opus.git
11 years agooptimisation: defined a reciprocal square root (celt_rsqrt) for use in
Jean-Marc Valin [Wed, 26 Mar 2008 05:42:42 +0000 (16:42 +1100)]
optimisation: defined a reciprocal square root (celt_rsqrt) for use in
find_spectral_pitch instead of using celt_rcp(celt_sqrt(x))

11 years agoFixed the rcp() testcase for new assumptions (x is positive)
Jean-Marc Valin [Wed, 26 Mar 2008 04:56:07 +0000 (15:56 +1100)]
Fixed the rcp() testcase for new assumptions (x is positive)

11 years agooptimisation: shaving a few cycles off prev_cwrs* by not computed the values
Jean-Marc Valin [Wed, 26 Mar 2008 04:46:51 +0000 (15:46 +1100)]
optimisation: shaving a few cycles off prev_cwrs* by not computed the values
we're not going to use.

11 years agooptimisations: faster handling of the zero for compute_band_energies() and
Jean-Marc Valin [Wed, 26 Mar 2008 03:36:07 +0000 (14:36 +1100)]
optimisations: faster handling of the zero for compute_band_energies() and
normalise_bands(). Also, another bunch of restrict qualifiers -- mainly
to tell the compiler there's no aliasing between the array bounds and the
data.

11 years agochanged 1*rcp(x) to just rcp(x)
Jean-Marc Valin [Wed, 26 Mar 2008 02:45:01 +0000 (13:45 +1100)]
changed 1*rcp(x) to just rcp(x)

11 years agooptimisation: intra_prediction() no longer needs to divide inside the search
Jean-Marc Valin [Wed, 26 Mar 2008 01:16:00 +0000 (12:16 +1100)]
optimisation: intra_prediction() no longer needs to divide inside the search

11 years agooptimisation: The "simple" Rxy/sqrt(Ryy) case in alg_quant no longer requires
Jean-Marc Valin [Tue, 25 Mar 2008 23:34:23 +0000 (10:34 +1100)]
optimisation: The "simple" Rxy/sqrt(Ryy) case in alg_quant no longer requires
a division

11 years agoDecision on whether to use pitch is now taken only based on energy in the
Jean-Marc Valin [Tue, 25 Mar 2008 21:10:27 +0000 (08:10 +1100)]
Decision on whether to use pitch is now taken only based on energy in the
first three bands (instead of total MDCT window energy)

11 years agoproperly defined EPSILON for the float case
Jean-Marc Valin [Tue, 25 Mar 2008 13:00:18 +0000 (00:00 +1100)]
properly defined EPSILON for the float case

11 years agoA bunch of pointers marked as "restrict" to ease the job of the compiler
Jean-Marc Valin [Tue, 25 Mar 2008 11:38:58 +0000 (22:38 +1100)]
A bunch of pointers marked as "restrict" to ease the job of the compiler

11 years agooptimisation: spreading_func now in-place with no branch in the loop and half
Jean-Marc Valin [Tue, 25 Mar 2008 11:09:32 +0000 (22:09 +1100)]
optimisation: spreading_func now in-place with no branch in the loop and half
the multiplications

11 years agooops. find_max32() now uses VERY_LARGE32 (instead of VERY_LARGE16)
Jean-Marc Valin [Tue, 25 Mar 2008 10:31:12 +0000 (21:31 +1100)]
oops. find_max32() now uses VERY_LARGE32 (instead of VERY_LARGE16)

11 years agooptimisations: Another bunch of simplifications to alg_quant(), mainly to
Jean-Marc Valin [Tue, 25 Mar 2008 10:28:40 +0000 (21:28 +1100)]
optimisations: Another bunch of simplifications to alg_quant(), mainly to
remove unnecessary copying and some conditional branches.

11 years agooptimisation: Making use of restrict in find_spectral_pitch() to disambiguate
Jean-Marc Valin [Tue, 25 Mar 2008 07:04:08 +0000 (18:04 +1100)]
optimisation: Making use of restrict in find_spectral_pitch() to disambiguate
X and Y.

