1//===-- VOP3PInstructions.td - Vector Instruction Defintions --------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9
10//===----------------------------------------------------------------------===//
11// VOP3P Classes
12//===----------------------------------------------------------------------===//
13
14class VOP3PInst<string OpName, VOPProfile P, SDPatternOperator node = null_frag> :
15  VOP3P_Pseudo<OpName, P,
16    !if(P.HasModifiers, getVOP3PModPat<P, node>.ret, getVOP3Pat<P, node>.ret)
17>;
18
19// Non-packed instructions that use the VOP3P encoding.
20// VOP3 neg/abs and VOP3P opsel/opsel_hi modifiers are allowed.
21class VOP3_VOP3PInst<string OpName, VOPProfile P, bit UseTiedOutput = 0,
22                     SDPatternOperator node = null_frag> :
23  VOP3P_Pseudo<OpName, P> {
24  // These 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.
28  let InOperandList =
29    !con(
30      !con(
31        (ins FP16InputMods:$src0_modifiers, VCSrc_f16:$src0,
32             FP16InputMods:$src1_modifiers, VCSrc_f16:$src1,
33             FP16InputMods:$src2_modifiers, VCSrc_f16:$src2,
34             clampmod:$clamp),
35         !if(UseTiedOutput, (ins VGPR_32:$vdst_in), (ins))),
36         (ins op_sel:$op_sel, op_sel_hi:$op_sel_hi));
37
38  let Constraints = !if(UseTiedOutput, "$vdst = $vdst_in", "");
39  let DisableEncoding = !if(UseTiedOutput, "$vdst_in", "");
40  let AsmOperands =
41    " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$op_sel$op_sel_hi$clamp";
42}
43
44let isCommutable = 1 in {
45def V_PK_MAD_I16 : VOP3PInst<"v_pk_mad_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
46def V_PK_MAD_U16 : VOP3PInst<"v_pk_mad_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
47
48let FPDPRounding = 1 in {
49def V_PK_FMA_F16 : VOP3PInst<"v_pk_fma_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, fma>;
50def V_PK_ADD_F16 : VOP3PInst<"v_pk_add_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fadd>;
51def V_PK_MUL_F16 : VOP3PInst<"v_pk_mul_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmul>;
52} // End FPDPRounding = 1
53def V_PK_MAX_F16 : VOP3PInst<"v_pk_max_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmaxnum_like>;
54def V_PK_MIN_F16 : VOP3PInst<"v_pk_min_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fminnum_like>;
55
56def V_PK_ADD_U16 : VOP3PInst<"v_pk_add_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, add>;
57def V_PK_ADD_I16 : VOP3PInst<"v_pk_add_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
58def V_PK_MUL_LO_U16 : VOP3PInst<"v_pk_mul_lo_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, mul>;
59
60def V_PK_MIN_I16 : VOP3PInst<"v_pk_min_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smin>;
61def V_PK_MIN_U16 : VOP3PInst<"v_pk_min_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umin>;
62def V_PK_MAX_I16 : VOP3PInst<"v_pk_max_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smax>;
63def V_PK_MAX_U16 : VOP3PInst<"v_pk_max_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umax>;
64}
65
66def V_PK_SUB_U16 : VOP3PInst<"v_pk_sub_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
67def V_PK_SUB_I16 : VOP3PInst<"v_pk_sub_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, sub>;
68
69def V_PK_LSHLREV_B16 : VOP3PInst<"v_pk_lshlrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshl_rev>;
70def V_PK_ASHRREV_I16 : VOP3PInst<"v_pk_ashrrev_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, ashr_rev>;
71def V_PK_LSHRREV_B16 : VOP3PInst<"v_pk_lshrrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshr_rev>;
72
73multiclass MadFmaMixPats<SDPatternOperator fma_like,
74                         Instruction mix_inst,
75                         Instruction mixlo_inst,
76                         Instruction mixhi_inst> {
77  def : GCNPat <
78    (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
79                            (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
80                            (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
81    (mixlo_inst $src0_modifiers, $src0,
82                $src1_modifiers, $src1,
83                $src2_modifiers, $src2,
84                DSTCLAMP.NONE,
85                (i32 (IMPLICIT_DEF)))
86  >;
87
88  // FIXME: Special case handling for maxhi (especially for clamp)
89  // because dealing with the write to high half of the register is
90  // difficult.
