[llvm] f5b2d24 - [NVPTX] Move roundeven on bf16 into the sm_90 has it but sm_80 doesn't bucket

[Benjamin Kramer via llvm-commits]

Author: Benjamin Kramer
Date: 2024-04-12T11:47:53+02:00
New Revision: f5b2d24b59f168eaeed08b5f45bf0c8dfcc1c292

URL: https://github.com/llvm/llvm-project/commit/f5b2d24b59f168eaeed08b5f45bf0c8dfcc1c292
DIFF: https://github.com/llvm/llvm-project/commit/f5b2d24b59f168eaeed08b5f45bf0c8dfcc1c292.diff

LOG: [NVPTX] Move roundeven on bf16 into the sm_90 has it but sm_80 doesn't bucket

Added: 
    

Modified: 
    llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
    llvm/test/CodeGen/NVPTX/bf16-instructions.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index e1cdbd6a4b4dc6..44b61a937d64e4 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -447,6 +447,7 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
       case ISD::FFLOOR:
       case ISD::FNEARBYINT:
       case ISD::FRINT:
+      case ISD::FROUNDEVEN:
       case ISD::FTRUNC:
         IsOpSupported = STI.getSmVersion() >= 90 && STI.getPTXVersion() >= 78;
         break;

diff  --git a/llvm/test/CodeGen/NVPTX/bf16-instructions.ll b/llvm/test/CodeGen/NVPTX/bf16-instructions.ll
index 884860712632d7..4cbab8bf24166c 100644
--- a/llvm/test/CodeGen/NVPTX/bf16-instructions.ll
+++ b/llvm/test/CodeGen/NVPTX/bf16-instructions.ll
@@ -330,3 +330,15 @@ define bfloat @test_uitofp_i64(i64 %a) {
   %r = uitofp i64 %a to bfloat
   ret bfloat %r
 }
+
+; CHECK-LABEL: test_roundeven(
+; CHECK:      ld.param.b16      [[A:%rs[0-9]+]], [test_roundeven_param_0];
+; SM80:       cvt.rni.f32.f32   [[F:%f[0-9]+]]
+; SM80:       cvt.rn.bf16.f32   [[R:%rs[0-9]+]], [[F]];
+; SM90:       cvt.rni.bf16.bf16 [[R:%rs[0-9]+]], [[A]];
+; CHECK:      st.param.b16      [func_retval0+0], [[R]];
+; CHECK:      ret;
+define bfloat @test_roundeven(bfloat %a) {
+  %r = call bfloat @llvm.roundeven.bf16(bfloat %a)
+  ret bfloat %r
+}