$OpenBSD: patch-src_video_out_yuv2rgb_mmx_c,v 1.2 2012/06/09 08:41:16 ajacoutot Exp $

Improve mmx_yuv2rgb().

--- src/video_out/yuv2rgb_mmx.c.orig	Tue Mar  9 17:17:05 2010
+++ src/video_out/yuv2rgb_mmx.c	Fri Jun  8 23:53:28 2012
@@ -50,7 +50,8 @@ do {				\
 typedef struct mmx_csc_s mmx_csc_t;
 
 struct mmx_csc_s {
-  mmx_t subYw;
+  mmx_t x00ffw;
+  mmx_t x0080w;
   mmx_t addYw;
   mmx_t U_green;
   mmx_t U_blue;
@@ -64,30 +65,24 @@ extern const int32_t Inverse_Table_6_9[8][4];
 void mmx_yuv2rgb_set_csc_levels(yuv2rgb_factory_t *this,
 				int brightness, int contrast, int saturation)
 {
-  int a,s,i;
   int crv, cbu, cgu, cgv, cty;
   mmx_csc_t *csc;
+  int i;
 
+  int yoffset = -16;
+  int ygain = ((1 << 16) * 255) / 219;
+
   /* 'table_mmx' is 64bit aligned for better performance */
   if (this->table_mmx == NULL) {
     this->table_mmx = xine_xmalloc_aligned (8, sizeof(mmx_csc_t), &this->table_mmx_base);
   }
 
-  if( brightness <= 16 ) {
-    a = 0;
-    s = 16 - brightness;
-  } else {
-    a = brightness - 16;
-    s = 0;
-  }
+  yoffset += brightness;
+  /* TV set behaviour: contrast affects color difference as well */
+  saturation = (contrast * saturation + 64) >> 7;
 
   csc = (mmx_csc_t *) this->table_mmx;
 
-  for( i = 0; i < 8; i++ ) {
-    csc->subYw.ub[i] = s;
-    csc->addYw.ub[i] = a;
-  }
-
   crv = Inverse_Table_6_9[this->matrix_coefficients][0];
   crv = (crv * saturation + 512) / 1024;
   cbu = Inverse_Table_6_9[this->matrix_coefficients][1];
@@ -97,114 +92,105 @@ void mmx_yuv2rgb_set_csc_levels(yuv2rgb_factory_t *thi
   cgu = (cgu * saturation + 512) / 1024;
   cgv = Inverse_Table_6_9[this->matrix_coefficients][3];
   cgv = (cgv * saturation + 512) / 1024;
-  cty = (76309 * contrast + 512) / 1024;
+  cty = (ygain * contrast + 512) / 1024;
 
+  /* the 8 is "+0,5" for later rounding */
+  yoffset = ((cty * (yoffset << 7)) >> 16) + 8;
+
   for (i=0; i < 4; i++) {
     csc->U_green.w[i] = -cgu;
     csc->U_blue.w[i]  =  cbu;
     csc->V_red.w[i]   =  crv;
     csc->V_green.w[i] = -cgv;
     csc->Y_coeff.w[i] =  cty;
+
+    csc->addYw.w[i]   = yoffset;
+
+    csc->x0080w.w[i]  = 128;
+    csc->x00ffw.w[i]  = 0xff;
   }
 }
 
 static inline void mmx_yuv2rgb (uint8_t * py, uint8_t * pu, uint8_t * pv, mmx_csc_t *csc)
 {
-    static mmx_t mmx_80w = {0x0080008000800080ULL};
-    static mmx_t mmx_00ffw = {0x00ff00ff00ff00ffULL};
 
+    /* OK what we're doing here is
+       y = ((cty * (y << 7)) >> 16) + yoffset;
+       u = (u - 128) << 7;
+       v = (v - 128) << 7;
+       r = (y + ((crv * v) >> 16)) >> 4;
+       g = (y + ((cgu * u) >> 16) + ((cgv * v) >> 16)) >> 4;
+       b = (y + ((cbu * u) >> 16)) >> 4; */
+
     movq_m2r (*py, mm6);		// mm6 = Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0
     pxor_r2r (mm4, mm4);		// mm4 = 0
-
-    psubusb_m2r (csc->subYw, mm6);	// Y -= 16
-    paddusb_m2r (csc->addYw, mm6);
-
     movd_m2r (*pu, mm0);		// mm0 = 00 00 00 00 u3 u2 u1 u0
+
     movq_r2r (mm6, mm7);		// mm7 = Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0
+    pand_m2r (csc->x00ffw, mm6);	// mm6 =    Y6    Y4    Y2    Y0
 
