Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | //==- BuiltinsAMDGPU.def - AMDGPU Builtin function database ------*- C++ -*-==// |
| 2 | // |
||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
||
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
||
| 6 | // |
||
| 7 | //===----------------------------------------------------------------------===// |
||
| 8 | // |
||
| 9 | // This file defines the AMDGPU-specific builtin function database. Users of |
||
| 10 | // this file must define the BUILTIN macro to make use of this information. |
||
| 11 | // |
||
| 12 | // Note: (unsigned) long int type should be avoided in builtin definitions |
||
| 13 | // since it has different size on Linux (64 bit) and Windows (32 bit). |
||
| 14 | // (unsigned) long long int type should also be avoided, which is 64 bit for |
||
| 15 | // C/C++/HIP but is 128 bit for OpenCL. Use `W` as width modifier in builtin |
||
| 16 | // definitions since it is fixed for 64 bit. |
||
| 17 | //===----------------------------------------------------------------------===// |
||
| 18 | |||
| 19 | // The format of this database matches clang/Basic/Builtins.def. |
||
| 20 | |||
| 21 | #if defined(BUILTIN) && !defined(TARGET_BUILTIN) |
||
| 22 | # define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) BUILTIN(ID, TYPE, ATTRS) |
||
| 23 | #endif |
||
| 24 | //===----------------------------------------------------------------------===// |
||
| 25 | // SI+ only builtins. |
||
| 26 | //===----------------------------------------------------------------------===// |
||
| 27 | |||
| 28 | BUILTIN(__builtin_amdgcn_dispatch_ptr, "v*4", "nc") |
||
| 29 | BUILTIN(__builtin_amdgcn_kernarg_segment_ptr, "v*4", "nc") |
||
| 30 | BUILTIN(__builtin_amdgcn_implicitarg_ptr, "v*4", "nc") |
||
| 31 | BUILTIN(__builtin_amdgcn_queue_ptr, "v*4", "nc") |
||
| 32 | |||
| 33 | BUILTIN(__builtin_amdgcn_workgroup_id_x, "Ui", "nc") |
||
| 34 | BUILTIN(__builtin_amdgcn_workgroup_id_y, "Ui", "nc") |
||
| 35 | BUILTIN(__builtin_amdgcn_workgroup_id_z, "Ui", "nc") |
||
| 36 | |||
| 37 | BUILTIN(__builtin_amdgcn_workitem_id_x, "Ui", "nc") |
||
| 38 | BUILTIN(__builtin_amdgcn_workitem_id_y, "Ui", "nc") |
||
| 39 | BUILTIN(__builtin_amdgcn_workitem_id_z, "Ui", "nc") |
||
| 40 | |||
| 41 | BUILTIN(__builtin_amdgcn_workgroup_size_x, "Us", "nc") |
||
| 42 | BUILTIN(__builtin_amdgcn_workgroup_size_y, "Us", "nc") |
||
| 43 | BUILTIN(__builtin_amdgcn_workgroup_size_z, "Us", "nc") |
||
| 44 | |||
| 45 | BUILTIN(__builtin_amdgcn_grid_size_x, "Ui", "nc") |
||
| 46 | BUILTIN(__builtin_amdgcn_grid_size_y, "Ui", "nc") |
||
| 47 | BUILTIN(__builtin_amdgcn_grid_size_z, "Ui", "nc") |
||
| 48 | |||
| 49 | BUILTIN(__builtin_amdgcn_mbcnt_hi, "UiUiUi", "nc") |
||
| 50 | BUILTIN(__builtin_amdgcn_mbcnt_lo, "UiUiUi", "nc") |
||
| 51 | |||
| 52 | TARGET_BUILTIN(__builtin_amdgcn_s_memtime, "WUi", "n", "s-memtime-inst") |
||
| 53 | |||
| 54 | //===----------------------------------------------------------------------===// |
||
| 55 | // Instruction builtins. |
||
| 56 | //===----------------------------------------------------------------------===// |
||
| 57 | BUILTIN(__builtin_amdgcn_s_getreg, "UiIi", "n") |
||
| 58 | BUILTIN(__builtin_amdgcn_s_setreg, "vIiUi", "n") |
||
| 59 | BUILTIN(__builtin_amdgcn_s_getpc, "WUi", "n") |
||
| 60 | BUILTIN(__builtin_amdgcn_s_waitcnt, "vIi", "n") |
||
| 61 | BUILTIN(__builtin_amdgcn_s_sendmsg, "vIiUi", "n") |
||
| 62 | BUILTIN(__builtin_amdgcn_s_sendmsghalt, "vIiUi", "n") |
||
| 63 | BUILTIN(__builtin_amdgcn_s_barrier, "v", "n") |
||
