ARM64: FP16 greater/less/greaterEquals/lessEquals intrinsics for ARMv8
This CL implements intrinsics for greater, greaterEquals, less,
lessEquals methods with ARMv8.2 FP16 instructions. This requires the
ARMv8.2 AArch64 asimd half precision extension.
The time required in milliseconds to execute the below code for the four
intrinsics on Pixel3 is (The code below is for FP16.less but is similar
for the rest of the intrinsics):
- Java implementation libcore.util.FP16.less():
- big cluster only: 19876
- little cluster only: 47525
- arm64 Intrinisic implementationi for less:
- big cluster only: 14526 (~27% faster)
- little cluster only: 45815 (~4% faster)
- Java implementation libcore.util.FP16.lessEquals():
- big cluster only: 19856
- little cluster only: 47419
- arm64 Intrinisic implementation for lessEquals:
- big cluster only: 14469 (~27% faster)
- little cluster only: 45762 (~4% faster)
- Java implementation libcore.util.FP16.greater():
- big cluster only: 19854
- little cluster only: 47623
- arm64 Intrinisic implementation for greater:
- big cluster only: 14519 (~27% faster)
- little cluster only: 45722 (~4% faster)
- Java implementation libcore.util.FP16.greaterEquals():
- big cluster only: 19865
- little cluster only: 47216
- arm64 Intrinisic implementation for greaterEquals:
- big cluster only: 14485 (~27% faster)
- little cluster only: 45729 (~4% faster)
public static boolean benchmarkComparison(){
boolean ret = false;
long before = 0;
long after = 0;
before = System.currentTimeMillis();
for(long i = 0; i < 1e9; i++){
// FP16.toHalf(12.3) = 0x4a26, FP16.toHalf(12.4) = 0x4a33
// FP16.toHalf(-12.3) = 0xca26, FP16.toHalf(-12.4) = 0xca33
ret |= FP16.less((short) 0x4a26,(short) 0x4a33);
ret |= FP16.less((short) 0x4a33,(short) 0x4a26);
ret |= FP16.less((short) 0xca26,(short) 0xca33);
ret |= FP16.less((short) 0xca33,(short) 0xca26);
}
after = System.currentTimeMillis();
System.out.println("Time of FP16.less (ms): " + (after - before));
System.out.println(ret);
return ret;
}
Test: 580-fp16
Test: art/test/testrunner/run_build_test_target.py -j80 art-test-javac
Change-Id: Id1a2c3e7328c82c798fcaf1fa74f5908a822cd0b
diff --git a/compiler/optimizing/common_arm64.h b/compiler/optimizing/common_arm64.h
index 9c80f32..36e68e4 100644
--- a/compiler/optimizing/common_arm64.h
+++ b/compiler/optimizing/common_arm64.h
@@ -107,6 +107,11 @@
return vixl::aarch64::VRegister::GetSRegFromCode(location.reg());
}
+inline vixl::aarch64::FPRegister HRegisterFrom(Location location) {
+ DCHECK(location.IsFpuRegister()) << location;
+ return vixl::aarch64::FPRegister::GetHRegFromCode(location.reg());
+}
+
inline vixl::aarch64::VRegister FPRegisterFrom(Location location, DataType::Type type) {
DCHECK(DataType::IsFloatingPointType(type)) << type;
return type == DataType::Type::kFloat64 ? DRegisterFrom(location) : SRegisterFrom(location);
diff --git a/compiler/optimizing/intrinsics_arm64.cc b/compiler/optimizing/intrinsics_arm64.cc
index 542bd17..bdeb6a4 100644
--- a/compiler/optimizing/intrinsics_arm64.cc
+++ b/compiler/optimizing/intrinsics_arm64.cc
@@ -54,6 +54,7 @@
using helpers::SRegisterFrom;
using helpers::WRegisterFrom;
using helpers::XRegisterFrom;
+using helpers::HRegisterFrom;
using helpers::InputRegisterAt;
using helpers::OutputRegister;
@@ -299,6 +300,14 @@
locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
}
+static void CreateIntIntToIntLocations(ArenaAllocator* allocator, HInvoke* invoke) {
+ LocationSummary* locations =
+ new (allocator) LocationSummary(invoke, LocationSummary::kNoCall, kIntrinsified);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+}
+
static void GenReverseBytes(LocationSummary* locations,
DataType::Type type,
MacroAssembler* masm) {
@@ -3304,6 +3313,92 @@
GenerateFP16Round(invoke, codegen_, masm, roundOp);
}
+template<typename OP>
+void GenerateFP16Compare(HInvoke* invoke,
+ CodeGeneratorARM64* codegen,
+ MacroAssembler* masm,
+ const OP compareOp) {
+ DCHECK(codegen->GetInstructionSetFeatures().HasFP16());
+ LocationSummary* locations = invoke->GetLocations();
+ Register out = WRegisterFrom(locations->Out());
+ VRegister half0 = HRegisterFrom(locations->GetTemp(0));
+ VRegister half1 = HRegisterFrom(locations->GetTemp(1));
+ __ Fmov(half0, WRegisterFrom(locations->InAt(0)));
+ __ Fmov(half1, WRegisterFrom(locations->InAt(1)));
+ compareOp(out, half0, half1);
+}
+
+static inline void GenerateFP16Compare(HInvoke* invoke,
+ CodeGeneratorARM64* codegen,
+ MacroAssembler* masm,
+ vixl::aarch64::Condition cond) {
+ auto compareOp = [masm, cond](const Register out, const VRegister& in0, const VRegister& in1) {
+ __ Fcmp(in0, in1);
+ __ Cset(out, cond);
+ };
+ GenerateFP16Compare(invoke, codegen, masm, compareOp);
+}
+
+void IntrinsicLocationsBuilderARM64::VisitFP16Greater(HInvoke* invoke) {
+ if (!codegen_->GetInstructionSetFeatures().HasFP16()) {
+ return;
+ }
+
+ CreateIntIntToIntLocations(allocator_, invoke);
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+}
+
+void IntrinsicCodeGeneratorARM64::VisitFP16Greater(HInvoke* invoke) {
+ MacroAssembler* masm = GetVIXLAssembler();
+ GenerateFP16Compare(invoke, codegen_, masm, gt);
+}
+
+void IntrinsicLocationsBuilderARM64::VisitFP16GreaterEquals(HInvoke* invoke) {
+ if (!codegen_->GetInstructionSetFeatures().HasFP16()) {
+ return;
+ }
+
+ CreateIntIntToIntLocations(allocator_, invoke);
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+}
+
+void IntrinsicCodeGeneratorARM64::VisitFP16GreaterEquals(HInvoke* invoke) {
+ MacroAssembler* masm = GetVIXLAssembler();
+ GenerateFP16Compare(invoke, codegen_, masm, ge);
+}
+
+void IntrinsicLocationsBuilderARM64::VisitFP16Less(HInvoke* invoke) {
+ if (!codegen_->GetInstructionSetFeatures().HasFP16()) {
+ return;
+ }
+
+ CreateIntIntToIntLocations(allocator_, invoke);
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+}
+
+void IntrinsicCodeGeneratorARM64::VisitFP16Less(HInvoke* invoke) {
+ MacroAssembler* masm = GetVIXLAssembler();
+ GenerateFP16Compare(invoke, codegen_, masm, mi);
+}
+
+void IntrinsicLocationsBuilderARM64::VisitFP16LessEquals(HInvoke* invoke) {
+ if (!codegen_->GetInstructionSetFeatures().HasFP16()) {
+ return;
+ }
+
+ CreateIntIntToIntLocations(allocator_, invoke);
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+ invoke->GetLocations()->AddTemp(Location::RequiresFpuRegister());
+}
+
+void IntrinsicCodeGeneratorARM64::VisitFP16LessEquals(HInvoke* invoke) {
+ MacroAssembler* masm = GetVIXLAssembler();
+ GenerateFP16Compare(invoke, codegen_, masm, ls);
+}
+
UNIMPLEMENTED_INTRINSIC(ARM64, ReferenceGetReferent)
UNIMPLEMENTED_INTRINSIC(ARM64, StringStringIndexOf);
