Skip memcpy() swizzles in SkPngCodec

src/codec/SkPngCodec.cpp

Index: src/codec/SkPngCodec.cpp
diff --git a/src/codec/SkPngCodec.cpp b/src/codec/SkPngCodec.cpp
index 232373c04407c63c2b391c148a7053db9e9d62e9..797178948fba8c6d05b07d96fabbb6e5a49811c8 100644
--- a/src/codec/SkPngCodec.cpp
+++ b/src/codec/SkPngCodec.cpp
@@ -426,6 +426,11 @@ SkCodec::Result SkPngCodec::initializeSwizzler(const SkImageInfo& requestedInfo,
             SkASSERT(false);
     }
 
+    // If we are not performing a subset decode, and the swizzle will not change the
+    // number of bytes per pixel, we can swizzle in place.
+    fInPlaceSwizzle = (SkSwizzler::BytesPerPixel(fSrcConfig) ==
+            SkColorTypeBytesPerPixel(requestedInfo.colorType()) && !options.fSubset);
+
     // Copy the color table to the client if they request kIndex8 mode
     copy_color_table(requestedInfo, fColorTable, ctable, ctableCount);
 
@@ -475,6 +480,11 @@ SkCodec::Result SkPngCodec::onGetPixels(const SkImageInfo& requestedInfo, void*
     if (result != kSuccess) {
         return result;
     }
+
+    const int width = requestedInfo.width();
+    const int height = requestedInfo.height();
+    const int bpp = SkSwizzler::BytesPerPixel(fSrcConfig);
+
     // FIXME: Could we use the return value of setjmp to specify the type of
     // error?
     int row = 0;
@@ -484,7 +494,7 @@ SkCodec::Result SkPngCodec::onGetPixels(const SkImageInfo& requestedInfo, void*
         // Assume that any error that occurs while reading rows is caused by an incomplete input.
         if (fNumberPasses > 1) {
             // FIXME (msarett): Handle incomplete interlaced pngs.
-            return kInvalidInput;
+            return row == height ? kSuccess : kInvalidInput;
         }
         // FIXME: We do a poor job on incomplete pngs compared to other decoders (ex: Chromium,
         // Ubuntu Image Viewer).  This is because we use the default buffer size in libpng (8192
@@ -498,25 +508,20 @@ SkCodec::Result SkPngCodec::onGetPixels(const SkImageInfo& requestedInfo, void*
         // it have on regular decode performance?  Should we investigate using a different API
         // instead of png_read_row(s)?  Chromium uses png_process_data.
         *rowsDecoded = row;
-        return kIncompleteInput;
+        return row == height ? kSuccess : kIncompleteInput;
     }
 
     // FIXME: We could split these out based on subclass.
     void* dstRow = dst;
     if (fNumberPasses > 1) {
-        const int width = requestedInfo.width();
-        const int height = requestedInfo.height();
-        const int bpp = SkSwizzler::BytesPerPixel(fSrcConfig);
-        const size_t srcRowBytes = width * bpp;
-
-        storage.reset(width * height * bpp);
-        uint8_t* const base = storage.get();
+        uint8_t* const base = fInPlaceSwizzle ? (uint8_t*) dstRow :
+                storage.reset(width * height * bpp);
+        const size_t srcRowBytes = fInPlaceSwizzle ? dstRowBytes : width * bpp;
 
         for (int i = 0; i < fNumberPasses; i++) {
             uint8_t* srcRow = base;
             for (int y = 0; y < height; y++) {
-                uint8_t* bmRow = srcRow;
-                png_read_rows(fPng_ptr, &bmRow, nullptr, 1);
+                png_read_rows(fPng_ptr, &srcRow, nullptr, 1);
                 srcRow += srcRowBytes;
             }
         }
@@ -529,21 +534,16 @@ SkCodec::Result SkPngCodec::onGetPixels(const SkImageInfo& requestedInfo, void*
             srcRow += srcRowBytes;
         }
     } else {
-        storage.reset(requestedInfo.width() * SkSwizzler::BytesPerPixel(fSrcConfig));
-        uint8_t* srcRow = storage.get();
+        uint8_t* srcRow = fInPlaceSwizzle ? (uint8_t*) dstRow : storage.reset(width * bpp);
+        const size_t srcRowBytes = fInPlaceSwizzle ? dstRowBytes : 0;
         for (; row < requestedInfo.height(); row++) {
             png_read_rows(fPng_ptr, &srcRow, nullptr, 1);
-            // FIXME: Only call IsOpaque once, outside the loop. Same for onGetScanlines.
             fSwizzler->swizzle(dstRow, srcRow);
             dstRow = SkTAddOffset<void>(dstRow, dstRowBytes);
+            srcRow += srcRowBytes;
         }
     }
 
-    if (setjmp(png_jmpbuf(fPng_ptr))) {
-        // We've already read all the scanlines. This is a success.
-        return kSuccess;
-    }
-
     // read rest of file, and get additional comment and time chunks in info_ptr
     png_read_end(fPng_ptr, fInfo_ptr);
 
@@ -595,8 +595,9 @@ public:
 
         void* dstRow = dst;
         for (; row < count; row++) {
-            png_read_rows(this->png_ptr(), &fSrcRow, nullptr, 1);
-            this->swizzler()->swizzle(dstRow, fSrcRow);
+            uint8_t* srcRow = this->inPlaceSwizzle() ? (uint8_t*) dstRow : fSrcRow;
+            png_read_rows(this->png_ptr(), &srcRow, nullptr, 1);
+            this->swizzler()->swizzle(dstRow, srcRow);
             dstRow = SkTAddOffset<void>(dstRow, rowBytes);
         }
 
@@ -690,8 +691,10 @@ public:
             // fail on the first pass, we can still report than some scanlines are initialized.
             return 0;
         }
-        SkAutoTMalloc<uint8_t> storage(count * fSrcRowBytes);
-        uint8_t* storagePtr = storage.get();
+        SkAutoTMalloc<uint8_t> storage;
+        uint8_t* const base = this->inPlaceSwizzle() ? (uint8_t*) dst :
+                storage.reset(count * fSrcRowBytes);
+        const size_t srcRowBytes = this->inPlaceSwizzle() ? dstRowBytes : fSrcRowBytes;
         uint8_t* srcRow;
         const int startRow = this->nextScanline();
         for (int i = 0; i < this->numberPasses(); i++) {
@@ -700,10 +703,10 @@ public:
                 png_read_rows(this->png_ptr(), &fGarbageRowPtr, nullptr, 1);
             }
             // read rows we care about into buffer
-            srcRow = storagePtr;
+            uint8_t* srcRow = base;
             for (int y = 0; y < count; y++) {
                 png_read_rows(this->png_ptr(), &srcRow, nullptr, 1);
-                srcRow += fSrcRowBytes;
+                srcRow += srcRowBytes;
             }
             // read rows we don't want into garbage buffer
             for (int y = 0; y < fHeight - startRow - count; y++) {
@@ -711,12 +714,12 @@ public:
             }
         }
         //swizzle the rows we care about
-        srcRow = storagePtr;
+        srcRow = base;
         void* dstRow = dst;
         for (int y = 0; y < count; y++) {
             this->swizzler()->swizzle(dstRow, srcRow);
             dstRow = SkTAddOffset<void>(dstRow, dstRowBytes);
-            srcRow += fSrcRowBytes;
+            srcRow += srcRowBytes;
         }
 
         return count;