| 64 | BUILTIN(__builtin_amdgcn_wave_barrier, "v", "n") |
||
| 65 | BUILTIN(__builtin_amdgcn_sched_barrier, "vIi", "n") |
||
| 66 | BUILTIN(__builtin_amdgcn_sched_group_barrier, "vIiIiIi", "n") |
||
| 67 | BUILTIN(__builtin_amdgcn_iglp_opt, "vIi", "n") |
||
| 68 | BUILTIN(__builtin_amdgcn_s_dcache_inv, "v", "n") |
||
| 69 | BUILTIN(__builtin_amdgcn_buffer_wbinvl1, "v", "n") |
||
| 70 | BUILTIN(__builtin_amdgcn_ds_gws_init, "vUiUi", "n") |
||
| 71 | BUILTIN(__builtin_amdgcn_ds_gws_barrier, "vUiUi", "n") |
||
| 72 | BUILTIN(__builtin_amdgcn_ds_gws_sema_v, "vUi", "n") |
||
| 73 | BUILTIN(__builtin_amdgcn_ds_gws_sema_br, "vUiUi", "n") |
||
| 74 | BUILTIN(__builtin_amdgcn_ds_gws_sema_p, "vUi", "n") |
||
| 75 | BUILTIN(__builtin_amdgcn_fence, "vUicC*", "n") |
||
| 76 | BUILTIN(__builtin_amdgcn_groupstaticsize, "Ui", "n") |
||
| 77 | |||
| 78 | BUILTIN(__builtin_amdgcn_atomic_inc32, "UZiUZiD*UZiUicC*", "n") |
||
| 79 | BUILTIN(__builtin_amdgcn_atomic_inc64, "UWiUWiD*UWiUicC*", "n") |
||
| 80 | |||
| 81 | BUILTIN(__builtin_amdgcn_atomic_dec32, "UZiUZiD*UZiUicC*", "n") |
||
| 82 | BUILTIN(__builtin_amdgcn_atomic_dec64, "UWiUWiD*UWiUicC*", "n") |
||
| 83 | |||
| 84 | // FIXME: Need to disallow constant address space. |
||
| 85 | BUILTIN(__builtin_amdgcn_div_scale, "dddbb*", "n") |
||
| 86 | BUILTIN(__builtin_amdgcn_div_scalef, "fffbb*", "n") |
||
| 87 | BUILTIN(__builtin_amdgcn_div_fmas, "ddddb", "nc") |
||
| 88 | BUILTIN(__builtin_amdgcn_div_fmasf, "ffffb", "nc") |
||
| 89 | BUILTIN(__builtin_amdgcn_div_fixup, "dddd", "nc") |
||
| 90 | BUILTIN(__builtin_amdgcn_div_fixupf, "ffff", "nc") |
||
| 91 | BUILTIN(__builtin_amdgcn_trig_preop, "ddi", "nc") |
||
| 92 | BUILTIN(__builtin_amdgcn_trig_preopf, "ffi", "nc") |
||
| 93 | BUILTIN(__builtin_amdgcn_rcp, "dd", "nc") |
||
| 94 | BUILTIN(__builtin_amdgcn_rcpf, "ff", "nc") |
||
| 95 | BUILTIN(__builtin_amdgcn_sqrt, "dd", "nc") |
||
| 96 | BUILTIN(__builtin_amdgcn_sqrtf, "ff", "nc") |
||
| 97 | BUILTIN(__builtin_amdgcn_rsq, "dd", "nc") |
||
| 98 | BUILTIN(__builtin_amdgcn_rsqf, "ff", "nc") |
||
| 99 | BUILTIN(__builtin_amdgcn_rsq_clamp, "dd", "nc") |
||
| 100 | BUILTIN(__builtin_amdgcn_rsq_clampf, "ff", "nc") |
||
| 101 | BUILTIN(__builtin_amdgcn_sinf, "ff", "nc") |
||
| 102 | BUILTIN(__builtin_amdgcn_cosf, "ff", "nc") |
||
| 103 | BUILTIN(__builtin_amdgcn_log_clampf, "ff", "nc") |
||
| 104 | BUILTIN(__builtin_amdgcn_ldexp, "ddi", "nc") |
||
| 105 | BUILTIN(__builtin_amdgcn_ldexpf, "ffi", "nc") |
||
| 106 | BUILTIN(__builtin_amdgcn_frexp_mant, "dd", "nc") |
||
| 107 | BUILTIN(__builtin_amdgcn_frexp_mantf, "ff", "nc") |
||
| 108 | BUILTIN(__builtin_amdgcn_frexp_exp, "id", "nc") |
||
| 109 | BUILTIN(__builtin_amdgcn_frexp_expf, "if", "nc") |
||
| 110 | BUILTIN(__builtin_amdgcn_fract, "dd", "nc") |
||
| 111 | BUILTIN(__builtin_amdgcn_fractf, "ff", "nc") |
||
| 112 | BUILTIN(__builtin_amdgcn_lerp, "UiUiUiUi", "nc") |
||
| 113 | BUILTIN(__builtin_amdgcn_class, "bdi", "nc") |
||
| 114 | BUILTIN(__builtin_amdgcn_classf, "bfi", "nc") |
||
| 115 | BUILTIN(__builtin_amdgcn_cubeid, "ffff", "nc") |
||
| 116 | BUILTIN(__builtin_amdgcn_cubesc, "ffff", "nc") |
||
| 117 | BUILTIN(__builtin_amdgcn_cubetc, "ffff", "nc") |
||
| 118 | BUILTIN(__builtin_amdgcn_cubema, "ffff", "nc") |
||
| 119 | BUILTIN(__builtin_amdgcn_s_sleep, "vIi", "n") |
||
| 120 | BUILTIN(__builtin_amdgcn_s_incperflevel, "vIi", "n") |
||
| 121 | BUILTIN(__builtin_amdgcn_s_decperflevel, "vIi", "n") |
||
| 122 | BUILTIN(__builtin_amdgcn_s_setprio, "vIs", "n") |
||
| 123 | BUILTIN(__builtin_amdgcn_ds_swizzle, "iiIi", "nc") |
||
| 124 | BUILTIN(__builtin_amdgcn_ds_permute, "iii", "nc") |
