From 0325b95938cac650ae7fd60243f96ca8896db0cc Mon Sep 17 00:00:00 2001
From: Johann <johannkoenig@google.com>
Date: Fri, 29 Jul 2016 08:29:55 -0700
Subject: [PATCH] Use common transpose for vpx_idct32x32_1024_add_neon

Change-Id: I6ef7970206d588761ebe80005aecd35365ec50ff
---
 vpx_dsp/arm/idct32x32_add_neon.c | 66 ++++++--------------------------
 1 file changed, 11 insertions(+), 55 deletions(-)

diff --git a/vpx_dsp/arm/idct32x32_add_neon.c b/vpx_dsp/arm/idct32x32_add_neon.c
index 04f51bfdd..c4d1e8473 100644
--- a/vpx_dsp/arm/idct32x32_add_neon.c
+++ b/vpx_dsp/arm/idct32x32_add_neon.c
@@ -11,6 +11,7 @@
 #include <arm_neon.h>
 
 #include "./vpx_config.h"
+#include "vpx_dsp/arm/transpose_neon.h"
 #include "vpx_dsp/txfm_common.h"
 
 #define LOAD_FROM_TRANSPOSED(prev, first, second) \
@@ -155,11 +156,7 @@ static INLINE void idct32_transpose_pair(int16_t *input, int16_t *t_buf) {
   int16_t *in;
   int i;
   const int stride = 32;
-  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
-  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
   int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
-  int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
-  int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
 
   for (i = 0; i < 4; i++, input += 8) {
     in = input;
@@ -179,65 +176,24 @@ static INLINE void idct32_transpose_pair(int16_t *input, int16_t *t_buf) {
     in += stride;
     q15s16 = vld1q_s16(in);
 
-    d16s16 = vget_low_s16(q8s16);
-    d17s16 = vget_high_s16(q8s16);
-    d18s16 = vget_low_s16(q9s16);
-    d19s16 = vget_high_s16(q9s16);
-    d20s16 = vget_low_s16(q10s16);
-    d21s16 = vget_high_s16(q10s16);
-    d22s16 = vget_low_s16(q11s16);
-    d23s16 = vget_high_s16(q11s16);
-    d24s16 = vget_low_s16(q12s16);
-    d25s16 = vget_high_s16(q12s16);
-    d26s16 = vget_low_s16(q13s16);
-    d27s16 = vget_high_s16(q13s16);
-    d28s16 = vget_low_s16(q14s16);
-    d29s16 = vget_high_s16(q14s16);
-    d30s16 = vget_low_s16(q15s16);
-    d31s16 = vget_high_s16(q15s16);
+    transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16,
+                      &q14s16, &q15s16);
 
-    q8s16 = vcombine_s16(d16s16, d24s16);   // vswp d17, d24
-    q9s16 = vcombine_s16(d18s16, d26s16);   // vswp d19, d26
-    q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
-    q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
-    q12s16 = vcombine_s16(d17s16, d25s16);
-    q13s16 = vcombine_s16(d19s16, d27s16);
-    q14s16 = vcombine_s16(d21s16, d29s16);
-    q15s16 = vcombine_s16(d23s16, d31s16);
-
-    q0x2s32 =
-        vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q10s16));
-    q1x2s32 =
-        vtrnq_s32(vreinterpretq_s32_s16(q9s16), vreinterpretq_s32_s16(q11s16));
-    q2x2s32 =
-        vtrnq_s32(vreinterpretq_s32_s16(q12s16), vreinterpretq_s32_s16(q14s16));
-    q3x2s32 =
-        vtrnq_s32(vreinterpretq_s32_s16(q13s16), vreinterpretq_s32_s16(q15s16));
-
-    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
-                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
-    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
-                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
-    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
-                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
-    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
-                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
-
-    vst1q_s16(t_buf, q0x2s16.val[0]);
+    vst1q_s16(t_buf, q8s16);
     t_buf += 8;
-    vst1q_s16(t_buf, q0x2s16.val[1]);
+    vst1q_s16(t_buf, q9s16);
     t_buf += 8;
-    vst1q_s16(t_buf, q1x2s16.val[0]);
+    vst1q_s16(t_buf, q10s16);
     t_buf += 8;
-    vst1q_s16(t_buf, q1x2s16.val[1]);
+    vst1q_s16(t_buf, q11s16);
     t_buf += 8;
-    vst1q_s16(t_buf, q2x2s16.val[0]);
+    vst1q_s16(t_buf, q12s16);
     t_buf += 8;
-    vst1q_s16(t_buf, q2x2s16.val[1]);
+    vst1q_s16(t_buf, q13s16);
     t_buf += 8;
-    vst1q_s16(t_buf, q3x2s16.val[0]);
+    vst1q_s16(t_buf, q14s16);
     t_buf += 8;
-    vst1q_s16(t_buf, q3x2s16.val[1]);
+    vst1q_s16(t_buf, q15s16);
     t_buf += 8;
   }
   return;