video-scaler: add ORC optimized versions

Add ORC optimized versions of 2 and 4tap vertical scaling. Provide
a high quality 12 bits and a low quality 6 bits version.

gst-libs/gst/video/video-orc-dist.h

@@ -149,6 +149,10 @@ void video_orc_convert_AYUV_ABGR (guint8 * ORC_RESTRICT d1, int d1_stride, const
 void video_orc_convert_AYUV_RGBA (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int p1, int p2, int p3, int p4, int p5, int n, int m);
 void video_orc_convert_I420_BGRA (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1, int p2, int p3, int p4, int p5, int n);
 void video_orc_matrix8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1, orc_int64 p1, orc_int64 p2, orc_int64 p3, int n);
+void video_orc_resample_v_2tap_8_lq (guint32 * ORC_RESTRICT d1, const guint32 * ORC_RESTRICT s1, const guint32 * ORC_RESTRICT s2, int p1, int n);
+void video_orc_resample_v_2tap_8 (guint32 * ORC_RESTRICT d1, const guint32 * ORC_RESTRICT s1, const guint32 * ORC_RESTRICT s2, int p1, int n);
+void video_orc_resample_v_4tap_8_lq (guint32 * ORC_RESTRICT d1, const guint32 * ORC_RESTRICT s1, const guint32 * ORC_RESTRICT s2, const guint32 * ORC_RESTRICT s3, const guint32 * ORC_RESTRICT s4, int p1, int p2, int p3, int p4, int n);
+void video_orc_resample_v_4tap_8 (guint32 * ORC_RESTRICT d1, const guint32 * ORC_RESTRICT s1, const guint32 * ORC_RESTRICT s2, const guint32 * ORC_RESTRICT s3, const guint32 * ORC_RESTRICT s4, int p1, int p2, int p3, int p4, int n);
 
 #ifdef __cplusplus
 }

gst-libs/gst/video/video-orc.orc

@@ -1284,3 +1284,156 @@ x4 addssw aq, aq, q1
 
 x4 convssswb ayuv2, aq
 x4 addb ayuv, ayuv2, c128
+
+#.function video_orc_resample_h_near_8888
+#.source 4 src guint32
+#.source 4 idx
+#.dest 4 dest guint32
+#.temp 4 t
+#
+#loadidxl t, src, idx
+#storel dest, t
+
+#.function video_orc_resample_h_2tap_8888_16
+#.source 4 src1 guint32
+#.source 4 src2 guint32
+#.source 8 coef1 guint64
+#.source 8 coef2 guint64
+#.source 4 idx
+#.dest 4 dest guint32
+#.temp 4 t1
+#.temp 4 t2
+#.temp 8 q1
+#.temp 8 q2
+#
+#loadidxl t1, src1, idx
+#x4 convubw q1, t1
+#x4 mulhuw q1, q1, coef1
+#
+#loadidxl t2, src2, idx
+#x4 convubw q2, t2
+#x4 mulhuw q2, q2, coef2
+#
+#x4 addw q2, q2, q1
+#x4 convuuswb dest, q2
+#
+#.function video_orc_resample_h_2tap_8888_lq
+#.source 4 src1 guint32
+#.source 4 src2 guint32
+#.source 8 coef1 guint64
+#.source 4 idx
+#.dest 4 dest guint32
+#.temp 4 t1
+#.temp 4 t2
+#.temp 8 q1
+#.temp 8 q2
+#
+#loadidxl t1, src1, idx
+#x4 convubw q1, t1
+#loadidxl t2, src2, idx
+#x4 convubw q2, t2
+#x4 subw q2, q2, q1
+#
+#x4 mullw q2, q2, coef1
+#x4 addw q2, q2, 128
+#x4 convhwb t2, q2
+#x4 addb dest, t2, t1
+
+.function video_orc_resample_v_2tap_8_lq
+.source 1 src1 guint32
+.source 1 src2 guint32
+.dest 1 dest guint32
+.param 2 p1
+.temp 1 t
+.temp 2 w1
+.temp 2 w2
+
+convubw w1, src1
+convubw w2, src2
+subw w2, w2, w1
+mullw w2, w2, p1
+addw w2, w2, 128
+convhwb t, w2
+addb dest, t, src1
+
+.function video_orc_resample_v_2tap_8
+.source 1 s1 guint32
+.source 1 s2 guint32
+.dest 1 d1 guint32
+.param 2 p1
+.temp 1 t
+.temp 2 w1
+.temp 2 w2
+.temp 4 t1
+.temp 4 t2
+
+convubw w1, s1
+convubw w2, s2
+subw w2, w2, w1
+mulswl t2, w2, p1
+addl t2, t2, 4095
+shrsl t2, t2, 12
+convlw w2, t2
+addw w2, w2, w1
+convsuswb d1, w2
+
+.function video_orc_resample_v_4tap_8_lq
+.source 1 s1 guint32
+.source 1 s2 guint32
+.source 1 s3 guint32
+.source 1 s4 guint32
+.dest 1 d1 guint32
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 w1
+.temp 2 w2
+
+convubw w1, s1
+mullw w1, w1, p1
+convubw w2, s2
+mullw w2, w2, p2
+addw w1, w1, w2
+convubw w2, s3
+mullw w2, w2, p3
+addw w1, w1, w2
+convubw w2, s4
+mullw w2, w2, p4
+addw w1, w1, w2
+addw w1, w1, 32
+shrsw w1, w1, 6
+convsuswb d1, w1
+
+.function video_orc_resample_v_4tap_8
+.source 1 s1 guint32
+.source 1 s2 guint32
+.source 1 s3 guint32
+.source 1 s4 guint32
+.dest 1 d1 guint32
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 w1
+.temp 2 w2
+.temp 4 t1
+.temp 4 t2
+
+convubw w1, s1
+mulswl t1, w1, p1
+convubw w2, s2
+mulswl t2, w2, p2
+addl t1, t1, t2
+convubw w2, s3
+mulswl t2, w2, p3
+addl t1, t1, t2
+convubw w2, s4
+mulswl t2, w2, p4
+addl t1, t1, t2
+addl t1, t1, 4095
+shrsl t1, t1, 12
+convlw w1, t1
+convsuswb d1, w1
+
+

