diff options
author | Nils Wallménius <nils@rockbox.org> | 2013-05-20 22:25:57 +0200 |
---|---|---|
committer | Nils Wallménius <nils@rockbox.org> | 2013-08-31 08:30:51 +0200 |
commit | 580b307fd791c0997a8831bc800bba87797bfb7e (patch) | |
tree | 807846056f06fd944a750ce41217a877910ebd59 /lib/rbcodec/codecs/libopus/celt/celt_lpc.c | |
parent | 74761b70acd96cecc0d35450dd56a98ad9ee7d3d (diff) | |
download | rockbox-580b307fd791c0997a8831bc800bba87797bfb7e.tar.gz rockbox-580b307fd791c0997a8831bc800bba87797bfb7e.zip |
Sync opus codec to upstream git
Sync opus codec to upstream commit
02fed471a4568852d6618e041c4f2af0d7730ee2 (August 30 2013)
This brings in a lot of optimizations but also makes the diff
between our codec and the upstream much smaller as most of our
optimizations have been upstreamed or supeceded.
Speedups across the board for CELT mode files:
64kbps 128kbps
H300 9.82MHz 15.48MHz
c200 4.86MHz 9.63MHz
fuze v1 10.32MHz 15.92MHz
For the silk mode test file (16kbps) arm targets get a speedup
of about 2MHz while the H300 is 7.8MHz slower, likely because it's
now using the pseudostack more rather than the real stack which
is in iram. Patches to get around that are upcomming.
Change-Id: Ifecf963e461c51ac42e09dac1e91bc4bc3b12fa3
Diffstat (limited to 'lib/rbcodec/codecs/libopus/celt/celt_lpc.c')
-rw-r--r-- | lib/rbcodec/codecs/libopus/celt/celt_lpc.c | 196 |
1 files changed, 158 insertions, 38 deletions
diff --git a/lib/rbcodec/codecs/libopus/celt/celt_lpc.c b/lib/rbcodec/codecs/libopus/celt/celt_lpc.c index 66aed1de09..7ffe90a357 100644 --- a/lib/rbcodec/codecs/libopus/celt/celt_lpc.c +++ b/lib/rbcodec/codecs/libopus/celt/celt_lpc.c | |||
@@ -26,12 +26,13 @@ | |||
26 | */ | 26 | */ |
27 | 27 | ||
28 | #ifdef HAVE_CONFIG_H | 28 | #ifdef HAVE_CONFIG_H |
29 | #include "opus_config.h" | 29 | #include "config.h" |
30 | #endif | 30 | #endif |
31 | 31 | ||
32 | #include "celt_lpc.h" | 32 | #include "celt_lpc.h" |
33 | #include "stack_alloc.h" | 33 | #include "stack_alloc.h" |
34 | #include "mathops.h" | 34 | #include "mathops.h" |
35 | #include "pitch.h" | ||
35 | 36 | ||
36 | void _celt_lpc( | 37 | void _celt_lpc( |
37 | opus_val16 *_lpc, /* out: [0...p-1] LPC coefficients */ | 38 | opus_val16 *_lpc, /* out: [0...p-1] LPC coefficients */ |
@@ -87,42 +88,71 @@ int p | |||
87 | #endif | 88 | #endif |
88 | } | 89 | } |
89 | 90 | ||
90 | void celt_fir(const opus_val16 *x, | 91 | void celt_fir(const opus_val16 *_x, |
91 | const opus_val16 *num, | 92 | const opus_val16 *num, |
92 | opus_val16 *y, | 93 | opus_val16 *_y, |
93 | int N, | 94 | int N, |
94 | int ord, | 95 | int ord, |
95 | opus_val16 *mem) | 96 | opus_val16 *mem) |
96 | { | 97 | { |
97 | int i,j; | 98 | int i,j; |
99 | VARDECL(opus_val16, rnum); | ||