-    pand_m2r (mmx_00ffw, mm6);		// mm6 =    Y6    Y4    Y2    Y0
     psrlw_i2r (8, mm7);			// mm7 =    Y7    Y5    Y3    Y1
-
+    psllw_i2r (7, mm6);			// promote precision
     movd_m2r (*pv, mm1);		// mm1 = 00 00 00 00 v3 v2 v1 v0
-    psllw_i2r (3, mm6);			// promote precision
 
     pmulhw_m2r (csc->Y_coeff, mm6);	// mm6 = luma_rgb even
-    psllw_i2r (3, mm7);			// promote precision
-
+    psllw_i2r (7, mm7);			// promote precision
     punpcklbw_r2r (mm4, mm0);		// mm0 = u3 u2 u1 u0
 
-    psubsw_m2r (mmx_80w, mm0);		// u -= 128
+    paddsw_m2r (csc->addYw, mm6);	// += yoffset
+    psubsw_m2r (csc->x0080w, mm0);	// u -= 128
     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
 
-    psubsw_m2r (mmx_80w, mm1);		// v -= 128
+    psllw_i2r (7, mm0);			// promote precision
+    psubsw_m2r (csc->x0080w, mm1);	// v -= 128
+
     movq_r2r (mm0, mm2);		// mm2 = u3 u2 u1 u0
+    psllw_i2r (7, mm1);			// promote precision
 
-    psllw_i2r (3, mm1);			// promote precision
-
     movq_r2r (mm1, mm4);		// mm4 = v3 v2 v1 v0
 
-    pmulhw_m2r (csc->U_blue, mm0);		// mm0 = chroma_b
-
-
-    // slot
-
-
-    // slot
-
-
+    paddsw_m2r (csc->addYw, mm7);	// += yoffset
+    pmulhw_m2r (csc->U_blue, mm0);	// mm0 = chroma_b
     pmulhw_m2r (csc->V_red, mm1);	// mm1 = chroma_r
+
     movq_r2r (mm0, mm3);		// mm3 = chroma_b
 
     paddsw_r2r (mm6, mm0);		// mm0 = B6 B4 B2 B0
     paddsw_r2r (mm7, mm3);		// mm3 = B7 B5 B3 B1
 
-    packuswb_r2r (mm0, mm0);		// saturate to 0-255
-
-
+    psraw_i2r (4, mm0);			// div round
     pmulhw_m2r (csc->U_green, mm2);	// mm2 = u * u_green
+    psraw_i2r (4, mm3);			// div round
 
-
+    packuswb_r2r (mm0, mm0);		// saturate to 0-255
     packuswb_r2r (mm3, mm3);		// saturate to 0-255
-
-
-    punpcklbw_r2r (mm3, mm0);		// mm0 = B7 B6 B5 B4 B3 B2 B1 B0
-
-
     pmulhw_m2r (csc->V_green, mm4);	// mm4 = v * v_green
 
-
-    // slot
-
-
-    // slot
-
-
+    punpcklbw_r2r (mm3, mm0);		// mm0 = B7 B6 B5 B4 B3 B2 B1 B0
     paddsw_r2r (mm4, mm2);		// mm2 = chroma_g
+    movq_r2r (mm1, mm4);		// mm4 = chroma_r
+
     movq_r2r (mm2, mm5);		// mm5 = chroma_g
 
-
-    movq_r2r (mm1, mm4);		// mm4 = chroma_r
     paddsw_r2r (mm6, mm2);		// mm2 = G6 G4 G2 G0
 
+    psraw_i2r (4, mm2);			// div round
+    paddsw_r2r (mm6, mm1);		// mm1 = R6 R4 R2 R0
 
     packuswb_r2r (mm2, mm2);		// saturate to 0-255
-    paddsw_r2r (mm6, mm1);		// mm1 = R6 R4 R2 R0
+    psraw_i2r (4, mm1);			// div round
+    paddsw_r2r (mm7, mm4);		// mm4 = R7 R5 R3 R1
 
     packuswb_r2r (mm1, mm1);		// saturate to 0-255
-    paddsw_r2r (mm7, mm4);		// mm4 = R7 R5 R3 R1
+    psraw_i2r (4, mm4);			// div round
+    paddsw_r2r (mm7, mm5);		// mm5 = G7 G5 G3 G1
 
     packuswb_r2r (mm4, mm4);		// saturate to 0-255
-    paddsw_r2r (mm7, mm5);		// mm5 = G7 G5 G3 G1
+    psraw_i2r (4, mm5);			// div round
 
-
-    packuswb_r2r (mm5, mm5);		// saturate to 0-255
-
-
     punpcklbw_r2r (mm4, mm1);		// mm1 = R7 R6 R5 R4 R3 R2 R1 R0
-
+    packuswb_r2r (mm5, mm5);		// saturate to 0-255
 
     punpcklbw_r2r (mm5, mm2);		// mm2 = G7 G6 G5 G4 G3 G2 G1 G0
 }
