SkScaledCodec class

src/codec/SkSwizzler.cpp

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkCodecPriv.h"
 #include "SkColorPriv.h"
+#include "SkScaledCodec.h"
 #include "SkSwizzler.h"
 #include "SkTemplates.h"
 #include "SkUtils.h"
 
 SkSwizzler::ResultAlpha SkSwizzler::GetResult(uint8_t zeroAlpha,
                                               uint8_t maxAlpha) {
     // In the transparent case, this returns 0x0000
     // In the opaque case, this returns 0xFFFF
     // If the row is neither transparent nor opaque, returns something else
     return (((uint16_t) maxAlpha) << 8) | zeroAlpha;
 }
 
+// samples the row. Does not do anything else but sampling
+static SkSwizzler::ResultAlpha sample565(void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src,
+        int width, int deltaSrc, const SkPMColor ctable[], int offset){
+
+    src += offset;
+    uint16_t* SK_RESTRICT dst = (uint16_t*) dstRow;
+    for (int x = 0; x < width; x++) {
+        dst[x] = src[1] << 8 | src[0];
+        src += deltaSrc;
+    }
+    // 565 is always opaque
+    return SkSwizzler::kOpaque_ResultAlpha;
+}
+
 // kBit
 // These routines exclusively choose between white and black
 
 #define GRAYSCALE_BLACK 0
 #define GRAYSCALE_WHITE 0xFF
 
+
 static SkSwizzler::ResultAlpha swizzle_bit_to_grayscale(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,

[scroggo, 2015/08/07 19:39:48]

Refresh my memory: is width the full width of the src? Or is it scaled, or does
it take the offset into account?

[emmaleer, 2015/08/11 20:10:35]

Width is the scaled dstWidth. 
The offset is used to choose the first pixel from the src.
Offset does not affect width in any way. 
Do you think we should change the variable width to dstWidth?
I could do that to make it more clear.

[scroggo, 2015/08/11 21:33:30]

Yes, I think that would make it more clear.

[emmaleer, 2015/08/12 14:04:33]

Acknowledged.

-        int /*bitsPerPixel*/, const SkPMColor* /*ctable*/) {
+        int deltaSrc, const SkPMColor* /*ctable*/, int offset) {
+
     uint8_t* SK_RESTRICT dst = (uint8_t*) dstRow;
 
-    // Determine how many full bytes are in the row
-    int bytesInRow = width >> 3;
-    int i;
-    for (i = 0; i < bytesInRow; i++) {
-        U8CPU currByte = src[i];
-        for (int j = 0; j < 8; j++) {
-            dst[j] = ((currByte >> (7 - j)) & 1) ? GRAYSCALE_WHITE : GRAYSCALE_BLACK;
+    int x = 0;
+    // increment src by byte offset and currBit by bit offset
+    src += offset / 8;

[scroggo, 2015/08/07 19:39:48]

We can do this divide much faster with:

offset >> 3

[emmaleer, 2015/08/11 20:10:35]

Acknowledged.

[mtklein, 2015/08/12 21:51:11]

Woah woah, there's absolutely no way the compiler's not doing that for you.  You
should never bother strength reducing a divide by a constant like this yourself.
 You're depriving the compiler of all its joy in life!
  src += offset / 8;
  int currBit = offset % 8;
will run identically as fast as
  src += offset >> 3;
  int currBit = offset & 7;
but using / and % is way more readable.

I would strongly encourage favoring the most readable version of any code you
can write.  This sort of speculation about speed is really best not speculated
on: profiling is the only good way to approach performance, and if we were
profiling this we'd see /8+%8 or >>3+&7 here is totally moot.

[emmaleer, 2015/08/12 22:19:20]

Okay, I've changed this back to use / and %.
Thanks for mentioning that! 

[scroggo, 2015/08/13 16:10:45]

Maybe we should use SkTDivMod then?

