Merge "Optimize 32x32 idct function" into experimental

1 files changed, 690 insertions, 0 deletions

diff --git a/vp9/common/x86/vp9_idct_x86.c b/vp9/common/x86/vp9_idct_x86.c
index 1a2c84a40..c8a3873f7 100644
--- a/vp9/common/x86/vp9_idct_x86.c
+++ b/vp9/common/x86/vp9_idct_x86.c
@@ -1506,4 +1506,694 @@ void vp9_short_idct10_16x16_sse2(int16_t *input, int16_t *output, int pitch) {
     output += 8;
   }
 }
+
+// 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); \
+  }
+
+void vp9_short_idct32x32_sse2(int16_t *input, int16_t *output, int pitch) {
+  const int half_pitch = pitch >> 1;
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+
+  // 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 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11, in12,
+          in13, in14, in15, in16, in17, in18, in19, in20, in21, in22, in23,
+          in24, in25, in26, in27, in28, in29, in30, in31;
+  __m128i col[128];
+  __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;
+
+  // We work on a 8x32 block each time, and loop 8 times for 2-D 32x32 idct.
+  for (i = 0; i < 8; i++) {
+    if (i < 4) {
+      // First 1-D idct
+      // Load input data.
+      in0 = _mm_load_si128((__m128i *)input);
+      in8 = _mm_load_si128((__m128i *)(input + 8 * 1));
+      in16 = _mm_load_si128((__m128i *)(input + 8 * 2));
+      in24 = _mm_load_si128((__m128i *)(input + 8 * 3));
+      in1 = _mm_load_si128((__m128i *)(input + 8 * 4));
+      in9 = _mm_load_si128((__m128i *)(input + 8 * 5));
+      in17 = _mm_load_si128((__m128i *)(input + 8 * 6));
+      in25 = _mm_load_si128((__m128i *)(input + 8 * 7));
+      in2 = _mm_load_si128((__m128i *)(input + 8 * 8));
+      in10 = _mm_load_si128((__m128i *)(input + 8 * 9));
+      in18 = _mm_load_si128((__m128i *)(input + 8 * 10));
+      in26 = _mm_load_si128((__m128i *)(input + 8 * 11));
+      in3 = _mm_load_si128((__m128i *)(input + 8 * 12));
+      in11 = _mm_load_si128((__m128i *)(input + 8 * 13));
+      in19 = _mm_load_si128((__m128i *)(input + 8 * 14));
+      in27 = _mm_load_si128((__m128i *)(input + 8 * 15));
+
+      in4 = _mm_load_si128((__m128i *)(input + 8 * 16));
+      in12 = _mm_load_si128((__m128i *)(input + 8 * 17));
+      in20 = _mm_load_si128((__m128i *)(input + 8 * 18));
+      in28 = _mm_load_si128((__m128i *)(input + 8 * 19));
+      in5 = _mm_load_si128((__m128i *)(input + 8 * 20));
+      in13 = _mm_load_si128((__m128i *)(input + 8 * 21));
+      in21 = _mm_load_si128((__m128i *)(input + 8 * 22));
+      in29 = _mm_load_si128((__m128i *)(input + 8 * 23));
+      in6 = _mm_load_si128((__m128i *)(input + 8 * 24));
+      in14 = _mm_load_si128((__m128i *)(input + 8 * 25));
+      in22 = _mm_load_si128((__m128i *)(input + 8 * 26));
+      in30 = _mm_load_si128((__m128i *)(input + 8 * 27));
+      in7 = _mm_load_si128((__m128i *)(input + 8 * 28));
+      in15 = _mm_load_si128((__m128i *)(input + 8 * 29));
+      in23 = _mm_load_si128((__m128i *)(input + 8 * 30));
+      in31 = _mm_load_si128((__m128i *)(input + 8 * 31));
+
+      input += 256;
+
+      // Transpose 32x8 block to 8x32 block
+      TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                    in4, in5, in6, in7);
+      TRANSPOSE_8X8(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,
+                    in10, in11, in12, in13, in14, in15);
+      TRANSPOSE_8X8(in16, in17, in18, in19, in20, in21, in22, in23, in16, in17,
+                    in18, in19, in20, in21, in22, in23);
+      TRANSPOSE_8X8(in24, in25, in26, in27, in28, in29, in30, in31, in24, in25,
