• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * This file is part of FFmpeg.
3  *
4  * FFmpeg is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2.1 of the License, or (at your option) any later version.
8  *
9  * FFmpeg is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with FFmpeg; if not, write to the Free Software
16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17  */
18 
19 #ifndef AVUTIL_TX_PRIV_H
20 #define AVUTIL_TX_PRIV_H
21 
22 #include "tx.h"
23 #include "thread.h"
24 #include "mem_internal.h"
25 #include "attributes.h"
26 
27 #ifdef TX_FLOAT
28 #define TX_TAB(x) x ## _float
29 #define TX_NAME(x) x ## _float_c
30 #define TX_NAME_STR(x) NULL_IF_CONFIG_SMALL(x "_float_c")
31 #define TX_TYPE(x) AV_TX_FLOAT_ ## x
32 #define TX_FN_NAME(fn, suffix) ff_tx_ ## fn ## _float_ ## suffix
33 #define TX_FN_NAME_STR(fn, suffix) NULL_IF_CONFIG_SMALL(#fn "_float_" #suffix)
34 #define MULT(x, m) ((x) * (m))
35 #define SCALE_TYPE float
36 typedef float TXSample;
37 typedef float TXUSample;
38 typedef AVComplexFloat TXComplex;
39 #elif defined(TX_DOUBLE)
40 #define TX_TAB(x) x ## _double
41 #define TX_NAME(x) x ## _double_c
42 #define TX_NAME_STR(x) NULL_IF_CONFIG_SMALL(x "_double_c")
43 #define TX_TYPE(x) AV_TX_DOUBLE_ ## x
44 #define TX_FN_NAME(fn, suffix) ff_tx_ ## fn ## _double_ ## suffix
45 #define TX_FN_NAME_STR(fn, suffix) NULL_IF_CONFIG_SMALL(#fn "_double_" #suffix)
46 #define MULT(x, m) ((x) * (m))
47 #define SCALE_TYPE double
48 typedef double TXSample;
49 typedef double TXUSample;
50 typedef AVComplexDouble TXComplex;
51 #elif defined(TX_INT32)
52 #define TX_TAB(x) x ## _int32
53 #define TX_NAME(x) x ## _int32_c
54 #define TX_NAME_STR(x) NULL_IF_CONFIG_SMALL(x "_int32_c")
55 #define TX_TYPE(x) AV_TX_INT32_ ## x
56 #define TX_FN_NAME(fn, suffix) ff_tx_ ## fn ## _int32_ ## suffix
57 #define TX_FN_NAME_STR(fn, suffix) NULL_IF_CONFIG_SMALL(#fn "_int32_" #suffix)
58 #define MULT(x, m) (((((int64_t)(x)) * (int64_t)(m)) + 0x40000000) >> 31)
59 #define SCALE_TYPE float
60 typedef int32_t TXSample;
61 typedef uint32_t TXUSample;
62 typedef AVComplexInt32 TXComplex;
63 #else
64 typedef void TXComplex;
65 #endif
66 
67 #define TX_DECL_FN(fn, suffix) \
68     void TX_FN_NAME(fn, suffix)(AVTXContext *s, void *o, void *i, ptrdiff_t st);
69 
70 #define TX_DEF(fn, tx_type, len_min, len_max, f1, f2,                          \
71                p, init_fn, suffix, cf, cd_flags, cf2)                          \
72     &(const FFTXCodelet){                                                      \
73         .name       = TX_FN_NAME_STR(fn, suffix),                              \
74         .function   = TX_FN_NAME(fn, suffix),                                  \
75         .type       = TX_TYPE(tx_type),                                        \
76         .flags      = FF_TX_ALIGNED | FF_TX_OUT_OF_PLACE | cd_flags,           \
77         .factors    = { f1, f2 },                                              \
78         .min_len    = len_min,                                                 \
79         .max_len    = len_max,                                                 \
80         .init       = init_fn,                                                 \
81         .cpu_flags  = cf2 | AV_CPU_FLAG_ ## cf,                                \
82         .prio       = p,                                                       \
83     }
84 
85 #if defined(TX_FLOAT) || defined(TX_DOUBLE)
86 
87 #define CMUL(dre, dim, are, aim, bre, bim)      \
88     do {                                        \
89         (dre) = (are) * (bre) - (aim) * (bim);  \
90         (dim) = (are) * (bim) + (aim) * (bre);  \
91     } while (0)
92 
93 #define SMUL(dre, dim, are, aim, bre, bim)      \
94     do {                                        \
