lavfi: port mp=fspp to a native libavfilter filter
authorArwa Arif <arwaarif1994@gmail.com>
Tue, 23 Dec 2014 16:54:37 +0000 (22:24 +0530)
committerStefano Sabatini <stefasab@gmail.com>
Wed, 24 Dec 2014 15:29:18 +0000 (16:29 +0100)
Signed-off-by: Stefano Sabatini <stefasab@gmail.com>
LICENSE.md
configure
doc/filters.texi
libavfilter/Makefile
libavfilter/allfilters.c
libavfilter/libmpcodecs/vf_fspp.c
libavfilter/version.h
libavfilter/vf_fspp.c [new file with mode: 0644]
libavfilter/vf_fspp.h [new file with mode: 0644]
libavfilter/x86/Makefile
libavfilter/x86/vf_fspp.c [new file with mode: 0644]

index 59cfad5..5659973 100644 (file)
@@ -31,6 +31,7 @@ Specifically, the GPL parts of FFmpeg are:
     - vf_cropdetect.c
     - vf_decimate.c
     - vf_delogo.c
+    - vf_fspp.c
     - vf_geq.c
     - vf_histeq.c
     - vf_hqdn3d.c
index 87be058..9085200 100755 (executable)
--- a/configure
+++ b/configure
@@ -2582,6 +2582,7 @@ ebur128_filter_deps="gpl"
 flite_filter_deps="libflite"
 frei0r_filter_deps="frei0r dlopen"
 frei0r_src_filter_deps="frei0r dlopen"
+fspp_filter_deps="gpl"
 geq_filter_deps="gpl"
 histeq_filter_deps="gpl"
 hqdn3d_filter_deps="gpl"
index 7fac8fb..ec7e85e 100644 (file)
@@ -5066,6 +5066,37 @@ frei0r=perspective:0.2/0.2|0.8/0.2
 For more information, see
 @url{http://frei0r.dyne.org}
 
+@section fspp
+
+Apply fast and simple postprocessing. It is a faster version of @ref{spp}.
+
+It splits (I)DCT into horizontal/vertical passes. Unlike the simple post-
+processing filter, one of them is performed once per block, not per pixel.
+This allows for much higher speed.
+
+The filter accepts the following options:
+
+@table @option
+@item quality
+Set quality. This option defines the number of levels for averaging. It accepts
+an integer in the range 4-5. Default value is @code{4}.
+
+@item qp
+Force a constant quantization parameter. It accepts an integer in range 0-63.
+If not set, the filter will use the QP from the video stream (if available).
+
+@item strength
+Set filter strength. It accepts an integer in range -15 to 32. Lower values mean
+more details but also more artifacts, while higher values make the image smoother
+but also blurrier. Default value is @code{0} − PSNR optimal.
+
+@item use_bframe_qp
+Enable the use of the QP from the B-Frames if set to @code{1}. Using this
+option may cause flicker since the B-Frames have often larger QP. Default is
+@code{0} (not enabled).
+
+@end table
+
 @section geq
 
 The filter accepts the following options:
@@ -8361,6 +8392,7 @@ stereo3d=abl:sbsr
 @end example
 @end itemize
 
+@anchor{spp}
 @section spp
 
 Apply a simple postprocessing filter that compresses and decompresses the image
index 67a7e4b..9cce2de 100644 (file)
@@ -126,6 +126,7 @@ OBJS-$(CONFIG_FRAMESTEP_FILTER)              += vf_framestep.o
 OBJS-$(CONFIG_FPS_FILTER)                    += vf_fps.o
 OBJS-$(CONFIG_FRAMEPACK_FILTER)              += vf_framepack.o
 OBJS-$(CONFIG_FREI0R_FILTER)                 += vf_frei0r.o
+OBJS-$(CONFIG_FSPP_FILTER)                   += vf_fspp.o
 OBJS-$(CONFIG_GEQ_FILTER)                    += vf_geq.o
 OBJS-$(CONFIG_GRADFUN_FILTER)                += vf_gradfun.o
 OBJS-$(CONFIG_HALDCLUT_FILTER)               += vf_lut3d.o dualinput.o framesync.o
index 066d9af..2911ed9 100644 (file)
@@ -142,6 +142,7 @@ void avfilter_register_all(void)
     REGISTER_FILTER(FRAMEPACK,      framepack,      vf);
     REGISTER_FILTER(FRAMESTEP,      framestep,      vf);
     REGISTER_FILTER(FREI0R,         frei0r,         vf);
+    REGISTER_FILTER(FSPP,           fspp,           vf);
     REGISTER_FILTER(GEQ,            geq,            vf);
     REGISTER_FILTER(GRADFUN,        gradfun,        vf);
     REGISTER_FILTER(HALDCLUT,       haldclut,       vf);
index d457859..3a80dc2 100644 (file)
@@ -710,8 +710,8 @@ const vf_info_t ff_vf_info_fspp = {
 #if HAVE_MMX_INLINE
 
 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_382683433)=FIX64(0.382683433, 14);
-DECLARE_ALIGNED(8, uint64_t, ff_MM_FIX_0_541196100)=FIX64(0.541196100, 14);
-DECLARE_ALIGNED(8, uint64_t, ff_MM_FIX_0_707106781)=FIX64(0.707106781, 14);
+extern uint64_t ff_MM_FIX_0_707106781;
+extern uint64_t ff_MM_FIX_0_541196100;
 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_306562965)=FIX64(1.306562965, 14);
 
 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562_A)=FIX64(1.414213562, 14);
index 8b6a09f..24279a2 100644 (file)
@@ -30,7 +30,7 @@
 #include "libavutil/version.h"
 
 #define LIBAVFILTER_VERSION_MAJOR  5
-#define LIBAVFILTER_VERSION_MINOR  4
+#define LIBAVFILTER_VERSION_MINOR  5
 #define LIBAVFILTER_VERSION_MICRO 100
 
 #define LIBAVFILTER_VERSION_INT AV_VERSION_INT(LIBAVFILTER_VERSION_MAJOR, \
diff --git a/libavfilter/vf_fspp.c b/libavfilter/vf_fspp.c
new file mode 100644 (file)
index 0000000..4a1b7e0
--- /dev/null
@@ -0,0 +1,698 @@
+/*
+ * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
+ * Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>
+ * Copyright (c) 2014 Arwa Arif <arwaarif1994@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU 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
+ * Fast Simple Post-processing filter
+ * This implementation is based on an algorithm described in
+ * "Aria Nosratinia Embedded Post-Processing for
+ * Enhancement of Compressed Images (1999)"
+ * (http://www.utdallas.edu/~aria/papers/vlsisp99.pdf)
+ * Further, with splitting (I)DCT into horizontal/vertical passes, one of
+ * them can be performed once per block, not per pixel. This allows for much
+ * higher speed.
+ *
+ * Originally written by Michael Niedermayer and Nikolaj for the MPlayer
+ * project, and ported by Arwa Arif for FFmpeg.
+ */
+
+#include "libavutil/avassert.h"
+#include "libavutil/imgutils.h"
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+#include "internal.h"
+#include "libavcodec/avcodec.h" //for reference to FF_QSCALE_TYPE
+#include "vf_fspp.h"
+
+#define OFFSET(x) offsetof(FSPPContext, x)
+#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
+static const AVOption fspp_options[] = {
+    { "quality",       "set quality",                          OFFSET(log2_count),    AV_OPT_TYPE_INT, {.i64 = 4},   4, MAX_LEVEL, FLAGS },
+    { "qp",            "force a constant quantizer parameter", OFFSET(qp),            AV_OPT_TYPE_INT, {.i64 = 0},   0, 64,        FLAGS },
+    { "strength",      "set filter strength",                  OFFSET(strength),      AV_OPT_TYPE_INT, {.i64 = 0}, -15, 32,        FLAGS },
+    { "use_bframe_qp", "use B-frames' QP",                     OFFSET(use_bframe_qp), AV_OPT_TYPE_INT, {.i64 = 0},   0, 1,         FLAGS },
+    { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(fspp);
+
+DECLARE_ALIGNED(32, static const uint8_t, dither)[8][8] = {
+    {  0,  48,  12,  60,   3,  51,  15,  63, },
+    { 32,  16,  44,  28,  35,  19,  47,  31, },
+    {  8,  56,   4,  52,  11,  59,   7,  55, },
+    { 40,  24,  36,  20,  43,  27,  39,  23, },
+    {  2,  50,  14,  62,   1,  49,  13,  61, },
+    { 34,  18,  46,  30,  33,  17,  45,  29, },
+    { 10,  58,   6,  54,   9,  57,   5,  53, },
+    { 42,  26,  38,  22,  41,  25,  37,  21, },
+};
+
+static const short custom_threshold[64] = {
+// values (296) can't be too high
+// -it causes too big quant dependence
+// or maybe overflow(check), which results in some flashing
+     71, 296, 295, 237,  71,  40,  38,  19,
+    245, 193, 185, 121, 102,  73,  53,  27,
+    158, 129, 141, 107,  97,  73,  50,  26,
+    102, 116, 109,  98,  82,  66,  45,  23,
+     71,  94,  95,  81,  70,  56,  38,  20,
+     56,  77,  74,  66,  56,  44,  30,  15,
+     38,  53,  50,  45,  38,  30,  21,  11,
+     20,  27,  26,  23,  20,  15,  11,   5
+};
+
+static inline int norm_qscale(int qscale, int type)
+{
+    switch (type) {
+    case FF_QSCALE_TYPE_MPEG1: return qscale;
+    case FF_QSCALE_TYPE_MPEG2: return qscale >> 1;
+    case FF_QSCALE_TYPE_H264:  return qscale >> 2;
+    case FF_QSCALE_TYPE_VP56:  return (63 - qscale + 2) >> 2;
+    }
+    return qscale;
+}
+
+//This func reads from 1 slice, 1 and clears 0 & 1
+static void store_slice_c(uint8_t *dst, int16_t *src,
+                          ptrdiff_t dst_stride, ptrdiff_t src_stride,
+                          ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)
+{
+    int y, x;
+#define STORE(pos)                                                             \
+    temp = (src[x + pos] + (d[pos] >> log2_scale)) >> (6 - log2_scale);        \
+    src[x + pos] = src[x + pos - 8 * src_stride] = 0;                          \
+    if (temp & 0x100) temp = ~(temp >> 31);                                    \
+    dst[x + pos] = temp;
+
+    for (y = 0; y < height; y++) {
+        const uint8_t *d = dither[y];
+        for (x = 0; x < width; x += 8) {
+            int temp;
+            STORE(0);
+            STORE(1);
+            STORE(2);
+            STORE(3);
+            STORE(4);
+            STORE(5);
+            STORE(6);
+            STORE(7);
+        }
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+//This func reads from 2 slices, 0 & 2  and clears 2-nd
+static void store_slice2_c(uint8_t *dst, int16_t *src,
+                           ptrdiff_t dst_stride, ptrdiff_t src_stride,
+                           ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)
+{
+    int y, x;
+#define STORE2(pos)                                                                                       \
+    temp = (src[x + pos] + src[x + pos + 16 * src_stride] + (d[pos] >> log2_scale)) >> (6 - log2_scale);  \
+    src[x + pos + 16 * src_stride] = 0;                                                                   \
+    if (temp & 0x100) temp = ~(temp >> 31);                                                               \
+    dst[x + pos] = temp;
+
+    for (y = 0; y < height; y++) {
+        const uint8_t *d = dither[y];
+        for (x = 0; x < width; x += 8) {
+            int temp;
+            STORE2(0);
+            STORE2(1);
+            STORE2(2);
+            STORE2(3);
+            STORE2(4);
+            STORE2(5);
+            STORE2(6);
+            STORE2(7);
+        }
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+static void mul_thrmat_c(FSPPContext *p, int q)
+{
+    int a;
+    for (a = 0; a < 64; a++)
+        ((int16_t *)p->threshold_mtx)[a] = q * ((int16_t *)p->threshold_mtx_noq)[a];//ints faster in C
+}
+
+static void filter(FSPPContext *p, uint8_t *dst, uint8_t *src,
+                   int dst_stride, int src_stride,
+                   int width, int height,
+                   uint8_t *qp_store, int qp_stride, int is_luma)
+{
+    int x, x0, y, es, qy, t;
+
+    const int stride = is_luma ? p->temp_stride : (width + 16);
+    const int step = 6 - p->log2_count;
+    const int qpsh = 4 - p->hsub * !is_luma;
+    const int qpsv = 4 - p->vsub * !is_luma;
+
+    DECLARE_ALIGNED(32, int32_t, block_align)[4 * 8 * BLOCKSZ + 4 * 8 * BLOCKSZ];
+    int16_t *block  = (int16_t *)block_align;
+    int16_t *block3 = (int16_t *)(block_align + 4 * 8 * BLOCKSZ);
+
+    memset(block3, 0, 4 * 8 * BLOCKSZ);
+
+    if (!src || !dst) return;
+
+    for (y = 0; y < height; y++) {
+        int index = 8 + 8 * stride + y * stride;
+        memcpy(p->src + index, src + y * src_stride, width);
+        for (x = 0; x < 8; x++) {
+            p->src[index         - x - 1] = p->src[index +         x    ];
+            p->src[index + width + x    ] = p->src[index + width - x - 1];
+        }
+    }
+
+    for (y = 0; y < 8; y++) {
+        memcpy(p->src + (     7 - y    ) * stride, p->src + (     y + 8    ) * stride, stride);
+        memcpy(p->src + (height + 8 + y) * stride, p->src + (height - y + 7) * stride, stride);
+    }
+    //FIXME (try edge emu)
+
+    for (y = 8; y < 24; y++)
+        memset(p->temp + 8 + y * stride, 0, width * sizeof(int16_t));
+
+    for (y = step; y < height + 8; y += step) {    //step= 1,2
+        const int y1 = y - 8 + step;                 //l5-7  l4-6;
+        qy = y - 4;
+
+        if (qy > height - 1) qy = height - 1;
+        if (qy < 0) qy = 0;
+
+        qy = (qy >> qpsv) * qp_stride;
+        p->row_fdct(block, p->src + y * stride + 2 - (y&1), stride, 2);
+
+        for (x0 = 0; x0 < width + 8 - 8 * (BLOCKSZ - 1); x0 += 8 * (BLOCKSZ - 1)) {
+            p->row_fdct(block + 8 * 8, p->src + y * stride + 8 + x0 + 2 - (y&1), stride, 2 * (BLOCKSZ - 1));
+
+            if (p->qp)
+                p->column_fidct((int16_t *)(&p->threshold_mtx[0]), block + 0 * 8, block3 + 0 * 8, 8 * (BLOCKSZ - 1)); //yes, this is a HOTSPOT
+            else
+                for (x = 0; x < 8 * (BLOCKSZ - 1); x += 8) {
+                    t = x + x0 - 2;                    //correct t=x+x0-2-(y&1), but its the same
+
+                    if (t < 0) t = 0;                   //t always < width-2
+
+                    t = qp_store[qy + (t >> qpsh)];
+                    t = norm_qscale(t, p->qscale_type);
+
+                    if (t != p->prev_q) p->prev_q = t, p->mul_thrmat(p, t);
+                    p->column_fidct((int16_t *)(&p->threshold_mtx[0]), block + x * 8, block3 + x * 8, 8); //yes, this is a HOTSPOT
+                }
+            p->row_idct(block3 + 0 * 8, p->temp + (y & 15) * stride + x0 + 2 - (y & 1), stride, 2 * (BLOCKSZ - 1));
+            memmove(block,  block  + (BLOCKSZ - 1) * 64, 8 * 8 * sizeof(int16_t)); //cycling
+            memmove(block3, block3 + (BLOCKSZ - 1) * 64, 6 * 8 * sizeof(int16_t));
+        }
+
+        es = width + 8 - x0; //  8, ...
