src/compiler/nir/nir_lower_alu_to_scalar.c

   1 /*
   2  * Copyright © 2014-2015 Broadcom
   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
  24 #include "nir.h"
  25 #include "nir_builder.h"
  26
  27 /** @file nir_lower_alu_to_scalar.c
  28  *
  29  * Replaces nir_alu_instr operations with more than one channel used in the
  30  * arguments with individual per-channel operations.
  31  */
  32
  33 static void
  34 nir_alu_ssa_dest_init(nir_alu_instr *alu, unsigned num_components,
  35                       unsigned bit_size)
  36 {
  37    nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components,
  38                      bit_size, NULL);
  39    alu->dest.write_mask = (1 << num_components) - 1;
  40 }
  41
  42 static void
  43 lower_reduction(nir_alu_instr *alu, nir_op chan_op, nir_op merge_op,
  44                 nir_builder *builder)
  45 {
  46    unsigned num_components = nir_op_infos[alu->op].input_sizes[0];
  47
  48    nir_ssa_def *last = NULL;
  49    for (unsigned i = 0; i < num_components; i++) {
  50       nir_alu_instr *chan = nir_alu_instr_create(builder->shader, chan_op);
  51       nir_alu_ssa_dest_init(chan, 1, alu->dest.dest.ssa.bit_size);
  52       nir_alu_src_copy(&chan->src[0], &alu->src[0], chan);
  53       chan->src[0].swizzle[0] = chan->src[0].swizzle[i];
  54       if (nir_op_infos[chan_op].num_inputs > 1) {
  55          assert(nir_op_infos[chan_op].num_inputs == 2);
  56          nir_alu_src_copy(&chan->src[1], &alu->src[1], chan);
  57          chan->src[1].swizzle[0] = chan->src[1].swizzle[i];
  58       }
  59       chan->exact = alu->exact;
  60
  61       nir_builder_instr_insert(builder, &chan->instr);
  62
  63       if (i == 0) {
  64          last = &chan->dest.dest.ssa;
  65       } else {
  66          last = nir_build_alu(builder, merge_op,
  67                               last, &chan->dest.dest.ssa, NULL, NULL);
  68       }
  69    }
  70
  71    assert(alu->dest.write_mask == 1);
  72    nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(last));
  73    nir_instr_remove(&alu->instr);
  74 }
  75
  76 static bool
  77 lower_alu_instr_scalar(nir_alu_instr *alu, nir_builder *b, BITSET_WORD *lower_set)
  78 {
  79    unsigned num_src = nir_op_infos[alu->op].num_inputs;
  80    unsigned i, chan;
  81
  82    assert(alu->dest.dest.is_ssa);
  83    assert(alu->dest.write_mask != 0);
  84
  85    b->cursor = nir_before_instr(&alu->instr);
  86    b->exact = alu->exact;
  87
  88    if (lower_set && !BITSET_TEST(lower_set, alu->op))
  89       return false;
  90
  91 #define LOWER_REDUCTION(name, chan, merge) \
  92    case name##2: \
  93    case name##3: \
  94    case name##4: \
  95       lower_reduction(alu, chan, merge, b); \
  96       return true;
  97
  98    switch (alu->op) {
  99    case nir_op_vec4:
 100    case nir_op_vec3:
 101    case nir_op_vec2:
 102    case nir_op_cube_face_coord:
 103    case nir_op_cube_face_index:
 104       /* We don't need to scalarize these ops, they're the ones generated to
 105        * group up outputs into a value that can be SSAed.
 106        */
 107       return false;
 108
 109    case nir_op_pack_half_2x16:
 110       if (!b->shader->options->lower_pack_half_2x16)
 111          return false;
 112
 113       nir_ssa_def *src_vec2 = nir_ssa_for_alu_src(b, alu, 0);
 114
 115       nir_ssa_def *val =
 116          nir_pack_half_2x16_split(b, nir_channel(b, src_vec2, 0),
 117                                      nir_channel(b, src_vec2, 1));
 118
 119       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(val));
 120       nir_instr_remove(&alu->instr);
 121       return true;
 122
 123    case nir_op_unpack_unorm_4x8:
 124    case nir_op_unpack_snorm_4x8:
 125    case nir_op_unpack_unorm_2x16:
 126    case nir_op_unpack_snorm_2x16:
 127       /* There is no scalar version of these ops, unless we were to break it
 128        * down to bitshifts and math (which is definitely not intended).
