src/panfrost/midgard/midgard_emit.c

   1 /*
   2  * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
   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
  24 #include "compiler.h"
  25 #include "midgard_ops.h"
  26
  27 /* Midgard IR only knows vector ALU types, but we sometimes need to actually
  28  * use scalar ALU instructions, for functional or performance reasons. To do
  29  * this, we just demote vector ALU payloads to scalar. */
  30
  31 static int
  32 component_from_mask(unsigned mask)
  33 {
  34         for (int c = 0; c < 8; ++c) {
  35                 if (mask & (1 << c))
  36                         return c;
  37         }
  38
  39         assert(0);
  40         return 0;
  41 }
  42
  43 static unsigned
  44 vector_to_scalar_source(unsigned u, bool is_int, bool is_full,
  45                 unsigned component)
  46 {
  47         midgard_vector_alu_src v;
  48         memcpy(&v, &u, sizeof(v));
  49
  50         /* TODO: Integers */
  51
  52         midgard_scalar_alu_src s = { 0 };
  53
  54         if (is_full) {
  55                 /* For a 32-bit op, just check the source half flag */
  56                 s.full = !v.half;
  57         } else if (!v.half) {
  58                 /* For a 16-bit op that's not subdivided, never full */
  59                 s.full = false;
  60         } else {
  61                 /* We can't do 8-bit scalar, abort! */
  62                 assert(0);
  63         }
  64
  65         /* Component indexing takes size into account */
  66
  67         if (s.full)
  68                 s.component = component << 1;
  69         else
  70                 s.component = component;
  71
  72         if (is_int) {
  73                 /* TODO */
  74         } else {
  75                 s.abs = v.mod & MIDGARD_FLOAT_MOD_ABS;
  76                 s.negate = v.mod & MIDGARD_FLOAT_MOD_NEG;
  77         }
  78
  79         unsigned o;
  80         memcpy(&o, &s, sizeof(s));
  81
  82         return o & ((1 << 6) - 1);
  83 }
  84
  85 static midgard_scalar_alu
  86 vector_to_scalar_alu(midgard_vector_alu v, midgard_instruction *ins)
  87 {
  88         bool is_int = midgard_is_integer_op(v.op);
  89         bool is_full = v.reg_mode == midgard_reg_mode_32;
  90         bool is_inline_constant = ins->has_inline_constant;
  91
  92         unsigned comp = component_from_mask(ins->mask);
  93
  94         /* The output component is from the mask */
  95         midgard_scalar_alu s = {
  96                 .op = v.op,
  97                 .src1 = vector_to_scalar_source(v.src1, is_int, is_full, ins->swizzle[0][comp]),
  98                 .src2 = !is_inline_constant ? vector_to_scalar_source(v.src2, is_int, is_full, ins->swizzle[1][comp]) : 0,
  99                 .unknown = 0,
 100                 .outmod = v.outmod,
 101                 .output_full = is_full,
 102                 .output_component = comp
 103         };
 104
 105         /* Full components are physically spaced out */
 106         if (is_full) {
 107                 assert(s.output_component < 4);
 108                 s.output_component <<= 1;
 109         }
 110
 111         /* Inline constant is passed along rather than trying to extract it
 112          * from v */
 113
 114         if (ins->has_inline_constant) {
 115                 uint16_t imm = 0;
 116                 int lower_11 = ins->inline_constant & ((1 << 12) - 1);
 117                 imm |= (lower_11 >> 9) & 3;
 118                 imm |= (lower_11 >> 6) & 4;
 119                 imm |= (lower_11 >> 2) & 0x38;
 120                 imm |= (lower_11 & 63) << 6;
 121
 122                 s.src2 = imm;
 123         }
 124
 125         return s;
 126 }
 127
 128 /* 64-bit swizzles are super easy since there are 2 components of 2 components
 129  * in an 8-bit field ... lots of duplication to go around!
