Refactor highbd idct 4x4 sse4.1 code and add highbd_inv_txfm_sse4.h

Also clean highbd_inv_txfm_sse2.h

BUG=webm:1412

Change-Id: I0722841d824ce602874019bd9779b10d49d10c0b

8 files changed, 139 insertions, 114 deletions

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index e8b89436b..badebc650 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -240,6 +240,7 @@ DSP_SRCS-$(HAVE_SSE2)  += x86/highbd_idct4x4_add_sse2.c
 DSP_SRCS-$(HAVE_SSE2)  += x86/highbd_idct8x8_add_sse2.c
 DSP_SRCS-$(HAVE_SSE2)  += x86/highbd_idct16x16_add_sse2.c
 DSP_SRCS-$(HAVE_SSE2)  += x86/highbd_idct32x32_add_sse2.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_inv_txfm_sse4.h
 DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_idct4x4_add_sse4.c
 endif  # !CONFIG_VP9_HIGHBITDEPTH
 
diff --git a/vpx_dsp/x86/highbd_idct16x16_add_sse2.c b/vpx_dsp/x86/highbd_idct16x16_add_sse2.c
index 1df91f08f..154f11ba3 100644
--- a/vpx_dsp/x86/highbd_idct16x16_add_sse2.c
+++ b/vpx_dsp/x86/highbd_idct16x16_add_sse2.c
@@ -105,8 +105,8 @@ void vpx_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint16_t *dest,
         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);
+        d[0] = add_clamp(d[0], inptr[i], bd);
+        d[1] = add_clamp(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]);
@@ -222,8 +222,8 @@ void vpx_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint16_t *dest,
         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);
+        d[0] = add_clamp(d[0], inptr[i], bd);
+        d[1] = add_clamp(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]);
diff --git a/vpx_dsp/x86/highbd_idct4x4_add_sse2.c b/vpx_dsp/x86/highbd_idct4x4_add_sse2.c
index ac6f73d83..1eb3a8347 100644
--- a/vpx_dsp/x86/highbd_idct4x4_add_sse2.c
+++ b/vpx_dsp/x86/highbd_idct4x4_add_sse2.c
@@ -191,15 +191,11 @@ void vpx_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint16_t *dest,
     io[1] = wraplow_16bit(io[2], io[3], _mm_set1_epi32(8));
   }
 
-  recon_and_store_4(dest, io, stride, bd);
+  recon_and_store_4(io, dest, stride, bd);
 }
 
 void vpx_highbd_idct4x4_1_add_sse2(const tran_low_t *input, uint16_t *dest,
                                    int stride, int bd) {
-  const __m128i zero = _mm_setzero_si128();
-  // Faster than _mm_set1_epi16((1 << bd) - 1).
-  const __m128i one = _mm_set1_epi16(1);
-  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
   int a1, i;
   tran_low_t out;
   __m128i dc, d;
@@ -211,7 +207,7 @@ void vpx_highbd_idct4x4_1_add_sse2(const tran_low_t *input, uint16_t *dest,
 
   for (i = 0; i < 4; ++i) {
     d = _mm_loadl_epi64((const __m128i *)dest);
-    d = add_dc_clamp(&zero, &max, &dc, &d);
+    d = add_clamp(d, dc, bd);
     _mm_storel_epi64((__m128i *)dest, d);
     dest += stride;
   }
diff --git a/vpx_dsp/x86/highbd_idct4x4_add_sse4.c b/vpx_dsp/x86/highbd_idct4x4_add_sse4.c
index 9d1c6f4b7..fc522c1c0 100644
--- a/vpx_dsp/x86/highbd_idct4x4_add_sse4.c
+++ b/vpx_dsp/x86/highbd_idct4x4_add_sse4.c
@@ -12,15 +12,10 @@
 
 #include "./vpx_dsp_rtcd.h"
 #include "vpx_dsp/x86/highbd_inv_txfm_sse2.h"
+#include "vpx_dsp/x86/highbd_inv_txfm_sse4.h"
 #include "vpx_dsp/x86/inv_txfm_sse2.h"
 #include "vpx_dsp/x86/transpose_sse2.h"
 
