Subset decoding benchmarks

bench/CodecSubsetBench.cpp

Index: bench/CodecSubsetBench.cpp
diff --git a/bench/CodecSubsetBench.cpp b/bench/CodecSubsetBench.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..188880548a4e2ab4f87e3b21c3d71b0216d85a9c
--- /dev/null
+++ b/bench/CodecSubsetBench.cpp
@@ -0,0 +1,340 @@
+/*
+ * 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 "CodecSubsetBench.h"
+#include "SkData.h"
+#include "SkCodec.h"
+#include "SkImageDecoder.h"
+#include "SkOSFile.h"
+#include "SkStream.h"
+
+/*
+ *
+ * This benchmark is designed to test the performance of subset decoding.
+ *
+ */
+CodecSubsetBench::CodecSubsetBench(SkString* path,
+                                   SubsetMode subsetMode,
+                                   SkColorType colorType,
+                                   uint32_t subsetWidth,
+                                   uint32_t subsetHeight,
+                                   uint32_t offsetLeft,
+                                   uint32_t offsetTop,
+                                   uint32_t divisor,
+                                   bool useCodec)
+    : fSubsetMode(subsetMode)
+    , fColorType(colorType)
+    , fSubsetWidth(subsetWidth)
+    , fSubsetHeight(subsetHeight)
+    , fOffsetLeft(offsetLeft)
+    , fOffsetTop(offsetTop)
+    , fDivisor(divisor)
+    , fUseCodec(useCodec)
+{
+    // Parse the filename
+    SkString baseName = SkOSPath::Basename(path->c_str());
+
+    // Choose an informative subset mode name
+    SkString modeName;
+    switch(fSubsetMode) {
+        case kSingle_SubsetMode:
+            SkASSERT(0 == divisor);
+            modeName.printf("Single_%dx%d +%d_+%d", fSubsetWidth, fSubsetHeight, fOffsetLeft,
+                    fOffsetTop);
+            break;
+        case kDivisor_SubsetMode:
+            SkASSERT(0 == subsetWidth);
+            SkASSERT(0 == subsetHeight);
+            SkASSERT(0 == offsetLeft);
+            SkASSERT(0 == offsetTop);
+            modeName.printf("Divisor_%dx%d", fDivisor, fDivisor);
+            break;
+        case kTranslate_SubsetMode:
+            SkASSERT(0 == divisor);
+            SkASSERT(0 == offsetLeft);
+            SkASSERT(0 == offsetTop);
+            modeName.printf("Size_%dx%d", fSubsetWidth, fSubsetHeight);
+            break;
+        case kScale_SubsetMode:
+            SkASSERT(0 == divisor);
+            SkASSERT(0 == offsetLeft);
+            SkASSERT(0 == offsetTop);
+            modeName.printf("Scale_%dx%d", fSubsetWidth, fSubsetHeight);
+            break;
+    }
+
+    // Choose an informative color name
+    const char* colorName;
+    switch(fColorType) {

[scroggo, 2015/05/29 15:36:02]

What about kIndex8?

[msarett, 2015/06/01 14:03:25]

Agreed!

+        case kN32_SkColorType:
+            colorName = "N32";
+            break;
+        case kRGB_565_SkColorType:
+            colorName = "565";
+            break;
+        case kGray_8_SkColorType:
+            colorName = "Gray8";
+            break;
+        case kAlpha_8_SkColorType:
+            colorName = "Alpha8";
+            break;
+        default:
+            colorName = "Unknown";
+    }
+    fName.printf("%sSubset%s_%s_%s", fUseCodec ? "Codec" : "Image", modeName.c_str(),
+            baseName.c_str(), colorName);
+    
+    // Perform the decode setup
+    SkAutoTUnref<SkData> encoded(SkData::NewFromFileName(path->c_str()));
+    fStream.reset(new SkMemoryStream(encoded));
+}
+
+const char* CodecSubsetBench::onGetName() {
+    return fName.c_str();
+}
+
+bool CodecSubsetBench::isSuitableFor(Backend backend) {
+    return kNonRendering_Backend == backend;
+}
+    
+void CodecSubsetBench::onDraw(const int n, SkCanvas* canvas) {
+    // Choose a different benchmark depending on the mode
+    switch(fSubsetMode) {

[scroggo, 2015/05/29 15:36:02]

This giant switch statement might argue for using a different subclass for each
mode - it looks like we don't share much code amongst the modes.

[msarett, 2015/06/01 14:03:25]

Agreed.