 130  *
 131  * Swizzles of 32-bit vectors accessed from 64-bit instructions are a little
 132  * funny -- pack them *as if* they were native 64-bit, using rep_* flags to
 133  * flag upper. For instance, xy would become 64-bit XY but that's just xyzw
 134  * native. Likewise, zz would become 64-bit XX with rep* so it would be xyxy
 135  * with rep. Pretty nifty, huh? */
 136
 137 static unsigned
 138 mir_pack_swizzle_64(unsigned *swizzle, unsigned max_component)
 139 {
 140         unsigned packed = 0;
 141
 142         for (unsigned i = 0; i < 2; ++i) {
 143                 assert(swizzle[i] <= max_component);
 144
 145                 unsigned a = (swizzle[i] & 1) ?
 146                         (COMPONENT_W << 2) | COMPONENT_Z :
 147                         (COMPONENT_Y << 2) | COMPONENT_X;
 148
 149                 packed |= a << (i * 4);
 150         }
 151
 152         return packed;
 153 }
 154
 155 static void
 156 mir_pack_mask_alu(midgard_instruction *ins)
 157 {
 158         unsigned effective = ins->mask;
 159
 160         /* If we have a destination override, we need to figure out whether to
 161          * override to the lower or upper half, shifting the effective mask in
 162          * the latter, so AAAA.... becomes AAAA */
 163
 164         unsigned upper_shift = mir_upper_override(ins);
 165
 166         if (upper_shift) {
 167                 effective >>= upper_shift;
 168                 ins->alu.dest_override = midgard_dest_override_upper;
 169         }
 170
 171         if (ins->alu.reg_mode == midgard_reg_mode_32)
 172                 ins->alu.mask = expand_writemask(effective, 4);
 173         else if (ins->alu.reg_mode == midgard_reg_mode_64)
 174                 ins->alu.mask = expand_writemask(effective, 2);
 175         else
 176                 ins->alu.mask = effective;
 177 }
 178
 179 static void
 180 mir_pack_swizzle_alu(midgard_instruction *ins)
 181 {
 182         midgard_vector_alu_src src[] = {
 183                 vector_alu_from_unsigned(ins->alu.src1),
 184                 vector_alu_from_unsigned(ins->alu.src2)
 185         };
 186
 187         for (unsigned i = 0; i < 2; ++i) {
 188                 unsigned packed = 0;
 189
 190                 if (ins->alu.reg_mode == midgard_reg_mode_64) {
 191                         unsigned sz = nir_alu_type_get_type_size(ins->src_types[i]);
 192                         unsigned components = 64 / sz;
 193
 194                         packed = mir_pack_swizzle_64(ins->swizzle[i], components);
 195
 196                         if (sz == 32) {
 197                                 bool lo = ins->swizzle[i][0] >= COMPONENT_Z;
 198                                 bool hi = ins->swizzle[i][1] >= COMPONENT_Z;
 199                                 unsigned mask = mir_bytemask(ins);
 200
 201                                 if (mask & 0xFF) {
 202                                         /* We can't mix halves... */
 203                                         if (mask & 0xFF00)
 204                                                 assert(lo == hi);
 205
 206                                         src[i].rep_low |= lo;
 207                                 } else {
 208                                         src[i].rep_low |= hi;
 209                                 }
 210                         } else if (sz < 32) {
 211                                 unreachable("Cannot encode 8/16 swizzle in 64-bit");
 212                         }
 213                 } else {
 214                         /* For 32-bit, swizzle packing is stupid-simple. For 16-bit,
 215                          * the strategy is to check whether the nibble we're on is
 216                          * upper or lower. We need all components to be on the same
 217                          * "side"; that much is enforced by the ISA and should have
 218                          * been lowered. TODO: 8-bit packing. TODO: vec8 */
 219
 220                         unsigned first = ins->mask ? ffs(ins->mask) - 1 : 0;
 221                         bool upper = ins->swizzle[i][first] > 3;
 222
 223                         if (upper && ins->mask)
 224                                 assert(nir_alu_type_get_type_size(ins->src_types[i]) <= 16);
 225
 226                         for (unsigned c = 0; c < 4; ++c) {
 227                                 unsigned v = ins->swizzle[i][c];
