Introduce a TextureDSP module

This module implements generic texture decompression from different
families (DXTC, RGTC, BCn) and texture compression DXTC 1, 3, and 5.

Signed-off-by: Vittorio Giovara <vittorio.giovara@gmail.com>

5 files changed, 1337 insertions, 0 deletions

diff --git a/configure b/configure
index 6d3fe62d45..d62fdb6c15 100755
--- a/configure
+++ b/configure
@@ -1626,6 +1626,8 @@ CONFIG_EXTRA="
     sinewin
     snappy
     startcode
+    texturedsp
+    texturedspenc
     tpeldsp
     videodsp
     vp3dsp
diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index 6e04ca62cb..1d34e2fd2d 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -93,6 +93,8 @@ OBJS-$(CONFIG_RDFT)                    += rdft.o $(RDFT-OBJS-yes)
 OBJS-$(CONFIG_SINEWIN)                 += sinewin.o
 OBJS-$(CONFIG_SNAPPY)                  += snappy.o
 OBJS-$(CONFIG_STARTCODE)               += startcode.o
+OBJS-$(CONFIG_TEXTUREDSP)              += texturedsp.o
+OBJS-$(CONFIG_TEXTUREDSPENC)           += texturedspenc.o
 OBJS-$(CONFIG_TPELDSP)                 += tpeldsp.o
 OBJS-$(CONFIG_VIDEODSP)                += videodsp.o
 OBJS-$(CONFIG_VP3DSP)                  += vp3dsp.o
diff --git a/libavcodec/texturedsp.c b/libavcodec/texturedsp.c
new file mode 100644
index 0000000000..78eb6fa0ce
--- /dev/null
+++ b/libavcodec/texturedsp.c
@@ -0,0 +1,610 @@
+/*
+ * Texture block decompression
+ * Copyright (C) 2009 Benjamin Dobell, Glass Echidna
+ * Copyright (C) 2012 Matthäus G. "Anteru" Chajdas (http://anteru.net)
+ * Copyright (C) 2015 Vittorio Giovara <vittorio.giovara@gmail.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stddef.h>
+#include <stdint.h>
+
+#include "libavutil/attributes.h"
+#include "libavutil/common.h"
+#include "libavutil/intreadwrite.h"
+
+#include "texturedsp.h"
+
+#define RGBA(r, g, b, a) (r) | ((g) << 8) | ((b) << 16) | ((a) << 24)
+
+static av_always_inline void extract_color(uint32_t colors[4],
+                                           uint16_t color0,
+                                           uint16_t color1,
+                                           int dxtn, int alpha)
+{
+    int tmp;
+    uint8_t r0, g0, b0, r1, g1, b1;
+    uint8_t a = dxtn ? 0 : 255;
+
+    tmp = (color0 >> 11) * 255 + 16;
+    r0  = (uint8_t) ((tmp / 32 + tmp) / 32);
+    tmp = ((color0 & 0x07E0) >> 5) * 255 + 32;
+    g0  = (uint8_t) ((tmp / 64 + tmp) / 64);
+    tmp = (color0 & 0x001F) * 255 + 16;
+    b0  = (uint8_t) ((tmp / 32 + tmp) / 32);
+
+    tmp = (color1 >> 11) * 255 + 16;
+    r1  = (uint8_t) ((tmp / 32 + tmp) / 32);
+    tmp = ((color1 & 0x07E0) >> 5) * 255 + 32;
+    g1  = (uint8_t) ((tmp / 64 + tmp) / 64);
+    tmp = (color1 & 0x001F) * 255 + 16;
+    b1  = (uint8_t) ((tmp / 32 + tmp) / 32);
+
+    if (dxtn || color0 > color1) {
+        colors[0] = RGBA(r0, g0, b0, a);
+        colors[1] = RGBA(r1, g1, b1, a);
+        colors[2] = RGBA((2 * r0 + r1) / 3,
+                         (2 * g0 + g1) / 3,
+                         (2 * b0 + b1) / 3,
+                         a);
+        colors[3] = RGBA((2 * r1 + r0) / 3,
+                         (2 * g1 + g0) / 3,
+                         (2 * b1 + b0) / 3,
+                         a);
+    } else {
+        colors[0] = RGBA(r0, g0, b0, a);
+        colors[1] = RGBA(r1, g1, b1, a);
+        colors[2] = RGBA((r0 + r1) / 2,
+                         (g0 + g1) / 2,
+                         (b0 + b1) / 2,
+                         a);
+        colors[3] = RGBA(0, 0, 0, alpha);
+    }
+}
+
+static inline void dxt1_block_internal(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *block, uint8_t alpha)
+{
+    int x, y;
+    uint32_t colors[4];
+    uint16_t color0 = AV_RL16(block + 0);
+    uint16_t color1 = AV_RL16(block + 2);
+    uint32_t code   = AV_RL32(block + 4);
+
+    extract_color(colors, color0, color1, 0, alpha);
+
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            uint32_t pixel = colors[code & 3];
+            code >>= 2;
+            AV_WL32(dst + x * 4, pixel);
+        }
+        dst += stride;
+    }
+}
+
+/**
+ * Decompress one block of a DXT1 texture and store the resulting
+ * RGBA pixels in 'dst'. Alpha component is fully opaque.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt1_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    dxt1_block_internal(dst, stride, block, 255);
+
+    return 8;
+}
+
+/**
+ * Decompress one block of a DXT1 with 1-bit alpha texture and store
+ * the resulting RGBA pixels in 'dst'. Alpha is either fully opaque or
+ * fully transparent.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt1a_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    dxt1_block_internal(dst, stride, block, 0);
+
+    return 8;
+}
+
+static inline void dxt3_block_internal(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *block)
+{
+    int x, y;
+    uint32_t colors[4];
+    uint16_t color0 = AV_RL16(block +  8);
+    uint16_t color1 = AV_RL16(block + 10);
+    uint32_t code   = AV_RL32(block + 12);
+
+    extract_color(colors, color0, color1, 1, 0);
+
+    for (y = 0; y < 4; y++) {
+        const uint16_t alpha_code = AV_RL16(block + 2 * y);
+        uint8_t alpha_values[4];
+
+        alpha_values[0] = ((alpha_code >>  0) & 0x0F) * 17;
+        alpha_values[1] = ((alpha_code >>  4) & 0x0F) * 17;
+        alpha_values[2] = ((alpha_code >>  8) & 0x0F) * 17;
+        alpha_values[3] = ((alpha_code >> 12) & 0x0F) * 17;
+
+        for (x = 0; x < 4; x++) {
+            uint8_t alpha = alpha_values[x];
+            uint32_t pixel = colors[code & 3] | (alpha << 24);
+            code >>= 2;
+
+            AV_WL32(dst + x * 4, pixel);
+        }
+        dst += stride;
+    }
+}
+
+/** Convert a premultiplied alpha pixel to a straigth alpha pixel. */
+static av_always_inline void premult2straight(uint8_t *src)
+{
+    int r = src[0];
+    int g = src[1];
+    int b = src[2];
+    int a = src[3]; /* unchanged */
+
+    src[0] = (uint8_t) r * a / 255;
+    src[1] = (uint8_t) g * a / 255;
+    src[2] = (uint8_t) b * a / 255;
+}
+
+/**
+ * Decompress one block of a DXT2 texture and store the resulting
+ * RGBA pixels in 'dst'.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt2_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    int x, y;
+
+    dxt3_block_internal(dst, stride, block);
+
+    /* This format is DXT3, but returns premultiplied alpha. It needs to be
+     * converted because it's what lavc outputs (and swscale expects). */
+    for (y = 0; y < 4; y++)
+        for (x = 0; x < 4; x++)
+            premult2straight(dst + x * 4 + y * stride);
+
+    return 16;
+}
+
+/**
+ * Decompress one block of a DXT3 texture and store the resulting
+ * RGBA pixels in 'dst'.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt3_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    dxt3_block_internal(dst, stride, block);
+
+    return 16;
+}
+
+/**
+ * Decompress a BC 16x3 index block stored as
+ *   h g f e
+ *   d c b a
+ *   p o n m
+ *   l k j i
+ *
+ * Bits packed as
+ *  | h | g | f | e | d | c | b | a | // Entry
+ *  |765 432 107 654 321 076 543 210| // Bit
+ *  |0000000000111111111112222222222| // Byte
+ *
+ * into 16 8-bit indices.
