[FFmpeg-devel] [PATCH 21/21] aarch64: hevc: Produce plain neon versions of qpel_bi_hv
Martin Storsjö
martin at martin.st
Mon Mar 25 17:02:43 EET 2024
As the plain neon qpel_h functions process two rows at a time,
we need to allocate storage for h+8 rows instead of h+7.
By allocating storage for h+8 rows, incrementing the stack
pointer won't end up at the right spot in the end. Store the
intended final stack pointer value in a register x14 which we
store on the stack.
AWS Graviton 3:
put_hevc_qpel_bi_hv4_8_c: 385.7
put_hevc_qpel_bi_hv4_8_neon: 131.0
put_hevc_qpel_bi_hv4_8_i8mm: 92.2
put_hevc_qpel_bi_hv6_8_c: 701.0
put_hevc_qpel_bi_hv6_8_neon: 239.5
put_hevc_qpel_bi_hv6_8_i8mm: 191.0
put_hevc_qpel_bi_hv8_8_c: 1162.0
put_hevc_qpel_bi_hv8_8_neon: 228.0
put_hevc_qpel_bi_hv8_8_i8mm: 225.2
put_hevc_qpel_bi_hv12_8_c: 2305.0
put_hevc_qpel_bi_hv12_8_neon: 558.0
put_hevc_qpel_bi_hv12_8_i8mm: 483.2
put_hevc_qpel_bi_hv16_8_c: 3965.2
put_hevc_qpel_bi_hv16_8_neon: 732.7
put_hevc_qpel_bi_hv16_8_i8mm: 656.5
put_hevc_qpel_bi_hv24_8_c: 8709.7
put_hevc_qpel_bi_hv24_8_neon: 1555.2
put_hevc_qpel_bi_hv24_8_i8mm: 1448.7
put_hevc_qpel_bi_hv32_8_c: 14818.0
put_hevc_qpel_bi_hv32_8_neon: 2763.7
put_hevc_qpel_bi_hv32_8_i8mm: 2468.0
put_hevc_qpel_bi_hv48_8_c: 32855.5
put_hevc_qpel_bi_hv48_8_neon: 6107.2
put_hevc_qpel_bi_hv48_8_i8mm: 5452.7
put_hevc_qpel_bi_hv64_8_c: 57591.5
put_hevc_qpel_bi_hv64_8_neon: 10660.2
put_hevc_qpel_bi_hv64_8_i8mm: 9580.0
---
libavcodec/aarch64/hevcdsp_init_aarch64.c | 5 +
libavcodec/aarch64/hevcdsp_qpel_neon.S | 164 +++++++++++++---------
2 files changed, 103 insertions(+), 66 deletions(-)
diff --git a/libavcodec/aarch64/hevcdsp_init_aarch64.c b/libavcodec/aarch64/hevcdsp_init_aarch64.c
index e9ee901322..e24dd0cbda 100644
--- a/libavcodec/aarch64/hevcdsp_init_aarch64.c
+++ b/libavcodec/aarch64/hevcdsp_init_aarch64.c
@@ -319,6 +319,10 @@ NEON8_FNPROTO(qpel_bi_v, (uint8_t *dst, ptrdiff_t dststride,
const uint8_t *src, ptrdiff_t srcstride, const int16_t *src2,
int height, intptr_t mx, intptr_t my, int width),);
+NEON8_FNPROTO(qpel_bi_hv, (uint8_t *dst, ptrdiff_t dststride,
+ const uint8_t *src, ptrdiff_t srcstride, const int16_t *src2,
+ int height, intptr_t mx, intptr_t my, int width),);
+
NEON8_FNPROTO(qpel_bi_hv, (uint8_t *dst, ptrdiff_t dststride,
const uint8_t *src, ptrdiff_t srcstride, const int16_t *src2,
int height, intptr_t mx, intptr_t my, int width), _i8mm);
@@ -452,6 +456,7 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth)
NEON8_FNASSIGN(c->put_hevc_qpel, 1, 1, qpel_hv,);
