1 //===-- VOP3PInstructions.td - Vector Instruction Defintions --------------===//
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
7 //===----------------------------------------------------------------------===//
9 //===----------------------------------------------------------------------===//
11 //===----------------------------------------------------------------------===//
13 class VOP3PInst<string OpName, VOPProfile P, SDPatternOperator node = null_frag> :
14 VOP3P_Pseudo<OpName, P,
15 !if(P.HasModifiers, getVOP3PModPat<P, node>.ret, getVOP3Pat<P, node>.ret)
18 // Non-packed instructions that use the VOP3P encoding.
19 // VOP3 neg/abs and VOP3P opsel/opsel_hi modifiers are allowed.
20 class VOP3_VOP3PInst<string OpName, VOPProfile P, bit UseTiedOutput = 0,
21 SDPatternOperator node = null_frag> :
22 VOP3P_Pseudo<OpName, P> {
23 // These operands are only sort of f16 operands. Depending on
24 // op_sel_hi, these may be interpreted as f32. The inline immediate
25 // values are really f16 converted to f32, so we treat these as f16
30 (ins FP16InputMods:$src0_modifiers, VCSrc_f16:$src0,
31 FP16InputMods:$src1_modifiers, VCSrc_f16:$src1,
32 FP16InputMods:$src2_modifiers, VCSrc_f16:$src2,
34 !if(UseTiedOutput, (ins VGPR_32:$vdst_in), (ins))),
35 (ins op_sel:$op_sel, op_sel_hi:$op_sel_hi));
37 let Constraints = !if(UseTiedOutput, "$vdst = $vdst_in", "");
38 let DisableEncoding = !if(UseTiedOutput, "$vdst_in", "");
40 " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$op_sel$op_sel_hi$clamp";
43 let isCommutable = 1 in {
44 def V_PK_MAD_I16 : VOP3PInst<"v_pk_mad_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
45 def V_PK_MAD_U16 : VOP3PInst<"v_pk_mad_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
47 let FPDPRounding = 1 in {
48 def V_PK_FMA_F16 : VOP3PInst<"v_pk_fma_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, fma>;
49 def V_PK_ADD_F16 : VOP3PInst<"v_pk_add_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fadd>;
50 def V_PK_MUL_F16 : VOP3PInst<"v_pk_mul_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmul>;
51 } // End FPDPRounding = 1
52 def V_PK_MAX_F16 : VOP3PInst<"v_pk_max_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmaxnum_like>;
53 def V_PK_MIN_F16 : VOP3PInst<"v_pk_min_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fminnum_like>;
55 def V_PK_ADD_U16 : VOP3PInst<"v_pk_add_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, add>;
56 def V_PK_ADD_I16 : VOP3PInst<"v_pk_add_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
57 def V_PK_MUL_LO_U16 : VOP3PInst<"v_pk_mul_lo_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, mul>;
59 def V_PK_MIN_I16 : VOP3PInst<"v_pk_min_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smin>;
60 def V_PK_MIN_U16 : VOP3PInst<"v_pk_min_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umin>;
61 def V_PK_MAX_I16 : VOP3PInst<"v_pk_max_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smax>;
62 def V_PK_MAX_U16 : VOP3PInst<"v_pk_max_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umax>;
65 def V_PK_SUB_U16 : VOP3PInst<"v_pk_sub_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
66 def V_PK_SUB_I16 : VOP3PInst<"v_pk_sub_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, sub>;
68 def V_PK_LSHLREV_B16 : VOP3PInst<"v_pk_lshlrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshl_rev>;
69 def V_PK_ASHRREV_I16 : VOP3PInst<"v_pk_ashrrev_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, ashr_rev>;
70 def V_PK_LSHRREV_B16 : VOP3PInst<"v_pk_lshrrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshr_rev>;
73 // Undo sub x, c -> add x, -c canonicalization since c is more likely
74 // an inline immediate than -c.
75 // The constant will be emitted as a mov, and folded later.
