123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619 |
- ;******************************************************************************
- ;* Copyright (c) 2012 Michael Niedermayer
- ;* Copyright (c) 2014 James Almer <jamrial <at> gmail.com>
- ;* Copyright (c) 2014 Ronald S. Bultje <rsbultje@gmail.com>
- ;*
- ;* This file is part of FFmpeg.
- ;*
- ;* FFmpeg is free software; you can redistribute it and/or
- ;* modify it under the terms of the GNU Lesser General Public
- ;* License as published by the Free Software Foundation; either
- ;* version 2.1 of the License, or (at your option) any later version.
- ;*
- ;* FFmpeg is distributed in the hope that it will be useful,
- ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
- ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- ;* Lesser General Public License for more details.
- ;*
- ;* You should have received a copy of the GNU Lesser General Public
- ;* License along with FFmpeg; if not, write to the Free Software
- ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- ;******************************************************************************
- %include "libavutil/x86/x86util.asm"
- %if ARCH_X86_64
- %define pointer resq
- %else
- %define pointer resd
- %endif
- struc ResampleContext
- .av_class: pointer 1
- .filter_bank: pointer 1
- .filter_length: resd 1
- .filter_alloc: resd 1
- .ideal_dst_incr: resd 1
- .dst_incr: resd 1
- .dst_incr_div: resd 1
- .dst_incr_mod: resd 1
- .index: resd 1
- .frac: resd 1
- .src_incr: resd 1
- .compensation_distance: resd 1
- .phase_count: resd 1
- ; there's a few more here but we only care about the first few
- endstruc
- SECTION_RODATA
- pf_1: dd 1.0
- pdbl_1: dq 1.0
- pd_0x4000: dd 0x4000
- SECTION .text
- ; FIXME remove unneeded variables (index_incr, phase_mask)
- %macro RESAMPLE_FNS 3-5 ; format [float or int16], bps, log2_bps, float op suffix [s or d], 1.0 constant
- ; int resample_common_$format(ResampleContext *ctx, $format *dst,
- ; const $format *src, int size, int update_ctx)
- %if ARCH_X86_64 ; unix64 and win64
- cglobal resample_common_%1, 0, 15, 2, ctx, dst, src, phase_count, index, frac, \
- dst_incr_mod, size, min_filter_count_x4, \
- min_filter_len_x4, dst_incr_div, src_incr, \
- phase_mask, dst_end, filter_bank
- ; use red-zone for variable storage
- %define ctx_stackq [rsp-0x8]
- %define src_stackq [rsp-0x10]
- %if WIN64
- %define update_context_stackd r4m
- %else ; unix64
- %define update_context_stackd [rsp-0x14]
- %endif
- ; load as many variables in registers as possible; for the rest, store
- ; on stack so that we have 'ctx' available as one extra register
- mov sized, r3d
- %if UNIX64
- mov update_context_stackd, r4d
- %endif
- mov indexd, [ctxq+ResampleContext.index]
- mov fracd, [ctxq+ResampleContext.frac]
- mov dst_incr_modd, [ctxq+ResampleContext.dst_incr_mod]
- mov filter_bankq, [ctxq+ResampleContext.filter_bank]
- mov src_incrd, [ctxq+ResampleContext.src_incr]
- mov ctx_stackq, ctxq
- mov min_filter_len_x4d, [ctxq+ResampleContext.filter_length]
- mov dst_incr_divd, [ctxq+ResampleContext.dst_incr_div]
- shl min_filter_len_x4d, %3
- lea dst_endq, [dstq+sizeq*%2]
- %if UNIX64
- mov ecx, [ctxq+ResampleContext.phase_count]
- mov edi, [ctxq+ResampleContext.filter_alloc]
- DEFINE_ARGS filter_alloc, dst, src, phase_count, index, frac, dst_incr_mod, \
- filter, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
- src_incr, phase_mask, dst_end, filter_bank
- %elif WIN64
- mov R9d, [ctxq+ResampleContext.filter_alloc]
- mov ecx, [ctxq+ResampleContext.phase_count]
