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 #define BIFROST_FMA_IADD_32 (0x4ff98 >> 3)
 119 #define BIFROST_FMA_ISUB_32 (0x4ffd8 >> 3)
 120
 121 struct bifrost_fma_2src {
 122         unsigned src0 : 3;
 123         unsigned src1 : 3;
 124         unsigned op   : 17;
 125 } __attribute__((packed));
 126
 127 #define BIFROST_FMA_OP_SEL8 (0x71)
 128
 129 struct bifrost_fma_sel8 {
 130         unsigned src0 : 3;
 131         unsigned src1 : 3;
 132         unsigned src2 : 3;
 133         unsigned src3 : 3;
 134         unsigned swizzle : 4;
 135         unsigned op   : 7;
 136 } __attribute__((packed));
 137
 138 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
 139
 140 struct bifrost_fma_mscale {
 141         unsigned src0 : 3;
 142         unsigned src1 : 3;
 143         unsigned src2 : 3;
 144         unsigned src3 : 3;
 145
 146         /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
 147          * regular outmod */
 148         unsigned mode : 2;
 149         unsigned mscale_mode : 1;
 150
 151         unsigned src0_abs : 1;
 152         unsigned src1_neg : 1;
 153         unsigned src2_neg : 1;
 154         unsigned op   : 5;
 155 } __attribute__((packed));
 156
 157 #define BIFROST_ADD_OP_BLEND (0x1952c)
 158 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
 159 #define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
 160 #define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
 161 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
 162 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
 163 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
 164 #define BIFROST_ADD_OP_LOG2_HELP  (0x0cc68)
 165 #define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
 166
 167 struct bifrost_add_inst {
 168         unsigned src0 : 3;
 169         unsigned op   : 17;
 170 } __attribute__((packed));
 171
 172 #define BIFROST_ADD_OP_DISCARD (0x19100 >> 8)
 173
 174 enum bifrost_discard_cond {
 175         BIFROST_DISCARD_FEQ = 0,
 176         BIFROST_DISCARD_FNE = 1,
 177         BIFROST_DISCARD_FLE = 2,
 178         BIFROST_DISCARD_FLT = 3,
 179 };
 180
 181 struct bifrost_add_discard {
 182         unsigned src0 : 3;
 183         unsigned src1 : 3;
 184         enum bifrost_discard_cond cond : 2;
 185         /* Zero for fp32 */
 186         unsigned src0_select : 1;
 187         unsigned src1_select : 1;
 188         unsigned fp32 : 1;
 189         unsigned op   : 9;
 190 } __attribute__((packed));
 191
 192 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
 193 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
 194 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
 195 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
 196 #define BIFROST_ADD_SEL_16(swiz) ((0xea60 >> 3) | (swiz))
 197
 198 #define BIFROST_ADD_IADD_8  (0x17880 >> 3)
 199 #define BIFROST_ADD_IADD_16 (0x17900 >> 3)
 200 #define BIFROST_ADD_IADD_32 (0x178c0 >> 3)
 201 #define BIFROST_ADD_ISUB_8  (0x17a80 >> 3)
 202 #define BIFROST_ADD_ISUB_16 (0x17b00 >> 3)
 203 #define BIFROST_ADD_ISUB_32 (0x17ac0 >> 3)
 204
 205 struct bifrost_add_2src {
 206         unsigned src0 : 3;
