src/panfrost/bifrost/bifrost.h

   1 /*
   2  * Copyright (C) 2019 Connor Abbott <cwabbott0@gmail.com>
   3  * Copyright (C) 2019 Lyude Paul <thatslyude@gmail.com>
   4  * Copyright (C) 2019 Ryan Houdek <Sonicadvance1@gmail.com>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  */
  25
  26 #ifndef __bifrost_h__
  27 #define __bifrost_h__
  28
  29 #include <stdint.h>
  30 #include <stdbool.h>
  31
  32 #define BIFROST_DBG_MSGS        0x0001
  33 #define BIFROST_DBG_SHADERS     0x0002
  34
  35 extern int bifrost_debug;
  36
  37 enum bifrost_clause_type {
  38         BIFROST_CLAUSE_NONE       = 0,
  39         BIFROST_CLAUSE_LOAD_VARY  = 1,
  40         BIFROST_CLAUSE_UBO        = 2,
  41         BIFROST_CLAUSE_TEX        = 3,
  42         BIFROST_CLAUSE_SSBO_LOAD  = 5,
  43         BIFROST_CLAUSE_SSBO_STORE = 6,
  44         BIFROST_CLAUSE_BLEND      = 9,
  45         BIFROST_CLAUSE_FRAGZ      = 12,
  46         BIFROST_CLAUSE_ATEST      = 13,
  47         BIFROST_CLAUSE_64BIT      = 15
  48 };
  49
  50 struct bifrost_header {
  51         unsigned unk0 : 7;
  52         // If true, convert any infinite result of any floating-point operation to
  53         // the biggest representable number.
  54         unsigned suppress_inf: 1;
  55         // Convert any NaN results to 0.
  56         unsigned suppress_nan : 1;
  57         unsigned unk1 : 2;
  58         // true if the execution mask of the next clause is the same as the mask of
  59         // the current clause.
  60         unsigned back_to_back : 1;
  61         unsigned no_end_of_shader: 1;
  62         unsigned unk2 : 2;
  63         // Set to true for fragment shaders, to implement this bit of spec text
  64         // from section 7.1.5 of the GLSL ES spec:
  65         //
  66         // "Stores to image and buffer variables performed by helper invocations
  67         // have no effect on the underlying image or buffer memory."
  68         //
  69         // Helper invocations are threads (invocations) corresponding to pixels in
  70         // a quad that aren't actually part of the triangle, but are included to
  71         // make derivatives work correctly. They're usually turned on, but they
  72         // need to be masked off for GLSL-level stores. This bit seems to be the
  73         // only bit that's actually different between fragment shaders and other
  74         // shaders, so this is probably what it's doing.
  75         unsigned elide_writes : 1;
  76         // If backToBack is off:
  77         // - true for conditional branches and fallthrough
  78         // - false for unconditional branches
  79         // The blob seems to always set it to true if back-to-back is on.
  80         unsigned branch_cond : 1;
  81         // This bit is set when the next clause writes to the data register of some
  82         // previous clause.
  83         unsigned datareg_writebarrier: 1;
  84         unsigned datareg : 6;
  85         unsigned scoreboard_deps: 8;
  86         unsigned scoreboard_index: 3;
  87         enum bifrost_clause_type clause_type: 4;
  88         unsigned unk3 : 1; // part of clauseType?
  89         enum bifrost_clause_type next_clause_type: 4;
  90         unsigned unk4 : 1; // part of nextClauseType?
