Isolate vp10's inv_txfm from vp9

1) copy following files from vpx_dsp/ to vp10/common/
vp10_inv_txfm.c
vp10_inv_txfm.h
vp10_inv_txfm_sse2.c
vp10_inv_txfm_sse2.h

2) change the function prefix "vpx_" to "vp10_" in above files

3) add unit test at vp10_inv_txfm_test.cc

Change-Id: I206f10f60c8b27d872c84b7482c3bb1d1cb4b913

diff --git a/test/test.mk b/test/test.mk
index 6bb08be..3361c91 100644 (file)
--- a/test/test.mk
+++ b/test/test.mk
@@ -167,6 +167,7 @@ LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += sad_test.cc
 TEST_INTRA_PRED_SPEED_SRCS-$(CONFIG_VP9) := test_intra_pred_speed.cc
 TEST_INTRA_PRED_SPEED_SRCS-$(CONFIG_VP9) += ../md5_utils.h ../md5_utils.c
 
+LIBVPX_TEST_SRCS-$(CONFIG_VP10) += vp10_inv_txfm_test.cc
 endif # CONFIG_SHARED
 
 include $(SRC_PATH_BARE)/test/test-data.mk

diff --git a/test/vp10_inv_txfm_test.cc b/test/vp10_inv_txfm_test.cc
new file mode 100644 (file)
index 0000000..c49081e
--- /dev/null
+++ b/test/vp10_inv_txfm_test.cc
@@ -0,0 +1,321 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp10_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp10/common/blockd.h"
+#include "vp10/common/scan.h"
+#include "vpx/vpx_integer.h"
+#include "vp10/common/vp10_inv_txfm.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+const double PI = 3.141592653589793238462643383279502884;
+const double kInvSqrt2 = 0.707106781186547524400844362104;
+
+void reference_idct_1d(const double *in, double *out, int size) {
+  for (int n = 0; n < size; ++n) {
+    out[n] = 0;
+    for (int k = 0; k < size; ++k) {
+      if (k == 0)
+        out[n] += kInvSqrt2 * in[k] * cos(PI * (2 * n + 1) * k / (2 * size));
+      else
+        out[n] += in[k] * cos(PI * (2 * n + 1) * k / (2 * size));
+    }
+  }
+}
+
+typedef void (*IdctFuncRef)(const double *in, double *out, int size);
+typedef void (*IdctFunc)(const tran_low_t *in, tran_low_t *out);
+
+class TransTestBase {
+ public:
+  virtual ~TransTestBase() {}
+
+ protected:
+  void RunInvAccuracyCheck() {
+    tran_low_t *input  = new tran_low_t[txfm_size_];
+    tran_low_t *output = new tran_low_t[txfm_size_];
+    double *ref_input  = new double[txfm_size_];
+    double *ref_output = new double[txfm_size_];
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+    for (int ti =  0; ti < count_test_block; ++ti) {
+      for (int ni = 0; ni < txfm_size_; ++ni) {
+        input[ni] = rnd.Rand8() - rnd.Rand8();
+        ref_input[ni] = static_cast<double>(input[ni]);
+      }
+
+      fwd_txfm_(input, output);
+      fwd_txfm_ref_(ref_input, ref_output, txfm_size_);
+
+      for (int ni = 0; ni < txfm_size_; ++ni) {
+        EXPECT_LE(
+            abs(output[ni] - static_cast<tran_low_t>(round(ref_output[ni]))),
+            max_error_);
+      }
+    }
+
+    delete[] input;
+    delete[] output;
+    delete[] ref_input;
+    delete[] ref_output;
+  }
+
+  double max_error_;
+  int txfm_size_;
+  IdctFunc fwd_txfm_;
+  IdctFuncRef fwd_txfm_ref_;
+};
+
+typedef std::tr1::tuple<IdctFunc, IdctFuncRef, int, int> IdctParam;
+class Vp10InvTxfm
+    : public TransTestBase,
+      public ::testing::TestWithParam<IdctParam> {
+ public:
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    fwd_txfm_ref_ = GET_PARAM(1);
+    txfm_size_ = GET_PARAM(2);
+    max_error_ = GET_PARAM(3);
+  }
+  virtual void TearDown() {}
+};
+
+TEST_P(Vp10InvTxfm, RunInvAccuracyCheck) {
+  RunInvAccuracyCheck();
+}
+
+INSTANTIATE_TEST_CASE_P(
+    C, Vp10InvTxfm,
+    ::testing::Values(
+        IdctParam(&vp10_idct4_c, &reference_idct_1d, 4, 1),
+        IdctParam(&vp10_idct8_c, &reference_idct_1d, 8, 2),
+        IdctParam(&vp10_idct16_c, &reference_idct_1d, 16, 4),
+        IdctParam(&vp10_idct32_c, &reference_idct_1d, 32, 6))
+);
+
+typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef std::tr1::tuple<FwdTxfmFunc,
+                        InvTxfmFunc,
+                        InvTxfmFunc,
+                        TX_SIZE, int> PartialInvTxfmParam;
+const int kMaxNumCoeffs = 1024;
+class Vp10PartialIDctTest
+    : public ::testing::TestWithParam<PartialInvTxfmParam> {
+ public:
+  virtual ~Vp10PartialIDctTest() {}
+  virtual void SetUp() {
+    ftxfm_ = GET_PARAM(0);
+    full_itxfm_ = GET_PARAM(1);
+    partial_itxfm_ = GET_PARAM(2);
+    tx_size_  = GET_PARAM(3);
+    last_nonzero_ = GET_PARAM(4);
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  int last_nonzero_;
+  TX_SIZE tx_size_;
+  FwdTxfmFunc ftxfm_;
+  InvTxfmFunc full_itxfm_;
+  InvTxfmFunc partial_itxfm_;
+};
+
+TEST_P(Vp10PartialIDctTest, RunQuantCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int size;
+  switch (tx_size_) {
+    case TX_4X4:
+      size = 4;
+      break;
+    case TX_8X8:
+      size = 8;
+      break;
+    case TX_16X16:
+      size = 16;
+      break;
+    case TX_32X32:
+      size = 32;
+      break;
+    default:
+      FAIL() << "Wrong Size!";
+      break;
+  }
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block2[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst2[kMaxNumCoeffs]);
+
+  const int count_test_block = 1000;
+  const int block_size = size * size;
+
+  DECLARE_ALIGNED(16, int16_t, input_extreme_block[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kMaxNumCoeffs]);
+
+  int max_error = 0;
+  for (int i = 0; i < count_test_block; ++i) {
+    // clear out destination buffer
+    memset(dst1, 0, sizeof(*dst1) * block_size);
+    memset(dst2, 0, sizeof(*dst2) * block_size);
+    memset(test_coef_block1, 0, sizeof(*test_coef_block1) * block_size);
+    memset(test_coef_block2, 0, sizeof(*test_coef_block2) * block_size);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-255, 255].
+      if (i == 0) {
+        for (int j = 0; j < block_size; ++j)
+          input_extreme_block[j] = 255;
+      } else if (i == 1) {
+        for (int j = 0; j < block_size; ++j)
+          input_extreme_block[j] = -255;
+      } else {
+        for (int j = 0; j < block_size; ++j) {
+          input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
+        }
+      }
+
+      ftxfm_(input_extreme_block, output_ref_block, size);
+
+      // quantization with maximum allowed step sizes
+      test_coef_block1[0] = (output_ref_block[0] / 1336) * 1336;
+      for (int j = 1; j < last_nonzero_; ++j)
+        test_coef_block1[vp10_default_scan_orders[tx_size_].scan[j]]
+                         = (output_ref_block[j] / 1828) * 1828;
+    }
+
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block1, dst2, size));
+
+    for (int j = 0; j < block_size; ++j) {
+      const int diff = dst1[j] - dst2[j];
+      const int error = diff * diff;
+      if (max_error < error)
+        max_error = error;
+    }
+  }
+
+  EXPECT_EQ(0, max_error)
+      << "Error: partial inverse transform produces different results";
+}
+
+TEST_P(Vp10PartialIDctTest, ResultsMatch) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int size;
+  switch (tx_size_) {
+    case TX_4X4:
+      size = 4;
+      break;
+    case TX_8X8:
+      size = 8;
+      break;
+    case TX_16X16:
+      size = 16;
+      break;
+    case TX_32X32:
+      size = 32;
+      break;
+    default:
+      FAIL() << "Wrong Size!";
+      break;
+  }
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block2[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst2[kMaxNumCoeffs]);
+  const int count_test_block = 1000;
+  const int max_coeff = 32766 / 4;
+  const int block_size = size * size;
+  int max_error = 0;
+  for (int i = 0; i < count_test_block; ++i) {
+    // clear out destination buffer
+    memset(dst1, 0, sizeof(*dst1) * block_size);
+    memset(dst2, 0, sizeof(*dst2) * block_size);
+    memset(test_coef_block1, 0, sizeof(*test_coef_block1) * block_size);
+    memset(test_coef_block2, 0, sizeof(*test_coef_block2) * block_size);
+    int max_energy_leftover = max_coeff * max_coeff;
+    for (int j = 0; j < last_nonzero_; ++j) {
+      int16_t coef = static_cast<int16_t>(sqrt(1.0 * max_energy_leftover) *
+                                          (rnd.Rand16() - 32768) / 65536);
+      max_energy_leftover -= coef * coef;
+      if (max_energy_leftover < 0) {
+        max_energy_leftover = 0;
+        coef = 0;
+      }
+      test_coef_block1[vp10_default_scan_orders[tx_size_].scan[j]] = coef;
+    }
+
+    memcpy(test_coef_block2, test_coef_block1,
+           sizeof(*test_coef_block2) * block_size);
+
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block2, dst2, size));
+
+    for (int j = 0; j < block_size; ++j) {
+      const int diff = dst1[j] - dst2[j];
+      const int error = diff * diff;
+      if (max_error < error)
+        max_error = error;
+    }
+  }
+
+  EXPECT_EQ(0, max_error)
+      << "Error: partial inverse transform produces different results";
+}
+using std::tr1::make_tuple;
+
+INSTANTIATE_TEST_CASE_P(
+    C, Vp10PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_c,
+                   &vp10_idct32x32_1024_add_c,
+                   &vp10_idct32x32_34_add_c,
+                   TX_32X32, 34),
+        make_tuple(&vpx_fdct32x32_c,
+                   &vp10_idct32x32_1024_add_c,
+                   &vp10_idct32x32_1_add_c,
+                   TX_32X32, 1),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vp10_idct16x16_256_add_c,
+                   &vp10_idct16x16_10_add_c,
+                   TX_16X16, 10),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vp10_idct16x16_256_add_c,
+                   &vp10_idct16x16_1_add_c,
+                   TX_16X16, 1),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vp10_idct8x8_64_add_c,
+                   &vp10_idct8x8_12_add_c,
+                   TX_8X8, 12),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vp10_idct8x8_64_add_c,
+                   &vp10_idct8x8_1_add_c,
+                   TX_8X8, 1),
+        make_tuple(&vpx_fdct4x4_c,
+                   &vp10_idct4x4_16_add_c,
+                   &vp10_idct4x4_1_add_c,
+                   TX_4X4, 1)));
+}  // namespace

diff --git a/vp10/common/scan.c b/vp10/common/scan.c
index ad849af..7217f6d 100644 (file)
--- a/vp10/common/scan.c
+++ b/vp10/common/scan.c
@@ -695,6 +695,13 @@ DECLARE_ALIGNED(16, static const int16_t, vp10_default_iscan_32x32[1024]) = {
   1023,
 };
 
+const scan_order vp10_default_scan_orders[TX_SIZES] = {
+  {default_scan_4x4,   vp10_default_iscan_4x4,   default_scan_4x4_neighbors},
+  {default_scan_8x8,   vp10_default_iscan_8x8,   default_scan_8x8_neighbors},
+  {default_scan_16x16, vp10_default_iscan_16x16, default_scan_16x16_neighbors},
+  {default_scan_32x32, vp10_default_iscan_32x32, default_scan_32x32_neighbors},
+};
+
 const scan_order vp10_scan_orders[TX_SIZES][TX_TYPES] = {
   {  // TX_4X4
     {default_scan_4x4, vp10_default_iscan_4x4, default_scan_4x4_neighbors},

diff --git a/vp10/common/scan.h b/vp10/common/scan.h
index ada67f4..f5a020f 100644 (file)
--- a/vp10/common/scan.h
+++ b/vp10/common/scan.h
@@ -29,6 +29,7 @@ typedef struct {
   const int16_t *neighbors;
 } scan_order;
 
+extern const scan_order vp10_default_scan_orders[TX_SIZES];
 extern const scan_order vp10_scan_orders[TX_SIZES][TX_TYPES];
 