 207         unsigned src1 : 3;
 208         unsigned op   : 14;
 209 } __attribute__((packed));
 210
 211 #define BIFROST_ADD_OP_FMAX32 (0x00)
 212 #define BIFROST_ADD_OP_FMIN32 (0x01)
 213 #define BIFROST_ADD_OP_FADD32 (0x02)
 214
 215 #define BIFROST_ADD_OP_FADD16 (0x0A)
 216
 217 struct bifrost_add_faddmin {
 218         unsigned src0 : 3;
 219         unsigned src1 : 3;
 220         unsigned src1_abs : 1;
 221         unsigned src0_neg : 1;
 222         unsigned src1_neg : 1;
 223         unsigned select : 2; /* swizzle_0 for fp16 */
 224         unsigned outmod : 2; /* swizzle_1 for fp16 */
 225         unsigned mode : 2;
 226         unsigned src0_abs : 1;
 227         unsigned op   : 4;
 228 } __attribute__((packed));
 229
 230 #define BIFROST_ADD_OP_FMAX16 (0x10)
 231 #define BIFROST_ADD_OP_FMIN16 (0x12)
 232
 233 struct bifrost_add_fmin16 {
 234         unsigned src0 : 3;
 235         unsigned src1 : 3;
 236         /* abs2 inferred as with FMA */
 237         unsigned abs1 : 1;
 238         unsigned src0_neg : 1;
 239         unsigned src1_neg : 1;
 240         unsigned src0_swizzle : 2;
 241         unsigned src1_swizzle : 2;
 242         unsigned mode : 2;
 243         unsigned op : 5;
 244 } __attribute__((packed));
 245
 246 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
 247
 248 struct bifrost_st_vary {
 249         unsigned src0 : 3;
 250         unsigned src1 : 3;
 251         unsigned src2 : 3;
 252         unsigned channels : 2;
 253         unsigned op   : 9;
 254 } __attribute__((packed));
 255
 256 #define BIFROST_ADD_OP_ATEST (0xc8f)
 257
 258 struct bifrost_add_atest {
 259         /* gl_SampleMask (R60) */
 260         unsigned src0 : 3;
 261
 262         /* Alpha value */
 263         unsigned src1 : 3;
 264
 265         /* If half, X/Y select. If !half, always set */
 266         unsigned component : 1;
 267         unsigned half : 1;
 268
 269         unsigned op   : 12;
 270 } __attribute__((packed));
 271
 272 enum bifrost_outmod {
 273         BIFROST_NONE = 0x0,
 274         BIFROST_POS = 0x1,
 275         BIFROST_SAT_SIGNED = 0x2,
 276         BIFROST_SAT = 0x3,
 277 };
 278
 279 enum bifrost_roundmode {
 280         BIFROST_RTE = 0x0, /* round to even */
 281         BIFROST_RTP = 0x1, /* round to positive */
 282         BIFROST_RTN = 0x2, /* round to negative */
 283         BIFROST_RTZ = 0x3 /* round to zero */
 284 };
 285
 286 /* NONE: Same as fmax() and fmin() -- return the other
 287  * number if any number is NaN.  Also always return +0 if
 288  * one argument is +0 and the other is -0.
 289  *
 290  * NAN_WINS: Instead of never returning a NaN, always return
 291  * one. The "greater"/"lesser" NaN is always returned, first
 292  * by checking the sign and then the mantissa bits.
 293  *
 294  * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
 295  * For min, implement src0 < src1 ? src0 : src1.  This
 296  * includes handling NaN's and signedness of 0 differently
 297  * from above, since +0 and -0 compare equal and comparisons
 298  * always return false for NaN's. As a result, this mode is
 299  * *not* commutative.
