aom_convolve.c (10044B)
1 /* 2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include <assert.h> 13 #include <string.h> 14 15 #include "config/aom_config.h" 16 #include "config/aom_dsp_rtcd.h" 17 18 #include "aom/aom_integer.h" 19 #include "aom_dsp/aom_dsp_common.h" 20 #include "aom_dsp/aom_filter.h" 21 #include "aom_ports/mem.h" 22 23 static inline int horz_scalar_product(const uint8_t *a, const int16_t *b) { 24 int sum = 0; 25 for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k]; 26 return sum; 27 } 28 29 static inline int vert_scalar_product(const uint8_t *a, ptrdiff_t a_stride, 30 const int16_t *b) { 31 int sum = 0; 32 for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k]; 33 return sum; 34 } 35 36 static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride, 37 uint8_t *dst, ptrdiff_t dst_stride, 38 const InterpKernel *x_filters, int x0_q4, 39 int x_step_q4, int w, int h) { 40 src -= SUBPEL_TAPS / 2 - 1; 41 for (int y = 0; y < h; ++y) { 42 int x_q4 = x0_q4; 43 for (int x = 0; x < w; ++x) { 44 const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; 45 const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; 46 const int sum = horz_scalar_product(src_x, x_filter); 47 dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)); 48 x_q4 += x_step_q4; 49 } 50 src += src_stride; 51 dst += dst_stride; 52 } 53 } 54 55 static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride, 56 uint8_t *dst, ptrdiff_t dst_stride, 57 const InterpKernel *y_filters, int y0_q4, 58 int y_step_q4, int w, int h) { 59 src -= src_stride * (SUBPEL_TAPS / 2 - 1); 60 61 for (int x = 0; x < w; ++x) { 62 int y_q4 = y0_q4; 63 for (int y = 0; y < h; ++y) { 64 const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; 65 const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; 66 const int sum = vert_scalar_product(src_y, src_stride, y_filter); 67 dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)); 68 y_q4 += y_step_q4; 69 } 70 ++src; 71 ++dst; 72 } 73 } 74 75 static const InterpKernel *get_filter_base(const int16_t *filter) { 76 // NOTE: This assumes that the filter table is 256-byte aligned. 77 return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF)); 78 } 79 80 static int get_filter_offset(const int16_t *f, const InterpKernel *base) { 81 return (int)((const InterpKernel *)(intptr_t)f - base); 82 } 83 84 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, 85 uint8_t *dst, ptrdiff_t dst_stride, 86 const int16_t *filter_x, int x_step_q4, 87 const int16_t *filter_y, int y_step_q4, int w, 88 int h) { 89 const InterpKernel *const filters_x = get_filter_base(filter_x); 90 const int x0_q4 = get_filter_offset(filter_x, filters_x); 91 92 (void)filter_y; 93 (void)y_step_q4; 94 95 convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4, 96 w, h); 97 } 98 99 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, 100 uint8_t *dst, ptrdiff_t dst_stride, 101 const int16_t *filter_x, int x_step_q4, 102 const int16_t *filter_y, int y_step_q4, int w, 103 int h) { 104 const InterpKernel *const filters_y = get_filter_base(filter_y); 105 const int y0_q4 = get_filter_offset(filter_y, filters_y); 106 107 (void)filter_x; 108 (void)x_step_q4; 109 110 convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4, 111 w, h); 112 } 113 114 void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, 115 ptrdiff_t dst_stride, const InterpKernel *filter, 116 int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, 117 int h) { 118 // Note: Fixed size intermediate buffer, temp, places limits on parameters. 119 // 2d filtering proceeds in 2 steps: 120 // (1) Interpolate horizontally into an intermediate buffer, temp. 121 // (2) Interpolate temp vertically to derive the sub-pixel result. 122 // Deriving the maximum number of rows in the temp buffer (135): 123 // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative). 124 // --Largest block size is 64x64 pixels. 125 // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the 126 // original frame (in 1/16th pixel units). 127 // --Must round-up because block may be located at sub-pixel position. 128 // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails. 129 // --((64 - 1) * 32 + 15) >> 4 + 8 = 135. 