+    int currBit = offset % 8;

[scroggo, 2015/08/07 19:39:48]

On the other hand, if you also want to do a mod, I think SkTDivMod (in SkMath.h)
is faster.

[emmaleer, 2015/08/11 20:10:35]

I looked at SkTDivMod, and it does a divide /, not a shift. I think it needs to
do that to work for all numbers, not just powers of 2.
I think the shift is faster than this, so I'm going to use the shift.
Also do do the mod, SkTDivMod does offset - byteOffset*8, so I changed my code
to do the same so it's faster.

[scroggo, 2015/08/11 21:33:30]

Right, of course! The fact that you are using a power of two makes the mod call
easier, too, right? I think we can just do:

src += offset >> 3;
int currBit = offset & 7;

[emmaleer, 2015/08/12 14:04:33]

Acknowledged.

+
+    U8CPU currByte = src[0];

[scroggo, 2015/08/07 19:39:48]

nit: curByte is a value, while currBit is an index. Maybe currBit would be
better named bitIndex?

[emmaleer, 2015/08/11 20:10:35]

Acknowledged.

+    int sample = deltaSrc;
+
+    while (x < width) {
+        if (currBit == 8) {

[scroggo, 2015/08/07 19:39:48]

nit: constant should go on the left

[emmaleer, 2015/08/11 20:10:35]

Acknowledged.

+            src++;
+            currByte = src[0];

[scroggo, 2015/08/07 19:39:48]

nit: This works, but is confusing to read - you've mixed pointer arithmetic with
indices. I think this is better as either

currByte = ++src;

or do not increment src and use

currByte = src[<some index>];

[emmaleer, 2015/08/11 20:10:35]

I changed this to:
src++
currByte = *src

[scroggo, 2015/08/11 21:33:30]

FWIW, this is the same as:

currByte = ++src;

You may find the two step version more readable though, which I think is fine.

[emmaleer, 2015/08/12 14:04:33]

I don't think it's the same thing.
Since src is a pointer and currByte is an integer.
It would be the same as:
currByte = *(++src);
which I thought was more clear in 2 steps.

[scroggo, 2015/08/12 14:35:22]

Oops; yes, I need to dereference the pointer. I'm fine with doing it in two
steps though.

+            currBit = 0;
         }
-        dst += 8;
-    }
-
-    // Finish the remaining bits
-    width &= 7;
-    if (width > 0) {
-        U8CPU currByte = src[i];
-        for (int j = 0; j < width; j++) {
-            dst[j] = ((currByte >> 7) & 1) ? GRAYSCALE_WHITE : GRAYSCALE_BLACK;
-            currByte <<= 1;
+        if (deltaSrc == sample) {
+            dst[x] = ((currByte >> (7-currBit)) & 1) ? GRAYSCALE_WHITE : GRAYSCALE_BLACK;
+            sample = 0;
+            x++;
         }
+        sample++;
+        currBit++;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 #undef GRAYSCALE_BLACK
 #undef GRAYSCALE_WHITE
 
 static SkSwizzler::ResultAlpha swizzle_bit_to_index(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int /*bitsPerPixel*/, const SkPMColor* /*ctable*/) {
+        int deltaSrc, const SkPMColor* /*ctable*/, int offset) {
     uint8_t* SK_RESTRICT dst = (uint8_t*) dstRow;
 
-    // Determine how many full bytes are in the row
-    int bytesInRow = width >> 3;
-    int i;
-    for (i = 0; i < bytesInRow; i++) {
-        U8CPU currByte = src[i];
-        for (int j = 0; j < 8; j++) {
-            dst[j] = (currByte >> (7 - j)) & 1;
+    int x = 0;
+    // increment src by byte offset and currBit by bit offset
+    src += offset / 8;
+    int currBit = offset % 8;
+
+    U8CPU currByte = src[0];
+    int sample = deltaSrc;
+
+    while (x < width) {
+        if (currBit == 8) {
+            src++;
+            currByte = src[0];
+            currBit = 0;
         }
-        dst += 8;
-    }
-
-    // Finish the remaining bits
-    width &= 7;
-    if (width > 0) {
-        U8CPU currByte = src[i];
-        for (int j = 0; j < width; j++) {
-            dst[j] = ((currByte >> 7) & 1);
-            currByte <<= 1;
+        if (deltaSrc == sample) {
+            dst[x] = ((currByte >> (7-currBit)) & 1);

[scroggo, 2015/08/07 19:39:48]

Is this the only line that is different in the three functions? I wonder if
there is a way to share more code?

