qdrawhelper_sse2.cpp

Absolute File Name:/home/qt/qt5_coco/qt5/qtbase/src/gui/painting/qdrawhelper_sse2.cpp
Source codeSwitch to Preprocessed file
LineSourceCount
1/****************************************************************************-
2**-
3** Copyright (C) 2015 The Qt Company Ltd.-
4** Contact: http://www.qt.io/licensing/-
5**-
6** This file is part of the QtGui module of the Qt Toolkit.-
7**-
8** $QT_BEGIN_LICENSE:LGPL21$-
9** Commercial License Usage-
10** Licensees holding valid commercial Qt licenses may use this file in-
11** accordance with the commercial license agreement provided with the-
12** Software or, alternatively, in accordance with the terms contained in-
13** a written agreement between you and The Qt Company. For licensing terms-
14** and conditions see http://www.qt.io/terms-conditions. For further-
15** information use the contact form at http://www.qt.io/contact-us.-
16**-
17** GNU Lesser General Public License Usage-
18** Alternatively, this file may be used under the terms of the GNU Lesser-
19** General Public License version 2.1 or version 3 as published by the Free-
20** Software Foundation and appearing in the file LICENSE.LGPLv21 and-
21** LICENSE.LGPLv3 included in the packaging of this file. Please review the-
22** following information to ensure the GNU Lesser General Public License-
23** requirements will be met: https://www.gnu.org/licenses/lgpl.html and-
24** http://www.gnu.org/licenses/old-licenses/lgpl-2.1.html.-
25**-
26** As a special exception, The Qt Company gives you certain additional-
27** rights. These rights are described in The Qt Company LGPL Exception-
28** version 1.1, included in the file LGPL_EXCEPTION.txt in this package.-
29**-
30** $QT_END_LICENSE$-
31**-
32****************************************************************************/-
33-
34#include <private/qdrawhelper_x86_p.h>-
35-
36#ifdef QT_COMPILER_SUPPORTS_SSE2-
37-
38#include <private/qdrawingprimitive_sse2_p.h>-
39#include <private/qpaintengine_raster_p.h>-
40-
41QT_BEGIN_NAMESPACE-
42-
43#ifndef QDRAWHELPER_AVX-
44// in AVX mode, we'll use the SSSE3 code-
45void qt_blend_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,-
46 const uchar *srcPixels, int sbpl,-
47 int w, int h,-
48 int const_alpha)-
49{-
50 const quint32 *src = (const quint32 *) srcPixels;-
51 quint32 *dst = (quint32 *) destPixels;-
52 if (const_alpha == 256) {
const_alpha == 256Description
TRUEnever evaluated
FALSEnever evaluated
0
53 const __m128i alphaMask = _mm_set1_epi32(0xff000000);-
54 const __m128i nullVector = _mm_set1_epi32(0);-
55 const __m128i half = _mm_set1_epi16(0x80);-
56 const __m128i one = _mm_set1_epi16(0xff);-
57 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
58 for (int y = 0; y < h; ++y) {
y < hDescription
TRUEnever evaluated
FALSEnever evaluated
0
59 BLEND_SOURCE_OVER_ARGB32_SSE2(dst, src, w, nullVector, half, one, colorMask, alphaMask);
never executed: dst[x] = s;
never executed: dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
never executed: dst[x] = s;
never executed: dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
never executed: end of block
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
_mm_movemask_e...or)) != 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
s >= 0xff000000Description
TRUEnever evaluated
FALSEnever evaluated
_mm_movemask_e...sk)) == 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
s >= 0xff000000Description
TRUEnever evaluated
FALSEnever evaluated
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
x < w-3Description
TRUEnever evaluated
FALSEnever evaluated
x < wDescription
TRUEnever evaluated
FALSEnever evaluated
0
60 dst = (quint32 *)(((uchar *) dst) + dbpl);-
61 src = (const quint32 *)(((const uchar *) src) + sbpl);-
62 }
never executed: end of block
0
63 } else if (const_alpha != 0) {
never executed: end of block
const_alpha != 0Description
TRUEnever evaluated
FALSEnever evaluated
0
64 // dest = (s + d * sia) * ca + d * cia-
65 // = s * ca + d * (sia * ca + cia)-
66 // = s * ca + d * (1 - sa*ca)-
67 const_alpha = (const_alpha * 255) >> 8;-
