src/gallium/drivers/vc4/vc4_nir_lower_io.c

   1 /*
   2  * Copyright © 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 "vc4_qir.h"
  25 #include "glsl/nir/nir_builder.h"
  26 #include "util/u_format.h"
  27
  28 /**
  29  * Walks the NIR generated by TGSI-to-NIR to lower its io intrinsics into
  30  * something amenable to the VC4 architecture.
  31  *
  32  * Currently, it split inputs, outputs, and uniforms into scalars, drops any
  33  * non-position outputs in coordinate shaders, and fixes up the addressing on
  34  * indirect uniform loads.
  35  */
  36
  37 static void
  38 replace_intrinsic_with_vec4(nir_builder *b, nir_intrinsic_instr *intr,
  39                             nir_ssa_def **comps)
  40 {
  41
  42         /* Batch things back together into a vec4.  This will get split by the
  43          * later ALU scalarization pass.
  44          */
  45         nir_ssa_def *vec = nir_vec4(b, comps[0], comps[1], comps[2], comps[3]);
  46
  47         /* Replace the old intrinsic with a reference to our reconstructed
  48          * vec4.
  49          */
  50         nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(vec));
  51         nir_instr_remove(&intr->instr);
  52 }
  53
  54 static nir_ssa_def *
  55 vc4_nir_unpack_8i(nir_builder *b, nir_ssa_def *src, unsigned chan)
  56 {
  57         return nir_ubitfield_extract(b,
  58                                      src,
  59                                      nir_imm_int(b, 8 * chan),
  60                                      nir_imm_int(b, 8));
  61 }
  62
  63 /** Returns the 16 bit field as a sign-extended 32-bit value. */
  64 static nir_ssa_def *
  65 vc4_nir_unpack_16i(nir_builder *b, nir_ssa_def *src, unsigned chan)
  66 {
  67         return nir_ibitfield_extract(b,
  68                                      src,
  69                                      nir_imm_int(b, 16 * chan),
  70                                      nir_imm_int(b, 16));
  71 }
  72
  73 /** Returns the 16 bit field as an unsigned 32 bit value. */
  74 static nir_ssa_def *
  75 vc4_nir_unpack_16u(nir_builder *b, nir_ssa_def *src, unsigned chan)
  76 {
  77         if (chan == 0) {
  78                 return nir_iand(b, src, nir_imm_int(b, 0xffff));
  79         } else {
  80                 return nir_ushr(b, src, nir_imm_int(b, 16));
  81         }
  82 }
  83
  84 static nir_ssa_def *
  85 vc4_nir_unpack_8f(nir_builder *b, nir_ssa_def *src, unsigned chan)
  86 {
  87         return nir_channel(b, nir_unpack_unorm_4x8(b, src), chan);
  88 }
  89
  90 static nir_ssa_def *
  91 vc4_nir_get_vattr_channel_vpm(struct vc4_compile *c,
  92                               nir_builder *b,
  93                               nir_ssa_def **vpm_reads,
  94                               uint8_t swiz,
  95                               const struct util_format_description *desc)
  96 {
  97         const struct util_format_channel_description *chan =
  98                 &desc->channel[swiz];
  99         nir_ssa_def *temp;
 100
 101         if (swiz > UTIL_FORMAT_SWIZZLE_W) {
 102                 return vc4_nir_get_swizzled_channel(b, vpm_reads, swiz);
 103         } else if (chan->size == 32 && chan->type == UTIL_FORMAT_TYPE_FLOAT) {
 104                 return vc4_nir_get_swizzled_channel(b, vpm_reads, swiz);
 105         } else if (chan->size == 32 && chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 106                 if (chan->normalized) {
 107                         return nir_fmul(b,
 108                                         nir_i2f(b, vpm_reads[swiz]),
 109                                         nir_imm_float(b,
 110                                                       1.0 / 0x7fffffff));
 111                 } else {
 112                         return nir_i2f(b, vpm_reads[swiz]);
 113                 }
 114         } else if (chan->size == 8 &&
 115                    (chan->type == UTIL_FORMAT_TYPE_UNSIGNED ||
 116                     chan->type == UTIL_FORMAT_TYPE_SIGNED)) {
 117                 nir_ssa_def *vpm = vpm_reads[0];
 118                 if (chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 119                         temp = nir_ixor(b, vpm, nir_imm_int(b, 0x80808080));
 120                         if (chan->normalized) {
 121                                 return nir_fsub(b, nir_fmul(b,
 122                                                             vc4_nir_unpack_8f(b, temp, swiz),
 123                                                             nir_imm_float(b, 2.0)),
 124                                                 nir_imm_float(b, 1.0));
 125                         } else {
 126                                 return nir_fadd(b,
 127                                                 nir_i2f(b,
 128                                                         vc4_nir_unpack_8i(b, temp,
 129                                                                           swiz)),
 130                                                 nir_imm_float(b, -128.0));
 131                         }
 132                 } else {
 133                         if (chan->normalized) {
 134                                 return vc4_nir_unpack_8f(b, vpm, swiz);
 135                         } else {
 136                                 return nir_i2f(b, vc4_nir_unpack_8i(b, vpm, swiz));
 137                         }
 138                 }
 139         } else if (chan->size == 16 &&
 140                    (chan->type == UTIL_FORMAT_TYPE_UNSIGNED ||
 141                     chan->type == UTIL_FORMAT_TYPE_SIGNED)) {
 142                 nir_ssa_def *vpm = vpm_reads[swiz / 2];
 143
 144                 /* Note that UNPACK_16F eats a half float, not ints, so we use
 145                  * UNPACK_16_I for all of these.
