Fixed S3 SIMD RGBX8888 output

Author: Laurence Bank

src/jpeg.inl

@@ -3133,15 +3133,15 @@ static void JPEGPixelRGB(uint32_t *pDest, int iY, int iCb, int iCr)
     i32 = ((iCBB + iY) >> 12);
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
-    u32Pixel |= (uint32_t)i32; // blue
+    u32Pixel |= (uint32_t)(i32<<16); // blue
     i32 = ((iCBG + iCRG + iY) >> 12); // green pixel
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
     u32Pixel |= (uint32_t)(i32 << 8);
     i32 = ((iCRR + iY) >> 12); // red pixel
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
-    u32Pixel |= (uint32_t)(i32 << 16);
+    u32Pixel |= (uint32_t)(i32);
     pDest[0] = u32Pixel;
 } /* JPEGPixelRGB() */
 
@@ -3221,28 +3221,28 @@ static void JPEGPixel2RGB(uint32_t *pDest, int32_t iY1, int32_t iY2, int32_t iCb
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
     u32Pixel1 = u32Pixel2 = 0xff000000; // Alpha = 255
-    u32Pixel1 |= (uint32_t)i32; // blue
+    u32Pixel1 |= (uint32_t)(i32<<16); // blue
     i32 = ((iCBG + iCRG + iY1) >> 12); // green pixel
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
     u32Pixel1 |= (uint32_t)(i32 << 8); // green
     i32 = ((iCRR + iY1) >> 12); // red pixel
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
-    u32Pixel1 |= (uint32_t)(i32 << 16); // red
+    u32Pixel1 |= (uint32_t)i32; // red
 
     i32 = ((iCBB + iY2) >> 12); // blue pixel
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
-    u32Pixel2 |= (uint32_t)i32;
+    u32Pixel2 |= (uint32_t)(i32<<16);
     i32 = ((iCBG + iCRG + iY2) >> 12); // green pixel
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
     u32Pixel2 |= (uint32_t)(i32 << 8);
     i32 = ((iCRR + iY2) >> 12); // red pixel
     if (i32 < 0) i32 = 0;
     else if (i32 > 255) i32 = 255;
-    u32Pixel2 |= (uint32_t)(i32 << 16);
+    u32Pixel2 |= (uint32_t)i32;
     pDest[0] = u32Pixel1;
     pDest[1] = u32Pixel2;
 } /* JPEGPixel2RGB() */

src/s3_simd_420.S

@@ -117,7 +117,7 @@
   ee.vzip.16 q7,q2  # create RGB8888 pixels
   ee.vst.128.ip q7,a5,16  # store 8 x RGB8888 pixels = 32 bytes
   ee.vst.128.ip q2,a5,16
-  addi.n a6,a6,-16         # restore pointer to start of 16-bit constants
+  addi.n a6,a6,-12         # restore pointer to start of 16-bit constants
   ee.vld.l.64.ip    q0,a2,0   # load right 8 Y values into Q0
   ee.movi.32.q q1,a3,0        # restore second 4 values of Cb
   ee.movi.32.q q2,a4,0        # restore second 4 values of Cr