vpx/vp9/common/vp9_entropy.c

362 lines
14 KiB
C
Raw Normal View History

/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
2010-05-18 17:58:33 +02:00
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
2010-05-18 17:58:33 +02:00
*/
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_blockd.h"
#include "vp9/common/vp9_onyxc_int.h"
#include "vp9/common/vp9_entropymode.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx/vpx_integer.h"
2010-05-18 17:58:33 +02:00
#define MODEL_NODES (ENTROPY_NODES - UNCONSTRAINED_NODES)
DECLARE_ALIGNED(16, const uint8_t, vp9_norm[256]) = {
0, 7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
};
DECLARE_ALIGNED(16, const uint8_t,
vp9_coefband_trans_8x8plus[MAXBAND_INDEX + 1]) = {
0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4,
4, 4, 4, 4, 4, 5
};
DECLARE_ALIGNED(16, const uint8_t,
vp9_coefband_trans_4x4[MAXBAND_INDEX + 1]) = {
0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 5, 5, 5,
5, 5, 5, 5, 5, 5
};
DECLARE_ALIGNED(16, const uint8_t, vp9_pt_energy_class[MAX_ENTROPY_TOKENS]) = {
0, 1, 2, 3, 3, 4, 4, 5, 5, 5, 5, 5
};
2010-05-18 17:58:33 +02:00
/* Array indices are identical to previously-existing CONTEXT_NODE indices */
const vp9_tree_index vp9_coef_tree[TREE_SIZE(MAX_ENTROPY_TOKENS)] = {
-DCT_EOB_TOKEN, 2, /* 0 = EOB */
-ZERO_TOKEN, 4, /* 1 = ZERO */
-ONE_TOKEN, 6, /* 2 = ONE */
8, 12, /* 3 = LOW_VAL */
-TWO_TOKEN, 10, /* 4 = TWO */
-THREE_TOKEN, -FOUR_TOKEN, /* 5 = THREE */
14, 16, /* 6 = HIGH_LOW */
-DCT_VAL_CATEGORY1, -DCT_VAL_CATEGORY2, /* 7 = CAT_ONE */
18, 20, /* 8 = CAT_THREEFOUR */
-DCT_VAL_CATEGORY3, -DCT_VAL_CATEGORY4, /* 9 = CAT_THREE */
-DCT_VAL_CATEGORY5, -DCT_VAL_CATEGORY6 /* 10 = CAT_FIVE */
2010-05-18 17:58:33 +02:00
};
struct vp9_token vp9_coef_encodings[MAX_ENTROPY_TOKENS];
2010-05-18 17:58:33 +02:00
/* Trees for extra bits. Probabilities are constant and
do not depend on previously encoded bits */
static const vp9_prob Pcat1[] = { 159};
static const vp9_prob Pcat2[] = { 165, 145};
static const vp9_prob Pcat3[] = { 173, 148, 140};
static const vp9_prob Pcat4[] = { 176, 155, 140, 135};
static const vp9_prob Pcat5[] = { 180, 157, 141, 134, 130};
static const vp9_prob Pcat6[] = {
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129
};
2010-05-18 17:58:33 +02:00
const vp9_tree_index vp9_coefmodel_tree[6] = {
-DCT_EOB_MODEL_TOKEN, 2, /* 0 = EOB */
-ZERO_TOKEN, 4, /* 1 = ZERO */
-ONE_TOKEN, -TWO_TOKEN,
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
};
// Model obtained from a 2-sided zero-centerd distribuition derived
// from a Pareto distribution. The cdf of the distribution is:
// cdf(x) = 0.5 + 0.5 * sgn(x) * [1 - {alpha/(alpha + |x|)} ^ beta]
//
// For a given beta and a given probablity of the 1-node, the alpha
// is first solved, and then the {alpha, beta} pair is used to generate
// the probabilities for the rest of the nodes.
// beta = 8
static const vp9_prob modelcoefprobs_pareto8[COEFPROB_MODELS][MODEL_NODES] = {
{ 3, 86, 128, 6, 86, 23, 88, 29},
{ 9, 86, 129, 17, 88, 61, 94, 76},
{ 15, 87, 129, 28, 89, 93, 100, 110},
{ 20, 88, 130, 38, 91, 118, 106, 136},
{ 26, 89, 131, 48, 92, 139, 111, 156},
{ 31, 90, 131, 58, 94, 156, 117, 171},
{ 37, 90, 132, 66, 95, 171, 122, 184},
{ 42, 91, 132, 75, 97, 183, 127, 194},
{ 47, 92, 133, 83, 98, 193, 132, 202},
{ 52, 93, 133, 90, 100, 201, 137, 208},
{ 57, 94, 134, 98, 101, 208, 142, 214},
{ 62, 94, 135, 105, 103, 214, 146, 218},
{ 66, 95, 135, 111, 104, 219, 151, 222},
{ 71, 96, 136, 117, 106, 224, 155, 225},
{ 76, 97, 136, 123, 107, 227, 159, 228},
{ 80, 98, 137, 129, 109, 231, 162, 231},
{ 84, 98, 138, 134, 110, 234, 166, 233},
{ 89, 99, 138, 140, 112, 236, 170, 235},
{ 93, 100, 139, 145, 113, 238, 173, 236},
{ 97, 101, 140, 149, 115, 240, 176, 238},
{101, 102, 140, 154, 116, 242, 179, 239},
{105, 103, 141, 158, 118, 243, 182, 240},
{109, 104, 141, 162, 119, 244, 185, 241},
{113, 104, 142, 166, 120, 245, 187, 242},
{116, 105, 143, 170, 122, 246, 190, 243},
{120, 106, 143, 173, 123, 247, 192, 244},
{123, 107, 144, 177, 125, 248, 195, 244},
{127, 108, 145, 180, 126, 249, 197, 245},
{130, 109, 145, 183, 128, 249, 199, 245},
{134, 110, 146, 186, 129, 250, 201, 246},
{137, 111, 147, 189, 131, 251, 203, 246},
{140, 112, 147, 192, 132, 251, 205, 247},
{143, 113, 148, 194, 133, 251, 207, 247},
{146, 114, 149, 197, 135, 252, 208, 248},
{149, 115, 149, 199, 136, 252, 210, 248},
{152, 115, 150, 201, 138, 252, 211, 248},
{155, 116, 151, 204, 139, 253, 213, 249},
{158, 117, 151, 206, 140, 253, 214, 249},
{161, 118, 152, 208, 142, 253, 216, 249},
{163, 119, 153, 210, 143, 253, 217, 249},
{166, 120, 153, 212, 144, 254, 218, 250},
{168, 121, 154, 213, 146, 254, 220, 250},
{171, 122, 155, 215, 147, 254, 221, 250},
{173, 123, 155, 217, 148, 254, 222, 250},
{176, 124, 156, 218, 150, 254, 223, 250},
{178, 125, 157, 220, 151, 254, 224, 251},
{180, 126, 157, 221, 152, 254, 225, 251},
{183, 127, 158, 222, 153, 254, 226, 251},
{185, 128, 159, 224, 155, 255, 227, 251},
{187, 129, 160, 225, 156, 255, 228, 251},
{189, 131, 160, 226, 157, 255, 228, 251},
{191, 132, 161, 227, 159, 255, 229, 251},
{193, 133, 162, 228, 160, 255, 230, 252},
{195, 134, 163, 230, 161, 255, 231, 252},
{197, 135, 163, 231, 162, 255, 231, 252},
{199, 136, 164, 232, 163, 255, 232, 252},
{201, 137, 165, 233, 165, 255, 233, 252},
{202, 138, 166, 233, 166, 255, 233, 252},
{204, 139, 166, 234, 167, 255, 234, 252},
{206, 140, 167, 235, 168, 255, 235, 252},
{207, 141, 168, 236, 169, 255, 235, 252},
{209, 142, 169, 237, 171, 255, 236, 252},
{210, 144, 169, 237, 172, 255, 236, 252},
{212, 145, 170, 238, 173, 255, 237, 252},
{214, 146, 171, 239, 174, 255, 237, 253},
{215, 147, 172, 240, 175, 255, 238, 253},
{216, 148, 173, 240, 176, 255, 238, 253},
{218, 149, 173, 241, 177, 255, 239, 253},
{219, 150, 174, 241, 179, 255, 239, 253},
{220, 152, 175, 242, 180, 255, 240, 253},
{222, 153, 176, 242, 181, 255, 240, 253},
{223, 154, 177, 243, 182, 255, 240, 253},
{224, 155, 178, 244, 183, 255, 241, 253},
{225, 156, 178, 244, 184, 255, 241, 253},
{226, 158, 179, 244, 185, 255, 242, 253},
{228, 159, 180, 245, 186, 255, 242, 253},
{229, 160, 181, 245, 187, 255, 242, 253},
{230, 161, 182, 246, 188, 255, 243, 253},
{231, 163, 183, 246, 189, 255, 243, 253},
{232, 164, 184, 247, 190, 255, 243, 253},
{233, 165, 185, 247, 191, 255, 244, 253},
{234, 166, 185, 247, 192, 255, 244, 253},
{235, 168, 186, 248, 193, 255, 244, 253},
{236, 169, 187, 248, 194, 255, 244, 253},
{236, 170, 188, 248, 195, 255, 245, 253},
{237, 171, 189, 249, 196, 255, 245, 254},
{238, 173, 190, 249, 197, 255, 245, 254},
{239, 174, 191, 249, 198, 255, 245, 254},
{240, 175, 192, 249, 199, 255, 246, 254},
{240, 177, 193, 250, 200, 255, 246, 254},
{241, 178, 194, 250, 201, 255, 246, 254},
{242, 179, 195, 250, 202, 255, 246, 254},
{242, 181, 196, 250, 203, 255, 247, 254},
{243, 182, 197, 251, 204, 255, 247, 254},
{244, 184, 198, 251, 205, 255, 247, 254},
{244, 185, 199, 251, 206, 255, 247, 254},
{245, 186, 200, 251, 207, 255, 247, 254},
{246, 188, 201, 252, 207, 255, 248, 254},
{246, 189, 202, 252, 208, 255, 248, 254},
{247, 191, 203, 252, 209, 255, 248, 254},
{247, 192, 204, 252, 210, 255, 248, 254},
{248, 194, 205, 252, 211, 255, 248, 254},
{248, 195, 206, 252, 212, 255, 249, 254},
{249, 197, 207, 253, 213, 255, 249, 254},
{249, 198, 208, 253, 214, 255, 249, 254},
{250, 200, 210, 253, 215, 255, 249, 254},
{250, 201, 211, 253, 215, 255, 249, 254},
{250, 203, 212, 253, 216, 255, 249, 254},
{251, 204, 213, 253, 217, 255, 250, 254},
{251, 206, 214, 254, 218, 255, 250, 254},
{252, 207, 216, 254, 219, 255, 250, 254},
{252, 209, 217, 254, 220, 255, 250, 254},
{252, 211, 218, 254, 221, 255, 250, 254},
{253, 213, 219, 254, 222, 255, 250, 254},
{253, 214, 221, 254, 223, 255, 250, 254},
{253, 216, 222, 254, 224, 255, 251, 254},
{253, 218, 224, 254, 225, 255, 251, 254},
{254, 220, 225, 254, 225, 255, 251, 254},
{254, 222, 227, 255, 226, 255, 251, 254},
{254, 224, 228, 255, 227, 255, 251, 254},
{254, 226, 230, 255, 228, 255, 251, 254},
{255, 228, 231, 255, 230, 255, 251, 254},
{255, 230, 233, 255, 231, 255, 252, 254},
{255, 232, 235, 255, 232, 255, 252, 254},
{255, 235, 237, 255, 233, 255, 252, 254},
{255, 238, 240, 255, 235, 255, 252, 255},
{255, 241, 243, 255, 236, 255, 252, 254},
{255, 246, 247, 255, 239, 255, 253, 255}
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
};
static void extend_model_to_full_distribution(vp9_prob p,
vp9_prob *tree_probs) {
const int l = (p - 1) / 2;
const vp9_prob (*model)[MODEL_NODES] = modelcoefprobs_pareto8;
if (p & 1) {
vpx_memcpy(tree_probs + UNCONSTRAINED_NODES,
model[l], MODEL_NODES * sizeof(vp9_prob));
} else {
// interpolate
int i;
for (i = UNCONSTRAINED_NODES; i < ENTROPY_NODES; ++i)
tree_probs[i] = (model[l][i - UNCONSTRAINED_NODES] +
model[l + 1][i - UNCONSTRAINED_NODES]) >> 1;
}
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
}
void vp9_model_to_full_probs(const vp9_prob *model, vp9_prob *full) {
if (full != model)
vpx_memcpy(full, model, sizeof(vp9_prob) * UNCONSTRAINED_NODES);
extend_model_to_full_distribution(model[PIVOT_NODE], full);
}
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
static vp9_tree_index cat1[2], cat2[4], cat3[6], cat4[8], cat5[10], cat6[28];
2010-05-18 17:58:33 +02:00
static void init_bit_tree(vp9_tree_index *p, int n) {
int i = 0;
2010-05-18 17:58:33 +02:00
while (++i < n) {
p[0] = p[1] = i << 1;
p += 2;
}
2010-05-18 17:58:33 +02:00
p[0] = p[1] = 0;
2010-05-18 17:58:33 +02:00
}
static void init_bit_trees() {
init_bit_tree(cat1, 1);
init_bit_tree(cat2, 2);
init_bit_tree(cat3, 3);
init_bit_tree(cat4, 4);
init_bit_tree(cat5, 5);
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
init_bit_tree(cat6, 14);
2010-05-18 17:58:33 +02:00
}
const vp9_extra_bit vp9_extra_bits[MAX_ENTROPY_TOKENS] = {
{ 0, 0, 0, 0},
{ 0, 0, 0, 1},
{ 0, 0, 0, 2},
{ 0, 0, 0, 3},
{ 0, 0, 0, 4},
{ cat1, Pcat1, 1, 5},
{ cat2, Pcat2, 2, 7},
{ cat3, Pcat3, 3, 11},
{ cat4, Pcat4, 4, 19},
{ cat5, Pcat5, 5, 35},
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
{ cat6, Pcat6, 14, 67},
{ 0, 0, 0, 0}
2010-05-18 17:58:33 +02:00
};
#include "vp9/common/vp9_default_coef_probs.h"
2010-05-18 17:58:33 +02:00
void vp9_default_coef_probs(VP9_COMMON *cm) {
vp9_copy(cm->fc.coef_probs[TX_4X4], default_coef_probs_4x4);
vp9_copy(cm->fc.coef_probs[TX_8X8], default_coef_probs_8x8);
vp9_copy(cm->fc.coef_probs[TX_16X16], default_coef_probs_16x16);
vp9_copy(cm->fc.coef_probs[TX_32X32], default_coef_probs_32x32);
}
2010-05-18 17:58:33 +02:00
void vp9_coef_tree_initialize() {
init_bit_trees();
vp9_tokens_from_tree(vp9_coef_encodings, vp9_coef_tree);
}
// #define COEF_COUNT_TESTING
#define COEF_COUNT_SAT 24
#define COEF_MAX_UPDATE_FACTOR 112
#define COEF_COUNT_SAT_KEY 24
#define COEF_MAX_UPDATE_FACTOR_KEY 112
#define COEF_COUNT_SAT_AFTER_KEY 24
#define COEF_MAX_UPDATE_FACTOR_AFTER_KEY 128
static void adapt_coef_probs(VP9_COMMON *cm, TX_SIZE tx_size,
unsigned int count_sat,
unsigned int update_factor) {
const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
vp9_coeff_probs_model *dst_coef_probs = cm->fc.coef_probs[tx_size];
const vp9_coeff_probs_model *pre_coef_probs = pre_fc->coef_probs[tx_size];
vp9_coeff_count_model *coef_counts = cm->counts.coef[tx_size];
unsigned int (*eob_branch_count)[REF_TYPES][COEF_BANDS][PREV_COEF_CONTEXTS] =
cm->counts.eob_branch[tx_size];
int i, j, k, l, m;
unsigned int branch_ct[UNCONSTRAINED_NODES][2];
for (i = 0; i < BLOCK_TYPES; ++i)
for (j = 0; j < REF_TYPES; ++j)
for (k = 0; k < COEF_BANDS; ++k)
for (l = 0; l < PREV_COEF_CONTEXTS; ++l) {
if (l >= 3 && k == 0)
continue;
vp9_tree_probs_from_distribution(vp9_coefmodel_tree, branch_ct,
coef_counts[i][j][k][l]);
branch_ct[0][1] = eob_branch_count[i][j][k][l] - branch_ct[0][0];
for (m = 0; m < UNCONSTRAINED_NODES; ++m)
dst_coef_probs[i][j][k][l][m] = merge_probs(
pre_coef_probs[i][j][k][l][m],
branch_ct[m],
count_sat, update_factor);
}
}
void vp9_adapt_coef_probs(VP9_COMMON *cm) {
TX_SIZE t;
unsigned int count_sat, update_factor;
if (frame_is_intra_only(cm)) {
update_factor = COEF_MAX_UPDATE_FACTOR_KEY;
count_sat = COEF_COUNT_SAT_KEY;
} else if (cm->last_frame_type == KEY_FRAME) {
update_factor = COEF_MAX_UPDATE_FACTOR_AFTER_KEY; /* adapt quickly */
count_sat = COEF_COUNT_SAT_AFTER_KEY;
} else {
update_factor = COEF_MAX_UPDATE_FACTOR;
count_sat = COEF_COUNT_SAT;
}
for (t = TX_4X4; t <= TX_32X32; t++)
adapt_coef_probs(cm, t, count_sat, update_factor);
}