Add SSSE3 implementation for SkPMFloat, with faster get() and set().

src/opts/SkPMFloat_SSSE3.h

Index: src/opts/SkPMFloat_SSSE3.h
diff --git a/src/opts/SkPMFloat_SSSE3.h b/src/opts/SkPMFloat_SSSE3.h
new file mode 100644
index 0000000000000000000000000000000000000000..fb5685c04547d618b6d23dffa07088dadbf084bc
--- /dev/null
+++ b/src/opts/SkPMFloat_SSSE3.h
@@ -0,0 +1,38 @@
+#include "SkColorPriv.h"
+#include <tmmintrin.h>
+
+// For set(), we widen our 8 bit components (fix8) to 8-bit components in 32 bits (fix8_32),
+// then convert those to floats.
+
+// get() does the opposite, working from floats to 8-bit-in-32-bits, then back to packed 8 bit.
+
+// clamped() is the same as _SSE2: floats to 8-in-32, to 8-in-16, to packed 8 bit, with
+// _mm_packus_epi16() both clamping and narrowing.
+
+inline void SkPMFloat::set(SkPMColor c) {
+    SkPMColorAssert(c);
+    const int _ = 255;  // _ means to zero that byte.
+    __m128i fix8    = _mm_set_epi32(0,0,0,c),
+            fix8_32 = _mm_shuffle_epi8(fix8, _mm_set_epi8(_,_,_,3, _,_,_,2, _,_,_,1, _,_,_,0));
+    _mm_store_ps(fColor, _mm_cvtepi32_ps(fix8_32));

[msarett, 2015/03/04 18:06:12]

As far as instructions, it looks to me like you have taken advantage of the best
ones.

I'm pretty sure that my main thoughts for improvements are the same things you
are already thinking.

Ideally, the code would be structured so that the convert hardware is always
busy (and the speed of the convert is the limiting factor).  If we pass in a
whole array of SkPMColor, setting the constant for the shuffle would be a one
time overhead.  We could perform four shuffles and four converts for every one
vector load.

If we unroll the loop appropriately to limit dependencies it is possible that
the load, shuffle, and convert could be performed in parallel (though I haven't
been able to confirm that the shuffle and convert use different hardware).  This
might get us closer to the optimum of only being constrained by the convert
step.

[mtklein, 2015/03/04 18:10:24]

+1, though keep in mind this code is inlined.   Setting the constant for the
shuffles _is_ a one time overhead if we call SkPMFloat::set() in a loop.  This
code is inlined and those non-loop-dependent ops are hoisted up.

+    SkASSERT(this->isValid());
+}
+
+inline SkPMColor SkPMFloat::get() const {
+    SkASSERT(this->isValid());
+    const int _ = 255;  // _ means to zero that byte.
+    __m128i fix8_32 = _mm_cvtps_epi32(_mm_load_ps(fColor)),  // _mm_cvtps_epi32 rounds for us!
+            fix8    = _mm_shuffle_epi8(fix8_32, _mm_set_epi8(_,_,_,_, _,_,_,_, _,_,_,_, 12,8,4,0));
+    SkPMColor c = _mm_cvtsi128_si32(fix8);
+    SkPMColorAssert(c);
+    return c;
+}
+
+inline SkPMColor SkPMFloat::clamped() const {
+    __m128i fix8_32 = _mm_cvtps_epi32(_mm_load_ps(fColor)),  // _mm_cvtps_epi32 rounds for us!
+            fix8_16 = _mm_packus_epi16(fix8_32, fix8_32),

[msarett, 2015/03/04 18:06:12]

For lack of anything to comment on, it might be more readable to use
_mm_packus_epi32 here.  But as far as I can tell, the behavior and performance
will be identical.

[mtklein, 2015/03/04 18:10:24]

Yeah.  I'd have done that, just for the readability, but sadly _mm_packus_epi32
is SSE 4.1.  We've only got _mm_packus_16 to play with until then.

+            fix8    = _mm_packus_epi16(fix8_16, fix8_16);
+    SkPMColor c = _mm_cvtsi128_si32(fix8);
+    SkPMColorAssert(c);
+    return c;
+}