openMSX
yuv2rgb.cc
Go to the documentation of this file.
1#include "yuv2rgb.hh"
2#include "RawFrame.hh"
3#include "Math.hh"
4#include "xrange.hh"
5#include <array>
6#include <cassert>
7#include <concepts>
8#include <cstdint>
9#ifdef __SSE2__
10#include <emmintrin.h>
11#endif
12
14
15using Pixel = uint32_t;
16
17#ifdef __SSE2__
18
19/*
20 * This implementation of yuv420 to rgb is based upon the corresponding routine
21 * from Mono. See this blog entry:
22 * http://blog.sublimeintervention.com/archive/2008/Mar-21.html
23 * Source code:
24 * http://anonsvn.mono-project.com/viewvc/trunk/moon/src/yuv-converter.cpp?revision=136072
25 * This code is GPL2 (only)
26 *
27 * Copyright 2008 Novell, Inc. (http://www.novell.com)
28 *
29 * There are other implementations:
30 * - ffmpeg
31 * - mythtv
32 * - pcsx2
33 * I have not done a comparison of these implementations.
34 */
35
36/* R = 1.164 * (Y - 16) + 1.596 * (V - 128)
37 * G = 1.164 * (Y - 16) - 0.813 * (V - 128) - 0.391 * (U - 128)
38 * B = 1.164 * (Y - 16) + 2.018 * (U - 128)
39 * OR
40 * R = 1.164 * Y + 1.596 * V - 222.921
41 * G = 1.164 * Y - 0.813 * V - 0.391 * U + 135.576
42 * B = 1.164 * Y + 2.018 * U - 276.836
43 */
44static inline void yuv2rgb_sse2(
45 const uint8_t* u_ , const uint8_t* v_,
46 const uint8_t* y0_, const uint8_t* y1_,
47 Pixel* out0_, Pixel* out1_)
48{
49 // This routine calculates 32x2 RGBA pixels. Each output pixel uses a
50 // unique corresponding input Y value, but a group of 2x2 ouput pixels
51 // shares the same U and V input value.
52 const auto* u = reinterpret_cast<const __m128i*>(u_);
53 const auto* v = reinterpret_cast<const __m128i*>(v_);
54 const auto* y0 = reinterpret_cast<const __m128i*>(y0_);
55 const auto* y1 = reinterpret_cast<const __m128i*>(y1_);
56 auto* out0 = reinterpret_cast< __m128i*>(out0_);
57 auto* out1 = reinterpret_cast< __m128i*>(out1_);
58
59 // constants
60 const __m128i ZERO = _mm_setzero_si128();
61 const __m128i ALPHA = _mm_set1_epi16( -1); // 0xFFFF
62 const __m128i RED_V = _mm_set1_epi16( 102); // 102/64 = 1.59
63 const __m128i GREEN_U = _mm_set1_epi16( -25); // -25/64 = -0.39
64 const __m128i GREEN_V = _mm_set1_epi16( -52); // -52/64 = -0.81
65 const __m128i BLUE_U = _mm_set1_epi16( 129); // 129/64 = 2.02
66 const __m128i COEF_Y = _mm_set1_epi16( 74); // 74/64 = 1.16
67 const __m128i CNST_R = _mm_set1_epi16( -223); // -222.921
68 const __m128i CNST_G = _mm_set1_epi16( 136); // 135.576
69 const __m128i CNST_B = _mm_set1_epi16( -277); // -276.836
70 const __m128i Y_MASK = _mm_set1_epi16(0x00FF);
71
72 // left
73 __m128i u0f = _mm_load_si128(u);
74 __m128i v0f = _mm_load_si128(v);
75 __m128i u07 = _mm_unpacklo_epi8(u0f, ZERO);
76 __m128i v07 = _mm_unpacklo_epi8(v0f, ZERO);
77 __m128i mr07 = _mm_srai_epi16(_mm_mullo_epi16(v07, RED_V), 6);
78 __m128i sg07 = _mm_mullo_epi16(v07, GREEN_V);
79 __m128i tg07 = _mm_mullo_epi16(u07, GREEN_U);
80 __m128i mg07 = _mm_srai_epi16(_mm_adds_epi16(sg07, tg07), 6);
81 __m128i mb07 = _mm_srli_epi16(_mm_mullo_epi16(u07, BLUE_U), 6); // logical shift
