Merge remote branch 'origin/master' into experimental

Change-Id: I8b0a76b3232c8cff15c0ca5289e18af6889e5095

5 files changed, 81 insertions, 30 deletions

diff --git a/build/make/ads2gas.pl b/build/make/ads2gas.pl
index be4658253..388133aa2 100755
--- a/build/make/ads2gas.pl
+++ b/build/make/ads2gas.pl
@@ -21,6 +21,9 @@ print "@ This file was created from a .asm file\n";
 print "@  using the ads2gas.pl script.\n";
 print "\t.equ DO1STROUNDING, 0\n";
 
+# Stack of procedure names.
+@proc_stack = ();
+
 while (<STDIN>)
 {
     # Load and store alignment
@@ -133,9 +136,23 @@ while (<STDIN>)
     # Strip PRESERVE8
     s/\sPRESERVE8/@ PRESERVE8/g;
 
-    # Strip PROC and ENDPROC
-    s/\sPROC/@/g;
-    s/\sENDP/@/g;
+    # Use PROC and ENDP to give the symbols a .size directive.
+    # This makes them show up properly in debugging tools like gdb and valgrind.
+    if (/\bPROC\b/)
+    {
+        my $proc;
+        /^_([\.0-9A-Z_a-z]\w+)\b/;
+        $proc = $1;
+        push(@proc_stack, $proc) if ($proc);
+        s/\bPROC\b/@ $&/;
+    }
+    if (/\bENDP\b/)
+    {
+        my $proc;
+        s/\bENDP\b/@ $&/;
+        $proc = pop(@proc_stack);
+        $_ = "\t.size $proc, .-$proc".$_ if ($proc);
+    }
 
     # EQU directive
     s/(.*)EQU(.*)/.equ $1, $2/;
@@ -154,3 +171,6 @@ while (<STDIN>)
     next if /^\s*END\s*$/;
     print;
 }
+
+# Mark that this object doesn't need an executable stack.
+printf ("\t.section\t.note.GNU-stack,\"\",\%\%progbits\n");
diff --git a/vp8/encoder/lookahead.c b/vp8/encoder/lookahead.c
index 3b86d4094..d7f85cba1 100644
--- a/vp8/encoder/lookahead.c
+++ b/vp8/encoder/lookahead.c
@@ -86,7 +86,8 @@ vp8_lookahead_init(unsigned int width,
         if(!ctx->buf)
             goto bail;
         for(i=0; i<depth; i++)
-            if (vp8_yv12_alloc_frame_buffer(&ctx->buf[i].img, width, height, 16))
+            if (vp8_yv12_alloc_frame_buffer(&ctx->buf[i].img,
+                                            width, height, VP8BORDERINPIXELS))
                 goto bail;
     }
     return ctx;
diff --git a/vp8/encoder/onyx_if.c b/vp8/encoder/onyx_if.c
index d936bac35..d3f3699c9 100644
--- a/vp8/encoder/onyx_if.c
+++ b/vp8/encoder/onyx_if.c
@@ -951,6 +951,10 @@ void vp8_set_speed_features(VP8_COMP *cpi)
 
             sf->improved_quant = 0;
             sf->improved_dct = 0;
+
+            sf->use_fastquant_for_pick = 1;
+            sf->no_skip_block4x4_search = 0;
+            sf->first_step = 1;
         }
 
         if (Speed > 1)
@@ -1309,7 +1313,7 @@ static void alloc_raw_frame_buffers(VP8_COMP *cpi)
 #if VP8_TEMPORAL_ALT_REF
 
     if (vp8_yv12_alloc_frame_buffer(&cpi->alt_ref_buffer,
-                                    width, height, 16))
+                                    width, height, VP8BORDERINPIXELS))
         vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
                            "Failed to allocate altref buffer");
 
@@ -1359,7 +1363,8 @@ void vp8_alloc_compressor_data(VP8_COMP *cpi)
         vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
                            "Failed to allocate last frame buffer");
 
-    if (vp8_yv12_alloc_frame_buffer(&cpi->scaled_source, width, height, 16))
+    if (vp8_yv12_alloc_frame_buffer(&cpi->scaled_source,
+                                    width, height, VP8BORDERINPIXELS))
         vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
                            "Failed to allocate scaled source buffer");
 
diff --git a/vp8/vp8_dx_iface.c b/vp8/vp8_dx_iface.c
index 58dc486de..13a072bff 100644
--- a/vp8/vp8_dx_iface.c
+++ b/vp8/vp8_dx_iface.c
@@ -301,6 +301,36 @@ update_error_state(vpx_codec_alg_priv_t                 *ctx,
     return res;
 }
 
