[NEON] Add vpx_highbd_subtract_block function

    Total gain for 12-bit encoding:
    * ~1% for best and rt profile

Change-Id: I4039120dc570baab1ae519a5e38b1acff38d81f0

diff --git a/vpx_dsp/arm/subtract_neon.c b/vpx_dsp/arm/subtract_neon.c
index 612897e247c27bc77fb783d3fe6518fdd9b6c2b6..2c008e48ab29d6c120860c85e224872aa42eafbd 100644 (file)
--- a/vpx_dsp/arm/subtract_neon.c
+++ b/vpx_dsp/arm/subtract_neon.c
@@ -79,3 +79,59 @@ void vpx_subtract_block_neon(int rows, int cols, int16_t *diff,
     } while (r);
   }
 }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_subtract_block_neon(int rows, int cols, int16_t *diff_ptr,
+                                    ptrdiff_t diff_stride,
+                                    const uint8_t *src8_ptr,
+                                    ptrdiff_t src_stride,
+                                    const uint8_t *pred8_ptr,
+                                    ptrdiff_t pred_stride, int bd) {
+  int r = rows, c;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8_ptr);
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8_ptr);
+  (void)bd;
+
+  if (cols >= 16) {
+    do {
+      for (c = 0; c < cols; c += 16) {
+        const uint16x8_t s0 = vld1q_u16(&src[c + 0]);
+        const uint16x8_t s1 = vld1q_u16(&src[c + 8]);
+        const uint16x8_t p0 = vld1q_u16(&pred[c + 0]);
+        const uint16x8_t p1 = vld1q_u16(&pred[c + 8]);
+        const uint16x8_t d0 = vsubq_u16(s0, p0);
+        const uint16x8_t d1 = vsubq_u16(s1, p1);
+        vst1q_s16(&diff_ptr[c + 0], vreinterpretq_s16_u16(d0));
+        vst1q_s16(&diff_ptr[c + 8], vreinterpretq_s16_u16(d1));
+      }
+      diff_ptr += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    } while (--r);
+  } else if (cols >= 8) {
+    do {
+      for (c = 0; c < cols; c += 8) {
+        const uint16x8_t s = vld1q_u16(&src[c]);
+        const uint16x8_t p = vld1q_u16(&pred[c]);
+        const uint16x8_t d0 = vsubq_u16(s, p);
+        vst1q_s16(&diff_ptr[c], vreinterpretq_s16_u16(d0));
+      }
+      diff_ptr += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    } while (--r);
+  } else if (cols >= 4) {
+    do {
+      for (c = 0; c < cols; c += 4) {
+        const uint16x4_t s = vld1_u16(&src[c]);
+        const uint16x4_t p = vld1_u16(&pred[c]);
+        const uint16x4_t v_diff = vsub_u16(s, p);
+        vst1_s16(&diff_ptr[c], vreinterpret_s16_u16(v_diff));
+      }
+      diff_ptr += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    } while (--r);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 72442ed41fb20f5b73b1c86d1c197a1ea86d5c46..6cd46129f026f8c5ebfb16e5d599908db2479f3c 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -935,6 +935,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   # Block subtraction
   #
   add_proto qw/void vpx_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src8_ptr, ptrdiff_t src_stride, const uint8_t *pred8_ptr, ptrdiff_t pred_stride, int bd";
+  specialize qw/vpx_highbd_subtract_block neon/;
 
   #
   # Single block SAD