[emmaleer, 2015/08/11 20:10:35]

Yes this is the only line that's different
I thought the point of the swizzler was to go fast, and copied code didn't
matter?
If copied code does in fact matter, I could try to create a function for the
shared parts.
Although, calling a function each time may make it slightly slower.

[scroggo, 2015/08/11 21:33:30]

Yes, the point is to go fast, and I think it's okay that there's some
duplication. The bit and small index functions seem to have a bit more duplicate
work than the others though. (This isn't new to your change, I'll admit.)
I agree that calling a function may make it slower. We could also use a macro,
but that has its own downsides. Maybe just add some comments that these are the
same?

[emmaleer, 2015/08/12 14:04:33]

Acknowledged.

+            sample = 0;
+            x++;
         }
+        sample++;
+        currBit++;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
+
 static SkSwizzler::ResultAlpha swizzle_bit_to_n32(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int /*bitsPerPixel*/, const SkPMColor* /*ctable*/) {
+        int deltaSrc, const SkPMColor* /*ctable*/, int offset) {
     SkPMColor* SK_RESTRICT dst = (SkPMColor*) dstRow;
 
-    // Determine how many full bytes are in the row
-    int bytesInRow = width >> 3;
-    int i;
-    for (i = 0; i < bytesInRow; i++) {
-        U8CPU currByte = src[i];
-        for (int j = 0; j < 8; j++) {
-            dst[j] = ((currByte >> (7 - j)) & 1) ? SK_ColorWHITE : SK_ColorBLACK;
+    int x = 0;
+    // increment src by byte offset and currBit by bit offset
+    src += offset / 8;
+    int currBit = offset % 8;
+
+    U8CPU currByte = src[0];
+    int sample = deltaSrc;
+
+    while (x < width) {
+        if (currBit == 8) {
+            src++;
+            currByte = src[0];
+            currBit = 0;
         }
-        dst += 8;
-    }
-
-    // Finish the remaining bits
-    width &= 7;
-    if (width > 0) {
-        U8CPU currByte = src[i];
-        for (int j = 0; j < width; j++) {
-            dst[j] = ((currByte >> 7) & 1) ? SK_ColorWHITE : SK_ColorBLACK;
-            currByte <<= 1;
+        if (deltaSrc == sample) {
+            dst[x] = ((currByte >> (7 - currBit)) & 1) ? SK_ColorWHITE : SK_ColorBLACK;
+            sample = 0;
+            x++;
         }
+        sample++;
+        currBit++;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 // kIndex1, kIndex2, kIndex4
 