+                    in26, in27, in28, in29, in30, in31);
+    } else {
+      // Second 1-D idct
+      j = i - 4;
+
+      // Transpose 32x8 block to 8x32 block
+      TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2],
+                    col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5],
+                    col[j * 8 + 6], col[j * 8 + 7], in0, in1, in2, in3, in4,
+                    in5, in6, in7);
+      j += 4;
+      TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2],
+                    col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5],
+                    col[j * 8 + 6], col[j * 8 + 7], in8, in9, in10,
+                    in11, in12, in13, in14, in15);
+      j += 4;
+      TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2],
+                    col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5],
+                    col[j * 8 + 6], col[j * 8 + 7], in16, in17, in18,
+                    in19, in20, in21, in22, in23);
+      j += 4;
+      TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2],
+                    col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5],
+                    col[j * 8 + 6], col[j * 8 + 7], in24, in25, in26, in27,
+                    in28, in29, in30, in31);
+    }
+
+    // Stage1
+    {
+      const __m128i lo_1_31 = _mm_unpacklo_epi16(in1, in31);
+      const __m128i hi_1_31 = _mm_unpackhi_epi16(in1, in31);
+      const __m128i lo_17_15 = _mm_unpacklo_epi16(in17, in15);
+      const __m128i hi_17_15 = _mm_unpackhi_epi16(in17, in15);
+
+      const __m128i lo_9_23 = _mm_unpacklo_epi16(in9, in23);
+      const __m128i hi_9_23 = _mm_unpackhi_epi16(in9, in23);
+      const __m128i lo_25_7= _mm_unpacklo_epi16(in25, in7);
+      const __m128i hi_25_7 = _mm_unpackhi_epi16(in25, in7);
+
+      const __m128i lo_5_27 = _mm_unpacklo_epi16(in5, in27);
+      const __m128i hi_5_27 = _mm_unpackhi_epi16(in5, in27);
+      const __m128i lo_21_11 = _mm_unpacklo_epi16(in21, in11);
+      const __m128i hi_21_11 = _mm_unpackhi_epi16(in21, in11);
+
+      const __m128i lo_13_19 = _mm_unpacklo_epi16(in13, in19);
+      const __m128i hi_13_19 = _mm_unpackhi_epi16(in13, in19);
+      const __m128i lo_29_3 = _mm_unpacklo_epi16(in29, in3);
+      const __m128i hi_29_3 = _mm_unpackhi_epi16(in29, in3);
+
+      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(in2, in30);
+      const __m128i hi_2_30 = _mm_unpackhi_epi16(in2, in30);
+      const __m128i lo_18_14 = _mm_unpacklo_epi16(in18, in14);
+      const __m128i hi_18_14 = _mm_unpackhi_epi16(in18, in14);
+
+      const __m128i lo_10_22 = _mm_unpacklo_epi16(in10, in22);
+      const __m128i hi_10_22 = _mm_unpackhi_epi16(in10, in22);
+      const __m128i lo_26_6 = _mm_unpacklo_epi16(in26, in6);
+      const __m128i hi_26_6 = _mm_unpackhi_epi16(in26, in6);
+
+      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(in4, in28);
+      const __m128i hi_4_28 = _mm_unpackhi_epi16(in4, in28);
+      const __m128i lo_20_12 = _mm_unpacklo_epi16(in20, in12);
+      const __m128i hi_20_12 = _mm_unpackhi_epi16(in20, in12);
+
+      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(in0, in16);
+      const __m128i hi_0_16 = _mm_unpackhi_epi16(in0, in16);
+      const __m128i lo_8_24 = _mm_unpacklo_epi16(in8, in24);
+      const __m128i hi_8_24 = _mm_unpackhi_epi16(in8, in24);
+
+      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;
+    }
+
+    // final stage
+    if (i < 4) {
+      // 1_D: Store 32 intermediate results for each 8x32 block.
