Use word-writing instead of dword-writing (with two cached but otherwise
[ffmpeg.git] / libavcodec / x86 / vp8dsp-init.c
index 4f40bdd..4bf4936 100644 (file)
@@ -346,7 +346,6 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
         VP8_BILINEAR_MC_FUNC(1, 8, sse2);
 
         c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_sse2;
-        c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse2;
 
         c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_sse2;
         c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_sse2;
@@ -358,6 +357,8 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
     if (mm_flags & FF_MM_SSE2) {
         c->vp8_idct_dc_add4y          = ff_vp8_idct_dc_add4y_sse2;
 
+        c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse2;
+
         c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_sse2;
         c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_sse2;
 
@@ -390,6 +391,7 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
     if (mm_flags & FF_MM_SSE4) {
         c->vp8_idct_dc_add                  = ff_vp8_idct_dc_add_sse4;
 
+        c->vp8_h_loop_filter_simple   = ff_vp8_h_loop_filter_simple_sse4;
         c->vp8_h_loop_filter16y       = ff_vp8_h_loop_filter16y_mbedge_sse4;
         c->vp8_h_loop_filter8uv       = ff_vp8_h_loop_filter8uv_mbedge_sse4;
     }