diff options
| -rw-r--r-- | modules/audio_processing/aec/Android.mk | 3 | ||||
| -rw-r--r-- | modules/audio_processing/aec/aec_rdft.c | 6 | ||||
| -rw-r--r-- | modules/audio_processing/aec/aec_rdft.h | 3 | ||||
| -rw-r--r-- | modules/audio_processing/aec/aec_rdft_neon.c | 77 | ||||
| -rw-r--r-- | modules/audio_processing/audio_processing.gypi | 1 |
5 files changed, 89 insertions, 1 deletions
diff --git a/modules/audio_processing/aec/Android.mk b/modules/audio_processing/aec/Android.mk index 181e87d9..b86ef5f9 100644 --- a/modules/audio_processing/aec/Android.mk +++ b/modules/audio_processing/aec/Android.mk @@ -53,6 +53,7 @@ include $(BUILD_STATIC_LIBRARY) ifeq ($(WEBRTC_BUILD_NEON_LIBS),true) LOCAL_SRC_FILES += \ - aec_core_neon.c + aec_core_neon.c \ + aec_rdft_neon.c endif # ifeq ($(WEBRTC_BUILD_NEON_LIBS),true) diff --git a/modules/audio_processing/aec/aec_rdft.c b/modules/audio_processing/aec/aec_rdft.c index 7731b37b..5b1c2210 100644 --- a/modules/audio_processing/aec/aec_rdft.c +++ b/modules/audio_processing/aec/aec_rdft.c @@ -287,6 +287,9 @@ static void cft1st_128_C(float* a) { float wk1r, wk1i, wk2r, wk2i, wk3r, wk3i; float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i; + // The processing of the first set of elements was simplified in C to avoid + // some operations (multiplication by zero or one, addition of two elements + // multiplied by the same weight, ...). x0r = a[0] + a[2]; x0i = a[1] + a[3]; x1r = a[0] - a[2]; @@ -660,6 +663,9 @@ void aec_rdft_init(void) { #if defined(MIPS_FPU_LE) aec_rdft_init_mips(); #endif +#if defined(WEBRTC_DETECT_ARM_NEON) || defined(WEBRTC_ARCH_ARM_NEON) + aec_rdft_init_neon(); +#endif // init library constants. makewt_32(); makect_32(); diff --git a/modules/audio_processing/aec/aec_rdft.h b/modules/audio_processing/aec/aec_rdft.h index 795c57d4..94301601 100644 --- a/modules/audio_processing/aec/aec_rdft.h +++ b/modules/audio_processing/aec/aec_rdft.h @@ -60,5 +60,8 @@ void aec_rdft_inverse_128(float* a); #if defined(MIPS_FPU_LE) void aec_rdft_init_mips(void); #endif +#if defined(WEBRTC_DETECT_ARM_NEON) || defined(WEBRTC_ARCH_ARM_NEON) +void aec_rdft_init_neon(void); +#endif #endif // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_RDFT_H_ diff --git a/modules/audio_processing/aec/aec_rdft_neon.c b/modules/audio_processing/aec/aec_rdft_neon.c new file mode 100644 index 00000000..79b23e2c --- /dev/null +++ b/modules/audio_processing/aec/aec_rdft_neon.c @@ -0,0 +1,77 @@ +/* + * Copyright (c) 2014 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +/* + * The rdft AEC algorithm, neon version of speed-critical functions. + * + * Based on the sse2 version. + */ + + +#include "webrtc/modules/audio_processing/aec/aec_rdft.h" + +#include <arm_neon.h> + +static const ALIGN16_BEG float ALIGN16_END + k_swap_sign[4] = {-1.f, 1.f, -1.f, 1.f}; + +static void cft1st_128_neon(float* a) { + const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign); + int j, k2; + + for (k2 = 0, j = 0; j < 128; j += 16, k2 += 4) { + float32x4_t a00v = vld1q_f32(&a[j + 0]); + float32x4_t a04v = vld1q_f32(&a[j + 4]); + float32x4_t a08v = vld1q_f32(&a[j + 8]); + float32x4_t a12v = vld1q_f32(&a[j + 12]); + float32x4_t a01v = vcombine_f32(vget_low_f32(a00v), vget_low_f32(a08v)); + float32x4_t a23v = vcombine_f32(vget_high_f32(a00v), vget_high_f32(a08v)); + float32x4_t a45v = vcombine_f32(vget_low_f32(a04v), vget_low_f32(a12v)); + float32x4_t a67v = vcombine_f32(vget_high_f32(a04v), vget_high_f32(a12v)); + const float32x4_t wk1rv = vld1q_f32(&rdft_wk1r[k2]); + const float32x4_t wk1iv = vld1q_f32(&rdft_wk1i[k2]); + const float32x4_t wk2rv = vld1q_f32(&rdft_wk2r[k2]); + const float32x4_t wk2iv = vld1q_f32(&rdft_wk2i[k2]); + const float32x4_t wk3rv = vld1q_f32(&rdft_wk3r[k2]); + const float32x4_t wk3iv = vld1q_f32(&rdft_wk3i[k2]); + float32x4_t x0v = vaddq_f32(a01v, a23v); + const float32x4_t x1v = vsubq_f32(a01v, a23v); + const float32x4_t x2v = vaddq_f32(a45v, a67v); + const float32x4_t x3v = vsubq_f32(a45v, a67v); + const float32x4_t x3w = vrev64q_f32(x3v); + float32x4_t x0w; + a01v = vaddq_f32(x0v, x2v); + x0v = vsubq_f32(x0v, x2v); + x0w = vrev64q_f32(x0v); + a45v = vmulq_f32(wk2rv, x0v); + a45v = vmlaq_f32(a45v, wk2iv, x0w); + x0v = vmlaq_f32(x1v, x3w, vec_swap_sign); + x0w = vrev64q_f32(x0v); + a23v = vmulq_f32(wk1rv, x0v); + a23v = vmlaq_f32(a23v, wk1iv, x0w); + x0v = vmlsq_f32(x1v, x3w, vec_swap_sign); + x0w = vrev64q_f32(x0v); + a67v = vmulq_f32(wk3rv, x0v); + a67v = vmlaq_f32(a67v, wk3iv, x0w); + a00v = vcombine_f32(vget_low_f32(a01v), vget_low_f32(a23v)); + a04v = vcombine_f32(vget_low_f32(a45v), vget_low_f32(a67v)); + a08v = vcombine_f32(vget_high_f32(a01v), vget_high_f32(a23v)); + a12v = vcombine_f32(vget_high_f32(a45v), vget_high_f32(a67v)); + vst1q_f32(&a[j + 0], a00v); + vst1q_f32(&a[j + 4], a04v); + vst1q_f32(&a[j + 8], a08v); + vst1q_f32(&a[j + 12], a12v); + } +} + +void aec_rdft_init_neon(void) { + cft1st_128 = cft1st_128_neon; +} + diff --git a/modules/audio_processing/audio_processing.gypi b/modules/audio_processing/audio_processing.gypi index b1d18c5b..98d437f8 100644 --- a/modules/audio_processing/audio_processing.gypi +++ b/modules/audio_processing/audio_processing.gypi @@ -200,6 +200,7 @@ ], 'sources': [ 'aec/aec_core_neon.c', + 'aec/aec_rdft_neon.c', 'aecm/aecm_core_neon.c', 'ns/nsx_core_neon.c', ], |