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 /* Tests all 64 combinations of floating point modifiers for a given
 194  * instruction / floating-type / test type */
 195
 196 static void
 197 bit_fmod_helper(struct panfrost_device *dev,
 198                 enum bi_class c, unsigned size, bool fma,
 199                 uint32_t *input, enum bit_debug debug, unsigned op)
 200 {
 201         bi_instruction ins = bit_ins(c, 2, nir_type_float, size);
 202
 203         bool fp16 = (size == 16);
 204         bool has_outmods = fma || !fp16;
 205
 206         for (unsigned outmod = 0; outmod < (has_outmods ? 4 : 1); ++outmod) {
 207                 for (unsigned inmod = 0; inmod < 16; ++inmod) {
 208                         ins.outmod = outmod;
 209                         ins.op.minmax = op;
 210                         ins.src_abs[0] = (inmod & 0x1);
 211                         ins.src_abs[1] = (inmod & 0x2);
 212                         ins.src_neg[0] = (inmod & 0x4);
 213                         ins.src_neg[1] = (inmod & 0x8);
 214
 215                         /* Skip over tests that cannot run */
 216                         if ((fma || c == BI_MINMAX) && fp16 && ins.src_abs[0] && ins.src_abs[1])
 217                                 continue;
 218
 219                         if (!bit_test_single(dev, &ins, input, fma, debug)) {
 220                                 fprintf(stderr, "FAIL: fmod.%s%u.%s%s.%u\n",
 221                                                 bi_class_name(c),
 222                                                 size,
 223                                                 fma ? "fma" : "add",
 224                                                 outmod ? bi_output_mod_name(outmod) : ".none",
 225                                                 inmod);
 226                         }
 227                 }
 228         }
 229 }
 230
 231 static void
 232 bit_fma_helper(struct panfrost_device *dev,
 233                 unsigned size, uint32_t *input, enum bit_debug debug)
 234 {
 235         bi_instruction ins = bit_ins(BI_FMA, 3, nir_type_float, size);
 236
 237         for (unsigned outmod = 0; outmod < 4; ++outmod) {
 238                 for (unsigned inmod = 0; inmod < 8; ++inmod) {
 239                         ins.outmod = outmod;
 240                         ins.src_neg[0] = (inmod & 0x1);
 241                         ins.src_neg[1] = (inmod & 0x2);
 242                         ins.src_neg[2] = (inmod & 0x4);
 243
 244                         if (!bit_test_single(dev, &ins, input, true, debug)) {
 245                                 fprintf(stderr, "FAIL: fma%u%s.%u\n",
 246                                                 size,
 247                                                 outmod ? bi_output_mod_name(outmod) : ".none",
 248                                                 inmod);
 249                         }
 250                 }
 251         }
 252 }
 253
 254 static void
 255 bit_fma_mscale_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 256 {
 257         bi_instruction ins = bit_ins(BI_FMA, 4, nir_type_float, 32);
 258         ins.op.mscale = true;
 259         ins.src_types[3] = nir_type_int32;
 260         ins.src[2] = ins.src[3]; /* Not enough ports! */
 261
 262         for (unsigned outmod = 0; outmod < 4; ++outmod) {
 263                 for (unsigned inmod = 0; inmod < 8; ++inmod) {
 264                         ins.outmod = outmod;
 265                         ins.src_abs[0] = (inmod & 0x1);
 266                         ins.src_neg[1] = (inmod & 0x2);
 267                         ins.src_neg[2] = (inmod & 0x4);
 268
 269                         if (!bit_test_single(dev, &ins, input, true, debug)) {
 270                                 fprintf(stderr, "FAIL: fma_mscale%s.%u\n",
 271                                                 outmod ? bi_output_mod_name(outmod) : ".none",
 272                                                 inmod);
 273                         }
 274                 }
 275         }
 276 }
 277
 278 static void
 279 bit_csel_helper(struct panfrost_device *dev,
 280                 unsigned size, uint32_t *input, enum bit_debug debug)
 281 {
 282         bi_instruction ins = bit_ins(BI_CSEL, 4, nir_type_uint, size);
 283
 284         /* SCHEDULER: We can only read 3 registers at once. */
 285         ins.src[2] = ins.src[0];
 286
 287         for (enum bi_cond cond = BI_COND_LT; cond <= BI_COND_NE; ++cond) {
 288                 ins.csel_cond = cond;
 289
 290                 if (!bit_test_single(dev, &ins, input, true, debug)) {
 291                         fprintf(stderr, "FAIL: csel%u.%s\n",
 292                                         size, bi_cond_name(cond));
 293                 }
 294         }
 295 }
 296
 297 static void
 298 bit_special_helper(struct panfrost_device *dev,
 299                 unsigned size, uint32_t *input, enum bit_debug debug)
 300 {
 301         bi_instruction ins = bit_ins(BI_SPECIAL, 2, nir_type_float, size);
 302         uint32_t exp_input[4];
 303
 304         for (enum bi_special_op op = BI_SPECIAL_FRCP; op <= BI_SPECIAL_EXP2_LOW; ++op) {
 305                 if (op == BI_SPECIAL_EXP2_LOW) {
 306                         /* exp2 only supported in fp32 mode */
 307                         if (size != 32)
 308                                 continue;
 309
 310                         /* Give expected input */
 311                         exp_input[1] = input[0];
 312                         float *ff = (float *) input;
 313                         exp_input[0] = (int) (ff[0] * (1 << 24));