 static INLINE int get_coef_context(const int16_t *neighbors,

diff --git a/vp10/common/vp10_inv_txfm.c b/vp10/common/vp10_inv_txfm.c
new file mode 100644 (file)
index 0000000..403b209
--- /dev/null
+++ b/vp10/common/vp10_inv_txfm.c
@@ -0,0 +1,2499 @@
+/*
+ *
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <string.h>
+
+#include "vp10/common/vp10_inv_txfm.h"
+
+void vp10_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+/* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+   0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = WRAPLOW(a1, 8);
+    op[1] = WRAPLOW(b1, 8);
+    op[2] = WRAPLOW(c1, 8);
+    op[3] = WRAPLOW(d1, 8);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] = clip_pixel_add(dest[stride * 0], a1);
+    dest[stride * 1] = clip_pixel_add(dest[stride * 1], b1);
+    dest[stride * 2] = clip_pixel_add(dest[stride * 2], c1);
+    dest[stride * 3] = clip_pixel_add(dest[stride * 3], d1);
+
+    ip++;
+    dest++;
+  }
+}
+
+void vp10_iwht4x4_1_add_c(const tran_low_t *in,
+                          uint8_t *dest,
+                          int dest_stride) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = WRAPLOW(a1, 8);
+  op[1] = op[2] = op[3] = WRAPLOW(e1, 8);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] = clip_pixel_add(dest[dest_stride * 0], a1);
+    dest[dest_stride * 1] = clip_pixel_add(dest[dest_stride * 1], e1);
+    dest[dest_stride * 2] = clip_pixel_add(dest[dest_stride * 2], e1);
+    dest[dest_stride * 3] = clip_pixel_add(dest[dest_stride * 3], e1);
+    ip++;
+    dest++;
+  }
+}
+
+void vp10_idct4_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step[1] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step[3] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 2
+  output[0] = WRAPLOW(step[0] + step[3], 8);
+  output[1] = WRAPLOW(step[1] + step[2], 8);
+  output[2] = WRAPLOW(step[1] - step[2], 8);
+  output[3] = WRAPLOW(step[0] - step[3], 8);
+}
+
+void vp10_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    vp10_idct4_c(input, outptr);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    vp10_idct4_c(temp_in, temp_out);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 4));
+    }
+  }
+}
+
+void vp10_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest,
+                         int dest_stride) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = clip_pixel_add(dest[0], a1);
+    dest[1] = clip_pixel_add(dest[1], a1);
+    dest[2] = clip_pixel_add(dest[2], a1);
+    dest[3] = clip_pixel_add(dest[3], a1);
+    dest += dest_stride;
+  }
+}
+
+void vp10_idct8_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[8], step2[8];
+  tran_high_t temp1, temp2;
+  // stage 1
+  step1[0] = input[0];
+  step1[2] = input[4];
+  step1[1] = input[2];
+  step1[3] = input[6];
+  temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+  temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+  temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 2
+  temp1 = (step1[0] + step1[2]) * cospi_16_64;
+  temp2 = (step1[0] - step1[2]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step1[1] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[1] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[4] = WRAPLOW(step1[4] + step1[5], 8);
+  step2[5] = WRAPLOW(step1[4] - step1[5], 8);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], 8);
+  step2[7] = WRAPLOW(step1[6] + step1[7], 8);
+
+  // stage 3
+  step1[0] = WRAPLOW(step2[0] + step2[3], 8);
+  step1[1] = WRAPLOW(step2[1] + step2[2], 8);
+  step1[2] = WRAPLOW(step2[1] - step2[2], 8);
+  step1[3] = WRAPLOW(step2[0] - step2[3], 8);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[7] = step2[7];
+
+  // stage 4
+  output[0] = WRAPLOW(step1[0] + step1[7], 8);
+  output[1] = WRAPLOW(step1[1] + step1[6], 8);
+  output[2] = WRAPLOW(step1[2] + step1[5], 8);
+  output[3] = WRAPLOW(step1[3] + step1[4], 8);
+  output[4] = WRAPLOW(step1[3] - step1[4], 8);
+  output[5] = WRAPLOW(step1[2] - step1[5], 8);
+  output[6] = WRAPLOW(step1[1] - step1[6], 8);
+  output[7] = WRAPLOW(step1[0] - step1[7], 8);
+}
+
+void vp10_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  for (i = 0; i < 8; ++i) {
+    vp10_idct8_c(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vp10_idct8_c(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void vp10_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+void vp10_iadst4_c(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[0];
+  tran_low_t x1 = input[1];
+  tran_low_t x2 = input[2];
+  tran_low_t x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = x0 - x2 + x3;
+
+  s0 = s0 + s3 + s5;
+  s1 = s1 - s4 - s6;
+  s3 = s2;
+  s2 = sinpi_3_9 * s7;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = WRAPLOW(dct_const_round_shift(s0 + s3), 8);
+  output[1] = WRAPLOW(dct_const_round_shift(s1 + s3), 8);
+  output[2] = WRAPLOW(dct_const_round_shift(s2), 8);
+  output[3] = WRAPLOW(dct_const_round_shift(s0 + s1 - s3), 8);
+}
+
+void vp10_iadst8_c(const tran_low_t *input, tran_low_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_high_t x0 = input[7];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[5];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[3];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[1];
+  tran_high_t x7 = input[6];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = (int)(cospi_2_64  * x0 + cospi_30_64 * x1);
+  s1 = (int)(cospi_30_64 * x0 - cospi_2_64  * x1);
+  s2 = (int)(cospi_10_64 * x2 + cospi_22_64 * x3);
+  s3 = (int)(cospi_22_64 * x2 - cospi_10_64 * x3);
+  s4 = (int)(cospi_18_64 * x4 + cospi_14_64 * x5);
+  s5 = (int)(cospi_14_64 * x4 - cospi_18_64 * x5);
+  s6 = (int)(cospi_26_64 * x6 + cospi_6_64  * x7);
+  s7 = (int)(cospi_6_64  * x6 - cospi_26_64 * x7);
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s4), 8);
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s5), 8);
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s6), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s7), 8);
+  x4 = WRAPLOW(dct_const_round_shift(s0 - s4), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s1 - s5), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s2 - s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s3 - s7), 8);
+
+  // stage 2
+  s0 = (int)x0;
+  s1 = (int)x1;
+  s2 = (int)x2;
+  s3 = (int)x3;
+  s4 = (int)(cospi_8_64 * x4 + cospi_24_64 * x5);
+  s5 = (int)(cospi_24_64 * x4 - cospi_8_64 * x5);
+  s6 = (int)(-cospi_24_64 * x6 + cospi_8_64 * x7);
+  s7 = (int)(cospi_8_64 * x6 + cospi_24_64 * x7);
+
+  x0 = WRAPLOW(s0 + s2, 8);
+  x1 = WRAPLOW(s1 + s3, 8);
+  x2 = WRAPLOW(s0 - s2, 8);
+  x3 = WRAPLOW(s1 - s3, 8);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7), 8);
+
+  // stage 3
+  s2 = (int)(cospi_16_64 * (x2 + x3));
+  s3 = (int)(cospi_16_64 * (x2 - x3));
+  s6 = (int)(cospi_16_64 * (x6 + x7));
+  s7 = (int)(cospi_16_64 * (x6 - x7));
+
+  x2 = WRAPLOW(dct_const_round_shift(s2), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s7), 8);
+
+  output[0] = WRAPLOW(x0, 8);
+  output[1] = WRAPLOW(-x4, 8);
+  output[2] = WRAPLOW(x6, 8);
+  output[3] = WRAPLOW(-x2, 8);
+  output[4] = WRAPLOW(x3, 8);
+  output[5] = WRAPLOW(-x7, 8);
+  output[6] = WRAPLOW(x5, 8);
+  output[7] = WRAPLOW(-x1, 8);
+}
+
+void vp10_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  // only first 4 row has non-zero coefs
+  for (i = 0; i < 4; ++i) {
+    vp10_idct8_c(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vp10_idct8_c(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void vp10_idct16_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[16], step2[16];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0/2];
+  step1[1] = input[16/2];
+  step1[2] = input[8/2];
+  step1[3] = input[24/2];
+  step1[4] = input[4/2];
+  step1[5] = input[20/2];
+  step1[6] = input[12/2];
+  step1[7] = input[28/2];
+  step1[8] = input[2/2];
+  step1[9] = input[18/2];
+  step1[10] = input[10/2];
+  step1[11] = input[26/2];
+  step1[12] = input[6/2];
+  step1[13] = input[22/2];
+  step1[14] = input[14/2];
+  step1[15] = input[30/2];
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], 8);
+  step1[9] = WRAPLOW(step2[8] - step2[9], 8);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], 8);
+  step1[11] = WRAPLOW(step2[10] + step2[11], 8);
+  step1[12] = WRAPLOW(step2[12] + step2[13], 8);
+  step1[13] = WRAPLOW(step2[12] - step2[13], 8);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], 8);
+  step1[15] = WRAPLOW(step2[14] + step2[15], 8);
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[4] = WRAPLOW(step1[4] + step1[5], 8);
+  step2[5] = WRAPLOW(step1[4] - step1[5], 8);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], 8);
+  step2[7] = WRAPLOW(step1[6] + step1[7], 8);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], 8);
+  step1[1] = WRAPLOW(step2[1] + step2[2], 8);
+  step1[2] = WRAPLOW(step2[1] - step2[2], 8);
+  step1[3] = WRAPLOW(step2[0] - step2[3], 8);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], 8);
+  step1[9] = WRAPLOW(step2[9] + step2[10], 8);
+  step1[10] = WRAPLOW(step2[9] - step2[10], 8);
+  step1[11] = WRAPLOW(step2[8] - step2[11], 8);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], 8);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], 8);
+  step1[14] = WRAPLOW(step2[13] + step2[14], 8);
+  step1[15] = WRAPLOW(step2[12] + step2[15], 8);
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], 8);
+  step2[1] = WRAPLOW(step1[1] + step1[6], 8);
+  step2[2] = WRAPLOW(step1[2] + step1[5], 8);
+  step2[3] = WRAPLOW(step1[3] + step1[4], 8);
+  step2[4] = WRAPLOW(step1[3] - step1[4], 8);
+  step2[5] = WRAPLOW(step1[2] - step1[5], 8);
+  step2[6] = WRAPLOW(step1[1] - step1[6], 8);
+  step2[7] = WRAPLOW(step1[0] - step1[7], 8);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  // stage 7
+  output[0] = WRAPLOW(step2[0] + step2[15], 8);
+  output[1] = WRAPLOW(step2[1] + step2[14], 8);
+  output[2] = WRAPLOW(step2[2] + step2[13], 8);
+  output[3] = WRAPLOW(step2[3] + step2[12], 8);
+  output[4] = WRAPLOW(step2[4] + step2[11], 8);
+  output[5] = WRAPLOW(step2[5] + step2[10], 8);
+  output[6] = WRAPLOW(step2[6] + step2[9], 8);
+  output[7] = WRAPLOW(step2[7] + step2[8], 8);
+  output[8] = WRAPLOW(step2[7] - step2[8], 8);
+  output[9] = WRAPLOW(step2[6] - step2[9], 8);
+  output[10] = WRAPLOW(step2[5] - step2[10], 8);
+  output[11] = WRAPLOW(step2[4] - step2[11], 8);
+  output[12] = WRAPLOW(step2[3] - step2[12], 8);
+  output[13] = WRAPLOW(step2[2] - step2[13], 8);
+  output[14] = WRAPLOW(step2[1] - step2[14], 8);
+  output[15] = WRAPLOW(step2[0] - step2[15], 8);
+}
+
+void vp10_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows
+  for (i = 0; i < 16; ++i) {
+    vp10_idct16_c(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    vp10_idct16_c(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp10_iadst16_c(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_high_t x0 = input[15];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[13];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[11];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[9];
+  tran_high_t x7 = input[6];
+  tran_high_t x8 = input[7];
+  tran_high_t x9 = input[8];
+  tran_high_t x10 = input[5];
+  tran_high_t x11 = input[10];
+  tran_high_t x12 = input[3];
+  tran_high_t x13 = input[12];
+  tran_high_t x14 = input[1];
+  tran_high_t x15 = input[14];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8
+           | x9 | x10 | x11 | x12 | x13 | x14 | x15)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = output[8]
+              = output[9] = output[10] = output[11] = output[12]
+              = output[13] = output[14] = output[15] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s8), 8);
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s9), 8);
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s10), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s11), 8);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s12), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s13), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s6 + s14), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s7 + s15), 8);
+  x8 = WRAPLOW(dct_const_round_shift(s0 - s8), 8);
+  x9 = WRAPLOW(dct_const_round_shift(s1 - s9), 8);
+  x10 = WRAPLOW(dct_const_round_shift(s2 - s10), 8);
+  x11 = WRAPLOW(dct_const_round_shift(s3 - s11), 8);
+  x12 = WRAPLOW(dct_const_round_shift(s4 - s12), 8);
+  x13 = WRAPLOW(dct_const_round_shift(s5 - s13), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s6 - s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s7 - s15), 8);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 =    x8 * cospi_4_64   + x9 * cospi_28_64;
+  s9 =    x8 * cospi_28_64  - x9 * cospi_4_64;
+  s10 =   x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 =   x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = - x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 =   x12 * cospi_4_64  + x13 * cospi_28_64;
+  s14 = - x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 =   x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = WRAPLOW(s0 + s4, 8);
+  x1 = WRAPLOW(s1 + s5, 8);
+  x2 = WRAPLOW(s2 + s6, 8);
+  x3 = WRAPLOW(s3 + s7, 8);
+  x4 = WRAPLOW(s0 - s4, 8);
+  x5 = WRAPLOW(s1 - s5, 8);
+  x6 = WRAPLOW(s2 - s6, 8);
+  x7 = WRAPLOW(s3 - s7, 8);
+  x8 = WRAPLOW(dct_const_round_shift(s8 + s12), 8);
+  x9 = WRAPLOW(dct_const_round_shift(s9 + s13), 8);
+  x10 = WRAPLOW(dct_const_round_shift(s10 + s14), 8);
+  x11 = WRAPLOW(dct_const_round_shift(s11 + s15), 8);
+  x12 = WRAPLOW(dct_const_round_shift(s8 - s12), 8);
+  x13 = WRAPLOW(dct_const_round_shift(s9 - s13), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s10 - s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s11 - s15), 8);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64  + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = - x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 =   x6 * cospi_8_64  + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64  + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = - x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 =   x14 * cospi_8_64  + x15 * cospi_24_64;
+
+  x0 = WRAPLOW(check_range(s0 + s2), 8);
+  x1 = WRAPLOW(check_range(s1 + s3), 8);
+  x2 = WRAPLOW(check_range(s0 - s2), 8);
+  x3 = WRAPLOW(check_range(s1 - s3), 8);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7), 8);
+  x8 = WRAPLOW(check_range(s8 + s10), 8);
+  x9 = WRAPLOW(check_range(s9 + s11), 8);
+  x10 = WRAPLOW(check_range(s8 - s10), 8);
+  x11 = WRAPLOW(check_range(s9 - s11), 8);
+  x12 = WRAPLOW(dct_const_round_shift(s12 + s14), 8);
+  x13 = WRAPLOW(dct_const_round_shift(s13 + s15), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s12 - s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s13 - s15), 8);
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (- x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (- x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = WRAPLOW(dct_const_round_shift(s2), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s7), 8);
+  x10 = WRAPLOW(dct_const_round_shift(s10), 8);
+  x11 = WRAPLOW(dct_const_round_shift(s11), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s15), 8);
+
+  output[0] = WRAPLOW(x0, 8);
+  output[1] = WRAPLOW(-x8, 8);
+  output[2] = WRAPLOW(x12, 8);
+  output[3] = WRAPLOW(-x4, 8);
+  output[4] = WRAPLOW(x6, 8);
+  output[5] = WRAPLOW(x14, 8);
+  output[6] = WRAPLOW(x10, 8);
+  output[7] = WRAPLOW(x2, 8);
+  output[8] = WRAPLOW(x3, 8);
+  output[9] = WRAPLOW(x11, 8);
+  output[10] = WRAPLOW(x15, 8);
+  output[11] = WRAPLOW(x7, 8);
+  output[12] = WRAPLOW(x5, 8);
+  output[13] = WRAPLOW(-x13, 8);
+  output[14] = WRAPLOW(x9, 8);
+  output[15] = WRAPLOW(-x1, 8);
+}
+
+void vp10_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  for (i = 0; i < 4; ++i) {
+    vp10_idct16_c(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j*16 + i];
+    vp10_idct16_c(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(
+          dest[j * stride + i],
+          ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp10_idct16x16_1_add_c(const tran_low_t *input,
+                            uint8_t *dest,
+                            int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  for (j = 0; j < 16; ++j) {
+    for (i = 0; i < 16; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+void vp10_idct32_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[32], step2[32];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0];
+  step1[1] = input[16];
+  step1[2] = input[8];
+  step1[3] = input[24];
+  step1[4] = input[4];
+  step1[5] = input[20];
+  step1[6] = input[12];
+  step1[7] = input[28];
+  step1[8] = input[2];
+  step1[9] = input[18];
+  step1[10] = input[10];
+  step1[11] = input[26];
+  step1[12] = input[6];
+  step1[13] = input[22];
+  step1[14] = input[14];
+  step1[15] = input[30];
+
+  temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64;
+  temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64;
+  step1[16] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[31] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64;
+  temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64;
+  temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64;
+  temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64;
+  temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64;
+  temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64;
+  temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64;
+  temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  step2[16] = WRAPLOW(step1[16] + step1[17], 8);
+  step2[17] = WRAPLOW(step1[16] - step1[17], 8);
+  step2[18] = WRAPLOW(-step1[18] + step1[19], 8);
+  step2[19] = WRAPLOW(step1[18] + step1[19], 8);
+  step2[20] = WRAPLOW(step1[20] + step1[21], 8);
+  step2[21] = WRAPLOW(step1[20] - step1[21], 8);
+  step2[22] = WRAPLOW(-step1[22] + step1[23], 8);
+  step2[23] = WRAPLOW(step1[22] + step1[23], 8);
+  step2[24] = WRAPLOW(step1[24] + step1[25], 8);
+  step2[25] = WRAPLOW(step1[24] - step1[25], 8);
+  step2[26] = WRAPLOW(-step1[26] + step1[27], 8);
+  step2[27] = WRAPLOW(step1[26] + step1[27], 8);
+  step2[28] = WRAPLOW(step1[28] + step1[29], 8);
+  step2[29] = WRAPLOW(step1[28] - step1[29], 8);
+  step2[30] = WRAPLOW(-step1[30] + step1[31], 8);
+  step2[31] = WRAPLOW(step1[30] + step1[31], 8);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], 8);
+  step1[9] = WRAPLOW(step2[8] - step2[9], 8);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], 8);
+  step1[11] = WRAPLOW(step2[10] + step2[11], 8);
+  step1[12] = WRAPLOW(step2[12] + step2[13], 8);
+  step1[13] = WRAPLOW(step2[12] - step2[13], 8);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], 8);
+  step1[15] = WRAPLOW(step2[14] + step2[15], 8);
+
+  step1[16] = step2[16];
+  step1[31] = step2[31];
+  temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64;
+  temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64;
+  temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[19] = step2[19];
+  step1[20] = step2[20];
+  temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64;
+  temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64;
+  temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[27] = step2[27];
+  step1[28] = step2[28];
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[4] = WRAPLOW(step1[4] + step1[5], 8);
+  step2[5] = WRAPLOW(step1[4] - step1[5], 8);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], 8);
+  step2[7] = WRAPLOW(step1[6] + step1[7], 8);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  step2[16] = WRAPLOW(step1[16] + step1[19], 8);
+  step2[17] = WRAPLOW(step1[17] + step1[18], 8);
+  step2[18] = WRAPLOW(step1[17] - step1[18], 8);
+  step2[19] = WRAPLOW(step1[16] - step1[19], 8);
+  step2[20] = WRAPLOW(-step1[20] + step1[23], 8);
+  step2[21] = WRAPLOW(-step1[21] + step1[22], 8);
+  step2[22] = WRAPLOW(step1[21] + step1[22], 8);
+  step2[23] = WRAPLOW(step1[20] + step1[23], 8);
+
+  step2[24] = WRAPLOW(step1[24] + step1[27], 8);
+  step2[25] = WRAPLOW(step1[25] + step1[26], 8);
+  step2[26] = WRAPLOW(step1[25] - step1[26], 8);
+  step2[27] = WRAPLOW(step1[24] - step1[27], 8);
+  step2[28] = WRAPLOW(-step1[28] + step1[31], 8);
+  step2[29] = WRAPLOW(-step1[29] + step1[30], 8);
+  step2[30] = WRAPLOW(step1[29] + step1[30], 8);
+  step2[31] = WRAPLOW(step1[28] + step1[31], 8);
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], 8);
+  step1[1] = WRAPLOW(step2[1] + step2[2], 8);
+  step1[2] = WRAPLOW(step2[1] - step2[2], 8);
+  step1[3] = WRAPLOW(step2[0] - step2[3], 8);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], 8);
+  step1[9] = WRAPLOW(step2[9] + step2[10], 8);
+  step1[10] = WRAPLOW(step2[9] - step2[10], 8);
+  step1[11] = WRAPLOW(step2[8] - step2[11], 8);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], 8);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], 8);
+  step1[14] = WRAPLOW(step2[13] + step2[14], 8);
+  step1[15] = WRAPLOW(step2[12] + step2[15], 8);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64;
+  temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64;
+  temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64;
+  temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64;
+  temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[22] = step2[22];
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[25] = step2[25];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], 8);
+  step2[1] = WRAPLOW(step1[1] + step1[6], 8);
+  step2[2] = WRAPLOW(step1[2] + step1[5], 8);
+  step2[3] = WRAPLOW(step1[3] + step1[4], 8);
+  step2[4] = WRAPLOW(step1[3] - step1[4], 8);
+  step2[5] = WRAPLOW(step1[2] - step1[5], 8);
+  step2[6] = WRAPLOW(step1[1] - step1[6], 8);
+  step2[7] = WRAPLOW(step1[0] - step1[7], 8);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  step2[16] = WRAPLOW(step1[16] + step1[23], 8);
+  step2[17] = WRAPLOW(step1[17] + step1[22], 8);
+  step2[18] = WRAPLOW(step1[18] + step1[21], 8);
+  step2[19] = WRAPLOW(step1[19] + step1[20], 8);
+  step2[20] = WRAPLOW(step1[19] - step1[20], 8);
+  step2[21] = WRAPLOW(step1[18] - step1[21], 8);
+  step2[22] = WRAPLOW(step1[17] - step1[22], 8);
+  step2[23] = WRAPLOW(step1[16] - step1[23], 8);
+
+  step2[24] = WRAPLOW(-step1[24] + step1[31], 8);
+  step2[25] = WRAPLOW(-step1[25] + step1[30], 8);
+  step2[26] = WRAPLOW(-step1[26] + step1[29], 8);
+  step2[27] = WRAPLOW(-step1[27] + step1[28], 8);
+  step2[28] = WRAPLOW(step1[27] + step1[28], 8);
+  step2[29] = WRAPLOW(step1[26] + step1[29], 8);
+  step2[30] = WRAPLOW(step1[25] + step1[30], 8);
+  step2[31] = WRAPLOW(step1[24] + step1[31], 8);
+
+  // stage 7
+  step1[0] = WRAPLOW(step2[0] + step2[15], 8);
+  step1[1] = WRAPLOW(step2[1] + step2[14], 8);
+  step1[2] = WRAPLOW(step2[2] + step2[13], 8);
+  step1[3] = WRAPLOW(step2[3] + step2[12], 8);
+  step1[4] = WRAPLOW(step2[4] + step2[11], 8);
+  step1[5] = WRAPLOW(step2[5] + step2[10], 8);
+  step1[6] = WRAPLOW(step2[6] + step2[9], 8);
+  step1[7] = WRAPLOW(step2[7] + step2[8], 8);
+  step1[8] = WRAPLOW(step2[7] - step2[8], 8);
+  step1[9] = WRAPLOW(step2[6] - step2[9], 8);
+  step1[10] = WRAPLOW(step2[5] - step2[10], 8);
+  step1[11] = WRAPLOW(step2[4] - step2[11], 8);
+  step1[12] = WRAPLOW(step2[3] - step2[12], 8);
+  step1[13] = WRAPLOW(step2[2] - step2[13], 8);
+  step1[14] = WRAPLOW(step2[1] - step2[14], 8);
+  step1[15] = WRAPLOW(step2[0] - step2[15], 8);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  step1[18] = step2[18];
+  step1[19] = step2[19];
+  temp1 = (-step2[20] + step2[27]) * cospi_16_64;
+  temp2 = (step2[20] + step2[27]) * cospi_16_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step2[21] + step2[26]) * cospi_16_64;
+  temp2 = (step2[21] + step2[26]) * cospi_16_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step2[22] + step2[25]) * cospi_16_64;
+  temp2 = (step2[22] + step2[25]) * cospi_16_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step2[23] + step2[24]) * cospi_16_64;
+  temp2 = (step2[23] + step2[24]) * cospi_16_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[28] = step2[28];
+  step1[29] = step2[29];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // final stage
+  output[0] = WRAPLOW(step1[0] + step1[31], 8);
+  output[1] = WRAPLOW(step1[1] + step1[30], 8);
+  output[2] = WRAPLOW(step1[2] + step1[29], 8);
+  output[3] = WRAPLOW(step1[3] + step1[28], 8);
+  output[4] = WRAPLOW(step1[4] + step1[27], 8);
+  output[5] = WRAPLOW(step1[5] + step1[26], 8);
+  output[6] = WRAPLOW(step1[6] + step1[25], 8);
+  output[7] = WRAPLOW(step1[7] + step1[24], 8);
+  output[8] = WRAPLOW(step1[8] + step1[23], 8);
+  output[9] = WRAPLOW(step1[9] + step1[22], 8);
+  output[10] = WRAPLOW(step1[10] + step1[21], 8);
+  output[11] = WRAPLOW(step1[11] + step1[20], 8);
+  output[12] = WRAPLOW(step1[12] + step1[19], 8);
+  output[13] = WRAPLOW(step1[13] + step1[18], 8);
+  output[14] = WRAPLOW(step1[14] + step1[17], 8);
+  output[15] = WRAPLOW(step1[15] + step1[16], 8);
+  output[16] = WRAPLOW(step1[15] - step1[16], 8);
+  output[17] = WRAPLOW(step1[14] - step1[17], 8);
+  output[18] = WRAPLOW(step1[13] - step1[18], 8);
+  output[19] = WRAPLOW(step1[12] - step1[19], 8);
+  output[20] = WRAPLOW(step1[11] - step1[20], 8);
+  output[21] = WRAPLOW(step1[10] - step1[21], 8);
+  output[22] = WRAPLOW(step1[9] - step1[22], 8);
+  output[23] = WRAPLOW(step1[8] - step1[23], 8);
+  output[24] = WRAPLOW(step1[7] - step1[24], 8);
+  output[25] = WRAPLOW(step1[6] - step1[25], 8);
+  output[26] = WRAPLOW(step1[5] - step1[26], 8);
+  output[27] = WRAPLOW(step1[4] - step1[27], 8);
+  output[28] = WRAPLOW(step1[3] - step1[28], 8);
+  output[29] = WRAPLOW(step1[2] - step1[29], 8);
+  output[30] = WRAPLOW(step1[1] - step1[30], 8);
+  output[31] = WRAPLOW(step1[0] - step1[31], 8);
+}
+
+void vp10_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  tran_low_t out[32 * 32];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    int16_t zero_coeff[16];
+    for (j = 0; j < 16; ++j)
+      zero_coeff[j] = input[2 * j] | input[2 * j + 1];
+    for (j = 0; j < 8; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 4; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 2; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+
+    if (zero_coeff[0] | zero_coeff[1])
+      vp10_idct32_c(input, outptr);
+    else
+      memset(outptr, 0, sizeof(tran_low_t) * 32);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    vp10_idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp10_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[32 * 32] = {0};
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  // only upper-left 8x8 has non-zero coeff
+  for (i = 0; i < 8; ++i) {
+    vp10_idct32_c(input, outptr);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    vp10_idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(
+          dest[j * stride + i],
+          ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp10_idct32x32_1_add_c(const tran_low_t *input,
+                            uint8_t *dest,
+                            int stride) {
+  int i, j;
+  tran_high_t a1;
+
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  for (j = 0; j < 32; ++j) {
+    for (i = 0; i < 32; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp10_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  /* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+     0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = WRAPLOW(a1, bd);
+    op[1] = WRAPLOW(b1, bd);
+    op[2] = WRAPLOW(c1, bd);
+    op[3] = WRAPLOW(d1, bd);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] = highbd_clip_pixel_add(dest[stride * 0], a1, bd);
+    dest[stride * 1] = highbd_clip_pixel_add(dest[stride * 1], b1, bd);
+    dest[stride * 2] = highbd_clip_pixel_add(dest[stride * 2], c1, bd);
+    dest[stride * 3] = highbd_clip_pixel_add(dest[stride * 3], d1, bd);
+
+    ip++;
+    dest++;
+  }
+}
+
+void vp10_highbd_iwht4x4_1_add_c(const tran_low_t *in, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  (void) bd;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = WRAPLOW(a1, bd);
+  op[1] = op[2] = op[3] = WRAPLOW(e1, bd);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] = highbd_clip_pixel_add(
+        dest[dest_stride * 0], a1, bd);
+    dest[dest_stride * 1] = highbd_clip_pixel_add(
+        dest[dest_stride * 1], e1, bd);
+    dest[dest_stride * 2] = highbd_clip_pixel_add(
+        dest[dest_stride * 2], e1, bd);
+    dest[dest_stride * 3] = highbd_clip_pixel_add(
+        dest[dest_stride * 3], e1, bd);
+    ip++;
+    dest++;
+  }
+}
+
+void vp10_highbd_idct4_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  (void) bd;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step[1] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step[3] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 2
+  output[0] = WRAPLOW(step[0] + step[3], bd);
+  output[1] = WRAPLOW(step[1] + step[2], bd);
+  output[2] = WRAPLOW(step[1] - step[2], bd);
+  output[3] = WRAPLOW(step[0] - step[3], bd);
+}
+
+void vp10_highbd_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    vp10_highbd_idct4_c(input, outptr, bd);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    vp10_highbd_idct4_c(temp_in, temp_out, bd);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+    }
+  }
+}
+
+void vp10_highbd_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = highbd_clip_pixel_add(dest[0], a1, bd);
+    dest[1] = highbd_clip_pixel_add(dest[1], a1, bd);
+    dest[2] = highbd_clip_pixel_add(dest[2], a1, bd);
+    dest[3] = highbd_clip_pixel_add(dest[3], a1, bd);
+    dest += dest_stride;
+  }
+}
+
+void vp10_highbd_idct8_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step1[8], step2[8];
+  tran_high_t temp1, temp2;
+  // stage 1
+  step1[0] = input[0];
+  step1[2] = input[4];
+  step1[1] = input[2];
+  step1[3] = input[6];
+  temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+  temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+  step1[4] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[7] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+  temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 2 & stage 3 - even half
+  vp10_highbd_idct4_c(step1, step1, bd);
+
+  // stage 2 - odd half
+  step2[4] = WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = WRAPLOW(step1[6] + step1[7], bd);
+
+  // stage 3 - odd half
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[7] = step2[7];
+
+  // stage 4
+  output[0] = WRAPLOW(step1[0] + step1[7], bd);
+  output[1] = WRAPLOW(step1[1] + step1[6], bd);
+  output[2] = WRAPLOW(step1[2] + step1[5], bd);
+  output[3] = WRAPLOW(step1[3] + step1[4], bd);
+  output[4] = WRAPLOW(step1[3] - step1[4], bd);
+  output[5] = WRAPLOW(step1[2] - step1[5], bd);
+  output[6] = WRAPLOW(step1[1] - step1[6], bd);
+  output[7] = WRAPLOW(step1[0] - step1[7], bd);
+}
+
+void vp10_highbd_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  for (i = 0; i < 8; ++i) {
+    vp10_highbd_idct8_c(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vp10_highbd_idct8_c(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+void vp10_highbd_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int bd) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+
+void vp10_highbd_iadst4_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[0];
+  tran_low_t x1 = input[1];
+  tran_low_t x2 = input[2];
+  tran_low_t x3 = input[3];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3)) {
+    memset(output, 0, 4 * sizeof(*output));
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = (tran_high_t)(x0 - x2 + x3);
+
+  s0 = s0 + s3 + s5;
+  s1 = s1 - s4 - s6;
+  s3 = s2;
+  s2 = sinpi_3_9 * s7;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = WRAPLOW(highbd_dct_const_round_shift(s0 + s3, bd), bd);
+  output[1] = WRAPLOW(highbd_dct_const_round_shift(s1 + s3, bd), bd);
+  output[2] = WRAPLOW(highbd_dct_const_round_shift(s2, bd), bd);
+  output[3] = WRAPLOW(highbd_dct_const_round_shift(s0 + s1 - s3, bd), bd);
+}
+
+void vp10_highbd_iadst8_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[7];
+  tran_low_t x1 = input[0];
+  tran_low_t x2 = input[5];
+  tran_low_t x3 = input[2];
+  tran_low_t x4 = input[3];
+  tran_low_t x5 = input[4];
+  tran_low_t x6 = input[1];
+  tran_low_t x7 = input[6];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    memset(output, 0, 8 * sizeof(*output));
+    return;
+  }
+
+  // stage 1
+  s0 = cospi_2_64  * x0 + cospi_30_64 * x1;
+  s1 = cospi_30_64 * x0 - cospi_2_64  * x1;
+  s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
+  s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
+  s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
+  s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
+  s6 = cospi_26_64 * x6 + cospi_6_64  * x7;
+  s7 = cospi_6_64  * x6 - cospi_26_64 * x7;
+
+  x0 = WRAPLOW(highbd_dct_const_round_shift(s0 + s4, bd), bd);
+  x1 = WRAPLOW(highbd_dct_const_round_shift(s1 + s5, bd), bd);
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2 + s6, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3 + s7, bd), bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s0 - s4, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s1 - s5, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s2 - s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s3 - s7, bd), bd);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 =  cospi_8_64  * x4 + cospi_24_64 * x5;
+  s5 =  cospi_24_64 * x4 - cospi_8_64  * x5;
+  s6 = -cospi_24_64 * x6 + cospi_8_64  * x7;
+  s7 =  cospi_8_64  * x6 + cospi_24_64 * x7;
+
+  x0 = WRAPLOW(s0 + s2, bd);
+  x1 = WRAPLOW(s1 + s3, bd);
+  x2 = WRAPLOW(s0 - s2, bd);
+  x3 = WRAPLOW(s1 - s3, bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s4 + s6, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s5 + s7, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s4 - s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s5 - s7, bd), bd);
+
+  // stage 3
+  s2 = cospi_16_64 * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (x6 - x7);
+
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s7, bd), bd);
+
+  output[0] = WRAPLOW(x0, bd);
+  output[1] = WRAPLOW(-x4, bd);
+  output[2] = WRAPLOW(x6, bd);
+  output[3] = WRAPLOW(-x2, bd);
+  output[4] = WRAPLOW(x3, bd);
+  output[5] = WRAPLOW(-x7, bd);
+  output[6] = WRAPLOW(x5, bd);
+  output[7] = WRAPLOW(-x1, bd);
+}
+
+void vp10_highbd_idct8x8_10_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  // Only first 4 row has non-zero coefs.
+  for (i = 0; i < 4; ++i) {
+    vp10_highbd_idct8_c(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+  // Then transform columns.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vp10_highbd_idct8_c(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+void vp10_highbd_idct16_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step1[16], step2[16];
+  tran_high_t temp1, temp2;
+  (void) bd;
+
+  // stage 1
+  step1[0] = input[0/2];
+  step1[1] = input[16/2];
+  step1[2] = input[8/2];
+  step1[3] = input[24/2];
+  step1[4] = input[4/2];
+  step1[5] = input[20/2];
+  step1[6] = input[12/2];
+  step1[7] = input[28/2];
+  step1[8] = input[2/2];
+  step1[9] = input[18/2];
+  step1[10] = input[10/2];
+  step1[11] = input[26/2];
+  step1[12] = input[6/2];
+  step1[13] = input[22/2];
+  step1[14] = input[14/2];
+  step1[15] = input[30/2];
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[15] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[7] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], bd);
+  step1[9] = WRAPLOW(step2[8] - step2[9], bd);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], bd);
+  step1[11] = WRAPLOW(step2[10] + step2[11], bd);
+  step1[12] = WRAPLOW(step2[12] + step2[13], bd);
+  step1[13] = WRAPLOW(step2[12] - step2[13], bd);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], bd);
+  step1[15] = WRAPLOW(step2[14] + step2[15], bd);
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[1] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[3] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[4] = WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = WRAPLOW(step1[6] + step1[7], bd);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], bd);
+  step1[1] = WRAPLOW(step2[1] + step2[2], bd);
+  step1[2] = WRAPLOW(step2[1] - step2[2], bd);
+  step1[3] = WRAPLOW(step2[0] - step2[3], bd);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], bd);
+  step1[9] = WRAPLOW(step2[9] + step2[10], bd);
+  step1[10] = WRAPLOW(step2[9] - step2[10], bd);
+  step1[11] = WRAPLOW(step2[8] - step2[11], bd);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], bd);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], bd);
+  step1[14] = WRAPLOW(step2[13] + step2[14], bd);
+  step1[15] = WRAPLOW(step2[12] + step2[15], bd);
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], bd);
+  step2[1] = WRAPLOW(step1[1] + step1[6], bd);
+  step2[2] = WRAPLOW(step1[2] + step1[5], bd);
+  step2[3] = WRAPLOW(step1[3] + step1[4], bd);
+  step2[4] = WRAPLOW(step1[3] - step1[4], bd);
+  step2[5] = WRAPLOW(step1[2] - step1[5], bd);
+  step2[6] = WRAPLOW(step1[1] - step1[6], bd);
+  step2[7] = WRAPLOW(step1[0] - step1[7], bd);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  // stage 7
+  output[0] = WRAPLOW(step2[0] + step2[15], bd);
+  output[1] = WRAPLOW(step2[1] + step2[14], bd);
+  output[2] = WRAPLOW(step2[2] + step2[13], bd);
+  output[3] = WRAPLOW(step2[3] + step2[12], bd);
+  output[4] = WRAPLOW(step2[4] + step2[11], bd);
+  output[5] = WRAPLOW(step2[5] + step2[10], bd);
+  output[6] = WRAPLOW(step2[6] + step2[9], bd);
+  output[7] = WRAPLOW(step2[7] + step2[8], bd);
+  output[8] = WRAPLOW(step2[7] - step2[8], bd);
+  output[9] = WRAPLOW(step2[6] - step2[9], bd);
+  output[10] = WRAPLOW(step2[5] - step2[10], bd);
+  output[11] = WRAPLOW(step2[4] - step2[11], bd);
+  output[12] = WRAPLOW(step2[3] - step2[12], bd);
+  output[13] = WRAPLOW(step2[2] - step2[13], bd);
+  output[14] = WRAPLOW(step2[1] - step2[14], bd);
+  output[15] = WRAPLOW(step2[0] - step2[15], bd);
+}
+
+void vp10_highbd_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  for (i = 0; i < 16; ++i) {
+    vp10_highbd_idct16_c(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    vp10_highbd_idct16_c(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i],
+          ROUND_POWER_OF_TWO(temp_out[j], 6),
+          bd);
+    }
+  }
+}
+
+void vp10_highbd_iadst16_c(const tran_low_t *input,
+                           tran_low_t *output,
+                           int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_low_t x0 = input[15];
+  tran_low_t x1 = input[0];
+  tran_low_t x2 = input[13];
+  tran_low_t x3 = input[2];
+  tran_low_t x4 = input[11];
+  tran_low_t x5 = input[4];
+  tran_low_t x6 = input[9];
+  tran_low_t x7 = input[6];
+  tran_low_t x8 = input[7];
+  tran_low_t x9 = input[8];
+  tran_low_t x10 = input[5];
+  tran_low_t x11 = input[10];
+  tran_low_t x12 = input[3];
+  tran_low_t x13 = input[12];
+  tran_low_t x14 = input[1];
+  tran_low_t x15 = input[14];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8
+           | x9 | x10 | x11 | x12 | x13 | x14 | x15)) {
+    memset(output, 0, 16 * sizeof(*output));
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = WRAPLOW(highbd_dct_const_round_shift(s0 + s8, bd), bd);
+  x1 = WRAPLOW(highbd_dct_const_round_shift(s1 + s9, bd), bd);
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2 + s10, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3 + s11, bd), bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s4 + s12, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s5 + s13, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s6 + s14, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s7 + s15, bd), bd);
+  x8  = WRAPLOW(highbd_dct_const_round_shift(s0 - s8, bd), bd);
+  x9  = WRAPLOW(highbd_dct_const_round_shift(s1 - s9, bd), bd);
+  x10 = WRAPLOW(highbd_dct_const_round_shift(s2 - s10, bd), bd);
+  x11 = WRAPLOW(highbd_dct_const_round_shift(s3 - s11, bd), bd);
+  x12 = WRAPLOW(highbd_dct_const_round_shift(s4 - s12, bd), bd);
+  x13 = WRAPLOW(highbd_dct_const_round_shift(s5 - s13, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s6 - s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s7 - s15, bd), bd);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
+  s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
+  s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
+  s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = WRAPLOW(s0 + s4, bd);
+  x1 = WRAPLOW(s1 + s5, bd);
+  x2 = WRAPLOW(s2 + s6, bd);
+  x3 = WRAPLOW(s3 + s7, bd);
+  x4 = WRAPLOW(s0 - s4, bd);
+  x5 = WRAPLOW(s1 - s5, bd);
+  x6 = WRAPLOW(s2 - s6, bd);
+  x7 = WRAPLOW(s3 - s7, bd);
+  x8 = WRAPLOW(highbd_dct_const_round_shift(s8 + s12, bd), bd);
+  x9 = WRAPLOW(highbd_dct_const_round_shift(s9 + s13, bd), bd);
+  x10 = WRAPLOW(highbd_dct_const_round_shift(s10 + s14, bd), bd);
+  x11 = WRAPLOW(highbd_dct_const_round_shift(s11 + s15, bd), bd);
+  x12 = WRAPLOW(highbd_dct_const_round_shift(s8 - s12, bd), bd);
+  x13 = WRAPLOW(highbd_dct_const_round_shift(s9 - s13, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s10 - s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s11 - s15, bd), bd);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
+
+  x0 = WRAPLOW(s0 + s2, bd);
+  x1 = WRAPLOW(s1 + s3, bd);
+  x2 = WRAPLOW(s0 - s2, bd);
+  x3 = WRAPLOW(s1 - s3, bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s4 + s6, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s5 + s7, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s4 - s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s5 - s7, bd), bd);
+  x8 = WRAPLOW(s8 + s10, bd);
+  x9 = WRAPLOW(s9 + s11, bd);
+  x10 = WRAPLOW(s8 - s10, bd);
+  x11 = WRAPLOW(s9 - s11, bd);
+  x12 = WRAPLOW(highbd_dct_const_round_shift(s12 + s14, bd), bd);
+  x13 = WRAPLOW(highbd_dct_const_round_shift(s13 + s15, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s12 - s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s13 - s15, bd), bd);
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (-x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (-x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s7, bd), bd);
+  x10 = WRAPLOW(highbd_dct_const_round_shift(s10, bd), bd);
+  x11 = WRAPLOW(highbd_dct_const_round_shift(s11, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s15, bd), bd);
+
+  output[0] = WRAPLOW(x0, bd);
+  output[1] = WRAPLOW(-x8, bd);
+  output[2] = WRAPLOW(x12, bd);
+  output[3] = WRAPLOW(-x4, bd);
+  output[4] = WRAPLOW(x6, bd);
+  output[5] = WRAPLOW(x14, bd);
+  output[6] = WRAPLOW(x10, bd);
+  output[7] = WRAPLOW(x2, bd);
+  output[8] = WRAPLOW(x3, bd);
+  output[9] = WRAPLOW(x11, bd);
+  output[10] = WRAPLOW(x15, bd);
+  output[11] = WRAPLOW(x7, bd);
+  output[12] = WRAPLOW(x5, bd);
+  output[13] = WRAPLOW(-x13, bd);
+  output[14] = WRAPLOW(x9, bd);
+  output[15] = WRAPLOW(-x1, bd);
+}
+
+void vp10_highbd_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int bd) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  for (i = 0; i < 4; ++i) {
+    vp10_highbd_idct16_c(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j*16 + i];
+    vp10_highbd_idct16_c(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vp10_highbd_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int bd) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  for (j = 0; j < 16; ++j) {
+    for (i = 0; i < 16; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+
+static void highbd_idct32_c(const tran_low_t *input,
+                            tran_low_t *output, int bd) {
+  tran_low_t step1[32], step2[32];
+  tran_high_t temp1, temp2;
+  (void) bd;
+
+  // stage 1
+  step1[0] = input[0];
+  step1[1] = input[16];
+  step1[2] = input[8];
+  step1[3] = input[24];
+  step1[4] = input[4];
+  step1[5] = input[20];
+  step1[6] = input[12];
+  step1[7] = input[28];
+  step1[8] = input[2];
+  step1[9] = input[18];
+  step1[10] = input[10];
+  step1[11] = input[26];
+  step1[12] = input[6];
+  step1[13] = input[22];
+  step1[14] = input[14];
+  step1[15] = input[30];
+
+  temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64;
+  temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64;
+  step1[16] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[31] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64;
+  temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64;
+  step1[17] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[30] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64;
+  temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64;
+  step1[18] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[29] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64;
+  temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64;
+  step1[19] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[28] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64;
+  temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64;
+  step1[20] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[27] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64;
+  temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64;
+  temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64;
+  step1[22] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[25] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64;
+  temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64;
+  step1[23] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[24] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[15] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  step2[16] = WRAPLOW(step1[16] + step1[17], bd);
+  step2[17] = WRAPLOW(step1[16] - step1[17], bd);
+  step2[18] = WRAPLOW(-step1[18] + step1[19], bd);
+  step2[19] = WRAPLOW(step1[18] + step1[19], bd);
+  step2[20] = WRAPLOW(step1[20] + step1[21], bd);
+  step2[21] = WRAPLOW(step1[20] - step1[21], bd);
+  step2[22] = WRAPLOW(-step1[22] + step1[23], bd);
+  step2[23] = WRAPLOW(step1[22] + step1[23], bd);
+  step2[24] = WRAPLOW(step1[24] + step1[25], bd);
+  step2[25] = WRAPLOW(step1[24] - step1[25], bd);
+  step2[26] = WRAPLOW(-step1[26] + step1[27], bd);
+  step2[27] = WRAPLOW(step1[26] + step1[27], bd);
+  step2[28] = WRAPLOW(step1[28] + step1[29], bd);
+  step2[29] = WRAPLOW(step1[28] - step1[29], bd);
+  step2[30] = WRAPLOW(-step1[30] + step1[31], bd);
+  step2[31] = WRAPLOW(step1[30] + step1[31], bd);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[7] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], bd);
+  step1[9] = WRAPLOW(step2[8] - step2[9], bd);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], bd);
+  step1[11] = WRAPLOW(step2[10] + step2[11], bd);
+  step1[12] = WRAPLOW(step2[12] + step2[13], bd);
+  step1[13] = WRAPLOW(step2[12] - step2[13], bd);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], bd);
+  step1[15] = WRAPLOW(step2[14] + step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[31] = step2[31];
+  temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64;
+  temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64;
+  step1[17] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[30] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64;
+  temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64;
+  step1[18] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[29] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[19] = step2[19];
+  step1[20] = step2[20];
+  temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64;
+  temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64;
+  temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64;
+  step1[22] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[25] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[27] = step2[27];
+  step1[28] = step2[28];
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[1] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[3] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[4] = WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = WRAPLOW(step1[6] + step1[7], bd);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  step2[16] = WRAPLOW(step1[16] + step1[19], bd);
+  step2[17] = WRAPLOW(step1[17] + step1[18], bd);
+  step2[18] = WRAPLOW(step1[17] - step1[18], bd);
+  step2[19] = WRAPLOW(step1[16] - step1[19], bd);
+  step2[20] = WRAPLOW(-step1[20] + step1[23], bd);
+  step2[21] = WRAPLOW(-step1[21] + step1[22], bd);
+  step2[22] = WRAPLOW(step1[21] + step1[22], bd);
+  step2[23] = WRAPLOW(step1[20] + step1[23], bd);
+
+  step2[24] = WRAPLOW(step1[24] + step1[27], bd);
+  step2[25] = WRAPLOW(step1[25] + step1[26], bd);
+  step2[26] = WRAPLOW(step1[25] - step1[26], bd);
+  step2[27] = WRAPLOW(step1[24] - step1[27], bd);
+  step2[28] = WRAPLOW(-step1[28] + step1[31], bd);
+  step2[29] = WRAPLOW(-step1[29] + step1[30], bd);
+  step2[30] = WRAPLOW(step1[29] + step1[30], bd);
+  step2[31] = WRAPLOW(step1[28] + step1[31], bd);
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], bd);
+  step1[1] = WRAPLOW(step2[1] + step2[2], bd);
+  step1[2] = WRAPLOW(step2[1] - step2[2], bd);
+  step1[3] = WRAPLOW(step2[0] - step2[3], bd);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], bd);
+  step1[9] = WRAPLOW(step2[9] + step2[10], bd);
+  step1[10] = WRAPLOW(step2[9] - step2[10], bd);
+  step1[11] = WRAPLOW(step2[8] - step2[11], bd);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], bd);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], bd);
+  step1[14] = WRAPLOW(step2[13] + step2[14], bd);
+  step1[15] = WRAPLOW(step2[12] + step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64;
+  temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64;
+  step1[18] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[29] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64;
+  temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64;
+  step1[19] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[28] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64;
+  temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64;
+  step1[20] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[27] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64;
+  temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[22] = step2[22];
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[25] = step2[25];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], bd);
+  step2[1] = WRAPLOW(step1[1] + step1[6], bd);
+  step2[2] = WRAPLOW(step1[2] + step1[5], bd);
+  step2[3] = WRAPLOW(step1[3] + step1[4], bd);
+  step2[4] = WRAPLOW(step1[3] - step1[4], bd);
+  step2[5] = WRAPLOW(step1[2] - step1[5], bd);
+  step2[6] = WRAPLOW(step1[1] - step1[6], bd);
+  step2[7] = WRAPLOW(step1[0] - step1[7], bd);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  step2[16] = WRAPLOW(step1[16] + step1[23], bd);
+  step2[17] = WRAPLOW(step1[17] + step1[22], bd);
+  step2[18] = WRAPLOW(step1[18] + step1[21], bd);
+  step2[19] = WRAPLOW(step1[19] + step1[20], bd);
+  step2[20] = WRAPLOW(step1[19] - step1[20], bd);
+  step2[21] = WRAPLOW(step1[18] - step1[21], bd);
+  step2[22] = WRAPLOW(step1[17] - step1[22], bd);
+  step2[23] = WRAPLOW(step1[16] - step1[23], bd);
+
+  step2[24] = WRAPLOW(-step1[24] + step1[31], bd);
+  step2[25] = WRAPLOW(-step1[25] + step1[30], bd);
+  step2[26] = WRAPLOW(-step1[26] + step1[29], bd);
+  step2[27] = WRAPLOW(-step1[27] + step1[28], bd);
+  step2[28] = WRAPLOW(step1[27] + step1[28], bd);
+  step2[29] = WRAPLOW(step1[26] + step1[29], bd);
+  step2[30] = WRAPLOW(step1[25] + step1[30], bd);
+  step2[31] = WRAPLOW(step1[24] + step1[31], bd);
+
+  // stage 7
+  step1[0] = WRAPLOW(step2[0] + step2[15], bd);
+  step1[1] = WRAPLOW(step2[1] + step2[14], bd);
+  step1[2] = WRAPLOW(step2[2] + step2[13], bd);
+  step1[3] = WRAPLOW(step2[3] + step2[12], bd);
+  step1[4] = WRAPLOW(step2[4] + step2[11], bd);
+  step1[5] = WRAPLOW(step2[5] + step2[10], bd);
+  step1[6] = WRAPLOW(step2[6] + step2[9], bd);
+  step1[7] = WRAPLOW(step2[7] + step2[8], bd);
+  step1[8] = WRAPLOW(step2[7] - step2[8], bd);
+  step1[9] = WRAPLOW(step2[6] - step2[9], bd);
+  step1[10] = WRAPLOW(step2[5] - step2[10], bd);
+  step1[11] = WRAPLOW(step2[4] - step2[11], bd);
+  step1[12] = WRAPLOW(step2[3] - step2[12], bd);
+  step1[13] = WRAPLOW(step2[2] - step2[13], bd);
+  step1[14] = WRAPLOW(step2[1] - step2[14], bd);
+  step1[15] = WRAPLOW(step2[0] - step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  step1[18] = step2[18];
+  step1[19] = step2[19];
+  temp1 = (-step2[20] + step2[27]) * cospi_16_64;
+  temp2 = (step2[20] + step2[27]) * cospi_16_64;
+  step1[20] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[27] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step2[21] + step2[26]) * cospi_16_64;
+  temp2 = (step2[21] + step2[26]) * cospi_16_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step2[22] + step2[25]) * cospi_16_64;
+  temp2 = (step2[22] + step2[25]) * cospi_16_64;
+  step1[22] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[25] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step2[23] + step2[24]) * cospi_16_64;
+  temp2 = (step2[23] + step2[24]) * cospi_16_64;
+  step1[23] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[24] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[28] = step2[28];
+  step1[29] = step2[29];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // final stage
+  output[0] = WRAPLOW(step1[0] + step1[31], bd);
+  output[1] = WRAPLOW(step1[1] + step1[30], bd);
+  output[2] = WRAPLOW(step1[2] + step1[29], bd);
+  output[3] = WRAPLOW(step1[3] + step1[28], bd);
+  output[4] = WRAPLOW(step1[4] + step1[27], bd);
+  output[5] = WRAPLOW(step1[5] + step1[26], bd);
+  output[6] = WRAPLOW(step1[6] + step1[25], bd);
+  output[7] = WRAPLOW(step1[7] + step1[24], bd);
+  output[8] = WRAPLOW(step1[8] + step1[23], bd);
+  output[9] = WRAPLOW(step1[9] + step1[22], bd);
+  output[10] = WRAPLOW(step1[10] + step1[21], bd);
+  output[11] = WRAPLOW(step1[11] + step1[20], bd);
+  output[12] = WRAPLOW(step1[12] + step1[19], bd);
+  output[13] = WRAPLOW(step1[13] + step1[18], bd);
+  output[14] = WRAPLOW(step1[14] + step1[17], bd);
+  output[15] = WRAPLOW(step1[15] + step1[16], bd);
+  output[16] = WRAPLOW(step1[15] - step1[16], bd);
+  output[17] = WRAPLOW(step1[14] - step1[17], bd);
+  output[18] = WRAPLOW(step1[13] - step1[18], bd);
+  output[19] = WRAPLOW(step1[12] - step1[19], bd);
+  output[20] = WRAPLOW(step1[11] - step1[20], bd);
+  output[21] = WRAPLOW(step1[10] - step1[21], bd);
+  output[22] = WRAPLOW(step1[9] - step1[22], bd);
+  output[23] = WRAPLOW(step1[8] - step1[23], bd);
+  output[24] = WRAPLOW(step1[7] - step1[24], bd);
+  output[25] = WRAPLOW(step1[6] - step1[25], bd);
+  output[26] = WRAPLOW(step1[5] - step1[26], bd);
+  output[27] = WRAPLOW(step1[4] - step1[27], bd);
+  output[28] = WRAPLOW(step1[3] - step1[28], bd);
+  output[29] = WRAPLOW(step1[2] - step1[29], bd);
+  output[30] = WRAPLOW(step1[1] - step1[30], bd);
+  output[31] = WRAPLOW(step1[0] - step1[31], bd);
+}
+
+void vp10_highbd_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest8,
+                                     int stride, int bd) {
+  tran_low_t out[32 * 32];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_low_t zero_coeff[16];
+    for (j = 0; j < 16; ++j)
+      zero_coeff[j] = input[2 * j] | input[2 * j + 1];
+    for (j = 0; j < 8; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 4; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 2; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+
+    if (zero_coeff[0] | zero_coeff[1])
+      highbd_idct32_c(input, outptr, bd);
+    else
+      memset(outptr, 0, sizeof(tran_low_t) * 32);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    highbd_idct32_c(temp_in, temp_out, bd);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vp10_highbd_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int bd) {
+  tran_low_t out[32 * 32] = {0};
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  // Only upper-left 8x8 has non-zero coeff.
+  for (i = 0; i < 8; ++i) {
+    highbd_idct32_c(input, outptr, bd);
+    input += 32;
+    outptr += 32;
+  }
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    highbd_idct32_c(temp_in, temp_out, bd);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vp10_highbd_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int bd) {
+  int i, j;
+  int a1;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  for (j = 0; j < 32; ++j) {
+    for (i = 0; i < 32; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH

diff --git a/vp10/common/vp10_inv_txfm.h b/vp10/common/vp10_inv_txfm.h
new file mode 100644 (file)
index 0000000..52611ac
--- /dev/null
+++ b/vp10/common/vp10_inv_txfm.h
@@ -0,0 +1,122 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_INV_TXFM_H_
+#define VPX_DSP_INV_TXFM_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE tran_low_t check_range(tran_high_t input) {
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid VP9 input streams, intermediate stage coefficients should always
+  // stay within the range of a signed 16 bit integer. Coefficients can go out
+  // of this range for invalid/corrupt VP9 streams. However, strictly checking
+  // this range for every intermediate coefficient can burdensome for a decoder,
+  // therefore the following assertion is only enabled when configured with
+  // --enable-coefficient-range-checking.
+  assert(INT16_MIN <= input);
+  assert(input <= INT16_MAX);
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+  return (tran_low_t)input;
+}
+
+static INLINE tran_low_t dct_const_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return check_range(rv);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE tran_low_t highbd_check_range(tran_high_t input,
+                                            int bd) {
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid highbitdepth VP9 streams, intermediate stage coefficients will
+  // stay within the ranges:
+  // - 8 bit: signed 16 bit integer
+  // - 10 bit: signed 18 bit integer
+  // - 12 bit: signed 20 bit integer
+  const int32_t int_max = (1 << (7 + bd)) - 1;
+  const int32_t int_min = -int_max - 1;
+  assert(int_min <= input);
+  assert(input <= int_max);
+  (void) int_min;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+  (void) bd;
+  return (tran_low_t)input;
+}
+
+static INLINE tran_low_t highbd_dct_const_round_shift(tran_high_t input,
+                                                      int bd) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return highbd_check_range(rv, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_EMULATE_HARDWARE
+// When CONFIG_EMULATE_HARDWARE is 1 the transform performs a
+// non-normative method to handle overflows. A stream that causes
+// overflows  in the inverse transform is considered invalid in VP9,
+// and a hardware implementer is free to choose any reasonable
+// method to handle overflows. However to aid in hardware
+// verification they can use a specific implementation of the
+// WRAPLOW() macro below that is identical to their intended
+// hardware implementation (and also use configure options to trigger
+// the C-implementation of the transform).
+//
+// The particular WRAPLOW implementation below performs strict
+// overflow wrapping to match common hardware implementations.
+// bd of 8 uses trans_low with 16bits, need to remove 16bits
+// bd of 10 uses trans_low with 18bits, need to remove 14bits
+// bd of 12 uses trans_low with 20bits, need to remove 12bits
+// bd of x uses trans_low with 8+x bits, need to remove 24-x bits
+#define WRAPLOW(x, bd) ((((int32_t)(x)) << (24 - bd)) >> (24 - bd))
+#else
+#define WRAPLOW(x, bd) ((int32_t)(x))
+#endif  // CONFIG_EMULATE_HARDWARE
+
+void vp10_idct4_c(const tran_low_t *input, tran_low_t *output);
+void vp10_idct8_c(const tran_low_t *input, tran_low_t *output);
+void vp10_idct16_c(const tran_low_t *input, tran_low_t *output);
+void vp10_idct32_c(const tran_low_t *input, tran_low_t *output);
+void vp10_iadst4_c(const tran_low_t *input, tran_low_t *output);
+void vp10_iadst8_c(const tran_low_t *input, tran_low_t *output);
+void vp10_iadst16_c(const tran_low_t *input, tran_low_t *output);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp10_highbd_idct4_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vp10_highbd_idct8_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vp10_highbd_idct16_c(const tran_low_t *input, tran_low_t *output, int bd);
+
+void vp10_highbd_iadst4_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vp10_highbd_iadst8_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vp10_highbd_iadst16_c(const tran_low_t *input, tran_low_t *output, int bd);
+
+static INLINE uint16_t highbd_clip_pixel_add(uint16_t dest, tran_high_t trans,
+                                             int bd) {
+  trans = WRAPLOW(trans, bd);
+  return clip_pixel_highbd(WRAPLOW(dest + trans, bd), bd);
+}
+#endif
+
+static INLINE uint8_t clip_pixel_add(uint8_t dest, tran_high_t trans) {
+  trans = WRAPLOW(trans, 8);
+  return clip_pixel(WRAPLOW(dest + trans, 8));
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // VPX_DSP_INV_TXFM_H_

