Sk2x::invert() and Sk2x::approxInvert()

BUG=skia:

Review URL: https://codereview.chromium.org/1024993002

5 files changed, 45 insertions, 4 deletions

diff --git a/src/core/Sk2x.h b/src/core/Sk2x.h
index 9b4e5ee61c..f49442563a 100644
--- a/src/core/Sk2x.h
+++ b/src/core/Sk2x.h
@@ -67,6 +67,9 @@ public:
     Sk2x rsqrt() const;   // Approximate 1/this->sqrt().
     Sk2x  sqrt() const;   // this->multiply(this->rsqrt()) may be faster, but less precise.
 
+    Sk2x       invert() const;  // 1/this.
+    Sk2x approxInvert() const;  // Approximate 1/this, usually faster but less precise.
+
     static Sk2x Min(const Sk2x&, const Sk2x&);
     static Sk2x Max(const Sk2x&, const Sk2x&);
 
diff --git a/src/opts/Sk2x_neon.h b/src/opts/Sk2x_neon.h
index ef61df4823..8e6e46164b 100644
--- a/src/opts/Sk2x_neon.h
+++ b/src/opts/Sk2x_neon.h
@@ -38,6 +38,18 @@ M(Sk2f&) operator=(const Sk2f& o) { fVec = o.fVec; return *this; }
 M(Sk2f) Load(const float vals[2]) { return vld1_f32(vals); }
 M(void) store(float vals[2]) const { vst1_f32(vals, fVec); }
 
+M(Sk2f) approxInvert() const {
+    float32x2_t est0 = vrecpe_f32(fVec),
+                est1 = vmul_f32(vrecps_f32(est0, fVec), est0);
+    return est1;
+}
+
+M(Sk2f) invert() const {
+    float32x2_t est1 = this->approxInvert().fVec,
+                est2 = vmul_f32(vrecps_f32(est1, fVec), est1);
+    return est2;
+}
+
 M(Sk2f)      add(const Sk2f& o) const { return vadd_f32(fVec, o.fVec); }
 M(Sk2f) subtract(const Sk2f& o) const { return vsub_f32(fVec, o.fVec); }
 M(Sk2f) multiply(const Sk2f& o) const { return vmul_f32(fVec, o.fVec); }
@@ -45,10 +57,7 @@ M(Sk2f)   divide(const Sk2f& o) const {
 #if defined(SK_CPU_ARM64)
     return vdiv_f32(fVec, o.fVec);
 #else
-    float32x2_t est0 = vrecpe_f32(o.fVec),
-                est1 = vmul_f32(vrecps_f32(est0, o.fVec), est0),
-                est2 = vmul_f32(vrecps_f32(est1, o.fVec), est1);
-    return vmul_f32(est2, fVec);
+    return vmul_f32(fVec, o.invert().fVec);
 #endif
 }
 
@@ -99,6 +108,19 @@ M(Sk2f)  sqrt() const {
     }
     M(Sk2d)  sqrt() const { return vsqrtq_f64(fVec); }
 
+    M(Sk2d) approxInvert() const {
+        float64x2_t est0 = vrecpeq_f64(fVec),
+                    est1 = vmulq_f64(vrecpsq_f64(est0, fVec), est0);
+        return est1;
+    }
+
+    M(Sk2d) invert() const {
+        float64x2_t est1 = this->approxInvert().fVec,
+                    est2 = vmulq_f64(vrecpsq_f64(est1, fVec), est1),
+                    est3 = vmulq_f64(vrecpsq_f64(est2, fVec), est2);
+        return est3;
+    }
+
 #else  // Scalar implementation for 32-bit chips, which don't have float64x2_t.
     M() Sk2x() {}
     M() Sk2x(double val)         { fVec[0] = fVec[1] = val; }
@@ -126,6 +148,9 @@ M(Sk2f)  sqrt() const {
 