 228
 229                                 bool t_upper = v > 3;
 230
 231                                 /* Ensure we're doing something sane */
 232
 233                                 if (ins->mask & (1 << c)) {
 234                                         assert(t_upper == upper);
 235                                         assert(v <= 7);
 236                                 }
 237
 238                                 /* Use the non upper part */
 239                                 v &= 0x3;
 240
 241                                 packed |= v << (2 * c);
 242                         }
 243
 244                         src[i].rep_high = upper;
 245
 246                         /* Replicate for now.. should really pick a side for
 247                          * dot products */
 248
 249                         if (ins->alu.reg_mode == midgard_reg_mode_16)
 250                                 src[i].rep_low = true;
 251                 }
 252
 253                 src[i].swizzle = packed;
 254         }
 255
 256         ins->alu.src1 = vector_alu_srco_unsigned(src[0]);
 257
 258         if (!ins->has_inline_constant)
 259                 ins->alu.src2 = vector_alu_srco_unsigned(src[1]);
 260 }
 261
 262 static void
 263 mir_pack_swizzle_ldst(midgard_instruction *ins)
 264 {
 265         /* TODO: non-32-bit, non-vec4 */
 266         for (unsigned c = 0; c < 4; ++c) {
 267                 unsigned v = ins->swizzle[0][c];
 268
 269                 /* Check vec4 */
 270                 assert(v <= 3);
 271
 272                 ins->load_store.swizzle |= v << (2 * c);
 273         }
 274
 275         /* TODO: arg_1/2 */
 276 }
 277
 278 static void
 279 mir_pack_swizzle_tex(midgard_instruction *ins)
 280 {
 281         for (unsigned i = 0; i < 2; ++i) {
 282                 unsigned packed = 0;
 283
 284                 for (unsigned c = 0; c < 4; ++c) {
 285                         unsigned v = ins->swizzle[i][c];
 286
 287                         /* Check vec4 */
 288                         assert(v <= 3);
 289
 290                         packed |= v << (2 * c);
 291                 }
 292
 293                 if (i == 0)
 294                         ins->texture.swizzle = packed;
 295                 else
 296                         ins->texture.in_reg_swizzle = packed;
 297         }
 298
 299         /* TODO: bias component */
 300 }
 301
 302 /* Load store masks are 4-bits. Load/store ops pack for that. vec4 is the
 303  * natural mask width; vec8 is constrained to be in pairs, vec2 is duplicated. TODO: 8-bit?
 304  */
 305
 306 static void
 307 mir_pack_ldst_mask(midgard_instruction *ins)
 308 {
 309         unsigned sz = nir_alu_type_get_type_size(ins->dest_type);
 310         unsigned packed = ins->mask;
 311
 312         if (sz == 64) {
 313                 packed = ((ins->mask & 0x2) ? (0x8 | 0x4) : 0) |
 314                          ((ins->mask & 0x1) ? (0x2 | 0x1) : 0);
 315         } else if (sz == 16) {
 316                 packed = 0;
 317
 318                 for (unsigned i = 0; i < 4; ++i) {
 319                         /* Make sure we're duplicated */
 320                         bool u = (ins->mask & (1 << (2*i + 0))) != 0;
 321                         bool v = (ins->mask & (1 << (2*i + 1))) != 0;
 322                         assert(u == v);
 323
 324                         packed |= (u << i);
 325                 }
 326         } else {
 327                 assert(sz == 32);
 328         }
 329
 330         ins->load_store.mask = packed;
 331 }
 332
 333 static void
 334 mir_lower_inverts(midgard_instruction *ins)
 335 {
 336         bool inv[3] = {
 337                 ins->src_invert[0],
 338                 ins->src_invert[1],
 339                 ins->src_invert[2]
 340         };
 341
 342         switch (ins->alu.op) {
 343         case midgard_alu_op_iand:
 344                 /* a & ~b = iandnot(a, b) */
 345                 /* ~a & ~b = ~(a | b) = inor(a, b) */
 346
 347                 if (inv[0] && inv[1])
 348                         ins->alu.op = midgard_alu_op_inor;
 349                 else if (inv[1])
 350                         ins->alu.op = midgard_alu_op_iandnot;
 351
 352                 break;
 353         case midgard_alu_op_ior:
 354                 /*  a | ~b = iornot(a, b) */
