tinyjpeg: modify to handle partial image blocks

A lot of common resolutions, e.g. 800x600, include partial blocks at
the side of the image.  Implement that properly.

Signed-off-by: H. Peter Anvin <hpa@zytor.com>

9 files changed, 374 insertions, 357 deletions

diff --git a/com32/lib/Makefile b/com32/lib/Makefile
index 5e0e15ae..1fb30cd7 100644
--- a/com32/lib/Makefile
+++ b/com32/lib/Makefile
@@ -74,7 +74,8 @@ LIBOBJS = \
 	libpng/pngerror.o libpng/pngpread.o				\
 	\
 	jpeg/tinyjpeg.o jpeg/jidctflt.o	jpeg/decode1.o jpeg/decode3.o   \
-	jpeg/rgb24.o jpeg/bgr24.o jpeg/yuv420p.o jpeg/grey.o		\
+	jpeg/grey.o							\
+	jpeg/rgb24.o jpeg/bgr24.o					\
 	jpeg/rgba32.o jpeg/bgra32.o					\
 	\
 	sys/x86_init_fpu.o math/pow.o math/strtod.o			\
diff --git a/com32/lib/jpeg/bgr24.c b/com32/lib/jpeg/bgr24.c
index ffdcbdf9..d72e29b2 100644
--- a/com32/lib/jpeg/bgr24.c
+++ b/com32/lib/jpeg/bgr24.c
@@ -69,7 +69,7 @@ static unsigned char clamp(int i)
  *  | 1 |
  *  `---'
  */
-static void YCrCB_to_BGR24_1x1(struct jdec_private *priv)
+static void YCrCB_to_BGR24_1x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -85,15 +85,14 @@ static void YCrCB_to_BGR24_1x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 8*3;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
+  for (i = sy; i > 0; i--) {
+    for (j = sx; j > 0; j--) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
        int r, g , b;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        cb = *Cb++ - 128;
        cr = *Cr++ - 128;
        add_r = FIX(1.40200) * cr + ONE_HALF;
@@ -107,6 +106,7 @@ static void YCrCB_to_BGR24_1x1(struct jdec_private *priv)
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
 
+       Y++;
     }
 
     p += offset_to_next_row;
@@ -125,7 +125,7 @@ static void YCrCB_to_BGR24_1x1(struct jdec_private *priv)
  *  | 1 | 2 |
  *  `-------'
  */
-static void YCrCB_to_BGR24_2x1(struct jdec_private *priv)
+static void YCrCB_to_BGR24_2x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -141,9 +141,8 @@ static void YCrCB_to_BGR24_2x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 16*3;
-  for (i=0; i<8; i++) {
-
-    for (j=0; j<8; j++) {
+  for (i = sy; i > 0; i--) {
+    for (j = sx; j > 0; j -= 2) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -155,15 +154,7 @@ static void YCrCB_to_BGR24_2x1(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
        g = (y + add_g) >> SCALEBITS;
@@ -171,6 +162,17 @@ static void YCrCB_to_BGR24_2x1(struct jdec_private *priv)
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
 
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+       }
+
+       Y += 2;
     }
 
     p += offset_to_next_row;
@@ -190,7 +192,7 @@ static void YCrCB_to_BGR24_2x1(struct jdec_private *priv)
  *  | 2 |
  *  `---'
  */
-static void YCrCB_to_BGR24_1x2(struct jdec_private *priv)
+static void YCrCB_to_BGR24_1x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -207,9 +209,8 @@ static void YCrCB_to_BGR24_1x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 8*3;
-  for (i=0; i<8; i++) {
-
-    for (j=0; j<8; j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; j > 0 ; j--) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -221,7 +222,7 @@ static void YCrCB_to_BGR24_1x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
        g = (y + add_g) >> SCALEBITS;
@@ -229,14 +230,17 @@ static void YCrCB_to_BGR24_1x2(struct jdec_private *priv)
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
 
-       y  = (Y[8-1]) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-
+       if (i > 1) {
+	   y  = Y[8] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+       }
+
+       Y++;
     }
     Y += 8;
     p += offset_to_next_row;
@@ -258,7 +262,7 @@ static void YCrCB_to_BGR24_1x2(struct jdec_private *priv)
  *  | 3 | 4 |
  *  `-------'
  */
-static void YCrCB_to_BGR24_2x2(struct jdec_private *priv)
+static void YCrCB_to_BGR24_2x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -275,9 +279,8 @@ static void YCrCB_to_BGR24_2x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 16*3;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; j > 0; j -= 2) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -289,7 +292,7 @@ static void YCrCB_to_BGR24_2x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
        g = (y + add_g) >> SCALEBITS;
@@ -297,29 +300,37 @@ static void YCrCB_to_BGR24_2x2(struct jdec_private *priv)
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
 