  91 } __attribute__((packed));
  92
  93 enum bifrost_packed_src {
  94         BIFROST_SRC_PORT0    = 0,
  95         BIFROST_SRC_PORT1    = 1,
  96         BIFROST_SRC_PORT3    = 2,
  97         BIFROST_SRC_STAGE    = 3,
  98         BIFROST_SRC_CONST_LO = 4,
  99         BIFROST_SRC_CONST_HI = 5,
 100         BIFROST_SRC_PASS_FMA = 6,
 101         BIFROST_SRC_PASS_ADD = 7,
 102 };
 103
 104 #define BIFROST_FMA_EXT (0xe0000)
 105 #define BIFROST_FMA_OP_MOV BIFROST_FMA_EXT | (0x32d)
 106 #define BIFROST_FMA_OP_FREXPE_LOG BIFROST_FMA_EXT | 0x3c5
 107 #define BIFROST_FMA_OP_ADD_FREXPM ((BIFROST_FMA_EXT | 0x1e80) >> 3)
 108 #define BIFROST_FMA_SEL_16(swiz) (((BIFROST_FMA_EXT | 0x1e00) >> 3) | (swiz))
 109
 110 #define BIFROST_FMA_ROUND_16(mode, swiz) (BIFROST_FMA_EXT | 0x1800 | (swiz) | ((mode) << 6))
 111 #define BIFROST_FMA_ROUND_32(mode) (BIFROST_FMA_EXT | 0x1805 | ((mode) << 6))
 112
 113 struct bifrost_fma_inst {
 114         unsigned src0 : 3;
 115         unsigned op   : 20;
 116 } __attribute__((packed));
 117
 118 struct bifrost_fma_2src {
 119         unsigned src0 : 3;
 120         unsigned src1 : 3;
 121         unsigned op   : 17;
 122 } __attribute__((packed));
 123
 124 #define BIFROST_FMA_OP_SEL8 (0x71)
 125
 126 struct bifrost_fma_sel8 {
 127         unsigned src0 : 3;
 128         unsigned src1 : 3;
 129         unsigned src2 : 3;
 130         unsigned src3 : 3;
 131         unsigned swizzle : 4;
 132         unsigned op   : 7;
 133 } __attribute__((packed));
 134
 135 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
 136
 137 struct bifrost_fma_mscale {
 138         unsigned src0 : 3;
 139         unsigned src1 : 3;
 140         unsigned src2 : 3;
 141         unsigned src3 : 3;
 142
 143         /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
 144          * regular outmod */
 145         unsigned mode : 2;
 146         unsigned mscale_mode : 1;
 147
 148         unsigned src0_abs : 1;
 149         unsigned src1_neg : 1;
 150         unsigned src2_neg : 1;
 151         unsigned op   : 5;
 152 } __attribute__((packed));
 153
 154 #define BIFROST_ADD_OP_BLEND (0x1952c)
 155 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
 156 #define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
 157 #define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
 158 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
 159 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
 160 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
 161 #define BIFROST_ADD_OP_LOG2_HELP  (0x0cc68)
 162 #define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
 163
 164 struct bifrost_add_inst {
 165         unsigned src0 : 3;
 166         unsigned op   : 17;
 167 } __attribute__((packed));
 168
 169 #define BIFROST_ADD_OP_DISCARD (0x19100 >> 8)
 170
 171 enum bifrost_discard_cond {
 172         BIFROST_DISCARD_FEQ = 0,
 173         BIFROST_DISCARD_FNE = 1,
 174         BIFROST_DISCARD_FLE = 2,
 175         BIFROST_DISCARD_FLT = 3,
 176 };
 177
 178 struct bifrost_add_discard {
 179         unsigned src0 : 3;
 180         unsigned src1 : 3;
 181         enum bifrost_discard_cond cond : 2;
 182         /* Zero for fp32 */
 183         unsigned src0_select : 1;
 184         unsigned src1_select : 1;
 185         unsigned fp32 : 1;
 186         unsigned op   : 9;
 187 } __attribute__((packed));
 188
 189 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
