--- /dev/null
+.init bench10_init
+
+#.init gst_volume_orc_init
+
+.function orc_scalarmultiply_f32_ns
+.dest 4 d1 float
+.floatparam 4 p1
+
+mulf d1, d1, p1
+
+
+.function orc_process_int16
+.dest 2 d1 gint16
+.param 2 p1
+.temp 4 t1
+
+mulswl t1, d1, p1
+shrsl t1, t1, 13
+convlw d1, t1
+
+
+.function orc_process_int16_clamp
+.dest 2 d1 gint16
+.param 2 p1
+.temp 4 t1
+
+mulswl t1, d1, p1
+shrsl t1, t1, 13
+convssslw d1, t1
+
+
+.function orc_process_int8
+.dest 1 d1 gint8
+.param 1 p1
+.temp 2 t1
+
+mulsbw t1, d1, p1
+shrsw t1, t1, 5
+convwb d1, t1
+
+
+.function orc_process_int8_clamp
+.dest 1 d1 gint8
+.param 1 p1
+.temp 2 t1
+
+mulsbw t1, d1, p1
+shrsw t1, t1, 5
+convssswb d1, t1
+
+
+
+#.init gst_audio_convert_orc_init
+
+.function orc_audio_convert_unpack_u8
+.dest 4 d1 gint32
+.source 1 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 2 t2
+.temp 4 t3
+
+convubw t2, s1
+convuwl t3, t2
+shll t3, t3, p1
+xorl d1, t3, c1
+
+
+.function orc_audio_convert_unpack_s8
+.dest 4 d1 gint32
+.source 1 s1 guint8
+.param 4 p1
+.temp 2 t2
+.temp 4 t3
+
+convubw t2, s1
+convuwl t3, t2
+shll d1, t3, p1
+
+
+.function orc_audio_convert_unpack_u16
+.dest 4 d1 gint32
+.source 2 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t2
+
+convuwl t2, s1
+shll t2, t2, p1
+xorl d1, t2, c1
+
+
+.function orc_audio_convert_unpack_s16
+.dest 4 d1 gint32
+.source 2 s1 guint8
+.param 4 p1
+.temp 4 t2
+
+convuwl t2, s1
+shll d1, t2, p1
+
+
+.function orc_audio_convert_unpack_u16_swap
+.dest 4 d1 gint32
+.source 2 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 2 t1
+.temp 4 t2
+
+swapw t1, s1
+convuwl t2, t1
+shll t2, t2, p1
+xorl d1, t2, c1
+
+
+.function orc_audio_convert_unpack_s16_swap
+.dest 4 d1 gint32
+.source 2 s1 guint8
+.param 4 p1
+.temp 2 t1
+.temp 4 t2
+
+swapw t1, s1
+convuwl t2, t1
+shll d1, t2, p1
+
+
+.function orc_audio_convert_unpack_u32
+.dest 4 d1 gint32
+.source 4 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+
+shll t1, s1, p1
+xorl d1, t1, c1
+
+
+.function orc_audio_convert_unpack_s32
+.dest 4 d1 gint32
+.source 4 s1 guint8
+.param 4 p1
+
+shll d1, s1, p1
+
+
+.function orc_audio_convert_unpack_u32_swap
+.dest 4 d1 gint32
+.source 4 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+
+swapl t1, s1
+shll t1, t1, p1
+xorl d1, t1, c1
+
+
+.function orc_audio_convert_unpack_s32_swap
+.dest 4 d1 gint32
+.source 4 s1 guint8
+.param 4 p1
+.temp 4 t1
+
+swapl t1, s1
+shll d1, t1, p1
+
+.function orc_audio_convert_unpack_float_s32
+.source 4 s1 gfloat
+.dest 4 d1 guint32
+.temp 4 t1
+
+loadl t1, s1
+# multiply with 2147483647.0
+mulf t1, t1, 0x4F000000
+# add 0.5 for rounding
+addf t1, t1, 0x3F000000
+convfl d1, t1
+
+.function orc_audio_convert_unpack_float_s32_swap
+.source 4 s1 gfloat
+.dest 4 d1 guint32
+.temp 4 t1
+
+swapl t1, s1
+# multiply with 2147483647.0
+mulf t1, t1, 0x4F000000
+# add 0.5 for rounding
+addf t1, t1, 0x3F000000
+convfl d1, t1
+
+.function orc_audio_convert_unpack_float_double
+.dest 8 d1 gdouble
+.source 4 s1 gfloat
+
+convfd d1, s1
+
+.function orc_audio_convert_unpack_float_double_swap
+.dest 8 d1 gdouble
+.source 4 s1 gfloat
+.temp 4 t1
+
+swapl t1, s1
+convfd d1, t1
+
+.function orc_audio_convert_unpack_double_double
+.dest 8 d1 gdouble
+.source 8 s1 gdouble
+
+copyq d1, s1
+
+.function orc_audio_convert_unpack_double_double_swap
+.dest 8 d1 gdouble
+.source 8 s1 gdouble
+
+swapq d1, s1
+
+.function orc_audio_convert_unpack_u8_double
+.dest 8 d1 gdouble
+.source 1 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 2 t2
+.temp 4 t3
+
+convubw t2, s1
+convuwl t3, t2
+shll t3, t3, p1
+xorl t3, t3, c1
+convld d1, t3
+
+.function orc_audio_convert_unpack_s8_double
+.dest 8 d1 gdouble
+.source 1 s1 guint8
+.param 4 p1
+.temp 2 t2
+.temp 4 t3
+
+convubw t2, s1
+convuwl t3, t2
+shll t3, t3, p1
+convld d1, t3
+
+.function orc_audio_convert_unpack_u16_double
+.dest 8 d1 gdouble
+.source 2 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t2
+
+convuwl t2, s1
+shll t2, t2, p1
+xorl t2, t2, c1
+convld d1, t2
+
+.function orc_audio_convert_unpack_s16_double
+.dest 8 d1 gdouble
+.source 2 s1 guint8
+.param 4 p1
+.temp 4 t2
+
+convuwl t2, s1
+shll t2, t2, p1
+convld d1, t2
+
+.function orc_audio_convert_unpack_u16_double_swap
+.dest 8 d1 gdouble
+.source 2 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 2 t1
+.temp 4 t2
+
+swapw t1, s1
+convuwl t2, t1
+shll t2, t2, p1
+xorl t2, t2, c1
+convld d1, t2
+
+.function orc_audio_convert_unpack_s16_double_swap
+.dest 8 d1 gdouble
+.source 2 s1 guint8
+.param 4 p1
+.temp 2 t1
+.temp 4 t2
+
+swapw t1, s1
+convuwl t2, t1
+shll t2, t2, p1
+convld d1, t2
+
+.function orc_audio_convert_unpack_u32_double
+.dest 8 d1 gdouble
+.source 4 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+
+shll t1, s1, p1
+xorl t1, t1, c1
+convld d1, t1
+
+.function orc_audio_convert_unpack_s32_double
+.dest 8 d1 gdouble
+.source 4 s1 guint8
+.param 4 p1
+.temp 4 t1
+
+shll t1, s1, p1
+convld d1, t1
+
+.function orc_audio_convert_unpack_u32_double_swap
+.dest 8 d1 gdouble
+.source 4 s1 guint8
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+
+swapl t1, s1
+shll t1, t1, p1
+xorl t1, t1, c1
+convld d1, t1
+
+.function orc_audio_convert_unpack_s32_double_swap
+.dest 8 d1 gdouble
+.source 4 s1 guint8
+.param 4 p1
+.temp 4 t1
+
+swapl t1, s1
+shll t1, t1, p1
+convld d1, t1
+
+.function orc_audio_convert_pack_u8
+.dest 1 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+.temp 2 t2
+
+xorl t1, s1, c1
+shrul t1, t1, p1
+convlw t2, t1
+convwb d1, t2
+
+
+.function orc_audio_convert_pack_s8
+.dest 1 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.temp 4 t1
+.temp 2 t2
+
+shrsl t1, s1, p1
+convlw t2, t1
+convwb d1, t2
+
+
+
+.function orc_audio_convert_pack_u16
+.dest 2 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+
+xorl t1, s1, c1
+shrul t1, t1, p1
+convlw d1, t1
+
+
+.function orc_audio_convert_pack_s16
+.dest 2 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.temp 4 t1
+
+shrsl t1, s1, p1
+convlw d1, t1
+
+
+.function orc_audio_convert_pack_u16_swap
+.dest 2 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+.temp 2 t2
+
+xorl t1, s1, c1
+shrul t1, t1, p1
+convlw t2, t1
+swapw d1, t2
+
+
+.function orc_audio_convert_pack_s16_swap
+.dest 2 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.temp 4 t1
+.temp 2 t2
+
+shrsl t1, s1, p1
+convlw t2, t1
+swapw d1, t2
+
+
+
+.function orc_audio_convert_pack_u32
+.dest 4 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+
+xorl t1, s1, c1
+shrul d1, t1, p1
+
+
+.function orc_audio_convert_pack_s32
+.dest 4 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+
+shrsl d1, s1, p1
+
+
+.function orc_audio_convert_pack_u32_swap
+.dest 4 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.const 4 c1 0x80000000
+.temp 4 t1
+
+xorl t1, s1, c1
+shrul t1, t1, p1
+swapl d1, t1
+
+
+.function orc_audio_convert_pack_s32_swap
+.dest 4 d1 guint8
+.source 4 s1 gint32
+.param 4 p1
+.temp 4 t1
+
+shrsl t1, s1, p1
+swapl d1, t1
+
+.function orc_audio_convert_pack_s32_float
+.dest 4 d1 gfloat
+.source 4 s1 gint32
+.temp 4 t1
+
+convlf t1, s1
+# divide by 2147483647.0
+divf t1, t1, 0x4F000000
+storel d1, t1
+
+.function orc_audio_convert_pack_s32_float_swap
+.dest 4 d1 gfloat
+.source 4 s1 gint32
+.temp 4 t1
+
+convlf t1, s1
+# divide by 2147483647.0
+divf t1, t1, 0x4F000000
+swapl d1, t1
+
+.function orc_audio_convert_pack_double_float
+.dest 4 d1 gfloat
+.source 8 s1 gdouble
+
+convdf d1, s1
+
+.function orc_audio_convert_pack_double_float_swap
+.dest 4 d1 gfloat
+.source 8 s1 gdouble
+.temp 4 t1
+
+convdf t1, s1
+swapl d1, t1
+
+.function orc_audio_convert_pack_double_s8
+.dest 1 d1 guint8
+.source 8 s1 gdouble
+.param 4 p1
+.temp 4 t1
+.temp 2 t2
+
+convdl t1, s1
+shrsl t1, t1, p1
+convlw t2, t1
+convwb d1, t2
+
+.function orc_audio_convert_pack_double_s16
+.dest 2 d1 guint8
+.source 8 s1 gdouble
+.param 4 p1
+.temp 4 t1
+
+convdl t1, s1
+shrsl t1, t1, p1
+convlw d1, t1
+
+.function orc_audio_convert_pack_double_s16_swap
+.dest 2 d1 guint8
+.source 8 s1 gdouble
+.param 4 p1
+.temp 4 t1
+.temp 2 t2
+
+convdl t1, s1
+shrsl t1, t1, p1
+convlw t2, t1
+swapw d1, t2
+
+.function orc_audio_convert_pack_double_s32
+.dest 4 d1 guint8
+.source 8 s1 gdouble
+.param 4 p1
+.temp 4 t1
+
+convdl t1, s1
+shrsl d1, t1, p1
+
+.function orc_audio_convert_pack_double_s32_swap
+.dest 4 d1 guint8
+.source 8 s1 gdouble
+.param 4 p1
+.temp 4 t1
+
+convdl t1, s1
+shrsl t1, t1, p1
+swapl d1, t1
+
+
+#.init gst_videotestsrc_orc_init
+
+.function gst_orc_splat_u8
+.dest 1 d1 guint8
+.param 1 p1
+
+copyb d1, p1
+
+
+.function gst_orc_splat_s16
+.dest 2 d1 gint8
+.param 2 p1
+
+copyw d1, p1
+
+
+.function gst_orc_splat_u16
+.dest 2 d1 guint8
+.param 2 p1
+
+copyw d1, p1
+
+
+.function gst_orc_splat_u32
+.dest 4 d1 guint8
+.param 4 p1
+
+copyl d1, p1
+
+
+
+#.init gst_videoscale_orc_init
+
+.function orc_merge_linear_u8
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+.param 1 p1
+.temp 2 t1
+.temp 2 t2
+.temp 1 a
+.temp 1 t
+
+loadb a, s1
+convubw t1, s1
+convubw t2, s2
+subw t2, t2, t1
+mullw t2, t2, p1
+addw t2, t2, 128
+convhwb t, t2
+addb d1, t, a
+
+
+
+.function orc_merge_linear_u16
+.dest 2 d1
+.source 2 s1
+.source 2 s2
+.param 2 p1
+.param 2 p2
+.temp 4 t1
+.temp 4 t2
+
+# This is slightly different thatn the u8 case, since muluwl
+# tends to be much faster than mulll
+muluwl t1, s1, p1
+muluwl t2, s2, p2
+addl t1, t1, t2
+shrul t1, t1, 16
+convlw d1, t1
+
+
+.function orc_splat_u16
+.dest 2 d1
+.param 2 p1
+
+copyw d1, p1
+
+
+.function orc_splat_u32
+.dest 4 d1
+.param 4 p1
+
+copyl d1, p1
+
+
+.function orc_downsample_u8
+.dest 1 d1 guint8
+.source 2 s1 guint8
+.temp 1 t1
+.temp 1 t2
+
+splitwb t1, t2, s1
+avgub d1, t1, t2
+
+
+.function orc_downsample_u16
+.dest 2 d1 guint16
+.source 4 s1 guint16
+.temp 2 t1
+.temp 2 t2
+
+splitlw t1, t2, s1
+avguw d1, t1, t2
+
+
+.function gst_videoscale_orc_downsample_u32
+.dest 4 d1 guint8
+.source 8 s1 guint8
+.temp 4 t1
+.temp 4 t2
+
+splitql t1, t2, s1
+x4 avgub d1, t1, t2
+
+
+.function gst_videoscale_orc_downsample_yuyv
+.dest 4 d1 guint8
+.source 8 s1 guint8
+.temp 4 yyyy
+.temp 4 uvuv
+.temp 2 t1
+.temp 2 t2
+.temp 2 yy
+.temp 2 uv
+
+x4 splitwb yyyy, uvuv, s1
+x2 splitwb t1, t2, yyyy
+x2 avgub yy, t1, t2
+splitlw t1, t2, uvuv
+x2 avgub uv, t1, t2
+x2 mergebw d1, yy, uv
+
+
+
+.function gst_videoscale_orc_resample_nearest_u8
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.param 4 p1
+.param 4 p2
+
+ldresnearb d1, s1, p1, p2
+
+
+.function gst_videoscale_orc_resample_bilinear_u8
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.param 4 p1
+.param 4 p2
+
+ldreslinb d1, s1, p1, p2
+
+
+.function gst_videoscale_orc_resample_nearest_u32
+.dest 4 d1 guint8
+.source 4 s1 guint8
+.param 4 p1
+.param 4 p2
+
+ldresnearl d1, s1, p1, p2
+
+
+.function gst_videoscale_orc_resample_bilinear_u32
+.dest 4 d1 guint8
+.source 4 s1 guint8
+.param 4 p1
+.param 4 p2
+
+ldreslinl d1, s1, p1, p2
+
+
+.function gst_videoscale_orc_resample_merge_bilinear_u32
+.dest 4 d1 guint8
+.dest 4 d2 guint8
+.source 4 s1 guint8
+.source 4 s2 guint8
+.temp 4 a
+.temp 4 b
+.temp 4 t
+.temp 8 t1
+.temp 8 t2
+.param 4 p1
+.param 4 p2
