Refactor texture compressors into separate files

src/utils/SkTextureCompressor_R11EAC.cpp

Index: src/utils/SkTextureCompressor_R11EAC.cpp
diff --git a/src/utils/SkTextureCompressor.cpp b/src/utils/SkTextureCompressor_R11EAC.cpp
similarity index 69%
copy from src/utils/SkTextureCompressor.cpp
copy to src/utils/SkTextureCompressor_R11EAC.cpp
index a593b36880e8588fe50a0eb0bfcea68afac2b320..f8b5346207fec8ccfc89f02fc8cc0bc32104f6f9 100644
--- a/src/utils/SkTextureCompressor.cpp
+++ b/src/utils/SkTextureCompressor_R11EAC.cpp
@@ -7,292 +7,8 @@
 
 #include "SkTextureCompressor.h"
 
-#include "SkBitmap.h"
-#include "SkData.h"
 #include "SkEndian.h"
 
-#include "SkTextureCompression_opts.h"
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// Utility Functions
-//
-////////////////////////////////////////////////////////////////////////////////
-
-// Absolute difference between two values. More correct than SkTAbs(a - b)
-// because it works on unsigned values.
-template <typename T> inline T abs_diff(const T &a, const T &b) {
-    return (a > b) ? (a - b) : (b - a);
-}
-
-static bool is_extremal(uint8_t pixel) {
-    return 0 == pixel || 255 == pixel;
-}
-
-typedef uint64_t (*A84x4To64BitProc)(const uint8_t block[]);
-
-// This function is used by both R11 EAC and LATC to compress 4x4 blocks
-// of 8-bit alpha into 64-bit values that comprise the compressed data.
-// For both formats, we need to make sure that the dimensions of the 
-// src pixels are divisible by 4, and copy 4x4 blocks one at a time
-// for compression.
-static bool compress_4x4_a8_to_64bit(uint8_t* dst, const uint8_t* src,
-                                     int width, int height, int rowBytes,
-                                     A84x4To64BitProc proc) {
-    // Make sure that our data is well-formed enough to be considered for compression
-    if (0 == width || 0 == height || (width % 4) != 0 || (height % 4) != 0) {
-        return false;
-    }
-
-    int blocksX = width >> 2;
-    int blocksY = height >> 2;
-
-    uint8_t block[16];
-    uint64_t* encPtr = reinterpret_cast<uint64_t*>(dst);
-    for (int y = 0; y < blocksY; ++y) {
-        for (int x = 0; x < blocksX; ++x) {
-            // Load block
-            for (int k = 0; k < 4; ++k) {
-                memcpy(block + k*4, src + k*rowBytes + 4*x, 4);
-            }
-
-            // Compress it
-            *encPtr = proc(block);
-            ++encPtr;
-        }
-        src += 4 * rowBytes;
-    }
-
-    return true;
-}
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// LATC compressor
-//
-////////////////////////////////////////////////////////////////////////////////
-
-// LATC compressed texels down into square 4x4 blocks
-static const int kLATCPaletteSize = 8;
-static const int kLATCBlockSize = 4;
-static const int kLATCPixelsPerBlock = kLATCBlockSize * kLATCBlockSize;
-
-// Generates an LATC palette. LATC constructs
-// a palette of eight colors from LUM0 and LUM1 using the algorithm:
-//
-// LUM0,              if lum0 > lum1 and code(x,y) == 0
-// LUM1,              if lum0 > lum1 and code(x,y) == 1
-// (6*LUM0+  LUM1)/7, if lum0 > lum1 and code(x,y) == 2
-// (5*LUM0+2*LUM1)/7, if lum0 > lum1 and code(x,y) == 3
-// (4*LUM0+3*LUM1)/7, if lum0 > lum1 and code(x,y) == 4
-// (3*LUM0+4*LUM1)/7, if lum0 > lum1 and code(x,y) == 5
-// (2*LUM0+5*LUM1)/7, if lum0 > lum1 and code(x,y) == 6
-// (  LUM0+6*LUM1)/7, if lum0 > lum1 and code(x,y) == 7
-//
-// LUM0,              if lum0 <= lum1 and code(x,y) == 0
-// LUM1,              if lum0 <= lum1 and code(x,y) == 1
