generic-library/webrtc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Clear 2 unused functions in audio processing aecm module.

Browse Source 
unused functions:
WebRtcAecm_WindowAndFFTNeon
WebRtcAecm_InverseFFTAndWindowNeon

BUG=3580
R=andrew@webrtc.org

Change-Id: I12c50a8706d40f9ea98208b5733c00ede7b1f435

Review URL: https://webrtc-codereview.appspot.com/30269004

Patch from Zhongwei Yao <zhongwei.yao@arm.com>.

git-svn-id: http://webrtc.googlecode.com/svn/trunk@7782 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
andrew@webrtc.org 2014-12-02 18:33:52 +00:00
parent beee9cec22
commit ac68ef9ad4
1 changed files with 0 additions and 198 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

198
webrtc/modules/audio_processing/aecm/aecm_core_neon.c
View File

@ -31,204 +31,6 @@ const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END = {
16111, 16179, 16237, 16286, 16325, 16354, 16373, 16384
};
// Square root of Hanning window in Q14, in reversed order.
static const ALIGN8_BEG int16_t kSqrtHanningReversed[] ALIGN8_END = {
16384, 16373, 16354, 16325, 16286, 16237, 16179, 16111,
16034, 15947, 15851, 15746, 15631, 15506, 15373, 15231,
15079, 14918, 14749, 14571, 14384, 14189, 13985, 13773,
13553, 13325, 13089, 12845, 12594, 12335, 12068, 11795,
11514, 11227, 10933, 10633, 10326, 10013, 9695, 9370,
9040, 8705, 8364, 8019, 7668, 7313, 6954, 6591,
6224, 5853, 5478, 5101, 4720, 4337, 3951, 3562,
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,
complex16_t* freq_signal,
int time_signal_scaling) {
int i = 0;
const int16_t* p_time_signal = time_signal;
const int16_t* p_time_signal_offset = &time_signal[PART_LEN];
const int16_t* p_hanning = WebRtcAecm_kSqrtHanning;
const int16_t* p_hanning_reversed = kSqrtHanningReversed;
int16_t* p_fft = fft;
int16_t* p_fft_offset = &fft[PART_LEN2];
assert((uintptr_t)p_time_signal % 8 == 0);
assert((uintptr_t)freq_signal % 32 == 0);
assert((uintptr_t)p_hanning % 8 == 0);
assert((uintptr_t)p_fft % 16 == 0);
__asm __volatile(
"vdup.16 d16, %0\n\t"
"vmov.i16 d21, #0\n\t"
"vmov.i16 d27, #0\n\t"
:
:"r"(time_signal_scaling)
: "d16", "d21", "d27"
);
for (i = 0; i < PART_LEN; i += 4) {
__asm __volatile(
"vld1.16 d0, [%[p_time_signal], :64]!\n\t"
"vld1.16 d22, [%[p_time_signal_offset], :64]!\n\t"
"vld1.16 d17, [%[p_hanning], :64]!\n\t"
"vld1.16 d23, [%[p_hanning_reversed], :64]!\n\t"
"vshl.s16 d18, d0, d16\n\t"
"vshl.s16 d22, d22, d16\n\t"
"vmull.s16 q9, d18, d17\n\t"
"vmull.s16 q12, d22, d23\n\t"
"vshrn.i32 d20, q9, #14\n\t"
"vshrn.i32 d26, q12, #14\n\t"
"vst2.16 {d20, d21}, [%[p_fft], :128]!\n\t"
"vst2.16 {d26, d27}, [%[p_fft_offset], :128]!\n\t"
:[p_time_signal]"+r"(p_time_signal),
[p_time_signal_offset]"+r"(p_time_signal_offset),
[p_hanning]"+r"(p_hanning),
