src/amd/compiler/aco_ir.cpp

   1 /*
   2  * Copyright © 2020 Valve Corporation
   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
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24 #include "aco_ir.h"
  25 #include "vulkan/radv_shader.h"
  26 #include "c11/threads.h"
  27 #include "util/debug.h"
  28
  29 namespace aco {
  30
  31 uint64_t debug_flags = 0;
  32
  33 static const struct debug_control aco_debug_options[] = {
  34    {"validateir", DEBUG_VALIDATE_IR},
  35    {"validatera", DEBUG_VALIDATE_RA},
  36    {"perfwarn", DEBUG_PERFWARN},
  37    {NULL, 0}
  38 };
  39
  40 static once_flag init_once_flag = ONCE_FLAG_INIT;
  41
  42 static void init_once()
  43 {
  44    debug_flags = parse_debug_string(getenv("ACO_DEBUG"), aco_debug_options);
  45
  46    #ifndef NDEBUG
  47    /* enable some flags by default on debug builds */
  48    debug_flags |= aco::DEBUG_VALIDATE_IR;
  49    #endif
  50 }
  51
  52 void init()
  53 {
  54    call_once(&init_once_flag, init_once);
  55 }
  56
  57 void init_program(Program *program, Stage stage, struct radv_shader_info *info,
  58                   enum chip_class chip_class, enum radeon_family family,
  59                   ac_shader_config *config)
  60 {
  61    program->stage = stage;
  62    program->config = config;
  63    program->info = info;
  64    program->chip_class = chip_class;
  65    if (family == CHIP_UNKNOWN) {
  66       switch (chip_class) {
  67       case GFX6:
  68          program->family = CHIP_TAHITI;
  69          break;
  70       case GFX7:
  71          program->family = CHIP_BONAIRE;
  72          break;
  73       case GFX8:
  74          program->family = CHIP_POLARIS10;
  75          break;
  76       case GFX9:
  77          program->family = CHIP_VEGA10;
  78          break;
  79       case GFX10:
  80          program->family = CHIP_NAVI10;
  81          break;
  82       default:
  83          program->family = CHIP_UNKNOWN;
  84          break;
  85       }
  86    } else {
  87       program->family = family;
  88    }
  89    program->wave_size = info->wave_size;
  90    program->lane_mask = program->wave_size == 32 ? s1 : s2;
  91
  92    program->lds_alloc_granule = chip_class >= GFX7 ? 512 : 256;
  93    program->lds_limit = chip_class >= GFX7 ? 65536 : 32768;
  94    /* apparently gfx702 also has 16-bank LDS but I can't find a family for that */
  95    program->has_16bank_lds = family == CHIP_KABINI || family == CHIP_STONEY;
  96
  97    program->vgpr_limit = 256;
  98    program->vgpr_alloc_granule = 3;
  99
 100    if (chip_class >= GFX10) {
 101       program->physical_sgprs = 2560; /* doesn't matter as long as it's at least 128 * 20 */
 102       program->sgpr_alloc_granule = 127;
 103       program->sgpr_limit = 106;
 104       if (chip_class >= GFX10_3)
 105          program->vgpr_alloc_granule = program->wave_size == 32 ? 15 : 7;
 106       else
 107          program->vgpr_alloc_granule = program->wave_size == 32 ? 7 : 3;
 108    } else if (program->chip_class >= GFX8) {
 109       program->physical_sgprs = 800;
 110       program->sgpr_alloc_granule = 15;
 111       if (family == CHIP_TONGA || family == CHIP_ICELAND)
 112          program->sgpr_limit = 94; /* workaround hardware bug */
 113       else
 114          program->sgpr_limit = 102;
 115    } else {
 116       program->physical_sgprs = 512;
 117       program->sgpr_alloc_granule = 7;
 118       program->sgpr_limit = 104;
 119    }
 120
 121    program->next_fp_mode.preserve_signed_zero_inf_nan32 = false;
 122    program->next_fp_mode.preserve_signed_zero_inf_nan16_64 = false;
 123    program->next_fp_mode.must_flush_denorms32 = false;