76 // TODO: We could directly encode the immediate now
78 (add (v2i16 (VOP3PMods0 v2i16:$src0, i32:$src0_modifiers, i1:$clamp)), NegSubInlineConstV216:$src1),
79 (V_PK_SUB_U16 $src0_modifiers, $src0, SRCMODS.OP_SEL_1, NegSubInlineConstV216:$src1, $clamp)
82 multiclass MadFmaMixPats<SDPatternOperator fma_like,
84 Instruction mixlo_inst,
85 Instruction mixhi_inst> {
87 (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
88 (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
89 (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
90 (mixlo_inst $src0_modifiers, $src0,
91 $src1_modifiers, $src1,
92 $src2_modifiers, $src2,
97 // FIXME: Special case handling for maxhi (especially for clamp)
98 // because dealing with the write to high half of the register is
101 (build_vector f16:$elt0, (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
102 (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
103 (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
104 (v2f16 (mixhi_inst $src0_modifiers, $src0,
105 $src1_modifiers, $src1,
106 $src2_modifiers, $src2,
114 (AMDGPUclamp (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
115 (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
116 (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers)))))),
117 (v2f16 (mixhi_inst $src0_modifiers, $src0,
118 $src1_modifiers, $src1,
119 $src2_modifiers, $src2,
125 (AMDGPUclamp (build_vector
126 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$lo_src0, i32:$lo_src0_modifiers)),
127 (f32 (VOP3PMadMixMods f16:$lo_src1, i32:$lo_src1_modifiers)),
128 (f32 (VOP3PMadMixMods f16:$lo_src2, i32:$lo_src2_modifiers)))),
129 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$hi_src0, i32:$hi_src0_modifiers)),
130 (f32 (VOP3PMadMixMods f16:$hi_src1, i32:$hi_src1_modifiers)),
131 (f32 (VOP3PMadMixMods f16:$hi_src2, i32:$hi_src2_modifiers)))))),
132 (v2f16 (mixhi_inst $hi_src0_modifiers, $hi_src0,
133 $hi_src1_modifiers, $hi_src1,
134 $hi_src2_modifiers, $hi_src2,
136 (mixlo_inst $lo_src0_modifiers, $lo_src0,
137 $lo_src1_modifiers, $lo_src1,
138 $lo_src2_modifiers, $lo_src2,
140 (i32 (IMPLICIT_DEF)))))
144 let SubtargetPredicate = HasMadMixInsts in {
145 // These are VOP3a-like opcodes which accept no omod.
146 // Size of src arguments (16/32) is controlled by op_sel.
147 // For 16-bit src arguments their location (hi/lo) are controlled by op_sel_hi.
148 let isCommutable = 1 in {
149 def V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
151 let FPDPRounding = 1 in {
152 // Clamp modifier is applied after conversion to f16.
153 def V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
155 let ClampLo = 0, ClampHi = 1 in {
156 def V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
158 } // End FPDPRounding = 1
161 defm : MadFmaMixPats<fmad, V_MAD_MIX_F32, V_MAD_MIXLO_F16, V_MAD_MIXHI_F16>;
162 } // End SubtargetPredicate = HasMadMixInsts
165 // Essentially the same as the mad_mix versions
166 let SubtargetPredicate = HasFmaMixInsts in {
167 let isCommutable = 1 in {
168 def V_FMA_MIX_F32 : VOP3_VOP3PInst<"v_fma_mix_f32", VOP3_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
170 let FPDPRounding = 1 in {
171 // Clamp modifier is applied after conversion to f16.
172 def V_FMA_MIXLO_F16 : VOP3_VOP3PInst<"v_fma_mixlo_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
174 let ClampLo = 0, ClampHi = 1 in {
175 def V_FMA_MIXHI_F16 : VOP3_VOP3PInst<"v_fma_mixhi_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
177 } // End FPDPRounding = 1
180 defm : MadFmaMixPats<fma, V_FMA_MIX_F32, V_FMA_MIXLO_F16, V_FMA_MIXHI_F16>;
183 // Defines patterns that extract signed 4bit from each Idx[0].
184 foreach Idx = [[0,28],[4,24],[8,20],[12,16],[16,12],[20,8],[24,4]] in
185 def ExtractSigned4bit_#Idx[0] : PatFrag<(ops node:$src),
186 (sra (shl node:$src, (i32 Idx[1])), (i32 28))>;
188 // Defines code pattern that extracts U(unsigned/signed) 4/8bit from FromBitIndex.