||
| 125 | BUILTIN(__builtin_amdgcn_ds_bpermute, "iii", "nc") |
||
| 126 | BUILTIN(__builtin_amdgcn_readfirstlane, "ii", "nc") |
||
| 127 | BUILTIN(__builtin_amdgcn_readlane, "iii", "nc") |
||
| 128 | BUILTIN(__builtin_amdgcn_fmed3f, "ffff", "nc") |
||
| 129 | BUILTIN(__builtin_amdgcn_ds_faddf, "ff*3fIiIiIb", "n") |
||
| 130 | BUILTIN(__builtin_amdgcn_ds_fminf, "ff*3fIiIiIb", "n") |
||
| 131 | BUILTIN(__builtin_amdgcn_ds_fmaxf, "ff*3fIiIiIb", "n") |
||
| 132 | BUILTIN(__builtin_amdgcn_ds_append, "ii*3", "n") |
||
| 133 | BUILTIN(__builtin_amdgcn_ds_consume, "ii*3", "n") |
||
| 134 | BUILTIN(__builtin_amdgcn_alignbit, "UiUiUiUi", "nc") |
||
| 135 | BUILTIN(__builtin_amdgcn_alignbyte, "UiUiUiUi", "nc") |
||
| 136 | BUILTIN(__builtin_amdgcn_ubfe, "UiUiUiUi", "nc") |
||
| 137 | BUILTIN(__builtin_amdgcn_sbfe, "UiUiUiUi", "nc") |
||
| 138 | BUILTIN(__builtin_amdgcn_cvt_pkrtz, "E2hff", "nc") |
||
| 139 | BUILTIN(__builtin_amdgcn_cvt_pknorm_i16, "E2sff", "nc") |
||
| 140 | BUILTIN(__builtin_amdgcn_cvt_pknorm_u16, "E2Usff", "nc") |
||
| 141 | BUILTIN(__builtin_amdgcn_cvt_pk_i16, "E2sii", "nc") |
||
| 142 | BUILTIN(__builtin_amdgcn_cvt_pk_u16, "E2UsUiUi", "nc") |
||
| 143 | BUILTIN(__builtin_amdgcn_cvt_pk_u8_f32, "UifUiUi", "nc") |
||
| 144 | BUILTIN(__builtin_amdgcn_sad_u8, "UiUiUiUi", "nc") |
||
| 145 | BUILTIN(__builtin_amdgcn_msad_u8, "UiUiUiUi", "nc") |
||
| 146 | BUILTIN(__builtin_amdgcn_sad_hi_u8, "UiUiUiUi", "nc") |
||
| 147 | BUILTIN(__builtin_amdgcn_sad_u16, "UiUiUiUi", "nc") |
||
| 148 | BUILTIN(__builtin_amdgcn_qsad_pk_u16_u8, "WUiWUiUiWUi", "nc") |
||
| 149 | BUILTIN(__builtin_amdgcn_mqsad_pk_u16_u8, "WUiWUiUiWUi", "nc") |
||
| 150 | BUILTIN(__builtin_amdgcn_mqsad_u32_u8, "V4UiWUiUiV4Ui", "nc") |
||
| 151 | |||
| 152 | //===----------------------------------------------------------------------===// |
||
| 153 | // Ballot builtins. |
||
| 154 | //===----------------------------------------------------------------------===// |
||
| 155 | |||
| 156 | TARGET_BUILTIN(__builtin_amdgcn_ballot_w32, "Uib", "nc", "wavefrontsize32") |
||
| 157 | TARGET_BUILTIN(__builtin_amdgcn_ballot_w64, "LUib", "nc", "wavefrontsize64") |
||
| 158 | |||
| 159 | // Deprecated intrinsics in favor of __builtin_amdgn_ballot_{w32|w64} |
||
| 160 | BUILTIN(__builtin_amdgcn_uicmp, "WUiUiUiIi", "nc") |
||
| 161 | BUILTIN(__builtin_amdgcn_uicmpl, "WUiWUiWUiIi", "nc") |
||
| 162 | BUILTIN(__builtin_amdgcn_sicmp, "WUiiiIi", "nc") |
||
| 163 | BUILTIN(__builtin_amdgcn_sicmpl, "WUiWiWiIi", "nc") |
||
| 164 | BUILTIN(__builtin_amdgcn_fcmp, "WUiddIi", "nc") |
||
| 165 | BUILTIN(__builtin_amdgcn_fcmpf, "WUiffIi", "nc") |
||
| 166 | |||
| 167 | //===----------------------------------------------------------------------===// |
||
| 168 | // Flat addressing builtins. |
||
| 169 | //===----------------------------------------------------------------------===// |
||
| 170 | BUILTIN(__builtin_amdgcn_is_shared, "bvC*0", "nc") |
||
| 171 | BUILTIN(__builtin_amdgcn_is_private, "bvC*0", "nc") |
||
| 172 | |||
| 173 | //===----------------------------------------------------------------------===// |
||
| 174 | // CI+ only builtins. |
||
| 175 | //===----------------------------------------------------------------------===// |
||
| 176 | TARGET_BUILTIN(__builtin_amdgcn_s_dcache_inv_vol, "v", "n", "ci-insts") |
||
| 177 | TARGET_BUILTIN(__builtin_amdgcn_buffer_wbinvl1_vol, "v", "n", "ci-insts") |
||
| 178 | TARGET_BUILTIN(__builtin_amdgcn_ds_gws_sema_release_all, "vUi", "n", "ci-insts") |
||
| 179 | |||
