| 1 | /* |
| 2 | * Copyright 2005 Balatoni Denes |
| 3 | * Copyright 2006 Loren Merritt |
| 4 | * |
| 5 | * This file is part of FFmpeg. |
| 6 | * |
| 7 | * FFmpeg is free software; you can redistribute it and/or |
| 8 | * modify it under the terms of the GNU Lesser General Public |
| 9 | * License as published by the Free Software Foundation; either |
| 10 | * version 2.1 of the License, or (at your option) any later version. |
| 11 | * |
| 12 | * FFmpeg is distributed in the hope that it will be useful, |
| 13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 15 | * Lesser General Public License for more details. |
| 16 | * |
| 17 | * You should have received a copy of the GNU Lesser General Public |
| 18 | * License along with FFmpeg; if not, write to the Free Software |
| 19 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 20 | */ |
| 21 | |
| 22 | #include "config.h" |
| 23 | #include "attributes.h" |
| 24 | #include "float_dsp.h" |
| 25 | |
| 26 | static void vector_fmul_c(float *dst, const float *src0, const float *src1, |
| 27 | int len) |
| 28 | { |
| 29 | int i; |
| 30 | for (i = 0; i < len; i++) |
| 31 | dst[i] = src0[i] * src1[i]; |
| 32 | } |
| 33 | |
| 34 | static void vector_fmac_scalar_c(float *dst, const float *src, float mul, |
| 35 | int len) |
| 36 | { |
| 37 | int i; |
| 38 | for (i = 0; i < len; i++) |
| 39 | dst[i] += src[i] * mul; |
| 40 | } |
| 41 | |
| 42 | static void vector_fmul_scalar_c(float *dst, const float *src, float mul, |
| 43 | int len) |
| 44 | { |
| 45 | int i; |
| 46 | for (i = 0; i < len; i++) |
| 47 | dst[i] = src[i] * mul; |
| 48 | } |
| 49 | |
| 50 | static void vector_dmul_scalar_c(double *dst, const double *src, double mul, |
| 51 | int len) |
| 52 | { |
| 53 | int i; |
| 54 | for (i = 0; i < len; i++) |
| 55 | dst[i] = src[i] * mul; |
| 56 | } |
| 57 | |
| 58 | static void vector_fmul_window_c(float *dst, const float *src0, |
| 59 | const float *src1, const float *win, int len) |
| 60 | { |
| 61 | int i, j; |
| 62 | |
| 63 | dst += len; |
| 64 | win += len; |
| 65 | src0 += len; |
| 66 | |
| 67 | for (i = -len, j = len - 1; i < 0; i++, j--) { |
| 68 | float s0 = src0[i]; |
| 69 | float s1 = src1[j]; |
| 70 | float wi = win[i]; |
| 71 | float wj = win[j]; |
| 72 | dst[i] = s0 * wj - s1 * wi; |
| 73 | dst[j] = s0 * wi + s1 * wj; |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | static void vector_fmul_add_c(float *dst, const float *src0, const float *src1, |
| 78 | const float *src2, int len){ |
| 79 | int i; |
| 80 | |
| 81 | for (i = 0; i < len; i++) |
| 82 | dst[i] = src0[i] * src1[i] + src2[i]; |
| 83 | } |
| 84 | |
| 85 | static void vector_fmul_reverse_c(float *dst, const float *src0, |
| 86 | const float *src1, int len) |
| 87 | { |
| 88 | int i; |
| 89 | |
| 90 | src1 += len-1; |
| 91 | for (i = 0; i < len; i++) |
| 92 | dst[i] = src0[i] * src1[-i]; |
| 93 | } |