 static SkSwizzler::ResultAlpha swizzle_small_index_to_index(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bitsPerPixel, const SkPMColor ctable[]) {
+        int bitsPerPixel, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     uint8_t* SK_RESTRICT dst = (uint8_t*) dstRow;
     INIT_RESULT_ALPHA;
     const uint32_t pixelsPerByte = 8 / bitsPerPixel;
     const size_t rowBytes = compute_row_bytes_ppb(width, pixelsPerByte);
     const uint8_t mask = (1 << bitsPerPixel) - 1;
     int x = 0;
     for (uint32_t byte = 0; byte < rowBytes; byte++) {
         uint8_t pixelData = src[byte];
         for (uint32_t p = 0; p < pixelsPerByte && x < width; p++) {
             uint8_t index = (pixelData >> (8 - bitsPerPixel)) & mask;
             UPDATE_RESULT_ALPHA(ctable[index] >> SK_A32_SHIFT);
             dst[x] = index;
             pixelData <<= bitsPerPixel;
             x++;
         }
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 static SkSwizzler::ResultAlpha swizzle_small_index_to_n32(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bitsPerPixel, const SkPMColor ctable[]) {
+        int bitsPerPixel, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*) dstRow;
     INIT_RESULT_ALPHA;
     const uint32_t pixelsPerByte = 8 / bitsPerPixel;
     const size_t rowBytes = compute_row_bytes_ppb(width, pixelsPerByte);
     const uint8_t mask = (1 << bitsPerPixel) - 1;
     int x = 0;
     for (uint32_t byte = 0; byte < rowBytes; byte++) {
         uint8_t pixelData = src[byte];
         for (uint32_t p = 0; p < pixelsPerByte && x < width; p++) {
             uint8_t index = (pixelData >> (8 - bitsPerPixel)) & mask;
             SkPMColor c = ctable[index];
             UPDATE_RESULT_ALPHA(c >> SK_A32_SHIFT);
             dst[x] = c;
             pixelData <<= bitsPerPixel;
             x++;
         }
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 // kIndex
 
 static SkSwizzler::ResultAlpha swizzle_index_to_index(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     uint8_t* SK_RESTRICT dst = (uint8_t*) dstRow;
-    memcpy(dst, src, width);
+    if (1 == deltaSrc) {
+        memcpy(dst, src, width);
+    } else {
+        for (int x = 0; x < width; x++) {
+            dst[x] = src[0];
+            src += deltaSrc;
+        }
+    }
     // TODO (msarett): Should we skip the loop here and guess that the row is opaque/not opaque?
     //                 SkScaledBitmap sampler just guesses that it is opaque.  This is dangerous
     //                 and probably wrong since gif and bmp (rarely) may have alpha.
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
         UPDATE_RESULT_ALPHA(ctable[src[x]] >> SK_A32_SHIFT);
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 static SkSwizzler::ResultAlpha swizzle_index_to_n32(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
-        SkPMColor c = ctable[src[x]];
+        SkPMColor c = ctable[*src];
         UPDATE_RESULT_ALPHA(c >> SK_A32_SHIFT);
         dst[x] = c;
+        src += deltaSrc;
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 static SkSwizzler::ResultAlpha swizzle_index_to_n32_skipZ(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
-        SkPMColor c = ctable[src[x]];
+        SkPMColor c = ctable[*src];
         UPDATE_RESULT_ALPHA(c >> SK_A32_SHIFT);
         if (c != 0) {
             dst[x] = c;
         }
+        src += deltaSrc;
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 static SkSwizzler::ResultAlpha swizzle_index_to_565(
       void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-      int bytesPerPixel, const SkPMColor ctable[]) {
+      int bytesPerPixel, const SkPMColor ctable[], int offset) {
     // FIXME: Support dithering? Requires knowing y, which I think is a bigger
     // change.
+    src += offset;
     uint16_t* SK_RESTRICT dst = (uint16_t*)dstRow;
     for (int x = 0; x < width; x++) {
         dst[x] = SkPixel32ToPixel16(ctable[*src]);
         src += bytesPerPixel;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 
 #undef A32_MASK_IN_PLACE
 