 190 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
 191 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
 192 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
 193 #define BIFROST_ADD_SEL_16(swiz) ((0xea60 >> 3) | (swiz))
 194
 195 struct bifrost_add_2src {
 196         unsigned src0 : 3;
 197         unsigned src1 : 3;
 198         unsigned op   : 14;
 199 } __attribute__((packed));
 200
 201 #define BIFROST_ADD_OP_FMAX32 (0x00)
 202 #define BIFROST_ADD_OP_FMIN32 (0x01)
 203 #define BIFROST_ADD_OP_FADD32 (0x02)
 204
 205 #define BIFROST_ADD_OP_FADD16 (0x0A)
 206
 207 struct bifrost_add_faddmin {
 208         unsigned src0 : 3;
 209         unsigned src1 : 3;
 210         unsigned src1_abs : 1;
 211         unsigned src0_neg : 1;
 212         unsigned src1_neg : 1;
 213         unsigned select : 2; /* swizzle_0 for fp16 */
 214         unsigned outmod : 2; /* swizzle_1 for fp16 */
 215         unsigned mode : 2;
 216         unsigned src0_abs : 1;
 217         unsigned op   : 4;
 218 } __attribute__((packed));
 219
 220 #define BIFROST_ADD_OP_FMAX16 (0x10)
 221 #define BIFROST_ADD_OP_FMIN16 (0x12)
 222
 223 struct bifrost_add_fmin16 {
 224         unsigned src0 : 3;
 225         unsigned src1 : 3;
 226         /* abs2 inferred as with FMA */
 227         unsigned abs1 : 1;
 228         unsigned src0_neg : 1;
 229         unsigned src1_neg : 1;
 230         unsigned src0_swizzle : 2;
 231         unsigned src1_swizzle : 2;
 232         unsigned mode : 2;
 233         unsigned op : 5;
 234 } __attribute__((packed));
 235
 236 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
 237
 238 struct bifrost_st_vary {
 239         unsigned src0 : 3;
 240         unsigned src1 : 3;
 241         unsigned src2 : 3;
 242         unsigned channels : 2;
 243         unsigned op   : 9;
 244 } __attribute__((packed));
 245
 246 #define BIFROST_ADD_OP_ATEST (0xc8f)
 247
 248 struct bifrost_add_atest {
 249         /* gl_SampleMask (R60) */
 250         unsigned src0 : 3;
 251
 252         /* Alpha value */
 253         unsigned src1 : 3;
 254
 255         /* If half, X/Y select. If !half, always set */
 256         unsigned component : 1;
 257         unsigned half : 1;
 258
 259         unsigned op   : 12;
 260 } __attribute__((packed));
 261
 262 enum bifrost_outmod {
 263         BIFROST_NONE = 0x0,
 264         BIFROST_POS = 0x1,
 265         BIFROST_SAT_SIGNED = 0x2,
 266         BIFROST_SAT = 0x3,
 267 };
 268
 269 enum bifrost_roundmode {
 270         BIFROST_RTE = 0x0, /* round to even */
 271         BIFROST_RTP = 0x1, /* round to positive */
 272         BIFROST_RTN = 0x2, /* round to negative */
 273         BIFROST_RTZ = 0x3 /* round to zero */
 274 };
 275
 276 /* NONE: Same as fmax() and fmin() -- return the other
 277  * number if any number is NaN.  Also always return +0 if
 278  * one argument is +0 and the other is -0.
 279  *
 280  * NAN_WINS: Instead of never returning a NaN, always return
 281  * one. The "greater"/"lesser" NaN is always returned, first
 282  * by checking the sign and then the mantissa bits.
 283  *
 284  * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
 285  * For min, implement src0 < src1 ? src0 : src1.  This
 286  * includes handling NaN's and signedness of 0 differently
 287  * from above, since +0 and -0 compare equal and comparisons
 288  * always return false for NaN's. As a result, this mode is
 289  * *not* commutative.
