DIRAC Decoder stable version, MMX support removed.

Look for MMX_DISABLED to find the disabled functions.

Authors of this code are Marco Gerards <marco@gnu.org> and David Conrad <lessen42@gmail.com>
With changes from Jordi Ortiz <nenjordi@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/Makefile

@@ -111,6 +111,8 @@ OBJS-$(CONFIG_CSCD_DECODER)            += cscd.o
 OBJS-$(CONFIG_CYUV_DECODER)            += cyuv.o
 OBJS-$(CONFIG_DCA_DECODER)             += dca.o synth_filter.o dcadsp.o
 OBJS-$(CONFIG_DCA_ENCODER)             += dcaenc.o
+OBJS-$(CONFIG_DIRAC_DECODER)           += diracdec.o dirac.o diracdsp.o \
+                                          dirac_arith.o mpeg12data.o dwt.o
 OBJS-$(CONFIG_DFA_DECODER)             += dfa.o
 OBJS-$(CONFIG_DNXHD_DECODER)           += dnxhddec.o dnxhddata.o
 OBJS-$(CONFIG_DNXHD_ENCODER)           += dnxhdenc.o dnxhddata.o       \

libavcodec/allcodecs.c

@@ -90,6 +90,7 @@ void avcodec_register_all(void)
     REGISTER_DECODER (CLJR, cljr);
     REGISTER_DECODER (CSCD, cscd);
     REGISTER_DECODER (CYUV, cyuv);
+    REGISTER_DECODER (DIRAC, dirac);
     REGISTER_DECODER (DFA, dfa);
     REGISTER_ENCDEC  (DNXHD, dnxhd);
     REGISTER_ENCDEC  (DPX, dpx);

libavcodec/dirac_arith.c

+/*
+ * Copyright (C) 2007 Marco Gerards <marco@gnu.org>
+ * Copyright (C) 2009 David Conrad
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file libavcodec/dirac_arith.c
+ * Arithmetic decoder for Dirac
+ * @author Marco Gerards <marco@gnu.org>
+ */
+
+#include "dirac_arith.h"
+
+const uint16_t ff_dirac_prob[256] = {
+    0,    2,    5,    8,    11,   15,   20,   24,
+    29,   35,   41,   47,   53,   60,   67,   74,
+    82,   89,   97,   106,  114,  123,  132,  141,
+    150,  160,  170,  180,  190,  201,  211,  222,
+    233,  244,  256,  267,  279,  291,  303,  315,
+    327,  340,  353,  366,  379,  392,  405,  419,
+    433,  447,  461,  475,  489,  504,  518,  533,
+    548,  563,  578,  593,  609,  624,  640,  656,
+    672,  688,  705,  721,  738,  754,  771,  788,
+    805,  822,  840,  857,  875,  892,  910,  928,
+    946,  964,  983,  1001, 1020, 1038, 1057, 1076,
+    1095, 1114, 1133, 1153, 1172, 1192, 1211, 1231,
+    1251, 1271, 1291, 1311, 1332, 1352, 1373, 1393,
+    1414, 1435, 1456, 1477, 1498, 1520, 1541, 1562,
+    1584, 1606, 1628, 1649, 1671, 1694, 1716, 1738,
+    1760, 1783, 1806, 1828, 1851, 1874, 1897, 1920,
+    1935, 1942, 1949, 1955, 1961, 1968, 1974, 1980,
+    1985, 1991, 1996, 2001, 2006, 2011, 2016, 2021,
+    2025, 2029, 2033, 2037, 2040, 2044, 2047, 2050,
+    2053, 2056, 2058, 2061, 2063, 2065, 2066, 2068,
+    2069, 2070, 2071, 2072, 2072, 2072, 2072, 2072,
+    2072, 2071, 2070, 2069, 2068, 2066, 2065, 2063,
+    2060, 2058, 2055, 2052, 2049, 2045, 2042, 2038,
+    2033, 2029, 2024, 2019, 2013, 2008, 2002, 1996,
+    1989, 1982, 1975, 1968, 1960, 1952, 1943, 1934,
+    1925, 1916, 1906, 1896, 1885, 1874, 1863, 1851,
+    1839, 1827, 1814, 1800, 1786, 1772, 1757, 1742,
+    1727, 1710, 1694, 1676, 1659, 1640, 1622, 1602,
+    1582, 1561, 1540, 1518, 1495, 1471, 1447, 1422,
+    1396, 1369, 1341, 1312, 1282, 1251, 1219, 1186,
+    1151, 1114, 1077, 1037, 995,  952,  906,  857,
+    805,  750,  690,  625,  553,  471,  376,  255
+};
+
+const uint8_t ff_dirac_next_ctx[DIRAC_CTX_COUNT] = {
+    [CTX_ZPZN_F1]   = CTX_ZP_F2,
+    [CTX_ZPNN_F1]   = CTX_ZP_F2,
+    [CTX_ZP_F2]     = CTX_ZP_F3,
+    [CTX_ZP_F3]     = CTX_ZP_F4,
+    [CTX_ZP_F4]     = CTX_ZP_F5,
+    [CTX_ZP_F5]     = CTX_ZP_F6,
+    [CTX_ZP_F6]     = CTX_ZP_F6,
+    [CTX_NPZN_F1]   = CTX_NP_F2,
+    [CTX_NPNN_F1]   = CTX_NP_F2,
+    [CTX_NP_F2]     = CTX_NP_F3,
+    [CTX_NP_F3]     = CTX_NP_F4,
+    [CTX_NP_F4]     = CTX_NP_F5,
+    [CTX_NP_F5]     = CTX_NP_F6,
+    [CTX_NP_F6]     = CTX_NP_F6,
+    [CTX_DELTA_Q_F] = CTX_DELTA_Q_F,
+};
+
+int16_t ff_dirac_prob_branchless[256][2];
+
+void ff_dirac_init_arith_decoder(DiracArith *c, GetBitContext *gb, int length)
+{
+    int i;
+    align_get_bits(gb);
+
+    length = FFMIN(length, get_bits_left(gb)/8);
+
+    c->bytestream     = gb->buffer + get_bits_count(gb)/8;
+    c->bytestream_end = c->bytestream + length;
+    skip_bits_long(gb, length*8);
+
+    c->low = 0;
+    for (i = 0; i < 4; i++) {
+        c->low <<= 8;
+        if (c->bytestream < c->bytestream_end)
+            c->low |= *c->bytestream++;
+        else
+            c->low |= 0xff;
+    }
+
+    c->counter = -16;
+    c->range   = 0xffff;
+
+    for (i = 0; i < 256; i++) {
+        ff_dirac_prob_branchless[i][0] =  ff_dirac_prob[255-i];
+        ff_dirac_prob_branchless[i][1] = -ff_dirac_prob[i];
+    }
+
+    for (i = 0; i < DIRAC_CTX_COUNT; i++)
+        c->contexts[i] = 0x8000;
+}