 300  *
 301  * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
 302  * For min, implement src0 > src1 ? src1 : src0
 303  */
 304
 305
 306 enum bifrost_minmax_mode {
 307         BIFROST_MINMAX_NONE = 0x0,
 308         BIFROST_NAN_WINS    = 0x1,
 309         BIFROST_SRC1_WINS   = 0x2,
 310         BIFROST_SRC0_WINS   = 0x3,
 311 };
 312
 313 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
 314 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
 315 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
 316
 317 struct bifrost_fma_add {
 318         unsigned src0 : 3;
 319         unsigned src1 : 3;
 320         unsigned src1_abs : 1;
 321         unsigned src0_neg : 1;
 322         unsigned src1_neg : 1;
 323         unsigned unk : 3;
 324         unsigned src0_abs : 1;
 325         enum bifrost_roundmode roundmode : 2;
 326         enum bifrost_outmod outmod : 2;
 327         unsigned op : 6;
 328 } __attribute__((packed));
 329
 330 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
 331 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
 332 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
 333
 334 struct bifrost_fma_add_minmax16 {
 335         unsigned src0 : 3;
 336         unsigned src1 : 3;
 337         /* abs2 inferred as (src1 < src0) */
 338         unsigned abs1 : 1;
 339         unsigned src0_neg : 1;
 340         unsigned src1_neg : 1;
 341         unsigned src0_swizzle : 2;
 342         unsigned src1_swizzle : 2;
 343         unsigned mode : 2;
 344         enum bifrost_outmod outmod : 2;
 345         /* roundmode for add, min/max mode for min/max */
 346         unsigned op : 6;
 347 } __attribute__((packed));
 348
 349 #define BIFROST_FMA_OP_FMA (0x00)
 350
 351 struct bifrost_fma_fma {
 352         unsigned src0 : 3;
 353         unsigned src1 : 3;
 354         unsigned src2 : 3;
 355         unsigned src_expand : 3;
 356         unsigned src0_abs : 1;
 357         enum bifrost_roundmode roundmode : 2;
 358         enum bifrost_outmod outmod : 2;
 359         unsigned src0_neg : 1; /* 14 */
 360         unsigned src2_neg : 1;
 361         unsigned src1_abs : 1;
 362         unsigned src2_abs : 1; /* 17 */
 363         unsigned op : 2;
 364 } __attribute__((packed));
 365
 366 #define BIFROST_FMA_OP_FMA16 (0x2)
 367
 368 struct bifrost_fma_fma16 {
 369         unsigned src0 : 3;
 370         unsigned src1 : 3;
 371         unsigned src2 : 3;
 372         unsigned swizzle_0 : 2;
 373         unsigned swizzle_1 : 2;
 374         enum bifrost_roundmode roundmode : 2;
 375         enum bifrost_outmod outmod : 2;
 376         unsigned src0_neg : 1;
 377         unsigned src2_neg : 1;
 378         unsigned swizzle_2 : 2;
 379         unsigned op : 2;
 380 } __attribute__((packed));
 381
 382 enum bifrost_csel_cond {
 383         BIFROST_FEQ_F = 0x0,
 384         BIFROST_FGT_F = 0x1,
 385         BIFROST_FGE_F = 0x2,
 386         BIFROST_IEQ_F = 0x3,
 387         BIFROST_IGT_I = 0x4,
 388         BIFROST_IGE_I = 0x5,
 389         BIFROST_UGT_I = 0x6,
 390         BIFROST_UGE_I = 0x7
 391 };
 392
 393 #define BIFROST_FMA_OP_CSEL4     (0x5c)
 394 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
 395
 396 struct bifrost_csel4 {
 397         unsigned src0 : 3;
 398         unsigned src1 : 3;
 399         unsigned src2 : 3;
 400         unsigned src3 : 3;
 401         enum bifrost_csel_cond cond : 3;
 402         unsigned op   : 8;
 403 } __attribute__((packed));
 404
 405 #define BIFROST_FMA_OP_RSHIFT_NAND     (0x60000 >> 12)
