allanger/ardour
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Activity

aarch64: fix unaligned count and peak computation

Browse Source 
Clamp misaligned prefix loops to the remaining frame count to avoid
nframes underflow and potential out-of-bounds access for small buffers.

Apply vabsq_f32() to all SIMD loads before peak reduction to ensure
correct absolute peak calculation in unrolled NEON paths.
This commit is contained in:
Ayan Shafqat
2025-12-29 20:12:43 -05:00
parent b673989763
commit 98eab68044
1 changed files with 26 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
libs/ardour/aarch64_neon_functions.cc
View File

@@ -24,6 +24,7 @@
#include <arm_neon.h>
#include <cstddef>
#include <algorithm>
/**
* @brief Aligns a pointer to the next 16-byte boundary
@@ -78,12 +79,18 @@ arm_neon_compute_peak(const float* src, uint32_t nframes, float current)
if (UNLIKELY(src_aligned != src))
{
size_t unaligned_count = src_aligned - src;
for (size_t i = 0; i < unaligned_count; i++)
// Handle small number of nframes
size_t count = std::min<size_t>(unaligned_count, nframes);
for (size_t i = 0; i < count; i++)
{
float32x4_t x0 = vld1q_dup_f32(src + i);
x0 = vabsq_f32(x0);
vmax = vmaxq_f32(vmax, x0);
}
nframes -= unaligned_count;
nframes -= count;
}
// Compute the number of SIMD frames
@@ -106,6 +113,11 @@ arm_neon_compute_peak(const float* src, uint32_t nframes, float current)
x2 = vld1q_f32(src_aligned + offset + (2 * 4));
x3 = vld1q_f32(src_aligned + offset + (3 * 4));
x0 = vabsq_f32(x0);
x1 = vabsq_f32(x1);
x2 = vabsq_f32(x2);
x3 = vabsq_f32(x3);
max0 = vmaxq_f32(x0, x1);
max1 = vmaxq_f32(x2, x3);
max2 = vmaxq_f32(max0, max1);
@@ -182,13 +194,18 @@ arm_neon_find_peaks(const float* src, uint32_t nframes, float* minf, float* maxf
if (UNLIKELY(src_aligned != src))
{
size_t unaligned_count = src_aligned - src;
for (size_t i = 0; i < unaligned_count; i++)
// Handle small number of nframes
size_t count = std::min<size_t>(unaligned_count, nframes);
for (size_t i = 0; i < count; i++)
{
float32x4_t x0 = vld1q_dup_f32(src + i);
vmax = vmaxq_f32(vmax, x0);
vmin = vminq_f32(vmin, x0);
}
nframes -= unaligned_count;
nframes -= count;
}
// Compute the number of SIMD frames
@@ -299,7 +316,9 @@ arm_neon_apply_gain_to_buffer(float* dst, uint32_t nframes, float gain)
if (UNLIKELY(dst_aligned != dst))
{
size_t unaligned_count = dst_aligned - dst;
for (size_t i = 0; i < unaligned_count; i++)
size_t count = std::min<size_t>(unaligned_count, nframes);
for (size_t i = 0; i < count; i++)
{
float32_t x0, y0;
@@ -307,7 +326,8 @@ arm_neon_apply_gain_to_buffer(float* dst, uint32_t nframes, float gain)
y0 = x0 * gain;
dst[i] = y0;
}
nframes -= unaligned_count;
nframes -= count;
}
// Compute the number of SIMD frames