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, int 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], num_components, swizzle, negate, abs);
  97       break;
  98
  99    case nir_op_fneg:
 100       alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs);
 101       *negate = !*negate;
 102       break;
 103
 104    case nir_op_fabs:
 105       alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs);
 106       *negate = false;
 107       *abs = true;
 108       break;
 109
 110    case nir_op_fmul:
 111       /* Only absorb a fmul into a ffma if the fmul is only used in fadd
 112        * operations.  This prevents us from being too aggressive with our
 113        * fusing which can actually lead to more instructions.
 114        */
 115       if (!are_all_uses_fadd(&alu->dest.dest.ssa))
 116          return NULL;
 117       break;
 118
 119    default:
 120       return NULL;
 121    }
 122
 123    if (!alu)
 124       return NULL;
 125
 126    /* Copy swizzle data before overwriting it to avoid setting a wrong swizzle.
 127     *
 128     * Example:
 129     *   Former swizzle[] = xyzw
 130     *   src->swizzle[] = zyxx
 131     *
 132     *   Expected output swizzle = zyxx
 133     *   If we reuse swizzle in the loop, then output swizzle would be zyzz.
 134     */
 135    memcpy(swizzle_tmp, swizzle, 4*sizeof(uint8_t));
 136    for (int i = 0; i < num_components; i++)
 137       swizzle[i] = swizzle_tmp[src->swizzle[i]];
 138
 139    return alu;
 140 }
 141
 142 /**
 143  * Given a list of (at least two) nir_alu_src's, tells if any of them is a
 144  * constant value and is used only once.
 145  */
 146 static bool
 147 any_alu_src_is_a_constant(nir_alu_src srcs[])
 148 {
 149    for (unsigned i = 0; i < 2; i++) {
 150       if (srcs[i].src.ssa->parent_instr->type == nir_instr_type_load_const) {
 151          nir_load_const_instr *load_const =
 152             nir_instr_as_load_const (srcs[i].src.ssa->parent_instr);
 153
 154          if (list_is_singular(&load_const->def.uses) &&
 155              list_empty(&load_const->def.if_uses)) {
 156             return true;
 157          }
 158       }
 159    }
 160
 161    return false;
 162 }
 163
 164 static bool
 165 brw_nir_opt_peephole_ffma_block(nir_builder *b, nir_block *block)
 166 {
 167    bool progress = false;
 168
 169    nir_foreach_instr_safe(instr, block) {
 170       if (instr->type != nir_instr_type_alu)
 171          continue;
 172
 173       nir_alu_instr *add = nir_instr_as_alu(instr);
 174       if (add->op != nir_op_fadd)
 175          continue;
 176
 177       assert(add->dest.dest.is_ssa);
 178       if (add->exact)
 179          continue;
 180
 181       assert(add->src[0].src.is_ssa && add->src[1].src.is_ssa);
 182
 183       /* This, is the case a + a.  We would rather handle this with an
 184        * algebraic reduction than fuse it.  Also, we want to only fuse
 185        * things where the multiply is used only once and, in this case,
 186        * it would be used twice by the same instruction.
 187        */
 188       if (add->src[0].src.ssa == add->src[1].src.ssa)
 189          continue;
 190
 191       nir_alu_instr *mul;
 192       uint8_t add_mul_src, swizzle[4];
 193       bool negate, abs;
 194       for (add_mul_src = 0; add_mul_src < 2; add_mul_src++) {
 195          for (unsigned i = 0; i < 4; i++)
 196             swizzle[i] = i;
 197
 198          negate = false;
 199          abs = false;
 200
 201          mul = get_mul_for_src(&add->src[add_mul_src],
 202                                add->dest.dest.ssa.num_components,
 203                                swizzle, &negate, &abs);
 204
 205          if (mul != NULL)
 206             break;
 207       }
 208
 209       if (mul == NULL)
 210          continue;
 211
 212       unsigned bit_size = add->dest.dest.ssa.bit_size;
 213
 214       nir_ssa_def *mul_src[2];
 215       mul_src[0] = mul->src[0].src.ssa;
 216       mul_src[1] = mul->src[1].src.ssa;
 217
 218       /* If any of the operands of the fmul and any of the fadd is a constant,
 219        * we bypass because it will be more efficient as the constants will be
 220        * propagated as operands, potentially saving two load_const instructions.
 221        */
 222       if (any_alu_src_is_a_constant(mul->src) &&
 223           any_alu_src_is_a_constant(add->src)) {
 224          continue;
 225       }
 226
 227       b->cursor = nir_before_instr(&add->instr);
 228
 229       if (abs) {
 230          for (unsigned i = 0; i < 2; i++)
 231             mul_src[i] = nir_fabs(b, mul_src[i]);
 232       }
 233
 234       if (negate)
 235          mul_src[0] = nir_fneg(b, mul_src[0]);
 236
 237       nir_alu_instr *ffma = nir_alu_instr_create(b->shader, nir_op_ffma);
 238       ffma->dest.saturate = add->dest.saturate;
 239       ffma->dest.write_mask = add->dest.write_mask;
 240
 241       for (unsigned i = 0; i < 2; i++) {
 242          ffma->src[i].src = nir_src_for_ssa(mul_src[i]);
 243          for (unsigned j = 0; j < add->dest.dest.ssa.num_components; j++)
 244             ffma->src[i].swizzle[j] = mul->src[i].swizzle[swizzle[j]];
 245       }
 246       nir_alu_src_copy(&ffma->src[2], &add->src[1 - add_mul_src], ffma);
 247
 248       assert(add->dest.dest.is_ssa);
 249
 250       nir_ssa_dest_init(&ffma->instr, &ffma->dest.dest,
 251                         add->dest.dest.ssa.num_components,
 252                         bit_size,
 253                         add->dest.dest.ssa.name);
 254       nir_ssa_def_rewrite_uses(&add->dest.dest.ssa,
 255                                nir_src_for_ssa(&ffma->dest.dest.ssa));
 256
 257       nir_builder_instr_insert(b, &ffma->instr);
 258       assert(list_empty(&add->dest.dest.ssa.uses));
 259       nir_instr_remove(&add->instr);
 260
 261       progress = true;
 262    }
 263
 264    return progress;
 265 }
 266
 267 static bool
 268 brw_nir_opt_peephole_ffma_impl(nir_function_impl *impl)
 269 {
 270    bool progress = false;
 271
 272    nir_builder builder;
 273    nir_builder_init(&builder, impl);
 274
 275    nir_foreach_block(block, impl) {
 276       progress |= brw_nir_opt_peephole_ffma_block(&builder, block);
 277    }
 278
 279    if (progress)
 280       nir_metadata_preserve(impl, nir_metadata_block_index |
 281                                   nir_metadata_dominance);
 282
 283    return progress;
 284 }
 285
 286 bool
 287 brw_nir_opt_peephole_ffma(nir_shader *shader)
 288 {
 289    bool progress = false;
 290
 291    nir_foreach_function(function, shader) {
 292       if (function->impl)
 293          progress |= brw_nir_opt_peephole_ffma_impl(function->impl);
 294    }
 295
 296    return progress;
 297 }