generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
89ac94f8fb
vpx/vp9/common/vp9_idctllm.c

2651 lines
83 KiB
C
Raw Normal View History

Initial WebM release
2010-05-18 17:58:33 +02:00
/*
Use WebM in copyright notice for consistency Changes 'The VP8 project' to 'The WebM project', for consistency with other webmproject.org repositories. Fixes issue #97. Change-Id: I37c13ed5fbdb9d334ceef71c6350e9febed9bbba
2010-09-09 14:16:39 +02:00
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
Initial WebM release
2010-05-18 17:58:33 +02:00
*
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
* Use of this source code is governed by a BSD-style license
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
* in the file PATENTS. All contributing project authors may
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
* be found in the AUTHORS file in the root of the source tree.
Initial WebM release
2010-05-18 17:58:33 +02:00
*/
/****************************************************************************
* Notes:
*
* This implementation makes use of 16 bit fixed point verio of two multiply
* constants:
* 1. sqrt(2) * cos (pi/8)
* 2. sqrt(2) * sin (pi/8)
* Becuase the first constant is bigger than 1, to maintain the same 16 bit
* fixed point precision as the second one, we use a trick of
* x * a = x + x*(a-1)
* so
* x * sqrt(2) * cos (pi/8) = x + x * (sqrt(2) *cos(pi/8)-1).
**************************************************************************/
Fixed style warnings Fixed some style warnings. Change-Id: I9eb77c3bf6d63628feef1b0c936e905bf4f2089d
2012-10-31 01:54:22 +01:00
#include <assert.h>
#include <math.h>
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
#include "vpx_ports/config.h"
Add vp9_ prefix to all vp9 files Support for gyp which doesn't support multiple objects in the same static library having the same basename. Change-Id: Ib947eefbaf68f8b177a796d23f875ccdfa6bc9dc
2012-11-27 22:59:17 +01:00
#include "vp9/common/vp9_systemdependent.h"
#include "vp9/common/vp9_blockd.h"
Consistently use get_prob(), clip_prob() and newly added clip_pixel(). Add a function clip_pixel() to clip a pixel value to the [0,255] range of allowed values, and use this where-ever appropriate (e.g. prediction, reconstruction). Likewise, consistently use the recently added function clip_prob(), which calculates a binary probability in the [1,255] range. If possible, try to use get_prob() or its sister get_binary_prob() to calculate binary probabilities, for consistency. Since in some places, this means that binary probability calculations are changed (we use {255,256}*count0/(total) in a range of places, and all of these are now changed to use 256*count0+(total>>1)/total), this changes the encoding result, so this patch warrants some extensive testing. Change-Id: Ibeeff8d886496839b8e0c0ace9ccc552351f7628
2012-12-10 21:09:07 +01:00
#include "vp9/common/vp9_common.h"
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
Initial WebM release
2010-05-18 17:58:33 +02:00
static const int cospi8sqrt2minus1 = 20091;
static const int sinpi8sqrt2 = 35468;
static const int rounding = 0;
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
// TODO: these transforms can be further converted into integer forms
// for complexity optimization
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const float idct_4[16] = {
Adds hybrid transform Adds ADST/DCT hybrid transform coding for Intra4x4 mode. The ADST is applied to directions in which the boundary pixels are used for prediction, while DCT applied to directions without corresponding boundary prediction. Adds enum TX_TYPE in b_mode_infor to indicate the transform type used. Make coding style consistent with google style. Fixed the commented issues. Experimental results in terms of bit-rate reduction: derf: 0.731% yt: 0.982% std-hd: 0.459% hd: 0.725% Will be looking at 8x8 transforms next. Change-Id: I46dbd7b80dbb3e8856e9c34fbc58cb3764a12fcf
2012-06-25 21:26:09 +02:00
0.500000000000000, 0.653281482438188, 0.500000000000000, 0.270598050073099,
0.500000000000000, 0.270598050073099, -0.500000000000000, -0.653281482438188,
0.500000000000000, -0.270598050073099, -0.500000000000000, 0.653281482438188,
0.500000000000000, -0.653281482438188, 0.500000000000000, -0.270598050073099
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const float iadst_4[16] = {
Adds hybrid transform Adds ADST/DCT hybrid transform coding for Intra4x4 mode. The ADST is applied to directions in which the boundary pixels are used for prediction, while DCT applied to directions without corresponding boundary prediction. Adds enum TX_TYPE in b_mode_infor to indicate the transform type used. Make coding style consistent with google style. Fixed the commented issues. Experimental results in terms of bit-rate reduction: derf: 0.731% yt: 0.982% std-hd: 0.459% hd: 0.725% Will be looking at 8x8 transforms next. Change-Id: I46dbd7b80dbb3e8856e9c34fbc58cb3764a12fcf
2012-06-25 21:26:09 +02:00
0.228013428883779, 0.577350269189626, 0.656538502008139, 0.428525073124360,
0.428525073124360, 0.577350269189626, -0.228013428883779, -0.656538502008139,
0.577350269189626, 0, -0.577350269189626, 0.577350269189626,
0.656538502008139, -0.577350269189626, 0.428525073124359, -0.228013428883779
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const float idct_8[64] = {
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
0.353553390593274, 0.490392640201615, 0.461939766255643, 0.415734806151273,
0.353553390593274, 0.277785116509801, 0.191341716182545, 0.097545161008064,
0.353553390593274, 0.415734806151273, 0.191341716182545, -0.097545161008064,
-0.353553390593274, -0.490392640201615, -0.461939766255643, -0.277785116509801,
0.353553390593274, 0.277785116509801, -0.191341716182545, -0.490392640201615,
-0.353553390593274, 0.097545161008064, 0.461939766255643, 0.415734806151273,
0.353553390593274, 0.097545161008064, -0.461939766255643, -0.277785116509801,
0.353553390593274, 0.415734806151273, -0.191341716182545, -0.490392640201615,
0.353553390593274, -0.097545161008064, -0.461939766255643, 0.277785116509801,
0.353553390593274, -0.415734806151273, -0.191341716182545, 0.490392640201615,
0.353553390593274, -0.277785116509801, -0.191341716182545, 0.490392640201615,
-0.353553390593274, -0.097545161008064, 0.461939766255643, -0.415734806151273,
0.353553390593274, -0.415734806151273, 0.191341716182545, 0.097545161008064,
-0.353553390593274, 0.490392640201615, -0.461939766255643, 0.277785116509801,
0.353553390593274, -0.490392640201615, 0.461939766255643, -0.415734806151273,
0.353553390593274, -0.277785116509801, 0.191341716182545, -0.097545161008064
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const float iadst_8[64] = {
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
0.089131608307533, 0.255357107325376, 0.387095214016349, 0.466553967085785,
0.483002021635509, 0.434217976756762, 0.326790388032145, 0.175227946595735,
0.175227946595735, 0.434217976756762, 0.466553967085785, 0.255357107325376,
-0.089131608307533, -0.387095214016348, -0.483002021635509, -0.326790388032145,
0.255357107325376, 0.483002021635509, 0.175227946595735, -0.326790388032145,
-0.466553967085785, -0.089131608307533, 0.387095214016349, 0.434217976756762,
0.326790388032145, 0.387095214016349, -0.255357107325376, -0.434217976756762,
0.175227946595735, 0.466553967085786, -0.089131608307534, -0.483002021635509,
0.387095214016349, 0.175227946595735, -0.483002021635509, 0.089131608307533,
0.434217976756762, -0.326790388032145, -0.255357107325377, 0.466553967085785,
0.434217976756762, -0.089131608307533, -0.326790388032145, 0.483002021635509,
-0.255357107325376, -0.175227946595735, 0.466553967085785, -0.387095214016348,
0.466553967085785, -0.326790388032145, 0.089131608307533, 0.175227946595735,
-0.387095214016348, 0.483002021635509, -0.434217976756762, 0.255357107325376,
0.483002021635509, -0.466553967085785, 0.434217976756762, -0.387095214016348,
0.326790388032145, -0.255357107325375, 0.175227946595736, -0.089131608307532
};
hybrid transform of 16x16 dimension Enable ADST/DCT of dimension 16x16 for I16X16 modes. This change provides benefits mostly for hd sequences. Set up the framework for selectable transform dimension. Also allowing quantization parameter threshold to control the use of hybrid transform (This is currently disabled by setting threshold always above the quantization parameter. Adaptive thresholding can be built upon this, which will further improve the coding performance.) The coding performance gains (with respect to the codec that has all other configuration settings turned on) are derf: 0.013 yt: 0.086 hd: 0.198 std-hd: 0.501 Change-Id: Ibb4263a61fc74e0b3c345f54d73e8c73552bf926
2012-08-29 20:25:38 +02:00
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const int16_t idct_i4[16] = {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
8192, 10703, 8192, 4433,
8192, 4433, -8192, -10703,
8192, -4433, -8192, 10703,
8192, -10703, 8192, -4433
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const int16_t iadst_i4[16] = {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
3736, 9459, 10757, 7021,
7021, 9459, -3736, -10757,
9459, 0, -9459, 9459,
10757, -9459, 7021, -3736
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const int16_t idct_i8[64] = {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
5793, 8035, 7568, 6811,
5793, 4551, 3135, 1598,
5793, 6811, 3135, -1598,
-5793, -8035, -7568, -4551,
5793, 4551, -3135, -8035,
-5793, 1598, 7568, 6811,
5793, 1598, -7568, -4551,
5793, 6811, -3135, -8035,
5793, -1598, -7568, 4551,
5793, -6811, -3135, 8035,
5793, -4551, -3135, 8035,
-5793, -1598, 7568, -6811,
5793, -6811, 3135, 1598,
-5793, 8035, -7568, 4551,
5793, -8035, 7568, -6811,
5793, -4551, 3135, -1598
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const int16_t iadst_i8[64] = {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
1460, 4184, 6342, 7644,
7914, 7114, 5354, 2871,
2871, 7114, 7644, 4184,
-1460, -6342, -7914, -5354,
4184, 7914, 2871, -5354,
-7644, -1460, 6342, 7114,
5354, 6342, -4184, -7114,
2871, 7644, -1460, -7914,
6342, 2871, -7914, 1460,
7114, -5354, -4184, 7644,
7114, -1460, -5354, 7914,
-4184, -2871, 7644, -6342,
7644, -5354, 1460, 2871,
-6342, 7914, -7114, 4184,
7914, -7644, 7114, -6342,
5354, -4184, 2871, -1460
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static float idct_16[256] = {
hybrid transform of 16x16 dimension Enable ADST/DCT of dimension 16x16 for I16X16 modes. This change provides benefits mostly for hd sequences. Set up the framework for selectable transform dimension. Also allowing quantization parameter threshold to control the use of hybrid transform (This is currently disabled by setting threshold always above the quantization parameter. Adaptive thresholding can be built upon this, which will further improve the coding performance.) The coding performance gains (with respect to the codec that has all other configuration settings turned on) are derf: 0.013 yt: 0.086 hd: 0.198 std-hd: 0.501 Change-Id: Ibb4263a61fc74e0b3c345f54d73e8c73552bf926
2012-08-29 20:25:38 +02:00
0.250000, 0.351851, 0.346760, 0.338330, 0.326641, 0.311806, 0.293969, 0.273300,
0.250000, 0.224292, 0.196424, 0.166664, 0.135299, 0.102631, 0.068975, 0.034654,
0.250000, 0.338330, 0.293969, 0.224292, 0.135299, 0.034654, -0.068975, -0.166664,
-0.250000, -0.311806, -0.346760, -0.351851, -0.326641, -0.273300, -0.196424, -0.102631,
0.250000, 0.311806, 0.196424, 0.034654, -0.135299, -0.273300, -0.346760, -0.338330,
-0.250000, -0.102631, 0.068975, 0.224292, 0.326641, 0.351851, 0.293969, 0.166664,
0.250000, 0.273300, 0.068975, -0.166664, -0.326641, -0.338330, -0.196424, 0.034654,
0.250000, 0.351851, 0.293969, 0.102631, -0.135299, -0.311806, -0.346760, -0.224292,
0.250000, 0.224292, -0.068975, -0.311806, -0.326641, -0.102631, 0.196424, 0.351851,
0.250000, -0.034654, -0.293969, -0.338330, -0.135299, 0.166664, 0.346760, 0.273300,
0.250000, 0.166664, -0.196424, -0.351851, -0.135299, 0.224292, 0.346760, 0.102631,
-0.250000, -0.338330, -0.068975, 0.273300, 0.326641, 0.034654, -0.293969, -0.311806,
0.250000, 0.102631, -0.293969, -0.273300, 0.135299, 0.351851, 0.068975, -0.311806,
-0.250000, 0.166664, 0.346760, 0.034654, -0.326641, -0.224292, 0.196424, 0.338330,
0.250000, 0.034654, -0.346760, -0.102631, 0.326641, 0.166664, -0.293969, -0.224292,
0.250000, 0.273300, -0.196424, -0.311806, 0.135299, 0.338330, -0.068975, -0.351851,
0.250000, -0.034654, -0.346760, 0.102631, 0.326641, -0.166664, -0.293969, 0.224292,
0.250000, -0.273300, -0.196424, 0.311806, 0.135299, -0.338330, -0.068975, 0.351851,
0.250000, -0.102631, -0.293969, 0.273300, 0.135299, -0.351851, 0.068975, 0.311806,
-0.250000, -0.166664, 0.346760, -0.034654, -0.326641, 0.224292, 0.196424, -0.338330,
0.250000, -0.166664, -0.196424, 0.351851, -0.135299, -0.224292, 0.346760, -0.102631,
-0.250000, 0.338330, -0.068975, -0.273300, 0.326641, -0.034654, -0.293969, 0.311806,
0.250000, -0.224292, -0.068975, 0.311806, -0.326641, 0.102631, 0.196424, -0.351851,
0.250000, 0.034654, -0.293969, 0.338330, -0.135299, -0.166664, 0.346760, -0.273300,
0.250000, -0.273300, 0.068975, 0.166664, -0.326641, 0.338330, -0.196424, -0.034654,
0.250000, -0.351851, 0.293969, -0.102631, -0.135299, 0.311806, -0.346760, 0.224292,
0.250000, -0.311806, 0.196424, -0.034654, -0.135299, 0.273300, -0.346760, 0.338330,
-0.250000, 0.102631, 0.068975, -0.224292, 0.326641, -0.351851, 0.293969, -0.166664,
0.250000, -0.338330, 0.293969, -0.224292, 0.135299, -0.034654, -0.068975, 0.166664,
-0.250000, 0.311806, -0.346760, 0.351851, -0.326641, 0.273300, -0.196424, 0.102631,
0.250000, -0.351851, 0.346760, -0.338330, 0.326641, -0.311806, 0.293969, -0.273300,
0.250000, -0.224292, 0.196424, -0.166664, 0.135299, -0.102631, 0.068975, -0.034654
};
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static float iadst_16[256] = {
