src/gallium/drivers/vc4/vc4_nir_lower_blend.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 /**
  25  * Implements most of the fixed function fragment pipeline in shader code.
  26  *
  27  * VC4 doesn't have any hardware support for blending, alpha test, logic ops,
  28  * or color mask.  Instead, you read the current contents of the destination
  29  * from the tile buffer after having waited for the scoreboard (which is
  30  * handled by vc4_qpu_emit.c), then do math using your output color and that
  31  * destination value, and update the output color appropriately.
  32  *
  33  * Once this pass is done, the color write will either have one component (for
  34  * single sample) with packed argb8888, or 4 components with the per-sample
  35  * argb8888 result.
  36  */
  37
  38 /**
  39  * Lowers fixed-function blending to a load of the destination color and a
  40  * series of ALU operations before the store of the output.
  41  */
  42 #include "util/u_format.h"
  43 #include "vc4_qir.h"
  44 #include "glsl/nir/nir_builder.h"
  45 #include "vc4_context.h"
  46
  47 static bool
  48 blend_depends_on_dst_color(struct vc4_compile *c)
  49 {
  50         return (c->fs_key->blend.blend_enable ||
  51                 c->fs_key->blend.colormask != 0xf ||
  52                 c->fs_key->logicop_func != PIPE_LOGICOP_COPY);
  53 }
  54
  55 /** Emits a load of the previous fragment color from the tile buffer. */
  56 static nir_ssa_def *
  57 vc4_nir_get_dst_color(nir_builder *b, int sample)
  58 {
  59         nir_intrinsic_instr *load =
  60                 nir_intrinsic_instr_create(b->shader,
  61                                            nir_intrinsic_load_input);
  62         load->num_components = 1;
  63         load->const_index[0] = VC4_NIR_TLB_COLOR_READ_INPUT + sample;
  64         load->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
  65         nir_ssa_dest_init(&load->instr, &load->dest, 1, NULL);
  66         nir_builder_instr_insert(b, &load->instr);
  67         return &load->dest.ssa;
  68 }
  69
  70 static  nir_ssa_def *
  71 vc4_nir_srgb_decode(nir_builder *b, nir_ssa_def *srgb)
  72 {
  73         nir_ssa_def *is_low = nir_flt(b, srgb, nir_imm_float(b, 0.04045));
  74         nir_ssa_def *low = nir_fmul(b, srgb, nir_imm_float(b, 1.0 / 12.92));
  75         nir_ssa_def *high = nir_fpow(b,
  76                                      nir_fmul(b,
  77                                               nir_fadd(b, srgb,
  78                                                        nir_imm_float(b, 0.055)),
  79                                               nir_imm_float(b, 1.0 / 1.055)),
  80                                      nir_imm_float(b, 2.4));
  81
  82         return nir_bcsel(b, is_low, low, high);
  83 }
  84
  85 static  nir_ssa_def *
  86 vc4_nir_srgb_encode(nir_builder *b, nir_ssa_def *linear)
  87 {
  88         nir_ssa_def *is_low = nir_flt(b, linear, nir_imm_float(b, 0.0031308));
  89         nir_ssa_def *low = nir_fmul(b, linear, nir_imm_float(b, 12.92));
  90         nir_ssa_def *high = nir_fsub(b,
  91                                      nir_fmul(b,
  92                                               nir_imm_float(b, 1.055),
  93                                               nir_fpow(b,
  94                                                        linear,
  95                                                        nir_imm_float(b, 0.41666))),
  96                                      nir_imm_float(b, 0.055));
  97
  98         return nir_bcsel(b, is_low, low, high);
  99 }
 100
 101 static nir_ssa_def *
 102 vc4_blend_channel_f(nir_builder *b,
 103                     nir_ssa_def **src,
 104                     nir_ssa_def **dst,
 105                     unsigned factor,
 106                     int channel)
 107 {
 108         switch(factor) {
 109         case PIPE_BLENDFACTOR_ONE:
 110                 return nir_imm_float(b, 1.0);
 111         case PIPE_BLENDFACTOR_SRC_COLOR:
 112                 return src[channel];
