1 | // Copyright (C) 2016 The Qt Company Ltd. |
2 | // SPDX-License-Identifier: LicenseRef-Qt-Commercial OR LGPL-3.0-only OR GPL-2.0-only OR GPL-3.0-only |
3 | |
4 | #ifndef QDRAWINGPRIMITIVE_SSE2_P_H |
5 | #define QDRAWINGPRIMITIVE_SSE2_P_H |
6 | |
7 | #include <QtGui/private/qtguiglobal_p.h> |
8 | #include <private/qsimd_p.h> |
9 | #include "qdrawhelper_x86_p.h" |
10 | #include "qrgba64_p.h" |
11 | |
12 | #ifdef __SSE2__ |
13 | |
14 | // |
15 | // W A R N I N G |
16 | // ------------- |
17 | // |
18 | // This file is not part of the Qt API. It exists purely as an |
19 | // implementation detail. This header file may change from version to |
20 | // version without notice, or even be removed. |
21 | // |
22 | // We mean it. |
23 | // |
24 | |
25 | QT_BEGIN_NAMESPACE |
26 | |
27 | /* |
28 | * Multiply the components of pixelVector by alphaChannel |
29 | * Each 32bits components of alphaChannel must be in the form 0x00AA00AA |
30 | * colorMask must have 0x00ff00ff on each 32 bits component |
31 | * half must have the value 128 (0x80) for each 32 bits component |
32 | */ |
33 | #define BYTE_MUL_SSE2(result, pixelVector, alphaChannel, colorMask, half) \ |
34 | { \ |
35 | /* 1. separate the colors in 2 vectors so each color is on 16 bits \ |
36 | (in order to be multiplied by the alpha \ |
37 | each 32 bit of dstVectorAG are in the form 0x00AA00GG \ |
38 | each 32 bit of dstVectorRB are in the form 0x00RR00BB */\ |
39 | __m128i pixelVectorAG = _mm_srli_epi16(pixelVector, 8); \ |
40 | __m128i pixelVectorRB = _mm_and_si128(pixelVector, colorMask); \ |
41 | \ |
42 | /* 2. multiply the vectors by the alpha channel */\ |
43 | pixelVectorAG = _mm_mullo_epi16(pixelVectorAG, alphaChannel); \ |
44 | pixelVectorRB = _mm_mullo_epi16(pixelVectorRB, alphaChannel); \ |
45 | \ |
46 | /* 3. divide by 255, that's the tricky part. \ |
47 | we do it like for BYTE_MUL(), with bit shift: X/255 ~= (X + X/256 + rounding)/256 */ \ |
48 | /** so first (X + X/256 + rounding) */\ |
49 | pixelVectorRB = _mm_add_epi16(pixelVectorRB, _mm_srli_epi16(pixelVectorRB, 8)); \ |
50 | pixelVectorRB = _mm_add_epi16(pixelVectorRB, half); \ |
51 | pixelVectorAG = _mm_add_epi16(pixelVectorAG, _mm_srli_epi16(pixelVectorAG, 8)); \ |
52 | pixelVectorAG = _mm_add_epi16(pixelVectorAG, half); \ |
53 | \ |
54 | /** second divide by 256 */\ |
55 | pixelVectorRB = _mm_srli_epi16(pixelVectorRB, 8); \ |
56 | /** for AG, we could >> 8 to divide followed by << 8 to put the \ |
57 | bytes in the correct position. By masking instead, we execute \ |
58 | only one instruction */\ |
59 | pixelVectorAG = _mm_andnot_si128(colorMask, pixelVectorAG); \ |
60 | \ |
61 | /* 4. combine the 2 pairs of colors */ \ |
62 | result = _mm_or_si128(pixelVectorAG, pixelVectorRB); \ |
63 | } |
64 | |
65 | /* |
66 | * Each 32bits components of alphaChannel must be in the form 0x00AA00AA |
67 | * oneMinusAlphaChannel must be 255 - alpha for each 32 bits component |
68 | * colorMask must have 0x00ff00ff on each 32 bits component |
69 | * half must have the value 128 (0x80) for each 32 bits component |
70 | */ |
71 | #define INTERPOLATE_PIXEL_255_SSE2(result, srcVector, dstVector, alphaChannel, oneMinusAlphaChannel, colorMask, half) { \ |
