Sk4x4f

   An API for loading and storing 4 Sk4f with transpose.
   This has SSSE3+ and portable versions.
   SSE2 and NEON versions to follow.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1825663002

Review URL: https://codereview.chromium.org/1825663002

3 files changed, 214 insertions, 53 deletions

diff --git a/src/core/Sk4x4f.h b/src/core/Sk4x4f.h
new file mode 100644
index 0000000000..f7c0566fc5
--- /dev/null
+++ b/src/core/Sk4x4f.h
@@ -0,0 +1,137 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef Sk4x4f_DEFINED
+#define Sk4x4f_DEFINED
+
+#include "SkNx.h"
+
+struct Sk4x4f {
+    Sk4f r,g,b,a;
+
+    static Sk4x4f Transpose(const Sk4f&, const Sk4f&, const Sk4f&, const Sk4f&);
+    static Sk4x4f Transpose(const   float[16]);
+    static Sk4x4f Transpose(const uint8_t[16]);
+
+    void transpose(Sk4f*, Sk4f*, Sk4f*, Sk4f*) const;
+    void transpose(  float[16]) const;
+    void transpose(uint8_t[16]) const;
+};
+
+// TODO: SSE2, NEON
+
+#if 1 && !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+
+inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, const Sk4f& w) {
+    auto r = x.fVec,
+         g = y.fVec,
+         b = z.fVec,
+         a = w.fVec;
+    _MM_TRANSPOSE4_PS(r,g,b,a);
+    return { r,g,b,a };
+}
+
+inline Sk4x4f Sk4x4f::Transpose(const float fs[16]) {
+    return Transpose(Sk4f::Load(fs+0), Sk4f::Load(fs+4), Sk4f::Load(fs+8), Sk4f::Load(fs+12));
+}
+
+inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) {
+    auto b16 = _mm_loadu_si128((const __m128i*)bs);
+    auto _ = ~0;  // Shuffles in a zero byte.
+    auto r = _mm_cvtepi32_ps(
+            _mm_shuffle_epi8(b16, _mm_setr_epi8(0,_,_,_,4,_,_,_, 8,_,_,_,12,_,_,_)));
+    auto g = _mm_cvtepi32_ps(
+            _mm_shuffle_epi8(b16, _mm_setr_epi8(1,_,_,_,5,_,_,_, 9,_,_,_,13,_,_,_)));
+    auto b = _mm_cvtepi32_ps(
+            _mm_shuffle_epi8(b16, _mm_setr_epi8(2,_,_,_,6,_,_,_,10,_,_,_,14,_,_,_)));
+    auto a = _mm_cvtepi32_ps(
+            _mm_shuffle_epi8(b16, _mm_setr_epi8(3,_,_,_,7,_,_,_,11,_,_,_,15,_,_,_)));
+    return { r,g,b,a };
+}
+
+inline void Sk4x4f::transpose(Sk4f* x, Sk4f* y, Sk4f* z, Sk4f* w) const {
+    auto R = r.fVec,
+         G = g.fVec,
+         B = b.fVec,
+         A = a.fVec;
+    _MM_TRANSPOSE4_PS(R,G,B,A);
+    *x = R;
+    *y = G;
+    *z = B;
+    *w = A;
+}
+
+inline void Sk4x4f::transpose(float fs[16]) const {
+    Sk4f x,y,z,w;
+    this->transpose(&x,&y,&z,&w);
+    x.store(fs+ 0);
+    y.store(fs+ 4);
+    z.store(fs+ 8);
+    w.store(fs+12);
+}
+
+inline void Sk4x4f::transpose(uint8_t bs[16]) const {
+    auto packed = _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(r.fVec),
+                                                    _mm_cvttps_epi32(g.fVec)),
+                                   _mm_packus_epi16(_mm_cvttps_epi32(b.fVec),
+                                                    _mm_cvttps_epi32(a.fVec)));
+    _mm_storeu_si128((__m128i*)bs, _mm_shuffle_epi8(packed, _mm_setr_epi8(0, 4,  8, 12,
+                                                                          1, 5,  9, 13,
+                                                                          2, 6, 10, 14,
+                                                                          3, 7, 11, 15)));
+}
+
+#else
+
+inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, const Sk4f& w) {
+    return {
+        { x[0], y[0], z[0], w[0] },
+        { x[1], y[1], z[1], w[1] },
+        { x[2], y[2], z[2], w[2] },
+        { x[3], y[3], z[3], w[3] },
+    };
+}
+
+inline Sk4x4f Sk4x4f::Transpose(const float fs[16]) {
+    return Transpose(Sk4f::Load(fs+0), Sk4f::Load(fs+4), Sk4f::Load(fs+8), Sk4f::Load(fs+12));
+}
+
+inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) {
+    return {
+        { (float)bs[0], (float)bs[4], (float)bs[ 8], (float)bs[12] },
+        { (float)bs[1], (float)bs[5], (float)bs[ 9], (float)bs[13] },
+        { (float)bs[2], (float)bs[6], (float)bs[10], (float)bs[14] },
+        { (float)bs[3], (float)bs[7], (float)bs[11], (float)bs[15] },
+    };
+}
+
+inline void Sk4x4f::transpose(Sk4f* x, Sk4f* y, Sk4f* z, Sk4f* w) const {
+    *x = { r[0], g[0], b[0], a[0] };
+    *y = { r[1], g[1], b[1], a[1] };
+    *z = { r[2], g[2], b[2], a[2] };
+    *w = { r[3], g[3], b[3], a[3] };
+}
+
+inline void Sk4x4f::transpose(float fs[16]) const {
+    Sk4f x,y,z,w;
+    this->transpose(&x,&y,&z,&w);
+    x.store(fs+ 0);
+    y.store(fs+ 4);
+    z.store(fs+ 8);
+    w.store(fs+12);
+}
+
+inline void Sk4x4f::transpose(uint8_t bs[16]) const {
+    bs[ 0] = (uint8_t)r[0]; bs[ 1] = (uint8_t)g[0]; bs[ 2] = (uint8_t)b[0]; bs[ 3] = (uint8_t)a[0];
+    bs[ 4] = (uint8_t)r[1]; bs[ 5] = (uint8_t)g[1]; bs[ 6] = (uint8_t)b[1]; bs[ 7] = (uint8_t)a[1];
+    bs[ 8] = (uint8_t)r[2]; bs[ 9] = (uint8_t)g[2]; bs[10] = (uint8_t)b[2]; bs[11] = (uint8_t)a[2];
+    bs[12] = (uint8_t)r[3]; bs[13] = (uint8_t)g[3]; bs[14] = (uint8_t)b[3]; bs[15] = (uint8_t)a[3];
+}
+
+#endif
+
+#endif//Sk4x4f_DEFINED
diff --git a/src/core/SkXfermode4f.cpp b/src/core/SkXfermode4f.cpp
index 4e444a791a..883cd12757 100644
--- a/src/core/SkXfermode4f.cpp
+++ b/src/core/SkXfermode4f.cpp
@@ -8,6 +8,7 @@
 #include "SkPM4fPriv.h"
 #include "SkUtils.h"
 #include "SkXfermode.h"
+#include "Sk4x4f.h"
 
