Add idct8

Change-Id: I8092a6f229b196c5c8b7dcd2dff8aaf68253e422

diff --git a/vp10/common/vp10_inv_txfm1d.c b/vp10/common/vp10_inv_txfm1d.c
index 55ad4a933e0cf6878264f65cf826ddbc1a60b98f..0e64fc109e273dfd6bd90c81556c0b09698222b3 100644 (file)
--- a/vp10/common/vp10_inv_txfm1d.c
+++ b/vp10/common/vp10_inv_txfm1d.c
@@ -81,3 +81,88 @@ void vp10_idct4_new(const int32_t *input, int32_t *output,
   bf1[3] = bf0[0] - bf0[3];
   range_check(stage, input, bf1, size, stage_range[stage]);
 }
+
+void vp10_idct8_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 8;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[8];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = input[4];
+  bf1[2] = input[2];
+  bf1[3] = input[6];
+  bf1[4] = input[1];
+  bf1[5] = input[5];
+  bf1[6] = input[3];
+  bf1[7] = input[7];
+  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] = 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]);
+  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] = 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];
+  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] + 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];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  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];
+  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 2ce4234817054b97da5e2da9402b1f8735698a54..d0763e4685dbbe6dd875a9912ab063738a248d8c 100644 (file)
--- a/vp10/common/vp10_inv_txfm1d.h
+++ b/vp10/common/vp10_inv_txfm1d.h
@@ -19,6 +19,8 @@ extern "C" {
 
 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);
 
 #ifdef __cplusplus
 }