diff --git a/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathCommon_simd.inl b/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathCommon_simd.inl index 770a8fa104..d06372256d 100644 --- a/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathCommon_simd.inl +++ b/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathCommon_simd.inl @@ -292,8 +292,13 @@ namespace AZ const typename VecType::FloatType cmp2 = VecType::AndNot(cmp0, cmp1); // -1/x + // this step is calculated for all values of x, but only used if x > Sqrt(2) + 1 + // in order to avoid a division by zero, detect if xabs is zero here and replace it with an arbitrary value + // if xabs does equal zero, the value here doesn't matter because the result will be thrown away + typename VecType::FloatType xabsSafe = + VecType::Add(xabs, VecType::And(VecType::CmpEq(xabs, VecType::ZeroFloat()), FastLoadConstant(Simd::g_vec1111))); const typename VecType::FloatType y0 = VecType::And(cmp0, FastLoadConstant(Simd::g_HalfPi)); - typename VecType::FloatType x0 = VecType::Div(FastLoadConstant(Simd::g_vec1111), xabs); + typename VecType::FloatType x0 = VecType::Div(FastLoadConstant(Simd::g_vec1111), xabsSafe); x0 = VecType::Xor(x0, VecType::CastToFloat(FastLoadConstant(Simd::g_negateMask))); const typename VecType::FloatType y1 = VecType::And(cmp2, FastLoadConstant(Simd::g_QuarterPi)); @@ -368,8 +373,12 @@ namespace AZ typename VecType::FloatType offset = VecType::And(x_lt_0, offset1); + // the result of this part of the computation is thrown away if x equals 0, + // but if x does equal 0, it will cause a division by zero + // so replace zero by an arbitrary value here in that case + typename VecType::FloatType xSafe = VecType::Add(x, VecType::And(x_eq_0, FastLoadConstant(Simd::g_vec1111))); const typename VecType::FloatType atan_mask = VecType::Not(VecType::Or(x_eq_0, y_eq_0)); - const typename VecType::FloatType atan_arg = VecType::Div(y, x); + const typename VecType::FloatType atan_arg = VecType::Div(y, xSafe); typename VecType::FloatType atan_result = VecType::Atan(atan_arg); atan_result = VecType::Add(atan_result, offset); atan_result = VecType::AndNot(pio2_mask, atan_result); diff --git a/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec2_sse.inl b/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec2_sse.inl index 8f35af258c..63e2dca8fd 100644 --- a/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec2_sse.inl +++ b/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec2_sse.inl @@ -471,6 +471,7 @@ namespace AZ AZ_MATH_INLINE Vec2::FloatType Vec2::Reciprocal(FloatArgType value) { + value = Sse::ReplaceFourth(Sse::ReplaceThird(value, 1.0f), 1.0f); return Sse::Reciprocal(value); } @@ -513,6 +514,7 @@ namespace AZ AZ_MATH_INLINE Vec2::FloatType Vec2::SqrtInv(FloatArgType value) { + value = Sse::ReplaceFourth(Sse::ReplaceThird(value, 1.0f), 1.0f); return Sse::SqrtInv(value); } diff --git a/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec3_sse.inl b/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec3_sse.inl index 78a0d5db66..75ee2ab7c5 100644 --- a/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec3_sse.inl +++ b/Code/Framework/AzCore/AzCore/Math/Internal/SimdMathVec3_sse.inl @@ -507,6 +507,7 @@ namespace AZ AZ_MATH_INLINE Vec3::FloatType Vec3::Reciprocal(FloatArgType value) { + value = Sse::ReplaceFourth(value, 1.0f); return Sse::Reciprocal(value); } @@ -549,6 +550,7 @@ namespace AZ AZ_MATH_INLINE Vec3::FloatType Vec3::SqrtInv(FloatArgType value) { + value = Sse::ReplaceFourth(value, 1.0f); return Sse::SqrtInv(value); }