hybrid transform of 16x16 dimension Enable ADST/DCT of dimension 16x16 for I16X16 modes. This change provides benefits mostly for hd sequences. Set up the framework for selectable transform dimension. Also allowing quantization parameter threshold to control the use of hybrid transform (This is currently disabled by setting threshold always above the quantization parameter. Adaptive thresholding can be built upon this, which will further improve the coding performance.) The coding performance gains (with respect to the codec that has all other configuration settings turned on) are derf: 0.013 yt: 0.086 hd: 0.198 std-hd: 0.501 Change-Id: Ibb4263a61fc74e0b3c345f54d73e8c73552bf926
2012-08-29 20:25:38 +02:00
0.033094, 0.098087, 0.159534, 0.215215, 0.263118, 0.301511, 0.329007, 0.344612,
0.347761, 0.338341, 0.316693, 0.283599, 0.240255, 0.188227, 0.129396, 0.065889,
0.065889, 0.188227, 0.283599, 0.338341, 0.344612, 0.301511, 0.215215, 0.098087,
-0.033094, -0.159534, -0.263118, -0.329007, -0.347761, -0.316693, -0.240255, -0.129396,
0.098087, 0.263118, 0.344612, 0.316693, 0.188227, 0.000000, -0.188227, -0.316693,
-0.344612, -0.263118, -0.098087, 0.098087, 0.263118, 0.344612, 0.316693, 0.188227,
0.129396, 0.316693, 0.329007, 0.159534, -0.098087, -0.301511, -0.338341, -0.188227,
0.065889, 0.283599, 0.344612, 0.215215, -0.033094, -0.263118, -0.347761, -0.240255,
0.159534, 0.344612, 0.240255, -0.065889, -0.316693, -0.301511, -0.033094, 0.263118,
0.338341, 0.129396, -0.188227, -0.347761, -0.215215, 0.098087, 0.329007, 0.283599,
0.188227, 0.344612, 0.098087, -0.263118, -0.316693, -0.000000, 0.316693, 0.263118,
-0.098087, -0.344612, -0.188227, 0.188227, 0.344612, 0.098087, -0.263118, -0.316693,
0.215215, 0.316693, -0.065889, -0.347761, -0.098087, 0.301511, 0.240255, -0.188227,
-0.329007, 0.033094, 0.344612, 0.129396, -0.283599, -0.263118, 0.159534, 0.338341,
0.240255, 0.263118, -0.215215, -0.283599, 0.188227, 0.301511, -0.159534, -0.316693,
0.129396, 0.329007, -0.098087, -0.338341, 0.065889, 0.344612, -0.033094, -0.347761,
0.263118, 0.188227, -0.316693, -0.098087, 0.344612, 0.000000, -0.344612, 0.098087,
0.316693, -0.188227, -0.263118, 0.263118, 0.188227, -0.316693, -0.098087, 0.344612,
0.283599, 0.098087, -0.347761, 0.129396, 0.263118, -0.301511, -0.065889, 0.344612,
-0.159534, -0.240255, 0.316693, 0.033094, -0.338341, 0.188227, 0.215215, -0.329007,
0.301511, 0.000000, -0.301511, 0.301511, 0.000000, -0.301511, 0.301511, 0.000000,
-0.301511, 0.301511, 0.000000, -0.301511, 0.301511, 0.000000, -0.301511, 0.301511,
0.316693, -0.098087, -0.188227, 0.344612, -0.263118, -0.000000, 0.263118, -0.344612,
0.188227, 0.098087, -0.316693, 0.316693, -0.098087, -0.188227, 0.344612, -0.263118,
0.329007, -0.188227, -0.033094, 0.240255, -0.344612, 0.301511, -0.129396, -0.098087,
0.283599, -0.347761, 0.263118, -0.065889, -0.159534, 0.316693, -0.338341, 0.215215,
0.338341, -0.263118, 0.129396, 0.033094, -0.188227, 0.301511, -0.347761, 0.316693,
-0.215215, 0.065889, 0.098087, -0.240255, 0.329007, -0.344612, 0.283599, -0.159534,
0.344612, -0.316693, 0.263118, -0.188227, 0.098087, 0.000000, -0.098087, 0.188227,
-0.263118, 0.316693, -0.344612, 0.344612, -0.316693, 0.263118, -0.188227, 0.098087,
0.347761, -0.344612, 0.338341, -0.329007, 0.316693, -0.301511, 0.283599, -0.263118,
0.240255, -0.215215, 0.188227, -0.159534, 0.129396, -0.098087, 0.065889, -0.033094
};
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const int16_t idct_i16[256] = {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
4096, 5765, 5681, 5543, 5352, 5109, 4816, 4478,
4096, 3675, 3218, 2731, 2217, 1682, 1130, 568,
4096, 5543, 4816, 3675, 2217, 568, -1130, -2731,
-4096, -5109, -5681, -5765, -5352, -4478, -3218, -1682,
4096, 5109, 3218, 568, -2217, -4478, -5681, -5543,
-4096, -1682, 1130, 3675, 5352, 5765, 4816, 2731,
4096, 4478, 1130, -2731, -5352, -5543, -3218, 568,
4096, 5765, 4816, 1682, -2217, -5109, -5681, -3675,
4096, 3675, -1130, -5109, -5352, -1682, 3218, 5765,
4096, -568, -4816, -5543, -2217, 2731, 5681, 4478,
4096, 2731, -3218, -5765, -2217, 3675, 5681, 1682,
-4096, -5543, -1130, 4478, 5352, 568, -4816, -5109,
4096, 1682, -4816, -4478, 2217, 5765, 1130, -5109,
-4096, 2731, 5681, 568, -5352, -3675, 3218, 5543,
4096, 568, -5681, -1682, 5352, 2731, -4816, -3675,
4096, 4478, -3218, -5109, 2217, 5543, -1130, -5765,
4096, -568, -5681, 1682, 5352, -2731, -4816, 3675,
4096, -4478, -3218, 5109, 2217, -5543, -1130, 5765,
4096, -1682, -4816, 4478, 2217, -5765, 1130, 5109,
-4096, -2731, 5681, -568, -5352, 3675, 3218, -5543,
4096, -2731, -3218, 5765, -2217, -3675, 5681, -1682,
-4096, 5543, -1130, -4478, 5352, -568, -4816, 5109,
4096, -3675, -1130, 5109, -5352, 1682, 3218, -5765,
4096, 568, -4816, 5543, -2217, -2731, 5681, -4478,
4096, -4478, 1130, 2731, -5352, 5543, -3218, -568,
4096, -5765, 4816, -1682, -2217, 5109, -5681, 3675,
4096, -5109, 3218, -568, -2217, 4478, -5681, 5543,
-4096, 1682, 1130, -3675, 5352, -5765, 4816, -2731,
4096, -5543, 4816, -3675, 2217, -568, -1130, 2731,
-4096, 5109, -5681, 5765, -5352, 4478, -3218, 1682,
4096, -5765, 5681, -5543, 5352, -5109, 4816, -4478,
4096, -3675, 3218, -2731, 2217, -1682, 1130, -568
};
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
static const int16_t iadst_i16[256] = {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
542, 1607, 2614, 3526, 4311, 4940, 5390, 5646,
5698, 5543, 5189, 4646, 3936, 3084, 2120, 1080,
1080, 3084, 4646, 5543, 5646, 4940, 3526, 1607,
-542, -2614, -4311, -5390, -5698, -5189, -3936, -2120,
1607, 4311, 5646, 5189, 3084, 0, -3084, -5189,
-5646, -4311, -1607, 1607, 4311, 5646, 5189, 3084,
2120, 5189, 5390, 2614, -1607, -4940, -5543, -3084,
1080, 4646, 5646, 3526, -542, -4311, -5698, -3936,
2614, 5646, 3936, -1080, -5189, -4940, -542, 4311,
5543, 2120, -3084, -5698, -3526, 1607, 5390, 4646,
3084, 5646, 1607, -4311, -5189, 0, 5189, 4311,
-1607, -5646, -3084, 3084, 5646, 1607, -4311, -5189,
3526, 5189, -1080, -5698, -1607, 4940, 3936, -3084,
-5390, 542, 5646, 2120, -4646, -4311, 2614, 5543,
3936, 4311, -3526, -4646, 3084, 4940, -2614, -5189,
2120, 5390, -1607, -5543, 1080, 5646, -542, -5698,
4311, 3084, -5189, -1607, 5646, 0, -5646, 1607,
5189, -3084, -4311, 4311, 3084, -5189, -1607, 5646,
4646, 1607, -5698, 2120, 4311, -4940, -1080, 5646,
-2614, -3936, 5189, 542, -5543, 3084, 3526, -5390,
4940, 0, -4940, 4940, 0, -4940, 4940, 0,
-4940, 4940, 0, -4940, 4940, 0, -4940, 4940,
5189, -1607, -3084, 5646, -4311, 0, 4311, -5646,
3084, 1607, -5189, 5189, -1607, -3084, 5646, -4311,
5390, -3084, -542, 3936, -5646, 4940, -2120, -1607,
4646, -5698, 4311, -1080, -2614, 5189, -5543, 3526,
5543, -4311, 2120, 542, -3084, 4940, -5698, 5189,
-3526, 1080, 1607, -3936, 5390, -5646, 4646, -2614,
5646, -5189, 4311, -3084, 1607, 0, -1607, 3084,
-4311, 5189, -5646, 5646, -5189, 4311, -3084, 1607,
5698, -5646, 5543, -5390, 5189, -4940, 4646, -4311,
3936, -3526, 3084, -2614, 2120, -1607, 1080, -542
};
Change common vp8_ public symbol prefixes to vp9_. Change-Id: Ic5a5f60e1ff9d9ccae4174160d36529466eeb509
2012-10-31 00:25:53 +01:00
void vp9_ihtllm_float_c(const int16_t *input, int16_t *output, int pitch,
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
TX_TYPE tx_type, int tx_dim) {
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
{
int i, j, k;
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
float bufa[256], bufb[256]; // buffers are for floating-point test purpose
// the implementation could be simplified in
// conjunction with integer transform
const int16_t *ip = input;
int16_t *op = output;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
int shortpitch = pitch >> 1;
float *pfa = &bufa[0];
float *pfb = &bufb[0];
// pointers to vertical and horizontal transforms
Change non-function symbol vp8_ prefixes to vp9_. For local symbols, make them static instead. Change-Id: I13d60947a46f711bc8991e16100cea2a13e3a22e
2012-10-31 01:12:12 +01:00
const float *ptv, *pth;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
Some cleanups and fixes. Separates the logic on transform type selection previously spread out over a number of files into a separate function. Currently the tx_type field in b_mode_info is not used, but still left in there to eventually use for signaling the transform type in the bitstream. Also, now for tx_type = DCT_DCT, the regular integer DCT is used, as opposed to the floating point DCT used in conjuction with hybrid transform. Results change somewhat due to the transform change, but are within reasonable limits. The hd/std-hd sets are slightly up, while derf/yt are slightly down. Change-Id: I5776840c2239ca2da31ca6cfd7fd1148dc5f9e0f
2012-10-16 01:41:41 +02:00
assert(tx_type != DCT_DCT);
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
// load and convert residual array into floating-point
for(j = 0; j < tx_dim; j++) {
for(i = 0; i < tx_dim; i++) {
pfa[i] = (float)ip[i];
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
}
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
pfa += tx_dim;
ip += tx_dim;
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
}
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
// vertical transformation
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
pfa = &bufa[0];
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
pfb = &bufb[0];
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
switch(tx_type) {
case ADST_ADST :
case ADST_DCT :
ptv = (tx_dim == 4) ? &iadst_4[0] :
((tx_dim == 8) ? &iadst_8[0] : &iadst_16[0]);
break;
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
default :
ptv = (tx_dim == 4) ? &idct_4[0] :
((tx_dim == 8) ? &idct_8[0] : &idct_16[0]);
break;
}
for(j = 0; j < tx_dim; j++) {
for(i = 0; i < tx_dim; i++) {
pfb[i] = 0 ;
for(k = 0; k < tx_dim; k++) {
pfb[i] += ptv[k] * pfa[(k * tx_dim)];
}
pfa += 1;
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
}
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
pfb += tx_dim;
ptv += tx_dim;
pfa = &bufa[0];
}
// horizontal transformation
pfa = &bufa[0];
pfb = &bufb[0];
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
switch(tx_type) {
case ADST_ADST :
case DCT_ADST :
hybrid transform of 16x16 dimension Enable ADST/DCT of dimension 16x16 for I16X16 modes. This change provides benefits mostly for hd sequences. Set up the framework for selectable transform dimension. Also allowing quantization parameter threshold to control the use of hybrid transform (This is currently disabled by setting threshold always above the quantization parameter. Adaptive thresholding can be built upon this, which will further improve the coding performance.) The coding performance gains (with respect to the codec that has all other configuration settings turned on) are derf: 0.013 yt: 0.086 hd: 0.198 std-hd: 0.501 Change-Id: Ibb4263a61fc74e0b3c345f54d73e8c73552bf926
2012-08-29 20:25:38 +02:00
pth = (tx_dim == 4) ? &iadst_4[0] :
((tx_dim == 8) ? &iadst_8[0] : &iadst_16[0]);
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
break;
default :
hybrid transform of 16x16 dimension Enable ADST/DCT of dimension 16x16 for I16X16 modes. This change provides benefits mostly for hd sequences. Set up the framework for selectable transform dimension. Also allowing quantization parameter threshold to control the use of hybrid transform (This is currently disabled by setting threshold always above the quantization parameter. Adaptive thresholding can be built upon this, which will further improve the coding performance.) The coding performance gains (with respect to the codec that has all other configuration settings turned on) are derf: 0.013 yt: 0.086 hd: 0.198 std-hd: 0.501 Change-Id: Ibb4263a61fc74e0b3c345f54d73e8c73552bf926
2012-08-29 20:25:38 +02:00
pth = (tx_dim == 4) ? &idct_4[0] :
((tx_dim == 8) ? &idct_8[0] : &idct_16[0]);
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
break;
}
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
for(j = 0; j < tx_dim; j++) {
for(i = 0; i < tx_dim; i++) {
pfa[i] = 0;
for(k = 0; k < tx_dim; k++) {
pfa[i] += pfb[k] * pth[k];
}
pth += tx_dim;
}
pfa += tx_dim;
pfb += tx_dim;
switch(tx_type) {
case ADST_ADST :
case DCT_ADST :
pth = (tx_dim == 4) ? &iadst_4[0] :
((tx_dim == 8) ? &iadst_8[0] : &iadst_16[0]);
break;
default :
pth = (tx_dim == 4) ? &idct_4[0] :
((tx_dim == 8) ? &idct_8[0] : &idct_16[0]);
break;
}
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
}
Refactoring hybrid transform coding The forward and inverse hybrid transforms are now performed using single function modules, where the dimension is sent as argument. Added an inline function clip8b to clip the reconstruction pixels into range of 0-255. Change-Id: Id7d870b3e1aefc092721c80c0af6f641eb5f3747
2012-08-06 23:48:11 +02:00
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
// convert to short integer format and load BLOCKD buffer
op = output;
pfa = &bufa[0];
for(j = 0; j < tx_dim; j++) {
for(i = 0; i < tx_dim; i++) {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
op[i] = (pfa[i] > 0 ) ? (int16_t)( pfa[i] / 8 + 0.49) :
-(int16_t)( - pfa[i] / 8 + 0.49);
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
}
Fixed style warnings Fixed some style warnings. Change-Id: I9eb77c3bf6d63628feef1b0c936e905bf4f2089d
2012-10-31 01:54:22 +01:00
op += shortpitch;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
pfa += tx_dim;
}
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
}
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
Replacing the 8x8 DCT with 8x8 ADST/DCT for I8x8 Fixed the code review comments. Under the htrans8x8 experiment the 8X8 DCT in the I8X8 mode is replaced with a combination of 8X8 ADST and DCT. Overall coding gains with the htrans8x8 experiment are: derf: 0.486 std-hd: 1.040 hd: 1.063 yt: 0.506 Note that part of the gain comes from bigger transforms (8x8 instead of 4x4) and part comes from replacing the DCT wth the ADST. Change-Id: I92ca6bbfce11b4165d612b81d9adfad4d010c775
2012-08-02 18:07:33 +02:00
}
