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 "compiler/nir/nir_builder.h"
  26 #include "util/format/u_format.h"
  27 #include "util/u_helpers.h"
  28
  29 /**
  30  * Walks the NIR generated by TGSI-to-NIR or GLSL-to-NIR to lower its io
  31  * intrinsics into something amenable to the VC4 architecture.
  32  *
  33  * Currently, it splits VS inputs and uniforms into scalars, drops any
  34  * non-position outputs in coordinate shaders, and fixes up the addressing on
  35  * indirect uniform loads.  FS input and VS output scalarization is handled by
  36  * nir_lower_io_to_scalar().
  37  */
  38
  39 static void
  40 replace_intrinsic_with_vec(nir_builder *b, nir_intrinsic_instr *intr,
  41                            nir_ssa_def **comps)
  42 {
  43
  44         /* Batch things back together into a vector.  This will get split by
  45          * the later ALU scalarization pass.
  46          */
  47         nir_ssa_def *vec = nir_vec(b, comps, intr->num_components);
  48
  49         /* Replace the old intrinsic with a reference to our reconstructed
  50          * vector.
  51          */
  52         nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(vec));
  53         nir_instr_remove(&intr->instr);
  54 }
  55
  56 static nir_ssa_def *
  57 vc4_nir_unpack_8i(nir_builder *b, nir_ssa_def *src, unsigned chan)
  58 {
  59         return nir_ubitfield_extract(b,
  60                                      src,
  61                                      nir_imm_int(b, 8 * chan),
  62                                      nir_imm_int(b, 8));
  63 }
  64
  65 /** Returns the 16 bit field as a sign-extended 32-bit value. */
  66 static nir_ssa_def *
  67 vc4_nir_unpack_16i(nir_builder *b, nir_ssa_def *src, unsigned chan)
  68 {
  69         return nir_ibitfield_extract(b,
  70                                      src,
  71                                      nir_imm_int(b, 16 * chan),
  72                                      nir_imm_int(b, 16));
  73 }
  74
  75 /** Returns the 16 bit field as an unsigned 32 bit value. */
  76 static nir_ssa_def *
  77 vc4_nir_unpack_16u(nir_builder *b, nir_ssa_def *src, unsigned chan)
  78 {
  79         if (chan == 0) {
  80                 return nir_iand(b, src, nir_imm_int(b, 0xffff));
  81         } else {
  82                 return nir_ushr(b, src, nir_imm_int(b, 16));
  83         }
  84 }
  85
  86 static nir_ssa_def *
  87 vc4_nir_unpack_8f(nir_builder *b, nir_ssa_def *src, unsigned chan)
  88 {
  89         return nir_channel(b, nir_unpack_unorm_4x8(b, src), chan);
  90 }
  91
  92 static nir_ssa_def *
  93 vc4_nir_get_vattr_channel_vpm(struct vc4_compile *c,
  94                               nir_builder *b,
  95                               nir_ssa_def **vpm_reads,
  96                               uint8_t swiz,
  97                               const struct util_format_description *desc)
  98 {
  99         const struct util_format_channel_description *chan =
 100                 &desc->channel[swiz];
 101         nir_ssa_def *temp;
 102
 103         if (swiz > PIPE_SWIZZLE_W) {
 104                 return vc4_nir_get_swizzled_channel(b, vpm_reads, swiz);
 105         } else if (chan->size == 32 && chan->type == UTIL_FORMAT_TYPE_FLOAT) {
 106                 return vc4_nir_get_swizzled_channel(b, vpm_reads, swiz);
 107         } else if (chan->size == 32 && chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 108                 if (chan->normalized) {
 109                         return nir_fmul(b,
 110                                         nir_i2f32(b, vpm_reads[swiz]),
 111                                         nir_imm_float(b,
 112                                                       1.0 / 0x7fffffff));
 113                 } else {
 114                         return nir_i2f32(b, vpm_reads[swiz]);
 115                 }
 116         } else if (chan->size == 8 &&
 117                    (chan->type == UTIL_FORMAT_TYPE_UNSIGNED ||
 118                     chan->type == UTIL_FORMAT_TYPE_SIGNED)) {
 119                 nir_ssa_def *vpm = vpm_reads[0];
 120                 if (chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 121                         temp = nir_ixor(b, vpm, nir_imm_int(b, 0x80808080));
 122                         if (chan->normalized) {
 123                                 return nir_fsub(b, nir_fmul(b,
 124                                                             vc4_nir_unpack_8f(b, temp, swiz),
 125                                                             nir_imm_float(b, 2.0)),
 126                                                 nir_imm_float(b, 1.0));
 127                         } else {
 128                                 return nir_fadd(b,
 129                                                 nir_i2f32(b,
 130                                                           vc4_nir_unpack_8i(b, temp,
 131                                                                             swiz)),
 132                                                 nir_imm_float(b, -128.0));
 133                         }
 134                 } else {
 135                         if (chan->normalized) {
 136                                 return vc4_nir_unpack_8f(b, vpm, swiz);
 137                         } else {
 138                                 return nir_i2f32(b, vc4_nir_unpack_8i(b, vpm, swiz));
 139                         }
 140                 }
 141         } else if (chan->size == 16 &&
 142                    (chan->type == UTIL_FORMAT_TYPE_UNSIGNED ||
 143                     chan->type == UTIL_FORMAT_TYPE_SIGNED)) {
 144                 nir_ssa_def *vpm = vpm_reads[swiz / 2];
 145
 146                 /* Note that UNPACK_16F eats a half float, not ints, so we use
 147                  * UNPACK_16_I for all of these.