 113         case PIPE_BLENDFACTOR_SRC_ALPHA:
 114                 return src[3];
 115         case PIPE_BLENDFACTOR_DST_ALPHA:
 116                 return dst[3];
 117         case PIPE_BLENDFACTOR_DST_COLOR:
 118                 return dst[channel];
 119         case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
 120                 if (channel != 3) {
 121                         return nir_fmin(b,
 122                                         src[3],
 123                                         nir_fsub(b,
 124                                                  nir_imm_float(b, 1.0),
 125                                                  dst[3]));
 126                 } else {
 127                         return nir_imm_float(b, 1.0);
 128                 }
 129         case PIPE_BLENDFACTOR_CONST_COLOR:
 130                 return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_X + channel);
 131         case PIPE_BLENDFACTOR_CONST_ALPHA:
 132                 return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_W);
 133         case PIPE_BLENDFACTOR_ZERO:
 134                 return nir_imm_float(b, 0.0);
 135         case PIPE_BLENDFACTOR_INV_SRC_COLOR:
 136                 return nir_fsub(b, nir_imm_float(b, 1.0), src[channel]);
 137         case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
 138                 return nir_fsub(b, nir_imm_float(b, 1.0), src[3]);
 139         case PIPE_BLENDFACTOR_INV_DST_ALPHA:
 140                 return nir_fsub(b, nir_imm_float(b, 1.0), dst[3]);
 141         case PIPE_BLENDFACTOR_INV_DST_COLOR:
 142                 return nir_fsub(b, nir_imm_float(b, 1.0), dst[channel]);
 143         case PIPE_BLENDFACTOR_INV_CONST_COLOR:
 144                 return nir_fsub(b, nir_imm_float(b, 1.0),
 145                                 vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_X + channel));
 146         case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
 147                 return nir_fsub(b, nir_imm_float(b, 1.0),
 148                                 vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_W));
 149
 150         default:
 151         case PIPE_BLENDFACTOR_SRC1_COLOR:
 152         case PIPE_BLENDFACTOR_SRC1_ALPHA:
 153         case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
 154         case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
 155                 /* Unsupported. */
 156                 fprintf(stderr, "Unknown blend factor %d\n", factor);
 157                 return nir_imm_float(b, 1.0);
 158         }
 159 }
 160
 161 static nir_ssa_def *
 162 vc4_nir_set_packed_chan(nir_builder *b, nir_ssa_def *src0, nir_ssa_def *src1,
 163                         int chan)
 164 {
 165         unsigned chan_mask = 0xff << (chan * 8);
 166         return nir_ior(b,
 167                        nir_iand(b, src0, nir_imm_int(b, ~chan_mask)),
 168                        nir_iand(b, src1, nir_imm_int(b, chan_mask)));
 169 }
 170
 171 static nir_ssa_def *
 172 vc4_blend_channel_i(nir_builder *b,
 173                     nir_ssa_def *src,
 174                     nir_ssa_def *dst,
 175                     nir_ssa_def *src_a,
 176                     nir_ssa_def *dst_a,
 177                     unsigned factor,
 178                     int a_chan)
 179 {
 180         switch (factor) {
 181         case PIPE_BLENDFACTOR_ONE:
 182                 return nir_imm_int(b, ~0);
 183         case PIPE_BLENDFACTOR_SRC_COLOR:
 184                 return src;
 185         case PIPE_BLENDFACTOR_SRC_ALPHA:
 186                 return src_a;
 187         case PIPE_BLENDFACTOR_DST_ALPHA:
 188                 return dst_a;
 189         case PIPE_BLENDFACTOR_DST_COLOR:
 190                 return dst;
 191         case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
 192                 return vc4_nir_set_packed_chan(b,
 193                                                nir_umin_4x8(b,
 194                                                             src_a,
 195                                                             nir_inot(b, dst_a)),