Adds hybrid transform Adds ADST/DCT hybrid transform coding for Intra4x4 mode. The ADST is applied to directions in which the boundary pixels are used for prediction, while DCT applied to directions without corresponding boundary prediction. Adds enum TX_TYPE in b_mode_infor to indicate the transform type used. Make coding style consistent with google style. Fixed the commented issues. Experimental results in terms of bit-rate reduction: derf: 0.731% yt: 0.982% std-hd: 0.459% hd: 0.725% Will be looking at 8x8 transforms next. Change-Id: I46dbd7b80dbb3e8856e9c34fbc58cb3764a12fcf
2012-06-25 21:26:09 +02:00
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
/* Converted the transforms to integer form. */
Fixed style warnings Fixed some style warnings. Change-Id: I9eb77c3bf6d63628feef1b0c936e905bf4f2089d
2012-10-31 01:54:22 +01:00
#define VERTICAL_SHIFT 14 // 16
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
#define VERTICAL_ROUNDING ((1 << (VERTICAL_SHIFT - 1)) - 1)
Fixed style warnings Fixed some style warnings. Change-Id: I9eb77c3bf6d63628feef1b0c936e905bf4f2089d
2012-10-31 01:54:22 +01:00
#define HORIZONTAL_SHIFT 17 // 15
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
#define HORIZONTAL_ROUNDING ((1 << (HORIZONTAL_SHIFT - 1)) - 1)
ihtllm moves to rtcd clears up some warnings Change-Id: I9899637497c6ad7519f098e055ab98580ae6d688
2012-11-29 16:19:38 +01:00
void vp9_ihtllm_c(const int16_t *input, int16_t *output, int pitch,
Adjust style to match Google Coding Style a little more closely. Most of these were picked up by jenkins in the commit that changed the vp8 namespace to vp9 in common/. Change-Id: I5cbd56ffc753b92ef805133cda6acc1713a13878
2012-11-01 00:09:17 +01:00
TX_TYPE tx_type, int tx_dim) {
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
int i, j, k;
int16_t imbuf[256];
const int16_t *ip = input;
int16_t *op = output;
int16_t *im = &imbuf[0];
/* pointers to vertical and horizontal transforms. */
const int16_t *ptv = NULL, *pth = NULL;
int shortpitch = pitch >> 1;
switch (tx_type) {
case ADST_ADST :
ptv = pth = (tx_dim == 4) ? &iadst_i4[0]
: ((tx_dim == 8) ? &iadst_i8[0]
: &iadst_i16[0]);
break;
case ADST_DCT :
ptv = (tx_dim == 4) ? &iadst_i4[0]
: ((tx_dim == 8) ? &iadst_i8[0] : &iadst_i16[0]);
pth = (tx_dim == 4) ? &idct_i4[0]
: ((tx_dim == 8) ? &idct_i8[0] : &idct_i16[0]);
break;
case DCT_ADST :
ptv = (tx_dim == 4) ? &idct_i4[0]
: ((tx_dim == 8) ? &idct_i8[0] : &idct_i16[0]);
pth = (tx_dim == 4) ? &iadst_i4[0]
: ((tx_dim == 8) ? &iadst_i8[0] : &iadst_i16[0]);
break;
case DCT_DCT :
ptv = pth = (tx_dim == 4) ? &idct_i4[0]
Fixed style warnings Fixed some style warnings. Change-Id: I9eb77c3bf6d63628feef1b0c936e905bf4f2089d
2012-10-31 01:54:22 +01:00
: ((tx_dim == 8) ? &idct_i8[0]
: &idct_i16[0]);
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
break;
default:
assert(0);
break;
}
/* vertical transformation */
for (j = 0; j < tx_dim; j++) {
for (i = 0; i < tx_dim; i++) {
int temp = 0;
for (k = 0; k < tx_dim; k++) {
temp += ptv[k] * ip[(k * tx_dim)];
}
im[i] = (int16_t)((temp + VERTICAL_ROUNDING) >> VERTICAL_SHIFT);
ip++;
}
Fixed style warnings Fixed some style warnings. Change-Id: I9eb77c3bf6d63628feef1b0c936e905bf4f2089d
2012-10-31 01:54:22 +01:00
im += tx_dim; // 16
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
ptv += tx_dim;
ip = input;
}
/* horizontal transformation */
im = &imbuf[0];
for (j = 0; j < tx_dim; j++) {
const int16_t *pthc = pth;
for (i = 0; i < tx_dim; i++) {
int temp = 0;
for (k = 0; k < tx_dim; k++) {
temp += im[k] * pthc[k];
}
op[i] = (int16_t)((temp + HORIZONTAL_ROUNDING) >> HORIZONTAL_SHIFT);
pthc += tx_dim;
}
Fixed style warnings Fixed some style warnings. Change-Id: I9eb77c3bf6d63628feef1b0c936e905bf4f2089d
2012-10-31 01:54:22 +01:00
im += tx_dim; // 16
Convert the transforms to integer forms. Converted the forward and inverse transforms to integer forms. Modify #define TEST_INT 1/0 in the code to call integer/float version of transforms. The tests showed that average OVERALL PSNR loss was less than 0.1%. Change-Id: I1dfa4eeab6412597e3b970ce299cf0e116a917e6
2012-10-19 01:31:59 +02:00
op += shortpitch;
}
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct4x4llm_c(int16_t *input, int16_t *output, int pitch) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
int a1, b1, c1, d1;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t *ip = input;
int16_t *op = output;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int temp1, temp2;
int shortpitch = pitch >> 1;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (i = 0; i < 4; i++) {
a1 = ip[0] + ip[8];
b1 = ip[0] - ip[8];
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
temp1 = (ip[4] * sinpi8sqrt2 + rounding) >> 16;
temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1 + rounding) >> 16);
c1 = temp1 - temp2;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
temp1 = ip[4] + ((ip[4] * cospi8sqrt2minus1 + rounding) >> 16);
temp2 = (ip[12] * sinpi8sqrt2 + rounding) >> 16;
d1 = temp1 + temp2;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
op[shortpitch * 0] = a1 + d1;
op[shortpitch * 3] = a1 - d1;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
op[shortpitch * 1] = b1 + c1;
op[shortpitch * 2] = b1 - c1;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
ip++;
op++;
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
ip = output;
op = output;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (i = 0; i < 4; i++) {
a1 = ip[0] + ip[2];
b1 = ip[0] - ip[2];
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
temp1 = (ip[1] * sinpi8sqrt2 + rounding) >> 16;
temp2 = ip[3] + ((ip[3] * cospi8sqrt2minus1 + rounding) >> 16);
c1 = temp1 - temp2;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
temp1 = ip[1] + ((ip[1] * cospi8sqrt2minus1 + rounding) >> 16);
temp2 = (ip[3] * sinpi8sqrt2 + rounding) >> 16;
d1 = temp1 + temp2;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
op[0] = (a1 + d1 + 16) >> 5;
op[3] = (a1 - d1 + 16) >> 5;
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
op[1] = (b1 + c1 + 16) >> 5;
op[2] = (b1 - c1 + 16) >> 5;
Initial WebM release
2010-05-18 17:58:33 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
ip += shortpitch;
op += shortpitch;
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct4x4llm_1_c(int16_t *input, int16_t *output, int pitch) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
int a1;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t *op = output;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int shortpitch = pitch >> 1;
a1 = ((input[0] + 16) >> 5);
for (i = 0; i < 4; i++) {
op[0] = a1;
op[1] = a1;
op[2] = a1;
op[3] = a1;
op += shortpitch;
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_dc_only_idct_add_c(int input_dc, uint8_t *pred_ptr,
uint8_t *dst_ptr, int pitch, int stride) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int a1 = ((input_dc + 16) >> 5);
int r, c;
Combine idct and reconstruction steps This moves the prediction step before the idct and combines the idct and reconstruction steps into a single step. Combining them seems to give an overall decoder performance improvement of about 1%. Change-Id: I90d8b167ec70d79c7ba2ee484106a78b3d16e318
2010-05-28 20:28:12 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (r = 0; r < 4; r++) {
for (c = 0; c < 4; c++) {
Consistently use get_prob(), clip_prob() and newly added clip_pixel(). Add a function clip_pixel() to clip a pixel value to the [0,255] range of allowed values, and use this where-ever appropriate (e.g. prediction, reconstruction). Likewise, consistently use the recently added function clip_prob(), which calculates a binary probability in the [1,255] range. If possible, try to use get_prob() or its sister get_binary_prob() to calculate binary probabilities, for consistency. Since in some places, this means that binary probability calculations are changed (we use {255,256}*count0/(total) in a range of places, and all of these are now changed to use 256*count0+(total>>1)/total), this changes the encoding result, so this patch warrants some extensive testing. Change-Id: Ibeeff8d886496839b8e0c0ace9ccc552351f7628
2012-12-10 21:09:07 +01:00
dst_ptr[c] = clip_pixel(a1 + pred_ptr[c]);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Combine idct and reconstruction steps This moves the prediction step before the idct and combines the idct and reconstruction steps into a single step. Combining them seems to give an overall decoder performance improvement of about 1%. Change-Id: I90d8b167ec70d79c7ba2ee484106a78b3d16e318
2010-05-28 20:28:12 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
dst_ptr += stride;
pred_ptr += pitch;
}
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_inv_walsh4x4_c(int16_t *input, int16_t *output) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
int a1, b1, c1, d1;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t *ip = input;
int16_t *op = output;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (i = 0; i < 4; i++) {
a1 = ((ip[0] + ip[3]));
b1 = ((ip[1] + ip[2]));
c1 = ((ip[1] - ip[2]));
d1 = ((ip[0] - ip[3]));
op[0] = (a1 + b1 + 1) >> 1;
op[1] = (c1 + d1) >> 1;
op[2] = (a1 - b1) >> 1;
op[3] = (d1 - c1) >> 1;
ip += 4;
op += 4;
}
ip = output;
op = output;
for (i = 0; i < 4; i++) {
a1 = ip[0] + ip[12];
b1 = ip[4] + ip[8];
c1 = ip[4] - ip[8];
d1 = ip[0] - ip[12];
op[0] = (a1 + b1 + 1) >> 1;
op[4] = (c1 + d1) >> 1;
op[8] = (a1 - b1) >> 1;
op[12] = (d1 - c1) >> 1;
ip++;
op++;
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_inv_walsh4x4_1_c(int16_t *in, int16_t *out) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t tmp[4];
int16_t *ip = in;
int16_t *op = tmp;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
op[0] = (ip[0] + 1) >> 1;
op[1] = op[2] = op[3] = (ip[0] >> 1);
ip = tmp;
op = out;
for (i = 0; i < 4; i++) {
op[0] = (ip[0] + 1) >> 1;
op[4] = op[8] = op[12] = (ip[0] >> 1);
ip++;
op++;
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
#if CONFIG_LOSSLESS
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_inv_walsh4x4_lossless_c(int16_t *input, int16_t *output) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
int a1, b1, c1, d1;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t *ip = input;
int16_t *op = output;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (i = 0; i < 4; i++) {
a1 = ((ip[0] + ip[3])) >> Y2_WHT_UPSCALE_FACTOR;
b1 = ((ip[1] + ip[2])) >> Y2_WHT_UPSCALE_FACTOR;
c1 = ((ip[1] - ip[2])) >> Y2_WHT_UPSCALE_FACTOR;
d1 = ((ip[0] - ip[3])) >> Y2_WHT_UPSCALE_FACTOR;
op[0] = (a1 + b1 + 1) >> 1;
op[1] = (c1 + d1) >> 1;
op[2] = (a1 - b1) >> 1;
op[3] = (d1 - c1) >> 1;
ip += 4;
op += 4;
}
ip = output;
op = output;
for (i = 0; i < 4; i++) {
a1 = ip[0] + ip[12];
b1 = ip[4] + ip[8];
c1 = ip[4] - ip[8];
d1 = ip[0] - ip[12];
op[0] = ((a1 + b1 + 1) >> 1) << Y2_WHT_UPSCALE_FACTOR;
op[4] = ((c1 + d1) >> 1) << Y2_WHT_UPSCALE_FACTOR;
op[8] = ((a1 - b1) >> 1) << Y2_WHT_UPSCALE_FACTOR;
op[12] = ((d1 - c1) >> 1) << Y2_WHT_UPSCALE_FACTOR;
ip++;
op++;
}
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_inv_walsh4x4_1_lossless_c(int16_t *in, int16_t *out) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t tmp[4];
int16_t *ip = in;
int16_t *op = tmp;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
op[0] = ((ip[0] >> Y2_WHT_UPSCALE_FACTOR) + 1) >> 1;
op[1] = op[2] = op[3] = ((ip[0] >> Y2_WHT_UPSCALE_FACTOR) >> 1);
ip = tmp;
op = out;
for (i = 0; i < 4; i++) {
op[0] = ((ip[0] + 1) >> 1) << Y2_WHT_UPSCALE_FACTOR;
op[4] = op[8] = op[12] = ((ip[0] >> 1)) << Y2_WHT_UPSCALE_FACTOR;
ip++;
op++;
}
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_inv_walsh4x4_x8_c(int16_t *input, int16_t *output, int pitch) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
int a1, b1, c1, d1;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t *ip = input;
int16_t *op = output;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int shortpitch = pitch >> 1;
for (i = 0; i < 4; i++) {
a1 = ((ip[0] + ip[3])) >> WHT_UPSCALE_FACTOR;
b1 = ((ip[1] + ip[2])) >> WHT_UPSCALE_FACTOR;
c1 = ((ip[1] - ip[2])) >> WHT_UPSCALE_FACTOR;
d1 = ((ip[0] - ip[3])) >> WHT_UPSCALE_FACTOR;
op[0] = (a1 + b1 + 1) >> 1;
op[1] = (c1 + d1) >> 1;
op[2] = (a1 - b1) >> 1;
op[3] = (d1 - c1) >> 1;
ip += 4;
op += shortpitch;
}
ip = output;
op = output;
for (i = 0; i < 4; i++) {
a1 = ip[shortpitch * 0] + ip[shortpitch * 3];
b1 = ip[shortpitch * 1] + ip[shortpitch * 2];
c1 = ip[shortpitch * 1] - ip[shortpitch * 2];
d1 = ip[shortpitch * 0] - ip[shortpitch * 3];
op[shortpitch * 0] = (a1 + b1 + 1) >> 1;
op[shortpitch * 1] = (c1 + d1) >> 1;
op[shortpitch * 2] = (a1 - b1) >> 1;
op[shortpitch * 3] = (d1 - c1) >> 1;
ip++;
op++;
}
}
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_inv_walsh4x4_1_x8_c(int16_t *in, int16_t *out, int pitch) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t tmp[4];
int16_t *ip = in;
int16_t *op = tmp;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int shortpitch = pitch >> 1;
op[0] = ((ip[0] >> WHT_UPSCALE_FACTOR) + 1) >> 1;
op[1] = op[2] = op[3] = ((ip[0] >> WHT_UPSCALE_FACTOR) >> 1);
ip = tmp;
op = out;
for (i = 0; i < 4; i++) {
op[shortpitch * 0] = (ip[0] + 1) >> 1;
op[shortpitch * 1] = op[shortpitch * 2] = op[shortpitch * 3] = ip[0] >> 1;
ip++;
op++;
}
}
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_dc_only_inv_walsh_add_c(short input_dc, uint8_t *pred_ptr,
uint8_t *dst_ptr,
Adjust style to match Google Coding Style a little more closely. Most of these were picked up by jenkins in the commit that changed the vp8 namespace to vp9 in common/. Change-Id: I5cbd56ffc753b92ef805133cda6acc1713a13878
2012-11-01 00:09:17 +01:00
int pitch, int stride) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int r, c;
short tmp[16];
Change common vp8_ public symbol prefixes to vp9_. Change-Id: Ic5a5f60e1ff9d9ccae4174160d36529466eeb509
2012-10-31 00:25:53 +01:00
vp9_short_inv_walsh4x4_1_x8_c(&input_dc, tmp, 4 << 1);
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (r = 0; r < 4; r++) {
for (c = 0; c < 4; c++) {