 124    program->next_fp_mode.must_flush_denorms16_64 = false;
 125    program->next_fp_mode.care_about_round32 = false;
 126    program->next_fp_mode.care_about_round16_64 = false;
 127    program->next_fp_mode.denorm16_64 = fp_denorm_keep;
 128    program->next_fp_mode.denorm32 = 0;
 129    program->next_fp_mode.round16_64 = fp_round_ne;
 130    program->next_fp_mode.round32 = fp_round_ne;
 131 }
 132
 133 memory_sync_info get_sync_info(const Instruction* instr)
 134 {
 135    switch (instr->format) {
 136    case Format::SMEM:
 137       return static_cast<const SMEM_instruction*>(instr)->sync;
 138    case Format::MUBUF:
 139       return static_cast<const MUBUF_instruction*>(instr)->sync;
 140    case Format::MIMG:
 141       return static_cast<const MIMG_instruction*>(instr)->sync;
 142    case Format::MTBUF:
 143       return static_cast<const MTBUF_instruction*>(instr)->sync;
 144    case Format::FLAT:
 145    case Format::GLOBAL:
 146    case Format::SCRATCH:
 147       return static_cast<const FLAT_instruction*>(instr)->sync;
 148    case Format::DS:
 149       return static_cast<const DS_instruction*>(instr)->sync;
 150    default:
 151       return memory_sync_info();
 152    }
 153 }
 154
 155 bool can_use_SDWA(chip_class chip, const aco_ptr<Instruction>& instr)
 156 {
 157    if (!instr->isVALU())
 158       return false;
 159
 160    if (chip < GFX8 || instr->isDPP())
 161       return false;
 162
 163    if (instr->isSDWA())
 164       return true;
 165
 166    if (instr->isVOP3()) {
 167       VOP3A_instruction *vop3 = static_cast<VOP3A_instruction*>(instr.get());
 168       if (instr->format == Format::VOP3)
 169          return false;
 170       if (vop3->clamp && instr->format == asVOP3(Format::VOPC) && chip != GFX8)
 171          return false;
 172       if (vop3->omod && chip < GFX9)
 173          return false;
 174
 175       //TODO: return true if we know we will use vcc
 176       if (instr->definitions.size() >= 2)
 177          return false;
 178
 179       for (unsigned i = 1; i < instr->operands.size(); i++) {
 180          if (instr->operands[i].isLiteral())
 181             return false;
 182          if (chip < GFX9 && !instr->operands[i].isOfType(RegType::vgpr))
 183             return false;
 184       }
 185    }
 186
 187    if (!instr->operands.empty()) {
 188       if (instr->operands[0].isLiteral())
 189          return false;
 190       if (chip < GFX9 && !instr->operands[0].isOfType(RegType::vgpr))
 191          return false;
 192    }
 193
 194    bool is_mac = instr->opcode == aco_opcode::v_mac_f32 ||
 195                  instr->opcode == aco_opcode::v_mac_f16 ||
 196                  instr->opcode == aco_opcode::v_fmac_f32 ||
 197                  instr->opcode == aco_opcode::v_fmac_f16;
 198
 199    if (chip != GFX8 && is_mac)
 200       return false;
 201
 202    //TODO: return true if we know we will use vcc
 203    if ((unsigned)instr->format & (unsigned)Format::VOPC)
 204       return false;
 205    if (instr->operands.size() >= 3 && !is_mac)
 206       return false;
 207
 208    return instr->opcode != aco_opcode::v_madmk_f32 &&
 209           instr->opcode != aco_opcode::v_madak_f32 &&
 210           instr->opcode != aco_opcode::v_madmk_f16 &&
 211           instr->opcode != aco_opcode::v_madak_f16 &&
 212           instr->opcode != aco_opcode::v_readfirstlane_b32 &&
 213           instr->opcode != aco_opcode::v_clrexcp &&
 214           instr->opcode != aco_opcode::v_swap_b32;
 215 }
 216
 217 /* updates "instr" and returns the old instruction (or NULL if no update was needed) */
 218 aco_ptr<Instruction> convert_to_SDWA(chip_class chip, aco_ptr<Instruction>& instr)
