[FFmpeg-cvslog] aarch64: vp9itxfm16: Restructure the idct32 store macros
Martin Storsjö
git at videolan.org
Sun Mar 19 23:02:01 EET 2017
ffmpeg | branch: master | Martin Storsjö <martin at martin.st> | Sun Feb 26 13:43:10 2017 +0200| [b76533f105cc01f6fb64199309fab84ba22da725] | committer: Martin Storsjö
aarch64: vp9itxfm16: Restructure the idct32 store macros
This avoids concatenation, which can't be used if the whole macro
is wrapped within another macro.
Signed-off-by: Martin Storsjö <martin at martin.st>
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=b76533f105cc01f6fb64199309fab84ba22da725
---
libavcodec/aarch64/vp9itxfm_16bpp_neon.S | 90 ++++++++++++++++----------------
1 file changed, 45 insertions(+), 45 deletions(-)
diff --git a/libavcodec/aarch64/vp9itxfm_16bpp_neon.S b/libavcodec/aarch64/vp9itxfm_16bpp_neon.S
index 86ea29e..a97c1b6 100644
--- a/libavcodec/aarch64/vp9itxfm_16bpp_neon.S
+++ b/libavcodec/aarch64/vp9itxfm_16bpp_neon.S
@@ -1244,27 +1244,27 @@ function idct32_1d_4x32_pass1_neon
.macro store_rev a, b, c, d
// There's no rev128 instruction, but we reverse each 64 bit
// half, and then flip them using an ext with 8 bytes offset.
- rev64 v7.4s, v\d\().4s
- st1 {v\a\().4s}, [x0], #16
+ rev64 v7.4s, \d
+ st1 {\a}, [x0], #16
ext v7.16b, v7.16b, v7.16b, #8
- st1 {v\b\().4s}, [x0], #16
- rev64 v6.4s, v\c\().4s
- st1 {v\c\().4s}, [x0], #16
+ st1 {\b}, [x0], #16
+ rev64 v6.4s, \c
+ st1 {\c}, [x0], #16
ext v6.16b, v6.16b, v6.16b, #8
- st1 {v\d\().4s}, [x0], #16
- rev64 v5.4s, v\b\().4s
+ st1 {\d}, [x0], #16
+ rev64 v5.4s, \b
st1 {v7.4s}, [x0], #16
ext v5.16b, v5.16b, v5.16b, #8
st1 {v6.4s}, [x0], #16
- rev64 v4.4s, v\a\().4s
+ rev64 v4.4s, \a
st1 {v5.4s}, [x0], #16
ext v4.16b, v4.16b, v4.16b, #8
st1 {v4.4s}, [x0], #16
.endm
- store_rev 16, 20, 24, 28
- store_rev 17, 21, 25, 29
- store_rev 18, 22, 26, 30
- store_rev 19, 23, 27, 31
+ store_rev v16.4s, v20.4s, v24.4s, v28.4s
+ store_rev v17.4s, v21.4s, v25.4s, v29.4s
+ store_rev v18.4s, v22.4s, v26.4s, v30.4s
+ store_rev v19.4s, v23.4s, v27.4s, v31.4s
sub x0, x0, #512
.purgem store_rev
@@ -1290,27 +1290,27 @@ function idct32_1d_4x32_pass1_neon
// Store the registers a, b, c, d horizontally,
// adding into the output first, and the mirrored,
// subtracted from the output.