+ */
+static void decompress_indices(uint8_t *dst, const uint8_t *src)
+{
+    int block, i;
+
+    for (block = 0; block < 2; block++) {
+        int tmp = AV_RL24(src);
+
+        /* Unpack 8x3 bit from last 3 byte block */
+        for (i = 0; i < 8; i++)
+            dst[i] = (tmp >> (i * 3)) & 0x7;
+
+        src += 3;
+        dst += 8;
+    }
+}
+
+static inline void dxt5_block_internal(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *block)
+{
+    int x, y;
+    uint32_t colors[4];
+    uint8_t alpha_indices[16];
+    uint16_t color0 = AV_RL16(block + 8);
+    uint16_t color1 = AV_RL16(block + 10);
+    uint32_t code   = AV_RL32(block + 12);
+    uint8_t alpha0  = *(block);
+    uint8_t alpha1  = *(block + 1);
+
+    decompress_indices(alpha_indices, block + 2);
+
+    extract_color(colors, color0, color1, 1, 0);
+
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            int alpha_code = alpha_indices[x + y * 4];
+            uint32_t pixel;
+            uint8_t alpha;
+
+            if (alpha_code == 0) {
+                alpha = alpha0;
+            } else if (alpha_code == 1) {
+                alpha = alpha1;
+            } else {
+                if (alpha0 > alpha1) {
+                    alpha = (uint8_t) (((8 - alpha_code) * alpha0 +
+                                        (alpha_code - 1) * alpha1) / 7);
+                } else {
+                    if (alpha_code == 6) {
+                        alpha = 0;
+                    } else if (alpha_code == 7) {
+                        alpha = 255;
+                    } else {
+                        alpha = (uint8_t) (((6 - alpha_code) * alpha0 +
+                                            (alpha_code - 1) * alpha1) / 5);
+                    }
+                }
+            }
+            pixel = colors[code & 3] | (alpha << 24);
+            code >>= 2;
+            AV_WL32(dst + x * 4, pixel);
+        }
+        dst += stride;
+    }
+}
+
+/**
+ * Decompress one block of a DXT4 texture and store the resulting
+ * RGBA pixels in 'dst'.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt4_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    int x, y;
+
+    dxt5_block_internal(dst, stride, block);
+
+    /* This format is DXT5, but returns premultiplied alpha. It needs to be
+     * converted because it's what lavc outputs (and swscale expects). */
+    for (y = 0; y < 4; y++)
+        for (x = 0; x < 4; x++)
+            premult2straight(dst + x * 4 + y * stride);
+
+    return 16;
+}
+
+/**
+ * Decompress one block of a DXT5 texture and store the resulting
+ * RGBA pixels in 'dst'.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt5_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    dxt5_block_internal(dst, stride, block);
+
+    return 16;
+}
+
+/**
+ * Convert a YCoCg buffer to RGBA.
+ *
+ * @param src    input buffer.
+ * @param scaled variant with scaled chroma components and opaque alpha.
+ */
+static av_always_inline void ycocg2rgba(uint8_t *src, int scaled)
+{
+    int r = src[0];
+    int g = src[1];
+    int b = src[2];
+    int a = src[3];
+
+    int s  = scaled ? (b >> 3) + 1 : 1;
+    int y  = a;
+    int co = (r - 128) / s;
+    int cg = (g - 128) / s;
+
+    src[0] = av_clip_uint8(y + co - cg);
+    src[1] = av_clip_uint8(y + cg);
+    src[2] = av_clip_uint8(y - co - cg);
+    src[3] = scaled ? 255 : b;
+}
+
+/**
+ * Decompress one block of a DXT5 texture with classic YCoCg and store
+ * the resulting RGBA pixels in 'dst'. Alpha component is fully opaque.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt5y_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    int x, y;
+
+    /* This format is basically DXT5, with luma stored in alpha.
+     * Run a normal decompress and then reorder the components. */
+    dxt5_block_internal(dst, stride, block);
+
+    for (y = 0; y < 4; y++)
+        for (x = 0; x < 4; x++)
+            ycocg2rgba(dst + x * 4 + y * stride, 0);
+
+    return 16;
+}
+
+/**
+ * Decompress one block of a DXT5 texture with scaled YCoCg and store
+ * the resulting RGBA pixels in 'dst'. Alpha component is fully opaque.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxt5ys_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    int x, y;
+
+    /* This format is basically DXT5, with luma stored in alpha.
+     * Run a normal decompress and then reorder the components. */
+    dxt5_block_internal(dst, stride, block);
+
+    for (y = 0; y < 4; y++)
+        for (x = 0; x < 4; x++)
+            ycocg2rgba(dst + x * 4 + y * stride, 1);
+
+    return 16;
+}
+
+static inline void rgtc_block_internal(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *block,
+                                       const int *color_tab)
+{
+    uint8_t indices[16];
+    int x, y;
+
+    decompress_indices(indices, block + 2);
+
+    /* Only one or two channels are stored at most, since it only used to
+     * compress specular (black and white) or normal (red and green) maps.
