[xine-cvs] HG: xine-lib-1.2: Improved mmx_yuv2rgb()

[Torsten J. <t....@gm...>]

# HG changeset patch
# User Torsten Jager <t....@gm...>
# Date 1334662056 -10800
# Node ID bd20c9edb6ae250d9d03fbfd2d0fff70ee685da2
# Branch  default
# Parent  9b5435b65fae102c139644f3e1cf0f4bf7a87276
Improved mmx_yuv2rgb()
yuv2rgb_mmx.c scales YUV and rounds them down to 8 bits
individually before the addition. That causes red and
blue to be off by up to 2, green even off by 3.

This little patch does the stuff using 10 bits per
component, plus correct rounding.

There seems to be no noticable impact on performance,
but color gradients come out much smoother now.

diff --git a/src/video_out/yuv2rgb_mmx.c b/src/video_out/yuv2rgb_mmx.c
--- a/src/video_out/yuv2rgb_mmx.c
+++ b/src/video_out/yuv2rgb_mmx.c
@@ -119,6 +119,7 @@
 {
     static mmx_t mmx_80w = {0x0080008000800080ULL};
     static mmx_t mmx_00ffw = {0x00ff00ff00ff00ffULL};
+    static mmx_t mmx_0002w = {0x0002000200020002ULL};
 
     movq_m2r (*py, mm6);		// mm6 = Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0
     pxor_r2r (mm4, mm4);		// mm4 = 0
@@ -133,10 +134,11 @@
     psrlw_i2r (8, mm7);			// mm7 =    Y7    Y5    Y3    Y1
 
     movd_m2r (*pv, mm1);		// mm1 = 00 00 00 00 v3 v2 v1 v0
-    psllw_i2r (3, mm6);			// promote precision
+    psllw_i2r (5, mm6);			// promote precision
 
     pmulhw_m2r (csc->Y_coeff, mm6);	// mm6 = luma_rgb even
-    psllw_i2r (3, mm7);			// promote precision
+    paddsw_m2r (mmx_0002w, mm6);	// +0.5 for later rounding
+    psllw_i2r (5, mm7);			// promote precision
 
     punpcklbw_r2r (mm4, mm0);		// mm0 = u3 u2 u1 u0
 
@@ -144,12 +146,13 @@
     punpcklbw_r2r (mm4, mm1);		// mm1 = v3 v2 v1 v0
 
     pmulhw_m2r (csc->Y_coeff, mm7);	// mm7 = luma_rgb odd
-    psllw_i2r (3, mm0);			// promote precision
+    paddsw_m2r (mmx_0002w, mm7);	// +0.5 for later rounding
+    psllw_i2r (5, mm0);			// promote precision
 
     psubsw_m2r (mmx_80w, mm1);		// v -= 128
     movq_r2r (mm0, mm2);		// mm2 = u3 u2 u1 u0
 
-    psllw_i2r (3, mm1);			// promote precision
+    psllw_i2r (5, mm1);			// promote precision
 
     movq_r2r (mm1, mm4);		// mm4 = v3 v2 v1 v0
 
@@ -168,12 +171,14 @@
     paddsw_r2r (mm6, mm0);		// mm0 = B6 B4 B2 B0
     paddsw_r2r (mm7, mm3);		// mm3 = B7 B5 B3 B1
 
+    psraw_i2r (2, mm0);			// div round
     packuswb_r2r (mm0, mm0);		// saturate to 0-255
 
 
     pmulhw_m2r (csc->U_green, mm2);	// mm2 = u * u_green
 
 
+    psraw_i2r (2, mm3);			// div round
     packuswb_r2r (mm3, mm3);		// saturate to 0-255
 
 
@@ -197,16 +202,20 @@
     paddsw_r2r (mm6, mm2);		// mm2 = G6 G4 G2 G0
 
 
+    psraw_i2r (2, mm2);			// div round
     packuswb_r2r (mm2, mm2);		// saturate to 0-255
     paddsw_r2r (mm6, mm1);		// mm1 = R6 R4 R2 R0
 
+    psraw_i2r (2, mm1);			// div round
     packuswb_r2r (mm1, mm1);		// saturate to 0-255
     paddsw_r2r (mm7, mm4);		// mm4 = R7 R5 R3 R1
 
+    psraw_i2r (2, mm4);			// div round
     packuswb_r2r (mm4, mm4);		// saturate to 0-255
     paddsw_r2r (mm7, mm5);		// mm5 = G7 G5 G3 G1
 
 
+    psraw_i2r (2, mm5);			// div round
     packuswb_r2r (mm5, mm5);		// saturate to 0-255