1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/encoder/vp9_quantize.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,348 @@
1.4 +/*
1.5 + * Copyright (c) 2010 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 <math.h>
1.15 +#include "vpx_mem/vpx_mem.h"
1.16 +
1.17 +#include "vp9/encoder/vp9_onyx_int.h"
1.18 +#include "vp9/encoder/vp9_rdopt.h"
1.19 +#include "vp9/encoder/vp9_quantize.h"
1.20 +#include "vp9/common/vp9_quant_common.h"
1.21 +
1.22 +#include "vp9/common/vp9_seg_common.h"
1.23 +
1.24 +#ifdef ENC_DEBUG
1.25 +extern int enc_debug;
1.26 +#endif
1.27 +
1.28 +void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t count,
1.29 + int skip_block,
1.30 + const int16_t *zbin_ptr, const int16_t *round_ptr,
1.31 + const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
1.32 + int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
1.33 + const int16_t *dequant_ptr,
1.34 + int zbin_oq_value, uint16_t *eob_ptr,
1.35 + const int16_t *scan, const int16_t *iscan) {
1.36 + int i, non_zero_count = count, eob = -1;
1.37 + const int zbins[2] = { zbin_ptr[0] + zbin_oq_value,
1.38 + zbin_ptr[1] + zbin_oq_value };
1.39 + const int nzbins[2] = { zbins[0] * -1,
1.40 + zbins[1] * -1 };
1.41 +
1.42 + vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
1.43 + vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
1.44 +
1.45 + if (!skip_block) {
1.46 + // Pre-scan pass
1.47 + for (i = count - 1; i >= 0; i--) {
1.48 + const int rc = scan[i];
1.49 + const int coeff = coeff_ptr[rc];
1.50 +
1.51 + if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
1.52 + non_zero_count--;
1.53 + else
1.54 + break;
1.55 + }
1.56 +
1.57 + // Quantization pass: All coefficients with index >= zero_flag are
1.58 + // skippable. Note: zero_flag can be zero.
1.59 + for (i = 0; i < non_zero_count; i++) {
1.60 + const int rc = scan[i];
1.61 + const int coeff = coeff_ptr[rc];
1.62 + const int coeff_sign = (coeff >> 31);
1.63 + const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
1.64 +
1.65 + if (abs_coeff >= zbins[rc != 0]) {
1.66 + int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
1.67 + tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
1.68 + quant_shift_ptr[rc != 0]) >> 16; // quantization
1.69 + qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
1.70 + dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
1.71 +
1.72 + if (tmp)
1.73 + eob = i;
1.74 + }
1.75 + }
1.76 + }
1.77 + *eob_ptr = eob + 1;
1.78 +}
1.79 +
1.80 +void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs,
1.81 + int skip_block,
1.82 + const int16_t *zbin_ptr, const int16_t *round_ptr,
1.83 + const int16_t *quant_ptr,
1.84 + const int16_t *quant_shift_ptr,
1.85 + int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
1.86 + const int16_t *dequant_ptr,
1.87 + int zbin_oq_value, uint16_t *eob_ptr,
1.88 + const int16_t *scan, const int16_t *iscan) {
1.89 + int i, rc, eob;
1.90 + int zbins[2], nzbins[2];
1.91 + int x, y, z, sz;
1.92 + int idx = 0;
1.93 + int idx_arr[1024];
1.94 +
1.95 + vpx_memset(qcoeff_ptr, 0, n_coeffs*sizeof(int16_t));
1.96 + vpx_memset(dqcoeff_ptr, 0, n_coeffs*sizeof(int16_t));
1.97 +
1.98 + eob = -1;
1.99 +
1.100 + // Base ZBIN
1.101 + zbins[0] = ROUND_POWER_OF_TWO(zbin_ptr[0] + zbin_oq_value, 1);
1.102 + zbins[1] = ROUND_POWER_OF_TWO(zbin_ptr[1] + zbin_oq_value, 1);
1.103 + nzbins[0] = zbins[0] * -1;
1.104 + nzbins[1] = zbins[1] * -1;
1.105 +
1.106 + if (!skip_block) {
1.107 + // Pre-scan pass
1.108 + for (i = 0; i < n_coeffs; i++) {
1.109 + rc = scan[i];
1.110 + z = coeff_ptr[rc];
1.111 +
1.112 + // If the coefficient is out of the base ZBIN range, keep it for
1.113 + // quantization.
