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

Committed: https://skia.googlesource.com/skia/+/86c6c4935171a1d2d6a9ffbff37ec6dac1326614

CQ_EXTRA_TRYBOTS=client.skia.android:Test-Android-GCC-Nexus7-GPU-Tegra3-Arm7-Release-Trybot,Test-Android-GCC-Nexus9-GPU-TegraK1-Arm64-Release-Trybot

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

bench/Sk4fBench.cpp

@@ -40,6 +40,22 @@ DEF_BENCH(return new Sk4fRoundtripBench<uint8_t>;)
 DEF_BENCH(return new Sk4fRoundtripBench<uint16_t>;)
 DEF_BENCH(return new Sk4fRoundtripBench<int>;)
 
+struct Sk4fFloorBench : public Benchmark {
+    Sk4fFloorBench() {}
+
+    const char* onGetName() override { return "Sk4f_floor"; }
+    bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; }
+
+    void onDraw(int loops, SkCanvas* canvas) override {
+        Sk4f fs(1,2,3,4);
+        while (loops --> 0) {
+            fs = fs.floor();
+        }
+        fs.store((float*)blackhole);
+    }
+};
+DEF_BENCH(return new Sk4fFloorBench;)
+
 struct Sk4fGradientBench : public Benchmark {
     const char* onGetName() override { return "Sk4f_gradient"; }
     bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; }

src/core/SkNx.h

@@ -50,7 +50,7 @@ public:
 #undef OP
 
 #define OP(op) SkNx op() const { return {fLo.op(), fHi.op()}; }
-    OP(abs)
+    OP(abs) OP(floor)
     OP(sqrt) OP(rsqrt0) OP(rsqrt1) OP(rsqrt2)
     OP(invert) OP(approxInvert)
 #undef OP
@@ -124,6 +124,7 @@ public:
     static SkNx Max(const SkNx& a, const SkNx& b) { return SkTMax(a.fVal, b.fVal); }
 
     SkNx abs() const { return SkTAbs(fVal); }
+    SkNx floor() const { return Floor(fVal); }
 
     SkNx  sqrt () const { return Sqrt(fVal); }
     SkNx rsqrt0() const { return this->sqrt().invert(); }
@@ -145,6 +146,8 @@ public:
     SkNx thenElse(const SkNx& t, const SkNx& e) const { return fVal != 0 ? t : e; }
 
 protected:
+    static double Floor(double val) { return ::floor (val); }
+    static float  Floor(float  val) { return ::floorf(val); }
     static double Sqrt(double val) { return ::sqrt (val); }
     static float  Sqrt(float  val) { return ::sqrtf(val); }
 

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 {

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); }

tests/SkNxTest.cpp

@@ -215,6 +215,14 @@ DEF_TEST(SkNx_abs, r) {
     REPORTER_ASSERT(r, fs.kth<3>() == 4.0f);
 }
 
+DEF_TEST(SkNx_floor, r) {
+    auto fs = Sk4f(0.4f, -0.4f, 0.6f, -0.6f).floor();
+    REPORTER_ASSERT(r, fs.kth<0>() ==  0.0f);
+    REPORTER_ASSERT(r, fs.kth<1>() == -1.0f);
+    REPORTER_ASSERT(r, fs.kth<2>() ==  0.0f);
+    REPORTER_ASSERT(r, fs.kth<3>() == -1.0f);
+}
+
 DEF_TEST(SkNx_shuffle, r) {
     Sk4f f4(0,10,20,30);