 129        */
 130       return false;
 131
 132    case nir_op_unpack_half_2x16: {
 133       if (!b->shader->options->lower_unpack_half_2x16)
 134          return false;
 135
 136       nir_ssa_def *packed = nir_ssa_for_alu_src(b, alu, 0);
 137
 138       nir_ssa_def *comps[2];
 139       comps[0] = nir_unpack_half_2x16_split_x(b, packed);
 140       comps[1] = nir_unpack_half_2x16_split_y(b, packed);
 141       nir_ssa_def *vec = nir_vec(b, comps, 2);
 142
 143       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(vec));
 144       nir_instr_remove(&alu->instr);
 145       return true;
 146    }
 147
 148    case nir_op_pack_uvec2_to_uint: {
 149       assert(b->shader->options->lower_pack_snorm_2x16 ||
 150              b->shader->options->lower_pack_unorm_2x16);
 151
 152       nir_ssa_def *word = nir_extract_u16(b, nir_ssa_for_alu_src(b, alu, 0),
 153                                              nir_imm_int(b, 0));
 154       nir_ssa_def *val =
 155          nir_ior(b, nir_ishl(b, nir_channel(b, word, 1), nir_imm_int(b, 16)),
 156                                 nir_channel(b, word, 0));
 157
 158       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(val));
 159       nir_instr_remove(&alu->instr);
 160       break;
 161    }
 162
 163    case nir_op_pack_uvec4_to_uint: {
 164       assert(b->shader->options->lower_pack_snorm_4x8 ||
 165              b->shader->options->lower_pack_unorm_4x8);
 166
 167       nir_ssa_def *byte = nir_extract_u8(b, nir_ssa_for_alu_src(b, alu, 0),
 168                                             nir_imm_int(b, 0));
 169       nir_ssa_def *val =
 170          nir_ior(b, nir_ior(b, nir_ishl(b, nir_channel(b, byte, 3), nir_imm_int(b, 24)),
 171                                nir_ishl(b, nir_channel(b, byte, 2), nir_imm_int(b, 16))),
 172                     nir_ior(b, nir_ishl(b, nir_channel(b, byte, 1), nir_imm_int(b, 8)),
 173                                nir_channel(b, byte, 0)));
 174
 175       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(val));
 176       nir_instr_remove(&alu->instr);
 177       break;
 178    }
 179
 180    case nir_op_fdph: {
 181       nir_ssa_def *src0_vec = nir_ssa_for_alu_src(b, alu, 0);
 182       nir_ssa_def *src1_vec = nir_ssa_for_alu_src(b, alu, 1);
 183
 184       nir_ssa_def *sum[4];
 185       for (unsigned i = 0; i < 3; i++) {
 186          sum[i] = nir_fmul(b, nir_channel(b, src0_vec, i),
 187                               nir_channel(b, src1_vec, i));
 188       }
 189       sum[3] = nir_channel(b, src1_vec, 3);
 190
 191       nir_ssa_def *val = nir_fadd(b, nir_fadd(b, sum[0], sum[1]),
 192                                      nir_fadd(b, sum[2], sum[3]));
 193
 194       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(val));
 195       nir_instr_remove(&alu->instr);
 196       return true;
 197    }
 198
 199    case nir_op_unpack_64_2x32:
 200    case nir_op_unpack_32_2x16:
 201       return false;
 202
 203       LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
 204       LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand);
 205       LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand);
 206       LOWER_REDUCTION(nir_op_bany_fnequal, nir_op_fne, nir_op_ior);
 207       LOWER_REDUCTION(nir_op_bany_inequal, nir_op_ine, nir_op_ior);
 208       LOWER_REDUCTION(nir_op_b32all_fequal, nir_op_feq32, nir_op_iand);
 209       LOWER_REDUCTION(nir_op_b32all_iequal, nir_op_ieq32, nir_op_iand);
 210       LOWER_REDUCTION(nir_op_b32any_fnequal, nir_op_fne32, nir_op_ior);
 211       LOWER_REDUCTION(nir_op_b32any_inequal, nir_op_ine32, nir_op_ior);
 212       LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fmin);
 213       LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_fmax);
 214
 215    default:
 216       break;
 217    }
 218
 219    if (alu->dest.dest.ssa.num_components == 1)
 220       return false;
 221
 222    unsigned num_components = alu->dest.dest.ssa.num_components;
 223    nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS] = { NULL };
 224
 225    for (chan = 0; chan < NIR_MAX_VEC_COMPONENTS; chan++) {
 226       if (!(alu->dest.write_mask & (1 << chan)))
 227          continue;
 228
 229       nir_alu_instr *lower = nir_alu_instr_create(b->shader, alu->op);
 230       for (i = 0; i < num_src; i++) {
 231          /* We only handle same-size-as-dest (input_sizes[] == 0) or scalar
 232           * args (input_sizes[] == 1).
 233           */
 234          assert(nir_op_infos[alu->op].input_sizes[i] < 2);
 235          unsigned src_chan = (nir_op_infos[alu->op].input_sizes[i] == 1 ?
 236                               0 : chan);
 237
 238          nir_alu_src_copy(&lower->src[i], &alu->src[i], lower);
 239          for (int j = 0; j < NIR_MAX_VEC_COMPONENTS; j++)
 240             lower->src[i].swizzle[j] = alu->src[i].swizzle[src_chan];
 241       }
 242
 243       nir_alu_ssa_dest_init(lower, 1, alu->dest.dest.ssa.bit_size);
 244       lower->dest.saturate = alu->dest.saturate;
 245       comps[chan] = &lower->dest.dest.ssa;
 246       lower->exact = alu->exact;
 247
 248       nir_builder_instr_insert(b, &lower->instr);
 249    }
 250
 251    nir_ssa_def *vec = nir_vec(b, comps, num_components);
 252
 253    nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(vec));
 254
 255    nir_instr_remove(&alu->instr);
 256    return true;
 257 }
 258
 259 static bool
 260 nir_lower_alu_to_scalar_impl(nir_function_impl *impl, BITSET_WORD *lower_set)
 261 {
 262    nir_builder builder;
 263    nir_builder_init(&builder, impl);
 264    bool progress = false;
 265
 266    nir_foreach_block(block, impl) {
 267       nir_foreach_instr_safe(instr, block) {
 268          if (instr->type == nir_instr_type_alu) {
 269             progress = lower_alu_instr_scalar(nir_instr_as_alu(instr),
 270                                               &builder,
 271                                               lower_set) || progress;
 272          }
 273       }
 274    }
 275
 276    nir_metadata_preserve(impl, nir_metadata_block_index |
 277                                nir_metadata_dominance);
 278
 279    return progress;
 280 }
 281
 282 bool
 283 nir_lower_alu_to_scalar(nir_shader *shader, BITSET_WORD *lower_set)
 284 {
 285    bool progress = false;
 286
 287    nir_foreach_function(function, shader) {
 288       if (function->impl)
 289          progress = nir_lower_alu_to_scalar_impl(function->impl,
 290                                                  lower_set) || progress;
 291    }
 292
 293    return progress;
 294 }