xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/VOP3PInstructions.td (revision 1db9f3b21e39176dd5b67cf8ac378633b172463e)
1//===-- VOP3PInstructions.td - Vector Instruction Definitions -------------===//
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//===----------------------------------------------------------------------===//
10// VOP3P Classes
11//===----------------------------------------------------------------------===//
12
13class VOP3P_Profile<VOPProfile P, VOP3Features Features = VOP3_REGULAR,
14                    bit HasDPP = 0> : VOP3_Profile<P, Features> {
15  let IsVOP3P = 1;
16  let HasExtVOP3DPP = HasDPP;
17  // We do not want to print src modifiers for vop3p because the bits are
18  // overloaded in meaning and the logic in printOperandAndFPInputMods is
19  // wrong for vop3p
20  let AsmVOP3Base = AsmVOP3P;
21}
22
23// Used for FMA_MIX* and MAD_MIX* insts
24// Their operands are only sort of f16 operands. Depending on
25// op_sel_hi, these may be interpreted as f32. The inline immediate
26// values are really f16 converted to f32, so we treat these as f16
27// operands.
28class VOP3P_Mix_Profile<VOPProfile P, VOP3Features Features = VOP3_REGULAR,
29                    bit useTiedOutput = 0> : VOP3P_Profile<P, Features, 1> {
30    bit UseTiedOutput = useTiedOutput;
31
32    dag srcs =
33          (ins FP16InputMods:$src0_modifiers, VCSrc_f16:$src0,
34               FP16InputMods:$src1_modifiers, VCSrc_f16:$src1,
35               FP16InputMods:$src2_modifiers, VCSrc_f16:$src2);
36    dag dpp_srcs =
37          (ins FPVRegInputMods:$src0_modifiers, VGPRSrc_32:$src0,
38               FPVRegInputMods:$src1_modifiers, VRegSrc_32:$src1,
39               FP16InputMods:$src2_modifiers, VCSrc_f16:$src2);
40
41           // FIXME: clampmod0 misbehaves with the non-default vdst_in
42           // following it. For now workaround this by requiring clamp
43           // in tied patterns. This should use undef_tied_input, but it
44           // seems underdeveloped and doesn't apply the right register
45           // class constraints.
46    dag mods = !con(!if(UseTiedOutput, (ins clampmod:$clamp, VGPR_32:$vdst_in),
47                        (ins clampmod0:$clamp)),
48                    (ins op_sel0:$op_sel, op_sel_hi0:$op_sel_hi));
49    // We use Ins64 because that is the one which populates InOperandList
50    // due to the logic in class VOP3_Pseudo
51    let Ins64 = !con(srcs, mods);
52    let InsVOP3Base = !con(dpp_srcs, mods);
53    let AsmVOP3Base =
54      "$vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$op_sel$op_sel_hi$clamp";
55}
56
57multiclass VOP3PInst<string OpName, VOPProfile P,
58                     SDPatternOperator node = null_frag, bit IsDOT = 0> {
59  def NAME : VOP3P_Pseudo<OpName, P,
60                          !if (P.HasModifiers,
61                               getVOP3PModPat<P, node, IsDOT, IsDOT>.ret,
62                               getVOP3Pat<P, node>.ret)>;
63  let SubtargetPredicate = isGFX11Plus in {
64  if P.HasExtVOP3DPP then
65    def _dpp : VOP3_DPP_Pseudo<OpName, P> {
66      let VOP3P = 1;
67      let PseudoInstr = OpName #"_dpp";
68    }
69  } // end SubtargetPredicate = isGFX11Plus
70}
71
72// Non-packed instructions that use the VOP3P encoding.
73// VOP3 neg/abs and VOP3P opsel/opsel_hi modifiers are allowed.
74multiclass VOP3_VOP3PInst<string OpName, VOP3P_Mix_Profile P> {
75  def NAME : VOP3P_Pseudo<OpName, P> {
76    let Constraints = !if(P.UseTiedOutput, "$vdst = $vdst_in", "");
77    let DisableEncoding = !if(P.UseTiedOutput, "$vdst_in", "");
78  }
79  let SubtargetPredicate = isGFX11Plus in {
80    if P.HasExtVOP3DPP then
81      def _dpp : VOP3_DPP_Pseudo<OpName, P> {
82        let VOP3P = 1;
83        let PseudoInstr = OpName#"_dpp";
84        let Constraints = !if(P.UseTiedOutput, "$vdst = $vdst_in", "");
85        let DisableEncoding = !if(P.UseTiedOutput, "$vdst_in", "");
86      }
87  } // end SubtargetPredicate = isGFX11Plus
88}
89
90let isReMaterializable = 1 in {
91let isCommutable = 1 in {
92defm V_PK_MAD_I16 : VOP3PInst<"v_pk_mad_i16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
93defm V_PK_MAD_U16 : VOP3PInst<"v_pk_mad_u16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
94
95let FPDPRounding = 1 in {
96defm V_PK_FMA_F16 : VOP3PInst<"v_pk_fma_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, any_fma>;
97defm V_PK_ADD_F16 : VOP3PInst<"v_pk_add_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16>, any_fadd>;
98defm V_PK_MUL_F16 : VOP3PInst<"v_pk_mul_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16>, any_fmul>;
99} // End FPDPRounding = 1
100defm V_PK_MAX_F16 : VOP3PInst<"v_pk_max_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16>, fmaxnum_like>;
101defm V_PK_MIN_F16 : VOP3PInst<"v_pk_min_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16>, fminnum_like>;
102
103defm V_PK_ADD_U16 : VOP3PInst<"v_pk_add_u16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, add>;
104defm V_PK_ADD_I16 : VOP3PInst<"v_pk_add_i16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>>;
105defm V_PK_MUL_LO_U16 : VOP3PInst<"v_pk_mul_lo_u16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, mul>;
106
107defm V_PK_MIN_I16 : VOP3PInst<"v_pk_min_i16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, smin>;
108defm V_PK_MIN_U16 : VOP3PInst<"v_pk_min_u16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, umin>;
109defm V_PK_MAX_I16 : VOP3PInst<"v_pk_max_i16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, smax>;
110defm V_PK_MAX_U16 : VOP3PInst<"v_pk_max_u16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, umax>;
111
112let SubtargetPredicate = isGFX12Plus, ReadsModeReg = 0 in {
113defm V_PK_MAXIMUM_F16 : VOP3PInst<"v_pk_maximum_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16>, fmaximum>;
114defm V_PK_MINIMUM_F16 : VOP3PInst<"v_pk_minimum_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16>, fminimum>;
115} // End SubtargetPredicate = isGFX12Plus, ReadsModeReg = 0
116}
117
118defm V_PK_SUB_U16 : VOP3PInst<"v_pk_sub_u16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>>;
119defm V_PK_SUB_I16 : VOP3PInst<"v_pk_sub_i16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, sub>;
120
121defm V_PK_LSHLREV_B16 : VOP3PInst<"v_pk_lshlrev_b16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, clshl_rev_16>;
122defm V_PK_ASHRREV_I16 : VOP3PInst<"v_pk_ashrrev_i16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, cashr_rev_16>;
123defm V_PK_LSHRREV_B16 : VOP3PInst<"v_pk_lshrrev_b16", VOP3P_Profile<VOP_V2I16_V2I16_V2I16>, clshr_rev_16>;
124} // End isReMaterializable = 1
125
126let SubtargetPredicate = HasVOP3PInsts in {
127
128// Integer operations with clamp bit set.
129class VOP3PSatPat<SDPatternOperator pat, Instruction inst> : GCNPat<
130  (pat (v2i16 (VOP3PMods v2i16:$src0, i32:$src0_modifiers)),
131       (v2i16 (VOP3PMods v2i16:$src1, i32:$src1_modifiers))),
132  (inst $src0_modifiers, $src0, $src1_modifiers, $src1, DSTCLAMP.ENABLE)
133>;
134
135def : VOP3PSatPat<uaddsat, V_PK_ADD_U16>;
136def : VOP3PSatPat<saddsat, V_PK_ADD_I16>;
137def : VOP3PSatPat<usubsat, V_PK_SUB_U16>;
138def : VOP3PSatPat<ssubsat, V_PK_SUB_I16>;
139} // End SubtargetPredicate = HasVOP3PInsts
140
141// TODO: Make sure we're doing the right thing with denormals. Note
142// that FMA and MAD will differ.
143multiclass MadFmaMixPats<SDPatternOperator fma_like,
144                         Instruction mix_inst,
145                         Instruction mixlo_inst,
146                         Instruction mixhi_inst> {
147  // At least one of the operands needs to be an fpextend of an f16
148  // for this to be worthwhile, so we need three patterns here.
149  // TODO: Could we use a predicate to inspect src1/2/3 instead?
150  def : GCNPat <
151    (f32 (fma_like (f32 (VOP3PMadMixModsExt f16:$src0, i32:$src0_mods)),
152                   (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_mods)),
153                   (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_mods)))),
154    (mix_inst $src0_mods, $src0, $src1_mods, $src1, $src2_mods, $src2,
155              DSTCLAMP.NONE)>;
156  def : GCNPat <
157    (f32 (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_mods)),
158                   (f32 (VOP3PMadMixModsExt f16:$src1, i32:$src1_mods)),
159                   (f32 (VOP3PMadMixMods f32:$src2, i32:$src2_mods)))),
160    (mix_inst $src0_mods, $src0, $src1_mods, $src1, $src2_mods, $src2,
161              DSTCLAMP.NONE)>;
162  def : GCNPat <
163    (f32 (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_mods)),
164                   (f32 (VOP3PMadMixMods f32:$src1, i32:$src1_mods)),
165                   (f32 (VOP3PMadMixModsExt f16:$src2, i32:$src2_mods)))),
166    (mix_inst $src0_mods, $src0, $src1_mods, $src1, $src2_mods, $src2,
167              DSTCLAMP.NONE)>;
168
169  def : GCNPat <
170    (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
171                            (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
172                            (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
173    (mixlo_inst $src0_modifiers, $src0,
174                $src1_modifiers, $src1,
175                $src2_modifiers, $src2,
176                DSTCLAMP.NONE,
177                (i32 (IMPLICIT_DEF)))
178  >;
179
180  // FIXME: Special case handling for maxhi (especially for clamp)
181  // because dealing with the write to high half of the register is
182  // difficult.
183  def : GCNPat <
184    (build_vector f16:$elt0, (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
185                                                     (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
186                                                     (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers)))))),
187    (v2f16 (mixhi_inst $src0_modifiers, $src0,
188                       $src1_modifiers, $src1,
189                       $src2_modifiers, $src2,
190                       DSTCLAMP.NONE,
191                       VGPR_32:$elt0))
192  >;
193
194  def : GCNPat <
195    (build_vector
196      f16:$elt0,
197      (AMDGPUclamp (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
198                                      (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
199                                      (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))))),
200    (v2f16 (mixhi_inst $src0_modifiers, $src0,
201                       $src1_modifiers, $src1,
202                       $src2_modifiers, $src2,
203                       DSTCLAMP.ENABLE,
204                       VGPR_32:$elt0))
205  >;
206
207  def : GCNPat <
208    (AMDGPUclamp (build_vector
209      (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$lo_src0, i32:$lo_src0_modifiers)),
210                         (f32 (VOP3PMadMixMods f16:$lo_src1, i32:$lo_src1_modifiers)),
211                         (f32 (VOP3PMadMixMods f16:$lo_src2, i32:$lo_src2_modifiers))))),
212      (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$hi_src0, i32:$hi_src0_modifiers)),
213                         (f32 (VOP3PMadMixMods f16:$hi_src1, i32:$hi_src1_modifiers)),
214                         (f32 (VOP3PMadMixMods f16:$hi_src2, i32:$hi_src2_modifiers))))))),
215    (v2f16 (mixhi_inst $hi_src0_modifiers, $hi_src0,
216                       $hi_src1_modifiers, $hi_src1,
217                       $hi_src2_modifiers, $hi_src2,
218                       DSTCLAMP.ENABLE,
219                       (mixlo_inst $lo_src0_modifiers, $lo_src0,
220                                   $lo_src1_modifiers, $lo_src1,
221                                   $lo_src2_modifiers, $lo_src2,
222                                   DSTCLAMP.ENABLE,
223                                   (i32 (IMPLICIT_DEF)))))
224  >;
225
226  def : GCNPat <
227    (f16 (fpround (fmul (f32 (VOP3PMadMixMods f32:$src0, i32:$src0_modifiers)),
228                        (f32 (VOP3PMadMixMods f32:$src1, i32:$src1_modifiers))))),
229    (mixlo_inst $src0_modifiers, $src0,
230                $src1_modifiers, $src1,
231                (i32 0), (i32 0),
232                DSTCLAMP.NONE,
233                (i32 (IMPLICIT_DEF)))
234  >;
235
236  def : GCNPat <
237    (build_vector f16:$elt0, (f16 (fpround (fmul (f32 (VOP3PMadMixMods f32:$src0, i32:$src0_modifiers)),
238                                            (f32 (VOP3PMadMixMods f32:$src1, i32:$src1_modifiers)))))),
239    (v2f16 (mixhi_inst $src0_modifiers, $src0,
240                       $src1_modifiers, $src1,
241                       (i32 0), (i32 0),
242                       DSTCLAMP.NONE,
243                       VGPR_32:$elt0))
244  >;
245}
246
247let SubtargetPredicate = HasMadMixInsts, OtherPredicates = [NoFP32Denormals] in {
248
249// These are VOP3a-like opcodes which accept no omod.
