From 6002b15cd111c015149f014c9185c82f2b930d58 Mon Sep 17 00:00:00 2001 From: cschuldt Date: Mon, 15 Nov 2021 11:00:07 +0100 Subject: [PATCH] Optimize ComputeFrequencyResponse(). Reducing pointer following. This will allow the compiler to optimize more efficiently with the "-fno-strict-aliasing" flag. Bug: None Change-Id: Ib1fd3a1cf3f89471b0ec87404650a6061eec5e2d Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/237782 Reviewed-by: Sam Zackrisson Commit-Queue: Christian Schuldt Cr-Commit-Position: refs/heads/main@{#35374} --- .../aec3/adaptive_fir_filter.cc | 32 +++++++++++-------- .../aec3/adaptive_fir_filter_avx2.cc | 16 ++++++---- 2 files changed, 27 insertions(+), 21 deletions(-) diff --git a/modules/audio_processing/aec3/adaptive_fir_filter.cc b/modules/audio_processing/aec3/adaptive_fir_filter.cc index bf3a7809f4..917aa951ee 100644 --- a/modules/audio_processing/aec3/adaptive_fir_filter.cc +++ b/modules/audio_processing/aec3/adaptive_fir_filter.cc @@ -68,19 +68,21 @@ void ComputeFrequencyResponse_Neon( RTC_DCHECK_EQ(H.size(), H2->capacity()); for (size_t p = 0; p < num_partitions; ++p) { RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size()); + auto& H2_p = (*H2)[p]; for (size_t ch = 0; ch < num_render_channels; ++ch) { + const FftData& H_p_ch = H[p][ch]; for (size_t j = 0; j < kFftLengthBy2; j += 4) { - const float32x4_t re = vld1q_f32(&H[p][ch].re[j]); - const float32x4_t im = vld1q_f32(&H[p][ch].im[j]); + const float32x4_t re = vld1q_f32(&H_p_ch.re[j]); + const float32x4_t im = vld1q_f32(&H_p_ch.im[j]); float32x4_t H2_new = vmulq_f32(re, re); H2_new = vmlaq_f32(H2_new, im, im); - float32x4_t H2_p_j = vld1q_f32(&(*H2)[p][j]); + float32x4_t H2_p_j = vld1q_f32(&H2_p[j]); H2_p_j = vmaxq_f32(H2_p_j, H2_new); - vst1q_f32(&(*H2)[p][j], H2_p_j); + vst1q_f32(&H2_p[j], H2_p_j); } - float H2_new = H[p][ch].re[kFftLengthBy2] * H[p][ch].re[kFftLengthBy2] + - H[p][ch].im[kFftLengthBy2] * H[p][ch].im[kFftLengthBy2]; - (*H2)[p][kFftLengthBy2] = std::max((*H2)[p][kFftLengthBy2], H2_new); + float H2_new = H_p_ch.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] + + H_p_ch.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2]; + H2_p[kFftLengthBy2] = std::max(H2_p[kFftLengthBy2], H2_new); } } } @@ -101,20 +103,22 @@ void ComputeFrequencyResponse_Sse2( // constexpr __mmmask8 kMaxMask = static_cast<__mmmask8>(256u); for (size_t p = 0; p < num_partitions; ++p) { RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size()); + auto& H2_p = (*H2)[p]; for (size_t ch = 0; ch < num_render_channels; ++ch) { + const FftData& H_p_ch = H[p][ch]; for (size_t j = 0; j < kFftLengthBy2; j += 4) { - const __m128 re = _mm_loadu_ps(&H[p][ch].re[j]); + const __m128 re = _mm_loadu_ps(&H_p_ch.re[j]); const __m128 re2 = _mm_mul_ps(re, re); - const __m128 im = _mm_loadu_ps(&H[p][ch].im[j]); + const __m128 im = _mm_loadu_ps(&H_p_ch.im[j]); const __m128 im2 = _mm_mul_ps(im, im); const __m128 H2_new = _mm_add_ps(re2, im2); - __m128 H2_k_j = _mm_loadu_ps(&(*H2)[p][j]); + __m128 H2_k_j = _mm_loadu_ps(&H2_p[j]); H2_k_j = _mm_max_ps(H2_k_j, H2_new); - _mm_storeu_ps(&(*H2)[p][j], H2_k_j); + _mm_storeu_ps(&H2_p[j], H2_k_j); } - float H2_new = H[p][ch].re[kFftLengthBy2] * H[p][ch].re[kFftLengthBy2] + - H[p][ch].im[kFftLengthBy2] * H[p][ch].im[kFftLengthBy2]; - (*H2)[p][kFftLengthBy2] = std::max((*H2)[p][kFftLengthBy2], H2_new); + float H2_new = H_p_ch.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] + + H_p_ch.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2]; + H2_p[kFftLengthBy2] = std::max(H2_p[kFftLengthBy2], H2_new); } } } diff --git a/modules/audio_processing/aec3/adaptive_fir_filter_avx2.cc b/modules/audio_processing/aec3/adaptive_fir_filter_avx2.cc index 245b45ac31..6c8c948026 100644 --- a/modules/audio_processing/aec3/adaptive_fir_filter_avx2.cc +++ b/modules/audio_processing/aec3/adaptive_fir_filter_avx2.cc @@ -31,19 +31,21 @@ void ComputeFrequencyResponse_Avx2( RTC_DCHECK_EQ(H.size(), H2->capacity()); for (size_t p = 0; p < num_partitions; ++p) { RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size()); + auto& H2_p = (*H2)[p]; for (size_t ch = 0; ch < num_render_channels; ++ch) { + const FftData& H_p_ch = H[p][ch]; for (size_t j = 0; j < kFftLengthBy2; j += 8) { - __m256 re = _mm256_loadu_ps(&H[p][ch].re[j]); + __m256 re = _mm256_loadu_ps(&H_p_ch.re[j]); __m256 re2 = _mm256_mul_ps(re, re); - __m256 im = _mm256_loadu_ps(&H[p][ch].im[j]); + __m256 im = _mm256_loadu_ps(&H_p_ch.im[j]); re2 = _mm256_fmadd_ps(im, im, re2); - __m256 H2_k_j = _mm256_loadu_ps(&(*H2)[p][j]); + __m256 H2_k_j = _mm256_loadu_ps(&H2_p[j]); H2_k_j = _mm256_max_ps(H2_k_j, re2); - _mm256_storeu_ps(&(*H2)[p][j], H2_k_j); + _mm256_storeu_ps(&H2_p[j], H2_k_j); } - float H2_new = H[p][ch].re[kFftLengthBy2] * H[p][ch].re[kFftLengthBy2] + - H[p][ch].im[kFftLengthBy2] * H[p][ch].im[kFftLengthBy2]; - (*H2)[p][kFftLengthBy2] = std::max((*H2)[p][kFftLengthBy2], H2_new); + float H2_new = H_p_ch.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] + + H_p_ch.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2]; + H2_p[kFftLengthBy2] = std::max(H2_p[kFftLengthBy2], H2_new); } } }