OXYGENHUB
◆ main openmsx / src / video / scalers / LineScalers.hh
16.6 KiB · C++ · last changed chore(): initial commit 1 month ago source blame history raw
1 
#ifndef LINESCALERS_HH
2 
#define LINESCALERS_HH
3
4 
#include "PixelOperations.hh"
5
6 
#include "ranges.hh"
7 
#include "xrange.hh"
8
9 
#include <bit>
10 
#include <cassert>
11 
#include <cstddef>
12 
#include <cstdint>
13 
#include <ranges>
14 
#include <span>
15 
#ifdef __SSE2__
16 
#include "emmintrin.h"
17 
#endif
18
19 
namespace openmsx {
20
21 
using Pixel = uint32_t;
22
23 
/**  Scale_XonY functions
24 
 * Transforms an input line of pixel to an output line (possibly) with
25 
 * a different width. X input pixels are mapped on Y output pixels.
26 
 * @param in Input line
27 
 * @param out Output line
28 
 */
29 
void scale_1on3(std::span<const Pixel> in, std::span<Pixel> out);
30 
void scale_1on4(std::span<const Pixel> in, std::span<Pixel> out);
31 
void scale_1on6(std::span<const Pixel> in, std::span<Pixel> out);
32 
void scale_1on2(std::span<const Pixel> in, std::span<Pixel> out);
33 
void scale_2on1(std::span<const Pixel> in, std::span<Pixel> out);
34 
void scale_6on1(std::span<const Pixel> in, std::span<Pixel> out);
35 
void scale_4on1(std::span<const Pixel> in, std::span<Pixel> out);
36 
void scale_3on1(std::span<const Pixel> in, std::span<Pixel> out);
37 
void scale_3on2(std::span<const Pixel> in, std::span<Pixel> out);
38 
void scale_3on4(std::span<const Pixel> in, std::span<Pixel> out);
39 
void scale_3on8(std::span<const Pixel> in, std::span<Pixel> out);
40 
void scale_2on3(std::span<const Pixel> in, std::span<Pixel> out);
41 
void scale_4on3(std::span<const Pixel> in, std::span<Pixel> out);
42 
void scale_8on3(std::span<const Pixel> in, std::span<Pixel> out);
43 
void scale_2on9(std::span<const Pixel> in, std::span<Pixel> out);
44 
void scale_4on9(std::span<const Pixel> in, std::span<Pixel> out);
45 
void scale_8on9(std::span<const Pixel> in, std::span<Pixel> out);
46
47 
/**  BlendLines functor
48 
 * Generate an output line that is an interpolation of two input lines.
49 
 * @param in1 First input line
50 
 * @param in2 Second input line
51 
 * @param out Output line
52 
 * @param width Width of the lines in pixels
53 
 */
54 
template<unsigned w1 = 1, unsigned w2 = 1>
55 
void blendLines(std::span<const Pixel> in1, std::span<const Pixel> in2,
56 
                std::span<Pixel> out);
57
58 
/**  AlphaBlendLines functor
59 
 * Generate an output line that is a per-pixel-alpha-blend of the two input
60 
 * lines. The first input line contains the alpha-value per pixel.