| 180 | //===----------------------------------------------------------------------===// |
||
| 181 | // Interpolation builtins. |
||
| 182 | //===----------------------------------------------------------------------===// |
||
| 183 | BUILTIN(__builtin_amdgcn_interp_p1_f16, "ffUiUibUi", "nc") |
||
| 184 | BUILTIN(__builtin_amdgcn_interp_p2_f16, "hffUiUibUi", "nc") |
||
| 185 | BUILTIN(__builtin_amdgcn_interp_p1, "ffUiUiUi", "nc") |
||
| 186 | BUILTIN(__builtin_amdgcn_interp_p2, "fffUiUiUi", "nc") |
||
| 187 | BUILTIN(__builtin_amdgcn_interp_mov, "fUiUiUiUi", "nc") |
||
| 188 | |||
| 189 | //===----------------------------------------------------------------------===// |
||
| 190 | // VI+ only builtins. |
||
| 191 | //===----------------------------------------------------------------------===// |
||
| 192 | |||
| 193 | TARGET_BUILTIN(__builtin_amdgcn_div_fixuph, "hhhh", "nc", "16-bit-insts") |
||
| 194 | TARGET_BUILTIN(__builtin_amdgcn_rcph, "hh", "nc", "16-bit-insts") |
||
| 195 | TARGET_BUILTIN(__builtin_amdgcn_sqrth, "hh", "nc", "16-bit-insts") |
||
| 196 | TARGET_BUILTIN(__builtin_amdgcn_rsqh, "hh", "nc", "16-bit-insts") |
||
| 197 | TARGET_BUILTIN(__builtin_amdgcn_sinh, "hh", "nc", "16-bit-insts") |
||
| 198 | TARGET_BUILTIN(__builtin_amdgcn_cosh, "hh", "nc", "16-bit-insts") |
||
| 199 | TARGET_BUILTIN(__builtin_amdgcn_ldexph, "hhi", "nc", "16-bit-insts") |
||
| 200 | TARGET_BUILTIN(__builtin_amdgcn_frexp_manth, "hh", "nc", "16-bit-insts") |
||
| 201 | TARGET_BUILTIN(__builtin_amdgcn_frexp_exph, "sh", "nc", "16-bit-insts") |
||
| 202 | TARGET_BUILTIN(__builtin_amdgcn_fracth, "hh", "nc", "16-bit-insts") |
||
| 203 | TARGET_BUILTIN(__builtin_amdgcn_classh, "bhi", "nc", "16-bit-insts") |
||
| 204 | TARGET_BUILTIN(__builtin_amdgcn_s_memrealtime, "WUi", "n", "s-memrealtime") |
||
| 205 | TARGET_BUILTIN(__builtin_amdgcn_mov_dpp, "iiIiIiIiIb", "nc", "dpp") |
||
| 206 | TARGET_BUILTIN(__builtin_amdgcn_update_dpp, "iiiIiIiIiIb", "nc", "dpp") |
||
| 207 | TARGET_BUILTIN(__builtin_amdgcn_s_dcache_wb, "v", "n", "gfx8-insts") |
||
| 208 | TARGET_BUILTIN(__builtin_amdgcn_perm, "UiUiUiUi", "nc", "gfx8-insts") |
||
| 209 | |||
| 210 | //===----------------------------------------------------------------------===// |
||
| 211 | // GFX9+ only builtins. |
||
| 212 | //===----------------------------------------------------------------------===// |
||
| 213 | |||
| 214 | TARGET_BUILTIN(__builtin_amdgcn_fmed3h, "hhhh", "nc", "gfx9-insts") |
||
| 215 | |||
| 216 | TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_f64, "dd*1d", "t", "gfx90a-insts") |
||
| 217 | TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_f32, "ff*1f", "t", "gfx90a-insts") |
||
| 218 | TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_v2f16, "V2hV2h*1V2h", "t", "gfx90a-insts") |
||
| 219 | TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fmin_f64, "dd*1d", "t", "gfx90a-insts") |
||
| 220 | TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fmax_f64, "dd*1d", "t", "gfx90a-insts") |
||
| 221 | |||
| 222 | TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_f64, "dd*0d", "t", "gfx90a-insts") |
||
| 223 | TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fmin_f64, "dd*0d", "t", "gfx90a-insts") |
||
| 224 | TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fmax_f64, "dd*0d", "t", "gfx90a-insts") |
||
| 225 | |||
| 226 | TARGET_BUILTIN(__builtin_amdgcn_ds_atomic_fadd_f64, "dd*3d", "t", "gfx90a-insts") |
||
| 227 | TARGET_BUILTIN(__builtin_amdgcn_ds_atomic_fadd_f32, "ff*3f", "t", "gfx8-insts") |
||
| 228 | |||
| 229 | TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_f32, "ff*0f", "t", "gfx940-insts") |
