start experimenting with 64bit frame-buffers

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1680523002

Review URL: https://codereview.chromium.org/1680523002

5 files changed, 452 insertions, 0 deletions

diff --git a/bench/XferU64Bench.cpp b/bench/XferU64Bench.cpp
new file mode 100644
index 0000000000..0cc2ba131a
--- /dev/null
+++ b/bench/XferU64Bench.cpp
@@ -0,0 +1,106 @@
+/*
+ * 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 "Benchmark.h"
+#include "SkString.h"
+#include "SkXfermode.h"
+
+#define USE_AA      (1 << 31)   // merge with Xfermode::PMFlags w/o conflict
+
+#define INNER_LOOPS 1000
+
+// Benchmark that draws non-AA rects or AA text with an SkXfermode::Mode.
+class XferU64Bench : public Benchmark {
+public:
+    XferU64Bench(bool doN, uint32_t flags)
+        : fDoN(doN)
+        , fFlags(flags & ~USE_AA)
+    {
+        SkXfermode::Mode mode = SkXfermode::kSrcOver_Mode;
+
+        fProc1 = SkXfermode::GetU64Proc1(mode, fFlags);
+        fProcN = SkXfermode::GetU64ProcN(mode, fFlags);
+        fName.printf("xferu64_%s_%c_%s_%s",
+                     (flags & USE_AA) ? "aa" : "bw",
+                     fDoN ? 'N' : '1',
+                     (flags & SkXfermode::kSrcIsOpaque_U64Flag) ? "opaque" : "alpha",
+                     (flags & SkXfermode::kDstIsFloat16_U64Flag) ? "f16" : "u16");
+
+        for (int i = 0; i < N; ++i) {
+            fSrc[i] = {{ 1, 1, 1, 1 }};
+            fDst[i] = 0;
+            fAAStorage[i] = i * 255 / (N - 1);
+        }
+        
+        if (flags & USE_AA) {
+            fAA = fAAStorage;
+        } else {
+            fAA = nullptr;
+        }
+    }
+
+protected:
+    bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; }
+
+    const char* onGetName() override { return fName.c_str(); }
+
+    void onDraw(int loops, SkCanvas*) override {
+        const SkXfermode::U64State state{ nullptr, fFlags };
+
+        for (int i = 0; i < loops * INNER_LOOPS; ++i) {
+            if (fDoN) {
+                fProcN(state, fDst, fSrc, N, fAA);
+            } else {
+                fProc1(state, fDst, fSrc[0], N, fAA);
+            }
+        }
+    }
+
+private:
+    SkString             fName;
+    SkXfermode::U64Proc1 fProc1;
+    SkXfermode::U64ProcN fProcN;
+    const SkAlpha*  fAA;
+    bool            fDoN;
+    uint32_t        fFlags;
+
+    enum {
+        N = 1000,
+    };
+    SkPM4f      fSrc[N];
+    uint64_t    fDst[N];
+    uint8_t     fAAStorage[N];
+
+    typedef Benchmark INHERITED;
+};
+
+#define F00 0
+#define F01 (SkXfermode::kSrcIsOpaque_U64Flag)
+#define F10 (SkXfermode::kDstIsFloat16_U64Flag)
+#define F11 (SkXfermode::kDstIsFloat16_U64Flag | SkXfermode::kSrcIsOpaque_U64Flag)
+
+#if 0
+DEF_BENCH( return new XferU64Bench(true,  F10 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(true,  F11 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(true,  F10); )
+DEF_BENCH( return new XferU64Bench(true,  F11); )
+
+DEF_BENCH( return new XferU64Bench(true,  F00 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(true,  F01 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(true,  F00); )
+DEF_BENCH( return new XferU64Bench(true,  F01); )
+#endif
+
+DEF_BENCH( return new XferU64Bench(false, F10 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(false, F11 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(false, F10); )
+DEF_BENCH( return new XferU64Bench(false, F11); )