82 __m128i dr07 = _mm_adds_epi16(mr07, CNST_R);
83 __m128i dg07 = _mm_adds_epi16(mg07, CNST_G);
84 __m128i db07 = _mm_adds_epi16(mb07, CNST_B);
85
86 // block top,left
87 __m128i y00_0f = _mm_load_si128(y0 + 0);
88 __m128i y00_even = _mm_and_si128(y00_0f, Y_MASK);
89 __m128i y00_odd = _mm_srli_epi16(y00_0f, 8);
90 __m128i dy00_even = _mm_srai_epi16(_mm_mullo_epi16(y00_even, COEF_Y), 6);
91 __m128i dy00_odd = _mm_srai_epi16(_mm_mullo_epi16(y00_odd, COEF_Y), 6);
92 __m128i r00_even = _mm_adds_epi16(dr07, dy00_even);
93 __m128i g00_even = _mm_adds_epi16(dg07, dy00_even);
94 __m128i b00_even = _mm_adds_epi16(db07, dy00_even);
95 __m128i r00_odd = _mm_adds_epi16(dr07, dy00_odd);
96 __m128i g00_odd = _mm_adds_epi16(dg07, dy00_odd);
97 __m128i b00_odd = _mm_adds_epi16(db07, dy00_odd);
98 __m128i r00_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r00_even, r00_even),
99 _mm_packus_epi16(r00_odd, r00_odd));
100 __m128i g00_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g00_even, g00_even),
101 _mm_packus_epi16(g00_odd, g00_odd));
102 __m128i b00_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b00_even, b00_even),
103 _mm_packus_epi16(b00_odd, b00_odd));
104 __m128i rb00_07 = _mm_unpacklo_epi8(r00_0f, b00_0f);
105 __m128i rb00_8f = _mm_unpackhi_epi8(r00_0f, b00_0f);
106 __m128i ga00_07 = _mm_unpacklo_epi8(g00_0f, ALPHA);
107 __m128i ga00_8f = _mm_unpackhi_epi8(g00_0f, ALPHA);
108 __m128i rgba00_03 = _mm_unpacklo_epi8(rb00_07, ga00_07);
109 __m128i rgba00_47 = _mm_unpackhi_epi8(rb00_07, ga00_07);
110 __m128i rgba00_8b = _mm_unpacklo_epi8(rb00_8f, ga00_8f);
111 __m128i rgba00_cf = _mm_unpackhi_epi8(rb00_8f, ga00_8f);
112 _mm_store_si128(out0 + 0, rgba00_03);
113 _mm_store_si128(out0 + 1, rgba00_47);
114 _mm_store_si128(out0 + 2, rgba00_8b);
115 _mm_store_si128(out0 + 3, rgba00_cf);
116
117 // block bottom,left
118 __m128i y10_0f = _mm_load_si128(y1 + 0);
119 __m128i y10_even = _mm_and_si128(y10_0f, Y_MASK);
120 __m128i y10_odd = _mm_srli_epi16(y10_0f, 8);
121 __m128i dy10_even = _mm_srai_epi16(_mm_mullo_epi16(y10_even, COEF_Y), 6);
122 __m128i dy10_odd = _mm_srai_epi16(_mm_mullo_epi16(y10_odd, COEF_Y), 6);
123 __m128i r10_even = _mm_adds_epi16(dr07, dy10_even);
124 __m128i g10_even = _mm_adds_epi16(dg07, dy10_even);
125 __m128i b10_even = _mm_adds_epi16(db07, dy10_even);
126 __m128i r10_odd = _mm_adds_epi16(dr07, dy10_odd);
127 __m128i g10_odd = _mm_adds_epi16(dg07, dy10_odd);
128 __m128i b10_odd = _mm_adds_epi16(db07, dy10_odd);
129 __m128i r10_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r10_even, r10_even),
130 _mm_packus_epi16(r10_odd, r10_odd));
131 __m128i g10_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g10_even, g10_even),
132 _mm_packus_epi16(g10_odd, g10_odd));
133 __m128i b10_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b10_even, b10_even),
134 _mm_packus_epi16(b10_odd, b10_odd));
135 __m128i rb10_07 = _mm_unpacklo_epi8(r10_0f, b10_0f);
136 __m128i rb10_8f = _mm_unpackhi_epi8(r10_0f, b10_0f);
137 __m128i ga10_07 = _mm_unpacklo_epi8(g10_0f, ALPHA);
