xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/VOP3PInstructions.td (revision 77013d11e6483b970af25e13c9b892075742f7e5)
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 VOP3PInst<string OpName, VOPProfile P,
14                SDPatternOperator node = null_frag,
15                bit HasExplicitClamp = 0> :
16  VOP3P_Pseudo<OpName, P,
17    !if(P.HasModifiers, getVOP3PModPat<P, node, HasExplicitClamp>.ret, getVOP3Pat<P, node>.ret)
18>;
19
20// Non-packed instructions that use the VOP3P encoding.
21// VOP3 neg/abs and VOP3P opsel/opsel_hi modifiers are allowed.
22class VOP3_VOP3PInst<string OpName, VOPProfile P, bit UseTiedOutput = 0,
23                     SDPatternOperator node = null_frag> :
24  VOP3P_Pseudo<OpName, P> {
25  // These operands are only sort of f16 operands. Depending on
26  // op_sel_hi, these may be interpreted as f32. The inline immediate
27  // values are really f16 converted to f32, so we treat these as f16
28  // operands.
29  let InOperandList =
30    !con(
31      !con(
32        (ins FP16InputMods:$src0_modifiers, VCSrc_f16:$src0,
33             FP16InputMods:$src1_modifiers, VCSrc_f16:$src1,
34             FP16InputMods:$src2_modifiers, VCSrc_f16:$src2),
35         // FIXME: clampmod0 misbehaves with the non-default vdst_in
36         // following it. For now workaround this by requiring clamp
37         // in tied patterns. This should use undef_tied_input, but it
38         // seems underdeveloped and doesn't apply the right register
39         // class constraints.
40         !if(UseTiedOutput, (ins clampmod:$clamp, VGPR_32:$vdst_in),
41                            (ins clampmod0:$clamp))),
42         (ins op_sel0:$op_sel, op_sel_hi0:$op_sel_hi));
43
44  let Constraints = !if(UseTiedOutput, "$vdst = $vdst_in", "");
45  let DisableEncoding = !if(UseTiedOutput, "$vdst_in", "");
46  let AsmOperands =
47    " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$op_sel$op_sel_hi$clamp";
48}
49
50let isCommutable = 1 in {
51def V_PK_MAD_I16 : VOP3PInst<"v_pk_mad_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
52def V_PK_MAD_U16 : VOP3PInst<"v_pk_mad_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
53
54let FPDPRounding = 1 in {
55def V_PK_FMA_F16 : VOP3PInst<"v_pk_fma_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, any_fma>;
56def V_PK_ADD_F16 : VOP3PInst<"v_pk_add_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, any_fadd>;
57def V_PK_MUL_F16 : VOP3PInst<"v_pk_mul_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, any_fmul>;
58} // End FPDPRounding = 1
59def V_PK_MAX_F16 : VOP3PInst<"v_pk_max_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmaxnum_like>;
60def V_PK_MIN_F16 : VOP3PInst<"v_pk_min_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fminnum_like>;
61
62def V_PK_ADD_U16 : VOP3PInst<"v_pk_add_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, add>;
63def V_PK_ADD_I16 : VOP3PInst<"v_pk_add_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
64def V_PK_MUL_LO_U16 : VOP3PInst<"v_pk_mul_lo_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, mul>;
65
66def V_PK_MIN_I16 : VOP3PInst<"v_pk_min_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smin>;
67def V_PK_MIN_U16 : VOP3PInst<"v_pk_min_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umin>;
68def V_PK_MAX_I16 : VOP3PInst<"v_pk_max_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smax>;
69def V_PK_MAX_U16 : VOP3PInst<"v_pk_max_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umax>;
70}
71
72def V_PK_SUB_U16 : VOP3PInst<"v_pk_sub_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
73def V_PK_SUB_I16 : VOP3PInst<"v_pk_sub_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, sub>;
74
75def V_PK_LSHLREV_B16 : VOP3PInst<"v_pk_lshlrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshl_rev>;
76def V_PK_ASHRREV_I16 : VOP3PInst<"v_pk_ashrrev_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, ashr_rev>;
77def V_PK_LSHRREV_B16 : VOP3PInst<"v_pk_lshrrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshr_rev>;
78
79
80let SubtargetPredicate = HasVOP3PInsts in {
81
82// Undo sub x, c -> add x, -c canonicalization since c is more likely
83// an inline immediate than -c.
84// The constant will be emitted as a mov, and folded later.
85// TODO: We could directly encode the immediate now
86def : GCNPat<
87  (add (v2i16 (VOP3PMods v2i16:$src0, i32:$src0_modifiers)), NegSubInlineConstV216:$src1),
88  (V_PK_SUB_U16 $src0_modifiers, $src0, SRCMODS.OP_SEL_1, NegSubInlineConstV216:$src1)
89>;
90
91// Integer operations with clamp bit set.
