1 /* Author(s): 2 * Connor Abbott 3 * Alyssa Rosenzweig 4 * 5 * Copyright (c) 2013 Connor Abbott (connor@abbott.cx) 6 * Copyright (c) 2018 Alyssa Rosenzweig (alyssa@rosenzweig.io) 7 * Copyright (C) 2019-2020 Collabora, Ltd. 8 * 9 * Permission is hereby granted, free of charge, to any person obtaining a copy 10 * of this software and associated documentation files (the "Software"), to deal 11 * in the Software without restriction, including without limitation the rights 12 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 13 * copies of the Software, and to permit persons to whom the Software is 14 * furnished to do so, subject to the following conditions: 15 * 16 * The above copyright notice and this permission notice shall be included in 17 * all copies or substantial portions of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 22 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 24 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 25 * THE SOFTWARE. 26 */ 27 28 #ifndef __midgard_h__ 29 #define __midgard_h__ 30 31 #include <stdint.h> 32 #include <stdbool.h> 33 34 #define MIDGARD_DBG_MSGS 0x0001 35 #define MIDGARD_DBG_SHADERS 0x0002 36 #define MIDGARD_DBG_SHADERDB 0x0004 37 #define MIDGARD_DBG_INORDER 0x0008 38 #define MIDGARD_DBG_VERBOSE 0x0010 39 #define MIDGARD_DBG_INTERNAL 0x0020 40 41 extern int midgard_debug; 42 43 typedef enum { 44 midgard_word_type_alu, 45 midgard_word_type_load_store, 46 midgard_word_type_texture, 47 midgard_word_type_unknown 48 } midgard_word_type; 49 50 typedef enum { 51 midgard_alu_vmul, 52 midgard_alu_sadd, 53 midgard_alu_smul, 54 midgard_alu_vadd, 55 midgard_alu_lut 56 } midgard_alu; 57 58 enum { 59 TAG_INVALID = 0x0, 60 TAG_BREAK = 0x1, 61 TAG_TEXTURE_4_VTX = 0x2, 62 TAG_TEXTURE_4 = 0x3, 63 TAG_TEXTURE_4_BARRIER = 0x4, 64 TAG_LOAD_STORE_4 = 0x5, 65 TAG_UNKNOWN_1 = 0x6, 66 TAG_UNKNOWN_2 = 0x7, 67 TAG_ALU_4 = 0x8, 68 TAG_ALU_8 = 0x9, 69 TAG_ALU_12 = 0xA, 70 TAG_ALU_16 = 0xB, 71 TAG_ALU_4_WRITEOUT = 0xC, 72 TAG_ALU_8_WRITEOUT = 0xD, 73 TAG_ALU_12_WRITEOUT = 0xE, 74 TAG_ALU_16_WRITEOUT = 0xF 75 }; 76 77 /* 78 * ALU words 79 */ 80 81 typedef enum { 82 midgard_alu_op_fadd = 0x10, /* round to even */ 83 midgard_alu_op_fadd_rtz = 0x11, 84 midgard_alu_op_fadd_rtn = 0x12, 85 midgard_alu_op_fadd_rtp = 0x13, 86 midgard_alu_op_fmul = 0x14, /* round to even */ 87 midgard_alu_op_fmul_rtz = 0x15, 88 midgard_alu_op_fmul_rtn = 0x16, 89 midgard_alu_op_fmul_rtp = 0x17, 90 91 midgard_alu_op_fmin = 0x28, /* if an operand is NaN, propagate the other */ 92 midgard_alu_op_fmin_nan = 0x29, /* if an operand is NaN, propagate it */ 93 midgard_alu_op_fabsmin = 0x2A, /* min(abs(a,b)) */ 94 midgard_alu_op_fabsmin_nan = 0x2B, /* min_nan(abs(a,b)) */ 95 midgard_alu_op_fmax = 0x2C, /* if an operand is NaN, propagate the other */ 96 midgard_alu_op_fmax_nan = 0x2D, /* if an operand is NaN, propagate it */ 97 midgard_alu_op_fabsmax = 0x2E, /* max(abs(a,b)) */ 98 midgard_alu_op_fabsmax_nan = 0x2F, /* max_nan(abs(a,b)) */ 99 100 midgard_alu_op_fmov = 0x30, /* fmov_rte */ 101 midgard_alu_op_fmov_rtz = 0x31, 102 midgard_alu_op_fmov_rtn = 0x32, 103 midgard_alu_op_fmov_rtp = 0x33, 104 midgard_alu_op_froundeven = 0x34, 105 midgard_alu_op_ftrunc = 0x35, 106 midgard_alu_op_ffloor = 0x36, 107 midgard_alu_op_fceil = 0x37, 108 midgard_alu_op_ffma = 0x38, /* rte */ 109 midgard_alu_op_ffma_rtz = 0x39, 110 midgard_alu_op_ffma_rtn = 0x3A, 111 midgard_alu_op_ffma_rtp = 0x3B, 112 midgard_alu_op_fdot3 = 0x3C, 113 midgard_alu_op_fdot3r = 0x3D, 114 midgard_alu_op_fdot4 = 0x3E, 115 midgard_alu_op_freduce = 0x3F, 116 117 midgard_alu_op_iadd = 0x40, 118 midgard_alu_op_ishladd = 0x41, /* (a<<1) + b */ 119 midgard_alu_op_isub = 0x46, 120 midgard_alu_op_ishlsub = 0x47, /* (a<<1) - b */ 121 midgard_alu_op_iaddsat = 0x48, 122 midgard_alu_op_uaddsat = 0x49, 123 midgard_alu_op_isubsat = 0x4E, 124 midgard_alu_op_usubsat = 0x4F, 125 126 midgard_alu_op_imul = 0x58, 127 /* Multiplies two ints and stores the result in the next larger datasize. */ 128 midgard_alu_op_iwmul = 0x59, /* sint * sint = sint */ 129 midgard_alu_op_uwmul = 0x5A, /* uint * uint = uint */ 130 midgard_alu_op_iuwmul = 0x5B, /* sint * uint = sint */ 131 132 midgard_alu_op_imin = 0x60, 133 midgard_alu_op_umin = 0x61, 134 midgard_alu_op_imax = 0x62, 135 midgard_alu_op_umax = 0x63, 136 midgard_alu_op_iavg = 0x64, 137 midgard_alu_op_uavg = 0x65, 138 midgard_alu_op_iravg = 0x66, 139 midgard_alu_op_uravg = 0x67, 140 midgard_alu_op_iasr = 0x68, 141 midgard_alu_op_ilsr = 0x69, 142 midgard_alu_op_ishlsat = 0x6C, 143 midgard_alu_op_ushlsat = 0x6D, 144 midgard_alu_op_ishl = 0x6E, 145 146 midgard_alu_op_iand = 0x70, 147 midgard_alu_op_ior = 0x71, 148 midgard_alu_op_inand = 0x72, /* ~(a & b), for inot let a = b */ 149 midgard_alu_op_inor = 0x73, /* ~(a | b) */ 150 midgard_alu_op_iandnot = 0x74, /* (a & ~b), used for not/b2f */ 151 midgard_alu_op_iornot = 0x75, /* (a | ~b) */ 152 midgard_alu_op_ixor = 0x76, 153 midgard_alu_op_inxor = 0x77, /* ~(a ^ b) */ 154 midgard_alu_op_iclz = 0x78, /* Number of zeroes on left */ 155 midgard_alu_op_ipopcnt = 0x7A, /* Population count */ 156 midgard_alu_op_imov = 0x7B, 157 midgard_alu_op_iabsdiff = 0x7C, 158 midgard_alu_op_uabsdiff = 0x7D, 159 midgard_alu_op_ichoose = 0x7E, /* vector, component number - dupe for shuffle() */ 160 161 midgard_alu_op_feq = 0x80, 162 midgard_alu_op_fne = 0x81, 163 midgard_alu_op_flt = 0x82, 164 midgard_alu_op_fle = 0x83, 165 midgard_alu_op_fball_eq = 0x88, 166 midgard_alu_op_fball_neq = 0x89, 167 midgard_alu_op_fball_lt = 0x8A, /* all(lessThan(.., ..)) */ 168 midgard_alu_op_fball_lte = 0x8B, /* all(lessThanEqual(.., ..)) */ 169 170 midgard_alu_op_fbany_eq = 0x90, 171 midgard_alu_op_fbany_neq = 0x91, 172 midgard_alu_op_fbany_lt = 0x92, /* any(lessThan(.., ..)) */ 173 midgard_alu_op_fbany_lte = 0x93, /* any(lessThanEqual(.., ..)) */ 174 175 midgard_alu_op_f2i_rte = 0x98, 176 midgard_alu_op_f2i_rtz = 0x99, 177 midgard_alu_op_f2i_rtn = 0x9A, 178 midgard_alu_op_f2i_rtp = 0x9B, 179 midgard_alu_op_f2u_rte = 0x9C, 180 midgard_alu_op_f2u_rtz = 0x9D, 181 midgard_alu_op_f2u_rtn = 0x9E, 182 midgard_alu_op_f2u_rtp = 0x9F, 183 184 midgard_alu_op_ieq = 0xA0, 185 midgard_alu_op_ine = 0xA1, 186 midgard_alu_op_ult = 0xA2, 187 midgard_alu_op_ule = 0xA3, 188 midgard_alu_op_ilt = 0xA4, 189 midgard_alu_op_ile = 0xA5, 190 midgard_alu_op_iball_eq = 0xA8, 191 midgard_alu_op_iball_neq = 0xA9, 192 midgard_alu_op_uball_lt = 0xAA, 193 midgard_alu_op_uball_lte = 0xAB, 194 midgard_alu_op_iball_lt = 0xAC, 195 midgard_alu_op_iball_lte = 0xAD, 196 197 midgard_alu_op_ibany_eq = 0xB0, 198 midgard_alu_op_ibany_neq = 0xB1, 199 midgard_alu_op_ubany_lt = 0xB2, 200 midgard_alu_op_ubany_lte = 0xB3, 201 midgard_alu_op_ibany_lt = 0xB4, /* any(lessThan(.., ..)) */ 202 midgard_alu_op_ibany_lte = 0xB5, /* any(lessThanEqual(.., ..)) */ 203 midgard_alu_op_i2f_rte = 0xB8, 204 midgard_alu_op_i2f_rtz = 0xB9, 205 midgard_alu_op_i2f_rtn = 0xBA, 206 midgard_alu_op_i2f_rtp = 0xBB, 207 midgard_alu_op_u2f_rte = 0xBC, 208 midgard_alu_op_u2f_rtz = 0xBD, 209 midgard_alu_op_u2f_rtn = 0xBE, 210 midgard_alu_op_u2f_rtp = 0xBF, 211 212 /* All csel* instructions use as a condition the output of the previous 213 * vector or scalar unit, thus it must run on the second pipeline stage 214 * and be scheduled to the same bundle as the opcode that it uses as a 215 * condition. */ 216 midgard_alu_op_icsel_v = 0xC0, 217 midgard_alu_op_icsel = 0xC1, 218 midgard_alu_op_fcsel_v = 0xC4, 219 midgard_alu_op_fcsel = 0xC5, 220 midgard_alu_op_froundaway = 0xC6, /* round to nearest away */ 221 222 midgard_alu_op_fatan2_pt2 = 0xE8, 223 midgard_alu_op_fpow_pt1 = 0xEC, 224 midgard_alu_op_fpown_pt1 = 0xED, 225 midgard_alu_op_fpowr_pt1 = 0xEE, 226 227 midgard_alu_op_frcp = 0xF0, 228 midgard_alu_op_frsqrt = 0xF2, 229 midgard_alu_op_fsqrt = 0xF3, 230 midgard_alu_op_fexp2 = 0xF4, 231 midgard_alu_op_flog2 = 0xF5, 232 midgard_alu_op_fsinpi = 0xF6, /* sin(pi * x) */ 233 midgard_alu_op_fcospi = 0xF7, /* cos(pi * x) */ 234 midgard_alu_op_fatan2_pt1 = 0xF9, 235 } midgard_alu_op; 236 237 typedef enum { 238 midgard_outmod_none = 0, 239 midgard_outmod_clamp_0_inf = 1, /* max(x, 0.0), NaNs become +0.0 */ 240 midgard_outmod_clamp_m1_1 = 2, /* clamp(x, -1.0, 1.0), NaNs become -1.0 */ 241 midgard_outmod_clamp_0_1 = 3 /* clamp(x, 0.0, 1.0), NaNs become +0.0 */ 242 } midgard_outmod_float; 243 244 /* These are applied to the resulting value that's going to be stored in the dest reg. 245 * This should be set to midgard_outmod_keeplo when shrink_mode is midgard_shrink_mode_none. */ 246 typedef enum { 247 midgard_outmod_ssat = 0, 248 midgard_outmod_usat = 1, 249 midgard_outmod_keeplo = 2, /* Keep low half */ 250 midgard_outmod_keephi = 3, /* Keep high half */ 251 } midgard_outmod_int; 252 253 typedef enum { 254 