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 28BUILTIN(__builtin_amdgcn_dispatch_ptr, "v*4", "nc") 29BUILTIN(__builtin_amdgcn_kernarg_segment_ptr, "v*4", "nc") 30BUILTIN(__builtin_amdgcn_implicitarg_ptr, "v*4", "nc") 31BUILTIN(__builtin_amdgcn_queue_ptr, "v*4", "nc") 32 33BUILTIN(__builtin_amdgcn_workgroup_id_x, "Ui", "nc") 34BUILTIN(__builtin_amdgcn_workgroup_id_y, "Ui", "nc") 35BUILTIN(__builtin_amdgcn_workgroup_id_z, "Ui", "nc") 36 37BUILTIN(__builtin_amdgcn_workitem_id_x, "Ui", "nc") 38BUILTIN(__builtin_amdgcn_workitem_id_y, "Ui", "nc") 39BUILTIN(__builtin_amdgcn_workitem_id_z, "Ui", "nc") 40 41BUILTIN(__builtin_amdgcn_workgroup_size_x, "Us", "nc") 42BUILTIN(__builtin_amdgcn_workgroup_size_y, "Us", "nc") 43BUILTIN(__builtin_amdgcn_workgroup_size_z, "Us", "nc") 44 45BUILTIN(__builtin_amdgcn_grid_size_x, "Ui", "nc") 46BUILTIN(__builtin_amdgcn_grid_size_y, "Ui", "nc") 47BUILTIN(__builtin_amdgcn_grid_size_z, "Ui", "nc") 48 49BUILTIN(__builtin_amdgcn_mbcnt_hi, "UiUiUi", "nc") 50BUILTIN(__builtin_amdgcn_mbcnt_lo, "UiUiUi", "nc") 51 52TARGET_BUILTIN(__builtin_amdgcn_s_memtime, "WUi", "n", "s-memtime-inst") 53 54//===----------------------------------------------------------------------===// 55// Instruction builtins. 56//===----------------------------------------------------------------------===// 57BUILTIN(__builtin_amdgcn_s_getreg, "UiIi", "n") 58BUILTIN(__builtin_amdgcn_s_setreg, "vIiUi", "n") 59BUILTIN(__builtin_amdgcn_s_getpc, "WUi", "n") 60BUILTIN(__builtin_amdgcn_s_waitcnt, "vIi", "n") 61BUILTIN(__builtin_amdgcn_s_sendmsg, "vIiUi", "n") 62BUILTIN(__builtin_amdgcn_s_sendmsghalt, "vIiUi", "n") 63BUILTIN(__builtin_amdgcn_s_barrier, "v", "n") 64BUILTIN(__builtin_amdgcn_wave_barrier, "v", "n") 65BUILTIN(__builtin_amdgcn_sched_barrier, "vIi", "n") 66BUILTIN(__builtin_amdgcn_sched_group_barrier, "vIiIiIi", "n") 67BUILTIN(__builtin_amdgcn_iglp_opt, "vIi", "n") 68BUILTIN(__builtin_amdgcn_s_dcache_inv, "v", "n") 69BUILTIN(__builtin_amdgcn_buffer_wbinvl1, "v", "n") 70BUILTIN(__builtin_amdgcn_fence, "vUicC*", "n") 71BUILTIN(__builtin_amdgcn_groupstaticsize, "Ui", "n") 72BUILTIN(__builtin_amdgcn_wavefrontsize, "Ui", "nc") 73 74BUILTIN(__builtin_amdgcn_atomic_inc32, "UZiUZiD*UZiUicC*", "n") 75BUILTIN(__builtin_amdgcn_atomic_inc64, "UWiUWiD*UWiUicC*", "n") 76 77BUILTIN(__builtin_amdgcn_atomic_dec32, "UZiUZiD*UZiUicC*", "n") 78BUILTIN(__builtin_amdgcn_atomic_dec64, "UWiUWiD*UWiUicC*", "n") 79 80// FIXME: Need to disallow constant address space. 81BUILTIN(__builtin_amdgcn_div_scale, "dddbb*", "n") 82BUILTIN(__builtin_amdgcn_div_scalef, "fffbb*", "n") 83BUILTIN(__builtin_amdgcn_div_fmas, "ddddb", "nc") 84BUILTIN(__builtin_amdgcn_div_fmasf, "ffffb", "nc") 85BUILTIN(__builtin_amdgcn_div_fixup, "dddd", "nc") 86BUILTIN(__builtin_amdgcn_div_fixupf, "ffff", "nc") 87BUILTIN(__builtin_amdgcn_trig_preop, "ddi", "nc") 88BUILTIN(__builtin_amdgcn_trig_preopf, "ffi", "nc") 89BUILTIN(__builtin_amdgcn_rcp, "dd", "nc") 90BUILTIN(__builtin_amdgcn_rcpf, "ff", "nc") 91BUILTIN(__builtin_amdgcn_sqrt, "dd", "nc") 92BUILTIN(__builtin_amdgcn_sqrtf, "ff", "nc") 93BUILTIN(__builtin_amdgcn_rsq, "dd", "nc") 94BUILTIN(__builtin_amdgcn_rsqf, "ff", "nc") 95BUILTIN(__builtin_amdgcn_rsq_clamp, "dd", "nc") 96BUILTIN(__builtin_amdgcn_rsq_clampf, "ff", "nc") 97BUILTIN(__builtin_amdgcn_sinf, "ff", "nc") 98BUILTIN(__builtin_amdgcn_cosf, "ff", "nc") 99BUILTIN(__builtin_amdgcn_logf, "ff", "nc") 100BUILTIN(__builtin_amdgcn_exp2f, "ff", "nc") 101BUILTIN(__builtin_amdgcn_log_clampf, "ff", "nc") 102BUILTIN(__builtin_amdgcn_ldexp, "ddi", "nc") 103BUILTIN(__builtin_amdgcn_ldexpf, "ffi", "nc") 104BUILTIN(__builtin_amdgcn_frexp_mant, "dd", "nc") 105BUILTIN(__builtin_amdgcn_frexp_mantf, "ff", "nc") 106BUILTIN(__builtin_amdgcn_frexp_exp, "id", "nc") 107BUILTIN(__builtin_amdgcn_frexp_expf, "if", "nc") 108BUILTIN(__builtin_amdgcn_fract, "dd", "nc") 109BUILTIN(__builtin_amdgcn_fractf, "ff", "nc") 110BUILTIN(__builtin_amdgcn_lerp, "UiUiUiUi", "nc") 111BUILTIN(__builtin_amdgcn_class, "bdi", "nc") 112BUILTIN(__builtin_amdgcn_classf, "bfi", "nc") 113BUILTIN(__builtin_amdgcn_cubeid, "ffff", "nc") 114BUILTIN(__builtin_amdgcn_cubesc, "ffff", "nc") 115BUILTIN(__builtin_amdgcn_cubetc, "ffff", "nc") 