+     * Although the standard says to zero out unused components, many
+     * implementations fill all of them with the same value. */
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            int i = indices[x + y * 4];
+            /* Interval expansion from [-1 1] or [0 1] to [0 255]. */
+            int c = color_tab[i];
+            uint32_t pixel = RGBA(c, c, c, 255);
+            AV_WL32(dst + x * 4 + y * stride, pixel);
+        }
+    }
+}
+
+static inline void rgtc1_block_internal(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *block, int sign)
+{
+    int color_table[8];
+    int r0, r1;
+
+    if (sign) {
+        /* signed data is in [-128 127] so just offset it to unsigned
+         * and it can be treated exactly the same */
+        r0 = ((int8_t) block[0]) + 128;
+        r1 = ((int8_t) block[1]) + 128;
+    } else {
+        r0 = block[0];
+        r1 = block[1];
+    }
+
+    color_table[0] = r0;
+    color_table[1] = r1;
+
+    if (r0 > r1) {
+        /* 6 interpolated color values */
+        color_table[2] = (6 * r0 + 1 * r1) / 7; // bit code 010
+        color_table[3] = (5 * r0 + 2 * r1) / 7; // bit code 011
+        color_table[4] = (4 * r0 + 3 * r1) / 7; // bit code 100
+        color_table[5] = (3 * r0 + 4 * r1) / 7; // bit code 101
+        color_table[6] = (2 * r0 + 5 * r1) / 7; // bit code 110
+        color_table[7] = (1 * r0 + 6 * r1) / 7; // bit code 111
+    } else {
+        /* 4 interpolated color values */
+        color_table[2] = (4 * r0 + 1 * r1) / 5; // bit code 010
+        color_table[3] = (3 * r0 + 2 * r1) / 5; // bit code 011
+        color_table[4] = (2 * r0 + 3 * r1) / 5; // bit code 100
+        color_table[5] = (1 * r0 + 4 * r1) / 5; // bit code 101
+        color_table[6] = 0;    /* min range */  // bit code 110
+        color_table[7] = 255;  /* max range */  // bit code 111
+    }
+
+    rgtc_block_internal(dst, stride, block, color_table);
+}
+
+/**
+ * Decompress one block of a RGRC1 texture with signed components
+ * and store the resulting RGBA pixels in 'dst'.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int rgtc1s_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    rgtc1_block_internal(dst, stride, block, 1);
+
+    return 8;
+}
+
+/**
+ * Decompress one block of a RGRC1 texture with unsigned components
+ * and store the resulting RGBA pixels in 'dst'.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int rgtc1u_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    rgtc1_block_internal(dst, stride, block, 0);
+
+    return 8;
+}
+
+static inline void rgtc2_block_internal(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *block, int sign)
+{
+    /* 4x4 block containing 4 component pixels. */
+    uint8_t c0[4 * 4 * 4];
+    uint8_t c1[4 * 4 * 4];
+    int x, y;
+
+    /* Decompress the two channels separately and interleave them afterwards. */
+    rgtc1_block_internal(c0, 16, block, sign);
+    rgtc1_block_internal(c1, 16, block + 8, sign);
+
+    /* B is rebuilt exactly like a normal map. */
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            uint8_t *p = dst + x * 4 + y * stride;
+            int r = c0[x * 4 + y * 16];
+            int g = c1[x * 4 + y * 16];
+            int b = 127;
+
+            int d = (255 * 255 - r * r - g * g) / 2;
+            if (d > 0)
+                b = rint(sqrtf(d));
+
+            p[0] = r;
+            p[1] = g;
+            p[2] = b;
+            p[3] = 255;
+        }
+    }
+}
+
+/**
+ * Decompress one block of a RGRC2 texture with signed components
+ * and store the resulting RGBA pixels in 'dst'. Alpha is fully opaque.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int rgtc2s_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    rgtc2_block_internal(dst, stride, block, 1);
+
+    return 16;
+}
+
+/**
+ * Decompress one block of a RGRC2 texture with unsigned components
+ * and store the resulting RGBA pixels in 'dst'. Alpha is fully opaque.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int rgtc2u_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    rgtc2_block_internal(dst, stride, block, 0);
+
+    return 16;
+}
+
+/**
+ * Decompress one block of a 3Dc texture with unsigned components
+ * and store the resulting RGBA pixels in 'dst'. Alpha is fully opaque.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to decompress.
+ * @return how much texture data has been consumed.
+ */
+static int dxn3dc_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    int x, y;
+    rgtc2_block_internal(dst, stride, block, 0);
+
+    /* This is the 3Dc variant of RGTC2, with swapped R and G. */
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            uint8_t *p = dst + x * 4 + y * stride;
+            FFSWAP(uint8_t, p[0], p[1]);
+        }
+    }
+
+    return 16;
+}
+
+av_cold void ff_texturedsp_init(TextureDSPContext *c)
+{
+    c->dxt1_block   = dxt1_block;
+    c->dxt1a_block  = dxt1a_block;
+    c->dxt2_block   = dxt2_block;
+    c->dxt3_block   = dxt3_block;
+    c->dxt4_block   = dxt4_block;
+    c->dxt5_block   = dxt5_block;
+    c->dxt5y_block  = dxt5y_block;
+    c->dxt5ys_block = dxt5ys_block;
+    c->rgtc1s_block = rgtc1s_block;
+    c->rgtc1u_block = rgtc1u_block;
+    c->rgtc2s_block = rgtc2s_block;
+    c->rgtc2u_block = rgtc2u_block;
+    c->dxn3dc_block = dxn3dc_block;
+}
diff --git a/libavcodec/texturedsp.h b/libavcodec/texturedsp.h
new file mode 100644
index 0000000000..fcbe7a4d26
--- /dev/null
+++ b/libavcodec/texturedsp.h
@@ -0,0 +1,64 @@
+/*
+ * Texture block module
+ * Copyright (C) 2015 Vittorio Giovara <vittorio.giovara@gmail.com>
+ *
+ * This file is part of Libav.
+ *
+ * Libav 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.
+ *
+ * Libav 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 Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * Texture block (4x4) module
+ *
+ * References:
+ *   https://www.opengl.org/wiki/S3_Texture_Compression
+ *   https://www.opengl.org/wiki/Red_Green_Texture_Compression
+ *   https://msdn.microsoft.com/en-us/library/bb694531%28v=vs.85%29.aspx
+ *
+ * All functions return how much data has been written or read.