 406 #define BIFROST_FMA_OP_RSHIFT_AND      (0x61000 >> 12)
 407 #define BIFROST_FMA_OP_LSHIFT_NAND     (0x62000 >> 12)
 408 #define BIFROST_FMA_OP_LSHIFT_AND      (0x63000 >> 12)
 409 #define BIFROST_FMA_OP_RSHIFT_XOR      (0x64000 >> 12)
 410 #define BIFROST_FMA_OP_LSHIFT_ADD_32   (0x65200 >> 6)
 411 #define BIFROST_FMA_OP_LSHIFT_SUB_32   (0x65600 >> 6)
 412 #define BIFROST_FMA_OP_LSHIFT_RSUB_32  (0x65a00 >> 6)
 413 #define BIFROST_FMA_OP_RSHIFT_ADD_32   (0x65e00 >> 6)
 414 #define BIFROST_FMA_OP_RSHIFT_SUB_32   (0x66200 >> 6)
 415 #define BIFROST_FMA_OP_RSHIFT_RSUB_32  (0x66600 >> 6)
 416
 417 struct bifrost_shift_fma {
 418         unsigned src0 : 3;
 419         unsigned src1 : 3;
 420         unsigned src2 : 3;
 421         unsigned half : 3;
 422         unsigned unk  : 1; /* always set? */
 423         unsigned invert_1 : 1; /* Inverts sources to combining op */
 424         /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
 425         unsigned invert_2 : 1;
 426         unsigned op : 8;
 427 } __attribute__((packed));
 428
 429 struct bifrost_shift_add {
 430         unsigned src0 : 3;
 431         unsigned src1 : 3;
 432         unsigned src2 : 3;
 433         unsigned zero : 2;
 434
 435         unsigned invert_1 : 1;
 436         unsigned invert_2 : 1;
 437
 438         unsigned op : 7;
 439 } __attribute__((packed));
 440
 441 enum bifrost_fcmp_cond {
 442         BIFROST_OEQ = 0,
 443         BIFROST_OGT = 1,
 444         BIFROST_OGE = 2,
 445         BIFROST_UNE = 3,
 446         BIFROST_OLT = 4,
 447         BIFROST_OLE = 5,
 448 };
 449
 450 #define BIFROST_FMA_OP_FCMP_GL (0x48000 >> 13)
 451 #define BIFROST_FMA_OP_FCMP_D3D (0x4c000 >> 13)
 452
 453 struct bifrost_fma_fcmp {
 454         unsigned src0 : 3;
 455         unsigned src1 : 3;
 456         unsigned src1_abs : 1;
 457         unsigned unk1 : 1;
 458         unsigned src1_neg : 1;
 459         unsigned src_expand : 3;
 460         unsigned src0_abs : 1;
 461         enum bifrost_fcmp_cond cond : 3;
 462         unsigned op   : 7;
 463 } __attribute__((packed));
 464
 465 struct bifrost_add_fcmp {
 466         unsigned src0 : 3;
 467         unsigned src1 : 3;
 468         enum bifrost_fcmp_cond cond : 3;
 469         unsigned src_expand : 2;
 470         unsigned src0_abs : 1;
 471         unsigned src1_abs : 1;
 472         unsigned src1_neg : 1;
 473         unsigned op   : 6;
 474 } __attribute__((packed));
 475
 476 #define BIFROST_FMA_OP_FCMP_GL_16 (0xc8000 >> 13)
 477 #define BIFROST_FMA_OP_FCMP_D3D_16 (0xcc000 >> 13)
 478
 479 struct bifrost_fma_fcmp16 {
 480         unsigned src0 : 3;
 481         unsigned src1 : 3;
 482
 483         /* abs2 inferred */
 484         unsigned abs1 : 1;
 485         unsigned unk : 2;
 486
 487         unsigned src0_swizzle : 2;
 488         unsigned src1_swizzle : 2;
 489
 490         enum bifrost_fcmp_cond cond : 3;
 491         unsigned op   : 7;
 492 } __attribute__((packed));
 493
 494 struct bifrost_add_fcmp16 {
 495         unsigned src0 : 3;
 496         unsigned src1 : 3;
 497         enum bifrost_fcmp_cond cond : 3;
 498
 499         unsigned src0_swizzle : 2;
 500         unsigned src1_swizzle : 2;
 501
 502         /* No abs mods */
 503         unsigned src0_neg : 1;
 504
 505         unsigned op   : 6;
 506 } __attribute__((packed));
 507
 508 enum bifrost_icmp_cond {
 509         BIFROST_ICMP_IGT = 0,
 510         BIFROST_ICMP_IGE = 1, /* swapped for 16-bit */
 511         BIFROST_ICMP_UGT = 2, /* swapped for 16-bit */
