[PATCH] Faster unorm_to_unorm for wide processing.

[Antti S. Lankila]

Because wide path doesn't have compile-time constants, unorm_to_unorm
compiles to rather complicated code that is evaluated per pixel. It's easier
to generate bit repetition according to multiplication and shift, for instance
from 10 bit value to 16-bit value you can do:

v = v * 0x401 >> 4

which is just two simple integer operations.
---
 pixman/pixman-utils.c |   31 +++++++++++++++++++++++++++----
 1 file changed, 27 insertions(+), 4 deletions(-)

diff --git a/pixman/pixman-utils.c b/pixman/pixman-utils.c
index 2ec2594..d33b10d 100644
--- a/pixman/pixman-utils.c
+++ b/pixman/pixman-utils.c
@@ -183,6 +183,21 @@ pixman_malloc_abc (unsigned int a,
 	return malloc (a * b * c);
 }
 
+void
+unorm_to_unorm_params(int inwidth, int outwidth, uint32_t *factor, int *shift)
+{
+        *factor = 0;
+        int w = 0;
+        while (inwidth != 0 && w < outwidth) {
+                *factor |= 1 << w;
+                w += inwidth;
+        }
+
+        /* Did we generate too many bits? */
+        *shift = w - outwidth;
+        return;
+}
+
 /*
  * This function expands images from ARGB8 format to ARGB16.  To preserve
  * precision, it needs to know the original source format.  For example, if the
@@ -214,6 +229,14 @@ pixman_expand (uint64_t *           dst,
                   b_mask = ~(~0 << b_size);
     int i;
 
+    uint32_t au_factor, ru_factor, gu_factor, bu_factor;
+    int au_shift, ru_shift, gu_shift, bu_shift;
+
+    unorm_to_unorm_params (a_size, 16, &au_factor, &au_shift);
+    unorm_to_unorm_params (r_size, 16, &ru_factor, &ru_shift);
+    unorm_to_unorm_params (g_size, 16, &gu_factor, &gu_shift);
+    unorm_to_unorm_params (b_size, 16, &bu_factor, &bu_shift);
+
     /* Start at the end so that we can do the expansion in place
      * when src == dst
      */
@@ -226,7 +249,7 @@ pixman_expand (uint64_t *           dst,
 	if (a_size)
 	{
 	    a = (pixel >> a_shift) & a_mask;
-	    a16 = unorm_to_unorm (a, a_size, 16);
+            a16 = a * au_factor >> au_shift;
 	}
 	else
 	{
@@ -238,9 +261,9 @@ pixman_expand (uint64_t *           dst,
 	    r = (pixel >> r_shift) & r_mask;
 	    g = (pixel >> g_shift) & g_mask;
 	    b = (pixel >> b_shift) & b_mask;
-	    r16 = unorm_to_unorm (r, r_size, 16);
-	    g16 = unorm_to_unorm (g, g_size, 16);
-	    b16 = unorm_to_unorm (b, b_size, 16);
+            r16 = r * ru_factor >> ru_shift;
+            g16 = g * gu_factor >> gu_shift;
+            b16 = b * bu_factor >> bu_shift;
 	}
 	else
 	{
-- 
1.7.9.5


--------------000002000301040300040601--