 196                                                nir_imm_int(b, ~0),
 197                                                a_chan);
 198         case PIPE_BLENDFACTOR_CONST_COLOR:
 199                 return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_RGBA);
 200         case PIPE_BLENDFACTOR_CONST_ALPHA:
 201                 return vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_AAAA);
 202         case PIPE_BLENDFACTOR_ZERO:
 203                 return nir_imm_int(b, 0);
 204         case PIPE_BLENDFACTOR_INV_SRC_COLOR:
 205                 return nir_inot(b, src);
 206         case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
 207                 return nir_inot(b, src_a);
 208         case PIPE_BLENDFACTOR_INV_DST_ALPHA:
 209                 return nir_inot(b, dst_a);
 210         case PIPE_BLENDFACTOR_INV_DST_COLOR:
 211                 return nir_inot(b, dst);
 212         case PIPE_BLENDFACTOR_INV_CONST_COLOR:
 213                 return nir_inot(b, vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_RGBA));
 214         case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
 215                 return nir_inot(b, vc4_nir_get_state_uniform(b, QUNIFORM_BLEND_CONST_COLOR_AAAA));
 216
 217         default:
 218         case PIPE_BLENDFACTOR_SRC1_COLOR:
 219         case PIPE_BLENDFACTOR_SRC1_ALPHA:
 220         case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
 221         case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
 222                 /* Unsupported. */
 223                 fprintf(stderr, "Unknown blend factor %d\n", factor);
 224                 return nir_imm_int(b, ~0);
 225         }
 226 }
 227
 228 static nir_ssa_def *
 229 vc4_blend_func_f(nir_builder *b, nir_ssa_def *src, nir_ssa_def *dst,
 230                  unsigned func)
 231 {
 232         switch (func) {
 233         case PIPE_BLEND_ADD:
 234                 return nir_fadd(b, src, dst);
 235         case PIPE_BLEND_SUBTRACT:
 236                 return nir_fsub(b, src, dst);
 237         case PIPE_BLEND_REVERSE_SUBTRACT:
 238                 return nir_fsub(b, dst, src);
 239         case PIPE_BLEND_MIN:
 240                 return nir_fmin(b, src, dst);
 241         case PIPE_BLEND_MAX:
 242                 return nir_fmax(b, src, dst);
 243
 244         default:
 245                 /* Unsupported. */
 246                 fprintf(stderr, "Unknown blend func %d\n", func);
 247                 return src;
 248
 249         }
 250 }
 251
 252 static nir_ssa_def *
 253 vc4_blend_func_i(nir_builder *b, nir_ssa_def *src, nir_ssa_def *dst,
 254                  unsigned func)
 255 {
 256         switch (func) {
 257         case PIPE_BLEND_ADD:
 258                 return nir_usadd_4x8(b, src, dst);
 259         case PIPE_BLEND_SUBTRACT:
 260                 return nir_ussub_4x8(b, src, dst);
 261         case PIPE_BLEND_REVERSE_SUBTRACT:
 262                 return nir_ussub_4x8(b, dst, src);
 263         case PIPE_BLEND_MIN:
 264                 return nir_umin_4x8(b, src, dst);
 265         case PIPE_BLEND_MAX:
 266                 return nir_umax_4x8(b, src, dst);
 267
 268         default:
 269                 /* Unsupported. */
 270                 fprintf(stderr, "Unknown blend func %d\n", func);
 271                 return src;
 272
 273         }
 274 }
 275
 276 static void
 277 vc4_do_blending_f(struct vc4_compile *c, nir_builder *b, nir_ssa_def **result,
 278                   nir_ssa_def **src_color, nir_ssa_def **dst_color)
 279 {
 280         struct pipe_rt_blend_state *blend = &c->fs_key->blend;
 281
 282         if (!blend->blend_enable) {
 283                 for (int i = 0; i < 4; i++)
 284                         result[i] = src_color[i];
 285                 return;
 286         }
 287
 288         /* Clamp the src color to [0, 1].  Dest is already clamped. */
 289         for (int i = 0; i < 4; i++)
 290                 src_color[i] = nir_fsat(b, src_color[i]);
 291
 292         nir_ssa_def *src_blend[4], *dst_blend[4];
