Correct reciprocal approximation for power-of-two values.
Intel's reciprocal approximation instruction is not exact for power-of-two
values. It provides 12 bits of mantissa precision and keeps a balance between
positive and negative errors, but the reciprocal of 2^x is not 2^-x. This
affects conformance tests which expect varyings not to be affected by the
perspective division. Correct for this by multiplying by the inverse.
Bug 27165393
Change-Id: Ie52ec511a14a4f447adc47ce9c875bbad03cd274
Reviewed-on: https://swiftshader-review.googlesource.com/4903
Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
diff --git a/src/Reactor/Nucleus.cpp b/src/Reactor/Nucleus.cpp
index 2325c7f..ce48b52 100644
--- a/src/Reactor/Nucleus.cpp
+++ b/src/Reactor/Nucleus.cpp
@@ -33,6 +33,7 @@
#include "Thread.hpp"
#include "Memory.hpp"
+#include <xmmintrin.h>
#include <fstream>
#if defined(__x86_64__) && defined(_WIN32)
@@ -4635,9 +4636,9 @@
Constant *shuffle[2];
shuffle[0] = Nucleus::createConstantInt(0);
shuffle[1] = Nucleus::createConstantInt(1);
-
+
Value *packed = Nucleus::createShuffleVector(Nucleus::createBitCast(lo.value, VectorType::get(Int::getType(), 1)), Nucleus::createBitCast(hi.value, VectorType::get(Int::getType(), 1)), Nucleus::createConstantVector(shuffle, 2));
-
+
storeValue(Nucleus::createBitCast(packed, Int2::getType()));
}
}
@@ -5199,7 +5200,7 @@
Value *element = Nucleus::createBitCast(cast.value, Long::getType());
long2 = Nucleus::createInsertElement(long2, element, 0);
RValue<Int4> vector = RValue<Int4>(Nucleus::createBitCast(long2, Int4::getType()));
-
+
if(CPUID::supportsSSE4_1())
{
storeValue(x86::pmovsxwd(vector).value);
@@ -6069,9 +6070,18 @@
return IfThenElse(x < y, x, y);
}
- RValue<Float> Rcp_pp(RValue<Float> x)
+ RValue<Float> Rcp_pp(RValue<Float> x, bool exactAtPow2)
{
- return x86::rcpss(x);
+ if(exactAtPow2)
+ {
+ // rcpss uses a piecewise-linear approximation which minimizes the relative error
+ // but is not exact at power-of-two values. Rectify by multiplying by the inverse.
+ return x86::rcpss(x) * Float(1.0f / _mm_cvtss_f32(_mm_rcp_ss(_mm_set_ps1(1.0f))));
+ }
+ else
+ {
+ return x86::rcpss(x);
+ }
}
RValue<Float> RcpSqrt_pp(RValue<Float> x)
@@ -6580,9 +6590,18 @@
return x86::minps(x, y);
}
- RValue<Float4> Rcp_pp(RValue<Float4> x)
+ RValue<Float4> Rcp_pp(RValue<Float4> x, bool exactAtPow2)
{
- return x86::rcpps(x);
+ if(exactAtPow2)
+ {
+ // rcpps uses a piecewise-linear approximation which minimizes the relative error
+ // but is not exact at power-of-two values. Rectify by multiplying by the inverse.
+ return x86::rcpps(x) * Float4(1.0f / _mm_cvtss_f32(_mm_rcp_ss(_mm_set_ps1(1.0f))));
+ }
+ else
+ {
+ return x86::rcpps(x);
+ }
}
RValue<Float4> RcpSqrt_pp(RValue<Float4> x)
diff --git a/src/Reactor/Nucleus.hpp b/src/Reactor/Nucleus.hpp
index e305b1a..6b6cdb4 100644
--- a/src/Reactor/Nucleus.hpp
+++ b/src/Reactor/Nucleus.hpp
@@ -1899,7 +1899,7 @@
RValue<Float> Abs(RValue<Float> x);
RValue<Float> Max(RValue<Float> x, RValue<Float> y);
RValue<Float> Min(RValue<Float> x, RValue<Float> y);
- RValue<Float> Rcp_pp(RValue<Float> val);
+ RValue<Float> Rcp_pp(RValue<Float> val, bool exactAtPow2 = false);
RValue<Float> RcpSqrt_pp(RValue<Float> val);
RValue<Float> Sqrt(RValue<Float> x);
RValue<Float> Round(RValue<Float> val);
@@ -2377,7 +2377,7 @@
RValue<Float4> Abs(RValue<Float4> x);
RValue<Float4> Max(RValue<Float4> x, RValue<Float4> y);
RValue<Float4> Min(RValue<Float4> x, RValue<Float4> y);
- RValue<Float4> Rcp_pp(RValue<Float4> val);
+ RValue<Float4> Rcp_pp(RValue<Float4> val, bool exactAtPow2 = false);
RValue<Float4> RcpSqrt_pp(RValue<Float4> val);
RValue<Float4> Sqrt(RValue<Float4> x);
RValue<Float4> Insert(const Float4 &val, RValue<Float> element, int i);
diff --git a/src/Shader/PixelRoutine.cpp b/src/Shader/PixelRoutine.cpp
index c43c87c..2b5f087 100644
--- a/src/Shader/PixelRoutine.cpp
+++ b/src/Shader/PixelRoutine.cpp
@@ -140,7 +140,7 @@
if(interpolateW())
{
w = interpolate(xxxx, Dw, rhw, primitive + OFFSET(Primitive,w), false, false);
- rhw = reciprocal(w);
+ rhw = reciprocal(w, false, false, true);
if(state.centroid)
{
@@ -518,7 +518,7 @@
zMask = SignMask(zTest) & cMask;
break;
}
-
+
if(state.stencilActive)
{
zMask &= sMask;
@@ -687,12 +687,12 @@
Int pitch;
