openMSX
yuv2rgb.cc
Go to the documentation of this file.
1#include "yuv2rgb.hh"
2#include "RawFrame.hh"
3#include "Math.hh"
4#include "PixelFormat.hh"
5#include "xrange.hh"
6#include <array>
7#include <cassert>
8#include <concepts>
9#include <cstdint>
10#ifdef __SSE2__
11#include <emmintrin.h>
12#endif
13
15
16#ifdef __SSE2__
17
18/*
19 * This implementation of yuv420 to rgb is based upon the corresponding routine
20 * from Mono. See this blog entry:
21 * http://blog.sublimeintervention.com/archive/2008/Mar-21.html
22 * Source code:
23 * http://anonsvn.mono-project.com/viewvc/trunk/moon/src/yuv-converter.cpp?revision=136072
24 * This code is GPL2 (only)
25 *
26 * Copyright 2008 Novell, Inc. (http://www.novell.com)
27 *
28 * There are other implementations:
29 * - ffmpeg
30 * - mythtv
31 * - pcsx2
32 * I have not done a comparison of these implementations.
33 */
34
35/* R = 1.164 * (Y - 16) + 1.596 * (V - 128)
36 * G = 1.164 * (Y - 16) - 0.813 * (V - 128) - 0.391 * (U - 128)
37 * B = 1.164 * (Y - 16) + 2.018 * (U - 128)
38 * OR
39 * R = 1.164 * Y + 1.596 * V - 222.921
40 * G = 1.164 * Y - 0.813 * V - 0.391 * U + 135.576
41 * B = 1.164 * Y + 2.018 * U - 276.836
42 */
43static inline void yuv2rgb_sse2(
44 const uint8_t* u_ , const uint8_t* v_,
45 const uint8_t* y0_, const uint8_t* y1_,
46 uint32_t* out0_, uint32_t* out1_)
47{
48 // This routine calculates 32x2 RGBA pixels. Each output pixel uses a
49 // unique corresponding input Y value, but a group of 2x2 ouput pixels
50 // shares the same U and V input value.
51 const auto* u = reinterpret_cast<const __m128i*>(u_);
52 const auto* v = reinterpret_cast<const __m128i*>(v_);
53 const auto* y0 = reinterpret_cast<const __m128i*>(y0_);
54 const auto* y1 = reinterpret_cast<const __m128i*>(y1_);
55 auto* out0 = reinterpret_cast< __m128i*>(out0_);
56 auto* out1 = reinterpret_cast< __m128i*>(out1_);
57
58 // constants
59 const __m128i ZERO = _mm_setzero_si128();
60 const __m128i ALPHA = _mm_set1_epi16( -1); // 0xFFFF
61 const __m128i RED_V = _mm_set1_epi16( 102); // 102/64 = 1.59
62 const __m128i GREEN_U = _mm_set1_epi16( -25); // -25/64 = -0.39
63 const __m128i GREEN_V = _mm_set1_epi16( -52); // -52/64 = -0.81
64 const __m128i BLUE_U = _mm_set1_epi16( 129); // 129/64 = 2.02
65 const __m128i COEF_Y = _mm_set1_epi16( 74); // 74/64 = 1.16
66 const __m128i CNST_R = _mm_set1_epi16( -223); // -222.921
67 const __m128i CNST_G = _mm_set1_epi16( 136); // 135.576
68 const __m128i CNST_B = _mm_set1_epi16( -277); // -276.836
69 const __m128i Y_MASK = _mm_set1_epi16(0x00FF);
70
71 // left
72 __m128i u0f = _mm_load_si128(u);
73 __m128i v0f = _mm_load_si128(v);
74 __m128i u07 = _mm_unpacklo_epi8(u0f, ZERO);
75 __m128i v07 = _mm_unpacklo_epi8(v0f, ZERO);
76 __m128i mr07 = _mm_srai_epi16(_mm_mullo_epi16(v07, RED_V), 6);
77 __m128i sg07 = _mm_mullo_epi16(v07, GREEN_V);
78 __m128i tg07 = _mm_mullo_epi16(u07, GREEN_U);
79 __m128i mg07 = _mm_srai_epi16(_mm_adds_epi16(sg07, tg07), 6);
80 __m128i mb07 = _mm_srli_epi16(_mm_mullo_epi16(u07, BLUE_U), 6); // logical shift
81 __m128i dr07 = _mm_adds_epi16(mr07, CNST_R);
82 __m128i dg07 = _mm_adds_epi16(mg07, CNST_G);