-static INLINE void extend_64bit(const __m128i in,
-                                __m128i *const out /*out[2]*/) {
-  out[0] = _mm_unpacklo_epi32(in, in);  // 0, 0, 1, 1
-  out[1] = _mm_unpackhi_epi32(in, in);  // 2, 2, 3, 3
-}
-
 static INLINE void highbd_idct4(__m128i *const io) {
   const __m128i cospi_p16_p16 =
       _mm_setr_epi32(cospi_16_64 << 2, 0, cospi_16_64 << 2, 0);
@@ -28,46 +23,19 @@ static INLINE void highbd_idct4(__m128i *const io) {
       _mm_setr_epi32(cospi_8_64 << 2, 0, cospi_8_64 << 2, 0);
   const __m128i cospi_p24_p24 =
       _mm_setr_epi32(cospi_24_64 << 2, 0, cospi_24_64 << 2, 0);
-  __m128i temp1[4], temp2[4], step[4];
+  __m128i temp1[4], step[4];
 
   transpose_32bit_4x4(&io[0], &io[1], &io[2], &io[3]);
 
   // stage 1
   temp1[0] = _mm_add_epi32(io[0], io[2]);  // input[0] + input[2]
-  temp2[0] = _mm_sub_epi32(io[0], io[2]);  // input[0] - input[2]
   extend_64bit(temp1[0], temp1);
-  extend_64bit(temp2[0], temp2);
-  temp1[0] = _mm_mul_epi32(temp1[0], cospi_p16_p16);
-  temp1[1] = _mm_mul_epi32(temp1[1], cospi_p16_p16);
-  temp2[0] = _mm_mul_epi32(temp2[0], cospi_p16_p16);
-  temp2[1] = _mm_mul_epi32(temp2[1], cospi_p16_p16);
-  temp1[0] = dct_const_round_shift_64bit(temp1[0]);
-  temp1[1] = dct_const_round_shift_64bit(temp1[1]);
-  temp2[0] = dct_const_round_shift_64bit(temp2[0]);
-  temp2[1] = dct_const_round_shift_64bit(temp2[1]);
-  step[0] = pack_4(temp1[0], temp1[1]);
-  step[1] = pack_4(temp2[0], temp2[1]);
-
-  extend_64bit(io[1], temp1);
-  extend_64bit(io[3], temp2);
-  temp1[2] = _mm_mul_epi32(temp1[0], cospi_p08_p08);
-  temp1[3] = _mm_mul_epi32(temp1[1], cospi_p08_p08);
-  temp1[0] = _mm_mul_epi32(temp1[0], cospi_p24_p24);
-  temp1[1] = _mm_mul_epi32(temp1[1], cospi_p24_p24);
-  temp2[2] = _mm_mul_epi32(temp2[0], cospi_p24_p24);
-  temp2[3] = _mm_mul_epi32(temp2[1], cospi_p24_p24);
-  temp2[0] = _mm_mul_epi32(temp2[0], cospi_p08_p08);
-  temp2[1] = _mm_mul_epi32(temp2[1], cospi_p08_p08);
-  temp1[0] = _mm_sub_epi64(temp1[0], temp2[0]);  // [1]*cospi_24 - [3]*cospi_8
-  temp1[1] = _mm_sub_epi64(temp1[1], temp2[1]);  // [1]*cospi_24 - [3]*cospi_8
-  temp2[0] = _mm_add_epi64(temp1[2], temp2[2]);  // [1]*cospi_8 + [3]*cospi_24
-  temp2[1] = _mm_add_epi64(temp1[3], temp2[3]);  // [1]*cospi_8 + [3]*cospi_24
-  temp1[0] = dct_const_round_shift_64bit(temp1[0]);
-  temp1[1] = dct_const_round_shift_64bit(temp1[1]);
-  temp2[0] = dct_const_round_shift_64bit(temp2[0]);
-  temp2[1] = dct_const_round_shift_64bit(temp2[1]);
-  step[2] = pack_4(temp1[0], temp1[1]);
-  step[3] = pack_4(temp2[0], temp2[1]);
+  step[0] = multiplication_round_shift(temp1, cospi_p16_p16);
+  temp1[0] = _mm_sub_epi32(io[0], io[2]);  // input[0] - input[2]
+  extend_64bit(temp1[0], temp1);
+  step[1] = multiplication_round_shift(temp1, cospi_p16_p16);
+  multiplication_and_add_2_ssse4_1(&io[1], &io[3], &cospi_p24_p24,
+                                   &cospi_p08_p08, &step[2], &step[3]);
 
