Handle gray alpha conversions in SkSwizzler

src/codec/SkSwizzler.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 "SkOpts.h"
@@ skipping 277 matching lines @@
         int bytesPerPixel, int deltaSrc, int offset, const SkPMColor ctable[]) {
 
     src += offset;
     uint16_t* SK_RESTRICT dst = (uint16_t*)dstRow;
     for (int x = 0; x < dstWidth; x++) {
         dst[x] = SkPack888ToRGB16(src[0], src[0], src[0]);
         src += deltaSrc;
     }
 }
 
+// kGrayAlpha
+
+static void swizzle_grayalpha_to_n32_unpremul(
+        void* dst, const uint8_t* src, int width, int bpp, int deltaSrc, int offset,
+        const SkPMColor ctable[]) {
+
+    src += offset;
+    SkPMColor* dst32 = (SkPMColor*) dst;
+    for (int x = 0; x < width; x++) {
+        dst32[x] = SkPackARGB32NoCheck(src[1], src[0], src[0], src[0]);
+        src += deltaSrc;
+    }
+}
+
+static void swizzle_grayalpha_to_n32_premul(
+        void* dst, const uint8_t* src, int width, int bpp, int deltaSrc, int offset,
+        const SkPMColor ctable[]) {
+
+    src += offset;
+    SkPMColor* dst32 = (SkPMColor*) dst;
+    for (int x = 0; x < width; x++) {
+        uint8_t pmgray = SkMulDiv255Round(src[1], src[0]);
+        dst32[x] = SkPackARGB32NoCheck(src[1], pmgray, pmgray, pmgray);
+        src += deltaSrc;
+    }
+}
+
 // kBGRX
 
 static void swizzle_bgrx_to_n32(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int dstWidth,
         int bpp, int deltaSrc, int offset, const SkPMColor ctable[]) {
 
     src += offset;
     SkPMColor* SK_RESTRICT dst = (SkPMColor*)dstRow;
     for (int x = 0; x < dstWidth; x++) {
         dst[x] = SkPackARGB32NoCheck(0xFF, src[2], src[1], src[0]);
@@ skipping 241 matching lines @@
         const uint8_t r = SkMulDiv255Round(src[0], src[3]);
         const uint8_t g = SkMulDiv255Round(src[1], src[3]);
         const uint8_t b = SkMulDiv255Round(src[2], src[3]);
 
         dst[x] = SkPack888ToRGB16(r, g, b);
         src += deltaSrc;
     }
 }
 
 template <SkSwizzler::RowProc proc>
+void SkSwizzler::SkipLeadingGrayAlphaZerosThen(
+        void* dst, const uint8_t* src, int width,
+        int bpp, int deltaSrc, int offset, const SkPMColor ctable[]) {
+    SkASSERT(!ctable);
+
+    const uint16_t* src16 = (const uint16_t*) (src + offset);
+    uint32_t* dst32 = (uint32_t*) dst;
+
+    // This may miss opportunities to skip when the output is premultiplied,
+    // e.g. for a src pixel 0x00FF which is not zero but becomes zero after premultiplication.
+    while (width > 0 && *src16 == 0x0000) {

[scroggo, 2016/02/03 16:11:33]

Should this have two exit conditions?

- width == 0 : return;
- *src16 == 0x0000 : break (to call proc)

Think that would be faster, slower, or the same?

[msarett, 2016/02/03 17:21:13]

I believe it would be the same.  I just checked the assembly and "proc" gets
inlined.  So the compiler can (and does) make the same optimization that you are
suggesting :).

+        width--;
+        dst32++;
+        src16 += deltaSrc / 2;
+    }
+    proc(dst32, (const uint8_t*)src16, width, bpp, deltaSrc, 0, ctable);
+}
+
+template <SkSwizzler::RowProc proc>
 void SkSwizzler::SkipLeading8888ZerosThen(
         void* SK_RESTRICT dstRow, const uint8_t* SK_RESTRICT src, int dstWidth,
         int bpp, int deltaSrc, int offset, const SkPMColor ctable[]) {
     SkASSERT(!ctable);
 
     auto src32 = (const uint32_t*)(src+offset);
     auto dst32 = (uint32_t*)dstRow;
 
     // This may miss opportunities to skip when the output is premultiplied,
     // e.g. for a src pixel 0x00FFFFFF which is not zero but becomes zero after premultiplication.
@@ skipping 89 matching lines @@
                     proc = &sample1;
                     fastProc = &copy;
                     break;
                 case kRGB_565_SkColorType:
                     proc = &swizzle_gray_to_565;
                     break;
                 default:
                     break;
             }
             break;
+        case kGrayAlpha:
+            switch (dstInfo.colorType()) {
+                case kN32_SkColorType:
+                    if (dstInfo.alphaType() == kUnpremul_SkAlphaType) {
+                        if (SkCodec::kYes_ZeroInitialized == zeroInit) {
+                            proc = &SkipLeadingGrayAlphaZerosThen
+                                    <swizzle_grayalpha_to_n32_unpremul>;
+                        } else {
+                            proc = &swizzle_grayalpha_to_n32_unpremul;
+                        }
+                    } else {
+                        if (SkCodec::kYes_ZeroInitialized == zeroInit) {
+                            proc = &SkipLeadingGrayAlphaZerosThen<swizzle_grayalpha_to_n32_premul>;
+                        } else {
+                            proc = &swizzle_grayalpha_to_n32_premul;
+                        }
+                    }
+                    break;
+                default:
+                    break;
+            }
+            break;
         case kBGR:
         case kBGRX:
             switch (dstInfo.colorType()) {
                 case kN32_SkColorType:
                     proc = &swizzle_bgrx_to_n32;
                     break;
                 case kRGB_565_SkColorType:
                     proc = &swizzle_bgrx_to_565;
                     break;
                 default:
@@ skipping 150 matching lines @@
     }
 
     return fAllocatedWidth;
 }
 
 void SkSwizzler::swizzle(void* dst, const uint8_t* SK_RESTRICT src) {
     SkASSERT(nullptr != dst && nullptr != src);
     fActualProc(SkTAddOffset<void>(dst, fDstOffsetBytes), src, fSwizzleWidth, fSrcBPP,
             fSampleX * fSrcBPP, fSrcOffsetUnits, fColorTable);
 }