 static SkPM4f rgba_to_pmcolor_order(const SkPM4f& x) {
 #ifdef SK_PMCOLOR_IS_BGRA
@@ -235,81 +236,50 @@ const SkXfermode::D32Proc gProcs_Dst[] = {
 
 
 static void srcover_n_srgb_bw(uint32_t dst[], const SkPM4f src[], int count) {
-#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3  // For _mm_shuffle_epi8
     while (count >= 4) {
         // Load 4 sRGB RGBA/BGRA 8888 dst pixels.
         // We'll write most of this as if they're RGBA, and just swizzle the src pixels to match.
-        __m128i d4 = _mm_loadu_si128((const __m128i*)dst);
-
-        // Transpose into planar and convert each plane to float.
-        auto _ = ~0;  // Shuffles in a zero byte.
-        auto dr = _mm_cvtepi32_ps(
-                _mm_shuffle_epi8(d4, _mm_setr_epi8(0,_,_,_, 4,_,_,_, 8,_,_,_,12,_,_,_)));
-        auto dg = _mm_cvtepi32_ps(
-                _mm_shuffle_epi8(d4, _mm_setr_epi8(1,_,_,_, 5,_,_,_, 9,_,_,_,13,_,_,_)));
-        auto db = _mm_cvtepi32_ps(
-                _mm_shuffle_epi8(d4, _mm_setr_epi8(2,_,_,_, 6,_,_,_,10,_,_,_,14,_,_,_)));
-        auto da = _mm_cvtepi32_ps(
-                _mm_shuffle_epi8(d4, _mm_setr_epi8(3,_,_,_, 7,_,_,_,11,_,_,_,15,_,_,_)));
+        auto d = Sk4x4f::Transpose((const uint8_t*)dst);
 