+        if (es > 8)
+            p->row_fdct(block + 8 * 8, p->src + y * stride + 8 + x0 + 2 - (y & 1), stride, (es - 4) >> 2);
+
+        p->column_fidct((int16_t *)(&p->threshold_mtx[0]), block, block3, es&(~1));
+        p->row_idct(block3 + 0 * 8, p->temp + (y & 15) * stride + x0 + 2 - (y & 1), stride, es >> 2);
+
+        if (!(y1 & 7) && y1) {
+            if (y1 & 8)
+                p->store_slice(dst + (y1 - 8) * dst_stride, p->temp + 8 + 8 * stride,
+                               dst_stride, stride, width, 8, 5 - p->log2_count);
+            else
+                p->store_slice2(dst + (y1 - 8) * dst_stride, p->temp + 8 + 0 * stride,
+                                dst_stride, stride, width, 8, 5 - p->log2_count);
+        }
+    }
+
+    if (y & 7) {  // == height & 7
+        if (y & 8)
+            p->store_slice(dst + ((y - 8) & ~7) * dst_stride, p->temp + 8 + 8 * stride,
+                           dst_stride, stride, width, y&7, 5 - p->log2_count);
+        else
+            p->store_slice2(dst + ((y - 8) & ~7) * dst_stride, p->temp + 8 + 0 * stride,
+                            dst_stride, stride, width, y&7, 5 - p->log2_count);
+    }
+}
+
+static void column_fidct_c(int16_t *thr_adr, int16_t *data, int16_t *output, int cnt)
+{
+    int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+    int_simd16_t tmp10, tmp11, tmp12, tmp13;
+    int_simd16_t z1,z2,z3,z4,z5, z10, z11, z12, z13;
+    int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7;
+
+    int16_t *dataptr;
+    int16_t *wsptr;
+    int16_t *threshold;
+    int ctr;
+
+    dataptr = data;
+    wsptr = output;
+
+    for (; cnt > 0; cnt -= 2) { //start positions
+        threshold = (int16_t *)thr_adr;//threshold_mtx
+        for (ctr = DCTSIZE; ctr > 0; ctr--) {
+            // Process columns from input, add to output.
+            tmp0 = dataptr[DCTSIZE * 0] + dataptr[DCTSIZE * 7];
+            tmp7 = dataptr[DCTSIZE * 0] - dataptr[DCTSIZE * 7];
+
+            tmp1 = dataptr[DCTSIZE * 1] + dataptr[DCTSIZE * 6];
+            tmp6 = dataptr[DCTSIZE * 1] - dataptr[DCTSIZE * 6];
+
+            tmp2 = dataptr[DCTSIZE * 2] + dataptr[DCTSIZE * 5];
+            tmp5 = dataptr[DCTSIZE * 2] - dataptr[DCTSIZE * 5];
+
+            tmp3 = dataptr[DCTSIZE * 3] + dataptr[DCTSIZE * 4];
+            tmp4 = dataptr[DCTSIZE * 3] - dataptr[DCTSIZE * 4];
+
+            // Even part of FDCT
+
+            tmp10 = tmp0 + tmp3;
+            tmp13 = tmp0 - tmp3;
+            tmp11 = tmp1 + tmp2;
+            tmp12 = tmp1 - tmp2;
+
+            d0 = tmp10 + tmp11;
+            d4 = tmp10 - tmp11;
+
+            z1 = MULTIPLY16H((tmp12 + tmp13) << 2, FIX_0_707106781);
+            d2 = tmp13 + z1;
+            d6 = tmp13 - z1;
+
+            // Even part of IDCT
+
+            THRESHOLD(tmp0, d0, threshold[0 * 8]);
+            THRESHOLD(tmp1, d2, threshold[2 * 8]);
+            THRESHOLD(tmp2, d4, threshold[4 * 8]);
+            THRESHOLD(tmp3, d6, threshold[6 * 8]);
+            tmp0 += 2;
+            tmp10 = (tmp0 + tmp2) >> 2;
+            tmp11 = (tmp0 - tmp2) >> 2;
+
+            tmp13 = (tmp1 + tmp3) >>2; //+2 !  (psnr decides)
+            tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2
+
+            tmp0 = tmp10 + tmp13; //->temps
+            tmp3 = tmp10 - tmp13; //->temps
+            tmp1 = tmp11 + tmp12; //->temps
+            tmp2 = tmp11 - tmp12; //->temps
+
+            // Odd part of FDCT
+
+            tmp10 = tmp4 + tmp5;
+            tmp11 = tmp5 + tmp6;
+            tmp12 = tmp6 + tmp7;
+
+            z5 = MULTIPLY16H((tmp10 - tmp12) << 2, FIX_0_382683433);
+            z2 = MULTIPLY16H(tmp10 << 2, FIX_0_541196100) + z5;
+            z4 = MULTIPLY16H(tmp12 << 2, FIX_1_306562965) + z5;
+            z3 = MULTIPLY16H(tmp11 << 2, FIX_0_707106781);
+
+            z11 = tmp7 + z3;
+            z13 = tmp7 - z3;
+
+            d5 = z13 + z2;
+            d3 = z13 - z2;
+            d1 = z11 + z4;
+            d7 = z11 - z4;
+
+            // Odd part of IDCT
+
+            THRESHOLD(tmp4, d1, threshold[1 * 8]);
+            THRESHOLD(tmp5, d3, threshold[3 * 8]);
+            THRESHOLD(tmp6, d5, threshold[5 * 8]);
+            THRESHOLD(tmp7, d7, threshold[7 * 8]);
+
+            //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0
+            z13 = tmp6 + tmp5;
+            z10 = (tmp6 - tmp5) << 1;
+            z11 = tmp4 + tmp7;
+            z12 = (tmp4 - tmp7) << 1;
+
+            tmp7  = (z11 + z13) >> 2; //+2 !
+            tmp11 = MULTIPLY16H((z11 - z13) << 1, FIX_1_414213562);
+            z5    = MULTIPLY16H(z10 + z12,        FIX_1_847759065);
+            tmp10 = MULTIPLY16H(z12,              FIX_1_082392200) - z5;
+            tmp12 = MULTIPLY16H(z10,              FIX_2_613125930) + z5; // - !!
+
+            tmp6 = tmp12 - tmp7;
+            tmp5 = tmp11 - tmp6;
+            tmp4 = tmp10 + tmp5;
+
+            wsptr[DCTSIZE * 0] +=  (tmp0 + tmp7);
+            wsptr[DCTSIZE * 1] +=  (tmp1 + tmp6);
+            wsptr[DCTSIZE * 2] +=  (tmp2 + tmp5);
+            wsptr[DCTSIZE * 3] +=  (tmp3 - tmp4);
+            wsptr[DCTSIZE * 4] +=  (tmp3 + tmp4);
+            wsptr[DCTSIZE * 5] +=  (tmp2 - tmp5);
+            wsptr[DCTSIZE * 6]  =  (tmp1 - tmp6);
+            wsptr[DCTSIZE * 7]  =  (tmp0 - tmp7);
+            //
+            dataptr++; //next column
+            wsptr++;
+            threshold++;
+        }
+        dataptr += 8; //skip each second start pos
+        wsptr   += 8;
+    }
+}
+
+static void row_idct_c(int16_t *workspace, int16_t *output_adr, int output_stride, int cnt)
+{
+    int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+    int_simd16_t tmp10, tmp11, tmp12, tmp13;
+    int_simd16_t z5, z10, z11, z12, z13;
+    int16_t *outptr;
+    int16_t *wsptr;
+
+    cnt *= 4;
+    wsptr = workspace;
+    outptr = output_adr;
+    for (; cnt > 0; cnt--) {
+        // Even part
+        //Simd version reads 4x4 block and transposes it
+        tmp10 = wsptr[2] +  wsptr[3];
+        tmp11 = wsptr[2] -  wsptr[3];
+
+        tmp13 = wsptr[0] +  wsptr[1];
+        tmp12 = (MULTIPLY16H(wsptr[0] - wsptr[1], FIX_1_414213562_A) << 2) - tmp13;//this shift order to avoid overflow
+
+        tmp0 = tmp10 + tmp13; //->temps
+        tmp3 = tmp10 - tmp13; //->temps
+        tmp1 = tmp11 + tmp12;
+        tmp2 = tmp11 - tmp12;
+
+        // Odd part
+        //Also transpose, with previous:
+        // ---- ----      ||||
+        // ---- ---- idct ||||
+        // ---- ---- ---> ||||
+        // ---- ----      ||||
+        z13 = wsptr[4] + wsptr[5];
+        z10 = wsptr[4] - wsptr[5];
+        z11 = wsptr[6] + wsptr[7];
+        z12 = wsptr[6] - wsptr[7];
+
+        tmp7 = z11 + z13;
+        tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562);
+
+        z5 =    MULTIPLY16H(z10 + z12, FIX_1_847759065);
+        tmp10 = MULTIPLY16H(z12,       FIX_1_082392200) - z5;
+        tmp12 = MULTIPLY16H(z10,       FIX_2_613125930) + z5; // - FIX_
+
+        tmp6 = (tmp12 << 3) - tmp7;
+        tmp5 = (tmp11 << 3) - tmp6;
+        tmp4 = (tmp10 << 3) + tmp5;
+
+        // Final output stage: descale and write column
+        outptr[0 * output_stride] += DESCALE(tmp0 + tmp7, 3);
+        outptr[1 * output_stride] += DESCALE(tmp1 + tmp6, 3);
+        outptr[2 * output_stride] += DESCALE(tmp2 + tmp5, 3);
+        outptr[3 * output_stride] += DESCALE(tmp3 - tmp4, 3);
+        outptr[4 * output_stride] += DESCALE(tmp3 + tmp4, 3);
+        outptr[5 * output_stride] += DESCALE(tmp2 - tmp5, 3);
+        outptr[6 * output_stride] += DESCALE(tmp1 - tmp6, 3); //no += ?
