History

lavc/aarch64: new optimization for 8-bit hevc_qpel_bi_hv

put_hevc_qpel_bi_hv4_8_c: 433.7 put_hevc_qpel_bi_hv4_8_i8mm: 117.9 put_hevc_qpel_bi_hv6_8_c: 803.9 put_hevc_qpel_bi_hv6_8_i8mm: 252.7 put_hevc_qpel_bi_hv8_8_c: 1296.4 put_hevc_qpel_bi_hv8_8_i8mm: 316.2 put_hevc_qpel_bi_hv12_8_c: 2867.4 put_hevc_qpel_bi_hv12_8_i8mm: 669.2 put_hevc_qpel_bi_hv16_8_c: 4709.4 put_hevc_qpel_bi_hv16_8_i8mm: 929.9 put_hevc_qpel_bi_hv24_8_c: 9639.7 put_hevc_qpel_bi_hv24_8_i8mm: 2072.4 put_hevc_qpel_bi_hv32_8_c: 16663.7 put_hevc_qpel_bi_hv32_8_i8mm: 3391.4 put_hevc_qpel_bi_hv48_8_c: 36972.9 put_hevc_qpel_bi_hv48_8_i8mm: 7505.7 put_hevc_qpel_bi_hv64_8_c: 64106.4 put_hevc_qpel_bi_hv64_8_i8mm: 13145.2

Logan Lyu committed 3 years ago

Commit fa047031 parent 595f9702 files · +304 −0Browse files

Showing the changes to libavcodec (2 of 2 files). Show all

Unified · Split

2 files
  1. libavcodec/aarch64/hevcdsp_init_aarch64.c +5 −0
  2. libavcodec/aarch64/hevcdsp_qpel_neon.S +299 −0
libavcodec/aarch64/hevcdsp_init_aarch64.c+5−0
@@ -255,6 +255,10 @@ NEON8_FNPROTO(qpel_bi_v, (uint8_t *dst, ptrdiff_t dststride, · 254 unmodified lines
255255 const uint8_t *src, ptrdiff_t srcstride, const int16_t *src2,
256256 int height, intptr_t mx, intptr_t my, int width),);
257257
258+NEON8_FNPROTO(qpel_bi_hv, (uint8_t *dst, ptrdiff_t dststride,
259+ const uint8_t *src, ptrdiff_t srcstride, const int16_t *src2,
260+ int height, intptr_t mx, intptr_t my, int width), _i8mm);
261+
258262 #define NEON8_FNASSIGN(member, v, h, fn, ext) \
259263 member[1][v][h] = ff_hevc_put_hevc_##fn##4_8_neon##ext; \
260264 member[2][v][h] = ff_hevc_put_hevc_##fn##6_8_neon##ext; \
@@ -370,6 +374,7 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth) · 109 unmodified lines
370374 NEON8_FNASSIGN(c->put_hevc_qpel_uni_w, 0, 1, qpel_uni_w_h, _i8mm);
371375 NEON8_FNASSIGN(c->put_hevc_epel_uni_w, 1, 1, epel_uni_w_hv, _i8mm);
372376 NEON8_FNASSIGN_PARTIAL_5(c->put_hevc_qpel_uni_w, 1, 1, qpel_uni_w_hv, _i8mm);
377+ NEON8_FNASSIGN(c->put_hevc_qpel_bi, 1, 1, qpel_bi_hv, _i8mm);
373378 }
374379
375380 }
libavcodec/aarch64/hevcdsp_qpel_neon.S+299−0
@@ -4200,5 +4200,304 @@ function ff_hevc_put_hevc_qpel_uni_w_hv64_8_neon_i8mm, export=1 · 4199 unmodified lines
42004200 ret
42014201 endfunc
42024202
4203+function ff_hevc_put_hevc_qpel_bi_hv4_8_neon_i8mm, export=1
4204+ add w10, w5, #7
4205+ lsl x10, x10, #7
4206+ sub sp, sp, x10 // tmp_array
4207+ stp x7, x30, [sp, #-48]!
