1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/common/vp9_convolve.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,305 @@
1.4 +/*
1.5 + * Copyright (c) 2013 The WebM project authors. All Rights Reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +#include <assert.h>
1.15 +
1.16 +#include "./vpx_config.h"
1.17 +#include "./vp9_rtcd.h"
1.18 +#include "vp9/common/vp9_common.h"
1.19 +#include "vp9/common/vp9_convolve.h"
1.20 +#include "vp9/common/vp9_filter.h"
1.21 +#include "vpx/vpx_integer.h"
1.22 +#include "vpx_ports/mem.h"
1.23 +
1.24 +static void convolve_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
1.25 + uint8_t *dst, ptrdiff_t dst_stride,
1.26 + const int16_t *filter_x0, int x_step_q4,
1.27 + const int16_t *filter_y, int y_step_q4,
1.28 + int w, int h, int taps) {
1.29 + int x, y, k;
1.30 +
1.31 + /* NOTE: This assumes that the filter table is 256-byte aligned. */
1.32 + /* TODO(agrange) Modify to make independent of table alignment. */
1.33 + const int16_t *const filter_x_base =
1.34 + (const int16_t *)(((intptr_t)filter_x0) & ~(intptr_t)0xff);
1.35 +
1.36 + /* Adjust base pointer address for this source line */
1.37 + src -= taps / 2 - 1;
1.38 +
1.39 + for (y = 0; y < h; ++y) {
1.40 + /* Initial phase offset */
1.41 + int x_q4 = (int)(filter_x0 - filter_x_base) / taps;
1.42 +
1.43 + for (x = 0; x < w; ++x) {
1.44 + /* Per-pixel src offset */
1.45 + const int src_x = x_q4 >> SUBPEL_BITS;
1.46 + int sum = 0;
1.47 +
1.48 + /* Pointer to filter to use */
1.49 + const int16_t *const filter_x = filter_x_base +
1.50 + (x_q4 & SUBPEL_MASK) * taps;
1.51 +
1.52 + for (k = 0; k < taps; ++k)
1.53 + sum += src[src_x + k] * filter_x[k];
1.54 +
1.55 + dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
1.56 +
1.57 + /* Move to the next source pixel */
1.58 + x_q4 += x_step_q4;
1.59 + }
1.60 + src += src_stride;
1.61 + dst += dst_stride;
1.62 + }
1.63 +}
1.64 +
1.65 +static void convolve_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
1.66 + uint8_t *dst, ptrdiff_t dst_stride,
1.67 + const int16_t *filter_x0, int x_step_q4,
1.68 + const int16_t *filter_y, int y_step_q4,
1.69 + int w, int h, int taps) {
1.70 + int x, y, k;
1.71 +
1.72 + /* NOTE: This assumes that the filter table is 256-byte aligned. */
1.73 + /* TODO(agrange) Modify to make independent of table alignment. */
1.74 + const int16_t *const filter_x_base =
1.75 + (const int16_t *)(((intptr_t)filter_x0) & ~(intptr_t)0xff);
1.76 +
1.77 + /* Adjust base pointer address for this source line */
1.78 + src -= taps / 2 - 1;
1.79 +
1.80 + for (y = 0; y < h; ++y) {
1.81 + /* Initial phase offset */
1.82 + int x_q4 = (int)(filter_x0 - filter_x_base) / taps;
1.83 +
1.84 + for (x = 0; x < w; ++x) {
1.85 + /* Per-pixel src offset */
1.86 + const int src_x = x_q4 >> SUBPEL_BITS;
1.87 + int sum = 0;
1.88 +
1.89 + /* Pointer to filter to use */
1.90 + const int16_t *const filter_x = filter_x_base +
1.91 + (x_q4 & SUBPEL_MASK) * taps;
1.92 +
1.93 + for (k = 0; k < taps; ++k)
1.94 + sum += src[src_x + k] * filter_x[k];
1.95 +
1.96 + dst[x] = ROUND_POWER_OF_TWO(dst[x] +
1.97 + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
1.98 +
1.99 + /* Move to the next source pixel */
1.100 + x_q4 += x_step_q4;
1.101 + }
1.102 + src += src_stride;
1.103 + dst += dst_stride;
1.104 + }
1.105 +}
1.106 +