+        outptr[7 * output_stride] += DESCALE(tmp0 - tmp7, 3); //no += ?
+        outptr++;
+
+        wsptr += DCTSIZE;       // advance pointer to next row
+    }
+}
+
+static void row_fdct_c(int16_t *data, const uint8_t *pixels, int line_size, int cnt)
+{
+    int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+    int_simd16_t tmp10, tmp11, tmp12, tmp13;
+    int_simd16_t z1, z2, z3, z4, z5, z11, z13;
+    int16_t *dataptr;
+
+    cnt *= 4;
+    // Pass 1: process rows.
+
+    dataptr = data;
+    for (; cnt > 0; cnt--) {
+        tmp0 = pixels[line_size * 0] + pixels[line_size * 7];
+        tmp7 = pixels[line_size * 0] - pixels[line_size * 7];
+        tmp1 = pixels[line_size * 1] + pixels[line_size * 6];
+        tmp6 = pixels[line_size * 1] - pixels[line_size * 6];
+        tmp2 = pixels[line_size * 2] + pixels[line_size * 5];
+        tmp5 = pixels[line_size * 2] - pixels[line_size * 5];
+        tmp3 = pixels[line_size * 3] + pixels[line_size * 4];
+        tmp4 = pixels[line_size * 3] - pixels[line_size * 4];
+
+        // Even part
+
+        tmp10 = tmp0 + tmp3;
+        tmp13 = tmp0 - tmp3;
+        tmp11 = tmp1 + tmp2;
+        tmp12 = tmp1 - tmp2;
+        //Even columns are written first, this leads to different order of columns
+        //in column_fidct(), but they are processed independently, so all ok.
+        //Later in the row_idct() columns readed at the same order.
+        dataptr[2] = tmp10 + tmp11;
+        dataptr[3] = tmp10 - tmp11;
+
+        z1 = MULTIPLY16H((tmp12 + tmp13) << 2, FIX_0_707106781);
+        dataptr[0] = tmp13 + z1;
+        dataptr[1] = tmp13 - z1;
+
+        // Odd part
+
+        tmp10 = (tmp4 + tmp5) << 2;
+        tmp11 = (tmp5 + tmp6) << 2;
+        tmp12 = (tmp6 + tmp7) << 2;
+
+        z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);
+        z2 = MULTIPLY16H(tmp10,         FIX_0_541196100) + z5;
+        z4 = MULTIPLY16H(tmp12,         FIX_1_306562965) + z5;
+        z3 = MULTIPLY16H(tmp11,         FIX_0_707106781);
+
+        z11 = tmp7 + z3;
+        z13 = tmp7 - z3;
+
+        dataptr[4] = z13 + z2;
+        dataptr[5] = z13 - z2;
+        dataptr[6] = z11 + z4;
+        dataptr[7] = z11 - z4;
+
+        pixels++;               // advance pointer to next column
+        dataptr += DCTSIZE;
+    }
+}
+
+static int query_formats(AVFilterContext *ctx)
+{
+    static const enum PixelFormat pix_fmts[] = {
+        AV_PIX_FMT_YUV444P,  AV_PIX_FMT_YUV422P,
+        AV_PIX_FMT_YUV420P,  AV_PIX_FMT_YUV411P,
+        AV_PIX_FMT_YUV410P,  AV_PIX_FMT_YUV440P,
+        AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ422P,
+        AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ440P,
+        AV_PIX_FMT_NONE
+    };
+    ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
+    return 0;
+}
+
+static int config_input(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx = inlink->dst;
+    FSPPContext *fspp = ctx->priv;
+    const int h = FFALIGN(inlink->h + 16, 16);
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
+
+    fspp->hsub = desc->log2_chroma_w;
+    fspp->vsub = desc->log2_chroma_h;
+
+    fspp->temp_stride = FFALIGN(inlink->w + 16, 16);
+    fspp->temp = av_malloc_array(fspp->temp_stride, h * sizeof(*fspp->temp));
+    fspp->src  = av_malloc_array(fspp->temp_stride, h * sizeof(*fspp->src));
+
+    if (!fspp->temp || !fspp->src)
+        return AVERROR(ENOMEM);
+
+    if (!fspp->use_bframe_qp && !fspp->qp) {
+        fspp->non_b_qp_alloc_size = FF_CEIL_RSHIFT(inlink->w, 4) * FF_CEIL_RSHIFT(inlink->h, 4);
+        fspp->non_b_qp_table = av_calloc(fspp->non_b_qp_alloc_size, sizeof(*fspp->non_b_qp_table));
+        if (!fspp->non_b_qp_table)
+            return AVERROR(ENOMEM);
+    }
+
+    fspp->store_slice  = store_slice_c;
+    fspp->store_slice2 = store_slice2_c;
+    fspp->mul_thrmat   = mul_thrmat_c;
+    fspp->column_fidct = column_fidct_c;
+    fspp->row_idct     = row_idct_c;
+    fspp->row_fdct     = row_fdct_c;
+
+    if (ARCH_X86)
+        ff_fspp_init_x86(fspp);
+
+    return 0;
+}
+
+static int filter_frame(AVFilterLink *inlink, AVFrame *in)
+{
+    AVFilterContext *ctx = inlink->dst;
+    FSPPContext *fspp = ctx->priv;
+    AVFilterLink *outlink = ctx->outputs[0];
+    AVFrame *out = in;
+
+    int qp_stride = 0;
+    uint8_t *qp_table = NULL;
+    int i, bias;
+    int custom_threshold_m[64];
+
+    bias = (1 << 4) + fspp->strength;
+
+    for (i = 0; i < 64; i++) //FIXME: tune custom_threshold[] and remove this !
+        custom_threshold_m[i] = (int)(custom_threshold[i] * (bias / 71.0) + 0.5);
+
+    for (i = 0; i < 8; i++) {
+        fspp->threshold_mtx_noq[2 * i] = (uint64_t)custom_threshold_m[i * 8 + 2]
+                                      |(((uint64_t)custom_threshold_m[i * 8 + 6]) << 16)
+                                      |(((uint64_t)custom_threshold_m[i * 8 + 0]) << 32)
+                                      |(((uint64_t)custom_threshold_m[i * 8 + 4]) << 48);
+
+        fspp->threshold_mtx_noq[2 * i + 1] = (uint64_t)custom_threshold_m[i * 8 + 5]
+                                          |(((uint64_t)custom_threshold_m[i * 8 + 3]) << 16)
+                                          |(((uint64_t)custom_threshold_m[i * 8 + 1]) << 32)
+                                          |(((uint64_t)custom_threshold_m[i * 8 + 7]) << 48);
+    }
+
+    if (fspp->qp)
+        fspp->prev_q = fspp->qp, fspp->mul_thrmat(fspp, fspp->qp);
+
+    /* if we are not in a constant user quantizer mode and we don't want to use
+     * the quantizers from the B-frames (B-frames often have a higher QP), we
+     * need to save the qp table from the last non B-frame; this is what the
+     * following code block does */
+    if (!fspp->qp) {
+        qp_table = av_frame_get_qp_table(in, &qp_stride, &fspp->qscale_type);
+
+        if (qp_table && !fspp->use_bframe_qp && in->pict_type != AV_PICTURE_TYPE_B) {
+            int w, h;
+
+            /* if the qp stride is not set, it means the QP are only defined on
+             * a line basis */
+           if (!qp_stride) {
+                w = FF_CEIL_RSHIFT(inlink->w, 4);
+                h = 1;
+            } else {
+                w = qp_stride;
+                h = FF_CEIL_RSHIFT(inlink->h, 4);
+            }
+            if (w * h > fspp->non_b_qp_alloc_size) {
+                int ret = av_reallocp_array(&fspp->non_b_qp_table, w, h);
+                if (ret < 0) {
+                    fspp->non_b_qp_alloc_size = 0;
+                    return ret;
+                }
+                fspp->non_b_qp_alloc_size = w * h;
+            }
+
+            av_assert0(w * h <= fspp->non_b_qp_alloc_size);
+            memcpy(fspp->non_b_qp_table, qp_table, w * h);
+        }
+    }
+
+    if (fspp->log2_count && !ctx->is_disabled) {
+        if (!fspp->use_bframe_qp && fspp->non_b_qp_table)
+            qp_table = fspp->non_b_qp_table;
+
+        if (qp_table || fspp->qp) {
+            const int cw = FF_CEIL_RSHIFT(inlink->w, fspp->hsub);
+            const int ch = FF_CEIL_RSHIFT(inlink->h, fspp->vsub);
+
+            /* get a new frame if in-place is not possible or if the dimensions
+             * are not multiple of 8 */
+            if (!av_frame_is_writable(in) || (inlink->w & 7) || (inlink->h & 7)) {
+                const int aligned_w = FFALIGN(inlink->w, 8);
+                const int aligned_h = FFALIGN(inlink->h, 8);
+
+                out = ff_get_video_buffer(outlink, aligned_w, aligned_h);
+                if (!out) {
+                    av_frame_free(&in);
+                    return AVERROR(ENOMEM);
+                }
+                av_frame_copy_props(out, in);
+            }
+
+            filter(fspp, out->data[0], in->data[0], out->linesize[0], in->linesize[0],
+                   inlink->w, inlink->h, qp_table, qp_stride, 1);
+            filter(fspp, out->data[1], in->data[1], out->linesize[1], in->linesize[1],
+                   cw,        ch,        qp_table, qp_stride, 0);
+            filter(fspp, out->data[2], in->data[2], out->linesize[2], in->linesize[2],
+                   cw,        ch,        qp_table, qp_stride, 0);
+            emms_c();
+        }
+    }
+
+    if (in != out) {
+        if (in->data[3])
+            av_image_copy_plane(out->data[3], out->linesize[3],
+                                in ->data[3], in ->linesize[3],
+                                inlink->w, inlink->h);
+        av_frame_free(&in);
+    }
+    return ff_filter_frame(outlink, out);
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    FSPPContext *fspp = ctx->priv;
+    av_freep(&fspp->temp);
+    av_freep(&fspp->src);
+    av_freep(&fspp->non_b_qp_table);
+}
+
+static const AVFilterPad fspp_inputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .config_props = config_input,
+        .filter_frame = filter_frame,
+    },
+    { NULL }
+};
+
+static const AVFilterPad fspp_outputs[] = {
+    {
+        .name = "default",
+        .type = AVMEDIA_TYPE_VIDEO,
+    },
+    { NULL }
+};
+
+AVFilter ff_vf_fspp = {
+    .name            = "fspp",
+    .description     = NULL_IF_CONFIG_SMALL("Apply Fast Simple Post-processing filter."),
+    .priv_size       = sizeof(FSPPContext),
+    .uninit          = uninit,
+    .query_formats   = query_formats,
+    .inputs          = fspp_inputs,
+    .outputs         = fspp_outputs,
+    .priv_class      = &fspp_class,
+    .flags           = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL,
+};
diff --git a/libavfilter/vf_fspp.h b/libavfilter/vf_fspp.h
new file mode 100644 (file)
index 0000000..db860c6
--- /dev/null
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
+ * Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>
+ * Copyright (c) 2014 Arwa Arif <arwaarif1994@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU 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 AVFILTER_FSPP_H
+#define AVFILTER_FSPP_H
+
+#include "avfilter.h"
+
+#define BLOCKSZ 12
+#define MAX_LEVEL 5
+
+#define DCTSIZE 8
+#define DCTSIZE_S "8"
+
+#define FIX(x,s)  ((int) ((x) * (1 << s) + 0.5) & 0xffff)
+#define C64(x)    ((uint64_t)((x) | (x) << 16)) <<32 | (uint64_t)(x) | (uint64_t)(x) << 16
+#define FIX64(x,s)  C64(FIX(x,s))
+
+#define MULTIPLY16H(x,k)   (((x) * (k)) >> 16)
+#define THRESHOLD(r,x,t)                         \
+    if(((unsigned)((x) + t)) > t * 2) r = (x);   \
+    else r = 0;
+#define DESCALE(x,n)  (((x) + (1 << ((n) - 1))) >> n)
+
+typedef int32_t int_simd16_t;
+static const int16_t FIX_0_382683433   = FIX(0.382683433, 14);
+static const int16_t FIX_0_541196100   = FIX(0.541196100, 14);
+static const int16_t FIX_0_707106781   = FIX(0.707106781, 14);