 290  *
 291  * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
 292  * For min, implement src0 > src1 ? src1 : src0
 293  */
 294
 295
 296 enum bifrost_minmax_mode {
 297         BIFROST_MINMAX_NONE = 0x0,
 298         BIFROST_NAN_WINS    = 0x1,
 299         BIFROST_SRC1_WINS   = 0x2,
 300         BIFROST_SRC0_WINS   = 0x3,
 301 };
 302
 303 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
 304 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
 305 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
 306
 307 struct bifrost_fma_add {
 308         unsigned src0 : 3;
 309         unsigned src1 : 3;
 310         unsigned src1_abs : 1;
 311         unsigned src0_neg : 1;
 312         unsigned src1_neg : 1;
 313         unsigned unk : 3;
 314         unsigned src0_abs : 1;
 315         enum bifrost_roundmode roundmode : 2;
 316         enum bifrost_outmod outmod : 2;
 317         unsigned op : 6;
 318 } __attribute__((packed));
 319
 320 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
 321 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
 322 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
 323
 324 struct bifrost_fma_add_minmax16 {
 325         unsigned src0 : 3;
 326         unsigned src1 : 3;
 327         /* abs2 inferred as (src1 < src0) */
 328         unsigned abs1 : 1;
 329         unsigned src0_neg : 1;
 330         unsigned src1_neg : 1;
 331         unsigned src0_swizzle : 2;
 332         unsigned src1_swizzle : 2;
 333         unsigned mode : 2;
 334         enum bifrost_outmod outmod : 2;
 335         /* roundmode for add, min/max mode for min/max */
 336         unsigned op : 6;
 337 } __attribute__((packed));
 338
 339 #define BIFROST_FMA_OP_FMA (0x00)
 340
 341 struct bifrost_fma_fma {
 342         unsigned src0 : 3;
 343         unsigned src1 : 3;
 344         unsigned src2 : 3;
 345         unsigned src_expand : 3;
 346         unsigned src0_abs : 1;
 347         enum bifrost_roundmode roundmode : 2;
 348         enum bifrost_outmod outmod : 2;
 349         unsigned src0_neg : 1; /* 14 */
 350         unsigned src2_neg : 1;
 351         unsigned src1_abs : 1;
 352         unsigned src2_abs : 1; /* 17 */
 353         unsigned op : 2;
 354 } __attribute__((packed));
 355
 356 #define BIFROST_FMA_OP_FMA16 (0x2)
 357
 358 struct bifrost_fma_fma16 {
 359         unsigned src0 : 3;
 360         unsigned src1 : 3;
 361         unsigned src2 : 3;
 362         unsigned swizzle_0 : 2;
 363         unsigned swizzle_1 : 2;
 364         enum bifrost_roundmode roundmode : 2;
 365         enum bifrost_outmod outmod : 2;
 366         unsigned src0_neg : 1;
 367         unsigned src2_neg : 1;
 368         unsigned swizzle_2 : 2;
 369         unsigned op : 2;
 370 } __attribute__((packed));
 371
 372 enum bifrost_csel_cond {
 373         BIFROST_FEQ_F = 0x0,
 374         BIFROST_FGT_F = 0x1,
 375         BIFROST_FGE_F = 0x2,
 376         BIFROST_IEQ_F = 0x3,
 377         BIFROST_IGT_I = 0x4,
 378         BIFROST_IGE_I = 0x5,
 379         BIFROST_UGT_I = 0x6,
 380         BIFROST_UGE_I = 0x7
 381 };
 382
 383 #define BIFROST_FMA_OP_CSEL4     (0x5c)
 384 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
 385
 386 struct bifrost_csel4 {
 387         unsigned src0 : 3;
 388         unsigned src1 : 3;
 389         unsigned src2 : 3;
 390         unsigned src3 : 3;
 391         enum bifrost_csel_cond cond : 3;
 392         unsigned op   : 8;
 393 } __attribute__((packed));
 394