61 
 * @param in1 First input line
62 
 * @param in2 Second input line
63 
 * @param out Output line
64 
 */
65 
void alphaBlendLines(std::span<const Pixel> in1, std::span<const Pixel> in2,
66 
                     std::span<Pixel> out);
67 
void alphaBlendLines(Pixel in1, std::span<const Pixel> in2,
68 
                     std::span<Pixel> out);
69
70
71 
// implementation
72
73 
template<unsigned N>
74 
static inline void scale_1onN(
75 
	std::span<const Pixel> in, std::span<Pixel> out)
76 
{
77 
	auto outWidth = out.size();
78 
	assert(in.size() == (outWidth / N));
79
80 
	size_t i = 0, j = 0;
81 
	for (/* */; i < (outWidth - (N - 1)); i += N, j += 1) {
82 
		Pixel pix = in[j];
83 
		for (auto k : xrange(N)) {
84 
			out[i + k] = pix;
85 
		}
86 
	}
87 
	for (auto k : xrange(N - 1)) {
88 
		if ((i + k) < outWidth) out[i + k] = 0;
89 
	}
90 
}
91
92 
inline void scale_1on3(std::span<const Pixel> in, std::span<Pixel> out)
93 
{
94 
	scale_1onN<3>(in, out);
95 
}
96
97 
inline void scale_1on4(std::span<const Pixel> in, std::span<Pixel> out)
98 
{
99 
	scale_1onN<4>(in, out);
100 
}
101
102 
inline void scale_1on6(std::span<const Pixel> in, std::span<Pixel> out)
103 
{
104 
	scale_1onN<6>(in, out);
105 
}
106
107 
#ifdef __SSE2__
108 
inline __m128i unpacklo(__m128i x, __m128i y)
109 
{
110 
	// 32bpp
111 
	return _mm_unpacklo_epi32(x, y);
112 
}
113 
inline __m128i unpackhi(__m128i x, __m128i y)
114 
{
115 
	// 32bpp
116 
	return _mm_unpackhi_epi32(x, y);
117 
}
118
119 
inline void scale_1on2_SSE(const Pixel* __restrict in_, Pixel* __restrict out_, size_t srcWidth)
120 
{
121 
	size_t bytes = srcWidth * sizeof(Pixel);
122 
	assert((bytes % (4 * sizeof(__m128i))) == 0);
123 
	assert(bytes != 0);
124
125 
	const auto* in  = std::bit_cast<const char*>(in_)  +     bytes;
126 
	      auto* out = std::bit_cast<      char*>(out_) + 2 * bytes;
127
128 
	auto x = -ptrdiff_t(bytes);
129 
	do {
130 
		__m128i a0 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x +  0));
131 
		__m128i a1 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x + 16));
132 
		__m128i a2 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x + 32));
133 
		__m128i a3 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + x + 48));
134 
		__m128i l0 = unpacklo(a0, a0);
135 
		__m128i h0 = unpackhi(a0, a0);
136 
		__m128i l1 = unpacklo(a1, a1);
137 
		__m128i h1 = unpackhi(a1, a1);
138 
		__m128i l2 = unpacklo(a2, a2);
139 
		__m128i h2 = unpackhi(a2, a2);
140 
		__m128i l3 = unpacklo(a3, a3);
141 
		__m128i h3 = unpackhi(a3, a3);
142 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +   0), l0);
143 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  16), h0);
144 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  32), l1);
145 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  48), h1);
146 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  64), l2);
147 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  80), h2);
148 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x +  96), l3);
149 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + 2*x + 112), h3);
150 
		x += 4 * sizeof(__m128i);
151 
	} while (x < 0);
152 
}
153 
#endif
154
155 
inline void scale_1on2(std::span<const Pixel> in, std::span<Pixel> out)
156 
{
157 
	// This is a fairly simple algorithm (output each input pixel twice).
158 
	// An ideal compiler should generate optimal (vector) code for it.
159 
	// I checked the 2013-05-29 dev snapshots of gcc-4.9 and clang-3.4:
160 
	// - Clang is not able to vectorize this loop. My best tuned C version
161 
	//   of this routine is a little over 4x slower than the tuned
162 
	//   SSE-intrinsics version.
163 
	// - Gcc can auto-vectorize this routine. Though my best tuned version
164 
	//   (I mean tuned to further improve the auto-vectorization, including
165 
	//   using the new __builtin_assume_aligned() intrinsic) still runs
166 
	//   approx 40% slower than the intrinsics version.
167 
	// Hopefully in some years the compilers have improved further so that
168 
	// the intrinsic version is no longer needed.
169 
	auto srcWidth = in.size();
170 
	assert((out.size() / 2) == srcWidth);
171
172 
#ifdef __SSE2__
173 
	size_t chunk = 4 * sizeof(__m128i) / sizeof(Pixel);
174 
	size_t srcWidth2 = srcWidth & ~(chunk - 1);
175 
	scale_1on2_SSE(in.data(), out.data(), srcWidth2);
176 
	in  = in .subspan(    srcWidth2);
177 
	out = out.subspan(2 * srcWidth2);
178 
	srcWidth -= srcWidth2;
179 
#endif
180
181 
	// C++ version. Used both on non-x86 machines and (possibly) on x86 for
182 
	// the last few pixels of the line.