 146                  */
 147                 if (chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 148                         temp = nir_i2f(b, vc4_nir_unpack_16i(b, vpm, swiz & 1));
 149                         if (chan->normalized) {
 150                                 return nir_fmul(b, temp,
 151                                                 nir_imm_float(b, 1/32768.0f));
 152                         } else {
 153                                 return temp;
 154                         }
 155                 } else {
 156                         temp = nir_i2f(b, vc4_nir_unpack_16u(b, vpm, swiz & 1));
 157                         if (chan->normalized) {
 158                                 return nir_fmul(b, temp,
 159                                                 nir_imm_float(b, 1 / 65535.0));
 160                         } else {
 161                                 return temp;
 162                         }
 163                 }
 164         } else {
 165                 return NULL;
 166         }
 167 }
 168
 169 static void
 170 vc4_nir_lower_vertex_attr(struct vc4_compile *c, nir_builder *b,
 171                           nir_intrinsic_instr *intr)
 172 {
 173         b->cursor = nir_before_instr(&intr->instr);
 174
 175         int attr = intr->const_index[0];
 176         enum pipe_format format = c->vs_key->attr_formats[attr];
 177         uint32_t attr_size = util_format_get_blocksize(format);
 178
 179         /* All TGSI-to-NIR inputs are vec4. */
 180         assert(intr->num_components == 4);
 181
 182         /* Generate dword loads for the VPM values (Since these intrinsics may
 183          * be reordered, the actual reads will be generated at the top of the
 184          * shader by ntq_setup_inputs().
 185          */
 186         nir_ssa_def *vpm_reads[4];
 187         for (int i = 0; i < align(attr_size, 4) / 4; i++) {
 188                 nir_intrinsic_instr *intr_comp =
 189                         nir_intrinsic_instr_create(c->s,
 190                                                    nir_intrinsic_load_input);
 191                 intr_comp->num_components = 1;
 192                 intr_comp->const_index[0] = intr->const_index[0] * 4 + i;
 193                 nir_ssa_dest_init(&intr_comp->instr, &intr_comp->dest, 1, NULL);
 194                 nir_builder_instr_insert(b, &intr_comp->instr);
 195
 196                 vpm_reads[i] = &intr_comp->dest.ssa;
 197         }
 198
 199         bool format_warned = false;
 200         const struct util_format_description *desc =
 201                 util_format_description(format);
 202
 203         nir_ssa_def *dests[4];
 204         for (int i = 0; i < 4; i++) {
 205                 uint8_t swiz = desc->swizzle[i];
 206                 dests[i] = vc4_nir_get_vattr_channel_vpm(c, b, vpm_reads, swiz,
 207                                                          desc);
 208
 209                 if (!dests[i]) {
 210                         if (!format_warned) {
 211                                 fprintf(stderr,
 212                                         "vtx element %d unsupported type: %s\n",
 213                                         attr, util_format_name(format));
 214                                 format_warned = true;
 215                         }
 216                         dests[i] = nir_imm_float(b, 0.0);
 217                 }
 218         }
 219
 220         replace_intrinsic_with_vec4(b, intr, dests);
 221 }
 222
 223 static void
 224 vc4_nir_lower_fs_input(struct vc4_compile *c, nir_builder *b,
 225                        nir_intrinsic_instr *intr)
 226 {
 227         b->cursor = nir_before_instr(&intr->instr);
 228
 229         if (intr->const_index[0] >= VC4_NIR_TLB_COLOR_READ_INPUT &&
 230             intr->const_index[0] < (VC4_NIR_TLB_COLOR_READ_INPUT +
 231                                     VC4_MAX_SAMPLES)) {
 232                 /* This doesn't need any lowering. */
 233                 return;
 234         }
 235
 236         nir_variable *input_var = NULL;
 237         nir_foreach_variable(var, &c->s->inputs) {
 238                 if (var->data.driver_location == intr->const_index[0]) {
 239                         input_var = var;
 240                         break;
 241                 }
 242         }
 243         assert(input_var);
 244
 245         /* All TGSI-to-NIR inputs are vec4. */
 246         assert(intr->num_components == 4);
 247