68 const __m128i nullVector = _mm_set1_epi32(0);-
69 const __m128i half = _mm_set1_epi16(0x80);-
70 const __m128i one = _mm_set1_epi16(0xff);-
71 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
72 const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);-
73 for (int y = 0; y < h; ++y) {
y < hDescription
TRUEnever evaluated
FALSEnever evaluated
0
74 BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2(dst, src, w, nullVector, half, one, colorMask, constAlphaVector)
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
_mm_movemask_e...or)) != 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
x < w-3Description
TRUEnever evaluated
FALSEnever evaluated
x < wDescription
TRUEnever evaluated
FALSEnever evaluated
0
75 dst = (quint32 *)(((uchar *) dst) + dbpl);-
76 src = (const quint32 *)(((const uchar *) src) + sbpl);-
77 }
never executed: end of block
0
78 }
never executed: end of block
0
79}
never executed: end of block
0
80#endif-
81-
82// qblendfunctions.cpp-
83void qt_blend_rgb32_on_rgb32(uchar *destPixels, int dbpl,-
84 const uchar *srcPixels, int sbpl,-
85 int w, int h,-
86 int const_alpha);-
87-
88void qt_blend_rgb32_on_rgb32_sse2(uchar *destPixels, int dbpl,-
89 const uchar *srcPixels, int sbpl,-
90 int w, int h,-
91 int const_alpha)-
92{-
93 const quint32 *src = (const quint32 *) srcPixels;-
94 quint32 *dst = (quint32 *) destPixels;-
95 if (const_alpha != 256) {
const_alpha != 256Description
TRUEnever evaluated
FALSEnever evaluated
0
96 if (const_alpha != 0) {
const_alpha != 0Description
TRUEnever evaluated
FALSEnever evaluated
0
97 const __m128i nullVector = _mm_set1_epi32(0);-
98 const __m128i half = _mm_set1_epi16(0x80);-
99 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
100-
101 const_alpha = (const_alpha * 255) >> 8;-
102 int one_minus_const_alpha = 255 - const_alpha;-
103 const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);-
104 const __m128i oneMinusConstAlpha = _mm_set1_epi16(one_minus_const_alpha);-
105 for (int y = 0; y < h; ++y) {
y < hDescription
TRUEnever evaluated
FALSEnever evaluated
0
106 int x = 0;-
107-
108 // First, align dest to 16 bytes:-
109 ALIGNMENT_PROLOGUE_16BYTES(dst, x, w) {
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
0
110 dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);-
111 }
never executed: end of block
0
112-
113 for (; x < w-3; x += 4) {
x < w-3Description
TRUEnever evaluated
FALSEnever evaluated
0
114 __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[x]);-
115 if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVector, nullVector)) != 0xffff) {
_mm_movemask_e...or)) != 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
0
116 const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);-
117 __m128i result;-
118 INTERPOLATE_PIXEL_255_SSE2(result, srcVector, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half);-
119 _mm_store_si128((__m128i *)&dst[x], result);-
120 }
never executed: end of block
0
121 }
never executed: end of block
0
122 for (; x<w; ++x) {
x<wDescription
TRUEnever evaluated
FALSEnever evaluated
0
123 dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);-
124 }
never executed: end of block
0
125 dst = (quint32 *)(((uchar *) dst) + dbpl);-
126 src = (const quint32 *)(((const uchar *) src) + sbpl);-
127 }
never executed: end of block
0
128 }
never executed: end of block
0
129 } else {
never executed: end of block
0
130 qt_blend_rgb32_on_rgb32(destPixels, dbpl, srcPixels, sbpl, w, h, const_alpha);-
131 }
never executed: end of block
0
132}-
133-
134void QT_FASTCALL comp_func_SourceOver_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha)-
135{-
136 Q_ASSERT(const_alpha < 256);-
137-
138 const quint32 *src = (const quint32 *) srcPixels;-
139 quint32 *dst = (quint32 *) destPixels;-
140-
141 const __m128i nullVector = _mm_set1_epi32(0);-
142 const __m128i half = _mm_set1_epi16(0x80);-
143 const __m128i one = _mm_set1_epi16(0xff);-
144 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
145 if (const_alpha == 255) {
const_alpha == 255Description
TRUEnever evaluated
FALSEnever evaluated
0
146 const __m128i alphaMask = _mm_set1_epi32(0xff000000);-