NEON8_FNASSIGN(c->put_hevc_qpel_uni, 1, 1, qpel_uni_hv,);
NEON8_FNASSIGN_PARTIAL_5(c->put_hevc_qpel_uni_w, 1, 1, qpel_uni_w_hv,);
+ NEON8_FNASSIGN(c->put_hevc_qpel_bi, 1, 1, qpel_bi_hv,);
if (have_i8mm(cpu_flags)) {
NEON8_FNASSIGN(c->put_hevc_epel, 0, 1, epel_h, _i8mm);
diff --git a/libavcodec/aarch64/hevcdsp_qpel_neon.S b/libavcodec/aarch64/hevcdsp_qpel_neon.S
index df7032b692..8ddaa32b70 100644
--- a/libavcodec/aarch64/hevcdsp_qpel_neon.S
+++ b/libavcodec/aarch64/hevcdsp_qpel_neon.S
@@ -4590,14 +4590,6 @@ endfunc
qpel_uni_w_hv neon
-#if HAVE_I8MM
-ENABLE_I8MM
-
-qpel_uni_w_hv neon_i8mm
-
-DISABLE_I8MM
-#endif
-
function hevc_put_hevc_qpel_bi_hv4_8_end_neon
mov x9, #(MAX_PB_SIZE * 2)
load_qpel_filterh x7, x6
@@ -4620,7 +4612,8 @@ function hevc_put_hevc_qpel_bi_hv4_8_end_neon
.endm
1: calc_all
.purgem calc
-2: ret
+2: mov sp, x14
+ ret
endfunc
function hevc_put_hevc_qpel_bi_hv6_8_end_neon
@@ -4650,7 +4643,8 @@ function hevc_put_hevc_qpel_bi_hv6_8_end_neon
.endm
1: calc_all
.purgem calc
-2: ret
+2: mov sp, x14
+ ret
endfunc
function hevc_put_hevc_qpel_bi_hv8_8_end_neon
@@ -4678,7 +4672,8 @@ function hevc_put_hevc_qpel_bi_hv8_8_end_neon
.endm
1: calc_all
.purgem calc
-2: ret
+2: mov sp, x14
+ ret
endfunc
function hevc_put_hevc_qpel_bi_hv16_8_end_neon
@@ -4723,83 +4718,87 @@ function hevc_put_hevc_qpel_bi_hv16_8_end_neon
subs x10, x10, #16
add x4, x4, #32
b.ne 0b
- add w10, w5, #7
- lsl x10, x10, #7
- sub x10, x10, x6, lsl #1 // part of first line
- add sp, sp, x10 // tmp_array without first line
+ mov sp, x14
ret
endfunc
-#if HAVE_I8MM
-ENABLE_I8MM
-
-function ff_hevc_put_hevc_qpel_bi_hv4_8_neon_i8mm, export=1
- add w10, w5, #7
+.macro qpel_bi_hv suffix
+function ff_hevc_put_hevc_qpel_bi_hv4_8_\suffix, export=1
+ add w10, w5, #8
lsl x10, x10, #7
+ mov x14, sp
sub sp, sp, x10 // tmp_array
- stp x7, x30, [sp, #-48]!
+ stp x7, x30, [sp, #-64]!
stp x4, x5, [sp, #16]
stp x0, x1, [sp, #32]
+ str x14, [sp, #48]
sub x1, x2, x3, lsl #1
sub x1, x1, x3
- add x0, sp, #48
+ add x0, sp, #64
mov x2, x3
add w3, w5, #7
mov x4, x6
- bl X(ff_hevc_put_hevc_qpel_h4_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_h4_8_\suffix)
ldp x4, x5, [sp, #16]
ldp x0, x1, [sp, #32]
- ldp x7, x30, [sp], #48
+ ldr x14, [sp, #48]
+ ldp x7, x30, [sp], #64
b hevc_put_hevc_qpel_bi_hv4_8_end_neon
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv6_8_neon_i8mm, export=1
- add w10, w5, #7
+function ff_hevc_put_hevc_qpel_bi_hv6_8_\suffix, export=1
+ add w10, w5, #8
lsl x10, x10, #7
+ mov x14, sp
sub sp, sp, x10 // tmp_array
- stp x7, x30, [sp, #-48]!