-// (4*LUM0+  LUM1)/5, if lum0 <= lum1 and code(x,y) == 2
-// (3*LUM0+2*LUM1)/5, if lum0 <= lum1 and code(x,y) == 3
-// (2*LUM0+3*LUM1)/5, if lum0 <= lum1 and code(x,y) == 4
-// (  LUM0+4*LUM1)/5, if lum0 <= lum1 and code(x,y) == 5
-// 0,                 if lum0 <= lum1 and code(x,y) == 6
-// 255,               if lum0 <= lum1 and code(x,y) == 7
-
-static void generate_latc_palette(uint8_t palette[], uint8_t lum0, uint8_t lum1) {
-    palette[0] = lum0;
-    palette[1] = lum1;
-    if (lum0 > lum1) {
-        for (int i = 1; i < 7; i++) {
-            palette[i+1] = ((7-i)*lum0 + i*lum1) / 7;
-        }
-    } else {
-        for (int i = 1; i < 5; i++) {
-            palette[i+1] = ((5-i)*lum0 + i*lum1) / 5;
-        }
-        palette[6] = 0;
-        palette[7] = 255;
-    }
-}
-
-// Compress a block by using the bounding box of the pixels. It is assumed that
-// there are no extremal pixels in this block otherwise we would have used
-// compressBlockBBIgnoreExtremal.
-static uint64_t compress_latc_block_bb(const uint8_t pixels[]) {
-    uint8_t minVal = 255;
-    uint8_t maxVal = 0;
-    for (int i = 0; i < kLATCPixelsPerBlock; ++i) {
-        minVal = SkTMin(pixels[i], minVal);
-        maxVal = SkTMax(pixels[i], maxVal);
-    }
-
-    SkASSERT(!is_extremal(minVal));
-    SkASSERT(!is_extremal(maxVal));
-
-    uint8_t palette[kLATCPaletteSize];
-    generate_latc_palette(palette, maxVal, minVal);
-
-    uint64_t indices = 0;
-    for (int i = kLATCPixelsPerBlock - 1; i >= 0; --i) {
-
-        // Find the best palette index
-        uint8_t bestError = abs_diff(pixels[i], palette[0]);
-        uint8_t idx = 0;
-        for (int j = 1; j < kLATCPaletteSize; ++j) {
-            uint8_t error = abs_diff(pixels[i], palette[j]);
-            if (error < bestError) {
-                bestError = error;
-                idx = j;
-            }
-        }
-
-        indices <<= 3;
-        indices |= idx;
-    }
-
-    return
-        SkEndian_SwapLE64(
-            static_cast<uint64_t>(maxVal) |
-            (static_cast<uint64_t>(minVal) << 8) |
-            (indices << 16));
-}
-
-// Compress a block by using the bounding box of the pixels without taking into
-// account the extremal values. The generated palette will contain extremal values
-// and fewer points along the line segment to interpolate.
-static uint64_t compress_latc_block_bb_ignore_extremal(const uint8_t pixels[]) {
-    uint8_t minVal = 255;
-    uint8_t maxVal = 0;
-    for (int i = 0; i < kLATCPixelsPerBlock; ++i) {
-        if (is_extremal(pixels[i])) {
-            continue;
-        }
-
-        minVal = SkTMin(pixels[i], minVal);
-        maxVal = SkTMax(pixels[i], maxVal);
-    }
-
-    SkASSERT(!is_extremal(minVal));
-    SkASSERT(!is_extremal(maxVal));
-
-    uint8_t palette[kLATCPaletteSize];
-    generate_latc_palette(palette, minVal, maxVal);
-
-    uint64_t indices = 0;
-    for (int i = kLATCPixelsPerBlock - 1; i >= 0; --i) {
-
-        // Find the best palette index
-        uint8_t idx = 0;
-        if (is_extremal(pixels[i])) {
-            if (0xFF == pixels[i]) {
-                idx = 7;
-            } else if (0 == pixels[i]) {
-                idx = 6;
-            } else {
-                SkFAIL("Pixel is extremal but not really?!");
-            }
-        } else {
-            uint8_t bestError = abs_diff(pixels[i], palette[0]);
-            for (int j = 1; j < kLATCPaletteSize - 2; ++j) {
-                uint8_t error = abs_diff(pixels[i], palette[j]);
-                if (error < bestError) {
-                    bestError = error;
-                    idx = j;
