Enabling building with NEON on ARM64
This patch enables NEON on ARM64 platform. Passed building both on
Android ARMv7 and Android ARM64.
BUG=3580
R=andrew@webrtc.org, jridges@masque.com
Review URL: https://webrtc-codereview.appspot.com/25069004
Patch from Zhongwei Yao <zhongwei.yao@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7751 4adac7df-926f-26a2-2b94-8c16560cd09d
diff --git a/webrtc/BUILD.gn b/webrtc/BUILD.gn
index 3f337fd..ea14644 100644
--- a/webrtc/BUILD.gn
+++ b/webrtc/BUILD.gn
@@ -122,15 +122,21 @@
if (cpu_arch == "arm64") {
defines += [ "WEBRTC_ARCH_ARM" ]
+ # TODO(zhongwei) Defining an unique WEBRTC_NEON and
+ # distinguishing ARMv7 NEON and ARM64 NEON by
+ # WEBRTC_ARCH_ARM_V7 and WEBRTC_ARCH_ARM64 should be better.
+
+ # This macro is used to distinguish ARMv7 NEON and ARM64 NEON
+ defines += [ "WEBRTC_ARCH_ARM64_NEON" ]
}
if (cpu_arch == "arm") {
defines += [ "WEBRTC_ARCH_ARM" ]
- if (arm_version == 7) {
+ if (arm_version >= 7) {
defines += [ "WEBRTC_ARCH_ARM_V7" ]
if (arm_use_neon) {
defines += [ "WEBRTC_ARCH_ARM_NEON" ]
- } else {
+ } else if (is_android) {
defines += [ "WEBRTC_DETECT_ARM_NEON" ]
}
}
diff --git a/webrtc/build/arm_neon.gypi b/webrtc/build/arm_neon.gypi
index 30d040b..037dd70 100644
--- a/webrtc/build/arm_neon.gypi
+++ b/webrtc/build/arm_neon.gypi
@@ -24,7 +24,14 @@
'-mfpu=vfpv3-d16',
],
'cflags': [
- '-mfpu=neon',
'-flax-vector-conversions',
],
+ 'conditions': [
+ # "-mfpu=neon" is not requried for arm64 in GCC.
+ ['target_arch!="arm64"', {
+ 'cflags': [
+ '-mfpu=neon',
+ ],
+ }],
+ ],
}
diff --git a/webrtc/build/common.gypi b/webrtc/build/common.gypi
index 6c3cc7b..307686d 100644
--- a/webrtc/build/common.gypi
+++ b/webrtc/build/common.gypi
@@ -153,7 +153,7 @@
'build_libjpeg%': 0,
'enable_protobuf%': 0,
}],
- ['target_arch=="arm" or target_arch=="armv7"', {
+ ['target_arch=="arm" or target_arch=="armv7" or target_arch=="arm64"', {
'prefer_fixed_point%': 1,
}],
['OS!="ios" and (target_arch!="arm" or arm_version>=7)', {
@@ -210,7 +210,7 @@
}, {
'conditions': [
['os_posix==1', {
- 'configurations': {
+ 'configurations': {
'Debug_Base': {
'defines': [
# Chromium's build/common.gypi defines this for all posix
@@ -254,6 +254,12 @@
['target_arch=="arm64"', {
'defines': [
'WEBRTC_ARCH_ARM',
+ # TODO(zhongwei) Defining an unique WEBRTC_NEON and
+ # distinguishing ARMv7 NEON and ARM64 NEON by
+ # WEBRTC_ARCH_ARM_V7 and WEBRTC_ARCH_ARM64 should be better.
+
+ # This macro is used to distinguish ARMv7 NEON and ARM64 NEON
+ 'WEBRTC_ARCH_ARM64_NEON',
],
}],
['target_arch=="arm" or target_arch=="armv7"', {
@@ -261,12 +267,13 @@
'WEBRTC_ARCH_ARM',
],
'conditions': [
- ['arm_version==7', {
+ ['arm_version>=7', {
'defines': ['WEBRTC_ARCH_ARM_V7',],
'conditions': [
['arm_neon==1', {
'defines': ['WEBRTC_ARCH_ARM_NEON',],
- }, {
+ }],
+ ['arm_neon==0 and OS=="android"', {
'defines': ['WEBRTC_DETECT_ARM_NEON',],
}],
],
diff --git a/webrtc/modules/audio_processing/BUILD.gn b/webrtc/modules/audio_processing/BUILD.gn
index 8f5186b..49b5627 100644
--- a/webrtc/modules/audio_processing/BUILD.gn
+++ b/webrtc/modules/audio_processing/BUILD.gn
@@ -129,7 +129,7 @@
deps += [ ":audio_processing_sse2" ]
}
- if (rtc_build_armv7_neon) {
+ if (rtc_build_armv7_neon || cpu_arch == "arm64") {
deps += [ ":audio_processing_neon" ]
}
@@ -187,11 +187,13 @@
}
}
-if (rtc_build_armv7_neon) {
+if (rtc_build_armv7_neon || cpu_arch == "arm64") {
source_set("audio_processing_neon") {
sources = [
"aec/aec_core_neon.c",
"aec/aec_rdft_neon.c",
+ "aecm/aecm_core_neon.c",
+ "ns/nsx_core_neon.c",
]
configs += [ "../..:common_config" ]
@@ -199,21 +201,6 @@
deps = [ "../../common_audio" ]
- if (is_android || is_ios) {
- sources += [
- # TODO(andrew): Re-enable these once webrtc:3580 is resolved.
