diff options
| author | Frank Barchard <fbarchard@google.com> | 2023-01-17 14:00:35 -0800 |
|---|---|---|
| committer | libyuv LUCI CQ <libyuv-scoped@luci-project-accounts.iam.gserviceaccount.com> | 2023-01-18 00:30:05 +0000 |
| commit | 0faf8dd0e004520a61a603a4d2996d5ecc80dc3f (patch) | |
| tree | 36a46d7763ec9e14c50a5e4af3227e2131b8894e /source/row_neon.cc | |
| parent | 541d8efbaf0319b1e068d9c31f6dbc653d5c72b8 (diff) | |
| download | libyuv-0faf8dd0e004520a61a603a4d2996d5ecc80dc3f.tar.gz | |
Fix for DivideRow_NEON functions
- was dup of 8h but mul of 4s. now use umull
Bug: libyuv:951
Change-Id: If6cb01f5f006c2235886b81ce120642d7e24a9bb
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4166563
Reviewed-by: Justin Green <greenjustin@google.com>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Diffstat (limited to 'source/row_neon.cc')
| -rw-r--r-- | source/row_neon.cc | 28 |
1 files changed, 12 insertions, 16 deletions
diff --git a/source/row_neon.cc b/source/row_neon.cc index 416f112f..37f6db0c 100644 --- a/source/row_neon.cc +++ b/source/row_neon.cc @@ -3911,21 +3911,17 @@ void DivideRow_16_NEON(const uint16_t* src_y, int scale, int width) { asm volatile( - "vdup.16 q6, %3 \n" - "1: \n" - "vld1.16 {q0, q1}, [%0]! \n" - "vmovl.u16 q2, d0 \n" - "vmovl.u16 q3, d1 \n" - "vmovl.u16 q4, d2 \n" - "vmovl.u16 q5, d3 \n" - "vmul.u32 q2, q2, q6 \n" - "vmul.u32 q3, q3, q6 \n" - "vmul.u32 q4, q4, q6 \n" - "vmul.u32 q5, q5, q6 \n" - "vshrn.u32 d0, q2, #16 \n" - "vshrn.u32 d1, q3, #16 \n" - "vshrn.u32 d2, q4, #16 \n" - "vshrn.u32 d3, q5, #16 \n" + "vdup.16 d8, %3 \n" + "1: \n" + "vld1.16 {q2, q3}, [%0]! \n" + "vmull.u16 q0, d4, d8 \n" + "vmull.u16 q1, d5, d8 \n" + "vmull.u16 q2, d6, d8 \n" + "vmull.u16 q3, d7, d8 \n" + "vshrn.u32 d0, q0, #16 \n" + "vshrn.u32 d1, q1, #16 \n" + "vshrn.u32 d2, q2, #16 \n" + "vshrn.u32 d3, q3, #16 \n" "vst1.16 {q0, q1}, [%1]! \n" // store 16 pixels "subs %2, %2, #16 \n" // 16 src pixels per loop "bgt 1b \n" @@ -3933,7 +3929,7 @@ void DivideRow_16_NEON(const uint16_t* src_y, "+r"(dst_y), // %1 "+r"(width) // %2 : "r"(scale) // %3 - : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6"); + : "cc", "memory", "q0", "q1", "q2", "q3", "d8"); } // Use scale to convert lsb formats to msb, depending how many bits there are: |