| 94 | |
| 95 | static void butterflies_float_c(float *av_restrict v1, float *av_restrict v2, |
| 96 | int len) |
| 97 | { |
| 98 | int i; |
| 99 | |
| 100 | for (i = 0; i < len; i++) { |
| 101 | float t = v1[i] - v2[i]; |
| 102 | v1[i] += v2[i]; |
| 103 | v2[i] = t; |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len) |
| 108 | { |
| 109 | float p = 0.0; |
| 110 | int i; |
| 111 | |
| 112 | for (i = 0; i < len; i++) |
| 113 | p += v1[i] * v2[i]; |
| 114 | |
| 115 | return p; |
| 116 | } |
| 117 | |
| 118 | av_cold void avpriv_float_dsp_init(AVFloatDSPContext *fdsp, int bit_exact) |
| 119 | { |
| 120 | fdsp->vector_fmul = vector_fmul_c; |
| 121 | fdsp->vector_fmac_scalar = vector_fmac_scalar_c; |
| 122 | fdsp->vector_fmul_scalar = vector_fmul_scalar_c; |
| 123 | fdsp->vector_dmul_scalar = vector_dmul_scalar_c; |
| 124 | fdsp->vector_fmul_window = vector_fmul_window_c; |
| 125 | fdsp->vector_fmul_add = vector_fmul_add_c; |
| 126 | fdsp->vector_fmul_reverse = vector_fmul_reverse_c; |
| 127 | fdsp->butterflies_float = butterflies_float_c; |
| 128 | fdsp->scalarproduct_float = avpriv_scalarproduct_float_c; |
| 129 | |
| 130 | if (ARCH_AARCH64) |
| 131 | ff_float_dsp_init_aarch64(fdsp); |
| 132 | if (ARCH_ARM) |
| 133 | ff_float_dsp_init_arm(fdsp); |
| 134 | if (ARCH_PPC) |
| 135 | ff_float_dsp_init_ppc(fdsp, bit_exact); |
| 136 | if (ARCH_X86) |
| 137 | ff_float_dsp_init_x86(fdsp); |
| 138 | if (ARCH_MIPS) |
| 139 | ff_float_dsp_init_mips(fdsp); |
| 140 | } |
| 141 | |
| 142 | #ifdef TEST |
| 143 | |
| 144 | #include <float.h> |
| 145 | #include <math.h> |
| 146 | #include <stdint.h> |
| 147 | #include <stdlib.h> |
| 148 | #include <string.h> |
| 149 | |
| 150 | #include "common.h" |
| 151 | #include "cpu.h" |
| 152 | #include "internal.h" |
| 153 | #include "lfg.h" |
| 154 | #include "log.h" |
| 155 | #include "mem.h" |
| 156 | #include "random_seed.h" |
| 157 | |
| 158 | #define LEN 240 |
| 159 | |
| 160 | static void fill_float_array(AVLFG *lfg, float *a, int len) |
| 161 | { |
| 162 | int i; |
| 163 | double bmg[2], stddev = 10.0, mean = 0.0; |
| 164 | |
| 165 | for (i = 0; i < len; i += 2) { |
| 166 | av_bmg_get(lfg, bmg); |
| 167 | a[i] = bmg[0] * stddev + mean; |
| 168 | a[i + 1] = bmg[1] * stddev + mean; |
| 169 | } |
| 170 | } |
| 171 | static int compare_floats(const float *a, const float *b, int len, |
| 172 | float max_diff) |
| 173 | { |
| 174 | int i; |
| 175 | for (i = 0; i < len; i++) { |
| 176 | if (fabsf(a[i] - b[i]) > max_diff) { |
| 177 | av_log(NULL, AV_LOG_ERROR, "%d: %- .12f - %- .12f = % .12g\n", |
| 178 | i, a[i], b[i], a[i] - b[i]); |
| 179 | return -1; |
| 180 | } |
| 181 | } |
| 182 | return 0; |
| 183 | } |
| 184 | |
| 185 | static void fill_double_array(AVLFG *lfg, double *a, int len) |
| 186 | { |
| 187 | int i; |
| 188 | double bmg[2], stddev = 10.0, mean = 0.0; |
| 189 | |