100 | VARDECL(opus_val16, x); | ||
101 | SAVE_STACK; | ||
98 | 102 | ||
103 | ALLOC(rnum, ord, opus_val16); | ||
104 | ALLOC(x, N+ord, opus_val16); | ||
105 | for(i=0;i<ord;i++) | ||
106 | rnum[i] = num[ord-i-1]; | ||
107 | for(i=0;i<ord;i++) | ||
108 | x[i] = mem[ord-i-1]; | ||
109 | for (i=0;i<N;i++) | ||
110 | x[i+ord]=_x[i]; | ||
111 | for(i=0;i<ord;i++) | ||
112 | mem[i] = _x[N-i-1]; | ||
113 | #ifdef SMALL_FOOTPRINT | ||
99 | for (i=0;i<N;i++) | 114 | for (i=0;i<N;i++) |
100 | { | 115 | { |
101 | opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT); | 116 | opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT); |
102 | for (j=0;j<ord;j++) | 117 | for (j=0;j<ord;j++) |
103 | { | 118 | { |
104 | sum += MULT16_16(num[j],mem[j]); | 119 | sum = MAC16_16(sum,rnum[j],x[i+j]); |
105 | } | ||
106 | for (j=ord-1;j>=1;j--) | ||
107 | { | ||
108 | mem[j]=mem[j-1]; | ||
109 | } | 120 | } |
110 | mem[0] = x[i]; | 121 | _y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT)); |
111 | y[i] = ROUND16(sum, SIG_SHIFT); | ||
112 | } | 122 | } |
123 | #else | ||
124 | for (i=0;i<N-3;i+=4) | ||
125 | { | ||
126 | opus_val32 sum[4]={0,0,0,0}; | ||
127 | xcorr_kernel(rnum, x+i, sum, ord); | ||
128 | _y[i ] = SATURATE16(ADD32(EXTEND32(_x[i ]), PSHR32(sum[0], SIG_SHIFT))); | ||
129 | _y[i+1] = SATURATE16(ADD32(EXTEND32(_x[i+1]), PSHR32(sum[1], SIG_SHIFT))); | ||
130 | _y[i+2] = SATURATE16(ADD32(EXTEND32(_x[i+2]), PSHR32(sum[2], SIG_SHIFT))); | ||
131 | _y[i+3] = SATURATE16(ADD32(EXTEND32(_x[i+3]), PSHR32(sum[3], SIG_SHIFT))); | ||
132 | } | ||
133 | for (;i<N;i++) | ||
134 | { | ||
135 | opus_val32 sum = 0; | ||
136 | for (j=0;j<ord;j++) | ||
137 | sum = MAC16_16(sum,rnum[j],x[i+j]); | ||
138 | _y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT))); | ||
139 | } | ||
140 | #endif | ||
141 | RESTORE_STACK; | ||
113 | } | 142 | } |
114 | 143 | ||
115 | void celt_iir(const opus_val32 *x, | 144 | void celt_iir(const opus_val32 *_x, |
116 | const opus_val16 *den, | 145 | const opus_val16 *den, |
117 | opus_val32 *y, | 146 | opus_val32 *_y, |
118 | int N, | 147 | int N, |
119 | int ord, | 148 | int ord, |
120 | opus_val16 *mem) | 149 | opus_val16 *mem) |
121 | { | 150 | { |
151 | #ifdef SMALL_FOOTPRINT | ||
122 | int i,j; | 152 | int i,j; |
123 | for (i=0;i<N;i++) | 153 | for (i=0;i<N;i++) |
124 | { | 154 | { |
125 | opus_val32 sum = x[i]; | 155 | opus_val32 sum = _x[i]; |
126 | for (j=0;j<ord;j++) | 156 | for (j=0;j<ord;j++) |
127 | { | 157 | { |
128 | sum -= MULT16_16(den[j],mem[j]); | 158 | sum -= MULT16_16(den[j],mem[j]); |
@@ -132,11 +162,65 @@ void celt_iir(const opus_val32 *x, | |||
132 | mem[j]=mem[j-1]; | 162 | mem[j]=mem[j-1]; |
133 | } | 163 | } |
134 | mem[0] = ROUND16(sum,SIG_SHIFT); | 164 | mem[0] = ROUND16(sum,SIG_SHIFT); |
