From: Zhao Zhili <zhiliz...@tencent.com>
On rpi5 (A76):
put_hevc_pel_bi_w_pixels4_8_c: 90.0 ( 1.00x)
put_hevc_pel_bi_w_pixels4_8_neon: 34.1 ( 2.64x)
put_hevc_pel_bi_w_pixels6_8_c: 188.3 ( 1.00x)
put_hevc_pel_bi_w_pixels6_8_neon: 73.5 ( 2.56x)
put_hevc_pel_bi_w_pixels8_8_c: 327.1 ( 1.00x)
put_hevc_pel_bi_w_pixels8_8_neon: 75.8 ( 4.32x)
put_hevc_pel_bi_w_pixels12_8_c: 728.8 ( 1.00x)
put_hevc_pel_bi_w_pixels12_8_neon: 186.1 ( 3.92x)
put_hevc_pel_bi_w_pixels16_8_c: 1288.1 ( 1.00x)
put_hevc_pel_bi_w_pixels16_8_neon: 268.5 ( 4.80x)
put_hevc_pel_bi_w_pixels24_8_c: 2855.5 ( 1.00x)
put_hevc_pel_bi_w_pixels24_8_neon: 723.8 ( 3.95x)
put_hevc_pel_bi_w_pixels32_8_c: 5095.3 ( 1.00x)
put_hevc_pel_bi_w_pixels32_8_neon: 1165.0 ( 4.37x)
put_hevc_pel_bi_w_pixels48_8_c: 11521.5 ( 1.00x)
put_hevc_pel_bi_w_pixels48_8_neon: 2856.0 ( 4.03x)
put_hevc_pel_bi_w_pixels64_8_c: 21020.5 ( 1.00x)
put_hevc_pel_bi_w_pixels64_8_neon: 4699.1 ( 4.47x)
---
libavcodec/aarch64/h26x/dsp.h | 5 +
libavcodec/aarch64/h26x/epel_neon.S | 373 ++++++++++++++++++++++
libavcodec/aarch64/hevcdsp_init_aarch64.c | 13 +
3 files changed, 391 insertions(+)
diff --git a/libavcodec/aarch64/h26x/dsp.h b/libavcodec/aarch64/h26x/dsp.h
index 6ea6a8d36a..6c91004301 100644
--- a/libavcodec/aarch64/h26x/dsp.h
+++ b/libavcodec/aarch64/h26x/dsp.h
@@ -92,6 +92,11 @@ NEON8_FNPROTO(pel_bi_pixels, (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(pel_bi_w_pixels, (uint8_t *_dst, ptrdiff_t _dststride,
+ const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2,
+ int height, int denom, int wx0, int wx1,
+ int ox0, int ox1, intptr_t mx, intptr_t my, int width),);
+
NEON8_FNPROTO(epel_bi_h, (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),);
diff --git a/libavcodec/aarch64/h26x/epel_neon.S
b/libavcodec/aarch64/h26x/epel_neon.S
index e44a448b1f..4cf6339171 100644
--- a/libavcodec/aarch64/h26x/epel_neon.S
+++ b/libavcodec/aarch64/h26x/epel_neon.S
@@ -473,6 +473,379 @@ function ff_hevc_put_hevc_pel_bi_pixels64_8_neon, export=1
ret
endfunc
+.macro load_bi_w_pixels_param
+ ldrsw x8, [sp] // wx1
+#if defined(__APPLE__)
+ ldpsw x9, x10, [sp, #4] // ox0, ox1
+ ldrsw x11, [sp, #32] // width
+#else
+ ldrsw x9, [sp, #8] // ox0
+ ldrsw x10, [sp, #16] // ox1
+ ldrsw x11, [sp, 40] // width
+#endif
+.endm
+
+function ff_hevc_put_hevc_pel_bi_w_pixels4_8_neon, export=1
+ load_bi_w_pixels_param
+ add w6, w6, #6 // log2Wd
+ dup v0.8h, w7 // wx0
+ dup v1.8h, w8 // wx1
+ add w9, w9, w10
+ add w9, w9, #1 // ox0 + ox1 + 1
+ lsl w9, w9, w6
+ add w7, w6, #1 // (log2Wd + 1)