183 
	for (auto x : xrange(srcWidth)) {
184 
		out[x * 2] = out[x * 2 + 1] = in[x];
185 
	}
186 
}
187
188 
#ifdef __SSE2__
189 
template<int IMM8> static inline __m128i shuffle(__m128i x, __m128i y)
190 
{
191 
	return _mm_castps_si128(_mm_shuffle_ps(
192 
		_mm_castsi128_ps(x), _mm_castsi128_ps(y), IMM8));
193 
}
194
195 
inline __m128i blend(__m128i x, __m128i y)
196 
{
197 
	// 32bpp
198 
	__m128i p = shuffle<0x88>(x, y);
199 
	__m128i q = shuffle<0xDD>(x, y);
200 
	return _mm_avg_epu8(p, q);
201 
}
202
203 
inline void scale_2on1_SSE(
204 
	const Pixel* __restrict in_, Pixel* __restrict out_, size_t dstBytes)
205 
{
206 
	assert((dstBytes % (4 * sizeof(__m128i))) == 0);
207 
	assert(dstBytes != 0);
208
209 
	const auto* in  = std::bit_cast<const char*>(in_)  + 2 * dstBytes;
210 
	      auto* out = std::bit_cast<      char*>(out_) +     dstBytes;
211
212 
	auto x = -ptrdiff_t(dstBytes);
213 
	do {
214 
		__m128i a0 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +   0));
215 
		__m128i a1 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  16));
216 
		__m128i a2 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  32));
217 
		__m128i a3 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  48));
218 
		__m128i a4 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  64));
219 
		__m128i a5 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  80));
220 
		__m128i a6 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x +  96));
221 
		__m128i a7 = _mm_loadu_si128(std::bit_cast<const __m128i*>(in + 2*x + 112));
222 
		__m128i b0 = blend(a0, a1);
223 
		__m128i b1 = blend(a2, a3);
224 
		__m128i b2 = blend(a4, a5);
225 
		__m128i b3 = blend(a6, a7);
226 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + x +  0), b0);
227 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + x + 16), b1);
228 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + x + 32), b2);
229 
		_mm_storeu_si128(std::bit_cast<__m128i*>(out + x + 48), b3);
230 
		x += 4 * sizeof(__m128i);
231 
	} while (x < 0);
232 
}
233 
#endif
234
235 
inline void scale_2on1(std::span<const Pixel> in, std::span<Pixel> out)
236 
{
237 
	assert((in.size() / 2) == out.size());
238 
	auto outWidth = out.size();
239 
#ifdef __SSE2__
240 
	auto n64 = (outWidth * sizeof(Pixel)) & ~63;
241 
	scale_2on1_SSE(in.data(), out.data(), n64); // process 64 byte chunks
242 
	outWidth &= ((64 / sizeof(Pixel)) - 1); // remaining pixels (if any)
243 
	if (outWidth == 0) [[likely]] return;
244 
	in  = in .subspan(2 * n64 / sizeof(Pixel));
245 
	out = out.subspan(    n64 / sizeof(Pixel));
246 
	// fallthrough to c++ version
247 
#endif
248 
	// pure C++ version
249 
	PixelOperations pixelOps;
250 
	for (auto i : xrange(outWidth)) {
251 
		out[i] = pixelOps.template blend<1, 1>(
252 
			in[2 * i + 0], in[2 * i + 1]);
253 
	}
254 
}
255
256 
inline void scale_6on1(std::span<const Pixel> in, std::span<Pixel> out)
257 
{
258 
	assert((in.size() / 6) == out.size());
259 
	PixelOperations pixelOps;
260 
	for (auto i : xrange(out.size())) {
261 
