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 struct alu_to_scalar_data {
  28    nir_instr_filter_cb cb;
  29    const void *data;
  30 };
  31
  32 /** @file nir_lower_alu_to_scalar.c
  33  *
  34  * Replaces nir_alu_instr operations with more than one channel used in the
  35  * arguments with individual per-channel operations.
  36  */
  37
  38 static bool
  39 inst_is_vector_alu(const nir_instr *instr, const void *_state)
  40 {
  41    if (instr->type != nir_instr_type_alu)
  42       return false;
  43
  44    nir_alu_instr *alu = nir_instr_as_alu(instr);
  45
  46    /* There is no ALU instruction which has a scalar destination, scalar
  47     * src[0], and some other vector source.
  48     */
  49    assert(alu->dest.dest.is_ssa);
  50    assert(alu->src[0].src.is_ssa);
  51    return alu->dest.dest.ssa.num_components > 1 ||
  52           nir_op_infos[alu->op].input_sizes[0] > 1;
  53 }
  54
  55 static void
  56 nir_alu_ssa_dest_init(nir_alu_instr *alu, unsigned num_components,
  57                       unsigned bit_size)
  58 {
  59    nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components,
  60                      bit_size, NULL);
  61    alu->dest.write_mask = (1 << num_components) - 1;
  62 }
  63
  64 static nir_ssa_def *
  65 lower_reduction(nir_alu_instr *alu, nir_op chan_op, nir_op merge_op,
  66                 nir_builder *builder)
  67 {
  68    unsigned num_components = nir_op_infos[alu->op].input_sizes[0];
  69
  70    nir_ssa_def *last = NULL;
  71    for (unsigned i = 0; i < num_components; i++) {
  72       nir_alu_instr *chan = nir_alu_instr_create(builder->shader, chan_op);
  73       nir_alu_ssa_dest_init(chan, 1, alu->dest.dest.ssa.bit_size);
  74       nir_alu_src_copy(&chan->src[0], &alu->src[0], chan);
  75       chan->src[0].swizzle[0] = chan->src[0].swizzle[i];
  76       if (nir_op_infos[chan_op].num_inputs > 1) {
  77          assert(nir_op_infos[chan_op].num_inputs == 2);
  78          nir_alu_src_copy(&chan->src[1], &alu->src[1], chan);
  79          chan->src[1].swizzle[0] = chan->src[1].swizzle[i];
  80       }
  81       chan->exact = alu->exact;
  82
  83       nir_builder_instr_insert(builder, &chan->instr);
  84
  85       if (i == 0) {
  86          last = &chan->dest.dest.ssa;
  87       } else {
  88          last = nir_build_alu(builder, merge_op,
  89                               last, &chan->dest.dest.ssa, NULL, NULL);
  90       }
  91    }
  92
  93    return last;
  94 }
  95
  96 static nir_ssa_def *
  97 lower_alu_instr_scalar(nir_builder *b, nir_instr *instr, void *_data)
  98 {
  99    struct alu_to_scalar_data *data = _data;
 100    nir_alu_instr *alu = nir_instr_as_alu(instr);
 101    unsigned num_src = nir_op_infos[alu->op].num_inputs;
 102    unsigned i, chan;
 103
 104    assert(alu->dest.dest.is_ssa);
 105    assert(alu->dest.write_mask != 0);
 106
 107    b->cursor = nir_before_instr(&alu->instr);
 108    b->exact = alu->exact;
 109
 110    if (data->cb && !data->cb(instr, data->data))
 111       return NULL;
 112
 113 #define LOWER_REDUCTION(name, chan, merge) \
 114    case name##2: \
 115    case name##3: \
 116    case name##4: \
 117       return lower_reduction(alu, chan, merge, b); \
 118
 119    switch (alu->op) {
 120    case nir_op_vec16:
 121    case nir_op_vec8:
 122    case nir_op_vec4:
 123    case nir_op_vec3:
 124    case nir_op_vec2:
 125    case nir_op_cube_face_coord:
 126    case nir_op_cube_face_index:
 127       /* We don't need to scalarize these ops, they're the ones generated to
 128        * group up outputs into a value that can be SSAed.