+.param 4 p3
+
+ldreslinl b, s2, p2, p3
+storel d2, b
+loadl a, s1
+x4 convubw t1, a
+x4 convubw t2, b
+x4 subw t2, t2, t1
+x4 mullw t2, t2, p1
+x4 convhwb t, t2
+x4 addb d1, t, a
+
+
+
+.function gst_videoscale_orc_merge_bicubic_u8
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.source 1 s4 guint8
+.param 4 p1
+.param 4 p2
+.param 4 p3
+.param 4 p4
+.temp 2 t1
+.temp 2 t2
+
+mulubw t1, s2, p2
+mulubw t2, s3, p3
+addw t1, t1, t2
+mulubw t2, s1, p1
+subw t1, t1, t2
+mulubw t2, s4, p4
+subw t1, t1, t2
+addw t1, t1, 32
+shrsw t1, t1, 6
+convsuswb d1, t1
+
+
+
+#.init gst_adder_orc_init
+
+.function add_int32
+.dest 4 d1 gint32
+.source 4 s1 gint32
+
+addssl d1, d1, s1
+
+
+.function add_int16
+.dest 2 d1 gint16
+.source 2 s1 gint16
+
+addssw d1, d1, s1
+
+
+.function add_int8
+.dest 1 d1 gint8
+.source 1 s1 gint8
+
+addssb d1, d1, s1
+
+
+.function add_uint32
+.dest 4 d1 guint32
+.source 4 s1 guint32
+
+addusl d1, d1, s1
+
+
+.function add_uint16
+.dest 2 d1 guint16
+.source 2 s1 guint16
+
+addusw d1, d1, s1
+
+
+.function add_uint8
+.dest 1 d1 guint8
+.source 1 s1 guint8
+
+addusb d1, d1, s1
+
+
+.function add_float32
+.dest 4 d1 float
+.source 4 s1 float
+
+addf d1, d1, s1
+
+
+.function orc_splat_u32
+.dest 4 d1 guint32
+.param 4 p1 guint32
+
+copyl d1, p1
+
+.function orc_memcpy_u32
+.dest 4 d1 guint32
+.source 4 s1 guint32
+
+copyl d1, s1
+
+.function orc_blend_u8
+.flags 2d
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.param 2 p1
+.temp 2 t1
+.temp 2 t2
+.const 1 c1 8
+
+convubw t1, d1
+convubw t2, s1
+subw t2, t2, t1
+mullw t2, t2, p1
+shlw t1, t1, c1
+addw t2, t1, t2
+shruw t2, t2, c1
+convsuswb d1, t2
+
+
+.function orc_blend_argb
+.flags 2d
+.dest 4 d guint8
+.source 4 s guint8
+.param 2 alpha
+.temp 4 t
+.temp 2 tw
+.temp 1 tb
+.temp 4 a
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 a_wide
+.const 4 a_alpha 0x000000ff
+
+loadl t, s
+convlw tw, t
+convwb tb, tw
+splatbl a, tb
+x4 convubw a_wide, a
+x4 mullw a_wide, a_wide, alpha
+x4 shruw a_wide, a_wide, 8
+x4 convubw s_wide, t
+loadl t, d
+x4 convubw d_wide, t
+x4 subw s_wide, s_wide, d_wide
+x4 mullw s_wide, s_wide, a_wide
+x4 div255w s_wide, s_wide
+x4 addw d_wide, d_wide, s_wide
+x4 convwb t, d_wide
+orl t, t, a_alpha
+storel d, t
+
+.function orc_blend_bgra
+.flags 2d
+.dest 4 d guint8
+.source 4 s guint8
+.param 2 alpha
+.temp 4 t
+.temp 4 t2
+.temp 2 tw
+.temp 1 tb
+.temp 4 a
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 a_wide
+.const 4 a_alpha 0xff000000
+
+loadl t, s
+shrul t2, t, 24
+convlw tw, t2
+convwb tb, tw
+splatbl a, tb
+x4 convubw a_wide, a
+x4 mullw a_wide, a_wide, alpha
+x4 shruw a_wide, a_wide, 8
+x4 convubw s_wide, t
+loadl t, d
+x4 convubw d_wide, t
+x4 subw s_wide, s_wide, d_wide
+x4 mullw s_wide, s_wide, a_wide
+x4 div255w s_wide, s_wide
+x4 addw d_wide, d_wide, s_wide
+x4 convwb t, d_wide
+orl t, t, a_alpha
+storel d, t
+
+
+.function orc_splat_u32
+.dest 4 d1 guint32
+.param 4 p1
+
+copyl d1, p1
+
+
+.function deinterlace_line_vfir
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.source 1 s4 guint8
+.source 1 s5 guint8
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+convubw t1, s1
+convubw t2, s5
+addw t1, t1, t2
+convubw t2, s2
+convubw t3, s4
+addw t2, t2, t3
+shlw t2, t2, 2
+convubw t3, s3
+shlw t3, t3, 1
+subw t2, t2, t1
+addw t2, t2, t3
+addw t2, t2, 4
+shrsw t2, t2, 3
+convsuswb d1, t2
+
+
+.function deinterlace_line_linear
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+
+avgub d1, s1, s2
+
+
+.function deinterlace_line_linear_blend
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+convubw t1, s1
+convubw t2, s2
+convubw t3, s3
+addw t1, t1, t2
+addw t3, t3, t3
+addw t1, t1, t3
+addw t1, t1, 2
+shrsw t1, t1, 2
+convsuswb d1, t1
+
+
+.function deinterlace_line_greedy
+.dest 1 d1
+.source 1 m0
+.source 1 t1
+.source 1 b1
+.source 1 m2
+.param 1 max_comb
+.temp 1 tm0
+.temp 1 tm2
+.temp 1 tb1
+.temp 1 tt1
+.temp 1 avg
+.temp 1 l2_diff
+.temp 1 lp2_diff
+.temp 1 t2
+.temp 1 t3
+.temp 1 best
+.temp 1 min
+.temp 1 max
+
+
+loadb tm0, m0
+loadb tm2, m2
+
+loadb tb1, b1
+loadb tt1, t1
+avgub avg, tt1, tb1
+
+maxub t2, tm0, avg
+minub t3, tm0, avg
+subb l2_diff, t2, t3
+
+maxub t2, tm2, avg
+minub t3, tm2, avg
+subb lp2_diff, t2, t3
+
+xorb l2_diff, l2_diff, 0x80
+xorb lp2_diff, lp2_diff, 0x80
+cmpgtsb t3, l2_diff, lp2_diff
+
+andb t2, tm2, t3
+andnb t3, t3, tm0
+orb best, t2, t3
+
+maxub max, tt1, tb1
+minub min, tt1, tb1
+addusb max, max, max_comb
+subusb min, min, max_comb
+minub best, best, max
+maxub d1, best, min
+
+
+
+
+.function cogorc_memcpy_2d
+.flags 2d
+.dest 1 d1 guint8
+.source 1 s1 guint8
+
+copyb d1, s1
+
+
+.function cogorc_downsample_horiz_cosite_1tap
+.dest 1 d1 guint8
+.source 2 s1 guint8
+
+select0wb d1, s1
+
+
+.function cogorc_downsample_horiz_cosite_3tap
+.dest 1 d1 guint8
+.source 2 s1 guint8
+.source 2 s2 guint8
+.temp 1 t1
+.temp 1 t2
+.temp 1 t3
+.temp 2 t4
+.temp 2 t5
+.temp 2 t6
+
+copyw t4, s1
+select0wb t1, t4
+select1wb t2, t4
+select0wb t3, s2
+convubw t4, t1
+convubw t5, t2
+convubw t6, t3
+mullw t5, t5, 2
+addw t4, t4, t6
+addw t4, t4, t5
+addw t4, t4, 2
+shrsw t4, t4, 2
+convsuswb d1, t4
+
+
+.function cogorc_downsample_420_jpeg
+.dest 1 d1 guint8
+.source 2 s1 guint8
+.source 2 s2 guint8
+.temp 2 t1
+.temp 1 t2
+.temp 1 t3
+.temp 1 t4
+.temp 1 t5
+
+copyw t1, s1
+select0wb t2, t1
+select1wb t3, t1
+avgub t2, t2, t3
+copyw t1, s2
+select0wb t4, t1
+select1wb t5, t1
+avgub t4, t4, t5
+avgub d1, t2, t4
+
+
+.function cogorc_downsample_vert_halfsite_2tap
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+
+avgub d1, s1, s2
+
+
+.function cogorc_downsample_vert_cosite_3tap
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+convubw t1, s1
+convubw t2, s2
+convubw t3, s3
+mullw t2, t2, 2
+addw t1, t1, t3
+addw t1, t1, t2
+addw t1, t1, 2
+shrsw t1, t1, 2
+convsuswb d1, t1
+
+
+
+.function cogorc_downsample_vert_halfsite_4tap
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.source 1 s4 guint8
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+.temp 2 t4
+
+convubw t1, s1
+convubw t2, s2
+convubw t3, s3
+convubw t4, s4
+addw t2, t2, t3
+mullw t2, t2, 26
+addw t1, t1, t4
+mullw t1, t1, 6
+addw t2, t2, t1
+addw t2, t2, 32
+shrsw t2, t2, 6
+convsuswb d1, t2
+
+
+.function cogorc_upsample_horiz_cosite_1tap
+.dest 2 d1 guint8 guint8
+.source 1 s1 guint8
+.temp 1 t1
+
+copyb t1, s1
+mergebw d1, t1, t1
+
+
+.function cogorc_upsample_horiz_cosite
+.dest 2 d1 guint8 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.temp 1 t1
+.temp 1 t2
+
+copyb t1, s1
+avgub t2, t1, s2
+mergebw d1, t1, t2
+
+
+.function cogorc_upsample_vert_avgub
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+
+avgub d1, s1, s2
+
+
+
+
+.function orc_unpack_yuyv_y
+.dest 1 d1 guint8
+.source 2 s1 guint8
+
+select0wb d1, s1
+
+
+.function orc_unpack_yuyv_u
+.dest 1 d1 guint8
+.source 4 s1 guint8
+.temp 2 t1
+
+select0lw t1, s1
+select1wb d1, t1
+
+
+.function orc_unpack_yuyv_v
+.dest 1 d1 guint8
+.source 4 s1 guint8
+.temp 2 t1
+
+select1lw t1, s1
+select1wb d1, t1
+
+
+.function orc_pack_yuyv
+.dest 4 d1 guint8
+.source 2 s1 guint8 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.temp 1 t1
+.temp 1 t2
+.temp 2 t3
+.temp 2 t4
+.temp 2 t5
+
+copyw t5, s1
+select0wb t1, t5
+select1wb t2, t5
+mergebw t3, t1, s2
+mergebw t4, t2, s3
+mergewl d1, t3, t4
+
+
+.function orc_unpack_uyvy_y
+.dest 1 d1 guint8
+.source 2 s1 guint8
+
+select1wb d1, s1
+
+
+.function orc_unpack_uyvy_u
+.dest 1 d1 guint8
+.source 4 s1 guint8
+.temp 2 t1
+
+select0lw t1, s1
+select0wb d1, t1
+
+
+.function orc_unpack_uyvy_v
+.dest 1 d1 guint8
+.source 4 s1 guint8
+.temp 2 t1
+
+select1lw t1, s1
+select0wb d1, t1
+
+
+.function orc_pack_uyvy
+.dest 4 d1 guint8
+.source 2 s1 guint8 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.temp 1 t1
+.temp 1 t2
+.temp 2 t3
+.temp 2 t4
+.temp 2 t5
+
+copyw t5, s1
+select0wb t1, t5
+select1wb t2, t5
+mergebw t3, s2, t1
+mergebw t4, s3, t2
+mergewl d1, t3, t4
+
+
+.function orc_matrix2_u8
+.dest 1 d1 guint8 guint8
+.source 1 s1 guint8 guint8
+.source 1 s2 guint8 guint8
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t1, t1, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+addw t1, t1, p3
+shrsw t1, t1, 6
+convsuswb d1, t1
+
+
+.function orc_matrix2_11_u8
+.dest 1 d1 guint8 guint8
+.source 1 s1 guint8 guint8
+.source 1 s2 guint8 guint8
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+.temp 2 t4
+
+convubw t1, s1
+subw t1, t1, 16
+mullw t3, t1, p1
+convubw t2, s2
+subw t2, t2, 128
+mullw t4, t2, p2
+addw t3, t3, t4
+addw t3, t3, 128
+shrsw t3, t3, 8
+addw t3, t3, t1
+addw t3, t3, t2
+convsuswb d1, t3
+
+
+.function orc_matrix2_12_u8
+.dest 1 d1 guint8 guint8
+.source 1 s1 guint8 guint8
+.source 1 s2 guint8 guint8
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+.temp 2 t4
+
+convubw t1, s1
+subw t1, t1, 16
+mullw t3, t1, p1
+convubw t2, s2
+subw t2, t2, 128
+mullw t4, t2, p2
+addw t3, t3, t4
+addw t3, t3, 128
+shrsw t3, t3, 8
+addw t3, t3, t1
+addw t3, t3, t2
+addw t3, t3, t2
+convsuswb d1, t3
+
+
+.function orc_matrix3_u8
+.dest 1 d1 guint8 guint8
+.source 1 s1 guint8 guint8
+.source 1 s2 guint8 guint8
+.source 1 s3 guint8 guint8
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t1, t1, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+convubw t2, s3
+mullw t2, t2, p3
+addw t1, t1, t2
+addw t1, t1, p4
+shrsw t1, t1, 6
+convsuswb d1, t1
+
+
+.function orc_matrix3_100_u8
+.dest 1 d1 guint8 guint8
+.source 1 s1 guint8 guint8
+.source 1 s2 guint8 guint8
+.source 1 s3 guint8 guint8
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+#.temp 2 t4
+
+convubw t1, s1
+subw t1, t1, 16
+mullw t3, t1, p1
+convubw t2, s2
+subw t2, t2, 128
+mullw t2, t2, p2
+addw t3, t3, t2
+convubw t2, s3
+subw t2, t2, 128
+mullw t2, t2, p3
+addw t3, t3, t2
+addw t3, t3, 128
+shrsw t3, t3, 8
+addw t3, t3, t1
+convsuswb d1, t3
+
+
+.function orc_matrix3_100_offset_u8
+.dest 1 d1 guint8 guint8
+.source 1 s1 guint8 guint8
+.source 1 s2 guint8 guint8
+.source 1 s3 guint8 guint8
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.param 2 p5
+#.param 2 p6
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+#.temp 2 t3
+#.temp 2 t4
+
+convubw t3, s1
+mullw t1, t3, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+convubw t2, s3
+mullw t2, t2, p3
+addw t1, t1, t2
+addw t1, t1, p4
+shrsw t1, t1, p5
+#addw t1, t1, p6
+addw t1, t1, t3
+convsuswb d1, t1
+
+
+
+.function orc_matrix3_000_u8
+.dest 1 d1 guint8 guint8
+.source 1 s1 guint8 guint8
+.source 1 s2 guint8 guint8
+.source 1 s3 guint8 guint8
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.param 2 p5
+#.param 2 p6
+.temp 2 t1
+.temp 2 t2
+#.temp 2 t3
+#.temp 2 t4
+
+convubw t1, s1
+mullw t1, t1, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+convubw t2, s3
+mullw t2, t2, p3
+addw t1, t1, t2
+addw t1, t1, p4
+shrsw t1, t1, p5
+#addw t1, t1, p6
+convwb d1, t1
+
+
+
+.function orc_pack_123x
+.dest 4 d1 guint32 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.param 1 p1
+.temp 2 t1
+.temp 2 t2
+
+mergebw t1, s1, s2
+mergebw t2, s3, p1
+mergewl d1, t1, t2
+
+
+.function orc_pack_x123
+.dest 4 d1 guint32 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.source 1 s3 guint8