midgard_reg_mode_8 = 0, 255 midgard_reg_mode_16 = 1, 256 midgard_reg_mode_32 = 2, 257 midgard_reg_mode_64 = 3 258 } midgard_reg_mode; 259 260 typedef enum { 261 midgard_shrink_mode_lower = 0, 262 midgard_shrink_mode_upper = 1, 263 midgard_shrink_mode_none = 2 264 } midgard_shrink_mode; 265 266 /* Only used if midgard_src_expand_mode is set to one of midgard_src_expand_*. */ 267 typedef enum { 268 midgard_int_sign_extend = 0, 269 midgard_int_zero_extend = 1, 270 midgard_int_replicate = 2, 271 midgard_int_left_shift = 3 272 } midgard_int_mod; 273 274 /* Unlike midgard_int_mod, fload modifiers are applied after the expansion happens, so 275 * they don't depend on midgard_src_expand_mode. */ 276 #define MIDGARD_FLOAT_MOD_ABS (1 << 0) 277 #define MIDGARD_FLOAT_MOD_NEG (1 << 1) 278 279 /* The expand options depend on both midgard_int_mod and midgard_reg_mode. For 280 * example, a vec4 with midgard_int_sign_extend and midgard_src_expand_low is 281 * treated as a vec8 and each 16-bit element from the low 64-bits is then sign 282 * extended, resulting in a vec4 where each 32-bit element corresponds to a 283 * 16-bit element from the low 64-bits of the input vector. */ 284 typedef enum { 285 midgard_src_passthrough = 0, 286 midgard_src_rep_low = 1, /* replicate lower 64 bits to higher 64 bits */ 287 midgard_src_rep_high = 2, /* replicate higher 64 bits to lower 64 bits */ 288 midgard_src_swap = 3, /* swap lower 64 bits with higher 64 bits */ 289 midgard_src_expand_low = 4, /* expand low 64 bits */ 290 midgard_src_expand_high = 5, /* expand high 64 bits */ 291 midgard_src_expand_low_swap = 6, /* expand low 64 bits, then swap */ 292 midgard_src_expand_high_swap = 7, /* expand high 64 bits, then swap */ 293 } midgard_src_expand_mode; 294 295 #define INPUT_EXPANDS(a) \ 296 (a >= midgard_src_expand_low && a <= midgard_src_expand_high_swap) 297 298 #define INPUT_SWAPS(a) \ 299 (a == midgard_src_swap || a >= midgard_src_expand_low_swap) 300 301 typedef struct 302 __attribute__((__packed__)) 303 { 304 /* Either midgard_int_mod or from midgard_float_mod_*, depending on the 305 * type of op */ 306 unsigned mod : 2; 307 midgard_src_expand_mode expand_mode : 3; 308 unsigned swizzle : 8; 309 } 310 midgard_vector_alu_src; 311 312 typedef struct 313 __attribute__((__packed__)) 314 { 315 midgard_alu_op op : 8; 316 midgard_reg_mode reg_mode : 2; 317 unsigned src1 : 13; 318 unsigned src2 : 13; 319 midgard_shrink_mode shrink_mode : 2; 320 unsigned outmod : 2; 321 unsigned mask : 8; 322 } 323 midgard_vector_alu; 324 325 typedef struct 326 __attribute__((__packed__)) 327 { 328 unsigned mod : 2; 329 bool full : 1; /* 0 = 16-bit, 1 = 32-bit */ 330 unsigned component : 3; 331 } 332 midgard_scalar_alu_src; 333 334 typedef struct 335 __attribute__((__packed__)) 336 { 337 midgard_alu_op op : 8; 338 unsigned src1 : 6; 339 /* last 5 bits are used when src2 is an immediate */ 340 unsigned src2 : 11; 341 unsigned unknown : 1; 342 unsigned outmod : 2; 343 bool output_full : 1; 344 unsigned output_component : 3; 345 } 346 midgard_scalar_alu; 347 348 typedef struct 349 __attribute__((__packed__)) 350 { 351 unsigned src1_reg : 5; 352 unsigned src2_reg : 5; 353 unsigned out_reg : 5; 354 bool src2_imm : 1; 355 } 356 midgard_reg_info; 357 358 /* In addition to conditional branches and jumps (unconditional branches), 359 * Midgard implements a bit of fixed function functionality used in fragment 360 * shaders via specially crafted branches. These have special branch opcodes, 361 * which perform a fixed-function operation and/or use the results of a 362 * fixed-function operation as the branch condition. */ 363 364 typedef enum { 365 /* Regular branches */ 366 midgard_jmp_writeout_op_branch_uncond = 1, 367 midgard_jmp_writeout_op_branch_cond = 2, 368 369 /* In a fragment shader, execute a discard_if instruction, with the 370 * corresponding condition code. Terminates the shader, so generally 371 * set the branch target to out of the shader */ 372 midgard_jmp_writeout_op_discard = 4, 373 374 /* Branch if the tilebuffer is not yet ready. At the beginning of a 375 * fragment shader that reads from the tile buffer, for instance via 376 * ARM_shader_framebuffer_fetch or EXT_pixel_local_storage, this branch 377 * operation should be used as a loop. An instruction like 378 * "br.tilebuffer.always -1" does the trick, corresponding to 379 * "while(!is_tilebuffer_ready) */ 