From: Angie Chiang <angiebird@google.com>
Date: Tue, 27 Oct 2015 23:31:42 +0000 (-0700)
Subject: Add fadst32
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=a9253a2029d90b6113ce090d15a9d082f52460a6;p=libvpx

Add fadst32

Change-Id: I77299f0e39fc7cef91e7e420513dbd05194f320a
---

diff --git a/vp10/common/vp10_fwd_txfm1d.c b/vp10/common/vp10_fwd_txfm1d.c
index fd8cba2a4..6e19e278d 100644
--- a/vp10/common/vp10_fwd_txfm1d.c
+++ b/vp10/common/vp10_fwd_txfm1d.c
@@ -1093,3 +1093,438 @@ void vp10_fadst16_new(const int32_t *input, int32_t *output,
   range_check(stage, input, bf1, size, stage_range[stage]);
 
 }
+
+void vp10_fadst32_new(const int32_t *input, int32_t *output,
+                      const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 32;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[32];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[31];
+  bf1[1] = input[0];
+  bf1[2] = input[29];
+  bf1[3] = input[2];
+  bf1[4] = input[27];
+  bf1[5] = input[4];
+  bf1[6] = input[25];
+  bf1[7] = input[6];
+  bf1[8] = input[23];
+  bf1[9] = input[8];
+  bf1[10] = input[21];
+  bf1[11] = input[10];
+  bf1[12] = input[19];
+  bf1[13] = input[12];
+  bf1[14] = input[17];
+  bf1[15] = input[14];
+  bf1[16] = input[15];
+  bf1[17] = input[16];
+  bf1[18] = input[13];
+  bf1[19] = input[18];
+  bf1[20] = input[11];
+  bf1[21] = input[20];
+  bf1[22] = input[9];
+  bf1[23] = input[22];
+  bf1[24] = input[7];
+  bf1[25] = input[24];
+  bf1[26] = input[5];
+  bf1[27] = input[26];
+  bf1[28] = input[3];
+  bf1[29] = input[28];
+  bf1[30] = input[1];
+  bf1[31] = input[30];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[1], bf0[0], cospi[63], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[1], bf0[1], cospi[63], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[5], bf0[2], cospi[59], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[5], bf0[3], cospi[59], bf0[2], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[9], bf0[4], cospi[55], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[9], bf0[5], cospi[55], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[13], bf0[6], cospi[51], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[13], bf0[7], cospi[51], bf0[6], cos_bit[stage]);
+  bf1[8] = half_btf(cospi[17], bf0[8], cospi[47], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(-cospi[17], bf0[9], cospi[47], bf0[8], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[21], bf0[10], cospi[43], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[21], bf0[11], cospi[43], bf0[10], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[25], bf0[12], cospi[39], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(-cospi[25], bf0[13], cospi[39], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[29], bf0[14], cospi[35], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(-cospi[29], bf0[15], cospi[35], bf0[14], cos_bit[stage]);
+  bf1[16] = half_btf(cospi[33], bf0[16], cospi[31], bf0[17], cos_bit[stage]);
+  bf1[17] = half_btf(-cospi[33], bf0[17], cospi[31], bf0[16], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[37], bf0[18], cospi[27], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[37], bf0[19], cospi[27], bf0[18], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[41], bf0[20], cospi[23], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[41], bf0[21], cospi[23], bf0[20], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[45], bf0[22], cospi[19], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[45], bf0[23], cospi[19], bf0[22], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[49], bf0[24], cospi[15], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(-cospi[49], bf0[25], cospi[15], bf0[24], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[53], bf0[26], cospi[11], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[53], bf0[27], cospi[11], bf0[26], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[57], bf0[28], cospi[7], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(-cospi[57], bf0[29], cospi[7], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[61], bf0[30], cospi[3], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(-cospi[61], bf0[31], cospi[3], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[16];
