src/panfrost/bifrost/compiler.h

   1 /*
   2  * Copyright (C) 2020 Collabora Ltd.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors (Collabora):
  24  *      Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  25  */
  26
  27 #ifndef __BIFROST_COMPILER_H
  28 #define __BIFROST_COMPILER_H
  29
  30 #include "bifrost.h"
  31 #include "compiler/nir/nir.h"
  32
  33 /* Bifrost opcodes are tricky -- the same op may exist on both FMA and
  34  * ADD with two completely different opcodes, and opcodes can be varying
  35  * length in some cases. Then we have different opcodes for int vs float
  36  * and then sometimes even for different typesizes. Further, virtually
  37  * every op has a number of flags which depend on the op. In constrast
  38  * to Midgard where you have a strict ALU/LDST/TEX division and within
  39  * ALU you have strict int/float and that's it... here it's a *lot* more
  40  * involved. As such, we use something much higher level for our IR,
  41  * encoding "classes" of operations, letting the opcode details get
  42  * sorted out at emit time.
  43  *
  44  * Please keep this list alphabetized. Please use a dictionary if you
  45  * don't know how to do that.
  46  */
  47
  48 enum bi_class {
  49         BI_ADD,
  50         BI_ATEST,
  51         BI_BRANCH,
  52         BI_CMP,
  53         BI_BLEND,
  54         BI_BITWISE,
  55         BI_CONVERT,
  56         BI_CSEL,
  57         BI_DISCARD,
  58         BI_FMA,
  59         BI_FREXP,
  60         BI_LOAD,
  61         BI_LOAD_ATTR,
  62         BI_LOAD_VAR,
  63         BI_LOAD_VAR_ADDRESS,
  64         BI_MINMAX,
  65         BI_MOV,
  66         BI_SHIFT,
  67         BI_STORE,
  68         BI_STORE_VAR,
  69         BI_SPECIAL, /* _FAST, _TABLE on supported GPUs */
  70         BI_TEX,
  71         BI_ROUND,
  72         BI_NUM_CLASSES
  73 };
  74
  75 /* Properties of a class... */
  76 extern unsigned bi_class_props[BI_NUM_CLASSES];
  77
  78 /* abs/neg/outmod valid for a float op */
  79 #define BI_MODS (1 << 0)
  80
  81 /* Generic enough that little class-specific information is required. In other
  82  * words, it acts as a "normal" ALU op, even if the encoding ends up being
  83  * irregular enough to warrant a separate class */
  84 #define BI_GENERIC (1 << 1)
  85
  86 /* Accepts a bifrost_roundmode */
  87 #define BI_ROUNDMODE (1 << 2)
  88
  89 /* Can be scheduled to FMA */
  90 #define BI_SCHED_FMA (1 << 3)
  91
  92 /* Can be scheduled to ADD */
  93 #define BI_SCHED_ADD (1 << 4)
  94
  95 /* Most ALU ops can do either, actually */
  96 #define BI_SCHED_ALL (BI_SCHED_FMA | BI_SCHED_ADD)
  97
  98 /* It can't get any worse than csel4... can it? */
  99 #define BIR_SRC_COUNT 4
 100
 101 /* Class-specific data for BI_LD_ATTR, BI_LD_VAR_ADDR */
 102 struct bi_load {
 103         /* Note: no indirects here */
 104         unsigned location;
 105
 106         /* Only for BI_LD_ATTR. But number of vector channels */
 107         unsigned channels;
 108 };
 109
 110 /* BI_LD_VARY */
 111 struct bi_load_vary {
 112         /* All parameters used here. Indirect location specified in
 113          * src1 and ignoring location, if present. */
 114         struct bi_load load;
 115
 116         enum bifrost_interp_mode interp_mode;
 117         bool reuse;
 118         bool flat;
 119 };
 120
 121 /* Opcodes within a class */
 122 enum bi_minmax_op {
 123         BI_MINMAX_MIN,
 124         BI_MINMAX_MAX
 125 };
 126
 127 enum bi_bitwise_op {
 128         BI_BITWISE_AND,
 129         BI_BITWISE_OR,
 130         BI_BITWISE_XOR
 131 };
 132
 133 enum bi_round_op {
 134         BI_ROUND_MODE, /* use round mode */
 135         BI_ROUND_ROUND /* i.e.: fround() */
 136 };
 137
 138 typedef struct {
 139         struct list_head link; /* Must be first */
 140         enum bi_class type;
 141
 142         /* Indices, see bir_ssa_index etc. Note zero is special cased
 143          * to "no argument" */
 144         unsigned dest;
 145         unsigned src[BIR_SRC_COUNT];
 146
 147         /* If one of the sources has BIR_INDEX_CONSTANT... */
 148         union {
 149                 uint64_t u64;
 150                 uint32_t u32;
 151                 uint16_t u16[2];
 152                 uint8_t u8[4];
 153         } constant;
 154
 155         /* Floating-point modifiers, type/class permitting. If not
 156          * allowed for the type/class, these are ignored. */
 157         enum bifrost_outmod outmod;
 158         bool src_abs[BIR_SRC_COUNT];
 159         bool src_neg[BIR_SRC_COUNT];
 160
 161         /* Round mode (requires BI_ROUNDMODE) */
 162         enum bifrost_roundmode roundmode;
 163
 164         /* Destination type. Usually the type of the instruction
 165          * itself, but if sources and destination have different
 166          * types, the type of the destination wins (so f2i would be
 167          * int). Zero if there is no destination. Bitsize included */
 168         nir_alu_type dest_type;
 169
 170         /* A class-specific op from which the actual opcode can be derived
 171          * (along with the above information) */
 172
 173         union {
 174                 enum bi_minmax_op minmax;
 175                 enum bi_bitwise_op bitwise;
 176                 enum bi_round_op round;
 177         } op;
 178
 179         /* Union for class-specific information */
 180         union {
 181                 enum bifrost_minmax_mode minmax;
 182                 struct bi_load load;
 183                 struct bi_load_vary load_vary;
 184         };
 185 } bi_instruction;
 186
 187 /* Scheduling takes place in two steps. Step 1 groups instructions within a
 188  * block into distinct clauses (bi_clause). Step 2 schedules instructions
 189  * within a clause into FMA/ADD pairs (bi_bundle).
