From 524edba9ddcda9f763ea0995407560cd334301eb Mon Sep 17 00:00:00 2001 From: Dmitry Sharshakov Date: Tue, 16 Jan 2024 14:30:09 +0300 Subject: [PATCH] audioconvert: fix rare unaligned load exceptions Supposed causes described in the issue. Also improve float semantics. Fixes #3790 --- spa/plugins/audioconvert/resample-native-avx.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/spa/plugins/audioconvert/resample-native-avx.c b/spa/plugins/audioconvert/resample-native-avx.c index c4b4c605b..a2de72705 100644 --- a/spa/plugins/audioconvert/resample-native-avx.c +++ b/spa/plugins/audioconvert/resample-native-avx.c @@ -16,18 +16,18 @@ static inline void inner_product_avx(float *d, const float * SPA_RESTRICT s, uint32_t n_taps4 = n_taps & ~0xf; for (; i < n_taps4; i += 16) { - ty = (__m256)_mm256_lddqu_si256((__m256i*)(s + i + 0)); + ty = (__m256)_mm256_loadu_ps(s + i + 0); sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(taps + i + 0), sy[0]); - ty = (__m256)_mm256_lddqu_si256((__m256i*)(s + i + 8)); + ty = (__m256)_mm256_loadu_ps(s + i + 8); sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(taps + i + 8), sy[1]); } sy[0] = _mm256_add_ps(sy[1], sy[0]); sx[1] = _mm256_extractf128_ps(sy[0], 1); sx[0] = _mm256_extractf128_ps(sy[0], 0); for (; i < n_taps; i += 8) { - tx = (__m128)_mm_lddqu_si128((__m128i*)(s + i + 0)); + tx = (__m128)_mm_loadu_ps(s + i + 0); sx[0] = _mm_fmadd_ps(tx, _mm_load_ps(taps + i + 0), sx[0]); - tx = (__m128)_mm_lddqu_si128((__m128i*)(s + i + 4)); + tx = (__m128)_mm_loadu_ps(s + i + 4); sx[1] = _mm_fmadd_ps(tx, _mm_load_ps(taps + i + 4), sx[1]); } sx[0] = _mm_add_ps(sx[0], sx[1]); @@ -45,10 +45,10 @@ static inline void inner_product_ip_avx(float *d, const float * SPA_RESTRICT s, uint32_t i, n_taps4 = n_taps & ~0xf; for (i = 0; i < n_taps4; i += 16) { - ty = (__m256)_mm256_lddqu_si256((__m256i*)(s + i + 0)); + ty = (__m256)_mm256_loadu_ps(s + i + 0); sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(t0 + i + 0), sy[0]); sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(t1 + i + 0), sy[1]); - ty = (__m256)_mm256_lddqu_si256((__m256i*)(s + i + 8)); + ty = (__m256)_mm256_loadu_ps(s + i + 8); sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(t0 + i + 8), sy[0]); sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(t1 + i + 8), sy[1]); } @@ -56,10 +56,10 @@ static inline void inner_product_ip_avx(float *d, const float * SPA_RESTRICT s, sx[1] = _mm_add_ps(_mm256_extractf128_ps(sy[1], 0), _mm256_extractf128_ps(sy[1], 1)); for (; i < n_taps; i += 8) { - tx = (__m128)_mm_lddqu_si128((__m128i*)(s + i + 0)); + tx = (__m128)_mm_loadu_ps(s + i + 0); sx[0] = _mm_fmadd_ps(tx, _mm_load_ps(t0 + i + 0), sx[0]); sx[1] = _mm_fmadd_ps(tx, _mm_load_ps(t1 + i + 0), sx[1]); - tx = (__m128)_mm_lddqu_si128((__m128i*)(s + i + 4)); + tx = (__m128)_mm_loadu_ps(s + i + 4); sx[0] = _mm_fmadd_ps(tx, _mm_load_ps(t0 + i + 4), sx[0]); sx[1] = _mm_fmadd_ps(tx, _mm_load_ps(t1 + i + 4), sx[1]); }