+static const int16_t FIX_1_306562965   = FIX(1.306562965, 14);
+static const int16_t FIX_1_414213562_A = FIX(1.414213562, 14);
+static const int16_t FIX_1_847759065   = FIX(1.847759065, 13);
+static const int16_t FIX_2_613125930   = FIX(-2.613125930, 13);
+static const int16_t FIX_1_414213562   = FIX(1.414213562, 13);
+static const int16_t FIX_1_082392200   = FIX(1.082392200, 13);
+
+typedef struct FSPPContext {
+    AVClass *class;
+    uint64_t threshold_mtx_noq[8 * 2];
+    uint64_t threshold_mtx[8 * 2];        //used in both C & MMX (& later SSE2) versions
+
+    int log2_count;
+    int strength;
+    int hsub;
+    int vsub;
+    int temp_stride;
+    int qp;
+    int qscale_type;
+    int prev_q;
+    uint8_t *src;
+    int16_t *temp;
+    uint8_t *non_b_qp_table;
+    int non_b_qp_alloc_size;
+    int use_bframe_qp;
+
+    void (*store_slice)(uint8_t *dst, int16_t *src,
+                        ptrdiff_t dst_stride, ptrdiff_t src_stride,
+                        ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale);
+
+    void (*store_slice2)(uint8_t *dst, int16_t *src,
+                         ptrdiff_t dst_stride, ptrdiff_t src_stride,
+                         ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale);
+
+    void (*mul_thrmat)(struct FSPPContext *fspp, int q);
+
+    void (*column_fidct)(int16_t *thr_adr, int16_t *data,
+                         int16_t *output, int cnt);
+
+    void (*row_idct)(int16_t *workspace, int16_t *output_adr,
+                     int output_stride, int cnt);
+
+    void (*row_fdct)(int16_t *data, const uint8_t *pixels,
+                     int line_size, int cnt);
+
+} FSPPContext;
+
+void ff_fspp_init_x86(FSPPContext *fspp);
+
+#endif /* AVFILTER_FSPP_H */
index 44765d2..4f9c83d 100644 (file)
@@ -1,3 +1,4 @@
+OBJS-$(CONFIG_FSPP_FILTER)                   += x86/vf_fspp.o
 OBJS-$(CONFIG_GRADFUN_FILTER)                += x86/vf_gradfun_init.o
 OBJS-$(CONFIG_HQDN3D_FILTER)                 += x86/vf_hqdn3d_init.o
 OBJS-$(CONFIG_IDET_FILTER)                   += x86/vf_idet_init.o
diff --git a/libavfilter/x86/vf_fspp.c b/libavfilter/x86/vf_fspp.c
new file mode 100644 (file)
index 0000000..14c1355
--- /dev/null
@@ -0,0 +1,1405 @@
+/*
+ * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
+ * Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU 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/attributes.h"
+#include "libavutil/cpu.h"
+#include "libavutil/mem.h"
+#include "libavutil/x86/asm.h"
+#include "libavfilter/vf_fspp.h"
+
+DECLARE_ALIGNED(32, static const uint8_t, dither)[8][8] = {
+    {  0,  48,  12,  60,   3,  51,  15,  63, },
+    { 32,  16,  44,  28,  35,  19,  47,  31, },
+    {  8,  56,   4,  52,  11,  59,   7,  55, },
+    { 40,  24,  36,  20,  43,  27,  39,  23, },
+    {  2,  50,  14,  62,   1,  49,  13,  61, },
+    { 34,  18,  46,  30,  33,  17,  45,  29, },
+    { 10,  58,   6,  54,   9,  57,   5,  53, },
+    { 42,  26,  38,  22,  41,  25,  37,  21, },
+};
+
+//This func reads from 1 slice, 1 and clears 0 & 1
+static void store_slice_mmx(uint8_t *dst, int16_t *src,
+                            ptrdiff_t dst_stride, ptrdiff_t src_stride,
+                            ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)
+{
+    const uint8_t *od = &dither[0][0];
+    const uint8_t *end = &dither[height][0];
+    width = (width + 7) & ~7;
+    dst_stride -= width;
+
+    __asm__ volatile(
+        "mov %5 , %%"REG_d"                \n\t"
+        "mov %6 , %%"REG_S"                \n\t"
+        "mov %7 , %%"REG_D"                \n\t"
+        "mov %1 , %%"REG_a"                \n\t"
+        "movd %%"REG_d" , %%mm5            \n\t"
+        "xor $-1 , %%"REG_d"               \n\t"
+        "mov %%"REG_a" , %%"REG_c"         \n\t"
+        "add $7 , %%"REG_d"                \n\t"
+        "neg %%"REG_a"                     \n\t"
+        "sub %0 , %%"REG_c"                \n\t"
+        "add %%"REG_c" , %%"REG_c"         \n\t"
+        "movd %%"REG_d" , %%mm2            \n\t"
+        "mov %%"REG_c" , %1                \n\t"
+        "mov %2 , %%"REG_d"                \n\t"
+        "shl $4 , %%"REG_a"                \n\t"
+
+        "2:                                \n\t"
+        "movq (%%"REG_d") , %%mm3          \n\t"
+        "movq %%mm3 , %%mm4                \n\t"
+        "pxor %%mm7 , %%mm7                \n\t"
+        "punpcklbw %%mm7 , %%mm3           \n\t"
+        "punpckhbw %%mm7 , %%mm4           \n\t"
+        "mov %0 , %%"REG_c"                \n\t"
+        "psraw %%mm5 , %%mm3               \n\t"
+        "psraw %%mm5 , %%mm4               \n\t"
+        "1:                                \n\t"
+        "movq %%mm7, (%%"REG_S",%%"REG_a") \n\t"
+        "movq (%%"REG_S") , %%mm0          \n\t"
+        "movq 8(%%"REG_S"), %%mm1          \n\t"
+
+        "movq %%mm7, 8(%%"REG_S",%%"REG_a")\n\t"
+        "paddw %%mm3, %%mm0                \n\t"
+        "paddw %%mm4, %%mm1                \n\t"
+
+        "movq %%mm7, (%%"REG_S")           \n\t"
+        "psraw %%mm2, %%mm0                \n\t"
+        "psraw %%mm2, %%mm1                \n\t"
+
+        "movq %%mm7, 8(%%"REG_S")          \n\t"
+        "packuswb %%mm1, %%mm0             \n\t"
+        "add $16, %%"REG_S"                \n\t"
+
+        "movq %%mm0, (%%"REG_D")           \n\t"
+        "add $8, %%"REG_D"                 \n\t"
+        "sub $8, %%"REG_c"                 \n\t"
+        "jg 1b                             \n\t"
+        "add %1, %%"REG_S"                 \n\t"
+        "add $8, %%"REG_d"                 \n\t"
+        "add %3, %%"REG_D"                 \n\t"
+        "cmp %4, %%"REG_d"                 \n\t"
+        "jl 2b                             \n\t"
+
+        :
+        : "m" (width),      "m" (src_stride), "erm" (od), "m" (dst_stride), "erm" (end),
+          "m" (log2_scale), "m" (src),        "m" (dst)                                     //input
+        : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
+        );
+}
+
+//This func reads from 2 slices, 0 & 2  and clears 2-nd
+static void store_slice2_mmx(uint8_t *dst, int16_t *src,
+                             ptrdiff_t dst_stride, ptrdiff_t src_stride,
+                             ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)
+{
+    const uint8_t *od = &dither[0][0];
+    const uint8_t *end = &dither[height][0];
+    width = (width + 7) & ~7;
+    dst_stride -= width;
+
+    __asm__ volatile(
+        "mov %5, %%"REG_d"                \n\t"
+        "mov %6, %%"REG_S"                \n\t"
+        "mov %7, %%"REG_D"                \n\t"
+        "mov %1, %%"REG_a"                \n\t"
+        "movd %%"REG_d", %%mm5            \n\t"
+        "xor $-1, %%"REG_d"               \n\t"
+        "mov %%"REG_a", %%"REG_c"         \n\t"
+        "add $7, %%"REG_d"                \n\t"
+        "sub %0, %%"REG_c"                \n\t"
+        "add %%"REG_c", %%"REG_c"         \n\t"
+        "movd %%"REG_d", %%mm2            \n\t"
+        "mov %%"REG_c", %1                \n\t"
+        "mov %2, %%"REG_d"                \n\t"
+        "shl $5, %%"REG_a"                \n\t"
+
+        "2:                               \n\t"
+        "movq (%%"REG_d"), %%mm3          \n\t"
+        "movq %%mm3, %%mm4                \n\t"
+        "pxor %%mm7, %%mm7                \n\t"
+        "punpcklbw %%mm7, %%mm3           \n\t"
+        "punpckhbw %%mm7, %%mm4           \n\t"
+        "mov %0, %%"REG_c"                \n\t"
+        "psraw %%mm5, %%mm3               \n\t"
+        "psraw %%mm5, %%mm4               \n\t"
+        "1:                               \n\t"
+        "movq (%%"REG_S"), %%mm0          \n\t"
+        "movq 8(%%"REG_S"), %%mm1         \n\t"
+        "paddw %%mm3, %%mm0               \n\t"
+
+        "paddw (%%"REG_S",%%"REG_a"),%%mm0\n\t"
+        "paddw %%mm4, %%mm1               \n\t"
+        "movq 8(%%"REG_S",%%"REG_a"),%%mm6\n\t"
+
+        "movq %%mm7, (%%"REG_S",%%"REG_a")\n\t"
+        "psraw %%mm2, %%mm0               \n\t"
+        "paddw %%mm6, %%mm1               \n\t"
+
+        "movq %%mm7,8(%%"REG_S",%%"REG_a")\n\t"
+        "psraw %%mm2, %%mm1               \n\t"
+        "packuswb %%mm1, %%mm0            \n\t"
+
+        "movq %%mm0, (%%"REG_D")          \n\t"
+        "add $16, %%"REG_S"               \n\t"
+        "add $8, %%"REG_D"                \n\t"
+        "sub $8, %%"REG_c"                \n\t"
+        "jg 1b                            \n\t"
+        "add %1, %%"REG_S"                \n\t"
+        "add $8, %%"REG_d"                \n\t"
+        "add %3, %%"REG_D"                \n\t"
+        "cmp %4, %%"REG_d"                \n\t"
+        "jl 2b                            \n\t"
+
+        :
+        : "m" (width),      "m" (src_stride), "erm" (od), "m" (dst_stride), "erm" (end),
+          "m" (log2_scale), "m" (src),        "m" (dst)                                     //input
+        : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_D, "%"REG_S
+        );
+}
+
+static void mul_thrmat_mmx(FSPPContext *p, int q)
+{
+    uint64_t *adr = &p->threshold_mtx_noq[0];
+
+    __asm__ volatile(
+        "movd %0, %%mm7                   \n\t"
+        "add $8*8*2, %%"REG_D"            \n\t"
+        "movq 0*8(%%"REG_S"), %%mm0       \n\t"
+        "punpcklwd %%mm7, %%mm7           \n\t"
+        "movq 1*8(%%"REG_S"), %%mm1       \n\t"
+        "punpckldq %%mm7, %%mm7           \n\t"
+        "pmullw %%mm7, %%mm0              \n\t"
+
+        "movq 2*8(%%"REG_S"), %%mm2       \n\t"
+        "pmullw %%mm7, %%mm1              \n\t"
+
+        "movq 3*8(%%"REG_S"), %%mm3       \n\t"
+        "pmullw %%mm7, %%mm2              \n\t"
+
+        "movq %%mm0, 0*8(%%"REG_D")       \n\t"
+        "movq 4*8(%%"REG_S"), %%mm4       \n\t"
+        "pmullw %%mm7, %%mm3              \n\t"
+
+        "movq %%mm1, 1*8(%%"REG_D")       \n\t"
+        "movq 5*8(%%"REG_S"), %%mm5       \n\t"
+        "pmullw %%mm7, %%mm4              \n\t"
+
+        "movq %%mm2, 2*8(%%"REG_D")       \n\t"
+        "movq 6*8(%%"REG_S"), %%mm6       \n\t"
+        "pmullw %%mm7, %%mm5              \n\t"
+
+        "movq %%mm3, 3*8(%%"REG_D")       \n\t"
+        "movq 7*8+0*8(%%"REG_S"), %%mm0   \n\t"
+        "pmullw %%mm7, %%mm6              \n\t"
+
+        "movq %%mm4, 4*8(%%"REG_D")       \n\t"
+        "movq 7*8+1*8(%%"REG_S"), %%mm1   \n\t"
+        "pmullw %%mm7, %%mm0              \n\t"
+
+        "movq %%mm5, 5*8(%%"REG_D")       \n\t"
+        "movq 7*8+2*8(%%"REG_S"), %%mm2   \n\t"
+        "pmullw %%mm7, %%mm1              \n\t"
+
+        "movq %%mm6, 6*8(%%"REG_D")       \n\t"