     M(Sk2d) rsqrt() const { return Sk2d(1.0/::sqrt(fVec[0]), 1.0/::sqrt(fVec[1])); }
     M(Sk2d)  sqrt() const { return Sk2d(    ::sqrt(fVec[0]),     ::sqrt(fVec[1])); }
+
+    M(Sk2d)       invert() const { return Sk2d(1.0 / fVec[0], 1.0 / fVec[1]); }
+    M(Sk2d) approxInvert() const { return this->invert(); }
 #endif
 
 #undef M
diff --git a/src/opts/Sk2x_none.h b/src/opts/Sk2x_none.h
index 65cc6703fb..2c68e736f4 100644
--- a/src/opts/Sk2x_none.h
+++ b/src/opts/Sk2x_none.h
@@ -54,6 +54,9 @@ M(Sk2x<T>) Max(const Sk2x<T>& a, const Sk2x<T>& b) {
     return Sk2x<T>(SkTMax(a.fVec[0], b.fVec[0]), SkTMax(a.fVec[1], b.fVec[1]));
 }
 
+M(Sk2x<T>)       invert() const { return Sk2x<T>((T)1.0 / fVec[0], (T)1.0 / fVec[1]); }
+M(Sk2x<T>) approxInvert() const { return this->invert(); }
+
 #undef M
 
 #define M template <> inline
diff --git a/src/opts/Sk2x_sse.h b/src/opts/Sk2x_sse.h
index 111d3c230e..1136f1d856 100644
--- a/src/opts/Sk2x_sse.h
+++ b/src/opts/Sk2x_sse.h
@@ -46,6 +46,9 @@ M(Sk2f) Max(const Sk2f& a, const Sk2f& b) { return _mm_max_ps(a.fVec, b.fVec); }
 M(Sk2f) rsqrt() const { return _mm_rsqrt_ps(fVec); }
 M(Sk2f)  sqrt() const { return _mm_sqrt_ps (fVec); }
 
+M(Sk2f)       invert() const { return Sk2f(1.0f) / *this; }
+M(Sk2f) approxInvert() const { return _mm_rcp_ps(fVec); }
+
 #undef M
 
 #define M(...) template <> inline __VA_ARGS__ Sk2x<double>::
@@ -70,6 +73,10 @@ M(Sk2d) Max(const Sk2d& a, const Sk2d& b) { return _mm_max_pd(a.fVec, b.fVec); }
 M(Sk2d) rsqrt() const { return _mm_cvtps_pd(_mm_rsqrt_ps(_mm_cvtpd_ps(fVec))); }
 M(Sk2d)  sqrt() const { return _mm_sqrt_pd(fVec); }
 
+// No _mm_rcp_pd, so do Sk2d::approxInvert() in floats.
+M(Sk2d)       invert() const { return Sk2d(1.0) / *this; }
+M(Sk2d) approxInvert() const { return _mm_cvtps_pd(_mm_rcp_ps(_mm_cvtpd_ps(fVec))); }
+
 #undef M
 
 #endif
diff --git a/tests/Sk2xTest.cpp b/tests/Sk2xTest.cpp
index b74cc5bf2f..40e90913a5 100644
--- a/tests/Sk2xTest.cpp
+++ b/tests/Sk2xTest.cpp
@@ -49,6 +49,9 @@ static void test(skiatest::Reporter* r) {
     REPORTER_ASSERT(r, nearly_eq(0.001, a.rsqrt(), 0.5, 0.5));
     REPORTER_ASSERT(r, eq(a.sqrt(), 2, 2));
 
+    REPORTER_ASSERT(r, nearly_eq(0.001, d.approxInvert(), 0.5, 0.2));
+    REPORTER_ASSERT(r, eq(d.invert(), 0.5, 0.2));
+
     REPORTER_ASSERT(r, eq(Sk2x<T>::Min(a, d), 2, 4));
     REPORTER_ASSERT(r, eq(Sk2x<T>::Max(a, d), 4, 5));