         // Scale to [0,1].
-        dr = _mm_mul_ps(dr, _mm_set1_ps(1/255.0f));
-        dg = _mm_mul_ps(dg, _mm_set1_ps(1/255.0f));
-        db = _mm_mul_ps(db, _mm_set1_ps(1/255.0f));
-        da = _mm_mul_ps(da, _mm_set1_ps(1/255.0f));
+        d.r *= 1/255.0f;
+        d.g *= 1/255.0f;
+        d.b *= 1/255.0f;
+        d.a *= 1/255.0f;
 
         // Apply approximate sRGB gamma correction to convert to linear (as if gamma were 2).
-        dr = _mm_mul_ps(dr, dr);
-        dg = _mm_mul_ps(dg, dg);
-        db = _mm_mul_ps(db, db);
+        d.r *= d.r;
+        d.g *= d.g;
+        d.b *= d.b;
 
         // Load 4 linear float src pixels.
-        auto s0 = _mm_loadu_ps(src[0].fVec),
-             s1 = _mm_loadu_ps(src[1].fVec),
-             s2 = _mm_loadu_ps(src[2].fVec),
-             s3 = _mm_loadu_ps(src[3].fVec);
-
-        // Transpose src pixels to planar too, and give the registers better names.
-        _MM_TRANSPOSE4_PS(s0, s1, s2, s3);
-        auto sr = s0,
-             sg = s1,
-             sb = s2,
-             sa = s3;
+        auto s = Sk4x4f::Transpose(src->fVec);
 
         // Match color order with destination, if necessary.
     #if defined(SK_PMCOLOR_IS_BGRA)
-        SkTSwap(sr, sb);
+        SkTSwap(s.r, s.b);
     #endif
 
         // Now, the meat of what we wanted to do... perform the srcover blend.
-        auto invSA = _mm_sub_ps(_mm_set1_ps(1), sa);
-        auto r = _mm_add_ps(sr, _mm_mul_ps(dr, invSA)),
-             g = _mm_add_ps(sg, _mm_mul_ps(dg, invSA)),
-             b = _mm_add_ps(sb, _mm_mul_ps(db, invSA)),
-             a = _mm_add_ps(sa, _mm_mul_ps(da, invSA));
+        auto invSA = 1.0f - s.a;
+        auto r = s.r + d.r * invSA,
+             g = s.g + d.g * invSA,
+             b = s.b + d.b * invSA,
+             a = s.a + d.a * invSA;
 
         // Convert back to sRGB and [0,255], again approximating sRGB as gamma == 2.
-        r = _mm_mul_ps(_mm_sqrt_ps(r), _mm_set1_ps(255));
-        g = _mm_mul_ps(_mm_sqrt_ps(g), _mm_set1_ps(255));
-        b = _mm_mul_ps(_mm_sqrt_ps(b), _mm_set1_ps(255));
-        a = _mm_mul_ps(           (a), _mm_set1_ps(255));
-
-        // Convert to int (with rounding) and pack back down to planar 8-bit.
-        __m128i x = _mm_packus_epi16(_mm_packus_epi16(_mm_cvtps_epi32(r), _mm_cvtps_epi32(g)),
-                                     _mm_packus_epi16(_mm_cvtps_epi32(b), _mm_cvtps_epi32(a)));
-
-        // Transpose back to interlaced RGBA and write back to dst.
-        x = _mm_shuffle_epi8(x, _mm_setr_epi8(0, 4,  8, 12,
-                                              1, 5,  9, 13,
-                                              2, 6, 10, 14,
-                                              3, 7, 11, 15));
-        _mm_storeu_si128((__m128i*)dst, x);
+        r = r.sqrt() * 255.0f + 0.5f;
+        g = g.sqrt() * 255.0f + 0.5f;
+        b = b.sqrt() * 255.0f + 0.5f;
+        a = a        * 255.0f + 0.5f;
+
+        Sk4x4f{r,g,b,a}.transpose((uint8_t*)dst);
 