147 BLEND_SOURCE_OVER_ARGB32_SSE2(dst, src, length, nullVector, half, one, colorMask, alphaMask);
never executed: dst[x] = s;
never executed: dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
never executed: dst[x] = s;
never executed: dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
never executed: end of block
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
_mm_movemask_e...or)) != 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
s >= 0xff000000Description
TRUEnever evaluated
FALSEnever evaluated
_mm_movemask_e...sk)) == 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
s >= 0xff000000Description
TRUEnever evaluated
FALSEnever evaluated
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
x < length-3Description
TRUEnever evaluated
FALSEnever evaluated
x < lengthDescription
TRUEnever evaluated
FALSEnever evaluated
0
148 } else {
never executed: end of block
0
149 const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);-
150 BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2(dst, src, length, nullVector, half, one, colorMask, constAlphaVector);
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
never executed: end of block
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
_mm_movemask_e...or)) != 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
s != 0Description
TRUEnever evaluated
FALSEnever evaluated
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
x < length-3Description
TRUEnever evaluated
FALSEnever evaluated
x < lengthDescription
TRUEnever evaluated
FALSEnever evaluated
0
151 }
never executed: end of block
0
152}-
153-
154void QT_FASTCALL comp_func_Plus_sse2(uint *dst, const uint *src, int length, uint const_alpha)-
155{-
156 int x = 0;-
157-
158 if (const_alpha == 255) {
const_alpha == 255Description
TRUEnever evaluated
FALSEnever evaluated
0
159 // 1) Prologue: align destination on 16 bytes-
160 ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
0
161 dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
never executed: dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
0
162-
163 // 2) composition with SSE2-
164 for (; x < length - 3; x += 4) {
x < length - 3Description
TRUEnever evaluated
FALSEnever evaluated
0
165 const __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[x]);-
166 const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);-
167-
168 const __m128i result = _mm_adds_epu8(srcVector, dstVector);-
169 _mm_store_si128((__m128i *)&dst[x], result);-
170 }
never executed: end of block
0
171-
172 // 3) Epilogue:-
173 for (; x < length; ++x)
x < lengthDescription
TRUEnever evaluated
FALSEnever evaluated
0
174 dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
never executed: dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
0
175 } else {
never executed: end of block
0
176 const int one_minus_const_alpha = 255 - const_alpha;-
177 const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);-
178 const __m128i oneMinusConstAlpha = _mm_set1_epi16(one_minus_const_alpha);-
179-
180 // 1) Prologue: align destination on 16 bytes-
181 ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
0
182 dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
never executed: dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
0
183-
184 const __m128i half = _mm_set1_epi16(0x80);-
185 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
186 // 2) composition with SSE2-
187 for (; x < length - 3; x += 4) {
x < length - 3Description
TRUEnever evaluated
FALSEnever evaluated
0
188 const __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[x]);-
189 const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);-
190-
191 __m128i result = _mm_adds_epu8(srcVector, dstVector);-
192 INTERPOLATE_PIXEL_255_SSE2(result, result, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half)-
193 _mm_store_si128((__m128i *)&dst[x], result);-
194 }
never executed: end of block
0
195-
196 // 3) Epilogue:-
197 for (; x < length; ++x)
x < lengthDescription
TRUEnever evaluated
FALSEnever evaluated
0
198 dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
