rv40dsp: implement prescaled versions for biweight.
[ffmpeg.git] / libavcodec / rv34dsp.h
index 6f53a09..58da59f 100644 (file)
 typedef void (*rv40_weight_func)(uint8_t *dst/*align width (8 or 16)*/,
                                  uint8_t *src1/*align width (8 or 16)*/,
                                  uint8_t *src2/*align width (8 or 16)*/,
-                                 int w1, int w2, int stride);
+                                 int w1, int w2, ptrdiff_t stride);
 
 typedef void (*rv34_inv_transform_func)(DCTELEM *block);
 
-typedef void (*rv40_weak_loop_filter_func)(uint8_t *src, int stride,
+typedef void (*rv34_idct_add_func)(uint8_t *dst, ptrdiff_t stride, DCTELEM *block);
+typedef void (*rv34_idct_dc_add_func)(uint8_t *dst, ptrdiff_t stride,
+                                      int   dc);
+
+typedef void (*rv40_weak_loop_filter_func)(uint8_t *src, ptrdiff_t stride,
                                            int filter_p1, int filter_q1,
                                            int alpha, int beta,
                                            int lims, int lim_q1, int lim_p1);
 
-typedef void (*rv40_strong_loop_filter_func)(uint8_t *src, int stride,
+typedef void (*rv40_strong_loop_filter_func)(uint8_t *src, ptrdiff_t stride,
                                              int alpha, int lims,
                                              int dmode, int chroma);
 
-typedef int (*rv40_loop_filter_strength_func)(uint8_t *src, int stride,
+typedef int (*rv40_loop_filter_strength_func)(uint8_t *src, ptrdiff_t stride,
                                               int beta, int beta2, int edge,
                                               int *p1, int *q1);
 
@@ -54,9 +58,16 @@ typedef struct RV34DSPContext {
     qpel_mc_func avg_pixels_tab[4][16];
     h264_chroma_mc_func put_chroma_pixels_tab[3];
     h264_chroma_mc_func avg_chroma_pixels_tab[3];
-    rv40_weight_func rv40_weight_pixels_tab[2];
-    rv34_inv_transform_func rv34_inv_transform_tab[2];
-    void (*rv34_inv_transform_dc_tab[2])(DCTELEM *block);
+    /**
+     * Biweight functions, first dimension is transform size (16/8),
+     * second is whether the weight is prescaled by 1/512 to skip
+     * the intermediate shifting.
+     */
+    rv40_weight_func rv40_weight_pixels_tab[2][2];
+    rv34_inv_transform_func rv34_inv_transform;
+    rv34_inv_transform_func rv34_inv_transform_dc;
+    rv34_idct_add_func rv34_idct_add;
+    rv34_idct_dc_add_func rv34_idct_dc_add;
     rv40_weak_loop_filter_func rv40_weak_loop_filter[2];
     rv40_strong_loop_filter_func rv40_strong_loop_filter[2];
     rv40_loop_filter_strength_func rv40_loop_filter_strength[2];