 148                  */
 149                 if (chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 150                         temp = nir_i2f32(b, vc4_nir_unpack_16i(b, vpm, swiz & 1));
 151                         if (chan->normalized) {
 152                                 return nir_fmul(b, temp,
 153                                                 nir_imm_float(b, 1/32768.0f));
 154                         } else {
 155                                 return temp;
 156                         }
 157                 } else {
 158                         temp = nir_i2f32(b, vc4_nir_unpack_16u(b, vpm, swiz & 1));
 159                         if (chan->normalized) {
 160                                 return nir_fmul(b, temp,
 161                                                 nir_imm_float(b, 1 / 65535.0));
 162                         } else {
 163                                 return temp;
 164                         }
 165                 }
 166         } else {
 167                 return NULL;
 168         }
 169 }
 170
 171 static void
 172 vc4_nir_lower_vertex_attr(struct vc4_compile *c, nir_builder *b,
 173                           nir_intrinsic_instr *intr)
 174 {
 175         b->cursor = nir_before_instr(&intr->instr);
 176
 177         int attr = nir_intrinsic_base(intr);
 178         enum pipe_format format = c->vs_key->attr_formats[attr];
 179         uint32_t attr_size = util_format_get_blocksize(format);
 180
 181         /* We only accept direct outputs and TGSI only ever gives them to us
 182          * with an offset value of 0.
 183          */
 184         assert(nir_src_as_uint(intr->src[0]) == 0);
 185
 186         /* Generate dword loads for the VPM values (Since these intrinsics may
 187          * be reordered, the actual reads will be generated at the top of the
 188          * shader by ntq_setup_inputs().
 189          */
 190         nir_ssa_def *vpm_reads[4];
 191         for (int i = 0; i < align(attr_size, 4) / 4; i++) {
 192                 nir_intrinsic_instr *intr_comp =
 193                         nir_intrinsic_instr_create(c->s,
 194                                                    nir_intrinsic_load_input);
 195                 intr_comp->num_components = 1;
 196                 nir_intrinsic_set_base(intr_comp, nir_intrinsic_base(intr));
 197                 nir_intrinsic_set_component(intr_comp, i);
 198                 intr_comp->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
 199                 nir_ssa_dest_init(&intr_comp->instr, &intr_comp->dest, 1, 32, NULL);
 200                 nir_builder_instr_insert(b, &intr_comp->instr);
 201
 202                 vpm_reads[i] = &intr_comp->dest.ssa;
 203         }
 204
 205         bool format_warned = false;
 206         const struct util_format_description *desc =
 207                 util_format_description(format);
 208
 209         nir_ssa_def *dests[4];
 210         for (int i = 0; i < intr->num_components; i++) {
 211                 uint8_t swiz = desc->swizzle[i];
 212                 dests[i] = vc4_nir_get_vattr_channel_vpm(c, b, vpm_reads, swiz,
 213                                                          desc);
 214
 215                 if (!dests[i]) {
 216                         if (!format_warned) {
 217                                 fprintf(stderr,
 218                                         "vtx element %d unsupported type: %s\n",
 219                                         attr, util_format_name(format));
 220                                 format_warned = true;
 221                         }
 222                         dests[i] = nir_imm_float(b, 0.0);
 223                 }
 224         }
 225
 226         replace_intrinsic_with_vec(b, intr, dests);
 227 }
 228
 229 static void
 230 vc4_nir_lower_fs_input(struct vc4_compile *c, nir_builder *b,
 231                        nir_intrinsic_instr *intr)
 232 {
 233         b->cursor = nir_after_instr(&intr->instr);
 234
 235         if (nir_intrinsic_base(intr) >= VC4_NIR_TLB_COLOR_READ_INPUT &&
 236             nir_intrinsic_base(intr) < (VC4_NIR_TLB_COLOR_READ_INPUT +
 237                                         VC4_MAX_SAMPLES)) {
 238                 /* This doesn't need any lowering. */
 239                 return;
 240         }
 241
 242         nir_variable *input_var =
 243                 nir_find_variable_with_driver_location(c->s, nir_var_shader_in,
 244                                                        nir_intrinsic_base(intr));
 245         assert(input_var);
 246
 247         int comp = nir_intrinsic_component(intr);
 248
 249         /* Lower away point coordinates, and fix up PNTC. */
 250         if (util_varying_is_point_coord(input_var->data.location,
 251                                         c->fs_key->point_sprite_mask)) {
 252                 assert(intr->num_components == 1);
 253
 254                 nir_ssa_def *result = &intr->dest.ssa;
 255
 256                 switch (comp) {
 257                 case 0:
 258                 case 1:
 259                         /* If we're not rendering points, we need to set a
 260                          * defined value for the input that would come from
 261                          * PNTC.