+ mov x8, #(2 * HEVC_MAX_PB_SIZE)
+ neg w7, w7
+ dup v2.4s, w9 // (ox0 + ox1 + 1) << logwWd
+ dup v6.4s, w7 // -(log2Wd + 1)
+1:
+ ld1 {v4.8b}, [x2], x3 // load src
+ ld1 {v5.8b}, [x4], x8 // load src2
+ subs w5, w5, #1
+ mov v3.16b, v2.16b
+ ushll v4.8h, v4.8b, #6
+ smlal v3.4s, v4.4h, v1.4h
+ smlal v3.4s, v5.4h, v0.4h
+ sshl v3.4s, v3.4s, v6.4s
+ sqxtn v3.4h, v3.4s
+ sqxtun v3.8b, v3.8h
+ st1 {v3.s}[0], [x0], x1
+ b.ne 1b
+ ret
+endfunc
+
+function ff_hevc_put_hevc_pel_bi_w_pixels6_8_neon, export=1
+ load_bi_w_pixels_param
+ add w6, w6, #6 // log2Wd
+ dup v0.8h, w7 // wx0
+ dup v1.8h, w8 // wx1
+ add w9, w9, w10
+ add w9, w9, #1 // ox0 + ox1 + 1
+ lsl w9, w9, w6
+ add w7, w6, #1 // (log2Wd + 1)
+ mov x8, #(2 * HEVC_MAX_PB_SIZE)
+ neg w7, w7
+ dup v2.4s, w9 // (ox0 + ox1 + 1) << logwWd
+ dup v6.4s, w7 // -(log2Wd + 1)
+ sub x1, x1, #4
+1:
+ ld1 {v4.8b}, [x2], x3 // load src
+ ld1 {v5.8h}, [x4], x8 // load src2
+ subs w5, w5, #1
+ mov v3.16b, v2.16b
+ mov v7.16b, v2.16b
+ ushll v4.8h, v4.8b, #6
+ smlal v3.4s, v4.4h, v1.4h
+ smlal v3.4s, v5.4h, v0.4h
+ smlal2 v7.4s, v4.8h, v1.8h
+ smlal2 v7.4s, v5.8h, v0.8h
+ sshl v3.4s, v3.4s, v6.4s
+ sshl v7.4s, v7.4s, v6.4s
+ sqxtn v3.4h, v3.4s
+ sqxtn2 v3.8h, v7.4s
+ sqxtun v3.8b, v3.8h
+ str s3, [x0], #4
+ st1 {v3.h}[2], [x0], x1
+ b.ne 1b
+ ret
+endfunc
+
+function ff_hevc_put_hevc_pel_bi_w_pixels8_8_neon, export=1
+ load_bi_w_pixels_param
+ add w6, w6, #6 // log2Wd
+ dup v0.8h, w7 // wx0
+ dup v1.8h, w8 // wx1
+ add w9, w9, w10
+ add w9, w9, #1 // ox0 + ox1 + 1
+ lsl w9, w9, w6
+ add w7, w6, #1 // (log2Wd + 1)
+ mov x8, #(2 * HEVC_MAX_PB_SIZE)
+ neg w7, w7
+ dup v2.4s, w9 // (ox0 + ox1 + 1) << logwWd
+ dup v6.4s, w7 // -(log2Wd + 1)
+1:
+ ld1 {v4.8b}, [x2], x3 // load src
+ ld1 {v5.8h}, [x4], x8 // load src2
+ subs w5, w5, #1
+ mov v3.16b, v2.16b
+ mov v7.16b, v2.16b
+ ushll v4.8h, v4.8b, #6
+ smlal v3.4s, v4.4h, v1.4h
+ smlal v3.4s, v5.4h, v0.4h
+ smlal2 v7.4s, v4.8h, v1.8h
+ smlal2 v7.4s, v5.8h, v0.8h
+ sshl v3.4s, v3.4s, v6.4s
+ sshl v7.4s, v7.4s, v6.4s
+ sqxtn v3.4h, v3.4s
+ sqxtn2 v3.8h, v7.4s
+ sqxtun v3.8b, v3.8h
+ st1 {v3.8b}, [x0], x1
+ b.ne 1b
+ ret
+endfunc
+
+function ff_hevc_put_hevc_pel_bi_w_pixels12_8_neon, export=1
+ load_bi_w_pixels_param
+ add w6, w6, #6 // log2Wd
+ dup v0.8h, w7 // wx0
+ dup v1.8h, w8 // wx1
+ add w9, w9, w10
+ add w9, w9, #1 // ox0 + ox1 + 1
+ lsl w9, w9, w6
+ add w7, w6, #1 // (log2Wd + 1)
+ mov x8, #(2 * HEVC_MAX_PB_SIZE)
+ neg w7, w7
+ dup v2.4s, w9 // (ox0 + ox1 + 1) << logwWd