-       y  = (*Y++) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-
-       y  = (Y[16-2]) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-
-       y  = (Y[16-1]) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+       }
+
+       if (i > 1) {
+	   y  = Y[16+0] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+
+	   if (j > 1) {
+	       y  = Y[16+1] << SCALEBITS;
+	       b = (y + add_b) >> SCALEBITS;
+	       *p2++ = clamp(b);
+	       g = (y + add_g) >> SCALEBITS;
+	       *p2++ = clamp(g);
+	       r = (y + add_r) >> SCALEBITS;
+	       *p2++ = clamp(r);
+	   }
+       }
+
+       Y += 2;
     }
     Y  += 16;
     p  += offset_to_next_row;
diff --git a/com32/lib/jpeg/bgra32.c b/com32/lib/jpeg/bgra32.c
index 7e2c72bc..1aa1835b 100644
--- a/com32/lib/jpeg/bgra32.c
+++ b/com32/lib/jpeg/bgra32.c
@@ -69,7 +69,7 @@ static unsigned char clamp(int i)
  *  | 1 |
  *  `---'
  */
-static void YCrCB_to_BGRA32_1x1(struct jdec_private *priv)
+static void YCrCB_to_BGRA32_1x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -85,15 +85,14 @@ static void YCrCB_to_BGRA32_1x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 8*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
+  for (i = sy; i > 0; i--) {
+    for (j = sx; j > 0; j--) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
        int r, g , b, a;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        cb = *Cb++ - 128;
        cr = *Cr++ - 128;
        add_r = FIX(1.40200) * cr + ONE_HALF;
@@ -109,6 +108,7 @@ static void YCrCB_to_BGRA32_1x1(struct jdec_private *priv)
        a = 255;
        *p++ = a;
 
+       Y++;
     }
 
     p += offset_to_next_row;
@@ -127,7 +127,7 @@ static void YCrCB_to_BGRA32_1x1(struct jdec_private *priv)
  *  | 1 | 2 |
  *  `-------'
  */
-static void YCrCB_to_BGRA32_2x1(struct jdec_private *priv)
+static void YCrCB_to_BGRA32_2x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -143,10 +143,8 @@ static void YCrCB_to_BGRA32_2x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 16*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0; j<8; j++) {
-
+  for (i = sy; i > 0; i--) {
+    for (j = sx; j > 0; j -= 2) {
        int y, cb, cr;
        int add_r, add_g, add_b;
        int r, g , b, a;
@@ -157,7 +155,7 @@ static void YCrCB_to_BGRA32_2x1(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
        g = (y + add_g) >> SCALEBITS;
@@ -166,17 +164,20 @@ static void YCrCB_to_BGRA32_2x1(struct jdec_private *priv)
        *p++ = clamp(r);
        a = 255;
        *p++ = a;
-
-       y  = (*Y++) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-       a = 255;
-       *p++ = a;
-
+       
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+	   a = 255;
+	   *p++ = a;
+       }
+
+       Y += 2;
     }
 
     p += offset_to_next_row;
@@ -196,7 +197,7 @@ static void YCrCB_to_BGRA32_2x1(struct jdec_private *priv)
  *  | 2 |
  *  `---'
  */
-static void YCrCB_to_BGRA32_1x2(struct jdec_private *priv)
+static void YCrCB_to_BGRA32_1x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -213,9 +214,8 @@ static void YCrCB_to_BGRA32_1x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 8*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0; j<8; j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; j > 0; j--) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -227,7 +227,7 @@ static void YCrCB_to_BGRA32_1x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
        g = (y + add_g) >> SCALEBITS;
@@ -237,16 +237,19 @@ static void YCrCB_to_BGRA32_1x2(struct jdec_private *priv)
        a = 255;
        *p++ = a;
 