-                }
-            }
-        }
-
-        indices <<= 3;
-        indices |= idx;
-    }
-
-    return
-        SkEndian_SwapLE64(
-            static_cast<uint64_t>(minVal) |
-            (static_cast<uint64_t>(maxVal) << 8) |
-            (indices << 16));
-}
-
-
-// Compress LATC block. Each 4x4 block of pixels is decompressed by LATC from two
-// values LUM0 and LUM1, and an index into the generated palette. Details of how
-// the palette is generated can be found in the comments of generatePalette above.
-//
-// We choose which palette type to use based on whether or not 'pixels' contains
-// any extremal values (0 or 255). If there are extremal values, then we use the
-// palette that has the extremal values built in. Otherwise, we use the full bounding
-// box.
-
-static uint64_t compress_latc_block(const uint8_t pixels[]) {
-    // Collect unique pixels
-    int nUniquePixels = 0;
-    uint8_t uniquePixels[kLATCPixelsPerBlock];
-    for (int i = 0; i < kLATCPixelsPerBlock; ++i) {
-        bool foundPixel = false;
-        for (int j = 0; j < nUniquePixels; ++j) {
-            foundPixel = foundPixel || uniquePixels[j] == pixels[i];
-        }
-
-        if (!foundPixel) {
-            uniquePixels[nUniquePixels] = pixels[i];
-            ++nUniquePixels;
-        }
-    }
-
-    // If there's only one unique pixel, then our compression is easy.
-    if (1 == nUniquePixels) {
-        return SkEndian_SwapLE64(pixels[0] | (pixels[0] << 8));
-
-    // Similarly, if there are only two unique pixels, then our compression is
-    // easy again: place the pixels in the block header, and assign the indices
-    // with one or zero depending on which pixel they belong to.
-    } else if (2 == nUniquePixels) {
-        uint64_t outBlock = 0;
-        for (int i = kLATCPixelsPerBlock - 1; i >= 0; --i) {
-            int idx = 0;
-            if (pixels[i] == uniquePixels[1]) {
-                idx = 1;
-            }
-
-            outBlock <<= 3;
-            outBlock |= idx;
-        }
-        outBlock <<= 16;
-        outBlock |= (uniquePixels[0] | (uniquePixels[1] << 8));
-        return SkEndian_SwapLE64(outBlock);
-    }
-
-    // Count non-maximal pixel values
-    int nonExtremalPixels = 0;
-    for (int i = 0; i < nUniquePixels; ++i) {
-        if (!is_extremal(uniquePixels[i])) {
-            ++nonExtremalPixels;
-        }
-    }
-
-    // If all the pixels are nonmaximal then compute the palette using
-    // the bounding box of all the pixels.
-    if (nonExtremalPixels == nUniquePixels) {
-        // This is really just for correctness, in all of my tests we
-        // never take this step. We don't lose too much perf here because
-        // most of the processing in this function is worth it for the 
-        // 1 == nUniquePixels optimization.
-        return compress_latc_block_bb(pixels);
-    } else {
-        return compress_latc_block_bb_ignore_extremal(pixels);
-    }
-}
-
-static inline bool compress_a8_to_latc(uint8_t* dst, const uint8_t* src,
-                                       int width, int height, int rowBytes) {
-    return compress_4x4_a8_to_64bit(dst, src, width, height, rowBytes, compress_latc_block);
-}
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// R11 EAC Compressor
-//
-////////////////////////////////////////////////////////////////////////////////
-
 // #define COMPRESS_R11_EAC_SLOW 1
 // #define COMPRESS_R11_EAC_FAST 1
 #define COMPRESS_R11_EAC_FASTEST 1
@@ -548,9 +264,52 @@ static uint64_t compress_r11eac_block(const uint8_t block[16]) {
 
