yading@10: /* yading@10: * (c) 2001 Fabrice Bellard yading@10: * 2007 Marc Hoffman yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: /** yading@10: * @file yading@10: * DCT test (c) 2001 Fabrice Bellard yading@10: * Started from sample code by Juan J. Sierralta P. yading@10: */ yading@10: yading@10: #include "config.h" yading@10: #include yading@10: #include yading@10: #include yading@10: #if HAVE_UNISTD_H yading@10: #include yading@10: #endif yading@10: #include yading@10: yading@10: #include "libavutil/cpu.h" yading@10: #include "libavutil/common.h" yading@10: #include "libavutil/lfg.h" yading@10: #include "libavutil/time.h" yading@10: yading@10: #include "dct.h" yading@10: #include "simple_idct.h" yading@10: #include "aandcttab.h" yading@10: #include "faandct.h" yading@10: #include "faanidct.h" yading@10: #include "x86/idct_xvid.h" yading@10: #include "dctref.h" yading@10: yading@10: #undef printf yading@10: yading@10: // BFIN yading@10: void ff_bfin_idct(int16_t *block); yading@10: void ff_bfin_fdct(int16_t *block); yading@10: yading@10: // ALTIVEC yading@10: void ff_fdct_altivec(int16_t *block); yading@10: yading@10: // ARM yading@10: void ff_j_rev_dct_arm(int16_t *data); yading@10: void ff_simple_idct_arm(int16_t *data); yading@10: void ff_simple_idct_armv5te(int16_t *data); yading@10: void ff_simple_idct_armv6(int16_t *data); yading@10: void ff_simple_idct_neon(int16_t *data); yading@10: yading@10: void ff_simple_idct_axp(int16_t *data); yading@10: yading@10: struct algo { yading@10: const char *name; yading@10: void (*func)(int16_t *block); yading@10: enum formattag { NO_PERM, MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM, yading@10: SSE2_PERM, PARTTRANS_PERM, TRANSPOSE_PERM } format; yading@10: int mm_support; yading@10: int nonspec; yading@10: }; yading@10: yading@10: static int cpu_flags; yading@10: yading@10: static const struct algo fdct_tab[] = { yading@10: { "REF-DBL", ff_ref_fdct, NO_PERM }, yading@10: { "FAAN", ff_faandct, NO_PERM }, yading@10: { "IJG-AAN-INT", ff_fdct_ifast, SCALE_PERM }, yading@10: { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, NO_PERM }, yading@10: yading@10: #if HAVE_MMX_INLINE yading@10: { "MMX", ff_fdct_mmx, NO_PERM, AV_CPU_FLAG_MMX }, yading@10: { "MMXEXT", ff_fdct_mmxext, NO_PERM, AV_CPU_FLAG_MMXEXT }, yading@10: { "SSE2", ff_fdct_sse2, NO_PERM, AV_CPU_FLAG_SSE2 }, yading@10: #endif yading@10: yading@10: #if HAVE_ALTIVEC yading@10: { "altivecfdct", ff_fdct_altivec, NO_PERM, AV_CPU_FLAG_ALTIVEC }, yading@10: #endif yading@10: yading@10: #if ARCH_BFIN yading@10: { "BFINfdct", ff_bfin_fdct, NO_PERM }, yading@10: #endif yading@10: yading@10: { 0 } yading@10: }; yading@10: yading@10: #if ARCH_X86_64 && HAVE_MMX && HAVE_YASM yading@10: void ff_prores_idct_put_10_sse2(uint16_t *dst, int linesize, yading@10: int16_t *block, int16_t *qmat); yading@10: yading@10: static void ff_prores_idct_put_10_sse2_wrap(int16_t *dst){ yading@10: DECLARE_ALIGNED(16, static int16_t, qmat)[64]; yading@10: DECLARE_ALIGNED(16, static int16_t, tmp)[64]; yading@10: int i; yading@10: yading@10: for(i=0; i<64; i++){ yading@10: qmat[i]=4; yading@10: tmp[i]= dst[i]; yading@10: } yading@10: ff_prores_idct_put_10_sse2(dst, 16, tmp, qmat); yading@10: } yading@10: #endif yading@10: yading@10: static const struct algo idct_tab[] = { yading@10: { "FAANI", ff_faanidct, NO_PERM }, yading@10: { "REF-DBL", ff_ref_idct, NO_PERM }, yading@10: { "INT", ff_j_rev_dct, MMX_PERM }, yading@10: { "SIMPLE-C", ff_simple_idct_8, NO_PERM }, yading@10: yading@10: #if HAVE_MMX_INLINE yading@10: { "SIMPLE-MMX", ff_simple_idct_mmx, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX }, yading@10: { "XVID-MMX", ff_idct_xvid_mmx, NO_PERM, AV_CPU_FLAG_MMX, 1 }, yading@10: { "XVID-MMXEXT", ff_idct_xvid_mmxext, NO_PERM, AV_CPU_FLAG_MMXEXT, 1 }, yading@10: { "XVID-SSE2", ff_idct_xvid_sse2, SSE2_PERM, AV_CPU_FLAG_SSE2, 1 }, yading@10: #if ARCH_X86_64 && HAVE_YASM yading@10: { "PR-SSE2", ff_prores_idct_put_10_sse2_wrap, TRANSPOSE_PERM, AV_CPU_FLAG_SSE2, 1 }, yading@10: #endif yading@10: #endif yading@10: yading@10: #if ARCH_BFIN yading@10: { "BFINidct", ff_bfin_idct, NO_PERM }, yading@10: #endif yading@10: yading@10: #if ARCH_ARM yading@10: { "SIMPLE-ARM", ff_simple_idct_arm, NO_PERM }, yading@10: { "INT-ARM", ff_j_rev_dct_arm, MMX_PERM }, yading@10: #endif yading@10: #if HAVE_ARMV5TE yading@10: { "SIMPLE-ARMV5TE", ff_simple_idct_armv5te,NO_PERM, AV_CPU_FLAG_ARMV5TE }, yading@10: #endif yading@10: #if HAVE_ARMV6 yading@10: { "SIMPLE-ARMV6", ff_simple_idct_armv6, MMX_PERM, AV_CPU_FLAG_ARMV6 }, yading@10: #endif yading@10: #if HAVE_NEON yading@10: { "SIMPLE-NEON", ff_simple_idct_neon, PARTTRANS_PERM, AV_CPU_FLAG_NEON }, yading@10: #endif yading@10: yading@10: #if ARCH_ALPHA yading@10: { "SIMPLE-ALPHA", ff_simple_idct_axp, NO_PERM }, yading@10: #endif yading@10: yading@10: { 0 } yading@10: }; yading@10: yading@10: #define AANSCALE_BITS 12 yading@10: yading@10: #define NB_ITS 20000 yading@10: #define NB_ITS_SPEED 50000 yading@10: yading@10: static short idct_mmx_perm[64]; yading@10: yading@10: static short idct_simple_mmx_perm[64] = { yading@10: 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D, yading@10: 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D, yading@10: 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D, yading@10: 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F, yading@10: 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F, yading@10: 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D, yading@10: 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F, yading@10: 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F, yading@10: }; yading@10: yading@10: static const uint8_t idct_sse2_row_perm[8] = { 0, 4, 1, 5, 2, 6, 3, 7 }; yading@10: yading@10: static void idct_mmx_init(void) yading@10: { yading@10: int i; yading@10: yading@10: /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */ yading@10: for (i = 0; i < 64; i++) { yading@10: idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2); yading@10: } yading@10: } yading@10: yading@10: DECLARE_ALIGNED(16, static int16_t, block)[64]; yading@10: DECLARE_ALIGNED(8, static int16_t, block1)[64]; yading@10: yading@10: static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng, int vals) yading@10: { yading@10: int i, j; yading@10: yading@10: memset(block, 0, 64 * sizeof(*block)); yading@10: yading@10: switch (test) { yading@10: case 0: yading@10: for (i = 0; i < 64; i++) yading@10: block[i] = (av_lfg_get(prng) % (2*vals)) -vals; yading@10: if (is_idct) { yading@10: ff_ref_fdct(block); yading@10: for (i = 0; i < 64; i++) yading@10: block[i] >>= 3; yading@10: } yading@10: break; yading@10: case 1: yading@10: j = av_lfg_get(prng) % 10 + 1; yading@10: for (i = 0; i < j; i++) { yading@10: int idx = av_lfg_get(prng) % 64; yading@10: block[idx] = av_lfg_get(prng) % (2*vals) -vals; yading@10: } yading@10: break; yading@10: case 2: yading@10: block[ 0] = av_lfg_get(prng) % (16*vals) - (8*vals); yading@10: block[63] = (block[0] & 1) ^ 1; yading@10: break; yading@10: } yading@10: } yading@10: yading@10: static void