| 190 | for (i = 0; i < len; i += 2) { |
| 191 | av_bmg_get(lfg, bmg); |
| 192 | a[i] = bmg[0] * stddev + mean; |
| 193 | a[i + 1] = bmg[1] * stddev + mean; |
| 194 | } |
| 195 | } |
| 196 | |
| 197 | static int compare_doubles(const double *a, const double *b, int len, |
| 198 | double max_diff) |
| 199 | { |
| 200 | int i; |
| 201 | |
| 202 | for (i = 0; i < len; i++) { |
| 203 | if (fabs(a[i] - b[i]) > max_diff) { |
| 204 | av_log(NULL, AV_LOG_ERROR, "%d: %- .12f - %- .12f = % .12g\n", |
| 205 | i, a[i], b[i], a[i] - b[i]); |
| 206 | return -1; |
| 207 | } |
| 208 | } |
| 209 | return 0; |
| 210 | } |
| 211 | |
| 212 | static int test_vector_fmul(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 213 | const float *v1, const float *v2) |
| 214 | { |
| 215 | LOCAL_ALIGNED(32, float, cdst, [LEN]); |
| 216 | LOCAL_ALIGNED(32, float, odst, [LEN]); |
| 217 | int ret; |
| 218 | |
| 219 | cdsp->vector_fmul(cdst, v1, v2, LEN); |
| 220 | fdsp->vector_fmul(odst, v1, v2, LEN); |
| 221 | |
| 222 | if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON)) |
| 223 | av_log(NULL, AV_LOG_ERROR, "vector_fmul failed\n"); |
| 224 | |
| 225 | return ret; |
| 226 | } |
| 227 | |
| 228 | #define ARBITRARY_FMAC_SCALAR_CONST 0.005 |
| 229 | static int test_vector_fmac_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 230 | const float *v1, const float *src0, float scale) |
| 231 | { |
| 232 | LOCAL_ALIGNED(32, float, cdst, [LEN]); |
| 233 | LOCAL_ALIGNED(32, float, odst, [LEN]); |
| 234 | int ret; |
| 235 | |
| 236 | memcpy(cdst, v1, LEN * sizeof(*v1)); |
| 237 | memcpy(odst, v1, LEN * sizeof(*v1)); |
| 238 | |
| 239 | cdsp->vector_fmac_scalar(cdst, src0, scale, LEN); |
| 240 | fdsp->vector_fmac_scalar(odst, src0, scale, LEN); |
| 241 | |
| 242 | if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMAC_SCALAR_CONST)) |
| 243 | av_log(NULL, AV_LOG_ERROR, "vector_fmac_scalar failed\n"); |
| 244 | |
| 245 | return ret; |
| 246 | } |
| 247 | |
| 248 | static int test_vector_fmul_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 249 | const float *v1, float scale) |
| 250 | { |
| 251 | LOCAL_ALIGNED(32, float, cdst, [LEN]); |
| 252 | LOCAL_ALIGNED(32, float, odst, [LEN]); |
| 253 | int ret; |
| 254 | |
| 255 | cdsp->vector_fmul_scalar(cdst, v1, scale, LEN); |
| 256 | fdsp->vector_fmul_scalar(odst, v1, scale, LEN); |
| 257 | |
| 258 | if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON)) |
| 259 | av_log(NULL, AV_LOG_ERROR, "vector_fmul_scalar failed\n"); |
| 260 | |
| 261 | return ret; |
| 262 | } |
| 263 | |
| 264 | static int test_vector_dmul_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 265 | const double *v1, double scale) |
| 266 | { |
| 267 | LOCAL_ALIGNED(32, double, cdst, [LEN]); |
| 268 | LOCAL_ALIGNED(32, double, odst, [LEN]); |
| 269 | int ret; |
| 270 | |
| 271 | cdsp->vector_dmul_scalar(cdst, v1, scale, LEN); |