libavcodec/dirac_arith.h

+/*
+ * Copyright (C) 2007 Marco Gerards <marco@gnu.org>
+ * Copyright (C) 2009 David Conrad
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file libavcodec/dirac_arith.h
+ * Arithmetic decoder for Dirac
+ * @author Marco Gerards <marco@gnu.org>
+ */
+
+#ifndef AVCODEC_DIRAC_ARITH_H
+#define AVCODEC_DIRAC_ARITH_H
+
+#include "bytestream.h"
+#include "get_bits.h"
+
+enum dirac_arith_contexts {
+    CTX_ZPZN_F1,
+    CTX_ZPNN_F1,
+    CTX_NPZN_F1,
+    CTX_NPNN_F1,
+    CTX_ZP_F2,
+    CTX_ZP_F3,
+    CTX_ZP_F4,
+    CTX_ZP_F5,
+    CTX_ZP_F6,
+    CTX_NP_F2,
+    CTX_NP_F3,
+    CTX_NP_F4,
+    CTX_NP_F5,
+    CTX_NP_F6,
+    CTX_COEFF_DATA,
+    CTX_SIGN_NEG,
+    CTX_SIGN_ZERO,
+    CTX_SIGN_POS,
+    CTX_ZERO_BLOCK,
+    CTX_DELTA_Q_F,
+    CTX_DELTA_Q_DATA,
+    CTX_DELTA_Q_SIGN,
+
+    DIRAC_CTX_COUNT
+};
+
+// Dirac resets the arith decoder between decoding various types of data,
+// so many contexts are never used simultaneously. Thus, we can reduce
+// the number of contexts needed by reusing them.
+#define CTX_SB_F1        CTX_ZP_F5
+#define CTX_SB_DATA      0
+#define CTX_PMODE_REF1   0
+#define CTX_PMODE_REF2   1
+#define CTX_GLOBAL_BLOCK 2
+#define CTX_MV_F1        CTX_ZP_F2
+#define CTX_MV_DATA      0
+#define CTX_DC_F1        CTX_ZP_F5
+#define CTX_DC_DATA      0
+
+typedef struct {
+    unsigned low;
+    uint16_t range;
+    int16_t  counter;
+
+    const uint8_t *bytestream;
+    const uint8_t *bytestream_end;
+
+    uint16_t contexts[DIRAC_CTX_COUNT];
+} DiracArith;
+
+extern const uint8_t ff_dirac_next_ctx[DIRAC_CTX_COUNT];
+extern const uint16_t ff_dirac_prob[256];
+extern int16_t ff_dirac_prob_branchless[256][2];
+
+static inline void renorm(DiracArith *c)
+{
+#if HAVE_FAST_CLZ
+    int shift = 14 - av_log2_16bit(c->range-1) + ((c->range-1)>>15);
+
+    c->low    <<= shift;
+    c->range  <<= shift;
+    c->counter += shift;
+#else
+    while (c->range <= 0x4000) {
+        c->low   <<= 1;
+        c->range <<= 1;
+        c->counter++;
+    }
+#endif
+}
+
+static inline void refill(DiracArith *c)
+{
+    int counter = c->counter;
+
+    if (counter >= 0) {
+        int new = bytestream_get_be16(&c->bytestream);
+
+        // the spec defines overread bits to be 1, and streams rely on this
+        if (c->bytestream > c->bytestream_end) {
+            new |= 0xff;
+            if (c->bytestream > c->bytestream_end+1)
+                new |= 0xff00;
+
+            c->bytestream = c->bytestream_end;
+        }
+
+        c->low += new << counter;
+        counter -= 16;
+    }
+    c->counter = counter;
+}
+
+static inline int dirac_get_arith_bit(DiracArith *c, int ctx)
+{
+    int prob_zero = c->contexts[ctx];
+    int range_times_prob, bit;
+    unsigned low = c->low;
+    int    range = c->range;
+
+    range_times_prob = (c->range * prob_zero) >> 16;
+
+#if HAVE_FAST_CMOV
+    low   -= range_times_prob << 16;
+    range -= range_times_prob;
+    bit = 0;
+    __asm__(
+        "cmpl   %5, %4 \n\t"
+        "setae  %b0    \n\t"
+        "cmovb  %3, %2 \n\t"
+        "cmovb  %5, %1 \n\t"
+        : "+q"(bit), "+r"(range), "+r"(low)
+        : "r"(c->low), "r"(c->low>>16),
+          "r"(range_times_prob)
+    );
+#else
+    bit = (low >> 16) >= range_times_prob;
+    if (bit) {
+        low   -= range_times_prob << 16;
+        range -= range_times_prob;
+    } else {
+        range  = range_times_prob;
+    }
+#endif
+
+    c->contexts[ctx] += ff_dirac_prob_branchless[prob_zero>>8][bit];
+    c->low   = low;
+    c->range = range;
+
+    renorm(c);
+    refill(c);
+    return bit;
+}
+
+static inline int dirac_get_arith_uint(DiracArith *c, int follow_ctx, int data_ctx)
+{
+    int ret = 1;
+    while (!dirac_get_arith_bit(c, follow_ctx)) {
+        ret <<= 1;
+        ret += dirac_get_arith_bit(c, data_ctx);
+        follow_ctx = ff_dirac_next_ctx[follow_ctx];
+    }
+    return ret-1;
+}
+
+static inline int dirac_get_arith_int(DiracArith *c, int follow_ctx, int data_ctx)
+{
+    int ret = dirac_get_arith_uint(c, follow_ctx, data_ctx);
+    if (ret && dirac_get_arith_bit(c, data_ctx+1))
+        ret = -ret;
+    return ret;
+}
+
+void ff_dirac_init_arith_decoder(DiracArith *c, GetBitContext *gb, int length);
+
+#endif /* AVCODEC_DIRAC_ARITH_H */

libavcodec/diracdsp.c

+/*
+ * Copyright (C) 2009 David Conrad
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "dsputil.h"
+#include "diracdsp.h"
+//MMX_DISABLE #include "libavcodec/x86/diracdsp_mmx.h"
+
+#define FILTER(src, stride) \
+  ((21*((src)[ 0*stride] + (src)[1*stride]) \
+    -7*((src)[-1*stride] + (src)[2*stride]) \
+    +3*((src)[-2*stride] + (src)[3*stride]) \
+    -1*((src)[-3*stride] + (src)[4*stride]) + 16) >> 5)
+
+static void dirac_hpel_filter(uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src,
+                              int stride, int width, int height)
+{
+    int x, y;
+
+    for (y = 0; y < height; y++) {
+        for (x = -3; x < width+5; x++)
+            dstv[x] = av_clip_uint8(FILTER(src+x, stride));
+
+        for (x = 0; x < width; x++)
+            dstc[x] = av_clip_uint8(FILTER(dstv+x, 1));
+
+        for (x = 0; x < width; x++)
+            dsth[x] = av_clip_uint8(FILTER(src+x, 1));
+
+        src  += stride;
+        dsth += stride;
+        dstv += stride;
+        dstc += stride;
+    }
+}
+
+#define PIXOP_BILINEAR(PFX, OP, WIDTH) \
+static void ff_ ## PFX ## _dirac_pixels ## WIDTH ## _bilinear_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    int x;\
+    const uint8_t *s0 = src[0];\
+    const uint8_t *s1 = src[1];\
+    const uint8_t *s2 = src[2];\
+    const uint8_t *s3 = src[3];\
+    const uint8_t *w  = src[4];\
+\
+    while (h--) {\
+        for (x = 0; x < WIDTH; x++) {\
+            OP(dst[x], (s0[x]*w[0] + s1[x]*w[1] + s2[x]*w[2] + s3[x]*w[3] + 8) >> 4);\
+        }\
+\
+        dst += stride;\
+        s0 += stride;\
+        s1 += stride;\
+        s2 += stride;\
+        s3 += stride;\
+    }\
+}
+
+#define OP_PUT(dst, val) (dst) = (val)
+#define OP_AVG(dst, val) (dst) = (((dst) + (val) + 1)>>1)
+
+PIXOP_BILINEAR(put, OP_PUT, 8)
+PIXOP_BILINEAR(put, OP_PUT, 16)
+PIXOP_BILINEAR(put, OP_PUT, 32)
+PIXOP_BILINEAR(avg, OP_AVG, 8)
+PIXOP_BILINEAR(avg, OP_AVG, 16)
+PIXOP_BILINEAR(avg, OP_AVG, 32)
+
+#define op_scale1(x)  block[x] = av_clip_uint8( (block[x]*weight + (1<<(log2_denom-1))) >> log2_denom)
+#define op_scale2(x)  dst[x] = av_clip_uint8( (src[x]*weights + dst[x]*weightd + (1<<(log2_denom-1))) >> log2_denom)
+
+#define DIRAC_WEIGHT(W) \
+static void weight_dirac_pixels ## W ## _c(uint8_t *block, int stride, int log2_denom, \
+                                           int weight, int h) { \
+    int x; \
+    while (h--) { \
+        for (x = 0; x < W; x++) { \
+            op_scale1(x); \
+            op_scale1(x+1); \
+        } \
+        block += stride; \
+    } \
+} \
+static void biweight_dirac_pixels ## W ## _c(uint8_t *dst, uint8_t *src, int stride, int log2_denom, \
+                                             int weightd, int weights, int h) { \
+    int x; \
+    while (h--) { \
+        for (x = 0; x < W; x++) { \
+            op_scale2(x); \
+            op_scale2(x+1); \
+        } \
+        dst += stride; \
+        src += stride; \
+    } \
+}
+
+DIRAC_WEIGHT(8)
+DIRAC_WEIGHT(16)
+DIRAC_WEIGHT(32)
+
+#define ADD_OBMC(xblen) \
+static void add_obmc ## xblen ## _c(uint16_t *dst, const uint8_t *src, int stride, \
+                                    const uint8_t *obmc_weight, int yblen) \
+{ \
+    int x; \
+    while (yblen--) { \
+        for (x = 0; x < xblen; x += 2) { \
+            dst[x  ] += src[x  ] * obmc_weight[x  ]; \
+            dst[x+1] += src[x+1] * obmc_weight[x+1]; \
+        } \
+        dst += stride; \
+        src += stride; \
+        obmc_weight += 32; \
+    } \
+}
+
+ADD_OBMC(8)
+ADD_OBMC(16)
+ADD_OBMC(32)
+
+static void put_signed_rect_clamped_c(uint8_t *dst, int dst_stride, const int16_t *src, int src_stride, int width, int height)
+{
+    int x, y;
+    for (y = 0; y < height; y++) {
+        for (x = 0; x < width; x+=4) {
+            dst[x  ] = av_clip_uint8(src[x  ] + 128);
+            dst[x+1] = av_clip_uint8(src[x+1] + 128);
+            dst[x+2] = av_clip_uint8(src[x+2] + 128);
+            dst[x+3] = av_clip_uint8(src[x+3] + 128);
+        }
+        dst += dst_stride;
+        src += src_stride;
+    }
+}
+
+static void add_rect_clamped_c(uint8_t *dst, const uint16_t *src, int stride,
+                               const int16_t *idwt, int idwt_stride,
+                               int width, int height)
+{
+    int x, y;
+
+    for (y = 0; y < height; y++) {
+        for (x = 0; x < width; x+=2) {
+            dst[x  ] = av_clip_uint8(((src[x  ]+32)>>6) + idwt[x  ]);
+            dst[x+1] = av_clip_uint8(((src[x+1]+32)>>6) + idwt[x+1]);
+        }
+        dst += stride;
+        src += stride;
+        idwt += idwt_stride;
+    }
+}
+
+#define PIXFUNC(PFX, WIDTH) \
+    c->PFX ## _dirac_pixels_tab[WIDTH>>4][0] = ff_ ## PFX ## _dirac_pixels ## WIDTH ## _c; \
+    c->PFX ## _dirac_pixels_tab[WIDTH>>4][1] = ff_ ## PFX ## _dirac_pixels ## WIDTH ## _l2_c; \
+    c->PFX ## _dirac_pixels_tab[WIDTH>>4][2] = ff_ ## PFX ## _dirac_pixels ## WIDTH ## _l4_c; \
+    c->PFX ## _dirac_pixels_tab[WIDTH>>4][3] = ff_ ## PFX ## _dirac_pixels ## WIDTH ## _bilinear_c
+
+void ff_diracdsp_init(DiracDSPContext *c)
+{
+    c->dirac_hpel_filter = dirac_hpel_filter;
+    c->add_rect_clamped = add_rect_clamped_c;
+    c->put_signed_rect_clamped = put_signed_rect_clamped_c;
+
+    c->add_dirac_obmc[0] = add_obmc8_c;
+    c->add_dirac_obmc[1] = add_obmc16_c;
+    c->add_dirac_obmc[2] = add_obmc32_c;
+
+    c->weight_dirac_pixels_tab[0] = weight_dirac_pixels8_c;
+    c->weight_dirac_pixels_tab[1] = weight_dirac_pixels16_c;
+    c->weight_dirac_pixels_tab[2] = weight_dirac_pixels32_c;
+    c->biweight_dirac_pixels_tab[0] = biweight_dirac_pixels8_c;
+    c->biweight_dirac_pixels_tab[1] = biweight_dirac_pixels16_c;
+    c->biweight_dirac_pixels_tab[2] = biweight_dirac_pixels32_c;
+
+    PIXFUNC(put, 8);
+    PIXFUNC(put, 16);
+    PIXFUNC(put, 32);
+    PIXFUNC(avg, 8);
+    PIXFUNC(avg, 16);
+    PIXFUNC(avg, 32);
+
+    //MMX_DISABLE    if (HAVE_MMX) ff_diracdsp_init_mmx(c);
+}

