[AMDGPU][True16][MC] true16 for minimummaximum/max/min/max3/min3 #124184

broxigarchen · 2025-01-23T20:28:16Z

true16 support for gfx12 instructions including:

v_minimummaximum_f16
v_maximumminimum_f16
v_maximum_f16
v_minimum_f16
v_maximum3_f16
v_minimum3_f16

github-actions · 2025-01-23T21:40:21Z

✅ With the latest revision this PR passed the C/C++ code formatter.

llvmbot · 2025-01-23T23:55:54Z

@llvm/pr-subscribers-mc

@llvm/pr-subscribers-backend-amdgpu

Author: Brox Chen (broxigarchen)

Changes

true16 support for gfx12 instructions including:

v_minimummaximum_f16
v_maximumminimum_f16
v_maximum_f16
v_minimum_f16
v_maximum3_f16
v_minimum3_f16

Patch is 352.37 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/124184.diff

10 Files Affected:

(modified) llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (+29-5)
(modified) llvm/lib/Target/AMDGPU/SIInstructions.td (+7-4)
(modified) llvm/lib/Target/AMDGPU/VOP3Instructions.td (+12-12)
(modified) llvm/lib/Target/AMDGPU/VOPInstructions.td (+10-3)
(modified) llvm/test/MC/AMDGPU/gfx12_asm_vop3.s (+246-180)
(modified) llvm/test/MC/AMDGPU/gfx12_asm_vop3_dpp16.s (+262-184)
(modified) llvm/test/MC/AMDGPU/gfx12_asm_vop3_dpp8.s (+194-128)
(modified) llvm/test/MC/Disassembler/AMDGPU/gfx12_dasm_vop3.txt (+492-90)
(modified) llvm/test/MC/Disassembler/AMDGPU/gfx12_dasm_vop3_dpp16.txt (+461-88)
(modified) llvm/test/MC/Disassembler/AMDGPU/gfx12_dasm_vop3_dpp8.txt (+276-60)

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 1d98d68a2ea5d4..26f6308d9317d0 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -5573,8 +5573,12 @@ unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) const {
   case AMDGPU::S_SUB_F16: return AMDGPU::V_SUB_F16_fake16_e64;
   case AMDGPU::S_MIN_F16: return AMDGPU::V_MIN_F16_fake16_e64;
   case AMDGPU::S_MAX_F16: return AMDGPU::V_MAX_F16_fake16_e64;
-  case AMDGPU::S_MINIMUM_F16: return AMDGPU::V_MINIMUM_F16_e64;
-  case AMDGPU::S_MAXIMUM_F16: return AMDGPU::V_MAXIMUM_F16_e64;
+  case AMDGPU::S_MINIMUM_F16:
+    return ST.useRealTrue16Insts() ? AMDGPU::V_MINIMUM_F16_t16_e64
+                                   : AMDGPU::V_MINIMUM_F16_fake16_e64;
+  case AMDGPU::S_MAXIMUM_F16:
+    return ST.useRealTrue16Insts() ? AMDGPU::V_MAXIMUM_F16_t16_e64
+                                   : AMDGPU::V_MAXIMUM_F16_fake16_e64;
   case AMDGPU::S_MUL_F16: return AMDGPU::V_MUL_F16_fake16_e64;
   case AMDGPU::S_CVT_PK_RTZ_F16_F32: return AMDGPU::V_CVT_PKRTZ_F16_F32_e64;
   case AMDGPU::S_FMAC_F32: return AMDGPU::V_FMAC_F32_e64;
@@ -7547,9 +7551,7 @@ void SIInstrInfo::moveToVALUImpl(SIInstrWorklist &Worklist,
     return;
   }
   case AMDGPU::S_MINIMUM_F32:
-  case AMDGPU::S_MAXIMUM_F32:
-  case AMDGPU::S_MINIMUM_F16:
-  case AMDGPU::S_MAXIMUM_F16: {
+  case AMDGPU::S_MAXIMUM_F32: {
     const DebugLoc &DL = Inst.getDebugLoc();
     Register NewDst = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
     MachineInstr *NewInstr = BuildMI(*MBB, Inst, DL, get(NewOpcode), NewDst)
@@ -7566,6 +7568,28 @@ void SIInstrInfo::moveToVALUImpl(SIInstrWorklist &Worklist,
     Inst.eraseFromParent();
     return;
   }
+  case AMDGPU::S_MINIMUM_F16:
+  case AMDGPU::S_MAXIMUM_F16: {
+    const DebugLoc &DL = Inst.getDebugLoc();
+    Register NewDst;
+    if (ST.useRealTrue16Insts())
+      NewDst = MRI.createVirtualRegister(&AMDGPU::VGPR_16RegClass);
+    else
+      NewDst = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+    MachineInstr *NewInstr = BuildMI(*MBB, Inst, DL, get(NewOpcode), NewDst)
+                                 .addImm(0) // src0_modifiers
+                                 .add(Inst.getOperand(1))
+                                 .addImm(0) // src1_modifiers
+                                 .add(Inst.getOperand(2))
+                                 .addImm(0)  // clamp
+                                 .addImm(0)  // omod
+                                 .addImm(0); // opsel0
+    MRI.replaceRegWith(Inst.getOperand(0).getReg(), NewDst);
+    legalizeOperands(*NewInstr, MDT);
+    addUsersToMoveToVALUWorklist(NewDst, MRI, Worklist);
+    Inst.eraseFromParent();
+    return;
+  }
   }
 