Consistently use get_prob(), clip_prob() and newly added clip_pixel(). Add a function clip_pixel() to clip a pixel value to the [0,255] range of allowed values, and use this where-ever appropriate (e.g. prediction, reconstruction). Likewise, consistently use the recently added function clip_prob(), which calculates a binary probability in the [1,255] range. If possible, try to use get_prob() or its sister get_binary_prob() to calculate binary probabilities, for consistency. Since in some places, this means that binary probability calculations are changed (we use {255,256}*count0/(total) in a range of places, and all of these are now changed to use 256*count0+(total>>1)/total), this changes the encoding result, so this patch warrants some extensive testing. Change-Id: Ibeeff8d886496839b8e0c0ace9ccc552351f7628
2012-12-10 21:09:07 +01:00
dst_ptr[c] = clip_pixel(tmp[r * 4 + c] + pred_ptr[c]);
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
}
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
dst_ptr += stride;
pred_ptr += pitch;
}
Add lossless compression mode. This commit adds lossless compression capability to the experimental branch. The lossless experiment can be enabled using --enable-lossless in configure. When the experiment is enabled, the encoder will use lossless compression mode by command line option --lossless, and the decoder automatically recognizes a losslessly encoded clip and decodes accordingly. To achieve the lossless coding, this commit has changed the following: 1. To encode at lossless mode, encoder forces the use of unit quantizer, i.e, Q 0, where effective quantization is 1. Encoder also disables the usage of 8x8 transform and allows only 4x4 transform; 2. At Q 0, the first order 4x4 DCT/IDCT have been switched over to a pair of forward and inverse Walsh-Hadamard Transform (http://goo.gl/EIsfy), with proper scaling applied to match the range of the original 4x4 DCT/IDCT pair; 3. At Q 0, the second order remains to use the previous walsh-hadamard transform pair. However, to maintain the reversibility in second order transform at Q 0, scaling down is applied to first order DC coefficients prior to forward transform, and scaling up is applied to the second order output prior to quantization. Symmetric upscaling and downscaling are added around inverse second order transform; 4. At lossless mode, encoder also disables a number of minor features to ensure no loss is introduced, these features includes: a. Trellis quantization optimization b. Loop filtering c. Aggressive zero-binning, rounding and zero-bin boosting d. Mode based zero-bin boosting Lossless coding test was performed on all clips within the derf set, to verify that the commit has achieved lossless compression for all clips. The average compression ratio is around 2.57 to 1. (http://goo.gl/dEShs) Change-Id: Ia3aba7dd09df40dd590f93b9aba134defbc64e34
2012-06-14 04:03:31 +02:00
}
#endif
Merge t8x8 experiments Change-Id: I8e9b6b154e1a0d0cb42d596366380d69c00ac15f
2012-02-29 02:11:12 +01:00
Change common vp8_ public symbol prefixes to vp9_. Change-Id: Ic5a5f60e1ff9d9ccae4174160d36529466eeb509
2012-10-31 00:25:53 +01:00
void vp9_dc_only_idct_add_8x8_c(short input_dc,
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
uint8_t *pred_ptr,
uint8_t *dst_ptr,
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int pitch, int stride) {
int a1 = ((input_dc + 16) >> 5);
int r, c, b;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
uint8_t *orig_pred = pred_ptr;
uint8_t *orig_dst = dst_ptr;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (b = 0; b < 4; b++) {
for (r = 0; r < 4; r++) {
for (c = 0; c < 4; c++) {
Consistently use get_prob(), clip_prob() and newly added clip_pixel(). Add a function clip_pixel() to clip a pixel value to the [0,255] range of allowed values, and use this where-ever appropriate (e.g. prediction, reconstruction). Likewise, consistently use the recently added function clip_prob(), which calculates a binary probability in the [1,255] range. If possible, try to use get_prob() or its sister get_binary_prob() to calculate binary probabilities, for consistency. Since in some places, this means that binary probability calculations are changed (we use {255,256}*count0/(total) in a range of places, and all of these are now changed to use 256*count0+(total>>1)/total), this changes the encoding result, so this patch warrants some extensive testing. Change-Id: Ibeeff8d886496839b8e0c0ace9ccc552351f7628
2012-12-10 21:09:07 +01:00
dst_ptr[c] = clip_pixel(a1 + pred_ptr[c]);
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
}
dst_ptr += stride;
pred_ptr += pitch;
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
dst_ptr = orig_dst + (b + 1) % 2 * 4 + (b + 1) / 2 * 4 * stride;
pred_ptr = orig_pred + (b + 1) % 2 * 4 + (b + 1) / 2 * 4 * pitch;
}
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
#define W1 2841 /* 2048*sqrt(2)*cos(1*pi/16) */
#define W2 2676 /* 2048*sqrt(2)*cos(2*pi/16) */
#define W3 2408 /* 2048*sqrt(2)*cos(3*pi/16) */
#define W5 1609 /* 2048*sqrt(2)*cos(5*pi/16) */
#define W6 1108 /* 2048*sqrt(2)*cos(6*pi/16) */
#define W7 565 /* 2048*sqrt(2)*cos(7*pi/16) */
/* row (horizontal) IDCT
*
* 7 pi 1 dst[k] = sum c[l] * src[l] * cos( -- *
* ( k + - ) * l ) l=0 8 2
*
* where: c[0] = 128 c[1..7] = 128*sqrt(2) */
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
static void idctrow(int *blk) {
int x0, x1, x2, x3, x4, x5, x6, x7, x8;
/* shortcut */
if (!((x1 = blk[4] << 11) | (x2 = blk[6]) | (x3 = blk[2]) |
(x4 = blk[1]) | (x5 = blk[7]) | (x6 = blk[5]) | (x7 = blk[3]))) {
blk[0] = blk[1] = blk[2] = blk[3] = blk[4]
= blk[5] = blk[6] = blk[7] = blk[0] << 3;
return;
}
x0 = (blk[0] << 11) + 128; /* for proper rounding in the fourth stage */
/* first stage */
x8 = W7 * (x4 + x5);
x4 = x8 + (W1 - W7) * x4;
x5 = x8 - (W1 + W7) * x5;
x8 = W3 * (x6 + x7);
x6 = x8 - (W3 - W5) * x6;
x7 = x8 - (W3 + W5) * x7;
/* second stage */
x8 = x0 + x1;
x0 -= x1;
x1 = W6 * (x3 + x2);
x2 = x1 - (W2 + W6) * x2;
x3 = x1 + (W2 - W6) * x3;
x1 = x4 + x6;
x4 -= x6;
x6 = x5 + x7;
x5 -= x7;
/* third stage */
x7 = x8 + x3;
x8 -= x3;
x3 = x0 + x2;
x0 -= x2;
x2 = (181 * (x4 + x5) + 128) >> 8;
x4 = (181 * (x4 - x5) + 128) >> 8;
/* fourth stage */
blk[0] = (x7 + x1) >> 8;
blk[1] = (x3 + x2) >> 8;
blk[2] = (x0 + x4) >> 8;
blk[3] = (x8 + x6) >> 8;
blk[4] = (x8 - x6) >> 8;
blk[5] = (x0 - x4) >> 8;
blk[6] = (x3 - x2) >> 8;
blk[7] = (x7 - x1) >> 8;
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
/* column (vertical) IDCT
*
* 7 pi 1 dst[8*k] = sum c[l] * src[8*l] *
* cos( -- * ( k + - ) * l ) l=0 8 2
*
* where: c[0] = 1/1024 c[1..7] = (1/1024)*sqrt(2) */
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
static void idctcol(int *blk) {
int x0, x1, x2, x3, x4, x5, x6, x7, x8;
a bit code clean-up Removed some transform code that is not in use. Change-Id: I9489af7e23d9d7fe052feb6c8bbafa62ebbda39c
2012-02-16 22:41:17 +01:00
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
/* shortcut */
if (!((x1 = (blk[8 * 4] << 8)) | (x2 = blk[8 * 6]) | (x3 = blk[8 * 2]) |
a bit code clean-up Removed some transform code that is not in use. Change-Id: I9489af7e23d9d7fe052feb6c8bbafa62ebbda39c
2012-02-16 22:41:17 +01:00
(x4 = blk[8 * 1]) | (x5 = blk[8 * 7]) | (x6 = blk[8 * 5]) |
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
(x7 = blk[8 * 3]))) {
blk[8 * 0] = blk[8 * 1] = blk[8 * 2] = blk[8 * 3]
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
= blk[8 * 4] = blk[8 * 5] = blk[8 * 6]
= blk[8 * 7] = ((blk[8 * 0] + 32) >> 6);
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
return;
}
x0 = (blk[8 * 0] << 8) + 16384;
/* first stage */
x8 = W7 * (x4 + x5) + 4;
x4 = (x8 + (W1 - W7) * x4) >> 3;
x5 = (x8 - (W1 + W7) * x5) >> 3;
x8 = W3 * (x6 + x7) + 4;
x6 = (x8 - (W3 - W5) * x6) >> 3;
x7 = (x8 - (W3 + W5) * x7) >> 3;
/* second stage */
x8 = x0 + x1;
x0 -= x1;
x1 = W6 * (x3 + x2) + 4;
x2 = (x1 - (W2 + W6) * x2) >> 3;
x3 = (x1 + (W2 - W6) * x3) >> 3;
x1 = x4 + x6;
x4 -= x6;
x6 = x5 + x7;
x5 -= x7;
/* third stage */
x7 = x8 + x3;
x8 -= x3;
x3 = x0 + x2;
x0 -= x2;
x2 = (181 * (x4 + x5) + 128) >> 8;
x4 = (181 * (x4 - x5) + 128) >> 8;
/* fourth stage */
blk[8 * 0] = (x7 + x1) >> 14;
blk[8 * 1] = (x3 + x2) >> 14;
blk[8 * 2] = (x0 + x4) >> 14;
blk[8 * 3] = (x8 + x6) >> 14;
blk[8 * 4] = (x8 - x6) >> 14;
blk[8 * 5] = (x0 - x4) >> 14;
blk[8 * 6] = (x3 - x2) >> 14;
blk[8 * 7] = (x7 - x1) >> 14;
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
#define TX_DIM 8
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct8x8_c(int16_t *coefs, int16_t *block, int pitch) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int X[TX_DIM * TX_DIM];
int i, j;
int shortpitch = pitch >> 1;
for (i = 0; i < TX_DIM; i++) {
for (j = 0; j < TX_DIM; j++) {
X[i * TX_DIM + j] = (int)(coefs[i * TX_DIM + j] + 1
+ (coefs[i * TX_DIM + j] < 0)) >> 2;
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
}
for (i = 0; i < 8; i++)
idctrow(X + 8 * i);
for (i = 0; i < 8; i++)
idctcol(X + i);
for (i = 0; i < TX_DIM; i++) {
for (j = 0; j < TX_DIM; j++) {
block[i * shortpitch + j] = X[i * TX_DIM + j] >> 1;
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
}
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
Optimize 8x8 dequant and idct Similar to 16x16 dequant and idct, based on the value of eobs, the 8x8 dequant and idct calculation was simplified to improve decorder performance. Combined vp9_dequant_idct_add_8x8 and vp9_dequant_dc_idct_add_8x8 to eliminate duplicate code. Change-Id: Ia58e50ab27f7012b7379c495837c9c0b5ba9cf7f
2012-11-10 02:50:13 +01:00
/* Row IDCT when only first 4 coefficients are non-zero. */
static void idctrow10(int *blk) {
int x0, x1, x2, x3, x4, x5, x6, x7, x8;
/* shortcut */
if (!((x1 = blk[4] << 11) | (x2 = blk[6]) | (x3 = blk[2]) |
(x4 = blk[1]) | (x5 = blk[7]) | (x6 = blk[5]) | (x7 = blk[3]))) {
blk[0] = blk[1] = blk[2] = blk[3] = blk[4]
= blk[5] = blk[6] = blk[7] = blk[0] << 3;
return;
}
x0 = (blk[0] << 11) + 128; /* for proper rounding in the fourth stage */
/* first stage */
x5 = W7 * x4;
x4 = W1 * x4;
x6 = W3 * x7;
x7 = -W5 * x7;
/* second stage */
x2 = W6 * x3;
x3 = W2 * x3;
x1 = x4 + x6;
x4 -= x6;
x6 = x5 + x7;
x5 -= x7;
/* third stage */
x7 = x0 + x3;
x8 = x0 - x3;
x3 = x0 + x2;
x0 -= x2;
x2 = (181 * (x4 + x5) + 128) >> 8;
x4 = (181 * (x4 - x5) + 128) >> 8;
/* fourth stage */
blk[0] = (x7 + x1) >> 8;
blk[1] = (x3 + x2) >> 8;
blk[2] = (x0 + x4) >> 8;
blk[3] = (x8 + x6) >> 8;
blk[4] = (x8 - x6) >> 8;
blk[5] = (x0 - x4) >> 8;
blk[6] = (x3 - x2) >> 8;
blk[7] = (x7 - x1) >> 8;
}
/* Column (vertical) IDCT when only first 4 coefficients are non-zero. */
static void idctcol10(int *blk) {
int x0, x1, x2, x3, x4, x5, x6, x7, x8;
/* shortcut */
if (!((x1 = (blk[8 * 4] << 8)) | (x2 = blk[8 * 6]) | (x3 = blk[8 * 2]) |
(x4 = blk[8 * 1]) | (x5 = blk[8 * 7]) | (x6 = blk[8 * 5]) |
(x7 = blk[8 * 3]))) {
blk[8 * 0] = blk[8 * 1] = blk[8 * 2] = blk[8 * 3]
= blk[8 * 4] = blk[8 * 5] = blk[8 * 6]
= blk[8 * 7] = ((blk[8 * 0] + 32) >> 6);
return;
}
x0 = (blk[8 * 0] << 8) + 16384;
/* first stage */
x5 = (W7 * x4 + 4) >> 3;
x4 = (W1 * x4 + 4) >> 3;
x6 = (W3 * x7 + 4) >> 3;
x7 = (-W5 * x7 + 4) >> 3;
/* second stage */
x2 = (W6 * x3 + 4) >> 3;
x3 = (W2 * x3 + 4) >> 3;
x1 = x4 + x6;
x4 -= x6;
x6 = x5 + x7;
x5 -= x7;
/* third stage */
x7 = x0 + x3;
x8 = x0 - x3;
x3 = x0 + x2;
x0 -= x2;
x2 = (181 * (x4 + x5) + 128) >> 8;
x4 = (181 * (x4 - x5) + 128) >> 8;
/* fourth stage */
blk[8 * 0] = (x7 + x1) >> 14;
blk[8 * 1] = (x3 + x2) >> 14;
blk[8 * 2] = (x0 + x4) >> 14;
blk[8 * 3] = (x8 + x6) >> 14;
blk[8 * 4] = (x8 - x6) >> 14;
blk[8 * 5] = (x0 - x4) >> 14;
blk[8 * 6] = (x3 - x2) >> 14;
blk[8 * 7] = (x7 - x1) >> 14;
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct10_8x8_c(int16_t *coefs, int16_t *block, int pitch) {
Optimize 8x8 dequant and idct Similar to 16x16 dequant and idct, based on the value of eobs, the 8x8 dequant and idct calculation was simplified to improve decorder performance. Combined vp9_dequant_idct_add_8x8 and vp9_dequant_dc_idct_add_8x8 to eliminate duplicate code. Change-Id: Ia58e50ab27f7012b7379c495837c9c0b5ba9cf7f
2012-11-10 02:50:13 +01:00
int X[TX_DIM * TX_DIM];
int i, j;
int shortpitch = pitch >> 1;
for (i = 0; i < TX_DIM; i++) {
for (j = 0; j < TX_DIM; j++) {
X[i * TX_DIM + j] = (int)(coefs[i * TX_DIM + j] + 1
+ (coefs[i * TX_DIM + j] < 0)) >> 2;
}
}
/* Do first 4 row idct only since non-zero dct coefficients are all in
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
* upper-left 4x4 area. */
Optimize 8x8 dequant and idct Similar to 16x16 dequant and idct, based on the value of eobs, the 8x8 dequant and idct calculation was simplified to improve decorder performance. Combined vp9_dequant_idct_add_8x8 and vp9_dequant_dc_idct_add_8x8 to eliminate duplicate code. Change-Id: Ia58e50ab27f7012b7379c495837c9c0b5ba9cf7f
2012-11-10 02:50:13 +01:00
for (i = 0; i < 4; i++)
idctrow10(X + 8 * i);
for (i = 0; i < 8; i++)
idctcol10(X + i);
for (i = 0; i < TX_DIM; i++) {
for (j = 0; j < TX_DIM; j++) {
block[i * shortpitch + j] = X[i * TX_DIM + j] >> 1;
}
}
}
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_ihaar2x2_c(int16_t *input, int16_t *output, int pitch) {
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
int i;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t *ip = input; // 0, 1, 4, 8
int16_t *op = output;
Restyle code Approximate the Google style guide[1] so that that there's a written document to follow and tools to check compliance[2]. [1]: http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml [2]: http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py Change-Id: Idf40e3d8dddcc72150f6af127b13e5dab838685f
2012-07-14 00:21:29 +02:00
for (i = 0; i < 16; i++) {
op[i] = 0;
}
op[0] = (ip[0] + ip[1] + ip[4] + ip[8] + 1) >> 1;
op[1] = (ip[0] - ip[1] + ip[4] - ip[8]) >> 1;
op[4] = (ip[0] + ip[1] - ip[4] - ip[8]) >> 1;
op[8] = (ip[0] - ip[1] - ip[4] + ip[8]) >> 1;
Add 8x8 transform to experimental branch Please refer to previous commit messages for detailed info: https://on2-git.corp.google.com/g/#change,5940 https://on2-git.corp.google.com/g/#change,6045 Change-Id: I8b16992f2f69c5a808ad40a3e32ef589cce7c59d
2011-02-14 23:18:18 +01:00
}
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
#if 0
// Keep a really bad float version as reference for now.
