maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/lib/Target/AMDGPU/VOP3PInstructions.td

178 lines
8.1 KiB
TableGen
Raw Normal View History

AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
//===-- VOP3PInstructions.td - Vector Instruction Defintions --------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// VOP3P Classes
//===----------------------------------------------------------------------===//
class VOP3PInst<string OpName, VOPProfile P, SDPatternOperator node = null_frag> :
VOP3P_Pseudo<OpName, P,
!if(P.HasModifiers, getVOP3PModPat<P, node>.ret, getVOP3Pat<P, node>.ret)
>;
[AMDGPU][mc][gfx9] Added support of op_sel/op_sel_hi for V_MAD_MIX* See https://bugs.llvm.org//show_bug.cgi?id=33595 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D35021 llvm-svn: 307402
2017-07-07 22:29:06 +08:00
// Non-packed instructions that use the VOP3P encoding.
// VOP3 neg/abs and VOP3P opsel/opsel_hi modifiers are allowed.
AMDGPU: Add tied operands to v_mad_mix{lo|hi}_f16 These write to the low and high half of the destination register and leave the other 16-bits unchanged. This is true for most 16-bit instructions on gfx9, but we don't use that now. llvm-svn: 313812
2017-09-21 04:53:49 +08:00
class VOP3_VOP3PInst<string OpName, VOPProfile P, bit UseTiedOutput = 0,
SDPatternOperator node = null_frag> :
[AMDGPU][mc][gfx9] Added support of op_sel/op_sel_hi for V_MAD_MIX* See https://bugs.llvm.org//show_bug.cgi?id=33595 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D35021 llvm-svn: 307402
2017-07-07 22:29:06 +08:00
VOP3P_Pseudo<OpName, P> {
AMDGPU: Correct operand types for v_mad_mix* These aren't really packed instructions, so the default op_sel_hi should be 0 since this indicates a conversion. The operand types are scalar values that behave similar to an f16 scalar that may be converted to f32. Doesn't change the default printing for op_sel_hi, just the parsing. llvm-svn: 312179
2017-08-31 06:18:40 +08:00
// These operands are only sort of f16 operands. Depending on
// op_sel_hi, these may be interpreted as f32. The inline immediate
// values are really f16 converted to f32, so we treat these as f16
// operands.
[AMDGPU][mc][gfx9] Added support of op_sel/op_sel_hi for V_MAD_MIX* See https://bugs.llvm.org//show_bug.cgi?id=33595 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D35021 llvm-svn: 307402
2017-07-07 22:29:06 +08:00
let InOperandList =
AMDGPU: Add tied operands to v_mad_mix{lo|hi}_f16 These write to the low and high half of the destination register and leave the other 16-bits unchanged. This is true for most 16-bit instructions on gfx9, but we don't use that now. llvm-svn: 313812
2017-09-21 04:53:49 +08:00
!con(
!con(
(ins FP16InputMods:$src0_modifiers, VCSrc_f16:$src0,
FP16InputMods:$src1_modifiers, VCSrc_f16:$src1,
FP16InputMods:$src2_modifiers, VCSrc_f16:$src2,
clampmod:$clamp),
!if(UseTiedOutput, (ins VGPR_32:$vdst_in), (ins))),
(ins op_sel:$op_sel, op_sel_hi:$op_sel_hi));
let Constraints = !if(UseTiedOutput, "$vdst = $vdst_in", "");
let DisableEncoding = !if(UseTiedOutput, "$vdst_in", "");
[AMDGPU][mc][gfx9] Added support of op_sel/op_sel_hi for V_MAD_MIX* See https://bugs.llvm.org//show_bug.cgi?id=33595 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D35021 llvm-svn: 307402
2017-07-07 22:29:06 +08:00
let AsmOperands =
" $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$op_sel$op_sel_hi$clamp";
}
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
let isCommutable = 1 in {
AMDGPU: Support v2i16/v2f16 packed operations llvm-svn: 296396
2017-02-28 06:15:25 +08:00
def V_PK_FMA_F16 : VOP3PInst<"v_pk_fma_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, fma>;
[AMDGPU][MC] Added missing VOP3P opcodes Added support of the following opcodes: v_pk_sub_u16 v_pk_mad_i16 v_pk_mad_u16 See Bug 33593: https://bugs.llvm.org//show_bug.cgi?id=33593 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D34890 llvm-svn: 308281
2017-07-18 17:24:10 +08:00
def V_PK_MAD_I16 : VOP3PInst<"v_pk_mad_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
