From c9b25dcae4f492b558f2025d96664f4d4b2b42ea Mon Sep 17 00:00:00 2001
From: hkuang <hkuang@google.com>
Date: Tue, 9 Jul 2013 12:06:21 -0700
Subject: [PATCH] Add neon optimize vp9_dc_only_idct_add.

Change-Id: Iae84ab945cc9662a0ddd839aa2b9ca59f2ae5423
---
 vp9/common/arm/neon/vp9_dc_only_idct_add_neon.asm | 69 +++++++++++++++++++++++
 vp9/common/vp9_rtcd_defs.sh                       |  2 +-
 vp9/vp9_common.mk                                 |  1 +
 3 files changed, 71 insertions(+), 1 deletion(-)
 create mode 100644 vp9/common/arm/neon/vp9_dc_only_idct_add_neon.asm

diff --git a/vp9/common/arm/neon/vp9_dc_only_idct_add_neon.asm b/vp9/common/arm/neon/vp9_dc_only_idct_add_neon.asm
new file mode 100644
index 0000000..60a0d98
--- /dev/null
+++ b/vp9/common/arm/neon/vp9_dc_only_idct_add_neon.asm
@@ -0,0 +1,69 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |vp9_dc_only_idct_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vp9_dc_only_idct_add_neon(int input_dc, uint8_t *pred_ptr,
+;                            uint8_t *dst_ptr, int pitch, int stride)
+;
+; r0  int input_dc
+; r1  uint8_t *pred_ptr
+; r2  uint8_t *dst_ptr
+; r3  int pitch
+; sp  int stride
+
+|vp9_dc_only_idct_add_neon| PROC
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    mul              r0, r0, r12               ; input_dc * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; ROUND_POWER_OF_TWO(out, 4)
+    add              r0, r0, #8                ; + (1 <<((4) - 1))
+    asr              r0, r0, #4                ; >> 4
+
+    vdup.16         q0, r0;                   ; duplicate a1
+    ldr              r12, [sp]                 ; load stride
+
+    vld1.32         {d2[0]}, [r1], r3
+    vld1.32         {d2[1]}, [r1], r3
+    vld1.32         {d4[0]}, [r1], r3
+    vld1.32         {d4[1]}, [r1]
+
+    vaddw.u8        q1, q0, d2                ; a1 + pred_ptr[c]
+    vaddw.u8        q2, q0, d4
+
+    vqmovun.s16     d2, q1                    ; clip_pixel
+    vqmovun.s16     d4, q2
+
+    vst1.32         {d2[0]}, [r2], r12
+    vst1.32         {d2[1]}, [r2], r12
+    vst1.32         {d4[0]}, [r2], r12
+    vst1.32         {d4[1]}, [r2]
+
+    bx               lr
+    ENDP             ; |vp9_dc_only_idct_add_neon|
+
+    END
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 56a2284..8d0eb51 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -338,7 +338,7 @@ specialize vp9_idct4_1d sse2
 # dct and add
 
 prototype void vp9_dc_only_idct_add "int input_dc, uint8_t *pred_ptr, uint8_t *dst_ptr, int pitch, int stride"
-specialize vp9_dc_only_idct_add sse2
+specialize vp9_dc_only_idct_add sse2 neon
 
 prototype void vp9_short_iwalsh4x4_1_add "int16_t *input, uint8_t *dest, int dest_stride"
 specialize vp9_short_iwalsh4x4_1_add
diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk
index ee744d5..75548c6 100644
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -89,5 +89,6 @@ endif
 VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.c
 
 VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_dc_only_idct_add_neon$(ASM)
 
 $(eval $(call rtcd_h_template,vp9_rtcd,vp9/common/vp9_rtcd_defs.sh))
-- 
2.7.4