91  def : GCNPat <
92    (build_vector f16:$elt0, (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
93                                                (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
94                                                (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
95    (v2f16 (mixhi_inst $src0_modifiers, $src0,
96                       $src1_modifiers, $src1,
97                       $src2_modifiers, $src2,
98                       DSTCLAMP.NONE,
99                       $elt0))
100  >;
101
102  def : GCNPat <
103    (build_vector
104      f16:$elt0,
105      (AMDGPUclamp (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
106                                      (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
107                                      (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers)))))),
108    (v2f16 (mixhi_inst $src0_modifiers, $src0,
109                       $src1_modifiers, $src1,
110                       $src2_modifiers, $src2,
111                       DSTCLAMP.ENABLE,
112                       $elt0))
113  >;
114
115  def : GCNPat <
116    (AMDGPUclamp (build_vector
117      (fpround (fma_like (f32 (VOP3PMadMixMods f16:$lo_src0, i32:$lo_src0_modifiers)),
118                         (f32 (VOP3PMadMixMods f16:$lo_src1, i32:$lo_src1_modifiers)),
119                         (f32 (VOP3PMadMixMods f16:$lo_src2, i32:$lo_src2_modifiers)))),
120      (fpround (fma_like (f32 (VOP3PMadMixMods f16:$hi_src0, i32:$hi_src0_modifiers)),
121                         (f32 (VOP3PMadMixMods f16:$hi_src1, i32:$hi_src1_modifiers)),
122                         (f32 (VOP3PMadMixMods f16:$hi_src2, i32:$hi_src2_modifiers)))))),
123    (v2f16 (mixhi_inst $hi_src0_modifiers, $hi_src0,
124                       $hi_src1_modifiers, $hi_src1,
125                       $hi_src2_modifiers, $hi_src2,
126                       DSTCLAMP.ENABLE,
127                       (mixlo_inst $lo_src0_modifiers, $lo_src0,
128                                   $lo_src1_modifiers, $lo_src1,
129                                   $lo_src2_modifiers, $lo_src2,
130                                   DSTCLAMP.ENABLE,
131                                   (i32 (IMPLICIT_DEF)))))
132  >;
133}
134
135let SubtargetPredicate = HasMadMixInsts in {
136// These are VOP3a-like opcodes which accept no omod.
137// Size of src arguments (16/32) is controlled by op_sel.
138// For 16-bit src arguments their location (hi/lo) are controlled by op_sel_hi.
139let isCommutable = 1 in {
140def V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
141
142let FPDPRounding = 1 in {
143// Clamp modifier is applied after conversion to f16.
144def V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
145
146let ClampLo = 0, ClampHi = 1 in {
147def V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
148}
149} // End FPDPRounding = 1
150}
151
152defm : MadFmaMixPats<fmad, V_MAD_MIX_F32, V_MAD_MIXLO_F16, V_MAD_MIXHI_F16>;
153} // End SubtargetPredicate = HasMadMixInsts
154
155
156// Essentially the same as the mad_mix versions
157let SubtargetPredicate = HasFmaMixInsts in {
158let isCommutable = 1 in {
159def V_FMA_MIX_F32 : VOP3_VOP3PInst<"v_fma_mix_f32", VOP3_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
160
161let FPDPRounding = 1 in {
162// Clamp modifier is applied after conversion to f16.
163def V_FMA_MIXLO_F16 : VOP3_VOP3PInst<"v_fma_mixlo_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
164
165let ClampLo = 0, ClampHi = 1 in {
166def V_FMA_MIXHI_F16 : VOP3_VOP3PInst<"v_fma_mixhi_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
167}
168} // End FPDPRounding = 1
169}
170
171defm : MadFmaMixPats<fma, V_FMA_MIX_F32, V_FMA_MIXLO_F16, V_FMA_MIXHI_F16>;
172}
173
174// Defines patterns that extract signed 4bit from each Idx[0].
