Make SkPMFloats store floats in [0,255] instead of [0,1].

src/opts/SkPMFloat_SSE2.h

Index: src/opts/SkPMFloat_SSE2.h
diff --git a/src/opts/SkPMFloat_SSE2.h b/src/opts/SkPMFloat_SSE2.h
index c3dbbf2cb50d655fb77d840a7ace132f5cec7cb7..f28608079f3a4ac9a41a6c4bcc5c941261c4dbf9 100644
--- a/src/opts/SkPMFloat_SSE2.h
+++ b/src/opts/SkPMFloat_SSE2.h
@@ -2,11 +2,10 @@
 #include <emmintrin.h>
 
 // For set(), we widen our 8 bit components (fix8) to 8-bit components in 16 bits (fix8_16),
-// then widen those to 8-bit-in-32-bits (fix8_32), convert those to floats (scaled),
-// then finally scale those down from [0.0f, 255.0f] to [0.0f, 1.0f] into fColor.
+// then widen those to 8-bit-in-32-bits (fix8_32), and finally convert those to floats.
 
-// get() and clamped() do the opposite, working from [0.0f, 1.0f] floats to [0.0f, 255.0f],
-// to 8-bit-in-32-bit, to 8-bit-in-16-bit, back down to 8-bit components.
+// get() and clamped() do the opposite, working from floats to 8-bit-in-32-bit,
+// to 8-bit-in-16-bit, back down to 8-bit components.
 // _mm_packus_epi16() gives us clamping for free while narrowing.
 
 inline void SkPMFloat::set(SkPMColor c) {
@@ -14,8 +13,7 @@ inline void SkPMFloat::set(SkPMColor c) {
     __m128i fix8    = _mm_set_epi32(0,0,0,c),
             fix8_16 = _mm_unpacklo_epi8 (fix8,    _mm_setzero_si128()),
             fix8_32 = _mm_unpacklo_epi16(fix8_16, _mm_setzero_si128());
-    __m128  scaled  = _mm_cvtepi32_ps(fix8_32);
-    _mm_store_ps(fColor, _mm_mul_ps(scaled, _mm_set1_ps(1.0f/255.0f)));
+    _mm_store_ps(fColor, _mm_cvtepi32_ps(fix8_32));

[msarett, 2015/03/03 14:34:28]

I think we might be able to improve performance a little bit if we convert
directly from 8-bit or 16-bit ints to floats (not 100% sure because I'm having
trouble finding documentation on the latency and throughput of the applicable
instructions).

Two instructions that might be useful:
__m128 _mm_cvtpi8_ps (__m64 a) // Converts 8-bit signed integers in lower half
of vector to floats
__m128 _mm_cvtpi16_ps (__m64 a) // Converts 16-bit signed integers (there is
also an unsigned version) to floats

The 16-bit version (either signed or unsigned) would certainly be workable and
would save the 16-bit to 32-bit step.  The 8-bit version is more interesting
because it would be faster, but the issue is that there is only a signed
version, and we need unsigned.  We might write something like this:
__m64 fix8 = _mm_set_pi32(0, c);
__m128i fix8_float = _mm_cvtpi8_ps(fix8);
_mm_store_ps(fColor, fix8_float);

The issue would be that input color components in the range [128, 255] would be
converted to the range [-128, -1].  I think we could trick our way around this
by adding 128 before and after the conversion (I can draw out how this might
work).  It's not clear if this would be faster than a 16-bit version or not but
it might be worth trying.

[mtklein, 2015/03/03 15:02:17]

The reason I've shied away from intrinsics like _mm_cvtpi8_ps is that they don't
compile to single instructions.
That'll be why you can't find latency and throughput numbers on them: they're
implemented as compound operations
that may vary from compiler to compiler.  Looking at my xmmintrin.h from GCC and
Clang, they do seem to vary, and they're quite a lot more work than what we've
got here.

It's certainly worth a try locally to see what code's generated and how it
performs.  I just tried myself with this code
   _mm_store_ps(fColor, _mm_cvtpu8_ps(_mm_set_pi32(0, c)));
and it ran about 4x slower (at about 4x the instructions) than my current best.

(As far as I can tell, _mm_cvtpu8_ps is what you'd want for unsigned 8 -> float
conversions.  But again, it's slow.)

As far as int->float conversions go, I really only see two options: cvtpi2ps
(_mm_cvtpi32_ps) to convert 2 at a time from SSE, or cvtepi2ps (_mm_cvtepi32_ps)
to convert 4 at a time from SSE2.  (AVX lets us go 8 at a time of course, but
that's not really realistic to target right now.)

     SkASSERT(this->isValid());

[msarett, 2015/03/03 14:34:28]

I'm starting another comment for another train of thought.  No matter what
strategy we choose, we are going to have to pay a price for the conversion from
ints to floats (latency = 3 cycles for _mm_cvtepi32_ps).  But some of this can
be offset by the high throughput of this instruction (throughput = 1 cycle for
_mm_cvtepi32_ps).  However, because we are only converting one pixel at a time
with each call to this function, it is likely that the latency is being paid
every time and that we are not taking advantage of the throughput.  There is a
chance that if we were to convert an array of pixels in a single function call
and use a partly unrolled loop, we could see a performance improvement.

[mtklein, 2015/03/03 15:02:17]

Yep, totally agree.  We're thinking the next logical place to go is to do
conversions of 4 pixels at a time, which also lets us do 128-bit reads and
writes of SkPMColors, and lets us get more work out of the _mm_pack
instructions.

 }
 
@@ -25,8 +23,7 @@ inline SkPMColor SkPMFloat::get() const {
 }
 
 inline SkPMColor SkPMFloat::clamped() const {
-    __m128  scaled  = _mm_mul_ps(_mm_load_ps(fColor), _mm_set1_ps(255.0f));
-    __m128i fix8_32 = _mm_cvtps_epi32(scaled),
+    __m128i fix8_32 = _mm_cvtps_epi32(_mm_load_ps(fColor)),
             fix8_16 = _mm_packus_epi16(fix8_32, fix8_32),
             fix8    = _mm_packus_epi16(fix8_16, fix8_16);
     SkPMColor c = _mm_cvtsi128_si32(fix8);