avcodec/mips: version 1 of vp8dsp optimizations for loongson mmi

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>

1 files changed, 89 insertions, 0 deletions

diff --git a/libavcodec/mips/vp8dsp_init_mips.c b/libavcodec/mips/vp8dsp_init_mips.c
index 58d1b6c..3fc5f8e 100644
--- a/libavcodec/mips/vp8dsp_init_mips.c
+++ b/libavcodec/mips/vp8dsp_init_mips.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2015 Manojkumar Bhosale (Manojkumar.Bhosale@imgtec.com)
+ * Copyright (c) 2016 Zhou Xiaoyong <zhouxiaoyong@loongson.cn>
  *
  * This file is part of FFmpeg.
  *
@@ -105,9 +106,97 @@ static av_cold void vp8dsp_init_msa(VP8DSPContext *dsp)
 }
 #endif  // #if HAVE_MSA
 
+#if HAVE_MMI
+static av_cold void vp8dsp_init_mmi(VP8DSPContext *dsp)
+{
+    dsp->vp8_luma_dc_wht    = ff_vp8_luma_dc_wht_mmi;
+    dsp->vp8_luma_dc_wht_dc = ff_vp8_luma_dc_wht_dc_mmi;
+    dsp->vp8_idct_add       = ff_vp8_idct_add_mmi;
+    dsp->vp8_idct_dc_add    = ff_vp8_idct_dc_add_mmi;
+    dsp->vp8_idct_dc_add4y  = ff_vp8_idct_dc_add4y_mmi;
+    dsp->vp8_idct_dc_add4uv = ff_vp8_idct_dc_add4uv_mmi;
+
+    dsp->put_vp8_epel_pixels_tab[0][0][1] = ff_put_vp8_epel16_h4_mmi;
+    dsp->put_vp8_epel_pixels_tab[0][0][2] = ff_put_vp8_epel16_h6_mmi;
+    dsp->put_vp8_epel_pixels_tab[0][1][0] = ff_put_vp8_epel16_v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[0][1][1] = ff_put_vp8_epel16_h4v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[0][1][2] = ff_put_vp8_epel16_h6v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[0][2][0] = ff_put_vp8_epel16_v6_mmi;
+    dsp->put_vp8_epel_pixels_tab[0][2][1] = ff_put_vp8_epel16_h4v6_mmi;
+    dsp->put_vp8_epel_pixels_tab[0][2][2] = ff_put_vp8_epel16_h6v6_mmi;
+
+    dsp->put_vp8_epel_pixels_tab[1][0][1] = ff_put_vp8_epel8_h4_mmi;
+    dsp->put_vp8_epel_pixels_tab[1][0][2] = ff_put_vp8_epel8_h6_mmi;
+    dsp->put_vp8_epel_pixels_tab[1][1][0] = ff_put_vp8_epel8_v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[1][1][1] = ff_put_vp8_epel8_h4v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[1][1][2] = ff_put_vp8_epel8_h6v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[1][2][0] = ff_put_vp8_epel8_v6_mmi;
+    dsp->put_vp8_epel_pixels_tab[1][2][1] = ff_put_vp8_epel8_h4v6_mmi;
+    dsp->put_vp8_epel_pixels_tab[1][2][2] = ff_put_vp8_epel8_h6v6_mmi;
+
+    dsp->put_vp8_epel_pixels_tab[2][0][1] = ff_put_vp8_epel4_h4_mmi;
+    dsp->put_vp8_epel_pixels_tab[2][0][2] = ff_put_vp8_epel4_h6_mmi;
+    dsp->put_vp8_epel_pixels_tab[2][1][0] = ff_put_vp8_epel4_v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[2][1][1] = ff_put_vp8_epel4_h4v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[2][1][2] = ff_put_vp8_epel4_h6v4_mmi;
+    dsp->put_vp8_epel_pixels_tab[2][2][0] = ff_put_vp8_epel4_v6_mmi;
+    dsp->put_vp8_epel_pixels_tab[2][2][1] = ff_put_vp8_epel4_h4v6_mmi;
+    dsp->put_vp8_epel_pixels_tab[2][2][2] = ff_put_vp8_epel4_h6v6_mmi;
+
+    dsp->put_vp8_bilinear_pixels_tab[0][0][1] = ff_put_vp8_bilinear16_h_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][0][2] = ff_put_vp8_bilinear16_h_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][1][0] = ff_put_vp8_bilinear16_v_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][1][1] = ff_put_vp8_bilinear16_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][1][2] = ff_put_vp8_bilinear16_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][2][0] = ff_put_vp8_bilinear16_v_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][2][1] = ff_put_vp8_bilinear16_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][2][2] = ff_put_vp8_bilinear16_hv_mmi;
+
+    dsp->put_vp8_bilinear_pixels_tab[1][0][1] = ff_put_vp8_bilinear8_h_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][0][2] = ff_put_vp8_bilinear8_h_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][1][0] = ff_put_vp8_bilinear8_v_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][1][1] = ff_put_vp8_bilinear8_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][1][2] = ff_put_vp8_bilinear8_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][2][0] = ff_put_vp8_bilinear8_v_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][2][1] = ff_put_vp8_bilinear8_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][2][2] = ff_put_vp8_bilinear8_hv_mmi;
+
+    dsp->put_vp8_bilinear_pixels_tab[2][0][1] = ff_put_vp8_bilinear4_h_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[2][0][2] = ff_put_vp8_bilinear4_h_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[2][1][0] = ff_put_vp8_bilinear4_v_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[2][1][1] = ff_put_vp8_bilinear4_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[2][1][2] = ff_put_vp8_bilinear4_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[2][2][0] = ff_put_vp8_bilinear4_v_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[2][2][1] = ff_put_vp8_bilinear4_hv_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[2][2][2] = ff_put_vp8_bilinear4_hv_mmi;
+
+    dsp->put_vp8_epel_pixels_tab[0][0][0]     = ff_put_vp8_pixels16_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_mmi;
+
+    dsp->put_vp8_epel_pixels_tab[1][0][0]     = ff_put_vp8_pixels8_mmi;
+    dsp->put_vp8_bilinear_pixels_tab[1][0][0] = ff_put_vp8_pixels8_mmi;
+
+    dsp->vp8_v_loop_filter16y = ff_vp8_v_loop_filter16_mmi;
+    dsp->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16_mmi;
+    dsp->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_mmi;
+    dsp->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mmi;
+
+    dsp->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16_inner_mmi;
+    dsp->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16_inner_mmi;
+    dsp->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_mmi;
+    dsp->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_mmi;
+
+    dsp->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmi;
+    dsp->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmi;
+}
+#endif /* HAVE_MMI */
+
 av_cold void ff_vp8dsp_init_mips(VP8DSPContext *dsp)
 {
 #if HAVE_MSA
     vp8dsp_init_msa(dsp);
 #endif  // #if HAVE_MSA
+#if HAVE_MMI
+    vp8dsp_init_mmi(dsp);
+#endif /* HAVE_MMI */
 }