175foreach Idx = [[0,28],[4,24],[8,20],[12,16],[16,12],[20,8],[24,4]] in
176  def ExtractSigned4bit_#Idx[0] : PatFrag<(ops node:$src),
177                                          (sra (shl node:$src, (i32 Idx[1])), (i32 28))>;
178
179// Defines code pattern that extracts U(unsigned/signed) 4/8bit from FromBitIndex.
180class Extract<int FromBitIndex, int BitMask, bit U>: PatFrag<
181  (ops node:$src),
182  !if (!or (!and (!eq (BitMask, 255), !eq (FromBitIndex, 24)), !eq (FromBitIndex, 28)), // last element
183       !if (U, (srl node:$src, (i32 FromBitIndex)), (sra node:$src, (i32 FromBitIndex))),
184       !if (!eq (FromBitIndex, 0), // first element
185            !if (U, (and node:$src, (i32 BitMask)),
186                 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
187                                         (sext_inreg node:$src, i8))),
188            !if (U, (and (srl node:$src, (i32 FromBitIndex)), (i32 BitMask)),
189                 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
190                      (sext_inreg (srl node:$src, (i32 FromBitIndex)), i8)))))>;
191
192
193foreach Type = ["I", "U"] in
194  foreach Index = 0-3 in {
195    // Defines patterns that extract each Index'ed 8bit from an unsigned
196    // 32bit scalar value;
197    def #Type#Index#"_8bit" : Extract<!shl(Index, 3), 255, !if (!eq (Type, "U"), 1, 0)>;
198
199    // Defines multiplication patterns where the multiplication is happening on each
200    // Index'ed 8bit of a 32bit scalar value.
201
202    def Mul#Type#_Elt#Index : PatFrag<
203      (ops node:$src0, node:$src1),
204      (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), AMDGPUmul_i24_oneuse, AMDGPUmul_u24_oneuse))
205                            (!cast<Extract>(#Type#Index#"_8bit") node:$src0),
206                            (!cast<Extract>(#Type#Index#"_8bit") node:$src1))>;
207  }
208
209// Different variants of dot8 patterns cause a huge increase in the compile time.
210// Define non-associative/commutative add/mul to prevent permutation in the dot8
211// pattern.
212def NonACAdd        : SDNode<"ISD::ADD"       , SDTIntBinOp>;
213def NonACAdd_oneuse : HasOneUseBinOp<NonACAdd>;
214
215def NonACAMDGPUmul_u24        : SDNode<"AMDGPUISD::MUL_U24"       , SDTIntBinOp>;
216def NonACAMDGPUmul_u24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_u24>;
217
218def NonACAMDGPUmul_i24        : SDNode<"AMDGPUISD::MUL_I24"       , SDTIntBinOp>;
219def NonACAMDGPUmul_i24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_i24>;
220
221foreach Type = ["I", "U"] in
222  foreach Index = 0-7 in {
223    // Defines patterns that extract each Index'ed 4bit from an unsigned
224    // 32bit scalar value;
225    def #Type#Index#"_4bit" : Extract<!shl(Index, 2), 15, !if (!eq (Type, "U"), 1, 0)>;
226
227    // Defines multiplication patterns where the multiplication is happening on each
228    // Index'ed 8bit of a 32bit scalar value.
