[FFmpeg-devel] [PATCH 3/3] aarch64/vvc: Add put_qpel_hv
Zhao Zhili
quinkblack at foxmail.com
Tue Sep 10 20:35:06 EEST 2024
From: Zhao Zhili <zhilizhao at tencent.com>
With Apple M1 (no i8mm):
put_luma_hv_8_4x4_c: 2.2 ( 1.00x)
put_luma_hv_8_4x4_neon: 0.8 ( 3.00x)
put_luma_hv_8_8x8_c: 7.0 ( 1.00x)
put_luma_hv_8_8x8_neon: 0.8 ( 9.33x)
put_luma_hv_8_16x16_c: 22.8 ( 1.00x)
put_luma_hv_8_16x16_neon: 2.5 ( 9.10x)
put_luma_hv_8_32x32_c: 84.8 ( 1.00x)
put_luma_hv_8_32x32_neon: 9.5 ( 8.92x)
put_luma_hv_8_64x64_c: 333.0 ( 1.00x)
put_luma_hv_8_64x64_neon: 35.5 ( 9.38x)
put_luma_hv_8_128x128_c: 1294.5 ( 1.00x)
put_luma_hv_8_128x128_neon: 137.8 ( 9.40x)
With Pixel 8 Pro:
put_luma_hv_8_4x4_c: 5.0 ( 1.00x)
put_luma_hv_8_4x4_neon: 0.8 ( 6.67x)
put_luma_hv_8_4x4_i8mm: 0.2 (20.00x)
put_luma_hv_8_8x8_c: 13.2 ( 1.00x)
put_luma_hv_8_8x8_neon: 1.2 (10.60x)
put_luma_hv_8_8x8_i8mm: 1.2 (10.60x)
put_luma_hv_8_16x16_c: 44.2 ( 1.00x)
put_luma_hv_8_16x16_neon: 4.5 ( 9.83x)
put_luma_hv_8_16x16_i8mm: 4.2 (10.41x)
put_luma_hv_8_32x32_c: 160.8 ( 1.00x)
put_luma_hv_8_32x32_neon: 17.5 ( 9.19x)
put_luma_hv_8_32x32_i8mm: 16.0 (10.05x)
put_luma_hv_8_64x64_c: 611.2 ( 1.00x)
put_luma_hv_8_64x64_neon: 68.0 ( 8.99x)
put_luma_hv_8_64x64_i8mm: 62.2 ( 9.82x)
put_luma_hv_8_128x128_c: 2384.8 ( 1.00x)
put_luma_hv_8_128x128_neon: 268.8 ( 8.87x)
put_luma_hv_8_128x128_i8mm: 245.8 ( 9.70x)
---
libavcodec/aarch64/h26x/dsp.h | 8 ++
libavcodec/aarch64/h26x/qpel_neon.S | 140 ++++++++++++++++++++++++++++
libavcodec/aarch64/vvc/dsp_init.c | 14 +++
3 files changed, 162 insertions(+)
diff --git a/libavcodec/aarch64/h26x/dsp.h b/libavcodec/aarch64/h26x/dsp.h
index 881091f39a..c54906dde2 100644
--- a/libavcodec/aarch64/h26x/dsp.h
+++ b/libavcodec/aarch64/h26x/dsp.h
@@ -282,4 +282,12 @@ void ff_vvc_put_qpel_v8_8_neon(int16_t *dst, const uint8_t *_src,
ptrdiff_t _srcstride, int height,
const int8_t *hf, const int8_t *vf, int width);
+NEON8_FNPROTO_PARTIAL_6(qpel_hv, (int16_t *dst,
+ const uint8_t *src, ptrdiff_t srcstride, int height,
+ const int8_t *hf, const int8_t *vf, int width),);
+
+NEON8_FNPROTO_PARTIAL_6(qpel_hv, (int16_t *dst,
+ const uint8_t *src, ptrdiff_t srcstride, int height,
+ const int8_t *hf, const int8_t *vf, int width), _i8mm);
+
#endif
diff --git a/libavcodec/aarch64/h26x/qpel_neon.S b/libavcodec/aarch64/h26x/qpel_neon.S