 395 #define BIFROST_FMA_OP_RSHIFT_NAND     (0x60000 >> 12)
 396 #define BIFROST_FMA_OP_RSHIFT_AND      (0x61000 >> 12)
 397 #define BIFROST_FMA_OP_LSHIFT_NAND     (0x62000 >> 12)
 398 #define BIFROST_FMA_OP_LSHIFT_AND      (0x63000 >> 12)
 399 #define BIFROST_FMA_OP_RSHIFT_XOR      (0x64000 >> 12)
 400 #define BIFROST_FMA_OP_LSHIFT_ADD_32   (0x65200 >> 6)
 401 #define BIFROST_FMA_OP_LSHIFT_SUB_32   (0x65600 >> 6)
 402 #define BIFROST_FMA_OP_LSHIFT_RSUB_32  (0x65a00 >> 6)
 403 #define BIFROST_FMA_OP_RSHIFT_ADD_32   (0x65e00 >> 6)
 404 #define BIFROST_FMA_OP_RSHIFT_SUB_32   (0x66200 >> 6)
 405 #define BIFROST_FMA_OP_RSHIFT_RSUB_32  (0x66600 >> 6)
 406
 407 struct bifrost_shift_fma {
 408         unsigned src0 : 3;
 409         unsigned src1 : 3;
 410         unsigned src2 : 3;
 411         unsigned half : 3;
 412         unsigned unk  : 1; /* always set? */
 413         unsigned invert_1 : 1; /* Inverts sources to combining op */
 414         /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
 415         unsigned invert_2 : 1;
 416         unsigned op : 8;
 417 } __attribute__((packed));
 418
 419 struct bifrost_shift_add {
 420         unsigned src0 : 3;
 421         unsigned src1 : 3;
 422         unsigned src2 : 3;
 423         unsigned zero : 2;
 424
 425         unsigned invert_1 : 1;
 426         unsigned invert_2 : 1;
 427
 428         unsigned op : 7;
 429 } __attribute__((packed));
 430
 431 enum bifrost_fcmp_cond {
 432         BIFROST_OEQ = 0,
 433         BIFROST_OGT = 1,
 434         BIFROST_OGE = 2,
 435         BIFROST_UNE = 3,
 436         BIFROST_OLT = 4,
 437         BIFROST_OLE = 5,
 438 };
 439
 440 #define BIFROST_FMA_OP_FCMP_GL (0x48000 >> 13)
 441 #define BIFROST_FMA_OP_FCMP_D3D (0x4c000 >> 13)
 442
 443 struct bifrost_fma_fcmp {
 444         unsigned src0 : 3;
 445         unsigned src1 : 3;
 446         unsigned src1_abs : 1;
 447         unsigned unk1 : 1;
 448         unsigned src1_neg : 1;
 449         unsigned src_expand : 3;
 450         unsigned src0_abs : 1;
 451         enum bifrost_fcmp_cond cond : 3;
 452         unsigned op   : 7;
 453 } __attribute__((packed));
 454
 455 struct bifrost_add_fcmp {
 456         unsigned src0 : 3;
 457         unsigned src1 : 3;
 458         enum bifrost_fcmp_cond cond : 3;
 459         unsigned src_expand : 2;
 460         unsigned src0_abs : 1;
 461         unsigned src1_abs : 1;
 462         unsigned src1_neg : 1;
 463         unsigned op   : 6;
 464 } __attribute__((packed));
 465
 466 #define BIFROST_FMA_OP_FCMP_GL_16 (0xc8000 >> 13)
 467 #define BIFROST_FMA_OP_FCMP_D3D_16 (0xcc000 >> 13)
 468
 469 struct bifrost_fma_fcmp16 {
 470         unsigned src0 : 3;
 471         unsigned src1 : 3;
 472
 473         /* abs2 inferred */
 474         unsigned abs1 : 1;
 475         unsigned unk : 2;
 476
 477         unsigned src0_swizzle : 2;
 478         unsigned src1_swizzle : 2;
 479
 480         enum bifrost_fcmp_cond cond : 3;
 481         unsigned op   : 7;
 482 } __attribute__((packed));
 483
 484 struct bifrost_add_fcmp16 {
 485         unsigned src0 : 3;
 486         unsigned src1 : 3;
 487         enum bifrost_fcmp_cond cond : 3;
 488
 489         unsigned src0_swizzle : 2;
 490         unsigned src1_swizzle : 2;
 491
 492         /* No abs mods */
 493         unsigned src0_neg : 1;
 494
 495         unsigned op   : 6;
 496 } __attribute__((packed));
 497
 498 enum bifrost_icmp_cond {
 499         BIFROST_ICMP_IGT = 0,
 500         BIFROST_ICMP_IGE = 1,
