Evas: Complete DDS loader with direct S3TC data load

This requires block flip (could be repeat, but flip is just as fast).

3 files changed, 236 insertions, 16 deletions

diff --git a/src/modules/evas/loaders/dds/evas_image_load_dds.c b/src/modules/evas/loaders/dds/evas_image_load_dds.c
index dc4fd629ba..cd3f87fb66 100644
--- a/src/modules/evas/loaders/dds/evas_image_load_dds.c
+++ b/src/modules/evas/loaders/dds/evas_image_load_dds.c
@@ -214,7 +214,7 @@ evas_image_load_file_head_dds(void *loader_data,
    height = _dword_read(&m);
    width = _dword_read(&m);
    pitchOrLinearSize = _dword_read(&m);
-   if (!width || !height || (width & 0x3) || (height & 0x3))
+   if (!width || !height)
      FAIL();
 
    // Skip depth & mipmap count + reserved[11]
@@ -302,6 +302,10 @@ evas_image_load_file_head_dds(void *loader_data,
 
    prop->h = height;
    prop->w = width;
+   prop->borders.l = 4;
+   prop->borders.t = 4;
+   prop->borders.r = 4 - (prop->w & 0x3);
+   prop->borders.b = 4 - (prop->h & 0x3);
    *error = EVAS_LOAD_ERROR_NONE;
 
 on_error:
@@ -309,6 +313,134 @@ on_error:
    return (*error == EVAS_LOAD_ERROR_NONE);
 }
 
+static Eina_Bool
+_dds_data_load(Evas_Loader_Internal *loader, Evas_Image_Property *prop,
+               unsigned char *map, void *pixels, int *error)
+{
+   const unsigned char *src;
+   int bsize = 16, srcstride, dststride, w, h;
+   unsigned char *dst;
+
+   void (* flip) (unsigned char *, const unsigned char *, int) = NULL;
+
+   *error = EVAS_LOAD_ERROR_GENERIC;
+
+   if (loader->format != prop->cspace)
+     FAIL();
+
+   switch (loader->format)
+     {
+      case EVAS_COLORSPACE_RGB_S3TC_DXT1:
+      case EVAS_COLORSPACE_RGBA_S3TC_DXT1:
+        flip = s3tc_encode_dxt1_flip;
+        bsize = 8;
+        break;
+      case EVAS_COLORSPACE_RGBA_S3TC_DXT2:
+        flip = s3tc_encode_dxt2_rgba_flip;
+        bsize = 16;
+        break;
+      case EVAS_COLORSPACE_RGBA_S3TC_DXT3:
+        flip = s3tc_encode_dxt3_rgba_flip;
+        bsize = 16;
+        break;
+      case EVAS_COLORSPACE_RGBA_S3TC_DXT4:
+        flip = s3tc_encode_dxt4_rgba_flip;
+        bsize = 16;
+        break;
+      case EVAS_COLORSPACE_RGBA_S3TC_DXT5:
+        flip = s3tc_encode_dxt5_rgba_flip;
+        bsize = 16;
+        break;
+      default: FAIL();
+     }
+
+   src = map + DDS_HEADER_SIZE;
+   w = prop->w;
+   h = prop->h;
+   srcstride = ((prop->w + 3) / 4) * bsize;
+   dststride = ((prop->w + prop->borders.l + prop->borders.r) / 4) * bsize;
+
+   // asserts
+   EINA_SAFETY_ON_FALSE_GOTO(prop->borders.l == 4, on_error);
+   EINA_SAFETY_ON_FALSE_GOTO(prop->borders.t == 4, on_error);
+   EINA_SAFETY_ON_FALSE_GOTO(prop->borders.r == (4 - (w & 0x3)), on_error);
+   EINA_SAFETY_ON_FALSE_GOTO(prop->borders.b == (4 - (h & 0x3)), on_error);
+
+   if (eina_file_size_get(loader->f) <
+       (size_t) (DDS_HEADER_SIZE + srcstride * h / 4))
+     {
+        *error = EVAS_LOAD_ERROR_CORRUPT_FILE;
+        goto on_error;
+     }
+
+   // First, copy the real data
+   for (int y = 0; y < h; y += 4)
+     {
+        dst = ((unsigned char *) pixels) + ((y / 4) + 1) * dststride + bsize;
+        memcpy(dst, src, srcstride);
+        src += srcstride;
+     }
+   // Top
+   for (int x = 0; x < w; x += 4)
+     {
+        src = map + DDS_HEADER_SIZE + (x / 4) * bsize;
+        dst = ((unsigned char *) pixels) + ((x / 4) + 1) * bsize;
+        flip(dst, src, EINA_TRUE);
+     }
+   // Left
+   for (int y = 0; y < h; y += 4)
+     {
+        src = map + DDS_HEADER_SIZE + (y / 4) * srcstride;
+        dst = ((unsigned char *) pixels) + ((y / 4) + 1) * dststride;
+        flip(dst, src, EINA_FALSE);
+     }
+   // Top-left
+   dst = pixels;
+   src = dst + bsize;
+   flip(dst, src, EINA_FALSE);
+   // Right
+   if ((prop->w & 0x3) == 0)
+     {
+        for (int y = 0; y < h; y += 4)
+          {
+             src = map + DDS_HEADER_SIZE + ((y / 4) + 1) * srcstride - bsize;
+             dst = ((unsigned char *) pixels) + ((y / 4) + 2) * dststride - bsize;
+             flip(dst, src, EINA_FALSE);
+          }
+        // Top-right
+        dst = ((unsigned char *) pixels) + dststride - bsize;
+        src = dst - bsize;
+        flip(dst, src, EINA_FALSE);
+     }
+   // Bottom
+   if ((prop->h & 0x3) == 0)
+     {
+        for (int x = 0; x < w; x += 4)
+          {
+             src = map + DDS_HEADER_SIZE + ((h / 4) - 1) * srcstride + (x / 4) * bsize;
+             dst = ((unsigned char *) pixels) + ((h / 4) + 1) * dststride + ((x / 4) + 1) * bsize;
+             flip(dst, src, EINA_TRUE);
+          }
+        // Bottom-left
+        dst = ((unsigned char *) pixels) + ((h / 4) + 1) * dststride;
+        src = dst + bsize;
+        flip(dst, src, EINA_FALSE);
+        if ((prop->w & 0x3) == 0)
+          {
+             // Bottom-right
+             dst = ((unsigned char *) pixels) + ((h / 4) + 2) * dststride - bsize;
+             src = dst - bsize;
+             flip(dst, src, EINA_FALSE);
+          }
+     }
+
+   *error = EVAS_LOAD_ERROR_NONE;
+
+on_error:
+   eina_file_map_free(loader->f, (void *) map);
+   return (*error == EVAS_LOAD_ERROR_NONE);
+}
+
 Eina_Bool
 evas_image_load_file_data_dds(void *loader_data,
                               Evas_Image_Property *prop,
@@ -332,14 +464,7 @@ evas_image_load_file_data_dds(void *loader_data,
      FAIL();
 