189 class Extract<int FromBitIndex, int BitMask, bit U>: PatFrag<
191 !if (!or (!and (!eq (BitMask, 255), !eq (FromBitIndex, 24)), !eq (FromBitIndex, 28)), // last element
192 !if (U, (srl node:$src, (i32 FromBitIndex)), (sra node:$src, (i32 FromBitIndex))),
193 !if (!eq (FromBitIndex, 0), // first element
194 !if (U, (and node:$src, (i32 BitMask)),
195 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
196 (sext_inreg node:$src, i8))),
197 !if (U, (and (srl node:$src, (i32 FromBitIndex)), (i32 BitMask)),
198 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
199 (sext_inreg (srl node:$src, (i32 FromBitIndex)), i8)))))>;
202 foreach Type = ["I", "U"] in
203 foreach Index = 0-3 in {
204 // Defines patterns that extract each Index'ed 8bit from an unsigned
205 // 32bit scalar value;
206 def #Type#Index#"_8bit" : Extract<!shl(Index, 3), 255, !if (!eq (Type, "U"), 1, 0)>;
208 // Defines multiplication patterns where the multiplication is happening on each
209 // Index'ed 8bit of a 32bit scalar value.
211 def Mul#Type#_Elt#Index : PatFrag<
212 (ops node:$src0, node:$src1),
213 (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), AMDGPUmul_i24_oneuse, AMDGPUmul_u24_oneuse))
214 (!cast<Extract>(#Type#Index#"_8bit") node:$src0),
215 (!cast<Extract>(#Type#Index#"_8bit") node:$src1))>;
218 // Different variants of dot8 patterns cause a huge increase in the compile time.
219 // Define non-associative/commutative add/mul to prevent permutation in the dot8
221 def NonACAdd : SDNode<"ISD::ADD" , SDTIntBinOp>;
222 def NonACAdd_oneuse : HasOneUseBinOp<NonACAdd>;
224 def NonACAMDGPUmul_u24 : SDNode<"AMDGPUISD::MUL_U24" , SDTIntBinOp>;
225 def NonACAMDGPUmul_u24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_u24>;
227 def NonACAMDGPUmul_i24 : SDNode<"AMDGPUISD::MUL_I24" , SDTIntBinOp>;
228 def NonACAMDGPUmul_i24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_i24>;
230 foreach Type = ["I", "U"] in
231 foreach Index = 0-7 in {
232 // Defines patterns that extract each Index'ed 4bit from an unsigned
233 // 32bit scalar value;
234 def #Type#Index#"_4bit" : Extract<!shl(Index, 2), 15, !if (!eq (Type, "U"), 1, 0)>;
236 // Defines multiplication patterns where the multiplication is happening on each
237 // Index'ed 8bit of a 32bit scalar value.
238 def Mul#Type#Index#"_4bit" : PatFrag<
239 (ops node:$src0, node:$src1),
240 (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), NonACAMDGPUmul_i24_oneuse, NonACAMDGPUmul_u24_oneuse))
241 (!cast<Extract>(#Type#Index#"_4bit") node:$src0),
242 (!cast<Extract>(#Type#Index#"_4bit") node:$src1))>;
245 class UDot2Pat<Instruction Inst> : GCNPat <
246 (add (add_oneuse (AMDGPUmul_u24_oneuse (srl i32:$src0, (i32 16)),
247 (srl i32:$src1, (i32 16))), i32:$src2),
248 (AMDGPUmul_u24_oneuse (and i32:$src0, (i32 65535)),
249 (and i32:$src1, (i32 65535)))
251 (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
252 let SubtargetPredicate = !cast<VOP_Pseudo>(Inst).SubtargetPredicate;
