src/glsl/nir/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 "nir.h"
  29
  30 /*
  31  * Implements a small peephole optimization that looks for a multiply that
  32  * is only ever used in an add and replaces both with an fma.
  33  */
  34
  35 struct peephole_ffma_state {
  36    void *mem_ctx;
  37    nir_function_impl *impl;
  38    bool progress;
  39 };
  40
  41 static inline bool
  42 are_all_uses_fadd(nir_ssa_def *def)
  43 {
  44    if (!list_empty(&def->if_uses))
  45       return false;
  46
  47    nir_foreach_use(def, use_src) {
  48       nir_instr *use_instr = use_src->parent_instr;
  49
  50       if (use_instr->type != nir_instr_type_alu)
  51          return false;
  52
  53       nir_alu_instr *use_alu = nir_instr_as_alu(use_instr);
  54       switch (use_alu->op) {
  55       case nir_op_fadd:
  56          break; /* This one's ok */
  57
  58       case nir_op_imov:
  59       case nir_op_fmov:
  60       case nir_op_fneg:
  61       case nir_op_fabs:
  62          assert(use_alu->dest.dest.is_ssa);
  63          if (!are_all_uses_fadd(&use_alu->dest.dest.ssa))
  64             return false;
  65          break;
  66
  67       default:
  68          return false;
  69       }
  70    }
  71
  72    return true;
  73 }
  74
  75 static nir_alu_instr *
  76 get_mul_for_src(nir_alu_src *src, int num_components,
  77                 uint8_t swizzle[4], bool *negate, bool *abs)
  78 {
  79    uint8_t swizzle_tmp[4];
  80    assert(src->src.is_ssa && !src->abs && !src->negate);
  81
  82    nir_instr *instr = src->src.ssa->parent_instr;
  83    if (instr->type != nir_instr_type_alu)
  84       return NULL;
  85
  86    nir_alu_instr *alu = nir_instr_as_alu(instr);
  87    switch (alu->op) {
  88    case nir_op_imov:
  89    case nir_op_fmov:
  90       alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs);
  91       break;
  92
  93    case nir_op_fneg:
  94       alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs);
  95       *negate = !*negate;
  96       break;
  97
  98    case nir_op_fabs:
  99       alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs);
 100       *negate = false;
 101       *abs = true;
 102       break;
 103
 104    case nir_op_fmul:
 105       /* Only absorb a fmul into a ffma if the fmul is is only used in fadd
 106        * operations.  This prevents us from being too aggressive with our
 107        * fusing which can actually lead to more instructions.
 108        */
 109       if (!are_all_uses_fadd(&alu->dest.dest.ssa))
 110          return NULL;
 111       break;
 112
 113    default:
 114       return NULL;
 115    }
 116
 117    if (!alu)
 118       return NULL;
 119
 120    /* Copy swizzle data before overwriting it to avoid setting a wrong swizzle.
 121     *
 122     * Example:
 123     *   Former swizzle[] = xyzw
 124     *   src->swizzle[] = zyxx
 125     *
 126     *   Expected output swizzle = zyxx
 127     *   If we reuse swizzle in the loop, then output swizzle would be zyzz.
 128     */
 129    memcpy(swizzle_tmp, swizzle, 4*sizeof(uint8_t));
 130    for (int i = 0; i < num_components; i++)
 131       swizzle[i] = swizzle_tmp[src->swizzle[i]];
 132
 133    return alu;
 134 }
 135
 136 static bool
 137 nir_opt_peephole_ffma_block(nir_block *block, void *void_state)
 138 {
 139    struct peephole_ffma_state *state = void_state;
 140
 141    nir_foreach_instr_safe(block, instr) {
 142       if (instr->type != nir_instr_type_alu)
 143          continue;
 144
 145       nir_alu_instr *add = nir_instr_as_alu(instr);
 146       if (add->op != nir_op_fadd)
 147          continue;
 148
 149       /* TODO: Maybe bail if this expression is considered "precise"? */
 150
 151       assert(add->src[0].src.is_ssa && add->src[1].src.is_ssa);
 152
 153       /* This, is the case a + a.  We would rather handle this with an
 154        * algebraic reduction than fuse it.  Also, we want to only fuse
 155        * things where the multiply is used only once and, in this case,
 156        * it would be used twice by the same instruction.