   if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 40a20fa9cb15ea..bf985f0d2b1720 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -3770,12 +3770,15 @@ let True16Predicate = UseFakeTrue16Insts in {
 let OtherPredicates = [isGFX12Plus] in {
 def : FPMinMaxPat<V_MINIMUMMAXIMUM_F32_e64, f32, DivergentBinFrag<fmaximum>, fminimum_oneuse>;
 def : FPMinMaxPat<V_MAXIMUMMINIMUM_F32_e64, f32, DivergentBinFrag<fminimum>, fmaximum_oneuse>;
-def : FPMinMaxPat<V_MINIMUMMAXIMUM_F16_e64, f16, DivergentBinFrag<fmaximum>, fminimum_oneuse>;
-def : FPMinMaxPat<V_MAXIMUMMINIMUM_F16_e64, f16, DivergentBinFrag<fminimum>, fmaximum_oneuse>;
 def : FPMinCanonMaxPat<V_MINIMUMMAXIMUM_F32_e64, f32, DivergentBinFrag<fmaximum>, fminimum_oneuse>;
 def : FPMinCanonMaxPat<V_MAXIMUMMINIMUM_F32_e64, f32, DivergentBinFrag<fminimum>, fmaximum_oneuse>;
-def : FPMinCanonMaxPat<V_MINIMUMMAXIMUM_F16_e64, f16, DivergentBinFrag<fmaximum>, fminimum_oneuse>;
-def : FPMinCanonMaxPat<V_MAXIMUMMINIMUM_F16_e64, f16, DivergentBinFrag<fminimum>, fmaximum_oneuse>;
+}
+
+let True16Predicate = UseFakeTrue16Insts, OtherPredicates = [isGFX12Plus] in {
+def : FPMinMaxPat<V_MINIMUMMAXIMUM_F16_fake16_e64, f16, DivergentBinFrag<fmaximum>, fminimum_oneuse>;
+def : FPMinMaxPat<V_MAXIMUMMINIMUM_F16_fake16_e64, f16, DivergentBinFrag<fminimum>, fmaximum_oneuse>;
+def : FPMinCanonMaxPat<V_MINIMUMMAXIMUM_F16_fake16_e64, f16, DivergentBinFrag<fmaximum>, fminimum_oneuse>;
+def : FPMinCanonMaxPat<V_MAXIMUMMINIMUM_F16_fake16_e64, f16, DivergentBinFrag<fminimum>, fmaximum_oneuse>;
 }
 
 // Convert a floating-point power of 2 to the integer exponent.
diff --git a/llvm/lib/Target/AMDGPU/VOP3Instructions.td b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
index 947ac5c27620f0..0de12e43fd36ea 100644
--- a/llvm/lib/Target/AMDGPU/VOP3Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
@@ -170,8 +170,8 @@ defm V_MUL_HI_I32 : VOP3Inst <"v_mul_hi_i32", V_MUL_PROF<VOP_I32_I32_I32>, mulhs
 let SubtargetPredicate = isGFX12Plus, ReadsModeReg = 0, AddedComplexity = 1 in {
 defm V_MINIMUM_F32 : VOP3Inst <"v_minimum_f32", VOP3_Profile<VOP_F32_F32_F32>, DivergentBinFrag<fminimum>>;
 defm V_MAXIMUM_F32 : VOP3Inst <"v_maximum_f32", VOP3_Profile<VOP_F32_F32_F32>, DivergentBinFrag<fmaximum>>;
-defm V_MINIMUM_F16 : VOP3Inst <"v_minimum_f16", VOP3_Profile<VOP_F16_F16_F16>, DivergentBinFrag<fminimum>>;
-defm V_MAXIMUM_F16 : VOP3Inst <"v_maximum_f16", VOP3_Profile<VOP_F16_F16_F16>, DivergentBinFrag<fmaximum>>;
+defm V_MINIMUM_F16 : VOP3Inst_t16 <"v_minimum_f16", VOP_F16_F16_F16, DivergentBinFrag<fminimum>>;
+defm V_MAXIMUM_F16 : VOP3Inst_t16 <"v_maximum_f16", VOP_F16_F16_F16, DivergentBinFrag<fmaximum>>;
 
 let SchedRW = [WriteDoubleAdd] in {
 defm V_MINIMUM_F64 : VOP3Inst <"v_minimum_f64", VOP3_Profile<VOP_F64_F64_F64>, fminimum>;
@@ -634,8 +634,8 @@ defm V_MAX3_I16 : VOP3Inst_t16 <"v_max3_i16", VOP_I16_I16_I16_I16, AMDGPUsmax3>;
 defm V_MAX3_U16 : VOP3Inst_t16 <"v_max3_u16", VOP_I16_I16_I16_I16, AMDGPUumax3>;
 