255 class SDot2Pat<Instruction Inst> : GCNPat <
256 (add (add_oneuse (AMDGPUmul_i24_oneuse (sra i32:$src0, (i32 16)),
257 (sra i32:$src1, (i32 16))), i32:$src2),
258 (AMDGPUmul_i24_oneuse (sext_inreg i32:$src0, i16),
259 (sext_inreg i32:$src1, i16))),
260 (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
261 let SubtargetPredicate = !cast<VOP_Pseudo>(Inst).SubtargetPredicate;
265 let SubtargetPredicate = HasDot2Insts in {
267 def V_DOT2_F32_F16 : VOP3PInst<"v_dot2_f32_f16", VOP3_Profile<VOP_F32_V2F16_V2F16_F32>>;
268 def V_DOT2_I32_I16 : VOP3PInst<"v_dot2_i32_i16", VOP3_Profile<VOP_I32_V2I16_V2I16_I32>>;
269 def V_DOT2_U32_U16 : VOP3PInst<"v_dot2_u32_u16", VOP3_Profile<VOP_I32_V2I16_V2I16_I32>>;
270 def V_DOT4_U32_U8 : VOP3PInst<"v_dot4_u32_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
271 def V_DOT8_U32_U4 : VOP3PInst<"v_dot8_u32_u4", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
273 } // End SubtargetPredicate = HasDot2Insts
275 let SubtargetPredicate = HasDot1Insts in {
277 def V_DOT4_I32_I8 : VOP3PInst<"v_dot4_i32_i8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
278 def V_DOT8_I32_I4 : VOP3PInst<"v_dot8_i32_i4", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
280 } // End SubtargetPredicate = HasDot1Insts
281 } // End let IsDOT = 1
283 multiclass DotPats<SDPatternOperator dot_op,
284 VOP3PInst dot_inst> {
285 let SubtargetPredicate = dot_inst.SubtargetPredicate in
287 (dot_op (dot_inst.Pfl.Src0VT (VOP3PMods0 dot_inst.Pfl.Src0VT:$src0, i32:$src0_modifiers)),
288 (dot_inst.Pfl.Src1VT (VOP3PMods dot_inst.Pfl.Src1VT:$src1, i32:$src1_modifiers)),
289 (dot_inst.Pfl.Src2VT (VOP3PMods dot_inst.Pfl.Src2VT:$src2, i32:$src2_modifiers)), i1:$clamp),
290 (dot_inst $src0_modifiers, $src0, $src1_modifiers, $src1, $src2_modifiers, $src2, (as_i1imm $clamp))>;
293 defm : DotPats<AMDGPUfdot2, V_DOT2_F32_F16>;
294 defm : DotPats<int_amdgcn_sdot2, V_DOT2_I32_I16>;
295 defm : DotPats<int_amdgcn_udot2, V_DOT2_U32_U16>;
296 defm : DotPats<int_amdgcn_sdot4, V_DOT4_I32_I8>;
297 defm : DotPats<int_amdgcn_udot4, V_DOT4_U32_U8>;
298 defm : DotPats<int_amdgcn_sdot8, V_DOT8_I32_I4>;
299 defm : DotPats<int_amdgcn_udot8, V_DOT8_U32_U4>;
301 def : UDot2Pat<V_DOT2_U32_U16>;
302 def : SDot2Pat<V_DOT2_I32_I16>;
304 foreach Type = ["U", "I"] in
305 let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT4_"#Type#"32_"#Type#8).SubtargetPredicate in
307 !cast<dag>(!foldl((i32 i32:$src2), [0, 1, 2, 3], lhs, y,
308 (add_oneuse lhs, (!cast<PatFrag>("Mul"#Type#"_Elt"#y) i32:$src0, i32:$src1)))),
309 (!cast<VOP3PInst>("V_DOT4_"#Type#"32_"#Type#8) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
311 foreach Type = ["U", "I"] in
312 let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).SubtargetPredicate in
314 !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
315 [1, 2, 3, 4, 5, 6, 7], lhs, y,
316 (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