   // stage 2
   io[0] = _mm_add_epi32(step[0], step[3]);  // step[0] + step[3]
@@ -103,5 +71,5 @@ void vpx_highbd_idct4x4_16_add_sse4_1(const tran_low_t *input, uint16_t *dest,
     io[1] = wraplow_16bit(io[2], io[3], _mm_set1_epi32(8));
   }
 
-  recon_and_store_4(dest, io, stride, bd);
+  recon_and_store_4(io, dest, stride, bd);
 }
diff --git a/vpx_dsp/x86/highbd_idct8x8_add_sse2.c b/vpx_dsp/x86/highbd_idct8x8_add_sse2.c
index c12e3e1b9..8eae17581 100644
--- a/vpx_dsp/x86/highbd_idct8x8_add_sse2.c
+++ b/vpx_dsp/x86/highbd_idct8x8_add_sse2.c
@@ -94,7 +94,7 @@ void vpx_highbd_idct8x8_64_add_sse2(const tran_low_t *input, uint16_t *dest,
         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);
+        d[i] = add_clamp(d[i], inptr[i], bd);
         // Store
         _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]);
       }
@@ -196,7 +196,7 @@ void vpx_highbd_idct8x8_12_add_sse2(const tran_low_t *input, uint16_t *dest,
         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);
+        d[i] = add_clamp(d[i], inptr[i], bd);
         // Store
         _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]);
       }
diff --git a/vpx_dsp/x86/highbd_inv_txfm_sse2.h b/vpx_dsp/x86/highbd_inv_txfm_sse2.h
index be740a8c7..5bdabc4d6 100644
--- a/vpx_dsp/x86/highbd_inv_txfm_sse2.h
+++ b/vpx_dsp/x86/highbd_inv_txfm_sse2.h
@@ -17,6 +17,12 @@
 #include "vpx_dsp/inv_txfm.h"
 #include "vpx_dsp/x86/txfm_common_sse2.h"
 
+static INLINE void extend_64bit(const __m128i in,
+                                __m128i *const out /*out[2]*/) {
+  out[0] = _mm_unpacklo_epi32(in, in);  // 0, 0, 1, 1
+  out[1] = _mm_unpackhi_epi32(in, in);  // 2, 2, 3, 3
+}
+
 static INLINE __m128i wraplow_16bit(const __m128i in0, const __m128i in1,
                                     const __m128i rounding) {
   __m128i temp[2];
@@ -40,24 +46,24 @@ static INLINE __m128i pack_4(const __m128i in0, const __m128i in1) {
   return _mm_unpacklo_epi32(t0, t1);                // 0, 1, 2, 3
 }
 
-static INLINE __m128i add_dc_clamp(const __m128i *const min,
-                                   const __m128i *const max,
-                                   const __m128i *const dc,
-                                   const __m128i *const in) {
-  __m128i out;
-  out = _mm_adds_epi16(*in, *dc);
-  out = _mm_max_epi16(out, *min);
-  out = _mm_min_epi16(out, *max);
-  return out;
+static INLINE __m128i add_clamp(const __m128i in0, const __m128i in1,
+                                const int bd) {
+  const __m128i zero = _mm_set1_epi16(0);
+  // Faster than _mm_set1_epi16((1 << bd) - 1).
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
+  __m128i d;
+
+  d = _mm_adds_epi16(in0, in1);
+  d = _mm_max_epi16(d, zero);
+  d = _mm_min_epi16(d, max);
+
+  return d;
 }
 
 static INLINE void highbd_idct_1_add_kernel(const tran_low_t *input,
                                             uint16_t *dest, int stride, int bd,
                                             const int size) {
-  const __m128i zero = _mm_setzero_si128();
-  // Faster than _mm_set1_epi16((1 << bd) - 1).
-  const __m128i one = _mm_set1_epi16(1);
-  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
   int a1, i, j;
   tran_low_t out;
   __m128i dc, d;
@@ -70,43 +76,31 @@ static INLINE void highbd_idct_1_add_kernel(const tran_low_t *input,
   for (i = 0; i < size; ++i) {
     for (j = 0; j < (size >> 3); ++j) {
       d = _mm_load_si128((const __m128i *)(&dest[j * 8]));
-      d = add_dc_clamp(&zero, &max, &dc, &d);
+      d = add_clamp(d, dc, bd);
       _mm_store_si128((__m128i *)(&dest[j * 8]), d);
     }
     dest += stride;
   }
 }
 