95         (dre) = (are) * (bre) - (aim) * (bim);  \
96         (dim) = (are) * (bim) - (aim) * (bre);  \
97     } while (0)
98 
99 #define UNSCALE(x) (x)
100 #define RESCALE(x) (x)
101 
102 #define FOLD(a, b) ((a) + (b))
103 
104 #define BF(x, y, a, b)  \
105     do {                \
106         x = (a) - (b);  \
107         y = (a) + (b);  \
108     } while (0)
109 
110 #elif defined(TX_INT32)
111 
112 /* Properly rounds the result */
113 #define CMUL(dre, dim, are, aim, bre, bim)             \
114     do {                                               \
115         int64_t accu;                                  \
116         (accu)  = (int64_t)(bre) * (are);              \
117         (accu) -= (int64_t)(bim) * (aim);              \
118         (dre)   = (int)(((accu) + 0x40000000) >> 31);  \
119         (accu)  = (int64_t)(bim) * (are);              \
120         (accu) += (int64_t)(bre) * (aim);              \
121         (dim)   = (int)(((accu) + 0x40000000) >> 31);  \
122     } while (0)
123 
124 #define SMUL(dre, dim, are, aim, bre, bim)             \
125     do {                                               \
126         int64_t accu;                                  \
127         (accu)  = (int64_t)(bre) * (are);              \
128         (accu) -= (int64_t)(bim) * (aim);              \
129         (dre)   = (int)(((accu) + 0x40000000) >> 31);  \
130         (accu)  = (int64_t)(bim) * (are);              \
131         (accu) -= (int64_t)(bre) * (aim);              \
132         (dim)   = (int)(((accu) + 0x40000000) >> 31);  \
133     } while (0)
134 
135 #define UNSCALE(x) ((double)(x)/2147483648.0)
136 #define RESCALE(x) (av_clip64(lrintf((x) * 2147483648.0), INT32_MIN, INT32_MAX))
137 
138 #define FOLD(x, y) ((int32_t)((x) + (unsigned)(y) + 32) >> 6)
139 
140 #define BF(x, y, a, b)  \
141     do {                \
142         x = (a) - (unsigned)(b);  \
143         y = (a) + (unsigned)(b);  \
144     } while (0)
145 
146 #endif /* TX_INT32 */
147 
148 #define CMUL3(c, a, b) CMUL((c).re, (c).im, (a).re, (a).im, (b).re, (b).im)
149 
150 /* Codelet flags, used to pick codelets. Must be a superset of enum AVTXFlags,
151  * but if it runs out of bits, it can be made separate. */
152 #define FF_TX_OUT_OF_PLACE (1ULL << 63) /* Can be OR'd with AV_TX_INPLACE             */
153 #define FF_TX_ALIGNED      (1ULL << 62) /* Cannot be OR'd with AV_TX_UNALIGNED        */
154 #define FF_TX_PRESHUFFLE   (1ULL << 61) /* Codelet expects permuted coeffs            */
155 #define FF_TX_INVERSE_ONLY (1ULL << 60) /* For non-orthogonal inverse-only transforms */
156 #define FF_TX_FORWARD_ONLY (1ULL << 59) /* For non-orthogonal forward-only transforms */
157 
158 typedef enum FFTXCodeletPriority {
159     FF_TX_PRIO_BASE = 0,               /* Baseline priority */
160 
161     /* For SIMD, set base prio to the register size in bits and increment in
162      * steps of 64 depending on faster/slower features, like FMA. */
163 
164     FF_TX_PRIO_MIN          = -131072, /* For naive implementations */
165     FF_TX_PRIO_MAX          =  32768,  /* For custom implementations/ASICs */
166 } FFTXCodeletPriority;
167 
168 /* Codelet options */
169 typedef struct FFTXCodeletOptions {
170     int invert_lookup;     /* If codelet is flagged as FF_TX_CODELET_PRESHUFFLE,
171                               invert the lookup direction for the map generated */
172 } FFTXCodeletOptions;
173 
174 /* Maximum amount of subtransform functions, subtransforms and factors. Arbitrary. */
175 #define TX_MAX_SUB 4
176 
177 typedef struct FFTXCodelet {
178     const char    *name;          /* Codelet name, for debugging */
179     av_tx_fn       function;      /* Codelet function, != NULL */
180     enum AVTXType  type;          /* Type of codelet transform */
