diff options
| author | Martin Storsjö <martin@martin.st> | 2019-02-27 12:29:14 +0300 |
|---|---|---|
| committer | Martin Storsjö <martin@martin.st> | 2019-03-07 00:36:08 +0300 |
| commit | 18ef9556b71e3b6b839c35ae614ef0bb5b6a2179 (patch) | |
| tree | d17b4c23ba9488c5a9b5d9d37a5040c8513ccc49 /src/arm/64/looprestoration.S | |
| parent | c204da0ff33a0d563d6c632b42799e4fbc48f402 (diff) | |
arm: looprestoration: Simplify a few padding cases in wiener_filter_h_neon
Diffstat (limited to 'src/arm/64/looprestoration.S')
| -rw-r--r-- | src/arm/64/looprestoration.S | 26 |
1 files changed, 10 insertions, 16 deletions
diff --git a/src/arm/64/looprestoration.S b/src/arm/64/looprestoration.S index cf86134..7fc34d9 100644 --- a/src/arm/64/looprestoration.S +++ b/src/arm/64/looprestoration.S @@ -224,31 +224,25 @@ function wiener_filter_h_neon, export=1 mov v3.16b, v28.16b mov v5.16b, v29.16b br x11 +44: // 4 pixels valid in v2/v4, fill the high half with padding. + ins v2.d[1], v3.d[0] + ins v4.d[1], v5.d[0] + b 88f // Shift v2 right, shifting out invalid pixels, // shift v2 left to the original offset, shifting in padding pixels. -44: // 4 pixels valid - ext v2.16b, v2.16b, v2.16b, #8 - ext v2.16b, v2.16b, v3.16b, #8 - ext v4.16b, v4.16b, v4.16b, #8 - ext v4.16b, v4.16b, v5.16b, #8 - b 88f 55: // 5 pixels valid ext v2.16b, v2.16b, v2.16b, #10 ext v2.16b, v2.16b, v3.16b, #6 ext v4.16b, v4.16b, v4.16b, #10 ext v4.16b, v4.16b, v5.16b, #6 b 88f -66: // 6 pixels valid - ext v2.16b, v2.16b, v2.16b, #12 - ext v2.16b, v2.16b, v3.16b, #4 - ext v4.16b, v4.16b, v4.16b, #12 - ext v4.16b, v4.16b, v5.16b, #4 +66: // 6 pixels valid, fill the upper 2 pixels with padding. + ins v2.s[3], v3.s[0] + ins v4.s[3], v5.s[0] b 88f -77: // 7 pixels valid - ext v2.16b, v2.16b, v2.16b, #14 - ext v2.16b, v2.16b, v3.16b, #2 - ext v4.16b, v4.16b, v4.16b, #14 - ext v4.16b, v4.16b, v5.16b, #2 +77: // 7 pixels valid, fill the last pixel with padding. + ins v2.h[7], v3.h[0] + ins v4.h[7], v5.h[0] b 88f L(variable_shift_tbl): |