+ dup v6.4s, w7 // -(log2Wd + 1)
+ sub x1, x1, #8
+1:
+ ld1 {v24.16b}, [x2], x3 // load src
+ ld1 {v20.16b, v21.16b}, [x4], x8 // load src2
+ subs w5, w5, #1
+
+ mov v16.16b, v2.16b
+ mov v17.16b, v2.16b
+ mov v18.16b, v2.16b
+
+ ushll v4.8h, v24.8b, #6
+ ushll2 v24.8h, v24.16b, #6
+
+ smlal v16.4s, v4.4h, v1.4h
+ smlal v16.4s, v20.4h, v0.4h
+ smlal2 v17.4s, v4.8h, v1.8h
+ smlal2 v17.4s, v20.8h, v0.8h
+ smlal v18.4s, v24.4h, v1.4h
+ smlal v18.4s, v21.4h, v0.4h
+
+ sshl v16.4s, v16.4s, v6.4s
+ sshl v17.4s, v17.4s, v6.4s
+ sshl v18.4s, v18.4s, v6.4s
+
+ sqxtn v16.4h, v16.4s
+ sqxtn2 v16.8h, v17.4s
+ sqxtn v18.4h, v18.4s
+ sqxtun v3.8b, v16.8h
+ sqxtun2 v3.16b, v18.8h
+ str d3, [x0], #8
+ st1 {v3.s}[2], [x0], x1
+ b.ne 1b
+ ret
+endfunc
+
+function ff_hevc_put_hevc_pel_bi_w_pixels16_8_neon, export=1
+ load_bi_w_pixels_param
+ add w6, w6, #6 // log2Wd
+ dup v0.8h, w7 // wx0
+ dup v1.8h, w8 // wx1
+ add w9, w9, w10
+ add w9, w9, #1 // ox0 + ox1 + 1
+ lsl w9, w9, w6
+ add w7, w6, #1 // (log2Wd + 1)
+ mov x8, #(2 * HEVC_MAX_PB_SIZE)
+ neg w7, w7
+ dup v2.4s, w9 // (ox0 + ox1 + 1) << logwWd
+ dup v6.4s, w7 // -(log2Wd + 1)
+1:
+ ld1 {v24.16b}, [x2], x3 // load src
+ ld1 {v20.16b, v21.16b}, [x4], x8 // load src2
+ subs w5, w5, #1
+
+ mov v16.16b, v2.16b
+ mov v17.16b, v2.16b
+ mov v18.16b, v2.16b
+ mov v19.16b, v2.16b
+
+ ushll v4.8h, v24.8b, #6
+ ushll2 v24.8h, v24.16b, #6
+
+ smlal v16.4s, v4.4h, v1.4h
+ smlal v16.4s, v20.4h, v0.4h
+ smlal2 v17.4s, v4.8h, v1.8h
+ smlal2 v17.4s, v20.8h, v0.8h
+ smlal v18.4s, v24.4h, v1.4h
+ smlal v18.4s, v21.4h, v0.4h
+ smlal2 v19.4s, v24.8h, v1.8h
+ smlal2 v19.4s, v21.8h, v0.8h
+
+ sshl v16.4s, v16.4s, v6.4s
+ sshl v17.4s, v17.4s, v6.4s
+ sshl v18.4s, v18.4s, v6.4s
+ sshl v19.4s, v19.4s, v6.4s
+
+ sqxtn v16.4h, v16.4s
+ sqxtn2 v16.8h, v17.4s
+ sqxtn v18.4h, v18.4s
+ sqxtn2 v18.8h, v19.4s
+ sqxtun v3.8b, v16.8h
+ sqxtun2 v3.16b, v18.8h
+ st1 {v3.16b}, [x0], x1
+ b.ne 1b
+ ret
+endfunc
+
+function ff_hevc_put_hevc_pel_bi_w_pixels24_8_neon, export=1
+ load_bi_w_pixels_param
+ add w6, w6, #6 // log2Wd
+ dup v0.8h, w7 // wx0
+ dup v1.8h, w8 // wx1
+ add w9, w9, w10
+ add w9, w9, #1 // ox0 + ox1 + 1
+ lsl w9, w9, w6
+ add w7, w6, #1 // (log2Wd + 1)
+ mov x8, #(2 * HEVC_MAX_PB_SIZE)
+ neg w7, w7
+ dup v2.4s, w9 // (ox0 + ox1 + 1) << logwWd
+ dup v6.4s, w7 // -(log2Wd + 1)
+ mov x7, #24
+ sub x3, x3, x11
+ sub x8, x8, x11, lsl #1
+ sub x1, x1, x11
+1:
+ mov w6, w11
+2:
+ ld1 {v24.16b, v25.16b}, [x2], x7
+ ld1 {v20.8h, v21.8h, v22.8h}, [x4], #48
+
+ subs w6, w6, #24
