Skip memcpy() swizzles in SkPngCodec

src/codec/SkPngCodec.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 "SkColorTable.h"
@@ skipping 355 matching lines @@
         png_destroy_read_struct(&fPng_ptr, &fInfo_ptr, nullptr);
         fPng_ptr = nullptr;
         fInfo_ptr = nullptr;
     }
 }
 
 ///////////////////////////////////////////////////////////////////////////////
 // Getting the pixels
 ///////////////////////////////////////////////////////////////////////////////
 
-SkCodec::Result SkPngCodec::initializeSwizzler(const SkImageInfo& requestedInfo,
-                                               const Options& options,
-                                               SkPMColor ctable[],
-                                               int* ctableCount) {
+void SkPngCodec::initializeSwizzler(const SkImageInfo& requestedInfo, const Options& options) {
+    const SkPMColor* colors = get_color_ptr(fColorTable.get());
+    fSwizzler.reset(SkSwizzler::CreateSwizzler(fSrcConfig, colors, requestedInfo, options));
+    SkASSERT(fSwizzler);
+}
+
+SkCodec::Result SkPngCodec::prepareToDecode(const SkImageInfo& requestedInfo,
+                                            const Options& options,
+                                            SkPMColor ctable[],
+                                            int* ctableCount) {
     // FIXME: Could we use the return value of setjmp to specify the type of
     // error?
     if (setjmp(png_jmpbuf(fPng_ptr))) {
         SkCodecPrintf("setjmp long jump!\n");
         return kInvalidInput;
     }
     png_read_update_info(fPng_ptr, fInfo_ptr);
 
     // suggestedColorType was determined in read_header() based on the encodedColorType
     const SkColorType suggestedColorType = this->getInfo().colorType();
 
+    // If the conversion provided by the swizzler would be a no-op, we may be able
+    // to skip the swizzle step.
+    bool skipSwizzle = false;
+
     switch (suggestedColorType) {
         case kIndex_8_SkColorType:
-            //decode palette to Skia format
             fSrcConfig = SkSwizzler::kIndex;
+            skipSwizzle = (kIndex_8_SkColorType == requestedInfo.colorType());
+
+            // Decode palette to Skia format
             if (!this->decodePalette(kPremul_SkAlphaType == requestedInfo.alphaType(),
                     ctableCount)) {
                 return kInvalidInput;
             }
             break;
         case kGray_8_SkColorType:
             fSrcConfig = SkSwizzler::kGray;
+            skipSwizzle = (kGray_8_SkColorType == requestedInfo.colorType());
             break;
         case kN32_SkColorType: {
             const uint8_t encodedColorType = png_get_color_type(fPng_ptr, fInfo_ptr);
             if (PNG_COLOR_TYPE_GRAY_ALPHA == encodedColorType ||
                     PNG_COLOR_TYPE_GRAY == encodedColorType) {
                 // If encodedColorType is GRAY, there must be a transparent chunk.
                 // Otherwise, suggestedColorType would be kGray.  We have already
                 // instructed libpng to convert the transparent chunk to alpha,
                 // so we can treat both GRAY and GRAY_ALPHA as kGrayAlpha.
                 SkASSERT(encodedColorType == PNG_COLOR_TYPE_GRAY_ALPHA ||
                         png_get_valid(fPng_ptr, fInfo_ptr, PNG_INFO_tRNS));
 
                 fSrcConfig = SkSwizzler::kGrayAlpha;
             } else {
                 if (this->getInfo().alphaType() == kOpaque_SkAlphaType) {
                     fSrcConfig = SkSwizzler::kRGB;
                 } else {
                     fSrcConfig = SkSwizzler::kRGBA;
+#ifdef SK_PMCOLOR_IS_RGBA
+                    skipSwizzle = (kUnpremul_SkAlphaType == this->getInfo().alphaType());
+#endif
                 }
             }
             break;
         }
         default:
             // We will always recommend one of the above colorTypes.
             SkASSERT(false);
     }
 
     // Copy the color table to the client if they request kIndex8 mode
     copy_color_table(requestedInfo, fColorTable, ctable, ctableCount);
 