 // kGray
 
 static SkSwizzler::ResultAlpha swizzle_gray_to_n32(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     for (int x = 0; x < width; x++) {
-        dst[x] = SkPackARGB32NoCheck(0xFF, src[x], src[x], src[x]);
+        dst[x] = SkPackARGB32NoCheck(0xFF, *src, *src, *src);
+        src += deltaSrc;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 static SkSwizzler::ResultAlpha swizzle_gray_to_gray(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
-    memcpy(dstRow, src, width);
+        int deltaSrc, const SkPMColor ctable[], int offset) {
+
+    src += offset;
+    uint8_t* SK_RESTRICT dst = (uint8_t*) dstRow;
+    if (1 == deltaSrc) {
+        memcpy(dstRow, src, width);
+    } else {
+        for (int x = 0; x < width; x++) {
+            dst[x] = src[0];
+            src += deltaSrc;
+        }
+    }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 static SkSwizzler::ResultAlpha swizzle_gray_to_565(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int bytesPerPixel, const SkPMColor ctable[], int offset) {
     // FIXME: Support dithering?
+    src += offset;
     uint16_t* SK_RESTRICT dst = (uint16_t*)dstRow;
     for (int x = 0; x < width; x++) {
         dst[x] = SkPack888ToRGB16(src[0], src[0], src[0]);
         src += bytesPerPixel;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 // kBGRX
 
 static SkSwizzler::ResultAlpha swizzle_bgrx_to_n32(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     for (int x = 0; x < width; x++) {
         dst[x] = SkPackARGB32NoCheck(0xFF, src[2], src[1], src[0]);
-        src += bytesPerPixel;
+        src += deltaSrc;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 // kBGRA
 
 static SkSwizzler::ResultAlpha swizzle_bgra_to_n32_unpremul(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
         uint8_t alpha = src[3];
         UPDATE_RESULT_ALPHA(alpha);
         dst[x] = SkPackARGB32NoCheck(alpha, src[2], src[1], src[0]);
-        src += bytesPerPixel;
+        src += deltaSrc;
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 static SkSwizzler::ResultAlpha swizzle_bgra_to_n32_premul(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
         uint8_t alpha = src[3];
         UPDATE_RESULT_ALPHA(alpha);
         dst[x] = SkPreMultiplyARGB(alpha, src[2], src[1], src[0]);
-        src += bytesPerPixel;
+        src += deltaSrc;
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 // kRGBX
 static SkSwizzler::ResultAlpha swizzle_rgbx_to_n32(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     for (int x = 0; x < width; x++) {
         dst[x] = SkPackARGB32(0xFF, src[0], src[1], src[2]);
-        src += bytesPerPixel;
+        src += deltaSrc;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 static SkSwizzler::ResultAlpha swizzle_rgbx_to_565(
        void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-       int bytesPerPixel, const SkPMColor ctable[]) {
+       int bytesPerPixel, const SkPMColor ctable[], int offset) {
     // FIXME: Support dithering?
+    src += offset;
     uint16_t* SK_RESTRICT dst = (uint16_t*)dstRow;
     for (int x = 0; x < width; x++) {
         dst[x] = SkPack888ToRGB16(src[0], src[1], src[2]);
         src += bytesPerPixel;
     }
     return SkSwizzler::kOpaque_ResultAlpha;
 }
 
 
 // kRGBA
 static SkSwizzler::ResultAlpha swizzle_rgba_to_n32_premul(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
         unsigned alpha = src[3];
         UPDATE_RESULT_ALPHA(alpha);
         dst[x] = SkPreMultiplyARGB(alpha, src[0], src[1], src[2]);
-        src += bytesPerPixel;
+        src += deltaSrc;
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 static SkSwizzler::ResultAlpha swizzle_rgba_to_n32_unpremul(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     uint32_t* SK_RESTRICT dst = reinterpret_cast<uint32_t*>(dstRow);
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
         unsigned alpha = src[3];
         UPDATE_RESULT_ALPHA(alpha);
         dst[x] = SkPackARGB32NoCheck(alpha, src[0], src[1], src[2]);
-        src += bytesPerPixel;
+        src += deltaSrc;
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 static SkSwizzler::ResultAlpha swizzle_rgba_to_n32_premul_skipZ(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int width,
-        int bytesPerPixel, const SkPMColor ctable[]) {
+        int deltaSrc, const SkPMColor ctable[], int offset) {
 
