replace FusedNegateMultiplyAdd with FusedMultiplyAdd when possible

interpretml · Oct 18, 2024 · 333f8de · 333f8de
1 parent dbadbaa
commit 333f8de
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 8 deletions.
diff --git a/shared/libebm/compute/math.hpp b/shared/libebm/compute/math.hpp
@@ -148,7 +148,7 @@ static INLINE_ALWAYS TFloat Exp32(const TFloat val) {
       rounded = Round(val * TFloat{1.44269504088896340736f});
       x = FusedMultiplyAdd(rounded, TFloat{-0.693359375f}, val);
    }
-   x = FusedNegateMultiplyAdd(rounded, TFloat{-2.12194440e-4f}, x);
+   x = FusedMultiplyAdd(rounded, TFloat{2.12194440e-4f}, x);
 
    const TFloat x2 = x * x;
    TFloat ret = Polynomial32(x,
@@ -233,7 +233,7 @@ static INLINE_ALWAYS TFloat Log32(const TFloat& val) noexcept {
    ret *= x2 * x;
 
    ret = FusedMultiplyAdd(exponentFloat, TFloat{-2.12194440E-4f}, ret);
-   ret += FusedNegateMultiplyAdd(x2, TFloat{0.5f}, x);
+   ret += FusedMultiplyAdd(x2, TFloat{-0.5f}, x);
 
    // exponentFloat must be a finite number, so use ret if we want an inf or NaN ret value
    if(bNaNPossible) {
@@ -298,7 +298,7 @@ static INLINE_ALWAYS TFloat Exp64(const TFloat val) {
       rounded = Round(val * TFloat{1.44269504088896340736});
       x = FusedMultiplyAdd(rounded, TFloat{-0.693145751953125}, val);
    }
-   x = FusedNegateMultiplyAdd(rounded, TFloat{1.42860682030941723212E-6}, x);
+   x = FusedMultiplyAdd(rounded, TFloat{-1.42860682030941723212E-6}, x);
 
    TFloat ret = Polynomial64(x,
          TFloat{1} / TFloat{2},
@@ -389,7 +389,7 @@ static INLINE_ALWAYS TFloat Log64(const TFloat& val) noexcept {
    TFloat ret = poly1 / poly2;
 
    ret = FusedMultiplyAdd(exponent, TFloat{-2.121944400546905827679E-4}, ret);
-   ret += FusedNegateMultiplyAdd(x2, TFloat{0.5}, x);
+   ret += FusedMultiplyAdd(x2, TFloat{-0.5}, x);
 
    // exponent must be a finite number, so use ret if we want an inf or NaN ret value
    if(bNaNPossible) {

diff --git a/shared/libebm/compute/objectives/TweedieDevianceRegressionObjective.hpp b/shared/libebm/compute/objectives/TweedieDevianceRegressionObjective.hpp
@@ -11,7 +11,7 @@ template<typename TFloat> struct TweedieDevianceRegressionObjective : Regression
 
    TFloat m_variancePowerParamSub1;
    TFloat m_variancePowerParamSub2;
-   TFloat m_inverseVariancePowerParamSub1;
+   TFloat m_negInverseVariancePowerParamSub1;
    TFloat m_inverseVariancePowerParamSub2;
 
    // The constructor parameters following config must match the RegisterObjective parameters in
@@ -40,7 +40,7 @@ template<typename TFloat> struct TweedieDevianceRegressionObjective : Regression
 
       m_variancePowerParamSub1 = variancePowerParamSub1;
       m_variancePowerParamSub2 = variancePowerParamSub2;
-      m_inverseVariancePowerParamSub1 = 1.0 / variancePowerParamSub1;
+      m_negInverseVariancePowerParamSub1 = (-1.0) / variancePowerParamSub1;
       m_inverseVariancePowerParamSub2 = 1.0 / variancePowerParamSub2;
    }
 
@@ -79,8 +79,8 @@ template<typename TFloat> struct TweedieDevianceRegressionObjective : Regression
    GPU_DEVICE inline TFloat CalcMetric(const TFloat& score, const TFloat& target) const noexcept {
       const TFloat exp1Score = Exp(m_variancePowerParamSub1 * score);
       const TFloat exp2Score = Exp(m_variancePowerParamSub2 * score);
-      const TFloat metric = FusedNegateMultiplyAdd(
-            target * m_inverseVariancePowerParamSub1, exp1Score, exp2Score * m_inverseVariancePowerParamSub2);
+      const TFloat metric = FusedMultiplyAdd(
+            target * m_negInverseVariancePowerParamSub1, exp1Score, exp2Score * m_inverseVariancePowerParamSub2);
       return metric;
    }