libavcodec/diracdsp.h

+/*
+ * Copyright (C) 2010 David Conrad
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_DIRACDSP_H
+#define AVCODEC_DIRACDSP_H
+
+typedef void (*dirac_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int h);
+typedef void (*dirac_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int h);
+
+typedef struct {
+    void (*dirac_hpel_filter)(uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src, int stride, int width, int height);
+    /**
+     * dirac_pixels_tab[width][subpel]
+     * width is 2 for 32, 1 for 16, 0 for 8
+     * subpel is 0 for fpel and hpel (only need to copy from the first plane in src)
+     *           1 if an average of the first 2 planes is needed (TODO: worth it?)
+     *           2 for general qpel (avg of 4)
+     *           3 for general epel (biweight of 4 using the weights in src[4])
+     * src[0-3] is each of the hpel planes
+     * src[4] is the 1/8 pel weights if needed
+     */
+    void (*put_dirac_pixels_tab[3][4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
+    void (*avg_dirac_pixels_tab[3][4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
+
+    void (*put_signed_rect_clamped)(uint8_t *dst/*align 16*/, int dst_stride, const int16_t *src/*align 16*/, int src_stride, int width, int height/*mod 2*/);
+    void (*put_rect_clamped)(uint8_t *dst/*align 16*/, int dst_stride, const int16_t *src/*align 16*/, int src_stride, int width, int height/*mod 2*/);
+    void (*add_rect_clamped)(uint8_t *dst/*align 16*/, const uint16_t *src/*align 16*/, int stride, const int16_t *idwt/*align 16*/, int idwt_stride, int width, int height/*mod 2*/);
+    void (*add_dirac_obmc[3])(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
+
+    dirac_weight_func weight_dirac_pixels_tab[3];
+    dirac_biweight_func biweight_dirac_pixels_tab[3];
+} DiracDSPContext;
+
+#define DECL_DIRAC_PIXOP(PFX, EXT) \
+    void ff_ ## PFX ## _dirac_pixels8_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h); \
+    void ff_ ## PFX ## _dirac_pixels16_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h); \
+    void ff_ ## PFX ## _dirac_pixels32_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)
+
+DECL_DIRAC_PIXOP(put, c);
+DECL_DIRAC_PIXOP(avg, c);
+DECL_DIRAC_PIXOP(put, l2_c);
+DECL_DIRAC_PIXOP(avg, l2_c);
+DECL_DIRAC_PIXOP(put, l4_c);
+DECL_DIRAC_PIXOP(avg, l4_c);
+
+void ff_diracdsp_init(DiracDSPContext *c);
+
+#endif

libavcodec/dsputil.c

@@ -38,6 +38,7 @@
 #include "config.h"
 #include "ac3dec.h"
 #include "vorbis.h"
+#include "diracdsp.h"
 
 uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP] = {0, };
 uint32_t ff_squareTbl[512] = {0, };
@@ -1328,6 +1329,51 @@ void ff_avg_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride){
 }
 #endif /* CONFIG_RV40_DECODER */
 
+#if CONFIG_DIRAC_DECODER
+#define DIRAC_MC(OPNAME)\
+void ff_ ## OPNAME ## _dirac_pixels8_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+     OPNAME ## _pixels8_8_c(dst, src[0], stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels16_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_8_c(dst, src[0], stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels32_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_8_c(dst   , src[0]   , stride, h);\
+    OPNAME ## _pixels16_8_c(dst+16, src[0]+16, stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels8_l2_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels8_l2_8(dst, src[0], src[1], stride, stride, stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels16_l2_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_l2_8(dst, src[0], src[1], stride, stride, stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels32_l2_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_l2_8(dst   , src[0]   , src[1]   , stride, stride, stride, h);\
+    OPNAME ## _pixels16_l2_8(dst+16, src[0]+16, src[1]+16, stride, stride, stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels8_l4_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels8_l4_8(dst, src[0], src[1], src[2], src[3], stride, stride, stride, stride, stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels16_l4_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_l4_8(dst, src[0], src[1], src[2], src[3], stride, stride, stride, stride, stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels32_l4_c(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_l4_8(dst   , src[0]   , src[1]   , src[2]   , src[3]   , stride, stride, stride, stride, stride, h);\
+    OPNAME ## _pixels16_l4_8(dst+16, src[0]+16, src[1]+16, src[2]+16, src[3]+16, stride, stride, stride, stride, stride, h);\
+}
+DIRAC_MC(put)
+DIRAC_MC(avg)
+#endif
+
 static void wmv2_mspel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int w){
     uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
     int i;

libavcodec/dwt.h

@@ -26,7 +26,12 @@
 typedef int DWTELEM;
 typedef short IDWTELEM;
 
+#define MAX_DWT_SUPPORT 8
+#define MAX_DECOMPOSITIONS 8
+
 typedef struct {
+    IDWTELEM *b[MAX_DWT_SUPPORT];
+
     IDWTELEM *b0;
     IDWTELEM *b1;
     IDWTELEM *b2;
@@ -45,13 +50,99 @@ typedef struct slice_buffer_s {
     IDWTELEM * base_buffer; ///< Buffer that this structure is caching.
 } slice_buffer;
 
