replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[NVPTX] Compute approx sqrt as 1/rsqrt(x) rather than x*rsqrt(x). 

x*rsqrt(x) returns NaN for x == 0, whereas 1/rsqrt(x) returns 0, as
desired.

Verified that the particular nvptx approximate instructions here do in
fact return 0 for x = 0.

llvm-svn: 293713
This commit is contained in:
Justin Lebar 2017-01-31 23:08:57 +00:00
parent d9953d9dd2
commit 06fcea4cd9
3 changed files with 15 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
View File

@ -1080,8 +1080,13 @@ SDValue NVPTXTargetLowering::getSqrtEstimate(SDValue Operand, SelectionDAG &DAG,
return MakeIntrinsicCall(Ftz ? Intrinsic::nvvm_sqrt_approx_ftz_f
: Intrinsic::nvvm_sqrt_approx_f);
else {
// There's no sqrt.approx.f64 instruction, so we emit x * rsqrt(x).
return DAG.getNode(ISD::FMUL, DL, VT, Operand,
// There's no sqrt.approx.f64 instruction, so we emit
// reciprocal(rsqrt(x)). This is faster than
// select(x == 0, 0, x * rsqrt(x)). (In fact, it's faster than plain
// x * rsqrt(x).)
return DAG.getNode(
ISD::INTRINSIC_WO_CHAIN, DL, VT,
DAG.getConstant(Intrinsic::nvvm_rcp_approx_ftz_d, DL, MVT::i32),
MakeIntrinsicCall(Intrinsic::nvvm_rsqrt_approx_d));
}
}

4
llvm/test/CodeGen/NVPTX/fast-math.ll
View File

@ -40,11 +40,11 @@ define float @sqrt_div_fast_ftz(float %a, float %b) #0 #1 {
}
; There are no fast-math or ftz versions of sqrt and div for f64. We use
; x * rsqrt(x) for sqrt(x), and emit a vanilla divide.
; reciprocal(rsqrt(x)) for sqrt(x), and emit a vanilla divide.
; CHECK-LABEL: sqrt_div_fast_ftz_f64(
; CHECK: rsqrt.approx.f64
; CHECK: mul.f64
; CHECK: rcp.approx.ftz.f64
; CHECK: div.rn.f64
define double @sqrt_div_fast_ftz_f64(double %a, double %b) #0 #1 {
%t1 = tail call double @llvm.sqrt.f64(double %a)

8
llvm/test/CodeGen/NVPTX/sqrt-approx.ll
View File

@ -59,9 +59,11 @@ define float @test_sqrt_ftz(float %a) #0 #1 {
; CHECK-LABEL test_sqrt64
define double @test_sqrt64(double %a) #0 {
; There's no sqrt.approx.f64 instruction; we emit x * rsqrt.approx.f64(x).
; There's no sqrt.approx.f64 instruction; we emit
; reciprocal(rsqrt.approx.f64(x)). There's no non-ftz approximate reciprocal,
; so we just use the ftz version.
; CHECK: rsqrt.approx.f64
; CHECK: mul.f64
; CHECK: rcp.approx.ftz.f64
%ret = tail call double @llvm.sqrt.f64(double %a)
ret double %ret
}
@ -70,7 +72,7 @@ define double @test_sqrt64(double %a) #0 {
define double @test_sqrt64_ftz(double %a) #0 #1 {
; There's no sqrt.approx.ftz.f64 instruction; we just use the non-ftz version.
; CHECK: rsqrt.approx.f64
; CHECK: mul.f64
; CHECK: rcp.approx.ftz.f64
%ret = tail call double @llvm.sqrt.f64(double %a)
ret double %ret
}