+    src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     INIT_RESULT_ALPHA;
     for (int x = 0; x < width; x++) {
         unsigned alpha = src[3];
         UPDATE_RESULT_ALPHA(alpha);
         if (0 != alpha) {
             dst[x] = SkPreMultiplyARGB(alpha, src[0], src[1], src[2]);
         }
-        src += bytesPerPixel;
+        src += deltaSrc;
     }
     return COMPUTE_RESULT_ALPHA;
 }
 
 /**
     FIXME: This was my idea to cheat in order to continue taking advantage of skipping zeroes.
     This would be fine for drawing normally, but not for drawing with transfer modes. Being
     honest means we can draw correctly with transfer modes, with the cost of not being able
     to take advantage of Android's free unwritten pages. Something to keep in mind when we
     decide whether to switch to unpremul default.
@@ skipping 13 matching lines @@
         }
         src += deltaSrc;
         alphaMask &= alpha;
     }
     return alphaMask != 0xFF;
 }
 */
 
 SkSwizzler* SkSwizzler::CreateSwizzler(SkSwizzler::SrcConfig sc,
                                        const SkPMColor* ctable,
-                                       const SkImageInfo& info,
-                                       SkCodec::ZeroInitialized zeroInit) {
-    if (info.colorType() == kUnknown_SkColorType || kUnknown == sc) {
+                                       const SkImageInfo& dstInfo, 
+                                       SkCodec::ZeroInitialized zeroInit, 
+                                       int srcWidth) {
+    if (dstInfo.colorType() == kUnknown_SkColorType || kUnknown == sc) {
         return NULL;
     }
     if ((kIndex == sc || kIndex4 == sc || kIndex2 == sc || kIndex1 == sc)
             && NULL == ctable) {
         return NULL;
     }
     RowProc proc = NULL;
+
     switch (sc) {
         case kBit:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     proc = &swizzle_bit_to_n32;
                     break;
                 case kIndex_8_SkColorType:
                     proc = &swizzle_bit_to_index;
                     break;
                 case kGray_8_SkColorType:
                     proc = &swizzle_bit_to_grayscale;
                     break;
                 default:
                     break;
             }
             break;
         case kIndex1:
         case kIndex2:
         case kIndex4:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     proc = &swizzle_small_index_to_n32;
                     break;
                 case kIndex_8_SkColorType:
                     proc = &swizzle_small_index_to_index;
                     break;
                 default:
                     break;
             }
             break;
         case kIndex:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     // We assume the color premultiplied ctable (or not) as desired.
                     if (SkCodec::kYes_ZeroInitialized == zeroInit) {
                         proc = &swizzle_index_to_n32_skipZ;
                         break;
                     } else {
                         proc = &swizzle_index_to_n32;
                         break;
                     }
                     break;
                 case kRGB_565_SkColorType:
                     proc = &swizzle_index_to_565;
                     break;
                 case kIndex_8_SkColorType:
                     proc = &swizzle_index_to_index;
                     break;
                 default:
                     break;
             }
             break;
         case kGray:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     proc = &swizzle_gray_to_n32;
                     break;
                 case kGray_8_SkColorType:
                     proc = &swizzle_gray_to_gray;
                     break;
                 case kRGB_565_SkColorType:
                     proc = &swizzle_gray_to_565;
                     break;
                 default:
                     break;
             }
             break;
         case kBGR:
         case kBGRX:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     proc = &swizzle_bgrx_to_n32;
                     break;
                 default:
                     break;
             }
             break;
         case kBGRA:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
-                    switch (info.alphaType()) {
+                    switch (dstInfo.alphaType()) {
                         case kUnpremul_SkAlphaType:
                             proc = &swizzle_bgra_to_n32_unpremul;
                             break;
                         case kPremul_SkAlphaType:
                             proc = &swizzle_bgra_to_n32_premul;
                             break;
                         default:
                             break;
                     }
                     break;
                 default:
                     break;
             }
             break;
         case kRGBX:
             // TODO: Support other swizzles.
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     proc = &swizzle_rgbx_to_n32;
                     break;
                 case kRGB_565_SkColorType:
                     proc = &swizzle_rgbx_to_565;
                 default:
                     break;
             }
             break;
         case kRGBA:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
-                    if (info.alphaType() == kUnpremul_SkAlphaType) {
+                    if (dstInfo.alphaType() == kUnpremul_SkAlphaType) {
                         // Respect zeroInit?
                         proc = &swizzle_rgba_to_n32_unpremul;
                     } else {
                         if (SkCodec::kYes_ZeroInitialized == zeroInit) {
                             proc = &swizzle_rgba_to_n32_premul_skipZ;
                         } else {
                             proc = &swizzle_rgba_to_n32_premul;
                         }
                     }
                     break;
                 default:
                     break;
             }
             break;
         case kRGB:
-            switch (info.colorType()) {
+            switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     proc = &swizzle_rgbx_to_n32;
                     break;
                 default:
                     break;
             }
             break;
+        case kRGB_565:
+            switch (dstInfo.colorType()) {
+                case kRGB_565_SkColorType:
+                    proc = &sample565;
+                    break;
+                default:
+                    break;
+            }
         default:
             break;
     }
     if (NULL == proc) {
         return NULL;
     }
 