+
+DEF_BENCH( return new XferU64Bench(false, F00 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(false, F01 | USE_AA); )
+DEF_BENCH( return new XferU64Bench(false, F00); )
+DEF_BENCH( return new XferU64Bench(false, F01); )
diff --git a/gm/xferu64.cpp b/gm/xferu64.cpp
new file mode 100644
index 0000000000..03fc8864b5
--- /dev/null
+++ b/gm/xferu64.cpp
@@ -0,0 +1,149 @@
+/*
+ * 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 "gm.h"
+#include "SkCanvas.h"
+#include "SkImageInfo.h"
+#include "SkXfermode.h"
+#include "SkHalf.h"
+#include "SkImage.h"
+
+static SkPMColor f16_to_pmcolor(uint64_t src) {
+    SkPMColor dst;
+    const SkHalf* sptr = reinterpret_cast<const SkHalf*>(&src);
+    uint8_t* dptr = reinterpret_cast<uint8_t*>(&dst);
+    for (int i = 0; i < 4; ++i) {
+        float f = SkHalfToFloat(sptr[i]);
+        dptr[i] = SkToU8((int)(f * 255 + 0.5f));
+    }
+    return dst;
+}
+
+static SkPMColor u16_to_pmcolor(uint64_t src) {
+    SkPMColor dst;
+    const uint16_t* sptr = reinterpret_cast<const uint16_t*>(&src);
+    uint8_t* dptr = reinterpret_cast<uint8_t*>(&dst);
+    for (int i = 0; i < 4; ++i) {
+        dptr[i] = sptr[i] >> 8;
+    }
+    return dst;
+}
+
+static SkImage* new_u64_image(const SkBitmap& src, uint32_t flags) {
+    SkBitmap dst;
+    dst.allocN32Pixels(src.width(), src.height());
+    SkPixmap srcPM, dstPM;
+    src.peekPixels(&srcPM);
+    dst.peekPixels(&dstPM);
+
+    for (int y = 0; y < srcPM.height(); ++y) {
+        for (int x = 0; x < srcPM.width(); ++x) {
+            uint64_t srcP = *srcPM.addr64(x, y);
+            uint32_t* dstP = dstPM.writable_addr32(x, y);
+
+            if (flags & SkXfermode::kDstIsFloat16_U64Flag) {
+                *dstP = f16_to_pmcolor(srcP);
+            } else {
+                *dstP = u16_to_pmcolor(srcP);
+            }
+        }
+    }
+    return SkImage::NewRasterCopy(dstPM.info(), dstPM.addr(), dstPM.rowBytes());
+}
+
+static void draw_rect(SkCanvas* canvas, const SkRect& r, SkColor c, uint32_t u64_flags,
+                      const SkAlpha aa[]) {
+    const SkIRect ir = r.round();
+    const SkImageInfo info = SkImageInfo::Make(ir.width(), ir.height(),
+                                               kRGBA_F16_SkColorType, kPremul_SkAlphaType);
+
+    SkBitmap bm;
+    bm.allocPixels(info);
+    SkPixmap pm;
+    bm.peekPixels(&pm);
+    memset(pm.writable_addr(), 0, pm.getSafeSize());
+
+    if (SkColorGetA(c) == 0xFF) {
+        u64_flags |= SkXfermode::kSrcIsOpaque_PM4fFlag;
+    }
+
+    const SkXfermode::U64State state { nullptr, u64_flags };
+
+    const SkPM4f src = SkColor4f::FromColor(c).premul();
+    auto proc1 = SkXfermode::GetU64Proc1(SkXfermode::kSrcOver_Mode, u64_flags);
+    for (int y = 0; y < ir.height()/2; ++y) {
+        proc1(state, pm.writable_addr64(0, y), src, ir.width(), aa);
+    }
+    
+    SkPM4f buffer[1000];
+    for (int i = 0; i < ir.width(); ++i) {
+        buffer[i] = src;
+    }
+    auto procN = SkXfermode::GetU64ProcN(SkXfermode::kSrcOver_Mode, u64_flags);
+    for (int y = ir.height()/2 + 1; y < ir.height(); ++y) {
+        procN(state, pm.writable_addr64(0, y), buffer, ir.width(), aa);
+    }
+
+    SkAutoTUnref<SkImage> image(new_u64_image(bm, u64_flags));
+    canvas->drawImage(image, r.left(), r.top(), nullptr);
+}
+
+/*
+ *  Test SkXfer4fProcs directly for src-over, comparing them to current SkColor blits.
