From: Angie Chiang <angiebird@google.com>
Date: Tue, 27 Oct 2015 23:29:29 +0000 (-0700)
Subject: Add fadst4
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=1ef285bfe9d5837a48d27e425ef36017d1067b3a;p=libvpx

Add fadst4

Change-Id: I320a1cf82d1705e5ec7fe1870327f70ce8493489
---

diff --git a/vp10/common/vp10_fwd_txfm1d.c b/vp10/common/vp10_fwd_txfm1d.c
index c4ed9e35d..49c0c8b13 100644
--- a/vp10/common/vp10_fwd_txfm1d.c
+++ b/vp10/common/vp10_fwd_txfm1d.c
@@ -698,3 +698,68 @@ void vp10_fdct32_new(const int32_t *input, int32_t *output,
   bf1[31] = bf0[31];
   range_check(stage, input, bf1, size, stage_range[stage]);
 }
+
+void vp10_fadst4_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 4;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[4];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[3];
+  bf1[1] = input[0];
+  bf1[2] = input[1];
+  bf1[3] = input[2];
+  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[8], bf0[0], cospi[56], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[8], bf0[1], cospi[56], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[40], bf0[2], cospi[24], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[40], bf0[3], cospi[24], bf0[2], 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[2] + bf0[0];
+  bf1[3] = -bf0[3] + bf0[1];
+  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] = 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]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = -bf0[2];
+  bf1[2] = bf0[3];
+  bf1[3] = -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 909782ee9..5f0056907 100644
--- a/vp10/common/vp10_fwd_txfm1d.h
+++ b/vp10/common/vp10_fwd_txfm1d.h
@@ -26,6 +26,9 @@ void vp10_fdct16_new(const int32_t *input, int32_t *output,
 void vp10_fdct32_new(const int32_t *input, int32_t *output,
                      const int8_t *cos_bit, const int8_t *stage_range);
 
+void vp10_fadst4_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range);
+
 #ifdef __cplusplus
 }
 #endif