Skip to content

Commit d6cb206

Browse files
tex3dtex3d
committed
[DirectX] Add atan2 intrinsic and expand for DXIL backend
This change is part of this proposal: https://discourse.llvm.org/t/rfc-all-the-math-intrinsics/78294 This preliminary work adds the intrinsic to llvm and expands using atan intrinsic for DXIL backend, since DXIL has no atan2 op.
1 parent 74405b9 commit d6cb206

File tree

5 files changed

+147
-0
lines changed

5 files changed

+147
-0
lines changed

‎llvm/docs/LangRef.rst

Lines changed: 37 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -15583,6 +15583,43 @@ trapping or setting ``errno``.
1558315583
When specified with the fast-math-flag 'afn', the result may be approximated
1558415584
using a less accurate calculation.
1558515585

15586+
'``llvm.atan2.*``' Intrinsic
15587+
^^^^^^^^^^^^^^^^^^^^^^^^^^^
15588+
15589+
Syntax:
15590+
"""""""
15591+
15592+
This is an overloaded intrinsic. You can use ``llvm.atan2`` on any
15593+
floating-point or vector of floating-point type. Not all targets support
15594+
all types however.
15595+
15596+
::
15597+
15598+
declare float @llvm.atan2.f32(float %X, float %Y)
15599+
declare double @llvm.atan2.f64(double %X, double %Y)
15600+
declare x86_fp80 @llvm.atan2.f80(x86_fp80 %X, x86_fp80 %Y)
15601+
declare fp128 @llvm.atan2.f128(fp128 %X, fp128 %Y)
15602+
declare ppc_fp128 @llvm.atan2.ppcf128(ppc_fp128 %X, ppc_fp128 %Y)
15603+
15604+
Overview:
15605+
"""""""""
15606+
15607+
The '``llvm.atan2.*``' intrinsics return the arctangent of the operand.
15608+
15609+
Arguments:
15610+
""""""""""
15611+
15612+
The arguments and return value are floating-point numbers of the same type.
15613+
15614+
Semantics:
15615+
""""""""""
15616+
15617+
Return the same value as a corresponding libm '``atan2``' function but without
15618+
trapping or setting ``errno``.
15619+
15620+
When specified with the fast-math-flag 'afn', the result may be approximated
15621+
using a less accurate calculation.
15622+
1558615623
'``llvm.sinh.*``' Intrinsic
1558715624
^^^^^^^^^^^^^^^^^^^^^^^^^^^
1558815625

‎llvm/include/llvm/IR/Intrinsics.td

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1016,6 +1016,7 @@ let IntrProperties = [IntrNoMem, IntrSpeculatable, IntrWillReturn] in {
10161016
def int_asin : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
10171017
def int_acos : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
10181018
def int_atan : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
1019+
def int_atan2 : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>]>;
10191020
def int_sin : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
10201021
def int_cos : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;
10211022
def int_tan : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>]>;

‎llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp

Lines changed: 46 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -36,6 +36,7 @@ using namespace llvm;
3636
static bool isIntrinsicExpansion(Function &F) {
3737
switch (F.getIntrinsicID()) {
3838
case Intrinsic::abs:
39+
case Intrinsic::atan2:
3940
case Intrinsic::exp:
4041
case Intrinsic::log:
4142
case Intrinsic::log10:
@@ -307,6 +308,48 @@ static Value *expandNormalizeIntrinsic(CallInst *Orig) {
307308
return Builder.CreateFMul(X, MultiplicandVec);
308309
}
309310