 let SubtargetPredicate = HasMinimum3Maximum3F16, ReadsModeReg = 0 in {
-  defm V_MINIMUM3_F16 : VOP3Inst <"v_minimum3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfminimum3>;
-  defm V_MAXIMUM3_F16 : VOP3Inst <"v_maximum3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmaximum3>;
+  defm V_MINIMUM3_F16 : VOP3Inst_t16 <"v_minimum3_f16", VOP_F16_F16_F16_F16, AMDGPUfminimum3>;
+  defm V_MAXIMUM3_F16 : VOP3Inst_t16 <"v_maximum3_f16", VOP_F16_F16_F16_F16, AMDGPUfmaximum3>;
 } // End SubtargetPredicate = isGFX12Plus, ReadsModeReg = 0
 
 defm V_ADD_I16 : VOP3Inst_t16 <"v_add_i16", VOP_I16_I16_I16>;
@@ -1440,8 +1440,8 @@ let SubtargetPredicate = HasF32ToF16BF16ConversionSRInsts in {
 let SubtargetPredicate = isGFX12Plus, ReadsModeReg = 0 in {
   defm V_MAXIMUMMINIMUM_F32 : VOP3Inst<"v_maximumminimum_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
   defm V_MINIMUMMAXIMUM_F32 : VOP3Inst<"v_minimummaximum_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
-  defm V_MAXIMUMMINIMUM_F16 : VOP3Inst<"v_maximumminimum_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>>;
-  defm V_MINIMUMMAXIMUM_F16 : VOP3Inst<"v_minimummaximum_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>>;
+  defm V_MAXIMUMMINIMUM_F16 : VOP3Inst_t16<"v_maximumminimum_f16", VOP_F16_F16_F16_F16>;
+  defm V_MINIMUMMAXIMUM_F16 : VOP3Inst_t16<"v_minimummaximum_f16", VOP_F16_F16_F16_F16>;
 } // End SubtargetPredicate = isGFX12Plus, ReadsModeReg = 0
 
 let SubtargetPredicate = HasDot9Insts, IsDOT=1 in {
@@ -1591,8 +1591,8 @@ defm V_MIN3_NUM_F16       : VOP3_Realtriple_t16_and_fake16_gfx12<0x22b, "v_min3_
 defm V_MAX3_NUM_F16       : VOP3_Realtriple_t16_and_fake16_gfx12<0x22c, "v_max3_num_f16", "V_MAX3_F16", "v_max3_f16">;
 defm V_MINIMUM3_F32       : VOP3Only_Realtriple_gfx12<0x22d>;
 defm V_MAXIMUM3_F32       : VOP3Only_Realtriple_gfx12<0x22e>;
-defm V_MINIMUM3_F16       : VOP3Only_Realtriple_t16_gfx12<0x22f>;
-defm V_MAXIMUM3_F16       : VOP3Only_Realtriple_t16_gfx12<0x230>;
+defm V_MINIMUM3_F16       : VOP3Only_Realtriple_t16_and_fake16_gfx12<0x22f, "v_minimum3_f16">;
+defm V_MAXIMUM3_F16       : VOP3Only_Realtriple_t16_and_fake16_gfx12<0x230, "v_maximum3_f16">;
 defm V_MED3_NUM_F32       : VOP3_Realtriple_with_name_gfx12<0x231, "V_MED3_F32", "v_med3_num_f32">;
 defm V_MED3_NUM_F16       : VOP3_Realtriple_t16_and_fake16_gfx12<0x232, "v_med3_num_f16", "V_MED3_F16", "v_med3_f16">;
 defm V_MINMAX_NUM_F32     : VOP3_Realtriple_with_name_gfx12<0x268, "V_MINMAX_F32", "v_minmax_num_f32">;
@@ -1601,8 +1601,8 @@ defm V_MINMAX_NUM_F16     : VOP3_Realtriple_t16_and_fake16_gfx12<0x26a, "v_minma
 defm V_MAXMIN_NUM_F16     : VOP3_Realtriple_t16_and_fake16_gfx12<0x26b, "v_maxmin_num_f16", "V_MAXMIN_F16", "v_maxmin_f16">;
 defm V_MINIMUMMAXIMUM_F32 : VOP3Only_Realtriple_gfx12<0x26c>;
 defm V_MAXIMUMMINIMUM_F32 : VOP3Only_Realtriple_gfx12<0x26d>;
-defm V_MINIMUMMAXIMUM_F16 : VOP3Only_Realtriple_t16_gfx12<0x26e>;
-defm V_MAXIMUMMINIMUM_F16 : VOP3Only_Realtriple_t16_gfx12<0x26f>;
+defm V_MINIMUMMAXIMUM_F16 : VOP3Only_Realtriple_t16_and_fake16_gfx12<0x26e, "v_minimummaximum_f16">;
+defm V_MAXIMUMMINIMUM_F16 : VOP3Only_Realtriple_t16_and_fake16_gfx12<0x26f, "v_maximumminimum_f16">;
 defm V_S_EXP_F32          : VOP3Only_Real_Base_gfx12<0x280>;
 defm V_S_EXP_F16          : VOP3Only_Real_Base_gfx12<0x281>;
 defm V_S_LOG_F32          : VOP3Only_Real_Base_gfx12<0x282>;
@@ -1619,8 +1619,8 @@ defm V_MINIMUM_F64        : VOP3Only_Real_Base_gfx12<0x341>;
 defm V_MAXIMUM_F64        : VOP3Only_Real_Base_gfx12<0x342>;
 defm V_MINIMUM_F32        : VOP3Only_Realtriple_gfx12<0x365>;
 defm V_MAXIMUM_F32        : VOP3Only_Realtriple_gfx12<0x366>;
-defm V_MINIMUM_F16        : VOP3Only_Realtriple_t16_gfx12<0x367>;
-defm V_MAXIMUM_F16        : VOP3Only_Realtriple_t16_gfx12<0x368>;
+defm V_MINIMUM_F16        : VOP3Only_Realtriple_t16_and_fake16_gfx12<0x367, "v_minimum_f16">;
+defm V_MAXIMUM_F16        : VOP3Only_Realtriple_t16_and_fake16_gfx12<0x368, "v_maximum_f16">;
 