 512         BIFROST_ICMP_UGE = 3,
 513         BIFROST_ICMP_EQ  = 4,
 514         BIFROST_ICMP_NEQ  = 5,
 515         BIFROST_ICMP_32_OR_8 = 6, /* nested */
 516         BIFROST_ICMP_64 = 7, /* nested */
 517 };
 518
 519 struct bifrost_fma_icmp32 {
 520         unsigned src0 : 3;
 521         unsigned src1 : 3;
 522         enum bifrost_icmp_cond cond : 3;
 523         unsigned unk1 : 1; /* set */
 524         unsigned d3d : 1;
 525         unsigned op : 12;
 526 } __attribute__((packed));
 527
 528 struct bifrost_fma_icmp16 {
 529         unsigned src0 : 3;
 530         unsigned src1 : 3;
 531         unsigned unk : 5; /* 11010 */
 532         enum bifrost_icmp_cond cond : 3;
 533         unsigned op : 9;
 534 } __attribute__((packed));
 535
 536 #define BIFROST_ADD_OP_ICMP_32 (0x0f600 >> 8)
 537 #define BIFROST_ADD_OP_ICMP_16 (0x0f000 >> 11)
 538
 539 struct bifrost_add_icmp {
 540         unsigned src0 : 3;
 541         unsigned src1 : 3;
 542         enum bifrost_icmp_cond cond : 3;
 543         unsigned sz : 1; /* 1 for 32, 0 for 8 */
 544         unsigned d3d : 1;
 545         unsigned op : 9;
 546 } __attribute__((packed));
 547
 548 struct bifrost_add_icmp16 {
 549         unsigned src0 : 3;
 550         unsigned src1 : 3;
 551         unsigned src0_swizzle : 2;
 552         unsigned src1_swizzle : 2;
 553         unsigned d3d : 1;
 554         enum bifrost_icmp_cond cond : 3;
 555         unsigned op : 6;
 556 } __attribute__((packed));
 557
 558 /* Two sources for vectorization */
 559 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
 560 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
 561
 562 enum bifrost_convert_mode {
 563         BIFROST_CONV_UNK0 = 0,
 564         BIFROST_CONV_F32_TO_I32 = 1,
 565         BIFROST_CONV_F16_TO_I16 = 2,
 566         BIFROST_CONV_I32_TO_F32 = 3,
 567         BIFROST_CONV_I16_TO_X32 = 4,
 568         BIFROST_CONV_F16_TO_F32 = 5,
 569         BIFROST_CONV_I16_TO_F16 = 6,
 570         BIFROST_CONV_UNK7 = 7
 571 };
 572
 573 /* i16 to x32 */
 574 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
 575         ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
 576          ((0x3) << 3) | ((4) << 5) | 0x100)
 577
 578 /* f16 to f32 */
 579 #define BIFROST_CONVERT_5(component) \
 580         ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
 581
 582 /* Other conversions */
 583 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
 584         ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
 585
 586 #define BIFROST_FMA_CONVERT (0xe0000)
 587 #define BIFROST_ADD_CONVERT (0x07800)
 588
 589 enum bifrost_ldst_type {
 590         BIFROST_LDST_F16 = 0,
 591         BIFROST_LDST_F32 = 1,
 592         BIFROST_LDST_I32 = 2,
 593         BIFROST_LDST_U32 = 3
 594 };
 595
 596 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
 597
 598 struct bifrost_ld_var_addr {
 599         unsigned src0 : 3;
 600         unsigned src1 : 3;
 601         unsigned location : 5;
 602         enum bifrost_ldst_type type : 2;
 603         unsigned op : 7;
 604 } __attribute__((packed));
 605
 606 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
 607
 608 struct bifrost_ld_attr {
 609         unsigned src0 : 3;
 610         unsigned src1 : 3;
 611         unsigned location : 5;
 612         unsigned channels : 2; /* MALI_POSITIVE */
 613         enum bifrost_ldst_type type : 2;
 614         unsigned op : 5;
 615 } __attribute__((packed));
 616
 617 enum bifrost_interp_mode {