+.param 1 p1
+.temp 2 t1
+.temp 2 t2
+
+mergebw t1, p1, s1
+mergebw t2, s2, s3
+mergewl d1, t1, t2
+
+
+.function cogorc_combine2_u8
+.dest 1 d1 guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t1, t1, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+shruw t1, t1, 8
+convsuswb d1, t1
+
+
+.function cogorc_convert_I420_UYVY
+.dest 4 d1 guint8
+.dest 4 d2 guint8
+.source 2 y1 guint8
+.source 2 y2 guint8
+.source 1 u guint8
+.source 1 v guint8
+.temp 2 uv
+
+mergebw uv, u, v
+x2 mergebw d1, uv, y1
+x2 mergebw d2, uv, y2
+
+
+.function cogorc_convert_I420_YUY2
+.dest 4 d1 guint8
+.dest 4 d2 guint8
+.source 2 y1 guint8
+.source 2 y2 guint8
+.source 1 u guint8
+.source 1 v guint8
+.temp 2 uv
+
+mergebw uv, u, v
+x2 mergebw d1, y1, uv
+x2 mergebw d2, y2, uv
+
+
+
+.function cogorc_convert_I420_AYUV
+.dest 4 d1 guint8
+.dest 4 d2 guint8
+.source 1 y1 guint8
+.source 1 y2 guint8
+.source 1 u guint8
+.source 1 v guint8
+.const 1 c255 255
+.temp 2 uv
+.temp 2 ay
+.temp 1 tu
+.temp 1 tv
+
+loadupdb tu, u
+loadupdb tv, v
+mergebw uv, tu, tv
+mergebw ay, c255, y1
+mergewl d1, ay, uv
+mergebw ay, c255, y2
+mergewl d2, ay, uv
+
+
+.function cogorc_convert_YUY2_I420
+.dest 2 y1 guint8
+.dest 2 y2 guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 4 yuv1 guint8
+.source 4 yuv2 guint8
+.temp 2 t1
+.temp 2 t2
+.temp 2 ty
+
+x2 splitwb t1, ty, yuv1
+storew y1, ty
+x2 splitwb t2, ty, yuv2
+storew y2, ty
+x2 avgub t1, t1, t2
+splitwb v, u, t1
+
+
+.function cogorc_convert_UYVY_YUY2
+.flags 2d
+.dest 4 yuy2 guint8
+.source 4 uyvy guint8
+
+x2 swapw yuy2, uyvy
+
+
+.function cogorc_planar_chroma_420_422
+.flags 2d
+.dest 1 d1 guint8
+.dest 1 d2 guint8
+.source 1 s guint8
+
+copyb d1, s
+copyb d2, s
+
+
+.function cogorc_planar_chroma_420_444
+.flags 2d
+.dest 2 d1 guint8
+.dest 2 d2 guint8
+.source 1 s guint8
+.temp 2 t
+
+splatbw t, s
+storew d1, t
+storew d2, t
+
+
+.function cogorc_planar_chroma_422_444
+.flags 2d
+.dest 2 d1 guint8
+.source 1 s guint8
+.temp 2 t
+
+splatbw t, s
+storew d1, t
+
+
+.function cogorc_planar_chroma_444_422
+.flags 2d
+.dest 1 d guint8
+.source 2 s guint8
+.temp 1 t1
+.temp 1 t2
+
+splitwb t1, t2, s
+avgub d, t1, t2
+
+
+.function cogorc_planar_chroma_444_420
+.flags 2d
+.dest 1 d guint8
+.source 2 s1 guint8
+.source 2 s2 guint8
+.temp 2 t
+.temp 1 t1
+.temp 1 t2
+
+x2 avgub t, s1, s2
+splitwb t1, t2, t
+avgub d, t1, t2
+
+
+.function cogorc_planar_chroma_422_420
+.flags 2d
+.dest 1 d guint8
+.source 1 s1 guint8
+.source 1 s2 guint8
+
+avgub d, s1, s2
+
+
+.function cogorc_convert_YUY2_AYUV
+.flags 2d
+.dest 8 ayuv guint8
+.source 4 yuy2 guint8
+.const 2 c255 0xff
+.temp 2 yy
+.temp 2 uv
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitwb uv, yy, yuy2
+x2 mergebw ayay, c255, yy
+mergewl uvuv, uv, uv
+x2 mergewl ayuv, ayay, uvuv
+
+
+.function cogorc_convert_UYVY_AYUV
+.flags 2d
+.dest 8 ayuv guint8
+.source 4 uyvy guint8
+.const 2 c255 0xff
+.temp 2 yy
+.temp 2 uv
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitwb yy, uv, uyvy
+x2 mergebw ayay, c255, yy
+mergewl uvuv, uv, uv
+x2 mergewl ayuv, ayay, uvuv
+
+
+.function cogorc_convert_YUY2_Y42B
+.flags 2d
+.dest 2 y guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 4 yuy2 guint8
+.temp 2 uv
+
+x2 splitwb uv, y, yuy2
+splitwb v, u, uv
+
+
+.function cogorc_convert_UYVY_Y42B
+.flags 2d
+.dest 2 y guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 4 uyvy guint8
+.temp 2 uv
+
+x2 splitwb y, uv, uyvy
+splitwb v, u, uv
+
+
+.function cogorc_convert_YUY2_Y444
+.flags 2d
+.dest 2 y guint8
+.dest 2 uu guint8
+.dest 2 vv guint8
+.source 4 yuy2 guint8
+.temp 2 uv
+.temp 1 u
+.temp 1 v
+
+x2 splitwb uv, y, yuy2
+splitwb v, u, uv
+splatbw uu, u
+splatbw vv, v
+
+
+.function cogorc_convert_UYVY_Y444
+.flags 2d
+.dest 2 y guint8
+.dest 2 uu guint8
+.dest 2 vv guint8
+.source 4 uyvy guint8
+.temp 2 uv
+.temp 1 u
+.temp 1 v
+
+x2 splitwb y, uv, uyvy
+splitwb v, u, uv
+splatbw uu, u
+splatbw vv, v
+
+
+.function cogorc_convert_UYVY_I420
+.dest 2 y1 guint8
+.dest 2 y2 guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 4 yuv1 guint8
+.source 4 yuv2 guint8
+.temp 2 t1
+.temp 2 t2
+.temp 2 ty
+
+x2 splitwb ty, t1, yuv1
+storew y1, ty
+x2 splitwb ty, t2, yuv2
+storew y2, ty
+x2 avgub t1, t1, t2
+splitwb v, u, t1
+
+
+
+.function cogorc_convert_AYUV_I420
+.flags 2d
+.dest 2 y1 guint8
+.dest 2 y2 guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 8 ayuv1 guint8
+.source 8 ayuv2 guint8
+.temp 4 ay
+.temp 4 uv1
+.temp 4 uv2
+.temp 4 uv
+.temp 2 uu
+.temp 2 vv
+.temp 1 t1
+.temp 1 t2
+
+x2 splitlw uv1, ay, ayuv1
+x2 select1wb y1, ay
+x2 splitlw uv2, ay, ayuv2
+x2 select1wb y2, ay
+x4 avgub uv, uv1, uv2
+x2 splitwb vv, uu, uv
+splitwb t1, t2, uu
+avgub u, t1, t2
+splitwb t1, t2, vv
+avgub v, t1, t2
+
+
+
+.function cogorc_convert_AYUV_YUY2
+.flags 2d
+.dest 4 yuy2 guint8
+.source 8 ayuv guint8
+.temp 2 yy
+.temp 2 uv1
+.temp 2 uv2
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitlw uvuv, ayay, ayuv
+splitlw uv1, uv2, uvuv
+x2 avgub uv1, uv1, uv2
+x2 select1wb yy, ayay
+x2 mergebw yuy2, yy, uv1
+
+
+.function cogorc_convert_AYUV_UYVY
+.flags 2d
+.dest 4 yuy2 guint8
+.source 8 ayuv guint8
+.temp 2 yy
+.temp 2 uv1
+.temp 2 uv2
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitlw uvuv, ayay, ayuv
+splitlw uv1, uv2, uvuv
+x2 avgub uv1, uv1, uv2
+x2 select1wb yy, ayay
+x2 mergebw yuy2, uv1, yy
+
+
+
+.function cogorc_convert_AYUV_Y42B
+.flags 2d
+.dest 2 y guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 8 ayuv guint8
+.temp 4 ayay
+.temp 4 uvuv
+.temp 2 uv1
+.temp 2 uv2
+
+x2 splitlw uvuv, ayay, ayuv
+splitlw uv1, uv2, uvuv
+x2 avgub uv1, uv1, uv2
+splitwb v, u, uv1
+x2 select1wb y, ayay
+
+
+.function cogorc_convert_AYUV_Y444
+.flags 2d
+.dest 1 y guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 4 ayuv guint8
+.temp 2 ay
+.temp 2 uv
+
+splitlw uv, ay, ayuv
+splitwb v, u, uv
+select1wb y, ay
+
+
+.function cogorc_convert_Y42B_YUY2
+.flags 2d
+.dest 4 yuy2 guint8
+.source 2 y guint8
+.source 1 u guint8
+.source 1 v guint8
+.temp 2 uv
+
+mergebw uv, u, v
+x2 mergebw yuy2, y, uv
+
+
+.function cogorc_convert_Y42B_UYVY
+.flags 2d
+.dest 4 uyvy guint8
+.source 2 y guint8
+.source 1 u guint8
+.source 1 v guint8
+.temp 2 uv
+
+mergebw uv, u, v
+x2 mergebw uyvy, uv, y
+
+
+.function cogorc_convert_Y42B_AYUV
+.flags 2d
+.dest 8 ayuv guint8
+.source 2 yy guint8
+.source 1 u guint8
+.source 1 v guint8
+.const 1 c255 255
+.temp 2 uv
+.temp 2 ay
+.temp 4 uvuv
+.temp 4 ayay
+
+mergebw uv, u, v
+x2 mergebw ayay, c255, yy
+mergewl uvuv, uv, uv
+x2 mergewl ayuv, ayay, uvuv
+
+
+.function cogorc_convert_Y444_YUY2
+.flags 2d
+.dest 4 yuy2 guint8
+.source 2 y guint8
+.source 2 u guint8
+.source 2 v guint8
+.temp 2 uv
+.temp 4 uvuv
+.temp 2 uv1
+.temp 2 uv2
+
+x2 mergebw uvuv, u, v
+splitlw uv1, uv2, uvuv
+x2 avgub uv, uv1, uv2
+x2 mergebw yuy2, y, uv
+
+
+.function cogorc_convert_Y444_UYVY
+.flags 2d
+.dest 4 uyvy guint8
+.source 2 y guint8
+.source 2 u guint8
+.source 2 v guint8
+.temp 2 uv
+.temp 4 uvuv
+.temp 2 uv1
+.temp 2 uv2
+
+x2 mergebw uvuv, u, v
+splitlw uv1, uv2, uvuv
+x2 avgub uv, uv1, uv2
+x2 mergebw uyvy, uv, y
+
+
+.function cogorc_convert_Y444_AYUV
+.flags 2d
+.dest 4 ayuv guint8
+.source 1 yy guint8
+.source 1 u guint8
+.source 1 v guint8
+.const 1 c255 255
+.temp 2 uv
+.temp 2 ay
+
+mergebw uv, u, v
+mergebw ay, c255, yy
+mergewl ayuv, ay, uv
+
+
+
+.function cogorc_convert_AYUV_ARGB
+.flags 2d
+.dest 4 argb guint8
+.source 4 ayuv guint8
+.temp 2 t1
+.temp 2 t2
+.temp 1 a
+.temp 1 y
+.temp 1 u
+.temp 1 v
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 4 x
+.const 1 c8 8
+
+x4 subb x, ayuv, 128
+splitlw t1, t2, x
+splitwb y, a, t2
+splitwb v, u, t1
+convsbw wy, y
+convsbw wu, u
+convsbw wv, v
+
+mullw t1, wy, 42
+shrsw t1, t1, c8
+addssw wy, wy, t1
+
+addssw wr, wy, wv
+mullw t1, wv, 103
+shrsw t1, t1, c8
+subssw wr, wr, t1
+addssw wr, wr, wv
+
+addssw wb, wy, wu
+addssw wb, wb, wu
+mullw t1, wu, 4
+shrsw t1, t1, c8
+addssw wb, wb, t1
+
+mullw t1, wu, 100
+shrsw t1, t1, c8
+subssw wg, wy, t1
+mullw t1, wv, 104
+shrsw t1, t1, c8
+subssw wg, wg, t1
+subssw wg, wg, t1
+
+convssswb r, wr
+convssswb g, wg
+convssswb b, wb
+
+mergebw t1, a, r
+mergebw t2, g, b
+mergewl x, t1, t2
+x4 addb argb, x, 128
+
+
+
+.function cogorc_convert_AYUV_BGRA
+.flags 2d
+.dest 4 argb guint8
+.source 4 ayuv guint8
+.temp 2 t1
+.temp 2 t2
+.temp 1 a
+.temp 1 y
+.temp 1 u
+.temp 1 v
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 4 x
+.const 1 c8 8
+
+x4 subb x, ayuv, 128
+splitlw t1, t2, x
+splitwb y, a, t2
+splitwb v, u, t1
+convsbw wy, y
+convsbw wu, u
+convsbw wv, v
+
+mullw t1, wy, 42
+shrsw t1, t1, c8
+addssw wy, wy, t1
+
+addssw wr, wy, wv
+mullw t1, wv, 103
+shrsw t1, t1, c8
+subssw wr, wr, t1
+addssw wr, wr, wv
+
+addssw wb, wy, wu
+addssw wb, wb, wu
+mullw t1, wu, 4
+shrsw t1, t1, c8
+addssw wb, wb, t1
+
+mullw t1, wu, 100
+shrsw t1, t1, c8
+subssw wg, wy, t1
+mullw t1, wv, 104
+shrsw t1, t1, c8
+subssw wg, wg, t1
+subssw wg, wg, t1
+
+convssswb r, wr
+convssswb g, wg
+convssswb b, wb
+
+mergebw t1, b, g
+mergebw t2, r, a
+mergewl x, t1, t2
+x4 addb argb, x, 128
+
+
+
+
+.function cogorc_convert_AYUV_ABGR
+.flags 2d
+.dest 4 argb guint8
+.source 4 ayuv guint8
+.temp 2 t1
+.temp 2 t2
+.temp 1 a
+.temp 1 y
+.temp 1 u
+.temp 1 v
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 4 x
+.const 1 c8 8
+
+x4 subb x, ayuv, 128
+splitlw t1, t2, x
+splitwb y, a, t2
+splitwb v, u, t1
+convsbw wy, y
+convsbw wu, u
+convsbw wv, v
+
+mullw t1, wy, 42
+shrsw t1, t1, c8
+addssw wy, wy, t1
+
+addssw wr, wy, wv
+mullw t1, wv, 103
+shrsw t1, t1, c8
+subssw wr, wr, t1
+addssw wr, wr, wv
+
+addssw wb, wy, wu
+addssw wb, wb, wu
+mullw t1, wu, 4
+shrsw t1, t1, c8
+addssw wb, wb, t1
+
+mullw t1, wu, 100
+shrsw t1, t1, c8
+subssw wg, wy, t1
+mullw t1, wv, 104
+shrsw t1, t1, c8
+subssw wg, wg, t1
+subssw wg, wg, t1
+
+convssswb r, wr
+convssswb g, wg
+convssswb b, wb
+
+mergebw t1, a, b
+mergebw t2, g, r
+mergewl x, t1, t2
+x4 addb argb, x, 128
+
+
+
+.function cogorc_convert_AYUV_RGBA
+.flags 2d
+.dest 4 argb guint8
+.source 4 ayuv guint8
+.temp 2 t1
+.temp 2 t2
+.temp 1 a
+.temp 1 y
+.temp 1 u
+.temp 1 v
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 4 x
+.const 1 c8 8
+
+x4 subb x, ayuv, 128
+splitlw t1, t2, x
+splitwb y, a, t2
+splitwb v, u, t1
+convsbw wy, y
+convsbw wu, u
+convsbw wv, v
+
+mullw t1, wy, 42
+shrsw t1, t1, c8
+addssw wy, wy, t1
+
+addssw wr, wy, wv
+mullw t1, wv, 103
+shrsw t1, t1, c8
+subssw wr, wr, t1
+addssw wr, wr, wv
+
+addssw wb, wy, wu
+addssw wb, wb, wu
+mullw t1, wu, 4
+shrsw t1, t1, c8
+addssw wb, wb, t1
+
+mullw t1, wu, 100
+shrsw t1, t1, c8
+subssw wg, wy, t1
+mullw t1, wv, 104
+shrsw t1, t1, c8
+subssw wg, wg, t1
+subssw wg, wg, t1
+
+convssswb r, wr
+convssswb g, wg
+convssswb b, wb
+
+mergebw t1, r, g
+mergebw t2, b, a
+mergewl x, t1, t2
+x4 addb argb, x, 128
+
+
+
+.function cogorc_convert_I420_BGRA
+.dest 4 argb guint8
+.source 1 y guint8
+.source 1 u guint8
+.source 1 v guint8
+.temp 2 t1
+.temp 2 t2
+.temp 1 t3
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 4 x
+.const 1 c8 8
+.const 1 c128 128