11 years agooptimisations: caching sign of x in alg_quant(), changed celt_div()/celt_rcp()
Jean-Marc Valin [Tue, 25 Mar 2008 05:07:05 +0000 (16:07 +1100)]
optimisations: caching sign of x in alg_quant(), changed celt_div()/celt_rcp()
to assume denominator is positive.

11 years agoOptimised intra prediction a bit -- removed a conditional branch and replaced
Jean-Marc Valin [Tue, 25 Mar 2008 04:25:08 +0000 (15:25 +1100)]
Optimised intra prediction a bit -- removed a conditional branch and replaced
some calls to ec_enc_uint() by calls to ec_enc_bits() (saves divisions)

11 years agoRemoved support for more than one MDCT blocks per frame. I don't think there's
Jean-Marc Valin [Tue, 25 Mar 2008 03:38:55 +0000 (14:38 +1100)]
Removed support for more than one MDCT blocks per frame. I don't think there's
a use for that anymore now that the overlap can be less than one frame.

11 years agoRemoved the "pitch compression" in the residual quantisation. Also, removed
Jean-Marc Valin [Tue, 25 Mar 2008 03:15:41 +0000 (14:15 +1100)]
Removed the "pitch compression" in the residual quantisation. Also, removed
the more complex "n-best search" and replaced it with a greedy search

11 years agoUnrolled the inner loop in vq_index() so that the codebook unpacking doesn't
Jean-Marc Valin [Mon, 24 Mar 2008 10:39:49 +0000 (21:39 +1100)]
Unrolled the inner loop in vq_index() so that the codebook unpacking doesn't
require a branch.

11 years agoMaking bits2pulses() use a fixed number of iterations to allow further
Jean-Marc Valin [Mon, 24 Mar 2008 06:06:08 +0000 (17:06 +1100)]
Making bits2pulses() use a fixed number of iterations to allow further
optimisations.

11 years agoinclude "dsplib.h" in fixed_c5x.h
Jean-Marc Valin [Sun, 23 Mar 2008 21:32:49 +0000 (08:32 +1100)]
include "dsplib.h" in fixed_c5x.h

11 years agoreplaced divisions by recoprocals in intra prediction and folding
Jean-Marc Valin [Sun, 23 Mar 2008 21:15:40 +0000 (08:15 +1100)]
replaced divisions by recoprocals in intra prediction and folding

11 years agodefined find_max16 and overrode it for C55x
Jean-Marc Valin [Sat, 22 Mar 2008 21:06:29 +0000 (08:06 +1100)]
defined find_max16 and overrode it for C55x

11 years agoMade a second version of ec_{en|de}code optimised for encoding bits (no div
Jean-Marc Valin [Sat, 22 Mar 2008 11:58:45 +0000 (22:58 +1100)]
Made a second version of ec_{en|de}code optimised for encoding bits (no div
required) and using it in ec_{en|de}c_bits()

11 years agoNo longer trying to save bits when encoding integers near the upper limit
Jean-Marc Valin [Sat, 22 Mar 2008 11:23:58 +0000 (22:23 +1100)]
No longer trying to save bits when encoding integers near the upper limit
(and fix for celt_div with 16-bit numerator on a 16-bit CPU)

11 years agofixed-point: added cheap celt_div() division using a reciprocal
Jean-Marc Valin [Sat, 22 Mar 2008 10:17:45 +0000 (21:17 +1100)]
fixed-point: added cheap celt_div() division using a reciprocal

11 years agomore use of restrict
Jean-Marc Valin [Sat, 22 Mar 2008 06:53:12 +0000 (17:53 +1100)]
more use of restrict

11 years agoUsing restrict to make it clear there's no aliasing issues in the mdct.
Jean-Marc Valin [Sat, 22 Mar 2008 00:23:57 +0000 (11:23 +1100)]
Using restrict to make it clear there's no aliasing issues in the mdct.