+        "movq 7*8+3*8(%%"REG_S"), %%mm3   \n\t"
+        "pmullw %%mm7, %%mm2              \n\t"
+
+        "movq %%mm0, 7*8+0*8(%%"REG_D")   \n\t"
+        "movq 7*8+4*8(%%"REG_S"), %%mm4   \n\t"
+        "pmullw %%mm7, %%mm3              \n\t"
+
+        "movq %%mm1, 7*8+1*8(%%"REG_D")   \n\t"
+        "movq 7*8+5*8(%%"REG_S"), %%mm5   \n\t"
+        "pmullw %%mm7, %%mm4              \n\t"
+
+        "movq %%mm2, 7*8+2*8(%%"REG_D")   \n\t"
+        "movq 7*8+6*8(%%"REG_S"), %%mm6   \n\t"
+        "pmullw %%mm7, %%mm5              \n\t"
+
+        "movq %%mm3, 7*8+3*8(%%"REG_D")   \n\t"
+        "movq 14*8+0*8(%%"REG_S"), %%mm0  \n\t"
+        "pmullw %%mm7, %%mm6              \n\t"
+
+        "movq %%mm4, 7*8+4*8(%%"REG_D")   \n\t"
+        "movq 14*8+1*8(%%"REG_S"), %%mm1  \n\t"
+        "pmullw %%mm7, %%mm0              \n\t"
+
+        "movq %%mm5, 7*8+5*8(%%"REG_D")   \n\t"
+        "pmullw %%mm7, %%mm1              \n\t"
+
+        "movq %%mm6, 7*8+6*8(%%"REG_D")   \n\t"
+        "movq %%mm0, 14*8+0*8(%%"REG_D")  \n\t"
+        "movq %%mm1, 14*8+1*8(%%"REG_D")  \n\t"
+
+        : "+g" (q), "+S" (adr), "+D" (adr)
+        :
+        );
+}
+
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_382683433)   = FIX64(0.382683433, 14);
+DECLARE_ALIGNED  (8, uint64_t, ff_MM_FIX_0_541196100)= FIX64(0.541196100, 14);
+DECLARE_ALIGNED  (8, uint64_t, ff_MM_FIX_0_707106781)= FIX64(0.707106781, 14);
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_306562965)   = FIX64(1.306562965, 14);
+
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562_A) = FIX64(1.414213562, 14);
+
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_847759065)   = FIX64(1.847759065, 13);
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_2_613125930)   = FIX64(-2.613125930, 13);
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562)   = FIX64(1.414213562, 13);
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_082392200)   = FIX64(1.082392200, 13);
+//for t3,t5,t7 == 0 shortcut
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_847759065)   = FIX64(0.847759065, 14);
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_566454497)   = FIX64(0.566454497, 14);
+DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_198912367)   = FIX64(0.198912367, 14);
+
+DECLARE_ASM_CONST(8, uint64_t, MM_DESCALE_RND)       = C64(4);
+DECLARE_ASM_CONST(8, uint64_t, MM_2)                 = C64(2);
+
+static void column_fidct_mmx(int16_t *thr_adr, int16_t *data, int16_t *output, int cnt)
+{
+    DECLARE_ALIGNED(8, uint64_t, temps)[4];
+
+    __asm__ volatile(
+
+        "1:                                       \n\t"
+        "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1   \n\t"
+        //
+        "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7   \n\t"
+        "movq %%mm1, %%mm0                        \n\t"
+
+        "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1  \n\t" //t0
+        "movq %%mm7, %%mm3                        \n\t"
+
+        "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7  \n\t" //t3
+        "movq %%mm1, %%mm5             \n\t"
+
+        "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6   \n\t"
+        "psubw %%mm7, %%mm1                       \n\t" //t13
+
+        "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2   \n\t"
+        "movq %%mm6, %%mm4                        \n\t"
+
+        "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6  \n\t" //t1
+        "paddw %%mm7, %%mm5                       \n\t" //t10
+
+        "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2  \n\t" //t2
+        "movq %%mm6, %%mm7                        \n\t"
+
+        "paddw %%mm2, %%mm6                       \n\t" //t11
+        "psubw %%mm2, %%mm7                       \n\t" //t12
+
+        "movq %%mm5, %%mm2                        \n\t"
+        "paddw %%mm6, %%mm5                       \n\t" //d0
+        // i0 t13 t12 i3 i1 d0 - d4
+        "psubw %%mm6, %%mm2                       \n\t" //d4
+        "paddw %%mm1, %%mm7                       \n\t"
+
+        "movq  4*16(%%"REG_d"), %%mm6             \n\t"
+        "psllw $2, %%mm7                          \n\t"
+
+        "psubw 0*16(%%"REG_d"), %%mm5             \n\t"
+        "psubw %%mm6, %%mm2                       \n\t"
+
+        "paddusw 0*16(%%"REG_d"), %%mm5           \n\t"
+        "paddusw %%mm6, %%mm2                     \n\t"
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_707106781)", %%mm7 \n\t"
+        //
+        "paddw 0*16(%%"REG_d"), %%mm5             \n\t"
+        "paddw %%mm6, %%mm2                       \n\t"
+
+        "psubusw 0*16(%%"REG_d"), %%mm5           \n\t"
+        "psubusw %%mm6, %%mm2                     \n\t"
+
+//This func is totally compute-bound,  operates at huge speed. So,  DC shortcut
+// at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
+//However,  typical numbers: nondc - 29%%,  dc - 46%%,  zero - 25%%. All <> 0 case is very rare.
+        "paddw "MANGLE(MM_2)", %%mm5              \n\t"
+        "movq %%mm2, %%mm6                        \n\t"
+
+        "paddw %%mm5, %%mm2                       \n\t"
+        "psubw %%mm6, %%mm5                       \n\t"
+
+        "movq %%mm1, %%mm6                        \n\t"
+        "paddw %%mm7, %%mm1                       \n\t" //d2
+
+        "psubw 2*16(%%"REG_d"), %%mm1             \n\t"
+        "psubw %%mm7, %%mm6                       \n\t" //d6
+
+        "movq 6*16(%%"REG_d"), %%mm7              \n\t"
+        "psraw $2, %%mm5                          \n\t"
+
+        "paddusw 2*16(%%"REG_d"), %%mm1           \n\t"
+        "psubw %%mm7, %%mm6                       \n\t"
+        // t7 d2 /t11 t4 t6 - d6 /t10
+
+        "paddw 2*16(%%"REG_d"), %%mm1             \n\t"
+        "paddusw %%mm7, %%mm6                     \n\t"
+
+        "psubusw 2*16(%%"REG_d"), %%mm1           \n\t"
+        "paddw %%mm7, %%mm6                       \n\t"
+
+        "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3  \n\t"
+        "psubusw %%mm7, %%mm6                     \n\t"
+
+        //movq [edi+"DCTSIZE_S"*2*2], mm1
+        //movq [edi+"DCTSIZE_S"*6*2], mm6
+        "movq %%mm1, %%mm7                        \n\t"
+        "psraw $2, %%mm2                          \n\t"
+
+        "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4  \n\t"
+        "psubw %%mm6, %%mm1                       \n\t"
+
+        "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0  \n\t"
+        "paddw %%mm7, %%mm6                       \n\t" //'t13
+
+        "psraw $2, %%mm6                          \n\t" //paddw mm6, MM_2 !!    ---
+        "movq %%mm2, %%mm7                        \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
+        "paddw %%mm6, %%mm2                       \n\t" //'t0
+
+        "movq %%mm2, 0*8+%3                       \n\t" //!
+        "psubw %%mm6, %%mm7                       \n\t" //'t3
+
+        "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2   \n\t"
+        "psubw %%mm6, %%mm1                       \n\t" //'t12
+
+        "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2  \n\t" //t5
+        "movq %%mm5, %%mm6                        \n\t"
+
+        "movq %%mm7, 3*8+%3                       \n\t"
+        "paddw %%mm2, %%mm3                       \n\t" //t10
+
+        "paddw %%mm4, %%mm2                       \n\t" //t11
+        "paddw %%mm0, %%mm4                       \n\t" //t12
+
+        "movq %%mm3, %%mm7                        \n\t"
+        "psubw %%mm4, %%mm3                       \n\t"
+
+        "psllw $2, %%mm3                          \n\t"
+        "psllw $2, %%mm7                          \n\t" //opt for P6
+
+        "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
+        "psllw $2, %%mm4                          \n\t"
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_541196100)", %%mm7 \n\t"
+        "psllw $2, %%mm2                          \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
+        "paddw %%mm1, %%mm5                       \n\t" //'t1
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_707106781)", %%mm2 \n\t"
+        "psubw %%mm1, %%mm6                       \n\t" //'t2
+        // t7 't12 't11 t4 t6 - 't13 't10   ---
+
+        "paddw %%mm3, %%mm7                       \n\t" //z2
+
+        "movq %%mm5, 1*8+%3                       \n\t"
+        "paddw %%mm3, %%mm4                       \n\t" //z4
+
+        "movq 3*16(%%"REG_d"), %%mm3              \n\t"
+        "movq %%mm0, %%mm1                        \n\t"
+
+        "movq %%mm6, 2*8+%3                       \n\t"
+        "psubw %%mm2, %%mm1                       \n\t" //z13
+
+//===
+        "paddw %%mm2, %%mm0                       \n\t" //z11
+        "movq %%mm1, %%mm5                        \n\t"
+
+        "movq 5*16(%%"REG_d"), %%mm2              \n\t"
+        "psubw %%mm7, %%mm1                       \n\t" //d3
+
+        "paddw %%mm7, %%mm5                       \n\t" //d5
+        "psubw %%mm3, %%mm1                       \n\t"
+
+        "movq 1*16(%%"REG_d"), %%mm7              \n\t"
+        "psubw %%mm2, %%mm5                       \n\t"
+
+        "movq %%mm0, %%mm6                        \n\t"
+        "paddw %%mm4, %%mm0                       \n\t" //d1
+
+        "paddusw %%mm3, %%mm1                     \n\t"
+        "psubw %%mm4, %%mm6                       \n\t" //d7
+
+        // d1 d3 - - - d5 d7 -
+        "movq 7*16(%%"REG_d"), %%mm4              \n\t"
+        "psubw %%mm7, %%mm0                       \n\t"
+
+        "psubw %%mm4, %%mm6                       \n\t"
+        "paddusw %%mm2, %%mm5                     \n\t"
+
+        "paddusw %%mm4, %%mm6                     \n\t"
+        "paddw %%mm3, %%mm1                       \n\t"
+
+        "paddw %%mm2, %%mm5                       \n\t"
+        "paddw %%mm4, %%mm6                       \n\t"
+
+        "psubusw %%mm3, %%mm1                     \n\t"
+        "psubusw %%mm2, %%mm5                     \n\t"
+
+        "psubusw %%mm4, %%mm6                     \n\t"
+        "movq %%mm1, %%mm4                        \n\t"
+
+        "por %%mm5, %%mm4                         \n\t"
+        "paddusw %%mm7, %%mm0                     \n\t"
+
+        "por %%mm6, %%mm4                         \n\t"
+        "paddw %%mm7, %%mm0                       \n\t"
+
+        "packssdw %%mm4, %%mm4                    \n\t"
+        "psubusw %%mm7, %%mm0                     \n\t"
+
+        "movd %%mm4, %%"REG_a"                    \n\t"
+        "or %%"REG_a", %%"REG_a"                  \n\t"
+        "jnz 2f                                   \n\t"