+
+subb t3, y, c128
+convsbw wy, t3
+loadupib t3, u
+subb t3, t3, c128
+convsbw wu, t3
+loadupib t3, v
+subb t3, t3, c128
+convsbw wv, t3
+
+mullw t1, wy, 42
+shrsw t1, t1, c8
+addssw wy, wy, t1
+
+addssw wr, wy, wv
+mullw t1, wv, 103
+shrsw t1, t1, c8
+subssw wr, wr, t1
+addssw wr, wr, wv
+
+addssw wb, wy, wu
+addssw wb, wb, wu
+mullw t1, wu, 4
+shrsw t1, t1, c8
+addssw wb, wb, t1
+
+mullw t1, wu, 100
+shrsw t1, t1, c8
+subssw wg, wy, t1
+mullw t1, wv, 104
+shrsw t1, t1, c8
+subssw wg, wg, t1
+subssw wg, wg, t1
+
+convssswb r, wr
+convssswb g, wg
+convssswb b, wb
+
+mergebw t1, b, g
+mergebw t2, r, 255
+mergewl x, t1, t2
+x4 addb argb, x, c128
+
+
+
+.function cogorc_convert_I420_BGRA_avg
+.dest 4 argb guint8
+.source 1 y guint8
+.source 1 u1 guint8
+.source 1 u2 guint8
+.source 1 v1 guint8
+.source 1 v2 guint8
+.temp 2 t1
+.temp 2 t2
+.temp 1 t3
+.temp 1 t4
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 4 x
+.const 1 c8 8
+.const 1 c128 128
+
+subb t3, y, c128
+convsbw wy, t3
+loadupib t3, u1
+loadupib t4, u2
+avgub t3, t3, t4
+subb t3, t3, c128
+convsbw wu, t3
+loadupib t3, v1
+loadupib t4, v2
+avgub t3, t3, t4
+subb t3, t3, c128
+convsbw wv, t3
+
+mullw t1, wy, 42
+shrsw t1, t1, c8
+addssw wy, wy, t1
+
+addssw wr, wy, wv
+mullw t1, wv, 103
+shrsw t1, t1, c8
+subssw wr, wr, t1
+addssw wr, wr, wv
+
+addssw wb, wy, wu
+addssw wb, wb, wu
+mullw t1, wu, 4
+shrsw t1, t1, c8
+addssw wb, wb, t1
+
+mullw t1, wu, 100
+shrsw t1, t1, c8
+subssw wg, wy, t1
+mullw t1, wv, 104
+shrsw t1, t1, c8
+subssw wg, wg, t1
+subssw wg, wg, t1
+
+convssswb r, wr
+convssswb g, wg
+convssswb b, wb
+
+mergebw t1, b, g
+mergebw t2, r, 255
+mergewl x, t1, t2
+x4 addb argb, x, c128
+
+
+
+.function cogorc_getline_I420
+.dest 4 d guint8
+.source 1 y guint8
+.source 1 u guint8
+.source 1 v guint8
+.const 1 c255 255
+.temp 2 uv
+.temp 2 ay
+.temp 1 tu
+.temp 1 tv
+
+loadupdb tu, u
+loadupdb tv, v
+mergebw uv, tu, tv
+mergebw ay, c255, y
+mergewl d, ay, uv
+
+
+.function cogorc_getline_YUY2
+.dest 8 ayuv guint8
+.source 4 yuy2 guint8
+.const 2 c255 0xff
+.temp 2 yy
+.temp 2 uv
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitwb uv, yy, yuy2
+x2 mergebw ayay, c255, yy
+mergewl uvuv, uv, uv
+x2 mergewl ayuv, ayay, uvuv
+
+
+.function cogorc_getline_UYVY
+.dest 8 ayuv guint8
+.source 4 uyvy guint8
+.const 2 c255 0xff
+.temp 2 yy
+.temp 2 uv
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitwb yy, uv, uyvy
+x2 mergebw ayay, c255, yy
+mergewl uvuv, uv, uv
+x2 mergewl ayuv, ayay, uvuv
+
+
+.function cogorc_getline_YVYU
+.dest 8 ayuv guint8
+.source 4 uyvy guint8
+.const 2 c255 0xff
+.temp 2 yy
+.temp 2 uv
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitwb yy, uv, uyvy
+x2 mergebw ayay, c255, yy
+mergewl uvuv, uv, uv
+x2 mergewl ayuv, ayay, uvuv
+
+
+.function cogorc_getline_Y42B
+.dest 8 ayuv guint8
+.source 2 yy guint8
+.source 1 u guint8
+.source 1 v guint8
+.const 1 c255 255
+.temp 2 uv
+.temp 2 ay
+.temp 4 uvuv
+.temp 4 ayay
+
+mergebw uv, u, v
+x2 mergebw ayay, c255, yy
+mergewl uvuv, uv, uv
+x2 mergewl ayuv, ayay, uvuv
+
+
+.function cogorc_getline_Y444
+.dest 4 ayuv guint8
+.source 1 y guint8
+.source 1 u guint8
+.source 1 v guint8
+.const 1 c255 255
+.temp 2 uv
+.temp 2 ay
+
+mergebw uv, u, v
+mergebw ay, c255, y
+mergewl ayuv, ay, uv
+
+
+.function cogorc_getline_Y800
+.dest 4 ayuv guint8
+.source 1 y guint8
+.const 1 c255 255
+.const 2 c0xffff 0xffff
+.temp 2 ay
+
+mergebw ay, c255, y
+mergewl ayuv, ay, c0xffff
+
+
+.function cogorc_getline_BGRA
+.dest 4 argb guint8
+.source 4 bgra guint8
+
+swapl argb, bgra
+
+
+.function cogorc_getline_ABGR
+.dest 4 argb guint8
+.source 4 abgr guint8
+.temp 1 a
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 2 gr
+.temp 2 ab
+.temp 2 ar
+.temp 2 gb
+
+splitlw gr, ab, abgr
+splitwb r, g, gr
+splitwb b, a, ab
+mergebw ar, a, r
+mergebw gb, g, b
+mergewl argb, ar, gb
+
+
+.function cogorc_getline_RGBA
+.dest 4 argb guint8
+.source 4 rgba guint8
+.temp 1 a
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 2 rg
+.temp 2 ba
+.temp 2 ar
+.temp 2 gb
+
+splitlw ba, rg, rgba
+splitwb g, r, rg
+splitwb a, b, ba
+mergebw ar, a, r
+mergebw gb, g, b
+mergewl argb, ar, gb
+
+
+.function cogorc_getline_NV12
+.dest 8 d guint8
+.source 2 y guint8
+.source 2 uv guint8
+.const 1 c255 255
+.temp 4 ay
+.temp 4 uvuv
+
+mergewl uvuv, uv, uv
+x2 mergebw ay, c255, y
+x2 mergewl d, ay, uvuv
+
+
+.function cogorc_getline_NV21
+.dest 8 d guint8
+.source 2 y guint8
+.source 2 vu guint8
+.const 1 c255 255
+.temp 2 uv
+.temp 4 ay
+.temp 4 uvuv
+
+swapw uv, vu
+mergewl uvuv, uv, uv
+x2 mergebw ay, c255, y
+x2 mergewl d, ay, uvuv
+
+
+.function cogorc_putline_I420
+.dest 2 y guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 8 ayuv guint8
+.temp 4 ay
+.temp 4 uv
+.temp 2 uu
+.temp 2 vv
+.temp 1 t1
+.temp 1 t2
+
+x2 splitlw uv, ay, ayuv
+x2 select1wb y, ay
+x2 splitwb vv, uu, uv
+splitwb t1, t2, uu
+avgub u, t1, t2
+splitwb t1, t2, vv
+avgub v, t1, t2
+
+
+
+.function cogorc_putline_YUY2
+.dest 4 yuy2 guint8
+.source 8 ayuv guint8
+.temp 2 yy
+.temp 2 uv1
+.temp 2 uv2
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitlw uvuv, ayay, ayuv
+splitlw uv1, uv2, uvuv
+x2 avgub uv1, uv1, uv2
+x2 select1wb yy, ayay
+x2 mergebw yuy2, yy, uv1
+
+
+.function cogorc_putline_YVYU
+.dest 4 yuy2 guint8
+.source 8 ayuv guint8
+.temp 2 yy
+.temp 2 uv1
+.temp 2 uv2
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitlw uvuv, ayay, ayuv
+splitlw uv1, uv2, uvuv
+x2 avgub uv1, uv1, uv2
+x2 select1wb yy, ayay
+swapw uv1, uv1
+x2 mergebw yuy2, yy, uv1
+
+
+.function cogorc_putline_UYVY
+.dest 4 yuy2 guint8
+.source 8 ayuv guint8
+.temp 2 yy
+.temp 2 uv1
+.temp 2 uv2
+.temp 4 ayay
+.temp 4 uvuv
+
+x2 splitlw uvuv, ayay, ayuv
+splitlw uv1, uv2, uvuv
+x2 avgub uv1, uv1, uv2
+x2 select1wb yy, ayay
+x2 mergebw yuy2, uv1, yy
+
+
+
+.function cogorc_putline_Y42B
+.dest 2 y guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 8 ayuv guint8
+.temp 4 ayay
+.temp 4 uvuv
+.temp 2 uv1
+.temp 2 uv2
+
+x2 splitlw uvuv, ayay, ayuv
+splitlw uv1, uv2, uvuv
+x2 avgub uv1, uv1, uv2
+splitwb v, u, uv1
+x2 select1wb y, ayay
+
+
+.function cogorc_putline_Y444
+.dest 1 y guint8
+.dest 1 u guint8
+.dest 1 v guint8
+.source 4 ayuv guint8
+.temp 2 ay
+.temp 2 uv
+
+splitlw uv, ay, ayuv
+splitwb v, u, uv
+select1wb y, ay
+
+
+.function cogorc_putline_Y800
+.dest 1 y guint8
+.source 4 ayuv guint8
+.temp 2 ay
+
+select0lw ay, ayuv
+select1wb y, ay
+
+
+.function cogorc_putline_BGRA
+.dest 4 bgra guint8
+.source 4 argb guint8
+
+swapl bgra, argb
+
+
+.function cogorc_putline_ABGR
+.dest 4 abgr guint8
+.source 4 argb guint8
+.temp 1 a
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 2 gr
+.temp 2 ab
+.temp 2 ar
+.temp 2 gb
+
+splitlw gb, ar, argb
+splitwb b, g, gb
+splitwb r, a, ar
+mergebw ab, a, b
+mergebw gr, g, r
+mergewl abgr, ab, gr
+
+
+.function cogorc_putline_RGBA
+.dest 4 rgba guint8
+.source 4 argb guint8
+.temp 1 a
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 2 rg
+.temp 2 ba
+.temp 2 ar
+.temp 2 gb
+
+splitlw gb, ar, argb
+splitwb b, g, gb
+splitwb r, a, ar
+mergebw ba, b, a
+mergebw rg, r, g
+mergewl rgba, rg, ba
+
+
+.function cogorc_putline_NV12
+.dest 2 y guint8
+.dest 2 uv guint8
+.source 8 ayuv guint8
+.temp 4 ay
+.temp 4 uvuv
+.temp 2 uv1
+.temp 2 uv2
+
+x2 splitlw uvuv, ay, ayuv
+x2 select1wb y, ay
+splitlw uv1, uv2, uvuv
+x2 avgub uv, uv1, uv2
+
+
+.function cogorc_putline_NV21
+.dest 2 y guint8
+.dest 2 vu guint8
+.source 8 ayuv guint8
+.temp 4 ay
+.temp 4 uvuv
+.temp 2 uv1
+.temp 2 uv2
+.temp 2 uv
+
+x2 splitlw uvuv, ay, ayuv
+x2 select1wb y, ay
+splitlw uv1, uv2, uvuv
+x2 avgub uv, uv1, uv2
+swapw vu, uv
+
+
+
+#.init schro_orc_init
+
+.function orc_add2_rshift_add_s16_22_op
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.source 2 s3 int16_t
+.temp 2 t1
+
+addw t1, s2, s3
+addw t1, t1, 2
+shrsw t1, t1, 2
+addw d1, s1, t1
+
+
+.function orc_add2_rshift_add_s16_22
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+
+addw t1, s1, s2
+addw t1, t1, 2
+shrsw t1, t1, 2
+addw d1, d1, t1
+
+
+.function orc_add2_rshift_sub_s16_22_op
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.source 2 s3 int16_t
+.temp 2 t1
+
+addw t1, s2, s3
+addw t1, t1, 2
+shrsw t1, t1, 2
+subw d1, s1, t1
+
+
+.function orc_add2_rshift_sub_s16_22
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+
+addw t1, s1, s2
+addw t1, t1, 2
+shrsw t1, t1, 2
+subw d1, d1, t1
+
+
+.function orc_add2_rshift_add_s16_11_op
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.source 2 s3 int16_t
+.temp 2 t1
+
+avgsw t1, s2, s3
+addw d1, s1, t1
+
+
+.function orc_add2_rshift_add_s16_11
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+
+avgsw t1, s1, s2
+addw d1, d1, t1
+
+
+.function orc_add2_rshift_sub_s16_11_op
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.source 2 s3 int16_t
+.temp 2 t1
+
+avgsw t1, s2, s3
+subw d1, s1, t1
+
+
+.function orc_add2_rshift_sub_s16_11
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+
+avgsw t1, s1, s2
+subw d1, d1, t1
+
+
+.function orc_add_const_rshift_s16_11
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.temp 2 t1
+
+addw t1, s1, 1
+shrsw d1, t1, 1
+
+
+.function orc_add_const_rshift_s16
+.dest 2 d1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+
+addw t1, d1, p1
+shrsw d1, t1, p2
+
+
+.function orc_add_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+
+addw d1, s1, s2
+
+
+.function orc_add_s16_2d
+.flags 2d
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+
+addw d1, d1, s1
+
+
+.function orc_addc_rshift_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.param 2 p1
+
+addw t1, s1, s2
+shrsw d1, t1, p1
+
+
+.function orc_lshift1_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+
+shlw d1, s1, 1
+
+
+.function orc_lshift2_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+
+shlw d1, s1, 2
+
+
+.function orc_lshift_s16_ip
+.dest 2 d1 int16_t
+.param 2 p1
+
+shlw d1, d1, p1
+
+
+.function orc_mas2_add_s16_op
+.dest 2 d1 int16_t
+.source 2 s0 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 4 t2
+.param 2 p1
+.param 4 p2
+.param 4 p3
+
+addw t1, s1, s2
+mulswl t2, t1, p1
+addl t2, t2, p2
+shrsl t2, t2, p3
+convlw t1, t2
+addw d1, s0, t1
+
+
+.function orc_mas2_add_s16_ip
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 4 t2
+.param 2 p1
+.param 4 p2
+.param 4 p3
+
+addw t1, s1, s2
+mulswl t2, t1, p1
+addl t2, t2, p2
+shrsl t2, t2, p3
+convlw t1, t2
+addw d1, d1, t1
+
+
+.function orc_mas2_sub_s16_op
+.dest 2 d1 int16_t
+.source 2 s0 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 4 t2
+.param 2 p1
+.param 4 p2
+.param 4 p3
+
+addw t1, s1, s2
+mulswl t2, t1, p1
+addl t2, t2, p2
+shrsl t2, t2, p3
+convlw t1, t2
+subw d1, s0, t1
+
+
+.function orc_mas2_sub_s16_ip
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 4 t2
+.param 2 p1
+.param 4 p2
+.param 4 p3
+
+addw t1, s1, s2
+mulswl t2, t1, p1
+addl t2, t2, p2
+shrsl t2, t2, p3
+convlw t1, t2
+subw d1, d1, t1
+
+
+.function orc_mas4_across_add_s16_1991_op
+.dest 2 d1 int16_t
+.source 2 s0 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.source 2 s3 int16_t
+.source 2 s4 int16_t
+.param 4 p1
+.param 4 p2