 355                 /* ~a | ~b = ~(a & b) = inand(a, b) */
 356
 357                 if (inv[0] && inv[1])
 358                         ins->alu.op = midgard_alu_op_inand;
 359                 else if (inv[1])
 360                         ins->alu.op = midgard_alu_op_iornot;
 361
 362                 break;
 363
 364         case midgard_alu_op_ixor:
 365                 /* ~a ^ b = a ^ ~b = ~(a ^ b) = inxor(a, b) */
 366                 /* ~a ^ ~b = a ^ b */
 367
 368                 if (inv[0] ^ inv[1])
 369                         ins->alu.op = midgard_alu_op_inxor;
 370
 371                 break;
 372
 373         default:
 374                 break;
 375         }
 376 }
 377
 378 static void
 379 emit_alu_bundle(compiler_context *ctx,
 380                 midgard_bundle *bundle,
 381                 struct util_dynarray *emission,
 382                 unsigned lookahead)
 383 {
 384         /* Emit the control word */
 385         util_dynarray_append(emission, uint32_t, bundle->control | lookahead);
 386
 387         /* Next up, emit register words */
 388         for (unsigned i = 0; i < bundle->instruction_count; ++i) {
 389                 midgard_instruction *ins = bundle->instructions[i];
 390
 391                 /* Check if this instruction has registers */
 392                 if (ins->compact_branch) continue;
 393
 394                 /* Otherwise, just emit the registers */
 395                 uint16_t reg_word = 0;
 396                 memcpy(&reg_word, &ins->registers, sizeof(uint16_t));
 397                 util_dynarray_append(emission, uint16_t, reg_word);
 398         }
 399
 400         /* Now, we emit the body itself */
 401         for (unsigned i = 0; i < bundle->instruction_count; ++i) {
 402                 midgard_instruction *ins = bundle->instructions[i];
 403
 404                 /* Where is this body */
 405                 unsigned size = 0;
 406                 void *source = NULL;
 407
 408                 /* In case we demote to a scalar */
 409                 midgard_scalar_alu scalarized;
 410
 411                 if (!ins->compact_branch)
 412                         mir_lower_inverts(ins);
 413
 414                 if (ins->unit & UNITS_ANY_VECTOR) {
 415                         mir_pack_mask_alu(ins);
 416                         mir_pack_swizzle_alu(ins);
 417                         size = sizeof(midgard_vector_alu);
 418                         source = &ins->alu;
 419                 } else if (ins->unit == ALU_ENAB_BR_COMPACT) {
 420                         size = sizeof(midgard_branch_cond);
 421                         source = &ins->br_compact;
 422                 } else if (ins->compact_branch) { /* misnomer */
 423                         size = sizeof(midgard_branch_extended);
 424                         source = &ins->branch_extended;
 425                 } else {
 426                         size = sizeof(midgard_scalar_alu);
 427                         scalarized = vector_to_scalar_alu(ins->alu, ins);
 428                         source = &scalarized;
 429                 }
 430
 431                 memcpy(util_dynarray_grow_bytes(emission, size, 1), source, size);
 432         }
 433
 434         /* Emit padding (all zero) */
 435         memset(util_dynarray_grow_bytes(emission, bundle->padding, 1), 0, bundle->padding);
 436
 437         /* Tack on constants */
 438
 439         if (bundle->has_embedded_constants)
 440                 util_dynarray_append(emission, midgard_constants, bundle->constants);
 441 }
 442
 443 /* Shift applied to the immediate used as an offset. Probably this is papering
 444  * over some other semantic distinction else well, but it unifies things in the
 445  * compiler so I don't mind. */
 446
 447 static unsigned
 448 mir_ldst_imm_shift(midgard_load_store_op op)
 449 {
 450         if (OP_IS_UBO_READ(op))
 451                 return 3;
 452         else
 453                 return 1;
 454 }
 455
 456 static enum mali_sampler_type
 457 midgard_sampler_type(nir_alu_type t) {
 458         switch (nir_alu_type_get_base_type(t))
 459         {
 460         case nir_type_float:
 461                 return MALI_SAMPLER_FLOAT;
 462         case nir_type_int:
 463                 return MALI_SAMPLER_SIGNED;
 464         case nir_type_uint:
 465                 return MALI_SAMPLER_UNSIGNED;
 466         default:
 467                 unreachable("Unknown sampler type");
 468         }
 469 }
 470
 471 /* After everything is scheduled, emit whole bundles at a time */
 472
 473 void
 474 emit_binary_bundle(compiler_context *ctx,
 475                    midgard_bundle *bundle,
 476                    struct util_dynarray *emission,
 477                    int next_tag)
 478 {
 479         int lookahead = next_tag << 4;
 480
 481         switch (bundle->tag) {
 482         case TAG_ALU_4:
 483         case TAG_ALU_8:
 484         case TAG_ALU_12:
 485         case TAG_ALU_16:
 486         case TAG_ALU_4 + 4:
 487         case TAG_ALU_8 + 4:
 488         case TAG_ALU_12 + 4:
 489         case TAG_ALU_16 + 4:
 490                 emit_alu_bundle(ctx, bundle, emission, lookahead);
 491                 break;
 492
 493         case TAG_LOAD_STORE_4: {
 494                 /* One or two composing instructions */
 495
 496                 uint64_t current64, next64 = LDST_NOP;
 497
 498                 /* Copy masks */
 499
 500                 for (unsigned i = 0; i < bundle->instruction_count; ++i) {
 501                         mir_pack_ldst_mask(bundle->instructions[i]);
 502
 503                         mir_pack_swizzle_ldst(bundle->instructions[i]);
 504
 505                         /* Apply a constant offset */
 506                         unsigned offset = bundle->instructions[i]->constants.u32[0];
 507
 508                         if (offset) {
 509                                 unsigned shift = mir_ldst_imm_shift(bundle->instructions[i]->load_store.op);
 510                                 unsigned upper_shift = 10 - shift;
 511
 512                                 bundle->instructions[i]->load_store.varying_parameters |= (offset & ((1 << upper_shift) - 1)) << shift;
 513                                 bundle->instructions[i]->load_store.address |= (offset >> upper_shift);
 514                         }
 515                 }
 516
 517                 memcpy(&current64, &bundle->instructions[0]->load_store, sizeof(current64));
 518
 519                 if (bundle->instruction_count == 2)
 520                         memcpy(&next64, &bundle->instructions[1]->load_store, sizeof(next64));
 521
 522                 midgard_load_store instruction = {
 523                         .type = bundle->tag,
 524                         .next_type = next_tag,
 525                         .word1 = current64,
 526                         .word2 = next64
 527                 };
 528
 529                 util_dynarray_append(emission, midgard_load_store, instruction);
 530
 531                 break;
 532         }
 533
 534         case TAG_TEXTURE_4:
 535         case TAG_TEXTURE_4_VTX:
 536         case TAG_TEXTURE_4_BARRIER: {
 537                 /* Texture instructions are easy, since there is no pipelining
 538                  * nor VLIW to worry about. We may need to set .cont/.last
 539                  * flags. */
 540
 541                 midgard_instruction *ins = bundle->instructions[0];
 542
 543                 ins->texture.type = bundle->tag;
 544                 ins->texture.next_type = next_tag;
 545                 ins->texture.mask = ins->mask;
 546                 mir_pack_swizzle_tex(ins);
 547
 548                 unsigned osz = nir_alu_type_get_type_size(ins->dest_type);
 549                 unsigned isz = nir_alu_type_get_type_size(ins->src_types[1]);
 550
 551                 assert(osz == 32 || osz == 16);
 552                 assert(isz == 32 || isz == 16);
 553
 554                 ins->texture.out_full = (osz == 32);
 555                 ins->texture.in_reg_full = (isz == 32);
 556                 ins->texture.sampler_type = midgard_sampler_type(ins->dest_type);
 557
 558                 if (mir_op_computes_derivatives(ctx->stage, ins->texture.op)) {
 559                         ins->texture.cont = !ins->helper_terminate;
 560                         ins->texture.last = ins->helper_terminate || ins->helper_execute;
 561                 } else {
 562                         ins->texture.cont = ins->texture.last = 1;
 563                 }
 564
 565                 util_dynarray_append(emission, midgard_texture_word, ins->texture);
 566                 break;
 567         }
 568
 569         default:
 570                 unreachable("Unknown midgard instruction type\n");
 571         }
 572 }