AMDGPU: Add support for V_CVT_PK_F16_F32 instruction for gfx950 (#118300)

Co-authored-by: Shilei Tian <[email protected]>

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -432,6 +432,12 @@ def FeatureAshrPkInsts : SubtargetFeature<"ashr-pk-insts",
   "Has Arithmetic Shift Pack instructions"
 >;
 
+def FeatureCvtPkF16F32Inst : SubtargetFeature<"cvt-pk-f16-f32-inst",
+  "HasCvtPkF16F32Inst",
+  "true",
+  "Has cvt_pk_f16_f32 instruction"
+>;
+
 def FeatureGFX950Insts : SubtargetFeature<"gfx950-insts",
   "GFX950Insts",
   "true",
@@ -445,8 +451,9 @@ def FeatureGFX950Insts : SubtargetFeature<"gfx950-insts",
    FeatureFP6BF6ConversionScaleInsts,
    FeatureF16BF16ToFP6BF6ConversionScaleInsts,
    FeatureF32ToF16BF16ConversionSRInsts,
+   FeatureCvtPkF16F32Inst,
    FeatureMinimum3Maximum3F32,
-   FeatureMinimum3Maximum3PKF16
+   FeatureMinimum3Maximum3PKF16,
    ]
 >;
 
@@ -2510,6 +2517,9 @@ def HasFP6BF6ConversionScaleInsts : Predicate<"Subtarget->hasFP6BF6ConversionSca
 def HasF16BF16ToFP6BF6ConversionScaleInsts : Predicate<"Subtarget->hasF16BF16ToFP6BF6ConversionScaleInsts()">,
   AssemblerPredicate<(all_of FeatureF16BF16ToFP6BF6ConversionScaleInsts)>;
 
+def HasCvtPkF16F32Inst : Predicate<"Subtarget->hasCvtPkF16F32Inst()">,
+  AssemblerPredicate<(all_of FeatureCvtPkF16F32Inst)>;
+
 def HasF32ToF16BF16ConversionSRInsts : Predicate<"Subtarget->hasF32ToF16BF16ConversionSRInsts()">,
   AssemblerPredicate<(all_of FeatureF32ToF16BF16ConversionSRInsts)>;
 

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -1040,10 +1040,13 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
       .lower();
   }
 
-  getActionDefinitionsBuilder(G_FPTRUNC)
-    .legalFor({{S32, S64}, {S16, S32}})
-    .scalarize(0)
-    .lower();
+  auto &FPTruncActions = getActionDefinitionsBuilder(G_FPTRUNC);
+  if (ST.hasCvtPkF16F32Inst())
+    FPTruncActions.legalFor(
+        {{S32, S64}, {S16, S32}, {V2S16, V2S32}, {V2S16, V2S64}});
+  else
+    FPTruncActions.legalFor({{S32, S64}, {S16, S32}});
+  FPTruncActions.scalarize(0).lower();
 
   getActionDefinitionsBuilder(G_FPEXT)
     .legalFor({{S64, S32}, {S32, S16}})

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

@@ -55,6 +55,7 @@ class AMDGPUSubtarget {
   bool HasFP4ConversionScaleInsts = false;
   bool HasFP6BF6ConversionScaleInsts = false;
   bool HasF16BF16ToFP6BF6ConversionScaleInsts = false;
+  bool HasCvtPkF16F32Inst = false;
   bool HasF32ToF16BF16ConversionSRInsts = false;
   bool EnableRealTrue16Insts = false;
   bool HasBF16ConversionInsts = false;
@@ -191,6 +192,8 @@ class AMDGPUSubtarget {
 
   bool hasF16BF16ToFP6BF6ConversionScaleInsts() const { return HasF16BF16ToFP6BF6ConversionScaleInsts; }
 
+  bool hasCvtPkF16F32Inst() const { return HasCvtPkF16F32Inst; }
+
   bool hasF32ToF16BF16ConversionSRInsts() const {
     return HasF32ToF16BF16ConversionSRInsts;
   }

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -902,6 +902,10 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::BUILD_VECTOR, MVT::v2bf16, Legal);
   }
 
+  if (Subtarget->hasCvtPkF16F32Inst()) {
+    setOperationAction(ISD::FP_ROUND, MVT::v2f16, Legal);
+  }
+
   setTargetDAGCombine({ISD::ADD,
                        ISD::UADDO_CARRY,
                        ISD::SUB,

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -1147,6 +1147,21 @@ let SubtargetPredicate = HasGFX950Insts, mayRaiseFPException = 0 in {
   defm V_CVT_SCALEF32_2XPK16_BF6_F32 : VOP3Inst<"v_cvt_scalef32_2xpk16_bf6_f32",  VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V6I32_V16F32_V16F32_F32>,  int_amdgcn_cvt_scalef32_2xpk16_bf6_f32>;
 }
 
+let SubtargetPredicate = HasCvtPkF16F32Inst in {
+  let ReadsModeReg = 0 in {
+    defm V_CVT_PK_F16_F32 : VOP3Inst<"v_cvt_pk_f16_f32", VOP3_Profile<VOP_V2F16_F32_F32>>;
+  }
+
+  def : GCNPat<(v2f16 (fpround v2f32:$src)),
+               (V_CVT_PK_F16_F32_e64 0, (EXTRACT_SUBREG VReg_64:$src, sub0), 0, (EXTRACT_SUBREG VReg_64:$src, sub1))>;
+  def : GCNPat<(v2f16 (fpround v2f64:$src)),
+               (V_CVT_PK_F16_F32_e64 0, (V_CVT_F32_F64_e64 0, (EXTRACT_SUBREG VReg_128:$src, sub0_sub1)),
+                                     0, (V_CVT_F32_F64_e64 0, (EXTRACT_SUBREG VReg_128:$src, sub2_sub3)))>;
+  def : GCNPat<(v2f16 (build_vector (f16 (fpround (f32 (VOP3Mods f32:$src0, i32:$src0_modifiers)))),
+                                    (f16 (fpround (f32 (VOP3Mods f32:$src1, i32:$src1_modifiers)))))),
+               (V_CVT_PK_F16_F32_e64 $src0_modifiers, $src0, $src1_modifiers, $src1)>;
+}
+
 class Cvt_Scale_FP4FP8BF8ToF16F32_Pat<SDPatternOperator node, VOP3_Pseudo inst, ValueType DstTy> : GCNPat<
     (DstTy (node i32:$src0, f32:$src1, timm:$index)),
     (inst (SrcAndDstSelToOpSelXForm_0_0 $index), $src0, (SrcAndDstSelToOpSelXForm_1_0 $index), $src1)
@@ -2274,6 +2289,9 @@ defm V_CVT_SR_BF16_F32: VOP3OpSel_Real_gfx9 <0x2a7>;
 
 defm V_ASHR_PK_I8_I32 : VOP3OpSel_Real_gfx9 <0x265>;
 defm V_ASHR_PK_U8_I32 : VOP3OpSel_Real_gfx9 <0x266>;
+let OtherPredicates = [HasCvtPkF16F32Inst] in {
+defm V_CVT_PK_F16_F32 : VOP3_Real_gfx9<0x267, "v_cvt_pk_f16_f32">;
+}
 
 defm V_CVT_SCALEF32_2XPK16_FP6_F32 : VOP3_Real_gfx9<0x252, "v_cvt_scalef32_2xpk16_fp6_f32">;
 defm V_CVT_SCALEF32_2XPK16_BF6_F32 : VOP3_Real_gfx9<0x253, "v_cvt_scalef32_2xpk16_bf6_f32">;