1.114 + if (z >= zbins[rc != 0] || z <= nzbins[rc != 0])
1.115 + idx_arr[idx++] = i;
1.116 + }
1.117 +
1.118 + // Quantization pass: only process the coefficients selected in
1.119 + // pre-scan pass. Note: idx can be zero.
1.120 + for (i = 0; i < idx; i++) {
1.121 + rc = scan[idx_arr[i]];
1.122 +
1.123 + z = coeff_ptr[rc];
1.124 + sz = (z >> 31); // sign of z
1.125 + x = (z ^ sz) - sz; // x = abs(z)
1.126 +
1.127 + x += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
1.128 + x = clamp(x, INT16_MIN, INT16_MAX);
1.129 + y = ((((x * quant_ptr[rc != 0]) >> 16) + x) *
1.130 + quant_shift_ptr[rc != 0]) >> 15; // quantize (x)
1.131 +
1.132 + x = (y ^ sz) - sz; // get the sign back
1.133 + qcoeff_ptr[rc] = x; // write to destination
1.134 + dqcoeff_ptr[rc] = x * dequant_ptr[rc != 0] / 2; // dequantized value
1.135 +
1.136 + if (y)
1.137 + eob = idx_arr[i]; // last nonzero coeffs
1.138 + }
1.139 + }
1.140 + *eob_ptr = eob + 1;
1.141 +}
1.142 +
1.143 +struct plane_block_idx {
1.144 + int plane;
1.145 + int block;
1.146 +};
1.147 +
1.148 +// TODO(jkoleszar): returning a struct so it can be used in a const context,
1.149 +// expect to refactor this further later.
1.150 +static INLINE struct plane_block_idx plane_block_idx(int y_blocks,
1.151 + int b_idx) {
1.152 + const int v_offset = y_blocks * 5 / 4;
1.153 + struct plane_block_idx res;
1.154 +
1.155 + if (b_idx < y_blocks) {
1.156 + res.plane = 0;
1.157 + res.block = b_idx;
1.158 + } else if (b_idx < v_offset) {
1.159 + res.plane = 1;
1.160 + res.block = b_idx - y_blocks;
1.161 + } else {
1.162 + assert(b_idx < y_blocks * 3 / 2);
1.163 + res.plane = 2;
1.164 + res.block = b_idx - v_offset;
1.165 + }
1.166 + return res;
1.167 +}
1.168 +
1.169 +void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int y_blocks, int b_idx,
1.170 + const int16_t *scan, const int16_t *iscan) {
1.171 + MACROBLOCKD *const xd = &x->e_mbd;
1.172 + const struct plane_block_idx pb_idx = plane_block_idx(y_blocks, b_idx);
1.173 + struct macroblock_plane* p = &x->plane[pb_idx.plane];
1.174 + struct macroblockd_plane* pd = &xd->plane[pb_idx.plane];
1.175 +
1.176 + vp9_quantize_b(BLOCK_OFFSET(p->coeff, pb_idx.block),
1.177 + 16, x->skip_block,
1.178 + p->zbin, p->round, p->quant, p->quant_shift,
1.179 + BLOCK_OFFSET(pd->qcoeff, pb_idx.block),
1.180 + BLOCK_OFFSET(pd->dqcoeff, pb_idx.block),
1.181 + pd->dequant, p->zbin_extra, &pd->eobs[pb_idx.block], scan, iscan);
1.182 +}
1.183 +
1.184 +static void invert_quant(int16_t *quant, int16_t *shift, int d) {
1.185 + unsigned t;
1.186 + int l;
1.187 + t = d;
1.188 + for (l = 0; t > 1; l++)
1.189 + t >>= 1;
1.190 + t = 1 + (1 << (16 + l)) / d;
1.191 + *quant = (int16_t)(t - (1 << 16));
1.192 + *shift = 1 << (16 - l);
1.193 +}
1.194 +
1.195 +void vp9_init_quantizer(VP9_COMP *cpi) {
1.196 + int i, q;
1.197 + VP9_COMMON *const cm = &cpi->common;
1.198 +
1.199 + for (q = 0; q < QINDEX_RANGE; q++) {
1.200 + const int qzbin_factor = q == 0 ? 64 : (vp9_dc_quant(q, 0) < 148 ? 84 : 80);