1.107 +static void convolve_vert_c(const uint8_t *src, ptrdiff_t src_stride,
1.108 + uint8_t *dst, ptrdiff_t dst_stride,
1.109 + const int16_t *filter_x, int x_step_q4,
1.110 + const int16_t *filter_y0, int y_step_q4,
1.111 + int w, int h, int taps) {
1.112 + int x, y, k;
1.113 +
1.114 + /* NOTE: This assumes that the filter table is 256-byte aligned. */
1.115 + /* TODO(agrange) Modify to make independent of table alignment. */
1.116 + const int16_t *const filter_y_base =
1.117 + (const int16_t *)(((intptr_t)filter_y0) & ~(intptr_t)0xff);
1.118 +
1.119 + /* Adjust base pointer address for this source column */
1.120 + src -= src_stride * (taps / 2 - 1);
1.121 +
1.122 + for (x = 0; x < w; ++x) {
1.123 + /* Initial phase offset */
1.124 + int y_q4 = (int)(filter_y0 - filter_y_base) / taps;
1.125 +
1.126 + for (y = 0; y < h; ++y) {
1.127 + /* Per-pixel src offset */
1.128 + const int src_y = y_q4 >> SUBPEL_BITS;
1.129 + int sum = 0;
1.130 +
1.131 + /* Pointer to filter to use */
1.132 + const int16_t *const filter_y = filter_y_base +
1.133 + (y_q4 & SUBPEL_MASK) * taps;
1.134 +
1.135 + for (k = 0; k < taps; ++k)
1.136 + sum += src[(src_y + k) * src_stride] * filter_y[k];
1.137 +
1.138 + dst[y * dst_stride] =
1.139 + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
1.140 +
1.141 + /* Move to the next source pixel */
1.142 + y_q4 += y_step_q4;
1.143 + }
1.144 + ++src;
1.145 + ++dst;
1.146 + }
1.147 +}
1.148 +
1.149 +static void convolve_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
1.150 + uint8_t *dst, ptrdiff_t dst_stride,
1.151 + const int16_t *filter_x, int x_step_q4,
1.152 + const int16_t *filter_y0, int y_step_q4,
1.153 + int w, int h, int taps) {
1.154 + int x, y, k;
1.155 +
1.156 + /* NOTE: This assumes that the filter table is 256-byte aligned. */
1.157 + /* TODO(agrange) Modify to make independent of table alignment. */
1.158 + const int16_t *const filter_y_base =
1.159 + (const int16_t *)(((intptr_t)filter_y0) & ~(intptr_t)0xff);
1.160 +
1.161 + /* Adjust base pointer address for this source column */
1.162 + src -= src_stride * (taps / 2 - 1);
1.163 +
1.164 + for (x = 0; x < w; ++x) {
1.165 + /* Initial phase offset */
1.166 + int y_q4 = (int)(filter_y0 - filter_y_base) / taps;
1.167 +
1.168 + for (y = 0; y < h; ++y) {
1.169 + /* Per-pixel src offset */
1.170 + const int src_y = y_q4 >> SUBPEL_BITS;
1.171 + int sum = 0;
1.172 +
1.173 + /* Pointer to filter to use */
1.174 + const int16_t *const filter_y = filter_y_base +
1.175 + (y_q4 & SUBPEL_MASK) * taps;
1.176 +
1.177 + for (k = 0; k < taps; ++k)
1.178 + sum += src[(src_y + k) * src_stride] * filter_y[k];
1.179 +
1.180 + dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
1.181 + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
1.182 +
1.183 + /* Move to the next source pixel */
1.184 + y_q4 += y_step_q4;
1.185 + }
1.186 + ++src;
1.187 + ++dst;
1.188 + }
1.189 +}
1.190 +
1.191 +static void convolve_c(const uint8_t *src, ptrdiff_t src_stride,
1.192 + uint8_t *dst, ptrdiff_t dst_stride,
1.193 + const int16_t *filter_x, int x_step_q4,
1.194 + const int16_t *filter_y, int y_step_q4,
1.195 + int w, int h, int taps) {
1.196 + /* Fixed size intermediate buffer places limits on parameters.
1.197 + * Maximum intermediate_height is 324, for y_step_q4 == 80,
1.198 + * h == 64, taps == 8.
