<div data-ntes="ntes_mail_body_root" style="line-height:1.7;color:#000000;font-size:14px;font-family:Arial"><div id="spnEditorContent"><p style="margin: 0;">Looks for dct32x32 and idct4x4 only, other size are similar or worse?</p><p style="margin: 0;"><br></p></div><pre>At 2024-11-24 12:34:48, gajjanag2@gmail.com wrote:
>From: Ganesh Ajjanagadde <gajjanag@alum.mit.edu>
>
>Apple silicon has 4 128 bit NEON execution units and benefits from unrolling.
>
>From ./TestBench on an M4 Mac Mini,
>
>before:
>dct8x8               |      2.32x |          205.12   |      476.62
>dct16x16     |      2.02x |          801.20   |      1619.62
>dct32x32     |      3.47x |          7566.39  |      26275.65
>idct4x4              |      0.90x |          175.80   |      157.90
>idct16x16    |      2.05x |          863.30   |      1771.80
>idct32x32    |      1.79x |          6344.33  |      11351.99
>
>after:
>dct8x8               |      2.33x |          204.72   |      476.53
>dct16x16     |      2.04x |          802.16   |      1637.39
>dct32x32     |      4.96x |          5181.02  |      25700.34
>idct4x4              |      1.08x |          162.09   |      174.40
>idct16x16    |      1.95x |          910.01   |      1771.61
>idct32x32    |      1.75x |          6350.72  |      11143.71
>
>~2% end to end encoding speedup
>---
> source/common/aarch64/dct-prim.cpp | 2 ++
> 1 file changed, 2 insertions(+)
>
>diff --git a/source/common/aarch64/dct-prim.cpp b/source/common/aarch64/dct-prim.cpp
>index 8b523ceb0..e6ee7005b 100644
>--- a/source/common/aarch64/dct-prim.cpp
>+++ b/source/common/aarch64/dct-prim.cpp
>@@ -435,6 +435,7 @@ static inline void partialButterfly32_neon(const int16_t *src, int16_t *dst)
>         for (int i = 0; i < line; i += 4)
>         {
>             int32x4_t t[4];
>+X265_PRAGMA_UNROLL(4)
>             for (int j = 0; j < 4; ++j) {
>                 t[j] = vmull_s16(c0, vget_low_s16(O[i + j][0]));
>                 t[j] = vmlal_s16(t[j], c1, vget_high_s16(O[i + j][0]));
>@@ -461,6 +462,7 @@ static inline void partialButterfly32_neon(const int16_t *src, int16_t *dst)
>         for (int i = 0; i < line; i += 4)
>         {
>             int32x4_t t[4];
>+X265_PRAGMA_UNROLL(4)
>             for (int j = 0; j < 4; ++j) {
>                 t[j] = vmulq_s32(c0, EO[i + j][0]);
>                 t[j] = vmlaq_s32(t[j], c1, EO[i + j][1]);
>-- 
>2.39.5 (Apple Git-154)
>
>_______________________________________________
>x265-devel mailing list
>x265-devel@videolan.org
>https://mailman.videolan.org/listinfo/x265-devel
</pre></div>