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