 defm V_PERMLANE16_VAR_B32  : VOP3Only_Real_Base_gfx12<0x30f>;
 defm V_PERMLANEX16_VAR_B32 : VOP3Only_Real_Base_gfx12<0x310>;
diff --git a/llvm/lib/Target/AMDGPU/VOPInstructions.td b/llvm/lib/Target/AMDGPU/VOPInstructions.td
index 3b5358b737aa4c..eb18cabe368ce0 100644
--- a/llvm/lib/Target/AMDGPU/VOPInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOPInstructions.td
@@ -1947,9 +1947,6 @@ multiclass VOP3Only_Realtriple_gfx12<bits<10> op, bit isSingle = 0> :
 multiclass VOP3Only_Real_Base_gfx12<bits<10> op> :
   VOP3_Real_Base<GFX12Gen, op, NAME, 1/*IsSingle*/>;
 
-multiclass VOP3Only_Realtriple_t16_gfx12<bits<10> op> :
-  VOP3Only_Realtriple<GFX12Gen, op>;
-
 multiclass VOP3_Realtriple_t16_gfx12<bits<10> op, string asmName, string opName = NAME,
                                      string pseudo_mnemonic = "", bit isSingle = 0> :
   VOP3_Realtriple_with_name<GFX12Gen, op, opName, asmName, pseudo_mnemonic, isSingle>;
@@ -1960,6 +1957,16 @@ multiclass VOP3_Realtriple_t16_and_fake16_gfx12<bits<10> op, string asmName, str
   defm _fake16:VOP3_Realtriple_t16_gfx12<op, asmName, opName#"_fake16", pseudo_mnemonic, isSingle>;
 }
 
+multiclass VOP3Only_Realtriple_t16_gfx12<bits<10> op, string asmName,
+                                     string opName = NAME, string pseudo_mnemonic = "">
+  : VOP3_Realtriple_t16_gfx12<op, asmName, opName, pseudo_mnemonic, 1>;
+
+multiclass VOP3Only_Realtriple_t16_and_fake16_gfx12<bits<10> op, string asmName,
+                                     string opName = NAME, string pseudo_mnemonic = ""> {
+  defm _t16 : VOP3Only_Realtriple_t16_gfx12<op, asmName, opName#"_t16", pseudo_mnemonic>;
+  defm _fake16 : VOP3Only_Realtriple_t16_gfx12<op, asmName, opName#"_fake16", pseudo_mnemonic>;
+}
+
 multiclass VOP3be_Real_with_name_gfx12<bits<10> op, string opName,
                                        string asmName, bit isSingle = 0> {
   defvar ps = !cast<VOP3_Pseudo>(opName#"_e64");
diff --git a/llvm/test/MC/AMDGPU/gfx12_asm_vop3.s b/llvm/test/MC/AMDGPU/gfx12_asm_vop3.s
index cd4ed2b9458e6c..ff24cbbbd2bea0 100644
--- a/llvm/test/MC/AMDGPU/gfx12_asm_vop3.s
+++ b/llvm/test/MC/AMDGPU/gfx12_asm_vop3.s
@@ -6445,101 +6445,119 @@ v_maximum_f32 v5, -src_scc, |vcc_lo|
 v_maximum_f32 v255, -|0xaf123456|, -|vcc_hi|
 // GFX12: v_maximum_f32 v255, -|0xaf123456|, -|vcc_hi| ; encoding: [0xff,0x03,0x66,0xd7,0xff,0xd6,0x00,0x60,0x56,0x34,0x12,0xaf]
 
-v_minimum_f16 v5, v1, v2
-// GFX12: v_minimum_f16 v5, v1, v2                ; encoding: [0x05,0x00,0x67,0xd7,0x01,0x05,0x02,0x00]
+v_minimum_f16 v5.l, v1.l, v2.l
+// GFX12: v_minimum_f16 v5.l, v1.l, v2.l          ; encoding: [0x05,0x00,0x67,0xd7,0x01,0x05,0x02,0x00]
 
-v_minimum_f16 v5, v255, v255
-// GFX12: v_minimum_f16 v5, v255, v255            ; encoding: [0x05,0x00,0x67,0xd7,0xff,0xff,0x03,0x00]
+v_minimum_f16 v5.l, v255.l, v255.l
+// GFX12: v_minimum_f16 v5.l, v255.l, v255.l      ; encoding: [0x05,0x00,0x67,0xd7,0xff,0xff,0x03,0x00]
 
-v_minimum_f16 v5, s1, s2
-// GFX12: v_minimum_f16 v5, s1, s2                ; encoding: [0x05,0x00,0x67,0xd7,0x01,0x04,0x00,0x00]
+v_minimum_f16 v5.l, s1, s2
+// GFX12: v_minimum_f16 v5.l, s1, s2              ; encoding: [0x05,0x00,0x67,0xd7,0x01,0x04,0x00,0x00]
 