92class VOP3PSatPat<SDPatternOperator pat, Instruction inst> : GCNPat<
93  (pat (v2i16 (VOP3PMods v2i16:$src0, i32:$src0_modifiers)),
94       (v2i16 (VOP3PMods v2i16:$src1, i32:$src1_modifiers))),
95  (inst $src0_modifiers, $src0, $src1_modifiers, $src1, DSTCLAMP.ENABLE)
96>;
97
98def : VOP3PSatPat<uaddsat, V_PK_ADD_U16>;
99def : VOP3PSatPat<saddsat, V_PK_ADD_I16>;
100def : VOP3PSatPat<usubsat, V_PK_SUB_U16>;
101def : VOP3PSatPat<ssubsat, V_PK_SUB_I16>;
102} // End SubtargetPredicate = HasVOP3PInsts
103
104multiclass MadFmaMixPats<SDPatternOperator fma_like,
105                         Instruction mix_inst,
106                         Instruction mixlo_inst,
107                         Instruction mixhi_inst> {
108  def : GCNPat <
109    (f16 (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
110                            (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
111                            (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
112    (mixlo_inst $src0_modifiers, $src0,
113                $src1_modifiers, $src1,
114                $src2_modifiers, $src2,
115                DSTCLAMP.NONE,
116                (i32 (IMPLICIT_DEF)))
117  >;
118
119  // FIXME: Special case handling for maxhi (especially for clamp)
120  // because dealing with the write to high half of the register is
121  // difficult.
122  def : GCNPat <
123    (build_vector f16:$elt0, (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
124                                                (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
125                                                (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
126    (v2f16 (mixhi_inst $src0_modifiers, $src0,
127                       $src1_modifiers, $src1,
128                       $src2_modifiers, $src2,
129                       DSTCLAMP.NONE,
130                       $elt0))
131  >;
132
133  def : GCNPat <
134    (build_vector
135      f16:$elt0,
136      (AMDGPUclamp (fpround (fma_like (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
137                                      (f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
138                                      (f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers)))))),
139    (v2f16 (mixhi_inst $src0_modifiers, $src0,
140                       $src1_modifiers, $src1,
141                       $src2_modifiers, $src2,
142                       DSTCLAMP.ENABLE,
143                       $elt0))
144  >;
145
146  def : GCNPat <
147    (AMDGPUclamp (build_vector
148      (fpround (fma_like (f32 (VOP3PMadMixMods f16:$lo_src0, i32:$lo_src0_modifiers)),
149                         (f32 (VOP3PMadMixMods f16:$lo_src1, i32:$lo_src1_modifiers)),
150                         (f32 (VOP3PMadMixMods f16:$lo_src2, i32:$lo_src2_modifiers)))),
151      (fpround (fma_like (f32 (VOP3PMadMixMods f16:$hi_src0, i32:$hi_src0_modifiers)),
152                         (f32 (VOP3PMadMixMods f16:$hi_src1, i32:$hi_src1_modifiers)),
153                         (f32 (VOP3PMadMixMods f16:$hi_src2, i32:$hi_src2_modifiers)))))),
154    (v2f16 (mixhi_inst $hi_src0_modifiers, $hi_src0,
155                       $hi_src1_modifiers, $hi_src1,
156                       $hi_src2_modifiers, $hi_src2,
157                       DSTCLAMP.ENABLE,
158                       (mixlo_inst $lo_src0_modifiers, $lo_src0,
159                                   $lo_src1_modifiers, $lo_src1,
160                                   $lo_src2_modifiers, $lo_src2,
161                                   DSTCLAMP.ENABLE,
162                                   (i32 (IMPLICIT_DEF)))))
163  >;
164}
165
166let SubtargetPredicate = HasMadMixInsts in {
167
168// These are VOP3a-like opcodes which accept no omod.
169// Size of src arguments (16/32) is controlled by op_sel.
170// For 16-bit src arguments their location (hi/lo) are controlled by op_sel_hi.
171let isCommutable = 1, mayRaiseFPException = 0 in {
172def V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
173
174let FPDPRounding = 1 in {
175// Clamp modifier is applied after conversion to f16.
176def V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
177
178let ClampLo = 0, ClampHi = 1 in {
179def V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
180}
181} // End FPDPRounding = 1
182}
183
184defm : MadFmaMixPats<fmad, V_MAD_MIX_F32, V_MAD_MIXLO_F16, V_MAD_MIXHI_F16>;
185} // End SubtargetPredicate = HasMadMixInsts
186
187
188// Essentially the same as the mad_mix versions
189let SubtargetPredicate = HasFmaMixInsts in {
190let isCommutable = 1 in {
191def V_FMA_MIX_F32 : VOP3_VOP3PInst<"v_fma_mix_f32", VOP3_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
192
193let FPDPRounding = 1 in {
194// Clamp modifier is applied after conversion to f16.