- DEFINE_ARGS phase_count, dst, src, filter_alloc, index, frac, dst_incr_mod, \
- filter, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
- src_incr, phase_mask, dst_end, filter_bank
- %endif
- neg min_filter_len_x4q
- sub filter_bankq, min_filter_len_x4q
- sub srcq, min_filter_len_x4q
- mov src_stackq, srcq
- %else ; x86-32
- cglobal resample_common_%1, 1, 7, 2, ctx, phase_count, dst, frac, \
- index, min_filter_length_x4, filter_bank
- ; push temp variables to stack
- %define ctx_stackq r0mp
- %define src_stackq r2mp
- %define update_context_stackd r4m
- mov dstq, r1mp
- mov r3, r3mp
- lea r3, [dstq+r3*%2]
- PUSH dword [ctxq+ResampleContext.dst_incr_div]
- PUSH dword [ctxq+ResampleContext.dst_incr_mod]
- PUSH dword [ctxq+ResampleContext.filter_alloc]
- PUSH r3
- PUSH dword [ctxq+ResampleContext.phase_count] ; unneeded replacement for phase_mask
- PUSH dword [ctxq+ResampleContext.src_incr]
- mov min_filter_length_x4d, [ctxq+ResampleContext.filter_length]
- mov indexd, [ctxq+ResampleContext.index]
- shl min_filter_length_x4d, %3
- mov fracd, [ctxq+ResampleContext.frac]
- neg min_filter_length_x4q
- mov filter_bankq, [ctxq+ResampleContext.filter_bank]
- sub r2mp, min_filter_length_x4q
- sub filter_bankq, min_filter_length_x4q
- PUSH min_filter_length_x4q
- PUSH filter_bankq
- mov phase_countd, [ctxq+ResampleContext.phase_count]
- DEFINE_ARGS src, phase_count, dst, frac, index, min_filter_count_x4, filter
- %define filter_bankq dword [rsp+0x0]
- %define min_filter_length_x4q dword [rsp+0x4]
- %define src_incrd dword [rsp+0x8]
- %define phase_maskd dword [rsp+0xc]
- %define dst_endq dword [rsp+0x10]
- %define filter_allocd dword [rsp+0x14]
- %define dst_incr_modd dword [rsp+0x18]
- %define dst_incr_divd dword [rsp+0x1c]
- mov srcq, r2mp
- %endif
- .loop:
- mov filterd, filter_allocd
- imul filterd, indexd
- %if ARCH_X86_64
- mov min_filter_count_x4q, min_filter_len_x4q
- lea filterq, [filter_bankq+filterq*%2]
- %else ; x86-32
- mov min_filter_count_x4q, filter_bankq
- lea filterq, [min_filter_count_x4q+filterq*%2]
- mov min_filter_count_x4q, min_filter_length_x4q
- %endif
- %ifidn %1, int16
- movd m0, [pd_0x4000]
- %else ; float/double
- xorps m0, m0, m0
- %endif
- align 16
- .inner_loop:
- movu m1, [srcq+min_filter_count_x4q*1]
- %ifidn %1, int16
- %if cpuflag(xop)
- vpmadcswd m0, m1, [filterq+min_filter_count_x4q*1], m0
- %else
- pmaddwd m1, [filterq+min_filter_count_x4q*1]
- paddd m0, m1
- %endif
- %else ; float/double
- %if cpuflag(fma4) || cpuflag(fma3)
- fmaddp%4 m0, m1, [filterq+min_filter_count_x4q*1], m0
- %else
- mulp%4 m1, m1, [filterq+min_filter_count_x4q*1]
- addp%4 m0, m0, m1
- %endif ; cpuflag
- %endif
- add min_filter_count_x4q, mmsize
- js .inner_loop
- %ifidn %1, int16
- HADDD m0, m1
- psrad m0, 15
- add fracd, dst_incr_modd
- packssdw m0, m0
- add indexd, dst_incr_divd
- movd [dstq], m0
- %else ; float/double
- ; horizontal sum & store
- %if mmsize == 32
- vextractf128 xm1, m0, 0x1
- addp%4 xm0, xm1
- %endif
- movhlps xm1, xm0
- %ifidn %1, float
- addps xm0, xm1
- shufps xm1, xm0, xm0, q0001
- %endif
- add fracd, dst_incr_modd
- addp%4 xm0, xm1
- add indexd, dst_incr_divd
- movs%4 [dstq], xm0
- %endif
- cmp fracd, src_incrd
- jl .skip
- sub fracd, src_incrd
- inc indexd
- %if UNIX64
- DEFINE_ARGS filter_alloc, dst, src, phase_count, index, frac, dst_incr_mod, \
- index_incr, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
- src_incr, phase_mask, dst_end, filter_bank
- %elif WIN64
- DEFINE_ARGS phase_count, dst, src, filter_alloc, index, frac, dst_incr_mod, \