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct16x16_c(int16_t *input, int16_t *output, int pitch) {
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
{
double x;
const int short_pitch = pitch >> 1;
int i, j, k, l;
for (l = 0; l < 16; ++l) {
for (k = 0; k < 16; ++k) {
double s = 0;
for (i = 0; i < 16; ++i) {
for (j = 0; j < 16; ++j) {
x=cos(PI*j*(l+0.5)/16.0)*cos(PI*i*(k+0.5)/16.0)*input[i*16+j]/32;
if (i != 0)
x *= sqrt(2.0);
if (j != 0)
x *= sqrt(2.0);
s += x;
}
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
output[k*short_pitch+l] = (short)round(s);
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
}
}
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
#endif
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
#define TEST_INT_16x16_IDCT 1
#if !TEST_INT_16x16_IDCT
A bit of temporary relief on encoder slowness The commit replaces run-time initialization of cosine constants with static constant values, which provides ~30% relief on slow speed. The real solution, however will be to implement integer versions of those functions that current use float/double. Change-Id: Ie3ff1793509653d78dd1aeaf88cc6737da1bc55f
2012-08-07 22:55:49 +02:00
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
static void butterfly_16x16_idct_1d(double input[16], double output[16]) {
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
static const double C1 = 0.995184726672197;
static const double C2 = 0.98078528040323;
static const double C3 = 0.956940335732209;
static const double C4 = 0.923879532511287;
static const double C5 = 0.881921264348355;
static const double C6 = 0.831469612302545;
static const double C7 = 0.773010453362737;
static const double C8 = 0.707106781186548;
static const double C9 = 0.634393284163646;
static const double C10 = 0.555570233019602;
static const double C11 = 0.471396736825998;
static const double C12 = 0.38268343236509;
static const double C13 = 0.290284677254462;
static const double C14 = 0.195090322016128;
static const double C15 = 0.098017140329561;
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
{
double step[16];
double intermediate[16];
double temp1, temp2;
// step 1 and 2
step[ 0] = input[0] + input[8];
step[ 1] = input[0] - input[8];
temp1 = input[4]*C12;
temp2 = input[12]*C4;
temp1 -= temp2;
temp1 *= C8;
step[ 2] = 2*(temp1);
temp1 = input[4]*C4;
temp2 = input[12]*C12;
temp1 += temp2;
temp1 = (temp1);
temp1 *= C8;
step[ 3] = 2*(temp1);
temp1 = input[2]*C8;
temp1 = 2*(temp1);
temp2 = input[6] + input[10];
step[ 4] = temp1 + temp2;
step[ 5] = temp1 - temp2;
temp1 = input[14]*C8;
temp1 = 2*(temp1);
temp2 = input[6] - input[10];
step[ 6] = temp2 - temp1;
step[ 7] = temp2 + temp1;
// for odd input
temp1 = input[3]*C12;
temp2 = input[13]*C4;
temp1 += temp2;
temp1 = (temp1);
temp1 *= C8;
intermediate[ 8] = 2*(temp1);
temp1 = input[3]*C4;
temp2 = input[13]*C12;
temp2 -= temp1;
temp2 = (temp2);
temp2 *= C8;
intermediate[ 9] = 2*(temp2);
intermediate[10] = 2*(input[9]*C8);
intermediate[11] = input[15] - input[1];
intermediate[12] = input[15] + input[1];
intermediate[13] = 2*((input[7]*C8));
temp1 = input[11]*C12;
temp2 = input[5]*C4;
temp2 -= temp1;
temp2 = (temp2);
temp2 *= C8;
intermediate[14] = 2*(temp2);
temp1 = input[11]*C4;
temp2 = input[5]*C12;
temp1 += temp2;
temp1 = (temp1);
temp1 *= C8;
intermediate[15] = 2*(temp1);
step[ 8] = intermediate[ 8] + intermediate[14];
step[ 9] = intermediate[ 9] + intermediate[15];
step[10] = intermediate[10] + intermediate[11];
step[11] = intermediate[10] - intermediate[11];
step[12] = intermediate[12] + intermediate[13];
step[13] = intermediate[12] - intermediate[13];
step[14] = intermediate[ 8] - intermediate[14];
step[15] = intermediate[ 9] - intermediate[15];
// step 3
output[0] = step[ 0] + step[ 3];
output[1] = step[ 1] + step[ 2];
output[2] = step[ 1] - step[ 2];
output[3] = step[ 0] - step[ 3];
temp1 = step[ 4]*C14;
temp2 = step[ 7]*C2;
temp1 -= temp2;
output[4] = (temp1);
temp1 = step[ 4]*C2;
temp2 = step[ 7]*C14;
temp1 += temp2;
output[7] = (temp1);
temp1 = step[ 5]*C10;
temp2 = step[ 6]*C6;
temp1 -= temp2;
output[5] = (temp1);
temp1 = step[ 5]*C6;
temp2 = step[ 6]*C10;
temp1 += temp2;
output[6] = (temp1);
output[8] = step[ 8] + step[11];
output[9] = step[ 9] + step[10];
output[10] = step[ 9] - step[10];
output[11] = step[ 8] - step[11];
output[12] = step[12] + step[15];
output[13] = step[13] + step[14];
output[14] = step[13] - step[14];
output[15] = step[12] - step[15];
// output 4
step[ 0] = output[0] + output[7];
step[ 1] = output[1] + output[6];
step[ 2] = output[2] + output[5];
step[ 3] = output[3] + output[4];
step[ 4] = output[3] - output[4];
step[ 5] = output[2] - output[5];
step[ 6] = output[1] - output[6];
step[ 7] = output[0] - output[7];
temp1 = output[8]*C7;
temp2 = output[15]*C9;
temp1 -= temp2;
step[ 8] = (temp1);
temp1 = output[9]*C11;
temp2 = output[14]*C5;
temp1 += temp2;
step[ 9] = (temp1);
temp1 = output[10]*C3;
temp2 = output[13]*C13;
temp1 -= temp2;
step[10] = (temp1);
temp1 = output[11]*C15;
temp2 = output[12]*C1;
temp1 += temp2;
step[11] = (temp1);
temp1 = output[11]*C1;
temp2 = output[12]*C15;
temp2 -= temp1;
step[12] = (temp2);
temp1 = output[10]*C13;
temp2 = output[13]*C3;
temp1 += temp2;
step[13] = (temp1);
temp1 = output[9]*C5;
temp2 = output[14]*C11;
temp2 -= temp1;
step[14] = (temp2);
temp1 = output[8]*C9;
temp2 = output[15]*C7;
temp1 += temp2;
step[15] = (temp1);
// step 5
output[0] = (step[0] + step[15]);
output[1] = (step[1] + step[14]);
output[2] = (step[2] + step[13]);
output[3] = (step[3] + step[12]);
output[4] = (step[4] + step[11]);
output[5] = (step[5] + step[10]);
output[6] = (step[6] + step[ 9]);
output[7] = (step[7] + step[ 8]);
output[15] = (step[0] - step[15]);
output[14] = (step[1] - step[14]);
output[13] = (step[2] - step[13]);
output[12] = (step[3] - step[12]);
output[11] = (step[4] - step[11]);
output[10] = (step[5] - step[10]);
output[9] = (step[6] - step[ 9]);
output[8] = (step[7] - step[ 8]);
}
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
// Remove once an int version of iDCT is written
#if 0
void reference_16x16_idct_1d(double input[16], double output[16]) {
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
{
const double kPi = 3.141592653589793238462643383279502884;
const double kSqrt2 = 1.414213562373095048801688724209698;
for (int k = 0; k < 16; k++) {
output[k] = 0.0;
for (int n = 0; n < 16; n++) {
output[k] += input[n]*cos(kPi*(2*k+1)*n/32.0);
if (n == 0)
output[k] = output[k]/kSqrt2;
}
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
}
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
#endif
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct16x16_c(int16_t *input, int16_t *output, int pitch) {
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
Fix SIMD unsafe use of floating point. This commit fixes unsafe simd / floating point interactions arising from the current hybrid and 16x16 transform implementation. These led to a raft of bugs and issues when the project was built using VS2008 for Win32 though they did not show up with the unix builds. Gerrit makes a meal out of presenting the fix but all I have actually done is indent the body of each function that uses floating point by one level and bracket with emms instructions using the function vp8_clear_system_state(). See below. function () { vp8_clear_system_state(); { ... function body } vp8_clear_system_state(); } This is almost certainly over the top in terms of number of emms instructions but is a temporary measure pending implementation of integer variants of each function to replace the floating point. Limited testing suggests that this fixes the problems that arose for Win32 VS2008 when the hybrid or 16x16 transforms were enabled. Change-Id: I7c9a72bd79315246ed880578dec51e2b7c178442
2012-10-05 12:16:46 +02:00
{
double out[16*16], out2[16*16];
const int short_pitch = pitch >> 1;
int i, j;
// First transform rows
for (i = 0; i < 16; ++i) {
double temp_in[16], temp_out[16];
for (j = 0; j < 16; ++j)
temp_in[j] = input[j + i*short_pitch];
butterfly_16x16_idct_1d(temp_in, temp_out);
for (j = 0; j < 16; ++j)
out[j + i*16] = temp_out[j];
}
// Then transform columns
for (i = 0; i < 16; ++i) {
double temp_in[16], temp_out[16];
for (j = 0; j < 16; ++j)
temp_in[j] = out[j*16 + i];
butterfly_16x16_idct_1d(temp_in, temp_out);
for (j = 0; j < 16; ++j)
out2[j*16 + i] = temp_out[j];
}
for (i = 0; i < 16*16; ++i)
output[i] = round(out2[i]/128);
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
Remove vp8 in local symbols. For non-static functions, change the prefix to vp9_. For static functions, remove the prefix. Also fix some comments, remove unused code or unused function prototypes. Change-Id: I1f8be05362f66060fe421c3d4c9a906fdf835de5
2012-10-31 22:40:53 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
16x16 DCT blocks. Set on all 16x16 intra/inter modes Features: - Butterfly fDCT/iDCT - Loop filter does not filter internal edges with 16x16 - Optimize coefficient function - Update coefficient probability function - RD - Entropy stats - 16x16 is a config option Have not tested with experiments. hd: 2.60% std-hd: 2.43% yt: 1.32% derf: 0.60% Change-Id: I96fb090517c30c5da84bad4fae602c3ec0c58b1c
2012-08-03 02:03:14 +02:00
}
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
#else
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
#define INITIAL_SHIFT 2
#define INITIAL_ROUNDING (1 << (INITIAL_SHIFT - 1))
#define RIGHT_SHIFT 14
#define RIGHT_ROUNDING (1 << (RIGHT_SHIFT - 1))
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
static const int16_t C1 = 16305;
static const int16_t C2 = 16069;
static const int16_t C3 = 15679;
static const int16_t C4 = 15137;
static const int16_t C5 = 14449;
static const int16_t C6 = 13623;
static const int16_t C7 = 12665;
static const int16_t C8 = 11585;
static const int16_t C9 = 10394;
static const int16_t C10 = 9102;
static const int16_t C11 = 7723;
static const int16_t C12 = 6270;
static const int16_t C13 = 4756;
static const int16_t C14 = 3196;
static const int16_t C15 = 1606;
static void butterfly_16x16_idct_1d(int16_t input[16], int16_t output[16],
int last_shift_bits) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
int16_t step[16];
int intermediate[16];
int temp1, temp2;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
int step1_shift = RIGHT_SHIFT + INITIAL_SHIFT;
int step1_rounding = 1 << (step1_shift - 1);
int last_rounding = 0;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
if (last_shift_bits > 0)
last_rounding = 1 << (last_shift_bits - 1);
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// step 1 and 2
step[ 0] = (input[0] + input[8] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[ 1] = (input[0] - input[8] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
temp1 = input[4] * C12;
temp2 = input[12] * C4;
temp1 = (temp1 - temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 *= C8;
step[ 2] = (2 * (temp1) + step1_rounding) >> step1_shift;
temp1 = input[4] * C4;
temp2 = input[12] * C12;
temp1 = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 *= C8;
step[ 3] = (2 * (temp1) + step1_rounding) >> step1_shift;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
temp1 = input[2] * C8;
temp1 = (2 * (temp1) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp2 = input[6] + input[10];
step[ 4] = (temp1 + temp2 + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[ 5] = (temp1 - temp2 + INITIAL_ROUNDING) >> INITIAL_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
temp1 = input[14] * C8;