+        //movq [edi+"DCTSIZE_S"*3*2], mm1
+        //movq [edi+"DCTSIZE_S"*5*2], mm5
+        //movq [edi+"DCTSIZE_S"*1*2], mm0
+        //movq [edi+"DCTSIZE_S"*7*2], mm6
+        // t4 t5 - - - t6 t7 -
+        //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
+//Typical numbers: nondc - 19%%,  dc - 26%%,  zero - 55%%. zero case alone isn't worthwhile
+        "movq 0*8+%3, %%mm4                      \n\t"
+        "movq %%mm0, %%mm1                       \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
+        "movq %%mm1, %%mm2                       \n\t"
+
+        "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
+        "movq %%mm2, %%mm3                      \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
+        "paddw %%mm4, %%mm5                     \n\t"
+
+        "movq 1*8+%3, %%mm6                     \n\t"
+        //paddw mm3, MM_2
+        "psraw $2, %%mm3                        \n\t" //tmp7
+
+        "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
+        "psubw %%mm3, %%mm4                     \n\t"
+
+        "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
+        "paddw %%mm3, %%mm5                     \n\t"
+
+        "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
+        "paddw %%mm6, %%mm7                     \n\t"
+
+        "movq 2*8+%3, %%mm3                     \n\t"
+        "psubw %%mm0, %%mm6                     \n\t"
+
+        "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
+        "paddw %%mm0, %%mm7                     \n\t"
+
+        "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
+        "paddw %%mm3, %%mm4                     \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
+        "psubw %%mm1, %%mm3                     \n\t"
+
+        "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
+        "paddw %%mm1, %%mm4                     \n\t"
+
+        "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
+        "paddw %%mm3, %%mm5                     \n\t"
+
+        "movq 3*8+%3, %%mm0                     \n\t"
+        "add $8, %%"REG_S"                      \n\t"
+
+        "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
+        "paddw %%mm0, %%mm6                     \n\t"
+
+        "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
+        "psubw %%mm2, %%mm0                     \n\t"
+
+        "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
+        "paddw %%mm2, %%mm6                     \n\t"
+
+        "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
+        "paddw %%mm0, %%mm7                     \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
+
+        "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
+        "add $8, %%"REG_D"                      \n\t"
+        "jmp 4f                                 \n\t"
+
+        "2:                                     \n\t"
+        //--- non DC2
+        //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1  (actually thr1, thr1, thr1-1)
+        //psraw mm5, 2
+        //psraw mm0, 2
+        //psraw mm6, 2
+        "movq %%mm5, %%mm3                      \n\t"
+        "psubw %%mm1, %%mm5                     \n\t"
+
+        "psllw $1, %%mm5                        \n\t" //'z10
+        "paddw %%mm1, %%mm3                     \n\t" //'z13
+
+        "movq %%mm0, %%mm2                      \n\t"
+        "psubw %%mm6, %%mm0                     \n\t"
+
+        "movq %%mm5, %%mm1                      \n\t"
+        "psllw $1, %%mm0                        \n\t" //'z12
+
+        "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
+        "paddw %%mm0, %%mm5                     \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
+        "paddw %%mm6, %%mm2                     \n\t" //'z11
+
+        "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
+        "movq %%mm2, %%mm7                      \n\t"
+
+        //---
+        "movq 0*8+%3, %%mm4                     \n\t"
+        "psubw %%mm3, %%mm2                     \n\t"
+
+        "psllw $1, %%mm2                        \n\t"
+        "paddw %%mm3, %%mm7                     \n\t" //'t7
+
+        "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
+        "movq %%mm4, %%mm6                      \n\t"
+        //paddw mm7, MM_2
+        "psraw $2, %%mm7                        \n\t"
+
+        "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4\n\t"
+        "psubw %%mm7, %%mm6                     \n\t"
+
+        "movq 1*8+%3, %%mm3                     \n\t"
+        "paddw %%mm7, %%mm4                     \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
+        "paddw %%mm5, %%mm1                     \n\t" //'t12
+
+        "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
+        "psubw %%mm7, %%mm1                     \n\t" //'t6
+
+        "movq 2*8+%3, %%mm7                     \n\t"
+        "psubw %%mm5, %%mm0                     \n\t" //'t10
+
+        "movq 3*8+%3, %%mm6                     \n\t"
+        "movq %%mm3, %%mm5                      \n\t"
+
+        "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3\n\t"
+        "psubw %%mm1, %%mm5                     \n\t"
+
+        "psubw %%mm1, %%mm2                     \n\t" //'t5
+        "paddw %%mm1, %%mm3                     \n\t"
+
+        "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
+        "movq %%mm7, %%mm4                      \n\t"
+
+        "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7\n\t"
+        "psubw %%mm2, %%mm4                     \n\t"
+
+        "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4\n\t"
+        "paddw %%mm2, %%mm7                     \n\t"
+
+        "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
+        "paddw %%mm2, %%mm0                     \n\t" //'t4
+
+        // 't4 't6 't5 - - - - 't7
+        "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
+        "movq %%mm6, %%mm1                      \n\t"
+
+        "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6\n\t"
+        "psubw %%mm0, %%mm1                     \n\t"
+
+        "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1\n\t"
+        "paddw %%mm0, %%mm6                     \n\t"
+
+        "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
+        "add $8, %%"REG_S"                      \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
+
+        "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
+        "add $8, %%"REG_D"                      \n\t"
+
+        "4:                                     \n\t"
+        "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
+        //
+        "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
+        "movq %%mm1, %%mm0                      \n\t"
+
+        "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1\n\t" //t0
+        "movq %%mm7, %%mm3                      \n\t"
+
+        "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7\n\t" //t3
+        "movq %%mm1, %%mm5                      \n\t"
+
+        "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
+        "psubw %%mm7, %%mm1                     \n\t" //t13
+
+        "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
+        "movq %%mm6, %%mm4                      \n\t"
+
+        "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6\n\t" //t1
+        "paddw %%mm7, %%mm5                     \n\t" //t10
+
+        "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2\n\t" //t2
+        "movq %%mm6, %%mm7                      \n\t"
+
+        "paddw %%mm2, %%mm6                     \n\t" //t11
+        "psubw %%mm2, %%mm7                     \n\t" //t12
+
+        "movq %%mm5, %%mm2                      \n\t"
+        "paddw %%mm6, %%mm5                     \n\t" //d0
+        // i0 t13 t12 i3 i1 d0 - d4
+        "psubw %%mm6, %%mm2                     \n\t" //d4
+        "paddw %%mm1, %%mm7                     \n\t"
+
+        "movq  1*8+4*16(%%"REG_d"), %%mm6       \n\t"
+        "psllw $2, %%mm7                        \n\t"
+
+        "psubw 1*8+0*16(%%"REG_d"), %%mm5       \n\t"
+        "psubw %%mm6, %%mm2                     \n\t"
+
+        "paddusw 1*8+0*16(%%"REG_d"), %%mm5     \n\t"
+        "paddusw %%mm6, %%mm2                   \n\t"
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_707106781)", %%mm7 \n\t"
+        //
+        "paddw 1*8+0*16(%%"REG_d"), %%mm5       \n\t"
+        "paddw %%mm6, %%mm2                     \n\t"
+
+        "psubusw 1*8+0*16(%%"REG_d"), %%mm5     \n\t"
+        "psubusw %%mm6, %%mm2                   \n\t"
+
+//This func is totally compute-bound,  operates at huge speed. So,  DC shortcut
+// at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
+//However,  typical numbers: nondc - 29%%,  dc - 46%%,  zero - 25%%. All <> 0 case is very rare.
+        "paddw "MANGLE(MM_2)", %%mm5            \n\t"
+        "movq %%mm2, %%mm6                      \n\t"
+
+        "paddw %%mm5, %%mm2                     \n\t"
+        "psubw %%mm6, %%mm5                     \n\t"
+
+        "movq %%mm1, %%mm6                      \n\t"
+        "paddw %%mm7, %%mm1                     \n\t" //d2
+
+        "psubw 1*8+2*16(%%"REG_d"), %%mm1       \n\t"
+        "psubw %%mm7, %%mm6                     \n\t" //d6
+
+        "movq 1*8+6*16(%%"REG_d"), %%mm7        \n\t"
+        "psraw $2, %%mm5                        \n\t"
+
+        "paddusw 1*8+2*16(%%"REG_d"), %%mm1     \n\t"
+        "psubw %%mm7, %%mm6                     \n\t"
+        // t7 d2 /t11 t4 t6 - d6 /t10
+
+        "paddw 1*8+2*16(%%"REG_d"), %%mm1       \n\t"
+        "paddusw %%mm7, %%mm6                   \n\t"
+
+        "psubusw 1*8+2*16(%%"REG_d"), %%mm1     \n\t"
+        "paddw %%mm7, %%mm6                     \n\t"
+
+        "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3\n\t"
+        "psubusw %%mm7, %%mm6                   \n\t"
+
+        //movq [edi+"DCTSIZE_S"*2*2], mm1
+        //movq [edi+"DCTSIZE_S"*6*2], mm6
+        "movq %%mm1, %%mm7                      \n\t"
+        "psraw $2, %%mm2                        \n\t"
+
+        "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4\n\t"
+        "psubw %%mm6, %%mm1                     \n\t"
+
+        "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0\n\t"
+        "paddw %%mm7, %%mm6                     \n\t" //'t13
+
+        "psraw $2, %%mm6                        \n\t" //paddw mm6, MM_2 !!    ---
+        "movq %%mm2, %%mm7                      \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
+        "paddw %%mm6, %%mm2                     \n\t" //'t0
+
+        "movq %%mm2, 0*8+%3                     \n\t" //!