def V_PK_MAD_U16 : VOP3PInst<"v_pk_mad_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_V2I16>>;
AMDGPU: Support v2i16/v2f16 packed operations llvm-svn: 296396
2017-02-28 06:15:25 +08:00
def V_PK_ADD_F16 : VOP3PInst<"v_pk_add_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fadd>;
def V_PK_MUL_F16 : VOP3PInst<"v_pk_mul_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmul>;
def V_PK_MAX_F16 : VOP3PInst<"v_pk_max_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmaxnum>;
def V_PK_MIN_F16 : VOP3PInst<"v_pk_min_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fminnum>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
AMDGPU: Support v2i16/v2f16 packed operations llvm-svn: 296396
2017-02-28 06:15:25 +08:00
def V_PK_ADD_U16 : VOP3PInst<"v_pk_add_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, add>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
def V_PK_ADD_I16 : VOP3PInst<"v_pk_add_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
AMDGPU: Support v2i16/v2f16 packed operations llvm-svn: 296396
2017-02-28 06:15:25 +08:00
def V_PK_MUL_LO_U16 : VOP3PInst<"v_pk_mul_lo_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, mul>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
AMDGPU: Support v2i16/v2f16 packed operations llvm-svn: 296396
2017-02-28 06:15:25 +08:00
def V_PK_MIN_I16 : VOP3PInst<"v_pk_min_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smin>;
def V_PK_MIN_U16 : VOP3PInst<"v_pk_min_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umin>;
def V_PK_MAX_I16 : VOP3PInst<"v_pk_max_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, smax>;
def V_PK_MAX_U16 : VOP3PInst<"v_pk_max_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, umax>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
}
[AMDGPU][MC] Added missing VOP3P opcodes Added support of the following opcodes: v_pk_sub_u16 v_pk_mad_i16 v_pk_mad_u16 See Bug 33593: https://bugs.llvm.org//show_bug.cgi?id=33593 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D34890 llvm-svn: 308281
2017-07-18 17:24:10 +08:00
def V_PK_SUB_U16 : VOP3PInst<"v_pk_sub_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;
def V_PK_SUB_I16 : VOP3PInst<"v_pk_sub_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, sub>;
AMDGPU: Support v2i16/v2f16 packed operations llvm-svn: 296396
2017-02-28 06:15:25 +08:00
def V_PK_LSHLREV_B16 : VOP3PInst<"v_pk_lshlrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshl_rev>;
def V_PK_ASHRREV_I16 : VOP3PInst<"v_pk_ashrrev_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, ashr_rev>;
def V_PK_LSHRREV_B16 : VOP3PInst<"v_pk_lshrrev_b16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, lshr_rev>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
[AMDGPU][mc][gfx9] Added support of op_sel/op_sel_hi for V_MAD_MIX* See https://bugs.llvm.org//show_bug.cgi?id=33595 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D35021 llvm-svn: 307402
2017-07-07 22:29:06 +08:00
// These are VOP3a-like opcodes which accept no omod.
// Size of src arguments (16/32) is controlled by op_sel.
// For 16-bit src arguments their location (hi/lo) are controlled by op_sel_hi.
AMDGPU: Correct operand types for v_mad_mix* These aren't really packed instructions, so the default op_sel_hi should be 0 since this indicates a conversion. The operand types are scalar values that behave similar to an f16 scalar that may be converted to f32. Doesn't change the default printing for op_sel_hi, just the parsing. llvm-svn: 312179
2017-08-31 06:18:40 +08:00
let isCommutable = 1 in {
AMDGPU: Fix encoding of op_sel for mad_mix* opcodes llvm-svn: 313797
2017-09-21 03:09:28 +08:00
def V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
AMDGPU: Start selecting v_mad_mixlo_f16 Also add some tests that should be able to use v_mad_mixhi_f16, but do not yet. This is trickier because we don't really model the partial update of the register done by 16-bit instructions. llvm-svn: 313806
2017-09-21 04:28:39 +08:00
// Clamp modifier is applied after conversion to f16.