1.199 + * y_step_q4 of 80 allows for 1/10 scale for 5 layer svc
1.200 + */
1.201 + uint8_t temp[64 * 324];
1.202 + int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + taps;
1.203 +
1.204 + assert(w <= 64);
1.205 + assert(h <= 64);
1.206 + assert(taps <= 8);
1.207 + assert(y_step_q4 <= 80);
1.208 + assert(x_step_q4 <= 80);
1.209 +
1.210 + if (intermediate_height < h)
1.211 + intermediate_height = h;
1.212 +
1.213 + convolve_horiz_c(src - src_stride * (taps / 2 - 1), src_stride, temp, 64,
1.214 + filter_x, x_step_q4, filter_y, y_step_q4, w,
1.215 + intermediate_height, taps);
1.216 + convolve_vert_c(temp + 64 * (taps / 2 - 1), 64, dst, dst_stride, filter_x,
1.217 + x_step_q4, filter_y, y_step_q4, w, h, taps);
1.218 +}
1.219 +
1.220 +void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
1.221 + uint8_t *dst, ptrdiff_t dst_stride,
1.222 + const int16_t *filter_x, int x_step_q4,
1.223 + const int16_t *filter_y, int y_step_q4,
1.224 + int w, int h) {
1.225 + convolve_horiz_c(src, src_stride, dst, dst_stride,
1.226 + filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
1.227 +}
1.228 +
1.229 +void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
1.230 + uint8_t *dst, ptrdiff_t dst_stride,
1.231 + const int16_t *filter_x, int x_step_q4,
1.232 + const int16_t *filter_y, int y_step_q4,
1.233 + int w, int h) {
1.234 + convolve_avg_horiz_c(src, src_stride, dst, dst_stride,
1.235 + filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
1.236 +}
1.237 +
1.238 +void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
1.239 + uint8_t *dst, ptrdiff_t dst_stride,
1.240 + const int16_t *filter_x, int x_step_q4,
1.241 + const int16_t *filter_y, int y_step_q4,
1.242 + int w, int h) {
1.243 + convolve_vert_c(src, src_stride, dst, dst_stride,
1.244 + filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
1.245 +}
1.246 +
1.247 +void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
1.248 + uint8_t *dst, ptrdiff_t dst_stride,
1.249 + const int16_t *filter_x, int x_step_q4,
1.250 + const int16_t *filter_y, int y_step_q4,
1.251 + int w, int h) {
1.252 + convolve_avg_vert_c(src, src_stride, dst, dst_stride,
1.253 + filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
1.254 +}
1.255 +
1.256 +void vp9_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
1.257 + uint8_t *dst, ptrdiff_t dst_stride,
1.258 + const int16_t *filter_x, int x_step_q4,
1.259 + const int16_t *filter_y, int y_step_q4,
1.260 + int w, int h) {
1.261 + convolve_c(src, src_stride, dst, dst_stride,
1.262 + filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
1.263 +}
1.264 +
1.265 +void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
1.266 + uint8_t *dst, ptrdiff_t dst_stride,
1.267 + const int16_t *filter_x, int x_step_q4,
1.268 + const int16_t *filter_y, int y_step_q4,
1.269 + int w, int h) {
1.270 + /* Fixed size intermediate buffer places limits on parameters. */
1.271 + DECLARE_ALIGNED_ARRAY(16, uint8_t, temp, 64 * 64);
1.272 + assert(w <= 64);
1.273 + assert(h <= 64);
1.274 +
1.275 + vp9_convolve8(src, src_stride, temp, 64,
1.276 + filter_x, x_step_q4, filter_y, y_step_q4, w, h);
1.277 + vp9_convolve_avg(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
1.278 +}
1.279 +
1.280 +void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,
1.281 + uint8_t *dst, ptrdiff_t dst_stride,
1.282 + const int16_t *filter_x, int filter_x_stride,
1.283 + const int16_t *filter_y, int filter_y_stride,
1.284 + int w, int h) {
1.285 + int r;
1.286 +
1.287 + for (r = h; r > 0; --r) {
1.288 + vpx_memcpy(dst, src, w);
1.289 + src += src_stride;
1.290 + dst += dst_stride;
1.291 + }
1.292 +}
1.293 +
1.294 +void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
1.295 + uint8_t *dst, ptrdiff_t dst_stride,
1.296 + const int16_t *filter_x, int filter_x_stride,
1.297 + const int16_t *filter_y, int filter_y_stride,
1.298 + int w, int h) {
1.299 + int x, y;
1.300 +
1.301 + for (y = 0; y < h; ++y) {
1.302 + for (x = 0; x < w; ++x)
1.303 + dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
1.304 +
1.305 + src += src_stride;
1.306 + dst += dst_stride;
1.307 + }
1.308 +}