||
| 230 | TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_v2f16, "V2hV2h*0V2h", "t", "gfx940-insts") |
||
| 231 | TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_v2bf16, "V2sV2s*0V2s", "t", "gfx940-insts") |
||
| 232 | TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_v2bf16, "V2sV2s*1V2s", "t", "gfx940-insts") |
||
| 233 | TARGET_BUILTIN(__builtin_amdgcn_ds_atomic_fadd_v2bf16, "V2sV2s*3V2s", "t", "gfx940-insts") |
||
| 234 | |||
| 235 | //===----------------------------------------------------------------------===// |
||
| 236 | // Deep learning builtins. |
||
| 237 | //===----------------------------------------------------------------------===// |
||
| 238 | |||
| 239 | TARGET_BUILTIN(__builtin_amdgcn_fdot2, "fV2hV2hfIb", "nc", "dot7-insts") |
||
| 240 | TARGET_BUILTIN(__builtin_amdgcn_fdot2_f16_f16, "hV2hV2hh", "nc", "dot9-insts") |
||
| 241 | TARGET_BUILTIN(__builtin_amdgcn_fdot2_bf16_bf16, "sV2sV2ss", "nc", "dot9-insts") |
||
| 242 | TARGET_BUILTIN(__builtin_amdgcn_fdot2_f32_bf16, "fV2sV2sfIb", "nc", "dot9-insts") |
||
| 243 | TARGET_BUILTIN(__builtin_amdgcn_sdot2, "SiV2SsV2SsSiIb", "nc", "dot2-insts") |
||
| 244 | TARGET_BUILTIN(__builtin_amdgcn_udot2, "UiV2UsV2UsUiIb", "nc", "dot2-insts") |
||
| 245 | TARGET_BUILTIN(__builtin_amdgcn_sdot4, "SiSiSiSiIb", "nc", "dot1-insts") |
||
| 246 | TARGET_BUILTIN(__builtin_amdgcn_udot4, "UiUiUiUiIb", "nc", "dot7-insts") |
||
| 247 | TARGET_BUILTIN(__builtin_amdgcn_sudot4, "iIbiIbiiIb", "nc", "dot8-insts") |
||
| 248 | TARGET_BUILTIN(__builtin_amdgcn_sdot8, "SiSiSiSiIb", "nc", "dot1-insts") |
||
| 249 | TARGET_BUILTIN(__builtin_amdgcn_udot8, "UiUiUiUiIb", "nc", "dot7-insts") |
||
| 250 | TARGET_BUILTIN(__builtin_amdgcn_sudot8, "iIbiIbiiIb", "nc", "dot8-insts") |
||
| 251 | |||
| 252 | //===----------------------------------------------------------------------===// |
||
| 253 | // GFX10+ only builtins. |
||
| 254 | //===----------------------------------------------------------------------===// |
||
| 255 | TARGET_BUILTIN(__builtin_amdgcn_permlane16, "UiUiUiUiUiIbIb", "nc", "gfx10-insts") |
||
| 256 | TARGET_BUILTIN(__builtin_amdgcn_permlanex16, "UiUiUiUiUiIbIb", "nc", "gfx10-insts") |
||
| 257 | TARGET_BUILTIN(__builtin_amdgcn_mov_dpp8, "UiUiIUi", "nc", "gfx10-insts") |
||
| 258 | |||
| 259 | //===----------------------------------------------------------------------===// |
||
| 260 | // Raytracing builtins. |
||
| 261 | // By default the 1st argument is i32 and the 4/5-th arguments are float4. |
||
| 262 | // Postfix l indicates the 1st argument is i64. |
||
| 263 | // Postfix h indicates the 4/5-th arguments are half4. |
||
| 264 | //===----------------------------------------------------------------------===// |
||
| 265 | TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray, "V4UiUifV4fV4fV4fV4Ui", "nc", "gfx10-insts") |
||
| 266 | TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray_h, "V4UiUifV4fV4hV4hV4Ui", "nc", "gfx10-insts") |
||
| 267 | TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray_l, "V4UiWUifV4fV4fV4fV4Ui", "nc", "gfx10-insts") |
||
| 268 | TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray_lh, "V4UiWUifV4fV4hV4hV4Ui", "nc", "gfx10-insts") |
||
| 269 | |||
| 270 | |||
| 271 | //===----------------------------------------------------------------------===// |
||
| 272 | // GFX11+ only builtins. |
||
| 273 | //===----------------------------------------------------------------------===// |
||
| 274 | |||
| 275 | // TODO: This is a no-op in wave32. Should the builtin require wavefrontsize64? |