 293         for (int i = 0; i < 4; i++) {
 294                 int src_factor = ((i != 3) ? blend->rgb_src_factor :
 295                                   blend->alpha_src_factor);
 296                 int dst_factor = ((i != 3) ? blend->rgb_dst_factor :
 297                                   blend->alpha_dst_factor);
 298                 src_blend[i] = nir_fmul(b, src_color[i],
 299                                         vc4_blend_channel_f(b,
 300                                                             src_color, dst_color,
 301                                                             src_factor, i));
 302                 dst_blend[i] = nir_fmul(b, dst_color[i],
 303                                         vc4_blend_channel_f(b,
 304                                                             src_color, dst_color,
 305                                                             dst_factor, i));
 306         }
 307
 308         for (int i = 0; i < 4; i++) {
 309                 result[i] = vc4_blend_func_f(b, src_blend[i], dst_blend[i],
 310                                              ((i != 3) ? blend->rgb_func :
 311                                               blend->alpha_func));
 312         }
 313 }
 314
 315 static nir_ssa_def *
 316 vc4_nir_splat(nir_builder *b, nir_ssa_def *src)
 317 {
 318         nir_ssa_def *or1 = nir_ior(b, src, nir_ishl(b, src, nir_imm_int(b, 8)));
 319         return nir_ior(b, or1, nir_ishl(b, or1, nir_imm_int(b, 16)));
 320 }
 321
 322 static nir_ssa_def *
 323 vc4_do_blending_i(struct vc4_compile *c, nir_builder *b,
 324                   nir_ssa_def *src_color, nir_ssa_def *dst_color,
 325                   nir_ssa_def *src_float_a)
 326 {
 327         struct pipe_rt_blend_state *blend = &c->fs_key->blend;
 328
 329         if (!blend->blend_enable)
 330                 return src_color;
 331
 332         enum pipe_format color_format = c->fs_key->color_format;
 333         const uint8_t *format_swiz = vc4_get_format_swizzle(color_format);
 334         nir_ssa_def *imm_0xff = nir_imm_int(b, 0xff);
 335         nir_ssa_def *src_a = nir_pack_unorm_4x8(b, src_float_a);
 336         nir_ssa_def *dst_a;
 337         int alpha_chan;
 338         for (alpha_chan = 0; alpha_chan < 4; alpha_chan++) {
 339                 if (format_swiz[alpha_chan] == 3)
 340                         break;
 341         }
 342         if (alpha_chan != 4) {
 343                 nir_ssa_def *shift = nir_imm_int(b, alpha_chan * 8);
 344                 dst_a = vc4_nir_splat(b, nir_iand(b, nir_ushr(b, dst_color,
 345                                                               shift), imm_0xff));
 346         } else {
 347                 dst_a = nir_imm_int(b, ~0);
 348         }
 349
 350         nir_ssa_def *src_factor = vc4_blend_channel_i(b,
 351                                                       src_color, dst_color,
 352                                                       src_a, dst_a,
 353                                                       blend->rgb_src_factor,
 354                                                       alpha_chan);
 355         nir_ssa_def *dst_factor = vc4_blend_channel_i(b,
 356                                                       src_color, dst_color,
 357                                                       src_a, dst_a,
 358                                                       blend->rgb_dst_factor,
 359                                                       alpha_chan);
 360
 361         if (alpha_chan != 4 &&
 362             blend->alpha_src_factor != blend->rgb_src_factor) {
 363                 nir_ssa_def *src_alpha_factor =
 364                         vc4_blend_channel_i(b,
 365                                             src_color, dst_color,
 366                                             src_a, dst_a,
 367                                             blend->alpha_src_factor,
 368                                             alpha_chan);
 369                 src_factor = vc4_nir_set_packed_chan(b, src_factor,
