1 /*
2 * Copyright (c) 2021 - 2024 the ThorVG project. All rights reserved.
3
4 * Permission is hereby granted, free of charge, to any person obtaining a copy
5 * of this software and associated documentation files (the "Software"), to deal
6 * in the Software without restriction, including without limitation the rights
7 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8 * copies of the Software, and to permit persons to whom the Software is
9 * furnished to do so, subject to the following conditions:
10
11 * The above copyright notice and this permission notice shall be included in all
12 * copies or substantial portions of the Software.
13
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
20 * SOFTWARE.
21 */
22
23 #include "../../lv_conf_internal.h"
24 #if LV_USE_THORVG_INTERNAL
25
26 #ifdef THORVG_AVX_VECTOR_SUPPORT
27
28 #include <immintrin.h>
29
30 #define N_32BITS_IN_128REG 4
31 #define N_32BITS_IN_256REG 8
32
ALPHA_BLEND(__m128i c,__m128i a)33 static inline __m128i ALPHA_BLEND(__m128i c, __m128i a)
34 {
35 //1. set the masks for the A/G and R/B channels
36 auto AG = _mm_set1_epi32(0xff00ff00);
37 auto RB = _mm_set1_epi32(0x00ff00ff);
38
39 //2. mask the alpha vector - originally quartet [a, a, a, a]
40 auto aAG = _mm_and_si128(a, AG);
41 auto aRB = _mm_and_si128(a, RB);
42
43 //3. calculate the alpha blending of the 2nd and 4th channel
44 //- mask the color vector
45 //- multiply it by the masked alpha vector
46 //- add the correction to compensate bit shifting used instead of dividing by 255
47 //- shift bits - corresponding to division by 256
48 auto even = _mm_and_si128(c, RB);
49 even = _mm_mullo_epi16(even, aRB);
50 even =_mm_add_epi16(even, RB);
51 even = _mm_srli_epi16(even, 8);
52
53 //4. calculate the alpha blending of the 1st and 3rd channel:
54 //- mask the color vector
55 //- multiply it by the corresponding masked alpha vector and store the high bits of the result
56 //- add the correction to compensate division by 256 instead of by 255 (next step)
57 //- remove the low 8 bits to mimic the division by 256
58 auto odd = _mm_and_si128(c, AG);
59 odd = _mm_mulhi_epu16(odd, aAG);
60 odd = _mm_add_epi16(odd, RB);
61 odd = _mm_and_si128(odd, AG);
62
63 //5. the final result
64 return _mm_or_si128(odd, even);
65 }
66
67
avxRasterGrayscale8(uint8_t * dst,uint8_t val,uint32_t offset,int32_t len)68 static void avxRasterGrayscale8(uint8_t* dst, uint8_t val, uint32_t offset, int32_t len)
69 {
70 dst += offset;
71
72 __m256i vecVal = _mm256_set1_epi8(val);
73
74 int32_t i = 0;
75 for (; i <= len - 32; i += 32) {
76 _mm256_storeu_si256((__m256i*)(dst + i), vecVal);
77 }
78
79 for (; i < len; ++i) {
80 dst[i] = val;
81 }
82 }
83
84
avxRasterPixel32(uint32_t * dst,uint32_t val,uint32_t offset,int32_t len)85 static void avxRasterPixel32(uint32_t *dst, uint32_t val, uint32_t offset, int32_t len)
86 {
87 //1. calculate how many iterations we need to cover the length
88 uint32_t iterations = len / N_32BITS_IN_256REG;
89 uint32_t avxFilled = iterations * N_32BITS_IN_256REG;
90
91 //2. set the beginning of the array
92 dst += offset;
93
94 //3. fill the octets
95 for (uint32_t i = 0; i < iterations; ++i, dst += N_32BITS_IN_256REG) {
96 _mm256_storeu_si256((__m256i*)dst, _mm256_set1_epi32(val));
97 }
98
99 //4. fill leftovers (in the first step we have to set the pointer to the place where the avx job is done)
100 int32_t leftovers = len - avxFilled;
101 while (leftovers--) *dst++ = val;
102 }
103
104
avxRasterTranslucentRect(SwSurface * surface,const SwBBox & region,uint8_t r,uint8_t g,uint8_t b,uint8_t a)105 static bool avxRasterTranslucentRect(SwSurface* surface, const SwBBox& region, uint8_t r, uint8_t g, uint8_t b, uint8_t a)
106 {
107 auto h = static_cast<uint32_t>(region.max.y - region.min.y);
108 auto w = static_cast<uint32_t>(region.max.x - region.min.x);
109
110 //32bits channels
111 if (surface->channelSize == sizeof(uint32_t)) {
112 auto color = surface->join(r, g, b, a);
113 auto buffer = surface->buf32 + (region.min.y * surface->stride) + region.min.x;
114
115 uint32_t ialpha = 255 - a;
116
117 auto avxColor = _mm_set1_epi32(color);