+struct DWTContext;
+
+// Possible prototypes for vertical_compose functions
+typedef void (*vertical_compose_2tap)(IDWTELEM *b0, IDWTELEM *b1, int width);
+typedef void (*vertical_compose_3tap)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width);
+typedef void (*vertical_compose_5tap)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, int width);
+typedef void (*vertical_compose_9tap)(IDWTELEM *dst, IDWTELEM *b[8], int width);
+
 typedef struct DWTContext {
+    IDWTELEM *buffer;
+    IDWTELEM *temp;
+    int width;
+    int height;
+    int stride;
+    int decomposition_count;
+    int support;
+
+    void (*spatial_compose)(struct DWTContext *cs, int level, int width, int height, int stride);
+    void (*vertical_compose_l0)();
+    void (*vertical_compose_h0)();
+    void (*vertical_compose_l1)();
+    void (*vertical_compose_h1)();
+    void (*vertical_compose)();     ///< one set of lowpass and highpass combined
+    void (*horizontal_compose)(IDWTELEM *b, IDWTELEM *tmp, int width);
+
     void (*vertical_compose97i)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width);
     void (*horizontal_compose97i)(IDWTELEM *b, int width);
     void (*inner_add_yblock)(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
+
+    DWTCompose cs[MAX_DECOMPOSITIONS];
 } DWTContext;
 
-#define MAX_DECOMPOSITIONS 8
+enum dwt_type {
+    DWT_SNOW_DAUB9_7,
+    DWT_SNOW_LEGALL5_3,
+    DWT_DIRAC_DD9_7,
+    DWT_DIRAC_LEGALL5_3,
+    DWT_DIRAC_DD13_7,
+    DWT_DIRAC_HAAR0,
+    DWT_DIRAC_HAAR1,
+    DWT_DIRAC_FIDELITY,
+    DWT_DIRAC_DAUB9_7,
+    DWT_NUM_TYPES
+};
+
+// -1 if an error occurred, e.g. the dwt_type isn't recognized
+int ff_spatial_idwt_init2(DWTContext *d, IDWTELEM *buffer, int width, int height,
+                          int stride, enum dwt_type type, int decomposition_count,
+                          IDWTELEM *temp);
+
+int ff_spatial_idwt2(IDWTELEM *buffer, int width, int height, int stride,
+                     enum dwt_type type, int decomposition_count, IDWTELEM *temp);
+
+void ff_spatial_idwt_slice2(DWTContext *d, int y);
+
+// shared stuff for simd optimiztions
+#define COMPOSE_53iL0(b0, b1, b2)\
+    (b1 - ((b0 + b2 + 2) >> 2))
+
+#define COMPOSE_DIRAC53iH0(b0, b1, b2)\
+    (b1 + ((b0 + b2 + 1) >> 1))
+
+#define COMPOSE_DD97iH0(b0, b1, b2, b3, b4)\
+    (b2 + ((-b0 + 9*b1 + 9*b3 - b4 + 8) >> 4))
+
+#define COMPOSE_DD137iL0(b0, b1, b2, b3, b4)\
+    (b2 - ((-b0 + 9*b1 + 9*b3 - b4 + 16) >> 5))
+
+#define COMPOSE_HAARiL0(b0, b1)\
+    (b0 - ((b1 + 1) >> 1))
+
+#define COMPOSE_HAARiH0(b0, b1)\
+    (b0 + b1)
+
+#define COMPOSE_FIDELITYiL0(b0, b1, b2, b3, b4, b5, b6, b7, b8)\
+    (b4 - ((-8*(b0+b8) + 21*(b1+b7) - 46*(b2+b6) + 161*(b3+b5) + 128) >> 8))
+
+#define COMPOSE_FIDELITYiH0(b0, b1, b2, b3, b4, b5, b6, b7, b8)\
+    (b4 + ((-2*(b0+b8) + 10*(b1+b7) - 25*(b2+b6) + 81*(b3+b5) + 128) >> 8))
+
+#define COMPOSE_DAUB97iL1(b0, b1, b2)\
+    (b1 - ((1817*(b0 + b2) + 2048) >> 12))
+
+#define COMPOSE_DAUB97iH1(b0, b1, b2)\
+    (b1 - (( 113*(b0 + b2) + 64) >> 7))
+
+#define COMPOSE_DAUB97iL0(b0, b1, b2)\
+    (b1 + (( 217*(b0 + b2) + 2048) >> 12))
+
+#define COMPOSE_DAUB97iH0(b0, b1, b2)\
+    (b1 + ((6497*(b0 + b2) + 2048) >> 12))
+
+
 
 #define DWT_97 0
 #define DWT_53 1

libavcodec/x86/Makefile

@@ -9,6 +9,8 @@ YASM-OBJS-FFT-$(HAVE_SSE)              += x86/fft_sse.o
 YASM-OBJS-$(CONFIG_FFT)                += x86/fft_mmx.o                 \
                                           $(YASM-OBJS-FFT-yes)
 
+YASM-OBJS-$(CONFIG_DWT)                += x86/dwt_yasm.o
+
 MMX-OBJS-$(CONFIG_H264DSP)             += x86/h264dsp_mmx.o
 YASM-OBJS-$(CONFIG_H264DSP)            += x86/h264_deblock.o            \
                                           x86/h264_deblock_10bit.o      \
@@ -25,6 +27,8 @@ MMX-OBJS-$(CONFIG_RV40_DECODER)        += x86/rv40dsp.o                 \
 
 YASM-OBJS-$(CONFIG_VC1_DECODER)        += x86/vc1dsp_yasm.o
 
+#DISABLE_MMX YASM-OBJS-$(CONFIG_DIRAC_DECODER)      += x86/diracdsp_mmx.o x86/diracdsp_yasm.o
+
 MMX-OBJS-$(CONFIG_AC3DSP)              += x86/ac3dsp_mmx.o
 YASM-OBJS-$(CONFIG_AC3DSP)             += x86/ac3dsp.o
 MMX-OBJS-$(CONFIG_CAVS_DECODER)        += x86/cavsdsp_mmx.o
@@ -38,7 +42,8 @@ YASM-OBJS-$(CONFIG_PRORES_LGPL_DECODER)     += x86/proresdsp.o
 MMX-OBJS-$(CONFIG_PRORES_LGPL_DECODER)      += x86/proresdsp-init.o
 YASM-OBJS-$(CONFIG_PRORES_DECODER)     += x86/proresdsp.o
 MMX-OBJS-$(CONFIG_PRORES_DECODER)      += x86/proresdsp-init.o
-MMX-OBJS-$(CONFIG_DWT)                 += x86/snowdsp_mmx.o
+MMX-OBJS-$(CONFIG_DWT)                 += x86/snowdsp_mmx.o \
+                                          x86/dwt.o
 YASM-OBJS-$(CONFIG_V210_DECODER)       += x86/v210.o
 MMX-OBJS-$(CONFIG_V210_DECODER)        += x86/v210-init.o
 MMX-OBJS-$(CONFIG_VC1_DECODER)         += x86/vc1dsp_mmx.o

