# HG changeset patch
# User melanson
# Date 1065890631 0
# Node ID e94e299aee4053393511a3f46628868cb0b744e2
# Parent  b1a0fc3ad09fe96aec9a3fc5ed1ddd26863c7a0e
optimised versions of the get_pixels, diff_pixels, and bswap_buf
functions and support for 4 pixel high blocks in the put_pixels8_xxx and
avg_pixels8_xxx functions, courtesy of Robin Kay (komadori
at gekkou.co.uk)

diff -r b1a0fc3ad09f -r e94e299aee40 mlib/dsputil_mlib.c
--- a/mlib/dsputil_mlib.c	Sat Oct 11 09:11:15 2003 +0000
+++ b/mlib/dsputil_mlib.c	Sat Oct 11 16:43:51 2003 +0000
@@ -23,196 +23,382 @@
 #include <mlib_types.h>
 #include <mlib_status.h>
 #include <mlib_sys.h>
+#include <mlib_algebra.h>
 #include <mlib_video.h>
 
-
-/* copy block, width 16 pixel, height 8/16 */
-
-static void put_pixels16_mlib (uint8_t * dest, const uint8_t * ref,
-			       int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoCopyRef_U8_U8_16x16(dest, (uint8_t *)ref, stride);
-    else
-	mlib_VideoCopyRef_U8_U8_16x8 (dest, (uint8_t *)ref, stride);
-}
+/* misc */
 
-static void put_pixels16_x2_mlib (uint8_t * dest, const uint8_t * ref,
-				  int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpX_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpX_U8_U8_16x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
-static void put_pixels16_y2_mlib (uint8_t * dest, const uint8_t * ref,
-				  int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpY_U8_U8_16x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
-static void put_pixels16_xy2_mlib(uint8_t * dest, const uint8_t * ref,
-				  int stride, int height)
+static void get_pixels_mlib(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
 {
-    assert(height == 16 || height == 8);
-    if (height == 16) 
-	mlib_VideoInterpXY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpXY_U8_U8_16x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
-
-/* copy block, width 8 pixel, height 8/16 */
-
-static void put_pixels8_mlib (uint8_t * dest, const uint8_t * ref,
-			      int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoCopyRef_U8_U8_8x16(dest, (uint8_t *)ref, stride);
-    else
-	mlib_VideoCopyRef_U8_U8_8x8 (dest, (uint8_t *)ref, stride);
-}
+  int i;
 
-static void put_pixels8_x2_mlib (uint8_t * dest, const uint8_t * ref,
-				 int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpX_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpX_U8_U8_8x8 (dest, (uint8_t *)ref, stride, stride);
-}
+  for (i=0;i<8;i++) {
+    mlib_VectorConvert_S16_U8_Mod((mlib_s16 *)block, (mlib_u8 *)pixels, 8);
 
-static void put_pixels8_y2_mlib (uint8_t * dest, const uint8_t * ref,
-				 int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpY_U8_U8_8x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
-static void put_pixels8_xy2_mlib(uint8_t * dest, const uint8_t * ref,
-				 int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16) 
-	mlib_VideoInterpXY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpXY_U8_U8_8x8 (dest, (uint8_t *)ref, stride, stride);
+    pixels += line_size;
+    block += 8;
+  }
 }
 
-
-/* average/merge dest+source block, width 16 pixel, height 8/16 */
-
-static void avg_pixels16_mlib (uint8_t * dest, const uint8_t * ref,
-			       int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoCopyRefAve_U8_U8_16x16(dest, (uint8_t *)ref, stride);
-    else
-	mlib_VideoCopyRefAve_U8_U8_16x8 (dest, (uint8_t *)ref, stride);
-}
-
-static void avg_pixels16_x2_mlib (uint8_t * dest, const uint8_t * ref,
-				  int stride, int height)
+static void diff_pixels_mlib(DCTELEM *restrict block, const uint8_t *s1, const uint8_t *s2, int line_size)
 {
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpAveX_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpAveX_U8_U8_16x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
-static void avg_pixels16_y2_mlib (uint8_t * dest, const uint8_t * ref,
-				  int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpAveY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpAveY_U8_U8_16x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
-static void avg_pixels16_xy2_mlib (uint8_t * dest, const uint8_t * ref,
-				   int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpAveXY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpAveXY_U8_U8_16x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
+  int i;
 
