src/panfrost/bifrost/test/bi_test_pack.c

   1 /*
   2  * Copyright (C) 2020 Collabora Ltd.
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors (Collabora):
  24  *      Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  25  */
  26
  27 #include "bit.h"
  28 #include "bi_print.h"
  29 #include "util/half_float.h"
  30 #include "bifrost/disassemble.h"
  31
  32 /* Instruction packing tests */
  33
  34 static bool
  35 bit_test_single(struct panfrost_device *dev,
  36                 bi_instruction *ins,
  37                 uint32_t input[4],
  38                 bool fma, enum bit_debug debug)
  39 {
  40         /* First, simulate the instruction */
  41         struct bit_state s = { 0 };
  42         memcpy(s.r, input, 16);
  43         bit_step(&s, ins, fma);
  44
  45         /* Next, wrap it up and pack it */
  46
  47         bi_instruction ldubo = {
  48                 .type = BI_LOAD_UNIFORM,
  49                 .src = {
  50                         BIR_INDEX_CONSTANT,
  51                         BIR_INDEX_ZERO
  52                 },
  53                 .src_types = {
  54                         nir_type_uint32,
  55                         nir_type_uint32,
  56                 },
  57                 .dest = BIR_INDEX_REGISTER | 0,
  58                 .dest_type = nir_type_uint32,
  59                 .vector_channels = 4,
  60         };
  61
  62         bi_instruction ldva = {
  63                 .type = BI_LOAD_VAR_ADDRESS,
  64                 .vector_channels = 3,
  65                 .dest = BIR_INDEX_REGISTER | 32,
  66                 .dest_type = nir_type_uint32,
  67                 .src = {
  68                         BIR_INDEX_CONSTANT,
  69                         BIR_INDEX_REGISTER | 61,
  70                         BIR_INDEX_REGISTER | 62,
  71                         0,
  72                 },
  73                 .src_types = {
  74                         nir_type_uint32,
  75                         nir_type_uint32,
  76                         nir_type_uint32,
  77                         nir_type_uint32,
  78                 }
  79         };
  80
  81         bi_instruction st = {
  82                 .type = BI_STORE_VAR,
  83                 .src = {
  84                         BIR_INDEX_REGISTER | 0,
  85                         ldva.dest, ldva.dest + 1, ldva.dest + 2,
  86                 },
  87                 .src_types = {
  88                         nir_type_uint32,
  89                         nir_type_uint32, nir_type_uint32, nir_type_uint32,
  90                 },
  91                 .vector_channels = 4
  92         };
  93
  94         bi_context *ctx = rzalloc(NULL, bi_context);
  95         ctx->stage = MESA_SHADER_VERTEX;
  96
  97         bi_block *blk = rzalloc(ctx, bi_block);
  98         blk->scheduled = true;
  99
 100         blk->base.predecessors = _mesa_set_create(blk,
 101                         _mesa_hash_pointer,
 102                         _mesa_key_pointer_equal);
 103
 104         list_inithead(&ctx->blocks);
 105         list_addtail(&blk->base.link, &ctx->blocks);
 106         list_inithead(&blk->clauses);
 107
 108         bi_clause *clauses[4] = {
 109                 rzalloc(ctx, bi_clause),
 110                 rzalloc(ctx, bi_clause),
 111                 rzalloc(ctx, bi_clause),
 112                 rzalloc(ctx, bi_clause)
 113         };
 114
 115         for (unsigned i = 0; i < 4; ++i) {
 116                 clauses[i]->bundle_count = 1;
 117                 list_addtail(&clauses[i]->link, &blk->clauses);
 118                 clauses[i]->scoreboard_id = (i & 1);
 119
 120                 if (i) {
 121                         clauses[i]->dependencies = 1 << (~i & 1);
 122                         clauses[i]->data_register_write_barrier = true;
 123                 }
 124         }
 125
 126         clauses[0]->bundles[0].add = &ldubo;
 127         clauses[0]->clause_type = BIFROST_CLAUSE_UBO;
 128
 129         if (fma)