 618         BIFROST_INTERP_PER_FRAG = 0x0,
 619         BIFROST_INTERP_CENTROID = 0x1,
 620         BIFROST_INTERP_DEFAULT  = 0x2,
 621         BIFROST_INTERP_EXPLICIT = 0x3
 622 };
 623
 624 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
 625 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
 626
 627 /* Fixed location for gl_FragCoord.zw */
 628 #define BIFROST_FRAGZ (23)
 629 #define BIFROST_FRAGW (22)
 630
 631 struct bifrost_ld_var {
 632         unsigned src0 : 3;
 633
 634         /* If top two bits set, indirect with src in bottom three */
 635         unsigned addr : 5;
 636
 637         unsigned channels : 2; /* MALI_POSITIVE */
 638         enum bifrost_interp_mode interp_mode : 2;
 639         unsigned reuse : 1;
 640         unsigned flat : 1;
 641         unsigned op : 6;
 642 } __attribute__((packed));
 643
 644 struct bifrost_tex_ctrl {
 645         unsigned sampler_index : 4; // also used to signal indirects
 646         unsigned tex_index : 7;
 647         bool no_merge_index : 1; // whether to merge (direct) sampler & texture indices
 648         bool filter : 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
 649         unsigned unk0 : 2;
 650         bool texel_offset : 1; // *Offset()
 651         bool is_shadow : 1;
 652         bool is_array : 1;
 653         unsigned tex_type : 2; // 2D, 3D, Cube, Buffer
 654         bool compute_lod : 1; // 0 for *Lod()
 655         bool not_supply_lod : 1; // 0 for *Lod() or when a bias is applied
 656         bool calc_gradients : 1; // 0 for *Grad()
 657         unsigned unk1 : 1;
 658         unsigned result_type : 4; // integer, unsigned, float TODO: why is this 4 bits?
 659         unsigned unk2 : 4;
 660 } __attribute__((packed));
 661
 662 struct bifrost_dual_tex_ctrl {
 663         unsigned sampler_index0 : 2;
 664         unsigned unk0 : 2;
 665         unsigned tex_index0 : 2;
 666         unsigned sampler_index1 : 2;
 667         unsigned tex_index1 : 2;
 668         unsigned unk1 : 22;
 669 } __attribute__((packed));
 670
 671 #define BIFROST_ADD_OP_TEX_COMPACT_F32(vtx) ((0x0b000 | ((vtx) ? (0x400) : (0))) >> 10)
 672 #define BIFROST_ADD_OP_TEX_COMPACT_F16(vtx) ((0x1b000 | ((vtx) ? (0x400) : (0))) >> 10)
 673
 674 struct bifrost_tex_compact {
 675         unsigned src0 : 3;
 676         unsigned src1 : 3;
 677         unsigned tex_index : 3;
 678         unsigned compute_lod : 1;
 679         unsigned sampler_index : 3;
 680         unsigned op   : 7;
 681 } __attribute__((packed));
 682
 683 enum branch_bit_size {
 684         BR_SIZE_32 = 0,
 685         BR_SIZE_16XX = 1,
 686         BR_SIZE_16YY = 2,
 687         // For the above combinations of bitsize and location, an extra bit is
 688         // encoded via comparing the sources. The only possible source of ambiguity
 689         // would be if the sources were the same, but then the branch condition
 690         // would be always true or always false anyways, so we can ignore it. But
 691         // this no longer works when comparing the y component to the x component,
 692         // since it's valid to compare the y component of a source against its own
 693         // x component. Instead, the extra bit is encoded via an extra bitsize.
 694         BR_SIZE_16YX0 = 3,
 695         BR_SIZE_16YX1 = 4,
 696         BR_SIZE_32_AND_16X = 5,
 697         BR_SIZE_32_AND_16Y = 6,
 698         // Used for comparisons with zero and always-true, see below. I think this
 699         // only works for integer comparisons.