317 (!cast<VOP3PInst>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
319 // Different variants of dot8 code-gen dag patterns are not generated through table-gen due to a huge increase
320 // in the compile time. Directly handle the pattern generated by the FE here.
321 foreach Type = ["U", "I"] in
322 let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).SubtargetPredicate in
324 !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
325 [7, 1, 2, 3, 4, 5, 6], lhs, y,
326 (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
327 (!cast<VOP3PInst>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
329 def ADst_32 : VOPDstOperand<AGPR_32>;
330 def ADst_128 : VOPDstOperand<AReg_128>;
331 def ADst_512 : VOPDstOperand<AReg_512>;
332 def ADst_1024 : VOPDstOperand<AReg_1024>;
334 def VOPProfileAccRead : VOP3_Profile<VOP_I32_I32, VOP3_MAI> {
335 let Src0RC64 = ARegSrc_32;
338 def VOPProfileAccWrite : VOP3_Profile<VOP_I32_I32, VOP3_MAI> {
340 let Src0RC64 = VISrc_b32;
343 class VOPProfileMAI<VOPProfile P, RegisterOperand _SrcRC, RegisterOperand _DstRC,
344 RegisterOperand SrcABRC = AVSrc_32>
345 : VOP3_Profile<P, VOP3_MAI> {
347 let Src0RC64 = SrcABRC;
348 let Src1RC64 = SrcABRC;
349 let Src2RC64 = _SrcRC;
352 let HasModifiers = 0;
353 let Asm64 = " $vdst, $src0, $src1, $src2$cbsz$abid$blgp";
354 let Ins64 = (ins Src0RC64:$src0, Src1RC64:$src1, Src2RC64:$src2, cbsz:$cbsz, abid:$abid, blgp:$blgp);
357 def VOPProfileMAI_F32_F32_X4 : VOPProfileMAI<VOP_V4F32_F32_F32_V4F32, AISrc_128_f32, ADst_128>;
358 def VOPProfileMAI_F32_F32_X16 : VOPProfileMAI<VOP_V16F32_F32_F32_V16F32, AISrc_512_f32, ADst_512>;
359 def VOPProfileMAI_F32_F32_X32 : VOPProfileMAI<VOP_V32F32_F32_F32_V32F32, AISrc_1024_f32, ADst_1024>;
360 def VOPProfileMAI_I32_I32_X4 : VOPProfileMAI<VOP_V4I32_I32_I32_V4I32, AISrc_128_b32, ADst_128>;
361 def VOPProfileMAI_I32_I32_X16 : VOPProfileMAI<VOP_V16I32_I32_I32_V16I32, AISrc_512_b32, ADst_512>;
362 def VOPProfileMAI_I32_I32_X32 : VOPProfileMAI<VOP_V32I32_I32_I32_V32I32, AISrc_1024_b32, ADst_1024>;
363 def VOPProfileMAI_F32_V2I16_X4 : VOPProfileMAI<VOP_V4F32_V2I16_V2I16_V4F32, AISrc_128_b32, ADst_128>;
364 def VOPProfileMAI_F32_V2I16_X16 : VOPProfileMAI<VOP_V16F32_V2I16_V2I16_V16F32, AISrc_512_b32, ADst_512>;
365 def VOPProfileMAI_F32_V2I16_X32 : VOPProfileMAI<VOP_V32F32_V2I16_V2I16_V32F32, AISrc_1024_b32, ADst_1024>;
366 def VOPProfileMAI_F32_V4F16_X4 : VOPProfileMAI<VOP_V4F32_V4F16_V4F16_V4F32, AISrc_128_b32, ADst_128, AVSrc_64>;
367 def VOPProfileMAI_F32_V4F16_X16 : VOPProfileMAI<VOP_V16F32_V4F16_V4F16_V16F32, AISrc_512_b32, ADst_512, AVSrc_64>;
368 def VOPProfileMAI_F32_V4F16_X32 : VOPProfileMAI<VOP_V32F32_V4F16_V4F16_V32F32, AISrc_1024_b32, ADst_1024, AVSrc_64>;
370 let Predicates = [HasMAIInsts] in {
371 def V_ACCVGPR_READ_B32 : VOP3Inst<"v_accvgpr_read_b32", VOPProfileAccRead>;
372 def V_ACCVGPR_WRITE_B32 : VOP3Inst<"v_accvgpr_write_b32", VOPProfileAccWrite> {