380 midgard_jmp_writeout_op_tilebuffer_pending = 6, 381 382 /* In a fragment shader, try to write out the value pushed to r0 to the 383 * tilebuffer, subject to unknown state in r1.z and r1.w. If this 384 * succeeds, the shader terminates. If it fails, it branches to the 385 * specified branch target. Generally, this should be used in a loop to 386 * itself, acting as "do { write(r0); } while(!write_successful);" */ 387 midgard_jmp_writeout_op_writeout = 7, 388 } midgard_jmp_writeout_op; 389 390 typedef enum { 391 midgard_condition_write0 = 0, 392 393 /* These condition codes denote a conditional branch on FALSE and on 394 * TRUE respectively */ 395 midgard_condition_false = 1, 396 midgard_condition_true = 2, 397 398 /* This condition code always branches. For a pure branch, the 399 * unconditional branch coding should be used instead, but for 400 * fixed-function branch opcodes, this is still useful */ 401 midgard_condition_always = 3, 402 } midgard_condition; 403 404 typedef struct 405 __attribute__((__packed__)) 406 { 407 midgard_jmp_writeout_op op : 3; /* == branch_uncond */ 408 unsigned dest_tag : 4; /* tag of branch destination */ 409 unsigned unknown : 2; 410 int offset : 7; 411 } 412 midgard_branch_uncond; 413 414 typedef struct 415 __attribute__((__packed__)) 416 { 417 midgard_jmp_writeout_op op : 3; /* == branch_cond */ 418 unsigned dest_tag : 4; /* tag of branch destination */ 419 int offset : 7; 420 midgard_condition cond : 2; 421 } 422 midgard_branch_cond; 423 424 typedef struct 425 __attribute__((__packed__)) 426 { 427 midgard_jmp_writeout_op op : 3; /* == branch_cond */ 428 unsigned dest_tag : 4; /* tag of branch destination */ 429 unsigned unknown : 2; 430 signed offset : 23; 431 432 /* Extended branches permit inputting up to 4 conditions loaded into 433 * r31 (two in r31.w and two in r31.x). In the most general case, we 434 * specify a function f(A, B, C, D) mapping 4 1-bit conditions to a 435 * single 1-bit branch criteria. Note that the domain of f has 2^(2^4) 436 * elements, each mapping to 1-bit of output, so we can trivially 437 * construct a Godel numbering of f as a (2^4)=16-bit integer. This 438 * 16-bit integer serves as a lookup table to compute f, subject to 439 * some swaps for ordering. 440 * 441 * Interesting, the standard 2-bit condition codes are also a LUT with 442 * the same format (2^1-bit), but it's usually easier to use enums. */ 443 444 unsigned cond : 16; 445 } 446 midgard_branch_extended; 447 448 typedef struct 449 __attribute__((__packed__)) 450 { 451 midgard_jmp_writeout_op op : 3; /* == writeout */ 452 unsigned unknown : 13; 453 } 454 midgard_writeout; 455 456 /* 457 * Load/store words 458 */ 459 460 typedef enum { 461 midgard_op_ld_st_noop = 0x03, 462 463 /* Unpacks a colour from a native format to <format> */ 464 midgard_op_unpack_colour_f32 = 0x04, 465 midgard_op_unpack_colour_f16 = 0x05, 466 midgard_op_unpack_colour_u32 = 0x06, 467 midgard_op_unpack_colour_s32 = 0x07, 468 469 /* Packs a colour from <format> to a native format */ 470 midgard_op_pack_colour_f32 = 0x08, 471 midgard_op_pack_colour_f16 = 0x09, 472 midgard_op_pack_colour_u32 = 0x0A, 473 midgard_op_pack_colour_s32 = 0x0B, 474 475 /* Computes the effective address of a mem address expression */ 476 midgard_op_lea = 0x0C, 477 478 /* Converts image coordinates into mem address */ 479 midgard_op_lea_image = 0x0D, 480 481 /* Unclear why this is on the L/S unit, but moves fp32 cube map 482 * coordinates in r27 to its cube map texture coordinate destination 483 * (e.g r29). */ 484 485 midgard_op_ld_cubemap_coords = 0x0E, 486 487 /* A mov between registers that the ldst pipeline can access */ 488 midgard_op_ldst_mov = 0x10, 489 490 /* The L/S unit can do perspective division a clock faster than the ALU 491 * if you're lucky. Put the vec4 in r27, and call with 0x24 as the 492 * unknown state; the output will be <x/w, y/w, z/w, 1>. Replace w with 493 * z for the z version */ 494 midgard_op_ldst_perspective_div_y = 0x11, 495 midgard_op_ldst_perspective_div_z = 0x12, 496 midgard_op_ldst_perspective_div_w = 0x13, 497 498 /* val in r27.y, address embedded, outputs result to argument. Invert val for sub. Let val = +-1 for inc/dec. */ 499 midgard_op_atomic_add = 0x40, 500 midgard_op_atomic_add64 = 0x41, 501 midgard_op_atomic_add_be = 0x42, 502 midgard_op_atomic_add64_be = 0x43, 503 504 midgard_op_atomic_and = 0x44, 505 midgard_op_atomic_and64 = 0x45, 506 midgard_op_atomic_and_be = 0x46, 507 midgard_op_atomic_and64_be = 0x47, 508 midgard_op_atomic_or = 0x48, 509 midgard_op_atomic_or64 = 0x49, 510 midgard_op_atomic_or_be = 0x4A, 511 midgard_op_atomic_or64_be = 0x4B, 512 midgard_op_atomic_xor = 0x4C, 513 midgard_op_atomic_xor64 = 0x4D, 514 midgard_op_atomic_xor_be = 0x4E, 515 midgard_op_atomic_xor64_be = 0x4F, 516 517 midgard_op_atomic_imin = 0x50, 518 midgard_op_atomic_imin64 = 0x51, 519 midgard_op_atomic_imin_be = 0x52, 520 midgard_op_atomic_imin64_be = 0x53, 521 midgard_op_atomic_umin = 0x54, 522 midgard_op_atomic_umin64 = 0x55, 523 midgard_op_atomic_umin_be = 0x56, 524 midgard_op_atomic_umin64_be = 0x57, 525 midgard_op_atomic_imax = 0x58, 526 midgard_op_atomic_imax64 = 0x59, 527 midgard_op_atomic_imax_be = 0x5A, 528 midgard_op_atomic_imax64_be = 0x5B, 529 midgard_op_atomic_umax = 0x5C, 530 midgard_op_atomic_umax64 = 0x5D, 531 midgard_op_atomic_umax_be = 0x5E, 532 midgard_op_atomic_umax64_be = 0x5F, 533 534 midgard_op_atomic_xchg = 0x60, 535 midgard_op_atomic_xchg64 = 0x61, 536 midgard_op_atomic_xchg_be = 0x62, 537 midgard_op_atomic_xchg64_be = 0x63, 538 539 midgard_op_atomic_cmpxchg = 0x64, 540 midgard_op_atomic_cmpxchg64 = 0x65, 541 midgard_op_atomic_cmpxchg_be = 0x66, 542 midgard_op_atomic_cmpxchg64_be = 0x67, 543 544 /* Used for compute shader's __global arguments, __local 545 * variables (or for register spilling) */ 546 547 midgard_op_ld_u8 = 0x80, /* zero extends */ 548 midgard_op_ld_i8 = 0x81, /* sign extends */ 549 midgard_op_ld_u16 = 0x84, /* zero extends */ 550 midgard_op_ld_i16 = 0x85, /* sign extends */ 551 midgard_op_ld_u16_be = 0x86, /* zero extends, big endian */ 552 midgard_op_ld_i16_be = 0x87, /* sign extends, big endian */ 553 midgard_op_ld_32 = 0x88, /* short2, int, float */ 554 midgard_op_ld_32_bswap2 = 0x89, /* 16-bit big endian vector */ 555 midgard_op_ld_32_bswap4 = 0x8A, /* 32-bit big endian scalar */ 556 midgard_op_ld_64 = 0x8C, /* int2, float2, long */ 557 midgard_op_ld_64_bswap2 = 0x8D, /* 16-bit big endian vector */ 558 midgard_op_ld_64_bswap4 = 0x8E, /* 32-bit big endian vector */ 559 midgard_op_ld_64_bswap8 = 0x8F, /* 64-bit big endian scalar */ 560 midgard_op_ld_128 = 0x90, /* float4, long2 */ 561 midgard_op_ld_128_bswap2 = 0x91, /* 16-bit big endian vector */ 562 midgard_op_ld_128_bswap4 = 0x92, /* 32-bit big endian vector */ 563 midgard_op_ld_128_bswap8 = 0x93, /* 64-bit big endian vector */ 564 565 midgard_op_ld_attr_32 = 0x94, 566 midgard_op_ld_attr_16 = 0x95, 567 midgard_op_ld_attr_32u = 0x96, 568 midgard_op_ld_attr_32i = 0x97, 569 midgard_op_ld_vary_32 = 0x98, 570 midgard_op_ld_vary_16 = 0x99, 571 midgard_op_ld_vary_32u = 0x9A, 572 midgard_op_ld_vary_32i = 0x9B, 573 574 /* This instruction behaves differently depending if the gpu is a v4 575 * or a newer gpu. The main difference hinges on which values of the 576 * second argument are valid for each gpu. 577 * TODO: properly document and decode each possible value for the 578 * second argument. */ 579 midgard_op_ld_special_32f = 0x9C, 580 midgard_op_ld_special_16f = 0x9D, 581 midgard_op_ld_special_32u = 0x9E, 582 midgard_op_ld_special_32i = 0x9F, 583 584 /* The distinction between these ops is the alignment 585 * requirement / accompanying shift. Thus, the offset to 586 * ld_ubo_128 is in 16-byte units and can load 128-bit. The 587 * offset to ld_ubo_64 is in 8-byte units; ld_ubo_32 in 4-byte 588 * units. */ 589 midgard_op_ld_ubo_u8 = 0xA0, /* theoretical */ 590 midgard_op_ld_ubo_i8 = 0xA1, /* theoretical */ 591 midgard_op_ld_ubo_u16 = 0xA4, /* theoretical */ 592 midgard_op_ld_ubo_i16 = 0xA5, /* theoretical */ 593 midgard_op_ld_ubo_u16_be = 0xA6, /* theoretical */ 594 midgard_op_ld_ubo_i16_be = 0xA7, /* theoretical */ 595 midgard_op_ld_ubo_32 = 0xA8, 596 midgard_op_ld_ubo_32_bswap2 = 0xA9, 597 midgard_op_ld_ubo_32_bswap4 = 0xAA, 598 midgard_op_ld_ubo_64 = 0xAC, 599 midgard_op_ld_ubo_64_bswap2 = 0xAD, 600 midgard_op_ld_ubo_64_bswap4 = 0xAE, 601 midgard_op_ld_ubo_64_bswap8 = 0xAF, 602 midgard_op_ld_ubo_128 = 0xB0, 603 midgard_op_ld_ubo_128_bswap2 = 0xB1, 604 midgard_op_ld_ubo_128_bswap4 = 0xB2, 605 midgard_op_ld_ubo_128_bswap8 = 0xB3, 606 607 midgard_op_ld_image_32f = 0xB4, 608 midgard_op_ld_image_16f = 0xB5, 609 midgard_op_ld_image_32u = 0xB6, 610 midgard_op_ld_image_32i = 0xB7, 611 612 /* Only works on v5 or newer. 