116BUILTIN(__builtin_amdgcn_cubema, "ffff", "nc") 117BUILTIN(__builtin_amdgcn_s_sleep, "vIi", "n") 118BUILTIN(__builtin_amdgcn_s_incperflevel, "vIi", "n") 119BUILTIN(__builtin_amdgcn_s_decperflevel, "vIi", "n") 120BUILTIN(__builtin_amdgcn_s_setprio, "vIs", "n") 121BUILTIN(__builtin_amdgcn_ds_swizzle, "iiIi", "nc") 122BUILTIN(__builtin_amdgcn_ds_permute, "iii", "nc") 123BUILTIN(__builtin_amdgcn_ds_bpermute, "iii", "nc") 124BUILTIN(__builtin_amdgcn_readfirstlane, "ii", "nc") 125BUILTIN(__builtin_amdgcn_readlane, "iii", "nc") 126BUILTIN(__builtin_amdgcn_fmed3f, "ffff", "nc") 127BUILTIN(__builtin_amdgcn_ds_faddf, "ff*3fIiIiIb", "n") 128BUILTIN(__builtin_amdgcn_ds_fminf, "ff*3fIiIiIb", "n") 129BUILTIN(__builtin_amdgcn_ds_fmaxf, "ff*3fIiIiIb", "n") 130BUILTIN(__builtin_amdgcn_ds_append, "ii*3", "n") 131BUILTIN(__builtin_amdgcn_ds_consume, "ii*3", "n") 132BUILTIN(__builtin_amdgcn_alignbit, "UiUiUiUi", "nc") 133BUILTIN(__builtin_amdgcn_alignbyte, "UiUiUiUi", "nc") 134BUILTIN(__builtin_amdgcn_ubfe, "UiUiUiUi", "nc") 135BUILTIN(__builtin_amdgcn_sbfe, "UiUiUiUi", "nc") 136BUILTIN(__builtin_amdgcn_cvt_pkrtz, "E2hff", "nc") 137BUILTIN(__builtin_amdgcn_cvt_pknorm_i16, "E2sff", "nc") 138BUILTIN(__builtin_amdgcn_cvt_pknorm_u16, "E2Usff", "nc") 139BUILTIN(__builtin_amdgcn_cvt_pk_i16, "E2sii", "nc") 140BUILTIN(__builtin_amdgcn_cvt_pk_u16, "E2UsUiUi", "nc") 141BUILTIN(__builtin_amdgcn_cvt_pk_u8_f32, "UifUiUi", "nc") 142BUILTIN(__builtin_amdgcn_sad_u8, "UiUiUiUi", "nc") 143BUILTIN(__builtin_amdgcn_msad_u8, "UiUiUiUi", "nc") 144BUILTIN(__builtin_amdgcn_sad_hi_u8, "UiUiUiUi", "nc") 145BUILTIN(__builtin_amdgcn_sad_u16, "UiUiUiUi", "nc") 146BUILTIN(__builtin_amdgcn_qsad_pk_u16_u8, "WUiWUiUiWUi", "nc") 147BUILTIN(__builtin_amdgcn_mqsad_pk_u16_u8, "WUiWUiUiWUi", "nc") 148BUILTIN(__builtin_amdgcn_mqsad_u32_u8, "V4UiWUiUiV4Ui", "nc") 149 150//===----------------------------------------------------------------------===// 151// Ballot builtins. 152//===----------------------------------------------------------------------===// 153 154TARGET_BUILTIN(__builtin_amdgcn_ballot_w32, "ZUib", "nc", "wavefrontsize32") 155BUILTIN(__builtin_amdgcn_ballot_w64, "WUib", "nc") 156 157// Deprecated intrinsics in favor of __builtin_amdgn_ballot_{w32|w64} 158BUILTIN(__builtin_amdgcn_uicmp, "WUiUiUiIi", "nc") 159BUILTIN(__builtin_amdgcn_uicmpl, "WUiWUiWUiIi", "nc") 160BUILTIN(__builtin_amdgcn_sicmp, "WUiiiIi", "nc") 161BUILTIN(__builtin_amdgcn_sicmpl, "WUiWiWiIi", "nc") 162BUILTIN(__builtin_amdgcn_fcmp, "WUiddIi", "nc") 163BUILTIN(__builtin_amdgcn_fcmpf, "WUiffIi", "nc") 164 165//===----------------------------------------------------------------------===// 166// Flat addressing builtins. 167//===----------------------------------------------------------------------===// 168BUILTIN(__builtin_amdgcn_is_shared, "bvC*0", "nc") 169BUILTIN(__builtin_amdgcn_is_private, "bvC*0", "nc") 170 171//===----------------------------------------------------------------------===// 172// GWS builtins. 173//===----------------------------------------------------------------------===// 174TARGET_BUILTIN(__builtin_amdgcn_ds_gws_init, "vUiUi", "n", "gws") 175TARGET_BUILTIN(__builtin_amdgcn_ds_gws_barrier, "vUiUi", "n", "gws") 176TARGET_BUILTIN(__builtin_amdgcn_ds_gws_sema_v, "vUi", "n", "gws") 177TARGET_BUILTIN(__builtin_amdgcn_ds_gws_sema_br, "vUiUi", "n", "gws") 178TARGET_BUILTIN(__builtin_amdgcn_ds_gws_sema_p, "vUi", "n", "gws") 179 180//===----------------------------------------------------------------------===// 181// CI+ only builtins. 182//===----------------------------------------------------------------------===// 183TARGET_BUILTIN(__builtin_amdgcn_s_dcache_inv_vol, "v", "n", "ci-insts") 184TARGET_BUILTIN(__builtin_amdgcn_buffer_wbinvl1_vol, "v", "n", "ci-insts") 185TARGET_BUILTIN(__builtin_amdgcn_ds_gws_sema_release_all, "vUi", "n", "ci-insts") 186 187//===----------------------------------------------------------------------===// 188// Interpolation builtins. 