 130                 clauses[1]->bundles[0].fma = ins;
 131         else
 132                 clauses[1]->bundles[0].add = ins;
 133
 134         clauses[0]->constant_count = 1;
 135         clauses[1]->constant_count = 1;
 136         clauses[1]->constants[0] = ins->constant.u64;
 137
 138         clauses[2]->bundles[0].add = &ldva;
 139         clauses[3]->bundles[0].add = &st;
 140
 141         clauses[2]->clause_type = BIFROST_CLAUSE_UBO;
 142         clauses[3]->clause_type = BIFROST_CLAUSE_SSBO_STORE;
 143
 144         panfrost_program prog;
 145         bi_pack(ctx, &prog.compiled);
 146
 147         bool succ = bit_vertex(dev, prog, input, 16, NULL, 0,
 148                         s.r, 16, debug);
 149
 150         if (debug >= BIT_DEBUG_ALL || (!succ && debug >= BIT_DEBUG_FAIL)) {
 151                 bi_print_shader(ctx, stderr);
 152                 disassemble_bifrost(stderr, prog.compiled.data, prog.compiled.size, true);
 153         }
 154
 155         return succ;
 156 }
 157
 158 /* Utilities for generating tests */
 159
 160 static void
 161 bit_generate_float4(float *mem)
 162 {
 163         for (unsigned i = 0; i < 4; ++i)
 164                 mem[i] = (float) ((rand() & 255) - 127) / 16.0;
 165 }
 166
 167 static void
 168 bit_generate_half8(uint16_t *mem)
 169 {
 170         for (unsigned i = 0; i < 8; ++i)
 171                 mem[i] = _mesa_float_to_half(((float) (rand() & 255) - 127) / 16.0);
 172 }
 173
 174 static bi_instruction
 175 bit_ins(enum bi_class C, unsigned argc, nir_alu_type base, unsigned size)
 176 {
 177         nir_alu_type T = base | size;
 178
 179         bi_instruction ins = {
 180                 .type = C,
 181                 .dest = BIR_INDEX_REGISTER | 0,
 182                 .dest_type = T,
 183         };
 184
 185         for (unsigned i = 0; i < argc; ++i) {
 186                 ins.src[i] = BIR_INDEX_REGISTER | i;
 187                 ins.src_types[i] = T;
 188         }
 189
 190         return ins;
 191 }
 192
 193 #define BIT_FOREACH_SWIZZLE(swz, args, sz) \
 194         for (unsigned swz = 0; swz < ((sz == 16) ? (1 << (2 * args)) : 1); ++swz)
 195
 196 static void
 197 bit_apply_swizzle(bi_instruction *ins, unsigned swz, unsigned args, unsigned sz)
 198 {
 199         unsigned slots_per_arg = (sz == 16) ? 4 : 1;
 200         unsigned slots_per_chan = (sz == 16) ? 1 : 0;
 201         unsigned mask = (sz == 16) ? 1 : 0;
 202
 203         for (unsigned i = 0; i < args; ++i) {
 204                 for (unsigned j = 0; j < (32 / sz); ++j) {
 205                         ins->swizzle[i][j] = ((swz >> (slots_per_arg * i)) >> (slots_per_chan * j)) & mask;
 206                 }
 207         }
 208 }
 209
 210 /* Tests all 64 combinations of floating point modifiers for a given
 211  * instruction / floating-type / test type */
 212
 213 static void
 214 bit_fmod_helper(struct panfrost_device *dev,
 215                 enum bi_class c, unsigned size, bool fma,
 216                 uint32_t *input, enum bit_debug debug, unsigned op)
 217 {
 218         bi_instruction ins = bit_ins(c, 2, nir_type_float, size);
 219
 220         bool fp16 = (size == 16);
 221         bool has_outmods = fma || !fp16;
 222
 223         for (unsigned outmod = 0; outmod < (has_outmods ? 4 : 1); ++outmod) {
 224         BIT_FOREACH_SWIZZLE(swz, 2, size) {
 225                 for (unsigned inmod = 0; inmod < 16; ++inmod) {
 226                         ins.outmod = outmod;
 227                         ins.op.minmax = op;
 228                         ins.src_abs[0] = (inmod & 0x1);
 229                         ins.src_abs[1] = (inmod & 0x2);
 230                         ins.src_neg[0] = (inmod & 0x4);
 231                         ins.src_neg[1] = (inmod & 0x8);
 232                         bit_apply_swizzle(&ins, swz, 2, size);
 233
 234                         if (!bit_test_single(dev, &ins, input, fma, debug)) {
 235                                 fprintf(stderr, "FAIL: fmod.%s%u.%s%s.%u\n",