if(!state.quadLayoutDepthBuffer)
- {
+ {
buffer = zBuffer + 4 * x;
pitch = *Pointer<Int>(data + OFFSET(DrawData,depthPitchB));
}
else
- {
+ {
buffer = zBuffer + 8 * x;
}
@@ -761,7 +761,7 @@
}
Byte8 bufferValue = As<Byte8>(Long1(*Pointer<UInt>(buffer)));
-
+
Byte8 newValue;
stencilOperation(newValue, bufferValue, state.stencilPassOperation, state.stencilZFailOperation, state.stencilFailOperation, false, zMask, sMask);
@@ -945,7 +945,7 @@
ASSERT(false);
}
}
-
+
void PixelRoutine::blendFactorAlpha(const Vector4s &blendFactor, const Vector4s ¤t, const Vector4s &pixel, BlendFactor blendFactorAlphaActive)
{
switch(blendFactorAlphaActive)
@@ -1170,7 +1170,7 @@
current.y = MulHigh(As<UShort4>(current.y), As<UShort4>(sourceFactor.y));
current.z = MulHigh(As<UShort4>(current.z), As<UShort4>(sourceFactor.z));
}
-
+
if(state.destBlendFactor != BLEND_ONE && state.destBlendFactor != BLEND_ZERO)
{
pixel.x = MulHigh(As<UShort4>(pixel.x), As<UShort4>(destFactor.x));
@@ -1229,7 +1229,7 @@
{
current.w = MulHigh(As<UShort4>(current.w), As<UShort4>(sourceFactor.w));
}
-
+
if(state.destBlendFactorAlpha != BLEND_ONE && state.destBlendFactorAlpha != BLEND_ZERO)
{
pixel.w = MulHigh(As<UShort4>(pixel.w), As<UShort4>(destFactor.w));
@@ -1828,7 +1828,7 @@
}
}
- void PixelRoutine::blendFactor(const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorActive)
+ void PixelRoutine::blendFactor(const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorActive)
{
switch(blendFactorActive)
{
@@ -1899,7 +1899,7 @@
}
}
- void PixelRoutine::blendFactorAlpha(const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorAlphaActive)
+ void PixelRoutine::blendFactorAlpha(const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorAlphaActive)
{
switch(blendFactorAlphaActive)
{
@@ -2041,7 +2041,7 @@
oC.y *= sourceFactor.y;
oC.z *= sourceFactor.z;
}
-
+
if(state.destBlendFactor != BLEND_ONE && state.destBlendFactor != BLEND_ZERO)
{
pixel.x *= destFactor.x;
@@ -2100,7 +2100,7 @@
{
oC.w *= sourceFactor.w;
}
-
+
if(state.destBlendFactorAlpha != BLEND_ONE && state.destBlendFactorAlpha != BLEND_ZERO)
{
pixel.w *= destFactor.w;
@@ -2118,10 +2118,10 @@
pixel.w -= oC.w;
oC.w = pixel.w;
break;
- case BLENDOP_MIN:
+ case BLENDOP_MIN:
oC.w = Min(oC.w, pixel.w);
break;
- case BLENDOP_MAX:
+ case BLENDOP_MAX:
oC.w = Max(oC.w, pixel.w);
break;
case BLENDOP_SOURCE:
@@ -2272,7 +2272,7 @@
masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X[rgbaWriteMask][0])));
oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(masked));
}
-
+
oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants,maskX0X) + xMask * 16, 16));
value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskX0X) + xMask * 16, 16));
oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(value));
@@ -2283,7 +2283,7 @@
value = *Pointer<Float4>(buffer + 16, 16);
if(rgbaWriteMask != 0x0000000F)
- {
+ {
Float4 masked = value;
oC.y = As<Float4>(As<Int4>(oC.y) & *Pointer<Int4>(constants + OFFSET(Constants,maskD4X[rgbaWriteMask][0])));
masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X[rgbaWriteMask][0])));
diff --git a/src/Shader/ShaderCore.cpp b/src/Shader/ShaderCore.cpp
index ad9f426..f177a68 100644
--- a/src/Shader/ShaderCore.cpp
+++ b/src/Shader/ShaderCore.cpp
@@ -271,7 +271,7 @@
return exponential2(log, pp);
}
- Float4 reciprocal(RValue<Float4> x, bool pp, bool finite)
+ Float4 reciprocal(RValue<Float4> x, bool pp, bool finite, bool exactAtPow2)
{
Float4 rcp;
@@ -281,7 +281,7 @@
}
else
{
- rcp = Rcp_pp(x);
+ rcp = Rcp_pp(x, exactAtPow2);
if(!pp)
{
diff --git a/src/Shader/ShaderCore.hpp b/src/Shader/ShaderCore.hpp
index fca88c1..3f4b22c 100644
--- a/src/Shader/ShaderCore.hpp
+++ b/src/Shader/ShaderCore.hpp
@@ -87,7 +87,7 @@
Float4 exponential(RValue<Float4> x, bool pp = false);
Float4 logarithm(RValue<Float4> x, bool abs, bool pp = false);
Float4 power(RValue<Float4> x, RValue<Float4> y, bool pp = false);
- Float4 reciprocal(RValue<Float4> x, bool pp = false, bool finite = false);
+ Float4 reciprocal(RValue<Float4> x, bool pp = false, bool finite = false, bool exactAtPow2 = false);
Float4 reciprocalSquareRoot(RValue<Float4> x, bool abs, bool pp = false);
Float4 modulo(RValue<Float4> x, RValue<Float4> y);
Float4 sine_pi(RValue<Float4> x, bool pp = false); // limited to [-pi, pi] range