    if (prop->cspace != EVAS_COLORSPACE_ARGB8888)
-     {
-        if (loader->format != prop->cspace)
-          FAIL();
-        memcpy(pixels, src, loader->data_size);
-        *error = EVAS_LOAD_ERROR_NONE;
-        eina_file_map_free(loader->f, (void *) src);
-        return EINA_TRUE;
-     }
+     return _dds_data_load(loader, prop, map, pixels, error);
 
    // Decode to BGRA
    switch (loader->format)
@@ -399,7 +524,7 @@ evas_image_load_file_data_dds(void *loader_data,
    *error = EVAS_LOAD_ERROR_NONE;
 
 on_error:
-   eina_file_map_free(loader->f, (void *) src);
+   eina_file_map_free(loader->f, (void *) map);
    return (*error == EVAS_LOAD_ERROR_NONE);
 }
 
diff --git a/src/modules/evas/loaders/dds/s3tc.h b/src/modules/evas/loaders/dds/s3tc.h
index 91f05f0b5a..3bc965db48 100644
--- a/src/modules/evas/loaders/dds/s3tc.h
+++ b/src/modules/evas/loaders/dds/s3tc.h
@@ -8,4 +8,10 @@ void s3tc_decode_dxt3_rgba(unsigned int *bgra, const unsigned char *s3tc);
 void s3tc_decode_dxt4_rgba(unsigned int *bgra, const unsigned char *s3tc);
 void s3tc_decode_dxt5_rgba(unsigned int *bgra, const unsigned char *s3tc);
 
+void s3tc_encode_dxt1_flip(unsigned char *dest, const unsigned char *orig, int vflip);
+void s3tc_encode_dxt2_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip);
+void s3tc_encode_dxt3_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip);
+void s3tc_encode_dxt4_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip);
+void s3tc_encode_dxt5_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip);
+
 #endif // EFL_S3TC_H
diff --git a/src/modules/evas/loaders/dds/s3tc_decoder.c b/src/modules/evas/loaders/dds/s3tc_decoder.c
index a3e09dc98e..f934fbd4a3 100644
--- a/src/modules/evas/loaders/dds/s3tc_decoder.c
+++ b/src/modules/evas/loaders/dds/s3tc_decoder.c
@@ -95,36 +95,125 @@ _decode_dxt_alpha(unsigned int *bgra, const unsigned char *s3tc)
      }
 }
 
-void s3tc_decode_dxt1_rgb(unsigned int *bgra, const unsigned char *s3tc)
+void
+s3tc_decode_dxt1_rgb(unsigned int *bgra, const unsigned char *s3tc)
 {
    _decode_dxt1_rgb(bgra, s3tc, 0xFF000000, EINA_TRUE, EINA_FALSE);
 }
 