138 __m128i ga10_8f = _mm_unpackhi_epi8(g10_0f, ALPHA);
139 __m128i rgba10_03 = _mm_unpacklo_epi8(rb10_07, ga10_07);
140 __m128i rgba10_47 = _mm_unpackhi_epi8(rb10_07, ga10_07);
141 __m128i rgba10_8b = _mm_unpacklo_epi8(rb10_8f, ga10_8f);
142 __m128i rgba10_cf = _mm_unpackhi_epi8(rb10_8f, ga10_8f);
143 _mm_store_si128(out1 + 0, rgba10_03);
144 _mm_store_si128(out1 + 1, rgba10_47);
145 _mm_store_si128(out1 + 2, rgba10_8b);
146 _mm_store_si128(out1 + 3, rgba10_cf);
147
148 // right
149 __m128i u8f = _mm_unpackhi_epi8(u0f, ZERO);
150 __m128i v8f = _mm_unpackhi_epi8(v0f, ZERO);
151 __m128i mr8f = _mm_srai_epi16(_mm_mullo_epi16(v8f, RED_V), 6);
152 __m128i sg8f = _mm_mullo_epi16(v8f, GREEN_V);
153 __m128i tg8f = _mm_mullo_epi16(u8f, GREEN_U);
154 __m128i mg8f = _mm_srai_epi16(_mm_adds_epi16(sg8f, tg8f), 6);
155 __m128i mb8f = _mm_srli_epi16(_mm_mullo_epi16(u8f, BLUE_U), 6); // logical shift
156 __m128i dr8f = _mm_adds_epi16(mr8f, CNST_R);
157 __m128i dg8f = _mm_adds_epi16(mg8f, CNST_G);
158 __m128i db8f = _mm_adds_epi16(mb8f, CNST_B);
159
160 // block top,right
161 __m128i y01_0f = _mm_load_si128(y0 + 1);
162 __m128i y01_even = _mm_and_si128(y01_0f, Y_MASK);
163 __m128i y01_odd = _mm_srli_epi16(y01_0f, 8);
164 __m128i dy01_even = _mm_srai_epi16(_mm_mullo_epi16(y01_even, COEF_Y), 6);
165 __m128i dy01_odd = _mm_srai_epi16(_mm_mullo_epi16(y01_odd, COEF_Y), 6);
166 __m128i r01_even = _mm_adds_epi16(dr8f, dy01_even);
167 __m128i g01_even = _mm_adds_epi16(dg8f, dy01_even);
168 __m128i b01_even = _mm_adds_epi16(db8f, dy01_even);
169 __m128i r01_odd = _mm_adds_epi16(dr8f, dy01_odd);
170 __m128i g01_odd = _mm_adds_epi16(dg8f, dy01_odd);
171 __m128i b01_odd = _mm_adds_epi16(db8f, dy01_odd);
172 __m128i r01_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r01_even, r01_even),
173 _mm_packus_epi16(r01_odd, r01_odd));
174 __m128i g01_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g01_even, g01_even),
175 _mm_packus_epi16(g01_odd, g01_odd));
176 __m128i b01_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b01_even, b01_even),
177 _mm_packus_epi16(b01_odd, b01_odd));
178 __m128i rb01_07 = _mm_unpacklo_epi8(r01_0f, b01_0f);
179 __m128i rb01_8f = _mm_unpackhi_epi8(r01_0f, b01_0f);
180 __m128i ga01_07 = _mm_unpacklo_epi8(g01_0f, ALPHA);
181 __m128i ga01_8f = _mm_unpackhi_epi8(g01_0f, ALPHA);
182 __m128i rgba01_03 = _mm_unpacklo_epi8(rb01_07, ga01_07);
183 __m128i rgba01_47 = _mm_unpackhi_epi8(rb01_07, ga01_07);
184 __m128i rgba01_8b = _mm_unpacklo_epi8(rb01_8f, ga01_8f);
185 __m128i rgba01_cf = _mm_unpackhi_epi8(rb01_8f, ga01_8f);
186 _mm_store_si128(out0 + 4, rgba01_03);
187 _mm_store_si128(out0 + 5, rgba01_47);
188 _mm_store_si128(out0 + 6, rgba01_8b);
189 _mm_store_si128(out0 + 7, rgba01_cf);
190
191 // block bottom,right
192 __m128i y11_0f = _mm_load_si128(y1 + 1);
193 __m128i y11_even = _mm_and_si128(y11_0f, Y_MASK);
194 __m128i y11_odd = _mm_srli_epi16(y11_0f, 8);
195 __m128i dy11_even = _mm_srai_epi16(_mm_mullo_epi16(y11_even, COEF_Y), 6);
196 __m128i dy11_odd = _mm_srai_epi16(_mm_mullo_epi16(y11_odd, COEF_Y), 6);