     return compress_heterogeneous_r11eac_block(block);
 }
+
+// This function is used by R11 EAC to compress 4x4 blocks
+// of 8-bit alpha into 64-bit values that comprise the compressed data.

[robertphillips, 2014/07/21 21:56:57]

both formats ?

[krajcevski, 2014/07/21 22:03:55]

Done.

+// For both formats, we need to make sure that the dimensions of the 
+// src pixels are divisible by 4, and copy 4x4 blocks one at a time
+// for compression.
+typedef uint64_t (*A84x4To64BitProc)(const uint8_t block[]);
+
+static bool compress_4x4_a8_to_64bit(uint8_t* dst, const uint8_t* src,
+                                     int width, int height, int rowBytes,
+                                     A84x4To64BitProc proc) {
+    // Make sure that our data is well-formed enough to be considered for compression
+    if (0 == width || 0 == height || (width % 4) != 0 || (height % 4) != 0) {
+        return false;
+    }
+
+    int blocksX = width >> 2;
+    int blocksY = height >> 2;
+
+    uint8_t block[16];
+    uint64_t* encPtr = reinterpret_cast<uint64_t*>(dst);
+    for (int y = 0; y < blocksY; ++y) {
+        for (int x = 0; x < blocksX; ++x) {
+            // Load block
+            for (int k = 0; k < 4; ++k) {
+                memcpy(block + k*4, src + k*rowBytes + 4*x, 4);
+            }
+
+            // Compress it
+            *encPtr = proc(block);
+            ++encPtr;
+        }
+        src += 4 * rowBytes;
+    }
+
+    return true;
+}
 #endif  // (COMPRESS_R11_EAC_SLOW) || (COMPRESS_R11_EAC_FAST)
 
 #if COMPRESS_R11_EAC_FASTEST
+template<unsigned shift>
+static inline uint64_t swap_shift(uint64_t x, uint64_t mask) {
+    const uint64_t t = (x ^ (x >> shift)) & mask;
+    return x ^ t ^ (t << shift);
+}
+
 static inline uint64_t interleave6(uint64_t topRows, uint64_t bottomRows) {
     // If our 3-bit block indices are laid out as:
     // a b c d
@@ -584,8 +343,7 @@ static inline uint64_t interleave6(uint64_t topRows, uint64_t bottomRows) {
 
     // x: 00 a e 00 b f 00 c g 00 d h 00 i m 00 j n 00 k o 00 l p
 
-    uint64_t t = (x ^ (x >> 10)) & 0x3FC0003FC00000ULL;
-    x = x ^ t ^ (t << 10);
+    x = swap_shift<10>(x, 0x3FC0003FC00000ULL);
 
     // x: b f 00 00 00 a e c g i m 00 00 00 d h j n 00 k o 00 l p
 
@@ -593,38 +351,30 @@ static inline uint64_t interleave6(uint64_t topRows, uint64_t bottomRows) {
 
     // x: 00 00 00 00 00 00 00 00 b f l p a e c g i m k o d h j n
 
-    t = (x ^ (x >> 6)) & 0xFC0000ULL;
-    x = x ^ t ^ (t << 6);
+    x = swap_shift<6>(x, 0xFC0000ULL);
 
 #if defined (SK_CPU_BENDIAN)
     // x: 00 00 00 00 00 00 00 00 b f l p a e i m c g k o d h j n
 
-    t = (x ^ (x >> 36)) & 0x3FULL;
-    x = x ^ t ^ (t << 36);
+    x = swap_shift<36>(x, 0x3FULL);
 
     // x: 00 00 00 00 00 00 00 00 b f j n a e i m c g k o d h l p
 
-    t = (x ^ (x >> 12)) & 0xFFF000000ULL;
-    x = x ^ t ^ (t << 12);
-
-    // x: 00 00 00 00 00 00 00 00 a e i m b f j n c g k o d h l p
-    return x;
+    x = swap_shift<12>(x, 0xFFF000000ULL);
 #else
     // If our CPU is little endian, then the above logic will
     // produce the following indices:
     // x: 00 00 00 00 00 00 00 00 c g i m d h l p b f j n a e k o
 
-    t = (x ^ (x >> 36)) & 0xFC0ULL;
-    x = x ^ t ^ (t << 36);
+    x = swap_shift<36>(x, 0xFC0ULL);
 