311+
static Value *expandAtan2Intrinsic(CallInst *Orig) {
312+
Value *Y = Orig->getOperand(0);
313+
Value *X = Orig->getOperand(1);
314+
Type *Ty = X->getType();
315+
IRBuilder<> Builder(Orig);
316+
317+
Value *Tan = Builder.CreateFDiv(Y, X);
318+
319+
Value *Atan =
320+
Builder.CreateIntrinsic(Ty, Intrinsic::atan, {Tan}, nullptr, "Elt.Atan");
321+
322+
Constant *Pi = ConstantFP::get(Ty, llvm::numbers::pi);
323+
Constant *HalfPi = ConstantFP::get(Ty, llvm::numbers::pi / 2);
324+
Constant *NegHalfPi = ConstantFP::get(Ty, -llvm::numbers::pi / 2);
325+
Constant *Zero = ConstantFP::get(Ty, 0);
326+
327+
Value *AtanAddPi = Builder.CreateFAdd(Atan, Pi);
328+
Value *AtanSubPi = Builder.CreateFSub(Atan, Pi);
329+
330+
Value *Result = Atan;
331+
332+
Value *XLt0 = Builder.CreateFCmpOLT(X, Zero);
333+
Value *XEq0 = Builder.CreateFCmpOEQ(X, Zero);
334+
335+
Value *YGe0 = Builder.CreateFCmpOGE(Y, Zero);
336+
Value *YLt0 = Builder.CreateFCmpOLT(Y, Zero);
337+
338+
Value *XLt0AndYGe0 = Builder.CreateAnd(XLt0, YGe0);
339+
Result = Builder.CreateSelect(XLt0AndYGe0, AtanAddPi, Result);
340+
341+
Value *XLt0AndYLt0 = Builder.CreateAnd(XLt0, YLt0);
342+
Result = Builder.CreateSelect(XLt0AndYLt0, AtanSubPi, Result);
343+
344+
Value *XEq0AndYLt0 = Builder.CreateAnd(XEq0, YLt0);
345+
Result = Builder.CreateSelect(XEq0AndYLt0, NegHalfPi, Result);
346+
347+
Value *XEq0AndYGe0 = Builder.CreateAnd(XEq0, YGe0);
348+
Result = Builder.CreateSelect(XEq0AndYGe0, HalfPi, Result);
349+
350+
return Result;
351+
}
352+
310353
static Value *expandPowIntrinsic(CallInst *Orig) {
311354

312355
Value *X = Orig->getOperand(0);
@@ -418,6 +461,9 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
418461
case Intrinsic::abs:
419462
Result = expandAbs(Orig);
420463
break;
464+
case Intrinsic::atan2:
465+
Result = expandAtan2Intrinsic(Orig);
466+
break;
421467
case Intrinsic::exp:
422468
Result = expandExpIntrinsic(Orig);
423469
break;