376 let isConvergent = 1 in {
377 def V_MFMA_F32_4X4X1F32 : VOP3Inst<"v_mfma_f32_4x4x1f32", VOPProfileMAI_F32_F32_X4, int_amdgcn_mfma_f32_4x4x1f32>;
378 def V_MFMA_F32_4X4X4F16 : VOP3Inst<"v_mfma_f32_4x4x4f16", VOPProfileMAI_F32_V4F16_X4, int_amdgcn_mfma_f32_4x4x4f16>;
379 def V_MFMA_I32_4X4X4I8 : VOP3Inst<"v_mfma_i32_4x4x4i8", VOPProfileMAI_I32_I32_X4, int_amdgcn_mfma_i32_4x4x4i8>;
380 def V_MFMA_F32_4X4X2BF16 : VOP3Inst<"v_mfma_f32_4x4x2bf16", VOPProfileMAI_F32_V2I16_X4, int_amdgcn_mfma_f32_4x4x2bf16>;
381 def V_MFMA_F32_16X16X1F32 : VOP3Inst<"v_mfma_f32_16x16x1f32", VOPProfileMAI_F32_F32_X16, int_amdgcn_mfma_f32_16x16x1f32>;
382 def V_MFMA_F32_16X16X4F32 : VOP3Inst<"v_mfma_f32_16x16x4f32", VOPProfileMAI_F32_F32_X4, int_amdgcn_mfma_f32_16x16x4f32>;
383 def V_MFMA_F32_16X16X4F16 : VOP3Inst<"v_mfma_f32_16x16x4f16", VOPProfileMAI_F32_V4F16_X16, int_amdgcn_mfma_f32_16x16x4f16>;
384 def V_MFMA_F32_16X16X16F16 : VOP3Inst<"v_mfma_f32_16x16x16f16", VOPProfileMAI_F32_V4F16_X4, int_amdgcn_mfma_f32_16x16x16f16>;
385 def V_MFMA_I32_16X16X4I8 : VOP3Inst<"v_mfma_i32_16x16x4i8", VOPProfileMAI_I32_I32_X16, int_amdgcn_mfma_i32_16x16x4i8>;
386 def V_MFMA_I32_16X16X16I8 : VOP3Inst<"v_mfma_i32_16x16x16i8", VOPProfileMAI_I32_I32_X4, int_amdgcn_mfma_i32_16x16x16i8>;
387 def V_MFMA_F32_16X16X2BF16 : VOP3Inst<"v_mfma_f32_16x16x2bf16", VOPProfileMAI_F32_V2I16_X16, int_amdgcn_mfma_f32_16x16x2bf16>;
388 def V_MFMA_F32_16X16X8BF16 : VOP3Inst<"v_mfma_f32_16x16x8bf16", VOPProfileMAI_F32_V2I16_X4, int_amdgcn_mfma_f32_16x16x8bf16>;
389 def V_MFMA_F32_32X32X1F32 : VOP3Inst<"v_mfma_f32_32x32x1f32", VOPProfileMAI_F32_F32_X32, int_amdgcn_mfma_f32_32x32x1f32>;
390 def V_MFMA_F32_32X32X2F32 : VOP3Inst<"v_mfma_f32_32x32x2f32", VOPProfileMAI_F32_F32_X16, int_amdgcn_mfma_f32_32x32x2f32>;
391 def V_MFMA_F32_32X32X4F16 : VOP3Inst<"v_mfma_f32_32x32x4f16", VOPProfileMAI_F32_V4F16_X32, int_amdgcn_mfma_f32_32x32x4f16>;
392 def V_MFMA_F32_32X32X8F16 : VOP3Inst<"v_mfma_f32_32x32x8f16", VOPProfileMAI_F32_V4F16_X16, int_amdgcn_mfma_f32_32x32x8f16>;
393 def V_MFMA_I32_32X32X4I8 : VOP3Inst<"v_mfma_i32_32x32x4i8", VOPProfileMAI_I32_I32_X32, int_amdgcn_mfma_i32_32x32x4i8>;
394 def V_MFMA_I32_32X32X8I8 : VOP3Inst<"v_mfma_i32_32x32x8i8", VOPProfileMAI_I32_I32_X16, int_amdgcn_mfma_i32_32x32x8i8>;
395 def V_MFMA_F32_32X32X2BF16 : VOP3Inst<"v_mfma_f32_32x32x2bf16", VOPProfileMAI_F32_V2I16_X32, int_amdgcn_mfma_f32_32x32x2bf16>;
396 def V_MFMA_F32_32X32X4BF16 : VOP3Inst<"v_mfma_f32_32x32x4bf16", VOPProfileMAI_F32_V2I16_X16, int_amdgcn_mfma_f32_32x32x4bf16>;
397 } // End isConvergent = 1
399 } // End SubtargetPredicate = HasMAIInsts
401 def : MnemonicAlias<"v_accvgpr_read", "v_accvgpr_read_b32">;
402 def : MnemonicAlias<"v_accvgpr_write", "v_accvgpr_write_b32">;
404 multiclass VOP3P_Real_vi<bits<10> op> {
405 def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
406 VOP3Pe <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