613 * Older cards must use ld_special with tilebuffer selectors. */ 614 midgard_op_ld_tilebuffer_32f = 0xB8, 615 midgard_op_ld_tilebuffer_16f = 0xB9, 616 midgard_op_ld_tilebuffer_raw = 0xBA, 617 618 midgard_op_st_u8 = 0xC0, /* zero extends */ 619 midgard_op_st_i8 = 0xC1, /* sign extends */ 620 midgard_op_st_u16 = 0xC4, /* zero extends */ 621 midgard_op_st_i16 = 0xC5, /* sign extends */ 622 midgard_op_st_u16_be = 0xC6, /* zero extends, big endian */ 623 midgard_op_st_i16_be = 0xC7, /* sign extends, big endian */ 624 midgard_op_st_32 = 0xC8, /* short2, int, float */ 625 midgard_op_st_32_bswap2 = 0xC9, /* 16-bit big endian vector */ 626 midgard_op_st_32_bswap4 = 0xCA, /* 32-bit big endian scalar */ 627 midgard_op_st_64 = 0xCC, /* int2, float2, long */ 628 midgard_op_st_64_bswap2 = 0xCD, /* 16-bit big endian vector */ 629 midgard_op_st_64_bswap4 = 0xCE, /* 32-bit big endian vector */ 630 midgard_op_st_64_bswap8 = 0xCF, /* 64-bit big endian scalar */ 631 midgard_op_st_128 = 0xD0, /* float4, long2 */ 632 midgard_op_st_128_bswap2 = 0xD1, /* 16-bit big endian vector */ 633 midgard_op_st_128_bswap4 = 0xD2, /* 32-bit big endian vector */ 634 midgard_op_st_128_bswap8 = 0xD3, /* 64-bit big endian vector */ 635 636 midgard_op_st_vary_32 = 0xD4, 637 midgard_op_st_vary_16 = 0xD5, 638 midgard_op_st_vary_32u = 0xD6, 639 midgard_op_st_vary_32i = 0xD7, 640 641 /* Value to st in r27, location r26.w as short2 */ 642 midgard_op_st_image_32f = 0xD8, 643 midgard_op_st_image_16f = 0xD9, 644 midgard_op_st_image_32u = 0xDA, 645 midgard_op_st_image_32i = 0xDB, 646 647 midgard_op_st_special_32f = 0xDC, 648 midgard_op_st_special_16f = 0xDD, 649 midgard_op_st_special_32u = 0xDE, 650 midgard_op_st_special_32i = 0xDF, 651 652 /* Only works on v5 or newer. 653 * Older cards must use ld_special with tilebuffer selectors. */ 654 midgard_op_st_tilebuffer_32f = 0xE8, 655 midgard_op_st_tilebuffer_16f = 0xE9, 656 midgard_op_st_tilebuffer_raw = 0xEA, 657 midgard_op_trap = 0xFC, 658 } midgard_load_store_op; 659 660 typedef enum { 661 midgard_interp_sample = 0, 662 midgard_interp_centroid = 1, 663 midgard_interp_default = 2 664 } midgard_interpolation; 665 666 typedef enum { 667 midgard_varying_mod_none = 0, 668 669 /* Take the would-be result and divide all components by its y/z/w 670 * (perspective division baked in with the load) */ 671 midgard_varying_mod_perspective_y = 1, 672 midgard_varying_mod_perspective_z = 2, 673 midgard_varying_mod_perspective_w = 3, 674 675 /* The result is a 64-bit cubemap descriptor to use with 676 * midgard_tex_op_normal or midgard_tex_op_gradient */ 677 midgard_varying_mod_cubemap = 4, 678 } midgard_varying_modifier; 679 680 typedef struct 681 __attribute__((__packed__)) 682 { 683 midgard_varying_modifier modifier : 3; 684 685 bool flat_shading : 1; 686 687 /* These are ignored if flat_shading is enabled. */ 688 bool perspective_correction : 1; 689 bool centroid_mapping : 1; 690 691 /* This is ignored if the shader only runs once per pixel. */ 692 bool interpolate_sample : 1; 693 694 bool zero0 : 1; /* Always zero */ 695 696 unsigned direct_sample_pos_x : 4; 697 unsigned direct_sample_pos_y : 4; 698 } 699 midgard_varying_params; 700 701 /* 8-bit register/etc selector for load/store ops */ 702 typedef struct 703 __attribute__((__packed__)) 704 { 705 /* Indexes into the register */ 706 unsigned component : 2; 707 708 /* Register select between r26/r27 */ 709 unsigned select : 1; 710 711 unsigned unknown : 2; 712 713 /* Like any good Arm instruction set, load/store arguments can be 714 * implicitly left-shifted... but only the second argument. Zero for no 715 * shifting, up to <<7 possible though. This is useful for indexing. 716 * 717 * For the first argument, it's unknown what these bits mean */ 718 unsigned shift : 3; 719 } 720 midgard_ldst_register_select; 721 722 typedef enum { 723 /* 0 is reserved */ 724 midgard_index_address_u64 = 1, 725 midgard_index_address_u32 = 2, 726 midgard_index_address_s32 = 3, 727 } midgard_index_address_format; 728 729 typedef struct 730 __attribute__((__packed__)) 731 { 732 midgard_load_store_op op : 8; 733 734 /* Source/dest reg */ 735 unsigned reg : 5; 736 737 /* Generally is a writemask. 738 * For ST_ATTR and ST_TEX, unused. 