diff options
| author | George Steed <george.steed@arm.com> | 2024-04-16 10:13:39 +0100 |
|---|---|---|
| committer | Frank Barchard <fbarchard@chromium.org> | 2024-04-18 19:02:43 +0000 |
| commit | 4838e7a194592c026e18cbe9f83a08e07a0ce95b (patch) | |
| tree | 621350012f3c7e3b2e6fb9d7aea578d90e757ac1 | |
| parent | 90070986aeac1129aa7632d986d636d3d29d5859 (diff) | |
| download | libyuv-4838e7a194592c026e18cbe9f83a08e07a0ce95b.tar.gz | |
[AArch64] Load full vectors in ARGB{Add,Subtract}Row
Using full vectors for Add and Subtract is a win across the board. Using
full vectors for the multiply is less obviously a win, especially for
smaller cores like Cortex-A53 or Cortex-A57, so is not considered for
this change.
Observed changes in performance with this change compared to the
existing Neon code:
| ARGBAddRow_NEON | ARGBSubtractRow_NEON
Cortex-A55 | -5.1% | -5.1%
Cortex-A510 | -18.4% | -18.4%
Cortex-A76 | -28.9% | -28.7%
Cortex-A720 | -36.1% | -36.2%
Cortex-X1 | -14.2% | -14.4%
Cortex-X2 | -12.5% | -12.5%
Bug: libyuv:976
Change-Id: I85316d4399c93b53baa62d0d43b2fa453517f5b4
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5457433
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
| -rw-r--r-- | source/row_neon64.cc | 24 |
1 files changed, 10 insertions, 14 deletions
diff --git a/source/row_neon64.cc b/source/row_neon64.cc index f9b34a49..dd0e7b77 100644 --- a/source/row_neon64.cc +++ b/source/row_neon64.cc @@ -3819,16 +3819,14 @@ void ARGBAddRow_NEON(const uint8_t* src_argb, asm volatile( // 8 pixel loop. "1: \n" - "ld1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB - "ld1 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 more + "ldp q0, q1, [%0], #32 \n" // load 8 ARGB + "ldp q4, q5, [%1], #32 \n" // load 8 more "subs %w3, %w3, #8 \n" // 8 processed per loop. - "uqadd v0.8b, v0.8b, v4.8b \n" "prfm pldl1keep, [%0, 448] \n" - "uqadd v1.8b, v1.8b, v5.8b \n" "prfm pldl1keep, [%1, 448] \n" - "uqadd v2.8b, v2.8b, v6.8b \n" - "uqadd v3.8b, v3.8b, v7.8b \n" - "st1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB + "uqadd v0.16b, v0.16b, v4.16b \n" + "uqadd v1.16b, v1.16b, v5.16b \n" + "stp q0, q1, [%2], #32 \n" // store 8 ARGB "b.gt 1b \n" : "+r"(src_argb), // %0 "+r"(src_argb1), // %1 @@ -3846,16 +3844,14 @@ void ARGBSubtractRow_NEON(const uint8_t* src_argb, asm volatile( // 8 pixel loop. "1: \n" - "ld1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB - "ld1 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 more + "ldp q0, q1, [%0], #32 \n" // load 8 ARGB + "ldp q4, q5, [%1], #32 \n" // load 8 more "subs %w3, %w3, #8 \n" // 8 processed per loop. - "uqsub v0.8b, v0.8b, v4.8b \n" "prfm pldl1keep, [%0, 448] \n" - "uqsub v1.8b, v1.8b, v5.8b \n" "prfm pldl1keep, [%1, 448] \n" - "uqsub v2.8b, v2.8b, v6.8b \n" - "uqsub v3.8b, v3.8b, v7.8b \n" - "st1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB + "uqsub v0.16b, v0.16b, v4.16b \n" + "uqsub v1.16b, v1.16b, v5.16b \n" + "stp q0, q1, [%2], #32 \n" // store 8 ARGB "b.gt 1b \n" : "+r"(src_argb), // %0 "+r"(src_argb1), // %1 |