189//===----------------------------------------------------------------------===// 190BUILTIN(__builtin_amdgcn_interp_p1_f16, "ffUiUibUi", "nc") 191BUILTIN(__builtin_amdgcn_interp_p2_f16, "hffUiUibUi", "nc") 192BUILTIN(__builtin_amdgcn_interp_p1, "ffUiUiUi", "nc") 193BUILTIN(__builtin_amdgcn_interp_p2, "fffUiUiUi", "nc") 194BUILTIN(__builtin_amdgcn_interp_mov, "fUiUiUiUi", "nc") 195 196//===----------------------------------------------------------------------===// 197// VI+ only builtins. 198//===----------------------------------------------------------------------===// 199 200TARGET_BUILTIN(__builtin_amdgcn_div_fixuph, "hhhh", "nc", "16-bit-insts") 201TARGET_BUILTIN(__builtin_amdgcn_rcph, "hh", "nc", "16-bit-insts") 202TARGET_BUILTIN(__builtin_amdgcn_sqrth, "hh", "nc", "16-bit-insts") 203TARGET_BUILTIN(__builtin_amdgcn_rsqh, "hh", "nc", "16-bit-insts") 204TARGET_BUILTIN(__builtin_amdgcn_sinh, "hh", "nc", "16-bit-insts") 205TARGET_BUILTIN(__builtin_amdgcn_cosh, "hh", "nc", "16-bit-insts") 206TARGET_BUILTIN(__builtin_amdgcn_ldexph, "hhi", "nc", "16-bit-insts") 207TARGET_BUILTIN(__builtin_amdgcn_frexp_manth, "hh", "nc", "16-bit-insts") 208TARGET_BUILTIN(__builtin_amdgcn_frexp_exph, "sh", "nc", "16-bit-insts") 209TARGET_BUILTIN(__builtin_amdgcn_fracth, "hh", "nc", "16-bit-insts") 210TARGET_BUILTIN(__builtin_amdgcn_classh, "bhi", "nc", "16-bit-insts") 211TARGET_BUILTIN(__builtin_amdgcn_s_memrealtime, "WUi", "n", "s-memrealtime") 212TARGET_BUILTIN(__builtin_amdgcn_mov_dpp, "iiIiIiIiIb", "nc", "dpp") 213TARGET_BUILTIN(__builtin_amdgcn_update_dpp, "iiiIiIiIiIb", "nc", "dpp") 214TARGET_BUILTIN(__builtin_amdgcn_s_dcache_wb, "v", "n", "gfx8-insts") 215TARGET_BUILTIN(__builtin_amdgcn_perm, "UiUiUiUi", "nc", "gfx8-insts") 216 217//===----------------------------------------------------------------------===// 218// GFX9+ only builtins. 219//===----------------------------------------------------------------------===// 220 221TARGET_BUILTIN(__builtin_amdgcn_fmed3h, "hhhh", "nc", "gfx9-insts") 222 223TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_f64, "dd*1d", "t", "gfx90a-insts") 224TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_f32, "ff*1f", "t", "atomic-fadd-rtn-insts") 225TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_v2f16, "V2hV2h*1V2h", "t", "atomic-buffer-global-pk-add-f16-insts") 226TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fmin_f64, "dd*1d", "t", "gfx90a-insts") 227TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fmax_f64, "dd*1d", "t", "gfx90a-insts") 228 229TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_f64, "dd*0d", "t", "gfx90a-insts") 230TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fmin_f64, "dd*0d", "t", "gfx90a-insts") 231TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fmax_f64, "dd*0d", "t", "gfx90a-insts") 232 233TARGET_BUILTIN(__builtin_amdgcn_ds_atomic_fadd_f64, "dd*3d", "t", "gfx90a-insts") 234TARGET_BUILTIN(__builtin_amdgcn_ds_atomic_fadd_f32, "ff*3f", "t", "gfx8-insts") 235 236TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_f32, "ff*0f", "t", "gfx940-insts") 237TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_v2f16, "V2hV2h*0V2h", "t", "atomic-flat-pk-add-16-insts") 238TARGET_BUILTIN(__builtin_amdgcn_flat_atomic_fadd_v2bf16, "V2sV2s*0V2s", "t", "atomic-flat-pk-add-16-insts") 239TARGET_BUILTIN(__builtin_amdgcn_global_atomic_fadd_v2bf16, "V2sV2s*1V2s", "t", "atomic-global-pk-add-bf16-inst") 240TARGET_BUILTIN(__builtin_amdgcn_ds_atomic_fadd_v2bf16, "V2sV2s*3V2s", "t", "atomic-ds-pk-add-16-insts") 241TARGET_BUILTIN(__builtin_amdgcn_ds_atomic_fadd_v2f16, "V2hV2h*3V2h", "t", "atomic-ds-pk-add-16-insts") 242 243//===----------------------------------------------------------------------===// 244// Deep learning builtins. 