 700         BR_SIZE_ZERO = 7,
 701 };
 702
 703 enum bifrost_reg_write_unit {
 704         REG_WRITE_NONE = 0, // don't write
 705         REG_WRITE_TWO, // write using reg2
 706         REG_WRITE_THREE, // write using reg3
 707 };
 708
 709 struct bifrost_regs {
 710         unsigned uniform_const : 8;
 711         unsigned reg2 : 6;
 712         unsigned reg3 : 6;
 713         unsigned reg0 : 5;
 714         unsigned reg1 : 6;
 715         unsigned ctrl : 4;
 716 } __attribute__((packed));
 717
 718 enum bifrost_branch_cond {
 719         BR_COND_LT = 0,
 720         BR_COND_LE = 1,
 721         BR_COND_GE = 2,
 722         BR_COND_GT = 3,
 723         // Equal vs. not-equal determined by src0/src1 comparison
 724         BR_COND_EQ = 4,
 725         // floating-point comparisons
 726         // Becomes UNE when you flip the arguments
 727         BR_COND_OEQ = 5,
 728         // TODO what happens when you flip the arguments?
 729         BR_COND_OGT = 6,
 730         BR_COND_OLT = 7,
 731 };
 732
 733 enum bifrost_branch_code {
 734         BR_ALWAYS = 63,
 735 };
 736
 737 struct bifrost_branch {
 738         unsigned src0 : 3;
 739
 740         /* For BR_SIZE_ZERO, upper two bits become ctrl */
 741         unsigned src1 : 3;
 742
 743         /* Offset source -- always uniform/const but
 744          * theoretically could support indirect jumps? */
 745         unsigned src2 : 3;
 746
 747         enum bifrost_branch_cond cond : 3;
 748         enum branch_bit_size size : 3;
 749
 750         unsigned op : 5;
 751 };
 752
 753 /* Clause packing */
 754
 755 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
 756 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
 757
 758 struct bifrost_fmt1 {
 759         unsigned ins_0 : 3;
 760         unsigned tag : 5;
 761         uint64_t ins_1 : 64;
 762         unsigned ins_2 : 11;
 763         uint64_t header : 45;
 764 } __attribute__((packed));
 765
 766 #define BIFROST_FMT1_INSTRUCTIONS    0b00101
 767 #define BIFROST_FMT1_FINAL           0b01001
 768 #define BIFROST_FMT1_CONSTANTS       0b00001
 769
 770 #define BIFROST_FMTC_CONSTANTS       0b0011
 771 #define BIFROST_FMTC_FINAL           0b0111
 772
 773 struct bifrost_fmt_constant {
 774         unsigned pos : 4;
 775         unsigned tag : 4;
 776         uint64_t imm_1 : 60;
 777         uint64_t imm_2 : 60;
 778 } __attribute__((packed));
 779
 780 enum bifrost_reg_control {
 781         BIFROST_WRITE_FMA_P2         = 1,
 782         BIFROST_WRITE_FMA_P2_READ_P3 = 2,
 783         BIFROST_FIRST_WRITE_FMA_P2_READ_P3 = 3,
 784         BIFROST_READ_P3              = 4,
 785         BIFROST_WRITE_ADD_P2         = 5,
 786         BIFROST_WRITE_ADD_P2_READ_P3 = 6,
 787         BIFROST_WRITE_ADD_P2_FMA_P3  = 7,
 788
 789         BIFROST_FIRST_NONE           = 8,
 790         BIFROST_FIRST_WRITE_FMA_P2   = 9,
 791         /* INSTR_INVALID_ENC */
 792         BIFROST_REG_NONE             = 11,
 793         BIFROST_FIRST_READ_P3        = 12,
 794         BIFROST_FIRST_WRITE_ADD_P2   = 13,
 795         BIFROST_FIRST_WRITE_ADD_P2_READ_P3 = 14,
 796         BIFROST_FIRST_WRITE_ADD_P2_FMA_P3  = 15
 797 };
 798
 799 #endif