+static void yuvconfig2image(vpx_image_t               *img,
+                            const YV12_BUFFER_CONFIG  *yv12,
+                            void                      *user_priv)
+{
+    /** vpx_img_wrap() doesn't allow specifying independent strides for
+      * the Y, U, and V planes, nor other alignment adjustments that
+      * might be representable by a YV12_BUFFER_CONFIG, so we just
+      * initialize all the fields.*/
+    img->fmt = yv12->clrtype == REG_YUV ?
+        VPX_IMG_FMT_I420 : VPX_IMG_FMT_VPXI420;
+    img->w = yv12->y_stride;
+    img->h = (yv12->y_height + 2 * VP8BORDERINPIXELS + 15) & ~15;
+    img->d_w = yv12->y_width;
+    img->d_h = yv12->y_height;
+    img->x_chroma_shift = 1;
+    img->y_chroma_shift = 1;
+    img->planes[VPX_PLANE_Y] = yv12->y_buffer;
+    img->planes[VPX_PLANE_U] = yv12->u_buffer;
+    img->planes[VPX_PLANE_V] = yv12->v_buffer;
+    img->planes[VPX_PLANE_ALPHA] = NULL;
+    img->stride[VPX_PLANE_Y] = yv12->y_stride;
+    img->stride[VPX_PLANE_U] = yv12->uv_stride;
+    img->stride[VPX_PLANE_V] = yv12->uv_stride;
+    img->stride[VPX_PLANE_ALPHA] = yv12->y_stride;
+    img->bps = 12;
+    img->user_priv = user_priv;
+    img->img_data = yv12->buffer_alloc;
+    img->img_data_owner = 0;
+    img->self_allocd = 0;
+}
 
 static vpx_codec_err_t vp8_decode(vpx_codec_alg_priv_t  *ctx,
                                   const uint8_t         *data,
@@ -429,21 +459,8 @@ static vpx_codec_err_t vp8_decode(vpx_codec_alg_priv_t  *ctx,
 
         if (!res && 0 == vp8dx_get_raw_frame(ctx->pbi, &sd, &time_stamp, &time_end_stamp, &flags))
         {
-            /* Align width/height */
-            unsigned int a_w = (sd.y_width + 15) & ~15;
-            unsigned int a_h = (sd.y_height + 15) & ~15;
-
-            vpx_img_wrap(&ctx->img, VPX_IMG_FMT_I420,
-                         a_w + 2 * VP8BORDERINPIXELS,
-                         a_h + 2 * VP8BORDERINPIXELS,
-                         1,
-                         sd.buffer_alloc);
-            vpx_img_set_rect(&ctx->img,
-                             VP8BORDERINPIXELS, VP8BORDERINPIXELS,
-                             sd.y_width, sd.y_height);
-            ctx->img.user_priv = user_priv;
+            yuvconfig2image(&ctx->img, &sd, user_priv);
             ctx->img_avail = 1;
-
         }
     }
 
diff --git a/vpx_scale/generic/yv12config.c b/vpx_scale/generic/yv12config.c
index d02cde28f..eff594e2d 100644
--- a/vpx_scale/generic/yv12config.c
+++ b/vpx_scale/generic/yv12config.c
@@ -49,25 +49,33 @@ vp8_yv12_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf, int width, int height, int
 
     if (ybf)
     {
+        int y_stride = ((width + 2 * border) + 31) & ~31;
+        int yplane_size = (height + 2 * border) * y_stride;
         int uv_width = width >> 1;
         int uv_height = height >> 1;
-        int yplane_size = (height + 2 * border) * (width + 2 * border);
-        int uvplane_size = (uv_height + border) * (uv_width + border);
+        /** There is currently a bunch of code which assumes
+          *  uv_stride == y_stride/2, so enforce this here. */
+        int uv_stride = y_stride >> 1;
+        int uvplane_size = (uv_height + border) * uv_stride;
 
         vp8_yv12_de_alloc_frame_buffer(ybf);
 
-        /* only support allocating buffers that have
-          a height and width that are multiples of 16 */
-        if ((width & 0xf) | (height & 0xf))
+        /** Only support allocating buffers that have a height and width that
+          *  are multiples of 16, and a border that's a multiple of 32.
+          * The border restriction is required to get 16-byte alignment of the
+          *  start of the chroma rows without intoducing an arbitrary gap
+          *  between planes, which would break the semantics of things like
+          *  vpx_img_set_rect(). */
+        if ((width & 0xf) | (height & 0xf) | (border & 0x1f))
             return -3;
 
         ybf->y_width  = width;
         ybf->y_height = height;
-        ybf->y_stride = width + 2 * border;
+        ybf->y_stride = y_stride;
 
         ybf->uv_width = uv_width;
         ybf->uv_height = uv_height;
-        ybf->uv_stride = uv_width + border;
+        ybf->uv_stride = uv_stride;
 
         ybf->border = border;
         ybf->frame_size = yplane_size + 2 * uvplane_size;
@@ -77,9 +85,9 @@ vp8_yv12_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf, int width, int height, int
         if (ybf->buffer_alloc == NULL)
             return -1;
 
-        ybf->y_buffer = ybf->buffer_alloc + (border * ybf->y_stride) + border;
-        ybf->u_buffer = ybf->buffer_alloc + yplane_size + (border / 2  * ybf->uv_stride) + border / 2;
-        ybf->v_buffer = ybf->buffer_alloc + yplane_size + uvplane_size + (border / 2  * ybf->uv_stride) + border / 2;
+        ybf->y_buffer = ybf->buffer_alloc + (border * y_stride) + border;
+        ybf->u_buffer = ybf->buffer_alloc + yplane_size + (border / 2  * uv_stride) + border / 2;
+        ybf->v_buffer = ybf->buffer_alloc + yplane_size + uvplane_size + (border / 2  * uv_stride) + border / 2;
 
         ybf->corrupted = 0; /* assume not currupted by errors */
     }