| 272 | fdsp->vector_dmul_scalar(odst, v1, scale, LEN); |
| 273 | |
| 274 | if (ret = compare_doubles(cdst, odst, LEN, DBL_EPSILON)) |
| 275 | av_log(NULL, AV_LOG_ERROR, "vector_dmul_scalar failed\n"); |
| 276 | |
| 277 | return ret; |
| 278 | } |
| 279 | |
| 280 | #define ARBITRARY_FMUL_WINDOW_CONST 0.008 |
| 281 | static int test_vector_fmul_window(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 282 | const float *v1, const float *v2, const float *v3) |
| 283 | { |
| 284 | LOCAL_ALIGNED(32, float, cdst, [LEN]); |
| 285 | LOCAL_ALIGNED(32, float, odst, [LEN]); |
| 286 | int ret; |
| 287 | |
| 288 | cdsp->vector_fmul_window(cdst, v1, v2, v3, LEN / 2); |
| 289 | fdsp->vector_fmul_window(odst, v1, v2, v3, LEN / 2); |
| 290 | |
| 291 | if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMUL_WINDOW_CONST)) |
| 292 | av_log(NULL, AV_LOG_ERROR, "vector_fmul_window failed\n"); |
| 293 | |
| 294 | return ret; |
| 295 | } |
| 296 | |
| 297 | #define ARBITRARY_FMUL_ADD_CONST 0.005 |
| 298 | static int test_vector_fmul_add(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 299 | const float *v1, const float *v2, const float *v3) |
| 300 | { |
| 301 | LOCAL_ALIGNED(32, float, cdst, [LEN]); |
| 302 | LOCAL_ALIGNED(32, float, odst, [LEN]); |
| 303 | int ret; |
| 304 | |
| 305 | cdsp->vector_fmul_add(cdst, v1, v2, v3, LEN); |
| 306 | fdsp->vector_fmul_add(odst, v1, v2, v3, LEN); |
| 307 | |
| 308 | if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMUL_ADD_CONST)) |
| 309 | av_log(NULL, AV_LOG_ERROR, "vector_fmul_add failed\n"); |
| 310 | |
| 311 | return ret; |
| 312 | } |
| 313 | |
| 314 | static int test_vector_fmul_reverse(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 315 | const float *v1, const float *v2) |
| 316 | { |
| 317 | LOCAL_ALIGNED(32, float, cdst, [LEN]); |
| 318 | LOCAL_ALIGNED(32, float, odst, [LEN]); |
| 319 | int ret; |
| 320 | |
| 321 | cdsp->vector_fmul_reverse(cdst, v1, v2, LEN); |
| 322 | fdsp->vector_fmul_reverse(odst, v1, v2, LEN); |
| 323 | |
| 324 | if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON)) |
| 325 | av_log(NULL, AV_LOG_ERROR, "vector_fmul_reverse failed\n"); |
| 326 | |
| 327 | return ret; |
| 328 | } |
| 329 | |
| 330 | static int test_butterflies_float(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 331 | const float *v1, const float *v2) |
| 332 | { |
| 333 | LOCAL_ALIGNED(32, float, cv1, [LEN]); |
| 334 | LOCAL_ALIGNED(32, float, cv2, [LEN]); |
| 335 | LOCAL_ALIGNED(32, float, ov1, [LEN]); |
| 336 | LOCAL_ALIGNED(32, float, ov2, [LEN]); |
| 337 | int ret; |
| 338 | |
| 339 | memcpy(cv1, v1, LEN * sizeof(*v1)); |
| 340 | memcpy(cv2, v2, LEN * sizeof(*v2)); |
| 341 | memcpy(ov1, v1, LEN * sizeof(*v1)); |
| 342 | memcpy(ov2, v2, LEN * sizeof(*v2)); |
| 343 | |
| 344 | cdsp->butterflies_float(cv1, cv2, LEN); |
| 345 | fdsp->butterflies_float(ov1, ov2, LEN); |
| 346 | |
