Merge changes Iaf8cbe95,I6748183d,I2a49811d

* changes:
  add vp9_satd_neon
  fix vp9_satd_sse2
  vp9_satd: return an int

test/vp9_avg_test.cc

@@ -194,6 +194,48 @@ class IntProColTest
   int16_t sum_c_;
 };
 
+typedef int (*SatdFunc)(const int16_t *coeffs, int length);
+typedef std::tr1::tuple<int, SatdFunc> SatdTestParam;
+
+class SatdTest
+    : public ::testing::Test,
+      public ::testing::WithParamInterface<SatdTestParam> {
+ protected:
+  virtual void SetUp() {
+    satd_size_ = GET_PARAM(0);
+    satd_func_ = GET_PARAM(1);
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    src_ = reinterpret_cast<int16_t*>(
+        vpx_memalign(16, sizeof(*src_) * satd_size_));
+    ASSERT_TRUE(src_ != NULL);
+  }
+
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+    vpx_free(src_);
+  }
+
+  void FillConstant(const int16_t val) {
+    for (int i = 0; i < satd_size_; ++i) src_[i] = val;
+  }
+
+  void FillRandom() {
+    for (int i = 0; i < satd_size_; ++i) src_[i] = rnd_.Rand16();
+  }
+
+  void Check(const int expected) {
+    int total;
+    ASM_REGISTER_STATE_CHECK(total = satd_func_(src_, satd_size_));
+    EXPECT_EQ(expected, total);
+  }
+
+  int satd_size_;
+
+ private:
+  int16_t *src_;
+  SatdFunc satd_func_;
+  ACMRandom rnd_;
+};
 
 uint8_t* AverageTestBase::source_data_ = NULL;
 
@@ -246,6 +288,36 @@ TEST_P(IntProColTest, Random) {
   RunComparison();
 }
 
+
+TEST_P(SatdTest, MinValue) {
+  const int kMin = -32640;
+  const int expected = -kMin * satd_size_;
+  FillConstant(kMin);
+  Check(expected);
+}
+
+TEST_P(SatdTest, MaxValue) {
+  const int kMax = 32640;
+  const int expected = kMax * satd_size_;
+  FillConstant(kMax);
+  Check(expected);
+}
+
+TEST_P(SatdTest, Random) {
+  int expected;
+  switch (satd_size_) {
+    case 16: expected = 205298; break;
+    case 64: expected = 1113950; break;
+    case 256: expected = 4268415; break;
+    case 1024: expected = 16954082; break;
+    default:
+      FAIL() << "Invalid satd size (" << satd_size_
+             << ") valid: 16/64/256/1024";
+  }
+  FillRandom();
+  Check(expected);
+}
+
 using std::tr1::make_tuple;
 
 INSTANTIATE_TEST_CASE_P(
@@ -254,6 +326,14 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(16, 16, 1, 8, &vp9_avg_8x8_c),
         make_tuple(16, 16, 1, 4, &vp9_avg_4x4_c)));
 
+INSTANTIATE_TEST_CASE_P(
+    C, SatdTest,
+    ::testing::Values(
+        make_tuple(16, &vp9_satd_c),
+        make_tuple(64, &vp9_satd_c),
+        make_tuple(256, &vp9_satd_c),
+        make_tuple(1024, &vp9_satd_c)));
+
 #if HAVE_SSE2
 INSTANTIATE_TEST_CASE_P(
     SSE2, AverageTest,
@@ -276,6 +356,14 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(16, &vp9_int_pro_col_sse2, &vp9_int_pro_col_c),
         make_tuple(32, &vp9_int_pro_col_sse2, &vp9_int_pro_col_c),
         make_tuple(64, &vp9_int_pro_col_sse2, &vp9_int_pro_col_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, SatdTest,
+    ::testing::Values(
+        make_tuple(16, &vp9_satd_sse2),
+        make_tuple(64, &vp9_satd_sse2),
+        make_tuple(256, &vp9_satd_sse2),
+        make_tuple(1024, &vp9_satd_sse2)));
 #endif
 
 #if HAVE_NEON
@@ -297,6 +385,14 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(16, &vp9_int_pro_col_neon, &vp9_int_pro_col_c),
         make_tuple(32, &vp9_int_pro_col_neon, &vp9_int_pro_col_c),
         make_tuple(64, &vp9_int_pro_col_neon, &vp9_int_pro_col_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, SatdTest,
+    ::testing::Values(
+        make_tuple(16, &vp9_satd_neon),
+        make_tuple(64, &vp9_satd_neon),
+        make_tuple(256, &vp9_satd_neon),
+        make_tuple(1024, &vp9_satd_neon)));
 #endif
 
 #if HAVE_MSA

vp9/common/vp9_rtcd_defs.pl

@@ -209,8 +209,8 @@ specialize qw/vp9_hadamard_8x8 sse2/, "$ssse3_x86_64_x86inc";
 add_proto qw/void vp9_hadamard_16x16/, "int16_t const *src_diff, int src_stride, int16_t *coeff";
 specialize qw/vp9_hadamard_16x16 sse2/;
 