407 let AssemblerPredicates = [HasVOP3PInsts];
408 let DecoderNamespace = "GFX8";
412 multiclass VOP3P_Real_MAI<bits<10> op> {
413 def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
414 VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
415 let AssemblerPredicates = [HasMAIInsts];
416 let DecoderNamespace = "GFX8";
420 defm V_PK_MAD_I16 : VOP3P_Real_vi <0x380>;
421 defm V_PK_MUL_LO_U16 : VOP3P_Real_vi <0x381>;
422 defm V_PK_ADD_I16 : VOP3P_Real_vi <0x382>;
423 defm V_PK_SUB_I16 : VOP3P_Real_vi <0x383>;
424 defm V_PK_LSHLREV_B16 : VOP3P_Real_vi <0x384>;
425 defm V_PK_LSHRREV_B16 : VOP3P_Real_vi <0x385>;
426 defm V_PK_ASHRREV_I16 : VOP3P_Real_vi <0x386>;
427 defm V_PK_MAX_I16 : VOP3P_Real_vi <0x387>;
428 defm V_PK_MIN_I16 : VOP3P_Real_vi <0x388>;
429 defm V_PK_MAD_U16 : VOP3P_Real_vi <0x389>;
431 defm V_PK_ADD_U16 : VOP3P_Real_vi <0x38a>;
432 defm V_PK_SUB_U16 : VOP3P_Real_vi <0x38b>;
433 defm V_PK_MAX_U16 : VOP3P_Real_vi <0x38c>;
434 defm V_PK_MIN_U16 : VOP3P_Real_vi <0x38d>;
435 defm V_PK_FMA_F16 : VOP3P_Real_vi <0x38e>;
436 defm V_PK_ADD_F16 : VOP3P_Real_vi <0x38f>;
437 defm V_PK_MUL_F16 : VOP3P_Real_vi <0x390>;
438 defm V_PK_MIN_F16 : VOP3P_Real_vi <0x391>;
439 defm V_PK_MAX_F16 : VOP3P_Real_vi <0x392>;
442 let SubtargetPredicate = HasMadMixInsts in {
443 defm V_MAD_MIX_F32 : VOP3P_Real_vi <0x3a0>;
444 defm V_MAD_MIXLO_F16 : VOP3P_Real_vi <0x3a1>;
445 defm V_MAD_MIXHI_F16 : VOP3P_Real_vi <0x3a2>;
448 let SubtargetPredicate = HasFmaMixInsts in {
449 let DecoderNamespace = "GFX9_DL" in {
450 // The mad_mix instructions were renamed and their behaviors changed,
451 // but the opcode stayed the same so we need to put these in a
452 // different DecoderNamespace to avoid the ambiguity.
453 defm V_FMA_MIX_F32 : VOP3P_Real_vi <0x3a0>;
454 defm V_FMA_MIXLO_F16 : VOP3P_Real_vi <0x3a1>;
455 defm V_FMA_MIXHI_F16 : VOP3P_Real_vi <0x3a2>;
460 let SubtargetPredicate = HasDot2Insts in {
462 defm V_DOT2_F32_F16 : VOP3P_Real_vi <0x3a3>;
463 defm V_DOT2_I32_I16 : VOP3P_Real_vi <0x3a6>;
464 defm V_DOT2_U32_U16 : VOP3P_Real_vi <0x3a7>;
465 defm V_DOT4_U32_U8 : VOP3P_Real_vi <0x3a9>;
466 defm V_DOT8_U32_U4 : VOP3P_Real_vi <0x3ab>;
468 } // End SubtargetPredicate = HasDot2Insts
470 let SubtargetPredicate = HasDot1Insts in {
472 defm V_DOT4_I32_I8 : VOP3P_Real_vi <0x3a8>;
473 defm V_DOT8_I32_I4 : VOP3P_Real_vi <0x3aa>;
475 } // End SubtargetPredicate = HasDot1Insts
477 let SubtargetPredicate = HasMAIInsts in {
479 defm V_ACCVGPR_READ_B32 : VOP3P_Real_MAI <0x3d8>;
480 defm V_ACCVGPR_WRITE_B32 : VOP3P_Real_MAI <0x3d9>;
481 defm V_MFMA_F32_32X32X1F32 : VOP3P_Real_MAI <0x3c0>;
482 defm V_MFMA_F32_16X16X1F32 : VOP3P_Real_MAI <0x3c1>;
483 defm V_MFMA_F32_4X4X1F32 : VOP3P_Real_MAI <0x3c2>;
484 defm V_MFMA_F32_32X32X2F32 : VOP3P_Real_MAI <0x3c4>;
485 defm V_MFMA_F32_16X16X4F32 : VOP3P_Real_MAI <0x3c5>;
486 defm V_MFMA_F32_32X32X4F16 : VOP3P_Real_MAI <0x3c8>;
487 defm V_MFMA_F32_16X16X4F16 : VOP3P_Real_MAI <0x3c9>;