+        case kSingle_SubsetMode:
+            if (fUseCodec) {
+                for (int count = 0; count < n; count++) {
+                    SkAutoTDelete<SkCodec> codec(SkCodec::NewFromStream(fStream->duplicate()));
+                    const SkImageInfo info = codec->getInfo();
+                    SkAutoTDeleteArray<uint8_t> row(SkNEW_ARRAY(uint8_t, info.minRowBytes()));
+                    SkScanlineDecoder* scanlineDecoder = codec->getScanlineDecoder(info);
+
+                    SkBitmap bitmap;
+                    if (!bitmap.tryAllocPixels(info.makeWH(fSubsetWidth, fSubsetHeight))) {
+                        SkDebugf("Could not allocate memory.  Aborting bench.\n");
+                        return;
+                    }
+
+                    scanlineDecoder->skipScanlines(fOffsetTop);
+                    uint32_t bpp = info.bytesPerPixel();
+                    for (uint32_t y = 0; y < fSubsetHeight; y++) {
+                        scanlineDecoder->getScanlines(row.get(), 1, 0);
+                        memcpy(bitmap.getAddr(0, y), row.get() + fOffsetLeft * bpp,
+                                fSubsetWidth * bpp);
+                    }
+                }
+            } else {
+                for (int count = 0; count < n; count++) {
+                    SkAutoTDelete<SkImageDecoder> decoder(SkImageDecoder::Factory(fStream));
+                    int width, height;
+                    decoder->buildTileIndex(fStream->duplicate(), &width, &height);

[scroggo, 2015/05/29 15:36:02]

Do we need to safely early exit in case buildTileIndex fails?

[msarett, 2015/06/01 14:03:24]

Let's check that the decode succeeds before creating the benchmark.

+                    SkBitmap bitmap;
+                    SkIRect rect = SkIRect::MakeXYWH(fOffsetLeft, fOffsetTop, fSubsetWidth,
+                            fSubsetHeight);
+                    decoder->decodeSubset(&bitmap, rect, fColorType);
+                }
+            }
+            break;
+        case kDivisor_SubsetMode:
+            if (fUseCodec) {
+                for (int count = 0; count < n; count++) {
+                    SkAutoTDelete<SkCodec> codec(SkCodec::NewFromStream(fStream->duplicate()));
+                    const SkImageInfo info = codec->getInfo();
+                    SkAutoTDeleteArray<uint8_t> row(SkNEW_ARRAY(uint8_t, info.minRowBytes()));
+                    SkScanlineDecoder* scanlineDecoder = codec->getScanlineDecoder(info);
+
+                    uint32_t subsetWidth = info.width() / fDivisor;

[scroggo, 2015/05/29 15:36:02]

These variables could all be const?

[msarett, 2015/06/01 14:03:25]

Done.

+                    uint32_t subsetHeight = info.height() / fDivisor;
+                    uint32_t maxSubsetWidth = subsetWidth + info.width() % fDivisor;
+                    uint32_t maxSubsetHeight = subsetHeight + info.height() % fDivisor;
+                    SkBitmap bitmap;
+                    // Note that we use the same bitmap for all of the subsets.
+                    // It might be slightly larger than necessary for some of the subsets.
+                    if (!bitmap.tryAllocPixels(info.makeWH(maxSubsetWidth, maxSubsetHeight))) {
+                        SkDebugf("Could not allocate memory.  Aborting bench.\n");
+                        return;
+                    }
+
+                    for (uint32_t blockX = 0; blockX < fDivisor; blockX++) {
+                        for (uint32_t blockY = 0; blockY < fDivisor; blockY++) {
+                            scanlineDecoder->skipScanlines(blockY * subsetHeight);
+                            uint32_t currSubsetWidth =
+                                    (blockX == fDivisor - 1) ? maxSubsetWidth : subsetWidth;
+                            uint32_t currSubsetHeight =
+                                    (blockY == fDivisor - 1) ? maxSubsetHeight : subsetHeight;
+                            uint32_t bpp = info.bytesPerPixel();
+                            for (uint32_t y = 0; y < currSubsetHeight; y++) {
+                                scanlineDecoder->getScanlines(row.get(), 1, 0);
+                                memcpy(bitmap.getAddr(0, y), row.get() + fOffsetLeft * bpp,
+                                        currSubsetWidth * bpp);
+                            }
+                        }
+                    }
+                }
+            } else {
+                for (int count = 0; count < n; count++) {
+                    int width, height;
+                    SkAutoTDelete<SkImageDecoder> decoder(SkImageDecoder::Factory(fStream));
+                    decoder->buildTileIndex(fStream->duplicate(), &width, &height);
+
+                    uint32_t subsetWidth = width / fDivisor;
+                    uint32_t subsetHeight = height / fDivisor;
+                    uint32_t maxSubsetWidth = subsetWidth + width % fDivisor;
+                    uint32_t maxSubsetHeight = subsetHeight + height % fDivisor;
+                    SkBitmap bitmap;
+                    // Note that we use the same bitmap for all of the subsets.
+                    // It might be slightly larger than necessary for some of the subsets.
+                    if (!bitmap.tryAllocPixels(SkImageInfo::Make(maxSubsetWidth, maxSubsetHeight,

[scroggo, 2015/05/29 15:36:02]

When using SkImageDecoder, you don't need to allocate any pixels - the decoder
will do that for you.