 248         /* Generate scalar loads equivalent to the original VEC4. */
 249         nir_ssa_def *dests[4];
 250         for (unsigned i = 0; i < intr->num_components; i++) {
 251                 nir_intrinsic_instr *intr_comp =
 252                         nir_intrinsic_instr_create(c->s, nir_intrinsic_load_input);
 253                 intr_comp->num_components = 1;
 254                 intr_comp->const_index[0] = intr->const_index[0] * 4 + i;
 255                 nir_ssa_dest_init(&intr_comp->instr, &intr_comp->dest, 1, NULL);
 256                 nir_builder_instr_insert(b, &intr_comp->instr);
 257
 258                 dests[i] = &intr_comp->dest.ssa;
 259         }
 260
 261         if (input_var->data.location == VARYING_SLOT_FACE) {
 262                 dests[0] = nir_fsub(b,
 263                                     nir_imm_float(b, 1.0),
 264                                     nir_fmul(b,
 265                                              nir_i2f(b, dests[0]),
 266                                              nir_imm_float(b, 2.0)));
 267                 dests[1] = nir_imm_float(b, 0.0);
 268                 dests[2] = nir_imm_float(b, 0.0);
 269                 dests[3] = nir_imm_float(b, 1.0);
 270         } else if (input_var->data.location >= VARYING_SLOT_VAR0) {
 271                 if (c->fs_key->point_sprite_mask &
 272                     (1 << (input_var->data.location -
 273                            VARYING_SLOT_VAR0))) {
 274                         if (!c->fs_key->is_points) {
 275                                 dests[0] = nir_imm_float(b, 0.0);
 276                                 dests[1] = nir_imm_float(b, 0.0);
 277                         }
 278                         if (c->fs_key->point_coord_upper_left) {
 279                                 dests[1] = nir_fsub(b,
 280                                                     nir_imm_float(b, 1.0),
 281                                                     dests[1]);
 282                         }
 283                         dests[2] = nir_imm_float(b, 0.0);
 284                         dests[3] = nir_imm_float(b, 1.0);
 285                 }
 286         }
 287
 288         replace_intrinsic_with_vec4(b, intr, dests);
 289 }
 290
 291 static void
 292 vc4_nir_lower_output(struct vc4_compile *c, nir_builder *b,
 293                      nir_intrinsic_instr *intr)
 294 {
 295         nir_variable *output_var = NULL;
 296         nir_foreach_variable(var, &c->s->outputs) {
 297                 if (var->data.driver_location == intr->const_index[0]) {
 298                         output_var = var;
 299                         break;
 300                 }
 301         }
 302         assert(output_var);
 303
 304         if (c->stage == QSTAGE_COORD &&
 305             output_var->data.location != VARYING_SLOT_POS &&
 306             output_var->data.location != VARYING_SLOT_PSIZ) {
 307                 nir_instr_remove(&intr->instr);
 308                 return;
 309         }
 310
 311         /* Color output is lowered by vc4_nir_lower_blend(). */
 312         if (c->stage == QSTAGE_FRAG &&
 313             (output_var->data.location == FRAG_RESULT_COLOR ||
 314              output_var->data.location == FRAG_RESULT_DATA0)) {
 315                 intr->const_index[0] *= 4;
 316                 return;
 317         }
 318
 319         /* All TGSI-to-NIR outputs are VEC4. */
 320         assert(intr->num_components == 4);
 321
 322         b->cursor = nir_before_instr(&intr->instr);
 323
 324         for (unsigned i = 0; i < intr->num_components; i++) {
 325                 nir_intrinsic_instr *intr_comp =
 326                         nir_intrinsic_instr_create(c->s, nir_intrinsic_store_output);
 327                 intr_comp->num_components = 1;
 328                 intr_comp->const_index[0] = intr->const_index[0] * 4 + i;
 329
 330                 assert(intr->src[0].is_ssa);
 331                 intr_comp->src[0] =
 332                         nir_src_for_ssa(nir_channel(b, intr->src[0].ssa, i));
 333                 nir_builder_instr_insert(b, &intr_comp->instr);
 334         }
 335
 336         nir_instr_remove(&intr->instr);
 337 }
 338
 339 static void
 340 vc4_nir_lower_uniform(struct vc4_compile *c, nir_builder *b,
 341                       nir_intrinsic_instr *intr)
 342 {
 343         /* All TGSI-to-NIR uniform loads are vec4, but we may create dword
 344          * loads in our lowering passes.