11 years agoAdded a few "restrict" keywords and changed some divisions to shifts
Jean-Marc Valin [Fri, 21 Mar 2008 22:01:50 +0000 (09:01 +1100)]
Added a few "restrict" keywords and changed some divisions to shifts

11 years agofixed TI fft code -- again
Jean-Marc Valin [Fri, 21 Mar 2008 11:54:41 +0000 (22:54 +1100)]
fixed TI fft code -- again

11 years agoRemoved potentially unused var in MDCT init
Jean-Marc Valin [Fri, 21 Mar 2008 02:11:16 +0000 (13:11 +1100)]
Removed potentially unused var in MDCT init

11 years agooops (TI FFT)
Jean-Marc Valin [Fri, 21 Mar 2008 02:02:53 +0000 (13:02 +1100)]
oops (TI FFT)

11 years agolocal var name maxval was shadowing the TI function used to compute it
Jean-Marc Valin [Fri, 21 Mar 2008 01:53:37 +0000 (12:53 +1100)]
local var name maxval was shadowing the TI function used to compute it

11 years agomake sure TRIG_UPSCALE is properly defined
Jean-Marc Valin [Fri, 21 Mar 2008 01:51:05 +0000 (12:51 +1100)]
make sure TRIG_UPSCALE is properly defined

11 years agofix for TI version of celt_maxabs16()
Jean-Marc Valin [Fri, 21 Mar 2008 01:49:36 +0000 (12:49 +1100)]
fix for TI version of celt_maxabs16()

11 years agofixed-point: fix for 32-bit TI FFT
Jean-Marc Valin [Fri, 21 Mar 2008 01:47:41 +0000 (12:47 +1100)]
fixed-point: fix for 32-bit TI FFT

11 years agofixed-point: Wrapper for the 32-bit complex FFT used in the MDCT so we can use
Jean-Marc Valin [Fri, 21 Mar 2008 01:09:27 +0000 (12:09 +1100)]
fixed-point: Wrapper for the 32-bit complex FFT used in the MDCT so we can use
the TI dsplib FFT.

11 years agofixed-point: defined celt_maxabs16() as basic operator
Jean-Marc Valin [Fri, 21 Mar 2008 00:13:51 +0000 (11:13 +1100)]
fixed-point: defined celt_maxabs16() as basic operator

11 years agofixed-point: MULT16_32_Q15 for TI DSP (not entirely happy with it)
Jean-Marc Valin [Thu, 20 Mar 2008 23:53:09 +0000 (10:53 +1100)]
fixed-point: MULT16_32_Q15 for TI DSP (not entirely happy with it)

11 years agos/ROUND/ROUND16/
Jean-Marc Valin [Thu, 20 Mar 2008 23:46:17 +0000 (10:46 +1100)]
s/ROUND/ROUND16/

11 years agofixed-point: using TI intrinsic for celt_ilog2() if available.
Jean-Marc Valin [Thu, 20 Mar 2008 23:40:43 +0000 (10:40 +1100)]
fixed-point: using TI intrinsic for celt_ilog2() if available.

11 years agoWrapper for the TI dsplib FFT
Jean-Marc Valin [Thu, 20 Mar 2008 12:17:18 +0000 (23:17 +1100)]
Wrapper for the TI dsplib FFT

11 years agoMaking the real/single FFT easier to replace
Jean-Marc Valin [Thu, 20 Mar 2008 11:23:54 +0000 (22:23 +1100)]
Making the real/single FFT easier to replace

11 years agoRandom numbers should work on 16-bit archs.
Jean-Marc Valin [Tue, 18 Mar 2008 22:33:49 +0000 (09:33 +1100)]
Random numbers should work on 16-bit archs.

11 years agofixed-point: more TI macros. Comments on the existing ones.
Jean-Marc Valin [Sun, 16 Mar 2008 12:33:02 +0000 (23:33 +1100)]
fixed-point: more TI macros. Comments on the existing ones.

11 years agoRemoved useless functions that caused warnings in skeleton.c
Jean-Marc Valin [Sun, 16 Mar 2008 10:39:05 +0000 (21:39 +1100)]
Removed useless functions that caused warnings in skeleton.c

11 years agoMaking sure not to use the C library calls directly
Jean-Marc Valin [Sat, 15 Mar 2008 20:55:29 +0000 (07:55 +1100)]
Making sure not to use the C library calls directly

11 years agoSome work on assertions.
Jean-Marc Valin [Sat, 15 Mar 2008 12:34:39 +0000 (23:34 +1100)]
Some work on assertions.

