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 struct bifrost_header {
  33         unsigned unk0 : 7;
  34         // If true, convert any infinite result of any floating-point operation to
  35         // the biggest representable number.
  36         unsigned suppress_inf: 1;
  37         // Convert any NaN results to 0.
  38         unsigned suppress_nan : 1;
  39         unsigned unk1 : 2;
  40         // true if the execution mask of the next clause is the same as the mask of
  41         // the current clause.
  42         unsigned back_to_back : 1;
  43         unsigned no_end_of_shader: 1;
  44         unsigned unk2 : 2;
  45         // Set to true for fragment shaders, to implement this bit of spec text
  46         // from section 7.1.5 of the GLSL ES spec:
  47         //
  48         // "Stores to image and buffer variables performed by helper invocations
  49         // have no effect on the underlying image or buffer memory."
  50         //
  51         // Helper invocations are threads (invocations) corresponding to pixels in
  52         // a quad that aren't actually part of the triangle, but are included to
  53         // make derivatives work correctly. They're usually turned on, but they
  54         // need to be masked off for GLSL-level stores. This bit seems to be the
  55         // only bit that's actually different between fragment shaders and other
  56         // shaders, so this is probably what it's doing.
  57         unsigned elide_writes : 1;
  58         // If backToBack is off:
  59         // - true for conditional branches and fallthrough
  60         // - false for unconditional branches
  61         // The blob seems to always set it to true if back-to-back is on.
  62         unsigned branch_cond : 1;
  63         // This bit is set when the next clause writes to the data register of some
  64         // previous clause.
  65         unsigned datareg_writebarrier: 1;
  66         unsigned datareg : 6;
  67         unsigned scoreboard_deps: 8;
  68         unsigned scoreboard_index: 3;
  69         unsigned clause_type: 4;
  70         unsigned unk3 : 1; // part of clauseType?
  71         unsigned next_clause_type: 4;
  72         unsigned unk4 : 1; // part of nextClauseType?
  73 } __attribute__((packed));
  74
  75 enum bifrost_packed_src {
  76         BIFROST_SRC_PORT0    = 0,
  77         BIFROST_SRC_PORT1    = 1,
  78         BIFROST_SRC_PORT3    = 2,
  79         BIFROST_SRC_STAGE    = 3,
  80         BIFROST_SRC_CONST_LO = 4,
  81         BIFROST_SRC_CONST_HI = 5,
  82         BIFROST_SRC_PASS_FMA = 6,
  83         BIFROST_SRC_PASS_ADD = 7,
  84 };
  85
  86 struct bifrost_fma_inst {
  87         unsigned src0 : 3;
  88         unsigned op   : 20;
  89 } __attribute__((packed));
  90
  91 struct bifrost_add_inst {
  92         unsigned src0 : 3;
  93         unsigned op   : 17;
  94 } __attribute__((packed));
  95
  96 enum bifrost_outmod {
  97         BIFROST_NONE = 0x0,
  98         BIFROST_POS = 0x1,
  99         BIFROST_SAT_SIGNED = 0x2,
 100         BIFROST_SAT = 0x3,
 101 };
 102
 103 enum bifrost_roundmode {
 104         BIFROST_RTE = 0x0, /* round to even */
 105         BIFROST_RTP = 0x1, /* round to positive */
 106         BIFROST_RTN = 0x2, /* round to negative */
 107         BIFROST_RTZ = 0x3 /* round to zero */
 108 };
 109
 110 /* NONE: Same as fmax() and fmin() -- return the other
 111  * number if any number is NaN.  Also always return +0 if
 112  * one argument is +0 and the other is -0.
 113  *
 114  * NAN_WINS: Instead of never returning a NaN, always return
 115  * one. The "greater"/"lesser" NaN is always returned, first
 116  * by checking the sign and then the mantissa bits.
 117  *
 118  * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
 119  * For min, implement src0 < src1 ? src0 : src1.  This
 120  * includes handling NaN's and signedness of 0 differently
 121  * from above, since +0 and -0 compare equal and comparisons
 122  * always return false for NaN's. As a result, this mode is
 123  * *not* commutative.