229    def Mul#Type#Index#"_4bit" : PatFrag<
230      (ops node:$src0, node:$src1),
231      (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), NonACAMDGPUmul_i24_oneuse, NonACAMDGPUmul_u24_oneuse))
232                             (!cast<Extract>(#Type#Index#"_4bit") node:$src0),
233                             (!cast<Extract>(#Type#Index#"_4bit") node:$src1))>;
234  }
235
236class UDot2Pat<Instruction Inst> : GCNPat <
237  (add (add_oneuse (AMDGPUmul_u24_oneuse (srl i32:$src0, (i32 16)),
238                                         (srl i32:$src1, (i32 16))), i32:$src2),
239       (AMDGPUmul_u24_oneuse (and i32:$src0, (i32 65535)),
240                             (and i32:$src1, (i32 65535)))
241   ),
242  (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))
243>;
244
245class SDot2Pat<Instruction Inst> : GCNPat <
246  (add (add_oneuse (AMDGPUmul_i24_oneuse (sra i32:$src0, (i32 16)),
247                                         (sra i32:$src1, (i32 16))), i32:$src2),
248       (AMDGPUmul_i24_oneuse (sext_inreg i32:$src0, i16),
249                             (sext_inreg i32:$src1, i16))),
250  (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))
251>;
252
253let SubtargetPredicate = HasDLInsts in {
254
255def V_DOT2_F32_F16 : VOP3PInst<"v_dot2_f32_f16", VOP3_Profile<VOP_F32_V2F16_V2F16_F32>>;
256def V_DOT2_I32_I16 : VOP3PInst<"v_dot2_i32_i16", VOP3_Profile<VOP_I32_V2I16_V2I16_I32>>;
257def V_DOT2_U32_U16 : VOP3PInst<"v_dot2_u32_u16", VOP3_Profile<VOP_I32_V2I16_V2I16_I32>>;
258def V_DOT4_I32_I8  : VOP3PInst<"v_dot4_i32_i8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
259def V_DOT4_U32_U8  : VOP3PInst<"v_dot4_u32_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
260def V_DOT8_I32_I4  : VOP3PInst<"v_dot8_i32_i4", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
261def V_DOT8_U32_U4  : VOP3PInst<"v_dot8_u32_u4", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
262
263multiclass DotPats<SDPatternOperator dot_op,
264                   VOP3PInst dot_inst> {
265  def : GCNPat <
266    (dot_op (dot_inst.Pfl.Src0VT (VOP3PMods0 dot_inst.Pfl.Src0VT:$src0, i32:$src0_modifiers)),
267            (dot_inst.Pfl.Src1VT (VOP3PMods dot_inst.Pfl.Src1VT:$src1, i32:$src1_modifiers)),
268            (dot_inst.Pfl.Src2VT (VOP3PMods dot_inst.Pfl.Src2VT:$src2, i32:$src2_modifiers)), i1:$clamp),
269    (dot_inst $src0_modifiers, $src0, $src1_modifiers, $src1, $src2_modifiers, $src2, (as_i1imm $clamp))>;
270}
271
272defm : DotPats<AMDGPUfdot2, V_DOT2_F32_F16>;
273defm : DotPats<int_amdgcn_sdot2, V_DOT2_I32_I16>;
274defm : DotPats<int_amdgcn_udot2, V_DOT2_U32_U16>;
275defm : DotPats<int_amdgcn_sdot4, V_DOT4_I32_I8>;
276defm : DotPats<int_amdgcn_udot4, V_DOT4_U32_U8>;
277defm : DotPats<int_amdgcn_sdot8, V_DOT8_I32_I4>;
278defm : DotPats<int_amdgcn_udot8, V_DOT8_U32_U4>;
279
280def : UDot2Pat<V_DOT2_U32_U16>;
281def : SDot2Pat<V_DOT2_I32_I16>;
282
283foreach Type = ["U", "I"] in
284  def : GCNPat <
285    !cast<dag>(!foldl((i32 i32:$src2), [0, 1, 2, 3], lhs, y,
286                      (add_oneuse lhs, (!cast<PatFrag>("Mul"#Type#"_Elt"#y) i32:$src0, i32:$src1)))),
287    (!cast<VOP3PInst>("V_DOT4_"#Type#"32_"#Type#8) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
288
289foreach Type = ["U", "I"] in
290  def : GCNPat <
291    !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
292                      [1, 2, 3, 4, 5, 6, 7], lhs, y,
293                      (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
294    (!cast<VOP3PInst>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
295
296// Different variants of dot8 code-gen dag patterns are not generated through table-gen due to a huge increase
297// in the compile time. Directly handle the pattern generated by the FE here.