11 years agoNew C55 macro
Jean-Marc Valin [Fri, 14 Mar 2008 21:12:49 +0000 (08:12 +1100)]
New C55 macro

11 years agofixed-point: Fixed a "shift too large" warning by using a 16-bit mul instead
Jean-Marc Valin [Fri, 14 Mar 2008 21:04:16 +0000 (08:04 +1100)]
fixed-point: Fixed a "shift too large" warning by using a 16-bit mul instead
of 32-bit in find_spectral_pitch()

11 years agofixed-point: playing it safe. SHL32() now automatically casts input to 32-bit
Jean-Marc Valin [Fri, 14 Mar 2008 06:58:33 +0000 (17:58 +1100)]
fixed-point: playing it safe. SHL32() now automatically casts input to 32-bit
to prevent surprises. Also, a few comments.

11 years agofixed-point: fixed a 16-bit issue in normalise_bands() where Q15ONE was directly
Jean-Marc Valin [Fri, 14 Mar 2008 05:28:28 +0000 (16:28 +1100)]
fixed-point: fixed a 16-bit issue in normalise_bands() where Q15ONE was directly
shifted left

11 years agoRemoved unnecessary header inclusions
Jean-Marc Valin [Fri, 14 Mar 2008 00:09:50 +0000 (11:09 +1100)]
Removed unnecessary header inclusions

11 years agoAdjusting/fixing warnings
Jean-Marc Valin [Thu, 13 Mar 2008 12:01:55 +0000 (23:01 +1100)]
Adjusting/fixing warnings

11 years agofixed-point: Using reciproqual instead of DIV32_16 for find_spectral_pitch()
Jean-Marc Valin [Thu, 13 Mar 2008 06:55:03 +0000 (17:55 +1100)]
fixed-point: Using reciproqual instead of DIV32_16 for find_spectral_pitch()

11 years agoEnabling warnings and restricting symbol visibility
Jean-Marc Valin [Thu, 13 Mar 2008 06:39:55 +0000 (17:39 +1100)]
Enabling warnings and restricting symbol visibility

11 years agoAdded mathops-test
Jean-Marc Valin [Thu, 13 Mar 2008 06:20:08 +0000 (17:20 +1100)]
Added mathops-test

11 years agofixed-point: simplification of the gain in mix_pitch_and_residual()
Jean-Marc Valin [Thu, 13 Mar 2008 00:36:45 +0000 (11:36 +1100)]
fixed-point: simplification of the gain in mix_pitch_and_residual()

11 years agofixed-point: simplifying the arithmetic in alg_quant()
Jean-Marc Valin [Thu, 13 Mar 2008 00:18:15 +0000 (11:18 +1100)]
fixed-point: simplifying the arithmetic in alg_quant()

11 years agoMore accurate sqrt approximation using MULT16_16_Q15() instead of Q14.
Jean-Marc Valin [Wed, 12 Mar 2008 12:00:42 +0000 (23:00 +1100)]
More accurate sqrt approximation using MULT16_16_Q15() instead of Q14.

11 years agoTODO update
Jean-Marc Valin [Wed, 12 Mar 2008 11:21:46 +0000 (22:21 +1100)]
TODO update

11 years agoIncreased accuracy of the reciprocal approximation and made it use
Jean-Marc Valin [Wed, 12 Mar 2008 10:34:50 +0000 (21:34 +1100)]
Increased accuracy of the reciprocal approximation and made it use
MULT16_16_Q15() instead of Q14 (should be faster).

11 years agoTestcases should now work even when symbols aren't visible in the dso
Jean-Marc Valin [Wed, 12 Mar 2008 07:04:27 +0000 (18:04 +1100)]
Testcases should now work even when symbols aren't visible in the dso

11 years agoAdded info for properly exporting symbols
Jean-Marc Valin [Wed, 12 Mar 2008 06:46:04 +0000 (17:46 +1100)]
Added info for properly exporting symbols

11 years agoChanged definition of VARDECL to make it more flexible. No actual code change.
Jean-Marc Valin [Wed, 12 Mar 2008 06:17:23 +0000 (17:17 +1100)]
Changed definition of VARDECL to make it more flexible. No actual code change.