diff --git a/vp10/common/vp10_rtcd_defs.pl b/vp10/common/vp10_rtcd_defs.pl
index 37b9622..ca3729c 100644 (file)
--- a/vp10/common/vp10_rtcd_defs.pl
+++ b/vp10/common/vp10_rtcd_defs.pl
@@ -289,6 +289,188 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp10_fwht4x4 msa/, "$mmx_x86inc";
 }
 
+# Inverse transform
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  # Note as optimized versions of these functions are added we need to add a check to ensure
+  # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
+  add_proto qw/void vp10_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct4x4_1_add/;
+
+  add_proto qw/void vp10_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct4x4_16_add/;
+
+  add_proto qw/void vp10_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct8x8_1_add/;
+
+  add_proto qw/void vp10_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct8x8_64_add/;
+
+  add_proto qw/void vp10_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct8x8_12_add/;
+
+  add_proto qw/void vp10_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct16x16_1_add/;
+
+  add_proto qw/void vp10_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct16x16_256_add/;
+
+  add_proto qw/void vp10_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct16x16_10_add/;
+
+  add_proto qw/void vp10_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct32x32_1024_add/;
+
+  add_proto qw/void vp10_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct32x32_34_add/;
+
+  add_proto qw/void vp10_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct32x32_1_add/;
+
+  add_proto qw/void vp10_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_iwht4x4_1_add/;
+
+  add_proto qw/void vp10_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_iwht4x4_16_add/;
+
+  add_proto qw/void vp10_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct4x4_1_add/;
+
+  add_proto qw/void vp10_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct8x8_1_add/;
+
+  add_proto qw/void vp10_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct16x16_1_add/;
+
+  add_proto qw/void vp10_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct32x32_1024_add/;
+
+  add_proto qw/void vp10_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct32x32_34_add/;
+
+  add_proto qw/void vp10_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct32x32_1_add/;
+
+  add_proto qw/void vp10_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_iwht4x4_1_add/;
+
+  add_proto qw/void vp10_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_iwht4x4_16_add/;
+
+  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vp10_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct4x4_16_add/;
+
+    add_proto qw/void vp10_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_64_add/;
+
+    add_proto qw/void vp10_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_10_add/;
+
+    add_proto qw/void vp10_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_256_add/;
+
+    add_proto qw/void vp10_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_10_add/;
+  } else {
+    add_proto qw/void vp10_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct4x4_16_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_64_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_10_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_256_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_10_add sse2/;
+  }  # CONFIG_EMULATE_HARDWARE
+} else {
+  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vp10_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_1_add/;
+
+    add_proto qw/void vp10_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_16_add/;
+
+    add_proto qw/void vp10_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_1_add/;
+
+    add_proto qw/void vp10_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_64_add/;
+
+    add_proto qw/void vp10_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_12_add/;
+
+    add_proto qw/void vp10_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_1_add/;
+
+    add_proto qw/void vp10_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_256_add/;
+
+    add_proto qw/void vp10_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_10_add/;
+
+    add_proto qw/void vp10_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1024_add/;
+
+    add_proto qw/void vp10_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_34_add/;
+
+    add_proto qw/void vp10_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1_add/;
+
+    add_proto qw/void vp10_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_1_add/;
+
+    add_proto qw/void vp10_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_16_add/;
+  } else {
+    add_proto qw/void vp10_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_1_add sse2/;
+
+    add_proto qw/void vp10_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_16_add sse2/;
+
+    add_proto qw/void vp10_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_1_add sse2/;
+
+    add_proto qw/void vp10_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_64_add sse2/;
+
+    add_proto qw/void vp10_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_12_add sse2/;
+
+    add_proto qw/void vp10_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_1_add sse2/;
+
+    add_proto qw/void vp10_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_256_add sse2/;
+
+    add_proto qw/void vp10_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_10_add sse2/;
+
+    add_proto qw/void vp10_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1024_add sse2/;
+
+    add_proto qw/void vp10_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_34_add sse2/;
+
+    add_proto qw/void vp10_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1_add sse2/;
+
+    add_proto qw/void vp10_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_1_add/;
+
+    add_proto qw/void vp10_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_16_add/;
+  }  # CONFIG_EMULATE_HARDWARE
+}  # CONFIG_VP9_HIGHBITDEPTH
+
 #
 # Motion search
 #