diff --git a/compiler/optimizing/intrinsics_arm_vixl.cc b/compiler/optimizing/intrinsics_arm_vixl.cc
index 1dfebdd..89e5203 100644
--- a/compiler/optimizing/intrinsics_arm_vixl.cc
+++ b/compiler/optimizing/intrinsics_arm_vixl.cc
@@ -3075,6 +3075,10 @@
UNIMPLEMENTED_INTRINSIC(ARMVIXL, FP16Floor)
UNIMPLEMENTED_INTRINSIC(ARMVIXL, FP16Ceil)
UNIMPLEMENTED_INTRINSIC(ARMVIXL, FP16Rint)
+UNIMPLEMENTED_INTRINSIC(ARMVIXL, FP16Greater)
+UNIMPLEMENTED_INTRINSIC(ARMVIXL, FP16GreaterEquals)
+UNIMPLEMENTED_INTRINSIC(ARMVIXL, FP16Less)
+UNIMPLEMENTED_INTRINSIC(ARMVIXL, FP16LessEquals)
UNIMPLEMENTED_INTRINSIC(ARMVIXL, StringStringIndexOf);
UNIMPLEMENTED_INTRINSIC(ARMVIXL, StringStringIndexOfAfter);
diff --git a/compiler/optimizing/intrinsics_mips.cc b/compiler/optimizing/intrinsics_mips.cc
index ea9c591..537255f 100644
--- a/compiler/optimizing/intrinsics_mips.cc
+++ b/compiler/optimizing/intrinsics_mips.cc
@@ -2712,6 +2712,10 @@
UNIMPLEMENTED_INTRINSIC(MIPS, FP16Floor)
UNIMPLEMENTED_INTRINSIC(MIPS, FP16Ceil)
UNIMPLEMENTED_INTRINSIC(MIPS, FP16Rint)
+UNIMPLEMENTED_INTRINSIC(MIPS, FP16Greater)
+UNIMPLEMENTED_INTRINSIC(MIPS, FP16GreaterEquals)
+UNIMPLEMENTED_INTRINSIC(MIPS, FP16Less)
+UNIMPLEMENTED_INTRINSIC(MIPS, FP16LessEquals)
UNIMPLEMENTED_INTRINSIC(MIPS, StringStringIndexOf);
UNIMPLEMENTED_INTRINSIC(MIPS, StringStringIndexOfAfter);
diff --git a/compiler/optimizing/intrinsics_mips64.cc b/compiler/optimizing/intrinsics_mips64.cc
index fd93902..5920394 100644
--- a/compiler/optimizing/intrinsics_mips64.cc
+++ b/compiler/optimizing/intrinsics_mips64.cc
@@ -2362,6 +2362,10 @@
UNIMPLEMENTED_INTRINSIC(MIPS64, FP16Floor)
UNIMPLEMENTED_INTRINSIC(MIPS64, FP16Ceil)
UNIMPLEMENTED_INTRINSIC(MIPS64, FP16Rint)
+UNIMPLEMENTED_INTRINSIC(MIPS64, FP16Greater)
+UNIMPLEMENTED_INTRINSIC(MIPS64, FP16GreaterEquals)
+UNIMPLEMENTED_INTRINSIC(MIPS64, FP16Less)
+UNIMPLEMENTED_INTRINSIC(MIPS64, FP16LessEquals)
UNIMPLEMENTED_INTRINSIC(MIPS64, StringStringIndexOf);
UNIMPLEMENTED_INTRINSIC(MIPS64, StringStringIndexOfAfter);
diff --git a/compiler/optimizing/intrinsics_x86.cc b/compiler/optimizing/intrinsics_x86.cc
index 9d3cecb..6d7462e 100644
--- a/compiler/optimizing/intrinsics_x86.cc
+++ b/compiler/optimizing/intrinsics_x86.cc
@@ -3086,6 +3086,10 @@
UNIMPLEMENTED_INTRINSIC(X86, FP16Floor)
UNIMPLEMENTED_INTRINSIC(X86, FP16Ceil)
UNIMPLEMENTED_INTRINSIC(X86, FP16Rint)
+UNIMPLEMENTED_INTRINSIC(X86, FP16Greater)
+UNIMPLEMENTED_INTRINSIC(X86, FP16GreaterEquals)
+UNIMPLEMENTED_INTRINSIC(X86, FP16Less)
+UNIMPLEMENTED_INTRINSIC(X86, FP16LessEquals)
UNIMPLEMENTED_INTRINSIC(X86, StringStringIndexOf);
UNIMPLEMENTED_INTRINSIC(X86, StringStringIndexOfAfter);
diff --git a/compiler/optimizing/intrinsics_x86_64.cc b/compiler/optimizing/intrinsics_x86_64.cc
index 1111a59..0f6b006 100644
--- a/compiler/optimizing/intrinsics_x86_64.cc
+++ b/compiler/optimizing/intrinsics_x86_64.cc
@@ -2753,6 +2753,10 @@
UNIMPLEMENTED_INTRINSIC(X86_64, FP16Floor)
UNIMPLEMENTED_INTRINSIC(X86_64, FP16Ceil)
UNIMPLEMENTED_INTRINSIC(X86_64, FP16Rint)
+UNIMPLEMENTED_INTRINSIC(X86_64, FP16Greater)
+UNIMPLEMENTED_INTRINSIC(X86_64, FP16GreaterEquals)
+UNIMPLEMENTED_INTRINSIC(X86_64, FP16Less)
+UNIMPLEMENTED_INTRINSIC(X86_64, FP16LessEquals)
UNIMPLEMENTED_INTRINSIC(X86_64, StringStringIndexOf);
UNIMPLEMENTED_INTRINSIC(X86_64, StringStringIndexOfAfter);