6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-01-12 17:32:19 +00:00
Code Issues Projects Releases Wiki Activity

R600/SI: Allow using f32 rcp / rsq when denormals not handled.

Browse Source 
These are precise enough to use for OpenCL unless denormals
are handled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213107 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Matt Arsenault 2014-07-15 23:50:10 +00:00
parent 6f7532bb97
commit a27bf373d9
6 changed files with 45 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
lib/Target/R600/SIISelLowering.cpp
View File

@ -936,16 +936,27 @@ SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res);
}
static SDValue performUnsafeFDIV(SDValue Op, SelectionDAG &DAG) {
// Catch division cases where we can use shortcuts with rcp and rsq
// instructions.
SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const {
SDLoc SL(Op);
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
EVT VT = Op.getValueType();
bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
if (CLHS->isExactlyValue(1.0)) {
if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
CLHS->isExactlyValue(1.0)) {
// v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
// the CI documentation has a worst case error of 1 ulp.
// OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
// use it as long as we aren't trying to use denormals.
// 1.0 / sqrt(x) -> rsq(x)
//
// XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
// error seems really high at 2^29 ULP.
if (RHS.getOpcode() == ISD::FSQRT)
return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
@ -954,15 +965,25 @@ static SDValue performUnsafeFDIV(SDValue Op, SelectionDAG &DAG) {
}
}
// Turn into multiply by the reciprocal
// x / y -> x * (1.0 / y)
SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip);
if (Unsafe) {
// Turn into multiply by the reciprocal.
// x / y -> x * (1.0 / y)
SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip);
}
return SDValue();
}
SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
if (DAG.getTarget().Options.UnsafeFPMath)
return performUnsafeFDIV(Op, DAG);
SDValue FastLowered = LowerFastFDIV(Op, DAG);
if (FastLowered.getNode())
return FastLowered;
// This uses v_rcp_f32 which does not handle denormals. Let this hit a
// selection error for now rather than do something incorrect.
if (Subtarget->hasFP32Denormals())
return SDValue();
SDLoc SL(Op);
SDValue LHS = Op.getOperand(0);

1
lib/Target/R600/SIISelLowering.h
View File

@ -27,6 +27,7 @@ class SITargetLowering : public AMDGPUTargetLowering {
SelectionDAG &DAG) const;
SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFDIV32(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFDIV64(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFDIV(SDValue Op, SelectionDAG &DAG) const;

3
lib/Target/R600/SIInstructions.td
View File

@ -1800,10 +1800,9 @@ def : Pat <
// VOP1 Patterns
//===----------------------------------------------------------------------===//
let Predicates = [UnsafeFPMath] in {
def : RcpPat<V_RCP_F64_e32, f64>;
defm : RsqPat<V_RSQ_F64_e32, f64>;
let Predicates = [UnsafeFPMath] in {
defm : RsqPat<V_RSQ_F32_e32, f32>;
}

3
test/CodeGen/R600/llvm.AMDGPU.rcp.f64.ll
View File

@ -20,7 +20,8 @@ define void @rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind {
}
; FUNC-LABEL: @rsq_rcp_pat_f64
; SI: V_RSQ_F64_e32
; SI-UNSAFE: V_RSQ_F64_e32
; SI-SAFE-NOT: V_RSQ_F64_e32
define void @rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind {
%sqrt = call double @llvm.sqrt.f64(double %src) nounwind readnone
%rcp = call double @llvm.AMDGPU.rcp.f64(double %sqrt) nounwind readnone

18
test/CodeGen/R600/llvm.AMDGPU.rcp.ll
View File

@ -1,5 +1,7 @@
; RUN: llc -march=r600 -mcpu=SI -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI -check-prefix=FUNC %s
; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI -check-prefix=FUNC %s
; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI -check-prefix=FUNC %s
; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI -check-prefix=FUNC %s
; XUN: llc -march=r600 -mcpu=SI -mattr=+fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE-SPDENORM -check-prefix=SI -check-prefix=FUNC %s
declare float @llvm.AMDGPU.rcp.f32(float) nounwind readnone
declare double @llvm.AMDGPU.rcp.f64(double) nounwind readnone
@ -25,15 +27,8 @@ define void @rcp_f64(double addrspace(1)* %out, double %src) nounwind {
}
; FUNC-LABEL: @rcp_pat_f32
; SI-UNSAFE-NOT: V_MUL_F32
; SI-UNSAFE: V_RCP_F32_e32
; SI-UNSAFE-NOT: V_MUL_F32
; Check for surrounding multiplies the correct divide has.
; SI-SAFE: V_MUL_F32
; SI-SAFE: V_RCP_F32_e32
; SI-SAFE: V_MUL_F32
; XSI-SAFE-SPDENORM-NOT: V_RCP_F32_e32
define void @rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind {
%rcp = fdiv float 1.0, %src
store float %rcp, float addrspace(1)* %out, align 4
@ -60,7 +55,8 @@ define void @rsq_rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind {
}
; FUNC-LABEL: @rsq_rcp_pat_f64
; SI: V_RSQ_F64_e32
; SI-UNSAFE: V_RSQ_F64_e32
; SI-SAFE-NOT: V_RSQ_F64_e32
define void @rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind {
%sqrt = call double @llvm.sqrt.f64(double %src) nounwind readnone
%rcp = call double @llvm.AMDGPU.rcp.f64(double %sqrt) nounwind readnone

10
test/CodeGen/R600/rsq.ll
View File

@ -1,12 +1,11 @@
; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI %s
; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI %s
; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI %s
; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI %s
declare float @llvm.sqrt.f32(float) nounwind readnone
declare double @llvm.sqrt.f64(double) nounwind readnone
; SI-LABEL: @rsq_f32
; SI-UNSAFE: V_RSQ_F32_e32
; SI-SAFE: V_SQRT_F32
; SI: V_RSQ_F32_e32
; SI: S_ENDPGM
define void @rsq_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
%val = load float addrspace(1)* %in, align 4
@ -17,7 +16,8 @@ define void @rsq_f32(float addrspace(1)* noalias %out, float addrspace(1)* noali
}
; SI-LABEL: @rsq_f64
; SI: V_RSQ_F64_e32
; SI-UNSAFE: V_RSQ_F64_e32
; SI-SAFE: V_SQRT_F64_e32
; SI: S_ENDPGM
define void @rsq_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
%val = load double addrspace(1)* %in, align 4