197 __m128i r11_even = _mm_adds_epi16(dr8f, dy11_even);
198 __m128i g11_even = _mm_adds_epi16(dg8f, dy11_even);
199 __m128i b11_even = _mm_adds_epi16(db8f, dy11_even);
200 __m128i r11_odd = _mm_adds_epi16(dr8f, dy11_odd);
201 __m128i g11_odd = _mm_adds_epi16(dg8f, dy11_odd);
202 __m128i b11_odd = _mm_adds_epi16(db8f, dy11_odd);
203 __m128i r11_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r11_even, r11_even),
204 _mm_packus_epi16(r11_odd, r11_odd));
205 __m128i g11_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g11_even, g11_even),
206 _mm_packus_epi16(g11_odd, g11_odd));
207 __m128i b11_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b11_even, b11_even),
208 _mm_packus_epi16(b11_odd, b11_odd));
209 __m128i rb11_07 = _mm_unpacklo_epi8(r11_0f, b11_0f);
210 __m128i rb11_8f = _mm_unpackhi_epi8(r11_0f, b11_0f);
211 __m128i ga11_07 = _mm_unpacklo_epi8(g11_0f, ALPHA);
212 __m128i ga11_8f = _mm_unpackhi_epi8(g11_0f, ALPHA);
213 __m128i rgba11_03 = _mm_unpacklo_epi8(rb11_07, ga11_07);
214 __m128i rgba11_47 = _mm_unpackhi_epi8(rb11_07, ga11_07);
215 __m128i rgba11_8b = _mm_unpacklo_epi8(rb11_8f, ga11_8f);
216 __m128i rgba11_cf = _mm_unpackhi_epi8(rb11_8f, ga11_8f);
217 _mm_store_si128(out1 + 4, rgba11_03);
218 _mm_store_si128(out1 + 5, rgba11_47);
219 _mm_store_si128(out1 + 6, rgba11_8b);
220 _mm_store_si128(out1 + 7, rgba11_cf);
221}
222
223static inline void convertHelperSSE2(
224 const th_ycbcr_buffer& buffer, RawFrame& output)
225{
226 const int width = buffer[0].width;
227 const size_t y_stride = buffer[0].stride;
228 const size_t uv_stride2 = buffer[1].stride / 2;
229
230 assert((width % 32) == 0);
231 assert((buffer[0].height % 2) == 0);
232
233 for (int y = 0; y < buffer[0].height; y += 2) {
234 const uint8_t* pY1 = buffer[0].data + (y + 0) * y_stride;
235 const uint8_t* pY2 = buffer[0].data + (y + 1) * y_stride;
236 const uint8_t* pCb = buffer[1].data + (y + 0) * uv_stride2;
237 const uint8_t* pCr = buffer[2].data + (y + 0) * uv_stride2;
238 auto out0 = output.getLineDirect(y + 0);
239 auto out1 = output.getLineDirect(y + 1);
240
241 for (int x = 0; x < width; x += 32) {
242 // convert a block of (32 x 2) pixels
243 yuv2rgb_sse2(pCb, pCr, pY1, pY2, &out0[x], &out1[x]);
244 pCb += 16;
245 pCr += 16;
246 pY1 += 32;
247 pY2 += 32;
248 }
249
250 output.setLineWidth(y + 0, width);
251 output.setLineWidth(y + 1, width);
252 }
253}
254
255#endif // __SSE2__
256
257static constexpr int PREC = 15;
258static constexpr int COEF_Y = int(1.164 * (1 << PREC) + 0.5); // prefer to use lrint() to round
259static constexpr int COEF_RV = int(1.596 * (1 << PREC) + 0.5); // but that's not (yet) constexpr
260static constexpr int COEF_GU = int(0.391 * (1 << PREC) + 0.5); // in current versions of c++
261static constexpr int COEF_GV = int(0.813 * (1 << PREC) + 0.5);
262static constexpr int COEF_BU = int(2.018 * (1 << PREC) + 0.5);
263
264struct Coefs {
265 std::array<int, 256> gu;
266 std::array<int, 256> gv;
267 std::array<int, 256> bu;
268 std::array<int, 256> rv;
