[babl/sse-conversions-2013] SSE Float: draft 4

[Daniel Sabo <daniels src gnome org>]

commit f222c8699d057c154cb8da001f9fa521280e58a7
Author: Daniel Sabo <DanielSabo gmail com>
Date:   Mon Apr 1 20:15:56 2013 -0700

    SSE Float: draft 4
    
    Don't divide by zero, more Bably handling of negative alpha.

 extensions/sse-float.c |   69 ++++++++++++++++++++++++++++-------------------
 1 files changed, 41 insertions(+), 28 deletions(-)
---
diff --git a/extensions/sse-float.c b/extensions/sse-float.c
index bc3c42e..96695d4 100644
--- a/extensions/sse-float.c
+++ b/extensions/sse-float.c
@@ -29,10 +29,13 @@
 
 #include "babl.h"
 #include "babl-cpuaccel.h"
+#include "base/util.h"
 #include "extensions/util.h"
 
 #define Q(a) { a, a, a, a }
 
+static const float BABL_ALPHA_THRESHOLD_FLOAT = (float)BABL_ALPHA_THRESHOLD;
+
 static long
 conv_rgbaF_linear_rgbAF_linear (const float *src, float *dst, long samples)
 {
@@ -99,33 +102,39 @@ conv_rgbAF_linear_rgbaF_linear (const float *src, float *dst, long samples)
 
   if (((uintptr_t)src % 16) + ((uintptr_t)dst % 16) == 0)
     {
-      const long    n = (samples / 2) * 2;
+      const long    n = samples;
       const __v4sf *s = (const __v4sf*) src;
             __v4sf *d = (__v4sf*)dst;
 
-      for ( ; i < n; i += 2)
+      for ( ; i < n; i += 1)
         {
-          __v4sf rbaa0, rbaa1;
+          __v4sf pre_rgba0, rgba0, rbaa0, raaaa0;
           
-          __v4sf rgba0 = *s++;
-          __v4sf rgba1 = *s++;
-
-          /* Expand alpha */
-          __v4sf aaaa0 = (__v4sf)_mm_shuffle_epi32((__m128i)rgba0, _MM_SHUFFLE(3, 3, 3, 3));
-          __v4sf aaaa1 = (__v4sf)_mm_shuffle_epi32((__m128i)rgba1, _MM_SHUFFLE(3, 3, 3, 3));
-          
-          /* Premultiply */
-          rgba0 = rgba0 / aaaa0;
-          rgba1 = rgba1 / aaaa1;
+          float alpha0 = ((float *)s)[3];
+          pre_rgba0 = *s;
           
+          if (alpha0 <= 0.0f)
+          {
+            /* Zero RGB */
+            rgba0 = _mm_setzero_ps();
+          }
+          else
+          {
+            float recip0 = 1.0f/alpha0;
+            
+            /* Expand reciprocal */
+            raaaa0 = _mm_load1_ps(&recip0);
+            
+            /* Un-Premultiply */
+            rgba0 = pre_rgba0 * raaaa0;
+          }
+            
           /* Shuffle the original alpha value back in */
-          rbaa0 = _mm_shuffle_ps(rgba0, aaaa0, _MM_SHUFFLE(0, 0, 2, 0));
-          rbaa1 = _mm_shuffle_ps(rgba1, aaaa1, _MM_SHUFFLE(0, 0, 2, 0));
+          rbaa0 = _mm_shuffle_ps(rgba0, pre_rgba0, _MM_SHUFFLE(3, 3, 2, 0));
           rgba0 = _mm_shuffle_ps(rgba0, rbaa0, _MM_SHUFFLE(2, 1, 1, 0));
-          rgba1 = _mm_shuffle_ps(rgba1, rbaa1, _MM_SHUFFLE(2, 1, 1, 0));
           
+          s++;
           *d++ = rgba0;
-          *d++ = rgba1;
         }
       _mm_empty ();
     }
@@ -134,17 +143,21 @@ conv_rgbAF_linear_rgbaF_linear (const float *src, float *dst, long samples)
   src += i * 4;
   remainder = samples - i;
   while (remainder--)
-  {
-    const float a = src[3];
-    const float a_term = 1.0f / a;
-    dst[0] = src[0] * a_term;
-    dst[1] = src[1] * a_term;
-    dst[2] = src[2] * a_term;
-    dst[3] = a;
-    
-    src += 4;
-    dst += 4;
-  }
+    {
+      float alpha = src[3];
+      float recip;
+      if (alpha <= 0.0f)
+        recip = 0.0f;
+      else
+        recip = 1.0f/alpha;
+      dst[0] = src[0] * recip;
+      dst[1] = src[1] * recip;
+      dst[2] = src[2] * recip;
+      dst[3] = alpha;
+      
+      src   += 4;
+      dst   += 4;
+    }
 
   return samples;
 }