195def V_FMA_MIXLO_F16 : VOP3_VOP3PInst<"v_fma_mixlo_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
196
197let ClampLo = 0, ClampHi = 1 in {
198def V_FMA_MIXHI_F16 : VOP3_VOP3PInst<"v_fma_mixhi_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
199}
200} // End FPDPRounding = 1
201}
202
203defm : MadFmaMixPats<fma, V_FMA_MIX_F32, V_FMA_MIXLO_F16, V_FMA_MIXHI_F16>;
204}
205
206// Defines patterns that extract signed 4bit from each Idx[0].
207foreach Idx = [[0,28],[4,24],[8,20],[12,16],[16,12],[20,8],[24,4]] in
208  def ExtractSigned4bit_#Idx[0] : PatFrag<(ops node:$src),
209                                          (sra (shl node:$src, (i32 Idx[1])), (i32 28))>;
210
211// Defines code pattern that extracts U(unsigned/signed) 4/8bit from FromBitIndex.
212class Extract<int FromBitIndex, int BitMask, bit U>: PatFrag<
213  (ops node:$src),
214  !if (!or (!and (!eq (BitMask, 255), !eq (FromBitIndex, 24)), !eq (FromBitIndex, 28)), // last element
215       !if (U, (srl node:$src, (i32 FromBitIndex)), (sra node:$src, (i32 FromBitIndex))),
216       !if (!eq (FromBitIndex, 0), // first element
217            !if (U, (and node:$src, (i32 BitMask)),
218                 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
219                                         (sext_inreg node:$src, i8))),
220            !if (U, (and (srl node:$src, (i32 FromBitIndex)), (i32 BitMask)),
221                 !if (!eq (BitMask, 15), (!cast<PatFrag>("ExtractSigned4bit_"#FromBitIndex) node:$src),
222                      (sext_inreg (srl node:$src, (i32 FromBitIndex)), i8)))))>;
223
224
225foreach Type = ["I", "U"] in
226  foreach Index = 0-3 in {
227    // Defines patterns that extract each Index'ed 8bit from an unsigned
228    // 32bit scalar value;
229    def Type#Index#"_8bit" : Extract<!shl(Index, 3), 255, !eq (Type, "U")>;
230
231    // Defines multiplication patterns where the multiplication is happening on each
232    // Index'ed 8bit of a 32bit scalar value.
233
234    def Mul#Type#_Elt#Index : PatFrag<
235      (ops node:$src0, node:$src1),
236      (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), AMDGPUmul_i24_oneuse, AMDGPUmul_u24_oneuse))
237                            (!cast<Extract>(Type#Index#"_8bit") node:$src0),
238                            (!cast<Extract>(Type#Index#"_8bit") node:$src1))>;
239  }
240
241// Different variants of dot8 patterns cause a huge increase in the compile time.
242// Define non-associative/commutative add/mul to prevent permutation in the dot8
243// pattern.
244def NonACAdd        : SDNode<"ISD::ADD"       , SDTIntBinOp>;
245def NonACAdd_oneuse : HasOneUseBinOp<NonACAdd>;
246
247def NonACAMDGPUmul_u24        : SDNode<"AMDGPUISD::MUL_U24"       , SDTIntBinOp>;
248def NonACAMDGPUmul_u24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_u24>;
249
250def NonACAMDGPUmul_i24        : SDNode<"AMDGPUISD::MUL_I24"       , SDTIntBinOp>;
251def NonACAMDGPUmul_i24_oneuse : HasOneUseBinOp<NonACAMDGPUmul_i24>;
252
253foreach Type = ["I", "U"] in
254  foreach Index = 0-7 in {
255    // Defines patterns that extract each Index'ed 4bit from an unsigned
256    // 32bit scalar value;
257    def Type#Index#"_4bit" : Extract<!shl(Index, 2), 15, !eq (Type, "U")>;
258
259    // Defines multiplication patterns where the multiplication is happening on each
260    // Index'ed 8bit of a 32bit scalar value.