250// Size of src arguments (16/32) is controlled by op_sel.
251// For 16-bit src arguments their location (hi/lo) are controlled by op_sel_hi.
252let isCommutable = 1, mayRaiseFPException = 0 in {
253let isReMaterializable = 1 in
254defm V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3P_Mix_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
255
256let FPDPRounding = 1 in {
257// Clamp modifier is applied after conversion to f16.
258defm V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3P_Mix_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL, 1>>;
259
260let ClampLo = 0, ClampHi = 1 in {
261defm V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3P_Mix_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL, 1>>;
262}
263} // End FPDPRounding = 1
264}
265
266defm : MadFmaMixPats<fmad, V_MAD_MIX_F32, V_MAD_MIXLO_F16, V_MAD_MIXHI_F16>;
267} // End SubtargetPredicate = HasMadMixInsts, OtherPredicates = [NoFP32Denormals]
268
269
270// Essentially the same as the mad_mix versions
271let SubtargetPredicate = HasFmaMixInsts in {
272let isCommutable = 1 in {
273
274let isReMaterializable = 1 in
275defm V_FMA_MIX_F32 : VOP3_VOP3PInst<"v_fma_mix_f32", VOP3P_Mix_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
276
277let FPDPRounding = 1 in {
278// Clamp modifier is applied after conversion to f16.
279defm V_FMA_MIXLO_F16 : VOP3_VOP3PInst<"v_fma_mixlo_f16", VOP3P_Mix_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL, 1>>;
280
281let ClampLo = 0, ClampHi = 1 in {
282defm V_FMA_MIXHI_F16 : VOP3_VOP3PInst<"v_fma_mixhi_f16", VOP3P_Mix_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL, 1>>;
283}
284} // End FPDPRounding = 1
285}
286
287defm : MadFmaMixPats<fma, V_FMA_MIX_F32, V_FMA_MIXLO_F16, V_FMA_MIXHI_F16>;
288}
289
290// Defines patterns that extract signed 4bit from each Idx[0].
291foreach Idx = [[0,28],[4,24],[8,20],[12,16],[16,12],[20,8],[24,4]] in
292  def ExtractSigned4bit_#Idx[0] : PatFrag<(ops node:$src),
293                                          (sra (shl node:$src, (i32 Idx[1])), (i32 28))>;
294
295// Defines code pattern that extracts U(unsigned/signed) 4/8bit from FromBitIndex.
296class Extract<int FromBitIndex, int BitMask, bit U>: PatFrag<
297  (ops node:$src),
298  !if (!or (!and (!eq (BitMask, 255), !eq (FromBitIndex, 24)), !eq (FromBitIndex, 28)), // last element
299       !if (U, (srl node:$src, (i32 FromBitIndex)), (sra node:$src, (i32 FromBitIndex))),
300       !if (!eq (FromBitIndex, 0), // first element
301            !if (U, (and node:$src, (i32 BitMask)),
302                 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
303                                         (sext_inreg node:$src, i8))),
304            !if (U, (and (srl node:$src, (i32 FromBitIndex)), (i32 BitMask)),
305                 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
306                      (sext_inreg (srl node:$src, (i32 FromBitIndex)), i8)))))>;
307
308
309foreach Type = ["I", "U"] in
310  foreach Index = 0-3 in {
311    // Defines patterns that extract each Index'ed 8bit from an unsigned
312    // 32bit scalar value;
313    def Type#Index#"_8bit" : Extract<!shl(Index, 3), 255, !eq (Type, "U")>;
314
315    // Defines multiplication patterns where the multiplication is happening on each
316    // Index'ed 8bit of a 32bit scalar value.
317
318    def Mul#Type#_Elt#Index : PatFrag<
319      (ops node:$src0, node:$src1),
320      (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), AMDGPUmul_i24_oneuse, AMDGPUmul_u24_oneuse))
321                            (!cast<Extract>(Type#Index#"_8bit") node:$src0),
322                            (!cast<Extract>(Type#Index#"_8bit") node:$src1))>;
323  }
324
325// Different variants of dot8 patterns cause a huge increase in the compile time.
326// Define non-associative/commutative add/mul to prevent permutation in the dot8
327// pattern.
328def NonACAdd        : SDNode<"ISD::ADD"       , SDTIntBinOp>;
329def NonACAdd_oneuse : HasOneUseBinOp<NonACAdd>;
330
331def NonACAMDGPUmul_u24        : SDNode<"AMDGPUISD::MUL_U24"       , SDTIntBinOp>;
332def NonACAMDGPUmul_u24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_u24>;
333
334def NonACAMDGPUmul_i24        : SDNode<"AMDGPUISD::MUL_I24"       , SDTIntBinOp>;
335def NonACAMDGPUmul_i24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_i24>;
336
337foreach Type = ["I", "U"] in
338  foreach Index = 0-7 in {
339    // Defines patterns that extract each Index'ed 4bit from an unsigned
340    // 32bit scalar value;
341    def Type#Index#"_4bit" : Extract<!shl(Index, 2), 15, !eq (Type, "U")>;
342
343    // Defines multiplication patterns where the multiplication is happening on each
344    // Index'ed 8bit of a 32bit scalar value.
345    def Mul#Type#Index#"_4bit" : PatFrag<
346      (ops node:$src0, node:$src1),
347      (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), NonACAMDGPUmul_i24_oneuse, NonACAMDGPUmul_u24_oneuse))
348                             (!cast<Extract>(Type#Index#"_4bit") node:$src0),
349                             (!cast<Extract>(Type#Index#"_4bit") node:$src1))>;
350  }
351
352class UDot2Pat<VOP_Pseudo Inst> : GCNPat <
353  (add (add_oneuse (AMDGPUmul_u24_oneuse (srl i32:$src0, (i32 16)),
354                                         (srl i32:$src1, (i32 16))), i32:$src2),
355       (AMDGPUmul_u24_oneuse (and i32:$src0, (i32 65535)),
356                             (and i32:$src1, (i32 65535)))
357   ),
358  (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
359  let Predicates = Inst.Predicates;
360}
361
362class SDot2Pat<VOP_Pseudo Inst> : GCNPat <
363  (add (add_oneuse (AMDGPUmul_i24_oneuse (sra i32:$src0, (i32 16)),
364                                         (sra i32:$src1, (i32 16))), i32:$src2),
365       (AMDGPUmul_i24_oneuse (sext_inreg i32:$src0, i16),
366                             (sext_inreg i32:$src1, i16))),
367  (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
368  let Predicates = Inst.Predicates;
369}
370
371let IsDOT = 1 in {
372let OtherPredicates = [HasDot2Insts] in {
373defm V_DOT2_I32_I16 : VOP3PInst<"v_dot2_i32_i16",
374  VOP3P_Profile<VOP_I32_V2I16_V2I16_I32>, int_amdgcn_sdot2, 1>;
375defm V_DOT2_U32_U16 : VOP3PInst<"v_dot2_u32_u16",
376  VOP3P_Profile<VOP_I32_V2I16_V2I16_I32>, int_amdgcn_udot2, 1>;
377} // End OtherPredicates = [HasDot2Insts]
378
379let OtherPredicates = [HasDot10Insts] in
380defm V_DOT2_F32_F16 : VOP3PInst<"v_dot2_f32_f16",
381  VOP3P_Profile<VOP_F32_V2F16_V2F16_F32, VOP3_REGULAR, /*HasDPP*/ 1>,
382  AMDGPUfdot2, 1/*ExplicitClamp*/>;
383
384let OtherPredicates = [HasDot7Insts] in {
385defm V_DOT4_U32_U8  : VOP3PInst<"v_dot4_u32_u8",
386  VOP3P_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_udot4, 1>;
387defm V_DOT8_U32_U4  : VOP3PInst<"v_dot8_u32_u4",
388  VOP3P_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_udot8, 1>;
389} // End OtherPredicates = [HasDot7Insts]
390
391let OtherPredicates = [HasDot1Insts] in {
392defm V_DOT4_I32_I8  : VOP3PInst<"v_dot4_i32_i8",
393  VOP3P_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_sdot4, 1>;
394defm V_DOT8_I32_I4  : VOP3PInst<"v_dot8_i32_i4",
395  VOP3P_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_sdot8, 1>;
396} // End OtherPredicates = [HasDot1Insts]
397
398def DOT2_BF16_Profile
399  : VOP3P_Profile<VOP_F32_V2I16_V2I16_F32, VOP3_REGULAR, /*HasDPP*/ 1> {
400  let HasSrc1Mods = 1;
401}
402
403let SubtargetPredicate = HasDot9Insts  in {
404
405defm V_DOT2_F32_BF16 : VOP3PInst<"v_dot2_f32_bf16", DOT2_BF16_Profile,
406  int_amdgcn_fdot2_f32_bf16, 1>;
407
408} // End SubtargetPredicate = HasDot9Insts
409
410} // End let IsDOT = 1
411
412multiclass VOP3PDOTIUInst <string OpName, SDPatternOperator intrinsic_node> {
413  let IsDOT = 1 in
414  defm NAME : VOP3PInst<OpName, VOP3P_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>,
415                        null_frag, 1>;
416  // Dot-iu instructions consider input as signed if imod neg bits are set. Thus
417  // Dot-iu Intrinsics have extra operands and require separate codegen pattern.
418  def : GCNPat < (intrinsic_node (DotIUVOP3PMods i32:$src0_mods), i32:$src0,
419                                 (DotIUVOP3PMods i32:$src1_mods), i32:$src1,
420                                 i32:$src2, (i1 timm:$clamp)),
421                 (!cast<Instruction>(NAME) $src0_mods, i32:$src0,
422                                           $src1_mods, i32:$src1,
423                                           (i32 8), i32:$src2, i1:$clamp)
424  >;
425}
426
427let SubtargetPredicate = HasDot8Insts  in {
428defm V_DOT4_I32_IU8 : VOP3PDOTIUInst<"v_dot4_i32_iu8", int_amdgcn_sudot4>;
429defm V_DOT8_I32_IU4 : VOP3PDOTIUInst<"v_dot8_i32_iu4", int_amdgcn_sudot8>;
430
431def : GCNPat < (int_amdgcn_sdot8 i32:$src0,
432                                 i32:$src1,
433                                 i32:$src2, (i1 timm:$clamp)),
434               (V_DOT8_I32_IU4  (i32 9), i32:$src0,
435                                (i32 9), i32:$src1, (i32 8), i32:$src2, i1:$clamp)
436>;
437
438def : GCNPat < (int_amdgcn_sdot4 i32:$src0,
439                                 i32:$src1,
440                                 i32:$src2, (i1 timm:$clamp)),
441               (V_DOT4_I32_IU8  (i32 9), i32:$src0,
442                                (i32 9), i32:$src1, (i32 8), i32:$src2, i1:$clamp)
443>;
444} // End SubtargetPredicate = HasDot8Insts
445
446def : UDot2Pat<V_DOT2_U32_U16>;
447def : SDot2Pat<V_DOT2_I32_I16>;
448
449foreach Type = ["U", "I"] in
450  let Predicates = !cast<VOP_Pseudo>("V_DOT4_"#Type#"32_"#Type#8).Predicates in
451  def : GCNPat <
452    !cast<dag>(!foldl((i32 i32:$src2), [0, 1, 2, 3], lhs, y,
453                      (add_oneuse lhs, (!cast<PatFrag>("Mul"#Type#"_Elt"#y) i32:$src0, i32:$src1)))),
454    (!cast<VOP3P_Pseudo>("V_DOT4_"#Type#"32_"#Type#8) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
455
456foreach Type = ["U", "I"] in
457  let Predicates = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).Predicates in
458  def : GCNPat <
459    !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
460                      [1, 2, 3, 4, 5, 6, 7], lhs, y,
461                      (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
462    (!cast<VOP3P_Pseudo>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
463
464// Different variants of dot8 code-gen dag patterns are not generated through table-gen due to a huge increase
465// in the compile time. Directly handle the pattern generated by the FE here.