245//===----------------------------------------------------------------------===// 246 247TARGET_BUILTIN(__builtin_amdgcn_fdot2, "fV2hV2hfIb", "nc", "dot10-insts") 248TARGET_BUILTIN(__builtin_amdgcn_fdot2_f16_f16, "hV2hV2hh", "nc", "dot9-insts") 249TARGET_BUILTIN(__builtin_amdgcn_fdot2_bf16_bf16, "sV2sV2ss", "nc", "dot9-insts") 250TARGET_BUILTIN(__builtin_amdgcn_fdot2_f32_bf16, "fV2sV2sfIb", "nc", "dot9-insts") 251TARGET_BUILTIN(__builtin_amdgcn_sdot2, "SiV2SsV2SsSiIb", "nc", "dot2-insts") 252TARGET_BUILTIN(__builtin_amdgcn_udot2, "UiV2UsV2UsUiIb", "nc", "dot2-insts") 253TARGET_BUILTIN(__builtin_amdgcn_sdot4, "SiSiSiSiIb", "nc", "dot1-insts") 254TARGET_BUILTIN(__builtin_amdgcn_udot4, "UiUiUiUiIb", "nc", "dot7-insts") 255TARGET_BUILTIN(__builtin_amdgcn_sudot4, "iIbiIbiiIb", "nc", "dot8-insts") 256TARGET_BUILTIN(__builtin_amdgcn_sdot8, "SiSiSiSiIb", "nc", "dot1-insts") 257TARGET_BUILTIN(__builtin_amdgcn_udot8, "UiUiUiUiIb", "nc", "dot7-insts") 258TARGET_BUILTIN(__builtin_amdgcn_sudot8, "iIbiIbiiIb", "nc", "dot8-insts") 259TARGET_BUILTIN(__builtin_amdgcn_dot4_f32_fp8_bf8, "fUiUif", "nc", "dot11-insts") 260TARGET_BUILTIN(__builtin_amdgcn_dot4_f32_bf8_fp8, "fUiUif", "nc", "dot11-insts") 261TARGET_BUILTIN(__builtin_amdgcn_dot4_f32_fp8_fp8, "fUiUif", "nc", "dot11-insts") 262TARGET_BUILTIN(__builtin_amdgcn_dot4_f32_bf8_bf8, "fUiUif", "nc", "dot11-insts") 263 264//===----------------------------------------------------------------------===// 265// GFX10+ only builtins. 266//===----------------------------------------------------------------------===// 267TARGET_BUILTIN(__builtin_amdgcn_permlane16, "UiUiUiUiUiIbIb", "nc", "gfx10-insts") 268TARGET_BUILTIN(__builtin_amdgcn_permlanex16, "UiUiUiUiUiIbIb", "nc", "gfx10-insts") 269TARGET_BUILTIN(__builtin_amdgcn_mov_dpp8, "UiUiIUi", "nc", "gfx10-insts") 270 271//===----------------------------------------------------------------------===// 272// Raytracing builtins. 273// By default the 1st argument is i32 and the 4/5-th arguments are float4. 274// Postfix l indicates the 1st argument is i64. 275// Postfix h indicates the 4/5-th arguments are half4. 276//===----------------------------------------------------------------------===// 277TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray, "V4UiUifV4fV4fV4fV4Ui", "nc", "gfx10-insts") 278TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray_h, "V4UiUifV4fV4hV4hV4Ui", "nc", "gfx10-insts") 279TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray_l, "V4UiWUifV4fV4fV4fV4Ui", "nc", "gfx10-insts") 280TARGET_BUILTIN(__builtin_amdgcn_image_bvh_intersect_ray_lh, "V4UiWUifV4fV4hV4hV4Ui", "nc", "gfx10-insts") 281 282 283//===----------------------------------------------------------------------===// 284// GFX11+ only builtins. 285//===----------------------------------------------------------------------===// 286 287// TODO: This is a no-op in wave32. Should the builtin require wavefrontsize64? 288TARGET_BUILTIN(__builtin_amdgcn_permlane64, "UiUi", "nc", "gfx11-insts") 289TARGET_BUILTIN(__builtin_amdgcn_s_wait_event_export_ready, "v", "n", "gfx11-insts") 290 291//===----------------------------------------------------------------------===// 292// WMMA builtins. 293// Postfix w32 indicates the builtin requires wavefront size of 32. 294// Postfix w64 indicates the builtin requires wavefront size of 64. 295//===----------------------------------------------------------------------===// 296TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w32, "V8fV16hV16hV8f", "nc", "gfx11-insts,wavefrontsize32") 297TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32, "V8fV16sV16sV8f", "nc", "gfx11-insts,wavefrontsize32") 298TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w32, "V16hV16hV16hV16hIb", "nc", "gfx11-insts,wavefrontsize32") 299TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32, "V16sV16sV16sV16sIb", "nc", "gfx11-insts,wavefrontsize32") 300TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_tied_w32, "V16hV16hV16hV16hIb", "nc", "gfx11-insts,wavefrontsize32") 301TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w32, "V16sV16sV16sV16sIb", "nc", "gfx11-insts,wavefrontsize32") 302TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32, "V8iIbV4iIbV4iV8iIb", "nc", "gfx11-insts,wavefrontsize32") 303TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32, "V8iIbV2iIbV2iV8iIb", "nc", "gfx11-insts,wavefrontsize32") 304 305TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w64, "V4fV16hV16hV4f", "nc", "gfx11-insts,wavefrontsize64") 306TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64, "V4fV16sV16sV4f", "nc", "gfx11-insts,wavefrontsize64") 307TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w64, "V8hV16hV16hV8hIb", "nc", "gfx11-insts,wavefrontsize64") 308TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64, "V8sV16sV16sV8sIb", "nc", "gfx11-insts,wavefrontsize64") 309TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_tied_w64, "V8hV16hV16hV8hIb", "nc", "gfx11-insts,wavefrontsize64") 310TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w64, "V8sV16sV16sV8sIb", "nc", "gfx11-insts,wavefrontsize64") 311TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64, "V4iIbV4iIbV4iV4iIb", "nc", "gfx11-insts,wavefrontsize64") 312TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64, "V4iIbV2iIbV2iV4iIb", "nc", "gfx11-insts,wavefrontsize64") 313 314TARGET_BUILTIN(__builtin_amdgcn_s_sendmsg_rtn, "UiUIi", "n", "gfx11-insts") 315TARGET_BUILTIN(__builtin_amdgcn_s_sendmsg_rtnl, "UWiUIi", "n", "gfx11-insts") 316 317TARGET_BUILTIN(__builtin_amdgcn_ds_bvh_stack_rtn, "V2UiUiUiV4UiIi", "n", "gfx11-insts") 318 319//===----------------------------------------------------------------------===// 320// Special builtins. 321//===----------------------------------------------------------------------===// 322BUILTIN(__builtin_amdgcn_read_exec, "WUi", "nc") 323BUILTIN(__builtin_amdgcn_read_exec_lo, "Ui", "nc") 324BUILTIN(__builtin_amdgcn_read_exec_hi, "Ui", "nc") 325 326BUILTIN(__builtin_amdgcn_endpgm, "v", "nr") 327 328BUILTIN(__builtin_amdgcn_get_fpenv, "WUi", "n") 329BUILTIN(__builtin_amdgcn_set_fpenv, "vWUi", "n") 330 331//===----------------------------------------------------------------------===// 332// R600-NI only builtins. 333//===----------------------------------------------------------------------===// 334 335BUILTIN(__builtin_r600_implicitarg_ptr, "Uc*7", "nc") 336 337BUILTIN(__builtin_r600_read_tgid_x, "Ui", "nc") 338BUILTIN(__builtin_r600_read_tgid_y, "Ui", "nc") 339BUILTIN(__builtin_r600_read_tgid_z, "Ui", "nc") 340 341BUILTIN(__builtin_r600_read_tidig_x, "Ui", "nc") 342BUILTIN(__builtin_r600_read_tidig_y, "Ui", "nc") 343BUILTIN(__builtin_r600_read_tidig_z, "Ui", "nc") 344 345BUILTIN(__builtin_r600_recipsqrt_ieee, "dd", "nc") 346BUILTIN(__builtin_r600_recipsqrt_ieeef, "ff", "nc") 347 348//===----------------------------------------------------------------------===// 349// MFMA builtins. 350//===----------------------------------------------------------------------===// 351 352TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x1f32, "V32fffV32fIiIiIi", "nc", "mai-insts") 353TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x1f32, "V16fffV16fIiIiIi", "nc", "mai-insts") 354TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x1f32, "V4fffV4fIiIiIi", "nc", "mai-insts") 355TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x2f32, "V16fffV16fIiIiIi", "nc", "mai-insts") 356TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x4f32, "V4fffV4fIiIiIi", "nc", "mai-insts") 357TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4f16, "V32fV4hV4hV32fIiIiIi", "nc", "mai-insts") 358TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x4f16, "V16fV4hV4hV16fIiIiIi", "nc", "mai-insts") 359TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x4f16, "V4fV4hV4hV4fIiIiIi", "nc", "mai-insts") 360TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x8f16, "V16fV4hV4hV16fIiIiIi", "nc", "mai-insts") 361TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x16f16, "V4fV4hV4hV4fIiIiIi", "nc", "mai-insts") 362TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_32x32x4i8, "V32iiiV32iIiIiIi", "nc", "mai-insts") 363TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_16x16x4i8, "V16iiiV16iIiIiIi", "nc", "mai-insts") 364TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_4x4x4i8, "V4iiiV4iIiIiIi", "nc", "mai-insts") 365TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_32x32x8i8, "V16iiiV16iIiIiIi", "nc", "mai-insts") 366TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_16x16x16i8, "V4iiiV4iIiIiIi", "nc", "mai-insts") 367TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x2bf16, "V32fV2sV2sV32fIiIiIi", "nc", "mai-insts") 368TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x2bf16, "V16fV2sV2sV16fIiIiIi", "nc", "mai-insts") 369TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x2bf16, "V4fV2sV2sV4fIiIiIi", "nc", "mai-insts") 370TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4bf16, "V16fV2sV2sV16fIiIiIi", "nc", "mai-insts") 371TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x8bf16, "V4fV2sV2sV4fIiIiIi", "nc", "mai-insts") 372 373TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4bf16_1k, "V32fV4sV4sV32fIiIiIi", "nc", "mai-insts") 374TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x4bf16_1k, "V16fV4sV4sV16fIiIiIi", "nc", "mai-insts") 375TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_4x4x4bf16_1k, "V4fV4sV4sV4fIiIiIi", "nc", "mai-insts") 376TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x8bf16_1k, "V16fV4sV4sV16fIiIiIi", "nc", "mai-insts") 377TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x16bf16_1k, "V4fV4sV4sV4fIiIiIi", "nc", "mai-insts") 378TARGET_BUILTIN(__builtin_amdgcn_mfma_f64_16x16x4f64, "V4dddV4dIiIiIi", "nc", "mai-insts") 379TARGET_BUILTIN(__builtin_amdgcn_mfma_f64_4x4x4f64, "ddddIiIiIi", "nc", "mai-insts") 380 381TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_16x16x32_i8, "V4iWiWiV4iIiIiIi", "nc", "mai-insts") 382TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_32x32x16_i8, "V16iWiWiV16iIiIiIi", "nc", "mai-insts") 383TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x8_xf32, "V4fV2fV2fV4fIiIiIi", "nc", "mai-insts") 384TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x4_xf32, "V16fV2fV2fV16fIiIiIi", "nc", "mai-insts") 385TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_bf8_bf8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") 386TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_bf8_fp8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") 387TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_fp8_bf8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") 388TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8, "V4fWiWiV4fIiIiIi", "nc", "fp8-insts") 389TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") 390TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") 391TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") 392TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8, "V16fWiWiV16fIiIiIi", "nc", "fp8-insts") 393TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x32_f16, "V4fV4hV8hV4fiIiIi", "nc", "mai-insts") 394TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x16_f16, "V16fV4hV8hV16fiIiIi", "nc", "mai-insts") 395TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x32_bf16, "V4fV4sV8sV4fiIiIi", "nc", "mai-insts") 396TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x16_bf16, "V16fV4sV8sV16fiIiIi", "nc", "mai-insts") 397TARGET_BUILTIN(__builtin_amdgcn_smfmac_i32_16x16x64_i8, "V4iV2iV4iV4iiIiIi", "nc", "mai-insts") 398TARGET_BUILTIN(__builtin_amdgcn_smfmac_i32_32x32x32_i8, "V16iV2iV4iV16iiIiIi", "nc", "mai-insts") 399TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_bf8_bf8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") 400TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_bf8_fp8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") 401TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_fp8_bf8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") 402TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_fp8_fp8, "V4fV2iV4iV4fiIiIi", "nc", "fp8-insts") 403TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_bf8_bf8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") 404TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_bf8_fp8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") 405TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_fp8_bf8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") 406TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_fp8_fp8, "V16fV2iV4iV16fiIiIi", "nc", "fp8-insts") 407 408TARGET_BUILTIN(__builtin_amdgcn_cvt_f32_bf8, "fiIi", "nc", "fp8-conversion-insts") 409TARGET_BUILTIN(__builtin_amdgcn_cvt_f32_fp8, "fiIi", "nc", "fp8-conversion-insts") 410TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_f32_bf8, "V2fiIb", "nc", "fp8-conversion-insts") 411TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_f32_fp8, "V2fiIb", "nc", "fp8-conversion-insts") 412TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_bf8_f32, "iffiIb", "nc", "fp8-conversion-insts") 413TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_fp8_f32, "iffiIb", "nc", "fp8-conversion-insts") 414TARGET_BUILTIN(__builtin_amdgcn_cvt_sr_bf8_f32, "ifiiIi", "nc", "fp8-conversion-insts") 415TARGET_BUILTIN(__builtin_amdgcn_cvt_sr_fp8_f32, "ifiiIi", "nc", "fp8-conversion-insts") 416 417//===----------------------------------------------------------------------===// 418// GFX12+ only builtins. 