+        "psubw %%mm6, %%mm7                     \n\t" //'t3
+
+        "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
+        "psubw %%mm6, %%mm1                     \n\t" //'t12
+
+        "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2\n\t" //t5
+        "movq %%mm5, %%mm6                      \n\t"
+
+        "movq %%mm7, 3*8+%3                     \n\t"
+        "paddw %%mm2, %%mm3                     \n\t" //t10
+
+        "paddw %%mm4, %%mm2                     \n\t" //t11
+        "paddw %%mm0, %%mm4                     \n\t" //t12
+
+        "movq %%mm3, %%mm7                      \n\t"
+        "psubw %%mm4, %%mm3                     \n\t"
+
+        "psllw $2, %%mm3                        \n\t"
+        "psllw $2, %%mm7                        \n\t" //opt for P6
+
+        "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
+        "psllw $2, %%mm4                        \n\t"
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_541196100)", %%mm7 \n\t"
+        "psllw $2, %%mm2                        \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
+        "paddw %%mm1, %%mm5                     \n\t" //'t1
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_707106781)", %%mm2 \n\t"
+        "psubw %%mm1, %%mm6                     \n\t" //'t2
+        // t7 't12 't11 t4 t6 - 't13 't10   ---
+
+        "paddw %%mm3, %%mm7                     \n\t" //z2
+
+        "movq %%mm5, 1*8+%3                     \n\t"
+        "paddw %%mm3, %%mm4                     \n\t" //z4
+
+        "movq 1*8+3*16(%%"REG_d"), %%mm3        \n\t"
+        "movq %%mm0, %%mm1                      \n\t"
+
+        "movq %%mm6, 2*8+%3                     \n\t"
+        "psubw %%mm2, %%mm1                     \n\t" //z13
+
+//===
+        "paddw %%mm2, %%mm0                     \n\t" //z11
+        "movq %%mm1, %%mm5                      \n\t"
+
+        "movq 1*8+5*16(%%"REG_d"), %%mm2        \n\t"
+        "psubw %%mm7, %%mm1                     \n\t" //d3
+
+        "paddw %%mm7, %%mm5                     \n\t" //d5
+        "psubw %%mm3, %%mm1                     \n\t"
+
+        "movq 1*8+1*16(%%"REG_d"), %%mm7        \n\t"
+        "psubw %%mm2, %%mm5                     \n\t"
+
+        "movq %%mm0, %%mm6                      \n\t"
+        "paddw %%mm4, %%mm0                     \n\t" //d1
+
+        "paddusw %%mm3, %%mm1                   \n\t"
+        "psubw %%mm4, %%mm6                     \n\t" //d7
+
+        // d1 d3 - - - d5 d7 -
+        "movq 1*8+7*16(%%"REG_d"), %%mm4        \n\t"
+        "psubw %%mm7, %%mm0                     \n\t"
+
+        "psubw %%mm4, %%mm6                     \n\t"
+        "paddusw %%mm2, %%mm5                   \n\t"
+
+        "paddusw %%mm4, %%mm6                   \n\t"
+        "paddw %%mm3, %%mm1                     \n\t"
+
+        "paddw %%mm2, %%mm5                     \n\t"
+        "paddw %%mm4, %%mm6                     \n\t"
+
+        "psubusw %%mm3, %%mm1                   \n\t"
+        "psubusw %%mm2, %%mm5                   \n\t"
+
+        "psubusw %%mm4, %%mm6                   \n\t"
+        "movq %%mm1, %%mm4                      \n\t"
+
+        "por %%mm5, %%mm4                       \n\t"
+        "paddusw %%mm7, %%mm0                   \n\t"
+
+        "por %%mm6, %%mm4                       \n\t"
+        "paddw %%mm7, %%mm0                     \n\t"
+
+        "packssdw %%mm4, %%mm4                  \n\t"
+        "psubusw %%mm7, %%mm0                   \n\t"
+
+        "movd %%mm4, %%"REG_a"                  \n\t"
+        "or %%"REG_a", %%"REG_a"                \n\t"
+        "jnz 3f                                 \n\t"
+        //movq [edi+"DCTSIZE_S"*3*2], mm1
+        //movq [edi+"DCTSIZE_S"*5*2], mm5
+        //movq [edi+"DCTSIZE_S"*1*2], mm0
+        //movq [edi+"DCTSIZE_S"*7*2], mm6
+        // t4 t5 - - - t6 t7 -
+        //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
+//Typical numbers: nondc - 19%%,  dc - 26%%,  zero - 55%%. zero case alone isn't worthwhile
+        "movq 0*8+%3, %%mm4                    \n\t"
+        "movq %%mm0, %%mm1                     \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
+        "movq %%mm1, %%mm2                     \n\t"
+
+        "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5\n\t"
+        "movq %%mm2, %%mm3                     \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
+        "paddw %%mm4, %%mm5                    \n\t"
+
+        "movq 1*8+%3, %%mm6                    \n\t"
+        //paddw mm3, MM_2
+        "psraw $2, %%mm3                       \n\t" //tmp7
+
+        "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
+        "psubw %%mm3, %%mm4                    \n\t"
+
+        "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7\n\t"
+        "paddw %%mm3, %%mm5                    \n\t"
+
+        "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D")\n\t"
+        "paddw %%mm6, %%mm7                    \n\t"
+
+        "movq 2*8+%3, %%mm3                    \n\t"
+        "psubw %%mm0, %%mm6                    \n\t"
+
+        "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4\n\t"
+        "paddw %%mm0, %%mm7                    \n\t"
+
+        "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D")\n\t"
+        "paddw %%mm3, %%mm4                    \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D")\n\t"
+        "psubw %%mm1, %%mm3                    \n\t"
+
+        "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5\n\t"
+        "paddw %%mm1, %%mm4                    \n\t"
+
+        "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6\n\t"
+        "paddw %%mm3, %%mm5                    \n\t"
+
+        "movq 3*8+%3, %%mm0                    \n\t"
+        "add $24, %%"REG_S"                    \n\t"
+
+        "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D")\n\t"
+        "paddw %%mm0, %%mm6                    \n\t"
+
+        "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D")\n\t"
+        "psubw %%mm2, %%mm0                    \n\t"
+
+        "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7\n\t"
+        "paddw %%mm2, %%mm6                    \n\t"
+
+        "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D")\n\t"
+        "paddw %%mm0, %%mm7                    \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D")\n\t"
+
+        "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D")\n\t"
+        "add $24, %%"REG_D"                    \n\t"
+        "sub $2, %%"REG_c"                     \n\t"
+        "jnz 1b                                \n\t"
+        "jmp 5f                                \n\t"
+
+        "3:                                    \n\t"
+        //--- non DC2
+        //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1  (actually thr1, thr1, thr1-1)
+        //psraw mm5, 2
+        //psraw mm0, 2
+        //psraw mm6, 2
+        "movq %%mm5, %%mm3                    \n\t"
+        "psubw %%mm1, %%mm5                   \n\t"
+
+        "psllw $1, %%mm5                      \n\t" //'z10
+        "paddw %%mm1, %%mm3                   \n\t" //'z13
+
+        "movq %%mm0, %%mm2                    \n\t"
+        "psubw %%mm6, %%mm0                   \n\t"
+
+        "movq %%mm5, %%mm1                    \n\t"
+        "psllw $1, %%mm0                      \n\t" //'z12
+
+        "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
+        "paddw %%mm0, %%mm5                   \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
+        "paddw %%mm6, %%mm2                   \n\t" //'z11
+
+        "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
+        "movq %%mm2, %%mm7                    \n\t"
+
+        //---
+        "movq 0*8+%3, %%mm4                   \n\t"
+        "psubw %%mm3, %%mm2                   \n\t"
+
+        "psllw $1, %%mm2                      \n\t"
+        "paddw %%mm3, %%mm7                   \n\t" //'t7
+
+        "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
+        "movq %%mm4, %%mm6                    \n\t"
+        //paddw mm7, MM_2
+        "psraw $2, %%mm7                      \n\t"
+
+        "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
+        "psubw %%mm7, %%mm6                   \n\t"
+
+        "movq 1*8+%3, %%mm3                   \n\t"
+        "paddw %%mm7, %%mm4                   \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
+        "paddw %%mm5, %%mm1                   \n\t" //'t12
+
+        "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
+        "psubw %%mm7, %%mm1                   \n\t" //'t6
+
+        "movq 2*8+%3, %%mm7                   \n\t"
+        "psubw %%mm5, %%mm0                   \n\t" //'t10
+
+        "movq 3*8+%3, %%mm6                   \n\t"
+        "movq %%mm3, %%mm5                    \n\t"
+
+        "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
+        "psubw %%mm1, %%mm5                   \n\t"
+
+        "psubw %%mm1, %%mm2                   \n\t" //'t5
+        "paddw %%mm1, %%mm3                   \n\t"
+
+        "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
+        "movq %%mm7, %%mm4                    \n\t"
+
+        "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
+        "psubw %%mm2, %%mm4                   \n\t"
+
+        "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
+        "paddw %%mm2, %%mm7                   \n\t"
+
+        "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
+        "paddw %%mm2, %%mm0                    \n\t" //'t4
+
+        // 't4 't6 't5 - - - - 't7
+        "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
+        "movq %%mm6, %%mm1                     \n\t"
+
+        "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
+        "psubw %%mm0, %%mm1                    \n\t"
+
+        "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
+        "paddw %%mm0, %%mm6                    \n\t"
+
+        "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
+        "add $24, %%"REG_S"                    \n\t"
+
+        "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
+
+        "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
+        "add $24, %%"REG_D"                    \n\t"
+        "sub $2, %%"REG_c"                     \n\t"
+        "jnz 1b                                \n\t"
+        "5:                                    \n\t"
+
+        : "+S"(data), "+D"(output), "+c"(cnt), "=o"(temps)
+        : "d"(thr_adr)
+          NAMED_CONSTRAINTS_ADD(ff_MM_FIX_0_707106781, MM_2,MM_FIX_1_414213562_A, MM_FIX_1_414213562, MM_FIX_0_382683433,
+                                ff_MM_FIX_0_541196100, MM_FIX_1_306562965, MM_FIX_0_847759065)
+          NAMED_CONSTRAINTS_ADD(MM_FIX_0_566454497, MM_FIX_0_198912367, MM_FIX_2_613125930, MM_FIX_1_847759065,
+                                MM_FIX_1_082392200)
+        : "%"REG_a
+        );
+}
+
+static void row_idct_mmx (int16_t *workspace, int16_t *output_adr, int output_stride, int cnt)
+{
+    DECLARE_ALIGNED(8, uint64_t, temps)[4];
+
+    __asm__ volatile(
+        "lea (%%"REG_a",%%"REG_a",2), %%"REG_d"    \n\t"
+        "1:                     \n\t"
+        "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm0    \n\t"
+        //
+
+        "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm1    \n\t"
+        "movq %%mm0, %%mm4                         \n\t"
+
+        "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2    \n\t"
+        "punpcklwd %%mm1, %%mm0                    \n\t"
+
+        "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm3    \n\t"
+        "punpckhwd %%mm1, %%mm4                    \n\t"
+
+        //transpose 4x4
+        "movq %%mm2, %%mm7                         \n\t"
+        "punpcklwd %%mm3, %%mm2                    \n\t"
+
+        "movq %%mm0, %%mm6                         \n\t"
+        "punpckldq %%mm2, %%mm0                    \n\t" //0
+
+        "punpckhdq %%mm2, %%mm6                    \n\t" //1
+        "movq %%mm0, %%mm5                         \n\t"
+
+        "punpckhwd %%mm3, %%mm7                    \n\t"
+        "psubw %%mm6, %%mm0                        \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm0 \n\t"
+        "movq %%mm4, %%mm2                         \n\t"
+
+        "punpckldq %%mm7, %%mm4                    \n\t" //2
+        "paddw %%mm6, %%mm5                        \n\t"
+
+        "punpckhdq %%mm7, %%mm2                    \n\t" //3
+        "movq %%mm4, %%mm1                         \n\t"
+
+        "psllw $2, %%mm0                           \n\t"
+        "paddw %%mm2, %%mm4                        \n\t" //t10
+
+        "movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_S"), %%mm3 \n\t"
+        "psubw %%mm2, %%mm1                        \n\t" //t11
+
+        "movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_S"), %%mm2 \n\t"
+        "psubw %%mm5, %%mm0                        \n\t"
+
+        "movq %%mm4, %%mm6                         \n\t"
+        "paddw %%mm5, %%mm4                        \n\t" //t0
+
+        "psubw %%mm5, %%mm6                        \n\t" //t3
+        "movq %%mm1, %%mm7                         \n\t"
+
+        "movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_S"), %%mm5 \n\t"
+        "paddw %%mm0, %%mm1                        \n\t" //t1
+
+        "movq %%mm4, 0*8+%3                        \n\t" //t0
+        "movq %%mm3, %%mm4                         \n\t"
+