466foreach Type = ["U", "I"] in
467  let Predicates = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).Predicates in
468  def : GCNPat <
469    !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
470                      [7, 1, 2, 3, 4, 5, 6], lhs, y,
471                      (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
472    (!cast<VOP3P_Pseudo>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
473
474def ADst_32   : VOPDstOperand<AGPR_32>;
475def ADst_64   : VOPDstOperand<AReg_64>;
476def ADst_128  : VOPDstOperand<AReg_128>;
477def ADst_256  : VOPDstOperand<AReg_256>;
478def ADst_512  : VOPDstOperand<AReg_512>;
479def ADst_1024 : VOPDstOperand<AReg_1024>;
480def VDst_64   : VOPDstOperand<VReg_64>;
481def VDst_128  : VOPDstOperand<VReg_128>;
482def VDst_256  : VOPDstOperand<VReg_256>;
483def VDst_512  : VOPDstOperand<VReg_512>;
484def VDst_1024 : VOPDstOperand<VReg_1024>;
485
486def VOPProfileAccRead : VOP3P_Profile<VOP_I32_I32, VOP3_MAI> {
487  let Src0RC64 = ARegSrc_32;
488}
489
490def VOPProfileAccWrite : VOP3P_Profile<VOP_I32_I32, VOP3_MAI> {
491  let DstRC = ADst_32;
492  let Src0RC64 = VCSrc_b32;
493}
494
495class VOPProfileMAI<VOPProfile P, RegisterOperand _SrcRC, RegisterOperand _DstRC,
496                    RegisterOperand SrcABRC = AVSrc_32>
497  : VOP3P_Profile<P, VOP3_MAI> {
498  let DstRC = _DstRC;
499  let Src0RC64 = SrcABRC;
500  let Src1RC64 = SrcABRC;
501  let Src2RC64 = _SrcRC;
502  let HasOpSel = 0;
503  let HasClamp = 0;
504  let HasIntClamp = 0;
505  let HasOMod = 0;
506  let HasModifiers = 0;
507  let AsmVOP3Base = "$vdst, $src0, $src1, $src2$cbsz$abid$blgp";
508  let Ins64 = (ins Src0RC64:$src0, Src1RC64:$src1, Src2RC64:$src2, cbsz:$cbsz, abid:$abid, blgp:$blgp);
509  let InsVOP3Base = Ins64;
510  // Dst and SrcC cannot partially overlap if SrcC/Dst is bigger than 4 VGPRs.
511  // We then create two versions of the instruction: with tied dst and src2
512  // and with the earlyclobber flag on the dst. This is stricter than the
513  // actual HW restriction. In particular earlyclobber also affects src0 and
514  // src1 allocation which is not required.
515  bit NoDstOverlap = !gt(DstVT.Size, 128);
516}
517
518class VOPProfileSMFMAC<VOPProfile P, RegisterOperand _DstRC,
519                       RegisterOperand _SrcARC, RegisterOperand _SrcBRC>
520  : VOPProfileMAI<P, _DstRC, _DstRC, _SrcARC> {
521  let Src1RC64 = _SrcBRC;
522  let Src2VT = DstVT;
523  let Asm64 = " $vdst, $src0, $src1, $idx$cbsz$abid";
524  let Outs64 = (outs DstRC:$vdst);
525  let Ins64 = (ins Src0RC64:$src0, Src1RC64:$src1, VRegSrc_32:$idx, cbsz:$cbsz, abid:$abid, Src2RC64:$src2);
526}
527
528def VOPProfileMAI_F32_F32_X4    : VOPProfileMAI<VOP_V4F32_F32_F32_V4F32,       AISrc_128_f32,  ADst_128>;
529def VOPProfileMAI_F32_F32_X16   : VOPProfileMAI<VOP_V16F32_F32_F32_V16F32,     AISrc_512_f32,  ADst_512>;
530def VOPProfileMAI_F32_F32_X32   : VOPProfileMAI<VOP_V32F32_F32_F32_V32F32,     AISrc_1024_f32, ADst_1024>;
531def VOPProfileMAI_I32_I32_X4    : VOPProfileMAI<VOP_V4I32_I32_I32_V4I32,       AISrc_128_b32,  ADst_128>;
532def VOPProfileMAI_I32_I32_X16   : VOPProfileMAI<VOP_V16I32_I32_I32_V16I32,     AISrc_512_b32,  ADst_512>;
533def VOPProfileMAI_I32_I32_X32   : VOPProfileMAI<VOP_V32I32_I32_I32_V32I32,     AISrc_1024_b32, ADst_1024>;
534def VOPProfileMAI_F32_V2I16_X4  : VOPProfileMAI<VOP_V4F32_V2I16_V2I16_V4F32,   AISrc_128_b32,  ADst_128>;
535def VOPProfileMAI_F32_V2I16_X16 : VOPProfileMAI<VOP_V16F32_V2I16_V2I16_V16F32, AISrc_512_b32,  ADst_512>;
536def VOPProfileMAI_F32_V2I16_X32 : VOPProfileMAI<VOP_V32F32_V2I16_V2I16_V32F32, AISrc_1024_b32, ADst_1024>;
537def VOPProfileMAI_F32_V4F16_X4  : VOPProfileMAI<VOP_V4F32_V4F16_V4F16_V4F32,   AISrc_128_b32,  ADst_128,  AVSrc_64>;
538def VOPProfileMAI_F32_V4F16_X16 : VOPProfileMAI<VOP_V16F32_V4F16_V4F16_V16F32, AISrc_512_b32,  ADst_512,  AVSrc_64>;
539def VOPProfileMAI_F32_V4F16_X32 : VOPProfileMAI<VOP_V32F32_V4F16_V4F16_V32F32, AISrc_1024_b32, ADst_1024, AVSrc_64>;
540def VOPProfileMAI_F32_V4I16_X4  : VOPProfileMAI<VOP_V4F32_V4I16_V4I16_V4F32,   AISrc_128_b32,  ADst_128,  AVSrc_64>;
541def VOPProfileMAI_F32_V4I16_X16 : VOPProfileMAI<VOP_V16F32_V4I16_V4I16_V16F32, AISrc_512_b32,  ADst_512,  AVSrc_64>;
542def VOPProfileMAI_F32_V4I16_X32 : VOPProfileMAI<VOP_V32F32_V4I16_V4I16_V32F32, AISrc_1024_b32, ADst_1024, AVSrc_64>;
543def VOPProfileMAI_F64_16X16X4F64 : VOPProfileMAI<VOP_V4F64_F64_F64_V4F64,      AISrc_256_f64,  ADst_256,  AVSrc_64>;
544def VOPProfileMAI_F64_4X4X4F64   : VOPProfileMAI<VOP_F64_F64_F64_F64,          AISrc_64_f64,   ADst_64,   AVSrc_64>;
545def VOPProfileMAI_I32_I64_X16   : VOPProfileMAI<VOP_V4I32_I64_I64_V4I32,       AISrc_128_b32,  ADst_128,  AVSrc_64>;
546def VOPProfileMAI_I32_I64_X32   : VOPProfileMAI<VOP_V16I32_I64_I64_V16I32,     AISrc_512_b32,  ADst_512,  AVSrc_64>;
547def VOPProfileMAI_F32_V2F32_X16 : VOPProfileMAI<VOP_V4F32_V2F32_V2F32_V4F32,   AISrc_128_b32,  ADst_128,  AVSrc_64>;
548def VOPProfileMAI_F32_V2F32_X32 : VOPProfileMAI<VOP_V16F32_V2F32_V2F32_V16F32, AISrc_512_b32,  ADst_512,  AVSrc_64>;
549def VOPProfileMAI_F32_I64_X32   : VOPProfileMAI<VOP_V4F32_I64_I64_V4F32,       AISrc_128_b32,  ADst_128,  AVSrc_64>;
550def VOPProfileMAI_F32_I64_X16   : VOPProfileMAI<VOP_V16F32_I64_I64_V16F32,     AISrc_512_b32,  ADst_512,  AVSrc_64>;
551
552def VOPProfileMAI_F32_F32_X4_VCD     : VOPProfileMAI<VOP_V4F32_F32_F32_V4F32,       VISrc_128_f32,  VDst_128>;
553def VOPProfileMAI_F32_F32_X16_VCD    : VOPProfileMAI<VOP_V16F32_F32_F32_V16F32,     VISrc_512_f32,  VDst_512>;
554def VOPProfileMAI_F32_F32_X32_VCD    : VOPProfileMAI<VOP_V32F32_F32_F32_V32F32,     VISrc_1024_f32, VDst_1024>;
555def VOPProfileMAI_I32_I32_X4_VCD     : VOPProfileMAI<VOP_V4I32_I32_I32_V4I32,       VISrc_128_b32,  VDst_128>;
556def VOPProfileMAI_I32_I32_X16_VCD    : VOPProfileMAI<VOP_V16I32_I32_I32_V16I32,     VISrc_512_b32,  VDst_512>;
557def VOPProfileMAI_I32_I32_X32_VCD    : VOPProfileMAI<VOP_V32I32_I32_I32_V32I32,     VISrc_1024_b32, VDst_1024>;
558def VOPProfileMAI_F32_V2I16_X4_VCD   : VOPProfileMAI<VOP_V4F32_V2I16_V2I16_V4F32,   VISrc_128_b32,  VDst_128>;
559def VOPProfileMAI_F32_V2I16_X16_VCD  : VOPProfileMAI<VOP_V16F32_V2I16_V2I16_V16F32, VISrc_512_b32,  VDst_512>;
560def VOPProfileMAI_F32_V2I16_X32_VCD  : VOPProfileMAI<VOP_V32F32_V2I16_V2I16_V32F32, VISrc_1024_b32, VDst_1024>;
561def VOPProfileMAI_F32_V4F16_X4_VCD   : VOPProfileMAI<VOP_V4F32_V4F16_V4F16_V4F32,   VISrc_128_b32,  VDst_128,  AVSrc_64>;
562def VOPProfileMAI_F32_V4F16_X16_VCD  : VOPProfileMAI<VOP_V16F32_V4F16_V4F16_V16F32, VISrc_512_b32,  VDst_512,  AVSrc_64>;
563def VOPProfileMAI_F32_V4F16_X32_VCD  : VOPProfileMAI<VOP_V32F32_V4F16_V4F16_V32F32, VISrc_1024_b32, VDst_1024, AVSrc_64>;
564def VOPProfileMAI_F32_V4I16_X4_VCD   : VOPProfileMAI<VOP_V4F32_V4I16_V4I16_V4F32,   VISrc_128_b32,  VDst_128,  AVSrc_64>;
565def VOPProfileMAI_F32_V4I16_X16_VCD  : VOPProfileMAI<VOP_V16F32_V4I16_V4I16_V16F32, VISrc_512_b32,  VDst_512,  AVSrc_64>;
566def VOPProfileMAI_F32_V4I16_X32_VCD  : VOPProfileMAI<VOP_V32F32_V4I16_V4I16_V32F32, VISrc_1024_b32, VDst_1024, AVSrc_64>;
567def VOPProfileMAI_F64_16X16X4F64_VCD : VOPProfileMAI<VOP_V4F64_F64_F64_V4F64,       VISrc_256_f64,  VDst_256,  AVSrc_64>;
568def VOPProfileMAI_F64_4X4X4F64_VCD   : VOPProfileMAI<VOP_F64_F64_F64_F64,           VISrc_64_f64,   VDst_64,   AVSrc_64>;
569def VOPProfileMAI_I32_I64_X16_VCD    : VOPProfileMAI<VOP_V4I32_I64_I64_V4I32,       VISrc_128_b32,  VDst_128,  AVSrc_64>;
570def VOPProfileMAI_I32_I64_X32_VCD    : VOPProfileMAI<VOP_V16I32_I64_I64_V16I32,     VISrc_512_b32,  VDst_512,  AVSrc_64>;
571def VOPProfileMAI_F32_V2F32_X16_VCD  : VOPProfileMAI<VOP_V4F32_V2F32_V2F32_V4F32,   VISrc_128_b32,  VDst_128,  AVSrc_64>;
572def VOPProfileMAI_F32_V2F32_X32_VCD  : VOPProfileMAI<VOP_V16F32_V2F32_V2F32_V16F32, VISrc_512_b32,  VDst_512,  AVSrc_64>;
573def VOPProfileMAI_F32_I64_X32_VCD    : VOPProfileMAI<VOP_V4F32_I64_I64_V4F32,       VISrc_128_b32,  VDst_128,  AVSrc_64>;
574def VOPProfileMAI_F32_I64_X16_VCD    : VOPProfileMAI<VOP_V16F32_I64_I64_V16F32,     VISrc_512_b32,  VDst_512,  AVSrc_64>;
575
576def VOPProfileSMFMAC_F32_16X16X32_F16 : VOPProfileSMFMAC<VOP_V4F32_V4F16_V8F16_I32,  AVDst_128, AVSrc_64, AVSrc_128>;
577def VOPProfileSMFMAC_F32_32X32X16_F16 : VOPProfileSMFMAC<VOP_V16F32_V4F16_V8F16_I32, AVDst_512, AVSrc_64, AVSrc_128>;
578def VOPProfileSMFMAC_F32_16X16X32_I16 : VOPProfileSMFMAC<VOP_V4F32_V4I16_V8I16_I32,  AVDst_128, AVSrc_64, AVSrc_128>;
579def VOPProfileSMFMAC_F32_32X32X16_I16 : VOPProfileSMFMAC<VOP_V16F32_V4I16_V8I16_I32, AVDst_512, AVSrc_64, AVSrc_128>;
580def VOPProfileSMFMAC_I32_16X16X64_I8  : VOPProfileSMFMAC<VOP_V4I32_V2I32_V4I32_I32,  AVDst_128, AVSrc_64, AVSrc_128>;
581def VOPProfileSMFMAC_I32_32X32X32_I8  : VOPProfileSMFMAC<VOP_V16I32_V2I32_V4I32_I32, AVDst_512, AVSrc_64, AVSrc_128>;
582def VOPProfileSMFMAC_F32_16X16X64_F8  : VOPProfileSMFMAC<VOP_V4F32_V2I32_V4I32_I32,  AVDst_128, AVSrc_64, AVSrc_128>;
583def VOPProfileSMFMAC_F32_32X32X32_F8  : VOPProfileSMFMAC<VOP_V16F32_V2I32_V4I32_I32, AVDst_512, AVSrc_64, AVSrc_128>;
584
585class MFMATable <bit is_mac, string Name> {
586  bit IsMac = is_mac;
587  string FMAOp = Name;
588}
589
590class MAIFrag<SDPatternOperator Op, code pred> : PatFrag <
591  (ops node:$src0, node:$src1, node:$src2, node:$cbsz, node:$abid, node:$blgp),
592  (Op $src0, $src1, $src2, $cbsz, $abid, $blgp),
593  pred
594>;
595
596let GISelPredicateCode = [{ return MF.getInfo<SIMachineFunctionInfo>()->mayNeedAGPRs(); }] in
597class AgprMAIFrag<SDPatternOperator Op> :
598  MAIFrag<Op, [{ return MF->getInfo<SIMachineFunctionInfo>()->mayNeedAGPRs(); }]>;
599
600let GISelPredicateCode = [{ return !MF.getInfo<SIMachineFunctionInfo>()->mayNeedAGPRs(); }] in
601class VgprMAIFrag<SDPatternOperator Op> :
602  MAIFrag<Op, [{ return !MF->getInfo<SIMachineFunctionInfo>()->mayNeedAGPRs(); }]>;
603
604let SubtargetPredicate = HasMAIInsts in {
605
606let isAsCheapAsAMove = 1, isReMaterializable = 1 in {
607  defm V_ACCVGPR_READ_B32  : VOP3Inst<"v_accvgpr_read_b32",  VOPProfileAccRead>;
608  let isMoveImm = 1 in {
609    defm V_ACCVGPR_WRITE_B32 : VOP3Inst<"v_accvgpr_write_b32", VOPProfileAccWrite>;
610  } // End isMoveImm = 1
611} // End isAsCheapAsAMove = 1, isReMaterializable = 1
612
613class MAIInst<string OpName, VOPProfile P, SDPatternOperator node>
614  : VOP3InstBase<OpName, P, node> {
615  Instruction Opcode = !cast<Instruction>(NAME);
616  bit is_dgemm = 0;
617  bit is_gfx940_xdl = 0;
618}
619
620multiclass MAIInst<string OpName, string P, SDPatternOperator node,
621                   bit NoDstOverlap = !cast<VOPProfileMAI>("VOPProfileMAI_" # P).NoDstOverlap> {
622  let isConvergent = 1, mayRaiseFPException = 0, ReadsModeReg = 1 in {
623    // FP32 denorm mode is respected, rounding mode is not. Exceptions are not supported.