+  bf1[1] = bf0[1] + bf0[17];
+  bf1[2] = bf0[2] + bf0[18];
+  bf1[3] = bf0[3] + bf0[19];
+  bf1[4] = bf0[4] + bf0[20];
+  bf1[5] = bf0[5] + bf0[21];
+  bf1[6] = bf0[6] + bf0[22];
+  bf1[7] = bf0[7] + bf0[23];
+  bf1[8] = bf0[8] + bf0[24];
+  bf1[9] = bf0[9] + bf0[25];
+  bf1[10] = bf0[10] + bf0[26];
+  bf1[11] = bf0[11] + bf0[27];
+  bf1[12] = bf0[12] + bf0[28];
+  bf1[13] = bf0[13] + bf0[29];
+  bf1[14] = bf0[14] + bf0[30];
+  bf1[15] = bf0[15] + bf0[31];
+  bf1[16] = -bf0[16] + bf0[0];
+  bf1[17] = -bf0[17] + bf0[1];
+  bf1[18] = -bf0[18] + bf0[2];
+  bf1[19] = -bf0[19] + bf0[3];
+  bf1[20] = -bf0[20] + bf0[4];
+  bf1[21] = -bf0[21] + bf0[5];
+  bf1[22] = -bf0[22] + bf0[6];
+  bf1[23] = -bf0[23] + bf0[7];
+  bf1[24] = -bf0[24] + bf0[8];
+  bf1[25] = -bf0[25] + bf0[9];
+  bf1[26] = -bf0[26] + bf0[10];
+  bf1[27] = -bf0[27] + bf0[11];
+  bf1[28] = -bf0[28] + bf0[12];
+  bf1[29] = -bf0[29] + bf0[13];
+  bf1[30] = -bf0[30] + bf0[14];
+  bf1[31] = -bf0[31] + bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = half_btf(cospi[4], bf0[16], cospi[60], bf0[17], cos_bit[stage]);
+  bf1[17] = half_btf(-cospi[4], bf0[17], cospi[60], bf0[16], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[20], bf0[18], cospi[44], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[20], bf0[19], cospi[44], bf0[18], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[36], bf0[20], cospi[28], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[36], bf0[21], cospi[28], bf0[20], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[52], bf0[22], cospi[12], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[52], bf0[23], cospi[12], bf0[22], cos_bit[stage]);
+  bf1[24] = half_btf(-cospi[60], bf0[24], cospi[4], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[60], bf0[25], cospi[4], bf0[24], cos_bit[stage]);
+  bf1[26] = half_btf(-cospi[44], bf0[26], cospi[20], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[44], bf0[27], cospi[20], bf0[26], cos_bit[stage]);
+  bf1[28] = half_btf(-cospi[28], bf0[28], cospi[36], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[28], bf0[29], cospi[36], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[12], bf0[30], cospi[52], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[12], bf0[31], cospi[52], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[8];
+  bf1[1] = bf0[1] + bf0[9];
+  bf1[2] = bf0[2] + bf0[10];
+  bf1[3] = bf0[3] + bf0[11];
+  bf1[4] = bf0[4] + bf0[12];
+  bf1[5] = bf0[5] + bf0[13];
+  bf1[6] = bf0[6] + bf0[14];
+  bf1[7] = bf0[7] + bf0[15];
+  bf1[8] = -bf0[8] + bf0[0];
+  bf1[9] = -bf0[9] + bf0[1];
+  bf1[10] = -bf0[10] + bf0[2];
+  bf1[11] = -bf0[11] + bf0[3];
+  bf1[12] = -bf0[12] + bf0[4];
+  bf1[13] = -bf0[13] + bf0[5];
+  bf1[14] = -bf0[14] + bf0[6];
+  bf1[15] = -bf0[15] + bf0[7];
+  bf1[16] = bf0[16] + bf0[24];
+  bf1[17] = bf0[17] + bf0[25];
+  bf1[18] = bf0[18] + bf0[26];
+  bf1[19] = bf0[19] + bf0[27];
+  bf1[20] = bf0[20] + bf0[28];
+  bf1[21] = bf0[21] + bf0[29];
+  bf1[22] = bf0[22] + bf0[30];
+  bf1[23] = bf0[23] + bf0[31];
+  bf1[24] = -bf0[24] + bf0[16];
+  bf1[25] = -bf0[25] + bf0[17];
+  bf1[26] = -bf0[26] + bf0[18];
+  bf1[27] = -bf0[27] + bf0[19];