-add_proto qw/int16_t vp9_satd/, "const int16_t *coeff, int length";
-specialize qw/vp9_satd sse2/;
+add_proto qw/int vp9_satd/, "const int16_t *coeff, int length";
+specialize qw/vp9_satd sse2 neon/;
 
 add_proto qw/void vp9_int_pro_row/, "int16_t *hbuf, uint8_t const *ref, const int ref_stride, const int height";
 specialize qw/vp9_int_pro_row sse2 neon/;

vp9/encoder/arm/neon/vp9_avg_neon.c

@@ -50,6 +50,33 @@ unsigned int vp9_avg_8x8_neon(const uint8_t *s, int p) {
   return (horizontal_add_u16x8(v_sum) + 32) >> 6;
 }
 
+// coeff: 16 bits, dynamic range [-32640, 32640].
+// length: value range {16, 64, 256, 1024}.
+int vp9_satd_neon(const int16_t *coeff, int length) {
+  const int16x4_t zero = vdup_n_s16(0);
+  int32x4_t accum = vdupq_n_s32(0);
+
+  do {
+    const int16x8_t src0 = vld1q_s16(coeff);
+    const int16x8_t src8 = vld1q_s16(coeff + 8);
+    accum = vabal_s16(accum, vget_low_s16(src0), zero);
+    accum = vabal_s16(accum, vget_high_s16(src0), zero);
+    accum = vabal_s16(accum, vget_low_s16(src8), zero);
+    accum = vabal_s16(accum, vget_high_s16(src8), zero);
+    length -= 16;
+    coeff += 16;
+  } while (length != 0);
+
+  {
+    // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
+    const int64x2_t s0 = vpaddlq_s32(accum);  // cascading summation of 'accum'.
+    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
+                                  vreinterpret_s32_s64(vget_high_s64(s0)));
+    const int satd = vget_lane_s32(s1, 0);
+    return satd;
+  }
+}
+
 void vp9_int_pro_row_neon(int16_t hbuf[16], uint8_t const *ref,
                           const int ref_stride, const int height) {
   int i;

vp9/encoder/vp9_avg.c

@@ -117,14 +117,14 @@ void vp9_hadamard_16x16_c(int16_t const *src_diff, int src_stride,
 
 // coeff: 16 bits, dynamic range [-32640, 32640].
 // length: value range {16, 64, 256, 1024}.
-int16_t vp9_satd_c(const int16_t *coeff, int length) {
+int vp9_satd_c(const int16_t *coeff, int length) {
   int i;
   int satd = 0;
   for (i = 0; i < length; ++i)
     satd += abs(coeff[i]);
 
   // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
-  return (int16_t)satd;
+  return satd;
 }
 
 // Integer projection onto row vectors.

vp9/encoder/vp9_pickmode.c

@@ -673,7 +673,7 @@ static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, int64_t *dist,
         if (*eob == 1)
           *rate += (int)abs(qcoeff[0]);
         else if (*eob > 1)
-          *rate += (int)vp9_satd((const int16_t *)qcoeff, step << 4);
+          *rate += vp9_satd((const int16_t *)qcoeff, step << 4);
 
         *dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> shift;
       }

vp9/encoder/x86/vp9_avg_intrin_sse2.c

@@ -283,31 +283,30 @@ void vp9_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride,
   }
 }
 
-int16_t vp9_satd_sse2(const int16_t *coeff, int length) {
+int vp9_satd_sse2(const int16_t *coeff, int length) {
   int i;
-  __m128i sum = _mm_load_si128((const __m128i *)coeff);
-  __m128i sign = _mm_srai_epi16(sum, 15);
-  __m128i val = _mm_xor_si128(sum, sign);
-  sum = _mm_sub_epi16(val, sign);
-  coeff += 8;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i accum = zero;
 
-  for (i = 8; i < length; i += 8) {
-    __m128i src_line = _mm_load_si128((const __m128i *)coeff);
-    sign = _mm_srai_epi16(src_line, 15);
-    val = _mm_xor_si128(src_line, sign);
-    val = _mm_sub_epi16(val, sign);
-    sum = _mm_add_epi16(sum, val);
+  for (i = 0; i < length; i += 8) {
+    const __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    const __m128i inv = _mm_sub_epi16(zero, src_line);
+    const __m128i abs = _mm_max_epi16(src_line, inv);  // abs(src_line)
+    const __m128i abs_lo = _mm_unpacklo_epi16(abs, zero);
+    const __m128i abs_hi = _mm_unpackhi_epi16(abs, zero);
+    const __m128i sum = _mm_add_epi32(abs_lo, abs_hi);
+    accum = _mm_add_epi32(accum, sum);
     coeff += 8;
   }
 
-  val = _mm_srli_si128(sum, 8);
-  sum = _mm_add_epi16(sum, val);
-  val = _mm_srli_epi64(sum, 32);
-  sum = _mm_add_epi16(sum, val);
-  val = _mm_srli_epi32(sum, 16);
-  sum = _mm_add_epi16(sum, val);
+  {  // cascading summation of accum
+    __m128i hi = _mm_srli_si128(accum, 8);
+    accum = _mm_add_epi32(accum, hi);
+    hi = _mm_srli_epi64(accum, 32);
+    accum = _mm_add_epi32(accum, hi);
+  }
 
-  return _mm_extract_epi16(sum, 0);
+  return _mm_cvtsi128_si32(accum);
 }
 
 void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,