+      col[i * 32 + 0] = _mm_add_epi16(stp1_0, stp1_31);
+      col[i * 32 + 1] = _mm_add_epi16(stp1_1, stp1_30);
+      col[i * 32 + 2] = _mm_add_epi16(stp1_2, stp1_29);
+      col[i * 32 + 3] = _mm_add_epi16(stp1_3, stp1_28);
+      col[i * 32 + 4] = _mm_add_epi16(stp1_4, stp1_27);
+      col[i * 32 + 5] = _mm_add_epi16(stp1_5, stp1_26);
+      col[i * 32 + 6] = _mm_add_epi16(stp1_6, stp1_25);
+      col[i * 32 + 7] = _mm_add_epi16(stp1_7, stp1_24);
+      col[i * 32 + 8] = _mm_add_epi16(stp1_8, stp1_23);
+      col[i * 32 + 9] = _mm_add_epi16(stp1_9, stp1_22);
+      col[i * 32 + 10] = _mm_add_epi16(stp1_10, stp1_21);
+      col[i * 32 + 11] = _mm_add_epi16(stp1_11, stp1_20);
+      col[i * 32 + 12] = _mm_add_epi16(stp1_12, stp1_19);
+      col[i * 32 + 13] = _mm_add_epi16(stp1_13, stp1_18);
+      col[i * 32 + 14] = _mm_add_epi16(stp1_14, stp1_17);
+      col[i * 32 + 15] = _mm_add_epi16(stp1_15, stp1_16);
+      col[i * 32 + 16] = _mm_sub_epi16(stp1_15, stp1_16);
+      col[i * 32 + 17] = _mm_sub_epi16(stp1_14, stp1_17);
+      col[i * 32 + 18] = _mm_sub_epi16(stp1_13, stp1_18);
+      col[i * 32 + 19] = _mm_sub_epi16(stp1_12, stp1_19);
+      col[i * 32 + 20] = _mm_sub_epi16(stp1_11, stp1_20);
+      col[i * 32 + 21] = _mm_sub_epi16(stp1_10, stp1_21);
+      col[i * 32 + 22] = _mm_sub_epi16(stp1_9, stp1_22);
+      col[i * 32 + 23] = _mm_sub_epi16(stp1_8, stp1_23);
+      col[i * 32 + 24] = _mm_sub_epi16(stp1_7, stp1_24);
+      col[i * 32 + 25] = _mm_sub_epi16(stp1_6, stp1_25);
+      col[i * 32 + 26] = _mm_sub_epi16(stp1_5, stp1_26);
+      col[i * 32 + 27] = _mm_sub_epi16(stp1_4, stp1_27);
+      col[i * 32 + 28] = _mm_sub_epi16(stp1_3, stp1_28);
+      col[i * 32 + 29] = _mm_sub_epi16(stp1_2, stp1_29);
+      col[i * 32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
+      col[i * 32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
+    } else {
+      // 2_D: Calculate the results and store them to destination.