+ stp x7, x30, [sp, #-64]!
stp x4, x5, [sp, #16]
stp x0, x1, [sp, #32]
+ str x14, [sp, #48]
sub x1, x2, x3, lsl #1
sub x1, x1, x3
- add x0, sp, #48
+ add x0, sp, #64
mov x2, x3
add x3, x5, #7
mov x4, x6
- bl X(ff_hevc_put_hevc_qpel_h6_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_h6_8_\suffix)
ldp x4, x5, [sp, #16]
ldp x0, x1, [sp, #32]
- ldp x7, x30, [sp], #48
+ ldr x14, [sp, #48]
+ ldp x7, x30, [sp], #64
b hevc_put_hevc_qpel_bi_hv6_8_end_neon
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv8_8_neon_i8mm, export=1
- add w10, w5, #7
+function ff_hevc_put_hevc_qpel_bi_hv8_8_\suffix, export=1
+ add w10, w5, #8
lsl x10, x10, #7
+ mov x14, sp
sub sp, sp, x10 // tmp_array
- stp x7, x30, [sp, #-48]!
+ stp x7, x30, [sp, #-64]!
stp x4, x5, [sp, #16]
stp x0, x1, [sp, #32]
+ str x14, [sp, #48]
sub x1, x2, x3, lsl #1
sub x1, x1, x3
- add x0, sp, #48
+ add x0, sp, #64
mov x2, x3
add x3, x5, #7
mov x4, x6
- bl X(ff_hevc_put_hevc_qpel_h8_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_h8_8_\suffix)
ldp x4, x5, [sp, #16]
ldp x0, x1, [sp, #32]
- ldp x7, x30, [sp], #48
+ ldr x14, [sp, #48]
+ ldp x7, x30, [sp], #64
b hevc_put_hevc_qpel_bi_hv8_8_end_neon
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv12_8_neon_i8mm, export=1
+function ff_hevc_put_hevc_qpel_bi_hv12_8_\suffix, export=1
stp x6, x7, [sp, #-80]!
stp x4, x5, [sp, #16]
stp x2, x3, [sp, #32]
stp x0, x1, [sp, #48]
str x30, [sp, #64]
- bl X(ff_hevc_put_hevc_qpel_bi_hv8_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_bi_hv8_8_\suffix)
ldp x4, x5, [sp, #16]
ldp x2, x3, [sp, #32]
ldp x0, x1, [sp, #48]
@@ -4807,39 +4806,42 @@ function ff_hevc_put_hevc_qpel_bi_hv12_8_neon_i8mm, export=1
add x4, x4, #16
add x2, x2, #8
add x0, x0, #8
- bl X(ff_hevc_put_hevc_qpel_bi_hv4_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_bi_hv4_8_\suffix)
ldr x30, [sp], #16
ret
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv16_8_neon_i8mm, export=1
- add w10, w5, #7
+function ff_hevc_put_hevc_qpel_bi_hv16_8_\suffix, export=1
+ add w10, w5, #8
lsl x10, x10, #7
+ mov x14, sp
sub sp, sp, x10 // tmp_array
- stp x7, x30, [sp, #-48]!
+ stp x7, x30, [sp, #-64]!
stp x4, x5, [sp, #16]
stp x0, x1, [sp, #32]
- add x0, sp, #48
+ str x14, [sp, #48]
+ add x0, sp, #64
sub x1, x2, x3, lsl #1
sub x1, x1, x3
mov x2, x3
add w3, w5, #7
mov x4, x6
- bl X(ff_hevc_put_hevc_qpel_h16_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_h16_8_\suffix)
ldp x4, x5, [sp, #16]
ldp x0, x1, [sp, #32]
- ldp x7, x30, [sp], #48
+ ldr x14, [sp, #48]
+ ldp x7, x30, [sp], #64
mov x6, #16 // width
b hevc_put_hevc_qpel_bi_hv16_8_end_neon
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv24_8_neon_i8mm, export=1
+function ff_hevc_put_hevc_qpel_bi_hv24_8_\suffix, export=1
stp x6, x7, [sp, #-80]!
stp x4, x5, [sp, #16]
stp x2, x3, [sp, #32]
stp x0, x1, [sp, #48]
str x30, [sp, #64]
- bl X(ff_hevc_put_hevc_qpel_bi_hv16_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_bi_hv16_8_\suffix)
ldp x4, x5, [sp, #16]
ldp x2, x3, [sp, #32]
ldp x0, x1, [sp, #48]
@@ -4847,73 +4849,103 @@ function ff_hevc_put_hevc_qpel_bi_hv24_8_neon_i8mm, export=1
add x4, x4, #32
add x2, x2, #16
add x0, x0, #16
- bl X(ff_hevc_put_hevc_qpel_bi_hv8_8_neon_i8mm)
+ bl X(ff_hevc_put_hevc_qpel_bi_hv8_8_\suffix)
ldr x30, [sp], #16
ret
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv32_8_neon_i8mm, export=1
- add w10, w5, #7
+function ff_hevc_put_hevc_qpel_bi_hv32_8_\suffix, export=1
+ add w10, w5, #8
lsl x10, x10, #7
+ mov x14, sp
sub sp, sp, x10 // tmp_array
- stp x7, x30, [sp, #-48]!