-.macro store_rev a, b, c, d
+.macro store_rev a, b, c, d, a16b, b16b
ld1 {v4.4s}, [x0]
- rev64 v9.4s, v\d\().4s
- add v4.4s, v4.4s, v\a\().4s
+ rev64 v9.4s, \d
+ add v4.4s, v4.4s, \a
st1 {v4.4s}, [x0], #16
- rev64 v8.4s, v\c\().4s
+ rev64 v8.4s, \c
ld1 {v4.4s}, [x0]
ext v9.16b, v9.16b, v9.16b, #8
- add v4.4s, v4.4s, v\b\().4s
+ add v4.4s, v4.4s, \b
st1 {v4.4s}, [x0], #16
ext v8.16b, v8.16b, v8.16b, #8
ld1 {v4.4s}, [x0]
- rev64 v\b\().4s, v\b\().4s
- add v4.4s, v4.4s, v\c\().4s
+ rev64 \b, \b
+ add v4.4s, v4.4s, \c
st1 {v4.4s}, [x0], #16
- rev64 v\a\().4s, v\a\().4s
+ rev64 \a, \a
ld1 {v4.4s}, [x0]
- ext v\b\().16b, v\b\().16b, v\b\().16b, #8
- add v4.4s, v4.4s, v\d\().4s
+ ext \b16b, \b16b, \b16b, #8
+ add v4.4s, v4.4s, \d
st1 {v4.4s}, [x0], #16
- ext v\a\().16b, v\a\().16b, v\a\().16b, #8
+ ext \a16b, \a16b, \a16b, #8
ld1 {v4.4s}, [x0]
sub v4.4s, v4.4s, v9.4s
st1 {v4.4s}, [x0], #16
@@ -1318,17 +1318,17 @@ function idct32_1d_4x32_pass1_neon
sub v4.4s, v4.4s, v8.4s
st1 {v4.4s}, [x0], #16
ld1 {v4.4s}, [x0]
- sub v4.4s, v4.4s, v\b\().4s
+ sub v4.4s, v4.4s, \b
st1 {v4.4s}, [x0], #16
ld1 {v4.4s}, [x0]
- sub v4.4s, v4.4s, v\a\().4s
+ sub v4.4s, v4.4s, \a
st1 {v4.4s}, [x0], #16
.endm
- store_rev 31, 27, 23, 19
- store_rev 30, 26, 22, 18
- store_rev 29, 25, 21, 17
- store_rev 28, 24, 20, 16
+ store_rev v31.4s, v27.4s, v23.4s, v19.4s, v31.16b, v27.16b
+ store_rev v30.4s, v26.4s, v22.4s, v18.4s, v30.16b, v26.16b
+ store_rev v29.4s, v25.4s, v21.4s, v17.4s, v29.16b, v25.16b
+ store_rev v28.4s, v24.4s, v20.4s, v16.4s, v28.16b, v24.16b
.purgem store_rev
ret
endfunc
@@ -1370,21 +1370,21 @@ function idct32_1d_4x32_pass2_neon
.if \neg == 0
ld1 {v4.4s}, [x2], x9
ld1 {v5.4s}, [x2], x9
- add v4.4s, v4.4s, v\a\().4s
+ add v4.4s, v4.4s, \a
ld1 {v6.4s}, [x2], x9
- add v5.4s, v5.4s, v\b\().4s
+ add v5.4s, v5.4s, \b
ld1 {v7.4s}, [x2], x9
- add v6.4s, v6.4s, v\c\().4s
- add v7.4s, v7.4s, v\d\().4s
+ add v6.4s, v6.4s, \c
+ add v7.4s, v7.4s, \d
.else
ld1 {v4.4s}, [x2], x7
ld1 {v5.4s}, [x2], x7
- sub v4.4s, v4.4s, v\a\().4s
+ sub v4.4s, v4.4s, \a
ld1 {v6.4s}, [x2], x7
- sub v5.4s, v5.4s, v\b\().4s
+ sub v5.4s, v5.4s, \b
ld1 {v7.4s}, [x2], x7
- sub v6.4s, v6.4s, v\c\().4s
- sub v7.4s, v7.4s, v\d\().4s
+ sub v6.4s, v6.4s, \c
+ sub v7.4s, v7.4s, \d
.endif
ld1 {v8.4h}, [x0], x1
ld1 {v8.d}[1], [x0], x1
@@ -1410,15 +1410,15 @@ function idct32_1d_4x32_pass2_neon
st1 {v5.4h}, [x0], x1
st1 {v5.d}[1], [x0], x1
.endm
- load_acc_store 31, 30, 29, 28
- load_acc_store 27, 26, 25, 24
- load_acc_store 23, 22, 21, 20
- load_acc_store 19, 18, 17, 16
+ load_acc_store v31.4s, v30.4s, v29.4s, v28.4s
+ load_acc_store v27.4s, v26.4s, v25.4s, v24.4s
+ load_acc_store v23.4s, v22.4s, v21.4s, v20.4s
+ load_acc_store v19.4s, v18.4s, v17.4s, v16.4s
sub x2, x2, x9
- load_acc_store 16, 17, 18, 19, 1
- load_acc_store 20, 21, 22, 23, 1
- load_acc_store 24, 25, 26, 27, 1
- load_acc_store 28, 29, 30, 31, 1
+ load_acc_store v16.4s, v17.4s, v18.4s, v19.4s, 1
+ load_acc_store v20.4s, v21.4s, v22.4s, v23.4s, 1
+ load_acc_store v24.4s, v25.4s, v26.4s, v27.4s, 1
+ load_acc_store v28.4s, v29.4s, v30.4s, v31.4s, 1
.purgem load_acc_store
ret
endfunc
More information about the ffmpeg-cvslog
mailing list