+      in0 = _mm_add_epi16(stp1_0, stp1_31);
+      in1 = _mm_add_epi16(stp1_1, stp1_30);
+      in2 = _mm_add_epi16(stp1_2, stp1_29);
+      in3 = _mm_add_epi16(stp1_3, stp1_28);
+      in4 = _mm_add_epi16(stp1_4, stp1_27);
+      in5 = _mm_add_epi16(stp1_5, stp1_26);
+      in6 = _mm_add_epi16(stp1_6, stp1_25);
+      in7 = _mm_add_epi16(stp1_7, stp1_24);
+      in8 = _mm_add_epi16(stp1_8, stp1_23);
+      in9 = _mm_add_epi16(stp1_9, stp1_22);
+      in10 = _mm_add_epi16(stp1_10, stp1_21);
+      in11 = _mm_add_epi16(stp1_11, stp1_20);
+      in12 = _mm_add_epi16(stp1_12, stp1_19);
+      in13 = _mm_add_epi16(stp1_13, stp1_18);
+      in14 = _mm_add_epi16(stp1_14, stp1_17);
+      in15 = _mm_add_epi16(stp1_15, stp1_16);
+      in16 = _mm_sub_epi16(stp1_15, stp1_16);
+      in17 = _mm_sub_epi16(stp1_14, stp1_17);
+      in18 = _mm_sub_epi16(stp1_13, stp1_18);
+      in19 = _mm_sub_epi16(stp1_12, stp1_19);
+      in20 = _mm_sub_epi16(stp1_11, stp1_20);
+      in21 = _mm_sub_epi16(stp1_10, stp1_21);
+      in22 = _mm_sub_epi16(stp1_9, stp1_22);
+      in23 = _mm_sub_epi16(stp1_8, stp1_23);
+      in24 = _mm_sub_epi16(stp1_7, stp1_24);
+      in25 = _mm_sub_epi16(stp1_6, stp1_25);
+      in26 = _mm_sub_epi16(stp1_5, stp1_26);
+      in27 = _mm_sub_epi16(stp1_4, stp1_27);
+      in28 = _mm_sub_epi16(stp1_3, stp1_28);
+      in29 = _mm_sub_epi16(stp1_2, stp1_29);
+      in30 = _mm_sub_epi16(stp1_1, stp1_30);
+      in31 = _mm_sub_epi16(stp1_0, stp1_31);
+
+      // 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);
+      in8 = _mm_adds_epi16(in8, final_rounding);
+      in9 = _mm_adds_epi16(in9, final_rounding);
+      in10 = _mm_adds_epi16(in10, final_rounding);
+      in11 = _mm_adds_epi16(in11, final_rounding);
+      in12 = _mm_adds_epi16(in12, final_rounding);
+      in13 = _mm_adds_epi16(in13, final_rounding);
+      in14 = _mm_adds_epi16(in14, final_rounding);
+      in15 = _mm_adds_epi16(in15, final_rounding);
+      in16 = _mm_adds_epi16(in16, final_rounding);
+      in17 = _mm_adds_epi16(in17, final_rounding);
+      in18 = _mm_adds_epi16(in18, final_rounding);
+      in19 = _mm_adds_epi16(in19, final_rounding);
+      in20 = _mm_adds_epi16(in20, final_rounding);
+      in21 = _mm_adds_epi16(in21, final_rounding);
+      in22 = _mm_adds_epi16(in22, final_rounding);
+      in23 = _mm_adds_epi16(in23, final_rounding);
+      in24 = _mm_adds_epi16(in24, final_rounding);
+      in25 = _mm_adds_epi16(in25, final_rounding);
+      in26 = _mm_adds_epi16(in26, final_rounding);
+      in27 = _mm_adds_epi16(in27, final_rounding);
+      in28 = _mm_adds_epi16(in28, final_rounding);
+      in29 = _mm_adds_epi16(in29, final_rounding);
+      in30 = _mm_adds_epi16(in30, final_rounding);
+      in31 = _mm_adds_epi16(in31, final_rounding);
+
+      in0 = _mm_srai_epi16(in0, 6);
+      in1 = _mm_srai_epi16(in1, 6);
+      in2 = _mm_srai_epi16(in2, 6);
+      in3 = _mm_srai_epi16(in3, 6);
+      in4 = _mm_srai_epi16(in4, 6);
+      in5 = _mm_srai_epi16(in5, 6);
+      in6 = _mm_srai_epi16(in6, 6);
+      in7 = _mm_srai_epi16(in7, 6);
+      in8 = _mm_srai_epi16(in8, 6);