 129        */
 130       return NULL;
 131
 132    case nir_op_pack_half_2x16: {
 133       if (!b->shader->options->lower_pack_half_2x16)
 134          return NULL;
 135
 136       nir_ssa_def *src_vec2 = nir_ssa_for_alu_src(b, alu, 0);
 137       return nir_pack_half_2x16_split(b, nir_channel(b, src_vec2, 0),
 138                                          nir_channel(b, src_vec2, 1));
 139    }
 140
 141    case nir_op_unpack_unorm_4x8:
 142    case nir_op_unpack_snorm_4x8:
 143    case nir_op_unpack_unorm_2x16:
 144    case nir_op_unpack_snorm_2x16:
 145       /* There is no scalar version of these ops, unless we were to break it
 146        * down to bitshifts and math (which is definitely not intended).
 147        */
 148       return NULL;
 149
 150    case nir_op_unpack_half_2x16_flush_to_zero:
 151    case nir_op_unpack_half_2x16: {
 152       if (!b->shader->options->lower_unpack_half_2x16)
 153          return NULL;
 154
 155       nir_ssa_def *packed = nir_ssa_for_alu_src(b, alu, 0);
 156       if (alu->op == nir_op_unpack_half_2x16_flush_to_zero) {
 157           return nir_vec2(b,
 158                           nir_unpack_half_2x16_split_x_flush_to_zero(b,
 159                                                                      packed),
 160                           nir_unpack_half_2x16_split_y_flush_to_zero(b,
 161                                                                      packed));
 162       } else {
 163           return nir_vec2(b,
 164                           nir_unpack_half_2x16_split_x(b, packed),
 165                           nir_unpack_half_2x16_split_y(b, packed));
 166       }
 167    }
 168
 169    case nir_op_pack_uvec2_to_uint: {
 170       assert(b->shader->options->lower_pack_snorm_2x16 ||
 171              b->shader->options->lower_pack_unorm_2x16);
 172
 173       nir_ssa_def *word = nir_extract_u16(b, nir_ssa_for_alu_src(b, alu, 0),
 174                                              nir_imm_int(b, 0));
 175       return nir_ior(b, nir_ishl(b, nir_channel(b, word, 1),
 176                                     nir_imm_int(b, 16)),
 177                         nir_channel(b, word, 0));
 178    }
 179
 180    case nir_op_pack_uvec4_to_uint: {
 181       assert(b->shader->options->lower_pack_snorm_4x8 ||
 182              b->shader->options->lower_pack_unorm_4x8);
 183
 184       nir_ssa_def *byte = nir_extract_u8(b, nir_ssa_for_alu_src(b, alu, 0),
 185                                             nir_imm_int(b, 0));
 186       return nir_ior(b, nir_ior(b, nir_ishl(b, nir_channel(b, byte, 3),
 187                                                nir_imm_int(b, 24)),
 188                                    nir_ishl(b, nir_channel(b, byte, 2),
 189                                                nir_imm_int(b, 16))),
 190                         nir_ior(b, nir_ishl(b, nir_channel(b, byte, 1),
 191                                                nir_imm_int(b, 8)),
 192                                    nir_channel(b, byte, 0)));
 193    }
 194
 195    case nir_op_fdph: {
 196       nir_ssa_def *src0_vec = nir_ssa_for_alu_src(b, alu, 0);
 197       nir_ssa_def *src1_vec = nir_ssa_for_alu_src(b, alu, 1);
 198
 199       nir_ssa_def *sum[4];
 200       for (unsigned i = 0; i < 3; i++) {
 201          sum[i] = nir_fmul(b, nir_channel(b, src0_vec, i),
 202                               nir_channel(b, src1_vec, i));
 203       }
 204       sum[3] = nir_channel(b, src1_vec, 3);
 205
 206       return nir_fadd(b, nir_fadd(b, sum[0], sum[1]),