 314                 }
 315
 316                 for (unsigned c = 0; c < ((size == 16) ? 2 : 1); ++c) {
 317                         ins.op.special = op;
 318                         ins.swizzle[0][0] = c;
 319
 320                         if (!bit_test_single(dev, &ins,
 321                                                 op == BI_SPECIAL_EXP2_LOW ? exp_input : input,
 322                                                 false, debug)) {
 323                                 fprintf(stderr, "FAIL: special%u.%s\n",
 324                                                 size, bi_special_op_name(op));
 325                         }
 326                 }
 327         }
 328 }
 329
 330 static void
 331 bit_table_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 332 {
 333         bi_instruction ins = bit_ins(BI_TABLE, 1, nir_type_float, 32);
 334
 335         for (enum bi_table_op op = 0; op <= BI_TABLE_LOG2_U_OVER_U_1_LOW; ++op) {
 336                 ins.op.table = op;
 337
 338                 if (!bit_test_single(dev, &ins, input, false, debug)) {
 339                         fprintf(stderr, "FAIL: table.%s\n",
 340                                         bi_table_op_name(op));
 341                 }
 342         }
 343 }
 344
 345 static void
 346 bit_frexp_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 347 {
 348         bi_instruction ins = bit_ins(BI_FREXP, 1, nir_type_float, 32);
 349         ins.dest_type = nir_type_int32;
 350
 351         for (enum bi_frexp_op op = 0; op <= BI_FREXPE_LOG; ++op) {
 352                 ins.op.frexp = op;
 353
 354                 if (!bit_test_single(dev, &ins, input, true, debug)) {
 355                         fprintf(stderr, "FAIL: frexp.%s\n",
 356                                         bi_frexp_op_name(op));
 357                 }
 358         }
 359 }
 360
 361 static void
 362 bit_reduce_helper(struct panfrost_device *dev, uint32_t *input, enum bit_debug debug)
 363 {
 364         bi_instruction ins = bit_ins(BI_REDUCE_FMA, 2, nir_type_float, 32);
 365
 366         for (enum bi_reduce_op op = 0; op <= BI_REDUCE_ADD_FREXPM; ++op) {
 367                 ins.op.reduce = op;
 368
 369                 if (!bit_test_single(dev, &ins, input, true, debug)) {
 370                         fprintf(stderr, "FAIL: reduce.%s\n",
 371                                         bi_reduce_op_name(op));
 372                 }
 373         }
 374 }
 375
 376 static void
 377 bit_convert_helper(struct panfrost_device *dev, unsigned from_size,
 378                 unsigned to_size, unsigned cx, unsigned cy, bool FMA,
 379                 enum bifrost_roundmode roundmode,
 380                 uint32_t *input, enum bit_debug debug)
 381 {
 382         bi_instruction ins = {
 383                 .type = BI_CONVERT,
 384                 .dest = BIR_INDEX_REGISTER | 0,
 385                 .src = { BIR_INDEX_REGISTER | 0 }
 386         };
 387
 388         nir_alu_type Ts[3] = { nir_type_float, nir_type_uint, nir_type_int };
 389
 390         for (unsigned from_base = 0; from_base < 3; ++from_base) {
 391                 for (unsigned to_base = 0; to_base < 3; ++to_base) {
 392                         /* Discard invalid combinations.. */
 393                         if ((from_size == to_size) && (from_base == to_base))
 394                                 continue;
 395
 396                         /* Can't switch signedness */
 397                         if (from_base && to_base)
 398                                 continue;
 399
 400                         /* No F16_TO_I32, etc */
 401                         if (from_size != to_size && from_base == 0 && to_base)
 402                                 continue;
 403
 404                         if (from_size != to_size && from_base && to_base == 0)
 405                                 continue;
 406
 407                         /* No need, just ignore the upper half */
 408                         if (from_size > to_size && from_base == to_base && from_base)
 409                                 continue;
 410
 411                         ins.dest_type = Ts[to_base] | to_size;
 412                         ins.src_types[0] = Ts[from_base] | from_size;
 413                         ins.roundmode = roundmode;
 414                         ins.swizzle[0][0] = cx;
 415                         ins.swizzle[0][1] = cy;
 416
 417                         if (!bit_test_single(dev, &ins, input, FMA, debug)) {
 418                                 fprintf(stderr, "FAIL: convert.%u-%u.%u-%u.%u%u\n",
 419                                                 from_base, from_size,
 420                                                 to_base, to_size,
 421                                                 cx, cy);
 422                         }
 423                 }
 424         }
 425 }
 426
 427 static void
 428 bit_constant_helper(struct panfrost_device *dev,
 429                 uint32_t *input, enum bit_debug debug)
 430 {
 431         enum bi_class C[3] = { BI_MOV, BI_ADD, BI_FMA };
 432
 433         for (unsigned doubled = 0; doubled < 2; ++doubled) {
 434                 for (unsigned count = 1; count <= 3; ++count) {
 435                         bi_instruction ins = bit_ins(C[count - 1], count, nir_type_float, 32);
 436
 437                         ins.src[0] = BIR_INDEX_CONSTANT | 0;
 438                         ins.src[1] = (count >= 2) ? BIR_INDEX_CONSTANT | (doubled ? 32 : 0) : 0;
 439                         ins.src[2] = (count >= 3) ? BIR_INDEX_ZERO : 0;
 440
 441                         ins.constant.u64 = doubled ?