permute(int16_t dst[64], const int16_t src[64], int perm) yading@10: { yading@10: int i; yading@10: yading@10: if (perm == MMX_PERM) { yading@10: for (i = 0; i < 64; i++) yading@10: dst[idct_mmx_perm[i]] = src[i]; yading@10: } else if (perm == MMX_SIMPLE_PERM) { yading@10: for (i = 0; i < 64; i++) yading@10: dst[idct_simple_mmx_perm[i]] = src[i]; yading@10: } else if (perm == SSE2_PERM) { yading@10: for (i = 0; i < 64; i++) yading@10: dst[(i & 0x38) | idct_sse2_row_perm[i & 7]] = src[i]; yading@10: } else if (perm == PARTTRANS_PERM) { yading@10: for (i = 0; i < 64; i++) yading@10: dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i]; yading@10: } else if (perm == TRANSPOSE_PERM) { yading@10: for (i = 0; i < 64; i++) yading@10: dst[(i>>3) | ((i<<3)&0x38)] = src[i]; yading@10: } else { yading@10: for (i = 0; i < 64; i++) yading@10: dst[i] = src[i]; yading@10: } yading@10: } yading@10: yading@10: static int dct_error(const struct algo *dct, int test, int is_idct, int speed, const int bits) yading@10: { yading@10: void (*ref)(int16_t *block) = is_idct ? ff_ref_idct : ff_ref_fdct; yading@10: int it, i, scale; yading@10: int err_inf, v; yading@10: int64_t err2, ti, ti1, it1, err_sum = 0; yading@10: int64_t sysErr[64], sysErrMax = 0; yading@10: int maxout = 0; yading@10: int blockSumErrMax = 0, blockSumErr; yading@10: AVLFG prng; yading@10: const int vals=1<format); yading@10: yading@10: dct->func(block); yading@10: emms_c(); yading@10: yading@10: if (dct->format == SCALE_PERM) { yading@10: for (i = 0; i < 64; i++) { yading@10: scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i]; yading@10: block[i] = (block[i] * scale) >> AANSCALE_BITS; yading@10: } yading@10: } yading@10: yading@10: ref(block1); yading@10: yading@10: blockSumErr = 0; yading@10: for (i = 0; i < 64; i++) { yading@10: int err = block[i] - block1[i]; yading@10: err_sum += err; yading@10: v = abs(err); yading@10: if (v > err_inf) yading@10: err_inf = v; yading@10: err2 += v * v; yading@10: sysErr[i] += block[i] - block1[i]; yading@10: blockSumErr += v; yading@10: if (abs(block[i]) > maxout) yading@10: maxout = abs(block[i]); yading@10: } yading@10: if (blockSumErrMax < blockSumErr) yading@10: blockSumErrMax = blockSumErr; yading@10: } yading@10: for (i = 0; i < 64; i++) yading@10: sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i])); yading@10: yading@10: for (i = 0; i < 64; i++) { yading@10: if (i % 8 == 0) yading@10: printf("\n"); yading@10: printf("%7d ", (int) sysErr[i]); yading@10: } yading@10: printf("\n"); yading@10: yading@10: omse = (double) err2 / NB_ITS / 64; yading@10: ome = (double) err_sum / NB_ITS / 64; yading@10: yading@10: spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015); yading@10: yading@10: printf("%s %s: max_err=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n", yading@10: is_idct ? "IDCT" : "DCT", dct->name, err_inf, yading@10: omse, ome, (double) sysErrMax / NB_ITS, yading@10: maxout, blockSumErrMax); yading@10: yading@10: if (spec_err && !dct->nonspec) yading@10: return 1; yading@10: yading@10: if (!speed) yading@10: return 0; yading@10: yading@10: /* speed test */ yading@10: yading@10: init_block(block, test, is_idct, &prng, vals); yading@10: permute(block1, block, dct->format); yading@10: yading@10: ti = av_gettime(); yading@10: it1 = 0; yading@10: do { yading@10: for (it = 0; it < NB_ITS_SPEED; it++) { yading@10: memcpy(block, block1, sizeof(block)); yading@10: dct->func(block); yading@10: } yading@10: emms_c(); yading@10: it1 += NB_ITS_SPEED; yading@10: ti1 = av_gettime() - ti; yading@10: } while (ti1 < 1000000); yading@10: yading@10: printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name, yading@10: (double) it1 * 1000.0 / (double) ti1); yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: DECLARE_ALIGNED(8, static uint8_t, img_dest)[64]; yading@10: DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64]; yading@10: yading@10: static void idct248_ref(uint8_t *dest, int linesize, int16_t *block) yading@10: { yading@10: static int init; yading@10: static double c8[8][8]; yading@10: static double c4[4][4]; yading@10: double block1[64], block2[64], block3[64]; yading@10: double s, sum, v; yading@10: int i, j, k; yading@10: yading@10: if (!init) { yading@10: init = 1; yading@10: yading@10: for (i = 0; i < 8; i++) { yading@10: sum = 0; yading@10: for (j = 0; j < 8; j++) { yading@10: s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0); yading@10: c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0); yading@10: sum += c8[i][j] * c8[i][j]; yading@10: } yading@10: } yading@10: yading@10: for (i = 0; i < 4; i++) { yading@10: sum = 0; yading@10: for (j = 0; j < 4; j++) { yading@10: s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0); yading@10: c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0); yading@10: sum += c4[i][j] * c4[i][j]; yading@10: } yading@10: } yading@10: } yading@10: yading@10: /* butterfly */ yading@10: s = 0.5 * sqrt(2.0); yading@10: for (i = 0; i < 4; i++) { yading@10: for (j = 0; j < 8; j++) { yading@10: block1[8 * (2 * i) + j] = yading@10: (block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s; yading@10: block1[8 * (2 * i + 1) + j] = yading@10: (block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s; yading@10: } yading@10: } yading@10: yading@10: /* idct8 on lines */ yading@10: for (i = 0; i < 8; i++) { yading@10: for (j = 0; j < 8; j++) { yading@10: sum = 0; yading@10: for (k = 0; k < 8; k++) yading@10: sum += c8[k][j] * block1[8 * i + k]; yading@10: block2[8 * i + j] = sum; yading@10: } yading@10: } yading@10: yading@10: /* idct4 */ yading@10: for (i = 0; i < 8; i++) { yading@10: for (j = 0; j < 4; j++) { yading@10: /* top */ yading@10: sum = 0; yading@10: for (k = 0; k < 4; k++) yading@10: sum += c4[k][j] * block2[8 * (2 * k) + i]; yading@10: block3[8 * (2 * j) + i] = sum; yading@10: yading@10: /* bottom */ yading@10: sum = 0; yading@10: for (k = 0; k < 4; k++) yading@10: sum += c4[k][j] * block2[8 * (2 * k + 1) + i]; yading@10: block3[8 * (2 * j + 1) + i] = sum; yading@10: } yading@10: } yading@10: yading@10: /* clamp and store the result */ yading@10: for (i = 0; i < 8; i++) { yading@10: for (j = 0; j < 8; j++) { yading@10: v = block3[8 * i + j]; yading@10: if (v < 0) v = 0; yading@10: else if (v > 255) v = 255; yading@10: dest[i * linesize + j] = (int) rint(v); yading@10: } yading@10: } yading@10: } yading@10: yading@10: static void idct248_error(const char *name, yading@10: void (*idct248_put)(uint8_t *dest, int line_size, yading@10: int16_t *block), yading@10: int speed) yading@10: { yading@10: int it, i, it1, ti, ti1, err_max, v; yading@10: AVLFG prng; yading@10: yading@10: av_lfg_init(&prng, 1); yading@10: yading@10: /* just one test to see if code is correct (precision is less yading@10: important here) */ yading@10: err_max = 0; yading@10: for (it = 0; it < NB_ITS; it++) { yading@10: /* XXX: use forward transform to generate values */ yading@10: for (i = 0; i < 64; i++) yading@10: block1[i] = av_lfg_get(&prng) % 256 - 128; yading@10: block1[0] += 1024; yading@10: yading@10: for (i = 0; i < 64; i++) yading@10: block[i] = block1[i]; yading@10: idct248_ref(img_dest1, 8, block); yading@10: yading@10: for (i = 0; i < 64; i++) yading@10: block[i] = block1[i]; yading@10: idct248_put(img_dest, 8, block); yading@10: yading@10: for (i = 0; i < 64; i++) { yading@10: v = abs((int) img_dest[i] - (int) img_dest1[i]); yading@10: if (v == 255) yading@10: printf("%d %d\n", img_dest[i], img_dest1[i]); yading@10: if (v > err_max) yading@10: err_max = v; yading@10: } yading@10: #if 0 yading@10: printf("ref=\n"); yading@10: for(i=0;i<8;i++) { yading@10: int j; yading@10: for(j=0;j<8;j++) { yading@10: printf(" %3d", img_dest1[i*8+j]); yading@10: } yading@10: printf("\n"); yading@10: } yading@10: yading@10: printf("out=\n"); yading@10: for(i=0;i<8;i++) { yading@10: int j; yading@10: for(j=0;j<8;j++) { yading@10: printf(" %3d", img_dest[i*8+j]); yading@10: } yading@10: printf("\n"); yading@10: } yading@10: #endif yading@10: } yading@10: printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max); yading@10: yading@10: if (!speed) yading@10: return; yading@10: yading@10: ti = av_gettime(); yading@10: it1 = 0; yading@10: do { yading@10: for (it = 0; it < NB_ITS_SPEED; it++) { yading@10: for (i = 0; i < 64; i++) yading@10: block[i] = block1[i]; yading@10: idct248_put(img_dest, 8, block); yading@10: } yading@10: emms_c(); yading@10: it1 += NB_ITS_SPEED; yading@10: ti1 = av_gettime() - ti; yading@10: } while (ti1 < 1000000); yading@10: yading@10: printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name, yading@10: (double) it1 * 1000.0 / (double) ti1); yading@10: } yading@10: yading@10: static void help(void) yading@10: { yading@10: printf("dct-test [-i] [] []\n" yading@10: "test-number 0 -> test with random matrixes\n" yading@10: " 1 -> test with random sparse matrixes\n" yading@10: " 2 -> do 3. test from mpeg4 std\n" yading@10: "bits Number of time domain bits to use, 8 is default\n" yading@10: "-i test IDCT implementations\n" yading@10: "-4 test IDCT248 implementations\n" yading@10: "-t speed test\n"); yading@10: } yading@10: yading@10: #if !HAVE_GETOPT yading@10: #include "compat/getopt.c" yading@10: #endif yading@10: yading@10: int main(int argc, char **argv) yading@10: { yading@10: int test_idct = 0, test_248_dct = 0; yading@10: int c, i; yading@10: int test = 1; yading@10: int speed = 0; yading@10: int err = 0; yading@10: int bits=8; yading@10: yading@10: cpu_flags = av_get_cpu_flags(); yading@10: yading@10: ff_ref_dct_init(); yading@10: idct_mmx_init(); yading@10: yading@10: for (;;) { yading@10: c = getopt(argc, argv, "ih4t"); yading@10: if (c == -1) yading@10: break; yading@10: switch (c) { yading@10: case 'i': yading@10: test_idct = 1; yading@10: break; yading@10: case '4': yading@10: test_248_dct = 1; yading@10: break; yading@10: case 't': yading@10: speed = 1; yading@10: break; yading@10: default: yading@10: case 'h': yading@10: help(); yading@10: return 0; yading@10: } yading@10: } yading@10: yading@10: if (optind < argc) yading@10: test = atoi(argv[optind]); yading@10: if(optind+1 < argc) bits= atoi(argv[optind+1]); yading@10: yading@10: printf("ffmpeg DCT/IDCT test\n"); yading@10: yading@10: if (test_248_dct) { yading@10: idct248_error("SIMPLE-C", ff_simple_idct248_put, speed); yading@10: } else { yading@10: const struct algo *algos = test_idct ? idct_tab : fdct_tab; yading@10: for (i = 0; algos[i].name; i++) yading@10: if (!(~cpu_flags & algos[i].mm_support)) { yading@10: err |= dct_error(&algos[i], test, test_idct, speed, bits); yading@10: } yading@10: } yading@10: yading@10: return err; yading@10: }