arm: Add NEON optimizations for 10 and 12 bit vp9 MC
[ffmpeg.git] / libavcodec / arm / vp9dsp_init_arm.c
index bd0ac37..f7b539e 100644 (file)
@@ -23,6 +23,7 @@
 #include "libavutil/attributes.h"
 #include "libavutil/arm/cpu.h"
 #include "libavcodec/vp9dsp.h"
+#include "vp9dsp_init.h"
 
 #define declare_fpel(type, sz)                                          \
 void ff_vp9_##type##sz##_neon(uint8_t *dst, ptrdiff_t dst_stride,       \
@@ -94,13 +95,10 @@ define_8tap_2d_funcs(8)
 define_8tap_2d_funcs(4)
 
 
-av_cold void ff_vp9dsp_init_arm(VP9DSPContext *dsp, int bpp)
+static av_cold void vp9dsp_mc_init_arm(VP9DSPContext *dsp)
 {
     int cpu_flags = av_get_cpu_flags();
 
-    if (bpp != 8)
-        return;
-
     if (have_neon(cpu_flags)) {
 #define init_fpel(idx1, idx2, sz, type)              \
     dsp->mc[idx1][FILTER_8TAP_SMOOTH ][idx2][0][0] = \
@@ -141,3 +139,118 @@ av_cold void ff_vp9dsp_init_arm(VP9DSPContext *dsp, int bpp)
         init_mc_funcs_dirs(4, 4);
     }
 }
+
+#define define_itxfm(type_a, type_b, sz)                                   \
+void ff_vp9_##type_a##_##type_b##_##sz##x##sz##_add_neon(uint8_t *_dst,    \
+                                                         ptrdiff_t stride, \
+                                                         int16_t *_block, int eob)
+
+#define define_itxfm_funcs(sz)      \
+    define_itxfm(idct,  idct,  sz); \
+    define_itxfm(iadst, idct,  sz); \
+    define_itxfm(idct,  iadst, sz); \
+    define_itxfm(iadst, iadst, sz)
+
+define_itxfm_funcs(4);
+define_itxfm_funcs(8);
+define_itxfm_funcs(16);
+define_itxfm(idct, idct, 32);
+define_itxfm(iwht, iwht, 4);
+
+
+static av_cold void vp9dsp_itxfm_init_arm(VP9DSPContext *dsp)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (have_neon(cpu_flags)) {
+#define init_itxfm(tx, sz)                                             \
+    dsp->itxfm_add[tx][DCT_DCT]   = ff_vp9_idct_idct_##sz##_add_neon;  \
+    dsp->itxfm_add[tx][DCT_ADST]  = ff_vp9_iadst_idct_##sz##_add_neon; \
+    dsp->itxfm_add[tx][ADST_DCT]  = ff_vp9_idct_iadst_##sz##_add_neon; \
+    dsp->itxfm_add[tx][ADST_ADST] = ff_vp9_iadst_iadst_##sz##_add_neon
+
+#define init_idct(tx, nm)           \
+    dsp->itxfm_add[tx][DCT_DCT]   = \
+    dsp->itxfm_add[tx][ADST_DCT]  = \
+    dsp->itxfm_add[tx][DCT_ADST]  = \
+    dsp->itxfm_add[tx][ADST_ADST] = ff_vp9_##nm##_add_neon
+
+        init_itxfm(TX_4X4, 4x4);
+        init_itxfm(TX_8X8, 8x8);
+        init_itxfm(TX_16X16, 16x16);
+        init_idct(TX_32X32, idct_idct_32x32);
+        init_idct(4, iwht_iwht_4x4);
+    }
+}
+
+#define define_loop_filter(dir, wd, size) \
+void ff_vp9_loop_filter_##dir##_##wd##_##size##_neon(uint8_t *dst, ptrdiff_t stride, int E, int I, int H)
+
+#define define_loop_filters(wd, size) \
+    define_loop_filter(h, wd, size);  \
+    define_loop_filter(v, wd, size)
+
+define_loop_filters(4, 8);
+define_loop_filters(8, 8);
+define_loop_filters(16, 8);
+define_loop_filters(16, 16);
+
+#define lf_mix_fn(dir, wd1, wd2, stridea)                                                         \
+static void loop_filter_##dir##_##wd1##wd2##_16_neon(uint8_t *dst,                                \
+                                                     ptrdiff_t stride,                            \
+                                                     int E, int I, int H)                         \
+{                                                                                                 \
+    ff_vp9_loop_filter_##dir##_##wd1##_8_neon(dst, stride, E & 0xff, I & 0xff, H & 0xff);         \
+    ff_vp9_loop_filter_##dir##_##wd2##_8_neon(dst + 8 * stridea, stride, E >> 8, I >> 8, H >> 8); \
+}
+
+#define lf_mix_fns(wd1, wd2)       \
+    lf_mix_fn(h, wd1, wd2, stride) \
+    lf_mix_fn(v, wd1, wd2, sizeof(uint8_t))
+
+lf_mix_fns(4, 4)
+lf_mix_fns(4, 8)
+lf_mix_fns(8, 4)
+lf_mix_fns(8, 8)
+
+static av_cold void vp9dsp_loopfilter_init_arm(VP9DSPContext *dsp)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (have_neon(cpu_flags)) {
+        dsp->loop_filter_8[0][1] = ff_vp9_loop_filter_v_4_8_neon;
+        dsp->loop_filter_8[0][0] = ff_vp9_loop_filter_h_4_8_neon;
+        dsp->loop_filter_8[1][1] = ff_vp9_loop_filter_v_8_8_neon;
+        dsp->loop_filter_8[1][0] = ff_vp9_loop_filter_h_8_8_neon;
+        dsp->loop_filter_8[2][1] = ff_vp9_loop_filter_v_16_8_neon;
+        dsp->loop_filter_8[2][0] = ff_vp9_loop_filter_h_16_8_neon;
+
+        dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_neon;
+        dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_neon;
+
+        dsp->loop_filter_mix2[0][0][0] = loop_filter_h_44_16_neon;
+        dsp->loop_filter_mix2[0][0][1] = loop_filter_v_44_16_neon;
+        dsp->loop_filter_mix2[0][1][0] = loop_filter_h_48_16_neon;
+        dsp->loop_filter_mix2[0][1][1] = loop_filter_v_48_16_neon;
+        dsp->loop_filter_mix2[1][0][0] = loop_filter_h_84_16_neon;
+        dsp->loop_filter_mix2[1][0][1] = loop_filter_v_84_16_neon;
+        dsp->loop_filter_mix2[1][1][0] = loop_filter_h_88_16_neon;
+        dsp->loop_filter_mix2[1][1][1] = loop_filter_v_88_16_neon;
+    }
+}
+
+av_cold void ff_vp9dsp_init_arm(VP9DSPContext *dsp, int bpp)
+{
+    if (bpp == 10) {
+        ff_vp9dsp_init_10bpp_arm(dsp);
+        return;
+    } else if (bpp == 12) {
+        ff_vp9dsp_init_12bpp_arm(dsp);
+        return;
+    } else if (bpp != 8)
+        return;
+
+    vp9dsp_mc_init_arm(dsp);
+    vp9dsp_loopfilter_init_arm(dsp);
+    vp9dsp_itxfm_init_arm(dsp);
+}