+  bf1[28] = -bf0[28] + bf0[20];
+  bf1[29] = -bf0[29] + bf0[21];
+  bf1[30] = -bf0[30] + bf0[22];
+  bf1[31] = -bf0[31] + bf0[23];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[8], bf0[8], cospi[56], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(-cospi[8], bf0[9], cospi[56], bf0[8], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[40], bf0[10], cospi[24], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[40], bf0[11], cospi[24], bf0[10], cos_bit[stage]);
+  bf1[12] = half_btf(-cospi[56], bf0[12], cospi[8], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[56], bf0[13], cospi[8], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[24], bf0[14], cospi[40], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[24], bf0[15], cospi[40], bf0[14], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = half_btf(cospi[8], bf0[24], cospi[56], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(-cospi[8], bf0[25], cospi[56], bf0[24], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[40], bf0[26], cospi[24], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[40], bf0[27], cospi[24], bf0[26], cos_bit[stage]);
+  bf1[28] = half_btf(-cospi[56], bf0[28], cospi[8], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[56], bf0[29], cospi[8], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[24], bf0[30], cospi[40], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[24], bf0[31], cospi[40], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[4];
+  bf1[1] = bf0[1] + bf0[5];
+  bf1[2] = bf0[2] + bf0[6];
+  bf1[3] = bf0[3] + bf0[7];
+  bf1[4] = -bf0[4] + bf0[0];
+  bf1[5] = -bf0[5] + bf0[1];
+  bf1[6] = -bf0[6] + bf0[2];
+  bf1[7] = -bf0[7] + bf0[3];
+  bf1[8] = bf0[8] + bf0[12];
+  bf1[9] = bf0[9] + bf0[13];
+  bf1[10] = bf0[10] + bf0[14];
+  bf1[11] = bf0[11] + bf0[15];
+  bf1[12] = -bf0[12] + bf0[8];
+  bf1[13] = -bf0[13] + bf0[9];
+  bf1[14] = -bf0[14] + bf0[10];
+  bf1[15] = -bf0[15] + bf0[11];
+  bf1[16] = bf0[16] + bf0[20];
+  bf1[17] = bf0[17] + bf0[21];
+  bf1[18] = bf0[18] + bf0[22];
+  bf1[19] = bf0[19] + bf0[23];
+  bf1[20] = -bf0[20] + bf0[16];
+  bf1[21] = -bf0[21] + bf0[17];
+  bf1[22] = -bf0[22] + bf0[18];
+  bf1[23] = -bf0[23] + bf0[19];
+  bf1[24] = bf0[24] + bf0[28];
+  bf1[25] = bf0[25] + bf0[29];
+  bf1[26] = bf0[26] + bf0[30];
+  bf1[27] = bf0[27] + bf0[31];
+  bf1[28] = -bf0[28] + bf0[24];
+  bf1[29] = -bf0[29] + bf0[25];
+  bf1[30] = -bf0[30] + bf0[26];
+  bf1[31] = -bf0[31] + bf0[27];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[16], bf0[5], cospi[48], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[48], bf0[7], cospi[16], bf0[6], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = half_btf(cospi[16], bf0[12], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(-cospi[16], bf0[13], cospi[48], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[48], bf0[14], cospi[16], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[48], bf0[15], cospi[16], bf0[14], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = half_btf(cospi[16], bf0[20], cospi[48], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[16], bf0[21], cospi[48], bf0[20], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[48], bf0[22], cospi[16], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[48], bf0[23], cospi[16], bf0[22], cos_bit[stage]);
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = bf0[26];
+  bf1[27] = bf0[27];
+  bf1[28] = half_btf(cospi[16], bf0[28], cospi[48], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(-cospi[16], bf0[29], cospi[48], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[48], bf0[30], cospi[16], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[48], bf0[31], cospi[16], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = -bf0[2] + bf0[0];