83 __m128i db07 = _mm_adds_epi16(mb07, CNST_B);
84
85 // block top,left
86 __m128i y00_0f = _mm_load_si128(y0 + 0);
87 __m128i y00_even = _mm_and_si128(y00_0f, Y_MASK);
88 __m128i y00_odd = _mm_srli_epi16(y00_0f, 8);
89 __m128i dy00_even = _mm_srai_epi16(_mm_mullo_epi16(y00_even, COEF_Y), 6);
90 __m128i dy00_odd = _mm_srai_epi16(_mm_mullo_epi16(y00_odd, COEF_Y), 6);
91 __m128i r00_even = _mm_adds_epi16(dr07, dy00_even);
92 __m128i g00_even = _mm_adds_epi16(dg07, dy00_even);
93 __m128i b00_even = _mm_adds_epi16(db07, dy00_even);
94 __m128i r00_odd = _mm_adds_epi16(dr07, dy00_odd);
95 __m128i g00_odd = _mm_adds_epi16(dg07, dy00_odd);
96 __m128i b00_odd = _mm_adds_epi16(db07, dy00_odd);
97 __m128i r00_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r00_even, r00_even),
98 _mm_packus_epi16(r00_odd, r00_odd));
99 __m128i g00_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g00_even, g00_even),
100 _mm_packus_epi16(g00_odd, g00_odd));
101 __m128i b00_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b00_even, b00_even),
102 _mm_packus_epi16(b00_odd, b00_odd));
103 __m128i rb00_07 = _mm_unpacklo_epi8(r00_0f, b00_0f);
104 __m128i rb00_8f = _mm_unpackhi_epi8(r00_0f, b00_0f);
105 __m128i ga00_07 = _mm_unpacklo_epi8(g00_0f, ALPHA);
106 __m128i ga00_8f = _mm_unpackhi_epi8(g00_0f, ALPHA);
107 __m128i rgba00_03 = _mm_unpacklo_epi8(rb00_07, ga00_07);
108 __m128i rgba00_47 = _mm_unpackhi_epi8(rb00_07, ga00_07);
109 __m128i rgba00_8b = _mm_unpacklo_epi8(rb00_8f, ga00_8f);
110 __m128i rgba00_cf = _mm_unpackhi_epi8(rb00_8f, ga00_8f);
111 _mm_store_si128(out0 + 0, rgba00_03);
112 _mm_store_si128(out0 + 1, rgba00_47);
113 _mm_store_si128(out0 + 2, rgba00_8b);
114 _mm_store_si128(out0 + 3, rgba00_cf);
115
116 // block bottom,left
117 __m128i y10_0f = _mm_load_si128(y1 + 0);
118 __m128i y10_even = _mm_and_si128(y10_0f, Y_MASK);
119 __m128i y10_odd = _mm_srli_epi16(y10_0f, 8);
120 __m128i dy10_even = _mm_srai_epi16(_mm_mullo_epi16(y10_even, COEF_Y), 6);
121 __m128i dy10_odd = _mm_srai_epi16(_mm_mullo_epi16(y10_odd, COEF_Y), 6);
122 __m128i r10_even = _mm_adds_epi16(dr07, dy10_even);
123 __m128i g10_even = _mm_adds_epi16(dg07, dy10_even);
124 __m128i b10_even = _mm_adds_epi16(db07, dy10_even);
125 __m128i r10_odd = _mm_adds_epi16(dr07, dy10_odd);
126 __m128i g10_odd = _mm_adds_epi16(dg07, dy10_odd);
127 __m128i b10_odd = _mm_adds_epi16(db07, dy10_odd);
128 __m128i r10_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r10_even, r10_even),
129 _mm_packus_epi16(r10_odd, r10_odd));
130 __m128i g10_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g10_even, g10_even),
131 _mm_packus_epi16(g10_odd, g10_odd));
132 __m128i b10_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b10_even, b10_even),
133 _mm_packus_epi16(b10_odd, b10_odd));
134 __m128i rb10_07 = _mm_unpacklo_epi8(r10_0f, b10_0f);
135 __m128i rb10_8f = _mm_unpackhi_epi8(r10_0f, b10_0f);
136 __m128i ga10_07 = _mm_unpacklo_epi8(g10_0f, ALPHA);
137 __m128i ga10_8f = _mm_unpackhi_epi8(g10_0f, ALPHA);
138 __m128i rgba10_03 = _mm_unpacklo_epi8(rb10_07, ga10_07);
139 __m128i rgba10_47 = _mm_unpackhi_epi8(rb10_07, ga10_07);