temp1 = (2 * (temp1) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp2 = input[6] - input[10];
step[ 6] = (temp2 - temp1 + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[ 7] = (temp2 + temp1 + INITIAL_ROUNDING) >> INITIAL_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// for odd input
temp1 = input[3] * C12;
temp2 = input[13] * C4;
temp1 = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 *= C8;
intermediate[ 8] = (2 * (temp1) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
temp1 = input[3] * C4;
temp2 = input[13] * C12;
temp2 = (temp2 - temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp2 *= C8;
intermediate[ 9] = (2 * (temp2) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
intermediate[10] = (2 * (input[9] * C8) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
intermediate[11] = input[15] - input[1];
intermediate[12] = input[15] + input[1];
intermediate[13] = (2 * (input[7] * C8) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
temp1 = input[11] * C12;
temp2 = input[5] * C4;
temp2 = (temp2 - temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp2 *= C8;
intermediate[14] = (2 * (temp2) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
temp1 = input[11] * C4;
temp2 = input[5] * C12;
temp1 = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 *= C8;
intermediate[15] = (2 * (temp1) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
step[ 8] = (intermediate[ 8] + intermediate[14] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
step[ 9] = (intermediate[ 9] + intermediate[15] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
step[10] = (intermediate[10] + intermediate[11] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
step[11] = (intermediate[10] - intermediate[11] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
step[12] = (intermediate[12] + intermediate[13] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
step[13] = (intermediate[12] - intermediate[13] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
step[14] = (intermediate[ 8] - intermediate[14] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
step[15] = (intermediate[ 9] - intermediate[15] + INITIAL_ROUNDING)
>> INITIAL_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// step 3
output[0] = step[ 0] + step[ 3];
output[1] = step[ 1] + step[ 2];
output[2] = step[ 1] - step[ 2];
output[3] = step[ 0] - step[ 3];
temp1 = step[ 4] * C14;
temp2 = step[ 7] * C2;
output[4] = (temp1 - temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = step[ 4] * C2;
temp2 = step[ 7] * C14;
output[7] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = step[ 5] * C10;
temp2 = step[ 6] * C6;
output[5] = (temp1 - temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = step[ 5] * C6;
temp2 = step[ 6] * C10;
output[6] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
output[8] = step[ 8] + step[11];
output[9] = step[ 9] + step[10];
output[10] = step[ 9] - step[10];
output[11] = step[ 8] - step[11];
output[12] = step[12] + step[15];
output[13] = step[13] + step[14];
output[14] = step[13] - step[14];
output[15] = step[12] - step[15];
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// output 4
step[ 0] = output[0] + output[7];
step[ 1] = output[1] + output[6];
step[ 2] = output[2] + output[5];
step[ 3] = output[3] + output[4];
step[ 4] = output[3] - output[4];
step[ 5] = output[2] - output[5];
step[ 6] = output[1] - output[6];
step[ 7] = output[0] - output[7];
temp1 = output[8] * C7;
temp2 = output[15] * C9;
step[ 8] = (temp1 - temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[9] * C11;
temp2 = output[14] * C5;
step[ 9] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[10] * C3;
temp2 = output[13] * C13;
step[10] = (temp1 - temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[11] * C15;
temp2 = output[12] * C1;
step[11] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[11] * C1;
temp2 = output[12] * C15;
step[12] = (temp2 - temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[10] * C13;
temp2 = output[13] * C3;
step[13] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[9] * C5;
temp2 = output[14] * C11;
step[14] = (temp2 - temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[8] * C9;
temp2 = output[15] * C7;
step[15] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// step 5
output[0] = (step[0] + step[15] + last_rounding) >> last_shift_bits;
output[1] = (step[1] + step[14] + last_rounding) >> last_shift_bits;
output[2] = (step[2] + step[13] + last_rounding) >> last_shift_bits;
output[3] = (step[3] + step[12] + last_rounding) >> last_shift_bits;
output[4] = (step[4] + step[11] + last_rounding) >> last_shift_bits;
output[5] = (step[5] + step[10] + last_rounding) >> last_shift_bits;
output[6] = (step[6] + step[ 9] + last_rounding) >> last_shift_bits;
output[7] = (step[7] + step[ 8] + last_rounding) >> last_shift_bits;
output[15] = (step[0] - step[15] + last_rounding) >> last_shift_bits;
output[14] = (step[1] - step[14] + last_rounding) >> last_shift_bits;
output[13] = (step[2] - step[13] + last_rounding) >> last_shift_bits;
output[12] = (step[3] - step[12] + last_rounding) >> last_shift_bits;
output[11] = (step[4] - step[11] + last_rounding) >> last_shift_bits;
output[10] = (step[5] - step[10] + last_rounding) >> last_shift_bits;
output[9] = (step[6] - step[ 9] + last_rounding) >> last_shift_bits;
output[8] = (step[7] - step[ 8] + last_rounding) >> last_shift_bits;
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
}
void vp9_short_idct16x16_c(int16_t *input, int16_t *output, int pitch) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
int16_t out[16 * 16];
int16_t *outptr = &out[0];
const int short_pitch = pitch >> 1;
int i, j;
int16_t temp_in[16], temp_out[16];
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// First transform rows
for (i = 0; i < 16; ++i) {
butterfly_16x16_idct_1d(input, outptr, 0);
input += short_pitch;
outptr += 16;
}
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// Then transform columns
for (i = 0; i < 16; ++i) {
for (j = 0; j < 16; ++j)
temp_in[j] = out[j * 16 + i];
butterfly_16x16_idct_1d(temp_in, temp_out, 3);
for (j = 0; j < 16; ++j)
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
output[j * 16 + i] = temp_out[j];
}
}
Optimize 16x16 dequant and idct As suggested by Yaowu, simplified 16x16 dequant and idct. In decoder, after detoken step, we know the number of non-zero dct coefficients (eobs) in a macroblock. Idct calculation can be skipped or simplified based on eobs, which improves the decoder performance. Change-Id: I9ffa1cb134bcb5a7d64fcf90c81871a96d1b4018
2012-11-07 01:06:22 +01:00
/* The following function is called when we know the maximum number of non-zero
* dct coefficients is less or equal 10.
*/
static void butterfly_16x16_idct10_1d(int16_t input[16], int16_t output[16],
int last_shift_bits) {
int16_t step[16] = {0};
int intermediate[16] = {0};
int temp1, temp2;
int last_rounding = 0;
if (last_shift_bits > 0)
last_rounding = 1 << (last_shift_bits - 1);
// step 1 and 2
step[ 0] = (input[0] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[ 1] = (input[0] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
temp1 = (2 * (input[2] * C8) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
step[ 4] = (temp1 + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[ 5] = (temp1 + INITIAL_ROUNDING) >> INITIAL_SHIFT;
// for odd input
temp1 = (input[3] * C12 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 *= C8;
intermediate[ 8] = (2 * (temp1) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = (-input[3] * C4 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 *= C8;
intermediate[ 9] = (2 * (temp1) + RIGHT_ROUNDING) >> RIGHT_SHIFT;
step[ 8] = (intermediate[ 8] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[ 9] = (intermediate[ 9] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[10] = (-input[1] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[11] = (input[1] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[12] = (input[1] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[13] = (input[1] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[14] = (intermediate[ 8] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
step[15] = (intermediate[ 9] + INITIAL_ROUNDING) >> INITIAL_SHIFT;
// step 3
output[0] = step[ 0];
output[1] = step[ 1];
output[2] = step[ 1];
output[3] = step[ 0];
temp1 = step[ 4] * C14;
output[4] = (temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = step[ 4] * C2;
output[7] = (temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = step[ 5] * C10;
output[5] = (temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = step[ 5] * C6;
output[6] = (temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
output[8] = step[ 8] + step[11];
output[9] = step[ 9] + step[10];
output[10] = step[ 9] - step[10];
output[11] = step[ 8] - step[11];
output[12] = step[12] + step[15];
output[13] = step[13] + step[14];
output[14] = step[13] - step[14];
output[15] = step[12] - step[15];
// output 4
step[ 0] = output[0] + output[7];
step[ 1] = output[1] + output[6];
step[ 2] = output[2] + output[5];
step[ 3] = output[3] + output[4];
step[ 4] = output[3] - output[4];
step[ 5] = output[2] - output[5];
step[ 6] = output[1] - output[6];
step[ 7] = output[0] - output[7];
temp1 = output[8] * C7;
temp2 = output[15] * C9;
step[ 8] = (temp1 - temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[9] * C11;
temp2 = output[14] * C5;
step[ 9] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[10] * C3;
temp2 = output[13] * C13;
step[10] = (temp1 - temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[11] * C15;
temp2 = output[12] * C1;
step[11] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[11] * C1;
temp2 = output[12] * C15;
step[12] = (temp2 - temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[10] * C13;
temp2 = output[13] * C3;
step[13] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[9] * C5;
temp2 = output[14] * C11;
step[14] = (temp2 - temp1 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
temp1 = output[8] * C9;
temp2 = output[15] * C7;
step[15] = (temp1 + temp2 + RIGHT_ROUNDING) >> RIGHT_SHIFT;
// step 5
output[0] = (step[0] + step[15] + last_rounding) >> last_shift_bits;
output[1] = (step[1] + step[14] + last_rounding) >> last_shift_bits;
output[2] = (step[2] + step[13] + last_rounding) >> last_shift_bits;
output[3] = (step[3] + step[12] + last_rounding) >> last_shift_bits;
output[4] = (step[4] + step[11] + last_rounding) >> last_shift_bits;
output[5] = (step[5] + step[10] + last_rounding) >> last_shift_bits;
output[6] = (step[6] + step[ 9] + last_rounding) >> last_shift_bits;
output[7] = (step[7] + step[ 8] + last_rounding) >> last_shift_bits;
output[15] = (step[0] - step[15] + last_rounding) >> last_shift_bits;
output[14] = (step[1] - step[14] + last_rounding) >> last_shift_bits;
output[13] = (step[2] - step[13] + last_rounding) >> last_shift_bits;
output[12] = (step[3] - step[12] + last_rounding) >> last_shift_bits;
output[11] = (step[4] - step[11] + last_rounding) >> last_shift_bits;
output[10] = (step[5] - step[10] + last_rounding) >> last_shift_bits;
output[9] = (step[6] - step[ 9] + last_rounding) >> last_shift_bits;
output[8] = (step[7] - step[ 8] + last_rounding) >> last_shift_bits;
}
void vp9_short_idct10_16x16_c(int16_t *input, int16_t *output, int pitch) {
int16_t out[16 * 16];
int16_t *outptr = &out[0];
const int short_pitch = pitch >> 1;
int i, j;
int16_t temp_in[16], temp_out[16];
/* First transform rows. Since all non-zero dct coefficients are in
* upper-left 4x4 area, we only need to calculate first 4 rows here.