         count -= 4;
         dst += 4;
         src += 4;
     }
-#endif
+
     // This should look just like the non-specialized case in srcover_n.
     for (int i = 0; i < count; ++i) {
         Sk4f s4 = src[i].to4f_pmorder();
diff --git a/tests/Sk4x4fTest.cpp b/tests/Sk4x4fTest.cpp
new file mode 100644
index 0000000000..100e13973f
--- /dev/null
+++ b/tests/Sk4x4fTest.cpp
@@ -0,0 +1,54 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "Sk4x4f.h"
+#include "Test.h"
+
+DEF_TEST(Sk4x4f, r) {
+    Sk4x4f f;
+
+    Sk4f x{ 0, 1, 2, 3},
+         y{ 4, 5, 6, 7},
+         z{ 8, 9,10,11},
+         w{12,13,14,15};
+    f = Sk4x4f::Transpose(x,y,z,w);
+    REPORTER_ASSERT(r, f.r[0] == 0 && f.r[1] == 4 && f.r[2] ==  8 && f.r[3] == 12);
+    REPORTER_ASSERT(r, f.g[0] == 1 && f.g[1] == 5 && f.g[2] ==  9 && f.g[3] == 13);
+    REPORTER_ASSERT(r, f.b[0] == 2 && f.b[1] == 6 && f.b[2] == 10 && f.b[3] == 14);
+    REPORTER_ASSERT(r, f.a[0] == 3 && f.a[1] == 7 && f.a[2] == 11 && f.a[3] == 15);
+
+    Sk4f s,t,u,v;
+    f.transpose(&s,&t,&u,&v);
+    REPORTER_ASSERT(r, (x == s).allTrue()
+                    && (y == t).allTrue()
+                    && (z == u).allTrue()
+                    && (w == v).allTrue());
+
+
+    float fs[16] = {0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15};
+    f = Sk4x4f::Transpose(fs);
+    REPORTER_ASSERT(r, f.r[0] == 0 && f.r[1] == 4 && f.r[2] ==  8 && f.r[3] == 12);
+    REPORTER_ASSERT(r, f.g[0] == 1 && f.g[1] == 5 && f.g[2] ==  9 && f.g[3] == 13);
+    REPORTER_ASSERT(r, f.b[0] == 2 && f.b[1] == 6 && f.b[2] == 10 && f.b[3] == 14);
+    REPORTER_ASSERT(r, f.a[0] == 3 && f.a[1] == 7 && f.a[2] == 11 && f.a[3] == 15);
+
+    float fs_back[16];
+    f.transpose(fs_back);
+    REPORTER_ASSERT(r, 0 == memcmp(fs, fs_back, sizeof(fs)));
+
+
+    uint8_t bs[16] = {0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15};
+    f = Sk4x4f::Transpose(bs);
+    REPORTER_ASSERT(r, f.r[0] == 0 && f.r[1] == 4 && f.r[2] ==  8 && f.r[3] == 12);
+    REPORTER_ASSERT(r, f.g[0] == 1 && f.g[1] == 5 && f.g[2] ==  9 && f.g[3] == 13);
+    REPORTER_ASSERT(r, f.b[0] == 2 && f.b[1] == 6 && f.b[2] == 10 && f.b[3] == 14);
+    REPORTER_ASSERT(r, f.a[0] == 3 && f.a[1] == 7 && f.a[2] == 11 && f.a[3] == 15);
+
+    uint8_t bs_back[16];
+    f.transpose(bs_back);
+    REPORTER_ASSERT(r, 0 == memcmp(bs, bs_back, sizeof(bs)));
+}