[msarett, 2015/06/01 14:03:24]

The reason I allocate here is so that SkImageDecoder will reuse the same bitmap
for all of the subsets decoded below.  This is what SkCodec does above.  I think
that allows for a more fair comparison.

[scroggo, 2015/06/01 17:25:14]

Got it. We talked about this in person and you convinced me this makes sense.
Can you add a comment, so I don't forget that it makes sense?

[msarett, 2015/06/01 20:37:01]

Done.

+                            fColorType, kOpaque_SkAlphaType))) {
+                        SkDebugf("Could not allocate memory.  Aborting bench.\n");
+                        return;
+                    }
+
+                    for (uint32_t blockX = 0; blockX < fDivisor; blockX++) {
+                        for (uint32_t blockY = 0; blockY < fDivisor; blockY++) {
+                            uint32_t currSubsetWidth =
+                                    (blockX == fDivisor - 1) ? maxSubsetWidth : subsetWidth;
+                            uint32_t currSubsetHeight =
+                                    (blockY == fDivisor - 1) ? maxSubsetHeight : subsetHeight;
+                            SkIRect rect = SkIRect::MakeXYWH(blockX * subsetWidth,
+                                    blockY * subsetHeight, currSubsetWidth, currSubsetHeight);
+                            decoder->decodeSubset(&bitmap, rect, fColorType);
+                        }
+                    }
+                }
+            }
+            break;
+        case kTranslate_SubsetMode:
+            if (fUseCodec) {
+                for (int count = 0; count < n; count++) {
+                    SkAutoTDelete<SkCodec> codec(SkCodec::NewFromStream(fStream->duplicate()));
+                    const SkImageInfo info = codec->getInfo();
+                    SkAutoTDeleteArray<uint8_t> row(SkNEW_ARRAY(uint8_t, info.minRowBytes()));
+                    SkScanlineDecoder* scanlineDecoder = codec->getScanlineDecoder(info);
+
+                    SkBitmap bitmap;
+                    // Note that we use the same bitmap for all of the subsets.
+                    // It might be larger than necessary for the end subsets.
+                    if (!bitmap.tryAllocPixels(info.makeWH(fSubsetWidth, fSubsetHeight))) {
+                        SkDebugf("Could not allocate memory.  Aborting bench.\n");
+                        return;
+                    }
+
+                    for (int x = 0; x < info.width(); x += fSubsetWidth) {
+                        for (int y = 0; y < info.height(); y += fSubsetHeight) {
+                            scanlineDecoder->skipScanlines(y);
+                            uint32_t currSubsetWidth = x + (int) fSubsetWidth > info.width() ?
+                                    info.width() - x : fSubsetWidth;
+                            uint32_t currSubsetHeight = y + (int) fSubsetHeight > info.height() ?
+                                    info.height() - y : fSubsetHeight;
+                            uint32_t bpp = info.bytesPerPixel();
+                            for (uint32_t y = 0; y < currSubsetHeight; y++) {
+                                scanlineDecoder->getScanlines(row.get(), 1, 0);
+                                memcpy(bitmap.getAddr(0, y), row.get() + x * bpp,
+                                        currSubsetWidth * bpp);
+                            }
+                        }
+                    }
+                }
+            } else {
+                for (int count = 0; count < n; count++) {
+                    int width, height;
+                    SkAutoTDelete<SkImageDecoder> decoder(SkImageDecoder::Factory(fStream));
+                    decoder->buildTileIndex(fStream->duplicate(), &width, &height);
+                    SkBitmap bitmap;
+                    // Note that we use the same bitmap for all of the subsets.
+                    // It might be larger than necessary for the end subsets.
+                    if (!bitmap.tryAllocPixels(SkImageInfo::Make(fSubsetWidth, fSubsetHeight,

[scroggo, 2015/05/29 15:36:02]

Again, this is not necessary.

[msarett, 2015/06/01 14:03:25]

Acknowledged.