 501         BIFROST_ICMP_UGT = 2,
 502         BIFROST_ICMP_UGE = 3,
 503         BIFROST_ICMP_EQ  = 4,
 504         BIFROST_ICMP_NEQ  = 5,
 505         BIFROST_ICMP_32_OR_8 = 6, /* nested */
 506         BIFROST_ICMP_64 = 7, /* nested */
 507 };
 508
 509 struct bifrost_fma_icmp32 {
 510         unsigned src0 : 3;
 511         unsigned src1 : 3;
 512         enum bifrost_icmp_cond cond : 3;
 513         unsigned unk1 : 1; /* set */
 514         unsigned d3d : 1;
 515         unsigned op : 12;
 516 } __attribute__((packed));
 517
 518 struct bifrost_fma_icmp16 {
 519         unsigned src0 : 3;
 520         unsigned src1 : 3;
 521         unsigned unk : 5; /* 11010 */
 522         enum bifrost_icmp_cond cond : 3;
 523         unsigned op : 9;
 524 } __attribute__((packed));
 525
 526 #define BIFROST_ADD_OP_ICMP_32 (0x0f600 >> 8)
 527 #define BIFROST_ADD_OP_ICMP_16 (0x0f000 >> 11)
 528
 529 struct bifrost_add_icmp {
 530         unsigned src0 : 3;
 531         unsigned src1 : 3;
 532         enum bifrost_icmp_cond cond : 3;
 533         unsigned sz : 1; /* 1 for 32, 0 for 8 */
 534         unsigned d3d : 1;
 535         unsigned op : 9;
 536 } __attribute__((packed));
 537
 538 struct bifrost_add_icmp16 {
 539         unsigned src0 : 3;
 540         unsigned src1 : 3;
 541         unsigned src0_swizzle : 2;
 542         unsigned src1_swizzle : 2;
 543         unsigned d3d : 1;
 544         enum bifrost_icmp_cond cond : 3;
 545         unsigned op : 6;
 546 } __attribute__((packed));
 547
 548 /* Two sources for vectorization */
 549 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
 550 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
 551
 552 enum bifrost_convert_mode {
 553         BIFROST_CONV_UNK0 = 0,
 554         BIFROST_CONV_F32_TO_I32 = 1,
 555         BIFROST_CONV_F16_TO_I16 = 2,
 556         BIFROST_CONV_I32_TO_F32 = 3,
 557         BIFROST_CONV_I16_TO_X32 = 4,
 558         BIFROST_CONV_F16_TO_F32 = 5,
 559         BIFROST_CONV_I16_TO_F16 = 6,
 560         BIFROST_CONV_UNK7 = 7
 561 };
 562
 563 /* i16 to x32 */
 564 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
 565         ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
 566          ((0x3) << 3) | ((4) << 5) | 0x100)
 567
 568 /* f16 to f32 */
 569 #define BIFROST_CONVERT_5(component) \
 570         ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
 571
 572 /* Other conversions */
 573 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
 574         ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
 575
 576 #define BIFROST_FMA_CONVERT (0xe0000)
 577 #define BIFROST_ADD_CONVERT (0x07800)
 578
 579 enum bifrost_ldst_type {
 580         BIFROST_LDST_F16 = 0,
 581         BIFROST_LDST_F32 = 1,
 582         BIFROST_LDST_I32 = 2,
 583         BIFROST_LDST_U32 = 3
 584 };
 585
 586 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
 587
 588 struct bifrost_ld_var_addr {
 589         unsigned src0 : 3;
 590         unsigned src1 : 3;
 591         unsigned location : 5;
 592         enum bifrost_ldst_type type : 2;
 593         unsigned op : 7;
 594 } __attribute__((packed));
 595
 596 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
 597
 598 struct bifrost_ld_attr {
 599         unsigned src0 : 3;
 600         unsigned src1 : 3;
 601         unsigned location : 5;
 602         unsigned channels : 2; /* MALI_POSITIVE */
 603         enum bifrost_ldst_type type : 2;
 604         unsigned op : 5;
 605 } __attribute__((packed));
 606
