summaryrefslogtreecommitdiffstats
path: root/third_party/aom/aom_dsp/arm/intrapred_neon.c
diff options
context:
space:
mode:
authortrav90 <travawine@palemoon.org>2018-10-19 23:00:02 -0500
committertrav90 <travawine@palemoon.org>2018-10-19 23:00:02 -0500
commitb8df135c97a854c2ff9b4394b016649c601177fa (patch)
tree802b7de5ad245f1a12adbcef835ab0d0687c1bf8 /third_party/aom/aom_dsp/arm/intrapred_neon.c
parenta4d3c59dcac642f6b9557dc09b60eda40b517630 (diff)
downloadUXP-b8df135c97a854c2ff9b4394b016649c601177fa.tar
UXP-b8df135c97a854c2ff9b4394b016649c601177fa.tar.gz
UXP-b8df135c97a854c2ff9b4394b016649c601177fa.tar.lz
UXP-b8df135c97a854c2ff9b4394b016649c601177fa.tar.xz
UXP-b8df135c97a854c2ff9b4394b016649c601177fa.zip
Update libaom to rev b25610052a1398032320008d69b51d2da94f5928
Diffstat (limited to 'third_party/aom/aom_dsp/arm/intrapred_neon.c')
-rw-r--r--third_party/aom/aom_dsp/arm/intrapred_neon.c60
1 files changed, 60 insertions, 0 deletions
diff --git a/third_party/aom/aom_dsp/arm/intrapred_neon.c b/third_party/aom/aom_dsp/arm/intrapred_neon.c
index 69470eeb0..c85b1e910 100644
--- a/third_party/aom/aom_dsp/arm/intrapred_neon.c
+++ b/third_party/aom/aom_dsp/arm/intrapred_neon.c
@@ -528,3 +528,63 @@ void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
}
}
}
+
+static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride, int bw,
+ const uint16_t *above,
+ const uint16_t *left) {
+ assert(bw >= 4);
+ assert(IS_POWER_OF_TWO(bw));
+ int expected_dc, sum = 0;
+ const int count = bw * 2;
+ uint32x4_t sum_q = vdupq_n_u32(0);
+ uint32x2_t sum_d;
+ uint16_t *dst_1;
+ if (bw >= 8) {
+ for (int i = 0; i < bw; i += 8) {
+ sum_q = vpadalq_u16(sum_q, vld1q_u16(above));
+ sum_q = vpadalq_u16(sum_q, vld1q_u16(left));
+ above += 8;
+ left += 8;
+ }
+ sum_d = vadd_u32(vget_low_u32(sum_q), vget_high_u32(sum_q));
+ sum = vget_lane_s32(vreinterpret_s32_u64(vpaddl_u32(sum_d)), 0);
+ expected_dc = (sum + (count >> 1)) / count;
+ const uint16x8_t dc = vdupq_n_u16((uint16_t)expected_dc);
+ for (int r = 0; r < bw; r++) {
+ dst_1 = dst;
+ for (int i = 0; i < bw; i += 8) {
+ vst1q_u16(dst_1, dc);
+ dst_1 += 8;
+ }
+ dst += stride;
+ }
+ } else { // 4x4
+ sum_q = vaddl_u16(vld1_u16(above), vld1_u16(left));
+ sum_d = vadd_u32(vget_low_u32(sum_q), vget_high_u32(sum_q));
+ sum = vget_lane_s32(vreinterpret_s32_u64(vpaddl_u32(sum_d)), 0);
+ expected_dc = (sum + (count >> 1)) / count;
+ const uint16x4_t dc = vdup_n_u16((uint16_t)expected_dc);
+ for (int r = 0; r < bw; r++) {
+ vst1_u16(dst, dc);
+ dst += stride;
+ }
+ }
+}
+
+#define intra_pred_highbd_sized_neon(type, width) \
+ void aom_highbd_##type##_predictor_##width##x##width##_neon( \
+ uint16_t *dst, ptrdiff_t stride, const uint16_t *above, \
+ const uint16_t *left, int bd) { \
+ (void)bd; \
+ highbd_##type##_predictor(dst, stride, width, above, left); \
+ }
+
+#define intra_pred_square(type) \
+ intra_pred_highbd_sized_neon(type, 4); \
+ intra_pred_highbd_sized_neon(type, 8); \
+ intra_pred_highbd_sized_neon(type, 16); \
+ intra_pred_highbd_sized_neon(type, 32); \
+ intra_pred_highbd_sized_neon(type, 64);
+
+intra_pred_square(dc);
+#undef intra_pred_square