624    let Constraints = !if(NoDstOverlap, "@earlyclobber $vdst", "") in {
625      def _e64 : MAIInst<OpName, !cast<VOPProfileMAI>("VOPProfileMAI_" # P),
626                         !if(!or(NoDstOverlap, !eq(node, null_frag)), null_frag, AgprMAIFrag<node>)>,
627                 MFMATable<0, NAME # "_e64">;
628
629      let SubtargetPredicate = isGFX90APlus, Mnemonic = OpName in
630      def _vgprcd_e64 : MAIInst<OpName # "_vgprcd", !cast<VOPProfileMAI>("VOPProfileMAI_" # P # "_VCD"),
631                                !if(!or(NoDstOverlap, !eq(node, null_frag)), null_frag, VgprMAIFrag<node>)>,
632                        MFMATable<0, NAME # "_vgprcd_e64">;
633    }
634
635    if NoDstOverlap then {
636      let Constraints = !if(NoDstOverlap, "$vdst = $src2", ""),
637          isConvertibleToThreeAddress = NoDstOverlap,
638          Mnemonic = OpName in {
639        def "_mac_e64" : MAIInst<OpName # "_mac", !cast<VOPProfileMAI>("VOPProfileMAI_" # P),
640                                 !if(!eq(node, null_frag), null_frag, AgprMAIFrag<node>)>,
641                         MFMATable<1, NAME # "_e64">;
642
643        let SubtargetPredicate = isGFX90APlus in
644        def _mac_vgprcd_e64 : MAIInst<OpName # "_mac_vgprcd", !cast<VOPProfileMAI>("VOPProfileMAI_" # P # "_VCD"),
645                                      !if(!eq(node, null_frag), null_frag, VgprMAIFrag<node>)>,
646                              MFMATable<1, NAME # "_vgprcd_e64">;
647      }
648    }
649  } // End isConvergent = 1, mayRaiseFPException = 0, ReadsModeReg = 1
650}
651
652defm V_MFMA_F32_4X4X1F32    : MAIInst<"v_mfma_f32_4x4x1f32",    "F32_F32_X4",    int_amdgcn_mfma_f32_4x4x1f32>;
653defm V_MFMA_F32_16X16X1F32  : MAIInst<"v_mfma_f32_16x16x1f32",  "F32_F32_X16",   int_amdgcn_mfma_f32_16x16x1f32>;
654defm V_MFMA_F32_16X16X4F32  : MAIInst<"v_mfma_f32_16x16x4f32",  "F32_F32_X4",    int_amdgcn_mfma_f32_16x16x4f32>;
655defm V_MFMA_F32_32X32X1F32  : MAIInst<"v_mfma_f32_32x32x1f32",  "F32_F32_X32",   int_amdgcn_mfma_f32_32x32x1f32>;
656defm V_MFMA_F32_32X32X2F32  : MAIInst<"v_mfma_f32_32x32x2f32",  "F32_F32_X16",   int_amdgcn_mfma_f32_32x32x2f32>;
657
658let is_gfx940_xdl = 1 in {
659defm V_MFMA_F32_4X4X4F16    : MAIInst<"v_mfma_f32_4x4x4f16",    "F32_V4F16_X4",  int_amdgcn_mfma_f32_4x4x4f16>;
660defm V_MFMA_I32_4X4X4I8     : MAIInst<"v_mfma_i32_4x4x4i8",     "I32_I32_X4",    int_amdgcn_mfma_i32_4x4x4i8>;
661defm V_MFMA_F32_16X16X4F16  : MAIInst<"v_mfma_f32_16x16x4f16",  "F32_V4F16_X16", int_amdgcn_mfma_f32_16x16x4f16>;
662defm V_MFMA_F32_16X16X16F16 : MAIInst<"v_mfma_f32_16x16x16f16", "F32_V4F16_X4",  int_amdgcn_mfma_f32_16x16x16f16>;
663defm V_MFMA_I32_16X16X4I8   : MAIInst<"v_mfma_i32_16x16x4i8",   "I32_I32_X16",   int_amdgcn_mfma_i32_16x16x4i8>;
664defm V_MFMA_F32_32X32X4F16  : MAIInst<"v_mfma_f32_32x32x4f16",  "F32_V4F16_X32", int_amdgcn_mfma_f32_32x32x4f16>;
665defm V_MFMA_F32_32X32X8F16  : MAIInst<"v_mfma_f32_32x32x8f16",  "F32_V4F16_X16", int_amdgcn_mfma_f32_32x32x8f16>;
666defm V_MFMA_I32_32X32X4I8   : MAIInst<"v_mfma_i32_32x32x4i8",   "I32_I32_X32",   int_amdgcn_mfma_i32_32x32x4i8>;
667}
668
669let Predicates = [isGFX908orGFX90A] in {
670defm V_MFMA_I32_16X16X16I8  : MAIInst<"v_mfma_i32_16x16x16i8",  "I32_I32_X4",    int_amdgcn_mfma_i32_16x16x16i8>;
671defm V_MFMA_I32_32X32X8I8   : MAIInst<"v_mfma_i32_32x32x8i8",   "I32_I32_X16",   int_amdgcn_mfma_i32_32x32x8i8>;
672defm V_MFMA_F32_4X4X2BF16   : MAIInst<"v_mfma_f32_4x4x2bf16",   "F32_V2I16_X4",  int_amdgcn_mfma_f32_4x4x2bf16>;
673defm V_MFMA_F32_16X16X2BF16 : MAIInst<"v_mfma_f32_16x16x2bf16", "F32_V2I16_X16", int_amdgcn_mfma_f32_16x16x2bf16>;
674defm V_MFMA_F32_16X16X8BF16 : MAIInst<"v_mfma_f32_16x16x8bf16", "F32_V2I16_X4",  int_amdgcn_mfma_f32_16x16x8bf16>;
675defm V_MFMA_F32_32X32X2BF16 : MAIInst<"v_mfma_f32_32x32x2bf16", "F32_V2I16_X32", int_amdgcn_mfma_f32_32x32x2bf16>;
676defm V_MFMA_F32_32X32X4BF16 : MAIInst<"v_mfma_f32_32x32x4bf16", "F32_V2I16_X16", int_amdgcn_mfma_f32_32x32x4bf16>;
677}
678
679} // End SubtargetPredicate = HasMAIInsts
680
681let Predicates = [isGFX90APlus] in {
682  let is_gfx940_xdl = 1 in {
683  defm V_MFMA_F32_32X32X4BF16_1K  : MAIInst<"v_mfma_f32_32x32x4bf16_1k",  "F32_V4I16_X32",  int_amdgcn_mfma_f32_32x32x4bf16_1k>;
684  defm V_MFMA_F32_16X16X4BF16_1K  : MAIInst<"v_mfma_f32_16x16x4bf16_1k",  "F32_V4I16_X16",  int_amdgcn_mfma_f32_16x16x4bf16_1k>;
685  defm V_MFMA_F32_4X4X4BF16_1K    : MAIInst<"v_mfma_f32_4x4x4bf16_1k",    "F32_V4I16_X4",   int_amdgcn_mfma_f32_4x4x4bf16_1k>;
686  defm V_MFMA_F32_32X32X8BF16_1K  : MAIInst<"v_mfma_f32_32x32x8bf16_1k",  "F32_V4I16_X16",  int_amdgcn_mfma_f32_32x32x8bf16_1k>;
687  defm V_MFMA_F32_16X16X16BF16_1K : MAIInst<"v_mfma_f32_16x16x16bf16_1k", "F32_V4I16_X4",   int_amdgcn_mfma_f32_16x16x16bf16_1k>;
688  }
689
690  let is_dgemm = 1 in {
691  defm V_MFMA_F64_16X16X4F64      : MAIInst<"v_mfma_f64_16x16x4f64",      "F64_16X16X4F64", int_amdgcn_mfma_f64_16x16x4f64>;
692  defm V_MFMA_F64_4X4X4F64        : MAIInst<"v_mfma_f64_4x4x4f64",        "F64_4X4X4F64",   int_amdgcn_mfma_f64_4x4x4f64>;
693  }
694} // End Predicates = [isGFX90APlus]
695
696let SubtargetPredicate = isGFX940Plus, is_gfx940_xdl = 1 in {
697  defm V_MFMA_I32_32X32X16I8       : MAIInst<"v_mfma_i32_32x32x16i8",       "I32_I64_X32",    int_amdgcn_mfma_i32_32x32x16_i8>;
698  defm V_MFMA_I32_16X16X32I8       : MAIInst<"v_mfma_i32_16x16x32i8",       "I32_I64_X16",    int_amdgcn_mfma_i32_16x16x32_i8>;
699  defm V_MFMA_F32_16X16X8XF32      : MAIInst<"v_mfma_f32_16x16x8xf32",      "F32_V2F32_X16",  int_amdgcn_mfma_f32_16x16x8_xf32>;
700  defm V_MFMA_F32_32X32X4XF32      : MAIInst<"v_mfma_f32_32x32x4xf32",      "F32_V2F32_X32",  int_amdgcn_mfma_f32_32x32x4_xf32>;
701  defm V_MFMA_F32_16X16X32_BF8_BF8 : MAIInst<"v_mfma_f32_16x16x32_bf8_bf8", "F32_I64_X32",    int_amdgcn_mfma_f32_16x16x32_bf8_bf8>;
702  defm V_MFMA_F32_16X16X32_BF8_FP8 : MAIInst<"v_mfma_f32_16x16x32_bf8_fp8", "F32_I64_X32",    int_amdgcn_mfma_f32_16x16x32_bf8_fp8>;
703  defm V_MFMA_F32_16X16X32_FP8_BF8 : MAIInst<"v_mfma_f32_16x16x32_fp8_bf8", "F32_I64_X32",    int_amdgcn_mfma_f32_16x16x32_fp8_bf8>;
704  defm V_MFMA_F32_16X16X32_FP8_FP8 : MAIInst<"v_mfma_f32_16x16x32_fp8_fp8", "F32_I64_X32",    int_amdgcn_mfma_f32_16x16x32_fp8_fp8>;
705  defm V_MFMA_F32_32X32X16_BF8_BF8 : MAIInst<"v_mfma_f32_32x32x16_bf8_bf8", "F32_I64_X16",    int_amdgcn_mfma_f32_32x32x16_bf8_bf8>;
706  defm V_MFMA_F32_32X32X16_BF8_FP8 : MAIInst<"v_mfma_f32_32x32x16_bf8_fp8", "F32_I64_X16",    int_amdgcn_mfma_f32_32x32x16_bf8_fp8>;
707  defm V_MFMA_F32_32X32X16_FP8_BF8 : MAIInst<"v_mfma_f32_32x32x16_fp8_bf8", "F32_I64_X16",    int_amdgcn_mfma_f32_32x32x16_fp8_bf8>;
708  defm V_MFMA_F32_32X32X16_FP8_FP8 : MAIInst<"v_mfma_f32_32x32x16_fp8_fp8", "F32_I64_X16",    int_amdgcn_mfma_f32_32x32x16_fp8_fp8>;
709} // End SubtargetPredicate = isGFX940Plus, is_gfx940_xdl = 1
710
711multiclass SMFMACInst<string OpName, string P, SDPatternOperator node> {
712  let Constraints = "$vdst = $src2", DisableEncoding = "$src2",
713      isConvergent = 1, mayRaiseFPException = 0, ReadsModeReg = 1, is_gfx940_xdl = 1 in {
714    def _e64 : MAIInst<OpName, !cast<VOPProfileSMFMAC>("VOPProfileSMFMAC_" # P), node>;
715  }
716}
717
718let SubtargetPredicate = isGFX940Plus in {
719defm V_SMFMAC_F32_16X16X32_F16     : SMFMACInst<"v_smfmac_f32_16x16x32_f16",     "F32_16X16X32_F16", int_amdgcn_smfmac_f32_16x16x32_f16>;
720defm V_SMFMAC_F32_32X32X16_F16     : SMFMACInst<"v_smfmac_f32_32x32x16_f16",     "F32_32X32X16_F16", int_amdgcn_smfmac_f32_32x32x16_f16>;
