Commit d3cee565 authored by Angie Chiang's avatar Angie Chiang

Add idct16

Change-Id: I8e5ba3a3f9b64ccbf038e371525e897774729b06
parent bd9db2f5
......@@ -166,3 +166,174 @@ void vp10_idct8_new(const int32_t *input, int32_t *output,
bf1[7] = bf0[0] - bf0[7];
range_check(stage, input, bf1, size, stage_range[stage]);
}
void vp10_idct16_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[8];
bf1[2] = input[4];
bf1[3] = input[12];
bf1[4] = input[2];
bf1[5] = input[10];
bf1[6] = input[6];
bf1[7] = input[14];
bf1[8] = input[1];
bf1[9] = input[9];
bf1[10] = input[5];
bf1[11] = input[13];
bf1[12] = input[3];
bf1[13] = input[11];
bf1[14] = input[7];
bf1[15] = input[15];
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] = 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[60], bf0[8], -cospi[4], bf0[15], cos_bit[stage]);
bf1[9] = half_btf(cospi[28], bf0[9], -cospi[36], bf0[14], cos_bit[stage]);
bf1[10] = half_btf(cospi[44], bf0[10], -cospi[20], bf0[13], cos_bit[stage]);
bf1[11] = half_btf(cospi[12], bf0[11], -cospi[52], bf0[12], cos_bit[stage]);
bf1[12] = half_btf(cospi[52], bf0[11], cospi[12], bf0[12], cos_bit[stage]);
bf1[13] = half_btf(cospi[20], bf0[10], cospi[44], bf0[13], cos_bit[stage]);
bf1[14] = half_btf(cospi[36], bf0[9], cospi[28], bf0[14], cos_bit[stage]);
bf1[15] = half_btf(cospi[4], bf0[8], cospi[60], 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];
bf1[1] = bf0[1];
bf1[2] = bf0[2];
bf1[3] = bf0[3];
bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit[stage]);
bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit[stage]);
bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit[stage]);
bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit[stage]);
bf1[8] = bf0[8] + bf0[9];
bf1[9] = bf0[8] - bf0[9];
bf1[10] = -bf0[10] + bf0[11];
bf1[11] = bf0[10] + bf0[11];
bf1[12] = bf0[12] + bf0[13];
bf1[13] = bf0[12] - bf0[13];
bf1[14] = -bf0[14] + bf0[15];
bf1[15] = bf0[14] + 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] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit[stage]);
bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit[stage]);
bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit[stage]);
bf1[4] = bf0[4] + bf0[5];
bf1[5] = bf0[4] - bf0[5];
bf1[6] = -bf0[6] + bf0[7];
bf1[7] = bf0[6] + bf0[7];
bf1[8] = bf0[8];
bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit[stage]);
bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit[stage]);
bf1[11] = bf0[11];
bf1[12] = bf0[12];
bf1[13] = half_btf(-cospi[16], bf0[10], cospi[48], bf0[13], cos_bit[stage]);
bf1[14] = half_btf(cospi[48], bf0[9], cospi[16], bf0[14], cos_bit[stage]);
bf1[15] = bf0[15];
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[3];
bf1[1] = bf0[1] + bf0[2];
bf1[2] = bf0[1] - bf0[2];
bf1[3] = bf0[0] - bf0[3];
bf1[4] = bf0[4];
bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
bf1[7] = bf0[7];
bf1[8] = bf0[8] + bf0[11];
bf1[9] = bf0[9] + bf0[10];
bf1[10] = bf0[9] - bf0[10];
bf1[11] = bf0[8] - bf0[11];
bf1[12] = -bf0[12] + bf0[15];
bf1[13] = -bf0[13] + bf0[14];
bf1[14] = bf0[13] + bf0[14];
bf1[15] = bf0[12] + 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] + bf0[7];
bf1[1] = bf0[1] + bf0[6];
bf1[2] = bf0[2] + bf0[5];
bf1[3] = bf0[3] + bf0[4];
bf1[4] = bf0[3] - bf0[4];
bf1[5] = bf0[2] - bf0[5];
bf1[6] = bf0[1] - bf0[6];
bf1[7] = bf0[0] - bf0[7];
bf1[8] = bf0[8];
bf1[9] = bf0[9];
bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
bf1[12] = half_btf(cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
bf1[13] = half_btf(cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
bf1[14] = bf0[14];
bf1[15] = bf0[15];
range_check(stage, input, bf1, size, stage_range[stage]);
// stage 7
stage++;
bf0 = step;
bf1 = output;
bf1[0] = bf0[0] + bf0[15];
bf1[1] = bf0[1] + bf0[14];
bf1[2] = bf0[2] + bf0[13];
bf1[3] = bf0[3] + bf0[12];
bf1[4] = bf0[4] + bf0[11];
bf1[5] = bf0[5] + bf0[10];
bf1[6] = bf0[6] + bf0[9];
bf1[7] = bf0[7] + bf0[8];
bf1[8] = bf0[7] - bf0[8];
bf1[9] = bf0[6] - bf0[9];
bf1[10] = bf0[5] - bf0[10];
bf1[11] = bf0[4] - bf0[11];
bf1[12] = bf0[3] - bf0[12];
bf1[13] = bf0[2] - bf0[13];
bf1[14] = bf0[1] - bf0[14];
bf1[15] = bf0[0] - bf0[15];
range_check(stage, input, bf1, size, stage_range[stage]);
}
......@@ -21,6 +21,8 @@ void vp10_idct4_new(const int32_t *input, int32_t *output,
const int8_t *cos_bit, const int8_t *stage_range);
void vp10_idct8_new(const int32_t *input, int32_t *output,
const int8_t *cos_bit, const int8_t *stage_range);
void vp10_idct16_new(const int32_t *input, int32_t *output,
const int8_t *cos_bit, const int8_t *stage_range);
#ifdef __cplusplus
}
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment