1 /* 2 * Copyright 2015 The LibYuv Project Authors. All rights reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include "libyuv/scale.h" 12 #include "libyuv/scale_row.h" 13 14 #include "libyuv/basic_types.h" 15 16 #ifdef __cplusplus 17 namespace libyuv { 18 extern "C" { 19 #endif 20 21 // Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols 22 #define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK) \ 23 void NAMEANY(uint8* dst_ptr, const uint8* src_ptr, \ 24 int dst_width, int x, int dx) { \ 25 int n = dst_width & ~MASK; \ 26 if (n > 0) { \ 27 TERP_SIMD(dst_ptr, src_ptr, n, x, dx); \ 28 } \ 29 TERP_C(dst_ptr + n * BPP, src_ptr, \ 30 dst_width & MASK, x + n * dx, dx); \ 31 } 32 33 #ifdef HAS_SCALEFILTERCOLS_NEON 34 CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7) 35 #endif 36 #ifdef HAS_SCALEARGBCOLS_NEON 37 CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7) 38 #endif 39 #ifdef HAS_SCALEARGBFILTERCOLS_NEON 40 CANY(ScaleARGBFilterCols_Any_NEON, ScaleARGBFilterCols_NEON, 41 ScaleARGBFilterCols_C, 4, 3) 42 #endif 43 #undef CANY 44 45 // Fixed scale down. 46 #define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK) \ 47 void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, \ 48 uint8* dst_ptr, int dst_width) { \ 49 int r = (int)((unsigned int)dst_width % (MASK + 1)); \ 50 int n = dst_width - r; \ 51 if (n > 0) { \ 52 SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n); \ 53 } \ 54 SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride, \ 55 dst_ptr + n * BPP, r); \ 56 } 57 58 // Fixed scale down for odd source width. Used by I420Blend subsampling. 59 // Since dst_width is (width + 1) / 2, this function scales one less pixel 60 // and copies the last pixel. 61 #define SDODD(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK) \ 62 void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, \ 63 uint8* dst_ptr, int dst_width) { \ 64 int r = (int)((unsigned int)(dst_width - 1) % (MASK + 1)); \ 65 int n = dst_width - r; \ 66 if (n > 0) { \ 67 SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n); \ 68 } \ 69 SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride, \ 70 dst_ptr + n * BPP, r); \ 71 } 72 73 #ifdef HAS_SCALEROWDOWN2_SSSE3 74 SDANY(ScaleRowDown2_Any_SSSE3, ScaleRowDown2_SSSE3, ScaleRowDown2_C, 2, 1, 15) 75 SDANY(ScaleRowDown2Linear_Any_SSSE3, ScaleRowDown2Linear_SSSE3, 76 ScaleRowDown2Linear_C, 2, 1, 15) 77 SDANY(ScaleRowDown2Box_Any_SSSE3, ScaleRowDown2Box_SSSE3, ScaleRowDown2Box_C, 78 2, 1, 15) 79 SDODD(ScaleRowDown2Box_Odd_SSSE3, ScaleRowDown2Box_SSSE3, 80 ScaleRowDown2Box_Odd_C, 2, 1, 15) 81 #endif 82 #ifdef HAS_SCALEROWDOWN2_AVX2 83 SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31) 84 