyzz.
 137     */
 138    memcpy(swizzle_tmp, swizzle, 4*sizeof(uint8_t));
 139    for (int i = 0; i < num_components; i++)
 140       swizzle[i] = swizzle_tmp[src->swizzle[i]];
 141
 142    return alu;
 143 }
 144
 145 /**
 146  * Given a list of (at least two) nir_alu_src's, tells if any of them is a
 147  * constant value and is used only once.
 148  */
 149 static bool
 150 any_alu_src_is_a_constant(nir_alu_src srcs[])
 151 {
 152    for (unsigned i = 0; i < 2; i++) {
 153       if (srcs[i].src.ssa->parent_instr->type == nir_instr_type_load_const) {
 154          nir_load_const_instr *load_const =
 155             nir_instr_as_load_const (srcs[i].src.ssa->parent_instr);
 156
 157          if (list_is_singular(&load_const->def.uses) &&
 158              list_empty(&load_const->def.if_uses)) {
 159             return true;
 160          }
 161       }
 162    }
 163
 164    return false;
 165 }
 166
 167 static bool
 168 brw_nir_opt_peephole_ffma_block(nir_builder *b, nir_block *block)
 169 {
 170    bool progress = false;
 171
 172    nir_foreach_instr_safe(instr, block) {
 173       if (instr->type != nir_instr_type_alu)
 174          continue;
 175
 176       nir_alu_instr *add = nir_instr_as_alu(instr);
 177       if (add->op != nir_op_fadd)
 178          continue;
 179
 180       assert(add->dest.dest.is_ssa);
 181       if (add->exact)
 182          continue;
 183
 184       assert(add->src[0].src.is_ssa && add->src[1].src.is_ssa);
 185
 186       /* This, is the case a + a.  We would rather handle this with an
 187        * algebraic reduction than fuse it.  Also, we want to only fuse
 188        * things where the multiply is used only once and, in this case,
 189        * it would be used twice by the same instruction.
 190        */
 191       if (add->src[0].src.ssa == add->src[1].src.ssa)
 192          continue;
 193
 194       nir_alu_instr *mul;
 195       uint8_t add_mul_src, swizzle[4];
 196       bool negate, abs;
 197       for (add_mul_src = 0; add_mul_src < 2; add_mul_src++) {
 198          for (unsigned i = 0; i < 4; i++)
 199             swizzle[i] = i;
 200
 201          negate = false;
 202          abs = false;
 203
 204          mul = get_mul_for_src(&add->src[add_mul_src],
 205                                add->dest.dest.ssa.num_components,
 206                                swizzle, &negate, &abs);
 207
 208          if (mul != NULL)
 209             break;
 210       }
 211
 212       if (mul == NULL)
 213          continue;
 214
 215       unsigned bit_size = add->dest.dest.ssa.bit_size;
 216
 217       nir_ssa_def *mul_src[2];
 218       mul_src[0] = mul->src[0].src.ssa;
 219       mul_src[1] = mul->src[1].src.ssa;
 220
 221       /* If any of the operands of the fmul and any of the fadd is a constant,
 222        * we bypass because it will be more efficient as the constants will be
 223        * propagated as operands, potentially saving two load_const instructions.
 224        */
 225       if (any_alu_src_is_a_constant(mul->src) &&
 226           any_alu_src_is_a_constant(add->src)) {
 227          continue;
 228       }
 229
 230       b->cursor = nir_before_instr(&add->instr);
 231
 232       if (abs) {
 233          for (unsigned i = 0; i < 2; i++)
 234             mul_src[i] = nir_fabs(b, mul_src[i]);
 235       }
 236
 237       if (negate)
 238          mul_src[0] = nir_fneg(b, mul_src[0]);
 239
 240       nir_alu_instr *ffma = nir_alu_instr_create(b->shader, nir_op_ffma);
 241       ffma->dest.saturate = add->dest.saturate;
 242       ffma->dest.write_mask = add->dest.write_mask;
 243
 244       for (unsigned i = 0; i < 2; i++) {
 245          ffma->src[i].src = nir_src_for_ssa(mul_src[i]);
 246          for (unsigned j = 0; j < add->dest.dest.ssa.num_components; j++)
 247             ffma->src[i].swizzle[j] = mul->src[i].swizzle[swizzle[j]];
 248       }
 249       nir_alu_src_copy(&ffma->src[2], &add->src[1 - add_mul_src], ffma);
 250
 251       assert(add->dest.dest.is_ssa);
 252
 253       nir_ssa_dest_init(&ffma->instr, &ffma->dest.dest,
 254                         add->dest.dest.ssa.num_components,
 255                         bit_size,
 256                         add->dest.dest.ssa.name);
 257       nir_ssa_def_rewrite_uses(&add->dest.dest.ssa,
 258                                nir_src_for_ssa(&ffma->dest.dest.ssa));
 259
 260       nir_builder_instr_insert(b, &ffma->instr);
 261       assert(list_empty(&add->dest.dest.ssa.uses));
 262       nir_instr_remove(&add->instr);
 263
 264       progress = true;
 265    }
 266
 267    return progress;
 268 }
 269
 270 static bool
 271 brw_nir_opt_peephole_ffma_impl(nir_function_impl *impl)
 272 {
 273    bool progress = false;
 274
 275    nir_builder builder;
 276    nir_builder_init(&builder, impl);
 277
 278    nir_foreach_block(block, impl) {
 279       progress |= brw_nir_opt_peephole_ffma_block(&builder, block);
 280    }
 281
 282    if (progress)
 283       nir_metadata_preserve(impl, nir_metadata_block_index |
 284                                   nir_metadata_dominance);
 285
 286    return progress;
 287 }
 288
 289 bool
 290 brw_nir_opt_peephole_ffma(nir_shader *shader)
 291 {
 292    bool progress = false;
 293
 294    nir_foreach_function(function, shader) {
 295       if (function->impl)
 296          progress |= brw_nir_opt_peephole_ffma_impl(function->impl);
 297    }
 298
 299    return progress;
 300 }