| changeset 37 | 758a864f9613 |
| parent 33 | 3e2da88830cd |
| 36:ef0373b55136 | 37:758a864f9613 |
|---|---|
141 // with shortcuts if fully opaque or fully transparent. |
141 // with shortcuts if fully opaque or fully transparent. |
142 #define BLEND_SOURCE_OVER_ARGB32_SSE2(dst, src, length, nullVector, half, one, colorMask, alphaMask) { \ |
142 #define BLEND_SOURCE_OVER_ARGB32_SSE2(dst, src, length, nullVector, half, one, colorMask, alphaMask) { \ |
143 int x = 0; \ |
143 int x = 0; \ |
144 \ |
144 \ |
145 /* First, get dst aligned. */ \ |
145 /* First, get dst aligned. */ \ |
146 const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;\ |
146 ALIGNMENT_PROLOGUE_16BYTES(dst, x, length) { \ |
147 const int prologLength = qMin(length, offsetToAlignOn16Bytes);\ |
|
148 for (; x < prologLength; ++x) { \ |
|
149 uint s = src[x]; \ |
147 uint s = src[x]; \ |
150 if (s >= 0xff000000) \ |
148 if (s >= 0xff000000) \ |
151 dst[x] = s; \ |
149 dst[x] = s; \ |
152 else if (s != 0) \ |
150 else if (s != 0) \ |
153 dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s)); \ |
151 dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s)); \ |
200 // = s * ca + d * (1 - sa*ca) |
198 // = s * ca + d * (1 - sa*ca) |
201 #define BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2(dst, src, length, nullVector, half, one, colorMask, constAlphaVector) \ |
199 #define BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2(dst, src, length, nullVector, half, one, colorMask, constAlphaVector) \ |
202 { \ |
200 { \ |
203 int x = 0; \ |
201 int x = 0; \ |
204 \ |
202 \ |
205 const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;\ |
203 ALIGNMENT_PROLOGUE_16BYTES(dst, x, length) { \ |
206 const int prologLength = qMin(length, offsetToAlignOn16Bytes);\ |
|
207 for (; x < prologLength; ++x) { \ |
|
208 quint32 s = src[x]; \ |
204 quint32 s = src[x]; \ |
209 if (s != 0) { \ |
205 if (s != 0) { \ |
210 s = BYTE_MUL(s, const_alpha); \ |
206 s = BYTE_MUL(s, const_alpha); \ |
211 dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s)); \ |
207 dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s)); \ |
212 } \ |
208 } \ |