181 #define TX_TYPE_ANY INT32_MAX     /* Special type to allow all types */
182 
183     uint64_t flags;               /* A combination of AVTXFlags and codelet
184                                    * flags that describe its properties. */
185 
186     int factors[TX_MAX_SUB];      /* Length factors */
187 #define TX_FACTOR_ANY -1          /* When used alone, signals that the codelet
188                                    * supports all factors. Otherwise, if other
189                                    * factors are present, it signals that whatever
190                                    * remains will be supported, as long as the
191                                    * other factors are a component of the length */
192 
193     int min_len;                  /* Minimum length of transform, must be >= 1 */
194     int max_len;                  /* Maximum length of transform */
195 #define TX_LEN_UNLIMITED -1       /* Special length value to permit all lengths */
196 
197     int (*init)(AVTXContext *s,   /* Optional callback for current context initialization. */
198                 const struct FFTXCodelet *cd,
199                 uint64_t flags,
200                 FFTXCodeletOptions *opts,
201                 int len, int inv,
202                 const void *scale);
203 
204     int (*uninit)(AVTXContext *s); /* Optional callback for uninitialization. */
205 
206     int cpu_flags;                 /* CPU flags. If any negative flags like
207                                     * SLOW are present, will avoid picking.
208                                     * 0x0 to signal it's a C codelet */
209 #define FF_TX_CPU_FLAGS_ALL 0x0    /* Special CPU flag for C */
210 
211     int prio;                      /* < 0 = least, 0 = no pref, > 0 = prefer */
212 } FFTXCodelet;
213 
214 struct AVTXContext {
215     /* Fields the root transform and subtransforms use or may use.
216      * NOTE: This section is used by assembly, do not reorder or change */
217     int                len;             /* Length of the transform */
218     int                inv;             /* If transform is inverse */
219     int               *map;             /* Lookup table(s) */
220     TXComplex         *exp;             /* Any non-pre-baked multiplication factors needed */
221     TXComplex         *tmp;             /* Temporary buffer, if needed */
222 
223     AVTXContext       *sub;             /* Subtransform context(s), if needed */
224     av_tx_fn           fn[TX_MAX_SUB];  /* Function(s) for the subtransforms */
225     int                nb_sub;          /* Number of subtransforms.
226                                          * The reason all of these are set here
227                                          * rather than in each separate context
228                                          * is to eliminate extra pointer
229                                          * dereferences. */
230 
231     /* Fields mainly useul/applicable for the root transform or initialization.
232      * Fields below are not used by assembly code. */
233     const FFTXCodelet *cd[TX_MAX_SUB];  /* Subtransform codelets */
234     const FFTXCodelet *cd_self;         /* Codelet for the current context */
235     enum AVTXType      type;            /* Type of transform */
236     uint64_t           flags;           /* A combination of AVTXFlags and
237                                          * codelet flags used when creating */
238     float              scale_f;
239     double             scale_d;
240     void              *opaque;          /* Free to use by implementations */
241 };
242 
243 /* Create a subtransform in the current context with the given parameters.
244  * The flags parameter from FFTXCodelet.init() should be preserved as much
245  * as that's possible.
