Specialize Sk2d for SSE2
Given the autovectorization we've seen, I wouldn't expect big speedups
from this, but it does give us a point of control over what's going on.
BUG=skia:
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review URL: https://codereview.chromium.org/1526923003
diff --git a/src/core/SkNx.h b/src/core/SkNx.h
index a28898b..684a31d 100644
--- a/src/core/SkNx.h
+++ b/src/core/SkNx.h
@@ -203,11 +203,16 @@
} // namespace
typedef SkNx<2, float> Sk2f;
-typedef SkNx<2, float> Sk2s;
typedef SkNx<4, float> Sk4f;
-typedef SkNx<4, float> Sk4s;
typedef SkNx<8, float> Sk8f;
-typedef SkNx<8, float> Sk8s;
+
+typedef SkNx<2, double> Sk2d;
+typedef SkNx<4, double> Sk4d;
+typedef SkNx<8, double> Sk8d;
+
+typedef SkNx<2, SkScalar> Sk2s;
+typedef SkNx<4, SkScalar> Sk4s;
+typedef SkNx<8, SkScalar> Sk8s;
typedef SkNx< 4, uint16_t> Sk4h;
typedef SkNx< 8, uint16_t> Sk8h;
diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index f3952a9..b3f5091 100644
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -65,6 +65,52 @@
};
template <>
+class SkNx<2, double> {
+public:
+ SkNx(const __m128d& vec) : fVec(vec) {}
+
+ SkNx() {}
+ SkNx(double val) : fVec(_mm_set1_pd(val)) {}
+ static SkNx Load(const double vals[2]) { return _mm_loadu_pd(vals); }
+ SkNx(double a, double b) : fVec(_mm_setr_pd(a,b)) {}
+
+ void store(double vals[2]) const { _mm_storeu_pd(vals, fVec); }
+
+ SkNx operator + (const SkNx& o) const { return _mm_add_pd(fVec, o.fVec); }
+ SkNx operator - (const SkNx& o) const { return _mm_sub_pd(fVec, o.fVec); }
+ SkNx operator * (const SkNx& o) const { return _mm_mul_pd(fVec, o.fVec); }
+ SkNx operator / (const SkNx& o) const { return _mm_div_pd(fVec, o.fVec); }
+
+ SkNx operator == (const SkNx& o) const { return _mm_cmpeq_pd (fVec, o.fVec); }
+ SkNx operator != (const SkNx& o) const { return _mm_cmpneq_pd(fVec, o.fVec); }
+ SkNx operator < (const SkNx& o) const { return _mm_cmplt_pd (fVec, o.fVec); }
+ SkNx operator > (const SkNx& o) const { return _mm_cmpgt_pd (fVec, o.fVec); }
+ SkNx operator <= (const SkNx& o) const { return _mm_cmple_pd (fVec, o.fVec); }
+ SkNx operator >= (const SkNx& o) const { return _mm_cmpge_pd (fVec, o.fVec); }
+
+ static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_pd(l.fVec, r.fVec); }
+ static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_pd(l.fVec, r.fVec); }
+
+ SkNx sqrt() const { return _mm_sqrt_pd(fVec); }
+
+ template <int k> double kth() const {
+ SkASSERT(0 <= k && k < 2);
+ union { __m128d v; double fs[2]; } pun = {fVec};
+ return pun.fs[k&1];
+ }
+
+ bool allTrue() const { return 0x3 == _mm_movemask_pd(fVec); }
+ bool anyTrue() const { return 0x0 != _mm_movemask_pd(fVec); }
+
+ SkNx thenElse(const SkNx& t, const SkNx& e) const {
+ return _mm_or_pd(_mm_and_pd (fVec, t.fVec),
+ _mm_andnot_pd(fVec, e.fVec));
+ }
+
+ __m128d fVec;
+};
+
+template <>
class SkNx<4, int> {
public:
SkNx(const __m128i& vec) : fVec(vec) {}