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+.temp 4 t4
+
+addw t1, s2, s3
+mulswl t3, t1, 9
+addw t2, s1, s4
+convswl t4, t2
+subl t3, t3, t4
+addl t3, t3, p1
+shrsl t3, t3, p2
+convlw t1, t3
+addw d1, s0, t1
+
+
+.function orc_mas4_across_add_s16_1991_ip
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.param 4 p1
+.param 4 p2
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+.temp 4 t4
+
+loadoffw t1, s1, 1
+loadoffw t2, s1, 2
+addw t1, t1, t2
+mulswl t3, t1, 9
+loadw t1, s1
+loadoffw t2, s1, 3
+addw t2, t1, t2
+convswl t4, t2
+subl t3, t3, t4
+addl t3, t3, p1
+shrsl t3, t3, p2
+convlw t1, t3
+addw d1, d1, t1
+
+
+.function orc_mas4_across_sub_s16_1991_op
+.dest 2 d1 int16_t
+.source 2 s0 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.source 2 s3 int16_t
+.source 2 s4 int16_t
+.param 4 p1
+.param 4 p2
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+.temp 4 t4
+
+addw t1, s2, s3
+mulswl t3, t1, 9
+addw t2, s1, s4
+convswl t4, t2
+subl t3, t3, t4
+addl t3, t3, p1
+shrsl t3, t3, p2
+convlw t1, t3
+subw d1, s0, t1
+
+
+.function orc_mas4_across_sub_s16_1991_ip
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.param 4 p1
+.param 4 p2
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+.temp 4 t4
+
+loadoffw t1, s1, 1
+loadoffw t2, s1, 2
+addw t1, t1, t2
+mulswl t3, t1, 9
+loadw t1, s1
+loadoffw t2, s1, 3
+addw t2, t1, t2
+convswl t4, t2
+subl t3, t3, t4
+addl t3, t3, p1
+shrsl t3, t3, p2
+convlw t1, t3
+subw d1, d1, t1
+
+
+.function orc_subtract_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+
+subw d1, s1, s2
+
+
+.function orc_add_s16_u8
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 1 s2
+.temp 2 t1
+
+convubw t1, s2
+addw d1, t1, s1
+
+
+.function orc_add_s16_u8_2d
+.flags 2d
+.dest 2 d1 int16_t
+.source 1 s1
+.temp 2 t1
+
+convubw t1, s1
+addw d1, d1, t1
+
+
+.function orc_convert_s16_u8
+.dest 2 d1
+.source 1 s1
+
+convubw d1, s1
+
+
+.function orc_convert_u8_s16
+.dest 1 d1
+.source 2 s1 int16_t
+
+convsuswb d1, s1
+
+
+.function orc_offsetconvert_u8_s16
+.dest 1 d1
+.source 2 s1 int16_t
+.temp 2 t1
+
+addw t1, s1, 128
+convsuswb d1, t1
+
+
+.function orc_offsetconvert_s16_u8
+.dest 2 d1 int16_t
+.source 1 s1
+.temp 2 t1
+
+convubw t1, s1
+subw d1, t1, 128
+
+
+.function orc_subtract_s16_u8
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 1 s2
+.temp 2 t1
+
+convubw t1, s2
+subw d1, s1, t1
+
+
+.function orc_multiply_and_add_s16_u8
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 1 s2
+.temp 2 t1
+
+convubw t1, s2
+mullw t1, t1, s1
+addw d1, d1, t1
+
+
+.function orc_splat_s16_ns
+.dest 2 d1 int16_t
+.param 2 p1
+
+copyw d1, p1
+
+
+.function orc_splat_s16_2d_4xn
+.n 4
+.flags 2d
+.dest 2 d1 int16_t
+.param 2 p1
+
+copyw d1, p1
+
+
+.function orc_splat_s16_2d_8xn
+.n 8
+.flags 2d
+.dest 2 d1 int16_t
+.param 2 p1
+
+copyw d1, p1
+
+
+.function orc_splat_s16_2d
+.flags 2d
+.dest 2 d1 int16_t
+.param 2 p1
+
+copyw d1, p1
+
+
+.function orc_splat_u8_ns
+.dest 1 d1
+.param 1 p1
+
+copyb d1, p1
+
+
+.function orc_splat_u8_2d
+.flags 2d
+.dest 1 d1
+.param 1 p1
+
+copyb d1, p1
+
+
+.function orc_average_u8
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+
+avgub d1, s1, s2
+
+
+.function orc_rrshift6_add_s16_2d
+.flags 2d
+.dest 1 d1 uint8_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+
+addw t1, s2, 32
+shrsw t1, t1, 6
+addw t1, s1, t1
+convsuswb d1, t1
+
+
+.function orc_rrshift6_sub_s16_2d
+.flags 2d
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.temp 2 t1
+
+subw t1, d2, 8160
+shrsw t1, t1, 6
+copyw d2, t1
+subw d1, d1, t1
+
+
+.function orc_rrshift6_s16_ip_2d
+.flags 2d
+.dest 2 d1 int16_t
+.temp 2 t1
+
+subw t1, d1, 8160
+shrsw d1, t1, 6
+
+
+.function orc_rrshift6_s16_ip
+.dest 2 d1 int16_t
+.temp 2 t1
+
+subw t1, d1, 8160
+shrsw d1, t1, 6
+
+
+.function orc_unpack_yuyv_y
+.dest 1 d1
+.source 2 s1
+
+select0wb d1, s1
+
+
+.function orc_unpack_yuyv_u
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select0lw t1, s1
+select1wb d1, t1
+
+
+.function orc_unpack_yuyv_v
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select1lw t1, s1
+select1wb d1, t1
+
+
+.function orc_packyuyv
+.dest 4 d1
+.source 2 s1 uint8_t
+.source 1 s2
+.source 1 s3
+.temp 1 t1
+.temp 1 t2
+.temp 2 t3
+.temp 2 t4
+.temp 2 t5
+
+copyw t5, s1
+select0wb t1, t5
+select1wb t2, t5
+mergebw t3, t1, s2
+mergebw t4, t2, s3
+mergewl d1, t3, t4
+
+
+.function orc_unpack_uyvy_y
+.dest 1 d1
+.source 2 s1
+
+select1wb d1, s1
+
+
+.function orc_unpack_uyvy_u
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select0lw t1, s1
+select0wb d1, t1
+
+
+.function orc_unpack_uyvy_v
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select1lw t1, s1
+select0wb d1, t1
+
+
+.function orc_interleave2_s16
+.dest 4 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+
+mergewl d1, s1, s2
+
+
+.function orc_interleave2_rrshift1_s16
+.dest 4 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+
+addw t1, s1, 1
+shrsw t1, t1, 1
+addw t2, s2, 1
+shrsw t2, t2, 1
+mergewl d1, t1, t2
+
+
+.function orc_deinterleave2_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.source 4 s1 int16_t
+.temp 4 t1
+
+copyl t1, s1
+select0lw d1, t1
+select1lw d2, t1
+
+
+.function orc_deinterleave2_lshift1_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.source 4 s1 int16_t
+.temp 4 t1
+.temp 2 t2
+.temp 2 t3
+
+copyl t1, s1
+select0lw t2, t1
+shlw d1, t2, 1
+select1lw t3, t1
+shlw d2, t3, 1
+
+
+.function orc_haar_deint_lshift1_split_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.source 4 s1 int16_t
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+
+copyl t3, s1
+select0lw t1, t3
+select1lw t2, t3
+shlw t1, t1, 1
+shlw t2, t2, 1
+subw t2, t2, t1
+copyw d2, t2
+avgsw t2, t2, 0
+addw d1, t1, t2
+
+
+.function orc_haar_deint_split_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.source 4 s1 int16_t
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+
+copyl t3, s1
+select0lw t1, t3
+select1lw t2, t3
+subw t2, t2, t1
+copyw d2, t2
+avgsw t2, t2, 0
+addw d1, t1, t2
+
+
+.function orc_haar_split_s16_lo
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, s1
+subw t2, s2, t1
+avgsw t2, t2, 0
+addw d1, t1, t2
+
+
+.function orc_haar_split_s16_hi
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+
+subw d1, s2, s1
+
+
+.function orc_haar_split_s16_op
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, s1
+subw t2, s2, t1
+copyw d2, t2
+avgsw t2, t2, 0
+addw d1, t1, t2
+
+
+.function orc_haar_split_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, d1
+copyw t2, d2
+subw t2, t2, t1
+copyw d2, t2
+avgsw t2, t2, 0
+addw d1, t1, t2
+
+
+.function orc_haar_synth_s16_lo
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+
+avgsw t1, s2, 0
+subw d1, s1, t1
+
+
+.function orc_haar_synth_s16_hi
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+copyw t2, s2
+avgsw t3, t2, 0
+subw t1, s1, t3
+addw d1, t2, t1
+
+
+.function orc_haar_synth_s16_op
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+copyw t2, s2
+avgsw t3, t2, 0
+subw t1, s1, t3
+copyw d1, t1
+addw d2, t2, t1
+
+
+.function orc_haar_synth_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+copyw t1, d1
+copyw t2, d2
+avgsw t3, t2, 0
+subw t1, t1, t3
+copyw d1, t1
+addw d2, t2, t1
+
+
+.function orc_haar_synth_rrshift1_int_s16
+.dest 4 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+
+copyw t2, s2
+avgsw t1, t2, 0
+subw t1, s1, t1
+addw t2, t2, t1
+avgsw t1, t1, 0
+avgsw t2, t2, 0
+mergewl d1, t1, t2
+
+
+.function orc_haar_synth_int_s16
+.dest 4 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+
+copyw t2, s2
+avgsw t1, t2, 0
+subw t1, s1, t1
+addw t2, t2, t1
+mergewl d1, t1, t2
+
+
+.function orc_haar_sub_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+
+subw d1, d1, s1
+
+
+.function orc_haar_add_half_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.temp 2 t1
+
+avgsw t1, s1, 0
+addw d1, d1, t1
+
+
+.function orc_haar_add_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+
+addw d1, d1, s1
+
+
+.function orc_haar_sub_half_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.temp 2 t1
+
+avgsw t1, s1, 0
+subw d1, d1, t1
+
+
+.function orc_sum_u8
+.accumulator 4 a1 int32_t
+.source 1 s1
+.temp 2 t1
+.temp 4 t2
+
+convubw t1, s1
+convuwl t2, t1
+accl a1, t2
+
+
+.function orc_sum_s16
+.accumulator 4 a1 int32_t
+.source 2 s1 int16_t
+.temp 4 t1
+
+convswl t1, s1
+accl a1, t1
+
+
+.function orc_sum_square_diff_u8
+.accumulator 4 a1 int32_t
+.source 1 s1
+.source 1 s2
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+
+convubw t1, s1
+convubw t2, s2
+subw t1, t1, t2
+mullw t1, t1, t1
+convuwl t3, t1
+accl a1, t3
+
+
+.function orc_dequantise_s16_2d_4xn
+.n 4
+.flags 2d
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, s1
+signw t2, t1
+absw t1, t1
+mullw t1, t1, p1
+addw t1, t1, p2
+shrsw t1, t1, 2
+mullw d1, t1, t2
+
+
+.function orc_dequantise_s16_2d_8xn
+.n 8
+.flags 2d
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, s1
+signw t2, t1
+absw t1, t1
+mullw t1, t1, p1
+addw t1, t1, p2
+shrsw t1, t1, 2
+mullw d1, t1, t2
+
+
+.function orc_dequantise_s16_ip_2d_8xn
+.n 8
+.flags 2d
+.dest 2 d1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, d1
+signw t2, t1
+absw t1, t1
+mullw t1, t1, p1
+addw t1, t1, p2
+shrsw t1, t1, 2
+mullw d1, t1, t2
+
+
+.function orc_dequantise_s16_ip_2d
+.flags 2d
+.dest 2 d1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, d1
+signw t2, t1
+absw t1, t1
+mullw t1, t1, p1
+addw t1, t1, p2
+shrsw t1, t1, 2
+mullw d1, t1, t2
+
+
+.function orc_dequantise_s16_ip
+.dest 2 d1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, d1
+signw t2, t1
+absw t1, t1
+mullw t1, t1, p1
+addw t1, t1, p2
+shrsw t1, t1, 2
+mullw d1, t1, t2
+
+
+.function orc_dequantise_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, s1
+signw t2, t1
+absw t1, t1
+mullw t1, t1, p1
+addw t1, t1, p2
+shrsw t1, t1, 2
+mullw d1, t1, t2
+
+
+.function orc_dequantise_var_s16_ip
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.source 2 s2 int16_t
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, d1
+signw t2, t1
+absw t1, t1
+mullw t1, t1, s1
+addw t1, t1, s2
+shrsw t1, t1, 2
+mullw d1, t1, t2
+
+
+# only works for values between -16384 and 16384
+.function orc_quantise1_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, s1
+signw t2, t1
+absw t1, t1
+shlw t1, t1, 2
+subw t1, t1, p2
+mulhuw t1, t1, p1
+shruw t1, t1, p3
+mullw d1, t1, t2
+
+
+# only works for values between -16384 and 16384
+.function orc_quantise2_s16
+.dest 2 d1 int16_t
+.source 2 s1 int16_t
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, s1
+signw t2, t1
+absw t1, t1
+shlw t1, t1, 2
+subw t1, t1, p2
+shruw t1, t1, p1
+mullw d1, t1, t2
+
+
+# only works for values between -16384 and 16384
+.function orc_quantdequant1_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.param 2 p5
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, d2
+signw t2, t1
+absw t1, t1
+shlw t1, t1, 2
+subw t1, t1, p2
+mulhuw t1, t1, p1
+shruw t1, t1, p3
+mullw t2, t1, t2
+copyw d1, t2
+signw t2, t2
+mullw t1, t1, p4
+addw t1, t1, p5
+shrsw t1, t1, 2
+mullw d2, t1, t2
+
+
+# only works for values between -16384 and 16384
+.function orc_quantdequant3_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.param 2 p5