721defm V_SMFMAC_F32_16X16X32_BF16    : SMFMACInst<"v_smfmac_f32_16x16x32_bf16",    "F32_16X16X32_I16", int_amdgcn_smfmac_f32_16x16x32_bf16>;
722defm V_SMFMAC_F32_32X32X16_BF16    : SMFMACInst<"v_smfmac_f32_32x32x16_bf16",    "F32_32X32X16_I16", int_amdgcn_smfmac_f32_32x32x16_bf16>;
723defm V_SMFMAC_I32_16X16X64_I8      : SMFMACInst<"v_smfmac_i32_16x16x64_i8",      "I32_16X16X64_I8",  int_amdgcn_smfmac_i32_16x16x64_i8>;
724defm V_SMFMAC_I32_32X32X32_I8      : SMFMACInst<"v_smfmac_i32_32x32x32_i8",      "I32_32X32X32_I8",  int_amdgcn_smfmac_i32_32x32x32_i8>;
725defm V_SMFMAC_F32_16X16X64_BF8_BF8 : SMFMACInst<"v_smfmac_f32_16x16x64_bf8_bf8", "F32_16X16X64_F8",  int_amdgcn_smfmac_f32_16x16x64_bf8_bf8>;
726defm V_SMFMAC_F32_16X16X64_BF8_FP8 : SMFMACInst<"v_smfmac_f32_16x16x64_bf8_fp8", "F32_16X16X64_F8",  int_amdgcn_smfmac_f32_16x16x64_bf8_fp8>;
727defm V_SMFMAC_F32_16X16X64_FP8_BF8 : SMFMACInst<"v_smfmac_f32_16x16x64_fp8_bf8", "F32_16X16X64_F8",  int_amdgcn_smfmac_f32_16x16x64_fp8_bf8>;
728defm V_SMFMAC_F32_16X16X64_FP8_FP8 : SMFMACInst<"v_smfmac_f32_16x16x64_fp8_fp8", "F32_16X16X64_F8",  int_amdgcn_smfmac_f32_16x16x64_fp8_fp8>;
729defm V_SMFMAC_F32_32X32X32_BF8_BF8 : SMFMACInst<"v_smfmac_f32_32x32x32_bf8_bf8", "F32_32X32X32_F8",  int_amdgcn_smfmac_f32_32x32x32_bf8_bf8>;
730defm V_SMFMAC_F32_32X32X32_BF8_FP8 : SMFMACInst<"v_smfmac_f32_32x32x32_bf8_fp8", "F32_32X32X32_F8",  int_amdgcn_smfmac_f32_32x32x32_bf8_fp8>;
731defm V_SMFMAC_F32_32X32X32_FP8_BF8 : SMFMACInst<"v_smfmac_f32_32x32x32_fp8_bf8", "F32_32X32X32_F8",  int_amdgcn_smfmac_f32_32x32x32_fp8_bf8>;
732defm V_SMFMAC_F32_32X32X32_FP8_FP8 : SMFMACInst<"v_smfmac_f32_32x32x32_fp8_fp8", "F32_32X32X32_F8",  int_amdgcn_smfmac_f32_32x32x32_fp8_fp8>;
733}
734
735def MAIInstInfoTable : GenericTable {
736  let FilterClass = "MAIInst";
737  let CppTypeName = "MAIInstInfo";
738  let Fields = [
739    "Opcode", "is_dgemm", "is_gfx940_xdl"
740  ];
741
742  let PrimaryKey = ["Opcode"];
743  let PrimaryKeyName = "getMAIInstInfoHelper";
744}
745
746let isCommutable = 1, isReMaterializable = 1 in {
747  let SubtargetPredicate = HasPackedFP32Ops in {
748    defm V_PK_FMA_F32 : VOP3PInst<"v_pk_fma_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fma>;
749    defm V_PK_MUL_F32 : VOP3PInst<"v_pk_mul_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fmul>;
750    defm V_PK_ADD_F32 : VOP3PInst<"v_pk_add_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fadd>;
751  } // End SubtargetPredicate = HasPackedFP32Ops
752
753  let SubtargetPredicate = HasPkMovB32 in
754  defm V_PK_MOV_B32 : VOP3PInst<"v_pk_mov_b32", VOP3P_Profile<VOP_V2I32_V2I32_V2I32, VOP3_PACKED>>;
755} // End isCommutable = 1, isReMaterializable = 1
756
757def : MnemonicAlias<"v_accvgpr_read",  "v_accvgpr_read_b32">;
758def : MnemonicAlias<"v_accvgpr_write", "v_accvgpr_write_b32">;
759
760class VOPProfileWMMA<VOPProfile P, string Suffix, RegisterOperand _Src01RC64, bit _HasClamp, bit _HasOpSel> : VOP3P_Profile<P> {
761  let DstRC = !if(!eq(Suffix, "_w32"), VDst_256, VDst_128);
762  let Src0RC64 = _Src01RC64;
763  let Src1RC64 = _Src01RC64;
764  let Src2RC64 = !if(!eq(Suffix, "_w32"), VISrc_256_f64, VISrc_128_f32);
765  let HasClamp = _HasClamp;
766  let HasOpSel = _HasOpSel;
767  let IsPacked = 1;
768  let IsWMMA = 1;
769}
770
771def VOP_V8F32_V16F16_V16F16_V8F32 : VOPProfile <[v8f32, v16f16, v16f16, v8f32]>;
772def VOP_V8F32_V16I16_V16I16_V8F32 : VOPProfile <[v8f32, v16i16, v16i16, v8f32]>;
773def VOP_V16F16_V16F16_V16F16_V16F16 : VOPProfile <[v16f16, v16f16, v16f16, v16f16]>;
774def VOP_V16I16_V16I16_V16I16_V16I16 : VOPProfile <[v16i16, v16i16, v16i16, v16i16]>;
775def VOP_V8I32_V4I32_V4I32_V8I32 : VOPProfile <[v8i32, v4i32, v4i32, v8i32]>;
776def VOP_V8I32_V2I32_V2I32_V8I32 : VOPProfile <[v8i32, v2i32, v2i32, v8i32]>;
777
778def VOP_V4F32_V16F16_V16F16_V4F32 : VOPProfile <[v4f32, v16f16, v16f16, v4f32]>;
779def VOP_V4F32_V16I16_V16I16_V4F32 : VOPProfile <[v4f32, v16i16, v16i16, v4f32]>;
780def VOP_V8F16_V16F16_V16F16_V8F16 : VOPProfile <[v8f16, v16f16, v16f16, v8f16]>;
781def VOP_V8I16_V16I16_V16I16_V8I16 : VOPProfile <[v8i16, v16i16, v16i16, v8i16]>;
782def VOP_V4I32_V4I32_V4I32_V4I32 : VOPProfile <[v4i32, v4i32, v4i32, v4i32]>;
783def VOP_V4I32_V2I32_V2I32_V4I32 : VOPProfile <[v4i32, v2i32, v2i32, v4i32]>;
784
785
786class WMMAType <bits<2> val> {
787  bit hasClamp = val{0};
788  bit hasOpsel = val{1};
789}
790
791def WMMARegular      : WMMAType<0b00>;
792def WMMAUIClamp      : WMMAType<0b01>;
793def WMMAOpSel        : WMMAType<0b10>;
794
795class WMMARegularPat<Instruction Inst, SDPatternOperator node, VOPProfile P> :
796  GCNPat < (P.DstVT (node
797                                (P.Src0VT (VOP3PMods P.Src0VT:$src0, i32:$src0_modifiers)),
798                                (P.Src1VT (VOP3PMods P.Src1VT:$src1, i32:$src1_modifiers)),
799                                (P.Src2VT (VOP3PMods P.Src2VT:$src2, i32:$src2_modifiers))
800                   )),
801                   (P.DstVT (Inst i32:$src0_modifiers, P.Src0VT:$src0, i32:$src1_modifiers, P.Src1VT:$src1, $src2_modifiers, P.Src2VT:$src2))
802>;
803
804class WMMAOpSelPat<Instruction Inst, SDPatternOperator node, VOPProfile P> :
805  GCNPat < (P.DstVT (node
806                                (P.Src0VT P.Src0VT:$src0),
807                                (P.Src1VT P.Src1VT:$src1),
808                                (P.Src2VT P.Src2VT:$src2), (WMMAOpSelVOP3PMods i32:$src2_modifiers)
809                   )),
810                   (P.DstVT (Inst (i32 8), P.Src0VT:$src0, (i32 8), P.Src1VT:$src1, i32:$src2_modifiers, P.Src2VT:$src2))
811>;
812
813class WMMAUIClampPat<Instruction Inst, SDPatternOperator node, VOPProfile P> :
814  GCNPat < (P.DstVT (node
815                                (DotIUVOP3PMods i32:$src0_modifiers), (P.Src0VT P.Src0VT:$src0),
816                                (DotIUVOP3PMods i32:$src1_modifiers), (P.Src1VT P.Src1VT:$src1),
817                                (P.Src2VT P.Src2VT:$src2), (i1 timm:$clamp)
818                   )),
819                   (P.DstVT (Inst i32:$src0_modifiers, P.Src0VT:$src0, i32:$src1_modifiers, P.Src1VT:$src1, (i32 8), P.Src2VT:$src2, i1:$clamp))
820>;
821
822class WMMAOpcodeMapping<Instruction TwoAddr, Instruction ThreeAddr> {
823  Instruction Opcode2Addr = TwoAddr;
824  Instruction Opcode3Addr = ThreeAddr;
825  Predicate WaveSizePredicate;
826}
827
828def WMMAOpcode : GenericEnum {
829  let FilterClass = "VOP3P_Pseudo";
830}
831
832class WMMAMappingTable : GenericTable {
833  let FilterClass = "WMMAOpcodeMapping";
834  let CppTypeName = "WMMAOpcodeMappingInfo";
835  let Fields = ["Opcode2Addr", "Opcode3Addr"];
836  string TypeOf_Opcode2Addr = "WMMAOpcode";
837  string TypeOf_Opcode3Addr = "WMMAOpcode";
838}
839
840def WMMAOpcode2AddrMappingTable : WMMAMappingTable {
841  let PrimaryKey = ["Opcode2Addr"];
842  let PrimaryKeyName = "getWMMAMappingInfoFrom2AddrOpcode";
843}
844
845def WMMAOpcode3AddrMappingTable : WMMAMappingTable {
846  let PrimaryKey = ["Opcode3Addr"];
847  let PrimaryKeyName = "getWMMAMappingInfoFrom3AddrOpcode";
848}
849
850// The WMMA instruction has extra constraints:
851// Matrices A and B cannot overlap with D. C cannot partially overlap with D,
852// but it is OK for them to be the same (which is a typical case).
853//
854// We implement it as follows:
855// 1) Map the intrinsic to the pseudo where D is tied to C ($vdst = $src2).
856// 2) The pass twoaddressinstruction checks if src2 is live and if that is the case
857//    it converts the default pseudo to the pseudo where src2 is not the same as vdst.
858// 3) @earlyclobber on the destination satisfies the constraint during RA.
859
860multiclass WMMAInst<string Suffix, string Instr, VOPProfile P, SDPatternOperator node = null_frag, RegisterOperand _Src01RC64 = VRegSrc_256, WMMAType Type, bit convertibleTo3Addr> {
861
862  defvar WMMAConstraints2Addr = "@earlyclobber $vdst,$vdst = $src2";
863  defvar WMMAConstraints3Addr = "@earlyclobber $vdst";
864
865  defvar WMMAProfile = VOPProfileWMMA<P, Suffix, _Src01RC64, Type.hasClamp, Type.hasOpsel>;
866  let Mnemonic = Instr, mayRaiseFPException = 0, ReadsModeReg = 0 in {
867    let Constraints = WMMAConstraints2Addr, isConvertibleToThreeAddress = convertibleTo3Addr in {
868      def _twoaddr # Suffix : VOP3P_Pseudo<Instr # Suffix, WMMAProfile>;
869    }
870  }
871  if convertibleTo3Addr then {
872    let Mnemonic = Instr, mayRaiseFPException = 0, ReadsModeReg = 0 in {
873      let Constraints = WMMAConstraints3Addr, SchedRW = [Write32Bit, Write32Bit] in {
874        def _threeaddr # Suffix : VOP3P_Pseudo<Instr # Suffix, WMMAProfile>;
875      }
876    }
877    def : WMMAOpcodeMapping<!cast<Instruction>(NAME # _twoaddr # Suffix),
878                          !cast<Instruction>(NAME # _threeaddr # Suffix)>;
879  }
880
881  if !eq(Type, WMMAOpSel) then {
882    def : WMMAOpSelPat<!cast<Instruction>(NAME # _twoaddr # Suffix), node, P>;
883  } else if !eq(Type, WMMAUIClamp) then {
884    def : WMMAUIClampPat<!cast<Instruction>(NAME # _twoaddr # Suffix), node, P>;
885  } else {
886    def : WMMARegularPat<!cast<Instruction>(NAME # _twoaddr # Suffix), node, P>;
887  }
888}
889
890
891let WaveSizePredicate = isWave32 in {
892  defm V_WMMA_F32_16X16X16_F16   : WMMAInst<"_w32", "v_wmma_f32_16x16x16_f16",  VOP_V8F32_V16F16_V16F16_V8F32, int_amdgcn_wmma_f32_16x16x16_f16, VRegSrc_256, WMMARegular, 1>;
893  defm V_WMMA_F32_16X16X16_BF16  : WMMAInst<"_w32", "v_wmma_f32_16x16x16_bf16", VOP_V8F32_V16I16_V16I16_V8F32, int_amdgcn_wmma_f32_16x16x16_bf16, VRegSrc_256, WMMARegular, 1>;
894  defm V_WMMA_F16_16X16X16_F16   : WMMAInst<"_w32", "v_wmma_f16_16x16x16_f16",   VOP_V16F16_V16F16_V16F16_V16F16, int_amdgcn_wmma_f16_16x16x16_f16, VRegSrc_256, WMMAOpSel, 1>;
895  defm V_WMMA_BF16_16X16X16_BF16 : WMMAInst<"_w32", "v_wmma_bf16_16x16x16_bf16", VOP_V16I16_V16I16_V16I16_V16I16, int_amdgcn_wmma_bf16_16x16x16_bf16, VRegSrc_256, WMMAOpSel, 1>;
896  defm V_WMMA_F16_16X16X16_F16_TIED   : WMMAInst<"_w32", "v_wmma_f16_16x16x16_f16",   VOP_V16F16_V16F16_V16F16_V16F16, int_amdgcn_wmma_f16_16x16x16_f16_tied, VRegSrc_256, WMMAOpSel, 0>;
897  defm V_WMMA_BF16_16X16X16_BF16_TIED : WMMAInst<"_w32", "v_wmma_bf16_16x16x16_bf16", VOP_V16I16_V16I16_V16I16_V16I16, int_amdgcn_wmma_bf16_16x16x16_bf16_tied, VRegSrc_256, WMMAOpSel, 0>;
898  defm V_WMMA_I32_16X16X16_IU8   : WMMAInst<"_w32", "v_wmma_i32_16x16x16_iu8",   VOP_V8I32_V4I32_V4I32_V8I32, int_amdgcn_wmma_i32_16x16x16_iu8, VRegSrc_128, WMMAUIClamp, 1>;
899  defm V_WMMA_I32_16X16X16_IU4   : WMMAInst<"_w32", "v_wmma_i32_16x16x16_iu4",   VOP_V8I32_V2I32_V2I32_V8I32, int_amdgcn_wmma_i32_16x16x16_iu4, VRegSrc_64,  WMMAUIClamp, 1>;
900}
901
902let WaveSizePredicate = isWave64 in {
903  defm V_WMMA_F32_16X16X16_F16   : WMMAInst<"_w64", "v_wmma_f32_16x16x16_f16",   VOP_V4F32_V16F16_V16F16_V4F32, int_amdgcn_wmma_f32_16x16x16_f16, VRegSrc_256, WMMARegular, 1>;
904  defm V_WMMA_F32_16X16X16_BF16  : WMMAInst<"_w64", "v_wmma_f32_16x16x16_bf16",  VOP_V4F32_V16I16_V16I16_V4F32, int_amdgcn_wmma_f32_16x16x16_bf16, VRegSrc_256, WMMARegular, 1>;
905  defm V_WMMA_F16_16X16X16_F16   : WMMAInst<"_w64", "v_wmma_f16_16x16x16_f16",   VOP_V8F16_V16F16_V16F16_V8F16, int_amdgcn_wmma_f16_16x16x16_f16, VRegSrc_256, WMMAOpSel, 1>;
906  defm V_WMMA_BF16_16X16X16_BF16 : WMMAInst<"_w64", "v_wmma_bf16_16x16x16_bf16", VOP_V8I16_V16I16_V16I16_V8I16, int_amdgcn_wmma_bf16_16x16x16_bf16, VRegSrc_256, WMMAOpSel, 1>;
907  defm V_WMMA_F16_16X16X16_F16_TIED   : WMMAInst<"_w64", "v_wmma_f16_16x16x16_f16",   VOP_V8F16_V16F16_V16F16_V8F16, int_amdgcn_wmma_f16_16x16x16_f16_tied, VRegSrc_256, WMMAOpSel, 0>;
908  defm V_WMMA_BF16_16X16X16_BF16_TIED : WMMAInst<"_w64", "v_wmma_bf16_16x16x16_bf16", VOP_V8I16_V16I16_V16I16_V8I16, int_amdgcn_wmma_bf16_16x16x16_bf16_tied, VRegSrc_256, WMMAOpSel, 0>;
909  defm V_WMMA_I32_16X16X16_IU8   : WMMAInst<"_w64", "v_wmma_i32_16x16x16_iu8",   VOP_V4I32_V4I32_V4I32_V4I32, int_amdgcn_wmma_i32_16x16x16_iu8, VRegSrc_128, WMMAUIClamp, 1>;
910  defm V_WMMA_I32_16X16X16_IU4   : WMMAInst<"_w64", "v_wmma_i32_16x16x16_iu4",   VOP_V4I32_V2I32_V2I32_V4I32, int_amdgcn_wmma_i32_16x16x16_iu4, VRegSrc_64, WMMAUIClamp, 1>;
911
912}
913
914//===----------------------------------------------------------------------===//
915// Begin Real Encodings
916//===----------------------------------------------------------------------===//
917
918class VOP3P_DPP16<bits<7> op, VOP_DPP_Pseudo ps, int subtarget,
919                  string opName = ps.OpName>
920    : VOP3P_DPP<op, opName, ps.Pfl, 1>, SIMCInstr<ps.PseudoInstr, subtarget> {
921  let hasSideEffects = ps.hasSideEffects;
922  let Defs = ps.Defs;
923  let SchedRW = ps.SchedRW;
924  let Uses = ps.Uses;
925  let AssemblerPredicate = HasDPP16;
926  let SubtargetPredicate = HasDPP16;
927  let OtherPredicates = ps.OtherPredicates;
928}
929
930class VOP3P_DPP8_Base<bits<7> op, VOP_Pseudo ps, string opName = ps.OpName>
931    : VOP3P_DPP8<op, opName, ps.Pfl> {
932  let hasSideEffects = ps.hasSideEffects;
933  let Defs = ps.Defs;
934  let SchedRW = ps.SchedRW;
935  let Uses = ps.Uses;
936  let OtherPredicates = ps.OtherPredicates;
937}
938
939//===----------------------------------------------------------------------===//
940// GFX11, GFX12
941//===----------------------------------------------------------------------===//
942
943multiclass VOP3P_Real_Base<GFXGen Gen, bits<7> op, string backing_ps_name = NAME,
944                      string asmName = !cast<VOP3P_Pseudo>(NAME).Mnemonic> {
945  def Gen.Suffix :
946    VOP3P_Real_Gen<!cast<VOP3P_Pseudo>(backing_ps_name), Gen, asmName>,
947    VOP3Pe_gfx11_gfx12<op, !cast<VOP3P_Pseudo>(backing_ps_name).Pfl>;
948}
949
950multiclass VOP3P_Real_with_name<GFXGen Gen, bits<7> op,
951                          string backing_ps_name = NAME,
952                          string asmName = !cast<VOP3P_Pseudo>(NAME).Mnemonic> {
953  defvar ps = !cast<VOP3P_Pseudo>(backing_ps_name);
954  let AsmString = asmName # ps.AsmOperands in
955    def Gen.Suffix :
956      VOP3P_Real_Gen<!cast<VOP3P_Pseudo>(backing_ps_name), Gen, asmName>,
957      VOP3Pe_gfx11_gfx12<op, !cast<VOP3P_Pseudo>(backing_ps_name).Pfl>,
958      MnemonicAlias<ps.Mnemonic, asmName>, Requires<[Gen.AssemblerPredicate]>;
959}
960
961multiclass VOP3P_Real_dpp<GFXGen Gen, bits<7> op, string backing_ps_name = NAME,
962                          string asmName = !cast<VOP3P_Pseudo>(NAME).Mnemonic> {
963  defvar ps = !cast<VOP3P_Pseudo>(backing_ps_name);
964  def _dpp#Gen.Suffix
965      : VOP3P_DPP16<op, !cast<VOP_DPP_Pseudo>(backing_ps_name #"_dpp"),
966                    Gen.Subtarget> {
967    let AsmString = asmName #ps.Pfl.AsmVOP3DPP16;
968    let DecoderNamespace = "DPP"#Gen.DecoderNamespace;
969    let AssemblerPredicate = Gen.AssemblerPredicate;
970  }
971}
972
973multiclass VOP3P_Real_dpp8<GFXGen Gen, bits<7> op, string backing_ps_name = NAME,
974                           string asmName = !cast<VOP3P_Pseudo>(NAME).Mnemonic> {
975  defvar ps = !cast<VOP3P_Pseudo>(backing_ps_name);
976  def _dpp8#Gen.Suffix : VOP3P_DPP8_Base<op, ps> {
977    let AsmString = asmName #ps.Pfl.AsmVOP3DPP8;
978    let DecoderNamespace = "DPP8"#Gen.DecoderNamespace;
979    let AssemblerPredicate = Gen.AssemblerPredicate;
980  }
981}
982
983multiclass VOP3P_Realtriple<GFXGen Gen, bits<7> op, string backing_ps_name = NAME,
984                            string asmName = !cast<VOP3P_Pseudo>(NAME).Mnemonic>
985    : VOP3P_Real_Base<Gen, op, backing_ps_name, asmName>,
986      VOP3P_Real_dpp<Gen, op, backing_ps_name, asmName>,
987      VOP3P_Real_dpp8<Gen, op, backing_ps_name, asmName>;
988
989//===----------------------------------------------------------------------===//
990// GFX12
991//===----------------------------------------------------------------------===//
992
993multiclass VOP3P_Real_gfx12<bits<7> op> : VOP3P_Real_Base<GFX12Gen, op>;