never executed: dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
0
199 }
never executed: end of block
0
200}-
201-
202void QT_FASTCALL comp_func_Source_sse2(uint *dst, const uint *src, int length, uint const_alpha)-
203{-
204 if (const_alpha == 255) {
const_alpha == 255Description
TRUEnever evaluated
FALSEnever evaluated
0
205 ::memcpy(dst, src, length * sizeof(uint));-
206 } else {
never executed: end of block
0
207 const int ialpha = 255 - const_alpha;-
208-
209 int x = 0;-
210-
211 // 1) prologue, align on 16 bytes-
212 ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
0
213 dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
never executed: dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
0
214-
215 // 2) interpolate pixels with SSE2-
216 const __m128i half = _mm_set1_epi16(0x80);-
217 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
218 const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);-
219 const __m128i oneMinusConstAlpha = _mm_set1_epi16(ialpha);-
220 for (; x < length - 3; x += 4) {
x < length - 3Description
TRUEnever evaluated
FALSEnever evaluated
0
221 const __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[x]);-
222 __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);-
223 INTERPOLATE_PIXEL_255_SSE2(dstVector, srcVector, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half)-
224 _mm_store_si128((__m128i *)&dst[x], dstVector);-
225 }
never executed: end of block
0
226-
227 // 3) Epilogue-
228 for (; x < length; ++x)
x < lengthDescription
TRUEnever evaluated
FALSEnever evaluated
0
229 dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
never executed: dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
0
230 }
never executed: end of block
0
231}-
232-
233void qt_memfill32(quint32 *dest, quint32 value, int count)-
234{-
235 if (count < 7) {
count < 7Description
TRUEnever evaluated
FALSEnever evaluated
0
236 switch (count) {-
237 case 6: *dest++ = value;
never executed: case 6:
0
238 case 5: *dest++ = value;
code before this statement never executed: case 5:
never executed: case 5:
0
239 case 4: *dest++ = value;
code before this statement never executed: case 4:
never executed: case 4:
0
240 case 3: *dest++ = value;
code before this statement never executed: case 3:
never executed: case 3:
0
241 case 2: *dest++ = value;
code before this statement never executed: case 2:
never executed: case 2:
0
242 case 1: *dest = value;
code before this statement never executed: case 1:
never executed: case 1:
0
243 }
never executed: end of block
0
244 return;
never executed: return;
0
245 };-
246-
247 const int align = (quintptr)(dest) & 0xf;-
248 switch (align) {-
249 case 4: *dest++ = value; --count;
never executed: case 4:
0
250 case 8: *dest++ = value; --count;
code before this statement never executed: case 8:
never executed: case 8:
0
251 case 12: *dest++ = value; --count;
code before this statement never executed: case 12:
never executed: case 12:
0
252 }
never executed: end of block
0
253-
254 const int rest = count & 0x3;-
255 if (rest) {
restDescription
TRUEnever evaluated
FALSEnever evaluated
0
256 switch (rest) {-
257 case 3: dest[count - 3] = value;
never executed: case 3:
0
258 case 2: dest[count - 2] = value;
code before this statement never executed: case 2:
never executed: case 2:
0
259 case 1: dest[count - 1] = value;
code before this statement never executed: case 1:
never executed: case 1:
0
260 }
never executed: end of block
0
261 }
never executed: end of block
0
262-
263 int count128 = count / 4;-
264 __m128i *dst128 = reinterpret_cast<__m128i*>(dest);-
265 __m128i *end128 = dst128 + count128;-
266 const __m128i value128 = _mm_set_epi32(value, value, value, value);-
267-
268 while (dst128 + 3 < end128) {
dst128 + 3 < end128Description
TRUEnever evaluated
FALSEnever evaluated
0
269 _mm_stream_si128(dst128 + 0, value128);-
270 _mm_stream_si128(dst128 + 1, value128);-
271 _mm_stream_si128(dst128 + 2, value128);-
272 _mm_stream_si128(dst128 + 3, value128);-
273 dst128 += 4;-
274 }
never executed: end of block
0
275-
276 switch (count128 & 0x3) {-
277 case 3: _mm_stream_si128(dst128++, value128);
never executed: case 3:
0
278 case 2: _mm_stream_si128(dst128++, value128);
code before this statement never executed: case 2:
never executed: case 2:
0
279 case 1: _mm_stream_si128(dst128++, value128);
code before this statement never executed: case 1:
never executed: case 1:
0
280 }
never executed: end of block
0
281}
never executed: end of block
0
282-
283void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, uint color, uint const_alpha)-
284{-
285 if ((const_alpha & qAlpha(color)) == 255) {
(const_alpha &...color)) == 255Description
TRUEnever evaluated
FALSEnever evaluated
0
286 qt_memfill32(destPixels, color, length);-
287 } else {
never executed: end of block
0
288 if (const_alpha != 255)
const_alpha != 255Description
TRUEnever evaluated
FALSEnever evaluated
0
289 color = BYTE_MUL(color, const_alpha);
never executed: color = BYTE_MUL(color, const_alpha);
0
290-
291 const quint32 minusAlphaOfColor = qAlpha(~color);-
292 int x = 0;-
293-
294 quint32 *dst = (quint32 *) destPixels;-
295 const __m128i colorVector = _mm_set1_epi32(color);-
296 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
297 const __m128i half = _mm_set1_epi16(0x80);-
298 const __m128i minusAlphaOfColorVector = _mm_set1_epi16(minusAlphaOfColor);-
299-
300 ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
0
301 destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
never executed: destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
0
302-
303 for (; x < length-3; x += 4) {
x < length-3Description
TRUEnever evaluated
FALSEnever evaluated
0
304 __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);-
305 BYTE_MUL_SSE2(dstVector, dstVector, minusAlphaOfColorVector, colorMask, half);-
306 dstVector = _mm_add_epi8(colorVector, dstVector);-
307 _mm_store_si128((__m128i *)&dst[x], dstVector);-
308 }
never executed: end of block
0
309 for (;x < length; ++x)
x < lengthDescription
TRUEnever evaluated
FALSEnever evaluated
0
310 destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
never executed: destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
0
311 }
never executed: end of block
0
312}-
313-
314void qt_memfill16(quint16 *dest, quint16 value, int count)-
315{-
316 if (count < 3) {
count < 3Description
TRUEnever evaluated
FALSEnever evaluated
0
317 switch (count) {-
318 case 2: *dest++ = value;
never executed: case 2:
0
319 case 1: *dest = value;
code before this statement never executed: case 1:
never executed: case 1:
0
320 }
never executed: end of block
0
321 return;
never executed: return;
0
322 }-
323-
324 const int align = (quintptr)(dest) & 0x3;-
325 switch (align) {-
326 case 2: *dest++ = value; --count;
never executed: case 2:
0
327 }
never executed: end of block
0
328-
329 const quint32 value32 = (value << 16) | value;-
330 qt_memfill32(reinterpret_cast<quint32*>(dest), value32, count / 2);-
331-
332 if (count & 0x1)
count & 0x1Description
TRUEnever evaluated
FALSEnever evaluated
0
333 dest[count - 1] = value;
never executed: dest[count - 1] = value;
0
334}
never executed: end of block
0
335-
336void qt_bitmapblit32_sse2_base(QRasterBuffer *rasterBuffer, int x, int y,-
337 quint32 color,-
338 const uchar *src, int width, int height, int stride)-
339{-
340 quint32 *dest = reinterpret_cast<quint32*>(rasterBuffer->scanLine(y)) + x;-
341 const int destStride = rasterBuffer->bytesPerLine() / sizeof(quint32);-
342-
343 const __m128i c128 = _mm_set1_epi32(color);-
344 const __m128i maskmask1 = _mm_set_epi32(0x10101010, 0x20202020,-
345 0x40404040, 0x80808080);-
346 const __m128i maskadd1 = _mm_set_epi32(0x70707070, 0x60606060,-
347 0x40404040, 0x00000000);-
348-
349 if (width > 4) {
width > 4Description
TRUEnever evaluated
FALSEnever evaluated
0
350 const __m128i maskmask2 = _mm_set_epi32(0x01010101, 0x02020202,-
351 0x04040404, 0x08080808);-
352 const __m128i maskadd2 = _mm_set_epi32(0x7f7f7f7f, 0x7e7e7e7e,-
353 0x7c7c7c7c, 0x78787878);-
354 while (height--) {
height--Description
TRUEnever evaluated
FALSEnever evaluated
0
355 for (int x = 0; x < width; x += 8) {
x < widthDescription
TRUEnever evaluated
FALSEnever evaluated
0
356 const quint8 s = src[x >> 3];-