-v_minimum_f16 v5, s105, s105
-// GFX12: v_minimum_f16 v5, s105, s105            ; encoding: [0x05,0x00,0x67,0xd7,0x69,0xd2,0x00,0x00]
+v_minimum_f16 v5.l, s105, s105
+// GFX12: v_minimum_f16 v5.l, s105, s105          ; encoding: [0x05,0x00,0x67,0xd7,0x69,0xd2,0x00,0x00]
 
-v_minimum_f16 v5, vcc_lo, ttmp15
-// GFX12: v_minimum_f16 v5, vcc_lo, ttmp15        ; encoding: [0x05,0x00,0x67,0xd7,0x6a,0xf6,0x00,0x00]
+v_minimum_f16 v5.l, vcc_lo, ttmp15
+// GFX12: v_minimum_f16 v5.l, vcc_lo, ttmp15      ; encoding: [0x05,0x00,0x67,0xd7,0x6a,0xf6,0x00,0x00]
 
-v_minimum_f16 v5, vcc_hi, 0xaf12
-// GFX12: v_minimum_f16 v5, vcc_hi, 0xaf12        ; encoding: [0x05,0x00,0x67,0xd7,0x6b,0xfe,0x01,0x00,0x12,0xaf,0x00,0x00]
+v_minimum_f16 v5.l, vcc_hi, 0xaf12
+// GFX12: v_minimum_f16 v5.l, vcc_hi, 0xaf12      ; encoding: [0x05,0x00,0x67,0xd7,0x6b,0xfe,0x01,0x00,0x12,0xaf,0x00,0x00]
 
-v_minimum_f16 v5, ttmp15, src_scc
-// GFX12: v_minimum_f16 v5, ttmp15, src_scc       ; encoding: [0x05,0x00,0x67,0xd7,0x7b,0xfa,0x01,0x00]
+v_minimum_f16 v5.l, ttmp15, src_scc
+// GFX12: v_minimum_f16 v5.l, ttmp15, src_scc     ; encoding: [0x05,0x00,0x67,0xd7,0x7b,0xfa,0x01,0x00]
 
-v_minimum_f16 v5, m0, 0.5
-// GFX12: v_minimum_f16 v5, m0, 0.5               ; encoding: [0x05,0x00,0x67,0xd7,0x7d,0xe0,0x01,0x00]
+v_minimum_f16 v5.l, m0, 0.5
+// GFX12: v_minimum_f16 v5.l, m0, 0.5             ; encoding: [0x05,0x00,0x67,0xd7,0x7d,0xe0,0x01,0x00]
 
-v_minimum_f16 v5, exec_lo, -1
-// GFX12: v_minimum_f16 v5, exec_lo, -1           ; encoding: [0x05,0x00,0x67,0xd7,0x7e,0x82,0x01,0x00]
+v_minimum_f16 v5.l, exec_lo, -1
+// GFX12: v_minimum_f16 v5.l, exec_lo, -1         ; encoding: [0x05,0x00,0x67,0xd7,0x7e,0x82,0x01,0x00]
 
-v_minimum_f16 v5, |exec_hi|, null
-// GFX12: v_minimum_f16 v5, |exec_hi|, null       ; encoding: [0x05,0x01,0x67,0xd7,0x7f,0xf8,0x00,0x00]
+v_minimum_f16 v5.l, |exec_hi|, null
+// GFX12: v_minimum_f16 v5.l, |exec_hi|, null     ; encoding: [0x05,0x01,0x67,0xd7,0x7f,0xf8,0x00,0x00]
 
-v_minimum_f16 v5, null, exec_lo
-// GFX12: v_minimum_f16 v5, null, exec_lo         ; encoding: [0x05,0x00,0x67,0xd7,0x7c,0xfc,0x00,0x00]
+v_minimum_f16 v5.l, null, exec_lo
+// GFX12: v_minimum_f16 v5.l, null, exec_lo       ; encoding: [0x05,0x00,0x67,0xd7,0x7c,0xfc,0x00,0x00]
 
-v_minimum_f16 v5, -1, exec_hi
-// GFX12: v_minimum_f16 v5, -1, exec_hi           ; encoding: [0x05,0x00,0x67,0xd7,0xc1,0xfe,0x00,0x00]
+v_minimum_f16 v5.l, -1, exec_hi
+// GFX12: v_minimum_f16 v5.l, -1, exec_hi         ; encoding: [0x05,0x00,0x67,0xd7,0xc1,0xfe,0x00,0x00]
 
-v_minimum_f16 v5, 0.5, -m0
-// GFX12: v_minimum_f16 v5, 0.5, -m0              ; encoding: [0x05,0x00,0x67,0xd7,0xf0,0xfa,0x00,0x40]
+v_minimum_f16 v5.l, 0.5, -m0
+// GFX12: v_minimum_f16 v5.l, 0.5, -m0            ; encoding: [0x05,0x00,0x67,0xd7,0xf0,0xfa,0x00,0x40]
 