 370                                                      src_alpha_factor,
 371                                                      alpha_chan);
 372         }
 373         if (alpha_chan != 4 &&
 374             blend->alpha_dst_factor != blend->rgb_dst_factor) {
 375                 nir_ssa_def *dst_alpha_factor =
 376                         vc4_blend_channel_i(b,
 377                                             src_color, dst_color,
 378                                             src_a, dst_a,
 379                                             blend->alpha_dst_factor,
 380                                             alpha_chan);
 381                 dst_factor = vc4_nir_set_packed_chan(b, dst_factor,
 382                                                      dst_alpha_factor,
 383                                                      alpha_chan);
 384         }
 385         nir_ssa_def *src_blend = nir_umul_unorm_4x8(b, src_color, src_factor);
 386         nir_ssa_def *dst_blend = nir_umul_unorm_4x8(b, dst_color, dst_factor);
 387
 388         nir_ssa_def *result =
 389                 vc4_blend_func_i(b, src_blend, dst_blend, blend->rgb_func);
 390         if (alpha_chan != 4 && blend->alpha_func != blend->rgb_func) {
 391                 nir_ssa_def *result_a = vc4_blend_func_i(b,
 392                                                          src_blend,
 393                                                          dst_blend,
 394                                                          blend->alpha_func);
 395                 result = vc4_nir_set_packed_chan(b, result, result_a,
 396                                                  alpha_chan);
 397         }
 398         return result;
 399 }
 400
 401 static nir_ssa_def *
 402 vc4_logicop(nir_builder *b, int logicop_func,
 403             nir_ssa_def *src, nir_ssa_def *dst)
 404 {
 405         switch (logicop_func) {
 406         case PIPE_LOGICOP_CLEAR:
 407                 return nir_imm_int(b, 0);
 408         case PIPE_LOGICOP_NOR:
 409                 return nir_inot(b, nir_ior(b, src, dst));
 410         case PIPE_LOGICOP_AND_INVERTED:
 411                 return nir_iand(b, nir_inot(b, src), dst);
 412         case PIPE_LOGICOP_COPY_INVERTED:
 413                 return nir_inot(b, src);
 414         case PIPE_LOGICOP_AND_REVERSE:
 415                 return nir_iand(b, src, nir_inot(b, dst));
 416         case PIPE_LOGICOP_INVERT:
 417                 return nir_inot(b, dst);
 418         case PIPE_LOGICOP_XOR:
 419                 return nir_ixor(b, src, dst);
 420         case PIPE_LOGICOP_NAND:
 421                 return nir_inot(b, nir_iand(b, src, dst));
 422         case PIPE_LOGICOP_AND:
 423                 return nir_iand(b, src, dst);
 424         case PIPE_LOGICOP_EQUIV:
 425                 return nir_inot(b, nir_ixor(b, src, dst));
 426         case PIPE_LOGICOP_NOOP:
 427                 return dst;
 428         case PIPE_LOGICOP_OR_INVERTED:
 429                 return nir_ior(b, nir_inot(b, src), dst);
 430         case PIPE_LOGICOP_OR_REVERSE:
 431                 return nir_ior(b, src, nir_inot(b, dst));
 432         case PIPE_LOGICOP_OR:
 433                 return nir_ior(b, src, dst);
 434         case PIPE_LOGICOP_SET:
 435                 return nir_imm_int(b, ~0);
 436         default:
 437                 fprintf(stderr, "Unknown logic op %d\n", logicop_func);
 438                 /* FALLTHROUGH */
 439         case PIPE_LOGICOP_COPY:
 440                 return src;
 441         }
 442 }
 443
 444 static nir_ssa_def *
 445 vc4_nir_pipe_compare_func(nir_builder *b, int func,
 446                           nir_ssa_def *src0, nir_ssa_def *src1)
 447 {
 448         switch (func) {
 449         default:
 450                 fprintf(stderr, "Unknown compare func %d\n", func);
 451                 /* FALLTHROUGH */
 452         case PIPE_FUNC_NEVER:
 453                 return nir_imm_int(b, 0);
 454         case PIPE_FUNC_ALWAYS:
 455                 return nir_imm_int(b, ~0);
 456         case PIPE_FUNC_EQUAL:
 457                 return nir_feq(b, src0, src1);
 458         case PIPE_FUNC_NOTEQUAL:
 459                 return nir_fne(b, src0, src1);
 460         case PIPE_FUNC_GREATER:
 461                 return nir_flt(b, src1, src0);
 462         case PIPE_FUNC_GEQUAL:
 463                 return nir_fge(b, src0, src1);
 464         case PIPE_FUNC_LESS:
 465                 return nir_flt(b, src0, src1);
 466         case PIPE_FUNC_LEQUAL:
 467                 return nir_fge(b, src1, src0);
 468         }
 469 }
 470
 471 static void
 472 vc4_nir_emit_alpha_test_discard(struct vc4_compile *c, nir_builder *b,
 473                                 nir_ssa_def *alpha)
 474 {
 475         if (!c->fs_key->alpha_test)
 476                 return;
 477
 478         nir_ssa_def *alpha_ref =
 479                 vc4_nir_get_state_uniform(b, QUNIFORM_ALPHA_REF);
 480         nir_ssa_def *condition =
 481                 vc4_nir_pipe_compare_func(b, c->fs_key->alpha_test_func,
 482                                           alpha, alpha_ref);
 483
 484         nir_intrinsic_instr *discard =
 485                 nir_intrinsic_instr_create(b->shader,
 486                                            nir_intrinsic_discard_if);
 487         discard->num_components = 1;
 488         discard->src[0] = nir_src_for_ssa(nir_inot(b, condition));
 489         nir_builder_instr_insert(b, &discard->instr);
 490 }
 491
 492 static nir_ssa_def *
 493 vc4_nir_swizzle_and_pack(struct vc4_compile *c, nir_builder *b,
 494                          nir_ssa_def **colors)
 495 {
 496         enum pipe_format color_format = c->fs_key->color_format;
 497         const uint8_t *format_swiz = vc4_get_format_swizzle(color_format);
 498
 499         nir_ssa_def *swizzled[4];
 500         for (int i = 0; i < 4; i++) {
 501                 swizzled[i] = vc4_nir_get_swizzled_channel(b, colors,
 502                                                            format_swiz[i]);
 503         }
 504
 505         return nir_pack_unorm_4x8(b,
 506                                   nir_vec4(b,
 507                                            swizzled[0], swizzled[1],
 508                                            swizzled[2], swizzled[3]));
 509
 510 }
 511
 512 static nir_ssa_def *
 513 vc4_nir_blend_pipeline(struct vc4_compile *c, nir_builder *b, nir_ssa_def *src,
 514                        int sample)
 515 {
 516         enum pipe_format color_format = c->fs_key->color_format;
 517         const uint8_t *format_swiz = vc4_get_format_swizzle(color_format);
 518         bool srgb = util_format_is_srgb(color_format);
 519
 520         /* Pull out the float src/dst color components. */
 521         nir_ssa_def *packed_dst_color = vc4_nir_get_dst_color(b, sample);
 522         nir_ssa_def *dst_vec4 = nir_unpack_unorm_4x8(b, packed_dst_color);
 523         nir_ssa_def *src_color[4], *unpacked_dst_color[4];
 524         for (unsigned i = 0; i < 4; i++) {
 525                 src_color[i] = nir_channel(b, src, i);
 526                 unpacked_dst_color[i] = nir_channel(b, dst_vec4, i);
 527         }
 528
 529         if (c->fs_key->sample_alpha_to_one && c->fs_key->msaa)
 530                 src_color[3] = nir_imm_float(b, 1.0);
 531
 532         vc4_nir_emit_alpha_test_discard(c, b, src_color[3]);
 533
 534         nir_ssa_def *packed_color;
 535         if (srgb) {
 536                 /* Unswizzle the destination color. */
 537                 nir_ssa_def *dst_color[4];
 538                 for (unsigned i = 0; i < 4; i++) {
 539                         dst_color[i] = vc4_nir_get_swizzled_channel(b,
 540                                                                     unpacked_dst_color,
 541                                                                     format_swiz[i]);
 542                 }
 543
 544                 /* Turn dst color to linear. */
 545                 for (int i = 0; i < 3; i++)
 546                         dst_color[i] = vc4_nir_srgb_decode(b, dst_color[i]);
 547
 548                 nir_ssa_def *blend_color[4];