libavcodec/x86/diracdsp_mmx.c

+/*
+ * Copyright (C) 2010 David Conrad
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "dsputil_mmx.h"
+#include "diracdsp_mmx.h"
+
+void ff_put_rect_clamped_mmx(uint8_t *dst, int dst_stride, const int16_t *src, int src_stride, int width, int height);
+void ff_put_rect_clamped_sse2(uint8_t *dst, int dst_stride, const int16_t *src, int src_stride, int width, int height);
+void ff_put_signed_rect_clamped_mmx(uint8_t *dst, int dst_stride, const int16_t *src, int src_stride, int width, int height);
+void ff_put_signed_rect_clamped_sse2(uint8_t *dst, int dst_stride, const int16_t *src, int src_stride, int width, int height);
+
+#define HPEL_FILTER(MMSIZE, EXT) \
+void ff_dirac_hpel_filter_v_ ## EXT(uint8_t *, uint8_t *, int, int);\
+void ff_dirac_hpel_filter_h_ ## EXT(uint8_t *, uint8_t *, int);\
+\
+static void dirac_hpel_filter_ ## EXT(uint8_t *dsth, uint8_t *dstv, uint8_t *dstc,\
+                                      uint8_t *src, int stride, int width, int height)\
+{\
+    while( height-- )\
+    {\
+        ff_dirac_hpel_filter_v_ ## EXT(dstv-MMSIZE, src-MMSIZE, stride, width+MMSIZE+5);\
+        ff_dirac_hpel_filter_h_ ## EXT(dsth, src, width);\
+        ff_dirac_hpel_filter_h_ ## EXT(dstc, dstv, width);\
+\
+        dsth += stride;\
+        dstv += stride;\
+        dstc += stride;\
+        src  += stride;\
+    }\
+}
+
+#if !ARCH_X86_64
+HPEL_FILTER(8, mmx)
+#endif
+HPEL_FILTER(16, sse2)
+
+#define PIXFUNC(PFX, IDX, EXT) \
+    c->PFX ## _dirac_pixels_tab[0][IDX] = ff_ ## PFX ## _dirac_pixels8_ ## EXT; \
+    c->PFX ## _dirac_pixels_tab[1][IDX] = ff_ ## PFX ## _dirac_pixels16_ ## EXT; \
+    c->PFX ## _dirac_pixels_tab[2][IDX] = ff_ ## PFX ## _dirac_pixels32_ ## EXT
+
+void ff_diracdsp_init_mmx(DiracDSPContext* c)
+{
+  int mm_flags = av_get_cpu_flags();;
+
+#if HAVE_YASM
+    c->add_dirac_obmc[0] = ff_add_dirac_obmc8_mmx;
+#if !ARCH_X86_64
+    c->add_dirac_obmc[1] = ff_add_dirac_obmc16_mmx;
+    c->add_dirac_obmc[2] = ff_add_dirac_obmc32_mmx;
+    c->dirac_hpel_filter = dirac_hpel_filter_mmx;
+    c->add_rect_clamped = ff_add_rect_clamped_mmx;
+    c->put_signed_rect_clamped = ff_put_signed_rect_clamped_mmx;
+#endif
+#endif
+
+    PIXFUNC(put, 0, mmx);
+    PIXFUNC(avg, 0, mmx);
+
+    if (mm_flags & AV_CPU_FLAG_MMX2) {
+        PIXFUNC(avg, 0, mmx2);
+    }
+
+    if (mm_flags & AV_CPU_FLAG_SSE2) {
+#if HAVE_YASM
+        c->dirac_hpel_filter = dirac_hpel_filter_sse2;
+        c->add_rect_clamped = ff_add_rect_clamped_sse2;
+        c->put_signed_rect_clamped = ff_put_signed_rect_clamped_sse2;
+
+        c->add_dirac_obmc[1] = ff_add_dirac_obmc16_sse2;
+        c->add_dirac_obmc[2] = ff_add_dirac_obmc32_sse2;
+#endif
+        c->put_dirac_pixels_tab[1][0] = ff_put_dirac_pixels16_sse2;
+        c->avg_dirac_pixels_tab[1][0] = ff_avg_dirac_pixels16_sse2;
+        c->put_dirac_pixels_tab[2][0] = ff_put_dirac_pixels32_sse2;
+        c->avg_dirac_pixels_tab[2][0] = ff_avg_dirac_pixels32_sse2;
+    }
+}

libavcodec/x86/diracdsp_mmx.h

+/*
+ * Copyright (c) 2010 David Conrad
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_X86_DIRACDSP_H
+#define AVCODEC_X86_DIRACDSP_H
+
+#include "libavcodec/diracdsp.h"
+
+void ff_diracdsp_init_mmx(DiracDSPContext* c);
+
+DECL_DIRAC_PIXOP(put, mmx);
+DECL_DIRAC_PIXOP(avg, mmx);
+DECL_DIRAC_PIXOP(avg, mmx2);
+
+void ff_put_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h);
+void ff_avg_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h);
+void ff_put_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h);
+void ff_avg_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h);
+
+void ff_add_rect_clamped_mmx(uint8_t *, const uint16_t *, int, const int16_t *, int, int, int);
+void ff_add_rect_clamped_sse2(uint8_t *, const uint16_t *, int, const int16_t *, int, int, int);
+
+void ff_add_dirac_obmc8_mmx(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
+void ff_add_dirac_obmc16_mmx(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
+void ff_add_dirac_obmc32_mmx(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
+
+void ff_add_dirac_obmc16_sse2(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
+void ff_add_dirac_obmc32_sse2(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
+
+#endif

libavcodec/x86/diracdsp_yasm.asm

+;******************************************************************************
+;* Copyright (c) 2010 David Conrad
+;*
+;* This file is part of FFmpeg.
+;*
+;* FFmpeg is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 of the License, or (at your option) any later version.
+;*
+;* FFmpeg is distributed in the hope that it will be useful,
+;* but WITHOUT ANY WARRANTY; without even the implied warranty of
+;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;* Lesser General Public License for more details.
+;*
+;* You should have received a copy of the GNU Lesser General Public
+;* License along with FFmpeg; if not, write to the Free Software
+;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+;******************************************************************************
+
+%include "x86inc.asm"
+
+SECTION_RODATA
+pw_3: times 8 dw 3
+pw_7: times 8 dw 7
+pw_16: times 8 dw 16
+pw_32: times 8 dw 32
+pb_128: times 16 db 128
+
+section .text
+
+%macro UNPACK_ADD 6
+    mov%5   %1, %3
+    mov%6   m5, %4
+    mova    m4, %1
+    mova    %2, m5
+    punpcklbw %1, m7
+    punpcklbw m5, m7
+    punpckhbw m4, m7
+    punpckhbw %2, m7
+    paddw   %1, m5
+    paddw   %2, m4
+%endmacro
+
+%macro HPEL_FILTER 1
+; dirac_hpel_filter_v_sse2(uint8_t *dst, uint8_t *src, int stride, int width);
+cglobal dirac_hpel_filter_v_%1, 4,6,8, dst, src, stride, width, src0, stridex3
+    mov     src0q, srcq
+    lea     stridex3q, [3*strideq]
+    sub     src0q, stridex3q
+    pxor    m7, m7
+.loop:
+    ; 7*(src[0] + src[1])
+    UNPACK_ADD m0, m1, [srcq], [srcq + strideq], a,a
+    pmullw  m0, [pw_7]
+    pmullw  m1, [pw_7]
+
+    ; 3*( ... + src[-2] + src[3])
+    UNPACK_ADD m2, m3, [src0q + strideq], [srcq + stridex3q], a,a
+    paddw   m0, m2
+    paddw   m1, m3
+    pmullw  m0, [pw_3]
+    pmullw  m1, [pw_3]
+
+    ; ... - 7*(src[-1] + src[2])
+    UNPACK_ADD m2, m3, [src0q + strideq*2], [srcq + strideq*2], a,a
+    pmullw  m2, [pw_7]
+    pmullw  m3, [pw_7]
+    psubw   m0, m2
+    psubw   m1, m3
+
+    ; ... - (src[-3] + src[4])
+    UNPACK_ADD m2, m3, [src0q], [srcq + strideq*4], a,a
+    psubw   m0, m2
+    psubw   m1, m3
+
+    paddw   m0, [pw_16]
+    paddw   m1, [pw_16]
+    psraw   m0, 5
+    psraw   m1, 5
+    packuswb m0, m1
+    mova    [dstq], m0
+    add     dstq, mmsize
+    add     srcq, mmsize
+    add     src0q, mmsize
+    sub     widthd, mmsize
+    jg      .loop
+    RET
+
+; dirac_hpel_filter_h_sse2(uint8_t *dst, uint8_t *src, int width);
+cglobal dirac_hpel_filter_h_%1, 3,3,8, dst, src, width
+    dec     widthd
+    pxor    m7, m7
+    and     widthd, ~(mmsize-1)
+.loop:
+    ; 7*(src[0] + src[1])
+    UNPACK_ADD m0, m1, [srcq + widthq], [srcq + widthq + 1], a,u
+    pmullw  m0, [pw_7]
+    pmullw  m1, [pw_7]
+
+    ; 3*( ... + src[-2] + src[3])
+    UNPACK_ADD m2, m3, [srcq + widthq - 2], [srcq + widthq + 3], u,u
+    paddw   m0, m2
+    paddw   m1, m3
+    pmullw  m0, [pw_3]
+    pmullw  m1, [pw_3]
+
+    ; ... - 7*(src[-1] + src[2])
+    UNPACK_ADD m2, m3, [srcq + widthq - 1], [srcq + widthq + 2], u,u
+    pmullw  m2, [pw_7]
+    pmullw  m3, [pw_7]
+    psubw   m0, m2
+    psubw   m1, m3
+
+    ; ... - (src[-3] + src[4])
+    UNPACK_ADD m2, m3, [srcq + widthq - 3], [srcq + widthq + 4], u,u
+    psubw   m0, m2
+    psubw   m1, m3
+
+    paddw   m0, [pw_16]
+    paddw   m1, [pw_16]
+    psraw   m0, 5
+    psraw   m1, 5
+    packuswb m0, m1
+    mova    [dstq + widthq], m0
+    sub     widthd, mmsize
+    jge     .loop
+    RET
+%endmacro
+
+%macro PUT_RECT 1
+; void put_rect_clamped(uint8_t *dst, int dst_stride, int16_t *src, int src_stride, int width, int height)
+cglobal put_signed_rect_clamped_%1, 5,7,3, dst, dst_stride, src, src_stride, w, dst2, src2
+    mova    m0, [pb_128]
+    add     wd, (mmsize-1)
+    and     wd, ~(mmsize-1)
+
+%ifdef ARCH_X86_64
+    mov   r10d, r5m
+    mov   r11d, wd
+    %define wspill r11d
+    %define hd r10d
+%else
+    mov    r4m, wd
+    %define wspill r4m
+    %define hd r5mp
+%endif
+
+.loopy
+    lea     src2q, [srcq+src_strideq*2]
+    lea     dst2q, [dstq+dst_strideq]
+.loopx:
+    sub      wd, mmsize
+    mova     m1, [srcq +2*wq]
+    mova     m2, [src2q+2*wq]
+    packsswb m1, [srcq +2*wq+mmsize]
+    packsswb m2, [src2q+2*wq+mmsize]
+    paddb    m1, m0
+    paddb    m2, m0
+    mova    [dstq +wq], m1
+    mova    [dst2q+wq], m2
+    jg      .loopx
+
+    lea   srcq, [srcq+src_strideq*4]
+    lea   dstq, [dstq+dst_strideq*2]
+    sub     hd, 2
+    mov     wd, wspill
+    jg      .loopy
+    RET
+%endm
+
+%macro ADD_RECT 1
+; void add_rect_clamped(uint8_t *dst, uint16_t *src, int stride, int16_t *idwt, int idwt_stride, int width, int height)
+cglobal add_rect_clamped_%1, 7,7,3, dst, src, stride, idwt, idwt_stride, w, h
+    mova    m0, [pw_32]
+    add     wd, (mmsize-1)
+    and     wd, ~(mmsize-1)
+
+%ifdef ARCH_X86_64
+    mov   r11d, wd
+    %define wspill r11d
+%else
+    mov    r5m, wd
+    %define wspill r5m
+%endif
+
+.loop:
+    sub     wd, mmsize
+    movu    m1, [srcq +2*wq] ; FIXME: ensure alignment
+    paddw   m1, m0
+    psraw   m1, 6
+    movu    m2, [srcq +2*wq+mmsize] ; FIXME: ensure alignment
+    paddw   m2, m0
+    psraw   m2, 6
+    paddw   m1, [idwtq+2*wq]
+    paddw   m2, [idwtq+2*wq+mmsize]
+    packuswb m1, m2
+    mova    [dstq +wq], m1
+    jg      .loop
+
+    lea   srcq, [srcq + 2*strideq]
+    add   dstq, strideq
+    lea  idwtq, [idwtq+ 2*idwt_strideq]
+    sub     hd, 1
+    mov     wd, wspill
+    jg      .loop
+    RET
+%endm
+
+%macro ADD_OBMC 2
+; void add_obmc(uint16_t *dst, uint8_t *src, int stride, uint8_t *obmc_weight, int yblen)
+cglobal add_dirac_obmc%1_%2, 6,6,5, dst, src, stride, obmc, yblen
+    pxor        m4, m4
+.loop:
+%assign i 0
+%rep %1 / mmsize
+    mova        m0, [srcq+i]
+    mova        m1, m0
+    punpcklbw   m0, m4
+    punpckhbw   m1, m4
+    mova        m2, [obmcq+i]
+    mova        m3, m2
+   punpcklbw   m2, m4
+    punpckhbw   m3, m4
+    pmullw      m0, m2
+    pmullw      m1, m3
+    movu        m2, [dstq+2*i]
+    movu        m3, [dstq+2*i+mmsize]
+    paddw       m0, m2
+    paddw       m1, m3
+    movu        [dstq+2*i], m0
+    movu        [dstq+2*i+mmsize], m1
+%assign i i+mmsize
+%endrep
+    lea         srcq, [srcq+strideq]
+    lea         dstq, [dstq+2*strideq]
+    add         obmcq, 32
+    sub         yblend, 1
+    jg          .loop
+    RET
+%endm
+
+INIT_MMX
+%ifndef ARCH_X86_64
+PUT_RECT mmx
+ADD_RECT mmx
+
+HPEL_FILTER mmx
+ADD_OBMC 32, mmx
+ADD_OBMC 16, mmx
+%endif
+ADD_OBMC 8, mmx
+
+INIT_XMM
+PUT_RECT sse2
+ADD_RECT sse2
+
+HPEL_FILTER sse2
+ADD_OBMC 32, sse2
+ADD_OBMC 16, sse2