+ */
+DEF_SIMPLE_GM(xfer_u64_srcover, canvas, 580, 760) {
+    const int IW = 50;
+    const SkScalar W = IW;
+    const SkScalar H = 100;
+
+    const int32_t flags[] = {
+        -1,                                 // normal
+        0,                                  // U16 components
+        SkXfermode::kDstIsFloat16_U64Flag,  // F16 components
+    };
+    const SkColor colors[] = {
+        SK_ColorBLACK, SK_ColorRED, SK_ColorGREEN, SK_ColorBLUE,
+        0x88000000, 0x88FF0000, 0x8800FF00, 0x880000FF
+    };
+    
+    uint8_t aa_scanline[IW];
+    for (int i = 0; i < IW; ++i) {
+        aa_scanline[i] = i * 255 / (IW - 1);
+    }
+    uint8_t const* aa_table[] = { nullptr, aa_scanline };
+
+    SkBitmap mask;
+    mask.installPixels(SkImageInfo::MakeA8(IW, 1), aa_scanline, IW);
+
+    canvas->translate(20, 20);
+
+    const SkRect r = SkRect::MakeWH(W, H);
+    for (const uint8_t* aa : aa_table) {
+        canvas->save();
+        for (auto flag : flags) {
+            canvas->save();
+            for (SkColor c : colors) {
+                if (flag < 0) {
+                    SkPaint p;
+                    p.setColor(c);
+                    if (aa) {
+                        canvas->drawBitmapRect(mask, r, &p);
+                    } else {
+                        canvas->drawRect(r, p);
+                    }
+                } else {
+                    draw_rect(canvas, r, c, flag, aa);
+                }
+                canvas->translate(W + 20, 0);
+            }
+            canvas->restore();
+            canvas->translate(0, H + 20);
+        }
+        canvas->restore();
+        canvas->translate(0, (H + 20) * SK_ARRAY_COUNT(flags) + 20);
+    }
+}
diff --git a/gyp/core.gypi b/gyp/core.gypi
index db54377723..1700a05331 100644
--- a/gyp/core.gypi
+++ b/gyp/core.gypi
@@ -302,6 +302,7 @@
         '<(skia_src_path)/core/SkWriter32.cpp',
         '<(skia_src_path)/core/SkXfermode.cpp',
         '<(skia_src_path)/core/SkXfermode4f.cpp',
+        '<(skia_src_path)/core/SkXfermodeU64.cpp',
         '<(skia_src_path)/core/SkXfermode_proccoeff.h',
         '<(skia_src_path)/core/SkXfermodeInterpretation.cpp',
         '<(skia_src_path)/core/SkXfermodeInterpretation.h',
diff --git a/include/core/SkXfermode.h b/include/core/SkXfermode.h
index 76c57a51dd..21d77f8747 100644
--- a/include/core/SkXfermode.h
+++ b/include/core/SkXfermode.h
@@ -233,6 +233,21 @@ public:
     virtual PM4fProc1 getPM4fProc1(uint32_t flags) const;
     virtual PM4fProcN getPM4fProcN(uint32_t flags) const;
 
+    enum U64Flags {
+        kSrcIsOpaque_U64Flag  = 1 << 0,
+        kDstIsFloat16_U64Flag = 1 << 1, // else U16 bit components
+    };
+    struct U64State {
+        const SkXfermode* fXfer;
+        uint32_t          fFlags;
+    };
+    typedef void (*U64Proc1)(const U64State&, uint64_t dst[], const SkPM4f& src, int count,
+                             const SkAlpha coverage[]);
+    typedef void (*U64ProcN)(const U64State&, uint64_t dst[], const SkPM4f src[], int count,
+                             const SkAlpha coverage[]);
+    static U64Proc1 GetU64Proc1(Mode, uint32_t flags);