+.param 4 p6
+.temp 2 t1
+.temp 2 t2
+.temp 4 t3
+
+copyw t1, d2
+signw t2, t1
+absw t1, t1
+shlw t1, t1, 2
+subw t1, t1, p2
+muluwl t3, t1, p1
+addl t3, t3, p6
+shrul t3, t3, p3
+convlw t1, t3
+mullw t2, t1, t2
+copyw d1, t2
+signw t2, t2
+mullw t1, t1, p4
+addw t1, t1, p5
+shrsw t1, t1, 2
+mullw d2, t1, t2
+
+
+# only works for values between -16384 and 16384
+.function orc_quantdequant2_s16
+.dest 2 d1 int16_t
+.dest 2 d2 int16_t
+.param 2 p1
+.param 2 p2
+.param 2 p4
+.param 2 p5
+.temp 2 t1
+.temp 2 t2
+
+copyw t1, d2
+signw t2, t1
+absw t1, t1
+shlw t1, t1, 2
+subw t1, t1, p2
+shruw t1, t1, p1
+mullw t2, t1, t2
+copyw d1, t2
+signw t2, t2
+mullw t1, t1, p4
+addw t1, t1, p5
+shrsw t1, t1, 2
+mullw d2, t1, t2
+
+
+
+.function orc_downsample_vert_u8
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+.source 1 s3
+.source 1 s4
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+convubw t1, s1
+convubw t2, s4
+addw t1, t1, t2
+mullw t1, t1, 6
+convubw t2, s2
+convubw t3, s3
+addw t2, t2, t3
+mullw t2, t2, 26
+addw t2, t2, t1
+addw t2, t2, 32
+shruw t2, t2, 6
+convwb d1, t2
+
+
+.function orc_downsample_horiz_u8
+.dest 1 d1
+.source 2 s1 uint8_t
+.temp 2 a
+.temp 2 b
+.temp 2 t1
+.temp 1 t2
+.temp 1 t3
+.temp 2 c
+
+loadw t1, s1
+select1wb t2, t1
+convubw a, t2
+loadoffw t1, s1, 2
+select0wb t2, t1
+convubw b, t2
+addw c, a, b
+mullw c, c, 6
+
+loadoffw t1, s1, 1
+splitwb t2, t3, t1
+convubw a, t2
+convubw b, t3
+addw a, a, b
+mullw a, a, 26
+addw c, c, a
+addw c, c, 32
+shruw c, c, 6
+convwb d1, c
+
+
+.function orc_stats_moment_s16
+.source 2 s1 int16_t
+.accumulator 4 a1 int32_t
+.temp 2 t1
+.temp 4 t2
+
+absw t1, s1
+subw t1, t1, 2
+maxsw t1, t1, 0
+convuwl t2, t1
+accl a1, t2
+
+
+.function orc_stats_above_s16
+.source 2 s1 int16_t
+.accumulator 4 a1 int32_t
+.temp 2 t1
+.temp 4 t2
+
+absw t1, s1
+subw t1, t1, 1
+maxsw t1, t1, 0
+minsw t1, t1, 1
+convuwl t2, t1
+accl a1, t2
+
+
+.function orc_accw
+.accumulator 2 a1 int
+.source 2 s1 int16_t
+.temp 2 t1
+
+absw t1, s1
+accw a1, t1
+
+
+.function orc_avg2_8xn_u8
+.flags 2d
+.n 8
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+avgub d1, s1, s2
+
+
+.function orc_avg2_12xn_u8
+.flags 2d
+.n 12
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+avgub d1, s1, s2
+
+
+.function orc_avg2_16xn_u8
+.flags 2d
+.n 16
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+avgub d1, s1, s2
+
+
+.function orc_avg2_32xn_u8
+.flags 2d
+.n 32
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+avgub d1, s1, s2
+
+
+.function orc_avg2_nxm_u8
+.flags 2d
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+avgub d1, s1, s2
+
+
+.function orc_combine4_8xn_u8
+.flags 2d
+.n 8
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.source 1 s3 uint8_t
+.source 1 s4 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t2, t1, p1
+convubw t1, s2
+mullw t1, t1, p2
+addw t2, t2, t1
+convubw t1, s3
+mullw t1, t1, p3
+addw t2, t2, t1
+convubw t1, s4
+mullw t1, t1, p4
+addw t2, t2, t1
+addw t2, t2, 8
+convsuswb d1, t2
+
+
+.function orc_combine4_12xn_u8
+.flags 2d
+.n 12
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.source 1 s3 uint8_t
+.source 1 s4 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t2, t1, p1
+convubw t1, s2
+mullw t1, t1, p2
+addw t2, t2, t1
+convubw t1, s3
+mullw t1, t1, p3
+addw t2, t2, t1
+convubw t1, s4
+mullw t1, t1, p4
+addw t2, t2, t1
+addw t2, t2, 8
+convsuswb d1, t2
+
+
+.function orc_combine4_16xn_u8
+.flags 2d
+.n 16
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.source 1 s3 uint8_t
+.source 1 s4 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t2, t1, p1
+convubw t1, s2
+mullw t1, t1, p2
+addw t2, t2, t1
+convubw t1, s3
+mullw t1, t1, p3
+addw t2, t2, t1
+convubw t1, s4
+mullw t1, t1, p4
+addw t2, t2, t1
+addw t2, t2, 8
+convsuswb d1, t2
+
+
+.function orc_combine4_24xn_u8
+.flags 2d
+.n 24
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.source 1 s3 uint8_t
+.source 1 s4 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t2, t1, p1
+convubw t1, s2
+mullw t1, t1, p2
+addw t2, t2, t1
+convubw t1, s3
+mullw t1, t1, p3
+addw t2, t2, t1
+convubw t1, s4
+mullw t1, t1, p4
+addw t2, t2, t1
+addw t2, t2, 8
+convsuswb d1, t2
+
+
+.function orc_combine4_32xn_u8
+.flags 2d
+.n 32
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.source 1 s3 uint8_t
+.source 1 s4 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t2, t1, p1
+convubw t1, s2
+mullw t1, t1, p2
+addw t2, t2, t1
+convubw t1, s3
+mullw t1, t1, p3
+addw t2, t2, t1
+convubw t1, s4
+mullw t1, t1, p4
+addw t2, t2, t1
+addw t2, t2, 8
+convsuswb d1, t2
+
+
+.function orc_combine4_nxm_u8
+.flags 2d
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.source 1 s3 uint8_t
+.source 1 s4 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t2, t1, p1
+convubw t1, s2
+mullw t1, t1, p2
+addw t2, t2, t1
+convubw t1, s3
+mullw t1, t1, p3
+addw t2, t2, t1
+convubw t1, s4
+mullw t1, t1, p4
+addw t2, t2, t1
+addw t2, t2, 8
+shrsw t2, t2, 4
+convsuswb d1, t2
+
+
+.function orc_combine2_8xn_u8
+.flags 2d
+.n 8
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+convubw t2, s2
+mullw t1, t1, p1
+mullw t2, t2, p2
+addw t1, t1, t2
+addw t1, t1, p3
+shrsw t1, t1, p4
+convsuswb d1, t1
+
+
+
+.function orc_combine2_12xn_u8
+.flags 2d
+.n 12
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+convubw t2, s2
+mullw t1, t1, p1
+mullw t2, t2, p2
+addw t1, t1, t2
+addw t1, t1, p3
+shrsw t1, t1, p4
+convsuswb d1, t1
+
+
+
+.function orc_combine2_16xn_u8
+.flags 2d
+.n 16
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+convubw t2, s2
+mullw t1, t1, p1
+mullw t2, t2, p2
+addw t1, t1, t2
+addw t1, t1, p3
+shrsw t1, t1, p4
+convsuswb d1, t1
+
+
+
+.function orc_combine2_nxm_u8
+.flags 2d
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+convubw t2, s2
+mullw t1, t1, p1
+mullw t2, t2, p2
+addw t1, t1, t2
+addw t1, t1, p3
+shrsw t1, t1, p4
+convsuswb d1, t1
+
+
+
+.function orc_sad_nxm_u8
+.flags 2d
+.accumulator 4 a1 uint32_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+accsadubl a1, s1, s2
+
+
+.function orc_sad_8x8_u8
+.flags 2d
+.n 8
+.m 8
+.accumulator 4 a1 uint32_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+accsadubl a1, s1, s2
+
+
+
+.function orc_sad_12x12_u8
+.flags 2d
+.n 12
+.m 12
+.accumulator 4 a1 uint32_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+accsadubl a1, s1, s2
+
+
+
+.function orc_sad_16xn_u8
+.flags 2d
+.n 16
+.accumulator 4 a1 uint32_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+accsadubl a1, s1, s2
+
+
+
+.function orc_sad_32xn_u8
+.flags 2d
+.n 32
+.accumulator 4 a1 uint32_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+
+accsadubl a1, s1, s2
+
+
+
--- /dev/null
+
+#include <orc/orc.h>
+#include <orc-test/orctest.h>
+#include <orc/orcparse.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+
+static char * read_file (const char *filename);
+void output_code (OrcProgram *p, FILE *output);
+void output_code_header (OrcProgram *p, FILE *output);
+void output_code_test (OrcProgram *p, FILE *output);
+
+int error = FALSE;
+
+double weights_ginger[];
+//double weights_preston[];
+double weights_n900[];
+
+int
+main (int argc, char *argv[])
+{
+ char *code;
+ int n;
+ int i;
+ OrcProgram **programs;
+ const char *filename = NULL;
+ double sum;
+
+ orc_init ();
+ orc_test_init ();
+
+ filename = "bench10.orc";
+ code = read_file (filename);
+ if (!code) {
+ printf("benchmorc needs bench10.orc file in current directory\n");
+ exit(1);
+ }
+
+ n = orc_parse (code, &programs);
+
+#if 0
+ sum = 0;
+ for(i=0;i<n;i++){
+ double perf;
+ double weight;
+
+ perf = orc_test_performance_full (programs[i], 0, NULL);
+
+ if (perf == 0) {
+ weight = 0;
+ } else {
+ weight = 1.0/perf/241.0;
+ sum++;
+ }
+ printf(" %g, /* %s */\n", weight, programs[i]->name);
+ }
+ printf("sum = %g\n", sum);
+#else
+ sum = 0;
+ for(i=0;i<n;i++){
+ double perf;
+ double weight;
+
+ perf = orc_test_performance_full (programs[i], 0, NULL);
+ //weight = weights_ginger[i];
+ weight = weights_n900[i];
+
+ sum += weight * perf;
+ }
+ printf("score %g\n", 100.0/sum);
+#endif
+
+ if (error) return 1;
+ return 0;
+}
+
+
+static char *
+read_file (const char *filename)
+{
+ FILE *file = NULL;
+ char *contents = NULL;
+ long size;
+ int ret;
+
+ file = fopen (filename, "r");
+ if (file == NULL) return NULL;
+
+ ret = fseek (file, 0, SEEK_END);
+ if (ret < 0) goto bail;
+
+ size = ftell (file);
+ if (size < 0) goto bail;
+
+ ret = fseek (file, 0, SEEK_SET);
+ if (ret < 0) goto bail;
+
+ contents = malloc (size + 1);
+ if (contents == NULL) goto bail;
+
+ ret = fread (contents, size, 1, file);
+ if (ret < 0) goto bail;
+
+ contents[size] = 0;
+
+ return contents;
+bail:
+ /* something failed */
+ if (file) fclose (file);
+ if (contents) free (contents);
+
+ return NULL;
+}
+
+
+
+/* tables */
+
+/* ginger Intel(R) Core(TM)2 CPU T7600 @ 2.33GHz */
+
+double weights_ginger[] = {
+ 0.00539898, /* orc_scalarmultiply_f32_ns */
+ 0.00173034, /* orc_process_int16 */
+ 0.00229296, /* orc_process_int16_clamp */
+ 0.00238334, /* orc_process_int8 */
+ 0.00286, /* orc_process_int8_clamp */
+ 0.00224671, /* orc_audio_convert_unpack_u8 */
+ 0.00223485, /* orc_audio_convert_unpack_s8 */
+ 0.00261931, /* orc_audio_convert_unpack_u16 */
+ 0.0026756, /* orc_audio_convert_unpack_s16 */
+ 0.00187791, /* orc_audio_convert_unpack_u16_swap */
+ 0.00188964, /* orc_audio_convert_unpack_s16_swap */
+ 0.0018846, /* orc_audio_convert_unpack_u32 */
+ 0.00208672, /* orc_audio_convert_unpack_s32 */
+ 0.00158413, /* orc_audio_convert_unpack_u32_swap */
+ 0.0016592, /* orc_audio_convert_unpack_s32_swap */
+ 0.00113724, /* orc_audio_convert_unpack_float_s32 */
+ 0.000966394, /* orc_audio_convert_unpack_float_s32_swap */
+ 0.00163051, /* orc_audio_convert_unpack_float_double */
+ 0.00129049, /* orc_audio_convert_unpack_float_double_swap */
+ 0.00328124, /* orc_audio_convert_unpack_double_double */
+ 0.0019506, /* orc_audio_convert_unpack_double_double_swap */
+ 0.000854422, /* orc_audio_convert_unpack_u8_double */
+ 0.000841177, /* orc_audio_convert_unpack_s8_double */
+ 0.0013211, /* orc_audio_convert_unpack_u16_double */
+ 0.0012878, /* orc_audio_convert_unpack_s16_double */
+ 0.000888871, /* orc_audio_convert_unpack_u16_double_swap */
+ 0.00113332, /* orc_audio_convert_unpack_s16_double_swap */
+ 0.00125976, /* orc_audio_convert_unpack_u32_double */
+ 0.00204625, /* orc_audio_convert_unpack_s32_double */
+ 0.0010244, /* orc_audio_convert_unpack_u32_double_swap */
+ 0.00148207, /* orc_audio_convert_unpack_s32_double_swap */
+ 0.00135233, /* orc_audio_convert_pack_u8 */
+ 0.0013869, /* orc_audio_convert_pack_s8 */
+ 0.0021164, /* orc_audio_convert_pack_u16 */
+ 0.00211852, /* orc_audio_convert_pack_s16 */
+ 0.00183715, /* orc_audio_convert_pack_u16_swap */
+ 0.00200417, /* orc_audio_convert_pack_s16_swap */
+ 0.00193889, /* orc_audio_convert_pack_u32 */
+ 0.00208307, /* orc_audio_convert_pack_s32 */
+ 0.00159261, /* orc_audio_convert_pack_u32_swap */
+ 0.00167437, /* orc_audio_convert_pack_s32_swap */
+ 0.00143194, /* orc_audio_convert_pack_s32_float */
+ 0.00118178, /* orc_audio_convert_pack_s32_float_swap */