||
| 276 | TARGET_BUILTIN(__builtin_amdgcn_permlane64, "UiUi", "nc", "gfx11-insts") |
||
| 277 | TARGET_BUILTIN(__builtin_amdgcn_s_wait_event_export_ready, "v", "n", "gfx11-insts") |
||
| 278 | |||
| 279 | //===----------------------------------------------------------------------===// |
||
| 280 | // WMMA builtins. |
||
| 281 | // Postfix w32 indicates the builtin requires wavefront size of 32. |
||
| 282 | // Postfix w64 indicates the builtin requires wavefront size of 64. |
||
| 283 | //===----------------------------------------------------------------------===// |
||
| 284 | TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w32, "V8fV16hV16hV8f", "nc", "gfx11-insts") |
||
| 285 | TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32, "V8fV16sV16sV8f", "nc", "gfx11-insts") |
||
| 286 | TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w32, "V16hV16hV16hV16hIb", "nc", "gfx11-insts") |
||
| 287 | TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32, "V16sV16sV16sV16sIb", "nc", "gfx11-insts") |
||
| 288 | TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32, "V8iIbV4iIbV4iV8iIb", "nc", "gfx11-insts") |
||
| 289 | TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32, "V8iIbV2iIbV2iV8iIb", "nc", "gfx11-insts") |
||
| 290 | |||
| 291 | TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w64, "V4fV16hV16hV4f", "nc", "gfx11-insts") |
||
| 292 | TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64, "V4fV16sV16sV4f", "nc", "gfx11-insts") |
||
| 293 | TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w64, "V8hV16hV16hV8hIb", "nc", "gfx11-insts") |
||
| 294 | TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64, "V8sV16sV16sV8sIb", "nc", "gfx11-insts") |
||
| 295 | TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64, "V4iIbV4iIbV4iV4iIb", "nc", "gfx11-insts") |
||
| 296 | TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64, "V4iIbV2iIbV2iV4iIb", "nc", "gfx11-insts") |
||
| 297 | |||
| 298 | TARGET_BUILTIN(__builtin_amdgcn_s_sendmsg_rtn, "UiUIi", "n", "gfx11-insts") |
||
| 299 | TARGET_BUILTIN(__builtin_amdgcn_s_sendmsg_rtnl, "UWiUIi", "n", "gfx11-insts") |
||
| 300 | |||
| 301 | TARGET_BUILTIN(__builtin_amdgcn_ds_bvh_stack_rtn, "V2UiUiUiV4UiIi", "n", "gfx11-insts") |
||
| 302 | |||
| 303 | //===----------------------------------------------------------------------===// |
||
| 304 | // Special builtins. |
||
| 305 | //===----------------------------------------------------------------------===// |
||
| 306 | BUILTIN(__builtin_amdgcn_read_exec, "WUi", "nc") |
||
| 307 | BUILTIN(__builtin_amdgcn_read_exec_lo, "Ui", "nc") |
||
| 308 | BUILTIN(__builtin_amdgcn_read_exec_hi, "Ui", "nc") |
||
| 309 | |||
| 310 | BUILTIN(__builtin_amdgcn_endpgm, "v", "nr") |
||
| 311 | |||
| 312 | //===----------------------------------------------------------------------===// |
||
| 313 | // R600-NI only builtins. |
||
| 314 | //===----------------------------------------------------------------------===// |
||
| 315 | |||
| 316 | BUILTIN(__builtin_r600_implicitarg_ptr, "Uc*7", "nc") |
||
| 317 | |||
| 318 | BUILTIN(__builtin_r600_read_tgid_x, "Ui", "nc") |
||
| 319 | BUILTIN(__builtin_r600_read_tgid_y, "Ui", "nc") |
||
| 320 | BUILTIN(__builtin_r600_read_tgid_z, "Ui", "nc") |
||
| 321 | |||
| 322 | BUILTIN(__builtin_r600_read_tidig_x, "Ui", "nc") |
||
| 323 | BUILTIN(__builtin_r600_read_tidig_y, "Ui", "nc") |
||
| 324 | BUILTIN(__builtin_r600_read_tidig_z, "Ui", "nc") |
||
| 325 | |||
| 326 | BUILTIN(__builtin_r600_recipsqrt_ieee, "dd", "nc") |
||
| 327 | BUILTIN(__builtin_r600_recipsqrt_ieeef, "ff", "nc") |