-v_minimum_f16 v5, -src_scc, |vcc_lo|
-// GFX12: v_minimum_f16 v5, -src_scc, |vcc_lo|    ; encoding: [0x05,0x02,0x67,0xd7,0xfd,0xd4,0x00,0x20]
+v_minimum_f16 v5.l, -src_scc, |vcc_lo|
+// GFX12: v_minimum_f16 v5.l, -src_scc, |vcc_lo|  ; encoding: [0x05,0x02,0x67,0xd7,0xfd,0xd4,0x00,0x20]
 
-v_minimum_f16 v255, -|0xaf12|, -|vcc_hi|
-// GFX12: v_minimum_f16 v255, -|0xaf12|, -|vcc_hi| ; encoding: [0xff,0x03,0x67,0xd7,0xff,0xd6,0x00,0x60,0x12,0xaf,0x00,0x00]
+v_minimum_f16 v255.l, -|0xaf12|, -|vcc_hi|
+// GFX12: v_minimum_f16 v255.l, -|0xaf12|, -|vcc_hi| ; encoding: [0xff,0x03,0x67,0xd7,0xff,0xd6,0x00,0x60,0x12,0xaf,0x00,0x00]
 
-v_minimum_f16 v205, v201, v200
-// GFX12: v_minimum_f16 v205, v201, v200          ; encoding: [0xcd,0x00,0x67,0xd7,0xc9,0x91,0x03,0x00]
+v_minimum_f16 v205.l, v201.l, v200.l
+// GFX12: v_minimum_f16 v205.l, v201.l, v200.l    ; encoding: [0xcd,0x00,0x67,0xd7,0xc9,0x91,0x03,0x00]
 
-v_maximum_f16 v5, v1, v2
-// GFX12: v_maximum_f16 v5, v1, v2                ; encoding: [0x05,0x00,0x68,0xd7,0x01,0x05,0x02,0x00]
+v_minimum_f16 v5.l, v1.h, v2.l
+// GFX12: v_minimum_f16 v5.l, v1.h, v2.l op_sel:[1,0,0] ; encoding: [0x05,0x08,0x67,0xd7,0x01,0x05,0x02,0x00]
 
-v_maximum_f16 v5, v255, v255
-// GFX12: v_maximum_f16 v5, v255, v255            ; encoding: [0x05,0x00,0x68,0xd7,0xff,0xff,0x03,0x00]
+v_minimum_f16 v5.l, v255.l, v255.h
+// GFX12: v_minimum_f16 v5.l, v255.l, v255.h op_sel:[0,1,0] ; encoding: [0x05,0x10,0x67,0xd7,0xff,0xff,0x03,0x00]
 
-v_maximum_f16 v5, s1, s2
-// GFX12: v_maximum_f16 v5, s1, s2                ; encoding: [0x05,0x00,0x68,0xd7,0x01,0x04,0x00,0x00]
+v_minimum_f16 v255.h, 0xfe0b, vcc_hi
+// GFX12: v_minimum_f16 v255.h, 0xfe0b, vcc_hi op_sel:[0,0,1] ; encoding: [0xff,0x40,0x67,0xd7,0xff,0xd6,0x00,0x00,0x0b,0xfe,0x00,0x00]
 
-v_maximum_f16 v5, s105, s105
-// GFX12: v_maximum_f16 v5, s105, s105            ; encoding: [0x05,0x00,0x68,0xd7,0x69,0xd2,0x00,0x00]
+v_maximum_f16 v5.l, v1.l, v2.l
+// GFX12: v_maximum_f16 v5.l, v1.l, v2.l          ; encoding: [0x05,0x00,0x68,0xd7,0x01,0x05,0x02,0x00]
 
-v_maximum_f16 v5, vcc_lo, ttmp15
-// GFX12: v_maximum_f16 v5, vcc_lo, ttmp15        ; encoding: [0x05,0x00,0x68,0xd7,0x6a,0xf6,0x00,0x00]
+v_maximum_f16 v5.l, v255.l, v255.l
+// GFX12: v_maximum_f16 v5.l, v255.l, v255.l      ; encoding: [0x05,0x00,0x68,0xd7,0xff,0xff,0x03,0x00]
 
-v_maximum_f16 v5, vcc_hi, 0xaf12
-// GFX12: v_maximum_f16 v5, vcc_hi, 0xaf12        ; encoding: [0x05,0x00,0x68,0xd7,0x6b,0xfe,0x01,0x00,0x12,0xaf,0x00,0x00]
+v_maximum_f16 v5.l, s1, s2
+// GFX12: v_maximum_f16 v5.l, s1, s2              ; encoding: [0x05,0x00,0x68,0xd7,0x01,0x04,0x00,0x00]
 
-v_maximum_f16 v5, ttmp15, src_scc
-// GFX12: v_maximum_f16 v5, ttmp15, src_scc       ; encoding: [0x05,0x00,0x68,0xd7,0x7b,0xfa,0x01,0x00]
+v_maximum_f16 v5.l, s105, s105
+// GFX12: v_maximum_f16 v5.l, s105, s105          ; encoding: [0x05,0x00,0x68,0xd7,0x69,0xd2,0x00,0x00]
 
-v_maximum_f16 v5, m0, 0.5
-// GFX12: v_maximum_f16 v5, m0, 0.5               ; encoding: [0x05,0x00,0x68,0xd7,0x7d,0xe0,0x01,0x00]
+v_maximum_f16 v5.l, vcc_lo, ttmp15
+// GFX12: v_maximum_f16 v5.l, vcc_lo, ttmp15      ; encoding: [0x05,0x00,0x68,0xd7,0x6a,0xf6,0x00,0x00]
 
