Sk4f: add floor()

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1685773002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review URL: https://codereview.chromium.org/1685773002

2 files changed, 31 insertions, 0 deletions

diff --git a/src/opts/SkNx_neon.h b/src/opts/SkNx_neon.h
index 2cb8eb348d..cdc4615849 100644
--- a/src/opts/SkNx_neon.h
+++ b/src/opts/SkNx_neon.h
@@ -10,6 +10,17 @@
 
 #define SKNX_IS_FAST
 
+// ARMv8 has vrndmq_f32 to floor 4 floats.  Here we emulate it:
+//   - round by adding (1<<23) with our sign, then subtracting it;
+//   - if that rounded value is bigger than our input, subtract 1.
+static inline float32x4_t armv7_vrndmq_f32(float32x4_t v) {
+    auto sign = vandq_u32((uint32x4_t)v, vdupq_n_u32(1<<31));
+    auto bias = (float32x4_t)(vorrq_u32((uint32x4_t)vdupq_n_f32(1<<23), sign));
+    auto rounded = vsubq_f32(vaddq_f32(v, bias), bias);
+    auto too_big = vcgtq_f32(rounded, v);
+    return vsubq_f32(rounded, (float32x4_t)vandq_u32(too_big, (uint32x4_t)vdupq_n_f32(1)));
+}
+
 // Well, this is absurd.  The shifts require compile-time constant arguments.
 
 #define SHIFT8(op, v, bits) switch(bits) { \
@@ -161,6 +172,14 @@ public:
     static SkNx Max(const SkNx& l, const SkNx& r) { return vmaxq_f32(l.fVec, r.fVec); }
 
     SkNx abs() const { return vabsq_f32(fVec); }
+    SkNx floor() const {
+    #if defined(SK_CPU_ARM64)
+        return vrndmq_f32(fVec);
+    #else
+        return armv7_vrndmq_f32(fVec);
+    #endif
+    }
+
 
     SkNx rsqrt0() const { return vrsqrteq_f32(fVec); }
     SkNx rsqrt1() const {
diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index 889253b9d8..def7ba1719 100644
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -13,6 +13,17 @@
 
 #define SKNX_IS_FAST
 
+// SSE 4.1 has _mm_floor_ps to floor 4 floats.  We emulate it:
+//   - round by adding (1<<23) with our sign, then subtracting it;
+//   - if that rounded value is bigger than our input, subtract 1.
+static inline __m128 sse2_mm_floor_ps(__m128 v) {
+    __m128 sign = _mm_and_ps(v, _mm_set1_ps(-0.0f));
+    __m128 bias = _mm_or_ps(sign, _mm_set1_ps(1<<23));
+    __m128 rounded = _mm_sub_ps(_mm_add_ps(v, bias), bias);
+    __m128 too_big = _mm_cmpgt_ps(rounded, v);
+    return _mm_sub_ps(rounded, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
+}
+
 template <>
 class SkNx<2, float> {
 public:
@@ -92,6 +103,7 @@ public:
     static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
 
     SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
+    SkNx floor() const { return sse2_mm_floor_ps(fVec); }
 
     SkNx  sqrt () const { return _mm_sqrt_ps (fVec);  }
     SkNx rsqrt0() const { return _mm_rsqrt_ps(fVec); }