261    def Mul#Type#Index#"_4bit" : PatFrag<
262      (ops node:$src0, node:$src1),
263      (!cast<HasOneUseBinOp>(!if (!eq (Type, "I"), NonACAMDGPUmul_i24_oneuse, NonACAMDGPUmul_u24_oneuse))
264                             (!cast<Extract>(Type#Index#"_4bit") node:$src0),
265                             (!cast<Extract>(Type#Index#"_4bit") node:$src1))>;
266  }
267
268class UDot2Pat<Instruction Inst> : GCNPat <
269  (add (add_oneuse (AMDGPUmul_u24_oneuse (srl i32:$src0, (i32 16)),
270                                         (srl i32:$src1, (i32 16))), i32:$src2),
271       (AMDGPUmul_u24_oneuse (and i32:$src0, (i32 65535)),
272                             (and i32:$src1, (i32 65535)))
273   ),
274  (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
275  let SubtargetPredicate = !cast<VOP_Pseudo>(Inst).SubtargetPredicate;
276}
277
278class SDot2Pat<Instruction Inst> : GCNPat <
279  (add (add_oneuse (AMDGPUmul_i24_oneuse (sra i32:$src0, (i32 16)),
280                                         (sra i32:$src1, (i32 16))), i32:$src2),
281       (AMDGPUmul_i24_oneuse (sext_inreg i32:$src0, i16),
282                             (sext_inreg i32:$src1, i16))),
283  (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
284  let SubtargetPredicate = !cast<VOP_Pseudo>(Inst).SubtargetPredicate;
285}
286
287let IsDOT = 1 in {
288let SubtargetPredicate = HasDot2Insts in {
289
290def V_DOT2_F32_F16 : VOP3PInst<"v_dot2_f32_f16",
291  VOP3_Profile<VOP_F32_V2F16_V2F16_F32>,
292  AMDGPUfdot2, 1/*ExplicitClamp*/>;
293def V_DOT2_I32_I16 : VOP3PInst<"v_dot2_i32_i16",
294  VOP3_Profile<VOP_I32_V2I16_V2I16_I32>, int_amdgcn_sdot2, 1>;
295def V_DOT2_U32_U16 : VOP3PInst<"v_dot2_u32_u16",
296  VOP3_Profile<VOP_I32_V2I16_V2I16_I32>, int_amdgcn_udot2, 1>;
297def V_DOT4_U32_U8  : VOP3PInst<"v_dot4_u32_u8",
298  VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_udot4, 1>;
299def V_DOT8_U32_U4  : VOP3PInst<"v_dot8_u32_u4",
300  VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_udot8, 1>;
301
302} // End SubtargetPredicate = HasDot2Insts
303
304let SubtargetPredicate = HasDot1Insts in {
305
306def V_DOT4_I32_I8  : VOP3PInst<"v_dot4_i32_i8",
307  VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_sdot4, 1>;
308def V_DOT8_I32_I4  : VOP3PInst<"v_dot8_i32_i4",
309  VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>, int_amdgcn_sdot8, 1>;
310
311} // End SubtargetPredicate = HasDot1Insts
312} // End let IsDOT = 1
313
314def : UDot2Pat<V_DOT2_U32_U16>;
315def : SDot2Pat<V_DOT2_I32_I16>;
316
317foreach Type = ["U", "I"] in
318  let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT4_"#Type#"32_"#Type#8).SubtargetPredicate in
319  def : GCNPat <
320    !cast<dag>(!foldl((i32 i32:$src2), [0, 1, 2, 3], lhs, y,
321                      (add_oneuse lhs, (!cast<PatFrag>("Mul"#Type#"_Elt"#y) i32:$src0, i32:$src1)))),
322    (!cast<VOP3PInst>("V_DOT4_"#Type#"32_"#Type#8) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
323
324foreach Type = ["U", "I"] in
325  let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).SubtargetPredicate in
326  def : GCNPat <
327    !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
328                      [1, 2, 3, 4, 5, 6, 7], lhs, y,
329                      (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
330    (!cast<VOP3PInst>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
331
332// Different variants of dot8 code-gen dag patterns are not generated through table-gen due to a huge increase
333// in the compile time. Directly handle the pattern generated by the FE here.