11 years agoUsing reciprocal approximation instead of full 32-bit division in alg_quant()
Jean-Marc Valin [Wed, 12 Mar 2008 02:26:37 +0000 (13:26 +1100)]
Using reciprocal approximation instead of full 32-bit division in alg_quant()

11 years agoconfigure update (MIXED_PRECISION and STATIC_MODES can now be used)
Jean-Marc Valin [Wed, 12 Mar 2008 00:47:19 +0000 (11:47 +1100)]
configure update (MIXED_PRECISION and STATIC_MODES can now be used)

11 years agoDefined the celt_rcp() reciprocal approximation
Jean-Marc Valin [Wed, 12 Mar 2008 00:28:26 +0000 (11:28 +1100)]
Defined the celt_rcp() reciprocal approximation

11 years ago__GNUC_PREREQ should be causing problems anymore.
Jean-Marc Valin [Tue, 11 Mar 2008 23:42:24 +0000 (10:42 +1100)]
__GNUC_PREREQ should be causing problems anymore.

11 years agoAdded macro definitions for the TI C5x family (untested)
Jean-Marc Valin [Tue, 11 Mar 2008 23:16:26 +0000 (10:16 +1100)]
Added macro definitions for the TI C5x family (untested)

11 years agoBetter heuristic for deciding when to use the 32-bit version of the cwrs code.
Jean-Marc Valin [Tue, 11 Mar 2008 04:24:07 +0000 (15:24 +1100)]
Better heuristic for deciding when to use the 32-bit version of the cwrs code.

11 years agoconst correctness. celt_mode_create() now makes a copy of the static mode
Jean-Marc Valin [Mon, 10 Mar 2008 23:43:06 +0000 (10:43 +1100)]
const correctness. celt_mode_create() now makes a copy of the static mode
(if STATIC_MODES is defined) instead of attempting to modify read-only
storage.

11 years agooops, forgot to disable the main()
Jean-Marc Valin [Mon, 10 Mar 2008 05:44:07 +0000 (16:44 +1100)]
oops, forgot to disable the main()

11 years agoChanged some of the mode data from int to celt_int16_t to save memory.
Jean-Marc Valin [Mon, 10 Mar 2008 05:37:37 +0000 (16:37 +1100)]
Changed some of the mode data from int to celt_int16_t to save memory.

11 years agoNo longer include redundent entries in the bit allocation cache
Jean-Marc Valin [Mon, 10 Mar 2008 05:10:15 +0000 (16:10 +1100)]
No longer include redundent entries in the bit allocation cache

11 years agoAllocation cache can now be pre-computed as well.
Jean-Marc Valin [Mon, 10 Mar 2008 04:42:35 +0000 (15:42 +1100)]
Allocation cache can now be pre-computed as well.

11 years agoPsychoacoustic decay coefficients can now be included in the static modes
Jean-Marc Valin [Mon, 10 Mar 2008 04:07:58 +0000 (15:07 +1100)]
Psychoacoustic decay coefficients can now be included in the static modes

11 years agoImprovements to static modes
Jean-Marc Valin [Mon, 10 Mar 2008 03:52:18 +0000 (14:52 +1100)]
Improvements to static modes

11 years agoMoved the psycoacoustics data to the mode struct
Jean-Marc Valin [Mon, 10 Mar 2008 02:34:52 +0000 (13:34 +1100)]
Moved the psycoacoustics data to the mode struct

11 years agoSimplified spreading function so that only one set of coefficients (the right
Jean-Marc Valin [Mon, 10 Mar 2008 02:17:58 +0000 (13:17 +1100)]
Simplified spreading function so that only one set of coefficients (the right
slope is required).

11 years agoInitial support for static modes (defined STATIC_MODES)
Jean-Marc Valin [Mon, 10 Mar 2008 01:13:23 +0000 (12:13 +1100)]
Initial support for static modes (defined STATIC_MODES)

11 years agofixed-point: done converting find_spectral_pitch()
Jean-Marc Valin [Sun, 9 Mar 2008 23:41:13 +0000 (10:41 +1100)]
fixed-point: done converting find_spectral_pitch()

11 years agofixed some build problems
Jean-Marc Valin [Sun, 9 Mar 2008 23:25:40 +0000 (10:25 +1100)]
fixed some build problems

11 years agotestcelt no longer attempts to use ALLOC()
Jean-Marc Valin [Sun, 9 Mar 2008 08:27:58 +0000 (19:27 +1100)]
testcelt no longer attempts to use ALLOC()

11 years agomanual stack allocator should now work for sizeof(char)!=1
Jean-Marc Valin [Sun, 9 Mar 2008 06:18:39 +0000 (17:18 +1100)]
manual stack allocator should now work for sizeof(char)!=1

11 years agofixed-point: stereo_mix() converted.
Jean-Marc Valin [Sat, 8 Mar 2008 09:17:12 +0000 (20:17 +1100)]
fixed-point: stereo_mix() converted.