-/* average/merge dest+source block, width 8 pixel, height 8/16 */
-
-static void avg_pixels8_mlib (uint8_t * dest, const uint8_t * ref,
-			      int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoCopyRefAve_U8_U8_8x16(dest, (uint8_t *)ref, stride);
-    else
-	mlib_VideoCopyRefAve_U8_U8_8x8 (dest, (uint8_t *)ref, stride);
-}
-
-static void avg_pixels8_x2_mlib (uint8_t * dest, const uint8_t * ref,
-				 int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpAveX_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpAveX_U8_U8_8x8 (dest, (uint8_t *)ref, stride, stride);
-}
+  for (i=0;i<8;i++) {
+    mlib_VectorSub_S16_U8_Mod((mlib_s16 *)block, (mlib_u8 *)s1, (mlib_u8 *)s2, 8);
 
-static void avg_pixels8_y2_mlib (uint8_t * dest, const uint8_t * ref,
-				 int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpAveY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpAveY_U8_U8_8x8 (dest, (uint8_t *)ref, stride, stride);
+    s1 += line_size;
+    s2 += line_size;
+    block += 8;
+  }
 }
 
-static void avg_pixels8_xy2_mlib (uint8_t * dest, const uint8_t * ref,
-				  int stride, int height)
-{
-    assert(height == 16 || height == 8);
-    if (height == 16)
-	mlib_VideoInterpAveXY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
-    else
-	mlib_VideoInterpAveXY_U8_U8_8x8 (dest, (uint8_t *)ref, stride, stride);
-}
-
-
-static void (*put_pixels_clamped)(const DCTELEM *block, uint8_t *pixels, int line_size);
-
-
 static void add_pixels_clamped_mlib(const DCTELEM *block, uint8_t *pixels, int line_size)
 {
     mlib_VideoAddBlock_U8_S16(pixels, (mlib_s16 *)block, line_size);
 }
 