-       y  = (Y[8-1]) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       a = 255;
-       *p2++ = a;
-
+       if (i > 1) {
+	   y  = Y[8] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+	   a = 255;
+	   *p2++ = a;
+       }
+
+       Y++;
     }
     Y += 8;
     p += offset_to_next_row;
@@ -268,7 +271,7 @@ static void YCrCB_to_BGRA32_1x2(struct jdec_private *priv)
  *  | 3 | 4 |
  *  `-------'
  */
-static void YCrCB_to_BGRA32_2x2(struct jdec_private *priv)
+static void YCrCB_to_BGRA32_2x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -285,9 +288,8 @@ static void YCrCB_to_BGRA32_2x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 16*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; j > 0; j -= 2) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -299,17 +301,7 @@ static void YCrCB_to_BGRA32_2x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-       a = 255;
-       *p++ = a;
-
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
        g = (y + add_g) >> SCALEBITS;
@@ -319,25 +311,43 @@ static void YCrCB_to_BGRA32_2x2(struct jdec_private *priv)
        a = 255;
        *p++ = a;
 
-       y  = (Y[16-2]) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       a = 255;
-       *p2++ = a;
-
-       y  = (Y[16-1]) << SCALEBITS;
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       a = 255;
-       *p2++ = a;
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+	   a = 255;
+	   *p++ = a;
+       }
+
+       if (i > 1) {
+	   y  = Y[16+0] << SCALEBITS;
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+	   a = 255;
+	   *p2++ = a;
+	   
+	   if (j > 1) {
+	       y  = Y[16+1] << SCALEBITS;
+	       b = (y + add_b) >> SCALEBITS;
+	       *p2++ = clamp(b);
+	       g = (y + add_g) >> SCALEBITS;
+	       *p2++ = clamp(g);
+	       r = (y + add_r) >> SCALEBITS;
+	       *p2++ = clamp(r);
+	       a = 255;
+	       *p2++ = a;
+	   }
+       }
+
+       Y += 2;
     }
     Y  += 16;
     p  += offset_to_next_row;
diff --git a/com32/lib/jpeg/grey.c b/com32/lib/jpeg/grey.c
index 0034ce2b..2f114744 100644
--- a/com32/lib/jpeg/grey.c
+++ b/com32/lib/jpeg/grey.c
@@ -40,12 +40,12 @@
 #include "tinyjpeg-internal.h"
 
 /**
- *  YCrCb -> Grey (1x1)
+ *  YCrCb -> Grey (1x1, 1x2)
  *  .---.
  *  | 1 |
  *  `---'
  */
-static void YCrCB_to_Grey_1x1(struct jdec_private *priv)
+static void YCrCB_to_Grey_1xN(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *y;
   unsigned char *p;
@@ -56,73 +56,20 @@ static void YCrCB_to_Grey_1x1(struct jdec_private *priv)
   y = priv->Y;
   offset_to_next_row = priv->bytes_per_row[0];
 
-  for (i=0; i<8; i++) {
-     memcpy(p, y, 8);
-     y+=8;
-     p += offset_to_next_row;
-  }
-}
-
-/**
- *  YCrCb -> Grey (2x1)
- *  .-------.
- *  | 1 | 2 |
- *  `-------'
- */
-static void YCrCB_to_Grey_2x1(struct jdec_private *priv)
-{
-  const unsigned char *y;
-  unsigned char *p;
-  unsigned int i;
-  int offset_to_next_row;
-
-  p = priv->plane[0];
-  y = priv->Y;
-  offset_to_next_row = priv->bytes_per_row[0];
-
-  for (i=0; i<8; i++) {
-     memcpy(p, y, 16);
-     y += 16;
-     p += offset_to_next_row;
-  }
-}
-
-
-/**
- *  YCrCb -> Grey (1x2)
- *  .---.
- *  | 1 |
- *  |---|
- *  | 2 |
- *  `---'
- */
-static void YCrCB_to_Grey_1x2(struct jdec_private *priv)
-{
-  const unsigned char *y;
-  unsigned char *p;
-  unsigned int i;
-  int offset_to_next_row;
-
-  p = priv->plane[0];
-  y = priv->Y;
-  offset_to_next_row = priv->bytes_per_row[0];
-
-  for (i=0; i<16; i++) {
-     memcpy(p, y, 8);
+  for (i = sy; i > 0; i--) {
+     memcpy(p, y, sx);
      y += 8;
      p += offset_to_next_row;
   }
 }
 
 /**
- *  YCrCb -> Grey (2x2)
+ *  YCrCb -> Grey (2x1, 2x2)
  *  .-------.
  *  | 1 | 2 |
- *  |---+---|
- *  | 3 | 4 |
  *  `-------'
  */
-static void YCrCB_to_Grey_2x2(struct jdec_private *priv)
+static void YCrCB_to_Grey_2xN(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *y;
   unsigned char *p;
@@ -133,8 +80,8 @@ static void YCrCB_to_Grey_2x2(struct jdec_private *priv)
   y = priv->Y;
   offset_to_next_row = priv->bytes_per_row[0];
 
