1// Copyright 2011 Google Inc. All Rights Reserved.
2//
3// Use of this source code is governed by a BSD-style license
4// that can be found in the COPYING file in the root of the source
5// tree. An additional intellectual property rights grant can be found
6// in the file PATENTS. All contributing project authors may
7// be found in the AUTHORS file in the root of the source tree.
8// -----------------------------------------------------------------------------
9//
10// SSE2 version of YUV to RGB upsampling functions.
11//
12// Author: somnath@google.com (Somnath Banerjee)
13
14#include "src/dsp/dsp.h"
15
16#if defined(WEBP_USE_SSE2)
17
18#include <assert.h>
19#include <emmintrin.h>
20#include <string.h>
21#include "src/dsp/yuv.h"
22
23#ifdef FANCY_UPSAMPLING
24
25// We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows
26// u = (9*a + 3*b + 3*c + d + 8) / 16
27// = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2
28// = (a + m + 1) / 2
29// where m = (a + 3*b + 3*c + d) / 8
30// = ((a + b + c + d) / 2 + b + c) / 4
31//
32// Let's say k = (a + b + c + d) / 4.
33// We can compute k as
34// k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1
35// where s = (a + d + 1) / 2 and t = (b + c + 1) / 2
36//
37// Then m can be written as
38// m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1
39
40// Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1
41#define GET_M(ij, in, out) do { \
42 const __m128i tmp0 = _mm_avg_epu8(k, (in)); /* (k + in + 1) / 2 */ \
43 const __m128i tmp1 = _mm_and_si128((ij), st); /* (ij) & (s^t) */ \
44 const __m128i tmp2 = _mm_xor_si128(k, (in)); /* (k^in) */ \
45 const __m128i tmp3 = _mm_or_si128(tmp1, tmp2); /* ((ij) & (s^t)) | (k^in) */\
46 const __m128i tmp4 = _mm_and_si128(tmp3, one); /* & 1 -> lsb_correction */ \
47 (out) = _mm_sub_epi8(tmp0, tmp4); /* (k + in + 1) / 2 - lsb_correction */ \
48} while (0)
49
50// pack and store two alternating pixel rows
51#define PACK_AND_STORE(a, b, da, db, out) do { \
52 const __m128i t_a = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \
53 const __m128i t_b = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \
54 const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b); \
55 const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b); \
56 _mm_store_si128(((__m128i*)(out)) + 0, t_1); \
57 _mm_store_si128(((__m128i*)(out)) + 1, t_2); \
58} while (0)
59
60// Loads 17 pixels each from rows r1 and r2 and generates 32 pixels.
61#define UPSAMPLE_32PIXELS(r1, r2, out) { \
62 const __m128i one = _mm_set1_epi8(1); \
63 const __m128i a = _mm_loadu_si128((const __m128i*)&(r1)[0]); \
64 const __m128i b = _mm_loadu_si128((const __m128i*)&(r1)[1]); \
65 const __m128i c = _mm_loadu_si128((const __m128i*)&(r2)[0]); \
66 const __m128i d = _mm_loadu_si128((const __m128i*)&(r2)[1]); \
67 \
68 const __m128i s = _mm_avg_epu8(a, d); /* s = (a + d + 1) / 2 */ \
69 const __m128i t = _mm_avg_epu8(b, c); /* t = (b + c + 1) / 2 */ \
70 const __m128i st = _mm_xor_si128(s, t); /* st = s^t */ \
71 \
72 const __m128i ad = _mm_xor_si128(a, d); /* ad = a^d */ \
73 const __m128i bc = _mm_xor_si128(b, c); /* bc = b^c */ \
74 \
75 const __m128i t1 = _mm_or_si128(ad, bc); /* (a^d) | (b^c) */ \
76 const __m128i t2 = _mm_or_si128(t1, st); /* (a^d) | (b^c) | (s^t) */ \
77 const __m128i t3 = _mm_and_si128(t2, one); /* (a^d) | (b^c) | (s^t) & 1 */ \
78 const __m128i t4 = _mm_avg_epu8(s, t); \
79 const __m128i k = _mm_sub_epi8(t4, t3); /* k = (a + b + c + d) / 4 */ \
80 __m128i diag1, diag2; \
81 \
82 GET_M(bc, t, diag1); /* diag1 = (a + 3b + 3c + d) / 8 */ \
83 GET_M(ad, s, diag2); /* diag2 = (3a + b + c + 3d) / 8 */ \
