Add texture compression utility

src/utils/SkTextureCompressor.cpp

Index: src/utils/SkTextureCompressor.cpp
diff --git a/src/utils/SkTextureCompressor.cpp b/src/utils/SkTextureCompressor.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..645de54e79ba14eb7004636c5767304100bbb28a
--- /dev/null
+++ b/src/utils/SkTextureCompressor.cpp
@@ -0,0 +1,189 @@
+/*
+ * Copyright 2014 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkTextureCompressor.h"
+
+#include "SkBitmap.h"
+#include "SkData.h"
+#include "SkEndian.h"
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// LATC compressor
+//
+////////////////////////////////////////////////////////////////////////////////
+
+// Return the squared minimum error cost of approximating 'pixel' using the
+// provided palette.
+static uint16_t compute_error(uint16_t pixel, uint8_t palette[8]) {
+    uint16_t error = 65535;
+    for (int i = 0; i < 8; ++i) {
+        uint16_t diff = SkTAbsDiff(static_cast<uint16_t>(palette[i]), pixel);
+        error = SkMin32(diff, error);
+    }
+    return error*error;
+}
+
+// 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. 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
+//
+// We compute the LATC palette using the following simple algorithm:
+// 1. Choose the minimum and maximum values in the block as LUM0 and LUM1
+// 2. Figure out which of the two possible palettes is better.
+
+static uint64_t compress_latc_block(uint8_t block[16]) {
+    // Just do a simple min/max but choose from which of the
+    // two palettes is better
+    uint8_t maxVal = 0;
+    uint8_t minVal = 255;
+    for (int i = 0; i < 16; ++i) {
+        maxVal = SkMax32(maxVal, block[i]);
+        minVal = SkMin32(minVal, block[i]);
+    }
+
+    // Generate palettes
+    uint8_t palettes[2][8];
+
+    // Straight linear ramp
+    palettes[0][0] = maxVal;
+    palettes[0][1] = minVal;
+    for (int i = 1; i < 7; ++i) {
+        palettes[0][i+1] = ((7-i)*maxVal + i*minVal) / 7;
+    }
+
+    // Smaller linear ramp with min and max byte values at the end.
+    palettes[1][0] = minVal;
+    palettes[1][1] = maxVal;
+    for (int i = 1; i < 5; ++i) {
+        palettes[1][i+1] = ((5-i)*maxVal + i*minVal) / 5;
+    }
+    palettes[1][6] = 0;
+    palettes[1][7] = 255;
+
+    // Figure out which of the two is better
+    uint16_t error[2] = { 0, 0 };
+    for (int i = 0; i < 16; ++i) {
+        uint16_t pixel = static_cast<uint16_t>(block[i]);
+        error[0] += compute_error(pixel, palettes[0]);
+        error[1] += compute_error(pixel, palettes[1]);
+    }
+    uint8_t *palette = (error[0] > error[1]) ? palettes[1] : palettes[0];
+
+    // Compress the chosen palette
+    uint64_t result = 0;
+    result |= static_cast<uint64_t>(palette[0]);
+    result |= static_cast<uint64_t>(palette[1]) << 8;
+
+    uint64_t indices = 0;
+    for (int i = 15; i >= 0; --i) {
+        uint8_t pixel = block[i];
+
+        // The palette is selected in the loop above. This loop
+        // is to determine what index into the palette to store for
+        // each pixel.
+        uint8_t minError = SkTAbsDiff(pixel, palette[0]);
+        uint8_t minErrorIndex = 0;
+        for (int j = 1; j < 8; ++j) {
+            uint8_t error = SkTAbsDiff(pixel, palette[j]);
+            if (error < minError) {
+                minError = error;
+                minErrorIndex = j;
+            }
+        }
+        SkASSERT(minErrorIndex < 8);
+
+        indices <<= 3;
+        indices |= minErrorIndex;
+    }
+    result |= indices << 16;
+
+    // We assume everything is little endian, if it's not then make it so.
+    return SkEndian_SwapLE64(result);
+}
+
+static SkData *compress_a8_to_latc(const SkBitmap &bm) {
+    // LATC compressed texels down into square 4x4 blocks
+    static const int kLATCBlockSize = 4;
+
+    // Make sure that our data is well-formed enough to be
+    // considered for LATC compression
+    if (bm.width() == 0 || bm.height() == 0 ||
+        (bm.width() % kLATCBlockSize) != 0 ||
+        (bm.height() % kLATCBlockSize) != 0 ||
+        (bm.config() != SkBitmap::kA8_Config)) {

[scroggo, 2014/06/09 20:54:18]

Use colorType() instead of config(). We are removing config().

[krajcevski, 2014/06/09 21:06:02]

Done.

+        return NULL;
+    }
+
+    // The LATC format is 64 bits per 4x4 block.
+    static const int kLATCEncodedBlockSize = 8;
+
+    int blocksX = bm.width() / kLATCBlockSize;
+    int blocksY = bm.height() / kLATCBlockSize;
+
+    int compressedDataSize = blocksX * blocksY * kLATCEncodedBlockSize;
+    uint64_t* dst = reinterpret_cast<uint64_t*>(sk_malloc_throw(compressedDataSize));
+
+    uint8_t block[16];
+    const uint8_t* row = reinterpret_cast<const uint8_t*>(bm.getPixels());
+    uint64_t* encPtr = dst;
+    for (int y = 0; y < blocksY; ++y) {
+        for (int x = 0; x < blocksX; ++x) {
+            memcpy(block, row + (kLATCBlockSize * x), 4);
+            memcpy(block + 4, row + bm.rowBytes() + (kLATCBlockSize * x), 4);
+            memcpy(block + 8, row + 2*bm.rowBytes() + (kLATCBlockSize * x), 4);
+            memcpy(block + 12, row + 3*bm.rowBytes() + (kLATCBlockSize * x), 4);
+
+            *encPtr = compress_latc_block(block);
+            ++encPtr;
+        }
+        row += kLATCBlockSize * bm.rowBytes();
+    }
+
+    return SkData::NewFromMalloc(dst, compressedDataSize);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+namespace SkTextureCompressor {
+
+typedef SkData *(*CompressBitmapProc)(const SkBitmap &bitmap);
+
+SkData *CompressBitmapToConfig(const SkBitmap &bitmap, Format config) {

[scroggo, 2014/06/09 20:54:18]

Shouldn't this be ToFormat, as in the header file?

[krajcevski, 2014/06/09 21:06:02]

Yes, whoops. Done.

+    CompressBitmapProc kProcMap[SkBitmap::kConfigCount][kFormatCnt];
+    memset(kProcMap, 0, sizeof(kProcMap));
+
+    // Map available bitmap configs to compression functions
+    kProcMap[SkBitmap::kA8_Config][kLATC_Format] = compress_a8_to_latc;
+
+    CompressBitmapProc proc = kProcMap[bitmap.config()][config];
+    if (NULL != proc) {
+        return proc(bitmap);
+    }
+
+    return NULL;
+}
+
+}  // namespace SkTextureCompressor