419//===----------------------------------------------------------------------===// 420 421TARGET_BUILTIN(__builtin_amdgcn_s_sleep_var, "vUi", "n", "gfx12-insts") 422TARGET_BUILTIN(__builtin_amdgcn_permlane16_var, "UiUiUiUiIbIb", "nc", "gfx12-insts") 423TARGET_BUILTIN(__builtin_amdgcn_permlanex16_var, "UiUiUiUiIbIb", "nc", "gfx12-insts") 424TARGET_BUILTIN(__builtin_amdgcn_s_barrier_signal, "vIi", "n", "gfx12-insts") 425TARGET_BUILTIN(__builtin_amdgcn_s_barrier_signal_var, "vi", "n", "gfx12-insts") 426TARGET_BUILTIN(__builtin_amdgcn_s_barrier_wait, "vIs", "n", "gfx12-insts") 427TARGET_BUILTIN(__builtin_amdgcn_s_barrier_signal_isfirst, "bIi", "n", "gfx12-insts") 428TARGET_BUILTIN(__builtin_amdgcn_s_barrier_signal_isfirst_var, "bi", "n", "gfx12-insts") 429TARGET_BUILTIN(__builtin_amdgcn_s_barrier_init, "vii", "n", "gfx12-insts") 430TARGET_BUILTIN(__builtin_amdgcn_s_barrier_join, "vi", "n", "gfx12-insts") 431TARGET_BUILTIN(__builtin_amdgcn_s_wakeup_barrier, "vi", "n", "gfx12-insts") 432TARGET_BUILTIN(__builtin_amdgcn_s_barrier_leave, "b", "n", "gfx12-insts") 433TARGET_BUILTIN(__builtin_amdgcn_s_get_barrier_state, "Uii", "n", "gfx12-insts") 434 435TARGET_BUILTIN(__builtin_amdgcn_global_load_tr_v2i32, "V2iV2i*1", "nc", "gfx12-insts,wavefrontsize32") 436TARGET_BUILTIN(__builtin_amdgcn_global_load_tr_v8i16, "V8sV8s*1", "nc", "gfx12-insts,wavefrontsize32") 437TARGET_BUILTIN(__builtin_amdgcn_global_load_tr_v8f16, "V8hV8h*1", "nc", "gfx12-insts,wavefrontsize32") 438 439TARGET_BUILTIN(__builtin_amdgcn_global_load_tr_i32, "ii*1", "nc", "gfx12-insts,wavefrontsize64") 440TARGET_BUILTIN(__builtin_amdgcn_global_load_tr_v4i16, "V4sV4s*1", "nc", "gfx12-insts,wavefrontsize64") 441TARGET_BUILTIN(__builtin_amdgcn_global_load_tr_v4f16, "V4hV4h*1", "nc", "gfx12-insts,wavefrontsize64") 442 443//===----------------------------------------------------------------------===// 444// WMMA builtins. 445// Postfix w32 indicates the builtin requires wavefront size of 32. 446// Postfix w64 indicates the builtin requires wavefront size of 64. 447// 448// Some of these are very similar to their GFX11 counterparts, but they don't 449// require replication of the A,B matrices, so they use fewer vector elements. 450// Therefore, we add an "_gfx12" suffix to distinguish them from the existing 451// builtins. 452//===----------------------------------------------------------------------===// 453TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12, "V8fV8hV8hV8f", "nc", "gfx12-insts,wavefrontsize32") 454TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12, "V8fV8sV8sV8f", "nc", "gfx12-insts,wavefrontsize32") 455TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w32_gfx12, "V8hV8hV8hV8h", "nc", "gfx12-insts,wavefrontsize32") 456TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32_gfx12, "V8sV8sV8sV8s", "nc", "gfx12-insts,wavefrontsize32") 457TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12, "V8iIbV2iIbV2iV8iIb", "nc", "gfx12-insts,wavefrontsize32") 458TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32_gfx12, "V8iIbiIbiV8iIb", "nc", "gfx12-insts,wavefrontsize32") 459// These are gfx12-only, but for consistency with the other WMMA variants we're 460// keeping the "_gfx12" suffix. 461TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_fp8_fp8_w32_gfx12, "V8fV2iV2iV8f", "nc", "gfx12-insts,wavefrontsize32") 462TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_fp8_bf8_w32_gfx12, "V8fV2iV2iV8f", "nc", "gfx12-insts,wavefrontsize32") 463TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf8_fp8_w32_gfx12, "V8fV2iV2iV8f", "nc", "gfx12-insts,wavefrontsize32") 464TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf8_bf8_w32_gfx12, "V8fV2iV2iV8f", "nc", "gfx12-insts,wavefrontsize32") 465TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x32_iu4_w32_gfx12, "V8iIbV2iIbV2iV8iIb", "nc", "gfx12-insts,wavefrontsize32") 466 467TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w64_gfx12, "V4fV4hV4hV4f", "nc", "gfx12-insts,wavefrontsize64") 468TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64_gfx12, "V4fV4sV4sV4f", "nc", "gfx12-insts,wavefrontsize64") 469TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w64_gfx12, "V4hV4hV4hV4h", "nc", "gfx12-insts,wavefrontsize64") 470TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64_gfx12, "V4sV4sV4sV4s", "nc", "gfx12-insts,wavefrontsize64") 471TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64_gfx12, "V4iIbiIbiV4iIb", "nc", "gfx12-insts,wavefrontsize64") 472TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64_gfx12, "V4iIbiIbiV4iIb", "nc", "gfx12-insts,wavefrontsize64") 473// These are gfx12-only, but for consistency with the other WMMA variants we're 474// keeping the "_gfx12" suffix. 475TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_fp8_fp8_w64_gfx12, "V4fiiV4f", "nc", "gfx12-insts,wavefrontsize64") 476TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_fp8_bf8_w64_gfx12, "V4fiiV4f", "nc", "gfx12-insts,wavefrontsize64") 477TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf8_fp8_w64_gfx12, "V4fiiV4f", "nc", "gfx12-insts,wavefrontsize64") 478TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf8_bf8_w64_gfx12, "V4fiiV4f", "nc", "gfx12-insts,wavefrontsize64") 479TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x32_iu4_w64_gfx12, "V4iIbiIbiV4iIb", "nc", "gfx12-insts,wavefrontsize64") 480 481TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_f16_w32, "V8fV8hV16hV8fs", "nc", "gfx12-insts,wavefrontsize32") 482TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf16_w32, "V8fV8sV16sV8fs", "nc", "gfx12-insts,wavefrontsize32") 483TARGET_BUILTIN(__builtin_amdgcn_swmmac_f16_16x16x32_f16_w32, "V8hV8hV16hV8hs", "nc", "gfx12-insts,wavefrontsize32") 484TARGET_BUILTIN(__builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w32, "V8sV8sV16sV8ss", "nc", "gfx12-insts,wavefrontsize32") 485TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu8_w32, "V8iIbV2iIbV4iV8isIb", "nc", "gfx12-insts,wavefrontsize32") 486TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu4_w32, "V8iIbiIbV2iV8isIb", "nc", "gfx12-insts,wavefrontsize32") 487TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x64_iu4_w32, "V8iIbV2iIbV4iV8isIb", "nc", "gfx12-insts,wavefrontsize32") 488TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w32, "V8fV2iV4iV8fs", "nc", "gfx12-insts,wavefrontsize32") 489TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w32, "V8fV2iV4iV8fs", "nc", "gfx12-insts,wavefrontsize32") 490TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w32, "V8fV2iV4iV8fs", "nc", "gfx12-insts,wavefrontsize32") 491TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w32, "V8fV2iV4iV8fs", "nc", "gfx12-insts,wavefrontsize32") 492 493TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_f16_w64, "V4fV4hV8hV4fs", "nc", "gfx12-insts,wavefrontsize64") 494TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf16_w64, "V4fV4sV8sV4fs", "nc", "gfx12-insts,wavefrontsize64") 495TARGET_BUILTIN(__builtin_amdgcn_swmmac_f16_16x16x32_f16_w64, "V4hV4hV8hV4hs", "nc", "gfx12-insts,wavefrontsize64") 496TARGET_BUILTIN(__builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w64, "V4sV4sV8sV4ss", "nc", "gfx12-insts,wavefrontsize64") 497TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu8_w64, "V4iIbiIbV2iV4isIb", "nc", "gfx12-insts,wavefrontsize64") 498TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu4_w64, "V4iIbiIbiV4isIb", "nc", "gfx12-insts,wavefrontsize64") 499TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x64_iu4_w64, "V4iIbiIbV2iV4isIb", "nc", "gfx12-insts,wavefrontsize64") 500TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w64, "V4fiV2iV4fs", "nc", "gfx12-insts,wavefrontsize64") 501TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w64, "V4fiV2iV4fs", "nc", "gfx12-insts,wavefrontsize64") 502TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w64, "V4fiV2iV4fs", "nc", "gfx12-insts,wavefrontsize64") 503TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w64, "V4fiV2iV4fs", "nc", "gfx12-insts,wavefrontsize64") 504 505#undef BUILTIN 506#undef TARGET_BUILTIN 507