Vectorize audio resampling for ARM NEON. #3745
Conversation
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [4171295]: BUILD STARTED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
| vgetq_lane_s32(i, 2), | ||
| vgetq_lane_s32(i, 3) | ||
| }; | ||
| float32x2_t c0 = vld1_f32(&lookup[idx[0]]); |
dali/kernels/signal/resampling.h
Outdated
| f4 = vfmaq_f32(f4, vld1q_f32(in_block_ptr + i), w4); | ||
| x4 = vaddq_f32(x4, vdupq_n_f32(4)); | ||
| } | ||
| // Reduce elements in f4 |
There was a problem hiding this comment.
You can add the same comment in L214 - it was unclear for me why is happening there.
dali/kernels/signal/resampling.h
Outdated
| float f = 0; | ||
| int i = i0; | ||
|
|
||
| #ifdef __ARM_NEON |
There was a problem hiding this comment.
Maybe we can extract this vectorized parts into separate functions, instead having a lot of variants for different architectures in one body.
There was a problem hiding this comment.
Possibly, but I'd have to re-evaluate the performance. There's quite a bit of variables that are modified in the loop and they all would need to be passed by reference.
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [4171320]: BUILD STARTED |
|
CI MESSAGE: [4173226]: BUILD STARTED |
|
CI MESSAGE: [4173226]: BUILD FAILED |
|
CI MESSAGE: [4180682]: BUILD STARTED |
dali/kernels/signal/resampling.h
Outdated
| int i = i_ref; | ||
| float32x4_t x4 = vaddq_f32(vdupq_n_f32(i - in_pos), _0123); | ||
|
|
||
| for (; i + 3 <= i1; i += 4) { |
There was a problem hiding this comment.
Shouldn't you change that as well like in L174?
mzient
force-pushed
the
AudioResamplingNEON
branch
from
e91d112
to
907fd20
Compare
|
CI MESSAGE: [4180791]: BUILD STARTED |
|
CI MESSAGE: [4180791]: BUILD PASSED |
| #ifdef __ARM_NEON | ||
|
|
||
| inline float32x4_t vsetq_f32(float x0, float x1, float x2, float x3) { | ||
| float32x4_t x; |
There was a problem hiding this comment.
Very little nitpick, could be float32x4_t x = vdubpq_n_f32(x0);
| #ifdef __SSE2__ | ||
| inline __m128 operator()(__m128 x) const { | ||
| __m128 fi = _mm_add_ps(x * _mm_set1_ps(scale), _mm_set1_ps(center)); | ||
| __m128i i = _mm_cvtps_epi32(fi); | ||
| __m128i i = _mm_cvttps_epi32(fi); |
There was a problem hiding this comment.
Was doing conversion without truncate an issue before?
There was a problem hiding this comment.
Yes, it was a bug - I detected it now that there's a slight difference between ARM and SSE and dug deeper to find out that it's SSE that's wrong.
* Vectorize audio resampling for ARM NEON. * Fix rounding mode in SSE vectorization. Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
* Vectorize audio resampling for ARM NEON. * Fix rounding mode in SSE vectorization. Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
Signed-off-by: Michał Zientkiewicz mzient@gmail.com
Category:
Other Performance optimization (main purpose)
Bug fix (additional, in the same file)
Description:
Implements vectorized single-channel audio resampling for ARM NEON.
Additional information:
Vectorization is done differently than on SSE (pairwise loads are faster on ARM).
Other changes include improved handling of floor function.
Bug fix: SSE implementation used rounding instead of truncation - fixed.
Affected modules and functionalities:
Audio resampling (audio decoder).
Key points relevant for the review:
Checklist
Tests
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: DALI-2651