334foreach Type = ["U", "I"] in
335  let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).SubtargetPredicate in
336  def : GCNPat <
337    !cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
338                      [7, 1, 2, 3, 4, 5, 6], lhs, y,
339                      (NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
340    (!cast<VOP3PInst>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
341
342def ADst_32   : VOPDstOperand<AGPR_32>;
343def ADst_128  : VOPDstOperand<AReg_128>;
344def ADst_512  : VOPDstOperand<AReg_512>;
345def ADst_1024 : VOPDstOperand<AReg_1024>;
346
347def VOPProfileAccRead : VOP3_Profile<VOP_I32_I32, VOP3_MAI> {
348  let Src0RC64 = ARegSrc_32;
349}
350
351def VOPProfileAccWrite : VOP3_Profile<VOP_I32_I32, VOP3_MAI> {
352  let DstRC = ADst_32;
353  let Src0RC64 = VISrc_b32;
354}
355
356class VOPProfileMAI<VOPProfile P, RegisterOperand _SrcRC, RegisterOperand _DstRC,
357                    RegisterOperand SrcABRC = AVSrc_32>
358  : VOP3_Profile<P, VOP3_MAI> {
359  let DstRC = _DstRC;
360  let Src0RC64 = SrcABRC;
361  let Src1RC64 = SrcABRC;
362  let Src2RC64 = _SrcRC;
363  let HasOpSel = 0;
364  let HasClamp = 0;
365  let Asm64 = " $vdst, $src0, $src1, $src2$cbsz$abid$blgp";
366  let Ins64 = (ins Src0RC64:$src0, Src1RC64:$src1, Src2RC64:$src2, cbsz:$cbsz, abid:$abid, blgp:$blgp);
367}
368
369def VOPProfileMAI_F32_F32_X4    : VOPProfileMAI<VOP_V4F32_F32_F32_V4F32,       AISrc_128_f32,  ADst_128>;
370def VOPProfileMAI_F32_F32_X16   : VOPProfileMAI<VOP_V16F32_F32_F32_V16F32,     AISrc_512_f32,  ADst_512>;
371def VOPProfileMAI_F32_F32_X32   : VOPProfileMAI<VOP_V32F32_F32_F32_V32F32,     AISrc_1024_f32, ADst_1024>;
372def VOPProfileMAI_I32_I32_X4    : VOPProfileMAI<VOP_V4I32_I32_I32_V4I32,       AISrc_128_b32,  ADst_128>;
373def VOPProfileMAI_I32_I32_X16   : VOPProfileMAI<VOP_V16I32_I32_I32_V16I32,     AISrc_512_b32,  ADst_512>;
374def VOPProfileMAI_I32_I32_X32   : VOPProfileMAI<VOP_V32I32_I32_I32_V32I32,     AISrc_1024_b32, ADst_1024>;
375def VOPProfileMAI_F32_V2I16_X4  : VOPProfileMAI<VOP_V4F32_V2I16_V2I16_V4F32,   AISrc_128_b32,  ADst_128>;
376def VOPProfileMAI_F32_V2I16_X16 : VOPProfileMAI<VOP_V16F32_V2I16_V2I16_V16F32, AISrc_512_b32,  ADst_512>;
377def VOPProfileMAI_F32_V2I16_X32 : VOPProfileMAI<VOP_V32F32_V2I16_V2I16_V32F32, AISrc_1024_b32, ADst_1024>;
378def VOPProfileMAI_F32_V4F16_X4  : VOPProfileMAI<VOP_V4F32_V4F16_V4F16_V4F32,   AISrc_128_b32,  ADst_128,  AVSrc_64>;
379def VOPProfileMAI_F32_V4F16_X16 : VOPProfileMAI<VOP_V16F32_V4F16_V4F16_V16F32, AISrc_512_b32,  ADst_512,  AVSrc_64>;
380def VOPProfileMAI_F32_V4F16_X32 : VOPProfileMAI<VOP_V32F32_V4F16_V4F16_V32F32, AISrc_1024_b32, ADst_1024, AVSrc_64>;
381
382let Predicates = [HasMAIInsts] in {
383
384let isAsCheapAsAMove = 1, isReMaterializable = 1 in {
385  defm V_ACCVGPR_READ_B32  : VOP3Inst<"v_accvgpr_read_b32",  VOPProfileAccRead>;
386  let isMoveImm = 1 in {
387    defm V_ACCVGPR_WRITE_B32 : VOP3Inst<"v_accvgpr_write_b32", VOPProfileAccWrite>;
388  } // End isMoveImm = 1
389} // End isAsCheapAsAMove = 1, isReMaterializable = 1
390
391// FP32 denorm mode is respected, rounding mode is not. Exceptions are not supported.