     // x: 00 00 00 00 00 00 00 00 a e i m d h l p b f j n c g k o
     
     x = (x & (0xFFFULL << 36)) | ((x & 0xFFFFFFULL) << 12) | ((x >> 24) & 0xFFFULL);
+#endif
 
     // x: 00 00 00 00 00 00 00 00 a e i m b f j n c g k o d h l p
-    
     return x;
-#endif
 }
 
 // This function converts an integer containing four bytes of alpha
@@ -732,6 +482,12 @@ static bool compress_a8_to_r11eac_fast(uint8_t* dst, const uint8_t* src,
 }
 #endif // COMPRESS_R11_EAC_FASTEST
 
+////////////////////////////////////////////////////////////////////////////////
+//
+// Utility functions used by the blitter
+//
+////////////////////////////////////////////////////////////////////////////////
+
 // The R11 EAC format expects that indices are given in column-major order. Since
 // we receive alpha values in raster order, this usually means that we have to use
 // pack6 above to properly pack our indices. However, if our indices come from the
@@ -795,43 +551,16 @@ static inline uint64_t compress_block_vertical(const uint32_t alphaColumn0,
         
 }
 
-static inline bool compress_a8_to_r11eac(uint8_t* dst, const uint8_t* src,
-                                         int width, int height, int rowBytes) {
-#if (COMPRESS_R11_EAC_SLOW) || (COMPRESS_R11_EAC_FAST)
-    return compress_4x4_a8_to_64bit(dst, src, width, height, rowBytes, compress_r11eac_block);
-#elif COMPRESS_R11_EAC_FASTEST
-    return compress_a8_to_r11eac_fast(dst, src, width, height, rowBytes);
-#else
-#error "Must choose R11 EAC algorithm"
-#endif
-}
-
 // Updates the block whose columns are stored in blockColN. curAlphai is expected
 // to store, as an integer, the four alpha values that will be placed within each
 // of the columns in the range [col, col+colsLeft).
-static inline void update_block_columns(
-    uint32_t* blockCol1, uint32_t* blockCol2, uint32_t* blockCol3, uint32_t* blockCol4,
-    const int col, const int colsLeft, const uint32_t curAlphai) {
-    SkASSERT(NULL != blockCol1);
-    SkASSERT(NULL != blockCol2);
-    SkASSERT(NULL != blockCol3);
-    SkASSERT(NULL != blockCol4);
+static inline void update_block_columns(uint32_t* block, const int col,
+                                        const int colsLeft, const uint32_t curAlphai) {
+    SkASSERT(NULL != block);
     SkASSERT(col + colsLeft <= 4);
+
     for (int i = col; i < (col + colsLeft); ++i) {
-        switch(i) {
-        case 0:
-            *blockCol1 = curAlphai;
-            break;
-        case 1:
-            *blockCol2 = curAlphai;
-            break;
-        case 2:
-            *blockCol3 = curAlphai;
-            break;
-        case 3:
-            *blockCol4 = curAlphai;
-            break;
-        }
+        block[i] = curAlphai;
     }
 }
 
@@ -839,78 +568,182 @@ static inline void update_block_columns(
 
 namespace SkTextureCompressor {
 
-static inline size_t get_compressed_data_size(Format fmt, int width, int height) {
-    switch (fmt) {
-        // These formats are 64 bits per 4x4 block.
-        case kR11_EAC_Format:
-        case kLATC_Format:
-        {
-            static const int kLATCEncodedBlockSize = 8;
-
-            const int blocksX = width / kLATCBlockSize;
-            const int blocksY = height / kLATCBlockSize;
-
-            return blocksX * blocksY * kLATCEncodedBlockSize;
-        }
+bool CompressA8ToR11EAC(uint8_t* dst, const uint8_t* src, int width, int height, int rowBytes) {
 
