upsampling_sse2.c (15689B)
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 #include <emmintrin.h> 18 19 #include <assert.h> 20 #include <string.h> 21 22 #include "src/webp/types.h" 23 #include "src/dsp/cpu.h" 24 #include "src/dsp/yuv.h" 25 #include "src/webp/decode.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 alternating 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) do { \ 66 const __m128i one = _mm_set1_epi8(1); \ 67 const __m128i a = _mm_loadu_si128((const __m128i*)&(r1)[0]); \ 68 const __m128i b = _mm_loadu_si128((const __m128i*)&(r1)[1]); \ 69 const __m128i c = _mm_loadu_si128((const __m128i*)&(r2)[0]); \ 70 const __m128i d = _mm_loadu_si128((const __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); /* store top */ \ 91 PACK_AND_STORE(c, d, diag2, diag1, (out) + 2 * 32); /* store bottom */ \ 92 } while (0) 93 94 // Turn the macro into a function for reducing code-size when non-critical 95 static void Upsample32Pixels_SSE2(const uint8_t* WEBP_RESTRICT const r1, 96 const uint8_t* WEBP_RESTRICT const r2, 97 uint8_t* WEBP_RESTRICT const out) { 98 UPSAMPLE_32PIXELS(r1, r2, out); 99 } 100 101 #define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \ 102 uint8_t r1[17], r2[17]; \ 103 memcpy(r1, (tb), (num_pixels)); \ 104 memcpy(r2, (bb), (num_pixels)); \ 105 /* replicate last byte */ \ 106 memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels)); \ 107 memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels)); \ 108 /* using the shared function instead of the macro saves ~3k code size */ \ 109 Upsample32Pixels_SSE2(r1, r2, out); \ 110 } 111 112 #define CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, \ 113 top_dst, bottom_dst, cur_x) do { \ 114 FUNC##32_SSE2((top_y) + (cur_x), r_u, r_v, (top_dst) + (cur_x) * (XSTEP)); \ 115 if ((bottom_y) != NULL) { \ 116 FUNC##32_SSE2((bottom_y) + (cur_x), r_u + 64, r_v + 64, \ 117 (bottom_dst) + (cur_x) * (XSTEP)); \ 118 } \ 119 } while (0) 120 121 #define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ 122 static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y, \ 123 const uint8_t* WEBP_RESTRICT bottom_y, \ 124 const uint8_t* WEBP_RESTRICT top_u, \ 125 const uint8_t* WEBP_RESTRICT top_v, \ 126 const uint8_t* WEBP_RESTRICT cur_u, \ 127 const uint8_t* WEBP_RESTRICT cur_v, \ 128 uint8_t* WEBP_RESTRICT top_dst, \ 129 uint8_t* WEBP_RESTRICT bottom_dst, int len) { \ 130 int uv_pos, pos; \ 131 /* 16byte-aligned array to cache reconstructed u and v */ \ 132 uint8_t uv_buf[14 * 32 + 15] = { 0 }; \ 133 uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~(uintptr_t)15); \ 134 uint8_t* const r_v = r_u + 32; \ 135 \ 136 assert(top_y != NULL); \ 137 { /* Treat the first pixel in regular way */ \ 138 const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \ 139 const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \ 140 const int u0_t = (top_u[0] + u_diag) >> 1; \ 141 const int v0_t = (top_v[0] + v_diag) >> 1; \ 142 FUNC(top_y[0], u0_t, v0_t, top_dst); \ 143 if (bottom_y != NULL) { \ 144 const int u0_b = (cur_u[0] + u_diag) >> 1; \ 145 const int v0_b = (cur_v[0] + v_diag) >> 1; \ 146 FUNC(bottom_y[0], u0_b, v0_b, bottom_dst); \ 147 } \ 148 } \ 149 /* For UPSAMPLE_32PIXELS, 17 u/v values must be read-able for each block */ \ 150 for (pos = 1, uv_pos = 0; pos + 32 + 1 <= len; pos += 32, uv_pos += 16) { \ 151 UPSAMPLE_32PIXELS(top_u + uv_pos, cur_u + uv_pos, r_u); \ 152 UPSAMPLE_32PIXELS(top_v + uv_pos, cur_v + uv_pos, r_v); \ 153 CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst, pos); \ 154 } \ 155 if (len > 1) { \ 156 const int left_over = ((len + 1) >> 1) - (pos >> 1); \ 157 uint8_t* const tmp_top_dst = r_u + 4 * 32; \ 158 uint8_t* const tmp_bottom_dst = tmp_top_dst + 4 * 32; \ 159 uint8_t* const tmp_top = tmp_bottom_dst + 4 * 32; \ 160 uint8_t* const tmp_bottom = (bottom_y == NULL) ? NULL : tmp_top + 32; \ 161 assert(left_over > 0); \ 162 UPSAMPLE_LAST_BLOCK(top_u + uv_pos, cur_u + uv_pos, left_over, r_u); \ 163 UPSAMPLE_LAST_BLOCK(top_v + uv_pos, cur_v + uv_pos, left_over, r_v); \ 164 memcpy(tmp_top, top_y + pos, len - pos); \ 165 if (bottom_y != NULL) memcpy(tmp_bottom, bottom_y + pos, len - pos); \ 166 CONVERT2RGB_32(FUNC, XSTEP, tmp_top, tmp_bottom, tmp_top_dst, \ 167 tmp_bottom_dst, 0); \ 168 memcpy(top_dst + pos * (XSTEP), tmp_top_dst, (len - pos) * (XSTEP)); \ 169 if (bottom_y != NULL) { \ 170 memcpy(bottom_dst + pos * (XSTEP), tmp_bottom_dst, \ 171 (len - pos) * (XSTEP)); \ 172 } \ 173 } \ 174 } 175 176 // SSE2 variants of the fancy upsampler. 