 549                 vc4_do_blending_f(c, b, blend_color, src_color, dst_color);
 550
 551                 /* sRGB encode the output color */
 552                 for (int i = 0; i < 3; i++)
 553                         blend_color[i] = vc4_nir_srgb_encode(b, blend_color[i]);
 554
 555                 packed_color = vc4_nir_swizzle_and_pack(c, b, blend_color);
 556         } else {
 557                 nir_ssa_def *packed_src_color =
 558                         vc4_nir_swizzle_and_pack(c, b, src_color);
 559
 560                 packed_color =
 561                         vc4_do_blending_i(c, b,
 562                                           packed_src_color, packed_dst_color,
 563                                           src_color[3]);
 564         }
 565
 566         packed_color = vc4_logicop(b, c->fs_key->logicop_func,
 567                                    packed_color, packed_dst_color);
 568
 569         /* If the bit isn't set in the color mask, then just return the
 570          * original dst color, instead.
 571          */
 572         uint32_t colormask = 0xffffffff;
 573         for (int i = 0; i < 4; i++) {
 574                 if (format_swiz[i] < 4 &&
 575                     !(c->fs_key->blend.colormask & (1 << format_swiz[i]))) {
 576                         colormask &= ~(0xff << (i * 8));
 577                 }
 578         }
 579
 580         return nir_ior(b,
 581                        nir_iand(b, packed_color,
 582                                 nir_imm_int(b, colormask)),
 583                        nir_iand(b, packed_dst_color,
 584                                 nir_imm_int(b, ~colormask)));
 585 }
 586
 587 static int
 588 vc4_nir_next_output_driver_location(nir_shader *s)
 589 {
 590         int maxloc = -1;
 591
 592         nir_foreach_variable(var, &s->inputs)
 593                 maxloc = MAX2(maxloc, var->data.driver_location);
 594
 595         return maxloc;
 596 }
 597
 598 static void
 599 vc4_nir_store_sample_mask(struct vc4_compile *c, nir_builder *b,
 600                           nir_ssa_def *val)
 601 {
 602         nir_variable *sample_mask = nir_variable_create(c->s, nir_var_shader_out,
 603                                                         glsl_uint_type(),
 604                                                         "sample_mask");
 605         sample_mask->data.driver_location =
 606                 vc4_nir_next_output_driver_location(c->s);
 607         sample_mask->data.location = FRAG_RESULT_SAMPLE_MASK;
 608         exec_list_push_tail(&c->s->outputs, &sample_mask->node);
 609
 610         nir_intrinsic_instr *intr =
 611                 nir_intrinsic_instr_create(c->s, nir_intrinsic_store_output);
 612         intr->num_components = 1;
 613         intr->const_index[0] = sample_mask->data.location;
 614
 615         intr->src[0] = nir_src_for_ssa(val);
 616         intr->src[1] = nir_src_for_ssa(nir_imm_int(b, 0));
 617         nir_builder_instr_insert(b, &intr->instr);
 618 }
 619
 620 static void
 621 vc4_nir_lower_blend_instr(struct vc4_compile *c, nir_builder *b,
 622                           nir_intrinsic_instr *intr)
 623 {
 624         nir_ssa_def *frag_color = intr->src[0].ssa;
 625
 626         if (c->fs_key->sample_coverage) {
 627                 nir_intrinsic_instr *load =
 628                         nir_intrinsic_instr_create(b->shader,
 629                                                    nir_intrinsic_load_sample_mask_in);
 630                 load->num_components = 1;
 631                 nir_ssa_dest_init(&load->instr, &load->dest, 1, NULL);
 632                 nir_builder_instr_insert(b, &load->instr);
 633
 634                 nir_ssa_def *bitmask = &load->dest.ssa;
 635
 636                 vc4_nir_store_sample_mask(c, b, bitmask);
 637         } else if (c->fs_key->sample_alpha_to_coverage) {
 638                 nir_ssa_def *a = nir_channel(b, frag_color, 3);
 639
 640                 /* XXX: We should do a nice dither based on the fragment
 641                  * coordinate, instead.