4208+ stp x4, x5, [sp, #16]
4209+ stp x0, x1, [sp, #32]
4210+ sub x1, x2, x3, lsl #1
4211+ sub x1, x1, x3
4212+ add x0, sp, #48
4213+ mov x2, x3
4214+ add w3, w5, #7
4215+ mov x4, x6
4216+ bl X(ff_hevc_put_hevc_qpel_h4_8_neon_i8mm)
4217+ ldp x4, x5, [sp, #16]
4218+ ldp x0, x1, [sp, #32]
4219+ ldp x7, x30, [sp], #48
4220+ mov x9, #(MAX_PB_SIZE * 2)
4221+ load_qpel_filterh x7, x6
4222+ ld1 {v16.4h}, [sp], x9
4223+ ld1 {v17.4h}, [sp], x9
4224+ ld1 {v18.4h}, [sp], x9
4225+ ld1 {v19.4h}, [sp], x9
4226+ ld1 {v20.4h}, [sp], x9
4227+ ld1 {v21.4h}, [sp], x9
4228+ ld1 {v22.4h}, [sp], x9
4229+.macro calc tmp, src0, src1, src2, src3, src4, src5, src6, src7
4230+ ld1 {\tmp\().4h}, [sp], x9
4231+ calc_qpelh v1, \src0, \src1, \src2, \src3, \src4, \src5, \src6, \src7, sshr
4232+ ld1 {v5.4h}, [x4], x9 // src2
4233+ saddw v1.4s, v1.4s, v5.4h
4234+ rshrn v1.4h, v1.4s, #7
4235+ sqxtun v1.8b, v1.8h
4236+ subs w5, w5, #1
4237+ st1 {v1.s}[0], [x0], x1
4238+.endm
4239+1: calc_all
4240+.purgem calc
4241+2: ret
4242+endfunc
4243+
4244+function ff_hevc_put_hevc_qpel_bi_hv6_8_neon_i8mm, export=1
4245+ add w10, w5, #7
4246+ lsl x10, x10, #7
4247+ sub sp, sp, x10 // tmp_array
4248+ stp x7, x30, [sp, #-48]!
4249+ stp x4, x5, [sp, #16]
4250+ stp x0, x1, [sp, #32]
4251+ sub x1, x2, x3, lsl #1
4252+ sub x1, x1, x3
4253+ add x0, sp, #48
4254+ mov x2, x3
4255+ add x3, x5, #7
4256+ mov x4, x6
4257+ bl X(ff_hevc_put_hevc_qpel_h6_8_neon_i8mm)
4258+ ldp x4, x5, [sp, #16]
4259+ ldp x0, x1, [sp, #32]
4260+ ldp x7, x30, [sp], #48
4261+ mov x9, #(MAX_PB_SIZE * 2)
4262+ load_qpel_filterh x7, x6
4263+ sub x1, x1, #4
4264+ ld1 {v16.8h}, [sp], x9
4265+ ld1 {v17.8h}, [sp], x9
4266+ ld1 {v18.8h}, [sp], x9
4267+ ld1 {v19.8h}, [sp], x9
4268+ ld1 {v20.8h}, [sp], x9
4269+ ld1 {v21.8h}, [sp], x9
4270+ ld1 {v22.8h}, [sp], x9
4271+.macro calc tmp, src0, src1, src2, src3, src4, src5, src6, src7
4272+ ld1 {\tmp\().8h}, [sp], x9
4273+ calc_qpelh v1, \src0, \src1, \src2, \src3, \src4, \src5, \src6, \src7, sshr
4274+ calc_qpelh2 v2, v2, \src0, \src1, \src2, \src3, \src4, \src5, \src6, \src7, sshr
4275+ ld1 {v5.8h}, [x4], x9 // src2
4276+ saddw v1.4s, v1.4s, v5.4h
4277+ saddw2 v2.4s, v2.4s, v5.8h
4278+ rshrn v1.4h, v1.4s, #7
4279+ rshrn2 v1.8h, v2.4s, #7
4280+ sqxtun v1.8b, v1.8h
4281+ st1 {v1.s}[0], [x0], #4
4282+ subs w5, w5, #1
4283+ st1 {v1.h}[2], [x0], x1
4284+.endm
4285+1: calc_all
4286+.purgem calc
4287+2: ret
4288+endfunc
4289+
4290+function ff_hevc_put_hevc_qpel_bi_hv8_8_neon_i8mm, export=1
4291+ add w10, w5, #7
4292+ lsl x10, x10, #7
4293+ sub sp, sp, x10 // tmp_array
4294+ stp x7, x30, [sp, #-48]!