-  for (i=0; i<16; i++) {
-     memcpy(p, y, 16);
+  for (i = sy; i > 0; i--) {
+     memcpy(p, y, sx);
      y += 16;
      p += offset_to_next_row;
   }
@@ -145,12 +92,12 @@ static int initialize_grey(struct jdec_private *priv,
 			   unsigned int *bytes_per_mcu)
 {
   if (priv->components[0] == NULL)
-    priv->components[0] = (uint8_t *)malloc(priv->width * priv->height * 3);
+    priv->components[0] = (uint8_t *)malloc(priv->width * priv->height);
   if (!priv->bytes_per_row[0])
-    priv->bytes_per_row[0] = priv->width * 3;
+    priv->bytes_per_row[0] = priv->width;
 
   bytes_per_blocklines[0] = priv->bytes_per_row[0];
-  bytes_per_mcu[0] = 3*8;
+  bytes_per_mcu[0] = 8;
 
   return !priv->components[0];
 }
@@ -158,10 +105,10 @@ static int initialize_grey(struct jdec_private *priv,
 static const struct tinyjpeg_colorspace format_grey =
   {
     {
-      YCrCB_to_Grey_1x1,
-      YCrCB_to_Grey_1x2,
-      YCrCB_to_Grey_2x1,
-      YCrCB_to_Grey_2x2,
+      YCrCB_to_Grey_1xN,
+      YCrCB_to_Grey_1xN,
+      YCrCB_to_Grey_2xN,
+      YCrCB_to_Grey_2xN,
     },
     tinyjpeg_decode_mcu_1comp_table,
     initialize_grey
diff --git a/com32/lib/jpeg/rgb24.c b/com32/lib/jpeg/rgb24.c
index e37cc77d..49a080de 100644
--- a/com32/lib/jpeg/rgb24.c
+++ b/com32/lib/jpeg/rgb24.c
@@ -69,7 +69,7 @@ static unsigned char clamp(int i)
  *  | 1 |
  *  `---'
  */
-static void YCrCB_to_RGB24_1x1(struct jdec_private *priv)
+static void YCrCB_to_RGB24_1x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -85,15 +85,13 @@ static void YCrCB_to_RGB24_1x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 8*3;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
-
+  for (i = sy; i > 0; i--) {
+    for (j = sx; j > 0; j++) {
        int y, cb, cr;
        int add_r, add_g, add_b;
        int r, g , b;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        cb = *Cb++ - 128;
        cr = *Cr++ - 128;
        add_r = FIX(1.40200) * cr + ONE_HALF;
@@ -107,6 +105,7 @@ static void YCrCB_to_RGB24_1x1(struct jdec_private *priv)
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
 
+       Y++;
     }
 
     p += offset_to_next_row;
@@ -124,7 +123,7 @@ static void YCrCB_to_RGB24_1x1(struct jdec_private *priv)
  *  | 1 | 2 |
  *  `-------'
  */
-static void YCrCB_to_RGB24_2x1(struct jdec_private *priv)
+static void YCrCB_to_RGB24_2x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -140,15 +139,15 @@ static void YCrCB_to_RGB24_2x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 16*3;
-  for (i=0; i<8; i++) {
+  for (i = sy; i > 0; i--) {
 
-    for (j=0; j<8; j++) {
+    for (j = sx; j > 0; j -= 2) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
        int r, g , b;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        cb = *Cb++ - 128;
        cr = *Cr++ - 128;
        add_r = FIX(1.40200) * cr + ONE_HALF;
@@ -162,14 +161,17 @@ static void YCrCB_to_RGB24_2x1(struct jdec_private *priv)
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
 
-       y  = (*Y++) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+       }
+
+       Y += 2;
     }
 
     p += offset_to_next_row;
@@ -190,7 +192,7 @@ static void YCrCB_to_RGB24_2x1(struct jdec_private *priv)
  *  | 2 |
  *  `---'
  */
-static void YCrCB_to_RGB24_1x2(struct jdec_private *priv)
+static void YCrCB_to_RGB24_1x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -207,9 +209,8 @@ static void YCrCB_to_RGB24_1x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 8*3;
-  for (i=0; i<8; i++) {
-
-    for (j=0; j<8; j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; j > 0; j--) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -221,7 +222,7 @@ static void YCrCB_to_RGB24_1x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
        g = (y + add_g) >> SCALEBITS;
@@ -229,14 +230,17 @@ static void YCrCB_to_RGB24_1x2(struct jdec_private *priv)
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
 
-       y  = (Y[8-1]) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-
+       if (i > 1) {
+	   y  = Y[8] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+       }
+
+       Y++;
     }
     Y += 8;
     p += offset_to_next_row;
@@ -257,7 +261,7 @@ static void YCrCB_to_RGB24_1x2(struct jdec_private *priv)
  *  | 3 | 4 |
  *  `-------'
  */
-static void YCrCB_to_RGB24_2x2(struct jdec_private *priv)
+static void YCrCB_to_RGB24_2x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -274,9 +278,8 @@ static void YCrCB_to_RGB24_2x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 16*3;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; j > 0; j -= 2) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -288,7 +291,7 @@ static void YCrCB_to_RGB24_2x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
        g = (y + add_g) >> SCALEBITS;
@@ -296,29 +299,37 @@ static void YCrCB_to_RGB24_2x2(struct jdec_private *priv)
        b = (y + add_b) >> SCALEBITS;
        *p++ = clamp(b);
 
-       y  = (*Y++) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-
-       y  = (Y[16-2]) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-
-       y  = (Y[16-1]) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+       }
+
+       if (i > 1) {
+	   y  = Y[16+0] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+
+	   if (j > 1) {
+	       y  = Y[16+1] << SCALEBITS;
+	       r = (y + add_r) >> SCALEBITS;
+	       *p2++ = clamp(r);
+	       g = (y + add_g) >> SCALEBITS;
+	       *p2++ = clamp(g);
+	       b = (y + add_b) >> SCALEBITS;
+	       *p2++ = clamp(b);
+	   }
+       }
+       
+       Y += 2;
     }
     Y  += 16;
     p  += offset_to_next_row;
diff --git a/com32/lib/jpeg/rgba32.c b/com32/lib/jpeg/rgba32.c
index 5fcbe992..da08344d 100644
--- a/com32/lib/jpeg/rgba32.c
+++ b/com32/lib/jpeg/rgba32.c
@@ -69,7 +69,7 @@ static unsigned char clamp(int i)
  *  | 1 |
  *  `---'
  */
-static void YCrCB_to_RGBA32_1x1(struct jdec_private *priv)
+static void YCrCB_to_RGBA32_1x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -85,15 +85,14 @@ static void YCrCB_to_RGBA32_1x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 8*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
+  for (i = sy; i > 0; i--) {
+    for (j = sx; j > 0; j--) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
        int r, g , b, a;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        cb = *Cb++ - 128;
        cr = *Cr++ - 128;
        add_r = FIX(1.40200) * cr + ONE_HALF;
@@ -108,6 +107,8 @@ static void YCrCB_to_RGBA32_1x1(struct jdec_private *priv)
        *p++ = clamp(b);
        a = 255;
        *p++ = a;
+
+       Y++;
     }
 
     p += offset_to_next_row;
@@ -125,7 +126,7 @@ static void YCrCB_to_RGBA32_1x1(struct jdec_private *priv)
  *  | 1 | 2 |
  *  `-------'
  */
-static void YCrCB_to_RGBA32_2x1(struct jdec_private *priv)
+static void YCrCB_to_RGBA32_2x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p;
@@ -141,15 +142,14 @@ static void YCrCB_to_RGBA32_2x1(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = priv->bytes_per_row[0] - 16*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0; j<8; j++) {
+  for (i = sy; i > 0; i--) {
+    for (j = sx; j > 0; j -= 2) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
        int r, g , b, a;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        cb = *Cb++ - 128;
        cr = *Cr++ - 128;
        add_r = FIX(1.40200) * cr + ONE_HALF;
@@ -165,15 +165,19 @@ static void YCrCB_to_RGBA32_2x1(struct jdec_private *priv)
        a = 255;
        *p++ = a;
 
-       y  = (*Y++) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-       a = 255;
-       *p++ = a;
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+	   a = 255;
+	   *p++ = a;
+       }
+
+       Y += 2;
     }
 
     p += offset_to_next_row;
@@ -194,7 +198,7 @@ static void YCrCB_to_RGBA32_2x1(struct jdec_private *priv)
  *  | 2 |
  *  `---'
  */
-static void YCrCB_to_RGBA32_1x2(struct jdec_private *priv)
+static void YCrCB_to_RGBA32_1x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -211,9 +215,8 @@ static void YCrCB_to_RGBA32_1x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 8*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0; j<8; j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; j > 0; j--) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -225,7 +228,7 @@ static void YCrCB_to_RGBA32_1x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
        g = (y + add_g) >> SCALEBITS;
@@ -235,16 +238,19 @@ static void YCrCB_to_RGBA32_1x2(struct jdec_private *priv)
        a = 255;
        *p++ = a;
 