index 671942109a..1b3da375ba 100644
--- a/libavcodec/aarch64/h26x/qpel_neon.S
+++ b/libavcodec/aarch64/h26x/qpel_neon.S
@@ -4142,9 +4142,15 @@ endfunc
DISABLE_I8MM
#endif
+function vvc_put_qpel_hv4_8_end_neon
+ vvc_load_qpel_filterh x5
+ mov x7, #(VVC_MAX_PB_SIZE * 2)
+ b 1f
+endfunc
function hevc_put_hevc_qpel_hv4_8_end_neon
load_qpel_filterh x5, x4
+1:
ldr d16, [sp]
ldr d17, [sp, x7]
add sp, sp, x7, lsl #1
@@ -4196,9 +4202,16 @@ function hevc_put_hevc_qpel_hv6_8_end_neon
ret
endfunc
+function vvc_put_qpel_hv8_8_end_neon
+ vvc_load_qpel_filterh x5
+ mov x7, #(VVC_MAX_PB_SIZE * 2)
+ b 1f
+endfunc
+
function hevc_put_hevc_qpel_hv8_8_end_neon
mov x7, #128
load_qpel_filterh x5, x4
+1:
ldr q16, [sp]
ldr q17, [sp, x7]
add sp, sp, x7, lsl #1
@@ -4249,9 +4262,16 @@ function hevc_put_hevc_qpel_hv12_8_end_neon
ret
endfunc
+function vvc_put_qpel_hv16_8_end_neon
+ vvc_load_qpel_filterh x5
+ mov x7, #(VVC_MAX_PB_SIZE * 2)
+ b 1f
+endfunc
+
function hevc_put_hevc_qpel_hv16_8_end_neon
mov x7, #128
load_qpel_filterh x5, x4
+1:
ld1 {v16.8h, v17.8h}, [sp], x7
ld1 {v18.8h, v19.8h}, [sp], x7
ld1 {v20.8h, v21.8h}, [sp], x7
@@ -4274,6 +4294,12 @@ function hevc_put_hevc_qpel_hv16_8_end_neon
ret
endfunc
+function vvc_put_qpel_hv32_8_end_neon
+ vvc_load_qpel_filterh x5
+ mov x7, #(VVC_MAX_PB_SIZE * 2)
+ b 0f
+endfunc
+
function hevc_put_hevc_qpel_hv32_8_end_neon
mov x7, #128
load_qpel_filterh x5, x4
@@ -4327,6 +4353,25 @@ function ff_hevc_put_hevc_qpel_hv4_8_\suffix, export=1
b hevc_put_hevc_qpel_hv4_8_end_neon
endfunc
+function ff_vvc_put_qpel_hv4_8_\suffix, export=1
+ add w10, w3, #8
+ lsl x10, x10, #8
+ mov x14, sp
+ sub sp, sp, x10 // tmp_array
+ stp x5, x30, [sp, #-48]!
+ stp x0, x3, [sp, #16]
+ str x14, [sp, #32]
+ add x0, sp, #48
+ sub x1, x1, x2, lsl #1
+ add x3, x3, #7
+ sub x1, x1, x2
+ bl X(ff_vvc_put_qpel_h4_8_\suffix)
+ ldr x14, [sp, #32]
+ ldp x0, x3, [sp, #16]
+ ldp x5, x30, [sp], #48
+ b vvc_put_qpel_hv4_8_end_neon
+endfunc
+
function ff_hevc_put_hevc_qpel_hv6_8_\suffix, export=1
add w10, w3, #8
mov x7, #128
@@ -4366,6 +4411,25 @@ function ff_hevc_put_hevc_qpel_hv8_8_\suffix, export=1
b hevc_put_hevc_qpel_hv8_8_end_neon
endfunc
+function ff_vvc_put_qpel_hv8_8_\suffix, export=1
+ add w10, w3, #8
+ lsl x10, x10, #8
+ sub x1, x1, x2, lsl #1
+ mov x14, sp
+ sub sp, sp, x10 // tmp_array
+ stp x5, x30, [sp, #-48]!