4295+ stp x4, x5, [sp, #16]
4296+ stp x0, x1, [sp, #32]
4297+ sub x1, x2, x3, lsl #1
4298+ sub x1, x1, x3
4299+ add x0, sp, #48
4300+ mov x2, x3
4301+ add x3, x5, #7
4302+ mov x4, x6
4303+ bl X(ff_hevc_put_hevc_qpel_h8_8_neon_i8mm)
4304+ ldp x4, x5, [sp, #16]
4305+ ldp x0, x1, [sp, #32]
4306+ ldp x7, x30, [sp], #48
4307+ mov x9, #(MAX_PB_SIZE * 2)
4308+ load_qpel_filterh x7, x6
4309+ ld1 {v16.8h}, [sp], x9
4310+ ld1 {v17.8h}, [sp], x9
4311+ ld1 {v18.8h}, [sp], x9
4312+ ld1 {v19.8h}, [sp], x9
4313+ ld1 {v20.8h}, [sp], x9
4314+ ld1 {v21.8h}, [sp], x9
4315+ ld1 {v22.8h}, [sp], x9
4316+.macro calc tmp, src0, src1, src2, src3, src4, src5, src6, src7
4317+ ld1 {\tmp\().8h}, [sp], x9
4318+ calc_qpelh v1, \src0, \src1, \src2, \src3, \src4, \src5, \src6, \src7, sshr
4319+ calc_qpelh2 v2, v2, \src0, \src1, \src2, \src3, \src4, \src5, \src6, \src7, sshr
4320+ ld1 {v5.8h}, [x4], x9 // src2
4321+ saddw v1.4s, v1.4s, v5.4h
4322+ saddw2 v2.4s, v2.4s, v5.8h
4323+ rshrn v1.4h, v1.4s, #7
4324+ rshrn2 v1.8h, v2.4s, #7
4325+ sqxtun v1.8b, v1.8h
4326+ subs w5, w5, #1
4327+ st1 {v1.8b}, [x0], x1
4328+.endm
4329+1: calc_all
4330+.purgem calc
4331+2: ret
4332+endfunc
4333+
4334+function ff_hevc_put_hevc_qpel_bi_hv12_8_neon_i8mm, export=1
4335+ stp x6, x7, [sp, #-80]!
4336+ stp x4, x5, [sp, #16]
4337+ stp x2, x3, [sp, #32]
4338+ stp x0, x1, [sp, #48]
4339+ str x30, [sp, #64]
4340+ bl X(ff_hevc_put_hevc_qpel_bi_hv8_8_neon_i8mm)
4341+ ldp x4, x5, [sp, #16]
4342+ ldp x2, x3, [sp, #32]
4343+ ldp x0, x1, [sp, #48]
4344+ ldp x6, x7, [sp], #64
4345+ add x4, x4, #16
4346+ add x2, x2, #8
4347+ add x0, x0, #8
4348+ bl X(ff_hevc_put_hevc_qpel_bi_hv4_8_neon_i8mm)
4349+ ldr x30, [sp], #16
4350+ ret
4351+endfunc
4352+
4353+function ff_hevc_put_hevc_qpel_bi_hv16_8_neon_i8mm, export=1
4354+ add w10, w5, #7
4355+ lsl x10, x10, #7
4356+ sub sp, sp, x10 // tmp_array
4357+ stp x7, x30, [sp, #-48]!