		out[i] = pixelOps.template blend<1, 1, 1, 1, 1, 1>(subspan<6>(in, 6 * i));
262 
	}
263 
}
264
265 
inline void scale_4on1(std::span<const Pixel> in, std::span<Pixel> out)
266 
{
267 
	assert((in.size() / 4) == out.size());
268 
	PixelOperations pixelOps;
269 
	for (auto i : xrange(out.size())) {
270 
		out[i] = pixelOps.template blend<1, 1, 1, 1>(subspan<4>(in, 4 * i));
271 
	}
272 
}
273
274 
inline void scale_3on1(std::span<const Pixel> in, std::span<Pixel> out)
275 
{
276 
	assert((in.size() / 3) == out.size());
277 
	PixelOperations pixelOps;
278 
	for (auto i : xrange(out.size())) {
279 
		out[i] = pixelOps.template blend<1, 1, 1>(subspan<3>(in, 3 * i));
280 
	}
281 
}
282
283 
inline void scale_3on2(std::span<const Pixel> in, std::span<Pixel> out)
284 
{
285 
	assert((in.size() / 3) == (out.size() / 2));
286 
	PixelOperations pixelOps;
287 
	size_t n = out.size();
288 
	size_t i = 0, j = 0;
289 
	for (/* */; i < (n - 1); i += 2, j += 3) {
290 
		out[i + 0] = pixelOps.template blend<2, 1>(subspan<2>(in, j + 0));
291 
		out[i + 1] = pixelOps.template blend<1, 2>(subspan<2>(in, j + 1));
292 
	}
293 
	if (i < n) out[i] = 0;
294 
}
295
296 
inline void scale_3on4(std::span<const Pixel> in, std::span<Pixel> out)
297 
{
298 
	assert((in.size() / 3) == (out.size() / 4));
299 
	PixelOperations pixelOps;
300 
	size_t n = out.size();
301 
	size_t i = 0, j = 0;
302 
	for (/* */; i < (n - 3); i += 4, j += 3) {
303 
		out[i + 0] = in[j + 0];
304 
		out[i + 1] = pixelOps.template blend<1, 2>(subspan<2>(in, j + 0));
305 
		out[i + 2] = pixelOps.template blend<2, 1>(subspan<2>(in, j + 1));
306 
		out[i + 3] = in[j + 2];
307 
	}
308 
	for (auto k : xrange(4 - 1)) {
309 
		if ((i + k) < n) out[i + k] = 0;
310 
	}
311 
}
312
313 
inline void scale_3on8(std::span<const Pixel> in, std::span<Pixel> out)
314 
{
315 
	assert((in.size() / 3) == (out.size() / 8));
316 
	PixelOperations pixelOps;
317 
	size_t n = out.size();
318 
	size_t i = 0, j = 0;
319 
	for (/* */; i < (n - 7); i += 8, j += 3) {
320 
		out[i + 0] = in[j + 0];
321 
		out[i + 1] = in[j + 0];
322 
		out[i + 2] = pixelOps.template blend<2, 1>(subspan<2>(in, j + 0));
323 
		out[i + 3] = in[j + 1];
324 
		out[i + 4] = in[j + 1];
325 
		out[i + 5] = pixelOps.template blend<1, 2>(subspan<2>(in, j + 1));
326 
		out[i + 6] = in[j + 2];
327 
		out[i + 7] = in[j + 2];
328 
	}
329 
	for (auto k : xrange(8 - 1)) {
330 
		if ((i + k) < n) out[i + k] = 0;
331 
	}
332 
}
333
334 
inline void scale_2on3(std::span<const Pixel> in, std::span<Pixel> out)
335 
{
336 
	assert((in.size() / 2) == (out.size() / 3));
337 
	PixelOperations pixelOps;
338 
	size_t n = out.size();
339 
	size_t i = 0, j = 0;
340 
	for (/* */; i < (n - 2); i += 3, j += 2) {
341 
		out[i + 0] = in[j + 0];
342 
		out[i + 1] = pixelOps.template blend<1, 1>(subspan<2>(in, j));
343 
		out[i + 2] = in[j + 1];
344 
	}
345 
	if ((i + 0) < n) out[i + 0] = 0;
346 
	if ((i + 1) < n) out[i + 1] = 0;
347 
}
348
349 
inline void scale_4on3(std::span<const Pixel> in, std::span<Pixel> out)
350 
{
351 
	assert((in.size() / 4) == (out.size() / 3));