SDANY(ScaleRowDown2Linear_Any_AVX2, ScaleRowDown2Linear_AVX2, 85 ScaleRowDown2Linear_C, 2, 1, 31) 86 SDANY(ScaleRowDown2Box_Any_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_C, 87 2, 1, 31) 88 SDODD(ScaleRowDown2Box_Odd_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_Odd_C, 89 2, 1, 31) 90 #endif 91 #ifdef HAS_SCALEROWDOWN2_NEON 92 SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15) 93 SDANY(ScaleRowDown2Linear_Any_NEON, ScaleRowDown2Linear_NEON, 94 ScaleRowDown2Linear_C, 2, 1, 15) 95 SDANY(ScaleRowDown2Box_Any_NEON, ScaleRowDown2Box_NEON, 96 ScaleRowDown2Box_C, 2, 1, 15) 97 SDODD(ScaleRowDown2Box_Odd_NEON, ScaleRowDown2Box_NEON, 98 ScaleRowDown2Box_Odd_C, 2, 1, 15) 99 #endif 100 #ifdef HAS_SCALEROWDOWN4_SSSE3 101 SDANY(ScaleRowDown4_Any_SSSE3, ScaleRowDown4_SSSE3, ScaleRowDown4_C, 4, 1, 7) 102 SDANY(ScaleRowDown4Box_Any_SSSE3, ScaleRowDown4Box_SSSE3, ScaleRowDown4Box_C, 103 4, 1, 7) 104 #endif 105 #ifdef HAS_SCALEROWDOWN4_AVX2 106 SDANY(ScaleRowDown4_Any_AVX2, ScaleRowDown4_AVX2, ScaleRowDown4_C, 4, 1, 15) 107 SDANY(ScaleRowDown4Box_Any_AVX2, ScaleRowDown4Box_AVX2, ScaleRowDown4Box_C, 108 4, 1, 15) 109 #endif 110 #ifdef HAS_SCALEROWDOWN4_NEON 111 SDANY(ScaleRowDown4_Any_NEON, ScaleRowDown4_NEON, ScaleRowDown4_C, 4, 1, 7) 112 SDANY(ScaleRowDown4Box_Any_NEON, ScaleRowDown4Box_NEON, ScaleRowDown4Box_C, 113 4, 1, 7) 114 #endif 115 #ifdef HAS_SCALEROWDOWN34_SSSE3 116 SDANY(ScaleRowDown34_Any_SSSE3, ScaleRowDown34_SSSE3, 117 ScaleRowDown34_C, 4 / 3, 1, 23) 118 SDANY(ScaleRowDown34_0_Box_Any_SSSE3, ScaleRowDown34_0_Box_SSSE3, 119 ScaleRowDown34_0_Box_C, 4 / 3, 1, 23) 120 SDANY(ScaleRowDown34_1_Box_Any_SSSE3, ScaleRowDown34_1_Box_SSSE3, 121 ScaleRowDown34_1_Box_C, 4 / 3, 1, 23) 122 #endif 123 #ifdef HAS_SCALEROWDOWN34_NEON 124 SDANY(ScaleRowDown34_Any_NEON, ScaleRowDown34_NEON, 125 ScaleRowDown34_C, 4 / 3, 1, 23) 126 SDANY(ScaleRowDown34_0_Box_Any_NEON, ScaleRowDown34_0_Box_NEON, 127 ScaleRowDown34_0_Box_C, 4 / 3, 1, 23) 128 SDANY(ScaleRowDown34_1_Box_Any_NEON, ScaleRowDown34_1_Box_NEON, 129 ScaleRowDown34_1_Box_C, 4 / 3, 1, 23) 130 #endif 131 #ifdef HAS_SCALEROWDOWN38_SSSE3 132 SDANY(ScaleRowDown38_Any_SSSE3, ScaleRowDown38_SSSE3, 133 ScaleRowDown38_C, 8 / 3, 1, 11) 134 SDANY(ScaleRowDown38_3_Box_Any_SSSE3, ScaleRowDown38_3_Box_SSSE3, 135 ScaleRowDown38_3_Box_C, 8 / 3, 1, 5) 136 SDANY(ScaleRowDown38_2_Box_Any_SSSE3, ScaleRowDown38_2_Box_SSSE3, 137 ScaleRowDown38_2_Box_C, 8 / 3, 1, 5) 138 #endif 139 #ifdef HAS_SCALEROWDOWN38_NEON 140 SDANY(ScaleRowDown38_Any_NEON, ScaleRowDown38_NEON, 141 ScaleRowDown38_C, 8 / 3, 1, 11) 142 SDANY(ScaleRowDown38_3_Box_Any_NEON, ScaleRowDown38_3_Box_NEON, 143 ScaleRowDown38_3_Box_C, 8 / 3, 1, 11) 144 SDANY(ScaleRowDown38_2_Box_Any_NEON, ScaleRowDown38_2_Box_NEON, 