+  bf1[3] = -bf0[3] + bf0[1];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = -bf0[6] + bf0[4];
+  bf1[7] = -bf0[7] + bf0[5];
+  bf1[8] = bf0[8] + bf0[10];
+  bf1[9] = bf0[9] + bf0[11];
+  bf1[10] = -bf0[10] + bf0[8];
+  bf1[11] = -bf0[11] + bf0[9];
+  bf1[12] = bf0[12] + bf0[14];
+  bf1[13] = bf0[13] + bf0[15];
+  bf1[14] = -bf0[14] + bf0[12];
+  bf1[15] = -bf0[15] + bf0[13];
+  bf1[16] = bf0[16] + bf0[18];
+  bf1[17] = bf0[17] + bf0[19];
+  bf1[18] = -bf0[18] + bf0[16];
+  bf1[19] = -bf0[19] + bf0[17];
+  bf1[20] = bf0[20] + bf0[22];
+  bf1[21] = bf0[21] + bf0[23];
+  bf1[22] = -bf0[22] + bf0[20];
+  bf1[23] = -bf0[23] + bf0[21];
+  bf1[24] = bf0[24] + bf0[26];
+  bf1[25] = bf0[25] + bf0[27];
+  bf1[26] = -bf0[26] + bf0[24];
+  bf1[27] = -bf0[27] + bf0[25];
+  bf1[28] = bf0[28] + bf0[30];
+  bf1[29] = bf0[29] + bf0[31];
+  bf1[30] = -bf0[30] + bf0[28];
+  bf1[31] = -bf0[31] + bf0[29];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 10
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[32], bf0[3], cospi[32], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[32], bf0[7], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(cospi[32], bf0[10], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[10], cos_bit[stage]);
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = half_btf(cospi[32], bf0[14], cospi[32], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(-cospi[32], bf0[15], cospi[32], bf0[14], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = half_btf(cospi[32], bf0[18], cospi[32], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[32], bf0[19], cospi[32], bf0[18], cos_bit[stage]);
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = half_btf(cospi[32], bf0[22], cospi[32], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[32], bf0[23], cospi[32], bf0[22], cos_bit[stage]);
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = half_btf(cospi[32], bf0[26], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[32], bf0[27], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = half_btf(cospi[32], bf0[30], cospi[32], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(-cospi[32], bf0[31], cospi[32], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 11
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = -bf0[16];
+  bf1[2] = bf0[24];
+  bf1[3] = -bf0[8];
+  bf1[4] = bf0[12];
+  bf1[5] = -bf0[28];
+  bf1[6] = bf0[20];
+  bf1[7] = -bf0[4];
+  bf1[8] = bf0[6];
+  bf1[9] = -bf0[22];
+  bf1[10] = bf0[30];
+  bf1[11] = -bf0[14];
+  bf1[12] = bf0[10];
+  bf1[13] = -bf0[26];
+  bf1[14] = bf0[18];
+  bf1[15] = -bf0[2];
+  bf1[16] = bf0[3];
+  bf1[17] = -bf0[19];
+  bf1[18] = bf0[27];
+  bf1[19] = -bf0[11];
+  bf1[20] = bf0[15];
+  bf1[21] = -bf0[31];
+  bf1[22] = bf0[23];
+  bf1[23] = -bf0[7];
+  bf1[24] = bf0[5];
+  bf1[25] = -bf0[21];
+  bf1[26] = bf0[29];
+  bf1[27] = -bf0[13];
+  bf1[28] = bf0[9];
+  bf1[29] = -bf0[25];
+  bf1[30] = bf0[17];
+  bf1[31] = -bf0[1];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
diff --git a/vp10/common/vp10_fwd_txfm1d.h b/vp10/common/vp10_fwd_txfm1d.h
index 2e9814961..d5b9f4014 100644
--- a/vp10/common/vp10_fwd_txfm1d.h
+++ b/vp10/common/vp10_fwd_txfm1d.h
@@ -32,6 +32,8 @@ void vp10_fadst8_new(const int32_t *input, int32_t *output,
                      const int8_t *cos_bit, const int8_t *stage_range);
 void vp10_fadst16_new(const int32_t *input, int32_t *output,
                       const int8_t *cos_bit, const int8_t *stage_range);
+void vp10_fadst32_new(const int32_t *input, int32_t *output,
+                      const int8_t *cos_bit, const int8_t *stage_range);
 
 #ifdef __cplusplus
 }