1 // Copyright 2011 Google Inc. All Rights Reserved.
2 //
3 // This code is licensed under the same terms as WebM:
4 // Software License Agreement: http://www.webmproject.org/license/software/
5 // Additional IP Rights Grant: http://www.webmproject.org/license/additional/
6 // -----------------------------------------------------------------------------
7 //
8 // SSE2 version of YUV to RGB upsampling functions.
9 //
10 // Author: somnath@google.com (Somnath Banerjee)
11
12 #include "./dsp.h"
13
14 #if defined(__cplusplus) || defined(c_plusplus)
15 extern "C" {
16 #endif
17
18 #if defined(WEBP_USE_SSE2)
19
20 #include <assert.h>
21 #include <emmintrin.h>
22 #include <string.h>
23 #include "./yuv.h"
24
25 #ifdef FANCY_UPSAMPLING
26
27 // We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows
28 // u = (9*a + 3*b + 3*c + d + 8) / 16
29 // = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2
30 // = (a + m + 1) / 2
31 // where m = (a + 3*b + 3*c + d) / 8
32 // = ((a + b + c + d) / 2 + b + c) / 4
33 //
34 // Let's say k = (a + b + c + d) / 4.
35 // We can compute k as
36 // k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1
37 // where s = (a + d + 1) / 2 and t = (b + c + 1) / 2
38 //
39 // Then m can be written as
40 // m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1
41
42 // Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1
43 #define GET_M(ij, in, out) do { \
44 const __m128i tmp0 = _mm_avg_epu8(k, (in)); /* (k + in + 1) / 2 */ \
45 const __m128i tmp1 = _mm_and_si128((ij), st); /* (ij) & (s^t) */ \
46 const __m128i tmp2 = _mm_xor_si128(k, (in)); /* (k^in) */ \
47 const __m128i tmp3 = _mm_or_si128(tmp1, tmp2); /* ((ij) & (s^t)) | (k^in) */\
48 const __m128i tmp4 = _mm_and_si128(tmp3, one); /* & 1 -> lsb_correction */ \
49 (out) = _mm_sub_epi8(tmp0, tmp4); /* (k + in + 1) / 2 - lsb_correction */ \
50 } while (0)
51
52 // pack and store two alterning pixel rows
53 #define PACK_AND_STORE(a, b, da, db, out) do { \
54 const __m128i t_a = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \
55 const __m128i t_b = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \
56 const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b); \
57 const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b); \
58 _mm_store_si128(((__m128i*)(out)) + 0, t_1); \
59 _mm_store_si128(((__m128i*)(out)) + 1, t_2); \
60 } while (0)
61
62 // Loads 17 pixels each from rows r1 and r2 and generates 32 pixels.
63 #define UPSAMPLE_32PIXELS(r1, r2, out) { \
64 const __m128i one = _mm_set1_epi8(1); \
65 const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]); \
66 const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]); \
67 const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]); \
68 const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]); \
69 \
70 const __m128i s = _mm_avg_epu8(a, d); /* s = (a + d + 1) / 2 */ \
71 const __m128i t = _mm_avg_epu8(b, c); /* t = (b + c + 1) / 2 */ \
72 const __m128i st = _mm_xor_si128(s, t); /* st = s^t */ \
73 \
74 const __m128i ad = _mm_xor_si128(a, d); /* ad = a^d */ \
75 const __m128i bc = _mm_xor_si128(b, c); /* bc = b^c */ \
76 \
77 const __m128i t1 = _mm_or_si128(ad, bc); /* (a^d) | (b^c) */ \
78 const __m128i t2 = _mm_or_si128(t1, st); /* (a^d) | (b^c) | (s^t) */ \
79 const __m128i t3 = _mm_and_si128(t2, one); /* (a^d) | (b^c) | (s^t) & 1 */ \
80 const __m128i t4 = _mm_avg_epu8(s, t); \
81 const __m128i k = _mm_sub_epi8(t4, t3); /* k = (a + b + c + d) / 4 */ \
82 __m128i diag1, diag2; \
83 \
84 GET_M(bc, t, diag1); /* diag1 = (a + 3b + 3c + d) / 8 */ \
85 GET_M(ad, s, diag2); /* diag2 = (3a + b + c + 3d) / 8 */ \
86 \
87 /* pack the alternate pixels */ \
88 PACK_AND_STORE(a, b, diag1, diag2, &(out)[0 * 32]); \
89 PACK_AND_STORE(c, d, diag2, diag1, &(out)[2 * 32]); \
90 }
91
92 // Turn the macro into a function for reducing code-size when non-critical
Upsample32Pixels(const uint8_t r1[],const uint8_t r2[],uint8_t * const out)93 static void Upsample32Pixels(const uint8_t r1[], const uint8_t r2[],
94 uint8_t* const out) {
95 UPSAMPLE_32PIXELS(r1, r2, out);
96 }
97
98 #define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \
99 uint8_t r1[17], r2[17]; \
100 memcpy(r1, (tb), (num_pixels)); \