 157        */
 158       if (add->src[0].src.ssa == add->src[1].src.ssa)
 159          continue;
 160
 161       nir_alu_instr *mul;
 162       uint8_t add_mul_src, swizzle[4];
 163       bool negate, abs;
 164       for (add_mul_src = 0; add_mul_src < 2; add_mul_src++) {
 165          for (unsigned i = 0; i < 4; i++)
 166             swizzle[i] = i;
 167
 168          negate = false;
 169          abs = false;
 170
 171          mul = get_mul_for_src(&add->src[add_mul_src],
 172                                add->dest.dest.ssa.num_components,
 173                                swizzle, &negate, &abs);
 174
 175          if (mul != NULL)
 176             break;
 177       }
 178
 179       if (mul == NULL)
 180          continue;
 181
 182       nir_ssa_def *mul_src[2];
 183       mul_src[0] = mul->src[0].src.ssa;
 184       mul_src[1] = mul->src[1].src.ssa;
 185
 186       if (abs) {
 187          for (unsigned i = 0; i < 2; i++) {
 188             nir_alu_instr *abs = nir_alu_instr_create(state->mem_ctx,
 189                                                       nir_op_fabs);
 190             abs->src[0].src = nir_src_for_ssa(mul_src[i]);
 191             nir_ssa_dest_init(&abs->instr, &abs->dest.dest,
 192                               mul_src[i]->num_components, NULL);
 193             abs->dest.write_mask = (1 << mul_src[i]->num_components) - 1;
 194             nir_instr_insert_before(&add->instr, &abs->instr);
 195             mul_src[i] = &abs->dest.dest.ssa;
 196          }
 197       }
 198
 199       if (negate) {
 200          nir_alu_instr *neg = nir_alu_instr_create(state->mem_ctx,
 201                                                    nir_op_fneg);
 202          neg->src[0].src = nir_src_for_ssa(mul_src[0]);
 203          nir_ssa_dest_init(&neg->instr, &neg->dest.dest,
 204                            mul_src[0]->num_components, NULL);
 205          neg->dest.write_mask = (1 << mul_src[0]->num_components) - 1;
 206          nir_instr_insert_before(&add->instr, &neg->instr);
 207          mul_src[0] = &neg->dest.dest.ssa;
 208       }
 209
 210       nir_alu_instr *ffma = nir_alu_instr_create(state->mem_ctx, nir_op_ffma);
 211       ffma->dest.saturate = add->dest.saturate;
 212       ffma->dest.write_mask = add->dest.write_mask;
 213
 214       for (unsigned i = 0; i < 2; i++) {
 215          ffma->src[i].src = nir_src_for_ssa(mul_src[i]);
 216          for (unsigned j = 0; j < add->dest.dest.ssa.num_components; j++)
 217             ffma->src[i].swizzle[j] = mul->src[i].swizzle[swizzle[j]];
 218       }
 219       nir_alu_src_copy(&ffma->src[2], &add->src[1 - add_mul_src], ffma);
 220
 221       assert(add->dest.dest.is_ssa);
 222
 223       nir_ssa_dest_init(&ffma->instr, &ffma->dest.dest,
 224                         add->dest.dest.ssa.num_components,
 225                         add->dest.dest.ssa.name);
 226       nir_ssa_def_rewrite_uses(&add->dest.dest.ssa,
 227                                nir_src_for_ssa(&ffma->dest.dest.ssa));
 228
 229       nir_instr_insert_before(&add->instr, &ffma->instr);
 230       assert(list_empty(&add->dest.dest.ssa.uses));
 231       nir_instr_remove(&add->instr);
 232
 233       state->progress = true;
 234    }
 235
 236    return true;
 237 }
 238
 239 static bool
 240 nir_opt_peephole_ffma_impl(nir_function_impl *impl)
 241 {
 242    struct peephole_ffma_state state;
 243
 244    state.mem_ctx = ralloc_parent(impl);
 245    state.impl = impl;
 246    state.progress = false;
 247
 248    nir_foreach_block(impl, nir_opt_peephole_ffma_block, &state);
 249
 250    if (state.progress)
 251       nir_metadata_preserve(impl, nir_metadata_block_index |
 252                                   nir_metadata_dominance);
 253
 254    return state.progress;
 255 }
 256
 257 bool
 258 nir_opt_peephole_ffma(nir_shader *shader)
 259 {
 260    bool progress = false;
 261
 262    nir_foreach_overload(shader, overload) {
 263       if (overload->impl)
 264          progress |= nir_opt_peephole_ffma_impl(overload->impl);
 265    }
 266
 267    return progress;
 268 }