SkNx miplevel building

All sizes approximately twice as fast.

Before:
    micros   	bench
   1649.35  	mipmap_build_512x512	nonrendering
   1824.42  	mipmap_build_511x512	nonrendering
   2100.66 ?	mipmap_build_512x511	nonrendering
   2375.94  	mipmap_build_511x511	nonrendering

After:
    micros   	bench
    730.32 !	mipmap_build_512x512	nonrendering
    922.12  	mipmap_build_511x512	nonrendering
    999.07  	mipmap_build_512x511	nonrendering
   1342.93 !	mipmap_build_511x511	nonrendering

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1606013003
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Committed: https://skia.googlesource.com/skia/+/3bd5aba2a0e165997f683cf3aa306661e71464f6

Review URL: https://codereview.chromium.org/1606013003

3 files changed, 70 insertions, 3 deletions

diff --git a/src/core/SkMipMap.cpp b/src/core/SkMipMap.cpp
index d468507ac7..14e87a33a3 100644
--- a/src/core/SkMipMap.cpp
+++ b/src/core/SkMipMap.cpp
@@ -8,6 +8,7 @@
 #include "SkMipMap.h"
 #include "SkBitmap.h"
 #include "SkColorPriv.h"
+#include "SkNx.h"
 
 //
 // ColorTypeFilter is the "Type" we pass to some downsample template functions.
@@ -18,12 +19,23 @@
 
 struct ColorTypeFilter_8888 {
     typedef uint32_t Type;
+#if defined(SKNX_IS_FAST)
+    static Sk4h Expand(uint32_t x) {
+        return SkNx_cast<uint16_t>(Sk4b::Load((const uint8_t*)&x));
+    }
+    static uint32_t Compact(const Sk4h& x) {
+        uint32_t r;
+        SkNx_cast<uint8_t>(x).store((uint8_t*)&r);
+        return r;
+    }
+#else
     static uint64_t Expand(uint32_t x) {
         return (x & 0xFF00FF) | ((uint64_t)(x & 0xFF00FF00) << 24);
     }
     static uint32_t Compact(uint64_t x) {
         return (uint32_t)((x & 0xFF00FF) | ((x >> 24) & 0xFF00FF00));
     }
+#endif
 };
 
 struct ColorTypeFilter_565 {
@@ -56,7 +68,7 @@ struct ColorTypeFilter_8 {
     }
 };
 
-template <typename T> T add_121(T a, T b, T c) {
+template <typename T> T add_121(const T& a, const T& b, const T& c) {
     return a + b + b + c;
 }
 
@@ -93,7 +105,7 @@ template <typename F> void downsample_3_2(void* dst, const void* src, size_t src
     auto p0 = static_cast<const typename F::Type*>(src);
     auto p1 = (const typename F::Type*)((const char*)p0 + srcRB);
     auto d = static_cast<typename F::Type*>(dst);
-    
+
     auto c02 = F::Expand(p0[0]);
     auto c12 = F::Expand(p1[0]);
     for (int i = 0; i < count; ++i) {
@@ -116,7 +128,7 @@ template <typename F> void downsample_2_3(void* dst, const void* src, size_t src
     auto p1 = (const typename F::Type*)((const char*)p0 + srcRB);
     auto p2 = (const typename F::Type*)((const char*)p1 + srcRB);
     auto d = static_cast<typename F::Type*>(dst);
-    
+
     for (int i = 0; i < count; ++i) {
         auto c00 = F::Expand(p0[0]);
         auto c01 = F::Expand(p0[1]);
diff --git a/src/opts/SkNx_neon.h b/src/opts/SkNx_neon.h
index c04f4d5f91..8adb276064 100644
--- a/src/opts/SkNx_neon.h
+++ b/src/opts/SkNx_neon.h
@@ -230,6 +230,45 @@ public:
     float32x4_t fVec;
 };
 
+// It's possible that for our current use cases, representing this as
+// half a uint16x8_t might be better than representing it as a uint16x4_t.
+// It'd make conversion to Sk4b one step simpler.
+template <>
+class SkNx<4, uint16_t> {
+public:
+    SkNx(const uint16x4_t& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint16_t val) : fVec(vdup_n_u16(val)) {}
+    static SkNx Load(const uint16_t vals[4]) { return vld1_u16(vals); }
+
+    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) {
+        fVec = (uint16x4_t) { a,b,c,d };
+    }
+
+    void store(uint16_t vals[4]) const { vst1_u16(vals, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return vadd_u16(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return vsub_u16(fVec, o.fVec); }
+    SkNx operator * (const SkNx& o) const { return vmul_u16(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { SHIFT16(vshl_n_u16, fVec, bits); }
+    SkNx operator >> (int bits) const { SHIFT16(vshr_n_u16, fVec, bits); }
+
+    static SkNx Min(const SkNx& a, const SkNx& b) { return vmin_u16(a.fVec, b.fVec); }
+
+    template <int k> uint16_t kth() const {
+        SkASSERT(0 <= k && k < 4);
+        return vget_lane_u16(fVec, k&3);
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+        return vbsl_u16(fVec, t.fVec, e.fVec);
+    }
+
+    uint16x4_t fVec;
+};
+
 template <>
 class SkNx<8, uint16_t> {
 public:
@@ -351,6 +390,14 @@ static inline void Sk4f_ToBytes(uint8_t bytes[16],
                                       (uint8x16_t)vcvtq_u32_f32(d.fVec)).val[0]).val[0]);
 }
 
+template<> inline Sk4h SkNx_cast<uint16_t, uint8_t, 4>(const Sk4b& src) {
+    return vget_low_u16(vmovl_u8(src.fVec));
+}
+
+template<> inline Sk4b SkNx_cast<uint8_t, uint16_t, 4>(const Sk4h& src) {
+    return vmovn_u16(vcombine_u16(src.fVec, src.fVec));
+}
+
 }  // namespace
 
 #endif//SkNx_neon_DEFINED
diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index c6163b6a80..a17d988ee7 100644
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -379,6 +379,14 @@ static inline void Sk4f_ToBytes(uint8_t bytes[16],
                                                        _mm_cvttps_epi32(d.fVec))));
 }
 
+template<> inline Sk4h SkNx_cast<uint16_t, uint8_t, 4>(const Sk4b& src) {
+    return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
+}
+
+template<> inline Sk4b SkNx_cast<uint8_t, uint16_t, 4>(const Sk4h& src) {
+    return _mm_packus_epi16(src.fVec, src.fVec);
+}
+
 
 }  // namespace