-       y  = (Y[8-1]) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       a = 255;
-       *p2++ = a;
-
+       if (i > 1) {
+	   y  = Y[8] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+	   a = 255;
+	   *p2++ = a;
+       }
+
+       Y++;
     }
     Y += 8;
     p += offset_to_next_row;
@@ -265,7 +271,7 @@ static void YCrCB_to_RGBA32_1x2(struct jdec_private *priv)
  *  | 3 | 4 |
  *  `-------'
  */
-static void YCrCB_to_RGBA32_2x2(struct jdec_private *priv)
+static void YCrCB_to_RGBA32_2x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *Y, *Cb, *Cr;
   unsigned char *p, *p2;
@@ -282,9 +288,8 @@ static void YCrCB_to_RGBA32_2x2(struct jdec_private *priv)
   Cb = priv->Cb;
   Cr = priv->Cr;
   offset_to_next_row = 2*priv->bytes_per_row[0] - 16*4;
-  for (i=0; i<8; i++) {
-
-    for (j=0;j<8;j++) {
+  for (i = sy; i > 0; i -= 2) {
+    for (j = sx; i > 0; j -= 2) {
 
        int y, cb, cr;
        int add_r, add_g, add_b;
@@ -296,7 +301,7 @@ static void YCrCB_to_RGBA32_2x2(struct jdec_private *priv)
        add_g = - FIX(0.34414) * cb - FIX(0.71414) * cr + ONE_HALF;
        add_b = FIX(1.77200) * cb + ONE_HALF;
 
-       y  = (*Y++) << SCALEBITS;
+       y  = Y[0] << SCALEBITS;
        r = (y + add_r) >> SCALEBITS;
        *p++ = clamp(r);
        g = (y + add_g) >> SCALEBITS;
@@ -306,35 +311,43 @@ static void YCrCB_to_RGBA32_2x2(struct jdec_private *priv)
        a = 255;
        *p++ = a;
 
-       y  = (*Y++) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p++ = clamp(b);
-       a = 255;
-       *p++ = a;
-
-       y  = (Y[16-2]) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       a = 255;
-       *p2++ = a;
-
-       y  = (Y[16-1]) << SCALEBITS;
-       r = (y + add_r) >> SCALEBITS;
-       *p2++ = clamp(r);
-       g = (y + add_g) >> SCALEBITS;
-       *p2++ = clamp(g);
-       b = (y + add_b) >> SCALEBITS;
-       *p2++ = clamp(b);
-       a = 255;
-       *p2++ = a;
+       if (j > 1) {
+	   y  = Y[1] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p++ = clamp(b);
+	   a = 255;
+	   *p++ = a;
+       }
+
+       if (i > 1) {
+	   y  = Y[16+0] << SCALEBITS;
+	   r = (y + add_r) >> SCALEBITS;
+	   *p2++ = clamp(r);
+	   g = (y + add_g) >> SCALEBITS;
+	   *p2++ = clamp(g);
+	   b = (y + add_b) >> SCALEBITS;
+	   *p2++ = clamp(b);
+	   a = 255;
+	   *p2++ = a;
+
+	   if (j > 1) {
+	       y  = Y[16+1] << SCALEBITS;
+	       r = (y + add_r) >> SCALEBITS;
+	       *p2++ = clamp(r);
+	       g = (y + add_g) >> SCALEBITS;
+	       *p2++ = clamp(g);
+	       b = (y + add_b) >> SCALEBITS;
+	       *p2++ = clamp(b);
+	       a = 255;
+	       *p2++ = a;
+	   }
+       }
+
+       Y += 2;
     }
     Y  += 16;
     p  += offset_to_next_row;
diff --git a/com32/lib/jpeg/tinyjpeg-internal.h b/com32/lib/jpeg/tinyjpeg-internal.h
index bd0e58da..bb5c6707 100644
--- a/com32/lib/jpeg/tinyjpeg-internal.h
+++ b/com32/lib/jpeg/tinyjpeg-internal.h
@@ -80,7 +80,7 @@ struct component
 
 
 typedef void (*decode_MCU_fct) (struct jdec_private *priv);
-typedef void (*convert_colorspace_fct) (struct jdec_private *priv);
+typedef void (*convert_colorspace_fct) (struct jdec_private *priv, int, int);
 
 struct jdec_private
 {
@@ -167,6 +167,16 @@ enum std_markers {
 #define trace(fmt, args...) do { } while (0)
 #endif
 
