From: Angie Chiang <angiebird@google.com>
Date: Tue, 27 Oct 2015 23:22:32 +0000 (-0700)
Subject: Add fdct8
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=3eb719f6b07840636d954830187cc22ca569e1be;p=libvpx

Add fdct8

Change-Id: I66b884bc831324b5295c7874aa23f62042157834
---

diff --git a/vp10/common/vp10_fwd_txfm1d.c b/vp10/common/vp10_fwd_txfm1d.c
index ebc9fcdce..8d855548c 100644
--- a/vp10/common/vp10_fwd_txfm1d.c
+++ b/vp10/common/vp10_fwd_txfm1d.c
@@ -81,3 +81,88 @@ void vp10_fdct4_new(const int32_t *input, int32_t *output,
   bf1[3] = bf0[3];
   range_check(stage, input, bf1, size, stage_range[stage]);
 }
+
+void vp10_fdct8_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] + input[7];
+  bf1[1] = input[1] + input[6];
+  bf1[2] = input[2] + input[5];
+  bf1[3] = input[3] + input[4];
+  bf1[4] = -input[4] + input[3];
+  bf1[5] = -input[5] + input[2];
+  bf1[6] = -input[6] + input[1];
+  bf1[7] = -input[7] + input[0];
+  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] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = -bf0[2] + bf0[1];
+  bf1[3] = -bf0[3] + bf0[0];
+  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[6], cospi[32], bf0[5], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  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[1], cospi[32], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[48], bf0[3], -cospi[16], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = -bf0[5] + bf0[4];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[7] + bf0[6];
+  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[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[24], bf0[6], -cospi[40], bf0[5], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[56], bf0[7], -cospi[8], bf0[4], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[4];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[6];
+  bf1[4] = bf0[1];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[3];
+  bf1[7] = bf0[7];
+  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 fe786a530..660242c51 100644
--- a/vp10/common/vp10_fwd_txfm1d.h
+++ b/vp10/common/vp10_fwd_txfm1d.h
@@ -19,6 +19,8 @@ extern "C" {
 
 void vp10_fdct4_new(const int32_t *input, int32_t *output,
                     const int8_t *cos_bit, const int8_t *stage_range);
+void vp10_fdct8_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
 
 #ifdef __cplusplus
 }