*/
vpx_memset(out, 0, sizeof(out));
for (i = 0; i < 4; ++i) {
butterfly_16x16_idct10_1d(input, outptr, 0);
input += short_pitch;
outptr += 16;
}
// Then transform columns
for (i = 0; i < 16; ++i) {
for (j = 0; j < 16; ++j)
temp_in[j] = out[j*16 + i];
butterfly_16x16_idct10_1d(temp_in, temp_out, 3);
for (j = 0; j < 16; ++j)
output[j*16 + i] = temp_out[j];
}
}
Convert 16x16 dct/idct to integer forms Converted vp9_short_fdct16x16_c and vp9_short_idct16x16_c to integer versions. Change-Id: Ie3ec985a890ac0f4f4f5818e6f0122e00c8af69f
2012-11-01 17:04:28 +01:00
#undef INITIAL_SHIFT
#undef INITIAL_ROUNDING
#undef RIGHT_SHIFT
#undef RIGHT_ROUNDING
#endif
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
#if CONFIG_TX32X32
#if !CONFIG_DWT32X32HYBRID
#define DownshiftMultiplyBy2(x) x * 2
#define DownshiftMultiply(x) x
static void idct16(double *input, double *output, int stride) {
static const double C1 = 0.995184726672197;
static const double C2 = 0.98078528040323;
static const double C3 = 0.956940335732209;
static const double C4 = 0.923879532511287;
static const double C5 = 0.881921264348355;
static const double C6 = 0.831469612302545;
static const double C7 = 0.773010453362737;
static const double C8 = 0.707106781186548;
static const double C9 = 0.634393284163646;
static const double C10 = 0.555570233019602;
static const double C11 = 0.471396736825998;
static const double C12 = 0.38268343236509;
static const double C13 = 0.290284677254462;
static const double C14 = 0.195090322016128;
static const double C15 = 0.098017140329561;
double step[16];
double intermediate[16];
double temp1, temp2;
// step 1 and 2
step[ 0] = input[stride*0] + input[stride*8];
step[ 1] = input[stride*0] - input[stride*8];
temp1 = input[stride*4]*C12;
temp2 = input[stride*12]*C4;
temp1 -= temp2;
temp1 = DownshiftMultiply(temp1);
temp1 *= C8;
step[ 2] = DownshiftMultiplyBy2(temp1);
temp1 = input[stride*4]*C4;
temp2 = input[stride*12]*C12;
temp1 += temp2;
temp1 = DownshiftMultiply(temp1);
temp1 *= C8;
step[ 3] = DownshiftMultiplyBy2(temp1);
temp1 = input[stride*2]*C8;
temp1 = DownshiftMultiplyBy2(temp1);
temp2 = input[stride*6] + input[stride*10];
step[ 4] = temp1 + temp2;
step[ 5] = temp1 - temp2;
temp1 = input[stride*14]*C8;
temp1 = DownshiftMultiplyBy2(temp1);
temp2 = input[stride*6] - input[stride*10];
step[ 6] = temp2 - temp1;
step[ 7] = temp2 + temp1;
// for odd input
temp1 = input[stride*3]*C12;
temp2 = input[stride*13]*C4;
temp1 += temp2;
temp1 = DownshiftMultiply(temp1);
temp1 *= C8;
intermediate[ 8] = DownshiftMultiplyBy2(temp1);
temp1 = input[stride*3]*C4;
temp2 = input[stride*13]*C12;
temp2 -= temp1;
temp2 = DownshiftMultiply(temp2);
temp2 *= C8;
intermediate[ 9] = DownshiftMultiplyBy2(temp2);
intermediate[10] = DownshiftMultiplyBy2(input[stride*9]*C8);
intermediate[11] = input[stride*15] - input[stride*1];
intermediate[12] = input[stride*15] + input[stride*1];
intermediate[13] = DownshiftMultiplyBy2((input[stride*7]*C8));
temp1 = input[stride*11]*C12;
temp2 = input[stride*5]*C4;
temp2 -= temp1;
temp2 = DownshiftMultiply(temp2);
temp2 *= C8;
intermediate[14] = DownshiftMultiplyBy2(temp2);
temp1 = input[stride*11]*C4;
temp2 = input[stride*5]*C12;
temp1 += temp2;
temp1 = DownshiftMultiply(temp1);
temp1 *= C8;
intermediate[15] = DownshiftMultiplyBy2(temp1);
step[ 8] = intermediate[ 8] + intermediate[14];
step[ 9] = intermediate[ 9] + intermediate[15];
step[10] = intermediate[10] + intermediate[11];
step[11] = intermediate[10] - intermediate[11];
step[12] = intermediate[12] + intermediate[13];
step[13] = intermediate[12] - intermediate[13];
step[14] = intermediate[ 8] - intermediate[14];
step[15] = intermediate[ 9] - intermediate[15];
// step 3
output[stride*0] = step[ 0] + step[ 3];
output[stride*1] = step[ 1] + step[ 2];
output[stride*2] = step[ 1] - step[ 2];
output[stride*3] = step[ 0] - step[ 3];
temp1 = step[ 4]*C14;
temp2 = step[ 7]*C2;
temp1 -= temp2;
output[stride*4] = DownshiftMultiply(temp1);
temp1 = step[ 4]*C2;
temp2 = step[ 7]*C14;
temp1 += temp2;
output[stride*7] = DownshiftMultiply(temp1);
temp1 = step[ 5]*C10;
temp2 = step[ 6]*C6;
temp1 -= temp2;
output[stride*5] = DownshiftMultiply(temp1);
temp1 = step[ 5]*C6;
temp2 = step[ 6]*C10;
temp1 += temp2;
output[stride*6] = DownshiftMultiply(temp1);
output[stride*8] = step[ 8] + step[11];
output[stride*9] = step[ 9] + step[10];
output[stride*10] = step[ 9] - step[10];
output[stride*11] = step[ 8] - step[11];
output[stride*12] = step[12] + step[15];
output[stride*13] = step[13] + step[14];
output[stride*14] = step[13] - step[14];
output[stride*15] = step[12] - step[15];
// output 4
step[ 0] = output[stride*0] + output[stride*7];
step[ 1] = output[stride*1] + output[stride*6];
step[ 2] = output[stride*2] + output[stride*5];
step[ 3] = output[stride*3] + output[stride*4];
step[ 4] = output[stride*3] - output[stride*4];
step[ 5] = output[stride*2] - output[stride*5];
step[ 6] = output[stride*1] - output[stride*6];
step[ 7] = output[stride*0] - output[stride*7];
temp1 = output[stride*8]*C7;
temp2 = output[stride*15]*C9;
temp1 -= temp2;
step[ 8] = DownshiftMultiply(temp1);
temp1 = output[stride*9]*C11;
temp2 = output[stride*14]*C5;
temp1 += temp2;
step[ 9] = DownshiftMultiply(temp1);
temp1 = output[stride*10]*C3;
temp2 = output[stride*13]*C13;
temp1 -= temp2;
step[10] = DownshiftMultiply(temp1);
temp1 = output[stride*11]*C15;
temp2 = output[stride*12]*C1;
temp1 += temp2;
step[11] = DownshiftMultiply(temp1);
temp1 = output[stride*11]*C1;
temp2 = output[stride*12]*C15;
temp2 -= temp1;
step[12] = DownshiftMultiply(temp2);
temp1 = output[stride*10]*C13;
temp2 = output[stride*13]*C3;
temp1 += temp2;
step[13] = DownshiftMultiply(temp1);
temp1 = output[stride*9]*C5;
temp2 = output[stride*14]*C11;
temp2 -= temp1;
step[14] = DownshiftMultiply(temp2);
temp1 = output[stride*8]*C9;
temp2 = output[stride*15]*C7;
temp1 += temp2;
step[15] = DownshiftMultiply(temp1);
// step 5
output[stride*0] = step[0] + step[15];
output[stride*1] = step[1] + step[14];
output[stride*2] = step[2] + step[13];
output[stride*3] = step[3] + step[12];
output[stride*4] = step[4] + step[11];
output[stride*5] = step[5] + step[10];
output[stride*6] = step[6] + step[ 9];
output[stride*7] = step[7] + step[ 8];
output[stride*15] = step[0] - step[15];
output[stride*14] = step[1] - step[14];
output[stride*13] = step[2] - step[13];
output[stride*12] = step[3] - step[12];
output[stride*11] = step[4] - step[11];
output[stride*10] = step[5] - step[10];
output[stride*9] = step[6] - step[ 9];
output[stride*8] = step[7] - step[ 8];
}
static void butterfly_32_idct_1d(double *input, double *output, int stride) {
static const double C1 = 0.998795456205; // cos(pi * 1 / 64)
static const double C3 = 0.989176509965; // cos(pi * 3 / 64)
static const double C5 = 0.970031253195; // cos(pi * 5 / 64)
static const double C7 = 0.941544065183; // cos(pi * 7 / 64)
static const double C9 = 0.903989293123; // cos(pi * 9 / 64)
static const double C11 = 0.857728610000; // cos(pi * 11 / 64)
static const double C13 = 0.803207531481; // cos(pi * 13 / 64)
static const double C15 = 0.740951125355; // cos(pi * 15 / 64)
static const double C16 = 0.707106781187; // cos(pi * 16 / 64)
static const double C17 = 0.671558954847; // cos(pi * 17 / 64)
static const double C19 = 0.595699304492; // cos(pi * 19 / 64)
static const double C21 = 0.514102744193; // cos(pi * 21 / 64)
static const double C23 = 0.427555093430; // cos(pi * 23 / 64)
static const double C25 = 0.336889853392; // cos(pi * 25 / 64)
static const double C27 = 0.242980179903; // cos(pi * 27 / 64)
static const double C29 = 0.146730474455; // cos(pi * 29 / 64)
static const double C31 = 0.049067674327; // cos(pi * 31 / 64)
double step1[32];
double step2[32];
step1[ 0] = input[stride*0];
step1[ 1] = input[stride*2];
step1[ 2] = input[stride*4];
step1[ 3] = input[stride*6];
step1[ 4] = input[stride*8];
step1[ 5] = input[stride*10];
step1[ 6] = input[stride*12];
step1[ 7] = input[stride*14];
step1[ 8] = input[stride*16];
step1[ 9] = input[stride*18];
step1[10] = input[stride*20];
step1[11] = input[stride*22];
step1[12] = input[stride*24];
step1[13] = input[stride*26];
step1[14] = input[stride*28];
step1[15] = input[stride*30];
step1[16] = DownshiftMultiplyBy2(input[stride*1]*C16);
step1[17] = (input[stride*3] + input[stride*1]);
step1[18] = (input[stride*5] + input[stride*3]);
step1[19] = (input[stride*7] + input[stride*5]);
step1[20] = (input[stride*9] + input[stride*7]);
step1[21] = (input[stride*11] + input[stride*9]);
step1[22] = (input[stride*13] + input[stride*11]);
step1[23] = (input[stride*15] + input[stride*13]);
step1[24] = (input[stride*17] + input[stride*15]);
step1[25] = (input[stride*19] + input[stride*17]);
step1[26] = (input[stride*21] + input[stride*19]);
step1[27] = (input[stride*23] + input[stride*21]);
step1[28] = (input[stride*25] + input[stride*23]);
step1[29] = (input[stride*27] + input[stride*25]);
step1[30] = (input[stride*29] + input[stride*27]);
step1[31] = (input[stride*31] + input[stride*29]);
idct16(step1, step2, 1);
idct16(step1 + 16, step2 + 16, 1);
step2[16] = DownshiftMultiply(step2[16] / (2*C1));
step2[17] = DownshiftMultiply(step2[17] / (2*C3));
step2[18] = DownshiftMultiply(step2[18] / (2*C5));
step2[19] = DownshiftMultiply(step2[19] / (2*C7));
step2[20] = DownshiftMultiply(step2[20] / (2*C9));
step2[21] = DownshiftMultiply(step2[21] / (2*C11));
step2[22] = DownshiftMultiply(step2[22] / (2*C13));
step2[23] = DownshiftMultiply(step2[23] / (2*C15));
step2[24] = DownshiftMultiply(step2[24] / (2*C17));
step2[25] = DownshiftMultiply(step2[25] / (2*C19));
step2[26] = DownshiftMultiply(step2[26] / (2*C21));
step2[27] = DownshiftMultiply(step2[27] / (2*C23));
step2[28] = DownshiftMultiply(step2[28] / (2*C25));
step2[29] = DownshiftMultiply(step2[29] / (2*C27));
step2[30] = DownshiftMultiply(step2[30] / (2*C29));
step2[31] = DownshiftMultiply(step2[31] / (2*C31));
output[stride* 0] = step2[ 0] + step2[16];
output[stride* 1] = step2[ 1] + step2[17];
output[stride* 2] = step2[ 2] + step2[18];
output[stride* 3] = step2[ 3] + step2[19];
output[stride* 4] = step2[ 4] + step2[20];
output[stride* 5] = step2[ 5] + step2[21];
output[stride* 6] = step2[ 6] + step2[22];
output[stride* 7] = step2[ 7] + step2[23];
output[stride* 8] = step2[ 8] + step2[24];
output[stride* 9] = step2[ 9] + step2[25];
output[stride*10] = step2[10] + step2[26];
output[stride*11] = step2[11] + step2[27];
output[stride*12] = step2[12] + step2[28];
output[stride*13] = step2[13] + step2[29];
output[stride*14] = step2[14] + step2[30];
output[stride*15] = step2[15] + step2[31];
output[stride*16] = step2[15] - step2[(31 - 0)];
output[stride*17] = step2[14] - step2[(31 - 1)];
output[stride*18] = step2[13] - step2[(31 - 2)];
output[stride*19] = step2[12] - step2[(31 - 3)];
output[stride*20] = step2[11] - step2[(31 - 4)];
output[stride*21] = step2[10] - step2[(31 - 5)];
output[stride*22] = step2[ 9] - step2[(31 - 6)];
output[stride*23] = step2[ 8] - step2[(31 - 7)];
output[stride*24] = step2[ 7] - step2[(31 - 8)];
output[stride*25] = step2[ 6] - step2[(31 - 9)];
output[stride*26] = step2[ 5] - step2[(31 - 10)];
output[stride*27] = step2[ 4] - step2[(31 - 11)];
output[stride*28] = step2[ 3] - step2[(31 - 12)];
output[stride*29] = step2[ 2] - step2[(31 - 13)];
output[stride*30] = step2[ 1] - step2[(31 - 14)];
output[stride*31] = step2[ 0] - step2[(31 - 15)];
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct32x32_c(int16_t *input, int16_t *output, int pitch) {
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
vp9_clear_system_state(); // Make it simd safe : __asm emms;
{
double out[32*32], out2[32*32];
const int short_pitch = pitch >> 1;
int i, j;
// First transform rows
for (i = 0; i < 32; ++i) {
double temp_in[32], temp_out[32];
for (j = 0; j < 32; ++j)
temp_in[j] = input[j + i*short_pitch];
butterfly_32_idct_1d(temp_in, temp_out, 1);
for (j = 0; j < 32; ++j)
out[j + i*32] = temp_out[j];
}
// Then transform columns
for (i = 0; i < 32; ++i) {
double temp_in[32], temp_out[32];
for (j = 0; j < 32; ++j)
temp_in[j] = out[j*32 + i];
butterfly_32_idct_1d(temp_in, temp_out, 1);
for (j = 0; j < 32; ++j)
out2[j*32 + i] = temp_out[j];
}
for (i = 0; i < 32*32; ++i)
output[i] = round(out2[i]/128);
}
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
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
#else // CONFIG_DWT32X32HYBRID
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
#define DWT_MAX_LENGTH 32
#define DWT_TYPE 26 // 26/53/97
#define DWT_PRECISION_BITS 2
#define DWT_PRECISION_RND ((1 << DWT_PRECISION_BITS) / 2)
#if DWT_TYPE == 53
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
// Note: block length must be even for this implementation
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
static void synthesis_53_row(int length, int16_t *lowpass, int16_t *highpass,
int16_t *x) {
int16_t r, *a, *b;
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
int n;
n = length >> 1;
b = highpass;
a = lowpass;
r = *highpass;
while (n--) {
*a++ -= (r + (*b) + 1) >> 1;
r = *b++;
}
n = length >> 1;
b = highpass;
a = lowpass;
while (--n) {
*x++ = ((r = *a++) + 1) >> 1;
*x++ = *b++ + ((r + (*a) + 2) >> 2);
}
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
*x++ = ((r = *a) + 1) >> 1;
*x++ = *b + ((r + 1) >> 1);
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
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
static void synthesis_53_col(int length, int16_t *lowpass, int16_t *highpass,
int16_t *x) {
int16_t r, *a, *b;
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
int n;
n = length >> 1;
b = highpass;
a = lowpass;
r = *highpass;
while (n--) {
*a++ -= (r + (*b) + 1) >> 1;
r = *b++;
}
n = length >> 1;
b = highpass;
a = lowpass;
while (--n) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
r = *a++;
*x++ = r;
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
*x++ = ((*b++) << 1) + ((r + (*a) + 1) >> 1);
}
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
*x++ = *a;
*x++ = ((*b) << 1) + *a;
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
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void dyadic_synthesize_53(int levels, int width, int height, int16_t *c,
int pitch_c, int16_t *x, int pitch_x) {
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
int th[16], tw[16], lv, i, j, nh, nw, hh = height, hw = width;
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
short buffer[2 * DWT_MAX_LENGTH];
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
th[0] = hh;
tw[0] = hw;
for (i = 1; i <= levels; i++) {
th[i] = (th[i - 1] + 1) >> 1;
tw[i] = (tw[i - 1] + 1) >> 1;
}
for (lv = levels - 1; lv >= 0; lv--) {
nh = th[lv];
nw = tw[lv];
hh = th[lv + 1];
hw = tw[lv + 1];
if ((nh < 2) || (nw < 2)) continue;
for (j = 0; j < nw; j++) {
for (i = 0; i < nh; i++)
buffer[i] = c[i * pitch_c + j];
synthesis_53_col(nh, buffer, buffer + hh, buffer + nh);
for (i = 0; i < nh; i++)
c[i * pitch_c + j] = buffer[i + nh];
}
for (i = 0; i < nh; i++) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
memcpy(buffer, &c[i * pitch_c], nw * sizeof(*buffer));
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
synthesis_53_row(nw, buffer, buffer + hw, &c[i * pitch_c]);
}
}
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
for (i = 0; i < height; i++) {
for (j = 0; j < width; j++) {
x[i * pitch_x + j] = c[i * pitch_c + j] >= 0 ?