+#ifndef __likely
+# define __likely(x) (!!(x))
+#endif
+#ifndef __unlikely
+# define __unlikely(x) (!!(x))
+#endif
+
+#define min(x, y) ((x) < (y) ? (x) : (y))
+#define max(x, y) ((x) > (y) ? (x) : (y))
+
 #if 0
 static char *print_bits(unsigned int value, char *bitstr)
 {
diff --git a/com32/lib/jpeg/tinyjpeg.c b/com32/lib/jpeg/tinyjpeg.c
index 657ff6e1..a9395a5d 100644
--- a/com32/lib/jpeg/tinyjpeg.c
+++ b/com32/lib/jpeg/tinyjpeg.c
@@ -546,12 +546,6 @@ static int parse_SOF(struct jdec_private *priv, const unsigned char *stream)
     error("Width and Height (%dx%d) seems suspicious\n", width, height);
   if (nr_components != 3)
     error("We only support YUV images\n");
-#if 0
-  if (height%16)
-    error("Height need to be a multiple of 16 (current height is %d)\n", height);
-  if (width%16)
-    error("Width need to be a multiple of 16 (current Width is %d)\n", width);
-#endif
 #endif
   stream += 8;
   for (i=0; i<nr_components; i++) {
@@ -883,18 +877,21 @@ int tinyjpeg_parse_header(struct jdec_private *priv, const unsigned char *buf, u
 int tinyjpeg_decode(struct jdec_private *priv,
 		    const struct tinyjpeg_colorspace *pixfmt)
 {
-  unsigned int x, y, xstride_by_mcu, ystride_by_mcu;
+  int x, y, sx, sy;
+  int xshift_by_mcu, yshift_by_mcu;
+  int xstride_by_mcu, ystride_by_mcu;
   unsigned int bytes_per_blocklines[3], bytes_per_mcu[3];
   decode_MCU_fct decode_MCU;
   const decode_MCU_fct *decode_mcu_table;
   convert_colorspace_fct convert_to_pixfmt;
+  uint8_t *pptr[3];
 
   decode_mcu_table = pixfmt->decode_mcu_table;
 
   /* Fix: check return value */
   pixfmt->initialize(priv, bytes_per_blocklines, bytes_per_mcu);
 
-  xstride_by_mcu = ystride_by_mcu = 8;
+  xshift_by_mcu = yshift_by_mcu = 3;
   if ((priv->component_infos[cY].Hfactor | priv->component_infos[cY].Vfactor) == 1) {
      decode_MCU = decode_mcu_table[0];
      convert_to_pixfmt = pixfmt->convert_colorspace[0];
@@ -902,43 +899,58 @@ int tinyjpeg_decode(struct jdec_private *priv,
   } else if (priv->component_infos[cY].Hfactor == 1) {
      decode_MCU = decode_mcu_table[1];
      convert_to_pixfmt = pixfmt->convert_colorspace[1];
-     ystride_by_mcu = 16;
+     yshift_by_mcu = 4;
      trace("Use decode 1x2 sampling (not supported)\n");
   } else if (priv->component_infos[cY].Vfactor == 2) {
      decode_MCU = decode_mcu_table[3];
      convert_to_pixfmt = pixfmt->convert_colorspace[3];
-     xstride_by_mcu = 16;
-     ystride_by_mcu = 16;
+     xshift_by_mcu = 4;
+     yshift_by_mcu = 4;
      trace("Use decode 2x2 sampling\n");
   } else {
      decode_MCU = decode_mcu_table[2];
      convert_to_pixfmt = pixfmt->convert_colorspace[2];
-     xstride_by_mcu = 16;
+     xshift_by_mcu = 4;
      trace("Use decode 2x1 sampling\n");
   }
 
   resync(priv);
 