 236                                                 bi_class_name(c),
 237                                                 size,
 238                                                 fma ? "fma" : "add",
 239                                                 outmod ? bi_output_mod_name(outmod) : ".none",
 240                                                 inmod);
 241                         }
 242                 }
 243         }
 244         }
 245 }
 246
 247 static void
 248 bit_fma_helper(struct panfrost_device *dev,
 249                 unsigned size, uint32_t *input, enum bit_debug debug)
 250 {
 251         bi_instruction ins = bit_ins(BI_FMA, 3, nir_type_float, size);
 252
 253         for (unsigned outmod = 0; outmod < 4; ++outmod) {
 254                 for (unsigned inmod = 0; inmod < 8; ++inmod) {
 255                         ins.outmod = outmod;
 256                         ins.src_neg[0] = (inmod & 0x1);
 257                         ins.src_neg[1] = (inmod & 0x2);
 258                         ins.src_neg[2] = (inmod & 0x4);
 259
 260                         if (!bit_test_single(dev, &ins, input, true, debug)) {
 261                                 fprintf(stderr, "FAIL: fma%u%s.%u\n",
 262                                                 size,
 263                                                 outmod ? bi_output_mod_name(outmod) : ".none",
 264                                                 inmod);
 265                         }
 266                 }
 267         }
 268 }
 269
 270 static void
 271 bit_fma_mscale_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 272 {
 273         bi_instruction ins = bit_ins(BI_FMA, 4, nir_type_float, 32);
 274         ins.op.mscale = true;
 275         ins.src_types[3] = nir_type_int32;
 276         ins.src[2] = ins.src[3]; /* Not enough ports! */
 277
 278         for (unsigned outmod = 0; outmod < 4; ++outmod) {
 279                 for (unsigned inmod = 0; inmod < 8; ++inmod) {
 280                         ins.outmod = outmod;
 281                         ins.src_abs[0] = (inmod & 0x1);
 282                         ins.src_neg[1] = (inmod & 0x2);
 283                         ins.src_neg[2] = (inmod & 0x4);
 284
 285                         if (!bit_test_single(dev, &ins, input, true, debug)) {
 286                                 fprintf(stderr, "FAIL: fma_mscale%s.%u\n",
 287                                                 outmod ? bi_output_mod_name(outmod) : ".none",
 288                                                 inmod);
 289                         }
 290                 }
 291         }
 292 }
 293
 294 static void
 295 bit_csel_helper(struct panfrost_device *dev,
 296                 unsigned size, uint32_t *input, enum bit_debug debug)
 297 {
 298         bi_instruction ins = bit_ins(BI_CSEL, 4, nir_type_uint, size);
 299
 300         /* SCHEDULER: We can only read 3 registers at once. */
 301         ins.src[2] = ins.src[0];
 302
 303         for (enum bi_cond cond = BI_COND_LT; cond <= BI_COND_NE; ++cond) {
 304                 ins.cond = cond;
 305
 306                 if (!bit_test_single(dev, &ins, input, true, debug)) {
 307                         fprintf(stderr, "FAIL: csel%u.%s\n",
 308                                         size, bi_cond_name(cond));
 309                 }
 310         }
 311 }
 312
 313 static void
 314 bit_special_helper(struct panfrost_device *dev,
 315                 unsigned size, uint32_t *input, enum bit_debug debug)
 316 {
 317         bi_instruction ins = bit_ins(BI_SPECIAL, 2, nir_type_float, size);
 318         uint32_t exp_input[4];
 319
 320         for (enum bi_special_op op = BI_SPECIAL_FRCP; op <= BI_SPECIAL_EXP2_LOW; ++op) {
 321                 if (op == BI_SPECIAL_EXP2_LOW) {
 322                         /* exp2 only supported in fp32 mode */
 323                         if (size != 32)
 324                                 continue;
 325
 326                         /* Give expected input */
