float components in xfermodes

src/core/SkXfermode4f.cpp

Index: src/core/SkXfermode4f.cpp
diff --git a/src/core/SkXfermode4f.cpp b/src/core/SkXfermode4f.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..0485a5e6edd4c0c65f884eae88b27beb8c9cd637
--- /dev/null
+++ b/src/core/SkXfermode4f.cpp
@@ -0,0 +1,327 @@
+/*
+ * 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 "SkPM4fPriv.h"
+#include "SkUtils.h"
+#include "SkXfermode.h"
+
+struct XferProcPair {
+    SkXfermode::PM4fProc1 fP1;
+    SkXfermode::PM4fProcN fPN;
+};
+
+enum DstType {
+    kLinear_Dst,
+    kSRGB_Dst,
+};
+
+static Sk4f scale_by_coverage(const Sk4f& x4, uint8_t coverage) {
+    return x4 * Sk4f(coverage * (1/255.0f));
+}
+
+static Sk4f lerp(const Sk4f& src, const Sk4f& dst, uint8_t srcCoverage) {
+    return dst + (src - dst) * Sk4f(srcCoverage * (1/255.0f));
+}
+
+template <DstType D> Sk4f load_dst(SkPMColor dstC) {
+    return (D == kSRGB_Dst) ? Sk4f_fromS32(dstC) : Sk4f_fromL32(dstC);
+}
+
+template <DstType D> uint32_t store_dst(const Sk4f& x4) {
+    return (D == kSRGB_Dst) ? Sk4f_toS32(x4) : Sk4f_toL32(x4);
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static Sk4f scale_255_round(const SkPM4f& pm4) {

[mtklein, 2016/02/01 22:53:52]

Generally I like to scope one-shot helpers like this as lambdas now, here
nestled into pm4f_to_linear_32.  Seeing it here makes me think i'll find it used
more later on down the file.

[reed1, 2016/02/02 14:20:54]

We can change. I find some value in naming these at times, as that documents the
(intent of) the transformation, something that is lost in a lambda. Perhaps the
need (or lack thereof) for micro documentation will change as we get more mature
in how we organize this new workflow.

[mtklein, 2016/02/02 14:24:10]

Lambdas can have names too...

auto scale_255_round = [](const SkPM4f& pm4) {
    return Sk4f::Load(pm4.fVec) * Sk4f(255) + Sk4f(0.5f);
};

Just suggesting to scope this tighter while it's only used in that tight scope.

+    return Sk4f::Load(pm4.fVec) * Sk4f(255) + Sk4f(0.5f);
+}
+
+static void pm4f_to_linear_32(SkPMColor dst[], const SkPM4f src[], int count) {
+    while (count >= 4) {
+        src[0].assertIsUnit();
+        src[1].assertIsUnit();
+        src[2].assertIsUnit();
+        src[3].assertIsUnit();
+        Sk4f_ToBytes((uint8_t*)dst,
+                     scale_255_round(src[0]), scale_255_round(src[1]),
+                     scale_255_round(src[2]), scale_255_round(src[3]));
+        src += 4;
+        dst += 4;
+        count -= 4;
+    }
+    for (int i = 0; i < count; ++i) {
+        src[i].assertIsUnit();
+        SkNx_cast<uint8_t>(scale_255_round(src[i])).store((uint8_t*)&dst[i]);
+    }
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// These are our fallback impl for the SkPM4f procs...
+//
+// They just convert the src color(s) into a linear SkPMColor value(s), and then
+// call the existing virtual xfer32. This clear throws away data (converting floats to bytes)
+// in the src, and ignores the sRGB flag, but should draw about the same as if the caller
+// had passed in SkPMColor values directly.
+//
+
+void xfer_pm4_proc_1(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f& src,
+                     int count, const SkAlpha aa[]) {
+    uint32_t pm;
+    pm4f_to_linear_32(&pm, &src, 1);
+
+    const int N = 128;
+    SkPMColor tmp[N];

[mtklein, 2016/02/01 22:53:52]

yeesh.  _1 variants are a really good idea.  We might want to backport the idea.