+ stp x0, x3, [sp, #16]
+ str x14, [sp, #32]
+ add x0, sp, #48
+ add x3, x3, #7
+ sub x1, x1, x2
+ bl X(ff_vvc_put_qpel_h8_8_\suffix)
+ ldr x14, [sp, #32]
+ ldp x0, x3, [sp, #16]
+ ldp x5, x30, [sp], #48
+ b vvc_put_qpel_hv8_8_end_neon
+endfunc
+
function ff_hevc_put_hevc_qpel_hv12_8_\suffix, export=1
add w10, w3, #8
lsl x10, x10, #7
@@ -4405,6 +4469,25 @@ function ff_hevc_put_hevc_qpel_hv16_8_\suffix, export=1
b hevc_put_hevc_qpel_hv16_8_end_neon
endfunc
+function ff_vvc_put_qpel_hv16_8_\suffix, export=1
+ add w10, w3, #8
+ lsl x10, x10, #8
+ sub x1, x1, x2, lsl #1
+ mov x14, sp
+ sub sp, sp, x10 // tmp_array
+ stp x5, x30, [sp, #-48]!
+ stp x0, x3, [sp, #16]
+ str x14, [sp, #32]
+ add x3, x3, #7
+ add x0, sp, #48
+ sub x1, x1, x2
+ bl X(ff_vvc_put_qpel_h16_8_\suffix)
+ ldr x14, [sp, #32]
+ ldp x0, x3, [sp, #16]
+ ldp x5, x30, [sp], #48
+ b vvc_put_qpel_hv16_8_end_neon
+endfunc
+
function ff_hevc_put_hevc_qpel_hv24_8_\suffix, export=1
stp x4, x5, [sp, #-64]!
stp x2, x3, [sp, #16]
@@ -4441,6 +4524,26 @@ function ff_hevc_put_hevc_qpel_hv32_8_\suffix, export=1
b hevc_put_hevc_qpel_hv32_8_end_neon
endfunc
+function ff_vvc_put_qpel_hv32_8_\suffix, export=1
+ add w10, w3, #8
+ sub x1, x1, x2, lsl #1
+ lsl x10, x10, #8
+ sub x1, x1, x2
+ mov x14, sp
+ sub sp, sp, x10 // tmp_array
+ stp x5, x30, [sp, #-48]!
+ stp x0, x3, [sp, #16]
+ str x14, [sp, #32]
+ add x3, x3, #7
+ add x0, sp, #48
+ mov w6, #32
+ bl X(ff_vvc_put_qpel_h32_8_\suffix)
+ ldr x14, [sp, #32]
+ ldp x0, x3, [sp, #16]
+ ldp x5, x30, [sp], #48
+ b vvc_put_qpel_hv32_8_end_neon
+endfunc
+
function ff_hevc_put_hevc_qpel_hv48_8_\suffix, export=1
stp x4, x5, [sp, #-64]!
stp x2, x3, [sp, #16]
@@ -4474,6 +4577,43 @@ function ff_hevc_put_hevc_qpel_hv64_8_\suffix, export=1
ldr x30, [sp], #16
ret
endfunc
+
+function ff_vvc_put_qpel_hv64_8_\suffix, export=1
+ stp x4, x5, [sp, #-64]!