 124  *
 125  * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
 126  * For min, implement src0 > src1 ? src1 : src0
 127  */
 128
 129
 130 enum bifrost_minmax_mode {
 131         BIFROST_MINMAX_NONE = 0x0,
 132         BIFROST_NAN_WINS    = 0x1,
 133         BIFROST_SRC1_WINS   = 0x2,
 134         BIFROST_SRC0_WINS   = 0x3,
 135 };
 136
 137 struct bifrost_fma_add {
 138         unsigned src0 : 3;
 139         unsigned src1 : 3;
 140         unsigned src1_abs : 1;
 141         unsigned src0_neg : 1;
 142         unsigned src1_neg : 1;
 143         unsigned unk : 3;
 144         unsigned src0_abs : 1;
 145         enum bifrost_outmod outmod : 2;
 146         enum bifrost_roundmode roundmode : 2;
 147         unsigned op : 6;
 148 } __attribute__((packed));
 149
 150 #define BIFROST_FMA_OP_FMA (0x00)
 151
 152 struct bifrost_fma_fma {
 153         unsigned src0 : 3;
 154         unsigned src1 : 3;
 155         unsigned src2 : 3;
 156         unsigned src_expand : 3;
 157         unsigned src0_abs : 1;
 158         unsigned unk : 4;
 159         unsigned src0_neg : 1;
 160         unsigned src2_neg : 1;
 161         unsigned src1_abs : 1;
 162         unsigned src2_abs : 1;
 163         unsigned op : 5;
 164 } __attribute__((packed));
 165
 166 enum bifrost_csel_cond {
 167         BIFROST_FEQ_F = 0x0,
 168         BIFROST_FGT_F = 0x1,
 169         BIFROST_FGE_F = 0x2,
 170         BIFROST_IEQ_F = 0x3,
 171         BIFROST_IGT_I = 0x4,
 172         BIFROST_IGE_I = 0x5,
 173         BIFROST_UGT_I = 0x6,
 174         BIFROST_UGE_I = 0x7
 175 };
 176
 177 struct bifrost_csel4 {
 178         unsigned src0 : 3;
 179         unsigned src1 : 3;
 180         unsigned src2 : 3;
 181         unsigned src3 : 3;
 182         enum bifrost_csel_cond cond : 3;
 183         unsigned op   : 8;
 184 } __attribute__((packed));
 185
 186 struct bifrost_shift_fma {
 187         unsigned src0 : 3;
 188         unsigned src1 : 3;
 189         unsigned src2 : 3;
 190         unsigned half : 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
 191         unsigned unk  : 1; /* always set? */
 192         unsigned invert_1 : 1; /* Inverts sources to combining op */
 193         /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
 194         unsigned invert_2 : 1;
 195         unsigned op : 8;
 196 } __attribute__((packed));
 197
 198 struct bifrost_shift_add {
 199         unsigned src0 : 3;
 200         unsigned src1 : 3;
 201         unsigned src2 : 3;
 202         unsigned zero : 2;
 203
 204         unsigned invert_1 : 1;
 205         unsigned invert_2 : 1;
 206
 207         unsigned op : 7;
 208 } __attribute__((packed));
 209
 210 enum bifrost_ldst_type {
 211         BIFROST_LDST_F16 = 0,
 212         BIFROST_LDST_F32 = 1,
 213         BIFROST_LDST_I32 = 2,
 214         BIFROST_LDST_U32 = 3
 215 };
 216
 217 struct bifrost_ld_var_addr {
 218         unsigned src0 : 3;
 219         unsigned src1 : 3;
 220         unsigned location : 5;
 221         enum bifrost_ldst_type type : 2;
 222         unsigned op : 7;
 223 } __attribute__((packed));
 224
 225 struct bifrost_ld_attr {
 226         unsigned src0 : 3;
 227         unsigned src1 : 3;
 228         unsigned location : 5;
 229         unsigned channels : 2; /* MALI_POSITIVE */
 230         enum bifrost_ldst_type type : 2;
 231         unsigned op : 5;
 232 } __attribute__((packed));
 233
 234 enum bifrost_interp_mode {
 235         BIFROST_INTERP_PER_FRAG = 0x0,
 236         BIFROST_INTERP_CENTROID = 0x1,
 237         BIFROST_INTERP_DEFAULT  = 0x2,
 238         BIFROST_INTERP_EXPLICIT = 0x3
 239 };
 240
 241 struct bifrost_ld_var {
 242         unsigned src0 : 3;
 243
 244         /* If top two bits set, indirect with src in bottom three */
 245         unsigned addr : 5;
 246
 247         unsigned channels : 2; /* MALI_POSITIVE */
 248         enum bifrost_interp_mode interp_mode : 2;
 249         unsigned reuse : 1;
 250         unsigned flat : 1;
 251         unsigned op : 6;
 252 } __attribute__((packed));
 253
 254 struct bifrost_tex_ctrl {
 255         unsigned sampler_index : 4; // also used to signal indirects
 256         unsigned tex_index : 7;
 257         bool no_merge_index : 1; // whether to merge (direct) sampler & texture indices
 258         bool filter : 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
 259         unsigned unk0 : 2;
 260         bool texel_offset : 1; // *Offset()
 261         bool is_shadow : 1;
 262         bool is_array : 1;
 263         unsigned tex_type : 2; // 2D, 3D, Cube, Buffer
 264         bool compute_lod : 1; // 0 for *Lod()
 265         bool not_supply_lod : 1; // 0 for *Lod() or when a bias is applied
 266         bool calc_gradients : 1; // 0 for *Grad()
 267         unsigned unk1 : 1;
 268         unsigned result_type : 4; // integer, unsigned, float TODO: why is this 4 bits?