392let isConvergent = 1, mayRaiseFPException = 0, ReadsModeReg = 1 in {
393defm V_MFMA_F32_4X4X1F32    : VOP3Inst<"v_mfma_f32_4x4x1f32",    VOPProfileMAI_F32_F32_X4,    int_amdgcn_mfma_f32_4x4x1f32>;
394defm V_MFMA_F32_4X4X4F16    : VOP3Inst<"v_mfma_f32_4x4x4f16",    VOPProfileMAI_F32_V4F16_X4,  int_amdgcn_mfma_f32_4x4x4f16>;
395defm V_MFMA_I32_4X4X4I8     : VOP3Inst<"v_mfma_i32_4x4x4i8",     VOPProfileMAI_I32_I32_X4,    int_amdgcn_mfma_i32_4x4x4i8>;
396defm V_MFMA_F32_4X4X2BF16   : VOP3Inst<"v_mfma_f32_4x4x2bf16",   VOPProfileMAI_F32_V2I16_X4,  int_amdgcn_mfma_f32_4x4x2bf16>;
397defm V_MFMA_F32_16X16X1F32  : VOP3Inst<"v_mfma_f32_16x16x1f32",  VOPProfileMAI_F32_F32_X16,   int_amdgcn_mfma_f32_16x16x1f32>;
398defm V_MFMA_F32_16X16X4F32  : VOP3Inst<"v_mfma_f32_16x16x4f32",  VOPProfileMAI_F32_F32_X4,    int_amdgcn_mfma_f32_16x16x4f32>;
399defm V_MFMA_F32_16X16X4F16  : VOP3Inst<"v_mfma_f32_16x16x4f16",  VOPProfileMAI_F32_V4F16_X16, int_amdgcn_mfma_f32_16x16x4f16>;
400defm V_MFMA_F32_16X16X16F16 : VOP3Inst<"v_mfma_f32_16x16x16f16", VOPProfileMAI_F32_V4F16_X4,  int_amdgcn_mfma_f32_16x16x16f16>;
401defm V_MFMA_I32_16X16X4I8   : VOP3Inst<"v_mfma_i32_16x16x4i8",   VOPProfileMAI_I32_I32_X16,   int_amdgcn_mfma_i32_16x16x4i8>;
402defm V_MFMA_I32_16X16X16I8  : VOP3Inst<"v_mfma_i32_16x16x16i8",  VOPProfileMAI_I32_I32_X4,    int_amdgcn_mfma_i32_16x16x16i8>;
403defm V_MFMA_F32_16X16X2BF16 : VOP3Inst<"v_mfma_f32_16x16x2bf16", VOPProfileMAI_F32_V2I16_X16, int_amdgcn_mfma_f32_16x16x2bf16>;
404defm V_MFMA_F32_16X16X8BF16 : VOP3Inst<"v_mfma_f32_16x16x8bf16", VOPProfileMAI_F32_V2I16_X4,  int_amdgcn_mfma_f32_16x16x8bf16>;
405defm V_MFMA_F32_32X32X1F32  : VOP3Inst<"v_mfma_f32_32x32x1f32",  VOPProfileMAI_F32_F32_X32,   int_amdgcn_mfma_f32_32x32x1f32>;
406defm V_MFMA_F32_32X32X2F32  : VOP3Inst<"v_mfma_f32_32x32x2f32",  VOPProfileMAI_F32_F32_X16,   int_amdgcn_mfma_f32_32x32x2f32>;
407defm V_MFMA_F32_32X32X4F16  : VOP3Inst<"v_mfma_f32_32x32x4f16",  VOPProfileMAI_F32_V4F16_X32, int_amdgcn_mfma_f32_32x32x4f16>;
408defm V_MFMA_F32_32X32X8F16  : VOP3Inst<"v_mfma_f32_32x32x8f16",  VOPProfileMAI_F32_V4F16_X16, int_amdgcn_mfma_f32_32x32x8f16>;
409defm V_MFMA_I32_32X32X4I8   : VOP3Inst<"v_mfma_i32_32x32x4i8",   VOPProfileMAI_I32_I32_X32,   int_amdgcn_mfma_i32_32x32x4i8>;
410defm V_MFMA_I32_32X32X8I8   : VOP3Inst<"v_mfma_i32_32x32x8i8",   VOPProfileMAI_I32_I32_X16,   int_amdgcn_mfma_i32_32x32x8i8>;
411defm V_MFMA_F32_32X32X2BF16 : VOP3Inst<"v_mfma_f32_32x32x2bf16", VOPProfileMAI_F32_V2I16_X32, int_amdgcn_mfma_f32_32x32x2bf16>;
412defm V_MFMA_F32_32X32X4BF16 : VOP3Inst<"v_mfma_f32_32x32x4bf16", VOPProfileMAI_F32_V2I16_X16, int_amdgcn_mfma_f32_32x32x4bf16>;
413} // End isConvergent = 1, mayRaiseFPException = 0, ReadsModeReg = 1
414
415} // End SubtargetPredicate = HasMAIInsts
416
417def : MnemonicAlias<"v_accvgpr_read",  "v_accvgpr_read_b32">;
418def : MnemonicAlias<"v_accvgpr_write", "v_accvgpr_write_b32">;
419
420//===----------------------------------------------------------------------===//
421// Begin Real Encodings
422//===----------------------------------------------------------------------===//
423
424//===----------------------------------------------------------------------===//
425// GFX8 (VI)
426//===----------------------------------------------------------------------===//
427
428multiclass VOP3P_Real_vi<bits<7> op> {
429  def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
430            VOP3Pe <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
431    let AssemblerPredicate = HasVOP3PInsts;
432    let DecoderNamespace = "GFX8";
433  }
434}
435
436multiclass VOP3P_Real_MAI<bits<7> op> {
437  def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
438            VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl> {
439    let AssemblerPredicate = HasMAIInsts;
440    let DecoderNamespace = "GFX8";
441    let Inst{14} = 1; // op_sel_hi(2) default value