libavcodec/x86/dsputil_mmx.c

@@ -31,6 +31,7 @@
 #include "libavcodec/ac3dec.h"
 #include "dsputil_mmx.h"
 #include "idct_xvid.h"
+#include "diracdsp_mmx.h"
 
 //#undef NDEBUG
 //#include <assert.h>
@@ -1902,6 +1903,46 @@ static void put_vp_no_rnd_pixels16_l2_mmx(uint8_t *dst, const uint8_t *a, const
     put_vp_no_rnd_pixels8_l2_mmx(dst+8, a+8, b+8, stride, h);
 }
 
+#if CONFIG_DIRAC_DECODER
+#define DIRAC_PIXOP(OPNAME, EXT)\
+void ff_ ## OPNAME ## _dirac_pixels8_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels8_ ## EXT(dst, src[0], stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels16_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_ ## EXT(dst, src[0], stride, h);\
+}\
+void ff_ ## OPNAME ## _dirac_pixels32_ ## EXT(uint8_t *dst, const uint8_t *src[5], int stride, int h)\
+{\
+    OPNAME ## _pixels16_ ## EXT(dst   , src[0]   , stride, h);\
+    OPNAME ## _pixels16_ ## EXT(dst+16, src[0]+16, stride, h);\
+}
+
+DIRAC_PIXOP(put, mmx)
+DIRAC_PIXOP(avg, mmx)
+DIRAC_PIXOP(avg, mmx2)
+
+void ff_put_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
+{
+    put_pixels16_sse2(dst, src[0], stride, h);
+}
+void ff_avg_dirac_pixels16_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
+{
+    avg_pixels16_sse2(dst, src[0], stride, h);
+}
+void ff_put_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
+{
+    put_pixels16_sse2(dst   , src[0]   , stride, h);
+    put_pixels16_sse2(dst+16, src[0]+16, stride, h);
+}
+void ff_avg_dirac_pixels32_sse2(uint8_t *dst, const uint8_t *src[5], int stride, int h)
+{
+    avg_pixels16_sse2(dst   , src[0]   , stride, h);
+    avg_pixels16_sse2(dst+16, src[0]+16, stride, h);
+}
+#endif
+
 /* XXX: those functions should be suppressed ASAP when all IDCTs are
    converted */
 #if CONFIG_GPL