-void s3tc_decode_dxt1_rgba(unsigned int *bgra, const unsigned char *s3tc)
+void
+s3tc_decode_dxt1_rgba(unsigned int *bgra, const unsigned char *s3tc)
 {
    _decode_dxt1_rgb(bgra, s3tc, 0xFF000000, EINA_TRUE, EINA_TRUE);
 }
 
-void s3tc_decode_dxt2_rgba(unsigned int *bgra, const unsigned char *s3tc)
+void
+s3tc_decode_dxt2_rgba(unsigned int *bgra, const unsigned char *s3tc)
 {
    _decode_dxt1_rgb(bgra, s3tc + 8, 0x0, EINA_FALSE, EINA_FALSE);
    _decode_alpha4(bgra, s3tc);
 }
 
-void s3tc_decode_dxt3_rgba(unsigned int *bgra, const unsigned char *s3tc)
+void
+s3tc_decode_dxt3_rgba(unsigned int *bgra, const unsigned char *s3tc)
 {
    _decode_dxt1_rgb(bgra, s3tc + 8, 0x0, EINA_FALSE, EINA_FALSE);
    _decode_alpha4(bgra, s3tc);
 }
 
-void s3tc_decode_dxt4_rgba(unsigned int *bgra, const unsigned char *s3tc)
+void
+s3tc_decode_dxt4_rgba(unsigned int *bgra, const unsigned char *s3tc)
 {
    _decode_dxt1_rgb(bgra, s3tc + 8, 0x0, EINA_FALSE, EINA_FALSE);
    _decode_dxt_alpha(bgra, s3tc);
 }
 
-void s3tc_decode_dxt5_rgba(unsigned int *bgra, const unsigned char *s3tc)
+void
+s3tc_decode_dxt5_rgba(unsigned int *bgra, const unsigned char *s3tc)
 {
    _decode_dxt1_rgb(bgra, s3tc + 8, 0x0, EINA_FALSE, EINA_FALSE);
    _decode_dxt_alpha(bgra, s3tc);
 }
+
+/* Fast re-encode functions to flip S3TC blocks */
+
+static inline unsigned char
+_byte_2222_flip(unsigned char src)
+{
+   return ((src & (0x3 << 6)) >> 6) |
+         ((src & (0x3 << 4)) >> 2) |
+         ((src & (0x3 << 2)) << 2) |
+         ((src & 0x3) << 6);
+}
+
+static inline unsigned char
+_byte_44_flip(unsigned char src)
+{
+   return ((src & 0xF0) >> 4) | ((src & 0x0F) << 4);
+}
+
+void
+s3tc_encode_dxt1_flip(unsigned char *dest, const unsigned char *orig, int vflip)
+{
+   dest[0] = orig[0];
+   dest[1] = orig[1];
+   dest[2] = orig[2];
+   dest[3] = orig[3];
+
+   if (vflip)
+     {
+        dest[4] = orig[7];
+        dest[5] = orig[6];
+        dest[6] = orig[5];
+        dest[7] = orig[4];
+     }
+   else
+     {
+        dest[4] = _byte_2222_flip(orig[4]);
+        dest[5] = _byte_2222_flip(orig[5]);
+        dest[6] = _byte_2222_flip(orig[6]);
+        dest[7] = _byte_2222_flip(orig[7]);
+     }
+}
+
+void
+s3tc_encode_dxt2_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip)
+{
+   if (vflip)
+     {
+        dest[0] = orig[6];
+        dest[1] = orig[7];
+        dest[2] = orig[4];
+        dest[3] = orig[5];
+        dest[4] = orig[2];
+        dest[5] = orig[3];
+        dest[6] = orig[0];
+        dest[7] = orig[1];
+     }
+   else
+     {
+        for (int k = 0; k < 8; k += 2)
+          {
+             dest[0+k] = _byte_44_flip(orig[1+k]);
+             dest[1+k] = _byte_44_flip(orig[0+k]);
+          }
+     }
+   s3tc_encode_dxt1_flip(dest + 8, orig + 8, vflip);
+}
+
+void
+s3tc_encode_dxt3_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip)
+{
+   s3tc_encode_dxt2_rgba_flip(dest, orig, vflip);
+}
+
+void s3tc_encode_dxt4_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip)
+{
+   s3tc_encode_dxt1_flip(dest, orig, vflip);
+   s3tc_encode_dxt1_flip(dest + 8, orig + 8, vflip);
+}
+
+void s3tc_encode_dxt5_rgba_flip(unsigned char *dest, const unsigned char *orig, int vflip)
+{
+   s3tc_encode_dxt4_rgba_flip(dest, orig, vflip);
+}