442    let Inst{59} = 1; // op_sel_hi(0) default value
443    let Inst{60} = 1; // op_sel_hi(1) default value
444  }
445}
446
447multiclass VOP3P_Real_MFMA<bits<7> op> {
448  def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
449            VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl> {
450    let AssemblerPredicate = HasMAIInsts;
451    let DecoderNamespace = "GFX8";
452  }
453}
454
455defm V_PK_MAD_I16 : VOP3P_Real_vi <0x00>;
456defm V_PK_MUL_LO_U16 : VOP3P_Real_vi <0x01>;
457defm V_PK_ADD_I16 : VOP3P_Real_vi <0x02>;
458defm V_PK_SUB_I16 : VOP3P_Real_vi <0x03>;
459defm V_PK_LSHLREV_B16 : VOP3P_Real_vi <0x04>;
460defm V_PK_LSHRREV_B16 : VOP3P_Real_vi <0x05>;
461defm V_PK_ASHRREV_I16 : VOP3P_Real_vi <0x06>;
462defm V_PK_MAX_I16 : VOP3P_Real_vi <0x07>;
463defm V_PK_MIN_I16 : VOP3P_Real_vi <0x08>;
464defm V_PK_MAD_U16 : VOP3P_Real_vi <0x09>;
465
466defm V_PK_ADD_U16 : VOP3P_Real_vi <0x0a>;
467defm V_PK_SUB_U16 : VOP3P_Real_vi <0x0b>;
468defm V_PK_MAX_U16 : VOP3P_Real_vi <0x0c>;
469defm V_PK_MIN_U16 : VOP3P_Real_vi <0x0d>;
470defm V_PK_FMA_F16 : VOP3P_Real_vi <0x0e>;
471defm V_PK_ADD_F16 : VOP3P_Real_vi <0x0f>;
472defm V_PK_MUL_F16 : VOP3P_Real_vi <0x10>;
473defm V_PK_MIN_F16 : VOP3P_Real_vi <0x11>;
474defm V_PK_MAX_F16 : VOP3P_Real_vi <0x12>;
475
476
477let SubtargetPredicate = HasMadMixInsts in {
478defm V_MAD_MIX_F32 : VOP3P_Real_vi <0x20>;
479defm V_MAD_MIXLO_F16 : VOP3P_Real_vi <0x21>;
480defm V_MAD_MIXHI_F16 : VOP3P_Real_vi <0x22>;
481}
482
483let SubtargetPredicate = HasFmaMixInsts in {
484let DecoderNamespace = "GFX9_DL" in {
485// The mad_mix instructions were renamed and their behaviors changed,
486// but the opcode stayed the same so we need to put these in a
487// different DecoderNamespace to avoid the ambiguity.
488defm V_FMA_MIX_F32 : VOP3P_Real_vi <0x20>;
489defm V_FMA_MIXLO_F16 : VOP3P_Real_vi <0x21>;
490defm V_FMA_MIXHI_F16 : VOP3P_Real_vi <0x22>;
491}
492}
493
494
495let SubtargetPredicate = HasDot2Insts in {
496
497defm V_DOT2_F32_F16 : VOP3P_Real_vi <0x23>;
498defm V_DOT2_I32_I16 : VOP3P_Real_vi <0x26>;
499defm V_DOT2_U32_U16 : VOP3P_Real_vi <0x27>;
500defm V_DOT4_U32_U8  : VOP3P_Real_vi <0x29>;
501defm V_DOT8_U32_U4  : VOP3P_Real_vi <0x2b>;
502
503} // End SubtargetPredicate = HasDot2Insts
504
505let SubtargetPredicate = HasDot1Insts in {
506
507defm V_DOT4_I32_I8  : VOP3P_Real_vi <0x28>;
508defm V_DOT8_I32_I4  : VOP3P_Real_vi <0x2a>;
509
510} // End SubtargetPredicate = HasDot1Insts
511
512let SubtargetPredicate = HasMAIInsts in {
513
514defm V_ACCVGPR_READ_B32  : VOP3P_Real_MAI <0x58>;
515defm V_ACCVGPR_WRITE_B32 : VOP3P_Real_MAI <0x59>;
516defm V_MFMA_F32_32X32X1F32  : VOP3P_Real_MFMA <0x40>;
517defm V_MFMA_F32_16X16X1F32  : VOP3P_Real_MFMA <0x41>;
518defm V_MFMA_F32_4X4X1F32    : VOP3P_Real_MFMA <0x42>;
519defm V_MFMA_F32_32X32X2F32  : VOP3P_Real_MFMA <0x44>;
520defm V_MFMA_F32_16X16X4F32  : VOP3P_Real_MFMA <0x45>;
521defm V_MFMA_F32_32X32X4F16  : VOP3P_Real_MFMA <0x48>;
522defm V_MFMA_F32_16X16X4F16  : VOP3P_Real_MFMA <0x49>;
523defm V_MFMA_F32_4X4X4F16    : VOP3P_Real_MFMA <0x4a>;
524defm V_MFMA_F32_32X32X8F16  : VOP3P_Real_MFMA <0x4c>;
525defm V_MFMA_F32_16X16X16F16 : VOP3P_Real_MFMA <0x4d>;
526defm V_MFMA_I32_32X32X4I8   : VOP3P_Real_MFMA <0x50>;
527defm V_MFMA_I32_16X16X4I8   : VOP3P_Real_MFMA <0x51>;
528defm V_MFMA_I32_4X4X4I8     : VOP3P_Real_MFMA <0x52>;
529defm V_MFMA_I32_16X16X16I8  : VOP3P_Real_MFMA <0x55>;
530defm V_MFMA_I32_32X32X8I8   : VOP3P_Real_MFMA <0x54>;
531defm V_MFMA_F32_32X32X2BF16 : VOP3P_Real_MFMA <0x68>;
532defm V_MFMA_F32_16X16X2BF16 : VOP3P_Real_MFMA <0x69>;
533defm V_MFMA_F32_4X4X2BF16   : VOP3P_Real_MFMA <0x6b>;
534defm V_MFMA_F32_32X32X4BF16 : VOP3P_Real_MFMA <0x6c>;
535defm V_MFMA_F32_16X16X8BF16 : VOP3P_Real_MFMA <0x6d>;
536
537} // End SubtargetPredicate = HasMAIInsts
538
539//===----------------------------------------------------------------------===//
540// GFX10.