+ stp x7, x30, [sp, #-64]!
stp x4, x5, [sp, #16]
stp x0, x1, [sp, #32]
- add x0, sp, #48
+ str x14, [sp, #48]
+ add x0, sp, #64
sub x1, x2, x3, lsl #1
mov x2, x3
sub x1, x1, x3
add w3, w5, #7
mov x4, x6
- bl X(ff_hevc_put_hevc_qpel_h32_8_neon_i8mm)
+ mov w6, #32
+ bl X(ff_hevc_put_hevc_qpel_h32_8_\suffix)
ldp x4, x5, [sp, #16]
ldp x0, x1, [sp, #32]
- ldp x7, x30, [sp], #48
+ ldr x14, [sp, #48]
+ ldp x7, x30, [sp], #64
mov x6, #32 // width
b hevc_put_hevc_qpel_bi_hv16_8_end_neon
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv48_8_neon_i8mm, export=1
- add w10, w5, #7
+function ff_hevc_put_hevc_qpel_bi_hv48_8_\suffix, export=1
+ add w10, w5, #8
lsl x10, x10, #7
+ mov x14, sp
sub sp, sp, x10 // tmp_array
- stp x7, x30, [sp, #-48]!
+ stp x7, x30, [sp, #-64]!
stp x4, x5, [sp, #16]
stp x0, x1, [sp, #32]
- add x0, sp, #48
+ str x14, [sp, #48]
+ add x0, sp, #64
sub x1, x2, x3, lsl #1
mov x2, x3
sub x1, x1, x3
add w3, w5, #7
mov x4, x6
- bl X(ff_hevc_put_hevc_qpel_h48_8_neon_i8mm)
+.ifc \suffix, neon
+ mov w6, #48
+ bl X(ff_hevc_put_hevc_qpel_h32_8_\suffix)
+.else
+ bl X(ff_hevc_put_hevc_qpel_h48_8_\suffix)
+.endif
ldp x4, x5, [sp, #16]
ldp x0, x1, [sp, #32]
- ldp x7, x30, [sp], #48
+ ldr x14, [sp, #48]
+ ldp x7, x30, [sp], #64
mov x6, #48 // width
b hevc_put_hevc_qpel_bi_hv16_8_end_neon
endfunc
-function ff_hevc_put_hevc_qpel_bi_hv64_8_neon_i8mm, export=1
- add w10, w5, #7
+function ff_hevc_put_hevc_qpel_bi_hv64_8_\suffix, export=1
+ add w10, w5, #8
lsl x10, x10, #7
+ mov x14, sp
sub sp, sp, x10 // tmp_array
- stp x7, x30, [sp, #-48]!
+ stp x7, x30, [sp, #-64]!
stp x4, x5, [sp, #16]
stp x0, x1, [sp, #32]
- add x0, sp, #48
+ str x14, [sp, #48]
+ add x0, sp, #64
sub x1, x2, x3, lsl #1
mov x2, x3
sub x1, x1, x3
add w3, w5, #7
mov x4, x6
- bl X(ff_hevc_put_hevc_qpel_h64_8_neon_i8mm)
+.ifc \suffix, neon
+ mov w6, #64
+ bl X(ff_hevc_put_hevc_qpel_h32_8_\suffix)
+.else
+ bl X(ff_hevc_put_hevc_qpel_h64_8_\suffix)
+.endif
ldp x4, x5, [sp, #16]
ldp x0, x1, [sp, #32]
- ldp x7, x30, [sp], #48
+ ldr x14, [sp, #48]
+ ldp x7, x30, [sp], #64
mov x6, #64 // width
b hevc_put_hevc_qpel_bi_hv16_8_end_neon
endfunc
+.endm
+
+qpel_bi_hv neon
+
+#if HAVE_I8MM
+ENABLE_I8MM
+
+qpel_uni_w_hv neon_i8mm
+
+qpel_bi_hv neon_i8mm
DISABLE_I8MM
#endif // HAVE_I8MM
--
2.39.3 (Apple Git-146)
More information about the ffmpeg-devel
mailing list