-        default:
-            SkFAIL("Unknown compressed format!");
-            return 0;
-    }
-}
-
-bool CompressBufferToFormat(uint8_t* dst, const uint8_t* src, SkColorType srcColorType,
-                            int width, int height, int rowBytes, Format format, bool opt) {
-    CompressionProc proc = NULL;
-    if (opt) {
-        proc = SkTextureCompressorGetPlatformProc(srcColorType, format);
-    }
+#if (COMPRESS_R11_EAC_SLOW) || (COMPRESS_R11_EAC_FAST)
 
-    if (NULL == proc) {
-        switch (srcColorType) {
-            case kAlpha_8_SkColorType:
-            {
-                switch (format) {
-                    case kLATC_Format:
-                        proc = compress_a8_to_latc;
-                        break;
-                    case kR11_EAC_Format:
-                        proc = compress_a8_to_r11eac;
-                        break;
-                    default:
-                        // Do nothing...
-                        break;
-                }
-            }
-            break;
+    return compress_4x4_a8_to_64bit(dst, src, width, height, rowBytes, compress_r11eac_block);
 
-            default:
-                // Do nothing...
-                break;
-        }
-    }
+#elif COMPRESS_R11_EAC_FASTEST
 
-    if (NULL != proc) {
-        return proc(dst, src, width, height, rowBytes);
-    }
+    return compress_a8_to_r11eac_fast(dst, src, width, height, rowBytes);
 