352 
	PixelOperations pixelOps;
353 
	size_t n = out.size();
354 
	size_t i = 0, j = 0;
355 
	for (/* */; i < (n - 2); i += 3, j += 4) {
356 
		out[i + 0] = pixelOps.template blend<3, 1>(subspan<2>(in, j + 0));
357 
		out[i + 1] = pixelOps.template blend<1, 1>(subspan<2>(in, j + 1));
358 
		out[i + 2] = pixelOps.template blend<1, 3>(subspan<2>(in, j + 2));
359 
	}
360 
	if ((i + 0) < n) out[i + 0] = 0;
361 
	if ((i + 1) < n) out[i + 1] = 0;
362 
}
363
364 
inline void scale_8on3(std::span<const Pixel> in, std::span<Pixel> out)
365 
{
366 
	assert((in.size() / 8) == (out.size() / 3));
367 
	PixelOperations pixelOps;
368 
	size_t n = out.size();
369 
	size_t i = 0, j = 0;
370 
	for (/* */; i < (n - 2); i += 3, j += 8) {
371 
		out[i + 0] = pixelOps.template blend<3, 3, 2>   (subspan<3>(in, j + 0));
372 
		out[i + 1] = pixelOps.template blend<1, 3, 3, 1>(subspan<4>(in, j + 2));
373 
		out[i + 2] = pixelOps.template blend<2, 3, 3>   (subspan<3>(in, j + 5));
374 
	}
375 
	if ((i + 0) < n) out[i + 0] = 0;
376 
	if ((i + 1) < n) out[i + 1] = 0;
377 
}
378
379 
inline void scale_2on9(std::span<const Pixel> in, std::span<Pixel> out)
380 
{
381 
	assert((in.size() / 2) == (out.size() / 9));
382 
	PixelOperations pixelOps;
383 
	size_t n = out.size();
384 
	size_t i = 0, j = 0;
385 
	for (/* */; i < (n - 8); i += 9, j += 2) {
386 
		out[i + 0] = in[j + 0];
387 
		out[i + 1] = in[j + 0];
388 
		out[i + 2] = in[j + 0];
389 
		out[i + 3] = in[j + 0];
390 
		out[i + 4] = pixelOps.template blend<1, 1>(subspan<2>(in, j));
391 
		out[i + 5] = in[j + 1];
392 
		out[i + 6] = in[j + 1];
393 
		out[i + 7] = in[j + 1];
394 
		out[i + 8] = in[j + 1];
395 
	}
396 
	if ((i + 0) < n) out[i + 0] = 0;
397 
	if ((i + 1) < n) out[i + 1] = 0;
398 
	if ((i + 2) < n) out[i + 2] = 0;
399 
	if ((i + 3) < n) out[i + 3] = 0;
400 
	if ((i + 4) < n) out[i + 4] = 0;
401 
	if ((i + 5) < n) out[i + 5] = 0;
402 
	if ((i + 6) < n) out[i + 6] = 0;
403 
	if ((i + 7) < n) out[i + 7] = 0;
404 
}
405
406 
inline void scale_4on9(std::span<const Pixel> in, std::span<Pixel> out)
407 
{
408 
	assert((in.size() / 4) == (out.size() / 9));
409 
	PixelOperations pixelOps;
410 
	size_t n = out.size();
411 
	size_t i = 0, j = 0;
412 
	for (/* */; i < (n - 8); i += 9, j += 4) {
413 
		out[i + 0] = in[j + 0];
414 
		out[i + 1] = in[j + 0];
415 
		out[i + 2] = pixelOps.template blend<1, 3>(subspan<2>(in, j + 0));
416 
		out[i + 3] = in[j + 1];
417 
		out[i + 4] = pixelOps.template blend<1, 1>(subspan<2>(in, j + 1));
418 
		out[i + 5] = in[j + 2];
419 
		out[i + 6] = pixelOps.template blend<3, 1>(subspan<2>(in, j + 2));
420 
		out[i + 7] = in[j + 3];
421 
		out[i + 8] = in[j + 3];
422 
	}
423 
	if ((i + 0) < n) out[i + 0] = 0;
424 
	if ((i + 1) < n) out[i + 1] = 0;
425 
	if ((i + 2) < n) out[i + 2] = 0;
426 
	if ((i + 3) < n) out[i + 3] = 0;
427 
	if ((i + 4) < n) out[i + 4] = 0;
428 
	if ((i + 5) < n) out[i + 5] = 0;
429 
	if ((i + 6) < n) out[i + 6] = 0;
430 
	if ((i + 7) < n) out[i + 7] = 0;
431 
}
432
433 
inline void scale_8on9(std::span<const Pixel> in, std::span<Pixel> out)