     // Store deltaSrc in bytes if it is an even multiple, otherwise use bits
-    int deltaSrc = SkIsAlign8(BitsPerPixel(sc)) ? BytesPerPixel(sc) :
-            BitsPerPixel(sc);
-    return SkNEW_ARGS(SkSwizzler, (proc, ctable, deltaSrc, info));
+    int deltaSrc = SkIsAlign8(BitsPerPixel(sc)) ? BytesPerPixel(sc) : BitsPerPixel(sc);
+
+    int sampleX = SkScaledCodec::GetSampleSize(srcWidth, dstInfo.width());
+    
+    return SkNEW_ARGS(SkSwizzler, (proc, ctable, deltaSrc, dstInfo, sampleX));
 }
 
 SkSwizzler::SkSwizzler(RowProc proc, const SkPMColor* ctable,
-                       int deltaSrc, const SkImageInfo& info)
+                       int deltaSrc, const SkImageInfo& info, int sampleX)
     : fRowProc(proc)
     , fColorTable(ctable)
     , fDeltaSrc(deltaSrc)
     , fDstInfo(info)
-{}
+    , fSampleX(sampleX)
+    , fX0(sampleX == 1 ? 0 : sampleX >> 1)
+{
+    // check that fX0 is less than original width
+    SkASSERT(fX0 >= 0 && fX0 < fDstInfo.width() * fSampleX);
+}
 
 SkSwizzler::ResultAlpha SkSwizzler::swizzle(void* dst, const uint8_t* SK_RESTRICT src) {
     SkASSERT(NULL != dst && NULL != src);
-    return fRowProc(dst, src, fDstInfo.width(), fDeltaSrc, fColorTable);
+    return fRowProc(dst, src, fDstInfo.width(), fSampleX * fDeltaSrc,
+                    fColorTable, fX0 * fDeltaSrc);
 }
 
 void SkSwizzler::Fill(void* dstStartRow, const SkImageInfo& dstInfo, size_t dstRowBytes,
         uint32_t numRows, uint32_t colorOrIndex, const SkPMColor* colorTable) {
     SkASSERT(dstStartRow != NULL);
     SkASSERT(numRows <= (uint32_t) dstInfo.height());
 
     // Calculate bytes to fill.  We use getSafeSize since the last row may not be padded.
     const size_t bytesToFill = dstInfo.makeWH(dstInfo.width(), numRows).getSafeSize(dstRowBytes);
 
@@ skipping 48 matching lines @@
             // bits of SK_ColorBLACK are identical to the 565 representation
             // for black.
             memset(dstStartRow, (uint16_t) colorOrIndex, bytesToFill);
             break;
         default:
             SkCodecPrintf("Error: Unsupported dst color type for fill().  Doing nothing.\n");
             SkASSERT(false);
             break;
     }
 }