diff --git a/vp10/common/x86/vp10_inv_txfm_sse2.c b/vp10/common/x86/vp10_inv_txfm_sse2.c
new file mode 100644 (file)
index 0000000..b25e22e
--- /dev/null
+++ b/vp10/common/x86/vp10_inv_txfm_sse2.c
@@ -0,0 +1,4058 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp10_rtcd.h"
+#include "vp10/common/x86/vp10_inv_txfm_sse2.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+
+#define RECON_AND_STORE4X4(dest, in_x) \
+{                                                     \
+  __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \
+  d0 = _mm_unpacklo_epi8(d0, zero); \
+  d0 = _mm_add_epi16(in_x, d0); \
+  d0 = _mm_packus_epi16(d0, d0); \
+  *(int *)(dest) = _mm_cvtsi128_si32(d0); \
+}
+
+void vp10_idct4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i cst = _mm_setr_epi16(
+      (int16_t)cospi_16_64, (int16_t)cospi_16_64, (int16_t)cospi_16_64,
+      (int16_t)-cospi_16_64, (int16_t)cospi_24_64, (int16_t)-cospi_8_64,
+      (int16_t)cospi_8_64, (int16_t)cospi_24_64);
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i input0, input1, input2, input3;
+
+  // Rows
+  input0 = _mm_load_si128((const __m128i *)input);
+  input2 = _mm_load_si128((const __m128i *)(input + 8));
+
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_shufflelo_epi16(input0, 0xd8);
+  input0 = _mm_shufflehi_epi16(input0, 0xd8);
+  input2 = _mm_shufflelo_epi16(input2, 0xd8);
+  input2 = _mm_shufflehi_epi16(input2, 0xd8);
+
+  input1 = _mm_unpackhi_epi32(input0, input0);
+  input0 = _mm_unpacklo_epi32(input0, input0);
+  input3 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpacklo_epi32(input2, input2);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input1);
+  input1 = _mm_packs_epi32(input2, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Columns
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_unpacklo_epi32(input2, input2);
+  input1 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpackhi_epi32(input3, input3);
+  input3 = _mm_unpacklo_epi32(input3, input3);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input2);
+  input1 = _mm_packs_epi32(input1, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Final round and shift
+  input2 = _mm_add_epi16(input2, eight);
+  input3 = _mm_add_epi16(input3, eight);
+
+  input2 = _mm_srai_epi16(input2, 4);
+  input3 = _mm_srai_epi16(input3, 4);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    d0 = _mm_unpacklo_epi32(d0,
+                            _mm_cvtsi32_si128(*(const int *)(dest + stride)));
+    d2 = _mm_unpacklo_epi32(
+        _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)), d2);
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d0 = _mm_add_epi16(d0, input2);
+    d2 = _mm_add_epi16(d2, input3);
+    d0 = _mm_packus_epi16(d0, d2);
+    // store input0
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    // store input1
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    // store input2
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+    // store input3
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+void vp10_idct4x4_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 4);
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE4X4(dest + 0 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 1 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 2 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 3 * stride, dc_value);
+}
+
+static INLINE void transpose_4x4(__m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
+
+  res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1);
+  res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1);
+}
+
+void vp10_idct4_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+
+  transpose_4x4(in);
+  // stage 1
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[1]);
+  u[1] = _mm_packs_epi32(v[3], v[2]);
+
+  // stage 2
+  in[0] = _mm_add_epi16(u[0], u[1]);
+  in[1] = _mm_sub_epi16(u[0], u[1]);
+  in[1] = _mm_shuffle_epi32(in[1], 0x4E);
+}
+
+void vp10_iadst4_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
+  const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8], in7;
+
+  transpose_4x4(in);
+  in7 = _mm_srli_si128(in[1], 8);
+  in7 = _mm_add_epi16(in7, in[0]);
+  in7 = _mm_sub_epi16(in7, in[1]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpackhi_epi16(in[0], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04);  // s0 + s3
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02);  // s2 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x2
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01);  // s1 - s4
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04);  // s2 - s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s2
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_add_epi32(v[3], v[4]);
+  u[2] = v[2];
+  u[3] = _mm_add_epi32(u[0], u[1]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_add_epi32(u[3], v[5]);
+  u[6] = _mm_sub_epi32(u[5], u[4]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[1] = _mm_packs_epi32(u[2], u[3]);
+}
+
+#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, \
+                      out0, out1, out2, out3, out4, out5, out6, out7) \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+    const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \
+    const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \
+    const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \
+    const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5); \
+    const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \
+                                                        \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); \
+                                                            \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
+    out4 = _mm_unpacklo_epi64(tr1_1, tr1_5); \
+    out5 = _mm_unpackhi_epi64(tr1_1, tr1_5); \
+    out6 = _mm_unpacklo_epi64(tr1_3, tr1_7); \
+    out7 = _mm_unpackhi_epi64(tr1_3, tr1_7); \
+  }
+
+#define TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, \
+                         out0, out1, out2, out3) \
+  {                                              \
+    const __m128i tr0_0 = _mm_unpackhi_epi16(tmp0, tmp1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(tmp1, tmp0); \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(tmp2, tmp3); \
+    const __m128i tr0_5 = _mm_unpackhi_epi16(tmp3, tmp2); \
+    \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
+    \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
+  }
+
+#define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \
+  {                                            \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+    out0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    out1 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+  }
+
+// Define Macro for multiplying elements by constants and adding them together.
+#define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, \
+                               cst0, cst1, cst2, cst3, res0, res1, res2, res3) \
+  {   \
+      tmp0 = _mm_madd_epi16(lo_0, cst0); \
+      tmp1 = _mm_madd_epi16(hi_0, cst0); \
+      tmp2 = _mm_madd_epi16(lo_0, cst1); \
+      tmp3 = _mm_madd_epi16(hi_0, cst1); \
+      tmp4 = _mm_madd_epi16(lo_1, cst2); \
+      tmp5 = _mm_madd_epi16(hi_1, cst2); \
+      tmp6 = _mm_madd_epi16(lo_1, cst3); \
+      tmp7 = _mm_madd_epi16(hi_1, cst3); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      tmp4 = _mm_add_epi32(tmp4, rounding); \
+      tmp5 = _mm_add_epi32(tmp5, rounding); \
+      tmp6 = _mm_add_epi32(tmp6, rounding); \
+      tmp7 = _mm_add_epi32(tmp7, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); \
+      tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); \
+      tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); \
+      tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); \
+      \
+      res0 = _mm_packs_epi32(tmp0, tmp1); \
+      res1 = _mm_packs_epi32(tmp2, tmp3); \
+      res2 = _mm_packs_epi32(tmp4, tmp5); \
+      res3 = _mm_packs_epi32(tmp6, tmp7); \
+  }
+
+#define MULTIPLICATION_AND_ADD_2(lo_0, hi_0, cst0, cst1, res0, res1) \
+  {   \
+      tmp0 = _mm_madd_epi16(lo_0, cst0); \
+      tmp1 = _mm_madd_epi16(hi_0, cst0); \
+      tmp2 = _mm_madd_epi16(lo_0, cst1); \
+      tmp3 = _mm_madd_epi16(hi_0, cst1); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      \
+      res0 = _mm_packs_epi32(tmp0, tmp1); \
+      res1 = _mm_packs_epi32(tmp2, tmp3); \
+  }
+
+#define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, \
+              out0, out1, out2, out3, out4, out5, out6, out7)  \
+  { \
+  /* Stage1 */      \
+  { \
+    const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \
+    const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \
+    const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); \
+    const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); \
+    \
+    MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, \
+                          stg1_1, stg1_2, stg1_3, stp1_4,      \
+                          stp1_7, stp1_5, stp1_6)              \
+  } \
+    \
+  /* Stage2 */ \
+  { \
+    const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4); \
+    const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4); \
+    const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); \
+    const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); \
+    \
+    MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, \
+                           stg2_1, stg2_2, stg2_3, stp2_0,     \
+                           stp2_1, stp2_2, stp2_3)             \
+    \
+    stp2_4 = _mm_adds_epi16(stp1_4, stp1_5); \
+    stp2_5 = _mm_subs_epi16(stp1_4, stp1_5); \
+    stp2_6 = _mm_subs_epi16(stp1_7, stp1_6); \
+    stp2_7 = _mm_adds_epi16(stp1_7, stp1_6); \
+  } \
+    \
+  /* Stage3 */ \
+  { \
+    const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+    const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+    \
+    stp1_0 = _mm_adds_epi16(stp2_0, stp2_3); \
+    stp1_1 = _mm_adds_epi16(stp2_1, stp2_2); \
+    stp1_2 = _mm_subs_epi16(stp2_1, stp2_2); \
+    stp1_3 = _mm_subs_epi16(stp2_0, stp2_3); \
+    \
+    tmp0 = _mm_madd_epi16(lo_56, stg2_1); \
+    tmp1 = _mm_madd_epi16(hi_56, stg2_1); \
+    tmp2 = _mm_madd_epi16(lo_56, stg2_0); \
+    tmp3 = _mm_madd_epi16(hi_56, stg2_0); \
+    \
+    tmp0 = _mm_add_epi32(tmp0, rounding); \
+    tmp1 = _mm_add_epi32(tmp1, rounding); \
+    tmp2 = _mm_add_epi32(tmp2, rounding); \
+    tmp3 = _mm_add_epi32(tmp3, rounding); \
+    \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+    \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  } \
+  \
+  /* Stage4  */ \
+  out0 = _mm_adds_epi16(stp1_0, stp2_7); \
+  out1 = _mm_adds_epi16(stp1_1, stp1_6); \
+  out2 = _mm_adds_epi16(stp1_2, stp1_5); \
+  out3 = _mm_adds_epi16(stp1_3, stp2_4); \
+  out4 = _mm_subs_epi16(stp1_3, stp2_4); \
+  out5 = _mm_subs_epi16(stp1_2, stp1_5); \
+  out6 = _mm_subs_epi16(stp1_1, stp1_6); \
+  out7 = _mm_subs_epi16(stp1_0, stp2_7); \
+  }
+
+void vp10_idct8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data.
+  in0 = _mm_load_si128((const __m128i *)input);
+  in1 = _mm_load_si128((const __m128i *)(input + 8 * 1));
+  in2 = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in3 = _mm_load_si128((const __m128i *)(input + 8 * 3));
+  in4 = _mm_load_si128((const __m128i *)(input + 8 * 4));
+  in5 = _mm_load_si128((const __m128i *)(input + 8 * 5));
+  in6 = _mm_load_si128((const __m128i *)(input + 8 * 6));
+  in7 = _mm_load_si128((const __m128i *)(input + 8 * 7));
+
+  // 2-D
+  for (i = 0; i < 2; i++) {
+    // 8x8 Transpose is copied from vp10_fdct8x8_sse2()
+    TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7,
+                  in0, in1, in2, in3, in4, in5, in6, in7);
+
+    // 4-stage 1D vp10_idct8x8
+    IDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+          in0, in1, in2, in3, in4, in5, in6, in7);
+  }
+
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+void vp10_idct8x8_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 5);
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE(dest + 0 * stride, dc_value);
+  RECON_AND_STORE(dest + 1 * stride, dc_value);
+  RECON_AND_STORE(dest + 2 * stride, dc_value);
+  RECON_AND_STORE(dest + 3 * stride, dc_value);
+  RECON_AND_STORE(dest + 4 * stride, dc_value);
+  RECON_AND_STORE(dest + 5 * stride, dc_value);
+  RECON_AND_STORE(dest + 6 * stride, dc_value);
+  RECON_AND_STORE(dest + 7 * stride, dc_value);
+}
+
+void vp10_idct8_sse2(__m128i *in) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // 8x8 Transpose is copied from vp10_fdct8x8_sse2()
+  TRANSPOSE_8X8(in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7],
+                in0, in1, in2, in3, in4, in5, in6, in7);
+
+  // 4-stage 1D vp10_idct8x8
+  IDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+        in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7]);
+}
+
+void vp10_iadst8_sse2(__m128i *in) {
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__const_0 = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+  // transpose
+  array_transpose_8x8(in, in);
+
+  // properly aligned for butterfly input
+  in0 = in[7];
+  in1 = in[0];
+  in2 = in[5];
+  in3 = in[2];
+  in4 = in[3];
+  in5 = in[4];
+  in6 = in[1];
+  in7 = in[6];
+
+  // column transformation
+  // stage 1
+  // interleave and multiply/add into 32-bit integer
+  s0 = _mm_unpacklo_epi16(in0, in1);
+  s1 = _mm_unpackhi_epi16(in0, in1);
+  s2 = _mm_unpacklo_epi16(in2, in3);
+  s3 = _mm_unpackhi_epi16(in2, in3);
+  s4 = _mm_unpacklo_epi16(in4, in5);
+  s5 = _mm_unpackhi_epi16(in4, in5);
+  s6 = _mm_unpacklo_epi16(in6, in7);
+  s7 = _mm_unpackhi_epi16(in6, in7);
+
+  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
+  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
+  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
+  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
+  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
+  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
+  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
+  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
+  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
+  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
+  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
+  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
+  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
+  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
+  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
+  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
+
+  // addition
+  w0 = _mm_add_epi32(u0, u8);
+  w1 = _mm_add_epi32(u1, u9);
+  w2 = _mm_add_epi32(u2, u10);
+  w3 = _mm_add_epi32(u3, u11);
+  w4 = _mm_add_epi32(u4, u12);
+  w5 = _mm_add_epi32(u5, u13);
+  w6 = _mm_add_epi32(u6, u14);
+  w7 = _mm_add_epi32(u7, u15);
+  w8 = _mm_sub_epi32(u0, u8);
+  w9 = _mm_sub_epi32(u1, u9);
+  w10 = _mm_sub_epi32(u2, u10);
+  w11 = _mm_sub_epi32(u3, u11);
+  w12 = _mm_sub_epi32(u4, u12);
+  w13 = _mm_sub_epi32(u5, u13);
+  w14 = _mm_sub_epi32(u6, u14);
+  w15 = _mm_sub_epi32(u7, u15);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
+  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
+  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
+  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
+  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
+  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
+  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
+  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
+  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
+  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
+  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
+  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
+  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
+  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
+  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
+
+  // back to 16-bit and pack 8 integers into __m128i
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[1] = _mm_packs_epi32(u2, u3);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[3] = _mm_packs_epi32(u6, u7);
+  in[4] = _mm_packs_epi32(u8, u9);
+  in[5] = _mm_packs_epi32(u10, u11);
+  in[6] = _mm_packs_epi32(u12, u13);
+  in[7] = _mm_packs_epi32(u14, u15);
+
+  // stage 2
+  s0 = _mm_add_epi16(in[0], in[2]);
+  s1 = _mm_add_epi16(in[1], in[3]);
+  s2 = _mm_sub_epi16(in[0], in[2]);
+  s3 = _mm_sub_epi16(in[1], in[3]);
+  u0 = _mm_unpacklo_epi16(in[4], in[5]);
+  u1 = _mm_unpackhi_epi16(in[4], in[5]);
+  u2 = _mm_unpacklo_epi16(in[6], in[7]);
+  u3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
+  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
+  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
+  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
+  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
+  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
+  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
+  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
+
+  w0 = _mm_add_epi32(v0, v4);
+  w1 = _mm_add_epi32(v1, v5);
+  w2 = _mm_add_epi32(v2, v6);
+  w3 = _mm_add_epi32(v3, v7);
+  w4 = _mm_sub_epi32(v0, v4);
+  w5 = _mm_sub_epi32(v1, v5);
+  w6 = _mm_sub_epi32(v2, v6);
+  w7 = _mm_sub_epi32(v3, v7);
+
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  // back to 16-bit intergers
+  s4 = _mm_packs_epi32(u0, u1);
+  s5 = _mm_packs_epi32(u2, u3);
+  s6 = _mm_packs_epi32(u4, u5);
+  s7 = _mm_packs_epi32(u6, u7);
+
+  // stage 3
+  u0 = _mm_unpacklo_epi16(s2, s3);
+  u1 = _mm_unpackhi_epi16(s2, s3);
+  u2 = _mm_unpacklo_epi16(s6, s7);
+  u3 = _mm_unpackhi_epi16(s6, s7);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
+  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
+  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
+  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
+
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  s2 = _mm_packs_epi32(v0, v1);
+  s3 = _mm_packs_epi32(v2, v3);
+  s6 = _mm_packs_epi32(v4, v5);
+  s7 = _mm_packs_epi32(v6, v7);
+
+  in[0] = s0;
+  in[1] = _mm_sub_epi16(k__const_0, s4);
+  in[2] = s6;
+  in[3] = _mm_sub_epi16(k__const_0, s2);
+  in[4] = s3;
+  in[5] = _mm_sub_epi16(k__const_0, s7);
+  in[6] = s5;
+  in[7] = _mm_sub_epi16(k__const_0, s1);
+}
+
+void vp10_idct8x8_12_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // Rows. Load 4-row input data.
+  in0 = _mm_load_si128((const __m128i *)input);
+  in1 = _mm_load_si128((const __m128i *)(input + 8 * 1));
+  in2 = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in3 = _mm_load_si128((const __m128i *)(input + 8 * 3));
+
+  // 8x4 Transpose
+  TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1);
+  // Stage1
+  {
+    const __m128i lo_17 = _mm_unpackhi_epi16(in0, zero);
+    const __m128i lo_35 = _mm_unpackhi_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_17, stg1_0);
+    tmp2 = _mm_madd_epi16(lo_17, stg1_1);
+    tmp4 = _mm_madd_epi16(lo_35, stg1_2);
+    tmp6 = _mm_madd_epi16(lo_35, stg1_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+    stp1_5 = _mm_packs_epi32(tmp4, tmp6);
+  }
+
+  // Stage2
+  {
+    const __m128i lo_04 = _mm_unpacklo_epi16(in0, zero);
+    const __m128i lo_26 = _mm_unpacklo_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_04, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_04, stg2_1);
+    tmp4 = _mm_madd_epi16(lo_26, stg2_2);
+    tmp6 = _mm_madd_epi16(lo_26, stg2_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp2_0 = _mm_packs_epi32(tmp0, tmp2);
+    stp2_2 = _mm_packs_epi32(tmp6, tmp4);
+
+    tmp0 = _mm_adds_epi16(stp1_4, stp1_5);
+    tmp1 = _mm_subs_epi16(stp1_4, stp1_5);
+
+    stp2_4 = tmp0;
+    stp2_5 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_6 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6);
+
+    tmp4 = _mm_adds_epi16(stp2_0, stp2_2);
+    tmp6 = _mm_subs_epi16(stp2_0, stp2_2);
+
+    stp1_2 = _mm_unpackhi_epi64(tmp6, tmp4);
+    stp1_3 = _mm_unpacklo_epi64(tmp6, tmp4);
+
+    tmp0 = _mm_madd_epi16(lo_56, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_56, stg2_0);  // stg3_1 = stg2_0
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_5 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  tmp0 = _mm_adds_epi16(stp1_3, stp2_4);
+  tmp1 = _mm_adds_epi16(stp1_2, stp1_5);
+  tmp2 = _mm_subs_epi16(stp1_3, stp2_4);
+  tmp3 = _mm_subs_epi16(stp1_2, stp1_5);
+
+  TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3)
+
+  IDCT8(in0, in1, in2, in3, zero, zero, zero, zero,
+        in0, in1, in2, in3, in4, in5, in6, in7);
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+#define IDCT16 \
+  /* Stage2 */ \
+  { \
+    const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], in[15]); \
+    const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], in[15]); \
+    const __m128i lo_9_7 = _mm_unpacklo_epi16(in[9], in[7]);   \
+    const __m128i hi_9_7 = _mm_unpackhi_epi16(in[9], in[7]);   \
+    const __m128i lo_5_11 = _mm_unpacklo_epi16(in[5], in[11]); \
+    const __m128i hi_5_11 = _mm_unpackhi_epi16(in[5], in[11]); \
+    const __m128i lo_13_3 = _mm_unpacklo_epi16(in[13], in[3]); \
+    const __m128i hi_13_3 = _mm_unpackhi_epi16(in[13], in[3]); \
+    \
+    MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, \
+                           stg2_0, stg2_1, stg2_2, stg2_3, \
+                           stp2_8, stp2_15, stp2_9, stp2_14) \
+    \
+    MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, \
+                           stg2_4, stg2_5, stg2_6, stg2_7, \
+                           stp2_10, stp2_13, stp2_11, stp2_12) \
+  } \
+    \
+  /* Stage3 */ \
+  { \
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], in[14]); \
+    const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], in[14]); \
+    const __m128i lo_10_6 = _mm_unpacklo_epi16(in[10], in[6]); \
+    const __m128i hi_10_6 = _mm_unpackhi_epi16(in[10], in[6]); \
+    \
+    MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, \
+                           stg3_0, stg3_1, stg3_2, stg3_3, \
+                           stp1_4, stp1_7, stp1_5, stp1_6) \
+    \
+    stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9);  \
+    stp1_9 = _mm_sub_epi16(stp2_8, stp2_9);    \
+    stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
+    stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
+    \
+    stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); \
+    stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
+    stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
+    stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
+  } \
+  \
+  /* Stage4 */ \
+  { \
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], in[8]); \
+    const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], in[8]); \
+    const __m128i lo_4_12 = _mm_unpacklo_epi16(in[4], in[12]); \
+    const __m128i hi_4_12 = _mm_unpackhi_epi16(in[4], in[12]); \
+    \
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+    \
+    MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, \
+                           stg4_0, stg4_1, stg4_2, stg4_3, \
+                           stp2_0, stp2_1, stp2_2, stp2_3) \
+    \
+    stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
+    stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
+    stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
+    stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
+    \
+    MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, \
+                           stg4_4, stg4_5, stg4_6, stg4_7, \
+                           stp2_9, stp2_14, stp2_10, stp2_13) \
+  } \
+    \
+  /* Stage5 */ \
+  { \
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+    const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+    \
+    stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
+    stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
+    stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
+    stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
+    \
+    tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+    tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+    tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+    tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+    \
+    tmp0 = _mm_add_epi32(tmp0, rounding); \
+    tmp1 = _mm_add_epi32(tmp1, rounding); \
+    tmp2 = _mm_add_epi32(tmp2, rounding); \
+    tmp3 = _mm_add_epi32(tmp3, rounding); \
+    \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+    \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+    \
+    stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);  \
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);    \
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);   \
+    stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
+    \
+    stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);   \
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);   \
+    stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
+  } \
+    \
+  /* Stage6 */ \
+  { \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+    const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+    \
+    stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
+    stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+    stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
+    stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+    stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+    stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
+    \
+    MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                           stg6_0, stg4_0, stg6_0, stg4_0, \
+                           stp2_10, stp2_13, stp2_11, stp2_12) \
+  }
+
+#define IDCT16_10 \
+    /* Stage2 */ \
+    { \
+      const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], zero); \
+      const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], zero); \
+      const __m128i lo_13_3 = _mm_unpacklo_epi16(zero, in[3]); \
+      const __m128i hi_13_3 = _mm_unpackhi_epi16(zero, in[3]); \
+      \
+      MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_13_3, hi_13_3, \
+                             stg2_0, stg2_1, stg2_6, stg2_7, \
+                             stp1_8_0, stp1_15, stp1_11, stp1_12_0) \
+    } \
+      \
+    /* Stage3 */ \
+    { \
+      const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], zero); \
+      const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], zero); \
+      \
+      MULTIPLICATION_AND_ADD_2(lo_2_14, hi_2_14, \
+                               stg3_0, stg3_1,  \
+                               stp2_4, stp2_7) \
+      \
+      stp1_9  =  stp1_8_0; \
+      stp1_10 =  stp1_11;  \
+      \
+      stp1_13 = stp1_12_0; \
+      stp1_14 = stp1_15;   \
+    } \
+    \
+    /* Stage4 */ \
+    { \
+      const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero); \
+      const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], zero); \
+      \
+      const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+      const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+      const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+      const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+      \
+      MULTIPLICATION_AND_ADD_2(lo_0_8, hi_0_8, \
+                               stg4_0, stg4_1, \
+                               stp1_0, stp1_1) \
+      stp2_5 = stp2_4; \
+      stp2_6 = stp2_7; \
+      \
+      MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, \
+                             stg4_4, stg4_5, stg4_6, stg4_7, \
+                             stp2_9, stp2_14, stp2_10, stp2_13) \
+    } \
+      \
+    /* Stage5 */ \
+    { \
+      const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+      const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+      \
+      stp1_2 = stp1_1; \
+      stp1_3 = stp1_0; \
+      \
+      tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+      tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+      tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+      tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      \
+      stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+      stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+      \
+      stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);  \
+      stp1_9 = _mm_add_epi16(stp2_9, stp2_10);    \
+      stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);   \
+      stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
+      \
+      stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
+      stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);   \
+      stp1_14 = _mm_add_epi16(stp2_14, stp2_13);   \
+      stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
+    } \
+      \
+    /* Stage6 */ \
+    { \
+      const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+      const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+      const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+      const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+      \
+      stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
+      stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+      stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+      stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
+      stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
+      stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+      stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+      stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
+      \
+      MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                             stg6_0, stg4_0, stg6_0, stg4_0, \
+                             stp2_10, stp2_13, stp2_11, stp2_12) \
+    }
+
+void vp10_idct16x16_256_add_sse2(const int16_t *input, uint8_t *dest,
+                                int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[16], l[16], r[16], *curr1;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  curr1 = l;
+  for (i = 0; i < 2; i++) {
+    // 1-D vp10_idct
+
+    // Load input data.
+    in[0] = _mm_load_si128((const __m128i *)input);
+    in[8] = _mm_load_si128((const __m128i *)(input + 8 * 1));
+    in[1] = _mm_load_si128((const __m128i *)(input + 8 * 2));
+    in[9] = _mm_load_si128((const __m128i *)(input + 8 * 3));
+    in[2] = _mm_load_si128((const __m128i *)(input + 8 * 4));
+    in[10] = _mm_load_si128((const __m128i *)(input + 8 * 5));
+    in[3] = _mm_load_si128((const __m128i *)(input + 8 * 6));
+    in[11] = _mm_load_si128((const __m128i *)(input + 8 * 7));
+    in[4] = _mm_load_si128((const __m128i *)(input + 8 * 8));
+    in[12] = _mm_load_si128((const __m128i *)(input + 8 * 9));
+    in[5] = _mm_load_si128((const __m128i *)(input + 8 * 10));
+    in[13] = _mm_load_si128((const __m128i *)(input + 8 * 11));
+    in[6] = _mm_load_si128((const __m128i *)(input + 8 * 12));
+    in[14] = _mm_load_si128((const __m128i *)(input + 8 * 13));
+    in[7] = _mm_load_si128((const __m128i *)(input + 8 * 14));
+    in[15] = _mm_load_si128((const __m128i *)(input + 8 * 15));
+
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+
+    IDCT16
+
+    // Stage7
+    curr1[0] = _mm_add_epi16(stp2_0, stp1_15);
+    curr1[1] = _mm_add_epi16(stp2_1, stp1_14);
+    curr1[2] = _mm_add_epi16(stp2_2, stp2_13);
+    curr1[3] = _mm_add_epi16(stp2_3, stp2_12);
+    curr1[4] = _mm_add_epi16(stp2_4, stp2_11);
+    curr1[5] = _mm_add_epi16(stp2_5, stp2_10);
+    curr1[6] = _mm_add_epi16(stp2_6, stp1_9);
+    curr1[7] = _mm_add_epi16(stp2_7, stp1_8);
+    curr1[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    curr1[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    curr1[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    curr1[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    curr1[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    curr1[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    curr1[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    curr1[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    curr1 = r;
+    input += 128;
+  }
+  for (i = 0; i < 2; i++) {
+    int j;
+    // 1-D vp10_idct
+    array_transpose_8x8(l + i * 8, in);
+    array_transpose_8x8(r + i * 8, in + 8);
+
+    IDCT16
+
+    // 2-D
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vp10_idct16x16_1_add_sse2(const int16_t *input,
+                               uint8_t *dest,
+                               int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, i;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (i = 0; i < 2; ++i) {
+    RECON_AND_STORE(dest +  0 * stride, dc_value);
+    RECON_AND_STORE(dest +  1 * stride, dc_value);
+    RECON_AND_STORE(dest +  2 * stride, dc_value);
+    RECON_AND_STORE(dest +  3 * stride, dc_value);
+    RECON_AND_STORE(dest +  4 * stride, dc_value);
+    RECON_AND_STORE(dest +  5 * stride, dc_value);
+    RECON_AND_STORE(dest +  6 * stride, dc_value);
+    RECON_AND_STORE(dest +  7 * stride, dc_value);
+    RECON_AND_STORE(dest +  8 * stride, dc_value);
+    RECON_AND_STORE(dest +  9 * stride, dc_value);
+    RECON_AND_STORE(dest + 10 * stride, dc_value);
+    RECON_AND_STORE(dest + 11 * stride, dc_value);
+    RECON_AND_STORE(dest + 12 * stride, dc_value);
+    RECON_AND_STORE(dest + 13 * stride, dc_value);
+    RECON_AND_STORE(dest + 14 * stride, dc_value);
+    RECON_AND_STORE(dest + 15 * stride, dc_value);
+    dest += 8;
+  }
+}
+
+static void vp10_iadst16_8col(__m128i *in) {
+  // perform 16x16 1-D ADST for 8 columns
+  __m128i s[16], x[16], u[32], v[32];
+  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+
+  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
+  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
+  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
+  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
+  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
+  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
+  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
+  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
+  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
+  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
+  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
+  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
+  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
+  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
+  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
+  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
+  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
+  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
+  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
+  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
+  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
+  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
+  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
+  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
+  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
+  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
+  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
+  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
+  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
+  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
+  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
+  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
+
+  u[0] = _mm_add_epi32(v[0], v[16]);
+  u[1] = _mm_add_epi32(v[1], v[17]);
+  u[2] = _mm_add_epi32(v[2], v[18]);
+  u[3] = _mm_add_epi32(v[3], v[19]);
+  u[4] = _mm_add_epi32(v[4], v[20]);
+  u[5] = _mm_add_epi32(v[5], v[21]);
+  u[6] = _mm_add_epi32(v[6], v[22]);
+  u[7] = _mm_add_epi32(v[7], v[23]);
+  u[8] = _mm_add_epi32(v[8], v[24]);
+  u[9] = _mm_add_epi32(v[9], v[25]);
+  u[10] = _mm_add_epi32(v[10], v[26]);
+  u[11] = _mm_add_epi32(v[11], v[27]);
+  u[12] = _mm_add_epi32(v[12], v[28]);
+  u[13] = _mm_add_epi32(v[13], v[29]);
+  u[14] = _mm_add_epi32(v[14], v[30]);
+  u[15] = _mm_add_epi32(v[15], v[31]);
+  u[16] = _mm_sub_epi32(v[0], v[16]);
+  u[17] = _mm_sub_epi32(v[1], v[17]);
+  u[18] = _mm_sub_epi32(v[2], v[18]);
+  u[19] = _mm_sub_epi32(v[3], v[19]);
+  u[20] = _mm_sub_epi32(v[4], v[20]);
+  u[21] = _mm_sub_epi32(v[5], v[21]);
+  u[22] = _mm_sub_epi32(v[6], v[22]);
+  u[23] = _mm_sub_epi32(v[7], v[23]);
+  u[24] = _mm_sub_epi32(v[8], v[24]);
+  u[25] = _mm_sub_epi32(v[9], v[25]);
+  u[26] = _mm_sub_epi32(v[10], v[26]);
+  u[27] = _mm_sub_epi32(v[11], v[27]);
+  u[28] = _mm_sub_epi32(v[12], v[28]);
+  u[29] = _mm_sub_epi32(v[13], v[29]);
+  u[30] = _mm_sub_epi32(v[14], v[30]);
+  u[31] = _mm_sub_epi32(v[15], v[31]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
+  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
+  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
+  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
+  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
+  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
+  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
+  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
+  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
+  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
+  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
+  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
+  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
+  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
+  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
+  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_packs_epi32(u[8], u[9]);
+  s[5] = _mm_packs_epi32(u[10], u[11]);
+  s[6] = _mm_packs_epi32(u[12], u[13]);
+  s[7] = _mm_packs_epi32(u[14], u[15]);
+  s[8] = _mm_packs_epi32(u[16], u[17]);
+  s[9] = _mm_packs_epi32(u[18], u[19]);
+  s[10] = _mm_packs_epi32(u[20], u[21]);
+  s[11] = _mm_packs_epi32(u[22], u[23]);
+  s[12] = _mm_packs_epi32(u[24], u[25]);
+  s[13] = _mm_packs_epi32(u[26], u[27]);
+  s[14] = _mm_packs_epi32(u[28], u[29]);
+  s[15] = _mm_packs_epi32(u[30], u[31]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
+  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], v[8]);
+  u[1] = _mm_add_epi32(v[1], v[9]);
+  u[2] = _mm_add_epi32(v[2], v[10]);
+  u[3] = _mm_add_epi32(v[3], v[11]);
+  u[4] = _mm_add_epi32(v[4], v[12]);
+  u[5] = _mm_add_epi32(v[5], v[13]);
+  u[6] = _mm_add_epi32(v[6], v[14]);
+  u[7] = _mm_add_epi32(v[7], v[15]);
+  u[8] = _mm_sub_epi32(v[0], v[8]);
+  u[9] = _mm_sub_epi32(v[1], v[9]);
+  u[10] = _mm_sub_epi32(v[2], v[10]);
+  u[11] = _mm_sub_epi32(v[3], v[11]);
+  u[12] = _mm_sub_epi32(v[4], v[12]);
+  u[13] = _mm_sub_epi32(v[5], v[13]);
+  u[14] = _mm_sub_epi32(v[6], v[14]);
+  u[15] = _mm_sub_epi32(v[7], v[15]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  x[0] = _mm_add_epi16(s[0], s[4]);
+  x[1] = _mm_add_epi16(s[1], s[5]);
+  x[2] = _mm_add_epi16(s[2], s[6]);
+  x[3] = _mm_add_epi16(s[3], s[7]);
+  x[4] = _mm_sub_epi16(s[0], s[4]);
+  x[5] = _mm_sub_epi16(s[1], s[5]);
+  x[6] = _mm_sub_epi16(s[2], s[6]);
+  x[7] = _mm_sub_epi16(s[3], s[7]);
+  x[8] = _mm_packs_epi32(u[0], u[1]);
+  x[9] = _mm_packs_epi32(u[2], u[3]);
+  x[10] = _mm_packs_epi32(u[4], u[5]);
+  x[11] = _mm_packs_epi32(u[6], u[7]);
+  x[12] = _mm_packs_epi32(u[8], u[9]);
+  x[13] = _mm_packs_epi32(u[10], u[11]);
+  x[14] = _mm_packs_epi32(u[12], u[13]);
+  x[15] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  u[0] = _mm_unpacklo_epi16(x[4], x[5]);
+  u[1] = _mm_unpackhi_epi16(x[4], x[5]);
+  u[2] = _mm_unpacklo_epi16(x[6], x[7]);
+  u[3] = _mm_unpackhi_epi16(x[6], x[7]);
+  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
+  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
+  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
+  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], v[4]);
+  u[1] = _mm_add_epi32(v[1], v[5]);
+  u[2] = _mm_add_epi32(v[2], v[6]);
+  u[3] = _mm_add_epi32(v[3], v[7]);
+  u[4] = _mm_sub_epi32(v[0], v[4]);
+  u[5] = _mm_sub_epi32(v[1], v[5]);
+  u[6] = _mm_sub_epi32(v[2], v[6]);
+  u[7] = _mm_sub_epi32(v[3], v[7]);
+  u[8] = _mm_add_epi32(v[8], v[12]);
+  u[9] = _mm_add_epi32(v[9], v[13]);
+  u[10] = _mm_add_epi32(v[10], v[14]);
+  u[11] = _mm_add_epi32(v[11], v[15]);
+  u[12] = _mm_sub_epi32(v[8], v[12]);
+  u[13] = _mm_sub_epi32(v[9], v[13]);
+  u[14] = _mm_sub_epi32(v[10], v[14]);
+  u[15] = _mm_sub_epi32(v[11], v[15]);
+
+  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_add_epi16(x[0], x[2]);
+  s[1] = _mm_add_epi16(x[1], x[3]);
+  s[2] = _mm_sub_epi16(x[0], x[2]);
+  s[3] = _mm_sub_epi16(x[1], x[3]);
+  s[4] = _mm_packs_epi32(v[0], v[1]);
+  s[5] = _mm_packs_epi32(v[2], v[3]);
+  s[6] = _mm_packs_epi32(v[4], v[5]);
+  s[7] = _mm_packs_epi32(v[6], v[7]);
+  s[8] = _mm_add_epi16(x[8], x[10]);
+  s[9] = _mm_add_epi16(x[9], x[11]);
+  s[10] = _mm_sub_epi16(x[8], x[10]);
+  s[11] = _mm_sub_epi16(x[9], x[11]);
+  s[12] = _mm_packs_epi32(v[8], v[9]);
+  s[13] = _mm_packs_epi32(v[10], v[11]);
+  s[14] = _mm_packs_epi32(v[12], v[13]);
+  s[15] = _mm_packs_epi32(v[14], v[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[0] = s[0];
+  in[1] = _mm_sub_epi16(kZero, s[8]);
+  in[2] = s[12];
+  in[3] = _mm_sub_epi16(kZero, s[4]);
+  in[4] = _mm_packs_epi32(v[4], v[5]);
+  in[5] = _mm_packs_epi32(v[12], v[13]);
+  in[6] = _mm_packs_epi32(v[8], v[9]);
+  in[7] = _mm_packs_epi32(v[0], v[1]);
+  in[8] = _mm_packs_epi32(v[2], v[3]);
+  in[9] = _mm_packs_epi32(v[10], v[11]);
+  in[10] = _mm_packs_epi32(v[14], v[15]);
+  in[11] = _mm_packs_epi32(v[6], v[7]);
+  in[12] = s[5];
+  in[13] = _mm_sub_epi16(kZero, s[13]);
+  in[14] = s[9];
+  in[15] = _mm_sub_epi16(kZero, s[1]);
+}
+
+static void vp10_idct16_8col(__m128i *in) {
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i v[16], u[16], s[16], t[16];
+
+  // stage 1
+  s[0] = in[0];
+  s[1] = in[8];
+  s[2] = in[4];
+  s[3] = in[12];
+  s[4] = in[2];
+  s[5] = in[10];
+  s[6] = in[6];
+  s[7] = in[14];
+  s[8] = in[1];
+  s[9] = in[9];
+  s[10] = in[5];
+  s[11] = in[13];
+  s[12] = in[3];
+  s[13] = in[11];
+  s[14] = in[7];
+  s[15] = in[15];
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[15]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[15]);
+  u[2] = _mm_unpacklo_epi16(s[9], s[14]);
+  u[3] = _mm_unpackhi_epi16(s[9], s[14]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[11], s[12]);
+  u[7] = _mm_unpackhi_epi16(s[11], s[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[8]  = _mm_packs_epi32(u[0], u[1]);
+  s[15] = _mm_packs_epi32(u[2], u[3]);
+  s[9]  = _mm_packs_epi32(u[4], u[5]);
+  s[14] = _mm_packs_epi32(u[6], u[7]);
+  s[10] = _mm_packs_epi32(u[8], u[9]);
+  s[13] = _mm_packs_epi32(u[10], u[11]);
+  s[11] = _mm_packs_epi32(u[12], u[13]);
+  s[12] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  t[0] = s[0];
+  t[1] = s[1];
+  t[2] = s[2];
+  t[3] = s[3];
+  u[0] = _mm_unpacklo_epi16(s[4], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[4], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[5], s[6]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[4] = _mm_packs_epi32(u[0], u[1]);
+  t[7] = _mm_packs_epi32(u[2], u[3]);
+  t[5] = _mm_packs_epi32(u[4], u[5]);
+  t[6] = _mm_packs_epi32(u[6], u[7]);
+  t[8] = _mm_add_epi16(s[8], s[9]);
+  t[9] = _mm_sub_epi16(s[8], s[9]);
+  t[10] = _mm_sub_epi16(s[11], s[10]);
+  t[11] = _mm_add_epi16(s[10], s[11]);
+  t[12] = _mm_add_epi16(s[12], s[13]);
+  t[13] = _mm_sub_epi16(s[12], s[13]);
+  t[14] = _mm_sub_epi16(s[15], s[14]);
+  t[15] = _mm_add_epi16(s[14], s[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(t[0], t[1]);
+  u[1] = _mm_unpackhi_epi16(t[0], t[1]);
+  u[2] = _mm_unpacklo_epi16(t[2], t[3]);
+  u[3] = _mm_unpackhi_epi16(t[2], t[3]);
+  u[4] = _mm_unpacklo_epi16(t[9], t[14]);
+  u[5] = _mm_unpackhi_epi16(t[9], t[14]);
+  u[6] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[7] = _mm_unpackhi_epi16(t[10], t[13]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_add_epi16(t[4], t[5]);
+  s[5] = _mm_sub_epi16(t[4], t[5]);
+  s[6] = _mm_sub_epi16(t[7], t[6]);
+  s[7] = _mm_add_epi16(t[6], t[7]);
+  s[8] = t[8];
+  s[15] = t[15];
+  s[9]  = _mm_packs_epi32(u[8], u[9]);
+  s[14] = _mm_packs_epi32(u[10], u[11]);
+  s[10] = _mm_packs_epi32(u[12], u[13]);
+  s[13] = _mm_packs_epi32(u[14], u[15]);
+  s[11] = t[11];
+  s[12] = t[12];
+
+  // stage 5
+  t[0] = _mm_add_epi16(s[0], s[3]);
+  t[1] = _mm_add_epi16(s[1], s[2]);
+  t[2] = _mm_sub_epi16(s[1], s[2]);
+  t[3] = _mm_sub_epi16(s[0], s[3]);
+  t[4] = s[4];
+  t[7] = s[7];
+
+  u[0] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[1] = _mm_unpackhi_epi16(s[5], s[6]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  t[5] = _mm_packs_epi32(u[0], u[1]);
+  t[6] = _mm_packs_epi32(u[2], u[3]);
+
+  t[8] = _mm_add_epi16(s[8], s[11]);
+  t[9] = _mm_add_epi16(s[9], s[10]);
+  t[10] = _mm_sub_epi16(s[9], s[10]);
+  t[11] = _mm_sub_epi16(s[8], s[11]);
+  t[12] = _mm_sub_epi16(s[15], s[12]);
+  t[13] = _mm_sub_epi16(s[14], s[13]);
+  t[14] = _mm_add_epi16(s[13], s[14]);
+  t[15] = _mm_add_epi16(s[12], s[15]);
+
+  // stage 6
+  s[0] = _mm_add_epi16(t[0], t[7]);
+  s[1] = _mm_add_epi16(t[1], t[6]);
+  s[2] = _mm_add_epi16(t[2], t[5]);
+  s[3] = _mm_add_epi16(t[3], t[4]);
+  s[4] = _mm_sub_epi16(t[3], t[4]);
+  s[5] = _mm_sub_epi16(t[2], t[5]);
+  s[6] = _mm_sub_epi16(t[1], t[6]);
+  s[7] = _mm_sub_epi16(t[0], t[7]);
+  s[8] = t[8];
+  s[9] = t[9];
+
+  u[0] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[1] = _mm_unpackhi_epi16(t[10], t[13]);
+  u[2] = _mm_unpacklo_epi16(t[11], t[12]);
+  u[3] = _mm_unpackhi_epi16(t[11], t[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  s[10] = _mm_packs_epi32(u[0], u[1]);
+  s[13] = _mm_packs_epi32(u[2], u[3]);
+  s[11] = _mm_packs_epi32(u[4], u[5]);
+  s[12] = _mm_packs_epi32(u[6], u[7]);
+  s[14] = t[14];
+  s[15] = t[15];
+
+  // stage 7
+  in[0] = _mm_add_epi16(s[0], s[15]);
+  in[1] = _mm_add_epi16(s[1], s[14]);
+  in[2] = _mm_add_epi16(s[2], s[13]);
+  in[3] = _mm_add_epi16(s[3], s[12]);
+  in[4] = _mm_add_epi16(s[4], s[11]);
+  in[5] = _mm_add_epi16(s[5], s[10]);
+  in[6] = _mm_add_epi16(s[6], s[9]);
+  in[7] = _mm_add_epi16(s[7], s[8]);
+  in[8] = _mm_sub_epi16(s[7], s[8]);
+  in[9] = _mm_sub_epi16(s[6], s[9]);
+  in[10] = _mm_sub_epi16(s[5], s[10]);
+  in[11] = _mm_sub_epi16(s[4], s[11]);
+  in[12] = _mm_sub_epi16(s[3], s[12]);
