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 (at) google.com (Somnath Banerjee) 13 14 #include "./dsp.h" 15 16 #if defined(__cplusplus) || defined(c_plusplus) 17 extern "C" { 18 #endif 19 20 #if defined(WEBP_USE_SSE2) 21 22 #include <assert.h> 23 #include <emmintrin.h> 24 #include <string.h> 25 #include "./yuv.h" 26 27 #ifdef FANCY_UPSAMPLING 28 29 // We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows 30 // u = (9*a + 3*b + 3*c + d + 8) / 16 31 // = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2 32 // = (a + m + 1) / 2 33 // where m = (a + 3*b + 3*c + d) / 8 34 // = ((a + b + c + d) / 2 + b + c) / 4 35 // 36 // Let's say k = (a + b + c + d) / 4. 37 // We can compute k as 38 // k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1 39 // where s = (a + d + 1) / 2 and t = (b + c + 1) / 2 40 // 41 // Then m can be written as 42 // m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1 43 44 // Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1 45 #define GET_M(ij, in, out) do { \ 46 const __m128i tmp0 = _mm_avg_epu8(k, (in)); /* (k + in + 1) / 2 */ \ 47 const __m128i tmp1 = _mm_and_si128((ij), st); /* (ij) & (s^t) */ \ 48 const __m128i tmp2 = _mm_xor_si128(k, (in)); /* (k^in) */ \ 49 const __m128i tmp3 = _mm_or_si128(tmp1, tmp2); /* ((ij) & (s^t)) | (k^in) */\ 50 const __m128i tmp4 = _mm_and_si128(tmp3, one); /* & 1 -> lsb_correction */ \ 51 (out) = _mm_sub_epi8(tmp0, tmp4); /* (k + in + 1) / 2 - lsb_correction */ \ 52 } while (0) 53 54 // pack and store two alterning pixel rows 55 #define PACK_AND_STORE(a, b, da, db, out) do { \ 56 const __m128i t_a = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \ 57 const __m128i t_b = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \ 58 const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b); \ 59 const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b); \ 60 _mm_store_si128(((__m128i*)(out)) + 0, t_1); \ 61 _mm_store_si128(((__m128i*)(out)) + 1, t_2); \ 62 } while (0) 63 64 // Loads 17 pixels each from rows r1 and r2 and generates 32 pixels. 65 #define UPSAMPLE_32PIXELS(r1, r2, out) { \ 66 const __m128i one = _mm_set1_epi8(1); \ 67 const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]); \ 68 const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]); \ 69 const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]); \ 70 const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]); \ 71 \ 72 const __m128i s = _mm_avg_epu8(a, d); /* s = (a + d + 1) / 2 */ \ 73 const __m128i t = _mm_avg_epu8(b, c); /* t = (b + c + 1) / 2 */ \ 74 const __m128i st = _mm_xor_si128(s, t); /* st = s^t */ \ 75 \ 76 const __m128i ad = _mm_xor_si128(a, d); /* ad = a^d */ \ 77 const __m128i bc = _mm_xor_si128(b, c); /* bc = b^c */ \ 78 \ 79 const __m128i t1 = _mm_or_si128(ad, bc); /* (a^d) | (b^c) */ \ 80 const __m128i t2 = _mm_or_si128(t1, st); /* (a^d) | (b^c) | (s^t) */ \ 81 const __m128i t3 = _mm_and_si128(t2, one); /* (a^d) | (b^c) | (s^t) & 1 */ \ 82 const __m128i t4 = _mm_avg_epu8(s, t); \ 83 const __m128i k = _mm_sub_epi8(t4, t3); /* k = (a + b + c + d) / 4 */ \ 84 __m128i diag1, diag2; \ 85 \ 86 GET_M(bc, t, diag1); /* diag1 = (a + 3b + 3c + d) / 8 */ \ 87 GET_M(ad, s, diag2); /* diag2 = (3a + b + c + 3d) / 8 */ \ 88 \ 89 /* pack the alternate pixels */ \ 90 PACK_AND_STORE(a, b, diag1, diag2, &(out)[0 * 32]); \ 91 PACK_AND_STORE(c, d, diag2, diag1, &(out)[2 * 32]); \ 92 } 93 94 // Turn the macro into a function for reducing code-size when non-critical 95 static void Upsample32Pixels(const uint8_t r1[], const uint8_t r2[], 96 uint8_t* const out) { 97 UPSAMPLE_32PIXELS(r1, r2, out); 98 } 99 100 #define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \ 101 uint8_t r1[17], r2[17]; \ 102 memcpy(r1, (tb), (num_pixels)); \ 103 memcpy(r2, (bb), (num_pixels)); \ 104 /* replicate last byte */ \ 105 memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels)); \ 106 memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels)); \ 