118 auto avxIalpha = _mm_set1_epi8(ialpha);
119
120 for (uint32_t y = 0; y < h; ++y) {
121 auto dst = &buffer[y * surface->stride];
122
123 //1. fill the not aligned memory (for 128-bit registers a 16-bytes alignment is required)
124 auto notAligned = ((uintptr_t)dst & 0xf) / 4;
125 if (notAligned) {
126 notAligned = (N_32BITS_IN_128REG - notAligned > w ? w : N_32BITS_IN_128REG - notAligned);
127 for (uint32_t x = 0; x < notAligned; ++x, ++dst) {
128 *dst = color + ALPHA_BLEND(*dst, ialpha);
129 }
130 }
131
132 //2. fill the aligned memory - N_32BITS_IN_128REG pixels processed at once
133 uint32_t iterations = (w - notAligned) / N_32BITS_IN_128REG;
134 uint32_t avxFilled = iterations * N_32BITS_IN_128REG;
135 auto avxDst = (__m128i*)dst;
136 for (uint32_t x = 0; x < iterations; ++x, ++avxDst) {
137 *avxDst = _mm_add_epi32(avxColor, ALPHA_BLEND(*avxDst, avxIalpha));
138 }
139
140 //3. fill the remaining pixels
141 int32_t leftovers = w - notAligned - avxFilled;
142 dst += avxFilled;
143 while (leftovers--) {
144 *dst = color + ALPHA_BLEND(*dst, ialpha);
145 dst++;
146 }
147 }
148 //8bit grayscale
149 } else if (surface->channelSize == sizeof(uint8_t)) {
150 TVGLOG("SW_ENGINE", "Require AVX Optimization, Channel Size = %d", surface->channelSize);
151 auto buffer = surface->buf8 + (region.min.y * surface->stride) + region.min.x;
152 auto ialpha = ~a;
153 for (uint32_t y = 0; y < h; ++y) {
154 auto dst = &buffer[y * surface->stride];
155 for (uint32_t x = 0; x < w; ++x, ++dst) {
156 *dst = a + MULTIPLY(*dst, ialpha);
157 }
158 }
159 }
160 return true;
161 }
162
163
avxRasterTranslucentRle(SwSurface * surface,const SwRle * rle,uint8_t r,uint8_t g,uint8_t b,uint8_t a)164 static bool avxRasterTranslucentRle(SwSurface* surface, const SwRle* rle, uint8_t r, uint8_t g, uint8_t b, uint8_t a)
165 {
166 auto span = rle->spans;
167
168 //32bit channels
169 if (surface->channelSize == sizeof(uint32_t)) {
170 auto color = surface->join(r, g, b, a);
171 uint32_t src;
172
173 for (uint32_t i = 0; i < rle->size; ++i) {
174 auto dst = &surface->buf32[span->y * surface->stride + span->x];
175
176 if (span->coverage < 255) src = ALPHA_BLEND(color, span->coverage);
177 else src = color;
178
179 auto ialpha = IA(src);
180
181 //1. fill the not aligned memory (for 128-bit registers a 16-bytes alignment is required)
182 auto notAligned = ((uintptr_t)dst & 0xf) / 4;
183 if (notAligned) {
184 notAligned = (N_32BITS_IN_128REG - notAligned > span->len ? span->len : N_32BITS_IN_128REG - notAligned);
185 for (uint32_t x = 0; x < notAligned; ++x, ++dst) {
186 *dst = src + ALPHA_BLEND(*dst, ialpha);
187 }
188 }
189
190 //2. fill the aligned memory using avx - N_32BITS_IN_128REG pixels processed at once
191 //In order to avoid unnecessary avx variables declarations a check is made whether there are any iterations at all
192 uint32_t iterations = (span->len - notAligned) / N_32BITS_IN_128REG;
193 uint32_t avxFilled = 0;
194 if (iterations > 0) {
195 auto avxSrc = _mm_set1_epi32(src);
196 auto avxIalpha = _mm_set1_epi8(ialpha);
197
198 avxFilled = iterations * N_32BITS_IN_128REG;
199 auto avxDst = (__m128i*)dst;
200 for (uint32_t x = 0; x < iterations; ++x, ++avxDst) {
201 *avxDst = _mm_add_epi32(avxSrc, ALPHA_BLEND(*avxDst, avxIalpha));
202 }
203 }
204
205 //3. fill the remaining pixels
206 int32_t leftovers = span->len - notAligned - avxFilled;
207 dst += avxFilled;
208 while (leftovers--) {
209 *dst = src + ALPHA_BLEND(*dst, ialpha);
210 dst++;
211 }
212
213 ++span;
214 }
215 //8bit grayscale
216 } else if (surface->channelSize == sizeof(uint8_t)) {
217 TVGLOG("SW_ENGINE", "Require AVX Optimization, Channel Size = %d", surface->channelSize);
218 uint8_t src;
219 for (uint32_t i = 0; i < rle->size; ++i, ++span) {
220 auto dst = &surface->buf8[span->y * surface->stride + span->x];
221 if (span->coverage < 255) src = MULTIPLY(span->coverage, a);
222 else src = a;
223 auto ialpha = ~a;
224 for (uint32_t x = 0; x < span->len; ++x, ++dst) {
225 *dst = src + MULTIPLY(*dst, ialpha);
226 }
227 }
228 }
229 return true;
230 }
231
232
233 #endif
234
235 #endif /* LV_USE_THORVG_INTERNAL */
236
237
