AMDGPU: Add V_CVT_PK_BF16_F32 for gfx950 (#116678)

Author: arsenm

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -889,6 +889,12 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
 
   setOperationAction(ISD::MUL, MVT::i1, Promote);
 
+  if (Subtarget->hasBF16ConversionInsts()) {
+    setOperationAction(ISD::FP_ROUND, MVT::v2bf16, Legal);
+    setOperationAction(ISD::FP_ROUND, MVT::bf16, Legal);
+    setOperationAction(ISD::BUILD_VECTOR, MVT::v2bf16, Legal);
+  }
+
   setTargetDAGCombine({ISD::ADD,
                        ISD::UADDO_CARRY,
                        ISD::SUB,

llvm/lib/Target/AMDGPU/SIInstrInfo.td

@@ -2787,6 +2787,7 @@ def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>;
 def VOP_I32_I32_I32_ARITH : VOPProfile <[i32, i32, i32, untyped], /*EnableClamp=*/1>;
 def VOP_V2F16_F32_F32 : VOPProfile <[v2f16, f32, f32, untyped]>;
 def VOP_F32_F16_F16_F16 : VOPProfile <[f32, f16, f16, f16]>;
+def VOP_V2BF16_F32_F32 : VOPProfile <[v2bf16, f32, f32, untyped]>;
 
 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
 def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>;

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -944,6 +944,30 @@ let SubtargetPredicate = isGFX11Plus in {
   defm V_CVT_PK_U16_F32 : VOP3Inst<"v_cvt_pk_u16_f32", VOP3_Profile<VOP_V2I16_F32_F32>>;
 } // End SubtargetPredicate = isGFX11Plus
 
+// FIXME: GlobalISel cannot distinguish f16 and bf16 and may start using bf16 patterns
+//        instead of less complex f16. Disable GlobalISel for these for now.
+def bf16_fpround : PatFrag <(ops node:$src0),  (fpround $src0), [{ return true; }]> {
+  let GISelPredicateCode = [{return false;}];
+}
+
+let SubtargetPredicate = HasBF16ConversionInsts in {
+  let ReadsModeReg = 0 in {
+    defm V_CVT_PK_BF16_F32    : VOP3Inst<"v_cvt_pk_bf16_f32", VOP3_Profile<VOP_V2BF16_F32_F32>>;
+  }
+  def : GCNPat<(v2bf16 (bf16_fpround v2f32:$src)),
+               (V_CVT_PK_BF16_F32_e64 0, (EXTRACT_SUBREG VReg_64:$src, sub0), 0, (EXTRACT_SUBREG VReg_64:$src, sub1))>;
+  def : GCNPat<(v2bf16 (bf16_fpround v2f64:$src)),
+               (V_CVT_PK_BF16_F32_e64 0, (V_CVT_F32_F64_e64 0, (EXTRACT_SUBREG VReg_128:$src, sub0_sub1)),
+                                      0, (V_CVT_F32_F64_e64 0, (EXTRACT_SUBREG VReg_128:$src, sub2_sub3)))>;
+  def : GCNPat<(v2bf16 (build_vector (bf16 (bf16_fpround (f32 (VOP3Mods f32:$src0, i32:$src0_modifiers)))),
+                                     (bf16 (bf16_fpround (f32 (VOP3Mods f32:$src1, i32:$src1_modifiers)))))),
+               (V_CVT_PK_BF16_F32_e64 $src0_modifiers, $src0, $src1_modifiers, $src1)>;
+  def : GCNPat<(bf16 (bf16_fpround (f32 (VOP3Mods f32:$src0, i32:$src0_modifiers)))),
+               (V_CVT_PK_BF16_F32_e64 $src0_modifiers, $src0, 0, (f32 (IMPLICIT_DEF)))>;
+  def : GCNPat<(bf16 (bf16_fpround (f64 (VOP3Mods f64:$src0, i32:$src0_modifiers)))),
+               (V_CVT_PK_BF16_F32_e64 0, (f32 (V_CVT_F32_F64_e64 $src0_modifiers, $src0)), 0, (f32 (IMPLICIT_DEF)))>;
+}
+
 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>>;
@@ -1721,5 +1745,6 @@ defm V_LSHL_ADD_U64 : VOP3_Real_vi <0x208>;
 
 defm V_CVT_PK_FP8_F32 : VOP3OpSel_Real_gfx9 <0x2a2>;
 defm V_CVT_PK_BF8_F32 : VOP3OpSel_Real_gfx9 <0x2a3>;
+defm V_CVT_PK_BF16_F32: VOP3OpSel_Real_gfx9 <0x268>;
 defm V_CVT_SR_FP8_F32 : VOP3OpSel_Real_gfx9_forced_opsel2 <0x2a4>;
 defm V_CVT_SR_BF8_F32 : VOP3OpSel_Real_gfx9_forced_opsel2 <0x2a5>;