135 | y[i] = sum; | 165 | _y[i] = sum; |
136 | } | 166 | } |
167 | #else | ||
168 | int i,j; | ||
169 | VARDECL(opus_val16, rden); | ||
170 | VARDECL(opus_val16, y); | ||
171 | SAVE_STACK; | ||
172 | |||
173 | celt_assert((ord&3)==0); | ||
174 | ALLOC(rden, ord, opus_val16); | ||
175 | ALLOC(y, N+ord, opus_val16); | ||
176 | for(i=0;i<ord;i++) | ||
177 | rden[i] = den[ord-i-1]; | ||
178 | for(i=0;i<ord;i++) | ||
179 | y[i] = -mem[ord-i-1]; | ||
180 | for(;i<N+ord;i++) | ||
181 | y[i]=0; | ||
182 | for (i=0;i<N-3;i+=4) | ||
183 | { | ||
184 | /* Unroll by 4 as if it were an FIR filter */ | ||
185 | opus_val32 sum[4]; | ||
186 | sum[0]=_x[i]; | ||
187 | sum[1]=_x[i+1]; | ||
188 | sum[2]=_x[i+2]; | ||
189 | sum[3]=_x[i+3]; | ||
190 | xcorr_kernel(rden, y+i, sum, ord); | ||
191 | |||
192 | /* Patch up the result to compensate for the fact that this is an IIR */ | ||
193 | y[i+ord ] = -ROUND16(sum[0],SIG_SHIFT); | ||
194 | _y[i ] = sum[0]; | ||
195 | sum[1] = MAC16_16(sum[1], y[i+ord ], den[0]); | ||
196 | y[i+ord+1] = -ROUND16(sum[1],SIG_SHIFT); | ||
197 | _y[i+1] = sum[1]; | ||
198 | sum[2] = MAC16_16(sum[2], y[i+ord+1], den[0]); | ||
199 | sum[2] = MAC16_16(sum[2], y[i+ord ], den[1]); | ||
200 | y[i+ord+2] = -ROUND16(sum[2],SIG_SHIFT); | ||
201 | _y[i+2] = sum[2]; | ||
202 | |||
203 | sum[3] = MAC16_16(sum[3], y[i+ord+2], den[0]); | ||
204 | sum[3] = MAC16_16(sum[3], y[i+ord+1], den[1]); | ||
205 | sum[3] = MAC16_16(sum[3], y[i+ord ], den[2]); | ||
206 | y[i+ord+3] = -ROUND16(sum[3],SIG_SHIFT); | ||
207 | _y[i+3] = sum[3]; | ||
208 | } | ||
209 | for (;i<N;i++) | ||
210 | { | ||
211 | opus_val32 sum = _x[i]; | ||
212 | for (j=0;j<ord;j++) | ||
213 | sum -= MULT16_16(rden[j],y[i+j]); | ||
214 | y[i+ord] = ROUND16(sum,SIG_SHIFT); | ||
215 | _y[i] = sum; | ||
216 | } | ||
217 | for(i=0;i<ord;i++) | ||
218 | mem[i] = _y[N-i-1]; | ||
219 | RESTORE_STACK; | ||
220 | #endif | ||
137 | } | 221 | } |
138 | 222 | ||
139 | void _celt_autocorr( | 223 | int _celt_autocorr( |
140 | const opus_val16 *x, /* in: [0...n-1] samples x */ | 224 | const opus_val16 *x, /* in: [0...n-1] samples x */ |
141 | opus_val32 *ac, /* out: [0...lag-1] ac values */ | 225 | opus_val32 *ac, /* out: [0...lag-1] ac values */ |
142 | const opus_val16 *window, | 226 | const opus_val16 *window, |
@@ -146,43 +230,79 @@ void _celt_autocorr( | |||
146 | ) | 230 | ) |
147 | { | 231 | { |
148 | opus_val32 d; | 232 | opus_val32 d; |
149 | int i; | 233 | int i, k; |
234 | int fastN=n-lag; | ||
235 | int shift; | ||
236 | const opus_val16 *xptr; | ||
150 | VARDECL(opus_val16, xx); | 237 | VARDECL(opus_val16, xx); |
151 | SAVE_STACK; | 238 | SAVE_STACK; |
152 | ALLOC(xx, n, opus_val16); | 239 | ALLOC(xx, n, opus_val16); |