298foreach Type = ["U", "I"] in
299  def : GCNPat <
300    !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
301                      [7, 1, 2, 3, 4, 5, 6], lhs, y,
302                      (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
303    (!cast<VOP3PInst>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
304
305} // End SubtargetPredicate = HasDLInsts
306
307multiclass VOP3P_Real_vi<bits<10> op> {
308  def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
309            VOP3Pe <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
310    let AssemblerPredicates = [HasVOP3PInsts];
311    let DecoderNamespace = "VI";
312  }
313}
314
315defm V_PK_MAD_I16 : VOP3P_Real_vi <0x380>;
316defm V_PK_MUL_LO_U16 : VOP3P_Real_vi <0x381>;
317defm V_PK_ADD_I16 : VOP3P_Real_vi <0x382>;
318defm V_PK_SUB_I16 : VOP3P_Real_vi <0x383>;
319defm V_PK_LSHLREV_B16 : VOP3P_Real_vi <0x384>;
320defm V_PK_LSHRREV_B16 : VOP3P_Real_vi <0x385>;
321defm V_PK_ASHRREV_I16 : VOP3P_Real_vi <0x386>;
322defm V_PK_MAX_I16 : VOP3P_Real_vi <0x387>;
323defm V_PK_MIN_I16 : VOP3P_Real_vi <0x388>;
324defm V_PK_MAD_U16 : VOP3P_Real_vi <0x389>;
325
326defm V_PK_ADD_U16 : VOP3P_Real_vi <0x38a>;
327defm V_PK_SUB_U16 : VOP3P_Real_vi <0x38b>;
328defm V_PK_MAX_U16 : VOP3P_Real_vi <0x38c>;
329defm V_PK_MIN_U16 : VOP3P_Real_vi <0x38d>;
330defm V_PK_FMA_F16 : VOP3P_Real_vi <0x38e>;
331defm V_PK_ADD_F16 : VOP3P_Real_vi <0x38f>;
332defm V_PK_MUL_F16 : VOP3P_Real_vi <0x390>;
333defm V_PK_MIN_F16 : VOP3P_Real_vi <0x391>;
334defm V_PK_MAX_F16 : VOP3P_Real_vi <0x392>;
335
336
337let SubtargetPredicate = HasMadMixInsts in {
338defm V_MAD_MIX_F32 : VOP3P_Real_vi <0x3a0>;
339defm V_MAD_MIXLO_F16 : VOP3P_Real_vi <0x3a1>;
340defm V_MAD_MIXHI_F16 : VOP3P_Real_vi <0x3a2>;
341}
342
343let SubtargetPredicate = HasFmaMixInsts in {
344let DecoderNamespace = "GFX9_DL" in {
345// The mad_mix instructions were renamed and their behaviors changed,
346// but the opcode stayed the same so we need to put these in a
347// different DecoderNamespace to avoid the ambiguity.
348defm V_FMA_MIX_F32 : VOP3P_Real_vi <0x3a0>;
349defm V_FMA_MIXLO_F16 : VOP3P_Real_vi <0x3a1>;
350defm V_FMA_MIXHI_F16 : VOP3P_Real_vi <0x3a2>;
351}
352}
353
354
355let SubtargetPredicate = HasDLInsts in {
356
357defm V_DOT2_F32_F16 : VOP3P_Real_vi <0x3a3>;
358defm V_DOT2_I32_I16 : VOP3P_Real_vi <0x3a6>;
359defm V_DOT2_U32_U16 : VOP3P_Real_vi <0x3a7>;
360defm V_DOT4_I32_I8  : VOP3P_Real_vi <0x3a8>;
361defm V_DOT4_U32_U8  : VOP3P_Real_vi <0x3a9>;
362defm V_DOT8_I32_I4  : VOP3P_Real_vi <0x3aa>;
363defm V_DOT8_U32_U4  : VOP3P_Real_vi <0x3ab>;
364
365} // End SubtargetPredicate = HasDLInsts
366