> On Feb 25, 2017, at 15:58, Marek Olšák <mar...@gmail.com> wrote:
>
> }
> +
> +LLVMValueRef ac_emit_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
> + LLVMValueRef args[2])
> +{
> + if (HAVE_LLVM >= 0x0500) {
> + LLVMTypeRef v2f16 =
> + LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
> + LLVMValueRef res =
> + ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz",
> + v2f16, args, 2,
> + AC_FUNC_ATTR_READNONE);
> + return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
> + }
> +
> + return ac_emit_llvm_intrinsic(ctx, "llvm.SI.packf16", ctx->i32, args, 2,
> + AC_FUNC_ATTR_READNONE |
> + AC_FUNC_ATTR_LEGACY);
> +}
> diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
> index f57acc2..e6bb
I would probably swap where you do the bitcast here, and cast the legacy
intrinsic to <2 x half>. With the pkrtz -> exp.compr, pattern you’ll be
emitting a cast to i32 and back. While that will be cleaned up, you’re wasting
some compile time/memory doing so.
-Matt
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
