[PATCH] D25975: AMDGPU/SI: Make f16 a legal type for VI subtargets

Konstantin Zhuravlyov via llvm-commits llvm-commits at lists.llvm.org
Wed Nov 9 10:27:55 PST 2016


kzhuravl added inline comments.


================
Comment at: lib/Target/AMDGPU/VOP2Instructions.td:326
+
+// FIXME: V_LDEXP_F16 requires a change to llvm.amdgcn.ldexp intrinsic.
+defm V_LDEXP_F16 : VOP2Inst <"v_ldexp_f16", VOP_F16_F16_I16/*, AMDGPUldexp*/>;
----------------
kzhuravl wrote:
> arsenm wrote:
> > Why is this necessary? It can already mangle the FP type. This should work if you change it to VOP_F16_F16_I32, the int type doesn't matter
> Would this be acceptable to use VOP_F16_F16_I32? Spec says:
> ```
> D.f16 = S0.f16 * (2 ** S1.i16)
> ```
> So it should be i16. I have a follow up patch that changes ldexp intrinsic.
Changed to VOP_F16_F16_I32.


https://reviews.llvm.org/D25975





More information about the llvm-commits mailing list