 607 enum bifrost_interp_mode {
 608         BIFROST_INTERP_PER_FRAG = 0x0,
 609         BIFROST_INTERP_CENTROID = 0x1,
 610         BIFROST_INTERP_DEFAULT  = 0x2,
 611         BIFROST_INTERP_EXPLICIT = 0x3
 612 };
 613
 614 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
 615 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
 616
 617 /* Fixed location for gl_FragCoord.zw */
 618 #define BIFROST_FRAGZ (23)
 619 #define BIFROST_FRAGW (22)
 620
 621 struct bifrost_ld_var {
 622         unsigned src0 : 3;
 623
 624         /* If top two bits set, indirect with src in bottom three */
 625         unsigned addr : 5;
 626
 627         unsigned channels : 2; /* MALI_POSITIVE */
 628         enum bifrost_interp_mode interp_mode : 2;
 629         unsigned reuse : 1;
 630         unsigned flat : 1;
 631         unsigned op : 6;
 632 } __attribute__((packed));
 633
 634 struct bifrost_tex_ctrl {
 635         unsigned sampler_index : 4; // also used to signal indirects
 636         unsigned tex_index : 7;
 637         bool no_merge_index : 1; // whether to merge (direct) sampler & texture indices
 638         bool filter : 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
 639         unsigned unk0 : 2;
 640         bool texel_offset : 1; // *Offset()
 641         bool is_shadow : 1;
 642         bool is_array : 1;
 643         unsigned tex_type : 2; // 2D, 3D, Cube, Buffer
 644         bool compute_lod : 1; // 0 for *Lod()
 645         bool not_supply_lod : 1; // 0 for *Lod() or when a bias is applied
 646         bool calc_gradients : 1; // 0 for *Grad()
 647         unsigned unk1 : 1;
 648         unsigned result_type : 4; // integer, unsigned, float TODO: why is this 4 bits?
 649         unsigned unk2 : 4;
 650 } __attribute__((packed));
 651
 652 struct bifrost_dual_tex_ctrl {
 653         unsigned sampler_index0 : 2;
 654         unsigned unk0 : 2;
 655         unsigned tex_index0 : 2;
 656         unsigned sampler_index1 : 2;
 657         unsigned tex_index1 : 2;
 658         unsigned unk1 : 22;
 659 } __attribute__((packed));
 660
 661 #define BIFROST_ADD_OP_TEX_COMPACT_F32 (0x0b000 >> 10)
 662 #define BIFROST_ADD_OP_TEX_COMPACT_F16 (0x1b000 >> 10)
 663
 664 struct bifrost_tex_compact {
 665         unsigned src0 : 3;
 666         unsigned src1 : 3;
 667         unsigned tex_index : 3;
 668         unsigned unknown : 1;
 669         unsigned sampler_index : 3;
 670         unsigned op   : 7;
 671 } __attribute__((packed));
 672
 673 enum branch_bit_size {
 674         BR_SIZE_32 = 0,
 675         BR_SIZE_16XX = 1,
 676         BR_SIZE_16YY = 2,
 677         // For the above combinations of bitsize and location, an extra bit is
 678         // encoded via comparing the sources. The only possible source of ambiguity
 679         // would be if the sources were the same, but then the branch condition
 680         // would be always true or always false anyways, so we can ignore it. But
 681         // this no longer works when comparing the y component to the x component,
 682         // since it's valid to compare the y component of a source against its own
 683         // x component. Instead, the extra bit is encoded via an extra bitsize.
 684         BR_SIZE_16YX0 = 3,
 685         BR_SIZE_16YX1 = 4,
 686         BR_SIZE_32_AND_16X = 5,
 687         BR_SIZE_32_AND_16Y = 6,
 688         // Used for comparisons with zero and always-true, see below. I think this
 689         // only works for integer comparisons.