-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_sub_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;
+static INLINE void recon_and_store_4_dual(const __m128i in,
+                                          uint16_t *const dest,
+                                          const int stride, const int bd) {
+  __m128i d;
+
+  d = _mm_loadl_epi64((const __m128i *)(dest + 0 * stride));
+  d = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(d), (const __m64 *)(dest + 1 * stride)));
+  d = add_clamp(d, in, bd);
+  _mm_storel_epi64((__m128i *)(dest + 0 * stride), d);
+  _mm_storeh_pi((__m64 *)(dest + 1 * stride), _mm_castsi128_ps(d));
 }
 
-static INLINE void recon_and_store_4(uint16_t *const dest,
-                                     const __m128i *const io, const int stride,
-                                     int bd) {
-  __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, io[0]), bd);
-  d2 = clamp_high_sse2(_mm_adds_epi16(d2, io[1]), bd);
-  _mm_storel_epi64((__m128i *)dest, d0);
-  d0 = _mm_srli_si128(d0, 8);
-  _mm_storel_epi64((__m128i *)(dest + stride), d0);
-  _mm_storel_epi64((__m128i *)(dest + stride * 2), d2);
-  d2 = _mm_srli_si128(d2, 8);
-  _mm_storel_epi64((__m128i *)(dest + stride * 3), d2);
+static INLINE void recon_and_store_4(const __m128i *const in, uint16_t *dest,
+                                     const int stride, const int bd) {
+  recon_and_store_4_dual(in[0], dest, stride, bd);
+  dest += 2 * stride;
+  recon_and_store_4_dual(in[1], dest, stride, bd);
 }
 
 #endif  // VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_
diff --git a/vpx_dsp/x86/highbd_inv_txfm_sse4.h b/vpx_dsp/x86/highbd_inv_txfm_sse4.h
new file mode 100644
index 000000000..72d3d5327
--- /dev/null
+++ b/vpx_dsp/x86/highbd_inv_txfm_sse4.h
@@ -0,0 +1,60 @@
+/*
+ *  Copyright (c) 2017 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_HIGHBD_INV_TXFM_SSE4_H_
+#define VPX_DSP_X86_HIGHBD_INV_TXFM_SSE4_H_
+
+#include <smmintrin.h>  // SSE4.1
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+
+static INLINE __m128i multiplication_round_shift(const __m128i *const in,
+                                                 const __m128i cospi) {
+  __m128i t0, t1;
+  t0 = _mm_mul_epi32(in[0], cospi);
+  t1 = _mm_mul_epi32(in[1], cospi);
+  t0 = dct_const_round_shift_64bit(t0);
+  t1 = dct_const_round_shift_64bit(t1);
+  return pack_4(t0, t1);
+}
+
+static INLINE void multiplication_and_add_2_ssse4_1(const __m128i *const in0,
+                                                    const __m128i *const in1,
+                                                    const __m128i *const cst0,
+                                                    const __m128i *const cst1,
+                                                    __m128i *const out0,
+                                                    __m128i *const out1) {
+  __m128i temp1[4], temp2[4];
+  extend_64bit(*in0, temp1);
+  extend_64bit(*in1, temp2);
+  temp1[2] = _mm_mul_epi32(temp1[0], *cst1);
+  temp1[3] = _mm_mul_epi32(temp1[1], *cst1);
+  temp1[0] = _mm_mul_epi32(temp1[0], *cst0);
+  temp1[1] = _mm_mul_epi32(temp1[1], *cst0);
+  temp2[2] = _mm_mul_epi32(temp2[0], *cst0);
+  temp2[3] = _mm_mul_epi32(temp2[1], *cst0);
+  temp2[0] = _mm_mul_epi32(temp2[0], *cst1);
+  temp2[1] = _mm_mul_epi32(temp2[1], *cst1);
+  temp1[0] = _mm_sub_epi64(temp1[0], temp2[0]);
+  temp1[1] = _mm_sub_epi64(temp1[1], temp2[1]);
+  temp2[0] = _mm_add_epi64(temp1[2], temp2[2]);
+  temp2[1] = _mm_add_epi64(temp1[3], temp2[3]);
+  temp1[0] = dct_const_round_shift_64bit(temp1[0]);
+  temp1[1] = dct_const_round_shift_64bit(temp1[1]);
+  temp2[0] = dct_const_round_shift_64bit(temp2[0]);
+  temp2[1] = dct_const_round_shift_64bit(temp2[1]);
+  *out0 = pack_4(temp1[0], temp1[1]);
+  *out1 = pack_4(temp2[0], temp2[1]);
+}
+
+#endif  // VPX_DSP_X86_HIGHBD_INV_TXFM_SSE4_H_
diff --git a/vpx_dsp/x86/inv_txfm_sse2.h b/vpx_dsp/x86/inv_txfm_sse2.h
index e6e1cd403..1499b59e5 100644
--- a/vpx_dsp/x86/inv_txfm_sse2.h
+++ b/vpx_dsp/x86/inv_txfm_sse2.h
@@ -152,28 +152,34 @@ static INLINE void recon_and_store(uint8_t *const dest, const __m128i in_x) {
   _mm_storel_epi64((__m128i *)(dest), d0);
 }
 