 327                         exp_input[1] = input[0];
 328                         float *ff = (float *) input;
 329                         exp_input[0] = (int) (ff[0] * (1 << 24));
 330                 }
 331
 332                 for (unsigned c = 0; c < ((size == 16) ? 2 : 1); ++c) {
 333                         ins.op.special = op;
 334                         ins.swizzle[0][0] = c;
 335
 336                         if (!bit_test_single(dev, &ins,
 337                                                 op == BI_SPECIAL_EXP2_LOW ? exp_input : input,
 338                                                 false, debug)) {
 339                                 fprintf(stderr, "FAIL: special%u.%s\n",
 340                                                 size, bi_special_op_name(op));
 341                         }
 342                 }
 343         }
 344 }
 345
 346 static void
 347 bit_table_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 348 {
 349         bi_instruction ins = bit_ins(BI_TABLE, 1, nir_type_float, 32);
 350
 351         for (enum bi_table_op op = 0; op <= BI_TABLE_LOG2_U_OVER_U_1_LOW; ++op) {
 352                 ins.op.table = op;
 353
 354                 if (!bit_test_single(dev, &ins, input, false, debug)) {
 355                         fprintf(stderr, "FAIL: table.%s\n",
 356                                         bi_table_op_name(op));
 357                 }
 358         }
 359 }
 360
 361 static void
 362 bit_frexp_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 363 {
 364         bi_instruction ins = bit_ins(BI_FREXP, 1, nir_type_float, 32);
 365         ins.dest_type = nir_type_int32;
 366
 367         for (enum bi_frexp_op op = 0; op <= BI_FREXPE_LOG; ++op) {
 368                 ins.op.frexp = op;
 369
 370                 if (!bit_test_single(dev, &ins, input, true, debug)) {
 371                         fprintf(stderr, "FAIL: frexp.%s\n",
 372                                         bi_frexp_op_name(op));
 373                 }
 374         }
 375 }
 376
 377 static void
 378 bit_round_helper(struct panfrost_device *dev, uint32_t *input, unsigned sz, bool FMA, enum bit_debug debug)
 379 {
 380         bi_instruction ins = bit_ins(BI_ROUND, 1, nir_type_float, sz);
 381
 382         for (enum bifrost_roundmode mode = 0; mode <= 3; ++mode) {
 383                 for (unsigned swizzle = 0; swizzle < (sz == 16 ? 4 : 1); ++swizzle) {
 384                         if (sz == 16) {
 385                                 for (unsigned i = 0; i < 2; ++i)
 386                                         ins.swizzle[0][i] = ((swizzle >> i) & 1) ? 1 : 0;
 387                         }
 388
 389                         ins.roundmode = mode;
 390
 391                         if (!bit_test_single(dev, &ins, input, FMA, debug)) {
 392                                 fprintf(stderr, "FAIL: round.%u.%u\n",
 393                                                 sz, mode);
 394                         }
 395                 }
 396         }
 397 }
 398
 399 static void
 400 bit_reduce_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 401 {
 402         bi_instruction ins = bit_ins(BI_REDUCE_FMA, 2, nir_type_float, 32);
 403
 404         for (enum bi_reduce_op op = 0; op <= BI_REDUCE_ADD_FREXPM; ++op) {
 405                 ins.op.reduce = op;
 406
 407                 if (!bit_test_single(dev, &ins, input, true, debug)) {
 408                         fprintf(stderr, "FAIL: reduce.%s\n",
 409                                         bi_reduce_op_name(op));
 410                 }
 411         }
 412 }
 413
 414 static void
 415 bit_select_helper(struct panfrost_device *dev, uint32_t *input, unsigned size, enum bit_debug debug)
 416 {
 417         unsigned C = 32 / size;
 418         bi_instruction ins = bit_ins(BI_SELECT, C, nir_type_uint, 32);
 419
 420         for (unsigned c = 0; c < C; ++c)
 421                 ins.src_types[c] = nir_type_uint | size;
 422
 423         if (size == 8) {
 424                 /* SCHEDULER: We can only read 3 registers at once. */