[p_hanning_reversed]"+r"(p_hanning_reversed),
[p_fft]"+r"(p_fft),
[p_fft_offset]"+r"(p_fft_offset)
:
:"d0", "d16", "d17", "d18", "d19", "d20", "d21",
"d22", "d23", "d24", "d25", "d26", "d27"
);
}
// Do forward FFT, then take only the first PART_LEN complex samples,
// and change signs of the imaginary parts.
WebRtcSpl_RealForwardFFT(aecm->real_fft, (int16_t*)fft,
(int16_t*)freq_signal);
for (i = 0; i < PART_LEN; i += 8) {
__asm __volatile(
"vld2.16 {d20, d21, d22, d23}, [%[freq_signal], :256]\n\t"
"vneg.s16 d22, d22\n\t"
"vneg.s16 d23, d23\n\t"
"vst2.16 {d20, d21, d22, d23}, [%[freq_signal], :256]!\n\t"
:[freq_signal]"+r"(freq_signal)
:
: "d20", "d21", "d22", "d23"
);
}
}
void WebRtcAecm_InverseFFTAndWindowNeon(AecmCore_t* aecm,
int16_t* fft,
complex16_t* efw,
int16_t* output,
const int16_t* nearendClean) {
int i, j, outCFFT;
assert((uintptr_t)efw % 32 == 0);
assert((uintptr_t)fft % 16 == 0);
assert((uintptr_t)output% 8 == 0);
assert((uintptr_t)WebRtcAecm_kSqrtHanning % 8 == 0);
assert((uintptr_t)kSqrtHanningReversed % 8 == 0);
assert((uintptr_t)(aecm->outBuf) % 8 == 0);
assert((uintptr_t)(aecm->xBuf) % 32 == 0);
assert((uintptr_t)(aecm->dBufNoisy) % 32 == 0);
assert((uintptr_t)(aecm->dBufClean) % 32 == 0);
// Synthesis
complex16_t* p_efw = efw;
int16_t* p_fft = fft;
int16_t* p_fft_offset = &fft[PART_LEN4 - 6];
for (i = 0, j = 0; i < PART_LEN; i += 4, j += 8) {
// We overwrite two more elements in fft[], but it's ok.
__asm __volatile(
"vld2.16 {q10}, [%[p_efw], :128]!\n\t"
"vmov q11, q10\n\t"
"vneg.s16 d23, d23\n\t"
"vst2.16 {d22, d23}, [%[p_fft], :128]!\n\t"
"vrev64.16 q10, q10\n\t"
"vst2.16 {q10}, [%[p_fft_offset]], %[offset]\n\t"
:[p_efw]"+r"(p_efw),
[p_fft]"+r"(p_fft),
[p_fft_offset]"+r"(p_fft_offset)
:[offset]"r"(-16)
:"d20", "d21", "d22", "d23"
);
}
fft[PART_LEN2] = efw[PART_LEN].real;
fft[PART_LEN2 + 1] = -efw[PART_LEN].imag;
// Inverse FFT. Then take only the real values, and keep outCFFT
// to scale the samples.
outCFFT = WebRtcSpl_RealInverseFFT(aecm->real_fft, fft, (int16_t*)efw);
int32x4_t tmp32x4_2;
__asm __volatile("vdup.32 %q0, %1" : "=w"(tmp32x4_2) : "r"((int32_t)
(outCFFT - aecm->dfaCleanQDomain)));
for (i = 0; i < PART_LEN; i += 4) {
int16x4_t tmp16x4_0;
int16x4_t tmp16x4_1;
int32x4_t tmp32x4_0;
int32x4_t tmp32x4_1;
//efw[i].real = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
// efw[i].real, WebRtcAecm_kSqrtHanning[i], 14);
__asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_0) : "r"(&efw[i].real));
__asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_1) : "r"(&WebRtcAecm_kSqrtHanning[i]));
__asm __volatile("vmull.s16 %q0, %P1, %P2" : "=w"(tmp32x4_0) : "w"(tmp16x4_0), "w"(tmp16x4_1));
__asm __volatile("vrshr.s32 %q0, %q1, #14" : "=w"(tmp32x4_0) : "0"(tmp32x4_0));
//tmp32no1 = WEBRTC_SPL_SHIFT_W32((int32_t)efw[i].real,