+ *
+ * Pixel input or output format is always AV_PIX_FMT_RGBA.
+ */
+
+#ifndef AVCODEC_TEXTUREDSP_H
+#define AVCODEC_TEXTUREDSP_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#define TEXTURE_BLOCK_W 4
+#define TEXTURE_BLOCK_H 4
+
+typedef struct TextureDSPContext {
+    int (*dxt1_block)  (uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxt1a_block) (uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxt2_block)  (uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxt3_block)  (uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxt4_block)  (uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxt5_block)  (uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxt5y_block) (uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxt5ys_block)(uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*rgtc1s_block)(uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*rgtc1u_block)(uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*rgtc2s_block)(uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*rgtc2u_block)(uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+    int (*dxn3dc_block)(uint8_t *dst, ptrdiff_t stride, const uint8_t *block);
+} TextureDSPContext;
+
+void ff_texturedsp_init(TextureDSPContext *c);
+void ff_texturedspenc_init(TextureDSPContext *c);
+
+#endif /* AVCODEC_TEXTUREDSP_H */
diff --git a/libavcodec/texturedspenc.c b/libavcodec/texturedspenc.c
new file mode 100644
index 0000000000..4a387c5edc
--- /dev/null
+++ b/libavcodec/texturedspenc.c
@@ -0,0 +1,659 @@
+/*
+ * Texture block compression
+ * Copyright (C) 2015 Vittorio Giovara <vittorio.giovara@gmail.com>
+ * Based on public domain code by Fabian Giesen, Sean Barrett and Yann Collet.
+ *
+ * This file is part of Libav
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <stddef.h>
+#include <stdint.h>
+
+#include "libavutil/attributes.h"
+#include "libavutil/common.h"
+#include "libavutil/intreadwrite.h"
+
+#include "texturedsp.h"
+
+const static uint8_t expand5[32] = {
+      0,   8,  16,  24,  33,  41,  49,  57,  66,  74,  82,  90,
+     99, 107, 115, 123, 132, 140, 148, 156, 165, 173, 181, 189,
+    198, 206, 214, 222, 231, 239, 247, 255,
+};
+
+const static uint8_t expand6[64] = {
+      0,   4,   8,  12,  16,  20,  24,  28,  32,  36,  40,  44,
+     48,  52,  56,  60,  65,  69,  73,  77,  81,  85,  89,  93,
+     97, 101, 105, 109, 113, 117, 121, 125, 130, 134, 138, 142,
+    146, 150, 154, 158, 162, 166, 170, 174, 178, 182, 186, 190,
+    195, 199, 203, 207, 211, 215, 219, 223, 227, 231, 235, 239,
+    243, 247, 251, 255,
+};
+
+const static uint8_t match5[256][2] = {
+    {  0,  0 }, {  0,  0 }, {  0,  1 }, {  0,  1 }, {  1,  0 }, {  1,  0 },
+    {  1,  0 }, {  1,  1 }, {  1,  1 }, {  2,  0 }, {  2,  0 }, {  0,  4 },
+    {  2,  1 }, {  2,  1 }, {  2,  1 }, {  3,  0 }, {  3,  0 }, {  3,  0 },
+    {  3,  1 }, {  1,  5 }, {  3,  2 }, {  3,  2 }, {  4,  0 }, {  4,  0 },
+    {  4,  1 }, {  4,  1 }, {  4,  2 }, {  4,  2 }, {  4,  2 }, {  3,  5 },
+    {  5,  1 }, {  5,  1 }, {  5,  2 }, {  4,  4 }, {  5,  3 }, {  5,  3 },
+    {  5,  3 }, {  6,  2 }, {  6,  2 }, {  6,  2 }, {  6,  3 }, {  5,  5 },
+    {  6,  4 }, {  6,  4 }, {  4,  8 }, {  7,  3 }, {  7,  3 }, {  7,  3 },
+    {  7,  4 }, {  7,  4 }, {  7,  4 }, {  7,  5 }, {  5,  9 }, {  7,  6 },
+    {  7,  6 }, {  8,  4 }, {  8,  4 }, {  8,  5 }, {  8,  5 }, {  8,  6 },
+    {  8,  6 }, {  8,  6 }, {  7,  9 }, {  9,  5 }, {  9,  5 }, {  9,  6 },
+    {  8,  8 }, {  9,  7 }, {  9,  7 }, {  9,  7 }, { 10,  6 }, { 10,  6 },
+    { 10,  6 }, { 10,  7 }, {  9,  9 }, { 10,  8 }, { 10,  8 }, {  8, 12 },
+    { 11,  7 }, { 11,  7 }, { 11,  7 }, { 11,  8 }, { 11,  8 }, { 11,  8 },
+    { 11,  9 }, {  9, 13 }, { 11, 10 }, { 11, 10 }, { 12,  8 }, { 12,  8 },
+    { 12,  9 }, { 12,  9 }, { 12, 10 }, { 12, 10 }, { 12, 10 }, { 11, 13 },
+    { 13,  9 }, { 13,  9 }, { 13, 10 }, { 12, 12 }, { 13, 11 }, { 13, 11 },