488 defm V_MFMA_F32_4X4X4F16 : VOP3P_Real_MAI <0x3ca>;
489 defm V_MFMA_F32_32X32X8F16 : VOP3P_Real_MAI <0x3cc>;
490 defm V_MFMA_F32_16X16X16F16 : VOP3P_Real_MAI <0x3cd>;
491 defm V_MFMA_I32_32X32X4I8 : VOP3P_Real_MAI <0x3d0>;
492 defm V_MFMA_I32_16X16X4I8 : VOP3P_Real_MAI <0x3d1>;
493 defm V_MFMA_I32_4X4X4I8 : VOP3P_Real_MAI <0x3d2>;
494 defm V_MFMA_I32_32X32X8I8 : VOP3P_Real_MAI <0x3d4>;
495 defm V_MFMA_I32_16X16X16I8 : VOP3P_Real_MAI <0x3d5>;
496 defm V_MFMA_F32_32X32X2BF16 : VOP3P_Real_MAI <0x3e8>;
497 defm V_MFMA_F32_16X16X2BF16 : VOP3P_Real_MAI <0x3e9>;
498 defm V_MFMA_F32_4X4X2BF16 : VOP3P_Real_MAI <0x3eb>;
499 defm V_MFMA_F32_32X32X4BF16 : VOP3P_Real_MAI <0x3ec>;
500 defm V_MFMA_F32_16X16X8BF16 : VOP3P_Real_MAI <0x3ed>;
502 } // End SubtargetPredicate = HasMAIInsts
504 //===----------------------------------------------------------------------===//
506 //===----------------------------------------------------------------------===//
508 let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in {
509 multiclass VOP3P_Real_gfx10<bits<10> op> {
510 def _gfx10 : VOP3P_Real<!cast<VOP3P_Pseudo>(NAME), SIEncodingFamily.GFX10>,
511 VOP3Pe_gfx10 <op, !cast<VOP3P_Pseudo>(NAME).Pfl>;
513 } // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10"
515 defm V_PK_MAD_I16 : VOP3P_Real_gfx10<0x000>;
516 defm V_PK_MUL_LO_U16 : VOP3P_Real_gfx10<0x001>;
517 defm V_PK_ADD_I16 : VOP3P_Real_gfx10<0x002>;
518 defm V_PK_SUB_I16 : VOP3P_Real_gfx10<0x003>;
519 defm V_PK_LSHLREV_B16 : VOP3P_Real_gfx10<0x004>;
520 defm V_PK_LSHRREV_B16 : VOP3P_Real_gfx10<0x005>;
521 defm V_PK_ASHRREV_I16 : VOP3P_Real_gfx10<0x006>;
522 defm V_PK_MAX_I16 : VOP3P_Real_gfx10<0x007>;
523 defm V_PK_MIN_I16 : VOP3P_Real_gfx10<0x008>;
524 defm V_PK_MAD_U16 : VOP3P_Real_gfx10<0x009>;
525 defm V_PK_ADD_U16 : VOP3P_Real_gfx10<0x00a>;
526 defm V_PK_SUB_U16 : VOP3P_Real_gfx10<0x00b>;
527 defm V_PK_MAX_U16 : VOP3P_Real_gfx10<0x00c>;
528 defm V_PK_MIN_U16 : VOP3P_Real_gfx10<0x00d>;
529 defm V_PK_FMA_F16 : VOP3P_Real_gfx10<0x00e>;
530 defm V_PK_ADD_F16 : VOP3P_Real_gfx10<0x00f>;
531 defm V_PK_MUL_F16 : VOP3P_Real_gfx10<0x010>;
532 defm V_PK_MIN_F16 : VOP3P_Real_gfx10<0x011>;
533 defm V_PK_MAX_F16 : VOP3P_Real_gfx10<0x012>;
534 defm V_FMA_MIX_F32 : VOP3P_Real_gfx10<0x020>;
535 defm V_FMA_MIXLO_F16 : VOP3P_Real_gfx10<0x021>;
536 defm V_FMA_MIXHI_F16 : VOP3P_Real_gfx10<0x022>;
538 let SubtargetPredicate = HasDot2Insts in {
540 defm V_DOT2_F32_F16 : VOP3P_Real_gfx10 <0x013>;
541 defm V_DOT2_I32_I16 : VOP3P_Real_gfx10 <0x014>;
542 defm V_DOT2_U32_U16 : VOP3P_Real_gfx10 <0x015>;
543 defm V_DOT4_U32_U8 : VOP3P_Real_gfx10 <0x017>;
544 defm V_DOT8_U32_U4 : VOP3P_Real_gfx10 <0x019>;
546 } // End SubtargetPredicate = HasDot2Insts
548 let SubtargetPredicate = HasDot1Insts in {
550 defm V_DOT4_I32_I8 : VOP3P_Real_gfx10 <0x016>;
551 defm V_DOT8_I32_I4 : VOP3P_Real_gfx10 <0x018>;
553 } // End SubtargetPredicate = HasDot1Insts