1.201 + const int qrounding_factor = q == 0 ? 64 : 48;
1.202 +
1.203 + // y
1.204 + for (i = 0; i < 2; ++i) {
1.205 + const int quant = i == 0 ? vp9_dc_quant(q, cm->y_dc_delta_q)
1.206 + : vp9_ac_quant(q, 0);
1.207 + invert_quant(&cpi->y_quant[q][i], &cpi->y_quant_shift[q][i], quant);
1.208 + cpi->y_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
1.209 + cpi->y_round[q][i] = (qrounding_factor * quant) >> 7;
1.210 + cm->y_dequant[q][i] = quant;
1.211 + }
1.212 +
1.213 + // uv
1.214 + for (i = 0; i < 2; ++i) {
1.215 + const int quant = i == 0 ? vp9_dc_quant(q, cm->uv_dc_delta_q)
1.216 + : vp9_ac_quant(q, cm->uv_ac_delta_q);
1.217 + invert_quant(&cpi->uv_quant[q][i], &cpi->uv_quant_shift[q][i], quant);
1.218 + cpi->uv_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
1.219 + cpi->uv_round[q][i] = (qrounding_factor * quant) >> 7;
1.220 + cm->uv_dequant[q][i] = quant;
1.221 + }
1.222 +
1.223 +#if CONFIG_ALPHA
1.224 + // alpha
1.225 + for (i = 0; i < 2; ++i) {
1.226 + const int quant = i == 0 ? vp9_dc_quant(q, cm->a_dc_delta_q)
1.227 + : vp9_ac_quant(q, cm->a_ac_delta_q);
1.228 + invert_quant(&cpi->a_quant[q][i], &cpi->a_quant_shift[q][i], quant);
1.229 + cpi->a_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
1.230 + cpi->a_round[q][i] = (qrounding_factor * quant) >> 7;
1.231 + cm->a_dequant[q][i] = quant;
1.232 + }
1.233 +#endif
1.234 +
1.235 + for (i = 2; i < 8; i++) {
1.236 + cpi->y_quant[q][i] = cpi->y_quant[q][1];
1.237 + cpi->y_quant_shift[q][i] = cpi->y_quant_shift[q][1];
1.238 + cpi->y_zbin[q][i] = cpi->y_zbin[q][1];
1.239 + cpi->y_round[q][i] = cpi->y_round[q][1];
1.240 + cm->y_dequant[q][i] = cm->y_dequant[q][1];
1.241 +
1.242 + cpi->uv_quant[q][i] = cpi->uv_quant[q][1];
1.243 + cpi->uv_quant_shift[q][i] = cpi->uv_quant_shift[q][1];
1.244 + cpi->uv_zbin[q][i] = cpi->uv_zbin[q][1];
1.245 + cpi->uv_round[q][i] = cpi->uv_round[q][1];
1.246 + cm->uv_dequant[q][i] = cm->uv_dequant[q][1];
1.247 +
1.248 +#if CONFIG_ALPHA
1.249 + cpi->a_quant[q][i] = cpi->a_quant[q][1];
1.250 + cpi->a_quant_shift[q][i] = cpi->a_quant_shift[q][1];
1.251 + cpi->a_zbin[q][i] = cpi->a_zbin[q][1];
1.252 + cpi->a_round[q][i] = cpi->a_round[q][1];
1.253 + cm->a_dequant[q][i] = cm->a_dequant[q][1];
1.254 +#endif
1.255 + }
1.256 + }
1.257 +}
1.258 +
1.259 +void vp9_mb_init_quantizer(VP9_COMP *cpi, MACROBLOCK *x) {
1.260 + int i;
1.261 + VP9_COMMON *const cm = &cpi->common;
1.262 + MACROBLOCKD *xd = &x->e_mbd;
1.263 + int zbin_extra;
1.264 + int segment_id = xd->mi_8x8[0]->mbmi.segment_id;
1.265 + const int qindex = vp9_get_qindex(&cpi->common.seg, segment_id,
1.266 + cpi->common.base_qindex);
1.267 +
1.268 + int rdmult = vp9_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q);
1.269 +
1.270 + // Y
1.271 + zbin_extra = (cpi->common.y_dequant[qindex][1] *
1.272 + (cpi->zbin_mode_boost + x->act_zbin_adj)) >> 7;
1.273 +
1.274 + x->plane[0].quant = cpi->y_quant[qindex];