357 if (!s)
!sDescription
TRUEnever evaluated
FALSEnever evaluated
0
358 continue;
never executed: continue;
0
359 __m128i mask1 = _mm_set1_epi8(s);-
360 __m128i mask2 = mask1;-
361-
362 mask1 = _mm_and_si128(mask1, maskmask1);-
363 mask1 = _mm_add_epi8(mask1, maskadd1);-
364 _mm_maskmoveu_si128(c128, mask1, (char*)(dest + x));-
365 mask2 = _mm_and_si128(mask2, maskmask2);-
366 mask2 = _mm_add_epi8(mask2, maskadd2);-
367 _mm_maskmoveu_si128(c128, mask2, (char*)(dest + x + 4));-
368 }
never executed: end of block
0
369 dest += destStride;-
370 src += stride;-
371 }
never executed: end of block
0
372 } else {
never executed: end of block
0
373 while (height--) {
height--Description
TRUEnever evaluated
FALSEnever evaluated
0
374 const quint8 s = *src;-
375 if (s) {
sDescription
TRUEnever evaluated
FALSEnever evaluated
0
376 __m128i mask1 = _mm_set1_epi8(s);-
377 mask1 = _mm_and_si128(mask1, maskmask1);-
378 mask1 = _mm_add_epi8(mask1, maskadd1);-
379 _mm_maskmoveu_si128(c128, mask1, (char*)(dest));-
380 }
never executed: end of block
0
381 dest += destStride;-
382 src += stride;-
383 }
never executed: end of block
0
384 }
never executed: end of block
0
385}-
386-
387void qt_bitmapblit32_sse2(QRasterBuffer *rasterBuffer, int x, int y,-
388 const QRgba64 &color,-
389 const uchar *src, int width, int height, int stride)-
390{-
391 qt_bitmapblit32_sse2_base(rasterBuffer, x, y, color.toArgb32(), src, width, height, stride);-
392}
never executed: end of block
0
393-
394void qt_bitmapblit8888_sse2(QRasterBuffer *rasterBuffer, int x, int y,-
395 const QRgba64 &color,-
396 const uchar *src, int width, int height, int stride)-
397{-
398 qt_bitmapblit32_sse2_base(rasterBuffer, x, y, ARGB2RGBA(color.toArgb32()), src, width, height, stride);-
399}
never executed: end of block
0
400-
401void qt_bitmapblit16_sse2(QRasterBuffer *rasterBuffer, int x, int y,-
402 const QRgba64 &color,-
403 const uchar *src, int width, int height, int stride)-
404{-
405 const quint16 c = qConvertRgb32To16(color.toArgb32());-
406 quint16 *dest = reinterpret_cast<quint16*>(rasterBuffer->scanLine(y)) + x;-
407 const int destStride = rasterBuffer->bytesPerLine() / sizeof(quint16);-
408-
409 const __m128i c128 = _mm_set1_epi16(c);-
410QT_WARNING_DISABLE_MSVC(4309) // truncation of constant value-
411 const __m128i maskmask = _mm_set_epi16(0x0101, 0x0202, 0x0404, 0x0808,-
412 0x1010, 0x2020, 0x4040, 0x8080);-
413 const __m128i maskadd = _mm_set_epi16(0x7f7f, 0x7e7e, 0x7c7c, 0x7878,-
414 0x7070, 0x6060, 0x4040, 0x0000);-
415-
416 while (height--) {
height--Description
TRUEnever evaluated
FALSEnever evaluated
0
417 for (int x = 0; x < width; x += 8) {
x < widthDescription
TRUEnever evaluated
FALSEnever evaluated
0
418 const quint8 s = src[x >> 3];-
419 if (!s)
!sDescription
TRUEnever evaluated
FALSEnever evaluated
0
420 continue;
never executed: continue;
0
421 __m128i mask = _mm_set1_epi8(s);-
422 mask = _mm_and_si128(mask, maskmask);-
423 mask = _mm_add_epi8(mask, maskadd);-
424 _mm_maskmoveu_si128(c128, mask, (char*)(dest + x));-
425 }
never executed: end of block
0
426 dest += destStride;-
427 src += stride;-
428 }
never executed: end of block
0
429}
never executed: end of block
0
430-
431class QSimdSse2-
432{-
433public:-
434 typedef __m128i Int32x4;-
435 typedef __m128 Float32x4;-
436-
437 union Vect_buffer_i { Int32x4 v; int i[4]; };-
438 union Vect_buffer_f { Float32x4 v; float f[4]; };-
439-
440 static inline Float32x4 v_dup(float x) { return _mm_set1_ps(x); }
never executed: return _mm_set1_ps(x);
0
441 static inline Float32x4 v_dup(double x) { return _mm_set1_ps(x); }
never executed: return _mm_set1_ps(x);
0
442 static inline Int32x4 v_dup(int x) { return _mm_set1_epi32(x); }
never executed: return _mm_set1_epi32(x);
0
443 static inline Int32x4 v_dup(uint x) { return _mm_set1_epi32(x); }
never executed: return _mm_set1_epi32(x);
0
444-
445 static inline Float32x4 v_add(Float32x4 a, Float32x4 b) { return _mm_add_ps(a, b); }
never executed: return _mm_add_ps(a, b);
0
446 static inline Int32x4 v_add(Int32x4 a, Int32x4 b) { return _mm_add_epi32(a, b); }
never executed: return _mm_add_epi32(a, b);