541//===----------------------------------------------------------------------===//
542
543let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in {
544  multiclass VOP3P_Real_gfx10<bits<7> op> {
545    def _gfx10 : VOP3P_Real<!cast<VOP3P_Pseudo>(NAME), SIEncodingFamily.GFX10>,
546                 VOP3Pe_gfx10 <op, !cast<VOP3P_Pseudo>(NAME).Pfl>;
547  }
548} // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10"
549
550defm V_PK_MAD_I16     : VOP3P_Real_gfx10<0x00>;
551defm V_PK_MUL_LO_U16  : VOP3P_Real_gfx10<0x01>;
552defm V_PK_ADD_I16     : VOP3P_Real_gfx10<0x02>;
553defm V_PK_SUB_I16     : VOP3P_Real_gfx10<0x03>;
554defm V_PK_LSHLREV_B16 : VOP3P_Real_gfx10<0x04>;
555defm V_PK_LSHRREV_B16 : VOP3P_Real_gfx10<0x05>;
556defm V_PK_ASHRREV_I16 : VOP3P_Real_gfx10<0x06>;
557defm V_PK_MAX_I16     : VOP3P_Real_gfx10<0x07>;
558defm V_PK_MIN_I16     : VOP3P_Real_gfx10<0x08>;
559defm V_PK_MAD_U16     : VOP3P_Real_gfx10<0x09>;
560defm V_PK_ADD_U16     : VOP3P_Real_gfx10<0x0a>;
561defm V_PK_SUB_U16     : VOP3P_Real_gfx10<0x0b>;
562defm V_PK_MAX_U16     : VOP3P_Real_gfx10<0x0c>;
563defm V_PK_MIN_U16     : VOP3P_Real_gfx10<0x0d>;
564defm V_PK_FMA_F16     : VOP3P_Real_gfx10<0x0e>;
565defm V_PK_ADD_F16     : VOP3P_Real_gfx10<0x0f>;
566defm V_PK_MUL_F16     : VOP3P_Real_gfx10<0x10>;
567defm V_PK_MIN_F16     : VOP3P_Real_gfx10<0x11>;
568defm V_PK_MAX_F16     : VOP3P_Real_gfx10<0x12>;
569defm V_FMA_MIX_F32    : VOP3P_Real_gfx10<0x20>;
570defm V_FMA_MIXLO_F16  : VOP3P_Real_gfx10<0x21>;
571defm V_FMA_MIXHI_F16  : VOP3P_Real_gfx10<0x22>;
572
573let SubtargetPredicate = HasDot2Insts in {
574
575defm V_DOT2_F32_F16 : VOP3P_Real_gfx10 <0x13>;
576defm V_DOT2_I32_I16 : VOP3P_Real_gfx10 <0x14>;
577defm V_DOT2_U32_U16 : VOP3P_Real_gfx10 <0x15>;
578defm V_DOT4_U32_U8  : VOP3P_Real_gfx10 <0x17>;
579defm V_DOT8_U32_U4  : VOP3P_Real_gfx10 <0x19>;
580
581} // End SubtargetPredicate = HasDot2Insts
582
583let SubtargetPredicate = HasDot1Insts in {
584
585defm V_DOT4_I32_I8  : VOP3P_Real_gfx10 <0x16>;
586defm V_DOT8_I32_I4  : VOP3P_Real_gfx10 <0x18>;
587
588} // End SubtargetPredicate = HasDot1Insts
589