michael@0: /*
michael@0:  * Copyright 2011 Google Inc.
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: #include "SkBlitRow.h"
michael@0: #include "SkBlitMask.h"
michael@0: #include "SkColorPriv.h"
michael@0: #include "SkUtils.h"
michael@0: 
michael@0: #define UNROLL
michael@0: 
michael@0: SkBlitRow::ColorRectProc PlatformColorRectProcFactory();
michael@0: 
michael@0: static void S32_Opaque_BlitRow32(SkPMColor* SK_RESTRICT dst,
michael@0:                                  const SkPMColor* SK_RESTRICT src,
michael@0:                                  int count, U8CPU alpha) {
michael@0:     SkASSERT(255 == alpha);
michael@0:     memcpy(dst, src, count * sizeof(SkPMColor));
michael@0: }
michael@0: 
michael@0: static void S32_Blend_BlitRow32(SkPMColor* SK_RESTRICT dst,
michael@0:                                 const SkPMColor* SK_RESTRICT src,
michael@0:                                 int count, U8CPU alpha) {
michael@0:     SkASSERT(alpha <= 255);
michael@0:     if (count > 0) {
michael@0:         unsigned src_scale = SkAlpha255To256(alpha);
michael@0:         unsigned dst_scale = 256 - src_scale;
michael@0: 
michael@0: #ifdef UNROLL
michael@0:         if (count & 1) {
michael@0:             *dst = SkAlphaMulQ(*(src++), src_scale) + SkAlphaMulQ(*dst, dst_scale);
michael@0:             dst += 1;
michael@0:             count -= 1;
michael@0:         }
michael@0: 
michael@0:         const SkPMColor* SK_RESTRICT srcEnd = src + count;
michael@0:         while (src != srcEnd) {
michael@0:             *dst = SkAlphaMulQ(*(src++), src_scale) + SkAlphaMulQ(*dst, dst_scale);
michael@0:             dst += 1;
michael@0:             *dst = SkAlphaMulQ(*(src++), src_scale) + SkAlphaMulQ(*dst, dst_scale);
michael@0:             dst += 1;
michael@0:         }
michael@0: #else
michael@0:         do {
michael@0:             *dst = SkAlphaMulQ(*src, src_scale) + SkAlphaMulQ(*dst, dst_scale);
michael@0:             src += 1;
michael@0:             dst += 1;
michael@0:         } while (--count > 0);
michael@0: #endif
michael@0:     }
michael@0: }
michael@0: 
michael@0: static void S32A_Opaque_BlitRow32(SkPMColor* SK_RESTRICT dst,
michael@0:                                   const SkPMColor* SK_RESTRICT src,
michael@0:                                   int count, U8CPU alpha) {
michael@0:     SkASSERT(255 == alpha);
michael@0:     if (count > 0) {
michael@0: #ifdef UNROLL
michael@0:         if (count & 1) {
michael@0:             *dst = SkPMSrcOver(*(src++), *dst);
michael@0:             dst += 1;
michael@0:             count -= 1;
michael@0:         }
michael@0: 
michael@0:         const SkPMColor* SK_RESTRICT srcEnd = src + count;
michael@0:         while (src != srcEnd) {
michael@0:             *dst = SkPMSrcOver(*(src++), *dst);
michael@0:             dst += 1;
michael@0:             *dst = SkPMSrcOver(*(src++), *dst);
michael@0:             dst += 1;
michael@0:         }
michael@0: #else
michael@0:         do {
michael@0:             *dst = SkPMSrcOver(*src, *dst);
michael@0:             src += 1;
michael@0:             dst += 1;
michael@0:         } while (--count > 0);
michael@0: #endif
michael@0:     }
michael@0: }
michael@0: 
michael@0: static void S32A_Blend_BlitRow32(SkPMColor* SK_RESTRICT dst,
michael@0:                                  const SkPMColor* SK_RESTRICT src,
michael@0:                                  int count, U8CPU alpha) {
michael@0:     SkASSERT(alpha <= 255);
michael@0:     if (count > 0) {
michael@0: #ifdef UNROLL
michael@0:         if (count & 1) {
michael@0:             *dst = SkBlendARGB32(*(src++), *dst, alpha);
michael@0:             dst += 1;
michael@0:             count -= 1;
michael@0:         }
michael@0: 
michael@0:         const SkPMColor* SK_RESTRICT srcEnd = src + count;
michael@0:         while (src != srcEnd) {
michael@0:             *dst = SkBlendARGB32(*(src++), *dst, alpha);
michael@0:             dst += 1;
michael@0:             *dst = SkBlendARGB32(*(src++), *dst, alpha);
michael@0:             dst += 1;
michael@0:         }
michael@0: #else
michael@0:         do {
michael@0:             *dst = SkBlendARGB32(*src, *dst, alpha);
michael@0:             src += 1;
michael@0:             dst += 1;
michael@0:         } while (--count > 0);
michael@0: #endif
michael@0:     }
michael@0: }
michael@0: 
michael@0: ///////////////////////////////////////////////////////////////////////////////
michael@0: 
michael@0: static const SkBlitRow::Proc32 gDefault_Procs32[] = {
michael@0:     S32_Opaque_BlitRow32,
michael@0:     S32_Blend_BlitRow32,
michael@0:     S32A_Opaque_BlitRow32,
michael@0:     S32A_Blend_BlitRow32
michael@0: };
michael@0: 
michael@0: SkBlitRow::Proc32 SkBlitRow::Factory32(unsigned flags) {
michael@0:     SkASSERT(flags < SK_ARRAY_COUNT(gDefault_Procs32));
michael@0:     // just so we don't crash