 425                 ins.src[2] = ins.src[0];
 426         }
 427
 428         /* Each argument has swizzle {lo, hi} so 2^C options */
 429         unsigned hi = (size == 16) ? 1 : 2;
 430
 431         for (unsigned add = 0; add < ((size == 16) ? 2 : 1); ++add) {
 432                 for (unsigned swizzle = 0; swizzle < (1 << C); ++swizzle) {
 433                         for (unsigned i = 0; i < C; ++i)
 434                                 ins.swizzle[i][0] = ((swizzle >> i) & 1) ? hi : 0;
 435
 436                         if (!bit_test_single(dev, &ins, input, !add, debug)) {
 437                                 fprintf(stderr, "FAIL: select.%u.%u\n",
 438                                                 size, swizzle);
 439                         }
 440                 }
 441         }
 442 }
 443
 444 static void
 445 bit_fcmp_helper(struct panfrost_device *dev, uint32_t *input, unsigned size, enum bit_debug debug, bool FMA)
 446 {
 447         bi_instruction ins = bit_ins(BI_CMP, 2, nir_type_float, size);
 448         ins.dest_type = nir_type_uint | size;
 449
 450         /* 16-bit has swizzles and abs. 32-bit has abs/neg mods. */
 451         unsigned max_mods = (size == 16) ? 64 : (size == 32) ? 16 : 1;
 452
 453         for (enum bi_cond cond = BI_COND_LT; cond <= BI_COND_NE; ++cond) {
 454                 for (unsigned mods = 0; mods < max_mods; ++mods) {
 455                         ins.cond = cond;
 456
 457                         if (size == 16) {
 458                                 for (unsigned i = 0; i < 2; ++i) {
 459                                         ins.swizzle[i][0] = ((mods >> (i * 2)) & 1) ? 1 : 0;
 460                                         ins.swizzle[i][1] = ((mods >> (i * 2)) & 2) ? 1 : 0;
 461                                 }
 462
 463                                 ins.src_abs[0] = (mods & 16) ? true : false;
 464                                 ins.src_abs[1] = (mods & 32) ? true : false;
 465                         } else if (size == 8) {
 466                                 for (unsigned i = 0; i < 2; ++i) {
 467                                         for (unsigned j = 0; j < 4; ++j)
 468                                                 ins.swizzle[i][j] = j;
 469                                 }
 470                         } else if (size == 32) {
 471                                 ins.src_abs[0] = (mods & 1) ? true : false;
 472                                 ins.src_abs[1] = (mods & 2) ? true : false;
 473                                 ins.src_neg[0] = (mods & 4) ? true : false;
 474                                 ins.src_neg[1] = (mods & 8) ? true : false;
 475                         }
 476
 477                         if (!bit_test_single(dev, &ins, input, FMA, debug)) {
 478                                 fprintf(stderr, "FAIL: cmp.%s.%u.%u.%u\n",
 479                                                 FMA ? "fma" : "add", size, mods, cond);
 480                         }
 481                 }
 482         }
 483 }
 484
 485 static void
 486 bit_icmp_helper(struct panfrost_device *dev, uint32_t *input, unsigned size, nir_alu_type T, enum bit_debug debug)
 487 {
 488         bi_instruction ins = bit_ins(BI_CMP, 2, T, size);
 489         ins.dest_type = nir_type_uint | size;
 490
 491         for (enum bi_cond cond = BI_COND_LT; cond <= BI_COND_NE; ++cond) {
 492         BIT_FOREACH_SWIZZLE(swz, 2, size) {
 493                 ins.cond = cond;
 494                 bit_apply_swizzle(&ins, swz, 2, size);
 495
 496                 if (!bit_test_single(dev, &ins, input, false, debug)) {
 497                         fprintf(stderr, "FAIL: icmp.%u.%u.%u\n",
 498                                         size, swz, cond);
 499                 }
 500         }
 501         }
 502 }
 503
 504
 505
 506 static void
 507 bit_convert_helper(struct panfrost_device *dev, unsigned from_size,
 508                 unsigned to_size, unsigned cx, unsigned cy, bool FMA,