246  * MUST be called during the sub() callback of each codelet. */
247 int ff_tx_init_subtx(AVTXContext *s, enum AVTXType type,
248                      uint64_t flags, FFTXCodeletOptions *opts,
249                      int len, int inv, const void *scale);
250 
251 /*
252  * Generates the PFA permutation table into AVTXContext->pfatab. The end table
253  * is appended to the start table.
254  */
255 int ff_tx_gen_compound_mapping(AVTXContext *s, int n, int m);
256 
257 /*
258  * Generates a standard-ish (slightly modified) Split-Radix revtab into
259  * AVTXContext->map. Invert lookup changes how the mapping needs to be applied.
260  * If it's set to 0, it has to be applied like out[map[i]] = in[i], otherwise
261  * if it's set to 1, has to be applied as out[i] = in[map[i]]
262  */
263 int ff_tx_gen_ptwo_revtab(AVTXContext *s, int invert_lookup);
264 
265 /*
266  * Generates an index into AVTXContext->inplace_idx that if followed in the
267  * specific order, allows the revtab to be done in-place. The sub-transform
268  * and its map should already be initialized.
269  */
270 int ff_tx_gen_ptwo_inplace_revtab_idx(AVTXContext *s);
271 
272 /*
273  * This generates a parity-based revtab of length len and direction inv.
274  *
275  * Parity means even and odd complex numbers will be split, e.g. the even
276  * coefficients will come first, after which the odd coefficients will be
277  * placed. For example, a 4-point transform's coefficients after reordering:
278  * z[0].re, z[0].im, z[2].re, z[2].im, z[1].re, z[1].im, z[3].re, z[3].im
279  *
280  * The basis argument is the length of the largest non-composite transform
281  * supported, and also implies that the basis/2 transform is supported as well,
282  * as the split-radix algorithm requires it to be.
283  *
284  * The dual_stride argument indicates that both the basis, as well as the
285  * basis/2 transforms support doing two transforms at once, and the coefficients
286  * will be interleaved between each pair in a split-radix like so (stride == 2):
287  * tx1[0], tx1[2], tx2[0], tx2[2], tx1[1], tx1[3], tx2[1], tx2[3]
288  * A non-zero number switches this on, with the value indicating the stride
289  * (how many values of 1 transform to put first before switching to the other).
290  * Must be a power of two or 0. Must be less than the basis.
291  * Value will be clipped to the transform size, so for a basis of 16 and a
292  * dual_stride of 8, dual 8-point transforms will be laid out as if dual_stride
293  * was set to 4.
294  * Usually you'll set this to half the complex numbers that fit in a single
295  * register or 0. This allows to reuse SSE functions as dual-transform
296  * functions in AVX mode.
297  *
298  * If length is smaller than basis/2 this function will not do anything.
299  */
300 int ff_tx_gen_split_radix_parity_revtab(AVTXContext *s, int invert_lookup,
301                                         int basis, int dual_stride);
302 
303 /* Typed init function to initialize shared tables. Will initialize all tables
304  * for all factors of a length. */
305 void ff_tx_init_tabs_float (int len);
306 void ff_tx_init_tabs_double(int len);
307 void ff_tx_init_tabs_int32 (int len);
308 
309 /* Typed init function to initialize an MDCT exptab in a context. */
310 int  ff_tx_mdct_gen_exp_float (AVTXContext *s);
311 int  ff_tx_mdct_gen_exp_double(AVTXContext *s);
312 int  ff_tx_mdct_gen_exp_int32 (AVTXContext *s);
313 
314 /* Lists of codelets */
315 extern const FFTXCodelet * const ff_tx_codelet_list_float_c       [];
316 extern const FFTXCodelet * const ff_tx_codelet_list_float_x86     [];
317 
318 extern const FFTXCodelet * const ff_tx_codelet_list_double_c      [];
319 
320 extern const FFTXCodelet * const ff_tx_codelet_list_int32_c       [];
321 
322 #endif /* AVUTIL_TX_PRIV_H */
323