107 /* using the shared function instead of the macro saves ~3k code size */ \ 108 Upsample32Pixels(r1, r2, out); \ 109 } 110 111 #define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, uv, \ 112 top_dst, bottom_dst, cur_x, num_pixels) { \ 113 int n; \ 114 if (top_y) { \ 115 for (n = 0; n < (num_pixels); ++n) { \ 116 FUNC(top_y[(cur_x) + n], (uv)[n], (uv)[32 + n], \ 117 top_dst + ((cur_x) + n) * XSTEP); \ 118 } \ 119 } \ 120 if (bottom_y) { \ 121 for (n = 0; n < (num_pixels); ++n) { \ 122 FUNC(bottom_y[(cur_x) + n], (uv)[64 + n], (uv)[64 + 32 + n], \ 123 bottom_dst + ((cur_x) + n) * XSTEP); \ 124 } \ 125 } \ 126 } 127 128 #define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ 129 static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ 130 const uint8_t* top_u, const uint8_t* top_v, \ 131 const uint8_t* cur_u, const uint8_t* cur_v, \ 132 uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ 133 int block; \ 134 /* 16 byte aligned array to cache reconstructed u and v */ \ 135 uint8_t uv_buf[4 * 32 + 15]; \ 136 uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ 137 const int uv_len = (len + 1) >> 1; \ 138 /* 17 pixels must be read-able for each block */ \ 139 const int num_blocks = (uv_len - 1) >> 4; \ 140 const int leftover = uv_len - num_blocks * 16; \ 141 const int last_pos = 1 + 32 * num_blocks; \ 142 \ 143 const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \ 144 const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \ 145 \ 146 assert(len > 0); \ 147 /* Treat the first pixel in regular way */ \ 148 if (top_y) { \ 149 const int u0 = (top_u[0] + u_diag) >> 1; \ 150 const int v0 = (top_v[0] + v_diag) >> 1; \ 151 FUNC(top_y[0], u0, v0, top_dst); \ 152 } \ 153 if (bottom_y) { \ 154 const int u0 = (cur_u[0] + u_diag) >> 1; \ 155 const int v0 = (cur_v[0] + v_diag) >> 1; \ 156 FUNC(bottom_y[0], u0, v0, bottom_dst); \ 157 } \ 158 \ 159 for (block = 0; block < num_blocks; ++block) { \ 160 UPSAMPLE_32PIXELS(top_u, cur_u, r_uv + 0 * 32); \ 161 UPSAMPLE_32PIXELS(top_v, cur_v, r_uv + 1 * 32); \ 162 CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst, \ 163 32 * block + 1, 32) \ 164 top_u += 16; \ 165 cur_u += 16; \ 166 top_v += 16; \ 167 cur_v += 16; \ 168 } \ 169 \ 170 UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv + 0 * 32); \ 171 UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 1 * 32); \ 172 CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst, \ 173 last_pos, len - last_pos); \ 174 } 175 176 // SSE2 variants of the fancy upsampler. 177 SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePairSSE2, VP8YuvToRgb, 3) 178 SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePairSSE2, VP8YuvToBgr, 3) 179 SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePairSSE2, VP8YuvToRgba, 4) 180 SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4) 181 182 #undef GET_M 183 #undef PACK_AND_STORE 184 #undef UPSAMPLE_32PIXELS 185 #undef UPSAMPLE_LAST_BLOCK 186 #undef CONVERT2RGB 187 #undef SSE2_UPSAMPLE_FUNC 188 189 #endif // FANCY_UPSAMPLING 190 191 #endif // WEBP_USE_SSE2 192 193 //------------------------------------------------------------------------------ 194 195 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; 196 197 void WebPInitUpsamplersSSE2(void) { 198 #if defined(WEBP_USE_SSE2) 199 WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePairSSE2; 200 WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2; 201 WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePairSSE2; 202 WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2; 203 #endif // WEBP_USE_SSE2 204 } 205 206 void WebPInitPremultiplySSE2(void) { 207 #if defined(WEBP_USE_SSE2) 208 WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2; 209 WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2; 210 #endif // WEBP_USE_SSE2 211 } 212 213 #if defined(__cplusplus) || defined(c_plusplus) 214 } // extern "C" 215 #endif 216 217 218