140 __m128i rgba10_8b = _mm_unpacklo_epi8(rb10_8f, ga10_8f);
141 __m128i rgba10_cf = _mm_unpackhi_epi8(rb10_8f, ga10_8f);
142 _mm_store_si128(out1 + 0, rgba10_03);
143 _mm_store_si128(out1 + 1, rgba10_47);
144 _mm_store_si128(out1 + 2, rgba10_8b);
145 _mm_store_si128(out1 + 3, rgba10_cf);
146
147 // right
148 __m128i u8f = _mm_unpackhi_epi8(u0f, ZERO);
149 __m128i v8f = _mm_unpackhi_epi8(v0f, ZERO);
150 __m128i mr8f = _mm_srai_epi16(_mm_mullo_epi16(v8f, RED_V), 6);
151 __m128i sg8f = _mm_mullo_epi16(v8f, GREEN_V);
152 __m128i tg8f = _mm_mullo_epi16(u8f, GREEN_U);
153 __m128i mg8f = _mm_srai_epi16(_mm_adds_epi16(sg8f, tg8f), 6);
154 __m128i mb8f = _mm_srli_epi16(_mm_mullo_epi16(u8f, BLUE_U), 6); // logical shift
155 __m128i dr8f = _mm_adds_epi16(mr8f, CNST_R);
156 __m128i dg8f = _mm_adds_epi16(mg8f, CNST_G);
157 __m128i db8f = _mm_adds_epi16(mb8f, CNST_B);
158
159 // block top,right
160 __m128i y01_0f = _mm_load_si128(y0 + 1);
161 __m128i y01_even = _mm_and_si128(y01_0f, Y_MASK);
162 __m128i y01_odd = _mm_srli_epi16(y01_0f, 8);
163 __m128i dy01_even = _mm_srai_epi16(_mm_mullo_epi16(y01_even, COEF_Y), 6);
164 __m128i dy01_odd = _mm_srai_epi16(_mm_mullo_epi16(y01_odd, COEF_Y), 6);
165 __m128i r01_even = _mm_adds_epi16(dr8f, dy01_even);
166 __m128i g01_even = _mm_adds_epi16(dg8f, dy01_even);
167 __m128i b01_even = _mm_adds_epi16(db8f, dy01_even);
168 __m128i r01_odd = _mm_adds_epi16(dr8f, dy01_odd);
169 __m128i g01_odd = _mm_adds_epi16(dg8f, dy01_odd);
170 __m128i b01_odd = _mm_adds_epi16(db8f, dy01_odd);
171 __m128i r01_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r01_even, r01_even),
172 _mm_packus_epi16(r01_odd, r01_odd));
173 __m128i g01_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g01_even, g01_even),
174 _mm_packus_epi16(g01_odd, g01_odd));
175 __m128i b01_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b01_even, b01_even),
176 _mm_packus_epi16(b01_odd, b01_odd));
177 __m128i rb01_07 = _mm_unpacklo_epi8(r01_0f, b01_0f);
178 __m128i rb01_8f = _mm_unpackhi_epi8(r01_0f, b01_0f);
179 __m128i ga01_07 = _mm_unpacklo_epi8(g01_0f, ALPHA);
180 __m128i ga01_8f = _mm_unpackhi_epi8(g01_0f, ALPHA);
181 __m128i rgba01_03 = _mm_unpacklo_epi8(rb01_07, ga01_07);
182 __m128i rgba01_47 = _mm_unpackhi_epi8(rb01_07, ga01_07);
183 __m128i rgba01_8b = _mm_unpacklo_epi8(rb01_8f, ga01_8f);
184 __m128i rgba01_cf = _mm_unpackhi_epi8(rb01_8f, ga01_8f);
185 _mm_store_si128(out0 + 4, rgba01_03);
186 _mm_store_si128(out0 + 5, rgba01_47);
187 _mm_store_si128(out0 + 6, rgba01_8b);
188 _mm_store_si128(out0 + 7, rgba01_cf);
189
190 // block bottom,right
191 __m128i y11_0f = _mm_load_si128(y1 + 1);
192 __m128i y11_even = _mm_and_si128(y11_0f, Y_MASK);
193 __m128i y11_odd = _mm_srli_epi16(y11_0f, 8);
194 __m128i dy11_even = _mm_srai_epi16(_mm_mullo_epi16(y11_even, COEF_Y), 6);
195 __m128i dy11_odd = _mm_srai_epi16(_mm_mullo_epi16(y11_odd, COEF_Y), 6);
196 __m128i r11_even = _mm_adds_epi16(dr8f, dy11_even);
197 __m128i g11_even = _mm_adds_epi16(dg8f, dy11_even);
198 __m128i b11_even = _mm_adds_epi16(db8f, dy11_even);
199 __m128i r11_odd = _mm_adds_epi16(dr8f, dy11_odd);
