1
From 9744978dd7b9ca9dda0846fa82faf6fa264a52a7 Mon Sep 17 00:00:00 2001
2
From: conrad <conrad@9553f0bf-9b14-0410-a0b8-cfaf0461ba5b>
3
Date: Wed, 29 Apr 2009 11:31:43 +0000
4
Subject: [PATCH 06/27] ARM: NEON put_pixels_clamped
6
git-svn-id: file:///var/local/repositories/ffmpeg/trunk@18712 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b
8
libavcodec/arm/dsputil_neon.c | 2 ++
9
libavcodec/arm/dsputil_neon_s.S | 24 ++++++++++++++++++++++++
10
2 files changed, 26 insertions(+), 0 deletions(-)
12
diff --git a/libavcodec/arm/dsputil_neon.c b/libavcodec/arm/dsputil_neon.c
13
index 37425a3..9b95130 100644
14
--- a/libavcodec/arm/dsputil_neon.c
15
+++ b/libavcodec/arm/dsputil_neon.c
16
@@ -42,6 +42,7 @@ void ff_put_pixels8_xy2_no_rnd_neon(uint8_t *, const uint8_t *, int, int);
17
void ff_avg_pixels16_neon(uint8_t *, const uint8_t *, int, int);
19
void ff_add_pixels_clamped_neon(const DCTELEM *, uint8_t *, int);
20
+void ff_put_pixels_clamped_neon(const DCTELEM *, uint8_t *, int);
21
void ff_put_signed_pixels_clamped_neon(const DCTELEM *, uint8_t *, int);
23
void ff_put_h264_qpel16_mc00_neon(uint8_t *, uint8_t *, int);
24
@@ -180,6 +181,7 @@ void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)
25
c->avg_pixels_tab[0][0] = ff_avg_pixels16_neon;
27
c->add_pixels_clamped = ff_add_pixels_clamped_neon;
28
+ c->put_pixels_clamped = ff_put_pixels_clamped_neon;
29
c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_neon;
31
c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_neon;
32
diff --git a/libavcodec/arm/dsputil_neon_s.S b/libavcodec/arm/dsputil_neon_s.S
33
index f16293d..a55e05f 100644
34
--- a/libavcodec/arm/dsputil_neon_s.S
35
+++ b/libavcodec/arm/dsputil_neon_s.S
36
@@ -273,6 +273,30 @@ function ff_put_h264_qpel8_mc00_neon, export=1
37
pixfunc2 put_ pixels8_y2, _no_rnd, vhadd.u8
38
pixfunc2 put_ pixels8_xy2, _no_rnd, vshrn.u16, 1
40
+function ff_put_pixels_clamped_neon, export=1
41
+ vld1.64 {d16-d19}, [r0,:128]!
43
+ vld1.64 {d20-d23}, [r0,:128]!
45
+ vld1.64 {d24-d27}, [r0,:128]!
47
+ vld1.64 {d28-d31}, [r0,:128]!
49
+ vst1.64 {d0}, [r1,:64], r2
51
+ vst1.64 {d1}, [r1,:64], r2
53
+ vst1.64 {d2}, [r1,:64], r2
55
+ vst1.64 {d3}, [r1,:64], r2
57
+ vst1.64 {d4}, [r1,:64], r2
58
+ vst1.64 {d5}, [r1,:64], r2
59
+ vst1.64 {d6}, [r1,:64], r2
60
+ vst1.64 {d7}, [r1,:64], r2
64
function ff_put_signed_pixels_clamped_neon, export=1
66
vld1.64 {d16-d17}, [r0,:128]!