+    { 13, 11 }, { 14, 10 }, { 14, 10 }, { 14, 10 }, { 14, 11 }, { 13, 13 },
+    { 14, 12 }, { 14, 12 }, { 12, 16 }, { 15, 11 }, { 15, 11 }, { 15, 11 },
+    { 15, 12 }, { 15, 12 }, { 15, 12 }, { 15, 13 }, { 13, 17 }, { 15, 14 },
+    { 15, 14 }, { 16, 12 }, { 16, 12 }, { 16, 13 }, { 16, 13 }, { 16, 14 },
+    { 16, 14 }, { 16, 14 }, { 15, 17 }, { 17, 13 }, { 17, 13 }, { 17, 14 },
+    { 16, 16 }, { 17, 15 }, { 17, 15 }, { 17, 15 }, { 18, 14 }, { 18, 14 },
+    { 18, 14 }, { 18, 15 }, { 17, 17 }, { 18, 16 }, { 18, 16 }, { 16, 20 },
+    { 19, 15 }, { 19, 15 }, { 19, 15 }, { 19, 16 }, { 19, 16 }, { 19, 16 },
+    { 19, 17 }, { 17, 21 }, { 19, 18 }, { 19, 18 }, { 20, 16 }, { 20, 16 },
+    { 20, 17 }, { 20, 17 }, { 20, 18 }, { 20, 18 }, { 20, 18 }, { 19, 21 },
+    { 21, 17 }, { 21, 17 }, { 21, 18 }, { 20, 20 }, { 21, 19 }, { 21, 19 },
+    { 21, 19 }, { 22, 18 }, { 22, 18 }, { 22, 18 }, { 22, 19 }, { 21, 21 },
+    { 22, 20 }, { 22, 20 }, { 20, 24 }, { 23, 19 }, { 23, 19 }, { 23, 19 },
+    { 23, 20 }, { 23, 20 }, { 23, 20 }, { 23, 21 }, { 21, 25 }, { 23, 22 },
+    { 23, 22 }, { 24, 20 }, { 24, 20 }, { 24, 21 }, { 24, 21 }, { 24, 22 },
+    { 24, 22 }, { 24, 22 }, { 23, 25 }, { 25, 21 }, { 25, 21 }, { 25, 22 },
+    { 24, 24 }, { 25, 23 }, { 25, 23 }, { 25, 23 }, { 26, 22 }, { 26, 22 },
+    { 26, 22 }, { 26, 23 }, { 25, 25 }, { 26, 24 }, { 26, 24 }, { 24, 28 },
+    { 27, 23 }, { 27, 23 }, { 27, 23 }, { 27, 24 }, { 27, 24 }, { 27, 24 },
+    { 27, 25 }, { 25, 29 }, { 27, 26 }, { 27, 26 }, { 28, 24 }, { 28, 24 },
+    { 28, 25 }, { 28, 25 }, { 28, 26 }, { 28, 26 }, { 28, 26 }, { 27, 29 },
+    { 29, 25 }, { 29, 25 }, { 29, 26 }, { 28, 28 }, { 29, 27 }, { 29, 27 },
+    { 29, 27 }, { 30, 26 }, { 30, 26 }, { 30, 26 }, { 30, 27 }, { 29, 29 },
+    { 30, 28 }, { 30, 28 }, { 30, 28 }, { 31, 27 }, { 31, 27 }, { 31, 27 },
+    { 31, 28 }, { 31, 28 }, { 31, 28 }, { 31, 29 }, { 31, 29 }, { 31, 30 },
+    { 31, 30 }, { 31, 30 }, { 31, 31 }, { 31, 31 },
+};
+
+const static uint8_t match6[256][2] = {
+    {  0,  0 }, {  0,  1 }, {  1,  0 }, {  1,  0 }, {  1,  1 }, {  2,  0 },
+    {  2,  1 }, {  3,  0 }, {  3,  0 }, {  3,  1 }, {  4,  0 }, {  4,  0 },
+    {  4,  1 }, {  5,  0 }, {  5,  1 }, {  6,  0 }, {  6,  0 }, {  6,  1 },
+    {  7,  0 }, {  7,  0 }, {  7,  1 }, {  8,  0 }, {  8,  1 }, {  8,  1 },
+    {  8,  2 }, {  9,  1 }, {  9,  2 }, {  9,  2 }, {  9,  3 }, { 10,  2 },
+    { 10,  3 }, { 10,  3 }, { 10,  4 }, { 11,  3 }, { 11,  4 }, { 11,  4 },
+    { 11,  5 }, { 12,  4 }, { 12,  5 }, { 12,  5 }, { 12,  6 }, { 13,  5 },
+    { 13,  6 }, {  8, 16 }, { 13,  7 }, { 14,  6 }, { 14,  7 }, {  9, 17 },
+    { 14,  8 }, { 15,  7 }, { 15,  8 }, { 11, 16 }, { 15,  9 }, { 15, 10 },
+    { 16,  8 }, { 16,  9 }, { 16, 10 }, { 15, 13 }, { 17,  9 }, { 17, 10 },
+    { 17, 11 }, { 15, 16 }, { 18, 10 }, { 18, 11 }, { 18, 12 }, { 16, 16 },
+    { 19, 11 }, { 19, 12 }, { 19, 13 }, { 17, 17 }, { 20, 12 }, { 20, 13 },
+    { 20, 14 }, { 19, 16 }, { 21, 13 }, { 21, 14 }, { 21, 15 }, { 20, 17 },
+    { 22, 14 }, { 22, 15 }, { 25, 10 }, { 22, 16 }, { 23, 15 }, { 23, 16 },
+    { 26, 11 }, { 23, 17 }, { 24, 16 }, { 24, 17 }, { 27, 12 }, { 24, 18 },
+    { 25, 17 }, { 25, 18 }, { 28, 13 }, { 25, 19 }, { 26, 18 }, { 26, 19 },
+    { 29, 14 }, { 26, 20 }, { 27, 19 }, { 27, 20 }, { 30, 15 }, { 27, 21 },
+    { 28, 20 }, { 28, 21 }, { 28, 21 }, { 28, 22 }, { 29, 21 }, { 29, 22 },
+    { 24, 32 }, { 29, 23 }, { 30, 22 }, { 30, 23 }, { 25, 33 }, { 30, 24 },
+    { 31, 23 }, { 31, 24 }, { 27, 32 }, { 31, 25 }, { 31, 26 }, { 32, 24 },
+    { 32, 25 }, { 32, 26 }, { 31, 29 }, { 33, 25 }, { 33, 26 }, { 33, 27 },
+    { 31, 32 }, { 34, 26 }, { 34, 27 }, { 34, 28 }, { 32, 32 }, { 35, 27 },
+    { 35, 28 }, { 35, 29 }, { 33, 33 }, { 36, 28 }, { 36, 29 }, { 36, 30 },
+    { 35, 32 }, { 37, 29 }, { 37, 30 }, { 37, 31 }, { 36, 33 }, { 38, 30 },
+    { 38, 31 }, { 41, 26 }, { 38, 32 }, { 39, 31 }, { 39, 32 }, { 42, 27 },
