Sk4f: floor() via int32_t roundtrip.

About 25% faster on both x86 and ARMv7.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1682953002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review URL: https://codereview.chromium.org/1682953002

2 files changed, 12 insertions, 14 deletions

diff --git a/src/opts/SkNx_neon.h b/src/opts/SkNx_neon.h
index cdc4615849..641e9d296b 100644
--- a/src/opts/SkNx_neon.h
+++ b/src/opts/SkNx_neon.h
@@ -11,14 +11,13 @@
 #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.
+//   - roundtrip through integers via truncation
+//   - subtract 1 if that's too big (possible for negative values).
+// This restricts the domain of our inputs to a maximum somehwere around 2^31.  Seems plenty big.
 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)));
+    auto roundtrip = vcvtq_f32_s32(vcvtq_s32_f32(v));
+    auto too_big = vcgtq_f32(roundtrip, v);
+    return vsubq_f32(roundtrip, (float32x4_t)vandq_u32(too_big, (uint32x4_t)vdupq_n_f32(1)));
 }
 
 // Well, this is absurd.  The shifts require compile-time constant arguments.
diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index def7ba1719..7d00ddf9ba 100644
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -14,14 +14,13 @@
 #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.
+//   - roundtrip through integers via truncation
+//   - subtract 1 if that's too big (possible for negative values).
+// This restricts the domain of our inputs to a maximum somehwere around 2^31.  Seems plenty big.
 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)));
+    __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
+    __m128 too_big = _mm_cmpgt_ps(roundtrip, v);
+    return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
 }
 
 template <>