michael@0:     flags &= kFlags32_Mask;
michael@0: 
michael@0:     SkBlitRow::Proc32 proc = PlatformProcs32(flags);
michael@0:     if (NULL == proc) {
michael@0:         proc = gDefault_Procs32[flags];
michael@0:     }
michael@0:     SkASSERT(proc);
michael@0:     return proc;
michael@0: }
michael@0: 
michael@0: SkBlitRow::Proc32 SkBlitRow::ColorProcFactory() {
michael@0:     SkBlitRow::ColorProc proc = PlatformColorProc();
michael@0:     if (NULL == proc) {
michael@0:         proc = Color32;
michael@0:     }
michael@0:     SkASSERT(proc);
michael@0:     return proc;
michael@0: }
michael@0: 
michael@0: void SkBlitRow::Color32(SkPMColor* SK_RESTRICT dst,
michael@0:                         const SkPMColor* SK_RESTRICT src,
michael@0:                         int count, SkPMColor color) {
michael@0:     if (count > 0) {
michael@0:         if (0 == color) {
michael@0:             if (src != dst) {
michael@0:                 memcpy(dst, src, count * sizeof(SkPMColor));
michael@0:             }
michael@0:             return;
michael@0:         }
michael@0:         unsigned colorA = SkGetPackedA32(color);
michael@0:         if (255 == colorA) {
michael@0:             sk_memset32(dst, color, count);
michael@0:         } else {
michael@0:             unsigned scale = 256 - SkAlpha255To256(colorA);
michael@0:             do {
michael@0:                 *dst = color + SkAlphaMulQ(*src, scale);
michael@0:                 src += 1;
michael@0:                 dst += 1;
michael@0:             } while (--count);
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: template <size_t N> void assignLoop(SkPMColor* dst, SkPMColor color) {
michael@0:     for (size_t i = 0; i < N; ++i) {
michael@0:         *dst++ = color;
michael@0:     }
michael@0: }
michael@0: 
michael@0: static inline void assignLoop(SkPMColor dst[], SkPMColor color, int count) {
michael@0:     while (count >= 4) {
michael@0:         *dst++ = color;
michael@0:         *dst++ = color;
michael@0:         *dst++ = color;
michael@0:         *dst++ = color;
michael@0:         count -= 4;
michael@0:     }
michael@0:     if (count >= 2) {
michael@0:         *dst++ = color;
michael@0:         *dst++ = color;
michael@0:         count -= 2;
michael@0:     }
michael@0:     if (count > 0) {
michael@0:         *dst++ = color;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void SkBlitRow::ColorRect32(SkPMColor* dst, int width, int height,
michael@0:                             size_t rowBytes, SkPMColor color) {
michael@0:     if (width <= 0 || height <= 0 || 0 == color) {
michael@0:         return;
michael@0:     }
michael@0: 
michael@0:     // Just made up this value, since I saw it once in a SSE2 file.
michael@0:     // We should consider writing some tests to find the optimimal break-point
michael@0:     // (or query the Platform proc?)
michael@0:     static const int MIN_WIDTH_FOR_SCANLINE_PROC = 32;
michael@0: 
michael@0:     bool isOpaque = (0xFF == SkGetPackedA32(color));
michael@0: 
michael@0:     if (!isOpaque || width >= MIN_WIDTH_FOR_SCANLINE_PROC) {
michael@0:         SkBlitRow::ColorProc proc = SkBlitRow::ColorProcFactory();
michael@0:         while (--height >= 0) {
michael@0:             (*proc)(dst, dst, width, color);
michael@0:             dst = (SkPMColor*) ((char*)dst + rowBytes);
michael@0:         }
michael@0:     } else {
michael@0:         switch (width) {
michael@0:             case 1:
michael@0:                 while (--height >= 0) {
michael@0:                     assignLoop<1>(dst, color);
michael@0:                     dst = (SkPMColor*) ((char*)dst + rowBytes);
michael@0:                 }
michael@0:                 break;
michael@0:             case 2:
michael@0:                 while (--height >= 0) {
michael@0:                     assignLoop<2>(dst, color);
michael@0:                     dst = (SkPMColor*) ((char*)dst + rowBytes);
michael@0:                 }
michael@0:                 break;
michael@0:             case 3:
michael@0:                 while (--height >= 0) {
michael@0:                     assignLoop<3>(dst, color);
michael@0:                     dst = (SkPMColor*) ((char*)dst + rowBytes);
michael@0:                 }
michael@0:                 break;
michael@0:             default:
michael@0:                 while (--height >= 0) {
michael@0:                     assignLoop(dst, color, width);
michael@0:                     dst = (SkPMColor*) ((char*)dst + rowBytes);
michael@0:                 }
michael@0:                 break;
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: SkBlitRow::ColorRectProc SkBlitRow::ColorRectProcFactory() {
michael@0:     SkBlitRow::ColorRectProc proc = PlatformColorRectProcFactory();
michael@0:     if (NULL == proc) {
michael@0:         proc = ColorRect32;
michael@0:     }
michael@0:     SkASSERT(proc);
michael@0:     return proc;
michael@0: }