994
995multiclass VOP3P_Real_with_name_gfx12<bits<7> op,
996                          string backing_ps_name = NAME,
997                          string asmName = !cast<VOP3P_Pseudo>(NAME).Mnemonic> :
998  VOP3P_Real_with_name<GFX12Gen, op, backing_ps_name, asmName>;
999
1000defm V_PK_MIN_NUM_F16 : VOP3P_Real_with_name_gfx12<0x1b, "V_PK_MIN_F16", "v_pk_min_num_f16">;
1001defm V_PK_MAX_NUM_F16 : VOP3P_Real_with_name_gfx12<0x1c, "V_PK_MAX_F16", "v_pk_max_num_f16">;
1002
1003defm V_PK_MINIMUM_F16 : VOP3P_Real_gfx12<0x1d>;
1004defm V_PK_MAXIMUM_F16 : VOP3P_Real_gfx12<0x1e>;
1005
1006//===----------------------------------------------------------------------===//
1007// GFX11
1008//===----------------------------------------------------------------------===//
1009
1010multiclass VOP3P_Real_gfx11_gfx12<bits<7> op> :
1011   VOP3P_Real_Base<GFX11Gen, op>, VOP3P_Real_Base<GFX12Gen, op>;
1012
1013defm V_DOT4_I32_IU8  : VOP3P_Real_gfx11_gfx12<0x16>;
1014defm V_DOT8_I32_IU4  : VOP3P_Real_gfx11_gfx12<0x18>;
1015defm V_DOT2_F32_BF16 : VOP3P_Real_gfx11_gfx12<0x1a>;
1016
1017multiclass VOP3P_Real_WMMA <bits<7> op> {
1018  let WaveSizePredicate = isWave32, DecoderNamespace = "GFX11" in {
1019    defm _twoaddr_w32 : VOP3P_Real_Base <GFX11Gen, op>;
1020  }
1021  let WaveSizePredicate = isWave64, DecoderNamespace = "WMMAGFX11" in {
1022    defm _twoaddr_w64 : VOP3P_Real_Base <GFX11Gen, op>;
1023  }
1024}
1025
1026defm V_WMMA_F32_16X16X16_F16   : VOP3P_Real_WMMA <0x040>;
1027defm V_WMMA_F32_16X16X16_BF16  : VOP3P_Real_WMMA <0x041>;
1028defm V_WMMA_F16_16X16X16_F16   : VOP3P_Real_WMMA <0x042>;
1029defm V_WMMA_BF16_16X16X16_BF16 : VOP3P_Real_WMMA <0x043>;
1030defm V_WMMA_I32_16X16X16_IU8   : VOP3P_Real_WMMA <0x044>;
1031defm V_WMMA_I32_16X16X16_IU4   : VOP3P_Real_WMMA <0x045>;
1032
1033//===----------------------------------------------------------------------===//
1034// GFX8 (VI)
1035//===----------------------------------------------------------------------===//
1036
1037multiclass VOP3P_Real_vi<bits<7> op> {
1038  def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
1039            VOP3Pe <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
1040    let AssemblerPredicate = HasVOP3PInsts;
1041    let DecoderNamespace = "GFX8";
1042    let VOP3P = 1;
1043  }
1044}
1045
1046multiclass VOP3P_Real_MAI<bits<7> op> {
1047  def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
1048            VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl, ?> {
1049    let AssemblerPredicate = HasMAIInsts;
1050    let DecoderNamespace = "GFX8";
1051    let Inst{14} = ?; // op_sel_hi(2)
1052    let Inst{59} = ?; // op_sel_hi(0)
1053    let Inst{60} = ?; // op_sel_hi(1)
1054  }
1055}
1056
1057let Constraints = "" in {
1058multiclass VOP3P_Real_MFMA_gfx90a<bits<7> op> {
1059  let SubtargetPredicate = isGFX90AOnly,
1060      AssemblerPredicate = isGFX90AOnly, DecoderNamespace = "GFX90A" in {
1061  def _gfx90a_acd : VOP3P_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX90A>,
1062             VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl, 1>;
1063
1064  def _gfx90a_vcd : VOP3P_Real<!cast<VOP3_Pseudo>(NAME # "_vgprcd" # "_e64"), SIEncodingFamily.GFX90A>,
1065             VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME # "_vgprcd" # "_e64").Pfl, 0>;
1066  } // End AssemblerPredicate = isGFX90AOnly, DecoderNamespace = "GFX90A"
1067}
1068}
1069
1070multiclass VOP3P_Real_MFMA_gfx940_aliases<string NameFrom, string NameTo, string Op,
1071                                          VOP3_Pseudo PS_ACD = !cast<VOP3_Pseudo>(Op # "_e64"),
1072                                          VOP3_Pseudo PS_VCD = !cast<VOP3_Pseudo>(Op # "_vgprcd" # "_e64"),
1073                                          VOPProfile Pfl_ACD = PS_ACD.Pfl,
1074                                          VOPProfile Pfl_VCD = PS_VCD.Pfl> {
1075  if !ne(NameFrom, NameTo) then {
1076    def : InstAlias <NameTo # " " # PS_ACD.AsmOperands,
1077                     (!cast<VOP3P_Real>(Op # "_gfx940_acd") Pfl_ACD.DstRC:$vdst,
1078                         Pfl_ACD.Src0RC64:$src0, Pfl_ACD.Src1RC64:$src1, Pfl_ACD.Src2RC64:$src2,
1079                         cbsz:$cbsz, abid:$abid, blgp:$blgp)>, PredicateControl;
1080    def : InstAlias <NameTo # " " # PS_VCD.AsmOperands,
1081                     (!cast<VOP3P_Real>(Op # "_gfx940_vcd") Pfl_VCD.DstRC:$vdst,
1082                         Pfl_VCD.Src0RC64:$src0, Pfl_VCD.Src1RC64:$src1, Pfl_VCD.Src2RC64:$src2,
1083                         cbsz:$cbsz, abid:$abid, blgp:$blgp)>, PredicateControl;
1084  }
1085}
1086
1087multiclass VOP3P_Real_MFMA_gfx940<bits<7> op, string Name = !cast<VOP3_Pseudo>(NAME#"_e64").Mnemonic,
1088                                  VOP3_Pseudo PS_ACD = !cast<VOP3_Pseudo>(NAME # "_e64"),
1089                                  VOP3_Pseudo PS_VCD = !cast<VOP3_Pseudo>(NAME # "_vgprcd" # "_e64")> {
1090  let SubtargetPredicate = isGFX940Plus,
1091      DecoderNamespace = "GFX940",
1092      AsmString = Name # PS_ACD.AsmOperands, Constraints = "" in {
1093  def _gfx940_acd : VOP3P_Real<PS_ACD, SIEncodingFamily.GFX940>,
1094                    VOP3Pe_MAI <op, PS_ACD.Pfl, 1>;
1095
1096  def _gfx940_vcd : VOP3P_Real<PS_VCD, SIEncodingFamily.GFX940>,
1097                    VOP3Pe_MAI <op, PS_VCD.Pfl, 0>;
1098  } // End AssemblerPredicate = isGFX940Plus, DecoderNamespace = "GFX940"
1099
1100  let SubtargetPredicate = isGFX940Plus in {
1101    defm : VOP3P_Real_MFMA_gfx940_aliases<Name, PS_ACD.Mnemonic, NAME>;
1102
1103    if !ne(!subst("_1k", "", PS_ACD.Mnemonic), PS_ACD.Mnemonic) then
1104    defm : VOP3P_Real_MFMA_gfx940_aliases<Name, !subst("_1k", "", PS_ACD.Mnemonic), NAME>;
1105  }
1106}
1107
1108multiclass VOP3P_Real_MFMA_vi<bits<7> op> {
1109  def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
1110            VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl, ?> {
1111    let SubtargetPredicate = isGFX8GFX9NotGFX90A;
1112    let AssemblerPredicate = HasMAIInsts;
1113    let DecoderNamespace = "GFX8";
1114    let Constraints = "";
1115  }
1116}
1117
1118multiclass VOP3P_Real_MFMA_vi_gfx90a<bits<7> op> :
1119  VOP3P_Real_MFMA_gfx90a <op>,
1120  VOP3P_Real_MFMA_vi <op>;
1121
1122multiclass VOP3P_Real_MFMA<bits<7> op, string GFX940Name = !cast<VOP3_Pseudo>(NAME#"_e64").Mnemonic> :
1123  VOP3P_Real_MFMA_vi_gfx90a <op>,
1124  VOP3P_Real_MFMA_gfx940 <op, GFX940Name>;
1125
1126multiclass VOP3P_Real_SMFMAC<bits<7> op, string alias> {
1127  def _gfx940 : VOP3P_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
1128                VOP3Pe_SMFMAC <op> {
1129    let AssemblerPredicate = isGFX940Plus;
1130    let DecoderNamespace = "GFX8";
1131  }
1132  def : MnemonicAlias<alias, !cast<VOP3_Pseudo>(NAME#"_e64").Mnemonic>;
1133}
1134
1135let SubtargetPredicate = isGFX8GFX9 in {
1136defm V_PK_MAD_I16 : VOP3P_Real_vi <0x00>;
1137defm V_PK_MUL_LO_U16 : VOP3P_Real_vi <0x01>;
1138defm V_PK_ADD_I16 : VOP3P_Real_vi <0x02>;
1139defm V_PK_SUB_I16 : VOP3P_Real_vi <0x03>;
1140defm V_PK_LSHLREV_B16 : VOP3P_Real_vi <0x04>;
1141defm V_PK_LSHRREV_B16 : VOP3P_Real_vi <0x05>;
1142defm V_PK_ASHRREV_I16 : VOP3P_Real_vi <0x06>;
1143defm V_PK_MAX_I16 : VOP3P_Real_vi <0x07>;
1144defm V_PK_MIN_I16 : VOP3P_Real_vi <0x08>;
1145defm V_PK_MAD_U16 : VOP3P_Real_vi <0x09>;
1146
1147defm V_PK_ADD_U16 : VOP3P_Real_vi <0x0a>;
1148defm V_PK_SUB_U16 : VOP3P_Real_vi <0x0b>;
1149defm V_PK_MAX_U16 : VOP3P_Real_vi <0x0c>;
1150defm V_PK_MIN_U16 : VOP3P_Real_vi <0x0d>;
1151defm V_PK_FMA_F16 : VOP3P_Real_vi <0x0e>;
1152defm V_PK_ADD_F16 : VOP3P_Real_vi <0x0f>;
1153defm V_PK_MUL_F16 : VOP3P_Real_vi <0x10>;
1154defm V_PK_MIN_F16 : VOP3P_Real_vi <0x11>;
1155defm V_PK_MAX_F16 : VOP3P_Real_vi <0x12>;
1156
1157let OtherPredicates = [HasMadMixInsts] in {
1158defm V_MAD_MIX_F32 : VOP3P_Real_vi <0x20>;
1159defm V_MAD_MIXLO_F16 : VOP3P_Real_vi <0x21>;
1160defm V_MAD_MIXHI_F16 : VOP3P_Real_vi <0x22>;
1161}
1162
1163let OtherPredicates = [HasFmaMixInsts],
1164    DecoderNamespace = "GFX9_DL" in {
1165// The mad_mix instructions were renamed and their behaviors changed,
1166// but the opcode stayed the same so we need to put these in a
1167// different DecoderNamespace to avoid the ambiguity.