+    sk_memset32(tmp, pm, SkMin32(count, N));
+    while (count > 0) {
+        const int n = SkMin32(count, N);
+        state.fXfer->xfer32(dst, tmp, n, aa);
+
+        dst += n;
+        if (aa) {
+            aa += n;
+        }
+        count -= n;
+    }
+}
+
+void xfer_pm4_proc_n(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f src[],
+                     int count, const SkAlpha aa[]) {
+    const int N = 128;
+    SkPMColor tmp[N];
+    while (count > 0) {
+        const int n = SkMin32(count, N);
+        pm4f_to_linear_32(tmp, src, n);
+        state.fXfer->xfer32(dst, tmp, n, aa);
+
+        src += n;
+        dst += n;
+        if (aa) {
+            aa += n;
+        }
+        count -= n;
+    }
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static void clear_linear_n(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f[],
+                           int count, const SkAlpha aa[]) {
+    if (aa) {
+        for (int i = 0; i < count; ++i) {
+            unsigned a = aa[i];
+            if (a) {
+                SkPMColor dstC = dst[i];
+                SkPMColor C = 0;
+                if (0xFF != a) {
+                    C = SkFourByteInterp(C, dstC, a);
+                }
+                dst[i] = C;
+            }
+        }
+    } else {
+        sk_bzero(dst, count * sizeof(SkPMColor));
+    }
+}
+
+static void clear_linear_1(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f&,
+                           int count, const SkAlpha coverage[]) {
+    clear_linear_n(state, dst, nullptr, count, coverage);

[mtklein, 2016/02/01 22:53:52]

We might want to make the _1 and _n variants take the same arguments (i.e. const
SkPM4f*).  It won't matter for speed, and it'll let us write some things like
this a bit more compactly.  We'll only have to deal with 1 proc type too.

[reed1, 2016/02/02 14:20:54]

I certainly pondered this. Can't tell if that helps readability at the call
sites or not. As it is, the caller is definitely aware that they are only giving
us 1 src. I think I prefer to error on the side of pedanticness for the moment,
and reconsider that once things are more fleshed-out. That said, I don't feel
terribly strongly either way.