+ 0.00268428, /* orc_audio_convert_pack_double_float */
+ 0.0014616, /* orc_audio_convert_pack_double_float_swap */
+ 0.000483737, /* orc_audio_convert_pack_double_s8 */
+ 0.000686549, /* orc_audio_convert_pack_double_s16 */
+ 0.000577306, /* orc_audio_convert_pack_double_s16_swap */
+ 0.00100781, /* orc_audio_convert_pack_double_s32 */
+ 0.000784434, /* orc_audio_convert_pack_double_s32_swap */
+ 0.0172065, /* gst_orc_splat_u8 */
+ 0.0121632, /* gst_orc_splat_s16 */
+ 0.01221, /* gst_orc_splat_u16 */
+ 0.00740001, /* gst_orc_splat_u32 */
+ 0.00205476, /* orc_merge_linear_u8 */
+ 0.000841177, /* orc_merge_linear_u16 */
+ 0.0120937, /* orc_splat_u16 */
+ 0.00740001, /* orc_splat_u32 */
+ 0.00367431, /* orc_downsample_u8 */
+ 0.00148717, /* orc_downsample_u16 */
+ 0.00182448, /* gst_videoscale_orc_downsample_u32 */
+ 0.000360484, /* gst_videoscale_orc_downsample_yuyv */
+ 0, /* gst_videoscale_orc_resample_nearest_u8 */
+ 0, /* gst_videoscale_orc_resample_bilinear_u8 */
+ 0.0011403, /* gst_videoscale_orc_resample_nearest_u32 */
+ 0.000404898, /* gst_videoscale_orc_resample_bilinear_u32 */
+ 0.00027372, /* gst_videoscale_orc_resample_merge_bilinear_u32 */
+ 0.000888871, /* gst_videoscale_orc_merge_bicubic_u8 */
+ 0.00122619, /* add_int32 */
+ 0.0044839, /* add_int16 */
+ 0.00820311, /* add_int8 */
+ 0.00138106, /* add_uint32 */
+ 0.0044839, /* add_uint16 */
+ 0.00807787, /* add_uint8 */
+ 0.00177998, /* add_float32 */
+ 0.00737422, /* orc_splat_u32 */
+ 0.00221845, /* orc_memcpy_u32 */
+ 0.00294308, /* orc_blend_u8 */
+ 0.000216663, /* orc_blend_argb */
+ 0.000213342, /* orc_blend_bgra */
+ 0.00737422, /* orc_splat_u32 */
+ 0.000987128, /* deinterlace_line_vfir */
+ 0.00386205, /* deinterlace_line_linear */
+ 0.00178901, /* deinterlace_line_linear_blend */
+ 0, /* deinterlace_line_greedy */
+ 0.00824266, /* cogorc_memcpy_2d */
+ 0.00483197, /* cogorc_downsample_horiz_cosite_1tap */
+ 0.00139512, /* cogorc_downsample_horiz_cosite_3tap */
+ 0.00157236, /* cogorc_downsample_420_jpeg */
+ 0.00387619, /* cogorc_downsample_vert_halfsite_2tap */
+ 0.00175199, /* cogorc_downsample_vert_cosite_3tap */
+ 0.00131699, /* cogorc_downsample_vert_halfsite_4tap */
+ 0.00594495, /* cogorc_upsample_horiz_cosite_1tap */
+ 0.003575, /* cogorc_upsample_horiz_cosite */
+ 0.00371299, /* cogorc_upsample_vert_avgub */
+ 0.00495645, /* orc_unpack_yuyv_y */
+ 0.00172345, /* orc_unpack_yuyv_u */
+ 0.0019524, /* orc_unpack_yuyv_v */
+ 0.000834873, /* orc_pack_yuyv */
+ 0.0059283, /* orc_unpack_uyvy_y */
+ 0.0015303, /* orc_unpack_uyvy_u */
+ 0.00172345, /* orc_unpack_uyvy_v */
+ 0.000837184, /* orc_pack_uyvy */
+ 0.00245522, /* orc_matrix2_u8 */
+ 0, /* orc_matrix2_11_u8 */
+ 0, /* orc_matrix2_12_u8 */
+ 0.00145158, /* orc_matrix3_u8 */
+ 0, /* orc_matrix3_100_u8 */
+ 0.00138146, /* orc_matrix3_100_offset_u8 */
+ 0.00134289, /* orc_matrix3_000_u8 */
+ 0.00164957, /* orc_pack_123x */
+ 0.00160943, /* orc_pack_x123 */
+ 0.00254681, /* cogorc_combine2_u8 */
+ 0.000633112, /* cogorc_convert_I420_UYVY */
+ 0.000635284, /* cogorc_convert_I420_YUY2 */
+ 0.000541003, /* cogorc_convert_I420_AYUV */
+ 0.00055886, /* cogorc_convert_YUY2_I420 */
+ 0.00187425, /* cogorc_convert_UYVY_YUY2 */
+ 0.00277488, /* cogorc_planar_chroma_420_422 */
+ 0.00147587, /* cogorc_planar_chroma_420_444 */
+ 0.00624044, /* cogorc_planar_chroma_422_444 */
+ 0.0038428, /* cogorc_planar_chroma_444_422 */
+ 0.00203705, /* cogorc_planar_chroma_444_420 */
+ 0.00434352, /* cogorc_planar_chroma_422_420 */
+ 0.0005254, /* cogorc_convert_YUY2_AYUV */
+ 0.000538103, /* cogorc_convert_UYVY_AYUV */
+ 0.00100143, /* cogorc_convert_YUY2_Y42B */
+ 0.00103223, /* cogorc_convert_UYVY_Y42B */
+ 0.000770722, /* cogorc_convert_YUY2_Y444 */
+ 0.000778117, /* cogorc_convert_UYVY_Y444 */
+ 0.00055886, /* cogorc_convert_UYVY_I420 */
+ 0, /* cogorc_convert_AYUV_I420 */
+ 0.000341915, /* cogorc_convert_AYUV_YUY2 */
+ 0.000348717, /* cogorc_convert_AYUV_UYVY */
+ 0.000312522, /* cogorc_convert_AYUV_Y42B */
+ 0.000873979, /* cogorc_convert_AYUV_Y444 */
+ 0.00161992, /* cogorc_convert_Y42B_YUY2 */
+ 0.00165063, /* cogorc_convert_Y42B_UYVY */
+ 0.000588503, /* cogorc_convert_Y42B_AYUV */
+ 0.000805998, /* cogorc_convert_Y444_YUY2 */
+ 0.000815565, /* cogorc_convert_Y444_UYVY */
+ 0.0014937, /* cogorc_convert_Y444_AYUV */
+ 0, /* cogorc_convert_AYUV_ARGB */
+ 0, /* cogorc_convert_AYUV_BGRA */
+ 0, /* cogorc_convert_AYUV_ABGR */
+ 0, /* cogorc_convert_AYUV_RGBA */
+ 0, /* cogorc_convert_I420_BGRA */
+ 0, /* cogorc_convert_I420_BGRA_avg */
+ 0.000819403, /* cogorc_getline_I420 */
+ 0.00057497, /* cogorc_getline_YUY2 */
+ 0.000551146, /* cogorc_getline_UYVY */
+ 0.000562873, /* cogorc_getline_YVYU */
+ 0.000552297, /* cogorc_getline_Y42B */
+ 0.00156539, /* cogorc_getline_Y444 */
+ 0.00260962, /* cogorc_getline_Y800 */
+ 0.00185722, /* cogorc_getline_BGRA */
+ 0.000454554, /* cogorc_getline_ABGR */
+ 0.000452223, /* cogorc_getline_RGBA */
+ 0.000903587, /* cogorc_getline_NV12 */
+ 0.000800455, /* cogorc_getline_NV21 */
+ 0.000286517, /* cogorc_putline_I420 */
+ 0.000349472, /* cogorc_putline_YUY2 */
+ 0.000391564, /* cogorc_putline_YVYU */
+ 0.00035675, /* cogorc_putline_UYVY */
+ 0.000320473, /* cogorc_putline_Y42B */
+ 0.000845209, /* cogorc_putline_Y444 */
+ 0.00172426, /* cogorc_putline_Y800 */
+ 0.00185867, /* cogorc_putline_BGRA */
+ 0.000454814, /* cogorc_putline_ABGR */
+ 0.000451794, /* cogorc_putline_RGBA */
+ 0.000369853, /* cogorc_putline_NV12 */
+ 0.000414899, /* cogorc_putline_NV21 */
+ 0.00161681, /* orc_add2_rshift_add_s16_22_op */
+ 0.00209961, /* orc_add2_rshift_add_s16_22 */
+ 0.00161557, /* orc_add2_rshift_sub_s16_22_op */
+ 0.00209961, /* orc_add2_rshift_sub_s16_22 */
+ 0.00156654, /* orc_add2_rshift_add_s16_11_op */
+ 0.00201776, /* orc_add2_rshift_add_s16_11 */
+ 0.00156654, /* orc_add2_rshift_sub_s16_11_op */
+ 0.00205078, /* orc_add2_rshift_sub_s16_11 */
+ 0.00428421, /* orc_add_const_rshift_s16_11 */
+ 0.00863838, /* orc_add_const_rshift_s16 */
+ 0.00281436, /* orc_add_s16 */
+ 0.00372344, /* orc_add_s16_2d */
+ 0.0022491, /* orc_addc_rshift_s16 */
+ 0.00466168, /* orc_lshift1_s16 */
+ 0.00469616, /* orc_lshift2_s16 */
+ 0.010175, /* orc_lshift_s16_ip */
+ 0.000855112, /* orc_mas2_add_s16_op */
+ 0.000850644, /* orc_mas2_add_s16_ip */
+ 0.000853044, /* orc_mas2_sub_s16_op */
+ 0.000851328, /* orc_mas2_sub_s16_ip */
+ 0.000639672, /* orc_mas4_across_add_s16_1991_op */
+ 0.000673799, /* orc_mas4_across_add_s16_1991_ip */
+ 0.000640558, /* orc_mas4_across_sub_s16_1991_op */
+ 0.000673584, /* orc_mas4_across_sub_s16_1991_ip */
+ 0.00275573, /* orc_subtract_s16 */
+ 0.00267222, /* orc_add_s16_u8 */
+ 0.00544385, /* orc_add_s16_u8_2d */
+ 0.00653211, /* orc_convert_s16_u8 */
+ 0.00678334, /* orc_convert_u8_s16 */
+ 0.00689382, /* orc_offsetconvert_u8_s16 */
+ 0.00584641, /* orc_offsetconvert_s16_u8 */
+ 0.00273437, /* orc_subtract_s16_u8 */
+ 0.00238334, /* orc_multiply_and_add_s16_u8 */
+ 0.012025, /* orc_splat_s16_ns */
+ 0.000952381, /* orc_splat_s16_2d_4xn */
+ 0.00172953, /* orc_splat_s16_2d_8xn */
+ 0.0121284, /* orc_splat_s16_2d */
+ 0.0176367, /* orc_splat_u8_ns */
+ 0.0286862, /* orc_splat_u8_2d */
+ 0.00372606, /* orc_average_u8 */
+ 0.00242965, /* orc_rrshift6_add_s16_2d */
+ 0.00107453, /* orc_rrshift6_sub_s16_2d */
+ 0.0139063, /* orc_rrshift6_s16_ip_2d */
+ 0.011565, /* orc_rrshift6_s16_ip */
+ 0.00493334, /* orc_unpack_yuyv_y */
+ 0.00171647, /* orc_unpack_yuyv_u */
+ 0.00194363, /* orc_unpack_yuyv_v */
+ 0.000834543, /* orc_packyuyv */
+ 0.00591174, /* orc_unpack_uyvy_y */
+ 0.00153214, /* orc_unpack_uyvy_u */
+ 0.00172627, /* orc_unpack_uyvy_v */
+ 0.0029935, /* orc_interleave2_s16 */
+ 0.00201094, /* orc_interleave2_rrshift1_s16 */
+ 0.00140531, /* orc_deinterleave2_s16 */
+ 0.0011584, /* orc_deinterleave2_lshift1_s16 */
+ 0.00131151, /* orc_haar_deint_lshift1_split_s16 */
+ 0.00136132, /* orc_haar_deint_split_s16 */
+ 0.00202333, /* orc_haar_split_s16_lo */
+ 0.00278474, /* orc_haar_split_s16_hi */
+ 0.00127417, /* orc_haar_split_s16_op */
+ 0.00123128, /* orc_haar_split_s16 */
+ 0.0020668, /* orc_haar_synth_s16_lo */
+ 0.00175199, /* orc_haar_synth_s16_hi */
+ 0.00124788, /* orc_haar_synth_s16_op */
+ 0.0010582, /* orc_haar_synth_s16 */
+ 0.00130965, /* orc_haar_synth_rrshift1_int_s16 */
+ 0.00173192, /* orc_haar_synth_int_s16 */
+ 0.00444622, /* orc_haar_sub_s16 */
+ 0.00300625, /* orc_haar_add_half_s16 */
+ 0.00446498, /* orc_haar_add_s16 */
+ 0.00297619, /* orc_haar_sub_half_s16 */
+ 0.00296415, /* orc_sum_u8 */
+ 0.00353913, /* orc_sum_s16 */
+ 0.00195601, /* orc_sum_square_diff_u8 */
+ 0.000907029, /* orc_dequantise_s16_2d_4xn */
+ 0.0010582, /* orc_dequantise_s16_2d_8xn */
+ 0.0010582, /* orc_dequantise_s16_ip_2d_8xn */
+ 0.00607507, /* orc_dequantise_s16_ip_2d */
+ 0.00544062, /* orc_dequantise_s16_ip */
+ 0.002788, /* orc_dequantise_s16 */
+ 0.00177551, /* orc_dequantise_var_s16_ip */
+ 0.00250462, /* orc_quantise1_s16 */
+ 0.00286, /* orc_quantise2_s16 */
+ 0.000653211, /* orc_quantdequant1_s16 */
+ 0.00033615, /* orc_quantdequant3_s16 */
+ 0.000700332, /* orc_quantdequant2_s16 */
+ 0.00117709, /* orc_downsample_vert_u8 */
+ 0.000427844, /* orc_downsample_horiz_u8 */
+ 0.00341907, /* orc_stats_moment_s16 */
+ 0.00300625, /* orc_stats_above_s16 */
+ 0.012025, /* orc_accw */
+ 0.000979968, /* orc_avg2_8xn_u8 */
+ 0.000970018, /* orc_avg2_12xn_u8 */
+ 0.0012075, /* orc_avg2_16xn_u8 */
+ 0.00257509, /* orc_avg2_32xn_u8 */
+ 0.0042909, /* orc_avg2_nxm_u8 */
+ 0.00106312, /* orc_combine4_8xn_u8 */
+ 0.00116315, /* orc_combine4_12xn_u8 */
+ 0.00147629, /* orc_combine4_16xn_u8 */
+ 0.0013289, /* orc_combine4_24xn_u8 */
+ 0.00158025, /* orc_combine4_32xn_u8 */
+ 0.00119665, /* orc_combine4_nxm_u8 */
+ 0.00118816, /* orc_combine2_8xn_u8 */
+ 0.00167189, /* orc_combine2_12xn_u8 */
+ 0.00178851, /* orc_combine2_16xn_u8 */
+ 0.00229296, /* orc_combine2_nxm_u8 */
+ 0.00304518, /* orc_sad_nxm_u8 */
+ 0.0010836, /* orc_sad_8x8_u8 */
+ 0.00171215, /* orc_sad_12x12_u8 */
+ 0.00228437, /* orc_sad_16xn_u8 */
+ 0.00310406, /* orc_sad_32xn_u8 */
+};
+
+double weights_n900[] = {
+ 0.00189692, /* orc_scalarmultiply_f32_ns */
+ 0.000655569, /* orc_process_int16 */
+ 0.000607111, /* orc_process_int16_clamp */
+ 0.00130039, /* orc_process_int8 */
+ 0.00121357, /* orc_process_int8_clamp */
+ 0, /* orc_audio_convert_unpack_u8 */
+ 0.000886761, /* orc_audio_convert_unpack_s8 */
+ 0, /* orc_audio_convert_unpack_u16 */
+ 0.000885403, /* orc_audio_convert_unpack_s16 */
+ 0, /* orc_audio_convert_unpack_u16_swap */
+ 0.000884163, /* orc_audio_convert_unpack_s16_swap */
+ 0, /* orc_audio_convert_unpack_u32 */
+ 0.000886388, /* orc_audio_convert_unpack_s32 */
+ 0, /* orc_audio_convert_unpack_u32_swap */
+ 0.000885463, /* orc_audio_convert_unpack_s32_swap */
+ 0.000887546, /* orc_audio_convert_unpack_float_s32 */
+ 0.00088498, /* orc_audio_convert_unpack_float_s32_swap */
+ 0, /* orc_audio_convert_unpack_float_double */
+ 0, /* orc_audio_convert_unpack_float_double_swap */
+ 0.000453781, /* orc_audio_convert_unpack_double_double */
+ 0.000455403, /* orc_audio_convert_unpack_double_double_swap */
+ 0, /* orc_audio_convert_unpack_u8_double */