1168defm V_FMA_MIX_F32 : VOP3P_Real_vi <0x20>;
1169defm V_FMA_MIXLO_F16 : VOP3P_Real_vi <0x21>;
1170defm V_FMA_MIXHI_F16 : VOP3P_Real_vi <0x22>;
1171}
1172
1173defm V_DOT2_I32_I16 : VOP3P_Real_vi <0x26>;
1174defm V_DOT2_U32_U16 : VOP3P_Real_vi <0x27>;
1175
1176defm V_DOT2_F32_F16 : VOP3P_Real_vi <0x23>;
1177defm V_DOT4_U32_U8  : VOP3P_Real_vi <0x29>;
1178defm V_DOT8_U32_U4  : VOP3P_Real_vi <0x2b>;
1179
1180defm V_DOT4_I32_I8  : VOP3P_Real_vi <0x28>;
1181defm V_DOT8_I32_I4  : VOP3P_Real_vi <0x2a>;
1182} // End SubtargetPredicate = isGFX8GFX9
1183
1184let OtherPredicates = [HasMAIInsts] in {
1185
1186defm V_ACCVGPR_READ_B32  : VOP3P_Real_MAI <0x58>;
1187defm V_ACCVGPR_WRITE_B32 : VOP3P_Real_MAI <0x59>;
1188defm V_MFMA_F32_32X32X1F32  : VOP3P_Real_MFMA <0x40, "v_mfma_f32_32x32x1_2b_f32">;
1189defm V_MFMA_F32_16X16X1F32  : VOP3P_Real_MFMA <0x41, "v_mfma_f32_16x16x1_4b_f32">;
1190defm V_MFMA_F32_4X4X1F32    : VOP3P_Real_MFMA <0x42, "v_mfma_f32_4x4x1_16b_f32">;
1191defm V_MFMA_F32_32X32X2F32  : VOP3P_Real_MFMA <0x44, "v_mfma_f32_32x32x2_f32">;
1192defm V_MFMA_F32_16X16X4F32  : VOP3P_Real_MFMA <0x45, "v_mfma_f32_16x16x4_f32">;
1193defm V_MFMA_F32_32X32X4F16  : VOP3P_Real_MFMA <0x48, "v_mfma_f32_32x32x4_2b_f16">;
1194defm V_MFMA_F32_16X16X4F16  : VOP3P_Real_MFMA <0x49, "v_mfma_f32_16x16x4_4b_f16">;
1195defm V_MFMA_F32_4X4X4F16    : VOP3P_Real_MFMA <0x4a, "v_mfma_f32_4x4x4_16b_f16">;
1196defm V_MFMA_F32_32X32X8F16  : VOP3P_Real_MFMA <0x4c, "v_mfma_f32_32x32x8_f16">;
1197defm V_MFMA_F32_16X16X16F16 : VOP3P_Real_MFMA <0x4d, "v_mfma_f32_16x16x16_f16">;
1198defm V_MFMA_I32_32X32X4I8   : VOP3P_Real_MFMA <0x50, "v_mfma_i32_32x32x4_2b_i8">;
1199defm V_MFMA_I32_16X16X4I8   : VOP3P_Real_MFMA <0x51, "v_mfma_i32_16x16x4_4b_i8">;
1200defm V_MFMA_I32_4X4X4I8     : VOP3P_Real_MFMA <0x52, "v_mfma_i32_4x4x4_16b_i8">;
1201
1202defm V_MFMA_I32_16X16X16I8  : VOP3P_Real_MFMA_vi_gfx90a <0x55>;
1203defm V_MFMA_I32_32X32X8I8   : VOP3P_Real_MFMA_vi_gfx90a <0x54>;
1204defm V_MFMA_F32_32X32X2BF16 : VOP3P_Real_MFMA_vi_gfx90a <0x68>;
1205defm V_MFMA_F32_16X16X2BF16 : VOP3P_Real_MFMA_vi_gfx90a <0x69>;
1206defm V_MFMA_F32_4X4X2BF16   : VOP3P_Real_MFMA_vi_gfx90a <0x6b>;
1207defm V_MFMA_F32_32X32X4BF16 : VOP3P_Real_MFMA_vi_gfx90a <0x6c>;
1208defm V_MFMA_F32_16X16X8BF16 : VOP3P_Real_MFMA_vi_gfx90a <0x6d>;
1209
1210} // End OtherPredicates = [HasMAIInsts]
1211
1212defm V_MFMA_F32_32X32X4BF16_1K  : VOP3P_Real_MFMA_gfx90a <0x63>;
1213defm V_MFMA_F32_16X16X4BF16_1K  : VOP3P_Real_MFMA_gfx90a <0x64>;
1214defm V_MFMA_F32_4X4X4BF16_1K    : VOP3P_Real_MFMA_gfx90a <0x65>;
1215defm V_MFMA_F32_32X32X8BF16_1K  : VOP3P_Real_MFMA_gfx90a <0x66>;
1216defm V_MFMA_F32_16X16X16BF16_1K : VOP3P_Real_MFMA_gfx90a <0x67>;
1217defm V_MFMA_F64_16X16X4F64      : VOP3P_Real_MFMA_gfx90a <0x6e>;
1218defm V_MFMA_F64_4X4X4F64        : VOP3P_Real_MFMA_gfx90a <0x6f>;
1219
1220defm V_MFMA_I32_32X32X16I8       : VOP3P_Real_MFMA_gfx940 <0x56, "v_mfma_i32_32x32x16_i8">;
1221defm V_MFMA_I32_16X16X32I8       : VOP3P_Real_MFMA_gfx940 <0x57, "v_mfma_i32_16x16x32_i8">;
1222defm V_MFMA_F32_16X16X8XF32      : VOP3P_Real_MFMA_gfx940 <0x3e, "v_mfma_f32_16x16x8_xf32">;
1223defm V_MFMA_F32_32X32X4XF32      : VOP3P_Real_MFMA_gfx940 <0x3f, "v_mfma_f32_32x32x4_xf32">;
1224defm V_MFMA_F32_16X16X32_BF8_BF8 : VOP3P_Real_MFMA_gfx940 <0x70>;
1225defm V_MFMA_F32_16X16X32_BF8_FP8 : VOP3P_Real_MFMA_gfx940 <0x71>;
1226defm V_MFMA_F32_16X16X32_FP8_BF8 : VOP3P_Real_MFMA_gfx940 <0x72>;
1227defm V_MFMA_F32_16X16X32_FP8_FP8 : VOP3P_Real_MFMA_gfx940 <0x73>;
1228defm V_MFMA_F32_32X32X16_BF8_BF8 : VOP3P_Real_MFMA_gfx940 <0x74>;
1229defm V_MFMA_F32_32X32X16_BF8_FP8 : VOP3P_Real_MFMA_gfx940 <0x75>;
1230defm V_MFMA_F32_32X32X16_FP8_BF8 : VOP3P_Real_MFMA_gfx940 <0x76>;
1231defm V_MFMA_F32_32X32X16_FP8_FP8 : VOP3P_Real_MFMA_gfx940 <0x77>;
1232
1233defm V_MFMA_F32_32X32X4BF16_1K   : VOP3P_Real_MFMA_gfx940 <0x5d, "v_mfma_f32_32x32x4_2b_bf16">;
1234defm V_MFMA_F32_16X16X4BF16_1K   : VOP3P_Real_MFMA_gfx940 <0x5e, "v_mfma_f32_16x16x4_4b_bf16">;
1235defm V_MFMA_F32_4X4X4BF16_1K     : VOP3P_Real_MFMA_gfx940 <0x5f, "v_mfma_f32_4x4x4_16b_bf16">;
1236defm V_MFMA_F32_32X32X8BF16_1K   : VOP3P_Real_MFMA_gfx940 <0x60, "v_mfma_f32_32x32x8_bf16">;
1237defm V_MFMA_F32_16X16X16BF16_1K  : VOP3P_Real_MFMA_gfx940 <0x61, "v_mfma_f32_16x16x16_bf16">;
1238
1239defm V_MFMA_F64_16X16X4F64       : VOP3P_Real_MFMA_gfx940 <0x6e, "v_mfma_f64_16x16x4_f64">;
1240defm V_MFMA_F64_4X4X4F64         : VOP3P_Real_MFMA_gfx940 <0x6f, "v_mfma_f64_4x4x4_4b_f64">;
1241
1242defm V_SMFMAC_F32_16X16X32_F16     : VOP3P_Real_SMFMAC <0x62, "v_smfmac_f32_16x16x32f16">;
1243defm V_SMFMAC_F32_32X32X16_F16     : VOP3P_Real_SMFMAC <0x64, "v_smfmac_f32_32x32x16f16">;
1244defm V_SMFMAC_F32_16X16X32_BF16    : VOP3P_Real_SMFMAC <0x66, "v_smfmac_f32_16x16x32bf16">;
1245defm V_SMFMAC_F32_32X32X16_BF16    : VOP3P_Real_SMFMAC <0x68, "v_smfmac_f32_32x32x16bf16">;
1246defm V_SMFMAC_I32_16X16X64_I8      : VOP3P_Real_SMFMAC <0x6a, "v_smfmac_i32_16x16x64i8">;
1247defm V_SMFMAC_I32_32X32X32_I8      : VOP3P_Real_SMFMAC <0x6c, "v_smfmac_i32_32x32x32i8">;
1248defm V_SMFMAC_F32_16X16X64_BF8_BF8 : VOP3P_Real_SMFMAC <0x78, "v_smfmac_f32_16x16x64bf8bf8">;
1249defm V_SMFMAC_F32_16X16X64_BF8_FP8 : VOP3P_Real_SMFMAC <0x79, "v_smfmac_f32_16x16x64bf8fp8">;
1250defm V_SMFMAC_F32_16X16X64_FP8_BF8 : VOP3P_Real_SMFMAC <0x7a, "v_smfmac_f32_16x16x64fp8bf8">;
1251defm V_SMFMAC_F32_16X16X64_FP8_FP8 : VOP3P_Real_SMFMAC <0x7b, "v_smfmac_f32_16x16x64fp8fp8">;
1252defm V_SMFMAC_F32_32X32X32_BF8_BF8 : VOP3P_Real_SMFMAC <0x7c, "v_smfmac_f32_32x32x32bf8bf8">;
1253defm V_SMFMAC_F32_32X32X32_BF8_FP8 : VOP3P_Real_SMFMAC <0x7d, "v_smfmac_f32_32x32x32bf8fp8">;
1254defm V_SMFMAC_F32_32X32X32_FP8_BF8 : VOP3P_Real_SMFMAC <0x7e, "v_smfmac_f32_32x32x32fp8bf8">;
1255defm V_SMFMAC_F32_32X32X32_FP8_FP8 : VOP3P_Real_SMFMAC <0x7f, "v_smfmac_f32_32x32x32fp8fp8">;
1256
1257defm V_PK_FMA_F32 : VOP3P_Real_vi <0x30>;
1258defm V_PK_MUL_F32 : VOP3P_Real_vi <0x31>;
1259defm V_PK_ADD_F32 : VOP3P_Real_vi <0x32>;
1260defm V_PK_MOV_B32 : VOP3P_Real_vi <0x33>;
1261
1262//===----------------------------------------------------------------------===//
1263// GFX10.
1264//===----------------------------------------------------------------------===//
1265
1266let AssemblerPredicate = isGFX10Only, DecoderNamespace = "GFX10", VOP3P = 1 in {
1267  multiclass VOP3P_Real_gfx10<bits<7> op> {
1268    def _gfx10 : VOP3P_Real<!cast<VOP3P_Pseudo>(NAME), SIEncodingFamily.GFX10>,
1269                 VOP3Pe_gfx10 <op, !cast<VOP3P_Pseudo>(NAME).Pfl>;
1270  }
1271} // End AssemblerPredicate = isGFX10Only, DecoderNamespace = "GFX10", VOP3P = 1
1272
1273multiclass VOP3P_Real_gfx10_gfx11<bits<7> op> :
1274  VOP3P_Real_gfx10<op>, VOP3P_Real_Base<GFX11Gen, op>;
1275
1276multiclass VOP3P_Real_gfx10_gfx11_gfx12<bits<7> op> :
1277  VOP3P_Real_gfx10_gfx11<op>, VOP3P_Real_Base<GFX12Gen, op>;
1278
1279multiclass VOP3P_Real_gfx10_gfx11_gfx12_Triple<bits<7> op> :
1280  VOP3P_Real_gfx10<op>, VOP3P_Realtriple<GFX11Gen, op>,
1281  VOP3P_Realtriple<GFX12Gen, op>;
1282
1283defm V_PK_MAD_I16     : VOP3P_Real_gfx10_gfx11_gfx12<0x00>;
1284defm V_PK_MUL_LO_U16  : VOP3P_Real_gfx10_gfx11_gfx12<0x01>;
1285defm V_PK_ADD_I16     : VOP3P_Real_gfx10_gfx11_gfx12<0x02>;
1286defm V_PK_SUB_I16     : VOP3P_Real_gfx10_gfx11_gfx12<0x03>;
1287defm V_PK_LSHLREV_B16 : VOP3P_Real_gfx10_gfx11_gfx12<0x04>;
1288defm V_PK_LSHRREV_B16 : VOP3P_Real_gfx10_gfx11_gfx12<0x05>;
1289defm V_PK_ASHRREV_I16 : VOP3P_Real_gfx10_gfx11_gfx12<0x06>;
1290defm V_PK_MAX_I16     : VOP3P_Real_gfx10_gfx11_gfx12<0x07>;
1291defm V_PK_MIN_I16     : VOP3P_Real_gfx10_gfx11_gfx12<0x08>;
1292defm V_PK_MAD_U16     : VOP3P_Real_gfx10_gfx11_gfx12<0x09>;
1293defm V_PK_ADD_U16     : VOP3P_Real_gfx10_gfx11_gfx12<0x0a>;
1294defm V_PK_SUB_U16     : VOP3P_Real_gfx10_gfx11_gfx12<0x0b>;
1295defm V_PK_MAX_U16     : VOP3P_Real_gfx10_gfx11_gfx12<0x0c>;
1296defm V_PK_MIN_U16     : VOP3P_Real_gfx10_gfx11_gfx12<0x0d>;
1297defm V_PK_FMA_F16     : VOP3P_Real_gfx10_gfx11_gfx12<0x0e>;
1298defm V_PK_ADD_F16     : VOP3P_Real_gfx10_gfx11_gfx12<0x0f>;
1299defm V_PK_MUL_F16     : VOP3P_Real_gfx10_gfx11_gfx12<0x10>;
1300defm V_PK_MIN_F16     : VOP3P_Real_gfx10_gfx11<0x11>;
1301defm V_PK_MAX_F16     : VOP3P_Real_gfx10_gfx11<0x12>;
1302defm V_FMA_MIX_F32    : VOP3P_Real_gfx10_gfx11_gfx12_Triple<0x20>;
1303defm V_FMA_MIXLO_F16  : VOP3P_Real_gfx10_gfx11_gfx12_Triple<0x21>;
1304defm V_FMA_MIXHI_F16  : VOP3P_Real_gfx10_gfx11_gfx12_Triple<0x22>;
1305
1306defm V_DOT2_I32_I16 : VOP3P_Real_gfx10 <0x14>;
1307defm V_DOT2_U32_U16 : VOP3P_Real_gfx10 <0x15>;
1308
1309defm V_DOT2_F32_F16 : VOP3P_Real_gfx10_gfx11_gfx12_Triple<0x13>;
1310defm V_DOT4_U32_U8  : VOP3P_Real_gfx10_gfx11_gfx12<0x17>;
1311defm V_DOT8_U32_U4  : VOP3P_Real_gfx10_gfx11_gfx12<0x19>;
1312
1313defm V_DOT4_I32_I8  : VOP3P_Real_gfx10 <0x16>;
1314defm V_DOT8_I32_I4  : VOP3P_Real_gfx10 <0x18>;
1315