Enable use of NEON intrinsics on ARM.
Bug b/37496338
Bug b/37495545
Bug b/37496078
Bug b/37496856
Bug b/37496321
Bug b/37496082
Change-Id: I173a296703fcc1f49810742d95ae6abe7c9bcf01
Reviewed-on: https://swiftshader-review.googlesource.com/12868
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>
diff --git a/src/Reactor/SubzeroReactor.cpp b/src/Reactor/SubzeroReactor.cpp
index 55a55c3..9429f62 100644
--- a/src/Reactor/SubzeroReactor.cpp
+++ b/src/Reactor/SubzeroReactor.cpp
@@ -119,7 +119,7 @@
const bool CPUID::ARM = CPUID::detectARM();
const bool CPUID::SSE4_1 = CPUID::detectSSE4_1();
- const bool emulateIntrinsics = CPUID::ARM;
+ const bool emulateIntrinsics = false;
const bool emulateMismatchedBitCast = CPUID::ARM;
}
@@ -2847,7 +2847,7 @@
RValue<Int> SignMask(RValue<Byte8> x)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
Byte8 xx = As<Byte8>(As<SByte8>(x) >> 7) & Byte8(0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80);
return Int(Extract(xx, 0)) | Int(Extract(xx, 1)) | Int(Extract(xx, 2)) | Int(Extract(xx, 3)) | Int(Extract(xx, 4)) | Int(Extract(xx, 5)) | Int(Extract(xx, 6)) | Int(Extract(xx, 7));
@@ -3123,7 +3123,7 @@
RValue<Int> SignMask(RValue<SByte8> x)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
SByte8 xx = (x >> 7) & SByte8(0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80);
return Int(Extract(xx, 0)) | Int(Extract(xx, 1)) | Int(Extract(xx, 2)) | Int(Extract(xx, 3)) | Int(Extract(xx, 4)) | Int(Extract(xx, 5)) | Int(Extract(xx, 6)) | Int(Extract(xx, 7));
@@ -3754,7 +3754,15 @@
{
if(CPUID::SSE4_1)
{
- Int4 int4(Min(cast, Float4(0xFFFF))); // packusdw takes care of 0x0000 saturation
+ // x86 produces 0x80000000 on 32-bit integer overflow/underflow.
+ // PackUnsigned takes care of 0x0000 saturation.
+ Int4 int4(Min(cast, Float4(0xFFFF)));
+ *this = As<UShort4>(PackUnsigned(int4, int4));
+ }
+ else if(CPUID::ARM)
+ {
+ // ARM saturates the 32-bit integer result on overflow/undeflow.
+ Int4 int4(cast);
*this = As<UShort4>(PackUnsigned(int4, int4));
}
else
@@ -4670,7 +4678,7 @@
RValue<Int> RoundInt(RValue<Float> cast)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
// Push the fractional part off the mantissa. Accurate up to +/-2^22.
return Int((cast + Float(0x00C00000)) - Float(0x00C00000));
@@ -5913,7 +5921,7 @@
RValue<Int4> RoundInt(RValue<Float4> cast)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
// Push the fractional part off the mantissa. Accurate up to +/-2^22.
return Int4((cast + Float4(0x00C00000)) - Float4(0x00C00000));
@@ -5963,7 +5971,17 @@
RValue<UShort8> PackUnsigned(RValue<Int4> x, RValue<Int4> y)
{
- if(CPUID::SSE4_1)
+ if(emulateIntrinsics || !(CPUID::SSE4_1 || CPUID::ARM))
+ {
+ RValue<Int4> sx = As<Int4>(x);
+ RValue<Int4> bx = (sx & ~(sx >> 31)) - Int4(0x8000);
+
+ RValue<Int4> sy = As<Int4>(y);
+ RValue<Int4> by = (sy & ~(sy >> 31)) - Int4(0x8000);
+
+ return As<UShort8>(PackSigned(bx, by) + Short8(0x8000u));
+ }
+ else
{
Ice::Variable *result = ::function->makeVariable(Ice::IceType_v8i16);
const Ice::Intrinsics::IntrinsicInfo intrinsic = {Ice::Intrinsics::VectorPackUnsigned, Ice::Intrinsics::SideEffects_F, Ice::Intrinsics::ReturnsTwice_F, Ice::Intrinsics::MemoryWrite_F};
@@ -5975,16 +5993,6 @@
return RValue<UShort8>(V(result));
}
- else
- {
- RValue<Int4> sx = As<Int4>(x);
- RValue<Int4> bx = (sx & ~(sx >> 31)) - Int4(0x8000);
-
- RValue<Int4> sy = As<Int4>(y);
- RValue<Int4> by = (sy & ~(sy >> 31)) - Int4(0x8000);
-
- return As<UShort8>(PackSigned(bx, by) + Short8(0x8000u));
- }
}
RValue<Int> Extract(RValue<Int4> x, int i)
@@ -5999,7 +6007,7 @@
RValue<Int> SignMask(RValue<Int4> x)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
Int4 xx = (x >> 31) & Int4(0x00000001, 0x00000002, 0x00000004, 0x00000008);
return Extract(xx, 0) | Extract(xx, 1) | Extract(xx, 2) | Extract(xx, 3);
@@ -6837,7 +6845,7 @@
RValue<Float4> Sqrt(RValue<Float4> x)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
Float4 result;
result.x = Sqrt(Float(Float4(x).x));
@@ -6911,7 +6919,7 @@
RValue<Int> SignMask(RValue<Float4> x)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
Int4 xx = (As<Int4>(x) >> 31) & Int4(0x00000001, 0x00000002, 0x00000004, 0x00000008);
return Extract(xx, 0) | Extract(xx, 1) | Extract(xx, 2) | Extract(xx, 3);
@@ -6961,7 +6969,7 @@
RValue<Float4> Round(RValue<Float4> x)
{
- if(emulateIntrinsics)
+ if(emulateIntrinsics || CPUID::ARM)
{
// Push the fractional part off the mantissa. Accurate up to +/-2^22.
return (x + Float4(0x00C00000)) - Float4(0x00C00000);
