starter procs for blending with pm4f

src/core/SkPM4fPriv.h

Index: src/core/SkPM4fPriv.h
diff --git a/src/core/SkPM4fPriv.h b/src/core/SkPM4fPriv.h
new file mode 100644
index 0000000000000000000000000000000000000000..a4cec5c449ac89e0bef824c54a8f552c9089e5c6
--- /dev/null
+++ b/src/core/SkPM4fPriv.h
@@ -0,0 +1,121 @@
+/*
+ * 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 "SkColorPriv.h"
+#include "SkNx.h"
+
+static inline float get_alpha(const Sk4f& f4) {
+    return f4.kth<SkPM4f::A>();
+}
+
+static inline Sk4f set_alpha(const Sk4f& f4, float alpha) {
+    static_assert(3 == SkPM4f::A, "");
+    return Sk4f(f4.kth<0>(), f4.kth<1>(), f4.kth<2>(), alpha);
+}
+
+static inline uint32_t to_4b(const Sk4f& f4) {
+    uint32_t b4;
+    SkNx_cast<uint8_t>(f4).store((uint8_t*)&b4);
+    return b4;
+}
+
+static inline Sk4f to_4f(uint32_t b4) {
+    return SkNx_cast<float>(Sk4b::Load((const uint8_t*)&b4));
+}
+
+static inline Sk4f s2l(const Sk4f& s4) {

[mtklein, 2016/01/29 15:23:00]

Had to think about what these were for a second.  Might wanna expand some of
these names out: sRGB_to_linear(), linear_to_sRGB().

[reed1, 2016/01/29 22:35:27]

Done.

+    return set_alpha(s4 * s4, get_alpha(s4));
+}
+
+static inline Sk4f l2s(const Sk4f& l4) {
+    return set_alpha(l4.rsqrt1() * l4, get_alpha(l4));
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static inline Sk4f Sk4f_fromL32(uint32_t src) {
+    return to_4f(src) * Sk4f(1.0f/255);
+}
+
+static inline Sk4f Sk4f_fromS32(uint32_t src) {
+    return s2l(to_4f(src) * Sk4f(1.0f/255));
+}
+
+static inline uint32_t Sk4f_toL32(const Sk4f& x4) {
+    return to_4b(x4 * Sk4f(255) + Sk4f(0.5f));
+}
+
+static inline uint32_t Sk4f_toS32(const Sk4f& x4) {
+    return to_4b(l2s(x4) * Sk4f(255) + Sk4f(0.5f));
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static Sk4f unit_to_l255_round(const SkPM4f& pm4) {

[mtklein, 2016/01/29 15:23:00]

I think this might be clearer if we start first with completely orthogonal
helper functions, e.g.

    Sk4f to_4f(uint32_t);
    uint32_t to_4b(Sk4f f255);

    Sk4f scale_255_to_1(Sk4f f255);
    Sk4f scale_1_to_255(Sk4f f1);    

    Sk4f to_sRGB(Sk4f f1);
    Sk4f to_linear(Sk4f f1);

then build the compound operations (sRGB 8888 -> linear unit float) out of those
exclusively, either as more static helper functions after a big slashline or as
little local lambdas.
   

[reed1, 2016/01/29 22:35:27]

Agree, this is definitely a worktable of experimental helpers. Will work to
refine/reduce over time.

+    return Sk4f::Load(pm4.fVec) * Sk4f(255) + Sk4f(0.5f);
+}
+
+static Sk4f unit_to_s255_round(const SkPM4f& pm4) {
+    return l2s(Sk4f::Load(pm4.fVec)) * Sk4f(255) + Sk4f(0.5f);
+}
+
+static inline void SkPM4f_l32_src_mode(SkPMColor dst[], const SkPM4f src[], int count) {
+    for (int i = 0; i < (count >> 2); ++i) {

[mtklein, 2016/01/29 15:23:00]

This looks right to me, but in my experience I've found a loop skeleton that
ticks `count` down is more glanceably correct when different strides are
involved:

while (count >= 4) {
   ...
   src += 4;
   dst += 4;
   count -= 4;
}
while (count >= 1) {
   ...
   src++;
   dst++;
   count--;
}

Usually these loops don't require us to name or think about an `i` either, which
is nice when it's not otherwise useful. 

Neither style measurably affects performance.

+        SkASSERT(src[0].isUnit());
+        SkASSERT(src[1].isUnit());
+        SkASSERT(src[2].isUnit());
+        SkASSERT(src[3].isUnit());
+        Sk4f_ToBytes((uint8_t*)dst,
+                     unit_to_l255_round(src[0]), unit_to_l255_round(src[1]),
+                     unit_to_l255_round(src[2]), unit_to_l255_round(src[3]));
+        src += 4;
+        dst += 4;
+    }
+    count &= 3;
+    for (int i = 0; i < count; ++i) {
+        SkASSERT(src[i].isUnit());
+        SkNx_cast<uint8_t>(unit_to_l255_round(src[i])).store((uint8_t*)&dst[i]);
+    }
+}
+
+static inline void SkPM4f_l32_srcover_mode(SkPMColor dst[], const SkPM4f src[], int count) {
+    for (int i = 0; i < count; ++i) {
+        SkASSERT(src[i].isUnit());
+        Sk4f s4 = Sk4f::Load(src[i].fVec);
+        Sk4f d4 = Sk4f_fromL32(dst[i]);
+        dst[i] = Sk4f_toL32(s4 + d4 * Sk4f(1 - get_alpha(s4)));
+    }
+}
+
+static inline void SkPM4f_s32_src_mode(SkPMColor dst[], const SkPM4f src[], int count) {

[mtklein, 2016/01/29 15:23:00]

I think we're going to have to come up with a new abbreviation for sRGB 8888, or
not abbreviate it.  I keep reading this as "uses 4 floats... 32-bit source...
src mode".  "s32" screams 32-bit source to me.

Let's just write out _sRGB_ and _linear_ ?

We could probably name these 
   SkPM4f_src_sRGB
   SkPM4f_src_linear
   SkPM4f_srcover_sRGB
   SkPM4f_srcover_linear
without losing any information.

[reed1, 2016/01/29 22:35:27]

Done.

+    for (int i = 0; i < (count >> 2); ++i) {
+        SkASSERT(src[0].isUnit());
+        SkASSERT(src[1].isUnit());
+        SkASSERT(src[2].isUnit());
+        SkASSERT(src[3].isUnit());
+        Sk4f_ToBytes((uint8_t*)dst,
+                     unit_to_s255_round(src[0]), unit_to_s255_round(src[1]),
+                     unit_to_s255_round(src[2]), unit_to_s255_round(src[3]));
+        src += 4;
+        dst += 4;
+    }
+    count &= 3;
+    for (int i = 0; i < count; ++i) {
+        SkASSERT(src[i].isUnit());
+        SkNx_cast<uint8_t>(unit_to_s255_round(src[i])).store((uint8_t*)&dst[i]);
+    }
+}
+
+static inline void SkPM4f_s32_srcover_mode(SkPMColor dst[], const SkPM4f src[], int count) {
+    for (int i = 0; i < count; ++i) {
+        SkASSERT(src[i].isUnit());
+        Sk4f s4 = Sk4f::Load(src[i].fVec);
+        Sk4f d4 = Sk4f_fromS32(dst[i]);
+        dst[i] = Sk4f_toS32(s4 + d4 * Sk4f(1 - get_alpha(s4)));
+    }
+}
+