+        "movq %%mm6, 1*8+%3                        \n\t" //t3
+        "punpcklwd %%mm2, %%mm3                    \n\t"
+
+        //transpose 4x4
+        "movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_S"), %%mm6 \n\t"
+        "punpckhwd %%mm2, %%mm4                    \n\t"
+
+        "movq %%mm5, %%mm2                         \n\t"
+        "punpcklwd %%mm6, %%mm5                    \n\t"
+
+        "psubw %%mm0, %%mm7                        \n\t" //t2
+        "punpckhwd %%mm6, %%mm2                    \n\t"
+
+        "movq %%mm3, %%mm0                         \n\t"
+        "punpckldq %%mm5, %%mm3                    \n\t" //4
+
+        "punpckhdq %%mm5, %%mm0                    \n\t" //5
+        "movq %%mm4, %%mm5                         \n\t"
+
+        //
+        "movq %%mm3, %%mm6                         \n\t"
+        "punpckldq %%mm2, %%mm4                    \n\t" //6
+
+        "psubw %%mm0, %%mm3                        \n\t" //z10
+        "punpckhdq %%mm2, %%mm5                    \n\t" //7
+
+        "paddw %%mm0, %%mm6                        \n\t" //z13
+        "movq %%mm4, %%mm2                         \n\t"
+
+        "movq %%mm3, %%mm0                         \n\t"
+        "psubw %%mm5, %%mm4                        \n\t" //z12
+
+        "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0\n\t" //-
+        "paddw %%mm4, %%mm3                        \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm3\n\t" //z5
+        "paddw %%mm5, %%mm2                        \n\t" //z11  >
+
+        "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4\n\t"
+        "movq %%mm2, %%mm5                         \n\t"
+
+        "psubw %%mm6, %%mm2                        \n\t"
+        "paddw %%mm6, %%mm5                        \n\t" //t7
+
+        "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2\n\t" //t11
+        "paddw %%mm3, %%mm0                        \n\t" //t12
+
+        "psllw $3, %%mm0                           \n\t"
+        "psubw %%mm3, %%mm4                        \n\t" //t10
+
+        "movq 0*8+%3, %%mm6                        \n\t"
+        "movq %%mm1, %%mm3                         \n\t"
+
+        "psllw $3, %%mm4                           \n\t"
+        "psubw %%mm5, %%mm0                        \n\t" //t6
+
+        "psllw $3, %%mm2                           \n\t"
+        "paddw %%mm0, %%mm1                        \n\t" //d1
+
+        "psubw %%mm0, %%mm2                        \n\t" //t5
+        "psubw %%mm0, %%mm3                        \n\t" //d6
+
+        "paddw %%mm2, %%mm4                        \n\t" //t4
+        "movq %%mm7, %%mm0                         \n\t"
+
+        "paddw %%mm2, %%mm7                        \n\t" //d2
+        "psubw %%mm2, %%mm0                        \n\t" //d5
+
+        "movq "MANGLE(MM_DESCALE_RND)", %%mm2      \n\t" //4
+        "psubw %%mm5, %%mm6                        \n\t" //d7
+
+        "paddw 0*8+%3, %%mm5                       \n\t" //d0
+        "paddw %%mm2, %%mm1                        \n\t"
+
+        "paddw %%mm2, %%mm5                        \n\t"
+        "psraw $3, %%mm1                           \n\t"
+
+        "paddw %%mm2, %%mm7                        \n\t"
+        "psraw $3, %%mm5                           \n\t"
+
+        "paddw (%%"REG_D"), %%mm5                  \n\t"
+        "psraw $3, %%mm7                           \n\t"
+
+        "paddw (%%"REG_D",%%"REG_a"), %%mm1        \n\t"
+        "paddw %%mm2, %%mm0                        \n\t"
+
+        "paddw (%%"REG_D",%%"REG_a",2), %%mm7      \n\t"
+        "paddw %%mm2, %%mm3                        \n\t"
+
+        "movq %%mm5, (%%"REG_D")                   \n\t"
+        "paddw %%mm2, %%mm6                        \n\t"
+
+        "movq %%mm1, (%%"REG_D",%%"REG_a")         \n\t"
+        "psraw $3, %%mm0                           \n\t"
+
+        "movq %%mm7, (%%"REG_D",%%"REG_a",2)       \n\t"
+        "add %%"REG_d", %%"REG_D"                  \n\t" //3*ls
+
+        "movq 1*8+%3, %%mm5                        \n\t" //t3
+        "psraw $3, %%mm3                           \n\t"
+
+        "paddw (%%"REG_D",%%"REG_a",2), %%mm0      \n\t"
+        "psubw %%mm4, %%mm5                        \n\t" //d3
+
+        "paddw (%%"REG_D",%%"REG_d"), %%mm3        \n\t"
+        "psraw $3, %%mm6                           \n\t"
+
+        "paddw 1*8+%3, %%mm4                       \n\t" //d4
+        "paddw %%mm2, %%mm5                        \n\t"
+
+        "paddw (%%"REG_D",%%"REG_a",4), %%mm6      \n\t"
+        "paddw %%mm2, %%mm4                        \n\t"
+
+        "movq %%mm0, (%%"REG_D",%%"REG_a",2)       \n\t"
+        "psraw $3, %%mm5                           \n\t"
+
+        "paddw (%%"REG_D"), %%mm5                  \n\t"
+        "psraw $3, %%mm4                           \n\t"
+
+        "paddw (%%"REG_D",%%"REG_a"), %%mm4        \n\t"
+        "add $"DCTSIZE_S"*2*4, %%"REG_S"           \n\t" //4 rows
+
+        "movq %%mm3, (%%"REG_D",%%"REG_d")         \n\t"
+        "movq %%mm6, (%%"REG_D",%%"REG_a",4)       \n\t"
+        "movq %%mm5, (%%"REG_D")                   \n\t"
+        "movq %%mm4, (%%"REG_D",%%"REG_a")         \n\t"
+
+        "sub %%"REG_d", %%"REG_D"                  \n\t"
+        "add $8, %%"REG_D"                         \n\t"
+        "dec %%"REG_c"                             \n\t"
+        "jnz 1b                                    \n\t"
+
+        : "+S"(workspace), "+D"(output_adr), "+c"(cnt), "=o"(temps)
+        : "a"(output_stride * sizeof(short))
+        NAMED_CONSTRAINTS_ADD(MM_FIX_1_414213562_A, MM_FIX_2_613125930, MM_FIX_1_847759065, MM_FIX_1_082392200,
+                              MM_FIX_1_414213562,MM_DESCALE_RND)
+        : "%"REG_d
+        );
+}
+
+static void row_fdct_mmx(int16_t *data, const uint8_t *pixels, int line_size, int cnt)
+{
+    DECLARE_ALIGNED(8, uint64_t, temps)[4];
+
+    __asm__ volatile(
+        "lea (%%"REG_a",%%"REG_a",2), %%"REG_d"    \n\t"
+        "6:                                        \n\t"
+        "movd (%%"REG_S"), %%mm0                   \n\t"
+        "pxor %%mm7, %%mm7                         \n\t"
+
+        "movd (%%"REG_S",%%"REG_a"), %%mm1         \n\t"
+        "punpcklbw %%mm7, %%mm0                    \n\t"
+
+        "movd (%%"REG_S",%%"REG_a",2), %%mm2       \n\t"
+        "punpcklbw %%mm7, %%mm1                    \n\t"
+
+        "punpcklbw %%mm7, %%mm2                    \n\t"
+        "add %%"REG_d", %%"REG_S"                  \n\t"
+
+        "movq %%mm0, %%mm5                         \n\t"
+        //
+
+        "movd (%%"REG_S",%%"REG_a",4), %%mm3       \n\t" //7  ;prefetch!
+        "movq %%mm1, %%mm6                         \n\t"
+
+        "movd (%%"REG_S",%%"REG_d"), %%mm4         \n\t" //6
+        "punpcklbw %%mm7, %%mm3                    \n\t"
+
+        "psubw %%mm3, %%mm5                        \n\t"
+        "punpcklbw %%mm7, %%mm4                    \n\t"
+
+        "paddw %%mm3, %%mm0                        \n\t"
+        "psubw %%mm4, %%mm6                        \n\t"
+
+        "movd (%%"REG_S",%%"REG_a",2), %%mm3       \n\t" //5
+        "paddw %%mm4, %%mm1                        \n\t"
+
+        "movq %%mm5, %3                            \n\t" //t7
+        "punpcklbw %%mm7, %%mm3                    \n\t"
+
+        "movq %%mm6, %4                            \n\t" //t6
+        "movq %%mm2, %%mm4                         \n\t"
+
+        "movd (%%"REG_S"), %%mm5                   \n\t" //3
+        "paddw %%mm3, %%mm2                        \n\t"
+
+        "movd (%%"REG_S",%%"REG_a"), %%mm6         \n\t" //4
+        "punpcklbw %%mm7, %%mm5                    \n\t"
+
+        "psubw %%mm3, %%mm4                        \n\t"
+        "punpcklbw %%mm7, %%mm6                    \n\t"
+
+        "movq %%mm5, %%mm3                         \n\t"
+        "paddw %%mm6, %%mm5                        \n\t" //t3
+
+        "psubw %%mm6, %%mm3                        \n\t" //t4  ; t0 t1 t2 t4 t5 t3 - -
+        "movq %%mm0, %%mm6                         \n\t"
+
+        "movq %%mm1, %%mm7                         \n\t"
+        "psubw %%mm5, %%mm0                        \n\t" //t13
+
+        "psubw %%mm2, %%mm1                        \n\t"
+        "paddw %%mm2, %%mm7                        \n\t" //t11
+
+        "paddw %%mm0, %%mm1                        \n\t"
+        "movq %%mm7, %%mm2                         \n\t"
+
+        "psllw $2, %%mm1                           \n\t"
+        "paddw %%mm5, %%mm6                        \n\t" //t10
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_707106781)", %%mm1 \n\t"
+        "paddw %%mm6, %%mm7                        \n\t" //d2
+
+        "psubw %%mm2, %%mm6                        \n\t" //d3
+        "movq %%mm0, %%mm5                         \n\t"
+
+        //transpose 4x4
+        "movq %%mm7, %%mm2                         \n\t"
+        "punpcklwd %%mm6, %%mm7                    \n\t"
+
+        "paddw %%mm1, %%mm0                        \n\t" //d0
+        "punpckhwd %%mm6, %%mm2                    \n\t"
+
+        "psubw %%mm1, %%mm5                        \n\t" //d1
+        "movq %%mm0, %%mm6                         \n\t"
+
+        "movq %4, %%mm1                            \n\t"
+        "punpcklwd %%mm5, %%mm0                    \n\t"
+
+        "punpckhwd %%mm5, %%mm6                    \n\t"
+        "movq %%mm0, %%mm5                         \n\t"
+
+        "punpckldq %%mm7, %%mm0                    \n\t" //0
+        "paddw %%mm4, %%mm3                        \n\t"
+
+        "punpckhdq %%mm7, %%mm5                    \n\t" //1
+        "movq %%mm6, %%mm7                         \n\t"
+
+        "movq %%mm0, "DCTSIZE_S"*0*2(%%"REG_D")    \n\t"
+        "punpckldq %%mm2, %%mm6                    \n\t" //2
+
+        "movq %%mm5, "DCTSIZE_S"*1*2(%%"REG_D")    \n\t"
+        "punpckhdq %%mm2, %%mm7                    \n\t" //3
+
+        "movq %%mm6, "DCTSIZE_S"*2*2(%%"REG_D")    \n\t"
+        "paddw %%mm1, %%mm4                        \n\t"
+
+        "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D")    \n\t"
+        "psllw $2, %%mm3                           \n\t" //t10
+
+        "movq %3, %%mm2                            \n\t"
+        "psllw $2, %%mm4                           \n\t" //t11
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_707106781)", %%mm4 \n\t" //z3
+        "paddw %%mm2, %%mm1                        \n\t"
+
+        "psllw $2, %%mm1                           \n\t" //t12
+        "movq %%mm3, %%mm0                         \n\t"
+
+        "pmulhw "MANGLE(ff_MM_FIX_0_541196100)", %%mm0 \n\t"
+        "psubw %%mm1, %%mm3                        \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" //z5
+        "movq %%mm2, %%mm5                         \n\t"
+
+        "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t"
+        "psubw %%mm4, %%mm2                        \n\t" //z13
+
+        "paddw %%mm4, %%mm5                        \n\t" //z11
+        "movq %%mm2, %%mm6                         \n\t"
+
+        "paddw %%mm3, %%mm0                        \n\t" //z2
+        "movq %%mm5, %%mm7                         \n\t"
+
+        "paddw %%mm0, %%mm2                        \n\t" //d4
+        "psubw %%mm0, %%mm6                        \n\t" //d5
+
+        "movq %%mm2, %%mm4                         \n\t"
+        "paddw %%mm3, %%mm1                        \n\t" //z4
+
+        //transpose 4x4
+        "punpcklwd %%mm6, %%mm2                    \n\t"
+        "paddw %%mm1, %%mm5                        \n\t" //d6
+
+        "punpckhwd %%mm6, %%mm4                    \n\t"
+        "psubw %%mm1, %%mm7                        \n\t" //d7
+
+        "movq %%mm5, %%mm6                         \n\t"
+        "punpcklwd %%mm7, %%mm5                    \n\t"
+
+        "punpckhwd %%mm7, %%mm6                    \n\t"
+        "movq %%mm2, %%mm7                         \n\t"
+
+        "punpckldq %%mm5, %%mm2                    \n\t" //4
+        "sub %%"REG_d", %%"REG_S"                  \n\t"
+
+        "punpckhdq %%mm5, %%mm7                    \n\t" //5
+        "movq %%mm4, %%mm5                         \n\t"
+
+        "movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_D") \n\t"
+        "punpckldq %%mm6, %%mm4                    \n\t" //6
+
+        "movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_D") \n\t"
+        "punpckhdq %%mm6, %%mm5                    \n\t" //7
+
+        "movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_D") \n\t"
+        "add $4, %%"REG_S"                         \n\t"
+
+        "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
+        "add $"DCTSIZE_S"*2*4, %%"REG_D"           \n\t" //4 rows
+        "dec %%"REG_c"                             \n\t"
+        "jnz 6b                                    \n\t"
+
+        : "+S"(pixels), "+D"(data), "+c"(cnt), "=o"(temps), "=o"(temps[1])
+        : "a"(line_size)
+        NAMED_CONSTRAINTS_ADD(ff_MM_FIX_0_707106781, ff_MM_FIX_0_541196100, MM_FIX_0_382683433, MM_FIX_1_306562965)
+        : "%"REG_d);
+}
+
+av_cold void ff_fspp_init_x86(FSPPContext *s)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (HAVE_MMX_INLINE && cpu_flags & AV_CPU_FLAG_MMX) {
+        s->store_slice  = store_slice_mmx;
+        s->store_slice2 = store_slice2_mmx;
+        s->mul_thrmat   = mul_thrmat_mmx;
+        s->column_fidct = column_fidct_mmx;
+        s->row_idct     = row_idct_mmx;
+        s->row_fdct     = row_fdct_mmx;
+    }
+}