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 {
  36    nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components, NULL);
  37    instr->dest.write_mask = (1 << num_components) - 1;
  38 }
  39
  40 static void
  41 lower_reduction(nir_alu_instr *instr, nir_op chan_op, nir_op merge_op,
  42                 nir_builder *builder)
  43 {
  44    unsigned num_components = nir_op_infos[instr->op].input_sizes[0];
  45
  46    nir_ssa_def *last = NULL;
  47    for (unsigned i = 0; i < num_components; i++) {
  48       nir_alu_instr *chan = nir_alu_instr_create(builder->shader, chan_op);
  49       nir_alu_ssa_dest_init(chan, 1);
  50       nir_alu_src_copy(&chan->src[0], &instr->src[0], chan);
  51       chan->src[0].swizzle[0] = chan->src[0].swizzle[i];
  52       if (nir_op_infos[chan_op].num_inputs > 1) {
  53          assert(nir_op_infos[chan_op].num_inputs == 2);
  54          nir_alu_src_copy(&chan->src[1], &instr->src[1], chan);
  55          chan->src[1].swizzle[0] = chan->src[1].swizzle[i];
  56       }
  57
  58       nir_builder_instr_insert(builder, &chan->instr);
  59
  60       if (i == 0) {
  61          last = &chan->dest.dest.ssa;
  62       } else {
  63          last = nir_build_alu(builder, merge_op,
  64                               last, &chan->dest.dest.ssa, NULL, NULL);
  65       }
  66    }
  67
  68    assert(instr->dest.write_mask == 1);
  69    nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(last));
  70    nir_instr_remove(&instr->instr);
  71 }
  72
  73 static void
  74 lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
  75 {
  76    unsigned num_src = nir_op_infos[instr->op].num_inputs;
  77    unsigned i, chan;
  78
  79    assert(instr->dest.dest.is_ssa);
  80    assert(instr->dest.write_mask != 0);
  81
  82    b->cursor = nir_before_instr(&instr->instr);
  83
  84 #define LOWER_REDUCTION(name, chan, merge) \
  85    case name##2: \
  86    case name##3: \
  87    case name##4: \
  88       lower_reduction(instr, chan, merge, b); \
  89       return;
  90
  91    switch (instr->op) {
  92    case nir_op_vec4:
  93    case nir_op_vec3:
  94    case nir_op_vec2:
  95       /* We don't need to scalarize these ops, they're the ones generated to
  96        * group up outputs into a value that can be SSAed.
  97        */
  98       return;
  99
 100    case nir_op_unpack_unorm_4x8:
 101    case nir_op_unpack_snorm_4x8:
 102    case nir_op_unpack_unorm_2x16:
 103    case nir_op_unpack_snorm_2x16:
 104       /* There is no scalar version of these ops, unless we were to break it
 105        * down to bitshifts and math (which is definitely not intended).
 106        */
 107       return;
 108
 109    case nir_op_unpack_half_2x16:
 110       /* We could split this into unpack_half_2x16_split_[xy], but should
 111        * we?
 112        */
 113       return;
 114
 115    case nir_op_fdph: {
 116       nir_ssa_def *sum[4];
 117       for (unsigned i = 0; i < 3; i++) {
 118          sum[i] = nir_fmul(b, nir_channel(b, instr->src[0].src.ssa,
 119                                           instr->src[0].swizzle[i]),
 120                               nir_channel(b, instr->src[1].src.ssa,
 121                                           instr->src[1].swizzle[i]));
 122       }
 123       sum[3] = nir_channel(b, instr->src[1].src.ssa, instr->src[1].swizzle[3]);
 124
 125       nir_ssa_def *val = nir_fadd(b, nir_fadd(b, sum[0], sum[1]),
 126                                      nir_fadd(b, sum[2], sum[3]));
 127
 128       nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
 129       nir_instr_remove(&instr->instr);
 130       return;
 131    }
 132
 133       LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
 134       LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand);
 135       LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand);
 136       LOWER_REDUCTION(nir_op_bany_fnequal, nir_op_fne, nir_op_ior);
 137       LOWER_REDUCTION(nir_op_bany_inequal, nir_op_ine, nir_op_ior);
 138       LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fand);
 139       LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_for);
 140
 141    default:
 142       break;
 143    }
 144
 145    if (instr->dest.dest.ssa.num_components == 1)
 146       return;
 147
 148    unsigned num_components = instr->dest.dest.ssa.num_components;
 149    nir_ssa_def *comps[] = { NULL, NULL, NULL, NULL };
 150
 151    for (chan = 0; chan < 4; chan++) {
 152       if (!(instr->dest.write_mask & (1 << chan)))
 153          continue;
 154
 155       nir_alu_instr *lower = nir_alu_instr_create(b->shader, instr->op);
 156       for (i = 0; i < num_src; i++) {
 157          /* We only handle same-size-as-dest (input_sizes[] == 0) or scalar
 158           * args (input_sizes[] == 1).
 159           */
 160          assert(nir_op_infos[instr->op].input_sizes[i] < 2);
 161          unsigned src_chan = (nir_op_infos[instr->op].input_sizes[i] == 1 ?
 162                               0 : chan);
 163
 164          nir_alu_src_copy(&lower->src[i], &instr->src[i], lower);
 165          for (int j = 0; j < 4; j++)
 166             lower->src[i].swizzle[j] = instr->src[i].swizzle[src_chan];
 167       }
 168
 169       nir_alu_ssa_dest_init(lower, 1);
 170       lower->dest.saturate = instr->dest.saturate;
 171       comps[chan] = &lower->dest.dest.ssa;
 172
 173       nir_builder_instr_insert(b, &lower->instr);
 174    }
 175
 176    nir_ssa_def *vec = nir_vec(b, comps, num_components);
 177
 178    nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
 179
 180    nir_instr_remove(&instr->instr);
 181 }
 182
 183 static bool
 184 lower_alu_to_scalar_block(nir_block *block, void *builder)
 185 {
 186    nir_foreach_instr_safe(block, instr) {
 187       if (instr->type == nir_instr_type_alu)
 188          lower_alu_instr_scalar(nir_instr_as_alu(instr), builder);
 189    }
 190
 191    return true;
 192 }
 193
 194 static void
 195 nir_lower_alu_to_scalar_impl(nir_function_impl *impl)
 196 {
 197    nir_builder builder;
 198    nir_builder_init(&builder, impl);
 199
 200    nir_foreach_block(impl, lower_alu_to_scalar_block, &builder);
 201 }
 202
 203 void
 204 nir_lower_alu_to_scalar(nir_shader *shader)
 205 {
 206    nir_foreach_function(shader, function) {
 207       if (function->impl)
 208          nir_lower_alu_to_scalar_impl(function->impl);
 209    }
 210 }