84 \
85 /* pack the alternate pixels */ \
86 PACK_AND_STORE(a, b, diag1, diag2, (out) + 0); /* store top */ \
87 PACK_AND_STORE(c, d, diag2, diag1, (out) + 2 * 32); /* store bottom */ \
88}
89
90// Turn the macro into a function for reducing code-size when non-critical
91static void Upsample32Pixels_SSE2(const uint8_t r1[], const uint8_t r2[],
92 uint8_t* const out) {
93 UPSAMPLE_32PIXELS(r1, r2, out);
94}
95
96#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \
97 uint8_t r1[17], r2[17]; \
98 memcpy(r1, (tb), (num_pixels)); \
99 memcpy(r2, (bb), (num_pixels)); \
100 /* replicate last byte */ \
101 memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels)); \
102 memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels)); \
103 /* using the shared function instead of the macro saves ~3k code size */ \
104 Upsample32Pixels_SSE2(r1, r2, out); \
105}
106
107#define CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, \
108 top_dst, bottom_dst, cur_x) do { \
109 FUNC##32_SSE2((top_y) + (cur_x), r_u, r_v, (top_dst) + (cur_x) * (XSTEP)); \
110 if ((bottom_y) != NULL) { \
111 FUNC##32_SSE2((bottom_y) + (cur_x), r_u + 64, r_v + 64, \
112 (bottom_dst) + (cur_x) * (XSTEP)); \
113 } \
114} while (0)
115
116#define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \
117static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \
118 const uint8_t* top_u, const uint8_t* top_v, \
119 const uint8_t* cur_u, const uint8_t* cur_v, \
120 uint8_t* top_dst, uint8_t* bottom_dst, int len) { \
121 int uv_pos, pos; \
122 /* 16byte-aligned array to cache reconstructed u and v */ \
123 uint8_t uv_buf[14 * 32 + 15] = { 0 }; \
124 uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~(uintptr_t)15); \
125 uint8_t* const r_v = r_u + 32; \
126 \
127 assert(top_y != NULL); \
128 { /* Treat the first pixel in regular way */ \
129 const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \
130 const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \
131 const int u0_t = (top_u[0] + u_diag) >> 1; \
132 const int v0_t = (top_v[0] + v_diag) >> 1; \
133 FUNC(top_y[0], u0_t, v0_t, top_dst); \
134 if (bottom_y != NULL) { \
135 const int u0_b = (cur_u[0] + u_diag) >> 1; \
136 const int v0_b = (cur_v[0] + v_diag) >> 1; \
137 FUNC(bottom_y[0], u0_b, v0_b, bottom_dst); \
138 } \
139 } \
140 /* For UPSAMPLE_32PIXELS, 17 u/v values must be read-able for each block */ \
141 for (pos = 1, uv_pos = 0; pos + 32 + 1 <= len; pos += 32, uv_pos += 16) { \
142 UPSAMPLE_32PIXELS(top_u + uv_pos, cur_u + uv_pos, r_u); \
143 UPSAMPLE_32PIXELS(top_v + uv_pos, cur_v + uv_pos, r_v); \
144 CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst, pos); \
145 } \
146 if (len > 1) { \
147 const int left_over = ((len + 1) >> 1) - (pos >> 1); \
148 uint8_t* const tmp_top_dst = r_u + 4 * 32; \
149 uint8_t* const tmp_bottom_dst = tmp_top_dst + 4 * 32; \
150 uint8_t* const tmp_top = tmp_bottom_dst + 4 * 32; \
151 uint8_t* const tmp_bottom = (bottom_y == NULL) ? NULL : tmp_top + 32; \
152 assert(left_over > 0); \
153 UPSAMPLE_LAST_BLOCK(top_u + uv_pos, cur_u + uv_pos, left_over, r_u); \
154 UPSAMPLE_LAST_BLOCK(top_v + uv_pos, cur_v + uv_pos, left_over, r_v); \
155 memcpy(tmp_top, top_y + pos, len - pos); \
156 if (bottom_y != NULL) memcpy(tmp_bottom, bottom_y + pos, len - pos); \
157 CONVERT2RGB_32(FUNC, XSTEP, tmp_top, tmp_bottom, tmp_top_dst, \
158 tmp_bottom_dst, 0); \
159 memcpy(top_dst + pos * (XSTEP), tmp_top_dst, (len - pos) * (XSTEP)); \
160 if (bottom_y != NULL) { \
161 memcpy(bottom_dst + pos * (XSTEP), tmp_bottom_dst, \
162 (len - pos) * (XSTEP)); \
163 } \
164 } \
165}
166
167// SSE2 variants of the fancy upsampler.
168SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePair_SSE2, VP8YuvToRgba, 4)
169SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePair_SSE2, VP8YuvToBgra, 4)
170
171#if !defined(WEBP_REDUCE_CSP)
172SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair_SSE2, VP8YuvToRgb, 3)
173SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair_SSE2, VP8YuvToBgr, 3)
174SSE2_UPSAMPLE_FUNC(UpsampleArgbLinePair_SSE2, VP8YuvToArgb, 4)
175SSE2_UPSAMPLE_FUNC(UpsampleRgba4444LinePair_SSE2, VP8YuvToRgba4444, 2)
176SSE2_UPSAMPLE_FUNC(UpsampleRgb565LinePair_SSE2, VP8YuvToRgb565, 2)
177#endif // WEBP_REDUCE_CSP
178
179#undef GET_M
180#undef PACK_AND_STORE
181#undef UPSAMPLE_32PIXELS
182#undef UPSAMPLE_LAST_BLOCK
183#undef CONVERT2RGB
184#undef CONVERT2RGB_32
185#undef SSE2_UPSAMPLE_FUNC
186
187//------------------------------------------------------------------------------
188// Entry point
189
190extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
191
192extern void WebPInitUpsamplersSSE2(void);
193
194WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE2(void) {
195 WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_SSE2;
196 WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair_SSE2;
197 WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair_SSE2;
198 WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair_SSE2;
199#if !defined(WEBP_REDUCE_CSP)
200 WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair_SSE2;
201 WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair_SSE2;
202 WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_SSE2;
203 WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair_SSE2;
204 WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair_SSE2;
205 WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_SSE2;
206 WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_SSE2;
207#endif // WEBP_REDUCE_CSP
208}
209
210#endif // FANCY_UPSAMPLING
211
212//------------------------------------------------------------------------------
213
214extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
215extern void WebPInitYUV444ConvertersSSE2(void);
216
217#define YUV444_FUNC(FUNC_NAME, CALL, CALL_C, XSTEP) \
218extern void CALL_C(const uint8_t* y, const uint8_t* u, const uint8_t* v, \
219 uint8_t* dst, int len); \
220static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \
221 uint8_t* dst, int len) { \
222 int i; \
223 const int max_len = len & ~31; \
224 for (i = 0; i < max_len; i += 32) { \
225 CALL(y + i, u + i, v + i, dst + i * (XSTEP)); \
226 } \
227 if (i < len) { /* C-fallback */ \
228 CALL_C(y + i, u + i, v + i, dst + i * (XSTEP), len - i); \
229 } \
230}
231
232YUV444_FUNC(Yuv444ToRgba_SSE2, VP8YuvToRgba32_SSE2, WebPYuv444ToRgba_C, 4);
233YUV444_FUNC(Yuv444ToBgra_SSE2, VP8YuvToBgra32_SSE2, WebPYuv444ToBgra_C, 4);
234#if !defined(WEBP_REDUCE_CSP)
235YUV444_FUNC(Yuv444ToRgb_SSE2, VP8YuvToRgb32_SSE2, WebPYuv444ToRgb_C, 3);
236YUV444_FUNC(Yuv444ToBgr_SSE2, VP8YuvToBgr32_SSE2, WebPYuv444ToBgr_C, 3);
237YUV444_FUNC(Yuv444ToArgb_SSE2, VP8YuvToArgb32_SSE2, WebPYuv444ToArgb_C, 4)
238YUV444_FUNC(Yuv444ToRgba4444_SSE2, VP8YuvToRgba444432_SSE2, \
239 WebPYuv444ToRgba4444_C, 2)
240YUV444_FUNC(Yuv444ToRgb565_SSE2, VP8YuvToRgb56532_SSE2, WebPYuv444ToRgb565_C, 2)
241#endif // WEBP_REDUCE_CSP
242
243WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE2(void) {
244 WebPYUV444Converters[MODE_RGBA] = Yuv444ToRgba_SSE2;
245 WebPYUV444Converters[MODE_BGRA] = Yuv444ToBgra_SSE2;
246 WebPYUV444Converters[MODE_rgbA] = Yuv444ToRgba_SSE2;
247 WebPYUV444Converters[MODE_bgrA] = Yuv444ToBgra_SSE2;
248#if !defined(WEBP_REDUCE_CSP)
249 WebPYUV444Converters[MODE_RGB] = Yuv444ToRgb_SSE2;
250 WebPYUV444Converters[MODE_BGR] = Yuv444ToBgr_SSE2;
251 WebPYUV444Converters[MODE_ARGB] = Yuv444ToArgb_SSE2;
252 WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444_SSE2;
253 WebPYUV444Converters[MODE_RGB_565] = Yuv444ToRgb565_SSE2;
254 WebPYUV444Converters[MODE_Argb] = Yuv444ToArgb_SSE2;
255 WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444_SSE2;
256#endif // WEBP_REDUCE_CSP
257}
258
259#else
260
261WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersSSE2)
262
263#endif // WEBP_USE_SSE2
264
265#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_SSE2))
266WEBP_DSP_INIT_STUB(WebPInitUpsamplersSSE2)
267#endif
268