From 2f6f955a178b120f89b3b554b3a12c4a8af82838 Mon Sep 17 00:00:00 2001
From: Johann <johannkoenig@google.com>
Date: Wed, 14 May 2014 11:33:47 -0700
Subject: [PATCH] Remove intermediate step in vp8_dequantize_b

With the intrinsics it is no longer necessary to have a stub/helper
function.

Change-Id: I3695961c3c94f1bb750d3b7b29716e509ebba482
---
 build/make/Android.mk                  |  5 +++--
 vp8/common/arm/dequantize_arm.c        | 17 -----------------
 vp8/common/arm/neon/dequantizeb_neon.c | 12 +++++-------
 vp8/common/rtcd_defs.pl                |  3 +--
 4 files changed, 9 insertions(+), 28 deletions(-)

diff --git a/build/make/Android.mk b/build/make/Android.mk
index 826ff2f81..816334e04 100644
--- a/build/make/Android.mk
+++ b/build/make/Android.mk
@@ -38,8 +38,9 @@
 # For this we import the 'cpufeatures' module from the NDK sources.
 # libvpx can also be configured without this runtime detection method.
 # Configuring with --disable-runtime-cpu-detect will assume presence of NEON.
-# Configuring with --disable-runtime-cpu-detect --disable-neon will remove any
-# NEON dependency.
+# Configuring with --disable-runtime-cpu-detect --disable-neon \
+#     --disable-neon-asm
+# will remove any NEON dependency.
 
 # To change to building armeabi, run ./libvpx/configure again, but with
 # --target=arm5te-android-gcc and modify the Application.mk file to
diff --git a/vp8/common/arm/dequantize_arm.c b/vp8/common/arm/dequantize_arm.c
index 3e37e0892..1f8157f0b 100644
--- a/vp8/common/arm/dequantize_arm.c
+++ b/vp8/common/arm/dequantize_arm.c
@@ -12,26 +12,9 @@
 #include "vpx_config.h"
 #include "vp8/common/blockd.h"
 
-#if HAVE_NEON_ASM
-extern void vp8_dequantize_b_loop_neon(short *Q, short *DQC, short *DQ);
-#endif
-
 #if HAVE_MEDIA
 extern void vp8_dequantize_b_loop_v6(short *Q, short *DQC, short *DQ);
-#endif
 
-#if HAVE_NEON_ASM
-
-void vp8_dequantize_b_neon(BLOCKD *d, short *DQC)
-{
-    short *DQ  = d->dqcoeff;
-    short *Q   = d->qcoeff;
-
-    vp8_dequantize_b_loop_neon(Q, DQC, DQ);
-}
-#endif
-
-#if HAVE_MEDIA
 void vp8_dequantize_b_v6(BLOCKD *d, short *DQC)
 {
     short *DQ  = d->dqcoeff;
diff --git a/vp8/common/arm/neon/dequantizeb_neon.c b/vp8/common/arm/neon/dequantizeb_neon.c
index 60f69c8db..54e709dd3 100644
--- a/vp8/common/arm/neon/dequantizeb_neon.c
+++ b/vp8/common/arm/neon/dequantizeb_neon.c
@@ -10,18 +10,16 @@
 
 #include <arm_neon.h>
 
-void vp8_dequantize_b_loop_neon(
-        int16_t *Q,
-        int16_t *DQC,
-        int16_t *DQ) {
+#include "vp8/common/blockd.h"
+
+void vp8_dequantize_b_neon(BLOCKD *d, short *DQC) {
     int16x8x2_t qQ, qDQC, qDQ;
 
-    qQ   = vld2q_s16(Q);
+    qQ   = vld2q_s16(d->qcoeff);
     qDQC = vld2q_s16(DQC);
 
     qDQ.val[0] = vmulq_s16(qQ.val[0], qDQC.val[0]);
     qDQ.val[1] = vmulq_s16(qQ.val[1], qDQC.val[1]);
 
-    vst2q_s16(DQ, qDQ);
-    return;
+    vst2q_s16(d->dqcoeff, qDQ);
 }
diff --git a/vp8/common/rtcd_defs.pl b/vp8/common/rtcd_defs.pl
index 502f49af9..6908a9728 100644
--- a/vp8/common/rtcd_defs.pl
+++ b/vp8/common/rtcd_defs.pl
@@ -29,9 +29,8 @@ $vp8_clear_system_state_mmx=vpx_reset_mmx_state;
 # Dequant
 #
 add_proto qw/void vp8_dequantize_b/, "struct blockd*, short *dqc";
-specialize qw/vp8_dequantize_b mmx media neon_asm/;
+specialize qw/vp8_dequantize_b mmx media neon/;
 $vp8_dequantize_b_media=vp8_dequantize_b_v6;
-$vp8_dequantize_b_neon_asm=vp8_dequantize_b_neon;
 
 add_proto qw/void vp8_dequant_idct_add/, "short *input, short *dq, unsigned char *output, int stride";
 specialize qw/vp8_dequant_idct_add mmx media neon dspr2/;