-v_maximum_f16 v5, exec_lo, -1
-// GFX12: v_maximum_f16 v5, exec_lo, -1           ; encoding: [0x05,0x00,0x68,0xd7,0x7e,0x82,0x01,0x00]
+v_maximum_f16 v5.l, vcc_hi, 0xaf12
+// GFX12: v_maximum_f16 v5.l, vcc_hi, 0xaf12      ; encoding: [0x05,0x00,0x68,0xd7,0x6b,0xfe,0x01,0x00,0x12,0xaf,0x00,0x00]
 
-v_maximum_f16 v5, |exec_hi|, null
-// GFX12: v_maximum_f16 v5, |exec_hi|, null       ; encoding: [0x05,0x01,0x68,0xd7,0x7f,0xf8,0x00,0x00]
+v_maximum_f16 v5.l, ttmp15, src_scc
+// GFX12: v_maximum_f16 v5.l, ttmp15, src_scc     ; encoding: [0x05,0x00,0x68,0xd7,0x7b,0xfa,0x01,0x00]
 
-v_maximum_f16 v5, null, exec_lo
-// GFX12: v_maximum_f16 v5, null, exec_lo         ; encoding: [0x05,0x00,0x68,0xd7,0x7c,0xfc,0x00,0x00]
+v_maximum_f16 v5.l, m0, 0.5
+// GFX12: v_maximum_f16 v5.l, m0, 0.5             ; encoding: [0x05,0x00,0x68,0xd7,0x7d,0xe0,0x01,0x00]
 
-v_maximum_f16 v5, -1, exec_hi
-// GFX12: v_maximum_f16 v5, -1, exec_hi           ; encoding: [0x05,0x00,0x68,0xd7,0xc1,0xfe,0x00,0x00]
+v_maximum_f16 v5.l, exec_lo, -1
+// GFX12: v_maximum_f16 v5.l, exec_lo, -1         ; encoding: [0x05,0x00,0x68,0xd7,0x7e,0x82,0x01,0x00]
 
-v_maximum_f16 v5, 0.5, -m0
-// GFX12: v_maximum_f16 v5, 0.5, -m0              ; encoding: [0x05,0x00,0x68,0xd7,0xf0,0xfa,0x00,0x40]
+v_maximum_f16 v5.l, |exec_hi|, null
+// GFX12: v_maximum_f16 v5.l, |exec_hi|, null     ; encoding: [0x05,0x01,0x...
[truncated]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

kosarev · 2025-01-27T12:46:13Z

llvm/lib/Target/AMDGPU/VOPInstructions.td

+
+multiclass VOP3Only_Realtriple_t16_and_fake16_gfx12<bits<10> op, string asmName,
+                                     string opName = NAME, string pseudo_mnemonic = ""> {
+  defm _t16 : VOP3Only_Realtriple_t16_gfx12<op, asmName, opName#"_t16", pseudo_mnemonic>;


Shouldn't VOP3Only_Realtriple_t16/fake16_gfx12 do the opName#"_t16/_fake16" thing themselves?

For now "t16_and_fake16" define t16 and fake16 together.

In the end the fake16 will be removed and the change will be something like VOP3Only_Realtriple_t16_and_fake16_gfx12 -> VOP3Only_Realtriple_t16_gfx12. There seems not necessary to add a fake16 function

Sisyph

LGTM

Sisyph · 2025-01-27T16:17:21Z

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

+  case AMDGPU::S_MINIMUM_F16:
+  case AMDGPU::S_MAXIMUM_F16: {
+    const DebugLoc &DL = Inst.getDebugLoc();
+    Register NewDst = MRI.createVirtualRegister(ST.useRealTrue16Insts()


Nit: As Matt suggested, you could do something like
Register NewDst = MRI.createVirtualRegister(MRI->getNamedOperand(NewOpcode, AMDGPU::vdst).getRegClass())). However, we still can't combine the 16 and 32 bit switch cases due to the opsel0 operand.

Sisyph · 2025-01-27T16:17:56Z

llvm/lib/Target/AMDGPU/SIInstructions.td

-def : FPMinCanonMaxPat<V_MAXIMUMMINIMUM_F16_e64, f16, DivergentBinFrag<fminimum>, fmaximum_oneuse>;
+}
+
+let True16Predicate = UseFakeTrue16Insts, OtherPredicates = [isGFX12Plus] in {


Suggested change

let True16Predicate = UseFakeTrue16Insts, OtherPredicates = [isGFX12Plus] in {

let True16Predicate = UseFakeTrue16Insts, SubtargetPredicate = isGFX12Plus in {

This is a nit

vvereschaka · 2025-01-29T03:58:53Z

Hi @broxigarchen,

these changes break LLVM::llvm.amdgcn.mfma.gfx950.ll test on the expensive check builders with the following messages:

warning: Instruction selection used fallback path for test_mfma_f32_16x16x32_bf16
warning: Instruction selection used fallback path for test_mfma_f32_16x16x32_bf16__flags
warning: Instruction selection used fallback path for test_mfma_f32_16x16x32_bf16_no_agpr__vgprcd
error: <unknown>:0:0: no registers from class available to allocate in function 'test_mfma_f32_16x16x32_bf16_no_agpr__vgprcd'

https://lab.llvm.org/buildbot/#/builders/187/builds/4011/steps/6/logs/FAIL__LLVM__llvm_amdgcn_mfma_gfx950_ll

the first failed build
https://lab.llvm.org/buildbot/#/builders/187/builds/4011

would you fix these as soon as possible or revert the changes?