4358+ stp x4, x5, [sp, #16]
4359+ stp x0, x1, [sp, #32]
4360+ add x0, sp, #48
4361+ sub x1, x2, x3, lsl #1
4362+ sub x1, x1, x3
4363+ mov x2, x3
4364+ add w3, w5, #7
4365+ mov x4, x6
4366+ bl X(ff_hevc_put_hevc_qpel_h16_8_neon_i8mm)
4367+ ldp x4, x5, [sp, #16]
4368+ ldp x0, x1, [sp, #32]
4369+ ldp x7, x30, [sp], #48
4370+ mov x6, #16 // width
4371+.Lqpel_bi_hv16_loop:
4372+ load_qpel_filterh x7, x8
4373+ mov x9, #(MAX_PB_SIZE * 2)
4374+ mov x10, x6
4375+0: mov x8, sp // src
4376+ ld1 {v16.8h, v17.8h}, [x8], x9
4377+ mov w11, w5 // height
4378+ ld1 {v18.8h, v19.8h}, [x8], x9
4379+ mov x12, x4 // src2
4380+ ld1 {v20.8h, v21.8h}, [x8], x9
4381+ mov x7, x0 // dst
4382+ ld1 {v22.8h, v23.8h}, [x8], x9
4383+ ld1 {v24.8h, v25.8h}, [x8], x9
4384+ ld1 {v26.8h, v27.8h}, [x8], x9
4385+ ld1 {v28.8h, v29.8h}, [x8], x9
4386+.macro calc tmp0, tmp1, src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10, src11, src12, src13, src14, src15
4387+ ld1 {\tmp0\().8h, \tmp1\().8h}, [x8], x9
4388+ calc_qpelh v1, \src0, \src1, \src2, \src3, \src4, \src5, \src6, \src7, sshr
4389+ calc_qpelh2 v2, v2, \src0, \src1, \src2, \src3, \src4, \src5, \src6, \src7, sshr
4390+ calc_qpelh v3, \src8, \src9, \src10, \src11, \src12, \src13, \src14, \src15, sshr
4391+ calc_qpelh2 v4, v4, \src8, \src9, \src10, \src11, \src12, \src13, \src14, \src15, sshr
4392+ ld1 {v5.8h, v6.8h}, [x12], x9 // src2
4393+ saddw v1.4s, v1.4s, v5.4h
4394+ saddw2 v2.4s, v2.4s, v5.8h
4395+ saddw v3.4s, v3.4s, v6.4h
4396+ saddw2 v4.4s, v4.4s, v6.8h
4397+ rshrn v1.4h, v1.4s, #7
4398+ rshrn2 v1.8h, v2.4s, #7
4399+ rshrn v2.4h, v3.4s, #7
4400+ rshrn2 v2.8h, v4.4s, #7
4401+ sqxtun v1.8b, v1.8h
4402+ sqxtun2 v1.16b, v2.8h
4403+ subs x11, x11, #1
4404+ st1 {v1.16b}, [x7], x1
4405+.endm
4406+1: calc_all2
4407+.purgem calc
4408+2: add x0, x0, #16
4409+ add sp, sp, #32
4410+ subs x10, x10, #16
4411+ add x4, x4, #32
4412+ b.ne 0b
4413+ add w10, w5, #7
4414+ lsl x10, x10, #7
4415+ sub x10, x10, x6, lsl #1 // part of first line
4416+ add sp, sp, x10 // tmp_array without first line
4417+ ret
4418+endfunc
4419+
4420+function ff_hevc_put_hevc_qpel_bi_hv24_8_neon_i8mm, export=1
4421+ stp x6, x7, [sp, #-80]!