1.275 + x->plane[0].quant_shift = cpi->y_quant_shift[qindex];
1.276 + x->plane[0].zbin = cpi->y_zbin[qindex];
1.277 + x->plane[0].round = cpi->y_round[qindex];
1.278 + x->plane[0].zbin_extra = (int16_t)zbin_extra;
1.279 + x->e_mbd.plane[0].dequant = cpi->common.y_dequant[qindex];
1.280 +
1.281 + // UV
1.282 + zbin_extra = (cpi->common.uv_dequant[qindex][1] *
1.283 + (cpi->zbin_mode_boost + x->act_zbin_adj)) >> 7;
1.284 +
1.285 + for (i = 1; i < 3; i++) {
1.286 + x->plane[i].quant = cpi->uv_quant[qindex];
1.287 + x->plane[i].quant_shift = cpi->uv_quant_shift[qindex];
1.288 + x->plane[i].zbin = cpi->uv_zbin[qindex];
1.289 + x->plane[i].round = cpi->uv_round[qindex];
1.290 + x->plane[i].zbin_extra = (int16_t)zbin_extra;
1.291 + x->e_mbd.plane[i].dequant = cpi->common.uv_dequant[qindex];
1.292 + }
1.293 +
1.294 +#if CONFIG_ALPHA
1.295 + x->plane[3].quant = cpi->a_quant[qindex];
1.296 + x->plane[3].quant_shift = cpi->a_quant_shift[qindex];
1.297 + x->plane[3].zbin = cpi->a_zbin[qindex];
1.298 + x->plane[3].round = cpi->a_round[qindex];
1.299 + x->plane[3].zbin_extra = (int16_t)zbin_extra;
1.300 + x->e_mbd.plane[3].dequant = cpi->common.a_dequant[qindex];
1.301 +#endif
1.302 +
1.303 + x->skip_block = vp9_segfeature_active(&cpi->common.seg, segment_id,
1.304 + SEG_LVL_SKIP);
1.305 +
1.306 + /* save this macroblock QIndex for vp9_update_zbin_extra() */
1.307 + x->q_index = qindex;
1.308 +
1.309 + /* R/D setup */
1.310 + cpi->mb.errorperbit = rdmult >> 6;
1.311 + cpi->mb.errorperbit += (cpi->mb.errorperbit == 0);
1.312 +
1.313 + vp9_initialize_me_consts(cpi, x->q_index);
1.314 +}
1.315 +
1.316 +void vp9_update_zbin_extra(VP9_COMP *cpi, MACROBLOCK *x) {
1.317 + const int qindex = x->q_index;
1.318 + const int y_zbin_extra = (cpi->common.y_dequant[qindex][1] *
1.319 + (cpi->zbin_mode_boost + x->act_zbin_adj)) >> 7;
1.320 + const int uv_zbin_extra = (cpi->common.uv_dequant[qindex][1] *
1.321 + (cpi->zbin_mode_boost + x->act_zbin_adj)) >> 7;
1.322 +
1.323 + x->plane[0].zbin_extra = (int16_t)y_zbin_extra;
1.324 + x->plane[1].zbin_extra = (int16_t)uv_zbin_extra;
1.325 + x->plane[2].zbin_extra = (int16_t)uv_zbin_extra;
1.326 +}
1.327 +
1.328 +void vp9_frame_init_quantizer(VP9_COMP *cpi) {
1.329 + // Clear Zbin mode boost for default case
1.330 + cpi->zbin_mode_boost = 0;
1.331 +
1.332 + // MB level quantizer setup
1.333 + vp9_mb_init_quantizer(cpi, &cpi->mb);
1.334 +}
1.335 +
1.336 +void vp9_set_quantizer(struct VP9_COMP *cpi, int q) {
1.337 + VP9_COMMON *cm = &cpi->common;
1.338 +
1.339 + cm->base_qindex = q;
1.340 +
1.341 + // if any of the delta_q values are changing update flag will
1.342 + // have to be set.
1.343 + cm->y_dc_delta_q = 0;
1.344 + cm->uv_dc_delta_q = 0;
1.345 + cm->uv_ac_delta_q = 0;
1.346 +
1.347 + // quantizer has to be reinitialized if any delta_q changes.
1.348 + // As there are not any here for now this is inactive code.
1.349 + // if(update)
1.350 + // vp9_init_quantizer(cpi);
1.351 +}
