diff options
author | bjornv@webrtc.org <bjornv@webrtc.org@4adac7df-926f-26a2-2b94-8c16560cd09d> | 2014-07-10 07:53:13 +0000 |
---|---|---|
committer | bjornv@webrtc.org <bjornv@webrtc.org@4adac7df-926f-26a2-2b94-8c16560cd09d> | 2014-07-10 07:53:13 +0000 |
commit | 71ba40d529e6643bf7e64a1d4b3e6d5e6efb56b9 (patch) | |
tree | c3b05b941b9233df8a77f211433609389e46e200 /modules | |
parent | 91b438940b4c2d3252fe67545d84f70cd54be37c (diff) | |
download | webrtc-71ba40d529e6643bf7e64a1d4b3e6d5e6efb56b9.tar.gz |
Neon version of rftbsub_128()
The performance gain on a Nexus 7 reported by audioproc is ~4.5%
The output is bit exact.
BUG=3131
TESTED=trybots and manually
R=bjornv@webrtc.org, cd@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/19919005
Patch from Scott LaVarnway <slavarnw@gmail.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk/webrtc@6646 4adac7df-926f-26a2-2b94-8c16560cd09d
Diffstat (limited to 'modules')
-rw-r--r-- | modules/audio_processing/aec/aec_rdft_neon.c | 112 |
1 files changed, 97 insertions, 15 deletions
diff --git a/modules/audio_processing/aec/aec_rdft_neon.c b/modules/audio_processing/aec/aec_rdft_neon.c index a9c79b7b..43b6a68c 100644 --- a/modules/audio_processing/aec/aec_rdft_neon.c +++ b/modules/audio_processing/aec/aec_rdft_neon.c @@ -187,19 +187,18 @@ __inline static float32x4_t reverse_order_f32x4(float32x4_t in) { static void rftfsub_128_neon(float* a) { const float* c = rdft_w + 32; - int j1, j2, k1, k2; - float wkr, wki, xr, xi, yr, yi; + int j1, j2; const float32x4_t mm_half = vdupq_n_f32(0.5f); // Vectorized code (four at once). // Note: commented number are indexes for the first iteration of the loop. for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) { // Load 'wk'. - const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4, - const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31, - const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31, - const float32x4_t wkr_ = reverse_order_f32x4(wkrt); - const float32x4_t wki_ = c_j1; // 1, 2, 3, 4, + const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4, + const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31, + const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31, + const float32x4_t wkr_ = reverse_order_f32x4(wkrt); // 31, 30, 29, 28, + const float32x4_t wki_ = c_j1; // 1, 2, 3, 4, // Load and shuffle 'a'. // 2, 4, 6, 8, 3, 5, 7, 9 float32x4x2_t a_j2_p = vld2q_f32(&a[0 + j2]); @@ -250,14 +249,14 @@ static void rftfsub_128_neon(float* a) { // Scalar code for the remaining items. for (; j2 < 64; j1 += 1, j2 += 2) { - k2 = 128 - j2; - k1 = 32 - j1; - wkr = 0.5f - c[k1]; - wki = c[j1]; - xr = a[j2 + 0] - a[k2 + 0]; - xi = a[j2 + 1] + a[k2 + 1]; - yr = wkr * xr - wki * xi; - yi = wkr * xi + wki * xr; + const int k2 = 128 - j2; + const int k1 = 32 - j1; + const float wkr = 0.5f - c[k1]; + const float wki = c[j1]; + const float xr = a[j2 + 0] - a[k2 + 0]; + const float xi = a[j2 + 1] + a[k2 + 1]; + const float yr = wkr * xr - wki * xi; + const float yi = wkr * xi + wki * xr; a[j2 + 0] -= yr; a[j2 + 1] -= yi; a[k2 + 0] += yr; @@ -265,9 +264,92 @@ static void rftfsub_128_neon(float* a) { } } +static void