0
447-
448 static inline Float32x4 v_max(Float32x4 a, Float32x4 b) { return _mm_max_ps(a, b); }
never executed: return _mm_max_ps(a, b);
0
449 static inline Float32x4 v_min(Float32x4 a, Float32x4 b) { return _mm_min_ps(a, b); }
never executed: return _mm_min_ps(a, b);
0
450 static inline Int32x4 v_min_16(Int32x4 a, Int32x4 b) { return _mm_min_epi16(a, b); }
never executed: return _mm_min_epi16(a, b);
0
451-
452 static inline Int32x4 v_and(Int32x4 a, Int32x4 b) { return _mm_and_si128(a, b); }
never executed: return _mm_and_si128(a, b);
0
453-
454 static inline Float32x4 v_sub(Float32x4 a, Float32x4 b) { return _mm_sub_ps(a, b); }
never executed: return _mm_sub_ps(a, b);
0
455 static inline Int32x4 v_sub(Int32x4 a, Int32x4 b) { return _mm_sub_epi32(a, b); }
never executed: return _mm_sub_epi32(a, b);
0
456-
457 static inline Float32x4 v_mul(Float32x4 a, Float32x4 b) { return _mm_mul_ps(a, b); }
never executed: return _mm_mul_ps(a, b);
0
458-
459 static inline Float32x4 v_sqrt(Float32x4 x) { return _mm_sqrt_ps(x); }
never executed: return _mm_sqrt_ps(x);
0
460-
461 static inline Int32x4 v_toInt(Float32x4 x) { return _mm_cvttps_epi32(x); }
never executed: return _mm_cvttps_epi32(x);
0
462-
463 static inline Int32x4 v_greaterOrEqual(Float32x4 a, Float32x4 b) { return _mm_castps_si128(_mm_cmpgt_ps(a, b)); }
never executed: return _mm_castps_si128(_mm_cmpgt_ps(a, b));
0
464};-
465-
466const uint * QT_FASTCALL qt_fetch_radial_gradient_sse2(uint *buffer, const Operator *op, const QSpanData *data,-
467 int y, int x, int length)-
468{-
469 return qt_fetch_radial_gradient_template<QRadialFetchSimd<QSimdSse2>,uint>(buffer, op, data, y, x, length);
never executed: return qt_fetch_radial_gradient_template<QRadialFetchSimd<QSimdSse2>,uint>(buffer, op, data, y, x, length);
0
470}-
471-
472void qt_scale_image_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,-
473 const uchar *srcPixels, int sbpl, int srch,-
474 const QRectF &targetRect,-
475 const QRectF &sourceRect,-
476 const QRect &clip,-
477 int const_alpha)-
478{-
479 if (const_alpha != 256) {
const_alpha != 256Description
TRUEnever evaluated
FALSEnever evaluated
0
480 // from qblendfunctions.cpp-
481 extern void qt_scale_image_argb32_on_argb32(uchar *destPixels, int dbpl,-
482 const uchar *srcPixels, int sbpl, int srch,-
483 const QRectF &targetRect,-
484 const QRectF &sourceRect,-
485 const QRect &clip,-
486 int const_alpha);-
487 return qt_scale_image_argb32_on_argb32(destPixels, dbpl, srcPixels, sbpl, srch, targetRect, sourceRect, clip, const_alpha);
never executed: return qt_scale_image_argb32_on_argb32(destPixels, dbpl, srcPixels, sbpl, srch, targetRect, sourceRect, clip, const_alpha);
0
488 }-
489-
490 qreal sx = targetRect.width() / (qreal) sourceRect.width();-
491 qreal sy = targetRect.height() / (qreal) sourceRect.height();-
492-
493 int ix = 0x00010000 / sx;-
494 int iy = 0x00010000 / sy;-
495-
496 int cx1 = clip.x();-
497 int cx2 = clip.x() + clip.width();-
498 int cy1 = clip.top();-
499 int cy2 = clip.y() + clip.height();-
500-
501 int tx1 = qRound(targetRect.left());-
502 int tx2 = qRound(targetRect.right());-
503 int ty1 = qRound(targetRect.top());-
504 int ty2 = qRound(targetRect.bottom());-
505-
506 if (tx2 < tx1)
tx2 < tx1Description
TRUEnever evaluated
FALSEnever evaluated
0
507 qSwap(tx2, tx1);
never executed: qSwap(tx2, tx1);
0
508 if (ty2 < ty1)
ty2 < ty1Description
TRUEnever evaluated
FALSEnever evaluated
0
509 qSwap(ty2, ty1);
never executed: qSwap(ty2, ty1);
0
510-
511 if (tx1 < cx1)
tx1 < cx1Description
TRUEnever evaluated
FALSEnever evaluated
0
512 tx1 = cx1;
never executed: tx1 = cx1;
0
513 if (tx2 >= cx2)
tx2 >= cx2Description
TRUEnever evaluated
FALSEnever evaluated
0
514 tx2 = cx2;
never executed: tx2 = cx2;
0
515-
516 if (tx1 >= tx2)
tx1 >= tx2Description
TRUEnever evaluated
FALSEnever evaluated
0
517 return;
never executed: return;
0
518-
519 if (ty1 < cy1)
ty1 < cy1Description
TRUEnever evaluated
FALSEnever evaluated
0
520 ty1 = cy1;
never executed: ty1 = cy1;
0
521 if (ty2 >= cy2)
ty2 >= cy2Description
TRUEnever evaluated
FALSEnever evaluated
0