72 | /* interpolate AG */\ |
73 | __m128i srcVectorAG = _mm_srli_epi16(srcVector, 8); \ |
74 | __m128i dstVectorAG = _mm_srli_epi16(dstVector, 8); \ |
75 | __m128i srcVectorAGalpha = _mm_mullo_epi16(srcVectorAG, alphaChannel); \ |
76 | __m128i dstVectorAGoneMinusAlphalpha = _mm_mullo_epi16(dstVectorAG, oneMinusAlphaChannel); \ |
77 | __m128i finalAG = _mm_add_epi16(srcVectorAGalpha, dstVectorAGoneMinusAlphalpha); \ |
78 | finalAG = _mm_add_epi16(finalAG, _mm_srli_epi16(finalAG, 8)); \ |
79 | finalAG = _mm_add_epi16(finalAG, half); \ |
80 | finalAG = _mm_andnot_si128(colorMask, finalAG); \ |
81 | \ |
82 | /* interpolate RB */\ |
83 | __m128i srcVectorRB = _mm_and_si128(srcVector, colorMask); \ |
84 | __m128i dstVectorRB = _mm_and_si128(dstVector, colorMask); \ |
85 | __m128i srcVectorRBalpha = _mm_mullo_epi16(srcVectorRB, alphaChannel); \ |
86 | __m128i dstVectorRBoneMinusAlphalpha = _mm_mullo_epi16(dstVectorRB, oneMinusAlphaChannel); \ |
87 | __m128i finalRB = _mm_add_epi16(srcVectorRBalpha, dstVectorRBoneMinusAlphalpha); \ |
88 | finalRB = _mm_add_epi16(finalRB, _mm_srli_epi16(finalRB, 8)); \ |
89 | finalRB = _mm_add_epi16(finalRB, half); \ |
90 | finalRB = _mm_srli_epi16(finalRB, 8); \ |
91 | \ |
92 | /* combine */\ |
93 | result = _mm_or_si128(finalAG, finalRB); \ |
94 | } |
95 | |
96 | // same as BLEND_SOURCE_OVER_ARGB32_SSE2, but for one vector srcVector |
97 | #define BLEND_SOURCE_OVER_ARGB32_SSE2_helper(dst, srcVector, nullVector, half, one, colorMask, alphaMask) { \ |
98 | const __m128i srcVectorAlpha = _mm_and_si128(srcVector, alphaMask); \ |
99 | if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVectorAlpha, alphaMask)) == 0xffff) { \ |
100 | /* all opaque */ \ |
101 | _mm_store_si128((__m128i *)&dst[x], srcVector); \ |
102 | } else if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVectorAlpha, nullVector)) != 0xffff) { \ |
103 | /* not fully transparent */ \ |
104 | /* extract the alpha channel on 2 x 16 bits */ \ |
105 | /* so we have room for the multiplication */ \ |
106 | /* each 32 bits will be in the form 0x00AA00AA */ \ |
107 | /* with A being the 1 - alpha */ \ |
108 | __m128i alphaChannel = _mm_srli_epi32(srcVector, 24); \ |
109 | alphaChannel = _mm_or_si128(alphaChannel, _mm_slli_epi32(alphaChannel, 16)); \ |
110 | alphaChannel = _mm_sub_epi16(one, alphaChannel); \ |
111 | \ |
112 | const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]); \ |
113 | __m128i destMultipliedByOneMinusAlpha; \ |
114 | BYTE_MUL_SSE2(destMultipliedByOneMinusAlpha, dstVector, alphaChannel, colorMask, half); \ |
115 | \ |
116 | /* result = s + d * (1-alpha) */\ |
117 | const __m128i result = _mm_add_epi8(srcVector, destMultipliedByOneMinusAlpha); \ |
118 | _mm_store_si128((__m128i *)&dst[x], result); \ |
119 | } \ |
120 | } |
121 | |
122 | |
123 | // Basically blend src over dst with the const alpha defined as constAlphaVector. |
124 | // nullVector, half, one, colorMask are constant across the whole image/texture, and should be defined as: |
125 | //const __m128i nullVector = _mm_set1_epi32(0); |
126 | //const __m128i half = _mm_set1_epi16(0x80); |
127 | //const __m128i one = _mm_set1_epi16(0xff); |