130 // When calling in frame scaling function, the smallest scaling factor is x1/4 131 // ==> y_step_q4 = 64. Since w and h are at most 16, the temp buffer is still 132 // big enough. 133 uint8_t temp[64 * 135]; 134 const int intermediate_height = 135 (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS; 136 137 assert(w <= 64); 138 assert(h <= 64); 139 assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32)); 140 assert(x_step_q4 <= 64); 141 142 convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64, 143 filter, x0_q4, x_step_q4, w, intermediate_height); 144 convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride, filter, 145 y0_q4, y_step_q4, w, h); 146 } 147 148 void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, 149 ptrdiff_t dst_stride, int w, int h) { 150 for (int r = h; r > 0; --r) { 151 memmove(dst, src, w); 152 src += src_stride; 153 dst += dst_stride; 154 } 155 } 156 157 #if CONFIG_AV1_HIGHBITDEPTH 158 static inline int highbd_vert_scalar_product(const uint16_t *a, 159 ptrdiff_t a_stride, 160 const int16_t *b) { 161 int sum = 0; 162 for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k]; 163 return sum; 164 } 165 166 static inline int highbd_horz_scalar_product(const uint16_t *a, 167 const int16_t *b) { 168 int sum = 0; 169 for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k]; 170 return sum; 171 } 172 173 static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride, 174 uint8_t *dst8, ptrdiff_t dst_stride, 175 const InterpKernel *x_filters, int x0_q4, 176 int x_step_q4, int w, int h, int bd) { 177 uint16_t *src = CONVERT_TO_SHORTPTR(src8); 178 uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); 179 src -= SUBPEL_TAPS / 2 - 1; 180 for (int y = 0; y < h; ++y) { 181 int x_q4 = x0_q4; 182 for (int x = 0; x < w; ++x) { 183 const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; 184 const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; 185 const int sum = highbd_horz_scalar_product(src_x, x_filter); 186 dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd); 187 x_q4 += x_step_q4; 188 } 189 src += src_stride; 190 dst += dst_stride; 191 } 192 } 193 194 static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride, 195 uint8_t *dst8, ptrdiff_t dst_stride, 196 const InterpKernel *y_filters, int y0_q4, 197 int y_step_q4, int w, int h, int bd) { 198 uint16_t *src = CONVERT_TO_SHORTPTR(src8); 199 uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); 200 src -= src_stride * (SUBPEL_TAPS / 2 - 1); 201 for (int x = 0; x < w; ++x) { 202 int y_q4 = y0_q4; 203 for (int y = 0; y < h; ++y) { 204 const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; 205 const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; 206 const int sum = highbd_vert_scalar_product(src_y, src_stride, y_filter); 207 dst[y * dst_stride] = 208 clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd); 209 y_q4 += y_step_q4; 210 } 211 ++src; 212 ++dst; 213 } 214 } 215 216 void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, 217 uint8_t *dst, ptrdiff_t dst_stride, 218 const int16_t *filter_x, int x_step_q4, 219 const int16_t *filter_y, int y_step_q4, int w, 220 int h, int bd) { 221 const InterpKernel *const filters_x = get_filter_base(filter_x); 222 const int x0_q4 = get_filter_offset(filter_x, filters_x); 223 (void)filter_y; 224 (void)y_step_q4; 225 226 highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, 227 x_step_q4, w, h, bd); 228 } 229 230 void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, 231 uint8_t *dst, ptrdiff_t dst_stride, 232 const int16_t *filter_x, int x_step_q4, 233 const int16_t *filter_y, int y_step_q4, int w, 234 int h, int bd) { 235 const InterpKernel *const filters_y = get_filter_base(filter_y); 236 const int y0_q4 = get_filter_offset(filter_y, filters_y); 237 (void)filter_x; 238 (void)x_step_q4; 239 240 highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, 241 y_step_q4, w, h, bd); 242 } 243 244 void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, 245 uint16_t *dst, ptrdiff_t dst_stride, int w, 246 int h) { 247 for (int y = 0; y < h; ++y) { 248 memmove(dst, src, w * sizeof(src[0])); 249 src += src_stride; 250 dst += dst_stride; 251 } 252 } 253 #endif // CONFIG_AV1_HIGHBITDEPTH