-    return false;
+#else
+#error "Must choose R11 EAC algorithm"
+#endif
 }
 
-SkData *CompressBitmapToFormat(const SkBitmap &bitmap, Format format) {
-    SkAutoLockPixels alp(bitmap);
-
-    int compressedDataSize = get_compressed_data_size(format, bitmap.width(), bitmap.height());
-    const uint8_t* src = reinterpret_cast<const uint8_t*>(bitmap.getPixels());
-    uint8_t* dst = reinterpret_cast<uint8_t*>(sk_malloc_throw(compressedDataSize));
-    if (CompressBufferToFormat(dst, src, bitmap.colorType(), bitmap.width(), bitmap.height(),
-                               bitmap.rowBytes(), format)) {
-        return SkData::NewFromMalloc(dst, compressedDataSize);
+// This class implements a blitter that blits directly into a buffer that will
+// be used as an R11 EAC compressed texture. We compute this buffer by
+// buffering four scan lines and then outputting them all at once. This blitter
+// is only expected to be used with alpha masks, i.e. kAlpha8_SkColorType.
+class R11_EACBlitter : public SkBlitter {
+public:
+    R11_EACBlitter(int width, int height, void *compressedBuffer);
+    virtual ~R11_EACBlitter() { this->flushRuns(); }
+
+    // Blit a horizontal run of one or more pixels.
+    virtual void blitH(int x, int y, int width) SK_OVERRIDE {
+        // This function is intended to be called from any standard RGB
+        // buffer, so we should never encounter it. However, if some code
+        // path does end up here, then this needs to be investigated.
+        SkFAIL("Not implemented!");
     }
+    
+    // Blit a horizontal run of antialiased pixels; runs[] is a *sparse*
+    // zero-terminated run-length encoding of spans of constant alpha values.
+    virtual void blitAntiH(int x, int y,
+                           const SkAlpha antialias[],
+                           const int16_t runs[]) SK_OVERRIDE;
+    
+    // Blit a vertical run of pixels with a constant alpha value.
+    virtual void blitV(int x, int y, int height, SkAlpha alpha) SK_OVERRIDE {
+        // This function is currently not implemented. It is not explicitly
+        // required by the contract, but if at some time a code path runs into
+        // this function (which is entirely possible), it needs to be implemented.
+        //
+        // TODO (krajcevski):
+        // This function will be most easily implemented in one of two ways:
+        // 1. Buffer each vertical column value and then construct a list
+        //    of alpha values and output all of the blocks at once. This only
+        //    requires a write to the compressed buffer
+        // 2. Replace the indices of each block with the proper indices based
+        //    on the alpha value. This requires a read and write of the compressed
+        //    buffer, but much less overhead.
+        SkFAIL("Not implemented!");
+    }
+
+    // Blit a solid rectangle one or more pixels wide.
+    virtual void blitRect(int x, int y, int width, int height) SK_OVERRIDE {
+        // Analogous to blitRow, this function is intended for RGB targets
+        // and should never be called by this blitter. Any calls to this function
+        // are probably a bug and should be investigated.
+        SkFAIL("Not implemented!");
+    }
+
+    // Blit a rectangle with one alpha-blended column on the left,
+    // width (zero or more) opaque pixels, and one alpha-blended column
+    // on the right. The result will always be at least two pixels wide.
+    virtual void blitAntiRect(int x, int y, int width, int height,
+                              SkAlpha leftAlpha, SkAlpha rightAlpha) SK_OVERRIDE {
+        // This function is currently not implemented. It is not explicitly
+        // required by the contract, but if at some time a code path runs into
+        // this function (which is entirely possible), it needs to be implemented.
+        //
+        // TODO (krajcevski):
+        // This function will be most easily implemented as follows:
+        // 1. If width/height are smaller than a block, then update the
+        //    indices of the affected blocks.
+        // 2. If width/height are larger than a block, then construct a 9-patch
+        //    of block encodings that represent the rectangle, and write them
+        //    to the compressed buffer as necessary. Whether or not the blocks
+        //    are overwritten by zeros or just their indices are updated is up
+        //    to debate.
+        SkFAIL("Not implemented!");
+    }
+
+    // Blit a pattern of pixels defined by a rectangle-clipped mask;
+    // typically used for text.
+    virtual void blitMask(const SkMask&, const SkIRect& clip) SK_OVERRIDE {
+        // This function is currently not implemented. It is not explicitly
+        // required by the contract, but if at some time a code path runs into
+        // this function (which is entirely possible), it needs to be implemented.
+        //
+        // TODO (krajcevski):
+        // This function will be most easily implemented in the same way as
+        // blitAntiRect above.
+        SkFAIL("Not implemented!");
+    }
+
+    // If the blitter just sets a single value for each pixel, return the
+    // bitmap it draws into, and assign value. If not, return NULL and ignore
+    // the value parameter.
+    virtual const SkBitmap* justAnOpaqueColor(uint32_t* value) SK_OVERRIDE {
+        return NULL;
+    }
+
+    /**
+     * Compressed texture blitters only really work correctly if they get
+     * four blocks at a time. That being said, this blitter tries it's best
+     * to preserve semantics if blitAntiH doesn't get called in too many
+     * weird ways...
+     */
+    virtual int requestRowsPreserved() const { return kR11_EACBlockSz; }
+
+protected:
+    virtual void onNotifyFinished() { this->flushRuns(); }
+
+private:
+    static const int kR11_EACBlockSz = 4;
+    static const int kPixelsPerBlock = kR11_EACBlockSz * kR11_EACBlockSz;
+
+    // The longest possible run of pixels that this blitter will receive.
+    // This is initialized in the constructor to 0x7FFE, which is one less
+    // than the largest positive 16-bit integer. We make sure that it's one
+    // less for debugging purposes. We also don't make this variable static
+    // in order to make sure that we can construct a valid pointer to it.
+    const int16_t kLongestRun;
+
+    // Usually used in conjunction with kLongestRun. This is initialized to
+    // zero.
+    const SkAlpha kZeroAlpha;
+
+    // This is the information that we buffer whenever we're asked to blit
+    // a row with this blitter.
+    struct BufferedRun {
+        const SkAlpha* fAlphas;
+        const int16_t* fRuns;
+        int fX, fY;
+    } fBufferedRuns[kR11_EACBlockSz];
+
+    // The next row (0-3) that we need to blit. This value should never exceed
+    // the number of rows that we have (kR11_EACBlockSz)
+    int fNextRun;
+
+    // The width and height of the image that we're blitting
+    const int fWidth;
+    const int fHeight;
+
+    // The R11 EAC buffer that we're blitting into. It is assumed that the buffer
+    // is large enough to store a compressed image of size fWidth*fHeight.
+    uint64_t* const fBuffer;
+
+    // Various utility functions
+    int blocksWide() const { return fWidth / kR11_EACBlockSz; }
+    int blocksTall() const { return fHeight / kR11_EACBlockSz; }
+    int totalBlocks() const { return (fWidth * fHeight) / kPixelsPerBlock; }
+
+    // Returns the block index for the block containing pixel (x, y). Block
+    // indices start at zero and proceed in raster order.
+    int getBlockOffset(int x, int y) const {
+        SkASSERT(x < fWidth);
+        SkASSERT(y < fHeight);
+        const int blockCol = x / kR11_EACBlockSz;
+        const int blockRow = y / kR11_EACBlockSz;
+        return blockRow * this->blocksWide() + blockCol;
+    }
+
+    // Returns a pointer to the block containing pixel (x, y)
+    uint64_t *getBlock(int x, int y) const {
+        return fBuffer + this->getBlockOffset(x, y);
+    }
+
+    // The following function writes the buffered runs to compressed blocks.
+    // If fNextRun < 4, then we fill the runs that we haven't buffered with
+    // the constant zero buffer.
+    void flushRuns();
+};
 