4422+ stp x4, x5, [sp, #16]
4423+ stp x2, x3, [sp, #32]
4424+ stp x0, x1, [sp, #48]
4425+ str x30, [sp, #64]
4426+ bl X(ff_hevc_put_hevc_qpel_bi_hv16_8_neon_i8mm)
4427+ ldp x4, x5, [sp, #16]
4428+ ldp x2, x3, [sp, #32]
4429+ ldp x0, x1, [sp, #48]
4430+ ldp x6, x7, [sp], #64
4431+ add x4, x4, #32
4432+ add x2, x2, #16
4433+ add x0, x0, #16
4434+ bl X(ff_hevc_put_hevc_qpel_bi_hv8_8_neon_i8mm)
4435+ ldr x30, [sp], #16
4436+ ret
4437+endfunc
4438+
4439+function ff_hevc_put_hevc_qpel_bi_hv32_8_neon_i8mm, export=1
4440+ add w10, w5, #7
4441+ lsl x10, x10, #7
4442+ sub sp, sp, x10 // tmp_array
4443+ stp x7, x30, [sp, #-48]!
4444+ stp x4, x5, [sp, #16]
4445+ stp x0, x1, [sp, #32]
4446+ add x0, sp, #48
4447+ sub x1, x2, x3, lsl #1
4448+ mov x2, x3
4449+ sub x1, x1, x3
4450+ add w3, w5, #7
4451+ mov x4, x6
4452+ bl X(ff_hevc_put_hevc_qpel_h32_8_neon_i8mm)
4453+ ldp x4, x5, [sp, #16]
4454+ ldp x0, x1, [sp, #32]
4455+ ldp x7, x30, [sp], #48
4456+ mov x6, #32 // width
4457+ b .Lqpel_bi_hv16_loop
4458+endfunc
4459+
4460+function ff_hevc_put_hevc_qpel_bi_hv48_8_neon_i8mm, export=1
4461+ add w10, w5, #7
4462+ lsl x10, x10, #7
4463+ sub sp, sp, x10 // tmp_array
4464+ stp x7, x30, [sp, #-48]!
4465+ stp x4, x5, [sp, #16]
4466+ stp x0, x1, [sp, #32]
4467+ add x0, sp, #48
4468+ sub x1, x2, x3, lsl #1
4469+ mov x2, x3
4470+ sub x1, x1, x3
4471+ add w3, w5, #7
4472+ mov x4, x6
4473+ bl X(ff_hevc_put_hevc_qpel_h48_8_neon_i8mm)
4474+ ldp x4, x5, [sp, #16]
4475+ ldp x0, x1, [sp, #32]
4476+ ldp x7, x30, [sp], #48
4477+ mov x6, #48 // width
4478+ b .Lqpel_bi_hv16_loop
4479+endfunc
4480+
4481+function ff_hevc_put_hevc_qpel_bi_hv64_8_neon_i8mm, export=1
4482+ add w10, w5, #7
4483+ lsl x10, x10, #7
4484+ sub sp, sp, x10 // tmp_array
4485+ stp x7, x30, [sp, #-48]!
4486+ stp x4, x5, [sp, #16]
4487+ stp x0, x1, [sp, #32]
4488+ add x0, sp, #48
4489+ sub x1, x2, x3, lsl #1
4490+ mov x2, x3
4491+ sub x1, x1, x3
4492+ add w3, w5, #7
4493+ mov x4, x6
4494+ bl X(ff_hevc_put_hevc_qpel_h64_8_neon_i8mm)
4495+ ldp x4, x5, [sp, #16]
4496+ ldp x0, x1, [sp, #32]
4497+ ldp x7, x30, [sp], #48
4498+ mov x6, #64 // width
4499+ b .Lqpel_bi_hv16_loop
4500+endfunc
4501+
42034502 DISABLE_I8MM
42044503 #endif // HAVE_I8MM

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t