libavcodec/x86/dwt.c

+/*
+ * MMX optimized discrete wavelet transform
+ * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
+ * Copyright (c) 2010 David Conrad
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/x86_cpu.h"
+#include "dsputil_mmx.h"
+#include "dwt.h"
+
+#define COMPOSE_VERTICAL(ext, align) \
+void ff_vertical_compose53iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width); \
+void ff_vertical_compose_dirac53iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width); \
+void ff_vertical_compose_dd137iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, int width); \
+void ff_vertical_compose_dd97iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, int width); \
+void ff_vertical_compose_haar##ext(IDWTELEM *b0, IDWTELEM *b1, int width); \
+\
+static void vertical_compose53iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width) \
+{ \
+    int i, width_align = width&~(align-1); \
+\
+    for(i=width_align; i<width; i++) \
+        b1[i] = COMPOSE_53iL0(b0[i], b1[i], b2[i]); \
+\
+    ff_vertical_compose53iL0##ext(b0, b1, b2, width_align); \
+} \
+\
+static void vertical_compose_dirac53iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width) \
+{ \
+    int i, width_align = width&~(align-1); \
+\
+    for(i=width_align; i<width; i++) \
+        b1[i] = COMPOSE_DIRAC53iH0(b0[i], b1[i], b2[i]); \
+\
+    ff_vertical_compose_dirac53iH0##ext(b0, b1, b2, width_align); \
+} \
+\
+static void vertical_compose_dd137iL0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, \
+                                           IDWTELEM *b3, IDWTELEM *b4, int width) \
+{ \
+    int i, width_align = width&~(align-1); \
+\
+    for(i=width_align; i<width; i++) \
+        b2[i] = COMPOSE_DD137iL0(b0[i], b1[i], b2[i], b3[i], b4[i]); \
+\
+    ff_vertical_compose_dd137iL0##ext(b0, b1, b2, b3, b4, width_align); \
+} \
+\
+static void vertical_compose_dd97iH0##ext(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, \
+                                          IDWTELEM *b3, IDWTELEM *b4, int width) \
+{ \
+    int i, width_align = width&~(align-1); \
+\
+    for(i=width_align; i<width; i++) \
+        b2[i] = COMPOSE_DD97iH0(b0[i], b1[i], b2[i], b3[i], b4[i]); \
+\
+    ff_vertical_compose_dd97iH0##ext(b0, b1, b2, b3, b4, width_align); \
+} \
+static void vertical_compose_haar##ext(IDWTELEM *b0, IDWTELEM *b1, int width) \
+{ \
+    int i, width_align = width&~(align-1); \
+\
+    for(i=width_align; i<width; i++) { \
+        b0[i] = COMPOSE_HAARiL0(b0[i], b1[i]); \
+        b1[i] = COMPOSE_HAARiH0(b1[i], b0[i]); \
+    } \
+\
+    ff_vertical_compose_haar##ext(b0, b1, width_align); \
+} \
+\
+
+#if HAVE_YASM
+#if !ARCH_X86_64
+COMPOSE_VERTICAL(_mmx, 4)
+#endif
+COMPOSE_VERTICAL(_sse2, 8)
+#endif
+
+
+void ff_horizontal_compose_dd97i_ssse3(IDWTELEM *b, IDWTELEM *tmp, int w);
+
+void ff_horizontal_compose_haar0i_mmx(IDWTELEM *b, IDWTELEM *tmp, int w);
+void ff_horizontal_compose_haar1i_mmx(IDWTELEM *b, IDWTELEM *tmp, int w);
+void ff_horizontal_compose_haar0i_sse2(IDWTELEM *b, IDWTELEM *tmp, int w);
+void ff_horizontal_compose_haar1i_sse2(IDWTELEM *b, IDWTELEM *tmp, int w);
+
+void ff_horizontal_compose_dd97i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x)
+{
+    for (; x < w2; x++) {
+        b[2*x  ] = (tmp[x] + 1)>>1;
+        b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
+    }
+}
+
+void ff_horizontal_compose_haar0i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x)
+{
+    for (; x < w2; x++) {
+        b[2*x  ] = tmp[x];
+        b[2*x+1] = COMPOSE_HAARiH0(b[x+w2], tmp[x]);
+    }
+}
+
+void ff_horizontal_compose_haar1i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x)
+{
+    for (; x < w2; x++) {
+        b[2*x  ] = (tmp[x] + 1)>>1;
+        b[2*x+1] = (COMPOSE_HAARiH0(b[x+w2], tmp[x]) + 1)>>1;
+    }
+}
+
+void ff_spatial_idwt_init_mmx(DWTContext *d, enum dwt_type type)
+{
+#if HAVE_YASM
+  int mm_flags = av_get_cpu_flags();;
+
+#if !ARCH_X86_64
+    if (!(mm_flags & AV_CPU_FLAG_MMX))
+        return;
+
+    switch (type) {
+    case DWT_DIRAC_DD9_7:
+        d->vertical_compose_l0 = vertical_compose53iL0_mmx;
+        d->vertical_compose_h0 = vertical_compose_dd97iH0_mmx;
+        break;
+    case DWT_DIRAC_LEGALL5_3:
+        d->vertical_compose_l0 = vertical_compose53iL0_mmx;
+        d->vertical_compose_h0 = vertical_compose_dirac53iH0_mmx;
+        break;
+    case DWT_DIRAC_DD13_7:
+        d->vertical_compose_l0 = vertical_compose_dd137iL0_mmx;
+        d->vertical_compose_h0 = vertical_compose_dd97iH0_mmx;
+        break;
+    case DWT_DIRAC_HAAR0:
+        d->vertical_compose   = vertical_compose_haar_mmx;
+        d->horizontal_compose = ff_horizontal_compose_haar0i_mmx;
+        break;
+    case DWT_DIRAC_HAAR1:
+        d->vertical_compose   = vertical_compose_haar_mmx;
+        d->horizontal_compose = ff_horizontal_compose_haar1i_mmx;
+        break;
+    }
+#endif
+
+    if (!(mm_flags & AV_CPU_FLAG_SSE2))
+        return;
+
+    switch (type) {
+    case DWT_DIRAC_DD9_7:
+        d->vertical_compose_l0 = vertical_compose53iL0_sse2;
+        d->vertical_compose_h0 = vertical_compose_dd97iH0_sse2;
+        break;
+    case DWT_DIRAC_LEGALL5_3:
+        d->vertical_compose_l0 = vertical_compose53iL0_sse2;
+        d->vertical_compose_h0 = vertical_compose_dirac53iH0_sse2;
+        break;
+    case DWT_DIRAC_DD13_7:
+        d->vertical_compose_l0 = vertical_compose_dd137iL0_sse2;
+        d->vertical_compose_h0 = vertical_compose_dd97iH0_sse2;
+        break;
+    case DWT_DIRAC_HAAR0:
+        d->vertical_compose   = vertical_compose_haar_sse2;
+        d->horizontal_compose = ff_horizontal_compose_haar0i_sse2;
+        break;
+    case DWT_DIRAC_HAAR1:
+        d->vertical_compose   = vertical_compose_haar_sse2;
+        d->horizontal_compose = ff_horizontal_compose_haar1i_sse2;
+        break;
+    }
+
+    if (!(mm_flags & AV_CPU_FLAG_SSSE3))
+        return;
+
+    switch (type) {
+    case DWT_DIRAC_DD9_7:
+        d->horizontal_compose = ff_horizontal_compose_dd97i_ssse3;
+        break;
+    }
+#endif // HAVE_YASM
+}

libavcodec/x86/dwt.h

+/*
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_X86_DWT_H
+#define AVCODEC_X86_DWT_H
+
+#include "libavcodec/dwt.h"
+
+void ff_horizontal_compose_dd97i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x);
+void ff_horizontal_compose_haar1i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x);
+void ff_horizontal_compose_haar0i_end_c(IDWTELEM *b, IDWTELEM *tmp, int w2, int x);
+
+void ff_spatial_idwt_init_mmx(DWTContext *d, enum dwt_type type);
+
+#endif