522 ty2 = cy2;
never executed: ty2 = cy2;
0
523 if (ty1 >= ty2)
ty1 >= ty2Description
TRUEnever evaluated
FALSEnever evaluated
0
524 return;
never executed: return;
0
525-
526 int h = ty2 - ty1;-
527 int w = tx2 - tx1;-
528-
529 quint32 basex;-
530 quint32 srcy;-
531-
532 if (sx < 0) {
sx < 0Description
TRUEnever evaluated
FALSEnever evaluated
0
533 int dstx = qFloor((tx1 + qreal(0.5) - targetRect.right()) * ix) + 1;-
534 basex = quint32(sourceRect.right() * 65536) + dstx;-
535 } else {
never executed: end of block
0
536 int dstx = qCeil((tx1 + qreal(0.5) - targetRect.left()) * ix) - 1;-
537 basex = quint32(sourceRect.left() * 65536) + dstx;-
538 }
never executed: end of block
0
539 if (sy < 0) {
sy < 0Description
TRUEnever evaluated
FALSEnever evaluated
0
540 int dsty = qFloor((ty1 + qreal(0.5) - targetRect.bottom()) * iy) + 1;-
541 srcy = quint32(sourceRect.bottom() * 65536) + dsty;-
542 } else {
never executed: end of block
0
543 int dsty = qCeil((ty1 + qreal(0.5) - targetRect.top()) * iy) - 1;-
544 srcy = quint32(sourceRect.top() * 65536) + dsty;-
545 }
never executed: end of block
0
546-
547 quint32 *dst = ((quint32 *) (destPixels + ty1 * dbpl)) + tx1;-
548-
549 const __m128i nullVector = _mm_set1_epi32(0);-
550 const __m128i half = _mm_set1_epi16(0x80);-
551 const __m128i one = _mm_set1_epi16(0xff);-
552 const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);-
553 const __m128i alphaMask = _mm_set1_epi32(0xff000000);-
554 const __m128i ixVector = _mm_set1_epi32(4*ix);-
555-
556 // this bounds check here is required as floating point rounding above might in some cases lead to-
557 // w/h values that are one pixel too large, falling outside of the valid image area.-
558 int yend = (srcy + iy * (h - 1)) >> 16;-
559 if (yend < 0 || yend >= srch)
yend < 0Description
TRUEnever evaluated
FALSEnever evaluated
yend >= srchDescription
TRUEnever evaluated
FALSEnever evaluated
0
560 --h;
never executed: --h;
0
561 int xend = (basex + ix * (w - 1)) >> 16;-
562 if (xend < 0 || xend >= (int)(sbpl/sizeof(quint32)))
xend < 0Description
TRUEnever evaluated
FALSEnever evaluated
xend >= (int)(...zeof(quint32))Description
TRUEnever evaluated
FALSEnever evaluated
0
563 --w;
never executed: --w;
0
564-
565 while (h--) {
h--Description
TRUEnever evaluated
FALSEnever evaluated
0
566 const uint *src = (const quint32 *) (srcPixels + (srcy >> 16) * sbpl);-
567 int srcx = basex;-
568 int x = 0;-
569-
570 ALIGNMENT_PROLOGUE_16BYTES(dst, x, w) {
x < static_cas...0x3)) & 0x3)))Description
TRUEnever evaluated
FALSEnever evaluated
0
571 uint s = src[srcx >> 16];-
572 dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));-
573 srcx += ix;-
574 }
never executed: end of block
0
575-
576 __m128i srcxVector = _mm_set_epi32(srcx, srcx + ix, srcx + ix + ix, srcx + ix + ix + ix);-
577-
578 for (; x<w - 3; x += 4) {
x<w - 3Description
TRUEnever evaluated
FALSEnever evaluated
0
579 union Vect_buffer { __m128i vect; quint32 i[4]; };-
580 Vect_buffer addr;-
581 addr.vect = _mm_srli_epi32(srcxVector, 16);-
582 srcxVector = _mm_add_epi32(srcxVector, ixVector);-
583-
584 const __m128i srcVector = _mm_set_epi32(src[addr.i[0]], src[addr.i[1]], src[addr.i[2]], src[addr.i[3]]);-
585 BLEND_SOURCE_OVER_ARGB32_SSE2_helper(dst, srcVector, nullVector, half, one, colorMask, alphaMask);
never executed: end of block
never executed: end of block
_mm_movemask_e...or)) != 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
_mm_movemask_e...sk)) == 0xffffDescription
TRUEnever evaluated
FALSEnever evaluated
0
586 }
never executed: end of block
0
587-
588 for (; x<w; x++) {
x<wDescription
TRUEnever evaluated
FALSEnever evaluated
0
589 uint s = src[(basex + x*ix) >> 16];-
590 dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));-
591 }
never executed: end of block
0
592 dst = (quint32 *)(((uchar *) dst) + dbpl);-
593 srcy += iy;-
594 }
never executed: end of block
0
595}
never executed: end of block
0
596-
597-
598QT_END_NAMESPACE-
599-
600#endif // QT_COMPILER_SUPPORTS_SSE2-
Source codeSwitch to Preprocessed file
Generated by Squish Coco Non-Commercial 4.3.0-BETA-master-30-08-2018-4cb69e9