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