739 * For other stores, each bit masks 1/4th of the output. */ 740 unsigned mask : 4; 741 742 /* Swizzle for stores, but for atomics it encodes also the source 743 * register. This fits because atomics dont need a swizzle since they 744 * are not vectorized instructions. */ 745 unsigned swizzle : 8; 746 747 /* Arg reg, meaning changes according to each opcode */ 748 unsigned arg_comp : 2; 749 unsigned arg_reg : 3; 750 751 /* 64-bit address enable 752 * 32-bit data type enable for CUBEMAP and perspective div. 753 * Explicit indexing enable for LD_ATTR. 754 * 64-bit coordinate enable for LD_IMAGE. */ 755 bool bitsize_toggle : 1; 756 757 /* These are mainly used for opcodes that have addresses. 758 * For cmpxchg, index_reg is used for the comparison value. 759 * For ops that access the attrib table, bit 1 encodes which table. 760 * For LD_VAR and LD/ST_ATTR, bit 0 enables dest/src type inferral. */ 761 midgard_index_address_format index_format : 2; 762 unsigned index_comp : 2; 763 unsigned index_reg : 3; 764 unsigned index_shift : 4; 765 766 /* Generaly is a signed offset, but has different bitsize and starts at 767 * different bits depending on the opcode, LDST_*_DISPLACEMENT helpers 768 * are recommended when packing/unpacking this attribute. 769 * For LD_UBO, bit 0 enables ubo index immediate. 770 * For LD_TILEBUFFER_RAW, bit 0 disables sample index immediate. */ 771 int signed_offset : 18; 772 } 773 midgard_load_store_word; 774 775 typedef struct 776 __attribute__((__packed__)) 777 { 778 unsigned type : 4; 779 unsigned next_type : 4; 780 uint64_t word1 : 60; 781 uint64_t word2 : 60; 782 } 783 midgard_load_store; 784 785 /* 8-bit register selector used in texture ops to select a bias/LOD/gradient 786 * register, shoved into the `bias` field */ 787 788 typedef struct 789 __attribute__((__packed__)) 790 { 791 /* 32-bit register, clear for half-register */ 792 unsigned full : 1; 793 794 /* Register select between r28/r29 */ 795 unsigned select : 1; 796 797 /* For a half-register, selects the upper half */ 798 unsigned upper : 1; 799 800 /* Indexes into the register */ 801 unsigned component : 2; 802 803 /* Padding to make this 8-bit */ 804 unsigned zero : 3; 805 } 806 midgard_tex_register_select; 807 808 /* Texture pipeline results are in r28-r29 */ 809 #define REG_TEX_BASE 28 810 811 enum mali_texture_op { 812 /* [texture + LOD bias] 813 * If the texture is mipmapped, barriers must be enabled in the 814 * instruction word in order for this opcode to compute the output 815 * correctly. */ 816 midgard_tex_op_normal = 1, 817 818 /* [texture + gradient for LOD and anisotropy] 819 * Unlike midgard_tex_op_normal, this opcode does not require barriers 820 * to compute the output correctly. */ 821 midgard_tex_op_gradient = 2, 822 823 /* [unfiltered texturing] 824 * Unlike midgard_tex_op_normal, this opcode does not require barriers 825 * to compute the output correctly. */ 826 midgard_tex_op_fetch = 4, 827 828 /* [gradient from derivative] */ 829 midgard_tex_op_grad_from_derivative = 9, 830 831 /* [mov] */ 832 midgard_tex_op_mov = 10, 833 834 /* [noop] 835 * Mostly used for barriers. */ 836 midgard_tex_op_barrier = 11, 837 838 /* [gradient from coords] */ 839 midgard_tex_op_grad_from_coords = 12, 840 841 /* [derivative] 842 * Computes derivatives in 2x2 fragment blocks. */ 843 midgard_tex_op_derivative = 13 844 }; 845 846 enum mali_sampler_type { 847 /* 0 is reserved */ 848 MALI_SAMPLER_FLOAT = 0x1, /* sampler */ 849 MALI_SAMPLER_UNSIGNED = 0x2, /* usampler */ 850 MALI_SAMPLER_SIGNED = 0x3, /* isampler */ 851 }; 852 853 /* Texture modes */ 854 enum mali_texture_mode { 855 TEXTURE_NORMAL = 1, 856 TEXTURE_SHADOW = 5, 857 TEXTURE_GATHER_SHADOW = 6, 858 TEXTURE_GATHER_X = 8, 859 TEXTURE_GATHER_Y = 9, 860 TEXTURE_GATHER_Z = 10, 861 TEXTURE_GATHER_W = 11, 862 }; 863 864 enum mali_derivative_mode { 865 TEXTURE_DFDX = 0, 866 TEXTURE_DFDY = 1, 867 }; 868 869 typedef struct 870 __attribute__((__packed__)) 871 { 872 unsigned type : 4; 873 unsigned next_type : 4; 874 875 enum mali_texture_op op : 4; 876 unsigned mode : 4; 877 878 /* A little obscure, but last is set for the last texture operation in 879 * a shader. cont appears to just be last's opposite (?). Yeah, I know, 880 * kind of funky.. BiOpen thinks it could do with memory hinting, or 881 * tile locking? */ 882 883 unsigned cont : 1; 884 unsigned last : 1; 885 886 unsigned format : 2; 887 888 /* Are sampler_handle/texture_handler respectively set by registers? If 889 * true, the lower 8-bits of the respective field is