libavcodec/x86/dwt_yasm.asm

+;******************************************************************************
+;* MMX optimized discrete wavelet trasnform
+;* Copyright (c) 2010 David Conrad
+;*
+;* This file is part of FFmpeg.
+;*
+;* FFmpeg is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 of the License, or (at your option) any later version.
+;*
+;* FFmpeg is distributed in the hope that it will be useful,
+;* but WITHOUT ANY WARRANTY; without even the implied warranty of
+;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;* Lesser General Public License for more details.
+;*
+;* You should have received a copy of the GNU Lesser General Public
+;* License along with FFmpeg; if not, write to the Free Software
+;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+;******************************************************************************
+
+%include "x86inc.asm"
+
+cextern horizontal_compose_dd97i_end_c
+cextern horizontal_compose_haar0i_end_c
+cextern horizontal_compose_haar1i_end_c
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+pw_2: times 8 dw 2
+pw_8: times 8 dw 8
+pw_16: times 8 dw 16
+pw_1991: times 4 dw 9,-1
+
+section .text
+
+; %1 -= (%2 + %3 + 2)>>2     %4 is pw_2
+%macro COMPOSE_53iL0 4
+    paddw   %2, %3
+    paddw   %2, %4
+    psraw   %2, 2
+    psubw   %1, %2
+%endm
+
+; m1 = %1 + (-m0 + 9*m1 + 9*%2 -%3 + 8)>>4
+; if %4 is supplied, %1 is loaded unaligned from there
+; m2: clobbered  m3: pw_8  m4: pw_1991
+%macro COMPOSE_DD97iH0 3-4
+    paddw   m0, %3
+    paddw   m1, %2
+    psubw   m0, m3
+    mova    m2, m1
+    punpcklwd m1, m0
+    punpckhwd m2, m0
+    pmaddwd m1, m4
+    pmaddwd m2, m4
+%if %0 > 3
+    movu    %1, %4
+%endif
+    psrad   m1, 4
+    psrad   m2, 4
+    packssdw m1, m2
+    paddw   m1, %1
+%endm
+
+%macro COMPOSE_VERTICAL 1
+; void vertical_compose53iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
+;                                  int width)
+cglobal vertical_compose53iL0_%1, 4,4,1, b0, b1, b2, width
+    mova    m2, [pw_2]
+.loop:
+    sub     widthd, mmsize/2
+    mova    m1, [b0q+2*widthq]
+    mova    m0, [b1q+2*widthq]
+    COMPOSE_53iL0 m0, m1, [b2q+2*widthq], m2
+    mova    [b1q+2*widthq], m0
+    jg      .loop
+    REP_RET
+
+; void vertical_compose_dirac53iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
+;                                  int width)
+cglobal vertical_compose_dirac53iH0_%1, 4,4,1, b0, b1, b2, width
+    mova    m1, [pw_1]
+.loop:
+    sub     widthd, mmsize/2
+    mova    m0, [b0q+2*widthq]
+    paddw   m0, [b2q+2*widthq]
+    paddw   m0, m1
+    psraw   m0, 1
+    paddw   m0, [b1q+2*widthq]
+    mova    [b1q+2*widthq], m0
+    jg      .loop
+    REP_RET
+
+; void vertical_compose_dd97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
+;                               IDWTELEM *b3, IDWTELEM *b4, int width)
+cglobal vertical_compose_dd97iH0_%1, 6,6,5, b0, b1, b2, b3, b4, width
+    mova    m3, [pw_8]
+    mova    m4, [pw_1991]
+.loop:
+    sub     widthd, mmsize/2
+    mova    m0, [b0q+2*widthq]
+    mova    m1, [b1q+2*widthq]
+    COMPOSE_DD97iH0 [b2q+2*widthq], [b3q+2*widthq], [b4q+2*widthq]
+    mova    [b2q+2*widthq], m1
+    jg      .loop
+    REP_RET
+
+; void vertical_compose_dd137iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
+;                                IDWTELEM *b3, IDWTELEM *b4, int width)
+cglobal vertical_compose_dd137iL0_%1, 6,6,6, b0, b1, b2, b3, b4, width
+    mova    m3, [pw_16]
+    mova    m4, [pw_1991]
+.loop:
+    sub     widthd, mmsize/2
+    mova    m0, [b0q+2*widthq]
+    mova    m1, [b1q+2*widthq]
+    mova    m5, [b2q+2*widthq]
+    paddw   m0, [b4q+2*widthq]
+    paddw   m1, [b3q+2*widthq]
+    psubw   m0, m3
+    mova    m2, m1
+    punpcklwd m1, m0
+    punpckhwd m2, m0
+    pmaddwd m1, m4
+    pmaddwd m2, m4
+    psrad   m1, 5
+    psrad   m2, 5
+    packssdw m1, m2
+    psubw   m5, m1
+    mova    [b2q+2*widthq], m5
+    jg      .loop
+    REP_RET
+
+; void vertical_compose_haar(IDWTELEM *b0, IDWTELEM *b1, int width)
+cglobal vertical_compose_haar_%1, 3,4,3, b0, b1, width
+    mova    m3, [pw_1]
+.loop:
+    sub     widthd, mmsize/2
+    mova    m1, [b1q+2*widthq]
+    mova    m0, [b0q+2*widthq]
+    mova    m2, m1
+    paddw   m1, m3
+    psraw   m1, 1
+    psubw   m0, m1
+    mova    [b0q+2*widthq], m0
+    paddw   m2, m0
+    mova    [b1q+2*widthq], m2
+    jg      .loop
+    REP_RET
+%endmacro
+
+; extend the left and right edges of the tmp array by %1 and %2 respectively
+%macro EDGE_EXTENSION 3
+    mov     %3, [tmpq]
+%assign %%i 1
+%rep %1
+    mov     [tmpq-2*%%i], %3
+    %assign %%i %%i+1
+%endrep
+    mov     %3, [tmpq+2*w2q-2]
+%assign %%i 0
+%rep %2
+    mov     [tmpq+2*w2q+2*%%i], %3
+    %assign %%i %%i+1
+%endrep
+%endmacro
+
+; On x86-64 this does a tail call to the C function to do the final bit
+; x86-32 doesn't because isn't enough stack space for the additional argument x
+%macro END_HORIZONTAL 1
+    shr     wd, 1
+%ifdef ARCH_X86_64
+    RET                                ;This RET was a CLEANUP call
+    jmp     %1
+%else
+    push    xd
+    push    wd
+    push    tmpd
+    push    bd
+    call    %1
+    add     esp, 16
+    RET
+%endif
+%endmacro
+
+%macro HAAR_HORIZONTAL 2
+; void horizontal_compose_haari(IDWTELEM *b, IDWTELEM *tmp, int width)
+cglobal horizontal_compose_haar%2i_%1, 3,6,4, b, tmp, w, x, w2, b_w2
+    mov    w2d, wd
+    xor     xd, xd
+    shr    w2d, 1
+    lea  b_w2q, [bq+wq]
+    mova    m3, [pw_1]
+.lowpass_loop:
+    movu    m1, [b_w2q + 2*xq]
+    mova    m0, [bq    + 2*xq]
+    paddw   m1, m3
+    psraw   m1, 1
+    psubw   m0, m1
+    mova    [tmpq + 2*xq], m0
+    add     xd, mmsize/2
+    cmp     xd, w2d
+    jl      .lowpass_loop
+
+    xor     xd, xd
+    and    w2d, ~(mmsize/2 - 1)
+    cmp    w2d, mmsize/2
+    jl      .end
+
+.highpass_loop:
+    mova    m1, [b_w2q + 2*xq]
+    mova    m0, [tmpq  + 2*xq]
+    paddw   m1, m0
+
+    ; shift and interleave
+%if %2 == 1
+    paddw   m0, m3
+    paddw   m1, m3
+    psraw   m0, 1
+    psraw   m1, 1
+%endif
+    mova    m2, m0
+    punpcklwd m0, m1
+    punpckhwd m2, m1
+    mova    [bq+4*xq], m0
+    mova    [bq+4*xq+mmsize], m2
+
+    add     xd, mmsize/2
+    cmp     xd, w2d
+    jl      .highpass_loop
+.end:
+    END_HORIZONTAL horizontal_compose_haar%2i_end_c
+%endmacro
+
+
+INIT_XMM
+; void horizontal_compose_dd97i(IDWTELEM *b, IDWTELEM *tmp, int width)
+cglobal horizontal_compose_dd97i_ssse3, 3,6,8, b, tmp, w, x, w2, b_w2
+    mov    w2d, wd
+    xor     xd, xd
+    shr    w2d, 1
+    lea  b_w2q, [bq+wq]
+    movu    m4, [bq+wq]
+    mova    m7, [pw_2]
+    pslldq  m4, 14
+.lowpass_loop:
+    movu    m1, [b_w2q + 2*xq]
+    mova    m0, [bq    + 2*xq]
+    mova    m2, m1
+    palignr m1, m4, 14
+    mova    m4, m2
+    COMPOSE_53iL0 m0, m1, m2, m7
+    mova    [tmpq + 2*xq], m0
+    add     xd, mmsize/2
+    cmp     xd, w2d
+    jl      .lowpass_loop
+
+    EDGE_EXTENSION 1, 2, xw
+    ; leave the last up to 7 (sse) or 3 (mmx) values for C
+    xor     xd, xd
+    and    w2d, ~(mmsize/2 - 1)
+    cmp    w2d, mmsize/2
+    jl      .end
+
+    mova    m7, [tmpq-mmsize]
+    mova    m0, [tmpq]
+    mova    m5, [pw_1]
+    mova    m3, [pw_8]
+    mova    m4, [pw_1991]
+.highpass_loop:
+    mova    m6, m0
+    palignr m0, m7, 14
+    mova    m7, [tmpq + 2*xq + 16]
+    mova    m1, m7
+    mova    m2, m7
+    palignr m1, m6, 2
+    palignr m2, m6, 4
+    COMPOSE_DD97iH0 m0, m6, m2, [b_w2q + 2*xq]
+    mova    m0, m7
+    mova    m7, m6
+
+    ; shift and interleave
+    paddw   m6, m5
+    paddw   m1, m5
+    psraw   m6, 1
+    psraw   m1, 1
+    mova    m2, m6
+    punpcklwd m6, m1
+    punpckhwd m2, m1
+    mova    [bq+4*xq], m6
+    mova    [bq+4*xq+mmsize], m2
+
+    add     xd, mmsize/2
+    cmp     xd, w2d
+    jl      .highpass_loop
+.end:
+    END_HORIZONTAL horizontal_compose_dd97i_end_c
+
+
+%ifndef ARCH_X86_64
+INIT_MMX
+COMPOSE_VERTICAL mmx
+HAAR_HORIZONTAL mmx, 0
+HAAR_HORIZONTAL mmx, 1
+%endif
+
+;;INIT_XMM
+INIT_XMM
+COMPOSE_VERTICAL sse2
+HAAR_HORIZONTAL sse2, 0
+HAAR_HORIZONTAL sse2, 1