+
+ mov v16.16b, v2.16b
+ mov v17.16b, v2.16b
+ mov v18.16b, v2.16b
+ mov v19.16b, v2.16b
+ mov v26.16b, v2.16b
+ mov v27.16b, v2.16b
+
+ ushll v4.8h, v24.8b, #6
+ ushll2 v24.8h, v24.16b, #6
+ ushll v5.8h, v25.8b, #6
+
+ smlal v16.4s, v4.4h, v1.4h
+ smlal v16.4s, v20.4h, v0.4h
+ smlal2 v17.4s, v4.8h, v1.8h
+ smlal2 v17.4s, v20.8h, v0.8h
+ smlal v18.4s, v24.4h, v1.4h
+ smlal v18.4s, v21.4h, v0.4h
+ smlal2 v19.4s, v24.8h, v1.8h
+ smlal2 v19.4s, v21.8h, v0.8h
+ smlal v26.4s, v5.4h, v1.4h
+ smlal v26.4s, v22.4h, v0.4h
+ smlal2 v27.4s, v5.8h, v1.8h
+ smlal2 v27.4s, v22.8h, v0.8h
+
+ sshl v16.4s, v16.4s, v6.4s
+ sshl v17.4s, v17.4s, v6.4s
+ sshl v18.4s, v18.4s, v6.4s
+ sshl v19.4s, v19.4s, v6.4s
+ sshl v26.4s, v26.4s, v6.4s
+ sshl v27.4s, v27.4s, v6.4s
+
+ sqxtn v16.4h, v16.4s
+ sqxtn2 v16.8h, v17.4s
+ sqxtn v18.4h, v18.4s
+ sqxtn2 v18.8h, v19.4s
+ sqxtn v26.4h, v26.4s
+ sqxtn2 v26.8h, v27.4s
+ sqxtun v3.8b, v16.8h
+ sqxtun2 v3.16b, v18.8h
+ sqxtun v4.8b, v26.8h
+ str q3, [x0], #16
+ str d4, [x0], #8
+ b.ne 2b
+
+ subs w5, w5, #1
+ add x0, x0, x1
+ add x2, x2, x3
+ add x4, x4, x8
+ b.ne 1b
+ ret
+endfunc
+
+function ff_hevc_put_hevc_pel_bi_w_pixels32_8_neon, export=1
+ load_bi_w_pixels_param
+ add w6, w6, #6 // log2Wd
+ dup v0.8h, w7 // wx0
+ dup v1.8h, w8 // wx1
+ add w9, w9, w10
+ add w9, w9, #1 // ox0 + ox1 + 1
+ lsl w9, w9, w6
+ add w7, w6, #1 // (log2Wd + 1)
+ mov x8, #(2 * HEVC_MAX_PB_SIZE)
+ neg w7, w7
+ dup v2.4s, w9 // (ox0 + ox1 + 1) << logwWd
+ dup v6.4s, w7 // -(log2Wd + 1)
+ sub x3, x3, x11
+ sub x8, x8, x11, lsl #1
+ sub x1, x1, x11
+1:
+ mov w6, w11
+2:
+ ld1 {v24.16b, v25.16b}, [x2], #32 //
load src
+ ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x4], #64 //
load src2
+
+ subs w6, w6, #32
+
+ mov v16.16b, v2.16b
+ mov v17.16b, v2.16b
+ mov v18.16b, v2.16b
+ mov v19.16b, v2.16b
+ mov v26.16b, v2.16b
+ mov v27.16b, v2.16b
+ mov v28.16b, v2.16b
+ mov v29.16b, v2.16b
+
+ ushll v4.8h, v24.8b, #6
+ ushll2 v24.8h, v24.16b, #6
+ ushll v5.8h, v25.8b, #6
+ ushll2 v25.8h, v25.16b, #6
+
+ smlal v16.4s, v4.4h, v1.4h
+ smlal v16.4s, v20.4h, v0.4h
+ smlal2 v17.4s, v4.8h, v1.8h
+ smlal2 v17.4s, v20.8h, v0.8h
+ smlal v18.4s, v24.4h, v1.4h
+ smlal v18.4s, v21.4h, v0.4h
+ smlal2 v19.4s, v24.8h, v1.8h
+ smlal2 v19.4s, v21.8h, v0.8h
+ smlal v26.4s, v5.4h, v1.4h
+ smlal v26.4s, v22.4h, v0.4h
+ smlal2 v27.4s, v5.8h, v1.8h
+ smlal2 v27.4s, v22.8h, v0.8h
+ smlal v28.4s, v25.4h, v1.4h
+ smlal v28.4s, v23.4h, v0.4h
+ smlal2 v29.4s, v25.8h, v1.8h
+ smlal2 v29.4s, v23.8h, v0.8h
+