+    static U64ProcN GetU64ProcN(Mode, uint32_t flags);
+
 protected:
     SkXfermode() {}
     /** The default implementation of xfer32/xfer16/xferA8 in turn call this
diff --git a/src/core/SkXfermodeU64.cpp b/src/core/SkXfermodeU64.cpp
new file mode 100644
index 0000000000..6435e6a2e2
--- /dev/null
+++ b/src/core/SkXfermodeU64.cpp
@@ -0,0 +1,181 @@
+/*
+ * 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 "SkHalf.h"
+#include "SkPM4fPriv.h"
+#include "SkUtils.h"
+#include "SkXfermode.h"
+
+static void sk_memset64(uint64_t dst[], uint64_t value, int count) {
+    for (int i = 0; i < count; ++i) {
+        dst[i] = value;
+    }
+}
+
+struct U64ProcPair {
+    SkXfermode::U64Proc1 fP1;
+    SkXfermode::U64ProcN fPN;
+};
+
+enum DstType {
+    kU16_Dst,
+    kF16_Dst,
+};
+
+static Sk4f lerp_by_coverage(const Sk4f& src, const Sk4f& dst, uint8_t srcCoverage) {
+    return dst + (src - dst) * Sk4f(srcCoverage * (1/255.0f));
+}
+
+template <DstType D> Sk4f unit_to_dst_bias(const Sk4f& x4) {
+    return (D == kU16_Dst) ? x4 * Sk4f(65535) : x4;
+}
+
+// returns value already biased by 65535
+static Sk4f load_from_u16(uint64_t value) {
+    return SkNx_cast<float>(Sk4h::Load(&value));
+}
+
+// takes floats already biased by 65535
+static uint64_t store_to_u16(const Sk4f& x4) {
+    uint64_t value;
+    SkNx_cast<uint16_t>(x4 + Sk4f(0.5f)).store(&value);
+    return value;
+}
+
+static Sk4f load_from_f16(uint64_t value) {
+    const uint16_t* u16 = reinterpret_cast<const uint16_t*>(&value);
+    float f4[4];
+    for (int i = 0; i < 4; ++i) {
+        f4[i] = SkHalfToFloat(u16[i]);
+    }
+    return Sk4f::Load(f4);
+}
+
+static uint64_t store_to_f16(const Sk4f& x4) {
+    uint64_t value;
+    uint16_t* u16 = reinterpret_cast<uint16_t*>(&value);
+
+    float f4[4];
+    x4.store(f4);
+    for (int i = 0; i < 4; ++i) {
+        u16[i] = SkFloatToHalf(f4[i]);
+    }
+    return value;
+}
+
+// Returns dst in its "natural" bias (either unit-float or 16bit int)
+//
+template <DstType D> Sk4f load_from_dst(uint64_t dst) {
+    return (D == kU16_Dst) ? load_from_u16(dst) : load_from_f16(dst);
+}
+
+// Assumes x4 is already in the "natural" bias (either unit-float or 16bit int)
+template <DstType D> uint64_t store_to_dst(const Sk4f& x4) {
+    return (D == kU16_Dst) ? store_to_u16(x4) : store_to_f16(x4);
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <DstType D> void src_1(const SkXfermode::U64State& state, uint64_t dst[],
+                                const SkPM4f& src, int count, const SkAlpha aa[]) {
+    const Sk4f s4 = unit_to_dst_bias<D>(Sk4f::Load(src.fVec));
+    if (aa) {
+        for (int i = 0; i < count; ++i) {
+            const Sk4f d4 = load_from_dst<D>(dst[i]);
+            dst[i] = store_to_dst<D>(lerp_by_coverage(s4, d4, aa[i]));
+        }
+    } else {
+        sk_memset64(dst, store_to_dst<D>(s4), count);
+    }
+}
+
+template <DstType D> void src_n(const SkXfermode::U64State& state, uint64_t dst[],
+                                const SkPM4f src[], int count, const SkAlpha aa[]) {
+    if (aa) {