200 __m128i g11_odd = _mm_adds_epi16(dg8f, dy11_odd);
201 __m128i b11_odd = _mm_adds_epi16(db8f, dy11_odd);
202 __m128i r11_0f = _mm_unpackhi_epi8(_mm_packus_epi16(r11_even, r11_even),
203 _mm_packus_epi16(r11_odd, r11_odd));
204 __m128i g11_0f = _mm_unpackhi_epi8(_mm_packus_epi16(g11_even, g11_even),
205 _mm_packus_epi16(g11_odd, g11_odd));
206 __m128i b11_0f = _mm_unpackhi_epi8(_mm_packus_epi16(b11_even, b11_even),
207 _mm_packus_epi16(b11_odd, b11_odd));
208 __m128i rb11_07 = _mm_unpacklo_epi8(r11_0f, b11_0f);
209 __m128i rb11_8f = _mm_unpackhi_epi8(r11_0f, b11_0f);
210 __m128i ga11_07 = _mm_unpacklo_epi8(g11_0f, ALPHA);
211 __m128i ga11_8f = _mm_unpackhi_epi8(g11_0f, ALPHA);
212 __m128i rgba11_03 = _mm_unpacklo_epi8(rb11_07, ga11_07);
213 __m128i rgba11_47 = _mm_unpackhi_epi8(rb11_07, ga11_07);
214 __m128i rgba11_8b = _mm_unpacklo_epi8(rb11_8f, ga11_8f);
215 __m128i rgba11_cf = _mm_unpackhi_epi8(rb11_8f, ga11_8f);
216 _mm_store_si128(out1 + 4, rgba11_03);
217 _mm_store_si128(out1 + 5, rgba11_47);
218 _mm_store_si128(out1 + 6, rgba11_8b);
219 _mm_store_si128(out1 + 7, rgba11_cf);
220}
221
222static inline void convertHelperSSE2(
223 const th_ycbcr_buffer& buffer, RawFrame& output)
224{
225 const int width = buffer[0].width;
226 const size_t y_stride = buffer[0].stride;
227 const size_t uv_stride2 = buffer[1].stride / 2;
228
229 assert((width % 32) == 0);
230 assert((buffer[0].height % 2) == 0);
231
232 for (int y = 0; y < buffer[0].height; y += 2) {
233 const uint8_t* pY1 = buffer[0].data + (y + 0) * y_stride;
234 const uint8_t* pY2 = buffer[0].data + (y + 1) * y_stride;
235 const uint8_t* pCb = buffer[1].data + (y + 0) * uv_stride2;
236 const uint8_t* pCr = buffer[2].data + (y + 0) * uv_stride2;
237 auto out0 = output.getLineDirect<uint32_t>(y + 0);
238 auto out1 = output.getLineDirect<uint32_t>(y + 1);
239
240 for (int x = 0; x < width; x += 32) {
241 // convert a block of (32 x 2) pixels
242 yuv2rgb_sse2(pCb, pCr, pY1, pY2, &out0[x], &out1[x]);
243 pCb += 16;
244 pCr += 16;
245 pY1 += 32;
246 pY2 += 32;
247 }
248
249 output.setLineWidth(y + 0, width);
250 output.setLineWidth(y + 1, width);
251 }
252}
253
254#endif // __SSE2__
255
256static constexpr int PREC = 15;
257static constexpr int COEF_Y = int(1.164 * (1 << PREC) + 0.5); // prefer to use lrint() to round
258static constexpr int COEF_RV = int(1.596 * (1 << PREC) + 0.5); // but that's not (yet) constexpr
259static constexpr int COEF_GU = int(0.391 * (1 << PREC) + 0.5); // in current versions of c++
260static constexpr int COEF_GV = int(0.813 * (1 << PREC) + 0.5);
261static constexpr int COEF_BU = int(2.018 * (1 << PREC) + 0.5);
262
263struct Coefs {
264 std::array<int, 256> gu;
265 std::array<int, 256> gv;
266 std::array<int, 256> bu;
267 std::array<int, 256> rv;
268 std::array<int, 256> y;
269};
270
271[[nodiscard]] static constexpr Coefs getCoefs()
272{
273 Coefs coefs = {};
274 for (auto i : xrange(256)) {
275 coefs.gu[i] = -COEF_GU * (i - 128);
276 coefs.gv[i] = -COEF_GV * (i - 128);
277 coefs.bu[i] = COEF_BU * (i - 128);
278 coefs.rv[i] = COEF_RV * (i - 128);