 690         BR_SIZE_ZERO = 7,
 691 };
 692
 693 enum bifrost_reg_write_unit {
 694         REG_WRITE_NONE = 0, // don't write
 695         REG_WRITE_TWO, // write using reg2
 696         REG_WRITE_THREE, // write using reg3
 697 };
 698
 699 struct bifrost_regs {
 700         unsigned uniform_const : 8;
 701         unsigned reg2 : 6;
 702         unsigned reg3 : 6;
 703         unsigned reg0 : 5;
 704         unsigned reg1 : 6;
 705         unsigned ctrl : 4;
 706 } __attribute__((packed));
 707
 708 enum bifrost_branch_cond {
 709         BR_COND_LT = 0,
 710         BR_COND_LE = 1,
 711         BR_COND_GE = 2,
 712         BR_COND_GT = 3,
 713         // Equal vs. not-equal determined by src0/src1 comparison
 714         BR_COND_EQ = 4,
 715         // floating-point comparisons
 716         // Becomes UNE when you flip the arguments
 717         BR_COND_OEQ = 5,
 718         // TODO what happens when you flip the arguments?
 719         BR_COND_OGT = 6,
 720         BR_COND_OLT = 7,
 721 };
 722
 723 enum bifrost_branch_code {
 724         BR_ALWAYS = 63,
 725 };
 726
 727 struct bifrost_branch {
 728         unsigned src0 : 3;
 729
 730         /* For BR_SIZE_ZERO, upper two bits become ctrl */
 731         unsigned src1 : 3;
 732
 733         /* Offset source -- always uniform/const but
 734          * theoretically could support indirect jumps? */
 735         unsigned src2 : 3;
 736
 737         enum bifrost_branch_cond cond : 3;
 738         enum branch_bit_size size : 3;
 739
 740         unsigned op : 5;
 741 };
 742
 743 /* Clause packing */
 744
 745 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
 746 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
 747
 748 struct bifrost_fmt1 {
 749         unsigned ins_0 : 3;
 750         unsigned tag : 5;
 751         uint64_t ins_1 : 64;
 752         unsigned ins_2 : 11;
 753         uint64_t header : 45;
 754 } __attribute__((packed));
 755
 756 #define BIFROST_FMT1_INSTRUCTIONS    0b00101
 757 #define BIFROST_FMT1_FINAL           0b01001
 758 #define BIFROST_FMT1_CONSTANTS       0b00001
 759
 760 #define BIFROST_FMTC_CONSTANTS       0b0011
 761 #define BIFROST_FMTC_FINAL           0b0111
 762
 763 struct bifrost_fmt_constant {
 764         unsigned pos : 4;
 765         unsigned tag : 4;
 766         uint64_t imm_1 : 60;
 767         uint64_t imm_2 : 60;
 768 } __attribute__((packed));
 769
 770 enum bifrost_reg_control {
 771         BIFROST_WRITE_FMA_P2         = 1,
 772         BIFROST_WRITE_FMA_P2_READ_P3 = 2,
 773         BIFROST_FIRST_WRITE_FMA_P2_READ_P3 = 3,
 774         BIFROST_READ_P3              = 4,
 775         BIFROST_WRITE_ADD_P2         = 5,
 776         BIFROST_WRITE_ADD_P2_READ_P3 = 6,
 777         BIFROST_WRITE_ADD_P2_FMA_P3  = 7,
 778
 779         BIFROST_FIRST_NONE           = 8,
 780         BIFROST_FIRST_WRITE_FMA_P2   = 9,
 781         /* INSTR_INVALID_ENC */
 782         BIFROST_REG_NONE             = 11,
 783         BIFROST_FIRST_READ_P3        = 12,
 784         BIFROST_FIRST_WRITE_ADD_P2   = 13,
 785         BIFROST_FIRST_WRITE_ADD_P2_READ_P3 = 14,
 786         BIFROST_FIRST_WRITE_ADD_P2_FMA_P3  = 15
 787 };
 788
 789 #endif