+        for (int i = 0; i < count; ++i) {
+            const Sk4f s4 = unit_to_dst_bias<D>(Sk4f::Load(src[i].fVec));
+            const Sk4f d4 = load_from_dst<D>(dst[i]);
+            dst[i] = store_to_dst<D>(lerp_by_coverage(s4, d4, aa[i]));
+        }
+    } else {
+        for (int i = 0; i < count; ++i) {
+            const Sk4f s4 = unit_to_dst_bias<D>(Sk4f::Load(src[i].fVec));
+            dst[i] = store_to_dst<D>(s4);
+        }
+    }
+}
+
+const U64ProcPair gU64Procs_Src[] = {
+    { src_1<kU16_Dst>, src_n<kU16_Dst>  },   // U16     alpha
+    { src_1<kU16_Dst>, src_n<kU16_Dst>  },   // U16     opaque
+    { src_1<kF16_Dst>, src_n<kF16_Dst>  },   // F16     alpha
+    { src_1<kF16_Dst>, src_n<kF16_Dst>  },   // F16     opaque
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <DstType D> void srcover_1(const SkXfermode::U64State& state, uint64_t dst[],
+                                    const SkPM4f& src, int count, const SkAlpha aa[]) {
+    const Sk4f s4 = Sk4f::Load(src.fVec);
+    const Sk4f dst_scale = Sk4f(1 - get_alpha(s4));
+    const Sk4f s4bias = unit_to_dst_bias<D>(s4);
+    for (int i = 0; i < count; ++i) {
+        const Sk4f d4bias = load_from_dst<D>(dst[i]);
+        const Sk4f r4bias = s4bias + d4bias * dst_scale;
+        if (aa) {
+            dst[i] = store_to_dst<D>(lerp_by_coverage(r4bias, d4bias, aa[i]));
+        } else {
+            dst[i] = store_to_dst<D>(r4bias);
+        }
+    }
+}
+
+template <DstType D> void srcover_n(const SkXfermode::U64State& state, uint64_t dst[],
+                                    const SkPM4f src[], int count, const SkAlpha aa[]) {
+    for (int i = 0; i < count; ++i) {
+        const Sk4f s4 = Sk4f::Load(src[i].fVec);
+        const Sk4f dst_scale = Sk4f(1 - get_alpha(s4));
+        const Sk4f s4bias = unit_to_dst_bias<D>(s4);
+        const Sk4f d4bias = load_from_dst<D>(dst[i]);
+        const Sk4f r4bias = s4bias + d4bias * dst_scale;
+        if (aa) {
+            dst[i] = store_to_dst<D>(lerp_by_coverage(r4bias, d4bias, aa[i]));
+        } else {
+            dst[i] = store_to_dst<D>(r4bias);
+        }
+    }
+}
+
+const U64ProcPair gU64Procs_SrcOver[] = {
+    { srcover_1<kU16_Dst>,  srcover_n<kU16_Dst> },   // U16     alpha
+    { src_1<kU16_Dst>,      src_n<kU16_Dst>     },   // U16     opaque
+    { srcover_1<kF16_Dst>,  srcover_n<kF16_Dst> },   // F16     alpha
+    { src_1<kF16_Dst>,      src_n<kF16_Dst>     },   // F16     opaque
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static U64ProcPair find_procs(SkXfermode::Mode mode, uint32_t flags) {
+    SkASSERT(0 == (flags & ~3));
+    flags &= 3;
+
+    switch (mode) {
+        case SkXfermode::kSrc_Mode:     return gU64Procs_Src[flags];
+        case SkXfermode::kSrcOver_Mode: return gU64Procs_SrcOver[flags];
+        default:
+            break;
+    }
+    return { nullptr, nullptr };
+}
+
+SkXfermode::U64Proc1 SkXfermode::GetU64Proc1(Mode mode, uint32_t flags) {
+    return find_procs(mode, flags).fP1;
+}
+
+SkXfermode::U64ProcN SkXfermode::GetU64ProcN(Mode mode, uint32_t flags) {
+    return find_procs(mode, flags).fPN;
+}