+ sshl v16.4s, v16.4s, v6.4s
+ sshl v17.4s, v17.4s, v6.4s
+ sshl v18.4s, v18.4s, v6.4s
+ sshl v19.4s, v19.4s, v6.4s
+ sshl v26.4s, v26.4s, v6.4s
+ sshl v27.4s, v27.4s, v6.4s
+ sshl v28.4s, v28.4s, v6.4s
+ sshl v29.4s, v29.4s, v6.4s
+
+ sqxtn v16.4h, v16.4s
+ sqxtn2 v16.8h, v17.4s
+ sqxtn v18.4h, v18.4s
+ sqxtn2 v18.8h, v19.4s
+ sqxtn v26.4h, v26.4s
+ sqxtn2 v26.8h, v27.4s
+ sqxtn v28.4h, v28.4s
+ sqxtn2 v28.8h, v29.4s
+ sqxtun v3.8b, v16.8h
+ sqxtun2 v3.16b, v18.8h
+ sqxtun v4.8b, v26.8h
+ sqxtun2 v4.16b, v28.8h
+ st1 {v3.16b, v4.16b}, [x0], #32
+ b.ne 2b
+
+ subs w5, w5, #1
+ add x0, x0, x1
+ add x2, x2, x3
+ add x4, x4, x8
+ b.ne 1b
+ ret
+endfunc
function ff_hevc_put_hevc_epel_bi_h4_8_neon, export=1
load_epel_filterb x6, x7
diff --git a/libavcodec/aarch64/hevcdsp_init_aarch64.c
b/libavcodec/aarch64/hevcdsp_init_aarch64.c
index 0b159d1886..1d37cd76d8 100644
--- a/libavcodec/aarch64/hevcdsp_init_aarch64.c
+++ b/libavcodec/aarch64/hevcdsp_init_aarch64.c
@@ -134,6 +134,17 @@ void ff_hevc_transform_luma_4x4_neon_8(int16_t *coeffs);
member[7][v][h] = ff_hevc_put_hevc_##fn##32_8_neon##ext; \
member[9][v][h] = ff_hevc_put_hevc_##fn##64_8_neon##ext;
+#define NEON8_FNASSIGN_PARTIAL_6(member, v, h, fn, ext) \
+ member[1][v][h] = ff_hevc_put_hevc_##fn##4_8_neon##ext; \
+ member[2][v][h] = ff_hevc_put_hevc_##fn##6_8_neon##ext; \
+ member[3][v][h] = ff_hevc_put_hevc_##fn##8_8_neon##ext; \
+ member[4][v][h] = ff_hevc_put_hevc_##fn##12_8_neon##ext; \
+ member[5][v][h] = ff_hevc_put_hevc_##fn##16_8_neon##ext; \
+ member[6][v][h] = ff_hevc_put_hevc_##fn##24_8_neon##ext; \
+ member[7][v][h] = ff_hevc_put_hevc_##fn##32_8_neon##ext; \
+ member[8][v][h] = ff_hevc_put_hevc_##fn##24_8_neon##ext; \
+ member[9][v][h] = ff_hevc_put_hevc_##fn##32_8_neon##ext;
+
av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth)
{
int cpu_flags = av_get_cpu_flags();
@@ -204,6 +215,8 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c,
const int bit_depth)
NEON8_FNASSIGN(c->put_hevc_epel_bi, 1, 0, epel_bi_v,);
NEON8_FNASSIGN(c->put_hevc_qpel_bi, 0, 0, pel_bi_pixels,);
NEON8_FNASSIGN(c->put_hevc_qpel_bi, 1, 0, qpel_bi_v,);
+ NEON8_FNASSIGN_PARTIAL_6(c->put_hevc_qpel_bi_w, 0, 0,
pel_bi_w_pixels,);
+ NEON8_FNASSIGN_PARTIAL_6(c->put_hevc_epel_bi_w, 0, 0,
pel_bi_w_pixels,);
NEON8_FNASSIGN(c->put_hevc_epel_uni, 0, 0, pel_uni_pixels,);
NEON8_FNASSIGN(c->put_hevc_epel_uni, 1, 0, epel_uni_v,);
NEON8_FNASSIGN(c->put_hevc_qpel_uni, 0, 0, pel_uni_pixels,);
--
2.46.0
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