 442                                 0x3f800000ull | (0x3f000000ull << 32ull) :
 443                                 0x3f800000ull;
 444
 445                         if (!bit_test_single(dev, &ins, input, true, debug)) {
 446                                 fprintf(stderr, "FAIL: constants.%s.%u\n",
 447                                                 doubled ? "two" : "one",
 448                                                 count);
 449                         }
 450                 }
 451         }
 452 }
 453
 454 void
 455 bit_packing(struct panfrost_device *dev, enum bit_debug debug)
 456 {
 457         float input32[4];
 458         uint16_t input16[8];
 459
 460         bit_generate_float4(input32);
 461         bit_generate_half8(input16);
 462
 463         bit_constant_helper(dev, (uint32_t *) input32, debug);
 464
 465         for (unsigned sz = 16; sz <= 32; sz *= 2) {
 466                 uint32_t *input =
 467                         (sz == 16) ? (uint32_t *) input16 :
 468                         (uint32_t *) input32;
 469
 470                 bit_fmod_helper(dev, BI_ADD, sz, true, input, debug, 0);
 471                 bit_fmod_helper(dev, BI_ADD, sz, false, input, debug, 0);
 472
 473                 bit_fmod_helper(dev, BI_MINMAX, sz, false, input, debug, BI_MINMAX_MIN);
 474                 bit_fmod_helper(dev, BI_MINMAX, sz, false, input, debug, BI_MINMAX_MAX);
 475
 476                 bit_fma_helper(dev, sz, input, debug);
 477         }
 478
 479         for (unsigned sz = 32; sz <= 32; sz *= 2)
 480                 bit_csel_helper(dev, sz, (uint32_t *) input32, debug);
 481
 482         float special[4] = { 0.9 };
 483         uint32_t special16[4] = { _mesa_float_to_half(special[0]) | (_mesa_float_to_half(0.2) << 16) };
 484
 485         bit_table_helper(dev, (uint32_t *) special, debug);
 486
 487         for (unsigned sz = 16; sz <= 32; sz *= 2) {
 488                 uint32_t *input =
 489                         (sz == 16) ? special16 :
 490                         (uint32_t *) special;
 491
 492                 bit_special_helper(dev, sz, input, debug);
 493         }
 494
 495         for (unsigned rm = 0; rm < 4; ++rm) {
 496                 bit_convert_helper(dev, 32, 32, 0, 0, false, rm, (uint32_t *) input32, debug);
 497
 498                 for (unsigned c = 0; c < 2; ++c)
 499                         bit_convert_helper(dev, 32, 16, c, 0, false, rm, (uint32_t *) input32, debug);
 500
 501                 bit_convert_helper(dev, 16, 32, 0, 0, false, rm, (uint32_t *) input16, debug);
 502
 503                 for (unsigned c = 0; c < 4; ++c)
 504                         bit_convert_helper(dev, 16, 16, c & 1, c >> 1, false, rm, (uint32_t *) input16, debug);
 505         }
 506
 507         bit_frexp_helper(dev, (uint32_t *) input32, debug);
 508         bit_reduce_helper(dev, (uint32_t *) input32, debug);
 509
 510         uint32_t mscale_input[4];
 511         memcpy(mscale_input, input32, sizeof(input32));
 512         mscale_input[3] = 0x7;
 513         bit_fma_mscale_helper(dev, mscale_input, debug);
 514 }