‎llvm/test/CodeGen/DirectX/atan2.ll

Lines changed: 52 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,52 @@
1+
; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
2+
3+
; Make sure correct dxil expansions for atan2 are generated for float and half.
4+
5+
define noundef float @atan2_float(float noundef %y, float noundef %x) {
6+
entry:
7+
; CHECK: [[DIV:%.+]] = fdiv float %y, %x
8+
; CHECK: [[TAN:%.+]] = call float @dx.op.unary.f32(i32 17, float [[DIV]])
9+
; CHECK-DAG: [[ADD_PI:%.+]] = fadd float [[TAN]], 0x400921FB60000000
10+
; CHECK-DAG: [[SUB_PI:%.+]] = fsub float [[TAN]], 0x400921FB60000000
11+
; CHECK-DAG: [[X_LT_0:%.+]] = fcmp olt float %x, 0.000000e+00
12+
; CHECK-DAG: [[X_EQ_0:%.+]] = fcmp oeq float %x, 0.000000e+00
13+
; CHECK-DAG: [[Y_GE_0:%.+]] = fcmp oge float %y, 0.000000e+00
14+
; CHECK-DAG: [[Y_LT_0:%.+]] = fcmp olt float %y, 0.000000e+00
15+
; CHECK: [[XLT0_AND_YGE0:%.+]] = and i1 [[X_LT_0]], [[Y_GE_0]]
16+
; CHECK: [[SELECT_ADD_PI:%.+]] = select i1 [[XLT0_AND_YGE0]], float [[ADD_PI]], float [[TAN]]
17+
; CHECK: [[XLT0_AND_YLT0:%.+]] = and i1 [[X_LT_0]], [[Y_LT_0]]
18+
; CHECK: [[SELECT_SUB_PI:%.+]] = select i1 [[XLT0_AND_YLT0]], float [[SUB_PI]], float [[SELECT_ADD_PI]]
19+
; CHECK: [[XEQ0_AND_YLT0:%.+]] = and i1 [[X_EQ_0]], [[Y_LT_0]]
20+
; CHECK: [[SELECT_NEGHPI:%.+]] = select i1 [[XEQ0_AND_YLT0]], float 0xBFF921FB60000000, float [[SELECT_SUB_PI]]
21+
; CHECK: [[XEQ0_AND_YGE0:%.+]] = and i1 [[X_EQ_0]], [[Y_GE_0]]
22+
; CHECK: [[SELECT_HPI:%.+]] = select i1 [[XEQ0_AND_YGE0]], float 0x3FF921FB60000000, float [[SELECT_NEGHPI]]
23+
; CHECK: ret float [[SELECT_HPI]]
24+
%elt.atan2 = call float @llvm.atan2.f32(float %y, float %x)
25+
ret float %elt.atan2
26+
}
27+
28+
define noundef half @atan2_half(half noundef %y, half noundef %x) {
29+
entry:
30+
; CHECK: [[DIV:%.+]] = fdiv half %y, %x
31+
; CHECK: [[TAN:%.+]] = call half @dx.op.unary.f16(i32 17, half [[DIV]])
32+
; CHECK-DAG: [[ADD_PI:%.+]] = fadd half [[TAN]], 0xH4248
33+
; CHECK-DAG: [[SUB_PI:%.+]] = fsub half [[TAN]], 0xH4248
34+
; CHECK-DAG: [[X_LT_0:%.+]] = fcmp olt half %x, 0xH0000
35+
; CHECK-DAG: [[X_EQ_0:%.+]] = fcmp oeq half %x, 0xH0000
36+
; CHECK-DAG: [[Y_GE_0:%.+]] = fcmp oge half %y, 0xH0000
37+
; CHECK-DAG: [[Y_LT_0:%.+]] = fcmp olt half %y, 0xH0000
38+
; CHECK: [[XLT0_AND_YGE0:%.+]] = and i1 [[X_LT_0]], [[Y_GE_0]]
39+
; CHECK: [[SELECT_ADD_PI:%.+]] = select i1 [[XLT0_AND_YGE0]], half [[ADD_PI]], half [[TAN]]
40+
; CHECK: [[XLT0_AND_YLT0:%.+]] = and i1 [[X_LT_0]], [[Y_LT_0]]
41+
; CHECK: [[SELECT_SUB_PI:%.+]] = select i1 [[XLT0_AND_YLT0]], half [[SUB_PI]], half [[SELECT_ADD_PI]]
42+
; CHECK: [[XEQ0_AND_YLT0:%.+]] = and i1 [[X_EQ_0]], [[Y_LT_0]]
43+
; CHECK: [[SELECT_NEGHPI:%.+]] = select i1 [[XEQ0_AND_YLT0]], half 0xHBE48, half [[SELECT_SUB_PI]]
44+
; CHECK: [[XEQ0_AND_YGE0:%.+]] = and i1 [[X_EQ_0]], [[Y_GE_0]]
45+
; CHECK: [[SELECT_HPI:%.+]] = select i1 [[XEQ0_AND_YGE0]], half 0xH3E48, half [[SELECT_NEGHPI]]
46+
; CHECK: ret half [[SELECT_HPI]]
47+
%elt.atan2 = call half @llvm.atan2.f16(half %y, half %x)
48+
ret half %elt.atan2
49+
}
50+
51+
declare half @llvm.atan2.f16(half, half)
52+
declare float @llvm.atan2.f32(float, float)

‎llvm/test/CodeGen/DirectX/atan2_error.ll

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,11 @@
1+
; RUN: not opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s 2>&1 | FileCheck %s
2+
3+
; DXIL operation atan does not support double overload type
4+
; CHECK: in function atan2_double
5+
; CHECK-SAME: Cannot create ATan operation: Invalid overload type
6+
7+
define noundef double @atan2_double(double noundef %a, double noundef %b) #0 {
8+
entry:
9+
%1 = call double @llvm.atan2.f64(double %a, double %b)
10+
ret double %1
11+
}

0 commit comments

Comments
 (0)