11 years agocleaning up some of the mode stuff
Jean-Marc Valin [Sat, 8 Mar 2008 04:21:24 +0000 (15:21 +1100)]
cleaning up some of the mode stuff

11 years agooops, forgot an int sampling rate somewhere
Jean-Marc Valin [Fri, 7 Mar 2008 23:02:30 +0000 (10:02 +1100)]
oops, forgot an int sampling rate somewhere

11 years agosampling rate has to be int32
Jean-Marc Valin [Fri, 7 Mar 2008 20:39:47 +0000 (07:39 +1100)]
sampling rate has to be int32

11 years agoAdding code to dump the contents of a mode struct to a C file that can be
Jean-Marc Valin [Fri, 7 Mar 2008 11:25:27 +0000 (22:25 +1100)]
Adding code to dump the contents of a mode struct to a C file that can be
compiled without having to use the initialisation code.

11 years agofixed-point: defined HALF32() and used it for the forward mdct.
Jean-Marc Valin [Fri, 7 Mar 2008 06:50:45 +0000 (17:50 +1100)]
fixed-point: defined HALF32() and used it for the forward mdct.

11 years agofixed-point: real FFT no longer needs float ops when MIXED_PRECISION is enabled.
Jean-Marc Valin [Fri, 7 Mar 2008 06:46:00 +0000 (17:46 +1100)]
fixed-point: real FFT no longer needs float ops when MIXED_PRECISION is enabled.

11 years agoBit of cleaning up. No real code change (well, I hope so!).
Jean-Marc Valin [Fri, 7 Mar 2008 06:38:58 +0000 (17:38 +1100)]
Bit of cleaning up. No real code change (well, I hope so!).

11 years agofixed-point: removed the last float op in celt.c
Jean-Marc Valin [Fri, 7 Mar 2008 06:33:58 +0000 (17:33 +1100)]
fixed-point: removed the last float op in celt.c

11 years agofixed-point: finished intra_prediction(). No float ops left in vq.c
Jean-Marc Valin [Fri, 7 Mar 2008 06:25:47 +0000 (17:25 +1100)]
fixed-point: finished intra_prediction(). No float ops left in vq.c

11 years agofixed-point: converting intra-frame predictor (checkpoint #1).
Jean-Marc Valin [Fri, 7 Mar 2008 06:17:37 +0000 (17:17 +1100)]
fixed-point: converting intra-frame predictor (checkpoint #1).

11 years agofixed-point: intra_fold() converted
Jean-Marc Valin [Fri, 7 Mar 2008 06:05:47 +0000 (17:05 +1100)]
fixed-point: intra_fold() converted

11 years agofixed-point: compute_band_energies() converted. Had to add some tiny bias to
Jean-Marc Valin [Fri, 7 Mar 2008 05:42:31 +0000 (16:42 +1100)]
fixed-point: compute_band_energies() converted. Had to add some tiny bias to
make sure never to end up with a pitch vector with >1 norm.

11 years agofixed-point: converted normalise_bands (had to split implementation)
Jean-Marc Valin [Fri, 7 Mar 2008 04:33:32 +0000 (15:33 +1100)]
fixed-point: converted normalise_bands (had to split implementation)

11 years agoRevert "compute_band_energies() merged with normalised_bands()"
Jean-Marc Valin [Fri, 7 Mar 2008 03:53:17 +0000 (14:53 +1100)]
Revert "compute_band_energies() merged with normalised_bands()"

This reverts commit a1bc18a38836e2b1360a26a1d7ef2571b2f3b694.
I think it's easier for now to deal with two functions here, even though they
could be merged.