153 | celt_assert(n>0); | 240 | celt_assert(n>0); |
154 | celt_assert(overlap>=0); | 241 | celt_assert(overlap>=0); |
155 | for (i=0;i<n;i++) | 242 | if (overlap == 0) |
156 | xx[i] = x[i]; | ||
157 | for (i=0;i<overlap;i++) | ||
158 | { | 243 | { |
159 | xx[i] = MULT16_16_Q15(x[i],window[i]); | 244 | xptr = x; |
160 | xx[n-i-1] = MULT16_16_Q15(x[n-i-1],window[i]); | 245 | } else { |
246 | for (i=0;i<n;i++) | ||
247 | xx[i] = x[i]; | ||
248 | for (i=0;i<overlap;i++) | ||
249 | { | ||
250 | xx[i] = MULT16_16_Q15(x[i],window[i]); | ||
251 | xx[n-i-1] = MULT16_16_Q15(x[n-i-1],window[i]); | ||
252 | } | ||
253 | xptr = xx; | ||
161 | } | 254 | } |
255 | shift=0; | ||
162 | #ifdef FIXED_POINT | 256 | #ifdef FIXED_POINT |
163 | { | 257 | { |
164 | opus_val32 ac0=0; | 258 | opus_val32 ac0; |
165 | int shift; | 259 | ac0 = 1+(n<<7); |
166 | for(i=0;i<n;i++) | 260 | if (n&1) ac0 += SHR32(MULT16_16(xptr[0],xptr[0]),9); |
167 | ac0 += SHR32(MULT16_16(xx[i],xx[i]),9); | 261 | for(i=(n&1);i<n;i+=2) |
168 | ac0 += 1+n; | 262 | { |
263 | ac0 += SHR32(MULT16_16(xptr[i],xptr[i]),9); | ||
264 | ac0 += SHR32(MULT16_16(xptr[i+1],xptr[i+1]),9); | ||
265 | } | ||
169 | 266 | ||
170 | shift = celt_ilog2(ac0)-30+10; | 267 | shift = celt_ilog2(ac0)-30+10; |
171 | shift = (shift+1)/2; | 268 | shift = (shift)/2; |
172 | for(i=0;i<n;i++) | 269 | if (shift>0) |
173 | xx[i] = VSHR32(xx[i], shift); | 270 | { |
271 | for(i=0;i<n;i++) | ||
272 | xx[i] = PSHR32(xptr[i], shift); | ||
273 | xptr = xx; | ||
274 | } else | ||
275 | shift = 0; | ||
174 | } | 276 | } |
175 | #endif | 277 | #endif |
176 | while (lag>=0) | 278 | celt_pitch_xcorr(xptr, xptr, ac, fastN, lag+1); |
279 | for (k=0;k<=lag;k++) | ||
177 | { | 280 | { |
178 | for (i = lag, d = 0; i < n; i++) | 281 | for (i = k+fastN, d = 0; i < n; i++) |
179 | d += xx[i] * xx[i-lag]; | 282 | d = MAC16_16(d, xptr[i], xptr[i-k]); |
180 | ac[lag] = d; | 283 | ac[k] += d; |
181 | /*printf ("%f ", ac[lag]);*/ | ||
182 | lag--; | ||
183 | } | 284 | } |
184 | /*printf ("\n");*/ | 285 | #ifdef FIXED_POINT |
185 | ac[0] += 10; | 286 | shift = 2*shift; |
287 | if (shift<=0) | ||
288 | ac[0] += SHL32((opus_int32)1, -shift); | ||
289 | if (ac[0] < 268435456) | ||
290 | { | ||
291 | int shift2 = 29 - EC_ILOG(ac[0]); | ||
292 | for (i=0;i<=lag;i++) | ||
293 | ac[i] = SHL32(ac[i], shift2); | ||
294 | shift -= shift2; | ||
295 | } else if (ac[0] >= 536870912) | ||
296 | { | ||
297 | int shift2=1; | ||
298 | if (ac[0] >= 1073741824) | ||
299 | shift2++; | ||
300 | for (i=0;i<=lag;i++) | ||
301 | ac[i] = SHR32(ac[i], shift2); | ||
302 | shift += shift2; | ||
303 | } | ||
304 | #endif | ||
186 | 305 | ||
187 | RESTORE_STACK; | 306 | RESTORE_STACK; |
307 | return shift; | ||
188 | } | 308 | } |