From: Linfeng Zhang <linfengz@google.com>
Date: Mon, 5 Mar 2018 23:16:49 +0000 (-0800)
Subject: Fix a bug in vp9_iht8x8_64_add_neon()
X-Git-Tag: v1.8.0~817^2~1
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=c244a862341826713c4ad17b1f7c637e8395cdfc;p=libvpx

Fix a bug in vp9_iht8x8_64_add_neon()

This bug was introduced in b14b616d.

BUG=webm:1403

Change-Id: I84b2733734982e52b66548850d61758c772b5494
---

diff --git a/test/dct_test.cc b/test/dct_test.cc
index 2a6fccb67..bc1afbbe2 100644
--- a/test/dct_test.cc
+++ b/test/dct_test.cc
@@ -639,10 +639,10 @@ static const FuncInfo ht_neon_func_info[] = {
 #endif
 #endif
   { &vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 1 },
-// TODO(linfengz): reenable these functions once test vector failures are
-// addressed.
-#if 0
+  // TODO(linfengz): reenable these functions once test vector failures are
+  // addressed.
   { &vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 1 },
+#if 0
   { &vp9_fht16x16_c, &iht_wrapper<vp9_iht16x16_256_add_neon>, 16, 1 }
 #endif
 };
diff --git a/test/fdct8x8_test.cc b/test/fdct8x8_test.cc
index 15033dbc1..d1b49dfa7 100644
--- a/test/fdct8x8_test.cc
+++ b/test/fdct8x8_test.cc
@@ -675,9 +675,8 @@ INSTANTIATE_TEST_CASE_P(NEON, FwdTrans8x8DCT,
                         ::testing::Values(make_tuple(&vpx_fdct8x8_neon,
                                                      &vpx_idct8x8_64_add_neon,
                                                      0, VPX_BITS_8)));
-// TODO(linfengz): reenable these functions once test vector failures are
-// addressed.
-#if 0   // !CONFIG_VP9_HIGHBITDEPTH
+
+#if !CONFIG_VP9_HIGHBITDEPTH
 INSTANTIATE_TEST_CASE_P(
     NEON, FwdTrans8x8HT,
     ::testing::Values(
diff --git a/vp9/common/arm/neon/vp9_iht_neon.h b/vp9/common/arm/neon/vp9_iht_neon.h
index e918ebc7f..965eff36b 100644
--- a/vp9/common/arm/neon/vp9_iht_neon.h
+++ b/vp9/common/arm/neon/vp9_iht_neon.h
@@ -59,14 +59,17 @@ static INLINE void iadst4(int16x8_t *const io) {
 
 static INLINE void iadst_half_butterfly_neon(int16x8_t *const x,
                                              const int16x4_t c) {
-  const int16x8_t sum = vaddq_s16(x[0], x[1]);
-  const int16x8_t sub = vsubq_s16(x[0], x[1]);
+  // Don't add/sub before multiply, which will overflow in iadst8.
+  const int32x4_t x0_lo = vmull_lane_s16(vget_low_s16(x[0]), c, 0);
+  const int32x4_t x0_hi = vmull_lane_s16(vget_high_s16(x[0]), c, 0);
+  const int32x4_t x1_lo = vmull_lane_s16(vget_low_s16(x[1]), c, 0);
+  const int32x4_t x1_hi = vmull_lane_s16(vget_high_s16(x[1]), c, 0);
   int32x4_t t0[2], t1[2];
 
-  t0[0] = vmull_lane_s16(vget_low_s16(sum), c, 0);
-  t0[1] = vmull_lane_s16(vget_high_s16(sum), c, 0);
-  t1[0] = vmull_lane_s16(vget_low_s16(sub), c, 0);
-  t1[1] = vmull_lane_s16(vget_high_s16(sub), c, 0);
+  t0[0] = vaddq_s32(x0_lo, x1_lo);
+  t0[1] = vaddq_s32(x0_hi, x1_hi);
+  t1[0] = vsubq_s32(x0_lo, x1_lo);
+  t1[1] = vsubq_s32(x0_hi, x1_hi);
   x[0] = dct_const_round_shift_low_8(t0);
   x[1] = dct_const_round_shift_low_8(t1);
 }
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 274edfd44..2b15b661c 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -68,7 +68,7 @@ if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
   # Note that there are more specializations appended when
   # CONFIG_VP9_HIGHBITDEPTH is off.
   specialize qw/vp9_iht4x4_16_add neon sse2/;
-  specialize qw/vp9_iht8x8_64_add sse2/;
+  specialize qw/vp9_iht8x8_64_add neon sse2/;
   specialize qw/vp9_iht16x16_256_add sse2/;
   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") ne "yes") {
     # Note that these specializations are appended to the above ones.