 269         unsigned unk2 : 4;
 270 } __attribute__((packed));
 271
 272 struct bifrost_dual_tex_ctrl {
 273         unsigned sampler_index0 : 2;
 274         unsigned unk0 : 2;
 275         unsigned tex_index0 : 2;
 276         unsigned sampler_index1 : 2;
 277         unsigned tex_index1 : 2;
 278         unsigned unk1 : 22;
 279 } __attribute__((packed));
 280
 281 enum branch_bit_size {
 282         BR_SIZE_32 = 0,
 283         BR_SIZE_16XX = 1,
 284         BR_SIZE_16YY = 2,
 285         // For the above combinations of bitsize and location, an extra bit is
 286         // encoded via comparing the sources. The only possible source of ambiguity
 287         // would be if the sources were the same, but then the branch condition
 288         // would be always true or always false anyways, so we can ignore it. But
 289         // this no longer works when comparing the y component to the x component,
 290         // since it's valid to compare the y component of a source against its own
 291         // x component. Instead, the extra bit is encoded via an extra bitsize.
 292         BR_SIZE_16YX0 = 3,
 293         BR_SIZE_16YX1 = 4,
 294         BR_SIZE_32_AND_16X = 5,
 295         BR_SIZE_32_AND_16Y = 6,
 296         // Used for comparisons with zero and always-true, see below. I think this
 297         // only works for integer comparisons.
 298         BR_SIZE_ZERO = 7,
 299 };
 300
 301 enum bifrost_reg_write_unit {
 302         REG_WRITE_NONE = 0, // don't write
 303         REG_WRITE_TWO, // write using reg2
 304         REG_WRITE_THREE, // write using reg3
 305 };
 306
 307 struct bifrost_regs {
 308         unsigned uniform_const : 8;
 309         unsigned reg2 : 6;
 310         unsigned reg3 : 6;
 311         unsigned reg0 : 5;
 312         unsigned reg1 : 6;
 313         unsigned ctrl : 4;
 314 } __attribute__((packed));
 315
 316 enum bifrost_branch_cond {
 317         BR_COND_LT = 0,
 318         BR_COND_LE = 1,
 319         BR_COND_GE = 2,
 320         BR_COND_GT = 3,
 321         // Equal vs. not-equal determined by src0/src1 comparison
 322         BR_COND_EQ = 4,
 323         // floating-point comparisons
 324         // Becomes UNE when you flip the arguments
 325         BR_COND_OEQ = 5,
 326         // TODO what happens when you flip the arguments?
 327         BR_COND_OGT = 6,
 328         BR_COND_OLT = 7,
 329 };
 330
 331 enum bifrost_branch_code {
 332         BR_ALWAYS = 63,
 333 };
 334
 335 struct bifrost_branch {
 336         unsigned src0 : 3;
 337
 338         /* For BR_SIZE_ZERO, upper two bits become ctrl */
 339         unsigned src1 : 3;
 340
 341         /* Offset source -- always uniform/const but
 342          * theoretically could support indirect jumps? */
 343         unsigned src2 : 3;
 344
 345         enum bifrost_branch_cond cond : 3;
 346         enum branch_bit_size size : 3;
 347
 348         unsigned op : 5;
 349 };
 350
 351 /* Clause packing */
 352
 353 #define BIFROST_FMA_NOP (0x701960)
 354 #define BIFROST_ADD_NOP (0x3D960)
 355
 356 struct bifrost_fmt1 {
 357         unsigned ins_0 : 3;
 358         unsigned tag : 5;
 359         uint64_t ins_1 : 64;
 360         unsigned ins_2 : 11;
 361         uint64_t header : 45;
 362 } __attribute__((packed));
 363
 364 #define BIFROST_FMT1_INSTRUCTIONS    0b00101
 365 #define BIFROST_FMT1_FINAL           0b01001
 366 #define BIFROST_FMT1_CONSTANTS       0b00001
 367
 368 enum bifrost_reg_control {
 369         BIFROST_WRITE_FMA_P2         = 1,
 370         BIFROST_WRITE_FMA_P2_READ_P3 = 2,
 371         BIFROST_WRITE_FMA_P2_READ_P3_ALT = 3,
 372         BIFROST_READ_P3              = 4,
 373         BIFROST_WRITE_ADD_P2         = 5,
 374         BIFROST_WRITE_ADD_P2_READ_P3 = 6,
 375         BIFROST_WRITE_ADD_P2_FMA_P3  = 7,
 376
 377         BIFROST_FIRST_NONE           = 8,
 378         BIFROST_FIRST_WRITE_FMA_P2   = 9,
 379         BIFROST_REG_NONE             = 11,
 380         BIFROST_FIRST_READ_P3        = 12,
 381         BIFROST_FIRST_WRITE_ADD_P2   = 13,
 382         BIFROST_FIRST_WRITE_ADD_P2_READ_P3 = 14,
 383         BIFROST_FIRST_WRITE_ADD_P2_FMA_P3  = 15
 384 };
 385
 386 #endif