+    // If this is a subset decode, we will need a swizzler.
+    if (options.fSubset) {
+        skipSwizzle = false;
+    }
+
     // Create the swizzler.  SkPngCodec retains ownership of the color table.
-    const SkPMColor* colors = get_color_ptr(fColorTable.get());
-    fSwizzler.reset(SkSwizzler::CreateSwizzler(fSrcConfig, colors, requestedInfo, options));
-    SkASSERT(fSwizzler);
+    if (!skipSwizzle) {
+        this->initializeSwizzler(requestedInfo, options);
+    }
 
     return kSuccess;
 }
 
+SkSampler* SkPngCodec::getSampler(bool createIfNecessary) {
+    if (!createIfNecessary || fSwizzler) {
+        return fSwizzler;
+    }
+
+    this->initializeSwizzler(this->dstInfo(), this->options());
+    return fSwizzler;
+}
 
 bool SkPngCodec::onRewind() {
     // This sets fPng_ptr and fInfo_ptr to nullptr. If read_header
     // succeeds, they will be repopulated, and if it fails, they will
     // remain nullptr. Any future accesses to fPng_ptr and fInfo_ptr will
     // come through this function which will rewind and again attempt
     // to reinitialize them.
     this->destroyReadStruct();
 
     png_structp png_ptr;
@@ skipping 14 matching lines @@
                                         int* rowsDecoded) {
     if (!conversion_possible(requestedInfo, this->getInfo())) {
         return kInvalidConversion;
     }
     if (options.fSubset) {
         // Subsets are not supported.
         return kUnimplemented;
     }
 
     // Note that ctable and ctableCount may be modified if there is a color table
-    const Result result = this->initializeSwizzler(requestedInfo, options, ctable, ctableCount);
+    const Result result = this->prepareToDecode(requestedInfo, options, ctable, ctableCount);
     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;
     // This must be declared above the call to setjmp to avoid memory leaks on incomplete images.
     SkAutoTMalloc<uint8_t> storage;
     if (setjmp(png_jmpbuf(fPng_ptr))) {
         // 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
         // bytes), and if we can't fill the buffer, we immediately fail.
         // For example, if we try to read 8192 bytes, and the image (incorrectly) only contains
         // half that, which may have been enough to contain a non-zero number of lines, we fail
         // when we could have decoded a few more lines and then failed.
         // The read function that we provide for libpng has no way of indicating that we have
         // made a partial read.
         // Making our buffer size smaller improves our incomplete decodes, but what impact does
         // 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;
+        const size_t rowBytes = fSwizzler ? srcRowBytes : dstRowBytes;
+        uint8_t* const base = fSwizzler ? storage.reset(width * height * bpp) : (uint8_t*) dst;
+        for (int i = 0; i < fNumberPasses; i++) {
+            uint8_t* rowPtr = base;
+            for (int y = 0; y < height; y++) {
+                png_read_rows(fPng_ptr, &rowPtr, nullptr, 1);
+                rowPtr += rowBytes;
+            }
+        }
 
-        storage.reset(width * height * bpp);
-        uint8_t* const base = storage.get();
-
-        for (int i = 0; i < fNumberPasses; i++) {
+        // Swizzle if necessary.
+        if (fSwizzler) {
             uint8_t* srcRow = base;
             for (int y = 0; y < height; y++) {
-                uint8_t* bmRow = srcRow;
-                png_read_rows(fPng_ptr, &bmRow, nullptr, 1);
+                fSwizzler->swizzle(dst, srcRow);
+                dst = SkTAddOffset<void>(dst, dstRowBytes);
                 srcRow += srcRowBytes;
             }
         }
+    } else {
+        uint8_t* rowPtr = fSwizzler ? storage.reset(width * bpp) : (uint8_t*) dst;
+        for (; row < requestedInfo.height(); row++) {
+            png_read_rows(fPng_ptr, &rowPtr, nullptr, 1);

[scroggo, 2016/02/05 21:36:52]

if (!swizzler), would it make sense to call this with an array of row pointers
and height for the last parameter, instead of looping?