||
| 328 | |||
| 329 | //===----------------------------------------------------------------------===// |
||
| 330 | // MFMA builtins. |
||
| 331 | //===----------------------------------------------------------------------===// |
||
| 332 | |||
| 333 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x1f32, "V32fffV32fIiIiIi", "nc", "mai-insts") |
||
| 334 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x1f32, "V16fffV16fIiIiIi", "nc", "mai-insts") |
||
| 335 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x1f32, "V4fffV4fIiIiIi", "nc", "mai-insts") |
||
| 336 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x2f32, "V16fffV16fIiIiIi", "nc", "mai-insts") |
||
| 337 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x4f32, "V4fffV4fIiIiIi", "nc", "mai-insts") |
||
| 338 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4f16, "V32fV4hV4hV32fIiIiIi", "nc", "mai-insts") |
||
| 339 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x4f16, "V16fV4hV4hV16fIiIiIi", "nc", "mai-insts") |
||
| 340 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x4f16, "V4fV4hV4hV4fIiIiIi", "nc", "mai-insts") |
||
| 341 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x8f16, "V16fV4hV4hV16fIiIiIi", "nc", "mai-insts") |
||
| 342 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x16f16, "V4fV4hV4hV4fIiIiIi", "nc", "mai-insts") |
||
| 343 | TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_32x32x4i8, "V32iiiV32iIiIiIi", "nc", "mai-insts") |
||
| 344 | TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_16x16x4i8, "V16iiiV16iIiIiIi", "nc", "mai-insts") |
||
| 345 | TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_4x4x4i8, "V4iiiV4iIiIiIi", "nc", "mai-insts") |
||
| 346 | TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_32x32x8i8, "V16iiiV16iIiIiIi", "nc", "mai-insts") |
||
| 347 | TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_16x16x16i8, "V4iiiV4iIiIiIi", "nc", "mai-insts") |
||
| 348 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x2bf16, "V32fV2sV2sV32fIiIiIi", "nc", "mai-insts") |
||
| 349 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x2bf16, "V16fV2sV2sV16fIiIiIi", "nc", "mai-insts") |
||
| 350 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x2bf16, "V4fV2sV2sV4fIiIiIi", "nc", "mai-insts") |
||
| 351 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4bf16, "V16fV2sV2sV16fIiIiIi", "nc", "mai-insts") |
||
| 352 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x8bf16, "V4fV2sV2sV4fIiIiIi", "nc", "mai-insts") |
||
| 353 | |||
| 354 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4bf16_1k, "V32fV4sV4sV32fIiIiIi", "nc", "mai-insts") |
||
| 355 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x4bf16_1k, "V16fV4sV4sV16fIiIiIi", "nc", "mai-insts") |
||
| 356 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x4bf16_1k, "V4fV4sV4sV4fIiIiIi", "nc", "mai-insts") |
||
| 357 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x8bf16_1k, "V16fV4sV4sV16fIiIiIi", "nc", "mai-insts") |
||
| 358 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x16bf16_1k, "V4fV4sV4sV4fIiIiIi", "nc", "mai-insts") |
||
| 359 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f64_16x16x4f64, "V4dddV4dIiIiIi", "nc", "mai-insts") |
||
| 360 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f64_4x4x4f64, "ddddIiIiIi", "nc", "mai-insts") |
||
| 361 | |||
| 362 | TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_16x16x32_i8, "V4iWiWiV4iIiIiIi", "nc", "mai-insts") |
||
| 363 | TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_32x32x16_i8, "V16iWiWiV16iIiIiIi", "nc", "mai-insts") |