 190  *
 191  * A bi_bundle contains two paired instruction pointers. If a slot is unfilled,
 192  * leave it NULL; the emitter will fill in a nop.
 193  */
 194
 195 typedef struct {
 196         bi_instruction *fma;
 197         bi_instruction *add;
 198 } bi_bundle;
 199
 200 typedef struct {
 201         struct list_head link;
 202
 203         /* A clause can have 8 instructions in bundled FMA/ADD sense, so there
 204          * can be 8 bundles. But each bundle can have both an FMA and an ADD,
 205          * so a clause can have up to 16 bi_instructions. Whether bundles or
 206          * instructions are used depends on where in scheduling we are. */
 207
 208         unsigned instruction_count;
 209         unsigned bundle_count;
 210
 211         union {
 212                 bi_instruction *instructions[16];
 213                 bi_bundle bundles[8];
 214         };
 215 } bi_clause;
 216
 217 typedef struct bi_block {
 218         struct list_head link; /* must be first */
 219         unsigned name; /* Just for pretty-printing */
 220
 221         /* If true, uses clauses; if false, uses instructions */
 222         bool scheduled;
 223
 224         union {
 225                 struct list_head instructions; /* pre-schedule, list of bi_instructions */
 226                 struct list_head clauses; /* list of bi_clause */
 227         };
 228 } bi_block;
 229
 230 typedef struct {
 231        nir_shader *nir;
 232        struct list_head blocks; /* list of bi_block */
 233 } bi_context;
 234
 235 /* So we can distinguish between SSA/reg/sentinel quickly */
 236 #define BIR_NO_ARG (0)
 237 #define BIR_IS_REG (1)
 238
 239 /* If high bits are set, instead of SSA/registers, we have specials indexed by
 240  * the low bits if necessary.
 241  *
 242  *  Fixed register: do not allocate register, do not collect $200.
 243  *  Uniform: access a uniform register given by low bits.
 244  *  Constant: access the specified constant
 245  *  Zero: special cased to avoid wasting a constant
 246  */
 247
 248 #define BIR_INDEX_REGISTER (1 << 31)
 249 #define BIR_INDEX_UNIFORM  (1 << 30)
 250 #define BIR_INDEX_CONSTANT (1 << 29)
 251 #define BIR_INDEX_ZERO     (1 << 28)
 252
 253 /* Keep me synced please so we can check src & BIR_SPECIAL */
 254
 255 #define BIR_SPECIAL        ((BIR_INDEX_REGISTER | BIR_INDEX_UNIFORM) | \
 256         (BIR_INDEX_CONSTANT | BIR_INDEX_ZERO)
 257
 258 static inline unsigned
 259 bir_ssa_index(nir_ssa_def *ssa)
 260 {
 261         /* Off-by-one ensures BIR_NO_ARG is skipped */
 262         return ((ssa->index + 1) << 1) | 0;
 263 }
 264
 265 static inline unsigned
 266 bir_src_index(nir_src *src)
 267 {
 268         if (src->is_ssa)
 269                 return bir_ssa_index(src->ssa);
 270         else {
 271                 assert(!src->reg.indirect);
 272                 return (src->reg.reg->index << 1) | BIR_IS_REG;
 273         }
 274 }
 275
 276 static inline unsigned
 277 bir_dest_index(nir_dest *dst)
 278 {
 279         if (dst->is_ssa)
 280                 return bir_ssa_index(&dst->ssa);
 281         else {
 282                 assert(!dst->reg.indirect);
 283                 return (dst->reg.reg->index << 1) | BIR_IS_REG;
 284         }
 285 }
 286
 287 #endif