// outCFFT - aecm->dfaCleanQDomain);
__asm __volatile("vshl.s32 %q0, %q1, %q2" : "=w"(tmp32x4_0) : "0"(tmp32x4_0), "w"(tmp32x4_2));
//efw[i].real = (int16_t)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX,
// tmp32no1 + aecm->outBuf[i], WEBRTC_SPL_WORD16_MIN);
// output[i] = efw[i].real;
__asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_0) : "r"(&aecm->outBuf[i]));
__asm __volatile("vmovl.s16 %q0, %P1" : "=w"(tmp32x4_1) : "w"(tmp16x4_0));
__asm __volatile("vadd.i32 %q0, %q1" : : "w"(tmp32x4_0), "w"(tmp32x4_1));
__asm __volatile("vqmovn.s32 %P0, %q1" : "=w"(tmp16x4_0) : "w"(tmp32x4_0));
__asm __volatile("vst1.16 %P0, [%1, :64]" : : "w"(tmp16x4_0), "r"(&efw[i].real));
__asm __volatile("vst1.16 %P0, [%1, :64]" : : "w"(tmp16x4_0), "r"(&output[i]));
// tmp32no1 = WEBRTC_SPL_MUL_16_16_RSFT(
// efw[PART_LEN + i].real, WebRtcAecm_kSqrtHanning[PART_LEN - i], 14);
__asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_0) : "r"(&efw[PART_LEN + i].real));
__asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_1) : "r"(&kSqrtHanningReversed[i]));
__asm __volatile("vmull.s16 %q0, %P1, %P2" : "=w"(tmp32x4_0) : "w"(tmp16x4_0), "w"(tmp16x4_1));
__asm __volatile("vshr.s32 %q0, %q1, #14" : "=w"(tmp32x4_0) : "0"(tmp32x4_0));
// tmp32no1 = WEBRTC_SPL_SHIFT_W32(tmp32no1, outCFFT - aecm->dfaCleanQDomain);
__asm __volatile("vshl.s32 %q0, %q1, %q2" : "=w"(tmp32x4_0) : "0"(tmp32x4_0), "w"(tmp32x4_2));
// aecm->outBuf[i] = (int16_t)WEBRTC_SPL_SAT(
// WEBRTC_SPL_WORD16_MAX, tmp32no1, WEBRTC_SPL_WORD16_MIN);
__asm __volatile("vqmovn.s32 %P0, %q1" : "=w"(tmp16x4_0) : "w"(tmp32x4_0));
__asm __volatile("vst1.16 %P0, [%1, :64]" : : "w"(tmp16x4_0), "r"(&aecm->outBuf[i]));
}
// Copy the current block to the old position (outBuf is shifted elsewhere).
for (i = 0; i < PART_LEN; i += 16) {
__asm __volatile("vld1.16 {d20, d21, d22, d23}, [%0, :256]" : :
"r"(&aecm->xBuf[i + PART_LEN]) : "q10");
__asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : : "r"(&aecm->xBuf[i]): "q10");
}
for (i = 0; i < PART_LEN; i += 16) {
__asm __volatile("vld1.16 {d20, d21, d22, d23}, [%0, :256]" : :
"r"(&aecm->dBufNoisy[i + PART_LEN]) : "q10");
__asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : :
"r"(&aecm->dBufNoisy[i]): "q10");
}
if (nearendClean != NULL) {
for (i = 0; i < PART_LEN; i += 16) {
__asm __volatile("vld1.16 {d20, d21, d22, d23}, [%0, :256]" : :
"r"(&aecm->dBufClean[i + PART_LEN]) : "q10");
__asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : :
"r"(&aecm->dBufClean[i]): "q10");
}
}
}
#endif //WEBRTC_ARCH_ARM64_NEON
static inline void AddLanes(uint32_t* ptr, uint32x4_t v) {
#if defined(__aarch64__)
*(ptr) = vaddvq_u32(v);