101 memcpy(r2, (bb), (num_pixels)); \
102 /* replicate last byte */ \
103 memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels)); \
104 memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels)); \
105 /* using the shared function instead of the macro saves ~3k code size */ \
106 Upsample32Pixels(r1, r2, out); \
107 }
108
109 #define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, uv, \
110 top_dst, bottom_dst, cur_x, num_pixels) { \
111 int n; \
112 if (top_y) { \
113 for (n = 0; n < (num_pixels); ++n) { \
114 FUNC(top_y[(cur_x) + n], (uv)[n], (uv)[32 + n], \
115 top_dst + ((cur_x) + n) * XSTEP); \
116 } \
117 } \
118 if (bottom_y) { \
119 for (n = 0; n < (num_pixels); ++n) { \
120 FUNC(bottom_y[(cur_x) + n], (uv)[64 + n], (uv)[64 + 32 + n], \
121 bottom_dst + ((cur_x) + n) * XSTEP); \
122 } \
123 } \
124 }
125
126 #define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \
127 static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \
128 const uint8_t* top_u, const uint8_t* top_v, \
129 const uint8_t* cur_u, const uint8_t* cur_v, \
130 uint8_t* top_dst, uint8_t* bottom_dst, int len) { \
131 int block; \
132 /* 16 byte aligned array to cache reconstructed u and v */ \
133 uint8_t uv_buf[4 * 32 + 15]; \
134 uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \
135 const int uv_len = (len + 1) >> 1; \
136 /* 17 pixels must be read-able for each block */ \
137 const int num_blocks = (uv_len - 1) >> 4; \
138 const int leftover = uv_len - num_blocks * 16; \
139 const int last_pos = 1 + 32 * num_blocks; \
140 \
141 const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \
142 const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \
143 \
144 assert(len > 0); \
145 /* Treat the first pixel in regular way */ \
146 if (top_y) { \
147 const int u0 = (top_u[0] + u_diag) >> 1; \
148 const int v0 = (top_v[0] + v_diag) >> 1; \
149 FUNC(top_y[0], u0, v0, top_dst); \
150 } \
151 if (bottom_y) { \
152 const int u0 = (cur_u[0] + u_diag) >> 1; \
153 const int v0 = (cur_v[0] + v_diag) >> 1; \
154 FUNC(bottom_y[0], u0, v0, bottom_dst); \
155 } \
156 \
157 for (block = 0; block < num_blocks; ++block) { \
158 UPSAMPLE_32PIXELS(top_u, cur_u, r_uv + 0 * 32); \
159 UPSAMPLE_32PIXELS(top_v, cur_v, r_uv + 1 * 32); \
160 CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst, \
161 32 * block + 1, 32) \
162 top_u += 16; \
163 cur_u += 16; \
164 top_v += 16; \
165 cur_v += 16; \
166 } \
167 \
168 UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv + 0 * 32); \
169 UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 1 * 32); \
170 CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst, \
171 last_pos, len - last_pos); \
172 }
173
174 // SSE2 variants of the fancy upsampler.
175 SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePairSSE2, VP8YuvToRgb, 3)
176 SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePairSSE2, VP8YuvToBgr, 3)
177 SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePairSSE2, VP8YuvToRgba, 4)
178 SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4)
179
180 #undef GET_M
181 #undef PACK_AND_STORE
182 #undef UPSAMPLE_32PIXELS
183 #undef UPSAMPLE_LAST_BLOCK
184 #undef CONVERT2RGB
185 #undef SSE2_UPSAMPLE_FUNC
186
187 #endif // FANCY_UPSAMPLING
188
189 #endif // WEBP_USE_SSE2
190
191 //------------------------------------------------------------------------------
192
193 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
194
WebPInitUpsamplersSSE2(void)195 void WebPInitUpsamplersSSE2(void) {
196 #if defined(WEBP_USE_SSE2)
197 WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePairSSE2;
198 WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2;
199 WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePairSSE2;
200 WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2;
201 #endif // WEBP_USE_SSE2
202 }
203
WebPInitPremultiplySSE2(void)204 void WebPInitPremultiplySSE2(void) {
205 #if defined(WEBP_USE_SSE2)
206 WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2;
207 WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2;
208 #endif // WEBP_USE_SSE2
209 }
210
211 #if defined(__cplusplus) || defined(c_plusplus)
212 } // extern "C"
213 #endif
214
215
216