+static INLINE void round_shift_8x8(const __m128i *const in,
+                                   __m128i *const out) {
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+
+  out[0] = _mm_add_epi16(in[0], final_rounding);
+  out[1] = _mm_add_epi16(in[1], final_rounding);
+  out[2] = _mm_add_epi16(in[2], final_rounding);
+  out[3] = _mm_add_epi16(in[3], final_rounding);
+  out[4] = _mm_add_epi16(in[4], final_rounding);
+  out[5] = _mm_add_epi16(in[5], final_rounding);
+  out[6] = _mm_add_epi16(in[6], final_rounding);
+  out[7] = _mm_add_epi16(in[7], final_rounding);
+
+  out[0] = _mm_srai_epi16(out[0], 5);
+  out[1] = _mm_srai_epi16(out[1], 5);
+  out[2] = _mm_srai_epi16(out[2], 5);
+  out[3] = _mm_srai_epi16(out[3], 5);
+  out[4] = _mm_srai_epi16(out[4], 5);
+  out[5] = _mm_srai_epi16(out[5], 5);
+  out[6] = _mm_srai_epi16(out[6], 5);
+  out[7] = _mm_srai_epi16(out[7], 5);
+}
+
 static INLINE void write_buffer_8x8(const __m128i *const in,
                                     uint8_t *const dest, const int stride) {
-  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
   __m128i t[8];
-  // Final rounding and shift
-  t[0] = _mm_adds_epi16(in[0], final_rounding);
-  t[1] = _mm_adds_epi16(in[1], final_rounding);
-  t[2] = _mm_adds_epi16(in[2], final_rounding);
-  t[3] = _mm_adds_epi16(in[3], final_rounding);
-  t[4] = _mm_adds_epi16(in[4], final_rounding);
-  t[5] = _mm_adds_epi16(in[5], final_rounding);
-  t[6] = _mm_adds_epi16(in[6], final_rounding);
-  t[7] = _mm_adds_epi16(in[7], final_rounding);
-
-  t[0] = _mm_srai_epi16(t[0], 5);
-  t[1] = _mm_srai_epi16(t[1], 5);
-  t[2] = _mm_srai_epi16(t[2], 5);
-  t[3] = _mm_srai_epi16(t[3], 5);
-  t[4] = _mm_srai_epi16(t[4], 5);
-  t[5] = _mm_srai_epi16(t[5], 5);
-  t[6] = _mm_srai_epi16(t[6], 5);
-  t[7] = _mm_srai_epi16(t[7], 5);
+
+  round_shift_8x8(in, t);
 
   recon_and_store(dest + 0 * stride, t[0]);
   recon_and_store(dest + 1 * stride, t[1]);