| 347 | if ((ret = compare_floats(cv1, ov1, LEN, FLT_EPSILON)) || |
| 348 | (ret = compare_floats(cv2, ov2, LEN, FLT_EPSILON))) |
| 349 | av_log(NULL, AV_LOG_ERROR, "butterflies_float failed\n"); |
| 350 | |
| 351 | return ret; |
| 352 | } |
| 353 | |
| 354 | #define ARBITRARY_SCALARPRODUCT_CONST 0.2 |
| 355 | static int test_scalarproduct_float(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp, |
| 356 | const float *v1, const float *v2) |
| 357 | { |
| 358 | float cprod, oprod; |
| 359 | int ret; |
| 360 | |
| 361 | cprod = cdsp->scalarproduct_float(v1, v2, LEN); |
| 362 | oprod = fdsp->scalarproduct_float(v1, v2, LEN); |
| 363 | |
| 364 | if (ret = compare_floats(&cprod, &oprod, 1, ARBITRARY_SCALARPRODUCT_CONST)) |
| 365 | av_log(NULL, AV_LOG_ERROR, "scalarproduct_float failed\n"); |
| 366 | |
| 367 | return ret; |
| 368 | } |
| 369 | |
| 370 | int main(int argc, char **argv) |
| 371 | { |
| 372 | int ret = 0; |
| 373 | uint32_t seed; |
| 374 | AVFloatDSPContext fdsp, cdsp; |
| 375 | AVLFG lfg; |
| 376 | |
| 377 | LOCAL_ALIGNED(32, float, src0, [LEN]); |
| 378 | LOCAL_ALIGNED(32, float, src1, [LEN]); |
| 379 | LOCAL_ALIGNED(32, float, src2, [LEN]); |
| 380 | LOCAL_ALIGNED(32, double, dbl_src0, [LEN]); |
| 381 | LOCAL_ALIGNED(32, double, dbl_src1, [LEN]); |
| 382 | |
| 383 | if (argc > 2 && !strcmp(argv[1], "-s")) |
| 384 | seed = strtoul(argv[2], NULL, 10); |
| 385 | else |
| 386 | seed = av_get_random_seed(); |
| 387 | |
| 388 | av_log(NULL, AV_LOG_INFO, "float_dsp-test: random seed %u\n", seed); |
| 389 | |
| 390 | av_lfg_init(&lfg, seed); |
| 391 | |
| 392 | fill_float_array(&lfg, src0, LEN); |
| 393 | fill_float_array(&lfg, src1, LEN); |
| 394 | fill_float_array(&lfg, src2, LEN); |
| 395 | |
| 396 | fill_double_array(&lfg, dbl_src0, LEN); |
| 397 | fill_double_array(&lfg, dbl_src1, LEN); |
| 398 | |
| 399 | avpriv_float_dsp_init(&fdsp, 1); |
| 400 | av_set_cpu_flags_mask(0); |
| 401 | avpriv_float_dsp_init(&cdsp, 1); |
| 402 | |
| 403 | if (test_vector_fmul(&fdsp, &cdsp, src0, src1)) |
| 404 | ret -= 1 << 0; |
| 405 | if (test_vector_fmac_scalar(&fdsp, &cdsp, src2, src0, src1[0])) |
| 406 | ret -= 1 << 1; |
| 407 | if (test_vector_fmul_scalar(&fdsp, &cdsp, src0, src1[0])) |
| 408 | ret -= 1 << 2; |
| 409 | if (test_vector_fmul_window(&fdsp, &cdsp, src0, src1, src2)) |
| 410 | ret -= 1 << 3; |
| 411 | if (test_vector_fmul_add(&fdsp, &cdsp, src0, src1, src2)) |
| 412 | ret -= 1 << 4; |
| 413 | if (test_vector_fmul_reverse(&fdsp, &cdsp, src0, src1)) |
| 414 | ret -= 1 << 5; |
| 415 | if (test_butterflies_float(&fdsp, &cdsp, src0, src1)) |
| 416 | ret -= 1 << 6; |
| 417 | if (test_scalarproduct_float(&fdsp, &cdsp, src0, src1)) |
| 418 | ret -= 1 << 7; |
| 419 | if (test_vector_dmul_scalar(&fdsp, &cdsp, dbl_src0, dbl_src1[0])) |
| 420 | ret -= 1 << 8; |
| 421 | |
| 422 | return ret; |
| 423 | } |
| 424 | |
| 425 | #endif /* TEST */ |