   /* Don't forget to that block can be either 8 or 16 lines */
-  bytes_per_blocklines[0] *= ystride_by_mcu;
-  bytes_per_blocklines[1] *= ystride_by_mcu;
-  bytes_per_blocklines[2] *= ystride_by_mcu;
+  bytes_per_blocklines[0] <<= yshift_by_mcu;
+  bytes_per_blocklines[1] <<= yshift_by_mcu;
+  bytes_per_blocklines[2] <<= yshift_by_mcu;
+
+  bytes_per_mcu[0] <<= xshift_by_mcu-3;
+  bytes_per_mcu[1] <<= xshift_by_mcu-3;
+  bytes_per_mcu[2] <<= xshift_by_mcu-3;
+
+  xstride_by_mcu = 1 << xshift_by_mcu;
+  ystride_by_mcu = 1 << yshift_by_mcu;
+
+  pptr[0] = priv->components[0];
+  pptr[1] = priv->components[1];
+  pptr[2] = priv->components[2];
 
-  bytes_per_mcu[0] *= xstride_by_mcu/8;
-  bytes_per_mcu[1] *= xstride_by_mcu/8;
-  bytes_per_mcu[2] *= xstride_by_mcu/8;
+  trace("bpbl = %d, bpmcu = %d\n",
+	bytes_per_blocklines[0], bytes_per_mcu[0]);
 
-  /* Just the decode the image by macroblock (size is 8x8, 8x16, or 16x16) */
-  for (y=0; y < priv->height/ystride_by_mcu; y++)
+  for (y = priv->height; y > 0; y -= ystride_by_mcu)
    {
-     //trace("Decoding row %d\n", y);
-     priv->plane[0] = priv->components[0] + (y * bytes_per_blocklines[0]);
-     priv->plane[1] = priv->components[1] + (y * bytes_per_blocklines[1]);
-     priv->plane[2] = priv->components[2] + (y * bytes_per_blocklines[2]);
-     for (x=0; x < priv->width; x+=xstride_by_mcu)
+     trace("Decoding row %d\n", priv->height-y);
+     priv->plane[0] = pptr[0];  pptr[0] += bytes_per_blocklines[0];
+     priv->plane[1] = pptr[1];  pptr[1] += bytes_per_blocklines[1];
+     priv->plane[2] = pptr[2];  pptr[2] += bytes_per_blocklines[2];
+
+     sy = min(y, ystride_by_mcu);
+
+     for (x = priv->width; x > 0; x -= xstride_by_mcu)
       {
+	sx = min(x, xstride_by_mcu);
+	trace("Block size: %dx%d\n", sx, sy);
+
 	decode_MCU(priv);
-	convert_to_pixfmt(priv);
+	convert_to_pixfmt(priv, sx, sy);
 	priv->plane[0] += bytes_per_mcu[0];
 	priv->plane[1] += bytes_per_mcu[1];
 	priv->plane[2] += bytes_per_mcu[2];
diff --git a/com32/lib/jpeg/yuv420p.c b/com32/lib/jpeg/yuv420p.c
index f8aeb447..0c84cada 100644
--- a/com32/lib/jpeg/yuv420p.c
+++ b/com32/lib/jpeg/yuv420p.c
@@ -35,6 +35,8 @@
  * yuv420p.c
  */
 
+/**** XXXX: THIS IS NOT YET CONVERTED TO HANDLE PARTIAL MCUS ****/
+
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -64,7 +66,7 @@
  *  | 1 |
  *  `---'
  */
-static void YCrCB_to_YUV420P_1x1(struct jdec_private *priv)
+static void YCrCB_to_YUV420P_1x1(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *s, *y;
   unsigned char *p;
@@ -106,7 +108,7 @@ static void YCrCB_to_YUV420P_1x1(struct jdec_private *priv)
  *  | 1 | 2 |
  *  `-------'
  */
-static void YCrCB_to_YUV420P_2x1(struct jdec_private *priv)
+static void YCrCB_to_YUV420P_2x1(struct jdec_private *priv, int sx, int sy)
 {
   unsigned char *p;
   const unsigned char *s, *y1;
@@ -149,7 +151,7 @@ static void YCrCB_to_YUV420P_2x1(struct jdec_private *priv)
  *  | 2 |
  *  `---'
  */
-static void YCrCB_to_YUV420P_1x2(struct jdec_private *priv)
+static void YCrCB_to_YUV420P_1x2(struct jdec_private *priv, int sx, int sy)
 {
   const unsigned char *s, *y;
   unsigned char *p;
@@ -191,7 +193,7 @@ static void YCrCB_to_YUV420P_1x2(struct jdec_private *priv)
  *  | 3 | 4 |
  *  `-------'
  */
-static void YCrCB_to_YUV420P_2x2(struct jdec_private *priv)
+static void YCrCB_to_YUV420P_2x2(struct jdec_private *priv, int sx, int sy)
 {
   unsigned char *p;
   const unsigned char *s, *y1;