+/* put block, width 16 pixel, height 8/16 */
 
-/* XXX: those functions should be suppressed ASAP when all IDCTs are
-   converted */
+static void put_pixels16_mlib (uint8_t * dest, const uint8_t * ref,
+			       int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoCopyRef_U8_U8_16x8(dest, (uint8_t *)ref, stride);
+    break;
+
+    case 16:
+      mlib_VideoCopyRef_U8_U8_16x16(dest, (uint8_t *)ref, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void put_pixels16_x2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoInterpX_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpX_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void put_pixels16_y2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoInterpY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void put_pixels16_xy2_mlib(uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoInterpXY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpXY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+/* put block, width 8 pixel, height 4/8/16 */
+
+static void put_pixels8_mlib (uint8_t * dest, const uint8_t * ref,
+			       int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoCopyRef_U8_U8_8x4(dest, (uint8_t *)ref, stride);
+    break;
+
+    case 8:
+      mlib_VideoCopyRef_U8_U8_8x8(dest, (uint8_t *)ref, stride);
+    break;
+
+    case 16:
+      mlib_VideoCopyRef_U8_U8_8x16(dest, (uint8_t *)ref, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void put_pixels8_x2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoInterpX_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 8:
+      mlib_VideoInterpX_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpX_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void put_pixels8_y2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoInterpY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 8:
+      mlib_VideoInterpY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void put_pixels8_xy2_mlib(uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoInterpXY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 8:
+      mlib_VideoInterpXY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpXY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+/* average block, width 16 pixel, height 8/16 */
+
+static void avg_pixels16_mlib (uint8_t * dest, const uint8_t * ref,
+			       int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoCopyRefAve_U8_U8_16x8(dest, (uint8_t *)ref, stride);
+    break;
+
+    case 16:
+      mlib_VideoCopyRefAve_U8_U8_16x16(dest, (uint8_t *)ref, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void avg_pixels16_x2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoInterpAveX_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpAveX_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void avg_pixels16_y2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoInterpAveY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpAveY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void avg_pixels16_xy2_mlib(uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 8:
+      mlib_VideoInterpAveXY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpAveXY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+/* average block, width 8 pixel, height 4/8/16 */
+
+static void avg_pixels8_mlib (uint8_t * dest, const uint8_t * ref,
+			       int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoCopyRefAve_U8_U8_8x4(dest, (uint8_t *)ref, stride);
+    break;
+
+    case 8:
+      mlib_VideoCopyRefAve_U8_U8_8x8(dest, (uint8_t *)ref, stride);
+    break;
+
+    case 16:
+      mlib_VideoCopyRefAve_U8_U8_8x16(dest, (uint8_t *)ref, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void avg_pixels8_x2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoInterpAveX_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 8:
+      mlib_VideoInterpAveX_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpAveX_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void avg_pixels8_y2_mlib (uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoInterpAveY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 8:
+      mlib_VideoInterpAveY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpAveY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+static void avg_pixels8_xy2_mlib(uint8_t * dest, const uint8_t * ref,
+				  int stride, int height)
+{
+  switch (height) {
+    case 4:
+      mlib_VideoInterpAveXY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 8:
+      mlib_VideoInterpAveXY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    case 16:
+      mlib_VideoInterpAveXY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
+    break;
+
+    default:
+      assert(0);
+  }
+}
+
+/* swap byte order of a buffer */
+
+static void bswap_buf_mlib(uint32_t *dst, uint32_t *src, int w)
+{
+  mlib_VectorReverseByteOrder_U32_U32(dst, src, w);
+}
+
+/* transformations */
+
 static void ff_idct_put_mlib(uint8_t *dest, int line_size, DCTELEM *data)
 {
+    int i;
+    uint8_t *cm = cropTbl + MAX_NEG_CROP;
+
     mlib_VideoIDCT8x8_S16_S16 (data, data);
-    put_pixels_clamped(data, dest, line_size);
+    
+    for(i=0;i<8;i++) {
+        dest[0] = cm[data[0]];
+        dest[1] = cm[data[1]];
+        dest[2] = cm[data[2]];
+        dest[3] = cm[data[3]];
+        dest[4] = cm[data[4]];
+        dest[5] = cm[data[5]];
+        dest[6] = cm[data[6]];
+        dest[7] = cm[data[7]];
+
+        dest += line_size;
+        data += 8;
+    }
 }
 
 static void ff_idct_add_mlib(uint8_t *dest, int line_size, DCTELEM *data)
@@ -233,6 +419,10 @@
 
 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx)
 {
+    c->get_pixels  = get_pixels_mlib;
+    c->diff_pixels = diff_pixels_mlib;
+    c->add_pixels_clamped = add_pixels_clamped_mlib;
+
     c->put_pixels_tab[0][0] = put_pixels16_mlib;
     c->put_pixels_tab[0][1] = put_pixels16_x2_mlib;
     c->put_pixels_tab[0][2] = put_pixels16_y2_mlib;
@@ -254,14 +444,11 @@
     c->put_no_rnd_pixels_tab[0][0] = put_pixels16_mlib;
     c->put_no_rnd_pixels_tab[1][0] = put_pixels8_mlib;
 
-    c->add_pixels_clamped = add_pixels_clamped_mlib;
-    put_pixels_clamped = c->put_pixels_clamped;
+    c->bswap_buf = bswap_buf_mlib;
 }
 
 void MPV_common_init_mlib(MpegEncContext *s)
 {
-    int i;
-
     if(s->avctx->dct_algo==FF_DCT_AUTO || s->avctx->dct_algo==FF_DCT_MLIB){
 	s->dsp.fdct = ff_fdct_mlib;
     }