a register word. 890 * If false, they are an immediate */ 891 892 unsigned sampler_register : 1; 893 unsigned texture_register : 1; 894 895 /* Is a register used to specify the 896 * LOD/bias/offset? If set, use the `bias` field as 897 * a register index. If clear, use the `bias` field 898 * as an immediate. */ 899 unsigned lod_register : 1; 900 901 /* Is a register used to specify an offset? If set, use the 902 * offset_reg_* fields to encode this, duplicated for each of the 903 * components. If clear, there is implcitly always an immediate offst 904 * specificed in offset_imm_* */ 905 unsigned offset_register : 1; 906 907 unsigned in_reg_full : 1; 908 unsigned in_reg_select : 1; 909 unsigned in_reg_upper : 1; 910 unsigned in_reg_swizzle : 8; 911 912 unsigned unknown8 : 2; 913 914 unsigned out_full : 1; 915 916 enum mali_sampler_type sampler_type : 2; 917 918 unsigned out_reg_select : 1; 919 unsigned out_upper : 1; 920 921 unsigned mask : 4; 922 923 /* Intriguingly, textures can take an outmod just like alu ops. Int 924 * outmods are not supported as far as I can tell, so this is only 925 * meaningful for float samplers */ 926 midgard_outmod_float outmod : 2; 927 928 unsigned swizzle : 8; 929 930 /* These indicate how many bundles after this texture op may be 931 * executed in parallel with this op. We may execute only ALU and 932 * ld/st in parallel (not other textures), and obviously there cannot 933 * be any dependency (the blob appears to forbid even accessing other 934 * channels of a given texture register). */ 935 936 unsigned out_of_order : 2; 937 unsigned unknown4 : 10; 938 939 /* In immediate mode, each offset field is an immediate range [0, 7]. 940 * 941 * In register mode, offset_x becomes a register (full, select, upper) 942 * triplet followed by a vec3 swizzle is splattered across 943 * offset_y/offset_z in a genuinely bizarre way. 944 * 945 * For texel fetches in immediate mode, the range is the full [-8, 7], 946 * but for normal texturing the top bit must be zero and a register 947 * used instead. It's not clear where this limitation is from. 948 * 949 * union { 950 * struct { 951 * signed offset_x : 4; 952 * signed offset_y : 4; 953 * signed offset_z : 4; 954 * } immediate; 955 * struct { 956 * bool full : 1; 957 * bool select : 1; 958 * bool upper : 1; 959 * unsigned swizzle : 8; 960 * unsigned zero : 1; 961 * } register; 962 * } 963 */ 964 965 unsigned offset : 12; 966 967 /* In immediate bias mode, for a normal texture op, this is 968 * texture bias, computed as int(2^8 * frac(biasf)), with 969 * bias_int = floor(bias). For a textureLod, it's that, but 970 * s/bias/lod. For a texel fetch, this is the LOD as-is. 971 * 972 * In register mode, this is a midgard_tex_register_select 973 * structure and bias_int is zero */ 974 975 unsigned bias : 8; 976 signed bias_int : 8; 977 978 /* If sampler/texture_register is set, the bottom 8-bits are 979 * midgard_tex_register_select and the top 8-bits are zero. If they are 980 * clear, they are immediate texture indices */ 981 982 unsigned sampler_handle : 16; 983 unsigned texture_handle : 16; 984 } 985 midgard_texture_word; 986 987 /* Technically barriers are texture instructions but it's less work to add them 988 * as an explicitly zeroed special case, since most fields are forced to go to 989 * zero */ 990 991 typedef struct 992 __attribute__((__packed__)) 993 { 994 unsigned type : 4; 995 unsigned next_type : 4; 996 997 /* op = TEXTURE_OP_BARRIER */ 998 unsigned op : 6; 999 unsigned zero1 : 2; 1000 1001 /* Since helper invocations don't make any sense, these are forced to one */ 1002 unsigned cont : 1; 1003 unsigned last : 1; 1004 unsigned zero2 : 14; 1005 1006 unsigned zero3 : 24; 1007 unsigned out_of_order : 4; 1008 unsigned zero4 : 4; 1009 1010 uint64_t zero5; 1011 } midgard_texture_barrier_word; 1012 1013 typedef union midgard_constants { 1014 double f64[2]; 1015 uint64_t u64[2]; 1016 int64_t i64[2]; 1017 float f32[4]; 1018 uint32_t u32[4]; 1019 int32_t i32[4]; 1020 uint16_t f16[8]; 1021 uint16_t u16[8]; 1022 int16_t i16[8]; 1023 uint8_t u8[16]; 1024 int8_t i8[16]; 1025 } 1026 midgard_constants; 1027 1028 enum midgard_roundmode { 1029 MIDGARD_RTE = 0x0, /* round to even */ 1030 MIDGARD_RTZ = 0x1, /* round to zero */ 1031 MIDGARD_RTN = 0x2, /* round to negative */ 1032 MIDGARD_RTP = 0x3, /* round to positive */ 1033 }; 1034 1035 #endif 1036