 509                 enum bifrost_roundmode roundmode,
 510                 uint32_t *input, enum bit_debug debug)
 511 {
 512         bi_instruction ins = {
 513                 .type = BI_CONVERT,
 514                 .dest = BIR_INDEX_REGISTER | 0,
 515                 .src = { BIR_INDEX_REGISTER | 0 }
 516         };
 517
 518         nir_alu_type Ts[3] = { nir_type_float, nir_type_uint, nir_type_int };
 519
 520         for (unsigned from_base = 0; from_base < 3; ++from_base) {
 521                 for (unsigned to_base = 0; to_base < 3; ++to_base) {
 522                         /* Discard invalid combinations.. */
 523                         if ((from_size == to_size) && (from_base == to_base))
 524                                 continue;
 525
 526                         /* Can't switch signedness */
 527                         if (from_base && to_base)
 528                                 continue;
 529
 530                         /* No F16_TO_I32, etc */
 531                         if (from_size != to_size && from_base == 0 && to_base)
 532                                 continue;
 533
 534                         if (from_size != to_size && from_base && to_base == 0)
 535                                 continue;
 536
 537                         /* No need, just ignore the upper half */
 538                         if (from_size > to_size && from_base == to_base && from_base)
 539                                 continue;
 540
 541                         ins.dest_type = Ts[to_base] | to_size;
 542                         ins.src_types[0] = Ts[from_base] | from_size;
 543                         ins.roundmode = roundmode;
 544                         ins.swizzle[0][0] = cx;
 545                         ins.swizzle[0][1] = cy;
 546
 547                         if (!bit_test_single(dev, &ins, input, FMA, debug)) {
 548                                 fprintf(stderr, "FAIL: convert.%u-%u.%u-%u.%u%u\n",
 549                                                 from_base, from_size,
 550                                                 to_base, to_size,
 551                                                 cx, cy);
 552                         }
 553                 }
 554         }
 555 }
 556
 557 static void
 558 bit_constant_helper(struct panfrost_device *dev,
 559                 uint32_t *input, enum bit_debug debug)
 560 {
 561         enum bi_class C[3] = { BI_MOV, BI_ADD, BI_FMA };
 562
 563         for (unsigned doubled = 0; doubled < 2; ++doubled) {
 564                 for (unsigned count = 1; count <= 3; ++count) {
 565                         bi_instruction ins = bit_ins(C[count - 1], count, nir_type_float, 32);
 566
 567                         ins.src[0] = BIR_INDEX_CONSTANT | 0;
 568                         ins.src[1] = (count >= 2) ? BIR_INDEX_CONSTANT | (doubled ? 32 : 0) : 0;
 569                         ins.src[2] = (count >= 3) ? BIR_INDEX_ZERO : 0;
 570
 571                         ins.constant.u64 = doubled ?
 572                                 0x3f800000ull | (0x3f000000ull << 32ull) :
 573                                 0x3f800000ull;
 574
 575                         if (!bit_test_single(dev, &ins, input, true, debug)) {
 576                                 fprintf(stderr, "FAIL: constants.%s.%u\n",
 577                                                 doubled ? "two" : "one",
 578                                                 count);
 579                         }
 580                 }
 581         }
 582 }
 583
 584 static void
 585 bit_bitwise_helper(struct panfrost_device *dev, uint32_t *input, unsigned size, enum bit_debug debug)
 586 {
 587         bi_instruction ins = bit_ins(BI_BITWISE, 3, nir_type_uint, size);
 588
 589         /* TODO: shifts */
 590         ins.src[2] = BIR_INDEX_ZERO;
 591
 592         /* Force identity swizzle -- bitwise is not swizzleable */
 593         for (unsigned i = 0; i < 2; ++i) {
 594                 for (unsigned j = 0; j < (32 / size); ++j)