gst-libs/gst/video/video-scaler.c

@@ -26,11 +26,15 @@
 #include <math.h>
 
 #include "resampler.h"
+#include <orc/orcfunctions.h>
+#include "video-orc.h"
 #include "video-scaler.h"
 
 #define S16_SCALE       12
 #define S16_SCALE_ROUND (1 << (S16_SCALE -1))
 
+#define LQ
+
 typedef void (*GstVideoScalerHFunc) (GstVideoScaler * scale,
     gpointer src, gpointer dest, guint dest_offset, guint width);
 typedef void (*GstVideoScalerVFunc) (GstVideoScaler * scale,
@@ -282,11 +286,42 @@ video_scale_h_near_8888 (GstVideoScaler * scale,
     d[i] = s[offset[i]];
 }
 
+#define BLEND_2TAP(a,b,p) (((((b)-(guint16)(a)) * p + S16_SCALE_ROUND) >> S16_SCALE) + (a))
+
 static void
-video_scale_v_near_8888 (GstVideoScaler * scale,
+video_scale_h_2tap_8888 (GstVideoScaler * scale,
-    gpointer srcs[], gpointer dest, guint dest_offset, guint width)
+    gpointer src, gpointer dest, guint dest_offset, guint width)
 {
-  memcpy (dest, srcs[0], 4 * width);
+  gint i, max_taps, sum0, sum1, sum2, sum3;
+  guint8 *s1, *s2, *d;
+  guint32 *offset, *phase;
+  gint16 *taps, *t;
+
+  if (scale->taps_s16 == NULL)
+    make_s16_taps (scale, S16_SCALE);
+
+  max_taps = scale->resampler.max_taps;
+  offset = scale->resampler.offset + dest_offset;
+  phase = scale->resampler.phase + dest_offset;
+  taps = scale->taps_s16;
+
+  d = (guint8 *) dest + 4 * dest_offset;
+
+  for (i = 0; i < width; i++) {
+    s1 = (guint8 *) src + 4 * offset[i];
+    s2 = s1 + 4;
+    t = taps + (phase[i] * max_taps);
+
+    sum0 = BLEND_2TAP (s1[0], s2[0], t[1]);
+    sum1 = BLEND_2TAP (s1[1], s2[1], t[1]);
+    sum2 = BLEND_2TAP (s1[2], s2[2], t[1]);
+    sum3 = BLEND_2TAP (s1[3], s2[3], t[1]);
+
+    d[i * 4 + 0] = CLAMP (sum0, 0, 255);
+    d[i * 4 + 1] = CLAMP (sum1, 0, 255);
+    d[i * 4 + 2] = CLAMP (sum2, 0, 255);
+    d[i * 4 + 3] = CLAMP (sum3, 0, 255);
+  }
 }
 