434 
{
435 
	assert((in.size() / 8) == (out.size() / 9));
436 
	PixelOperations pixelOps;
437 
	size_t n = out.size();
438 
	size_t i = 0, j = 0;
439 
	for (/* */; i < (n - 8); i += 9, j += 8) {
440 
		out[i + 0] = in[j + 0];
441 
		out[i + 1] = pixelOps.template blend<1, 7>(subspan<2>(in, j + 0));
442 
		out[i + 2] = pixelOps.template blend<1, 3>(subspan<2>(in, j + 1));
443 
		out[i + 3] = pixelOps.template blend<3, 5>(subspan<2>(in, j + 2));
444 
		out[i + 4] = pixelOps.template blend<1, 1>(subspan<2>(in, j + 3));
445 
		out[i + 5] = pixelOps.template blend<5, 3>(subspan<2>(in, j + 4));
446 
		out[i + 6] = pixelOps.template blend<3, 1>(subspan<2>(in, j + 5));
447 
		out[i + 7] = pixelOps.template blend<7, 1>(subspan<2>(in, j + 6));
448 
		out[i + 8] = in[j + 7];
449 
	}
450 
	if ((i + 0) < n) out[i + 0] = 0;
451 
	if ((i + 1) < n) out[i + 1] = 0;
452 
	if ((i + 2) < n) out[i + 2] = 0;
453 
	if ((i + 3) < n) out[i + 3] = 0;
454 
	if ((i + 4) < n) out[i + 4] = 0;
455 
	if ((i + 5) < n) out[i + 5] = 0;
456 
	if ((i + 6) < n) out[i + 6] = 0;
457 
	if ((i + 7) < n) out[i + 7] = 0;
458 
}
459
460 
template<unsigned w1, unsigned w2>
461 
void blendLines(std::span<const Pixel> in1, std::span<const Pixel> in2, std::span<Pixel> out)
462 
{
463 
	// It _IS_ allowed that the output is the same as one of the inputs.
464 
	// TODO SSE optimizations
465 
	// pure C++ version
466 
	assert(in1.size() == in2.size());
467 
	assert(in1.size() == out.size());
468 
	PixelOperations pixelOps;
469 
	for (auto [i1, i2, o] : std::views::zip(in1, in2, out)) {
470 
		o = pixelOps.template blend<w1, w2>(i1, i2);
471 
	}
472 
}
473
474 
inline void alphaBlendLines(
475 
	std::span<const Pixel> in1, std::span<const Pixel> in2, std::span<Pixel> out)
476 
{
477 
	// It _IS_ allowed that the output is the same as one of the inputs.
478 
	assert(in1.size() == in2.size());
479 
	assert(in1.size() == out.size());
480 
	PixelOperations pixelOps;
481 
	for (auto [i1, i2, o] : std::views::zip(in1, in2, out)) {
482 
		o = pixelOps.alphaBlend(i1, i2);
483 
	}
484 
}
485
486 
inline void alphaBlendLines(
487 
	Pixel in1, std::span<const Pixel> in2, std::span<Pixel> out)
488 
{
489 
	// It _IS_ allowed that the output is the same as the input.
490
491 
	// ATM this routine is only called when 'in1' is not fully opaque nor
492 
	// fully transparent.
493 
	assert(in2.size() == out.size());
494
495 
	PixelOperations pixelOps;
496 
	unsigned alpha = pixelOps.alpha(in1);
497
498 
	// When one of the two colors is loop-invariant, using the
499 
	// pre-multiplied-alpha-blending equation is a tiny bit more efficient
500 
	// than using alphaBlend() or even lerp().
501 
	//    for (auto i : xrange(width)) {
502 
	//        out[i] = pixelOps.lerp(in1, in2[i], alpha);
503 
	//    }
504 
	Pixel in1M = pixelOps.multiply(in1, alpha);
505 
	unsigned alpha2 = 256 - alpha;
506 
	for (auto [i2, o] : std::views::zip(in2, out)) {
507 
		o = in1M + pixelOps.multiply(i2, alpha2);
508 
	}
509 
}
510
511 
} // namespace openmsx
512
513 
#endif