+ stp x2, x3, [sp, #16]
+ stp x0, x1, [sp, #32]
+ str x30, [sp, #48]
+ mov x6, #32
+ bl X(ff_vvc_put_qpel_hv32_8_\suffix)
+ ldp x0, x1, [sp, #32]
+ ldp x2, x3, [sp, #16]
+ ldp x4, x5, [sp], #48
+ add x1, x1, #32
+ add x0, x0, #64
+ mov x6, #32
+ bl X(ff_vvc_put_qpel_hv32_8_\suffix)
+ ldr x30, [sp], #16
+ ret
+endfunc
+
+function ff_vvc_put_qpel_hv128_8_\suffix, export=1
+ stp x4, x5, [sp, #-64]!
+ stp x2, x3, [sp, #16]
+ stp x0, x1, [sp, #32]
+ str x30, [sp, #48]
+ mov x6, #64
+ bl X(ff_vvc_put_qpel_hv64_8_\suffix)
+ ldp x0, x1, [sp, #32]
+ ldp x2, x3, [sp, #16]
+ ldp x4, x5, [sp], #48
+ add x1, x1, #64
+ add x0, x0, #128
+ mov x6, #64
+ bl X(ff_vvc_put_qpel_hv64_8_\suffix)
+ ldr x30, [sp], #16
+ ret
+endfunc
+
.endm
qpel_hv neon
diff --git a/libavcodec/aarch64/vvc/dsp_init.c b/libavcodec/aarch64/vvc/dsp_init.c
index ba3a49aa1a..934d918ffd 100644
--- a/libavcodec/aarch64/vvc/dsp_init.c
+++ b/libavcodec/aarch64/vvc/dsp_init.c
@@ -67,6 +67,13 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd)
c->inter.put[0][5][1][0] =
c->inter.put[0][6][1][0] = ff_vvc_put_qpel_v8_8_neon;
+ c->inter.put[0][1][1][1] = ff_vvc_put_qpel_hv4_8_neon;
+ c->inter.put[0][2][1][1] = ff_vvc_put_qpel_hv8_8_neon;
+ c->inter.put[0][3][1][1] = ff_vvc_put_qpel_hv16_8_neon;
+ c->inter.put[0][4][1][1] = ff_vvc_put_qpel_hv32_8_neon;
+ c->inter.put[0][5][1][1] = ff_vvc_put_qpel_hv64_8_neon;
+ c->inter.put[0][6][1][1] = ff_vvc_put_qpel_hv128_8_neon;
+
c->inter.put_uni[0][1][0][0] = ff_vvc_put_pel_uni_pixels4_8_neon;
c->inter.put_uni[0][2][0][0] = ff_vvc_put_pel_uni_pixels8_8_neon;
c->inter.put_uni[0][3][0][0] = ff_vvc_put_pel_uni_pixels16_8_neon;
@@ -103,6 +110,13 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd)
c->inter.put[0][4][0][1] = ff_vvc_put_qpel_h32_8_neon_i8mm;
c->inter.put[0][5][0][1] = ff_vvc_put_qpel_h64_8_neon_i8mm;
c->inter.put[0][6][0][1] = ff_vvc_put_qpel_h128_8_neon_i8mm;
+
+ c->inter.put[0][1][1][1] = ff_vvc_put_qpel_hv4_8_neon_i8mm;
+ c->inter.put[0][2][1][1] = ff_vvc_put_qpel_hv8_8_neon_i8mm;
+ c->inter.put[0][3][1][1] = ff_vvc_put_qpel_hv16_8_neon_i8mm;
+ c->inter.put[0][4][1][1] = ff_vvc_put_qpel_hv32_8_neon_i8mm;
+ c->inter.put[0][5][1][1] = ff_vvc_put_qpel_hv64_8_neon_i8mm;
+ c->inter.put[0][6][1][1] = ff_vvc_put_qpel_hv128_8_neon_i8mm;
}
} else if (bd == 10) {
c->alf.filter[LUMA] = alf_filter_luma_10_neon;
--
2.42.0
More information about the ffmpeg-devel
mailing list