- #"aecm/aecm_core_neon.S",
- #"ns/nsx_core_neon.S",
- ]
-
- include_dirs = [ target_out_dir ]
- } else {
- sources += [
- "aecm/aecm_core_neon.c",
- "ns/nsx_core_neon.c",
- ]
- }
-
# Enable compilation for the ARM v7 Neon instruction set. This is needed
# since //build/config/arm.gni only enables Neon for iOS, not Android.
# This provides the same functionality as webrtc/build/arm_neon.gypi.
@@ -223,9 +210,13 @@
configs -= [ "//build/config/compiler:compiler_arm_fpu" ]
cflags = [
"-flax-vector-conversions",
- "-mfpu=neon",
]
+ # "-mfpu=neon" is not requried for arm64 in GCC.
+ if (cpu_arch != "arm64") {
+ cflags += [ "-mfpu=neon" ]
+ }
+
# Disable LTO in audio_processing_neon target due to compiler bug.
if (rtc_use_lto) {
cflags -= [
diff --git a/webrtc/modules/audio_processing/aec/aec_core_neon.c b/webrtc/modules/audio_processing/aec/aec_core_neon.c
index a21a954..f2b8a88 100644
--- a/webrtc/modules/audio_processing/aec/aec_core_neon.c
+++ b/webrtc/modules/audio_processing/aec/aec_core_neon.c
@@ -77,6 +77,8 @@
}
}
+// ARM64's arm_neon.h has already defined vdivq_f32 vsqrtq_f32.
+#if !defined (WEBRTC_ARCH_ARM64_NEON)
static float32x4_t vdivq_f32(float32x4_t a, float32x4_t b) {
int i;
float32x4_t x = vrecpeq_f32(b);
@@ -119,6 +121,8 @@
return vmulq_f32(s, x);;
}
+#endif // WEBRTC_ARCH_ARM64_NEON
+
static void ScaleErrorSignalNEON(AecCore* aec, float ef[2][PART_LEN1]) {
const float mu = aec->extended_filter_enabled ? kExtendedMu : aec->normal_mu;
const float error_threshold = aec->extended_filter_enabled ?
diff --git a/webrtc/modules/audio_processing/aecm/aecm_core.c b/webrtc/modules/audio_processing/aecm/aecm_core.c
index 03a0ecf..fc590f7 100644
--- a/webrtc/modules/audio_processing/aecm/aecm_core.c
+++ b/webrtc/modules/audio_processing/aecm/aecm_core.c
@@ -378,7 +378,8 @@
}
// Initialize function pointers for ARM Neon platform.
-#if (defined WEBRTC_DETECT_ARM_NEON || defined WEBRTC_ARCH_ARM_NEON)
+#if (defined WEBRTC_DETECT_ARM_NEON || defined WEBRTC_ARCH_ARM_NEON || \
+ defined WEBRTC_ARCH_ARM64_NEON)
static void WebRtcAecm_InitNeon(void)
{
WebRtcAecm_StoreAdaptiveChannel = WebRtcAecm_StoreAdaptiveChannelNeon;
@@ -532,7 +533,7 @@
{
WebRtcAecm_InitNeon();
}
-#elif defined(WEBRTC_ARCH_ARM_NEON)
+#elif defined(WEBRTC_ARCH_ARM_NEON) || defined(WEBRTC_ARCH_ARM64_NEON)
WebRtcAecm_InitNeon();
#endif
diff --git a/webrtc/modules/audio_processing/aecm/aecm_core.h b/webrtc/modules/audio_processing/aecm/aecm_core.h
index e56ede6..a7f2695 100644
--- a/webrtc/modules/audio_processing/aecm/aecm_core.h
+++ b/webrtc/modules/audio_processing/aecm/aecm_core.h
@@ -416,7 +416,8 @@
// For the above function pointers, functions for generic platforms are declared
// and defined as static in file aecm_core.c, while those for ARM Neon platforms
// are declared below and defined in file aecm_core_neon.s.