AMDGPU: Add tied operands to v_mad_mix{lo|hi}_f16 These write to the low and high half of the destination register and leave the other 16-bits unchanged. This is true for most 16-bit instructions on gfx9, but we don't use that now. llvm-svn: 313812
2017-09-21 04:53:49 +08:00
def V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
AMDGPU: Start selecting v_mad_mixhi_f16 llvm-svn: 313814
2017-09-21 05:01:24 +08:00
let ClampLo = 0, ClampHi = 1 in {
AMDGPU: Add tied operands to v_mad_mix{lo|hi}_f16 These write to the low and high half of the destination register and leave the other 16-bits unchanged. This is true for most 16-bit instructions on gfx9, but we don't use that now. llvm-svn: 313812
2017-09-21 04:53:49 +08:00
def V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, 1>;
AMDGPU: Correct operand types for v_mad_mix* These aren't really packed instructions, so the default op_sel_hi should be 0 since this indicates a conversion. The operand types are scalar values that behave similar to an f16 scalar that may be converted to f32. Doesn't change the default printing for op_sel_hi, just the parsing. llvm-svn: 312179
2017-08-31 06:18:40 +08:00
}
AMDGPU: Start selecting v_mad_mixhi_f16 llvm-svn: 313814
2017-09-21 05:01:24 +08:00
}
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
AMDGPU: Start selecting v_mad_mixlo_f16 Also add some tests that should be able to use v_mad_mixhi_f16, but do not yet. This is trickier because we don't really model the partial update of the register done by 16-bit instructions. llvm-svn: 313806
2017-09-21 04:28:39 +08:00
let Predicates = [HasMadMix] in {
def : Pat <
(f16 (fpround (fmad (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
(f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
(f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
(V_MAD_MIXLO_F16 $src0_modifiers, $src0,
$src1_modifiers, $src1,
$src2_modifiers, $src2,
AMDGPU: Start selecting v_mad_mixhi_f16 llvm-svn: 313814
2017-09-21 05:01:24 +08:00
DSTCLAMP.NONE,
AMDGPU: Add tied operands to v_mad_mix{lo|hi}_f16 These write to the low and high half of the destination register and leave the other 16-bits unchanged. This is true for most 16-bit instructions on gfx9, but we don't use that now. llvm-svn: 313812
2017-09-21 04:53:49 +08:00
(i32 (IMPLICIT_DEF)))
AMDGPU: Start selecting v_mad_mixlo_f16 Also add some tests that should be able to use v_mad_mixhi_f16, but do not yet. This is trickier because we don't really model the partial update of the register done by 16-bit instructions. llvm-svn: 313806
2017-09-21 04:28:39 +08:00
>;
AMDGPU: Start selecting v_mad_mixhi_f16 llvm-svn: 313814
2017-09-21 05:01:24 +08:00
// FIXME: Special case handling for maxhi (especially for clamp)
// because dealing with the write to high half of the register is
// difficult.
def : Pat <
(build_vector f16:$elt0, (fpround (fmad (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
(f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
(f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers))))),
(v2f16 (V_MAD_MIXHI_F16 $src0_modifiers, $src0,
$src1_modifiers, $src1,
$src2_modifiers, $src2,
DSTCLAMP.NONE,
$elt0))
>;
def : Pat <
(build_vector
f16:$elt0,
(AMDGPUclamp (fpround (fmad (f32 (VOP3PMadMixMods f16:$src0, i32:$src0_modifiers)),
(f32 (VOP3PMadMixMods f16:$src1, i32:$src1_modifiers)),
(f32 (VOP3PMadMixMods f16:$src2, i32:$src2_modifiers)))))),
(v2f16 (V_MAD_MIXHI_F16 $src0_modifiers, $src0,
$src1_modifiers, $src1,
$src2_modifiers, $src2,
DSTCLAMP.ENABLE,
$elt0))
>;
def : Pat <
(AMDGPUclamp (build_vector
(fpround (fmad (f32 (VOP3PMadMixMods f16:$lo_src0, i32:$lo_src0_modifiers)),
(f32 (VOP3PMadMixMods f16:$lo_src1, i32:$lo_src1_modifiers)),
(f32 (VOP3PMadMixMods f16:$lo_src2, i32:$lo_src2_modifiers)))),
(fpround (fmad (f32 (VOP3PMadMixMods f16:$hi_src0, i32:$hi_src0_modifiers)),
(f32 (VOP3PMadMixMods f16:$hi_src1, i32:$hi_src1_modifiers)),
(f32 (VOP3PMadMixMods f16:$hi_src2, i32:$hi_src2_modifiers)))))),
(v2f16 (V_MAD_MIXHI_F16 $hi_src0_modifiers, $hi_src0,
$hi_src1_modifiers, $hi_src1,
$hi_src2_modifiers, $hi_src2,
DSTCLAMP.ENABLE,
(V_MAD_MIXLO_F16 $lo_src0_modifiers, $lo_src0,
$lo_src1_modifiers, $lo_src1,
$lo_src2_modifiers, $lo_src2,
DSTCLAMP.ENABLE,
(i32 (IMPLICIT_DEF)))))
>;
AMDGPU: Start selecting v_mad_mixlo_f16 Also add some tests that should be able to use v_mad_mixhi_f16, but do not yet. This is trickier because we don't really model the partial update of the register done by 16-bit instructions. llvm-svn: 313806
2017-09-21 04:28:39 +08:00
} // End Predicates = [HasMadMix]
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
multiclass VOP3P_Real_vi<bits<10> op> {
def _vi : VOP3P_Real<!cast<VOP3P_Pseudo>(NAME), SIEncodingFamily.VI>,
VOP3Pe <op, !cast<VOP3P_Pseudo>(NAME).Pfl> {
let AssemblerPredicates = [HasVOP3PInsts];
let DecoderNamespace = "VI";
}
}
[AMDGPU][MC] Added missing VOP3P opcodes Added support of the following opcodes: v_pk_sub_u16 v_pk_mad_i16 v_pk_mad_u16 See Bug 33593: https://bugs.llvm.org//show_bug.cgi?id=33593 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D34890 llvm-svn: 308281
2017-07-18 17:24:10 +08:00
defm V_PK_MAD_I16 : VOP3P_Real_vi <0x380>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
defm V_PK_MUL_LO_U16 : VOP3P_Real_vi <0x381>;
defm V_PK_ADD_I16 : VOP3P_Real_vi <0x382>;
defm V_PK_SUB_I16 : VOP3P_Real_vi <0x383>;
defm V_PK_LSHLREV_B16 : VOP3P_Real_vi <0x384>;
defm V_PK_LSHRREV_B16 : VOP3P_Real_vi <0x385>;
defm V_PK_ASHRREV_I16 : VOP3P_Real_vi <0x386>;
defm V_PK_MAX_I16 : VOP3P_Real_vi <0x387>;
defm V_PK_MIN_I16 : VOP3P_Real_vi <0x388>;
[AMDGPU][MC] Added missing VOP3P opcodes Added support of the following opcodes: v_pk_sub_u16 v_pk_mad_i16 v_pk_mad_u16 See Bug 33593: https://bugs.llvm.org//show_bug.cgi?id=33593 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D34890 llvm-svn: 308281
2017-07-18 17:24:10 +08:00
defm V_PK_MAD_U16 : VOP3P_Real_vi <0x389>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
defm V_PK_ADD_U16 : VOP3P_Real_vi <0x38a>;
[AMDGPU][MC] Added missing VOP3P opcodes Added support of the following opcodes: v_pk_sub_u16 v_pk_mad_i16 v_pk_mad_u16 See Bug 33593: https://bugs.llvm.org//show_bug.cgi?id=33593 Reviewers: vpykhtin, artem.tamazov, arsenm Differential Revision: https://reviews.llvm.org/D34890 llvm-svn: 308281
2017-07-18 17:24:10 +08:00
defm V_PK_SUB_U16 : VOP3P_Real_vi <0x38b>;
AMDGPU: Add VOP3P instruction format Add a few non-VOP3P but instructions related to packed. Includes hack with dummy operands for the benefit of the assembler llvm-svn: 296368
2017-02-28 02:49:11 +08:00
defm V_PK_MAX_U16 : VOP3P_Real_vi <0x38c>;
defm V_PK_MIN_U16 : VOP3P_Real_vi <0x38d>;
defm V_PK_FMA_F16 : VOP3P_Real_vi <0x38e>;
defm V_PK_ADD_F16 : VOP3P_Real_vi <0x38f>;
defm V_PK_MUL_F16 : VOP3P_Real_vi <0x390>;
defm V_PK_MIN_F16 : VOP3P_Real_vi <0x391>;
defm V_PK_MAX_F16 : VOP3P_Real_vi <0x392>;
defm V_MAD_MIX_F32 : VOP3P_Real_vi <0x3a0>;
defm V_MAD_MIXLO_F16 : VOP3P_Real_vi <0x3a1>;
defm V_MAD_MIXHI_F16 : VOP3P_Real_vi <0x3a2>;