+    { 39, 33 }, { 40, 32 }, { 40, 33 }, { 43, 28 }, { 40, 34 }, { 41, 33 },
+    { 41, 34 }, { 44, 29 }, { 41, 35 }, { 42, 34 }, { 42, 35 }, { 45, 30 },
+    { 42, 36 }, { 43, 35 }, { 43, 36 }, { 46, 31 }, { 43, 37 }, { 44, 36 },
+    { 44, 37 }, { 44, 37 }, { 44, 38 }, { 45, 37 }, { 45, 38 }, { 40, 48 },
+    { 45, 39 }, { 46, 38 }, { 46, 39 }, { 41, 49 }, { 46, 40 }, { 47, 39 },
+    { 47, 40 }, { 43, 48 }, { 47, 41 }, { 47, 42 }, { 48, 40 }, { 48, 41 },
+    { 48, 42 }, { 47, 45 }, { 49, 41 }, { 49, 42 }, { 49, 43 }, { 47, 48 },
+    { 50, 42 }, { 50, 43 }, { 50, 44 }, { 48, 48 }, { 51, 43 }, { 51, 44 },
+    { 51, 45 }, { 49, 49 }, { 52, 44 }, { 52, 45 }, { 52, 46 }, { 51, 48 },
+    { 53, 45 }, { 53, 46 }, { 53, 47 }, { 52, 49 }, { 54, 46 }, { 54, 47 },
+    { 57, 42 }, { 54, 48 }, { 55, 47 }, { 55, 48 }, { 58, 43 }, { 55, 49 },
+    { 56, 48 }, { 56, 49 }, { 59, 44 }, { 56, 50 }, { 57, 49 }, { 57, 50 },
+    { 60, 45 }, { 57, 51 }, { 58, 50 }, { 58, 51 }, { 61, 46 }, { 58, 52 },
+    { 59, 51 }, { 59, 52 }, { 62, 47 }, { 59, 53 }, { 60, 52 }, { 60, 53 },
+    { 60, 53 }, { 60, 54 }, { 61, 53 }, { 61, 54 }, { 61, 54 }, { 61, 55 },
+    { 62, 54 }, { 62, 55 }, { 62, 55 }, { 62, 56 }, { 63, 55 }, { 63, 56 },
+    { 63, 56 }, { 63, 57 }, { 63, 58 }, { 63, 59 }, { 63, 59 }, { 63, 60 },
+    { 63, 61 }, { 63, 62 }, { 63, 62 }, { 63, 63 },
+};
+
+/* Multiplication over 8 bit emulation */
+#define mul8(a, b) (a * b + 128 + ((a * b + 128) >> 8)) >> 8
+
+/* Conversion from rgb24 to rgb565 */
+#define rgb2rgb565(r, g, b) \
+    (mul8(r, 31) << 11) | (mul8(g, 63) << 5) | (mul8(b, 31) << 0)
+
+/* Linear interpolation at 1/3 point between a and b */
+#define lerp13(a, b) (2 * a + b) / 3
+
+/* Linear interpolation on an RGB pixel */
+static inline void lerp13rgb(uint8_t *out, uint8_t *p1, uint8_t *p2)
+{
+    out[0] = lerp13(p1[0], p2[0]);
+    out[1] = lerp13(p1[1], p2[1]);
+    out[2] = lerp13(p1[2], p2[2]);
+}
+
+/* Conversion from rgb565 to rgb24 */
+static inline void rgb5652rgb(uint8_t *out, uint16_t v)
+{
+    int rv = (v & 0xf800) >> 11;
+    int gv = (v & 0x07e0) >> 5;
+    int bv = (v & 0x001f) >> 0;
+
+    out[0] = expand5[rv];
+    out[1] = expand6[gv];
+    out[2] = expand5[bv];
+    out[3] = 0;
+}
+
+/* Color matching function */
+static unsigned int match_colors(const uint8_t *block, ptrdiff_t stride,
+                                 uint16_t c0, uint16_t c1)
+{
+    uint32_t mask = 0;
+    int dirr, dirg, dirb;
+    int dots[16];
+    int stops[4];
+    int x, y, k = 0;
+    int c0_point, half_point, c3_point;
+    uint8_t color[16];
+    const int indexMap[8] = {
+        0 << 30, 2 << 30, 0 << 30, 2 << 30,
+        3 << 30, 3 << 30, 1 << 30, 1 << 30,
+    };
+
+    /* Fill color and compute direction for each component */
+    rgb5652rgb(color + 0, c0);
+    rgb5652rgb(color + 4, c1);
+    lerp13rgb(color + 8, color + 0, color + 4);
+    lerp13rgb(color + 12, color + 4, color + 0);
+
+    dirr = color[0 * 4 + 0] - color[1 * 4 + 0];
+    dirg = color[0 * 4 + 1] - color[1 * 4 + 1];
+    dirb = color[0 * 4 + 2] - color[1 * 4 + 2];
+
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++)
+            dots[k++] = block[0 + x * 4 + y * stride] * dirr +
+                        block[1 + x * 4 + y * stride] * dirg +
+                        block[2 + x * 4 + y * stride] * dirb;
+
+        stops[y] = color[0 + y * 4] * dirr +
+                   color[1 + y * 4] * dirg +
+                   color[2 + y * 4] * dirb;
+    }
+
+    /* Think of the colors as arranged on a line; project point onto that line,
+     * then choose next color out of available ones. we compute the crossover
+     * points for 'best color in top half'/'best in bottom half' and then
+     * the same inside that subinterval.
+     *
+     * Relying on this 1d approximation isn't always optimal in terms of
+     * euclidean distance, but it's very close and a lot faster.