 262                          */
 263                         if (!c->fs_key->is_points)
 264                                 result = nir_imm_float(b, 0.0);
 265                         break;
 266                 case 2:
 267                         result = nir_imm_float(b, 0.0);
 268                         break;
 269                 case 3:
 270                         result = nir_imm_float(b, 1.0);
 271                         break;
 272                 }
 273
 274                 if (c->fs_key->point_coord_upper_left && comp == 1)
 275                         result = nir_fsub(b, nir_imm_float(b, 1.0), result);
 276
 277                 if (result != &intr->dest.ssa) {
 278                         nir_ssa_def_rewrite_uses_after(&intr->dest.ssa,
 279                                                        nir_src_for_ssa(result),
 280                                                        result->parent_instr);
 281                 }
 282         }
 283 }
 284
 285 static void
 286 vc4_nir_lower_output(struct vc4_compile *c, nir_builder *b,
 287                      nir_intrinsic_instr *intr)
 288 {
 289         nir_variable *output_var =
 290                 nir_find_variable_with_driver_location(c->s, nir_var_shader_out,
 291                                                        nir_intrinsic_base(intr));
 292         assert(output_var);
 293
 294         if (c->stage == QSTAGE_COORD &&
 295             output_var->data.location != VARYING_SLOT_POS &&
 296             output_var->data.location != VARYING_SLOT_PSIZ) {
 297                 nir_instr_remove(&intr->instr);
 298                 return;
 299         }
 300 }
 301
 302 static void
 303 vc4_nir_lower_uniform(struct vc4_compile *c, nir_builder *b,
 304                       nir_intrinsic_instr *intr)
 305 {
 306         b->cursor = nir_before_instr(&intr->instr);
 307
 308         /* Generate scalar loads equivalent to the original vector. */
 309         nir_ssa_def *dests[4];
 310         for (unsigned i = 0; i < intr->num_components; i++) {
 311                 nir_intrinsic_instr *intr_comp =
 312                         nir_intrinsic_instr_create(c->s, intr->intrinsic);
 313                 intr_comp->num_components = 1;
 314                 nir_ssa_dest_init(&intr_comp->instr, &intr_comp->dest, 1,
 315                                   intr->dest.ssa.bit_size, NULL);
 316
 317                 /* Convert the uniform offset to bytes.  If it happens
 318                  * to be a constant, constant-folding will clean up
 319                  * the shift for us.
 320                  */
 321                 nir_intrinsic_set_base(intr_comp,
 322                                        nir_intrinsic_base(intr) * 16 +
 323                                        i * 4);
 324                 nir_intrinsic_set_range(intr_comp,
 325                                         nir_intrinsic_range(intr) * 16 - i * 4);
 326
 327                 intr_comp->src[0] =
 328                         nir_src_for_ssa(nir_ishl(b, intr->src[0].ssa,
 329                                                  nir_imm_int(b, 4)));
 330
 331                 dests[i] = &intr_comp->dest.ssa;
 332
 333                 nir_builder_instr_insert(b, &intr_comp->instr);
 334         }
 335
 336         replace_intrinsic_with_vec(b, intr, dests);
 337 }
 338
 339 static void
 340 vc4_nir_lower_io_instr(struct vc4_compile *c, nir_builder *b,
 341                        struct nir_instr *instr)
 342 {
 343         if (instr->type != nir_instr_type_intrinsic)
 344                 return;
 345         nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 346
 347         switch (intr->intrinsic) {
 348         case nir_intrinsic_load_input:
 349                 if (c->stage == QSTAGE_FRAG)
 350                         vc4_nir_lower_fs_input(c, b, intr);
 351                 else
 352                         vc4_nir_lower_vertex_attr(c, b, intr);
 353                 break;
 354
 355         case nir_intrinsic_store_output:
 356                 vc4_nir_lower_output(c, b, intr);
 357                 break;
 358
 359         case nir_intrinsic_load_uniform:
 360                 vc4_nir_lower_uniform(c, b, intr);
 361                 break;
 362
 363         case nir_intrinsic_load_user_clip_plane:
 364         default:
 365                 break;
 366         }
 367 }
 368
 369 static bool
 370 vc4_nir_lower_io_impl(struct vc4_compile *c, nir_function_impl *impl)
 371 {
 372         nir_builder b;
 373         nir_builder_init(&b, impl);
 374
 375         nir_foreach_block(block, impl) {
 376                 nir_foreach_instr_safe(instr, block)
 377                         vc4_nir_lower_io_instr(c, &b, instr);
 378         }
 379
 380         nir_metadata_preserve(impl, nir_metadata_block_index |
 381                               nir_metadata_dominance);
 382
 383         return true;
 384 }
 385
 386 void
 387 vc4_nir_lower_io(nir_shader *s, struct vc4_compile *c)
 388 {
 389         nir_foreach_function(function, s) {
 390                 if (function->impl)
 391                         vc4_nir_lower_io_impl(c, function->impl);
 392         }
 393 }