||
| 364 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x8_xf32, "V4fV2fV2fV4fIiIiIi", "nc", "mai-insts") |
||
| 365 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4_xf32, "V16fV2fV2fV16fIiIiIi", "nc", "mai-insts") |
||
| 366 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_bf8_bf8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") |
||
| 367 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_bf8_fp8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") |
||
| 368 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_fp8_bf8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") |
||
| 369 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") |
||
| 370 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") |
||
| 371 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") |
||
| 372 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") |
||
| 373 | TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") |
||
| 374 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x32_f16, "V4fV4hV8hV4fiIiIi", "nc", "mai-insts") |
||
| 375 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x16_f16, "V16fV4hV8hV16fiIiIi", "nc", "mai-insts") |
||
| 376 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x32_bf16, "V4fV4sV8sV4fiIiIi", "nc", "mai-insts") |
||
| 377 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x16_bf16, "V16fV4sV8sV16fiIiIi", "nc", "mai-insts") |
||
| 378 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_i32_16x16x64_i8, "V4iV2iV4iV4iiIiIi", "nc", "mai-insts") |
||
| 379 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_i32_32x32x32_i8, "V16iV2iV4iV16iiIiIi", "nc", "mai-insts") |
||
| 380 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_bf8_bf8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") |
||
| 381 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_bf8_fp8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") |
||
| 382 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_fp8_bf8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") |
||
| 383 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_fp8_fp8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") |
||
| 384 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_bf8_bf8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") |
||
| 385 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_bf8_fp8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") |
||
| 386 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_fp8_bf8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") |
||
| 387 | TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_fp8_fp8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") |
||
| 388 | |||
| 389 | TARGET_BUILTIN(__builtin_amdgcn_cvt_f32_bf8, "fiIi", "nc", "fp8-insts") |
||
| 390 | TARGET_BUILTIN(__builtin_amdgcn_cvt_f32_fp8, "fiIi", "nc", "fp8-insts") |
||
| 391 | TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_f32_bf8, "V2fiIb", "nc", "fp8-insts") |
||
| 392 | TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_f32_fp8, "V2fiIb", "nc", "fp8-insts") |
||
| 393 | TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_bf8_f32, "iffiIb", "nc", "fp8-insts") |
||
| 394 | TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_fp8_f32, "iffiIb", "nc", "fp8-insts") |
||
| 395 | TARGET_BUILTIN(__builtin_amdgcn_cvt_sr_bf8_f32, "ifiiIi", "nc", "fp8-insts") |
||
| 396 | TARGET_BUILTIN(__builtin_amdgcn_cvt_sr_fp8_f32, "ifiiIi", "nc", "fp8-insts") |
||
| 397 | |||
| 398 | #undef BUILTIN |
||
| 399 | #undef TARGET_BUILTIN |