+     *
+     * http://cbloomrants.blogspot.com/2008/12/12-08-08-dxtc-summary.html */
+    c0_point   = (stops[1] + stops[3]) >> 1;
+    half_point = (stops[3] + stops[2]) >> 1;
+    c3_point   = (stops[2] + stops[0]) >> 1;
+
+    for (x = 0; x < 16; x++) {
+        int dot  = dots[x];
+        int bits = (dot < half_point ? 4 : 0) |
+                   (dot < c0_point   ? 2 : 0) |
+                   (dot < c3_point   ? 1 : 0);
+
+        mask >>= 2;
+        mask  |= indexMap[bits];
+    }
+
+    return mask;
+}
+
+/* Color optimization function */
+static void optimize_colors(const uint8_t *block, ptrdiff_t stride,
+                            uint16_t *pmax16, uint16_t *pmin16)
+{
+    const uint8_t *minp;
+    const uint8_t *maxp;
+    const int iter_power = 4;
+    double magn;
+    int v_r, v_g, v_b;
+    float covf[6], vfr, vfg, vfb;
+    int mind, maxd;
+    int cov[6] = { 0 };
+    int mu[3], min[3], max[3];
+    int ch, iter, x, y;
+
+    /* Determine color distribution */
+    for (ch = 0; ch < 3; ch++) {
+        const uint8_t *bp = &block[ch];
+        int muv, minv, maxv;
+
+        muv = minv = maxv = bp[0];
+        for (y = 0; y < 4; y++) {
+            for (x = 4; x < 4; x += 4) {
+                muv += bp[x * 4 + y * stride];
+                if (bp[x] < minv)
+                    minv = bp[x * 4 + y * stride];
+                else if (bp[x] > maxv)
+                    maxv = bp[x * 4 + y * stride];
+            }
+        }
+
+        mu[ch]  = (muv + 8) >> 4;
+        min[ch] = minv;
+        max[ch] = maxv;
+    }
+
+    /* Determine covariance matrix */
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            int r = block[x * 4 + stride * y + 0] - mu[0];
+            int g = block[x * 4 + stride * y + 1] - mu[1];
+            int b = block[x * 4 + stride * y + 2] - mu[2];
+
+            cov[0] += r * r;
+            cov[1] += r * g;
+            cov[2] += r * b;
+            cov[3] += g * g;
+            cov[4] += g * b;
+            cov[5] += b * b;
+        }
+    }
+
+    /* Convert covariance matrix to float, find principal axis via power iter */
+    for (x = 0; x < 6; x++)
+        covf[x] = cov[x] / 255.0f;
+
+    vfr = (float) (max[0] - min[0]);
+    vfg = (float) (max[1] - min[1]);
+    vfb = (float) (max[2] - min[2]);
+
+    for (iter = 0; iter < iter_power; iter++) {
+        float r = vfr * covf[0] + vfg * covf[1] + vfb * covf[2];
+        float g = vfr * covf[1] + vfg * covf[3] + vfb * covf[4];
+        float b = vfr * covf[2] + vfg * covf[4] + vfb * covf[5];
+
+        vfr = r;
+        vfg = g;
+        vfb = b;
+    }
+
+    magn = fabs(vfr);
+    if (fabs(vfg) > magn)
+        magn = fabs(vfg);
+    if (fabs(vfb) > magn)
+        magn = fabs(vfb);
+
+    /* if magnitudo is too small, default to luminance */
+    if (magn < 4.0f) {
+        /* JPEG YCbCr luma coefs, scaled by 1000 */
+        v_r = 299;
+        v_g = 587;
+        v_b = 114;
+    } else {
+        magn = 512.0 / magn;
+        v_r  = (int) (vfr * magn);
+        v_g  = (int) (vfg * magn);
+        v_b  = (int) (vfb * magn);
+    }
+
+    /* Pick colors at extreme points */
+    mind = maxd = block[0] * v_r + block[1] * v_g + block[2] * v_b;
+    minp = maxp = block;
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            int dot = block[x * 4 + y * stride + 0] * v_r +
+                      block[x * 4 + y * stride + 1] * v_g +
+                      block[x * 4 + y * stride + 2] * v_b;
+
+            if (dot < mind) {
+                mind = dot;
+                minp = block + x * 4 + y * stride;
+            } else if (dot > maxd) {
+                maxd = dot;
+                maxp = block + x * 4 + y * stride;
+            }
+        }
+    }
+
+    *pmax16 = rgb2rgb565(maxp[0], maxp[1], maxp[2]);
+    *pmin16 = rgb2rgb565(minp[0], minp[1], minp[2]);
+}
+
+/* Try to optimize colors to suit block contents better, by solving
+ * a least squares system via normal equations + Cramer's rule. */
+static int refine_colors(const uint8_t *block, ptrdiff_t stride,
+                         uint16_t *pmax16, uint16_t *pmin16, uint32_t mask)
+{
+    uint32_t cm = mask;
+    uint16_t oldMin = *pmin16;
+    uint16_t oldMax = *pmax16;
+    uint16_t min16, max16;
+    int x, y;
+
+    /* Additional magic to save a lot of multiplies in the accumulating loop.
+     * The tables contain precomputed products of weights for least squares
+     * system, accumulated inside one 32-bit register */
+    const int w1tab[4] = { 3, 0, 2, 1 };
+    const int prods[4] = { 0x090000, 0x000900, 0x040102, 0x010402 };
+
+    /* Check if all pixels have the same index */
+    if (mask ^ (mask << 2) < 4) {
+        /* If so, linear system would be singular; solve using optimal
+         * single-color match on average color. */
+        int r = 8, g = 8, b = 8;
+        for (y = 0; y < 4; y++) {
+            for (x = 0; x < 4; x++) {
+                r += block[0 + x * 4 + y * stride];
+                g += block[1 + x * 4 + y * stride];
+                b += block[2 + x * 4 + y * stride];
+            }
+        }
+
+        r >>= 4;
+        g >>= 4;
+        b >>= 4;
+
+        max16 = (match5[r][0] << 11) | (match6[g][0] << 5) | match5[b][0];
+        min16 = (match5[r][1] << 11) | (match6[g][1] << 5) | match5[b][1];
+    } else {
+        float fr, fg, fb;
+        int at1_r = 0, at1_g = 0, at1_b = 0;
+        int at2_r = 0, at2_g = 0, at2_b = 0;
+        int akku = 0;
+        int xx, xy, yy;
+
+        for (y = 0; y < 4; y++) {
+            for (x = 0; x < 4; x++) {
+                int step = cm & 3;
+                int w1 = w1tab[step];
+                int r = block[0 + x * 4 + y * stride];
+                int g = block[1 + x * 4 + y * stride];
+                int b = block[2 + x * 4 + y * stride];
+
+                akku  += prods[step];
+                at1_r += w1 * r;
+                at1_g += w1 * g;
+                at1_b += w1 * b;
+                at2_r += r;
+                at2_g += g;
+                at2_b += b;
+
+                cm >>= 2;
+            }
+        }
+
+        at2_r = 3 * at2_r - at1_r;
+        at2_g = 3 * at2_g - at1_g;
+        at2_b = 3 * at2_b - at1_b;
+
+        /* Extract solutions and decide solvability */
+        xx =  akku >> 16;
+        yy = (akku >>  8) & 0xFF;
+        xy = (akku >>  0) & 0xFF;
+
+        fr = 3.0f * 31.0f / 255.0f / (xx * yy - xy * xy);
+        fg = fr * 63.0f / 31.0f;
+        fb = fr;
+
+        /* Solve */
+        max16  = av_clip_uintp2((at1_r * yy - at2_r * xy) * fr + 0.5f, 5) << 11;