269 std::array<int, 256> y;
270};
271
272[[nodiscard]] static constexpr Coefs getCoefs()
273{
274 Coefs coefs = {};
275 for (auto i : xrange(256)) {
276 coefs.gu[i] = -COEF_GU * (i - 128);
277 coefs.gv[i] = -COEF_GV * (i - 128);
278 coefs.bu[i] = COEF_BU * (i - 128);
279 coefs.rv[i] = COEF_RV * (i - 128);
280 coefs.y[i] = COEF_Y * (i - 16) + (PREC / 2);
281 }
282 return coefs;
283}
284
285[[nodiscard]] static inline Pixel calc(
286 int y, int ruv, int guv, int buv)
287{
288 uint8_t r = Math::clipIntToByte((y + ruv) >> PREC);
289 uint8_t g = Math::clipIntToByte((y + guv) >> PREC);
290 uint8_t b = Math::clipIntToByte((y + buv) >> PREC);
291 return (r << 0) | (g << 8) | (b << 16);
292}
293
294static void convertHelper(const th_ycbcr_buffer& buffer, RawFrame& output)
295{
296 assert(buffer[1].width * 2 == buffer[0].width);
297 assert(buffer[1].height * 2 == buffer[0].height);
298
299 static constexpr Coefs coefs = getCoefs();
300
301 const int width = buffer[0].width;
302 const size_t y_stride = buffer[0].stride;
303 const size_t uv_stride2 = buffer[1].stride / 2;
304
305 for (int y = 0; y < buffer[0].height; y += 2) {
306 const uint8_t* pY = buffer[0].data + y * y_stride;
307 const uint8_t* pCb = buffer[1].data + y * uv_stride2;
308 const uint8_t* pCr = buffer[2].data + y * uv_stride2;
309 auto out0 = output.getLineDirect(y + 0);
310 auto out1 = output.getLineDirect(y + 1);
311
312 for (int x = 0; x < width; x += 2, pY += 2, ++pCr, ++pCb) {
313 int ruv = coefs.rv[*pCr];
314 int guv = coefs.gu[*pCb] + coefs.gv[*pCr];
315 int buv = coefs.bu[*pCb];
316
317 int Y00 = coefs.y[pY[0]];
318 out0[x + 0] = calc(Y00, ruv, guv, buv);
319
320 int Y01 = coefs.y[pY[1]];
321 out0[x + 1] = calc(Y01, ruv, guv, buv);
322
323 int Y10 = coefs.y[pY[y_stride + 0]];
324 out1[x + 0] = calc(Y10, ruv, guv, buv);
325
326 int Y11 = coefs.y[pY[y_stride + 1]];
327 out1[x + 1] = calc(Y11, ruv, guv, buv);
328 }
329
330 output.setLineWidth(y + 0, width);
331 output.setLineWidth(y + 1, width);
332 }
333}
334
335void convert(const th_ycbcr_buffer& input, RawFrame& output)
336{
337#ifdef __SSE2__
338 convertHelperSSE2(input, output);
339 return;
340#endif
341 convertHelper(input, output);
342}
343
344} // namespace openmsx::yuv2rgb
int g
A video frame as output by the VDP scanline conversion unit, before any postprocessing filters are ap...
Definition: RawFrame.hh:15
std::span< Pixel > getLineDirect(unsigned y)
Definition: RawFrame.hh:21
void setLineWidth(unsigned line, unsigned width)
Definition: RawFrame.hh:31
uint8_t clipIntToByte(int x)
Clip x to range [0,255].
Definition: Math.hh:61
uint32_t Pixel
Definition: yuv2rgb.cc:15
void convert(const th_ycbcr_buffer &input, RawFrame &output)
Definition: yuv2rgb.cc:335
std::array< int, 256 > gu
Definition: yuv2rgb.cc:265
std::array< int, 256 > gv
Definition: yuv2rgb.cc:266
std::array< int, 256 > y
Definition: yuv2rgb.cc:269
std::array< int, 256 > rv
Definition: yuv2rgb.cc:268
std::array< int, 256 > bu
Definition: yuv2rgb.cc:267
constexpr auto xrange(T e)
Definition: xrange.hh:132