((c[i * pitch_c + j] + DWT_PRECISION_RND) >> DWT_PRECISION_BITS) :
-((-c[i * pitch_c + j] + DWT_PRECISION_RND) >> DWT_PRECISION_BITS);
}
}
}
#elif DWT_TYPE == 26
// Note: block length must be even for this implementation
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
static void synthesis_26_row(int length, int16_t *lowpass, int16_t *highpass,
int16_t *x) {
int16_t r, s, *a, *b;
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
int i, n = length >> 1;
if (n >= 4) {
a = lowpass;
b = highpass;
r = *lowpass;
while (--n) {
*b++ += (r - a[1] + 4) >> 3;
r = *a++;
}
*b += (r - *a + 4) >> 3;
}
a = lowpass;
b = highpass;
for (i = length >> 1; i; i--) {
s = *b++;
r = *a++;
*x++ = (r + s + 1) >> 1;
*x++ = (r - s + 1) >> 1;
}
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
static void synthesis_26_col(int length, int16_t *lowpass, int16_t *highpass,
int16_t *x) {
int16_t r, s, *a, *b;
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
int i, n = length >> 1;
if (n >= 4) {
a = lowpass;
b = highpass;
r = *lowpass;
while (--n) {
*b++ += (r - a[1] + 4) >> 3;
r = *a++;
}
*b += (r - *a + 4) >> 3;
}
a = lowpass;
b = highpass;
for (i = length >> 1; i; i--) {
s = *b++;
r = *a++;
*x++ = r + s;
*x++ = r - s;
}
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void dyadic_synthesize_26(int levels, int width, int height, int16_t *c,
int pitch_c, int16_t *x, int pitch_x) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
int th[16], tw[16], lv, i, j, nh, nw, hh = height, hw = width;
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t buffer[2 * DWT_MAX_LENGTH];
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
th[0] = hh;
tw[0] = hw;
for (i = 1; i <= levels; i++) {
th[i] = (th[i - 1] + 1) >> 1;
tw[i] = (tw[i - 1] + 1) >> 1;
}
for (lv = levels - 1; lv >= 0; lv--) {
nh = th[lv];
nw = tw[lv];
hh = th[lv + 1];
hw = tw[lv + 1];
if ((nh < 2) || (nw < 2)) continue;
for (j = 0; j < nw; j++) {
for (i = 0; i < nh; i++)
buffer[i] = c[i * pitch_c + j];
synthesis_26_col(nh, buffer, buffer + hh, buffer + nh);
for (i = 0; i < nh; i++)
c[i * pitch_c + j] = buffer[i + nh];
}
for (i = 0; i < nh; i++) {
memcpy(buffer, &c[i * pitch_c], nw * sizeof(*buffer));
synthesis_26_row(nw, buffer, buffer + hw, &c[i * pitch_c]);
}
}
for (i = 0; i < height; i++) {
for (j = 0; j < width; j++) {
x[i * pitch_x + j] = c[i * pitch_c + j] >= 0 ?
((c[i * pitch_c + j] + DWT_PRECISION_RND) >> DWT_PRECISION_BITS) :
-((-c[i * pitch_c + j] + DWT_PRECISION_RND) >> DWT_PRECISION_BITS);
}
}
}
#elif DWT_TYPE == 97
static void synthesis_97(int length, double *lowpass, double *highpass,
double *x) {
static const double a_predict1 = -1.586134342;
static const double a_update1 = -0.05298011854;
static const double a_predict2 = 0.8829110762;
static const double a_update2 = 0.4435068522;
static const double s_low = 1.149604398;
static const double s_high = 1/1.149604398;
static const double inv_s_low = 1 / s_low;
static const double inv_s_high = 1 / s_high;
int i;
double y[DWT_MAX_LENGTH];
// Undo pack and scale
for (i = 0; i < length / 2; i++) {
y[i * 2] = lowpass[i] * inv_s_low;
y[i * 2 + 1] = highpass[i] * inv_s_high;
}
memcpy(x, y, sizeof(*y) * length);
// Undo update 2
for (i = 2; i < length; i += 2) {
x[i] -= a_update2 * (x[i-1] + x[i+1]);
}
x[0] -= 2 * a_update2 * x[1];
// Undo predict 2
for (i = 1; i < length - 2; i += 2) {
x[i] -= a_predict2 * (x[i - 1] + x[i + 1]);
}
x[length - 1] -= 2 * a_predict2 * x[length - 2];
// Undo update 1
for (i = 2; i < length; i += 2) {
x[i] -= a_update1 * (x[i - 1] + x[i + 1]);
}
x[0] -= 2 * a_update1 * x[1];
// Undo predict 1
for (i = 1; i < length - 2; i += 2) {
x[i] -= a_predict1 * (x[i - 1] + x[i + 1]);
}
x[length - 1] -= 2 * a_predict1 * x[length - 2];
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void dyadic_synthesize_97(int levels, int width, int height, int16_t *c,
int pitch_c, int16_t *x, int pitch_x) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
int th[16], tw[16], lv, i, j, nh, nw, hh = height, hw = width;
double buffer[2 * DWT_MAX_LENGTH];
double y[DWT_MAX_LENGTH * DWT_MAX_LENGTH];
th[0] = hh;
tw[0] = hw;
for (i = 1; i <= levels; i++) {
th[i] = (th[i - 1] + 1) >> 1;
tw[i] = (tw[i - 1] + 1) >> 1;
}
for (lv = levels - 1; lv >= 0; lv--) {
nh = th[lv];
nw = tw[lv];
hh = th[lv + 1];
hw = tw[lv + 1];
if ((nh < 2) || (nw < 2)) continue;
for (j = 0; j < nw; j++) {
for (i = 0; i < nh; i++)
buffer[i] = c[i * pitch_c + j];
synthesis_97(nh, buffer, buffer + hh, buffer + nh);
for (i = 0; i < nh; i++)
y[i * DWT_MAX_LENGTH + j] = buffer[i + nh];
}
for (i = 0; i < nh; i++) {
memcpy(buffer, &y[i * DWT_MAX_LENGTH], nw * sizeof(*buffer));
synthesis_97(nw, buffer, buffer + hw, &y[i * DWT_MAX_LENGTH]);
}
}
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
for (i = 0; i < height; i++)
for (j = 0; j < width; j++)
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
x[i * pitch_x + j] = round(y[i * DWT_MAX_LENGTH + j] /
(1 << DWT_PRECISION_BITS));
}
#endif // DWT_TYPE
// TODO(debargha): Implement scaling differently so as not to have to use the
// floating point 16x16 dct
static void butterfly_16x16_idct_1d_f(double input[16], double output[16]) {
static const double C1 = 0.995184726672197;
static const double C2 = 0.98078528040323;
static const double C3 = 0.956940335732209;
static const double C4 = 0.923879532511287;
static const double C5 = 0.881921264348355;
static const double C6 = 0.831469612302545;
static const double C7 = 0.773010453362737;
static const double C8 = 0.707106781186548;
static const double C9 = 0.634393284163646;
static const double C10 = 0.555570233019602;
static const double C11 = 0.471396736825998;
static const double C12 = 0.38268343236509;
static const double C13 = 0.290284677254462;
static const double C14 = 0.195090322016128;
static const double C15 = 0.098017140329561;
vp9_clear_system_state(); // Make it simd safe : __asm emms;
{
double step[16];
double intermediate[16];
double temp1, temp2;
// step 1 and 2
step[ 0] = input[0] + input[8];
step[ 1] = input[0] - input[8];
temp1 = input[4]*C12;
temp2 = input[12]*C4;
temp1 -= temp2;
temp1 *= C8;
step[ 2] = 2*(temp1);
temp1 = input[4]*C4;
temp2 = input[12]*C12;
temp1 += temp2;
temp1 = (temp1);
temp1 *= C8;
step[ 3] = 2*(temp1);
temp1 = input[2]*C8;
temp1 = 2*(temp1);
temp2 = input[6] + input[10];
step[ 4] = temp1 + temp2;
step[ 5] = temp1 - temp2;
temp1 = input[14]*C8;
temp1 = 2*(temp1);
temp2 = input[6] - input[10];
step[ 6] = temp2 - temp1;
step[ 7] = temp2 + temp1;
// for odd input
temp1 = input[3]*C12;
temp2 = input[13]*C4;
temp1 += temp2;
temp1 = (temp1);
temp1 *= C8;
intermediate[ 8] = 2*(temp1);
temp1 = input[3]*C4;
temp2 = input[13]*C12;
temp2 -= temp1;
temp2 = (temp2);
temp2 *= C8;
intermediate[ 9] = 2*(temp2);
intermediate[10] = 2*(input[9]*C8);
intermediate[11] = input[15] - input[1];
intermediate[12] = input[15] + input[1];
intermediate[13] = 2*((input[7]*C8));
temp1 = input[11]*C12;
temp2 = input[5]*C4;
temp2 -= temp1;
temp2 = (temp2);
temp2 *= C8;
intermediate[14] = 2*(temp2);
temp1 = input[11]*C4;
temp2 = input[5]*C12;
temp1 += temp2;
temp1 = (temp1);
temp1 *= C8;
intermediate[15] = 2*(temp1);
step[ 8] = intermediate[ 8] + intermediate[14];
step[ 9] = intermediate[ 9] + intermediate[15];
step[10] = intermediate[10] + intermediate[11];
step[11] = intermediate[10] - intermediate[11];
step[12] = intermediate[12] + intermediate[13];
step[13] = intermediate[12] - intermediate[13];
step[14] = intermediate[ 8] - intermediate[14];
step[15] = intermediate[ 9] - intermediate[15];
// step 3
output[0] = step[ 0] + step[ 3];
output[1] = step[ 1] + step[ 2];
output[2] = step[ 1] - step[ 2];
output[3] = step[ 0] - step[ 3];
temp1 = step[ 4]*C14;
temp2 = step[ 7]*C2;
temp1 -= temp2;
output[4] = (temp1);
temp1 = step[ 4]*C2;
temp2 = step[ 7]*C14;
temp1 += temp2;
output[7] = (temp1);
temp1 = step[ 5]*C10;
temp2 = step[ 6]*C6;
temp1 -= temp2;
output[5] = (temp1);
temp1 = step[ 5]*C6;
temp2 = step[ 6]*C10;
temp1 += temp2;
output[6] = (temp1);
output[8] = step[ 8] + step[11];
output[9] = step[ 9] + step[10];
output[10] = step[ 9] - step[10];
output[11] = step[ 8] - step[11];
output[12] = step[12] + step[15];
output[13] = step[13] + step[14];
output[14] = step[13] - step[14];
output[15] = step[12] - step[15];
// output 4
step[ 0] = output[0] + output[7];
step[ 1] = output[1] + output[6];
step[ 2] = output[2] + output[5];
step[ 3] = output[3] + output[4];
step[ 4] = output[3] - output[4];
step[ 5] = output[2] - output[5];
step[ 6] = output[1] - output[6];
step[ 7] = output[0] - output[7];
temp1 = output[8]*C7;
temp2 = output[15]*C9;
temp1 -= temp2;
step[ 8] = (temp1);
temp1 = output[9]*C11;
temp2 = output[14]*C5;
temp1 += temp2;
step[ 9] = (temp1);
temp1 = output[10]*C3;
temp2 = output[13]*C13;
temp1 -= temp2;
step[10] = (temp1);
temp1 = output[11]*C15;
temp2 = output[12]*C1;
temp1 += temp2;
step[11] = (temp1);
temp1 = output[11]*C1;
temp2 = output[12]*C15;
temp2 -= temp1;
step[12] = (temp2);
temp1 = output[10]*C13;
temp2 = output[13]*C3;
temp1 += temp2;
step[13] = (temp1);
temp1 = output[9]*C5;
temp2 = output[14]*C11;
temp2 -= temp1;
step[14] = (temp2);
temp1 = output[8]*C9;
temp2 = output[15]*C7;
temp1 += temp2;
step[15] = (temp1);
// step 5
output[0] = (step[0] + step[15]);
output[1] = (step[1] + step[14]);
output[2] = (step[2] + step[13]);
output[3] = (step[3] + step[12]);
output[4] = (step[4] + step[11]);
output[5] = (step[5] + step[10]);
output[6] = (step[6] + step[ 9]);
output[7] = (step[7] + step[ 8]);
output[15] = (step[0] - step[15]);
output[14] = (step[1] - step[14]);
output[13] = (step[2] - step[13]);
output[12] = (step[3] - step[12]);
output[11] = (step[4] - step[11]);
output[10] = (step[5] - step[10]);
output[9] = (step[6] - step[ 9]);
output[8] = (step[7] - step[ 8]);
}
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct16x16_c_f(int16_t *input, int16_t *output, int pitch) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
vp9_clear_system_state(); // Make it simd safe : __asm emms;
{
double out[16*16], out2[16*16];
const int short_pitch = pitch >> 1;
int i, j;
// First transform rows
for (i = 0; i < 16; ++i) {
double temp_in[16], temp_out[16];
for (j = 0; j < 16; ++j)
temp_in[j] = input[j + i*short_pitch];
butterfly_16x16_idct_1d_f(temp_in, temp_out);
for (j = 0; j < 16; ++j)
out[j + i*16] = temp_out[j];
}
// Then transform columns
for (i = 0; i < 16; ++i) {
double temp_in[16], temp_out[16];
for (j = 0; j < 16; ++j)
temp_in[j] = out[j*16 + i];
butterfly_16x16_idct_1d_f(temp_in, temp_out);
for (j = 0; j < 16; ++j)
out2[j*16 + i] = temp_out[j];
}
for (i = 0; i < 16*16; ++i)
output[i] = round(out2[i] / (64 >> DWT_PRECISION_BITS));
}
vp9_clear_system_state(); // Make it simd safe : __asm emms;
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
}
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
void vp9_short_idct32x32_c(int16_t *input, int16_t *output, int pitch) {
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
// assume out is a 32x32 buffer
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// Temporary buffer to hold a 16x16 block for 16x16 inverse dct
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t buffer[16 * 16];
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// Temporary buffer to hold a 32x32 block for inverse 32x32 dwt
Use standard integer types for pixel values and coefficients. For coefficients, use int16_t (instead of short); for pixel values in 16-bit intermediates, use uint16_t (instead of unsigned short); for all others, use uint8_t (instead of unsigned char). Change-Id: I3619cd9abf106c3742eccc2e2f5e89a62774f7da
2012-12-19 00:31:19 +01:00
int16_t buffer2[32 * 32];
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
// Note: pitch is in bytes, short_pitch is in short units
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
const int short_pitch = pitch >> 1;
int i;
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
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
// TODO(debargha): Implement more efficiently by adding output pitch
// argument to the idct16x16 function
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
vp9_short_idct16x16_c_f(input, buffer, pitch);
for (i = 0; i < 16; ++i) {
vpx_memcpy(buffer2 + i * 32, buffer + i * 16, sizeof(*buffer2) * 16);
}
vp9_short_idct16x16_c_f(input + 16, buffer, pitch);
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
for (i = 0; i < 16; ++i) {
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
vpx_memcpy(buffer2 + i * 32 + 16, buffer + i * 16, sizeof(*buffer2) * 16);
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
}
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
vp9_short_idct16x16_c_f(input + 16 * short_pitch, buffer, pitch);
for (i = 0; i < 16; ++i) {
vpx_memcpy(buffer2 + i * 32 + 16 * 32, buffer + i * 16,
sizeof(*buffer2) * 16);
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
}
Further improvements on the hybrid dwt/dct expt Modifies the scanning pattern and uses a floating point 16x16 dct implementation for now to handle scaling better. Also experiments are in progress with 2/6 and 9/7 wavelets. Results have improved to within ~0.25% of 32x32 dct for std-hd and about 0.03% for derf. This difference can probably be bridged by re-optimizing the entropy stats for these transforms. Currently the stats used are common between 32x32 dct and dwt/dct. Experiments are in progress with various scan pattern - wavelet combinations. Ideally the subbands should be tokenized separately, and an experiment will be condcuted next on that. Change-Id: Ia9cbfc2d63cb7a47e562b2cd9341caf962bcc110
2012-12-12 02:06:35 +01:00
vp9_short_idct16x16_c_f(input + 16 * short_pitch + 16, buffer, pitch);
for (i = 0; i < 16; ++i) {
vpx_memcpy(buffer2 + i * 32 + 16 * 33, buffer + i * 16,
sizeof(*buffer2) * 16);
}
#if DWT_TYPE == 26
dyadic_synthesize_26(1, 32, 32, buffer2, 32, output, 32);
#elif DWT_TYPE == 97
dyadic_synthesize_97(1, 32, 32, buffer2, 32, output, 32);
#elif DWT_TYPE == 53
dyadic_synthesize_53(1, 32, 32, buffer2, 32, output, 32);
#endif
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
}
#endif // CONFIG_DWT32X32HYBRID
#endif // CONFIG_TX32X32
Reference in New Issue Copy Permalink