+  in[13] = _mm_sub_epi16(s[2], s[13]);
+  in[14] = _mm_sub_epi16(s[1], s[14]);
+  in[15] = _mm_sub_epi16(s[0], s[15]);
+}
+
+void vp10_idct16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  vp10_idct16_8col(in0);
+  vp10_idct16_8col(in1);
+}
+
+void vp10_iadst16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  vp10_iadst16_8col(in0);
+  vp10_iadst16_8col(in1);
+}
+
+void vp10_idct16x16_10_add_sse2(const int16_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  __m128i in[16], l[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+  // First 1-D inverse DCT
+  // Load input data.
+  in[0] = _mm_load_si128((const __m128i *)input);
+  in[1] = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in[2] = _mm_load_si128((const __m128i *)(input + 8 * 4));
+  in[3] = _mm_load_si128((const __m128i *)(input + 8 * 6));
+
+  TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]);
+
+  // Stage2
+  {
+    const __m128i lo_1_15 = _mm_unpackhi_epi16(in[0], zero);
+    const __m128i lo_13_3 = _mm_unpackhi_epi16(zero, in[1]);
+
+    tmp0 = _mm_madd_epi16(lo_1_15, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_1_15, stg2_1);
+    tmp5 = _mm_madd_epi16(lo_13_3, stg2_6);
+    tmp7 = _mm_madd_epi16(lo_13_3, stg2_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp2_8  = _mm_packs_epi32(tmp0, tmp2);
+    stp2_11 = _mm_packs_epi32(tmp5, tmp7);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[1], zero);
+
+    tmp0 = _mm_madd_epi16(lo_2_14, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_2_14, stg3_1);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_13 = _mm_unpackhi_epi64(stp2_11, zero);
+    stp1_14 = _mm_unpackhi_epi64(stp2_8, zero);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  {
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero);
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13);
+
+    tmp0 = _mm_madd_epi16(lo_0_8, stg4_0);
+    tmp2 = _mm_madd_epi16(lo_0_8, stg4_1);
+    tmp1 = _mm_madd_epi16(lo_9_14, stg4_4);
+    tmp3 = _mm_madd_epi16(lo_9_14, stg4_5);
+    tmp5 = _mm_madd_epi16(lo_10_13, stg4_6);
+    tmp7 = _mm_madd_epi16(lo_10_13, stg4_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp1_0 = _mm_packs_epi32(tmp0, tmp0);
+    stp1_1 = _mm_packs_epi32(tmp2, tmp2);
+    stp2_9 = _mm_packs_epi32(tmp1, tmp3);
+    stp2_10 = _mm_packs_epi32(tmp5, tmp7);
+
+    stp2_6 = _mm_unpackhi_epi64(stp1_4, zero);
+  }
+
+  // Stage5 and Stage6
+  {
+    tmp0 = _mm_add_epi16(stp2_8, stp2_11);
+    tmp1 = _mm_sub_epi16(stp2_8, stp2_11);
+    tmp2 = _mm_add_epi16(stp2_9, stp2_10);
+    tmp3 = _mm_sub_epi16(stp2_9, stp2_10);
+
+    stp1_9  = _mm_unpacklo_epi64(tmp2, zero);
+    stp1_10 = _mm_unpacklo_epi64(tmp3, zero);
+    stp1_8  = _mm_unpacklo_epi64(tmp0, zero);
+    stp1_11 = _mm_unpacklo_epi64(tmp1, zero);
+
+    stp1_13 = _mm_unpackhi_epi64(tmp3, zero);
+    stp1_14 = _mm_unpackhi_epi64(tmp2, zero);
+    stp1_12 = _mm_unpackhi_epi64(tmp1, zero);
+    stp1_15 = _mm_unpackhi_epi64(tmp0, zero);
+  }
+
+  // Stage6
+  {
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp1_4);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);
+
+    tmp1 = _mm_madd_epi16(lo_6_5, stg4_1);
+    tmp3 = _mm_madd_epi16(lo_6_5, stg4_0);
+    tmp0 = _mm_madd_epi16(lo_10_13, stg6_0);
+    tmp2 = _mm_madd_epi16(lo_10_13, stg4_0);
+    tmp4 = _mm_madd_epi16(lo_11_12, stg6_0);
+    tmp6 = _mm_madd_epi16(lo_11_12, stg4_0);
+
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_6 = _mm_packs_epi32(tmp3, tmp1);
+
+    stp2_10 = _mm_packs_epi32(tmp0, zero);
+    stp2_13 = _mm_packs_epi32(tmp2, zero);
+    stp2_11 = _mm_packs_epi32(tmp4, zero);
+    stp2_12 = _mm_packs_epi32(tmp6, zero);
+
+    tmp0 = _mm_add_epi16(stp1_0, stp1_4);
+    tmp1 = _mm_sub_epi16(stp1_0, stp1_4);
+    tmp2 = _mm_add_epi16(stp1_1, stp1_6);
+    tmp3 = _mm_sub_epi16(stp1_1, stp1_6);
+
+    stp2_0 = _mm_unpackhi_epi64(tmp0, zero);
+    stp2_1 = _mm_unpacklo_epi64(tmp2, zero);
+    stp2_2 = _mm_unpackhi_epi64(tmp2, zero);
+    stp2_3 = _mm_unpacklo_epi64(tmp0, zero);
+    stp2_4 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_5 = _mm_unpackhi_epi64(tmp3, zero);
+    stp2_6 = _mm_unpacklo_epi64(tmp3, zero);
+    stp2_7 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage7. Left 8x16 only.
+  l[0] = _mm_add_epi16(stp2_0, stp1_15);
+  l[1] = _mm_add_epi16(stp2_1, stp1_14);
+  l[2] = _mm_add_epi16(stp2_2, stp2_13);
+  l[3] = _mm_add_epi16(stp2_3, stp2_12);
+  l[4] = _mm_add_epi16(stp2_4, stp2_11);
+  l[5] = _mm_add_epi16(stp2_5, stp2_10);
+  l[6] = _mm_add_epi16(stp2_6, stp1_9);
+  l[7] = _mm_add_epi16(stp2_7, stp1_8);
+  l[8] = _mm_sub_epi16(stp2_7, stp1_8);
+  l[9] = _mm_sub_epi16(stp2_6, stp1_9);
+  l[10] = _mm_sub_epi16(stp2_5, stp2_10);
+  l[11] = _mm_sub_epi16(stp2_4, stp2_11);
+  l[12] = _mm_sub_epi16(stp2_3, stp2_12);
+  l[13] = _mm_sub_epi16(stp2_2, stp2_13);
+  l[14] = _mm_sub_epi16(stp2_1, stp1_14);
+  l[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+  // Second 1-D inverse transform, performed per 8x16 block
+  for (i = 0; i < 2; i++) {
+    int j;
+    array_transpose_4X8(l + 8 * i, in);
+
+    IDCT16_10
+
+    // Stage7
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+#define LOAD_DQCOEFF(reg, input) \
+  {  \
+    reg = _mm_load_si128((const __m128i *) input); \
+    input += 8; \
+  }  \
+
+#define IDCT32_34 \
+/* Stage1 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], zero); \
+  const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], zero); \
+  \
+  const __m128i lo_25_7= _mm_unpacklo_epi16(zero, in[7]); \
+  const __m128i hi_25_7 = _mm_unpackhi_epi16(zero, in[7]); \
+  \
+  const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], zero); \
+  const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], zero); \
+  \
+  const __m128i lo_29_3 = _mm_unpacklo_epi16(zero, in[3]); \
+  const __m128i hi_29_3 = _mm_unpackhi_epi16(zero, in[3]); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_1_31, hi_1_31, stg1_0, \
+                         stg1_1, stp1_16, stp1_31); \
+  MULTIPLICATION_AND_ADD_2(lo_25_7, hi_25_7, stg1_6, \
+                         stg1_7, stp1_19, stp1_28); \
+  MULTIPLICATION_AND_ADD_2(lo_5_27, hi_5_27, stg1_8, \
+                         stg1_9, stp1_20, stp1_27); \
+  MULTIPLICATION_AND_ADD_2(lo_29_3, hi_29_3, stg1_14, \
+                         stg1_15, stp1_23, stp1_24); \
+} \
+\
+/* Stage2 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], zero); \
+  const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], zero); \
+  \
+  const __m128i lo_26_6 = _mm_unpacklo_epi16(zero, in[6]); \
+  const __m128i hi_26_6 = _mm_unpackhi_epi16(zero, in[6]); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_2_30, hi_2_30, stg2_0, \
+                         stg2_1, stp2_8, stp2_15); \
+  MULTIPLICATION_AND_ADD_2(lo_26_6, hi_26_6, stg2_6, \
+                         stg2_7, stp2_11, stp2_12); \
+  \
+  stp2_16 = stp1_16; \
+  stp2_19 = stp1_19; \
+  \
+  stp2_20 = stp1_20; \
+  stp2_23 = stp1_23; \
+  \
+  stp2_24 = stp1_24; \
+  stp2_27 = stp1_27; \
+  \
+  stp2_28 = stp1_28; \
+  stp2_31 = stp1_31; \
+} \
+\
+/* Stage3 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], zero); \
+  const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], zero); \
+  \
+  const __m128i lo_17_30 = _mm_unpacklo_epi16(stp1_16, stp1_31); \
+  const __m128i hi_17_30 = _mm_unpackhi_epi16(stp1_16, stp1_31); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp1_19, stp1_28); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp1_19, stp1_28); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp1_20, stp1_27); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp1_20, stp1_27); \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp1_23, stp1_24); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp1_23, stp2_24); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_4_28, hi_4_28, stg3_0, \
+                         stg3_1, stp1_4, stp1_7); \
+  \
+  stp1_8 = stp2_8; \
+  stp1_11 = stp2_11; \
+  stp1_12 = stp2_12; \
+  stp1_15 = stp2_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
+                         stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, \
+                         stp1_18, stp1_29) \
+  MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
+                         stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, \
+                         stp1_22, stp1_25) \
+  \
+  stp1_16 = stp2_16; \
+  stp1_31 = stp2_31; \
+  stp1_19 = stp2_19; \
+  stp1_20 = stp2_20; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_27 = stp2_27; \
+  stp1_28 = stp2_28; \
+} \
+\
+/* Stage4 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], zero); \
+  const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], zero); \
+  \
+  const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp2_15); \
+  const __m128i hi_9_14 = _mm_unpackhi_epi16(stp2_8, stp2_15); \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp2_12); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp2_11, stp2_12); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_0_16, hi_0_16, stg4_0, \
+                         stg4_1, stp2_0, stp2_1); \
+  \
+  stp2_4 = stp1_4; \
+  stp2_5 = stp1_4; \
+  stp2_6 = stp1_7; \
+  stp2_7 = stp1_7; \
+  \
+  MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
+                         stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, \
+                         stp2_10, stp2_13) \
+  \
+  stp2_8 = stp1_8; \
+  stp2_15 = stp1_15; \
+  stp2_11 = stp1_11; \
+  stp2_12 = stp1_12; \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
+  stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
+  stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
+  stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
+  stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
+  stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
+  stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
+  stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
+  stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
+} \
+\
+/* Stage5 */ \
+{ \
+  const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+  const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
+  const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  stp1_0 = stp2_0; \
+  stp1_1 = stp2_1; \
+  stp1_2 = stp2_1; \
+  stp1_3 = stp2_0; \
+  \
+  tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+  tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+  tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+  tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+  \
+  tmp0 = _mm_add_epi32(tmp0, rounding); \
+  tmp1 = _mm_add_epi32(tmp1, rounding); \
+  tmp2 = _mm_add_epi32(tmp2, rounding); \
+  tmp3 = _mm_add_epi32(tmp3, rounding); \
+  \
+  tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+  tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+  \
+  stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+  stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  \
+  stp1_4 = stp2_4; \
+  stp1_7 = stp2_7; \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
+  stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
+  stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
+  stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  \
+  MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
+                         stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
+                         stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  \
+  stp1_22 = stp2_22; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_25 = stp2_25; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+} \
+\
+/* Stage6 */ \
+{ \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+  const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+  \
+  stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
+  stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+  stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+  stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
+  stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
+  stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+  stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
+  \
+  stp2_8 = stp1_8; \
+  stp2_9 = stp1_9; \
+  stp2_14 = stp1_14; \
+  stp2_15 = stp1_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                         stg6_0, stg4_0, stg6_0, stg4_0, stp2_10, \
+                         stp2_13, stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
+  stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
+  stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
+  stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
+  \
+  stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
+  stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
+  stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
+  stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
+} \
+\
+/* Stage7 */ \
+{ \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
+  const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
+  stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
+  stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
+  stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
+  stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
+  stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
+  stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
+  stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
+  stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
+  stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  stp1_18 = stp2_18; \
+  stp1_19 = stp2_19; \
+  \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+  \
+  stp1_28 = stp2_28; \
+  stp1_29 = stp2_29; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+}
+
+
+#define IDCT32 \
+/* Stage1 */ \
+{ \
+  const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], in[31]); \
+  const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], in[31]); \
+  const __m128i lo_17_15 = _mm_unpacklo_epi16(in[17], in[15]); \
+  const __m128i hi_17_15 = _mm_unpackhi_epi16(in[17], in[15]); \
+  \
+  const __m128i lo_9_23 = _mm_unpacklo_epi16(in[9], in[23]); \
+  const __m128i hi_9_23 = _mm_unpackhi_epi16(in[9], in[23]); \
+  const __m128i lo_25_7= _mm_unpacklo_epi16(in[25], in[7]); \
+  const __m128i hi_25_7 = _mm_unpackhi_epi16(in[25], in[7]); \
+  \
+  const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], in[27]); \
+  const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], in[27]); \
+  const __m128i lo_21_11 = _mm_unpacklo_epi16(in[21], in[11]); \
+  const __m128i hi_21_11 = _mm_unpackhi_epi16(in[21], in[11]); \
+  \
+  const __m128i lo_13_19 = _mm_unpacklo_epi16(in[13], in[19]); \
+  const __m128i hi_13_19 = _mm_unpackhi_epi16(in[13], in[19]); \
+  const __m128i lo_29_3 = _mm_unpacklo_epi16(in[29], in[3]); \
+  const __m128i hi_29_3 = _mm_unpackhi_epi16(in[29], in[3]); \
+  \
+  MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0, \
+                         stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, \
+                         stp1_17, stp1_30) \
+  MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, \
+                         stg1_5, stg1_6, stg1_7, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8, \
+                         stg1_9, stg1_10, stg1_11, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12, \
+                         stg1_13, stg1_14, stg1_15, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+} \
+\
+/* Stage2 */ \
+{ \
+  const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], in[30]); \
+  const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], in[30]); \
+  const __m128i lo_18_14 = _mm_unpacklo_epi16(in[18], in[14]); \
+  const __m128i hi_18_14 = _mm_unpackhi_epi16(in[18], in[14]); \
+  \
+  const __m128i lo_10_22 = _mm_unpacklo_epi16(in[10], in[22]); \
+  const __m128i hi_10_22 = _mm_unpackhi_epi16(in[10], in[22]); \
+  const __m128i lo_26_6 = _mm_unpacklo_epi16(in[26], in[6]); \
+  const __m128i hi_26_6 = _mm_unpackhi_epi16(in[26], in[6]); \
+  \
+  MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0, \
+                         stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, \
+                         stp2_14) \
+  MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4, \
+                         stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, \
+                         stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_17); \
+  stp2_17 = _mm_sub_epi16(stp1_16, stp1_17); \
+  stp2_18 = _mm_sub_epi16(stp1_19, stp1_18); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_18); \
+  \
+  stp2_20 = _mm_add_epi16(stp1_20, stp1_21); \
+  stp2_21 = _mm_sub_epi16(stp1_20, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_23, stp1_22); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_22); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_25); \
+  stp2_25 = _mm_sub_epi16(stp1_24, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_27, stp1_26); \
+  stp2_27 = _mm_add_epi16(stp1_27, stp1_26); \
+  \
+  stp2_28 = _mm_add_epi16(stp1_28, stp1_29); \
+  stp2_29 = _mm_sub_epi16(stp1_28, stp1_29); \
+  stp2_30 = _mm_sub_epi16(stp1_31, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_31, stp1_30); \
+} \
+\
+/* Stage3 */ \
+{ \
+  const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], in[28]); \
+  const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], in[28]); \
+  const __m128i lo_20_12 = _mm_unpacklo_epi16(in[20], in[12]); \
+  const __m128i hi_20_12 = _mm_unpackhi_epi16(in[20], in[12]); \
+  \
+  const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30); \
+  const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  \
+  MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0, \
+                         stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, \
+                         stp1_6) \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_9); \
+  stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
+  stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
+  stp1_12 = _mm_add_epi16(stp2_12, stp2_13); \
+  stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
+  \
+  MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
+                         stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, \
+                         stp1_18, stp1_29) \
+  MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
+                         stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, \
+                         stp1_22, stp1_25) \
+  \
+  stp1_16 = stp2_16; \
+  stp1_31 = stp2_31; \
+  stp1_19 = stp2_19; \
+  stp1_20 = stp2_20; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_27 = stp2_27; \
+  stp1_28 = stp2_28; \
+} \
+\
+/* Stage4 */ \
+{ \
+  const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], in[16]); \
+  const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], in[16]); \
+  const __m128i lo_8_24 = _mm_unpacklo_epi16(in[8], in[24]); \
+  const __m128i hi_8_24 = _mm_unpackhi_epi16(in[8], in[24]); \
+  \
+  const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+  const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  \
+  MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, \
+                         stg4_1, stg4_2, stg4_3, stp2_0, stp2_1, \
+                         stp2_2, stp2_3) \
+  \
+  stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
+  stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
+  stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
+  \
+  MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
+                         stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, \
+                         stp2_10, stp2_13) \
+  \
+  stp2_8 = stp1_8; \
+  stp2_15 = stp1_15; \
+  stp2_11 = stp1_11; \
+  stp2_12 = stp1_12; \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
+  stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
+  stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
+  stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
+  stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
+  stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
+  stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
+  stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
+  stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
+} \
+\
+/* Stage5 */ \
+{ \
+  const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+  const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
+  const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
+  stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
+  stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
+  \
+  tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+  tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+  tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+  tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+  \
+  tmp0 = _mm_add_epi32(tmp0, rounding); \
+  tmp1 = _mm_add_epi32(tmp1, rounding); \
+  tmp2 = _mm_add_epi32(tmp2, rounding); \
+  tmp3 = _mm_add_epi32(tmp3, rounding); \
+  \
+  tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+  tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+  \
+  stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+  stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  \
+  stp1_4 = stp2_4; \
+  stp1_7 = stp2_7; \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
+  stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
+  stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
+  stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  \
+  MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
+                         stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
+                         stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  \
+  stp1_22 = stp2_22; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_25 = stp2_25; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+} \
+\
+/* Stage6 */ \
+{ \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+  const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+  \
+  stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
+  stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+  stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+  stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
+  stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
+  stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+  stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
+  \
+  stp2_8 = stp1_8; \
+  stp2_9 = stp1_9; \
+  stp2_14 = stp1_14; \
+  stp2_15 = stp1_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                         stg6_0, stg4_0, stg6_0, stg4_0, stp2_10, \
+                         stp2_13, stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
+  stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
+  stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
+  stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
+  \
+  stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
+  stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
+  stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
+  stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
+} \
+\
+/* Stage7 */ \
+{ \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
+  const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
+  stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
+  stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
+  stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
+  stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
+  stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
+  stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
+  stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
+  stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
+  stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  stp1_18 = stp2_18; \
+  stp1_19 = stp2_19; \
+  \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+  \
+  stp1_28 = stp2_28; \
+  stp1_29 = stp2_29; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+}
+
+// Only upper-left 8x8 has non-zero coeff
+void vp10_idct32x32_34_add_sse2(const int16_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+
+  // vp10_idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[32];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22,
+          stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29,
+          stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+          stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22,
+          stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29,
+          stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data. Only need to load the top left 8x8 block.
+  in[0] = _mm_load_si128((const __m128i *)input);
+  in[1] = _mm_load_si128((const __m128i *)(input + 32));
+  in[2] = _mm_load_si128((const __m128i *)(input + 64));
+  in[3] = _mm_load_si128((const __m128i *)(input + 96));
+  in[4] = _mm_load_si128((const __m128i *)(input + 128));
+  in[5] = _mm_load_si128((const __m128i *)(input + 160));
+  in[6] = _mm_load_si128((const __m128i *)(input + 192));
+  in[7] = _mm_load_si128((const __m128i *)(input + 224));
+
+  for (i = 8; i < 32; ++i) {
+    in[i] = _mm_setzero_si128();
+  }
+
+  array_transpose_8x8(in, in);
+  // TODO(hkuang): Following transposes are unnecessary. But remove them will
+  // lead to performance drop on some devices.
+  array_transpose_8x8(in + 8, in + 8);
+  array_transpose_8x8(in + 16, in + 16);
+  array_transpose_8x8(in + 24, in + 24);
+
+  IDCT32_34
+
+  // 1_D: Store 32 intermediate results for each 8x32 block.
+  col[0] = _mm_add_epi16(stp1_0, stp1_31);
+  col[1] = _mm_add_epi16(stp1_1, stp1_30);
+  col[2] = _mm_add_epi16(stp1_2, stp1_29);
+  col[3] = _mm_add_epi16(stp1_3, stp1_28);
+  col[4] = _mm_add_epi16(stp1_4, stp1_27);
+  col[5] = _mm_add_epi16(stp1_5, stp1_26);
+  col[6] = _mm_add_epi16(stp1_6, stp1_25);
+  col[7] = _mm_add_epi16(stp1_7, stp1_24);
+  col[8] = _mm_add_epi16(stp1_8, stp1_23);
+  col[9] = _mm_add_epi16(stp1_9, stp1_22);
+  col[10] = _mm_add_epi16(stp1_10, stp1_21);
+  col[11] = _mm_add_epi16(stp1_11, stp1_20);
+  col[12] = _mm_add_epi16(stp1_12, stp1_19);
+  col[13] = _mm_add_epi16(stp1_13, stp1_18);
+  col[14] = _mm_add_epi16(stp1_14, stp1_17);
+  col[15] = _mm_add_epi16(stp1_15, stp1_16);
+  col[16] = _mm_sub_epi16(stp1_15, stp1_16);
+  col[17] = _mm_sub_epi16(stp1_14, stp1_17);
+  col[18] = _mm_sub_epi16(stp1_13, stp1_18);
+  col[19] = _mm_sub_epi16(stp1_12, stp1_19);
+  col[20] = _mm_sub_epi16(stp1_11, stp1_20);
+  col[21] = _mm_sub_epi16(stp1_10, stp1_21);
+  col[22] = _mm_sub_epi16(stp1_9, stp1_22);
+  col[23] = _mm_sub_epi16(stp1_8, stp1_23);
+  col[24] = _mm_sub_epi16(stp1_7, stp1_24);
+  col[25] = _mm_sub_epi16(stp1_6, stp1_25);
+  col[26] = _mm_sub_epi16(stp1_5, stp1_26);
+  col[27] = _mm_sub_epi16(stp1_4, stp1_27);
+  col[28] = _mm_sub_epi16(stp1_3, stp1_28);
+  col[29] = _mm_sub_epi16(stp1_2, stp1_29);
+  col[30] = _mm_sub_epi16(stp1_1, stp1_30);
+  col[31] = _mm_sub_epi16(stp1_0, stp1_31);
+  for (i = 0; i < 4; i++) {
+    int j;
+    const __m128i zero = _mm_setzero_si128();
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + i * 8, in);
+    IDCT32_34
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vp10_idct32x32_1024_add_sse2(const int16_t *input, uint8_t *dest,
+                                 int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  // vp10_idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[128], zero_idx[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22,
+          stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29,
+          stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+          stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22,
+          stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29,
+          stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i, j, i32;
+
+  for (i = 0; i < 4; i++) {
+    i32 = (i << 5);
+    // First 1-D vp10_idct
+    // Load input data.
+    LOAD_DQCOEFF(in[0], input);
+    LOAD_DQCOEFF(in[8], input);
+    LOAD_DQCOEFF(in[16], input);
+    LOAD_DQCOEFF(in[24], input);
+    LOAD_DQCOEFF(in[1], input);
+    LOAD_DQCOEFF(in[9], input);
+    LOAD_DQCOEFF(in[17], input);
+    LOAD_DQCOEFF(in[25], input);
+    LOAD_DQCOEFF(in[2], input);
+    LOAD_DQCOEFF(in[10], input);
+    LOAD_DQCOEFF(in[18], input);
+    LOAD_DQCOEFF(in[26], input);
+    LOAD_DQCOEFF(in[3], input);
+    LOAD_DQCOEFF(in[11], input);
+    LOAD_DQCOEFF(in[19], input);
+    LOAD_DQCOEFF(in[27], input);
+
+    LOAD_DQCOEFF(in[4], input);
+    LOAD_DQCOEFF(in[12], input);
+    LOAD_DQCOEFF(in[20], input);
+    LOAD_DQCOEFF(in[28], input);
+    LOAD_DQCOEFF(in[5], input);
+    LOAD_DQCOEFF(in[13], input);
+    LOAD_DQCOEFF(in[21], input);
+    LOAD_DQCOEFF(in[29], input);
+    LOAD_DQCOEFF(in[6], input);
+    LOAD_DQCOEFF(in[14], input);
+    LOAD_DQCOEFF(in[22], input);
+    LOAD_DQCOEFF(in[30], input);
+    LOAD_DQCOEFF(in[7], input);
+    LOAD_DQCOEFF(in[15], input);
+    LOAD_DQCOEFF(in[23], input);
+    LOAD_DQCOEFF(in[31], input);
+
+    // checking if all entries are zero
+    zero_idx[0] = _mm_or_si128(in[0], in[1]);
+    zero_idx[1] = _mm_or_si128(in[2], in[3]);
+    zero_idx[2] = _mm_or_si128(in[4], in[5]);
+    zero_idx[3] = _mm_or_si128(in[6], in[7]);
+    zero_idx[4] = _mm_or_si128(in[8], in[9]);
+    zero_idx[5] = _mm_or_si128(in[10], in[11]);
+    zero_idx[6] = _mm_or_si128(in[12], in[13]);
+    zero_idx[7] = _mm_or_si128(in[14], in[15]);
+    zero_idx[8] = _mm_or_si128(in[16], in[17]);
+    zero_idx[9] = _mm_or_si128(in[18], in[19]);
+    zero_idx[10] = _mm_or_si128(in[20], in[21]);
+    zero_idx[11] = _mm_or_si128(in[22], in[23]);
+    zero_idx[12] = _mm_or_si128(in[24], in[25]);
+    zero_idx[13] = _mm_or_si128(in[26], in[27]);
+    zero_idx[14] = _mm_or_si128(in[28], in[29]);
+    zero_idx[15] = _mm_or_si128(in[30], in[31]);
+
+    zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+    zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]);
+
+    zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+
+    if (_mm_movemask_epi8(_mm_cmpeq_epi32(zero_idx[14], zero)) == 0xFFFF) {
+      col[i32 + 0] = _mm_setzero_si128();
+      col[i32 + 1] = _mm_setzero_si128();
+      col[i32 + 2] = _mm_setzero_si128();
+      col[i32 + 3] = _mm_setzero_si128();
+      col[i32 + 4] = _mm_setzero_si128();
+      col[i32 + 5] = _mm_setzero_si128();
+      col[i32 + 6] = _mm_setzero_si128();
+      col[i32 + 7] = _mm_setzero_si128();
+      col[i32 + 8] = _mm_setzero_si128();
+      col[i32 + 9] = _mm_setzero_si128();
+      col[i32 + 10] = _mm_setzero_si128();
+      col[i32 + 11] = _mm_setzero_si128();
+      col[i32 + 12] = _mm_setzero_si128();
+      col[i32 + 13] = _mm_setzero_si128();
+      col[i32 + 14] = _mm_setzero_si128();
+      col[i32 + 15] = _mm_setzero_si128();
+      col[i32 + 16] = _mm_setzero_si128();
+      col[i32 + 17] = _mm_setzero_si128();
+      col[i32 + 18] = _mm_setzero_si128();
+      col[i32 + 19] = _mm_setzero_si128();
+      col[i32 + 20] = _mm_setzero_si128();
+      col[i32 + 21] = _mm_setzero_si128();
+      col[i32 + 22] = _mm_setzero_si128();
+      col[i32 + 23] = _mm_setzero_si128();
+      col[i32 + 24] = _mm_setzero_si128();
+      col[i32 + 25] = _mm_setzero_si128();
+      col[i32 + 26] = _mm_setzero_si128();
+      col[i32 + 27] = _mm_setzero_si128();
+      col[i32 + 28] = _mm_setzero_si128();
+      col[i32 + 29] = _mm_setzero_si128();
+      col[i32 + 30] = _mm_setzero_si128();
+      col[i32 + 31] = _mm_setzero_si128();
+      continue;
+    }
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+    array_transpose_8x8(in + 16, in + 16);
+    array_transpose_8x8(in + 24, in + 24);
+
+    IDCT32
+
+    // 1_D: Store 32 intermediate results for each 8x32 block.
+    col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31);
+    col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30);
+    col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29);
+    col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28);
+    col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27);
+    col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26);
+    col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25);
+    col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24);
+    col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23);
+    col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22);
+    col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21);
+    col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20);
+    col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19);
+    col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18);
+    col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17);
+    col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16);
+    col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16);
+    col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17);
+    col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18);
+    col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19);
+    col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20);
+    col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21);
+    col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22);
+    col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23);
+    col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24);
+    col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25);
+    col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26);
+    col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27);
+    col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28);
+    col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29);
+    col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
+    col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
+  }
+  for (i = 0; i < 4; i++) {
+    // Second 1-D vp10_idct
+    j = i << 3;
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + j, in);
+    array_transpose_8x8(col + j + 32, in + 8);
+    array_transpose_8x8(col + j + 64, in + 16);
+    array_transpose_8x8(col + j + 96, in + 24);
+
+    IDCT32
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vp10_idct32x32_1_add_sse2(const int16_t *input,
+                               uint8_t *dest,
+                               int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, i;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (i = 0; i < 4; ++i) {
+    int j;
+    for (j = 0; j < 32; ++j) {
+      RECON_AND_STORE(dest + j * stride, dc_value);
+    }
+    dest += 8;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE __m128i clamp_high_sse2(__m128i value, int bd) {
+  __m128i ubounded, retval;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_subs_epi16(_mm_slli_epi16(one, bd), one);
+  ubounded = _mm_cmpgt_epi16(value, max);
+  retval = _mm_andnot_si128(ubounded, value);
+  ubounded = _mm_and_si128(ubounded, max);
+  retval = _mm_or_si128(retval, ubounded);
+  retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero));
+  return retval;
+}
+
+void vp10_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  __m128i inptr[4];
+  __m128i sign_bits[2];
+  __m128i temp_mm, min_input, max_input;
+  int test;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  int optimised_cols = 0;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i max = _mm_set1_epi16(12043);
+  const __m128i min = _mm_set1_epi16(-12043);
+  // Load input into __m128i
+  inptr[0] = _mm_loadu_si128((const __m128i *)input);
+  inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4));
+  inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8));
+  inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12));
+
+  // Pack to 16 bits
+  inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]);
+  inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]);
+
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp_mm = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp_mm);
+
+  if (!test) {
+    // Do the row transform
+    vp10_idct4_sse2(inptr);
+
+    // Check the min & max values
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp_mm = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp_mm);
+
+    if (test) {
+      transpose_4x4(inptr);
+      sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero);
+      sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero);
+      inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]);
+      inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]);
+      inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]);
+      inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]);
+      _mm_storeu_si128((__m128i *)outptr, inptr[0]);
+      _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]);
+      _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]);
+      _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]);
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vp10_highbd_idct4_c(input, outptr, bd);
+      input += 4;
+      outptr += 4;
+    }
+  }
+
+  if (optimised_cols) {
+    vp10_idct4_sse2(inptr);
+
+    // Final round and shift
+    inptr[0] = _mm_add_epi16(inptr[0], eight);
+    inptr[1] = _mm_add_epi16(inptr[1], eight);
+
+    inptr[0] = _mm_srai_epi16(inptr[0], 4);
+    inptr[1] = _mm_srai_epi16(inptr[1], 4);
+
+    // Reconstruction and Store
+    {
+      __m128i d0 = _mm_loadl_epi64((const __m128i *)dest);
+      __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2));
+      d0 = _mm_unpacklo_epi64(
+          d0, _mm_loadl_epi64((const __m128i *)(dest + stride)));
+      d2 = _mm_unpacklo_epi64(
+          d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3)));
+      d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd);
+      d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd);
+      // store input0
+      _mm_storel_epi64((__m128i *)dest, d0);
+      // store input1
+      d0 = _mm_srli_si128(d0, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride), d0);
+      // store input2
+      _mm_storel_epi64((__m128i *)(dest + stride * 2), d2);
+      // store input3
+      d2 = _mm_srli_si128(d2, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride * 3), d2);
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[4], temp_out[4];
+    // Columns
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        temp_in[j] = out[j * 4 + i];
+      vp10_highbd_idct4_c(temp_in, temp_out, bd);
+      for (j = 0; j < 4; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+      }
+    }
+  }
+}
+
+void vp10_highbd_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[8];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  const __m128i max = _mm_set1_epi16(6201);
+  const __m128i min = _mm_set1_epi16(-6201);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 8; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 8; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    vp10_idct8_sse2(inptr);
+
+    // Find the min & max for the column transform
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 8; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      array_transpose_8x8(inptr, inptr);
+      for (i = 0; i < 8; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 8; ++i) {
+      vp10_highbd_idct8_c(input, outptr, bd);
+      input += 8;
+      outptr += 8;
+    }
+  }
+
+  if (optimised_cols) {
+    vp10_idct8_sse2(inptr);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[8];
+      for (i = 0; i < 8; i++) {
+        inptr[i] = _mm_add_epi16(inptr[i], sixteen);
+        d[i] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        inptr[i] = _mm_srai_epi16(inptr[i], 5);
+        d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[i]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[8], temp_out[8];
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j * 8 + i];
+      vp10_highbd_idct8_c(temp_in, temp_out, bd);
+      for (j = 0; j < 8; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+      }
+    }
+  }
+}
+
+void vp10_highbd_idct8x8_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[8];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  const __m128i max = _mm_set1_epi16(6201);
+  const __m128i min = _mm_set1_epi16(-6201);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 8; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  // only first 4 row has non-zero coefs
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 4; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    vp10_idct8_sse2(inptr);
+
+    // Find the min & max for the column transform
+    // N.B. Only first 4 cols contain non-zero coeffs
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 8; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      // Use fact only first 4 rows contain non-zero coeffs
+      array_transpose_4X8(inptr, inptr);
+      for (i = 0; i < 4; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vp10_highbd_idct8_c(input, outptr, bd);
+      input += 8;
+      outptr += 8;
+    }
+  }
+
+  if (optimised_cols) {
+    vp10_idct8_sse2(inptr);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[8];
+      for (i = 0; i < 8; i++) {
+        inptr[i] = _mm_add_epi16(inptr[i], sixteen);
+        d[i] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        inptr[i] = _mm_srai_epi16(inptr[i], 5);
+        d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[i]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[8], temp_out[8];
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j * 8 + i];
+      vp10_highbd_idct8_c(temp_in, temp_out, bd);
+      for (j = 0; j < 8; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+      }
+    }
+  }
+}
+
+void vp10_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                       int stride, int bd) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[32];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i rounding = _mm_set1_epi16(32);
+  const __m128i max = _mm_set1_epi16(3155);
+  const __m128i min = _mm_set1_epi16(-3155);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 16; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
+    inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 32; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    vp10_idct16_sse2(inptr, inptr + 16);
+
+    // Find the min & max for the column transform
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 32; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      array_transpose_16x16(inptr, inptr + 16);
+      for (i = 0; i < 16; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
+        sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 16; ++i) {
+      vp10_highbd_idct16_c(input, outptr, bd);
+      input += 16;
+      outptr += 16;
+    }
+  }
+
+  if (optimised_cols) {
+    vp10_idct16_sse2(inptr, inptr + 16);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[2];
+      for (i = 0; i < 16; i++) {
+        inptr[i   ] = _mm_add_epi16(inptr[i   ], rounding);
+        inptr[i+16] = _mm_add_epi16(inptr[i+16], rounding);
+        d[0] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        d[1] = _mm_loadu_si128((const __m128i *)(dest + stride*i + 8));
+        inptr[i   ] = _mm_srai_epi16(inptr[i   ], 6);
+        inptr[i+16] = _mm_srai_epi16(inptr[i+16], 6);
+        d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i   ]), bd);
+        d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i+16]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[0]);
+        _mm_storeu_si128((__m128i *)(dest + stride*i + 8), d[1]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[16], temp_out[16];
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j * 16 + i];
+      vp10_highbd_idct16_c(temp_in, temp_out, bd);
+      for (j = 0; j < 16; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+      }
+    }
+  }
+}
+
+void vp10_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                      int stride, int bd) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[32];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i rounding = _mm_set1_epi16(32);
+  const __m128i max = _mm_set1_epi16(3155);
+  const __m128i min = _mm_set1_epi16(-3155);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 16; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
+    inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  // Since all non-zero dct coefficients are in upper-left 4x4 area,
+  // we only need to consider first 4 rows here.
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 4; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform (N.B. This transposes inptr)
+    vp10_idct16_sse2(inptr, inptr + 16);
+
+    // Find the min & max for the column transform
+    // N.B. Only first 4 cols contain non-zero coeffs
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 16; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      // Use fact only first 4 rows contain non-zero coeffs
+      array_transpose_8x8(inptr, inptr);
+      array_transpose_8x8(inptr + 8, inptr + 16);
+      for (i = 0; i < 4; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
+        sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vp10_highbd_idct16_c(input, outptr, bd);
+      input += 16;
+      outptr += 16;
+    }
+  }
+
+  if (optimised_cols) {
+    vp10_idct16_sse2(inptr, inptr + 16);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[2];
+      for (i = 0; i < 16; i++) {
+        inptr[i   ] = _mm_add_epi16(inptr[i   ], rounding);
+        inptr[i+16] = _mm_add_epi16(inptr[i+16], rounding);
+        d[0] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        d[1] = _mm_loadu_si128((const __m128i *)(dest + stride*i + 8));
+        inptr[i   ] = _mm_srai_epi16(inptr[i   ], 6);
+        inptr[i+16] = _mm_srai_epi16(inptr[i+16], 6);
+        d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i   ]), bd);
+        d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i+16]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[0]);
+        _mm_storeu_si128((__m128i *)(dest + stride*i + 8), d[1]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[16], temp_out[16];
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j * 16 + i];
+      vp10_highbd_idct16_c(temp_in, temp_out, bd);
+      for (j = 0; j < 16; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+      }
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH

diff --git a/vp10/common/x86/vp10_inv_txfm_sse2.h b/vp10/common/x86/vp10_inv_txfm_sse2.h
new file mode 100644 (file)
index 0000000..b79781a
--- /dev/null
+++ b/vp10/common/x86/vp10_inv_txfm_sse2.h
@@ -0,0 +1,184 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_X86_INV_TXFM_SSE2_H_
+#define VPX_DSP_X86_INV_TXFM_SSE2_H_
+
+#include <emmintrin.h>  // SSE2
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vp10/common/vp10_inv_txfm.h"
+
+// perform 8x8 transpose
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+}
+
+#define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+                                                        \
+    in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);  /* i1 i0 */  \
+    in1 = _mm_unpackhi_epi32(tr0_0, tr0_1);  /* i3 i2 */  \
+  }
+
+static INLINE void array_transpose_4X8(__m128i *in, __m128i * out) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+
+  out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+static INLINE void load_buffer_8x16(const int16_t *input, __m128i *in) {
+  in[0]  = _mm_load_si128((const __m128i *)(input + 0 * 16));
+  in[1]  = _mm_load_si128((const __m128i *)(input + 1 * 16));
+  in[2]  = _mm_load_si128((const __m128i *)(input + 2 * 16));
+  in[3]  = _mm_load_si128((const __m128i *)(input + 3 * 16));
+  in[4]  = _mm_load_si128((const __m128i *)(input + 4 * 16));
+  in[5]  = _mm_load_si128((const __m128i *)(input + 5 * 16));
+  in[6]  = _mm_load_si128((const __m128i *)(input + 6 * 16));
+  in[7]  = _mm_load_si128((const __m128i *)(input + 7 * 16));
+
+  in[8]  = _mm_load_si128((const __m128i *)(input + 8 * 16));
+  in[9]  = _mm_load_si128((const __m128i *)(input + 9 * 16));
+  in[10]  = _mm_load_si128((const __m128i *)(input + 10 * 16));
+  in[11]  = _mm_load_si128((const __m128i *)(input + 11 * 16));
+  in[12]  = _mm_load_si128((const __m128i *)(input + 12 * 16));
+  in[13]  = _mm_load_si128((const __m128i *)(input + 13 * 16));
+  in[14]  = _mm_load_si128((const __m128i *)(input + 14 * 16));
+  in[15]  = _mm_load_si128((const __m128i *)(input + 15 * 16));
+}
+
+#define RECON_AND_STORE(dest, in_x) \
+  {                                                     \
+     __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
+      d0 = _mm_unpacklo_epi8(d0, zero); \
+      d0 = _mm_add_epi16(in_x, d0); \
+      d0 = _mm_packus_epi16(d0, d0); \
+      _mm_storel_epi64((__m128i *)(dest), d0); \
+  }
+
+static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+  in[8] = _mm_adds_epi16(in[8], final_rounding);
+  in[9] = _mm_adds_epi16(in[9], final_rounding);
+  in[10] = _mm_adds_epi16(in[10], final_rounding);
+  in[11] = _mm_adds_epi16(in[11], final_rounding);
+  in[12] = _mm_adds_epi16(in[12], final_rounding);
+  in[13] = _mm_adds_epi16(in[13], final_rounding);
+  in[14] = _mm_adds_epi16(in[14], final_rounding);
+  in[15] = _mm_adds_epi16(in[15], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 6);
+  in[1] = _mm_srai_epi16(in[1], 6);
+  in[2] = _mm_srai_epi16(in[2], 6);
+  in[3] = _mm_srai_epi16(in[3], 6);
+  in[4] = _mm_srai_epi16(in[4], 6);
+  in[5] = _mm_srai_epi16(in[5], 6);
+  in[6] = _mm_srai_epi16(in[6], 6);
+  in[7] = _mm_srai_epi16(in[7], 6);
+  in[8] = _mm_srai_epi16(in[8], 6);
+  in[9] = _mm_srai_epi16(in[9], 6);
+  in[10] = _mm_srai_epi16(in[10], 6);
+  in[11] = _mm_srai_epi16(in[11], 6);
+  in[12] = _mm_srai_epi16(in[12], 6);
+  in[13] = _mm_srai_epi16(in[13], 6);
+  in[14] = _mm_srai_epi16(in[14], 6);
+  in[15] = _mm_srai_epi16(in[15], 6);
+
+  RECON_AND_STORE(dest +  0 * stride, in[0]);
+  RECON_AND_STORE(dest +  1 * stride, in[1]);
+  RECON_AND_STORE(dest +  2 * stride, in[2]);
+  RECON_AND_STORE(dest +  3 * stride, in[3]);
+  RECON_AND_STORE(dest +  4 * stride, in[4]);
+  RECON_AND_STORE(dest +  5 * stride, in[5]);
+  RECON_AND_STORE(dest +  6 * stride, in[6]);
+  RECON_AND_STORE(dest +  7 * stride, in[7]);
+  RECON_AND_STORE(dest +  8 * stride, in[8]);
+  RECON_AND_STORE(dest +  9 * stride, in[9]);
+  RECON_AND_STORE(dest + 10 * stride, in[10]);
+  RECON_AND_STORE(dest + 11 * stride, in[11]);
+  RECON_AND_STORE(dest + 12 * stride, in[12]);
+  RECON_AND_STORE(dest + 13 * stride, in[13]);
+  RECON_AND_STORE(dest + 14 * stride, in[14]);
+  RECON_AND_STORE(dest + 15 * stride, in[15]);
+}
+
+void idct4_sse2(__m128i *in);
+void idct8_sse2(__m128i *in);
+void idct16_sse2(__m128i *in0, __m128i *in1);
+void iadst4_sse2(__m128i *in);
+void iadst8_sse2(__m128i *in);
+void iadst16_sse2(__m128i *in0, __m128i *in1);
+
+#endif  // VPX_DSP_X86_INV_TXFM_SSE2_H_