279 coefs.y[i] = COEF_Y * (i - 16) + (PREC / 2);
280 }
281 return coefs;
282}
283
284template<std::unsigned_integral Pixel>
285[[nodiscard]] static inline Pixel calc(
286 const PixelFormat& format, 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 if constexpr (sizeof(Pixel) == 4) {
292 return (r << 0) | (g << 8) | (b << 16);
293 } else {
294 return static_cast<Pixel>(format.map(r, g, b));
295 }
296}
297
298template<std::unsigned_integral Pixel>
299static void convertHelper(const th_ycbcr_buffer& buffer, RawFrame& output,
300 const PixelFormat& format)
301{
302 assert(buffer[1].width * 2 == buffer[0].width);
303 assert(buffer[1].height * 2 == buffer[0].height);
304
305 static constexpr Coefs coefs = getCoefs();
306
307 const int width = buffer[0].width;
308 const size_t y_stride = buffer[0].stride;
309 const size_t uv_stride2 = buffer[1].stride / 2;
310
311 for (int y = 0; y < buffer[0].height; y += 2) {
312 const uint8_t* pY = buffer[0].data + y * y_stride;
313 const uint8_t* pCb = buffer[1].data + y * uv_stride2;
314 const uint8_t* pCr = buffer[2].data + y * uv_stride2;
315 auto out0 = output.getLineDirect<Pixel>(y + 0);
316 auto out1 = output.getLineDirect<Pixel>(y + 1);
317
318 for (int x = 0; x < width; x += 2, pY += 2, ++pCr, ++pCb) {
319 int ruv = coefs.rv[*pCr];
320 int guv = coefs.gu[*pCb] + coefs.gv[*pCr];
321 int buv = coefs.bu[*pCb];
322
323 int Y00 = coefs.y[pY[0]];
324 out0[x + 0] = calc<Pixel>(format, Y00, ruv, guv, buv);
325
326 int Y01 = coefs.y[pY[1]];
327 out0[x + 1] = calc<Pixel>(format, Y01, ruv, guv, buv);
328
329 int Y10 = coefs.y[pY[y_stride + 0]];
330 out1[x + 0] = calc<Pixel>(format, Y10, ruv, guv, buv);
331
332 int Y11 = coefs.y[pY[y_stride + 1]];
333 out1[x + 1] = calc<Pixel>(format, Y11, ruv, guv, buv);
334 }
335
336 output.setLineWidth(y + 0, width);
337 output.setLineWidth(y + 1, width);
338 }
339}
340
341void convert(const th_ycbcr_buffer& input, RawFrame& output)
342{
343 const PixelFormat& format = output.getPixelFormat();
344 if (format.getBytesPerPixel() == 4) {
345#ifdef __SSE2__
346 convertHelperSSE2(input, output);
347#else
348 convertHelper<uint32_t>(input, output, format);
349#endif
350 } else {
351 assert(format.getBytesPerPixel() == 2);
352 convertHelper<uint16_t>(input, output, format);
353 }
354}
355
356} // namespace openmsx::yuv2rgb
int g
const PixelFormat & getPixelFormat() const
Definition: FrameSource.hh:151
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:20
void setLineWidth(unsigned line, unsigned width)
Definition: RawFrame.hh:30
uint8_t clipIntToByte(int x)
Clip x to range [0,255].
Definition: Math.hh:60
void format(SectorAccessibleDisk &disk, bool dos1)
Format the given disk (= a single partition).
void convert(const th_ycbcr_buffer &input, RawFrame &output)
Definition: yuv2rgb.cc:341
uint32_t Pixel
std::array< int, 256 > gu
Definition: yuv2rgb.cc:264
std::array< int, 256 > gv
Definition: yuv2rgb.cc:265
std::array< int, 256 > y
Definition: yuv2rgb.cc:268
std::array< int, 256 > rv
Definition: yuv2rgb.cc:267
std::array< int, 256 > bu
Definition: yuv2rgb.cc:266
constexpr auto xrange(T e)
Definition: xrange.hh:133