128 | //const __m128i colorMask = _mm_set1_epi32(0x00ff00ff); |
129 | //const __m128i alphaMask = _mm_set1_epi32(0xff000000); |
130 | // |
131 | // The computation being done is: |
132 | // result = s + d * (1-alpha) |
133 | // with shortcuts if fully opaque or fully transparent. |
134 | #define BLEND_SOURCE_OVER_ARGB32_SSE2(dst, src, length, nullVector, half, one, colorMask, alphaMask) { \ |
135 | int x = 0; \ |
136 | \ |
137 | /* First, get dst aligned. */ \ |
138 | ALIGNMENT_PROLOGUE_16BYTES(dst, x, length) { \ |
139 | blend_pixel(dst[x], src[x]); \ |
140 | } \ |
141 | \ |
142 | for (; x < length-3; x += 4) { \ |
143 | const __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[x]); \ |
144 | BLEND_SOURCE_OVER_ARGB32_SSE2_helper(dst, srcVector, nullVector, half, one, colorMask, alphaMask) \ |
145 | } \ |
146 | SIMD_EPILOGUE(x, length, 3) { \ |
147 | blend_pixel(dst[x], src[x]); \ |
148 | } \ |
149 | } |
150 | |
151 | // Basically blend src over dst with the const alpha defined as constAlphaVector. |
152 | // nullVector, half, one, colorMask are constant across the whole image/texture, and should be defined as: |
153 | //const __m128i nullVector = _mm_set1_epi32(0); |
154 | //const __m128i half = _mm_set1_epi16(0x80); |
155 | //const __m128i one = _mm_set1_epi16(0xff); |
156 | //const __m128i colorMask = _mm_set1_epi32(0x00ff00ff); |
157 | // |
158 | // The computation being done is: |
159 | // dest = (s + d * sia) * ca + d * cia |
160 | // = s * ca + d * (sia * ca + cia) |
161 | // = s * ca + d * (1 - sa*ca) |
162 | #define BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2(dst, src, length, nullVector, half, one, colorMask, constAlphaVector) \ |
163 | { \ |
164 | int x = 0; \ |
165 | \ |
166 | ALIGNMENT_PROLOGUE_16BYTES(dst, x, length) { \ |
167 | blend_pixel(dst[x], src[x], const_alpha); \ |
168 | } \ |
169 | \ |
170 | for (; x < length-3; x += 4) { \ |
171 | __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[x]); \ |
172 | if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVector, nullVector)) != 0xffff) { \ |
173 | BYTE_MUL_SSE2(srcVector, srcVector, constAlphaVector, colorMask, half); \ |
174 | \ |
175 | __m128i alphaChannel = _mm_srli_epi32(srcVector, 24); \ |
176 | alphaChannel = _mm_or_si128(alphaChannel, _mm_slli_epi32(alphaChannel, 16)); \ |
177 | alphaChannel = _mm_sub_epi16(one, alphaChannel); \ |
178 | \ |
179 | const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]); \ |
180 | __m128i destMultipliedByOneMinusAlpha; \ |
181 | BYTE_MUL_SSE2(destMultipliedByOneMinusAlpha, dstVector, alphaChannel, colorMask, half); \ |
182 | \ |
183 | const __m128i result = _mm_add_epi8(srcVector, destMultipliedByOneMinusAlpha); \ |
184 | _mm_store_si128((__m128i *)&dst[x], result); \ |
185 | } \ |
186 | } \ |
187 | SIMD_EPILOGUE(x, length, 3) { \ |
188 | blend_pixel(dst[x], src[x], const_alpha); \ |
189 | } \ |
190 | } |
191 | |
192 | QT_END_NAMESPACE |
193 | |
194 | #endif // __SSE2__ |
195 | |
196 | QT_BEGIN_NAMESPACE |
197 | #if QT_COMPILER_SUPPORTS_HERE(SSE4_1) |
198 | QT_FUNCTION_TARGET(SSE2) |
199 | static inline void Q_DECL_VECTORCALL reciprocal_mul_ss(__m128 &ia, const __m128 a, float mul) |
200 | { |