+ 0, /* orc_audio_convert_unpack_s8_double */
+ 0, /* orc_audio_convert_unpack_u16_double */
+ 0, /* orc_audio_convert_unpack_s16_double */
+ 0, /* orc_audio_convert_unpack_u16_double_swap */
+ 0, /* orc_audio_convert_unpack_s16_double_swap */
+ 0, /* orc_audio_convert_unpack_u32_double */
+ 0, /* orc_audio_convert_unpack_s32_double */
+ 0, /* orc_audio_convert_unpack_u32_double_swap */
+ 0, /* orc_audio_convert_unpack_s32_double_swap */
+ 0, /* orc_audio_convert_pack_u8 */
+ 0.000923392, /* orc_audio_convert_pack_s8 */
+ 0, /* orc_audio_convert_pack_u16 */
+ 0.0010451, /* orc_audio_convert_pack_s16 */
+ 0, /* orc_audio_convert_pack_u16_swap */
+ 0.000928539, /* orc_audio_convert_pack_s16_swap */
+ 0, /* orc_audio_convert_pack_u32 */
+ 0.000889152, /* orc_audio_convert_pack_s32 */
+ 0, /* orc_audio_convert_pack_u32_swap */
+ 0.000878315, /* orc_audio_convert_pack_s32_swap */
+ 0, /* orc_audio_convert_pack_s32_float */
+ 0, /* orc_audio_convert_pack_s32_float_swap */
+ 0, /* orc_audio_convert_pack_double_float */
+ 0, /* orc_audio_convert_pack_double_float_swap */
+ 0, /* orc_audio_convert_pack_double_s8 */
+ 0, /* orc_audio_convert_pack_double_s16 */
+ 0, /* orc_audio_convert_pack_double_s16_swap */
+ 0, /* orc_audio_convert_pack_double_s32 */
+ 0, /* orc_audio_convert_pack_double_s32_swap */
+ 0.00364483, /* gst_orc_splat_u8 */
+ 0.00176666, /* gst_orc_splat_s16 */
+ 0.00176698, /* gst_orc_splat_u16 */
+ 0.000887025, /* gst_orc_splat_u32 */
+ 0.00120903, /* orc_merge_linear_u8 */
+ 0.000840928, /* orc_merge_linear_u16 */
+ 0.00176687, /* orc_splat_u16 */
+ 0.000886406, /* orc_splat_u32 */
+ 0.00206071, /* orc_downsample_u8 */
+ 0.00111783, /* orc_downsample_u16 */
+ 0.00057776, /* gst_videoscale_orc_downsample_u32 */
+ 0.000282225, /* gst_videoscale_orc_downsample_yuyv */
+ 0, /* gst_videoscale_orc_resample_nearest_u8 */
+ 0, /* gst_videoscale_orc_resample_bilinear_u8 */
+ 0, /* gst_videoscale_orc_resample_nearest_u32 */
+ 0, /* gst_videoscale_orc_resample_bilinear_u32 */
+ 0, /* gst_videoscale_orc_resample_merge_bilinear_u32 */
+ 0.000810787, /* gst_videoscale_orc_merge_bicubic_u8 */
+ 0.00190264, /* add_int32 */
+ 0.00361624, /* add_int16 */
+ 0.00507259, /* add_int8 */
+ 0.00191598, /* add_uint32 */
+ 0.00361182, /* add_uint16 */
+ 0.00507259, /* add_uint8 */
+ 0.00153987, /* add_float32 */
+ 0.00124016, /* orc_splat_u32 */
+ 0.00113814, /* orc_memcpy_u32 */
+ 0.000908623, /* orc_blend_u8 */
+ 0.000133396, /* orc_blend_argb */
+ 0, /* orc_blend_bgra */
+ 0.000901581, /* orc_splat_u32 */
+ 0.000844679, /* deinterlace_line_vfir */
+ 0.00413285, /* deinterlace_line_linear */
+ 0.00123704, /* deinterlace_line_linear_blend */
+ 0.00109371, /* deinterlace_line_greedy */
+ 0.00381793, /* cogorc_memcpy_2d */
+ 0.00377314, /* cogorc_downsample_horiz_cosite_1tap */
+ 0.0010716, /* cogorc_downsample_horiz_cosite_3tap */
+ 0.00183775, /* cogorc_downsample_420_jpeg */
+ 0.00354723, /* cogorc_downsample_vert_halfsite_2tap */
+ 0.00110979, /* cogorc_downsample_vert_cosite_3tap */
+ 0.000901003, /* cogorc_downsample_vert_halfsite_4tap */
+ 0.00170136, /* cogorc_upsample_horiz_cosite_1tap */
+ 0.00165625, /* cogorc_upsample_horiz_cosite */
+ 0.00447476, /* cogorc_upsample_vert_avgub */
+ 0.00463618, /* orc_unpack_yuyv_y */
+ 0.00171047, /* orc_unpack_yuyv_u */
+ 0.00154902, /* orc_unpack_yuyv_v */
+ 0.000832261, /* orc_pack_yuyv */
+ 0.00380119, /* orc_unpack_uyvy_y */
+ 0.00190989, /* orc_unpack_uyvy_u */
+ 0.00170967, /* orc_unpack_uyvy_v */
+ 0.000832185, /* orc_pack_uyvy */
+ 0.00104047, /* orc_matrix2_u8 */
+ 0.000862238, /* orc_matrix2_11_u8 */
+ 0.00082022, /* orc_matrix2_12_u8 */
+ 0.00078049, /* orc_matrix3_u8 */
+ 0.000649602, /* orc_matrix3_100_u8 */
+ 0.000727013, /* orc_matrix3_100_offset_u8 */
+ 0.000760795, /* orc_matrix3_000_u8 */
+ 0.000790078, /* orc_pack_123x */
+ 0.000792433, /* orc_pack_x123 */
+ 0.00117391, /* cogorc_combine2_u8 */
+ 0.000196525, /* cogorc_convert_I420_UYVY */
+ 0.00019252, /* cogorc_convert_I420_YUY2 */
+ 0, /* cogorc_convert_I420_AYUV */
+ 0.000114783, /* cogorc_convert_YUY2_I420 */
+ 0.000977855, /* cogorc_convert_UYVY_YUY2 */
+ 0.000740782, /* cogorc_planar_chroma_420_422 */
+ 0.000600582, /* cogorc_planar_chroma_420_444 */
+ 0.0018566, /* cogorc_planar_chroma_422_444 */
+ 0.00219326, /* cogorc_planar_chroma_444_422 */
+ 0.00130591, /* cogorc_planar_chroma_444_420 */
+ 0.00211275, /* cogorc_planar_chroma_422_420 */
+ 0.000442738, /* cogorc_convert_YUY2_AYUV */
+ 0.000448218, /* cogorc_convert_UYVY_AYUV */
+ 0.000126284, /* cogorc_convert_YUY2_Y42B */
+ 0.00013561, /* cogorc_convert_UYVY_Y42B */
+ 0.000141577, /* cogorc_convert_YUY2_Y444 */
+ 0.000186874, /* cogorc_convert_UYVY_Y444 */
+ 9.1806e-05, /* cogorc_convert_UYVY_I420 */
+ 4.92179e-05, /* cogorc_convert_AYUV_I420 */
+ 0.000387252, /* cogorc_convert_AYUV_YUY2 */
+ 0.000386643, /* cogorc_convert_AYUV_UYVY */
+ 7.10203e-05, /* cogorc_convert_AYUV_Y42B */
+ 0.000121778, /* cogorc_convert_AYUV_Y444 */
+ 0.000693003, /* cogorc_convert_Y42B_YUY2 */
+ 0.00068451, /* cogorc_convert_Y42B_UYVY */
+ 0.000366249, /* cogorc_convert_Y42B_AYUV */
+ 0.000602631, /* cogorc_convert_Y444_YUY2 */
+ 0.00059332, /* cogorc_convert_Y444_UYVY */
+ 0.000742439, /* cogorc_convert_Y444_AYUV */
+ 0.00016071, /* cogorc_convert_AYUV_ARGB */
+ 0.00015913, /* cogorc_convert_AYUV_BGRA */
+ 0.000159045, /* cogorc_convert_AYUV_ABGR */
+ 0.000159024, /* cogorc_convert_AYUV_RGBA */
+ 0, /* cogorc_convert_I420_BGRA */
+ 0, /* cogorc_convert_I420_BGRA_avg */
+ 0, /* cogorc_getline_I420 */
+ 0.000447772, /* cogorc_getline_YUY2 */
+ 0.000449395, /* cogorc_getline_UYVY */
+ 0.000448533, /* cogorc_getline_YVYU */
+ 0.000385634, /* cogorc_getline_Y42B */
+ 0.000794574, /* cogorc_getline_Y444 */
+ 0, /* cogorc_getline_Y800 */
+ 0.000886484, /* cogorc_getline_BGRA */
+ 0.000554285, /* cogorc_getline_ABGR */
+ 0.000558078, /* cogorc_getline_RGBA */
+ 0.000405223, /* cogorc_getline_NV12 */
+ 0.000448563, /* cogorc_getline_NV21 */
+ 8.13877e-05, /* cogorc_putline_I420 */
+ 0.000387218, /* cogorc_putline_YUY2 */
+ 0.00035472, /* cogorc_putline_YVYU */
+ 0.000388082, /* cogorc_putline_UYVY */
+ 7.53593e-05, /* cogorc_putline_Y42B */
+ 0.000140356, /* cogorc_putline_Y444 */
+ 0.0017132, /* cogorc_putline_Y800 */
+ 0.000887432, /* cogorc_putline_BGRA */
+ 0.000554304, /* cogorc_putline_ABGR */
+ 0.000558475, /* cogorc_putline_RGBA */
+ 0.000325961, /* cogorc_putline_NV12 */
+ 0.000302852, /* cogorc_putline_NV21 */
+ 0.00165479, /* orc_add2_rshift_add_s16_22_op */
+ 0.00203387, /* orc_add2_rshift_add_s16_22 */
+ 0.00178764, /* orc_add2_rshift_sub_s16_22_op */
+ 0.00191923, /* orc_add2_rshift_sub_s16_22 */
+ 0.00240863, /* orc_add2_rshift_add_s16_11_op */
+ 0.00236533, /* orc_add2_rshift_add_s16_11 */
+ 0.00238607, /* orc_add2_rshift_sub_s16_11_op */
+ 0.00215832, /* orc_add2_rshift_sub_s16_11 */
+ 0.00181178, /* orc_add_const_rshift_s16_11 */
+ 0.00276915, /* orc_add_const_rshift_s16 */
+ 0.00281042, /* orc_add_s16 */
+ 0.00166994, /* orc_add_s16_2d */
+ 0.00213569, /* orc_addc_rshift_s16 */
+ 0.00285013, /* orc_lshift1_s16 */
+ 0.00245598, /* orc_lshift2_s16 */
+ 0.00370756, /* orc_lshift_s16_ip */
+ 0.000576921, /* orc_mas2_add_s16_op */
+ 0.000574115, /* orc_mas2_add_s16_ip */
+ 0.000577379, /* orc_mas2_sub_s16_op */
+ 0.000572396, /* orc_mas2_sub_s16_ip */
+ 0.000463221, /* orc_mas4_across_add_s16_1991_op */
+ 0.000454726, /* orc_mas4_across_add_s16_1991_ip */
+ 0.000459198, /* orc_mas4_across_sub_s16_1991_op */
+ 0.000454726, /* orc_mas4_across_sub_s16_1991_ip */
+ 0.00228059, /* orc_subtract_s16 */
+ 0.00221536, /* orc_add_s16_u8 */
+ 0.00207322, /* orc_add_s16_u8_2d */
+ 0.00292092, /* orc_convert_s16_u8 */
+ 0.0036697, /* orc_convert_u8_s16 */
+ 0.00262931, /* orc_offsetconvert_u8_s16 */
+ 0.00169662, /* orc_offsetconvert_s16_u8 */
+ 0.00238882, /* orc_subtract_s16_u8 */
+ 0.00214581, /* orc_multiply_and_add_s16_u8 */
+ 0.00171415, /* orc_splat_s16_ns */
+ 0.000393839, /* orc_splat_s16_2d_4xn */
+ 0.00108717, /* orc_splat_s16_2d_8xn */
+ 0.00238507, /* orc_splat_s16_2d */
+ 0.00710336, /* orc_splat_u8_ns */
+ 0.00403526, /* orc_splat_u8_2d */
+ 0.0050664, /* orc_average_u8 */
+ 0.00134097, /* orc_rrshift6_add_s16_2d */
+ 0.000807761, /* orc_rrshift6_sub_s16_2d */
+ 0.00307394, /* orc_rrshift6_s16_ip_2d */
+ 0.00316505, /* orc_rrshift6_s16_ip */
+ 0.00414213, /* orc_unpack_yuyv_y */
+ 0.00171048, /* orc_unpack_yuyv_u */
+ 0.00152237, /* orc_unpack_yuyv_v */
+ 0.000832094, /* orc_packyuyv */
+ 0.00399968, /* orc_unpack_uyvy_y */
+ 0.00191216, /* orc_unpack_uyvy_u */
+ 0.00171008, /* orc_unpack_uyvy_v */
+ 0.000897314, /* orc_interleave2_s16 */
+ 0.000901892, /* orc_interleave2_rrshift1_s16 */
+ 0.00076676, /* orc_deinterleave2_s16 */
+ 0.000698548, /* orc_deinterleave2_lshift1_s16 */
+ 0.00052955, /* orc_haar_deint_lshift1_split_s16 */
+ 0.000541159, /* orc_haar_deint_split_s16 */
+ 0.00193144, /* orc_haar_split_s16_lo */
+ 0.00199633, /* orc_haar_split_s16_hi */
+ 0.00100157, /* orc_haar_split_s16_op */
+ 0.00110899, /* orc_haar_split_s16 */
+ 0.00250888, /* orc_haar_synth_s16_lo */
+ 0.00237495, /* orc_haar_synth_s16_hi */
+ 0.00102733, /* orc_haar_synth_s16_op */
+ 0.000903217, /* orc_haar_synth_s16 */
+ 0.000876559, /* orc_haar_synth_rrshift1_int_s16 */
+ 0.000887662, /* orc_haar_synth_int_s16 */
+ 0.00410694, /* orc_haar_sub_s16 */
+ 0.0032058, /* orc_haar_add_half_s16 */
+ 0.00388831, /* orc_haar_add_s16 */
+ 0.00291108, /* orc_haar_sub_half_s16 */
+ 0.000801367, /* orc_sum_u8 */
+ 0.00114909, /* orc_sum_s16 */
+ 0.000370646, /* orc_sum_square_diff_u8 */
+ 0.00013075, /* orc_dequantise_s16_2d_4xn */
+ 0.000133912, /* orc_dequantise_s16_2d_8xn */
+ 9.03212e-05, /* orc_dequantise_s16_ip_2d_8xn */
+ 0.00113719, /* orc_dequantise_s16_ip_2d */
+ 0.00110444, /* orc_dequantise_s16_ip */
+ 0.0010195, /* orc_dequantise_s16 */
+ 0.00102223, /* orc_dequantise_var_s16_ip */
+ 0.000730524, /* orc_quantise1_s16 */
+ 0.00110444, /* orc_quantise2_s16 */
+ 0.000335265, /* orc_quantdequant1_s16 */
+ 0.000192182, /* orc_quantdequant3_s16 */
+ 0.000417592, /* orc_quantdequant2_s16 */
+ 0.000898304, /* orc_downsample_vert_u8 */
+ 0.000362124, /* orc_downsample_horiz_u8 */
+ 0.000544868, /* orc_stats_moment_s16 */
+ 0.000453732, /* orc_stats_above_s16 */
+ 0.00292441, /* orc_accw */
+ 0.000359527, /* orc_avg2_8xn_u8 */
+ 0.000354736, /* orc_avg2_12xn_u8 */
+ 0.00149989, /* orc_avg2_16xn_u8 */
+ 0.00233937, /* orc_avg2_32xn_u8 */
+ 0.00314972, /* orc_avg2_nxm_u8 */
+ 0.00035942, /* orc_combine4_8xn_u8 */
+ 0.000194902, /* orc_combine4_12xn_u8 */
+ 0.000479254, /* orc_combine4_16xn_u8 */
+ 0.000570636, /* orc_combine4_24xn_u8 */
+ 0.00024299, /* orc_combine4_32xn_u8 */
+ 0.00065699, /* orc_combine4_nxm_u8 */
+ 0.000471688, /* orc_combine2_8xn_u8 */
+ 0.000324071, /* orc_combine2_12xn_u8 */
+ 0.000253177, /* orc_combine2_16xn_u8 */
+ 0.000954874, /* orc_combine2_nxm_u8 */
+ 0.000585976, /* orc_sad_nxm_u8 */
+ 0.000462648, /* orc_sad_8x8_u8 */
+ 0.000516877, /* orc_sad_12x12_u8 */
+ 0.000539588, /* orc_sad_16xn_u8 */
+ 0.000534326, /* orc_sad_32xn_u8 */
+};
+
projects / platform / upstream / orc.git / commitdiff