 static void
@@ -330,6 +365,79 @@ video_scale_h_ntap_8888 (GstVideoScaler * scale,
   }
 }
 
+
+static void
+video_scale_v_near_8888 (GstVideoScaler * scale,
+    gpointer srcs[], gpointer dest, guint dest_offset, guint width)
+{
+  orc_memcpy (dest, srcs[0], 4 * width);
+}
+
+static void
+video_scale_v_2tap_8888 (GstVideoScaler * scale,
+    gpointer srcs[], gpointer dest, guint dest_offset, guint width)
+{
+  gint max_taps;
+  guint32 *s1, *s2, *d;
+  guint64 p1;
+
+  if (scale->taps_s16 == NULL)
+#ifdef LQ
+    make_s16_taps (scale, 8);
+#else
+    make_s16_taps (scale, S16_SCALE);
+#endif
+
+  max_taps = scale->resampler.max_taps;
+
+  d = (guint32 *) dest;
+  s1 = (guint32 *) srcs[0];
+  s2 = (guint32 *) srcs[1];
+  p1 = scale->taps_s16[dest_offset * max_taps + 1];
+
+#ifdef LQ
+  video_orc_resample_v_2tap_8_lq (d, s1, s2, p1, width * 4);
+#else
+  video_orc_resample_v_2tap_8 (d, s1, s2, p1, width * 4);
+#endif
+}
+
+static void
+video_scale_v_4tap_8888 (GstVideoScaler * scale,
+    gpointer srcs[], gpointer dest, guint dest_offset, guint width)
+{
+  gint max_taps;
+  guint32 *s1, *s2, *s3, *s4, *d;
+  gint p1, p2, p3, p4;
+  gint16 *taps;
+
+  if (scale->taps_s16 == NULL)
+#ifdef LQ
+    make_s16_taps (scale, 6);
+#else
+    make_s16_taps (scale, S16_SCALE);
+#endif
+
+  max_taps = scale->resampler.max_taps;
+  taps = scale->taps_s16 + dest_offset * max_taps;
+
+  d = (guint32 *) dest;
+  s1 = (guint32 *) srcs[0];
+  s2 = (guint32 *) srcs[1];
+  s3 = (guint32 *) srcs[2];
+  s4 = (guint32 *) srcs[3];
+  p1 = taps[0];
+  p2 = taps[1];
+  p3 = taps[2];
+  p4 = taps[3];
+
+#ifdef LQ
+  video_orc_resample_v_4tap_8_lq (d, s1, s2, s3, s4, p1, p2, p3, p4, width * 4);
+#else
+  video_orc_resample_v_4tap_8 (d, s1, s2, s3, s4, p1, p2, p3, p4, width * 4);
+#endif
+}
+
 static void
 video_scale_v_ntap_8888 (GstVideoScaler * scale,
     gpointer srcs[], gpointer dest, guint dest_offset, guint width)
@@ -400,6 +508,9 @@ gst_video_scaler_horizontal (GstVideoScaler * scale, GstVideoFormat format,
     case 1:
       func = video_scale_h_near_8888;
       break;
+    case 2:
+      func = video_scale_h_2tap_8888;
+      break;
     default:
       func = video_scale_h_ntap_8888;
       break;
@@ -437,6 +548,12 @@ gst_video_scaler_vertical (GstVideoScaler * scale, GstVideoFormat format,
     case 1:
       func = video_scale_v_near_8888;
       break;
+    case 2:
+      func = video_scale_v_2tap_8888;
+      break;
+    case 4:
+      func = video_scale_v_4tap_8888;
+      break;
     default:
       func = video_scale_v_ntap_8888;
       break;