 595                         ins.swizzle[i][j] = j;
 596         }
 597
 598         for (unsigned op = BI_BITWISE_AND; op <= BI_BITWISE_XOR; ++op) {
 599                 ins.op.bitwise = op;
 600
 601                 for (unsigned mods = 0; mods < 4; ++mods) {
 602                         ins.bitwise.src_invert[0] = mods & 1;
 603                         ins.bitwise.src_invert[1] = mods & 2;
 604
 605                         if (!bit_test_single(dev, &ins, input, true, debug)) {
 606                                 fprintf(stderr, "FAIL: bitwise.%u.%u.%u\n",
 607                                                 size, op, mods);
 608                         }
 609                 }
 610         }
 611 }
 612
 613 void
 614 bit_packing(struct panfrost_device *dev, enum bit_debug debug)
 615 {
 616         float input32[4];
 617         uint16_t input16[8];
 618
 619         bit_generate_float4(input32);
 620         bit_generate_half8(input16);
 621
 622         bit_constant_helper(dev, (uint32_t *) input32, debug);
 623
 624         for (unsigned sz = 16; sz <= 32; sz *= 2) {
 625                 uint32_t *input =
 626                         (sz == 16) ? (uint32_t *) input16 :
 627                         (uint32_t *) input32;
 628
 629                 bit_fmod_helper(dev, BI_ADD, sz, true, input, debug, 0);
 630                 bit_fmod_helper(dev, BI_ADD, sz, false, input, debug, 0);
 631                 bit_round_helper(dev, (uint32_t *) input32, sz, true, debug);
 632
 633                 bit_fmod_helper(dev, BI_MINMAX, sz, false, input, debug, BI_MINMAX_MIN);
 634                 bit_fmod_helper(dev, BI_MINMAX, sz, false, input, debug, BI_MINMAX_MAX);
 635
 636                 bit_fma_helper(dev, sz, input, debug);
 637                 bit_icmp_helper(dev, input, sz, nir_type_uint, debug);
 638                 bit_icmp_helper(dev, input, sz, nir_type_int, debug);
 639         }
 640
 641         for (unsigned sz = 32; sz <= 32; sz *= 2)
 642                 bit_csel_helper(dev, sz, (uint32_t *) input32, debug);
 643
 644         float special[4] = { 0.9 };
 645         uint32_t special16[4] = { _mesa_float_to_half(special[0]) | (_mesa_float_to_half(0.2) << 16) };
 646
 647         bit_table_helper(dev, (uint32_t *) special, debug);
 648
 649         for (unsigned sz = 16; sz <= 32; sz *= 2) {
 650                 uint32_t *input =
 651                         (sz == 16) ? special16 :
 652                         (uint32_t *) special;
 653
 654                 bit_special_helper(dev, sz, input, debug);
 655         }
 656
 657         for (unsigned rm = 0; rm < 4; ++rm) {
 658                 bit_convert_helper(dev, 32, 32, 0, 0, false, rm, (uint32_t *) input32, debug);
 659
 660                 for (unsigned c = 0; c < 2; ++c)
 661                         bit_convert_helper(dev, 32, 16, c, 0, false, rm, (uint32_t *) input32, debug);
 662
 663                 bit_convert_helper(dev, 16, 32, 0, 0, false, rm, (uint32_t *) input16, debug);
 664
 665                 for (unsigned c = 0; c < 4; ++c)
 666                         bit_convert_helper(dev, 16, 16, c & 1, c >> 1, false, rm, (uint32_t *) input16, debug);
 667         }
 668
 669         bit_frexp_helper(dev, (uint32_t *) input32, debug);
 670         bit_reduce_helper(dev, (uint32_t *) input32, debug);
 671
 672         uint32_t mscale_input[4];
 673         memcpy(mscale_input, input32, sizeof(input32));
 674         mscale_input[3] = 0x7;
 675         bit_fma_mscale_helper(dev, mscale_input, debug);
 676
 677         for (unsigned sz = 8; sz <= 16; sz *= 2) {
 678                 bit_select_helper(dev, (uint32_t *) input32, sz, debug);
 679         }
 680
 681         bit_fcmp_helper(dev, (uint32_t *) input32, 32, debug, true);
 682         bit_fcmp_helper(dev, (uint32_t *) input32, 16, debug, true);
 683
 684         for (unsigned sz = 8; sz <= 32; sz *= 2)
 685                 bit_bitwise_helper(dev, (uint32_t *) input32, sz, debug);
 686 }