rftbsub_128_neon(float* a) { + const float* c = rdft_w + 32; + int j1, j2; + const float32x4_t mm_half = vdupq_n_f32(0.5f); + + a[1] = -a[1]; + // Vectorized code (four at once). + // Note: commented number are indexes for the first iteration of the loop. + for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) { + // Load 'wk'. + const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4, + const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31, + const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31, + const float32x4_t wkr_ = reverse_order_f32x4(wkrt); // 31, 30, 29, 28, + const float32x4_t wki_ = c_j1; // 1, 2, 3, 4, + // Load and shuffle 'a'. + // 2, 4, 6, 8, 3, 5, 7, 9 + float32x4x2_t a_j2_p = vld2q_f32(&a[0 + j2]); + // 120, 122, 124, 126, 121, 123, 125, 127, + const float32x4x2_t k2_0_4 = vld2q_f32(&a[122 - j2]); + // 126, 124, 122, 120 + const float32x4_t a_k2_p0 = reverse_order_f32x4(k2_0_4.val[0]); + // 127, 125, 123, 121 + const float32x4_t a_k2_p1 = reverse_order_f32x4(k2_0_4.val[1]); + // Calculate 'x'. + const float32x4_t xr_ = vsubq_f32(a_j2_p.val[0], a_k2_p0); + // 2-126, 4-124, 6-122, 8-120, + const float32x4_t xi_ = vaddq_f32(a_j2_p.val[1], a_k2_p1); + // 3-127, 5-125, 7-123, 9-121, + // Calculate product into 'y'. + // yr = wkr * xr - wki * xi; + // yi = wkr * xi + wki * xr; + const float32x4_t a_ = vmulq_f32(wkr_, xr_); + const float32x4_t b_ = vmulq_f32(wki_, xi_); + const float32x4_t c_ = vmulq_f32(wkr_, xi_); + const float32x4_t d_ = vmulq_f32(wki_, xr_); + const float32x4_t yr_ = vaddq_f32(a_, b_); // 2-126, 4-124, 6-122, 8-120, + const float32x4_t yi_ = vsubq_f32(c_, d_); // 3-127, 5-125, 7-123, 9-121, + // Update 'a'. + // a[j2 + 0] -= yr; + // a[j2 + 1] -= yi; + // a[k2 + 0] += yr; + // a[k2 + 1] -= yi; + // 126, 124, 122, 120, + const float32x4_t a_k2_p0n = vaddq_f32(a_k2_p0, yr_); + // 127, 125, 123, 121, + const float32x4_t a_k2_p1n = vsubq_f32(yi_, a_k2_p1); + // Shuffle in right order and store. + // 2, 3, 4, 5, 6, 7, 8, 9, + const float32x4_t a_k2_p0nr = vrev64q_f32(a_k2_p0n); + const float32x4_t a_k2_p1nr = vrev64q_f32(a_k2_p1n); + // 124, 125, 126, 127, 120, 121, 122, 123 + const float32x4x2_t a_k2_n = vzipq_f32(a_k2_p0nr, a_k2_p1nr); + // 2, 4, 6, 8, + a_j2_p.val[0] = vsubq_f32(a_j2_p.val[0], yr_); + // 3, 5, 7, 9, + a_j2_p.val[1] = vsubq_f32(yi_, a_j2_p.val[1]); + // 2, 3, 4, 5, 6, 7, 8, 9, + vst2q_f32(&a[0 + j2], a_j2_p); + + vst1q_f32(&a[122 - j2], a_k2_n.val[1]); + vst1q_f32(&a[126 - j2], a_k2_n.val[0]); + } + + // Scalar code for the remaining items. + for (; j2 < 64; j1 += 1, j2 += 2) { + const int k2 = 128 - j2; + const int k1 = 32 - j1; + const float wkr = 0.5f - c[k1]; + const float wki = c[j1]; + const float xr = a[j2 + 0] - a[k2 + 0]; + const float xi = a[j2 + 1] + a[k2 + 1]; + const float yr = wkr * xr + wki * xi; + const float yi = wkr * xi - wki * xr; + a[j2 + 0] = a[j2 + 0] - yr; + a[j2 + 1] = yi - a[j2 + 1]; + a[k2 + 0] = yr + a[k2 + 0]; + a[k2 + 1] = yi - a[k2 + 1]; + } + a[65] = -a[65]; +} + void aec_rdft_init_neon(void) { cft1st_128 = cft1st_128_neon; cftmdl_128 = cftmdl_128_neon; rftfsub_128 = rftfsub_128_neon; + rftbsub_128 = rftbsub_128_neon; } |