 207                          nir_fadd(b, sum[2], sum[3]));
 208    }
 209
 210    case nir_op_unpack_64_2x32:
 211    case nir_op_unpack_64_4x16:
 212    case nir_op_unpack_32_2x16:
 213       return NULL;
 214
 215       LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
 216       LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand);
 217       LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand);
 218       LOWER_REDUCTION(nir_op_bany_fnequal, nir_op_fne, nir_op_ior);
 219       LOWER_REDUCTION(nir_op_bany_inequal, nir_op_ine, nir_op_ior);
 220       LOWER_REDUCTION(nir_op_b8all_fequal, nir_op_feq8, nir_op_iand);
 221       LOWER_REDUCTION(nir_op_b8all_iequal, nir_op_ieq8, nir_op_iand);
 222       LOWER_REDUCTION(nir_op_b8any_fnequal, nir_op_fne8, nir_op_ior);
 223       LOWER_REDUCTION(nir_op_b8any_inequal, nir_op_ine8, nir_op_ior);
 224       LOWER_REDUCTION(nir_op_b16all_fequal, nir_op_feq16, nir_op_iand);
 225       LOWER_REDUCTION(nir_op_b16all_iequal, nir_op_ieq16, nir_op_iand);
 226       LOWER_REDUCTION(nir_op_b16any_fnequal, nir_op_fne16, nir_op_ior);
 227       LOWER_REDUCTION(nir_op_b16any_inequal, nir_op_ine16, nir_op_ior);
 228       LOWER_REDUCTION(nir_op_b32all_fequal, nir_op_feq32, nir_op_iand);
 229       LOWER_REDUCTION(nir_op_b32all_iequal, nir_op_ieq32, nir_op_iand);
 230       LOWER_REDUCTION(nir_op_b32any_fnequal, nir_op_fne32, nir_op_ior);
 231       LOWER_REDUCTION(nir_op_b32any_inequal, nir_op_ine32, nir_op_ior);
 232       LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fmin);
 233       LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_fmax);
 234
 235    default:
 236       break;
 237    }
 238
 239    if (alu->dest.dest.ssa.num_components == 1)
 240       return NULL;
 241
 242    unsigned num_components = alu->dest.dest.ssa.num_components;
 243    nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS] = { NULL };
 244
 245    for (chan = 0; chan < NIR_MAX_VEC_COMPONENTS; chan++) {
 246       if (!(alu->dest.write_mask & (1 << chan)))
 247          continue;
 248
 249       nir_alu_instr *lower = nir_alu_instr_create(b->shader, alu->op);
 250       for (i = 0; i < num_src; i++) {
 251          /* We only handle same-size-as-dest (input_sizes[] == 0) or scalar
 252           * args (input_sizes[] == 1).
 253           */
 254          assert(nir_op_infos[alu->op].input_sizes[i] < 2);
 255          unsigned src_chan = (nir_op_infos[alu->op].input_sizes[i] == 1 ?
 256                               0 : chan);
 257
 258          nir_alu_src_copy(&lower->src[i], &alu->src[i], lower);
 259          for (int j = 0; j < NIR_MAX_VEC_COMPONENTS; j++)
 260             lower->src[i].swizzle[j] = alu->src[i].swizzle[src_chan];
 261       }
 262
 263       nir_alu_ssa_dest_init(lower, 1, alu->dest.dest.ssa.bit_size);
 264       lower->dest.saturate = alu->dest.saturate;
 265       comps[chan] = &lower->dest.dest.ssa;
 266       lower->exact = alu->exact;
 267
 268       nir_builder_instr_insert(b, &lower->instr);
 269    }
 270
 271    return nir_vec(b, comps, num_components);
 272 }
 273
 274 bool
 275 nir_lower_alu_to_scalar(nir_shader *shader, nir_instr_filter_cb cb, const void *_data)
 276 {
 277    struct alu_to_scalar_data data = {
 278       .cb = cb,
 279       .data = _data,
 280    };
 281
 282    return nir_shader_lower_instructions(shader,
 283                                         inst_is_vector_alu,
 284                                         lower_alu_instr_scalar,
 285                                         &data);
 286 }