Add iadst16

Change-Id: I093881aacaf9a070f78cc4eea2e8a6ede8a71792

diff --git a/vp10/common/vp10_inv_txfm1d.c b/vp10/common/vp10_inv_txfm1d.c
index 957ebeda344026a8b5850e17e40af014f329b66b..0f10462c83fd59212e643b66fe95a3ec9aadf6ff 100644 (file)
--- a/vp10/common/vp10_inv_txfm1d.c
+++ b/vp10/common/vp10_inv_txfm1d.c
@@ -878,3 +878,220 @@ void vp10_iadst8_new(const int32_t *input, int32_t *output,
   bf1[7] = bf0[0];
   range_check(stage, input, bf1, size, stage_range[stage]);
 }
+
+void vp10_iadst16_new(const int32_t *input, int32_t *output,
+                      const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 16;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[16];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = -input[15];
+  bf1[2] = -input[7];
+  bf1[3] = input[8];
+  bf1[4] = -input[3];
+  bf1[5] = input[12];
+  bf1[6] = input[4];
+  bf1[7] = -input[11];
+  bf1[8] = -input[1];
+  bf1[9] = input[14];
+  bf1[10] = input[6];
+  bf1[11] = -input[9];
+  bf1[12] = input[2];
+  bf1[13] = -input[13];
+  bf1[14] = -input[5];
+  bf1[15] = input[10];
+  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] = 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[2], -cospi[32], bf0[3], 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[6], -cospi[32], bf0[7], 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[10], -cospi[32], bf0[11], 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[14], -cospi[32], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = bf0[0] - bf0[2];
+  bf1[3] = bf0[1] - bf0[3];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = bf0[4] - bf0[6];
+  bf1[7] = bf0[5] - bf0[7];
+  bf1[8] = bf0[8] + bf0[10];
+  bf1[9] = bf0[9] + bf0[11];
+  bf1[10] = bf0[8] - bf0[10];
+  bf1[11] = bf0[9] - bf0[11];
+  bf1[12] = bf0[12] + bf0[14];
+  bf1[13] = bf0[13] + bf0[15];
+  bf1[14] = bf0[12] - bf0[14];
+  bf1[15] = bf0[13] - 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] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[48], bf0[4], -cospi[16], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[16], bf0[6], cospi[48], bf0[7], 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[48], bf0[12], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[48], bf0[14], cospi[16], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[16], bf0[14], cospi[48], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  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[0] - bf0[4];
+  bf1[5] = bf0[1] - bf0[5];
+  bf1[6] = bf0[2] - bf0[6];
+  bf1[7] = bf0[3] - bf0[7];
+  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[8] - bf0[12];
+  bf1[13] = bf0[9] - bf0[13];
+  bf1[14] = bf0[10] - bf0[14];
+  bf1[15] = bf0[11] - bf0[15];
+  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[56], bf0[8], -cospi[8], bf0[9], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[40], bf0[10], cospi[24], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[24], bf0[10], -cospi[40], bf0[11], cos_bit[stage]);
+  bf1[12] = half_btf(-cospi[56], bf0[12], cospi[8], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[8], bf0[12], cospi[56], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[24], bf0[14], cospi[40], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[40], bf0[14], cospi[24], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  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[0] - bf0[8];
+  bf1[9] = bf0[1] - bf0[9];
+  bf1[10] = bf0[2] - bf0[10];
+  bf1[11] = bf0[3] - bf0[11];
+  bf1[12] = bf0[4] - bf0[12];
+  bf1[13] = bf0[5] - bf0[13];
+  bf1[14] = bf0[6] - bf0[14];
+  bf1[15] = bf0[7] - bf0[15];
+  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] = half_btf(cospi[2], bf0[0], cospi[62], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[62], bf0[0], -cospi[2], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[10], bf0[2], cospi[54], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[54], bf0[2], -cospi[10], bf0[3], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[18], bf0[4], cospi[46], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[46], bf0[4], -cospi[18], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[26], bf0[6], cospi[38], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[38], bf0[6], -cospi[26], bf0[7], cos_bit[stage]);
+  bf1[8] = half_btf(cospi[34], bf0[8], cospi[30], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[30], bf0[8], -cospi[34], bf0[9], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[42], bf0[10], cospi[22], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[22], bf0[10], -cospi[42], bf0[11], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[50], bf0[12], cospi[14], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[14], bf0[12], -cospi[50], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[58], bf0[14], cospi[6], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[6], bf0[14], -cospi[58], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[1];
+  bf1[1] = bf0[14];
+  bf1[2] = bf0[3];
+  bf1[3] = bf0[12];
+  bf1[4] = bf0[5];
+  bf1[5] = bf0[10];
+  bf1[6] = bf0[7];
+  bf1[7] = bf0[8];
+  bf1[8] = bf0[9];
+  bf1[9] = bf0[6];
+  bf1[10] = bf0[11];
+  bf1[11] = bf0[4];
+  bf1[12] = bf0[13];
+  bf1[13] = bf0[2];
+  bf1[14] = bf0[15];
+  bf1[15] = bf0[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}

diff --git a/vp10/common/vp10_inv_txfm1d.h b/vp10/common/vp10_inv_txfm1d.h
index 72aa65575adad124406a3037927cd136b7aab76a..eb1ec0fc1c78102c0cb59d492fa50c25bcafefa5 100644 (file)
--- a/vp10/common/vp10_inv_txfm1d.h
+++ b/vp10/common/vp10_inv_txfm1d.h
@@ -30,6 +30,8 @@ void vp10_iadst4_new(const int32_t *input, int32_t *output,
                      const int8_t *cos_bit, const int8_t *stage_range);
 void vp10_iadst8_new(const int32_t *input, int32_t *output,
                      const int8_t *cos_bit, const int8_t *stage_range);
+void vp10_iadst16_new(const int32_t *input, int32_t *output,
+                      const int8_t *cos_bit, const int8_t *stage_range);
 
 #ifdef __cplusplus
 }