arsenm · 2025-01-29T04:02:52Z

Hi @broxigarchen,

these changes break LLVM::llvm.amdgcn.mfma.gfx950.ll test on the expensive check builders with the following messages:
warning: Instruction selection used fallback path for test_mfma_f32_16x16x32_bf16
warning: Instruction selection used fallback path for test_mfma_f32_16x16x32_bf16__flags
warning: Instruction selection used fallback path for test_mfma_f32_16x16x32_bf16_no_agpr__vgprcd
error: <unknown>:0:0: no registers from class available to allocate in function 'test_mfma_f32_16x16x32_bf16_no_agpr__vgprcd'
https://lab.llvm.org/buildbot/#/builders/187/builds/4011/steps/6/logs/FAIL__LLVM__llvm_amdgcn_mfma_gfx950_ll

the first failed build https://lab.llvm.org/buildbot/#/builders/187/builds/4011

would you fix these as soon as possible or revert the changes?

I think you have the wrong patch, this is #124799

vvereschaka · 2025-01-29T04:16:41Z

hm, #124799 has been merged 10 hours ago, but the first test failure has occurred apprx 24 hours ago.
I'll better check

arsenm · 2025-01-29T04:19:19Z

hm, #124799 has been merged 10 hours ago, but the first test failure has occurred apprx 24 hours ago. I'll better check

It is not merged yet

vvereschaka · 2025-01-29T04:26:39Z

yes, my bad. Sorry for false alarm.
Looks like #124113 breaks the test and #124799 is a possible fix.

broxigarchen force-pushed the main-merge-true16-vop3-mc-more-instructions-11 branch from 536f8ea to 190473a Compare January 23, 2025 21:36

broxigarchen force-pushed the main-merge-true16-vop3-mc-more-instructions-11 branch from 190473a to 097f93a Compare January 23, 2025 21:42

broxigarchen marked this pull request as ready for review January 23, 2025 23:54

broxigarchen requested review from Sisyph, arsenm and kosarev January 23, 2025 23:54

llvmbot added backend:AMDGPU llvm:mc Machine (object) code labels Jan 23, 2025

arsenm reviewed Jan 24, 2025

View reviewed changes

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp Outdated Show resolved Hide resolved

kosarev reviewed Jan 27, 2025

View reviewed changes

broxigarchen force-pushed the main-merge-true16-vop3-mc-more-instructions-11 branch from 4e6b113 to 5721a81 Compare January 27, 2025 16:11

broxigarchen requested review from arsenm and kosarev January 27, 2025 16:12

kosarev approved these changes Jan 27, 2025

View reviewed changes

Sisyph approved these changes Jan 27, 2025

View reviewed changes

broxigarchen force-pushed the main-merge-true16-vop3-mc-more-instructions-11 branch 2 times, most recently from cfa1485 to 538f29c Compare January 27, 2025 18:38

true16 for minimummaximum/maximumminimum/maximum/minimum

9c6da9b

broxigarchen force-pushed the main-merge-true16-vop3-mc-more-instructions-11 branch from 538f29c to 9c6da9b Compare January 27, 2025 19:46

broxigarchen merged commit 5d1c596 into llvm:main Jan 27, 2025
6 of 7 checks passed

	let True16Predicate = UseFakeTrue16Insts, OtherPredicates = [isGFX12Plus] in {
	let True16Predicate = UseFakeTrue16Insts, SubtargetPredicate = isGFX12Plus in {

[AMDGPU][True16][MC] true16 for minimummaximum/max/min/max3/min3 #124184

[AMDGPU][True16][MC] true16 for minimummaximum/max/min/max3/min3 #124184

Uh oh!

Conversation

broxigarchen commented Jan 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Jan 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Jan 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

kosarev Jan 27, 2025

Choose a reason for hiding this comment

Uh oh!

broxigarchen Jan 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kosarev Jan 27, 2025

Choose a reason for hiding this comment

Uh oh!

Sisyph left a comment

Choose a reason for hiding this comment

Uh oh!

Sisyph Jan 27, 2025

Choose a reason for hiding this comment

Uh oh!

Sisyph Jan 27, 2025

Choose a reason for hiding this comment

Uh oh!

Sisyph Jan 27, 2025

Choose a reason for hiding this comment

Uh oh!

broxigarchen Jan 27, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

vvereschaka commented Jan 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

arsenm commented Jan 29, 2025

Uh oh!

vvereschaka commented Jan 29, 2025

Uh oh!

arsenm commented Jan 29, 2025

Uh oh!

vvereschaka commented Jan 29, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

broxigarchen commented Jan 23, 2025 •

edited

Loading

github-actions bot commented Jan 23, 2025 •

edited

Loading

llvmbot commented Jan 23, 2025 •

edited

Loading

broxigarchen Jan 27, 2025 •

edited

Loading

vvereschaka commented Jan 29, 2025 •

edited

Loading