 642                  */
 643                 nir_ssa_def *num_samples = nir_imm_float(b, VC4_MAX_SAMPLES);
 644                 nir_ssa_def *num_bits = nir_f2i(b, nir_fmul(b, a, num_samples));
 645                 nir_ssa_def *bitmask = nir_isub(b,
 646                                                 nir_ishl(b,
 647                                                          nir_imm_int(b, 1),
 648                                                          num_bits),
 649                                                 nir_imm_int(b, 1));
 650                 vc4_nir_store_sample_mask(c, b, bitmask);
 651         }
 652
 653         /* The TLB color read returns each sample in turn, so if our blending
 654          * depends on the destination color, we're going to have to run the
 655          * blending function separately for each destination sample value, and
 656          * then output the per-sample color using TLB_COLOR_MS.
 657          */
 658         nir_ssa_def *blend_output;
 659         if (c->fs_key->msaa && blend_depends_on_dst_color(c)) {
 660                 c->msaa_per_sample_output = true;
 661
 662                 nir_ssa_def *samples[4];
 663                 for (int i = 0; i < VC4_MAX_SAMPLES; i++)
 664                         samples[i] = vc4_nir_blend_pipeline(c, b, frag_color, i);
 665                 blend_output = nir_vec4(b,
 666                                         samples[0], samples[1],
 667                                         samples[2], samples[3]);
 668         } else {
 669                 blend_output = vc4_nir_blend_pipeline(c, b, frag_color, 0);
 670         }
 671
 672         nir_instr_rewrite_src(&intr->instr, &intr->src[0],
 673                               nir_src_for_ssa(blend_output));
 674         intr->num_components = blend_output->num_components;
 675 }
 676
 677 static bool
 678 vc4_nir_lower_blend_block(nir_block *block, void *state)
 679 {
 680         struct vc4_compile *c = state;
 681
 682         nir_foreach_instr_safe(block, instr) {
 683                 if (instr->type != nir_instr_type_intrinsic)
 684                         continue;
 685                 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 686                 if (intr->intrinsic != nir_intrinsic_store_output)
 687                         continue;
 688
 689                 nir_variable *output_var = NULL;
 690                 nir_foreach_variable(var, &c->s->outputs) {
 691                         if (var->data.driver_location == intr->const_index[0]) {
 692                                 output_var = var;
 693                                 break;
 694                         }
 695                 }
 696                 assert(output_var);
 697
 698                 if (output_var->data.location != FRAG_RESULT_COLOR &&
 699                     output_var->data.location != FRAG_RESULT_DATA0) {
 700                         continue;
 701                 }
 702
 703                 nir_function_impl *impl =
 704                         nir_cf_node_get_function(&block->cf_node);
 705                 nir_builder b;
 706                 nir_builder_init(&b, impl);
 707                 b.cursor = nir_before_instr(&intr->instr);
 708                 vc4_nir_lower_blend_instr(c, &b, intr);
 709         }
 710         return true;
 711 }
 712
 713 void
 714 vc4_nir_lower_blend(struct vc4_compile *c)
 715 {
 716         nir_foreach_overload(c->s, overload) {
 717                 if (overload->impl) {
 718                         nir_foreach_block(overload->impl,
 719                                           vc4_nir_lower_blend_block, c);
 720
 721                         nir_metadata_preserve(overload->impl,
 722                                               nir_metadata_block_index |
 723                                               nir_metadata_dominance);
 724                 }
 725         }
 726 }