201 | ia = _mm_rcp_ss(a: a); // Approximate 1/a |
202 | // Improve precision of ia using Newton-Raphson |
203 | ia = _mm_sub_ss(a: _mm_add_ss(a: ia, b: ia), b: _mm_mul_ss(a: ia, b: _mm_mul_ss(a: ia, b: a))); |
204 | ia = _mm_mul_ss(a: ia, b: _mm_set_ss(w: mul)); |
205 | ia = _mm_shuffle_ps(ia, ia, _MM_SHUFFLE(0,0,0,0)); |
206 | } |
207 | |
208 | QT_FUNCTION_TARGET(SSE4_1) |
209 | static inline QRgb qUnpremultiply_sse4(QRgb p) |
210 | { |
211 | const uint alpha = qAlpha(rgb: p); |
212 | if (alpha == 255) |
213 | return p; |
214 | if (alpha == 0) |
215 | return 0; |
216 | const __m128 va = _mm_set1_ps(w: alpha); |
217 | __m128 via; |
218 | reciprocal_mul_ss(ia&: via, a: va, mul: 255.0f); // Approximate 1/a |
219 | __m128i vl = _mm_cvtepu8_epi32(V: _mm_cvtsi32_si128(a: p)); |
220 | vl = _mm_cvtps_epi32(a: _mm_mul_ps(a: _mm_cvtepi32_ps(a: vl), b: via)); |
221 | vl = _mm_packus_epi32(V1: vl, V2: vl); |
222 | vl = _mm_insert_epi16(vl, alpha, 3); |
223 | vl = _mm_packus_epi16(a: vl, b: vl); |
224 | return _mm_cvtsi128_si32(a: vl); |
225 | } |
226 | |
227 | template<enum QtPixelOrder PixelOrder> |
228 | QT_FUNCTION_TARGET(SSE4_1) |
229 | static inline uint qConvertArgb32ToA2rgb30_sse4(QRgb p) |
230 | { |
231 | const uint alpha = qAlpha(rgb: p); |
232 | if (alpha == 255) |
233 | return qConvertRgb32ToRgb30<PixelOrder>(p); |
234 | if (alpha == 0) |
235 | return 0; |
236 | constexpr float mult = 1023.0f / (255 >> 6); |
237 | const uint newalpha = (alpha >> 6); |
238 | const __m128 va = _mm_set1_ps(w: alpha); |
239 | __m128 via; |
240 | reciprocal_mul_ss(ia&: via, a: va, mul: mult * newalpha); |
241 | __m128i vl = _mm_cvtsi32_si128(a: p); |
242 | vl = _mm_cvtepu8_epi32(V: vl); |
243 | vl = _mm_cvtps_epi32(a: _mm_mul_ps(a: _mm_cvtepi32_ps(a: vl), b: via)); |
244 | vl = _mm_packus_epi32(V1: vl, V2: vl); |
245 | uint rgb30 = (newalpha << 30); |
246 | rgb30 |= ((uint)_mm_extract_epi16(vl, 1)) << 10; |
247 | if (PixelOrder == PixelOrderRGB) { |
248 | rgb30 |= ((uint)_mm_extract_epi16(vl, 2)) << 20; |
249 | rgb30 |= ((uint)_mm_extract_epi16(vl, 0)); |
250 | } else { |
251 | rgb30 |= ((uint)_mm_extract_epi16(vl, 0)) << 20; |
252 | rgb30 |= ((uint)_mm_extract_epi16(vl, 2)); |
253 | } |
254 | return rgb30; |
255 | } |
256 | |
257 | template<enum QtPixelOrder PixelOrder> |
258 | QT_FUNCTION_TARGET(SSE4_1) |
259 | static inline uint qConvertRgba64ToRgb32_sse4(QRgba64 p) |
260 | { |
261 | if (p.isTransparent()) |
262 | return 0; |
263 | __m128i vl = _mm_loadl_epi64(p: reinterpret_cast<const __m128i *>(&p)); |
264 | if (!p.isOpaque()) { |
265 | const __m128 va = _mm_set1_ps(w: p.alpha()); |
266 | __m128 via; |
267 | reciprocal_mul_ss(ia&: via, a: va, mul: 65535.0f); |
268 | vl = _mm_unpacklo_epi16(a: vl, b: _mm_setzero_si128()); |
269 | vl = _mm_cvtps_epi32(a: _mm_mul_ps(a: _mm_cvtepi32_ps(a: vl) , b: via)); |
270 | vl = _mm_packus_epi32(V1: vl, V2: vl); |
271 | vl = _mm_insert_epi16(vl, p.alpha(), 3); |
272 | } |
273 | if (PixelOrder == PixelOrderBGR) |
274 | vl = _mm_shufflelo_epi16(vl, _MM_SHUFFLE(3, 0, 1, 2)); |
275 | return toArgb32(v: vl); |
276 | } |
277 | #endif |
278 | QT_END_NAMESPACE |
279 | |
280 | #endif // QDRAWINGPRIMITIVE_SSE2_P_H |
281 | |