145 ScaleRowDown38_2_Box_C, 8 / 3, 1, 11) 146 #endif 147 148 #ifdef HAS_SCALEARGBROWDOWN2_SSE2 149 SDANY(ScaleARGBRowDown2_Any_SSE2, ScaleARGBRowDown2_SSE2, 150 ScaleARGBRowDown2_C, 2, 4, 3) 151 SDANY(ScaleARGBRowDown2Linear_Any_SSE2, ScaleARGBRowDown2Linear_SSE2, 152 ScaleARGBRowDown2Linear_C, 2, 4, 3) 153 SDANY(ScaleARGBRowDown2Box_Any_SSE2, ScaleARGBRowDown2Box_SSE2, 154 ScaleARGBRowDown2Box_C, 2, 4, 3) 155 #endif 156 #ifdef HAS_SCALEARGBROWDOWN2_NEON 157 SDANY(ScaleARGBRowDown2_Any_NEON, ScaleARGBRowDown2_NEON, 158 ScaleARGBRowDown2_C, 2, 4, 7) 159 SDANY(ScaleARGBRowDown2Linear_Any_NEON, ScaleARGBRowDown2Linear_NEON, 160 ScaleARGBRowDown2Linear_C, 2, 4, 7) 161 SDANY(ScaleARGBRowDown2Box_Any_NEON, ScaleARGBRowDown2Box_NEON, 162 ScaleARGBRowDown2Box_C, 2, 4, 7) 163 #endif 164 #undef SDANY 165 166 // Scale down by even scale factor. 167 #define SDAANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, BPP, MASK) \ 168 void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, int src_stepx, \ 169 uint8* dst_ptr, int dst_width) { \ 170 int r = (int)((unsigned int)dst_width % (MASK + 1)); \ 171 int n = dst_width - r; \ 172 if (n > 0) { \ 173 SCALEROWDOWN_SIMD(src_ptr, src_stride, src_stepx, dst_ptr, n); \ 174 } \ 175 SCALEROWDOWN_C(src_ptr + (n * src_stepx) * BPP, src_stride, \ 176 src_stepx, dst_ptr + n * BPP, r); \ 177 } 178 179 #ifdef HAS_SCALEARGBROWDOWNEVEN_SSE2 180 SDAANY(ScaleARGBRowDownEven_Any_SSE2, ScaleARGBRowDownEven_SSE2, 181 ScaleARGBRowDownEven_C, 4, 3) 182 SDAANY(ScaleARGBRowDownEvenBox_Any_SSE2, ScaleARGBRowDownEvenBox_SSE2, 183 ScaleARGBRowDownEvenBox_C, 4, 3) 184 #endif 185 #ifdef HAS_SCALEARGBROWDOWNEVEN_NEON 186 SDAANY(ScaleARGBRowDownEven_Any_NEON, ScaleARGBRowDownEven_NEON, 187 ScaleARGBRowDownEven_C, 4, 3) 188 SDAANY(ScaleARGBRowDownEvenBox_Any_NEON, ScaleARGBRowDownEvenBox_NEON, 189 ScaleARGBRowDownEvenBox_C, 4, 3) 190 #endif 191 192 // Add rows box filter scale down. 193 #define SAANY(NAMEANY, SCALEADDROW_SIMD, SCALEADDROW_C, MASK) \ 194 void NAMEANY(const uint8* src_ptr, uint16* dst_ptr, int src_width) { \ 195 int n = src_width & ~MASK; \ 196 if (n > 0) { \ 197 SCALEADDROW_SIMD(src_ptr, dst_ptr, n); \ 198 } \ 199 SCALEADDROW_C(src_ptr + n, dst_ptr + n, src_width & MASK); \ 200 } 201 202 #ifdef HAS_SCALEADDROW_SSE2 203 SAANY(ScaleAddRow_Any_SSE2, ScaleAddRow_SSE2, ScaleAddRow_C, 15) 204 #endif 205 #ifdef HAS_SCALEADDROW_AVX2 206 SAANY(ScaleAddRow_Any_AVX2, ScaleAddRow_AVX2, ScaleAddRow_C, 31) 207 #endif 208 #ifdef HAS_SCALEADDROW_NEON 209 SAANY(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, ScaleAddRow_C, 15) 210 #endif 211 #undef SAANY 212 213 #ifdef __cplusplus 214 } // extern "C" 215 } // namespace libyuv 216 #endif 217 218 219 220 221 222