 345          */
 346         if (intr->num_components == 1)
 347                 return;
 348         assert(intr->num_components == 4);
 349
 350         b->cursor = nir_before_instr(&intr->instr);
 351
 352         /* Generate scalar loads equivalent to the original VEC4. */
 353         nir_ssa_def *dests[4];
 354         for (unsigned i = 0; i < intr->num_components; i++) {
 355                 nir_intrinsic_instr *intr_comp =
 356                         nir_intrinsic_instr_create(c->s, intr->intrinsic);
 357                 intr_comp->num_components = 1;
 358                 nir_ssa_dest_init(&intr_comp->instr, &intr_comp->dest, 1, NULL);
 359
 360                 if (intr->intrinsic == nir_intrinsic_load_uniform_indirect) {
 361                         /* Convert the variable TGSI register index to a byte
 362                          * offset.
 363                          */
 364                         intr_comp->src[0] =
 365                                 nir_src_for_ssa(nir_ishl(b,
 366                                                          intr->src[0].ssa,
 367                                                          nir_imm_int(b, 4)));
 368
 369                         /* Convert the offset to be a byte index, too. */
 370                         intr_comp->const_index[0] = (intr->const_index[0] * 16 +
 371                                                      i * 4);
 372                 } else {
 373                         /* We want a dword index for non-indirect uniform
 374                          * loads.
 375                          */
 376                         intr_comp->const_index[0] = (intr->const_index[0] * 4 +
 377                                                      i);
 378                 }
 379
 380                 dests[i] = &intr_comp->dest.ssa;
 381
 382                 nir_builder_instr_insert(b, &intr_comp->instr);
 383         }
 384
 385         replace_intrinsic_with_vec4(b, intr, dests);
 386 }
 387
 388 static void
 389 vc4_nir_lower_io_instr(struct vc4_compile *c, nir_builder *b,
 390                        struct nir_instr *instr)
 391 {
 392         if (instr->type != nir_instr_type_intrinsic)
 393                 return;
 394         nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 395
 396         switch (intr->intrinsic) {
 397         case nir_intrinsic_load_input:
 398                 if (c->stage == QSTAGE_FRAG)
 399                         vc4_nir_lower_fs_input(c, b, intr);
 400                 else
 401                         vc4_nir_lower_vertex_attr(c, b, intr);
 402                 break;
 403
 404         case nir_intrinsic_store_output:
 405                 vc4_nir_lower_output(c, b, intr);
 406                 break;
 407
 408         case nir_intrinsic_load_uniform:
 409         case nir_intrinsic_load_uniform_indirect:
 410         case nir_intrinsic_load_user_clip_plane:
 411                 vc4_nir_lower_uniform(c, b, intr);
 412                 break;
 413
 414         default:
 415                 break;
 416         }
 417 }
 418
 419 static bool
 420 vc4_nir_lower_io_block(nir_block *block, void *arg)
 421 {
 422         struct vc4_compile *c = arg;
 423         nir_function_impl *impl =
 424                 nir_cf_node_get_function(&block->cf_node);
 425
 426         nir_builder b;
 427         nir_builder_init(&b, impl);
 428
 429         nir_foreach_instr_safe(block, instr)
 430                 vc4_nir_lower_io_instr(c, &b, instr);
 431
 432         return true;
 433 }
 434
 435 static bool
 436 vc4_nir_lower_io_impl(struct vc4_compile *c, nir_function_impl *impl)
 437 {
 438         nir_foreach_block(impl, vc4_nir_lower_io_block, c);
 439
 440         nir_metadata_preserve(impl, nir_metadata_block_index |
 441                               nir_metadata_dominance);
 442
 443         return true;
 444 }
 445
 446 void
 447 vc4_nir_lower_io(struct vc4_compile *c)
 448 {
 449         nir_foreach_overload(c->s, overload) {
 450                 if (overload->impl)
 451                         vc4_nir_lower_io_impl(c, overload->impl);
 452         }
 453 }