+}
+
+static void clear_srgb_n(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f[],
+                           int count, const SkAlpha aa[]) {
+    if (aa) {
+        for (int i = 0; i < count; ++i) {
+            unsigned a = aa[i];
+            if (a) {
+                Sk4f d = Sk4f_fromS32(dst[i]) * Sk4f((255 - a) * (1/255.0f));
+                dst[i] = Sk4f_toS32(d);
+            }
+        }
+    } else {
+        sk_bzero(dst, count * sizeof(SkPMColor));
+    }
+}
+
+static void clear_srgb_1(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f&,
+                           int count, const SkAlpha coverage[]) {
+    clear_srgb_n(state, dst, nullptr, count, coverage);
+}
+
+const XferProcPair gProcs_Clear[] = {
+    { clear_linear_1, clear_linear_n },       // linear   [alpha]
+    { clear_linear_1, clear_linear_n },       // linear   [opaque]
+    { clear_srgb_1,   clear_srgb_n   },       // srgb     [alpha]
+    { clear_srgb_1,   clear_srgb_n   },       // srgb     [opaque]
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <DstType D> void src_n(const SkXfermode::PM4fState& state, uint32_t dst[],
+                                const SkPM4f src[], int count, const SkAlpha aa[]) {
+    for (int i = 0; i < count; ++i) {
+        unsigned a = 0xFF;
+        if (aa) {
+            a = aa[i];
+            if (0 == a) {
+                continue;
+            }
+        }
+        Sk4f r4 = Sk4f::Load(src[i].fVec);   // src always overrides dst
+        if (a != 0xFF) {
+            Sk4f d4 = load_dst<D>(dst[i]);
+            r4 = lerp(r4, d4, a);
+        }
+        dst[i] = store_dst<D>(r4);
+    }
+}
+
+template <DstType D> void src_1(const SkXfermode::PM4fState& state, uint32_t dst[],
+                                const SkPM4f& src, int count, const SkAlpha aa[]) {
+    const Sk4f r4 = Sk4f::Load(src.fVec);   // src always overrides dst
+    const uint32_t r32 = store_dst<D>(r4);
+
+    if (aa) {
+        for (int i = 0; i < count; ++i) {
+            unsigned a = aa[i];
+            if (0 == a) {
+                continue;
+            }
+            if (a != 0xFF) {
+                Sk4f d4 = load_dst<D>(dst[i]);
+                dst[i] = store_dst<D>(lerp(r4, d4, a));
+            } else {
+                dst[i] = r32;
+            }
+        }
+    } else {
+        sk_memset32(dst, r32, count);
+    }
+}
+
+const XferProcPair gProcs_Src[] = {
+    { src_1<kLinear_Dst>, src_n<kLinear_Dst> },       // linear   [alpha]
+    { src_1<kLinear_Dst>, src_n<kLinear_Dst> },       // linear   [opaque]
+    { src_1<kSRGB_Dst>,   src_n<kSRGB_Dst>   },       // srgb     [alpha]
+    { src_1<kSRGB_Dst>,   src_n<kSRGB_Dst>   },       // srgb     [opaque]
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static void dst_n(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f[],
+                         int count, const SkAlpha aa[]) {}
+
+static void dst_1(const SkXfermode::PM4fState& state, uint32_t dst[], const SkPM4f&,
+                  int count, const SkAlpha coverage[]) {}
+
+const XferProcPair gProcs_Dst[] = {
+    { dst_1, dst_n },
+    { dst_1, dst_n },
+    { dst_1, dst_n },
+    { dst_1, dst_n },
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <DstType D> void srcover_n(const SkXfermode::PM4fState& state, uint32_t dst[],
+                                    const SkPM4f src[], int count, const SkAlpha aa[]) {
+    if (aa) {
+        for (int i = 0; i < count; ++i) {
+            unsigned a = aa[i];
+            if (0 == a) {
+                continue;
+            }
+            Sk4f s4 = Sk4f::Load(src[i].fVec);
+            Sk4f d4 = load_dst<D>(dst[i]);
+            if (a != 0xFF) {
+                s4 = scale_by_coverage(s4, a);
+            }
+            Sk4f r4 = s4 + d4 * Sk4f(1 - get_alpha(s4));
+            dst[i] = store_dst<D>(r4);
+        }
+    } else {
+        for (int i = 0; i < count; ++i) {
+            Sk4f s4 = Sk4f::Load(src[i].fVec);
+            Sk4f d4 = load_dst<D>(dst[i]);
+            Sk4f r4 = s4 + d4 * Sk4f(1 - get_alpha(s4));
+            dst[i] = store_dst<D>(r4);
+        }
+    }
+}
+
+template <DstType D> void srcover_1(const SkXfermode::PM4fState& state, uint32_t dst[],
+                                    const SkPM4f& src, int count, const SkAlpha aa[]) {
+    Sk4f s4 = Sk4f::Load(src.fVec);
+    Sk4f scale = Sk4f(1 - get_alpha(s4));
+
+    if (aa) {
+        for (int i = 0; i < count; ++i) {
+            unsigned a = aa[i];
+            if (0 == a) {
+                continue;
+            }
+            Sk4f d4 = load_dst<D>(dst[i]);
+            Sk4f r4;
+            if (a != 0xFF) {
+                s4 = scale_by_coverage(s4, a);
+                r4 = s4 + d4 * Sk4f(1 - get_alpha(s4));
+            } else {
+                r4 = s4 + d4 * scale;
+            }
+            dst[i] = store_dst<D>(r4);
+        }
+    } else {
+        for (int i = 0; i < count; ++i) {
+            Sk4f d4 = load_dst<D>(dst[i]);
+            Sk4f r4 = s4 + d4 * scale;
+            dst[i] = store_dst<D>(r4);
+        }
+    }
+}
+
+const XferProcPair gProcs_SrcOver[] = {
+    { srcover_1<kLinear_Dst>,   srcover_n<kLinear_Dst> },   // linear   alpha
+    { src_1<kLinear_Dst>,       src_n<kLinear_Dst>     },   // linear   opaque [ we are src-mode ]
+    { srcover_1<kSRGB_Dst>,     srcover_n<kSRGB_Dst>   },   // srgb     alpha
+    { src_1<kSRGB_Dst>,         src_n<kSRGB_Dst>       },   // srgb     opaque [ we are src-mode ]
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static XferProcPair find_procs(SkXfermode::Mode mode, uint32_t flags) {
+    SkASSERT(0 == (flags & ~3));
+    flags &= 3;
+
+    switch (mode) {
+        case SkXfermode::kClear_Mode:   return gProcs_Clear[flags];
+        case SkXfermode::kSrc_Mode:     return gProcs_Src[flags];
+        case SkXfermode::kDst_Mode:     return gProcs_Dst[flags];
+        case SkXfermode::kSrcOver_Mode: return gProcs_SrcOver[flags];
+        default:
+            break;
+    }
+    return { xfer_pm4_proc_1, xfer_pm4_proc_n };
+}
+
+SkXfermode::PM4fProc1 SkXfermode::GetPM4fProc1(Mode mode, uint32_t flags) {
+    return find_procs(mode, flags).fP1;
+}
+
+SkXfermode::PM4fProcN SkXfermode::GetPM4fProcN(Mode mode, uint32_t flags) {
+    return find_procs(mode, flags).fPN;
+}
+
+SkXfermode::PM4fProc1 SkXfermode::getPM4fProc1(uint32_t flags) const {
+    Mode mode;
+    return this->asMode(&mode) ? GetPM4fProc1(mode, flags) : xfer_pm4_proc_1;
+}
+
+SkXfermode::PM4fProcN SkXfermode::getPM4fProcN(uint32_t flags) const {
+    Mode mode;
+    return this->asMode(&mode) ? GetPM4fProcN(mode, flags) : xfer_pm4_proc_n;
+}