 219 {
 220    if (instr->isSDWA())
 221       return NULL;
 222
 223    aco_ptr<Instruction> tmp = std::move(instr);
 224    Format format = (Format)(((uint16_t)tmp->format & ~(uint16_t)Format::VOP3) | (uint16_t)Format::SDWA);
 225    instr.reset(create_instruction<SDWA_instruction>(tmp->opcode, format, tmp->operands.size(), tmp->definitions.size()));
 226    std::copy(tmp->operands.cbegin(), tmp->operands.cend(), instr->operands.begin());
 227    std::copy(tmp->definitions.cbegin(), tmp->definitions.cend(), instr->definitions.begin());
 228
 229    SDWA_instruction *sdwa = static_cast<SDWA_instruction*>(instr.get());
 230
 231    if (tmp->isVOP3()) {
 232       VOP3A_instruction *vop3 = static_cast<VOP3A_instruction*>(tmp.get());
 233       memcpy(sdwa->neg, vop3->neg, sizeof(sdwa->neg));
 234       memcpy(sdwa->abs, vop3->abs, sizeof(sdwa->abs));
 235       sdwa->omod = vop3->omod;
 236       sdwa->clamp = vop3->clamp;
 237    }
 238
 239    for (unsigned i = 0; i < instr->operands.size(); i++) {
 240       switch (instr->operands[i].bytes()) {
 241       case 1:
 242          sdwa->sel[i] = sdwa_ubyte;
 243          break;
 244       case 2:
 245          sdwa->sel[i] = sdwa_uword;
 246          break;
 247       case 4:
 248          sdwa->sel[i] = sdwa_udword;
 249          break;
 250       }
 251    }
 252    switch (instr->definitions[0].bytes()) {
 253    case 1:
 254       sdwa->dst_sel = sdwa_ubyte;
 255       sdwa->dst_preserve = true;
 256       break;
 257    case 2:
 258       sdwa->dst_sel = sdwa_uword;
 259       sdwa->dst_preserve = true;
 260       break;
 261    case 4:
 262       sdwa->dst_sel = sdwa_udword;
 263       break;
 264    }
 265
 266    if (instr->definitions[0].getTemp().type() == RegType::sgpr && chip == GFX8)
 267       instr->definitions[0].setFixed(vcc);
 268    if (instr->definitions.size() >= 2)
 269       instr->definitions[1].setFixed(vcc);
 270    if (instr->operands.size() >= 3)
 271       instr->operands[2].setFixed(vcc);
 272
 273    return tmp;
 274 }
 275
 276 bool can_use_opsel(chip_class chip, aco_opcode op, int idx, bool high)
 277 {
 278    /* opsel is only GFX9+ */
 279    if ((high || idx == -1) && chip < GFX9)
 280       return false;
 281
 282    switch (op) {
 283    case aco_opcode::v_div_fixup_f16:
 284    case aco_opcode::v_fma_f16:
 285    case aco_opcode::v_mad_f16:
 286    case aco_opcode::v_mad_u16:
 287    case aco_opcode::v_mad_i16:
 288    case aco_opcode::v_med3_f16:
 289    case aco_opcode::v_med3_i16:
 290    case aco_opcode::v_med3_u16:
 291    case aco_opcode::v_min3_f16:
 292    case aco_opcode::v_min3_i16:
 293    case aco_opcode::v_min3_u16:
 294    case aco_opcode::v_max3_f16:
 295    case aco_opcode::v_max3_i16:
 296    case aco_opcode::v_max3_u16:
 297    case aco_opcode::v_max_u16_e64:
 298    case aco_opcode::v_max_i16_e64:
 299    case aco_opcode::v_min_u16_e64:
 300    case aco_opcode::v_min_i16_e64:
 301    case aco_opcode::v_add_i16:
 302    case aco_opcode::v_sub_i16:
 303    case aco_opcode::v_add_u16_e64:
 304    case aco_opcode::v_sub_u16_e64:
 305    case aco_opcode::v_cvt_pknorm_i16_f16:
 306    case aco_opcode::v_cvt_pknorm_u16_f16:
 307    case aco_opcode::v_lshlrev_b16_e64:
 308    case aco_opcode::v_lshrrev_b16_e64:
 309    case aco_opcode::v_ashrrev_i16_e64:
 310    case aco_opcode::v_mul_lo_u16_e64:
 311       return true;
 312    case aco_opcode::v_pack_b32_f16:
 313       return idx != -1;
 314    case aco_opcode::v_mad_u32_u16:
 315    case aco_opcode::v_mad_i32_i16:
 316       return idx >= 0 && idx < 2;
 317    default:
 318       return false;
 319    }
 320 }
 321
 322 }