+      in9 = _mm_srai_epi16(in9, 6);
+      in10 = _mm_srai_epi16(in10, 6);
+      in11 = _mm_srai_epi16(in11, 6);
+      in12 = _mm_srai_epi16(in12, 6);
+      in13 = _mm_srai_epi16(in13, 6);
+      in14 = _mm_srai_epi16(in14, 6);
+      in15 = _mm_srai_epi16(in15, 6);
+      in16 = _mm_srai_epi16(in16, 6);
+      in17 = _mm_srai_epi16(in17, 6);
+      in18 = _mm_srai_epi16(in18, 6);
+      in19 = _mm_srai_epi16(in19, 6);
+      in20 = _mm_srai_epi16(in20, 6);
+      in21 = _mm_srai_epi16(in21, 6);
+      in22 = _mm_srai_epi16(in22, 6);
+      in23 = _mm_srai_epi16(in23, 6);
+      in24 = _mm_srai_epi16(in24, 6);
+      in25 = _mm_srai_epi16(in25, 6);
+      in26 = _mm_srai_epi16(in26, 6);
+      in27 = _mm_srai_epi16(in27, 6);
+      in28 = _mm_srai_epi16(in28, 6);
+      in29 = _mm_srai_epi16(in29, 6);
+      in30 = _mm_srai_epi16(in30, 6);
+      in31 = _mm_srai_epi16(in31, 6);
+
+      // Store results
+      _mm_store_si128((__m128i *)output, in0);
+      _mm_store_si128((__m128i *)(output + half_pitch * 1), in1);
+      _mm_store_si128((__m128i *)(output + half_pitch * 2), in2);
+      _mm_store_si128((__m128i *)(output + half_pitch * 3), in3);
+      _mm_store_si128((__m128i *)(output + half_pitch * 4), in4);
+      _mm_store_si128((__m128i *)(output + half_pitch * 5), in5);
+      _mm_store_si128((__m128i *)(output + half_pitch * 6), in6);
+      _mm_store_si128((__m128i *)(output + half_pitch * 7), in7);
+      _mm_store_si128((__m128i *)(output + half_pitch * 8), in8);
+      _mm_store_si128((__m128i *)(output + half_pitch * 9), in9);
+      _mm_store_si128((__m128i *)(output + half_pitch * 10), in10);
+      _mm_store_si128((__m128i *)(output + half_pitch * 11), in11);
+      _mm_store_si128((__m128i *)(output + half_pitch * 12), in12);
+      _mm_store_si128((__m128i *)(output + half_pitch * 13), in13);
+      _mm_store_si128((__m128i *)(output + half_pitch * 14), in14);
+      _mm_store_si128((__m128i *)(output + half_pitch * 15), in15);
+      _mm_store_si128((__m128i *)(output + half_pitch * 16), in16);
+      _mm_store_si128((__m128i *)(output + half_pitch * 17), in17);
+      _mm_store_si128((__m128i *)(output + half_pitch * 18), in18);
+      _mm_store_si128((__m128i *)(output + half_pitch * 19), in19);
+      _mm_store_si128((__m128i *)(output + half_pitch * 20), in20);
+      _mm_store_si128((__m128i *)(output + half_pitch * 21), in21);
+      _mm_store_si128((__m128i *)(output + half_pitch * 22), in22);
+      _mm_store_si128((__m128i *)(output + half_pitch * 23), in23);
+      _mm_store_si128((__m128i *)(output + half_pitch * 24), in24);
+      _mm_store_si128((__m128i *)(output + half_pitch * 25), in25);
+      _mm_store_si128((__m128i *)(output + half_pitch * 26), in26);
+      _mm_store_si128((__m128i *)(output + half_pitch * 27), in27);
+      _mm_store_si128((__m128i *)(output + half_pitch * 28), in28);
+      _mm_store_si128((__m128i *)(output + half_pitch * 29), in29);
+      _mm_store_si128((__m128i *)(output + half_pitch * 30), in30);
+      _mm_store_si128((__m128i *)(output + half_pitch * 31), in31);
+
+      output += 8;
+    }
+  }
+}
 #endif