[msarett, 2016/02/05 22:14:16]

This looks like it makes things another half percent faster...

 
-        // Now swizzle it.
-        uint8_t* srcRow = base;
-        for (int y = 0; y < height; y++) {
-            fSwizzler->swizzle(dstRow, srcRow);
-            dstRow = SkTAddOffset<void>(dstRow, dstRowBytes);
-            srcRow += srcRowBytes;
-        }
-    } else {
-        storage.reset(requestedInfo.width() * SkSwizzler::BytesPerPixel(fSrcConfig));
-        uint8_t* srcRow = storage.get();
-        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);
+            if (fSwizzler) {
+                fSwizzler->swizzle(dst, rowPtr);
+                dst = SkTAddOffset<void>(dst, dstRowBytes);
+            } else {
+                rowPtr += dstRowBytes;
+            }
         }
     }
 
-    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);
 
     return kSuccess;
 }
 
 uint32_t SkPngCodec::onGetFillValue(SkColorType colorType) const {
     const SkPMColor* colorPtr = get_color_ptr(fColorTable.get());
     if (colorPtr) {
         return get_color_table_fill_value(colorType, colorPtr, 0);
     }
     return INHERITED::onGetFillValue(colorType);
 }
 
 // Subclass of SkPngCodec which supports scanline decoding
 class SkPngScanlineDecoder : public SkPngCodec {
 public:
     SkPngScanlineDecoder(const SkImageInfo& srcInfo, SkStream* stream,
             SkPngChunkReader* chunkReader, png_structp png_ptr, png_infop info_ptr, int bitDepth)
         : INHERITED(srcInfo, stream, chunkReader, png_ptr, info_ptr, bitDepth, 1)
         , fSrcRow(nullptr)
     {}
 
     Result onStartScanlineDecode(const SkImageInfo& dstInfo, const Options& options,
             SkPMColor ctable[], int* ctableCount) override {
         if (!conversion_possible(dstInfo, this->getInfo())) {
             return kInvalidConversion;
         }
 
-        const Result result = this->initializeSwizzler(dstInfo, options, ctable,
-                                                       ctableCount);
+        const Result result = this->prepareToDecode(dstInfo, options, ctable, ctableCount);
         if (result != kSuccess) {
             return result;
         }
 
         fStorage.reset(this->getInfo().width() * SkSwizzler::BytesPerPixel(this->srcConfig()));
         fSrcRow = fStorage.get();
 
         return kSuccess;
     }
 
     int onGetScanlines(void* dst, int count, size_t rowBytes) override {
         // Assume that an error in libpng indicates an incomplete input.
         int row = 0;
         if (setjmp(png_jmpbuf(this->png_ptr()))) {
             SkCodecPrintf("setjmp long jump!\n");
             return row;
         }
 
-        void* dstRow = dst;
+        uint8_t* rowPtr;
+        if (this->swizzler()) {
+            rowPtr = fSrcRow;
+        } else {
+            // Write decoded pixels directly to dst.
+            rowPtr = (uint8_t*) dst;
+        }
+
         for (; row < count; row++) {
-            png_read_rows(this->png_ptr(), &fSrcRow, nullptr, 1);
-            this->swizzler()->swizzle(dstRow, fSrcRow);
-            dstRow = SkTAddOffset<void>(dstRow, rowBytes);
+            png_read_rows(this->png_ptr(), &rowPtr, nullptr, 1);

[scroggo, 2016/02/05 21:36:52]

Again, instead of looping when there's no swizzler, would it make sense to call
this once with an array of row pointers?

I guess this means a separate code path and we still need to set up the
pointers.