-    sk_free(dst);
-    return NULL;
-}
 
 R11_EACBlitter::R11_EACBlitter(int width, int height, void *latcBuffer)
     // 0x7FFE is one minus the largest positive 16-bit int. We use it for
@@ -1027,11 +860,7 @@ void R11_EACBlitter::flushRuns() {
     // end of a loop.
 
     // Setup:
-    uint32_t c1 = 0;
-    uint32_t c2 = 0;
-    uint32_t c3 = 0;
-    uint32_t c4 = 0;
-
+    uint32_t c[4] = { 0, 0, 0, 0 };
     uint32_t curAlphaColumn = 0;
     SkAlpha *curAlpha = reinterpret_cast<SkAlpha*>(&curAlphaColumn);
 
@@ -1066,10 +895,10 @@ void R11_EACBlitter::flushRuns() {
             const int colsLeft = 4 - col;
             SkASSERT(curX + colsLeft <= finalX);
 
-            update_block_columns(&c1, &c2, &c3, &c4, col, colsLeft, curAlphaColumn);
+            update_block_columns(c, col, colsLeft, curAlphaColumn);
 
             // Write this block
-            *outPtr = compress_block_vertical(c1, c2, c3, c4);
+            *outPtr = compress_block_vertical(c[0], c[1], c[2], c[3]);

[robertphillips, 2014/07/21 21:56:57]

Why not pass the pointer into compress_block_vertical too?

[krajcevski, 2014/07/21 22:03:55]

I tried that, it seemed like it was a tad slower, albeit within margin of error.
Not sure why.

             ++outPtr;
             curX += colsLeft;
         }
@@ -1081,10 +910,10 @@ void R11_EACBlitter::flushRuns() {
             const int col = 0;
             const int colsLeft = kR11_EACBlockSz;
 
-            update_block_columns(&c1, &c2, &c3, &c4, col, colsLeft, curAlphaColumn);
+            update_block_columns(c, col, colsLeft, curAlphaColumn);
 
             // While we can keep advancing, just keep writing the block.
-            uint64_t lastBlock = compress_block_vertical(c1, c2, c3, c4);
+            uint64_t lastBlock = compress_block_vertical(c[0], c[1], c[2], c[3]);
             while((finalX - curX) >= kR11_EACBlockSz) {
                 *outPtr = lastBlock;
                 ++outPtr;
@@ -1097,7 +926,7 @@ void R11_EACBlitter::flushRuns() {
             const int col = curX & 3;
             const int colsLeft = finalX - curX;
 
-            update_block_columns(&c1, &c2, &c3, &c4, col, colsLeft, curAlphaColumn);
+            update_block_columns(c, col, colsLeft, curAlphaColumn);
 
             curX += colsLeft;
         }
@@ -1123,10 +952,14 @@ void R11_EACBlitter::flushRuns() {
 
     // If we didn't land on a block boundary, output the block...
     if ((curX & 3) > 1) {
-        *outPtr = compress_block_vertical(c1, c2, c3, c4);
+        *outPtr = compress_block_vertical(c[0], c[1], c[2], c[3]);
     }
 
     fNextRun = 0;
 }
 
+SkBlitter* CreateR11EACBlitter(int width, int height, void* outputBuffer) {
+    return new R11_EACBlitter(width, height, outputBuffer);
+}
+
 }  // namespace SkTextureCompressor