+                            fColorType, kOpaque_SkAlphaType))) {
+                        SkDebugf("Could not allocate memory.  Aborting bench.\n");
+                        return;
+                    }
+
+                    for (int x = 0; x < width; x += fSubsetWidth) {
+                        for (int y = 0; y < height; y += fSubsetHeight) {
+                            uint32_t currSubsetWidth = x + (int) fSubsetWidth > width ?
+                                    width - x : fSubsetWidth;
+                            uint32_t currSubsetHeight = y + (int) fSubsetHeight > height ?
+                                    height - y : fSubsetHeight;
+                            SkIRect rect = SkIRect::MakeXYWH(x, y, currSubsetWidth,
+                                    currSubsetHeight);
+                            decoder->decodeSubset(&bitmap, rect, fColorType);
+                        }
+                    }
+                }
+            }
+            break;
+        case kScale_SubsetMode:
+            if (fUseCodec) {
+                for (int count = 0; count < n; count++) {
+                    SkAutoTDelete<SkCodec> codec(SkCodec::NewFromStream(fStream->duplicate()));
+                    const SkImageInfo info = codec->getInfo();
+                    SkAutoTDeleteArray<uint8_t> row(SkNEW_ARRAY(uint8_t, info.minRowBytes()));
+                    SkScanlineDecoder* scanlineDecoder = codec->getScanlineDecoder(info);
+
+                    int centerX = info.width() / 2;
+                    int centerY = info.height() / 2;
+                    int w = fSubsetWidth;
+                    int h = fSubsetHeight;
+                    do {
+                        int subsetStartX = SkTMax(0, centerX - w / 2);
+                        int subsetStartY = SkTMax(0, centerY - h / 2);
+                        int subsetWidth = SkTMin(w, info.width() - subsetStartX);
+                        int subsetHeight = SkTMin(h, info.height() - subsetStartY);
+                        // Note that if we subseted and scaled in a single step, we could use the

[scroggo, 2015/05/29 15:36:02]

subsetted? (Technically not a verb, but it seems like it should have two 't's if
we add an 'ed'.)

[msarett, 2015/06/01 14:03:25]

Done.

+                        // same bitmap - as is often done in actual use cases.
+                        SkBitmap bitmap;
+                        if (!bitmap.tryAllocPixels(info.makeWH(subsetWidth, subsetHeight))) {
+                            SkDebugf("Could not allocate memory.  Aborting bench.\n");
+                            return;
+                        }
+
+                        uint32_t bpp = info.bytesPerPixel();
+                        scanlineDecoder->skipScanlines(subsetStartY);
+                        for (int y = 0; y < subsetHeight; y++) {
+                            scanlineDecoder->getScanlines(row.get(), 1, 0);
+                            memcpy(bitmap.getAddr(0, y), row.get() + subsetStartX * bpp,
+                                    subsetWidth * bpp);
+                        }
+                        w <<= 1;
+                        h <<= 1;
+                    } while (w < 2 * info.width() || h < 2 * info.height());
+                }
+            } else {
+                for (int count = 0; count < n; count++) {
+                    int width, height;
+                    SkAutoTDelete<SkImageDecoder> decoder(SkImageDecoder::Factory(fStream));
+                    decoder->buildTileIndex(fStream->duplicate(), &width, &height);
+
+                    int centerX = width / 2;
+                    int centerY = height / 2;
+                    int w = fSubsetWidth;
+                    int h = fSubsetHeight;
+                    do {
+                        int subsetStartX = SkTMax(0, centerX - w / 2);
+                        int subsetStartY = SkTMax(0, centerY - h / 2);
+                        int subsetWidth = SkTMin(w, width - subsetStartX);
+                        int subsetHeight = SkTMin(h, height - subsetStartY);
+                        // Note that if we subseted and scaled in a single step, we could use the
+                        // same bitmap - as is often done in actual use cases.
+                        SkBitmap bitmap;
+                        if (!bitmap.tryAllocPixels(SkImageInfo::Make(subsetWidth, subsetHeight,

[scroggo, 2015/05/29 15:36:03]

Again, not needed. There are some cases where a bitmap is passed in already
allocated (which actually means it may make sense to test that), but they
wouldn't be allocating a new pixelref each time (I think?)

[msarett, 2015/06/01 14:03:24]

Agreed that this allocation is unnecessary here.  We use a different bitmap for
each subset anyway.

+                                fColorType, kOpaque_SkAlphaType))) {
+                            SkDebugf("Could not allocate memory.  Aborting bench.\n");
+                            return;
+                        }
+
+                        SkIRect rect = SkIRect::MakeXYWH(subsetStartX, subsetStartY, subsetWidth,
+                                subsetHeight);
+                        decoder->decodeSubset(&bitmap, rect, fColorType);
+                        w <<= 1;
+                        h <<= 1;
+                    } while (w < 2 * width || h < 2 * height);
+                }
+            }
+            break;
+    }
+}