+
+            if (this->swizzler()) {
+                this->swizzler()->swizzle(dst, rowPtr);
+                dst = SkTAddOffset<void>(dst, rowBytes);
+            } else {
+                rowPtr += rowBytes;
+            }
         }
 
         return row;
     }
 
     bool onSkipScanlines(int count) override {
         // Assume that an error in libpng indicates an incomplete input.
         if (setjmp(png_jmpbuf(this->png_ptr()))) {
             SkCodecPrintf("setjmp long jump!\n");
             return false;
@@ skipping 26 matching lines @@
     {
         SkASSERT(numberPasses != 1);
     }
 
     Result onStartScanlineDecode(const SkImageInfo& dstInfo, const Options& options,
             SkPMColor ctable[], int* ctableCount) override {
         if (!conversion_possible(dstInfo, this->getInfo())) {
             return kInvalidConversion;
         }
 
-        const Result result = this->initializeSwizzler(dstInfo, options, ctable,
-                                                       ctableCount);
+        const Result result = this->prepareToDecode(dstInfo, options, ctable, ctableCount);
         if (result != kSuccess) {
             return result;
         }
 
         fHeight = dstInfo.height();
         // FIXME: This need not be called on a second call to onStartScanlineDecode.
         fSrcRowBytes = this->getInfo().width() * SkSwizzler::BytesPerPixel(this->srcConfig());
         fGarbageRow.reset(fSrcRowBytes);
         fGarbageRowPtr = static_cast<uint8_t*>(fGarbageRow.get());
         fCanSkipRewind = true;
@@ skipping 24 matching lines @@
             this->updateCurrScanline(currScanline);
         }
 
         if (setjmp(png_jmpbuf(this->png_ptr()))) {
             SkCodecPrintf("setjmp long jump!\n");
             // FIXME (msarett): Returning 0 is pessimistic.  If we can complete a single pass,
             // we may be able to report that all of the memory has been initialized.  Even if we
             // 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();
-        uint8_t* srcRow;
+        SkAutoTMalloc<uint8_t> storage;
+        uint8_t* const base = this->swizzler() ?
+                storage.reset(count * fSrcRowBytes) : (uint8_t*) dst;
+        const size_t rowBytes = this->swizzler() ? fSrcRowBytes : dstRowBytes;
         const int startRow = this->nextScanline();
         for (int i = 0; i < this->numberPasses(); i++) {
             // read rows we planned to skip into garbage row
             for (int y = 0; y < startRow; y++){
                 png_read_rows(this->png_ptr(), &fGarbageRowPtr, nullptr, 1);
             }
             // read rows we care about into buffer
-            srcRow = storagePtr;
+            uint8_t* rowPtr = base;
             for (int y = 0; y < count; y++) {
-                png_read_rows(this->png_ptr(), &srcRow, nullptr, 1);
-                srcRow += fSrcRowBytes;
+                png_read_rows(this->png_ptr(), &rowPtr, nullptr, 1);
+                rowPtr += rowBytes;
             }
             // read rows we don't want into garbage buffer
             for (int y = 0; y < fHeight - startRow - count; y++) {
                 png_read_rows(this->png_ptr(), &fGarbageRowPtr, nullptr, 1);
             }
         }
-        //swizzle the rows we care about
-        srcRow = storagePtr;
-        void* dstRow = dst;
-        for (int y = 0; y < count; y++) {
-            this->swizzler()->swizzle(dstRow, srcRow);
-            dstRow = SkTAddOffset<void>(dstRow, dstRowBytes);
-            srcRow += fSrcRowBytes;
+
+        // Swizzle if necessary
+        if (this->swizzler()) {
+            uint8_t* srcRow = storage.get();
+            void* dstRow = dst;
+            for (int y = 0; y < count; y++) {
+                this->swizzler()->swizzle(dstRow, srcRow);
+                dstRow = SkTAddOffset<void>(dstRow, dstRowBytes);
+                srcRow += fSrcRowBytes;
+            }
         }
 
         return count;
     }
 
     bool onSkipScanlines(int count) override {
         // The non-virtual version will update fCurrScanline.
         return true;
     }
 
@@ skipping 33 matching lines @@
     }
 
     if (1 == numberPasses) {
         return new SkPngScanlineDecoder(imageInfo, streamDeleter.detach(), chunkReader,
                                         png_ptr, info_ptr, bitDepth);
     }
 
     return new SkPngInterlacedScanlineDecoder(imageInfo, streamDeleter.detach(), chunkReader,
                                               png_ptr, info_ptr, bitDepth, numberPasses);
 }