+        max16 |= av_clip_uintp2((at1_g * yy - at2_g * xy) * fg + 0.5f, 6) <<  5;
+        max16 |= av_clip_uintp2((at1_b * yy - at2_b * xy) * fb + 0.5f, 5) <<  0;
+
+        min16  = av_clip_uintp2((at2_r * xx - at1_r * xy) * fr + 0.5f, 5) << 11;
+        min16 |= av_clip_uintp2((at2_g * xx - at1_g * xy) * fg + 0.5f, 6) <<  5;
+        min16 |= av_clip_uintp2((at2_b * xx - at1_b * xy) * fb + 0.5f, 5) <<  0;
+    }
+
+    *pmin16 = min16;
+    *pmax16 = max16;
+    return oldMin != min16 || oldMax != max16;
+}
+
+/* Check if input block is a constant color */
+static int constant_color(const uint8_t *block, ptrdiff_t stride)
+{
+    int x, y;
+    uint32_t first = AV_RL32(block);
+
+    for (y = 0; y < 4; y++)
+        for (x = 0; x < 4; x++)
+            if (first != AV_RL32(block + x * 4 + y * stride))
+                return 0;
+    return 1;
+}
+
+/* Main color compression function */
+static void compress_color(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    uint32_t mask;
+    uint16_t max16, min16;
+    int constant = constant_color(block, stride);
+
+    /* Constant color will load values from tables */
+    if (constant) {
+        int r = block[0];
+        int g = block[1];
+        int b = block[2];
+        mask  = 0xAAAAAAAA;
+        max16 = (match5[r][0] << 11) | (match6[g][0] << 5) | match5[b][0];
+        min16 = (match5[r][1] << 11) | (match6[g][1] << 5) | match5[b][1];
+    } else {
+        int refine;
+
+        /* Otherwise find pca and map along principal axis */
+        optimize_colors(block, stride, &max16, &min16);
+        if (max16 != min16)
+            mask = match_colors(block, stride, max16, min16);
+        else
+            mask = 0;
+
+        /* One pass refinement */
+        refine  = refine_colors(block, stride, &max16, &min16, mask);
+        if (refine) {
+            if (max16 != min16)
+                mask = match_colors(block, stride, max16, min16);
+            else
+                mask = 0;
+        }
+    }
+
+    /* Finally write the color block */
+    if (max16 < min16) {
+        FFSWAP(uint16_t, min16, max16);
+        mask ^= 0x55555555;
+    }
+
+    AV_WL16(dst + 0, max16);
+    AV_WL16(dst + 2, min16);
+    AV_WL32(dst + 4, mask);
+}
+
+/* Alpha compression function */
+static void compress_alpha(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    int x, y;
+    int dist, bias, dist4, dist2;
+    int mn, mx;
+    int bits = 0;
+    int mask = 0;
+
+    memset(dst, 0, 8);
+
+    /* Find min/max color */
+    mn = mx = block[3];
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            int val = block[3 + x * 4 + y * stride];
+            if (val < mn)
+                mn = val;
+            else if (val > mx)
+                mx = val;
+        }
+    }
+
+    /* Encode them */
+    dst[0] = (uint8_t) mx;
+    dst[1] = (uint8_t) mn;
+    dst += 2;
+
+    /* Mono-alpha shortcut */
+    if (mn == mx)
+        return;
+
+    /* Determine bias and emit color indices.
+     * Given the choice of mx/mn, these indices are optimal:
+     * fgiesen.wordpress.com/2009/12/15/dxt5-alpha-block-index-determination */
+    dist = mx - mn;
+
+    dist4 = dist * 4;
+    dist2 = dist * 2;
+    if (dist < 8)
+        bias = dist - 1 - mn * 7;
+    else
+        bias = dist / 2 + 2 - mn * 7;
+
+    for (y = 0; y < 4; y++) {
+        for (x = 0; x < 4; x++) {
+            int alp = block[3 + x * 4 + y * stride] * 7 + bias;
+            int ind, tmp;
+
+            /* This is a "linear scale" lerp factor between 0 (val=min)
+             * and 7 (val=max) to select index. */
+            tmp  = (alp >= dist4) ? -1 : 0;
+            ind  = tmp & 4;
+            alp -= dist4 & tmp;
+            tmp  = (alp >= dist2) ? -1 : 0;
+            ind += tmp & 2;
+            alp -= dist2 & tmp;
+            ind += (alp >= dist);
+
+            /* Turn linear scale into DXT index (0/1 are extreme points) */
+            ind  = -ind & 7;
+            ind ^= (2 > ind);
+
+            /* Write index */
+            mask |= ind << bits;
+            bits += 3;
+            if (bits >= 8) {
+                *dst++ = mask;
+                mask >>= 8;
+                bits  -= 8;
+            }
+        }
+    }
+}
+
+/**
+ * Convert a RGBA buffer to unscaled YCoCg.
+ * Scale is usually introduced to avoid banding over a certain range of colors,
+ * but this version of the algorithm does not introduce it as much as other
+ * implementations, allowing for a simpler and faster conversion.
+ */
+static void rgba2ycocg(uint8_t *dst, const uint8_t *pixel)
+{
+    int r =  pixel[0];
+    int g = (pixel[1] + 1) >> 1;
+    int b =  pixel[2];
+    int t = (2 + r + b) >> 2;
+
+    int y  = av_clip_uint8(g + t);
+    int co = av_clip_uint8(128 + ((r - b + 1) >> 1));
+    int cg = av_clip_uint8(128 + g - t);
+
+    dst[0] = (uint8_t) co;
+    dst[1] = (uint8_t) cg;
+    dst[2] = 0;
+    dst[3] = (uint8_t) y;
+}
+
+/**
+ * Compress one block of RGBA pixels in a DXT1 texture and store the
+ * resulting bytes in 'dst'. Alpha is not preserved.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to compress.
+ * @return how much texture data has been written.
+ */
+static int dxt1_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    compress_color(dst, stride, block);
+
+    return 8;
+}
+
+/**
+ * Compress one block of RGBA pixels in a DXT5 texture and store the
+ * resulting bytes in 'dst'. Alpha is preserved.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to compress.
+ * @return how much texture data has been written.
+ */
+static int dxt5_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    compress_alpha(dst, stride, block);
+    compress_color(dst + 8, stride, block);
+
+    return 16;
+}
+
+/**
+ * Compress one block of RGBA pixels in a DXT5-YCoCg texture and store the
+ * resulting bytes in 'dst'. Alpha is not preserved.
+ *
+ * @param dst    output buffer.
+ * @param stride scanline in bytes.
+ * @param block  block to compress.
+ * @return how much texture data has been written.
+ */
+static int dxt5ys_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
+{
+    int x, y;
+    uint8_t reorder[64];
+
+    /* Reorder the components and then run a normal DXT5 compression. */
+    for (y = 0; y < 4; y++)
+        for (x = 0; x < 4; x++)
+            rgba2ycocg(reorder + x * 4 + y * 16, block + x * 4 + y * stride);
+
+    compress_alpha(dst + 0, 16, reorder);
+    compress_color(dst + 8, 16, reorder);
+
+    return 16;
+}
+
+av_cold void ff_texturedspenc_init(TextureDSPContext *c)
+{
+    c->dxt1_block   = dxt1_block;
+    c->dxt5_block   = dxt5_block;
+    c->dxt5ys_block = dxt5ys_block;
+}