-#if (defined WEBRTC_DETECT_ARM_NEON) || defined (WEBRTC_ARCH_ARM_NEON)
+#if (defined WEBRTC_DETECT_ARM_NEON) || defined (WEBRTC_ARCH_ARM_NEON) || \
+ defined (WEBRTC_ARCH_ARM64_NEON)
void WebRtcAecm_CalcLinearEnergiesNeon(AecmCore_t* aecm,
const uint16_t* far_spectrum,
int32_t* echo_est,
diff --git a/webrtc/modules/audio_processing/aecm/aecm_core_neon.c b/webrtc/modules/audio_processing/aecm/aecm_core_neon.c
index ff733b7..bc1b77b 100644
--- a/webrtc/modules/audio_processing/aecm/aecm_core_neon.c
+++ b/webrtc/modules/audio_processing/aecm/aecm_core_neon.c
@@ -43,6 +43,7 @@
3172, 2780, 2386, 1990, 1594, 1196, 798, 399
};
+#ifndef WEBRTC_ARCH_ARM64_NEON
void WebRtcAecm_WindowAndFFTNeon(AecmCore_t* aecm,
int16_t* fft,
const int16_t* time_signal,
@@ -226,6 +227,7 @@
}
}
}
+#endif //WEBRTC_ARCH_ARM64_NEON
static inline void AddLanes(uint32_t* ptr, uint32x4_t v) {
#if defined(__aarch64__)
diff --git a/webrtc/modules/audio_processing/audio_processing.gypi b/webrtc/modules/audio_processing/audio_processing.gypi
index 3ac87b8..b2c9bb4 100644
--- a/webrtc/modules/audio_processing/audio_processing.gypi
+++ b/webrtc/modules/audio_processing/audio_processing.gypi
@@ -138,7 +138,7 @@
['target_arch=="ia32" or target_arch=="x64"', {
'dependencies': ['audio_processing_sse2',],
}],
- ['(target_arch=="arm" and arm_version==7) or target_arch=="armv7"', {
+ ['(target_arch=="arm" and arm_version==7) or target_arch=="armv7" or target_arch=="arm64"', {
'dependencies': ['audio_processing_neon',],
}],
['target_arch=="mipsel" and mips_arch_variant!="r6" and android_webview_build==0', {
@@ -197,7 +197,7 @@
},
],
}],
- ['(target_arch=="arm" and arm_version==7) or target_arch=="armv7"', {
+ ['(target_arch=="arm" and arm_version==7) or target_arch=="armv7" or target_arch=="arm64"', {
'targets': [{
'target_name': 'audio_processing_neon',
'type': 'static_library',
@@ -212,7 +212,7 @@
'ns/nsx_core_neon.c',
],
'conditions': [
- ['OS=="android" or OS=="ios"', {
+ ['(OS=="android" or OS=="ios") and target_arch!="arm64"', {
'dependencies': [
'<(gen_core_neon_offsets_gyp):*',
],
diff --git a/webrtc/modules/audio_processing/ns/nsx_core.c b/webrtc/modules/audio_processing/ns/nsx_core.c
index 28e9dc5..05efa3a 100644
--- a/webrtc/modules/audio_processing/ns/nsx_core.c
+++ b/webrtc/modules/audio_processing/ns/nsx_core.c
@@ -557,7 +557,8 @@
Denormalize WebRtcNsx_Denormalize;
NormalizeRealBuffer WebRtcNsx_NormalizeRealBuffer;
-#if (defined WEBRTC_DETECT_ARM_NEON || defined WEBRTC_ARCH_ARM_NEON)
+#if (defined WEBRTC_DETECT_ARM_NEON || defined WEBRTC_ARCH_ARM_NEON || \
+ defined WEBRTC_ARCH_ARM64_NEON)
// Initialize function pointers for ARM Neon platform.
static void WebRtcNsx_InitNeon(void) {
WebRtcNsx_NoiseEstimation = WebRtcNsx_NoiseEstimationNeon;
@@ -775,7 +776,7 @@
if ((features & kCPUFeatureNEON) != 0) {
WebRtcNsx_InitNeon();
}
-#elif defined(WEBRTC_ARCH_ARM_NEON)
+#elif defined(WEBRTC_ARCH_ARM_NEON) || defined(WEBRTC_ARCH_ARM64_NEON)
WebRtcNsx_InitNeon();
#endif
diff --git a/webrtc/modules/audio_processing/ns/nsx_core.h b/webrtc/modules/audio_processing/ns/nsx_core.h
index 5b3c5e7..9a619b4 100644
--- a/webrtc/modules/audio_processing/ns/nsx_core.h
+++ b/webrtc/modules/audio_processing/ns/nsx_core.h
@@ -218,10 +218,11 @@
uint32_t* priorLocSnr,
uint32_t* postLocSnr);
-#if (defined WEBRTC_DETECT_ARM_NEON) || defined (WEBRTC_ARCH_ARM_NEON)
+#if (defined WEBRTC_DETECT_ARM_NEON || defined WEBRTC_ARCH_ARM_NEON || \
+ defined WEBRTC_ARCH_ARM64_NEON)
// For the above function pointers, functions for generic platforms are declared
// and defined as static in file nsx_core.c, while those for ARM Neon platforms
-// are declared below and defined in file nsx_core_neon.S.
+// are declared below and defined in file nsx_core_neon.c.
void WebRtcNsx_NoiseEstimationNeon(NsxInst_t* inst,
uint16_t* magn,
uint32_t* noise,