177 SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePair_SSE2, VP8YuvToRgba, 4) 178 SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePair_SSE2, VP8YuvToBgra, 4) 179 180 #if !defined(WEBP_REDUCE_CSP) 181 SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair_SSE2, VP8YuvToRgb, 3) 182 SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair_SSE2, VP8YuvToBgr, 3) 183 SSE2_UPSAMPLE_FUNC(UpsampleArgbLinePair_SSE2, VP8YuvToArgb, 4) 184 SSE2_UPSAMPLE_FUNC(UpsampleRgba4444LinePair_SSE2, VP8YuvToRgba4444, 2) 185 SSE2_UPSAMPLE_FUNC(UpsampleRgb565LinePair_SSE2, VP8YuvToRgb565, 2) 186 #endif // WEBP_REDUCE_CSP 187 188 #undef GET_M 189 #undef PACK_AND_STORE 190 #undef UPSAMPLE_32PIXELS 191 #undef UPSAMPLE_LAST_BLOCK 192 #undef CONVERT2RGB 193 #undef CONVERT2RGB_32 194 #undef SSE2_UPSAMPLE_FUNC 195 196 //------------------------------------------------------------------------------ 197 // Entry point 198 199 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; 200 201 extern void WebPInitUpsamplersSSE2(void); 202 203 WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE2(void) { 204 WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_SSE2; 205 WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair_SSE2; 206 WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair_SSE2; 207 WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair_SSE2; 208 #if !defined(WEBP_REDUCE_CSP) 209 WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair_SSE2; 210 WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair_SSE2; 211 WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_SSE2; 212 WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair_SSE2; 213 WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair_SSE2; 214 WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_SSE2; 215 WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_SSE2; 216 #endif // WEBP_REDUCE_CSP 217 } 218 219 #endif // FANCY_UPSAMPLING 220 221 //------------------------------------------------------------------------------ 222 223 extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */]; 224 extern void WebPInitYUV444ConvertersSSE2(void); 225 226 #define YUV444_FUNC(FUNC_NAME, CALL, CALL_C, XSTEP) \ 227 extern void CALL_C(const uint8_t* WEBP_RESTRICT y, \ 228 const uint8_t* WEBP_RESTRICT u, \ 229 const uint8_t* WEBP_RESTRICT v, \ 230 uint8_t* WEBP_RESTRICT dst, int len); \ 231 static void FUNC_NAME(const uint8_t* WEBP_RESTRICT y, \ 232 const uint8_t* WEBP_RESTRICT u, \ 233 const uint8_t* WEBP_RESTRICT v, \ 234 uint8_t* WEBP_RESTRICT dst, int len) { \ 235 int i; \ 236 const int max_len = len & ~31; \ 237 for (i = 0; i < max_len; i += 32) { \ 238 CALL(y + i, u + i, v + i, dst + i * (XSTEP)); \ 239 } \ 240 if (i < len) { /* C-fallback */ \ 241 CALL_C(y + i, u + i, v + i, dst + i * (XSTEP), len - i); \ 242 } \ 243 } 244 245 YUV444_FUNC(Yuv444ToRgba_SSE2, VP8YuvToRgba32_SSE2, WebPYuv444ToRgba_C, 4) 246 YUV444_FUNC(Yuv444ToBgra_SSE2, VP8YuvToBgra32_SSE2, WebPYuv444ToBgra_C, 4) 247 #if !defined(WEBP_REDUCE_CSP) 248 YUV444_FUNC(Yuv444ToRgb_SSE2, VP8YuvToRgb32_SSE2, WebPYuv444ToRgb_C, 3) 249 YUV444_FUNC(Yuv444ToBgr_SSE2, VP8YuvToBgr32_SSE2, WebPYuv444ToBgr_C, 3) 250 YUV444_FUNC(Yuv444ToArgb_SSE2, VP8YuvToArgb32_SSE2, WebPYuv444ToArgb_C, 4) 251 YUV444_FUNC(Yuv444ToRgba4444_SSE2, VP8YuvToRgba444432_SSE2, \ 252 WebPYuv444ToRgba4444_C, 2) 253 YUV444_FUNC(Yuv444ToRgb565_SSE2, VP8YuvToRgb56532_SSE2, WebPYuv444ToRgb565_C, 2) 254 #endif // WEBP_REDUCE_CSP 255 256 WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE2(void) { 257 WebPYUV444Converters[MODE_RGBA] = Yuv444ToRgba_SSE2; 258 WebPYUV444Converters[MODE_BGRA] = Yuv444ToBgra_SSE2; 259 WebPYUV444Converters[MODE_rgbA] = Yuv444ToRgba_SSE2; 260 WebPYUV444Converters[MODE_bgrA] = Yuv444ToBgra_SSE2; 261 #if !defined(WEBP_REDUCE_CSP) 262 WebPYUV444Converters[MODE_RGB] = Yuv444ToRgb_SSE2; 263 WebPYUV444Converters[MODE_BGR] = Yuv444ToBgr_SSE2; 264 WebPYUV444Converters[MODE_ARGB] = Yuv444ToArgb_SSE2; 265 WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444_SSE2; 266 WebPYUV444Converters[MODE_RGB_565] = Yuv444ToRgb565_SSE2; 267 WebPYUV444Converters[MODE_Argb] = Yuv444ToArgb_SSE2; 268 WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444_SSE2; 269 #endif // WEBP_REDUCE_CSP 270 } 271 272 #else 273 274 WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersSSE2) 275 276 #endif // WEBP_USE_SSE2 277 278 #if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_SSE2)) 279 WEBP_DSP_INIT_STUB(WebPInitUpsamplersSSE2) 280 #endif