diff --git a/vp10/vp10_common.mk b/vp10/vp10_common.mk
index 16a723c..dc44507 100644 (file)
--- a/vp10/vp10_common.mk
+++ b/vp10/vp10_common.mk
@@ -19,7 +19,6 @@ VP10_COMMON_SRCS-yes += common/entropymode.c
 VP10_COMMON_SRCS-yes += common/entropymv.c
 VP10_COMMON_SRCS-yes += common/frame_buffers.c
 VP10_COMMON_SRCS-yes += common/frame_buffers.h
-VP10_COMMON_SRCS-yes += common/idct.c
 VP10_COMMON_SRCS-yes += common/alloccommon.h
 VP10_COMMON_SRCS-yes += common/blockd.h
 VP10_COMMON_SRCS-yes += common/common.h
@@ -30,6 +29,9 @@ VP10_COMMON_SRCS-yes += common/enums.h
 VP10_COMMON_SRCS-yes += common/filter.h
 VP10_COMMON_SRCS-yes += common/filter.c
 VP10_COMMON_SRCS-yes += common/idct.h
+VP10_COMMON_SRCS-yes += common/idct.c
+VP10_COMMON_SRCS-yes += common/vp10_inv_txfm.h
+VP10_COMMON_SRCS-yes += common/vp10_inv_txfm.c
 VP10_COMMON_SRCS-yes += common/loopfilter.h
 VP10_COMMON_SRCS-yes += common/thread_common.h
 VP10_COMMON_SRCS-yes += common/mv.h
@@ -91,4 +93,7 @@ VP10_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/iht4x4_add_neon.c
 VP10_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/iht8x8_add_neon.c
 endif
 
+VP10_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp10_inv_txfm_sse2.c
+VP10_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp10_inv_txfm_sse2.h
+
 $(eval $(call rtcd_h_template,vp10_rtcd,vp10/common/vp10_rtcd_defs.pl))