- index_incr, min_filter_count_x4, min_filter_len_x4, dst_incr_div, \
- src_incr, phase_mask, dst_end, filter_bank
- %else ; x86-32
- DEFINE_ARGS src, phase_count, dst, frac, index, index_incr
- %endif
- .skip:
- add dstq, %2
- cmp indexd, phase_countd
- jb .index_skip
- .index_while:
- sub indexd, phase_countd
- lea srcq, [srcq+%2]
- cmp indexd, phase_countd
- jnb .index_while
- .index_skip:
- cmp dstq, dst_endq
- jne .loop
- %if ARCH_X86_64
- DEFINE_ARGS ctx, dst, src, phase_count, index, frac
- %else ; x86-32
- DEFINE_ARGS src, ctx, update_context, frac, index
- %endif
- cmp dword update_context_stackd, 0
- jz .skip_store
- ; strictly speaking, the function should always return the consumed
- ; number of bytes; however, we only use the value if update_context
- ; is true, so let's just leave it uninitialized otherwise
- mov ctxq, ctx_stackq
- movifnidn rax, srcq
- mov [ctxq+ResampleContext.frac ], fracd
- sub rax, src_stackq
- mov [ctxq+ResampleContext.index], indexd
- shr rax, %3
- .skip_store:
- %if ARCH_X86_32
- ADD rsp, 0x20
- %endif
- RET
- ; int resample_linear_$format(ResampleContext *ctx, float *dst,
- ; const float *src, int size, int update_ctx)
- %if ARCH_X86_64 ; unix64 and win64
- %if UNIX64
- cglobal resample_linear_%1, 0, 15, 5, ctx, dst, phase_mask, phase_count, index, frac, \
- size, dst_incr_mod, min_filter_count_x4, \
- min_filter_len_x4, dst_incr_div, src_incr, \
- src, dst_end, filter_bank
- mov srcq, r2mp
- %else ; win64
- cglobal resample_linear_%1, 0, 15, 5, ctx, phase_mask, src, phase_count, index, frac, \
- size, dst_incr_mod, min_filter_count_x4, \
- min_filter_len_x4, dst_incr_div, src_incr, \
- dst, dst_end, filter_bank
- mov dstq, r1mp
- %endif
- ; use red-zone for variable storage
- %define ctx_stackq [rsp-0x8]
- %define src_stackq [rsp-0x10]
- %define phase_mask_stackd [rsp-0x14]
- %if WIN64
- %define update_context_stackd r4m
- %else ; unix64
- %define update_context_stackd [rsp-0x18]
- %endif
- ; load as many variables in registers as possible; for the rest, store
- ; on stack so that we have 'ctx' available as one extra register
- mov sized, r3d
- %if UNIX64
- mov update_context_stackd, r4d
- %endif
- mov indexd, [ctxq+ResampleContext.index]
- mov fracd, [ctxq+ResampleContext.frac]
- mov dst_incr_modd, [ctxq+ResampleContext.dst_incr_mod]
- mov filter_bankq, [ctxq+ResampleContext.filter_bank]
- mov src_incrd, [ctxq+ResampleContext.src_incr]
- mov ctx_stackq, ctxq
- mov min_filter_len_x4d, [ctxq+ResampleContext.filter_length]
- %ifidn %1, int16
- movd m4, [pd_0x4000]
- %else ; float/double
- cvtsi2s%4 xm0, src_incrd
- movs%4 xm4, [%5]
- divs%4 xm4, xm0
- %endif
- mov dst_incr_divd, [ctxq+ResampleContext.dst_incr_div]
- shl min_filter_len_x4d, %3
- lea dst_endq, [dstq+sizeq*%2]
- %if UNIX64
- mov ecx, [ctxq+ResampleContext.phase_count]
- mov edi, [ctxq+ResampleContext.filter_alloc]
- DEFINE_ARGS filter_alloc, dst, filter2, phase_count, index, frac, filter1, \
- dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
- dst_incr_div, src_incr, src, dst_end, filter_bank
- %elif WIN64
- mov R9d, [ctxq+ResampleContext.filter_alloc]
- mov ecx, [ctxq+ResampleContext.phase_count]
- DEFINE_ARGS phase_count, filter2, src, filter_alloc, index, frac, filter1, \
- dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
- dst_incr_div, src_incr, dst, dst_end, filter_bank
- %endif
- neg min_filter_len_x4q
- sub filter_bankq, min_filter_len_x4q
- sub srcq, min_filter_len_x4q
- mov src_stackq, srcq
- %else ; x86-32
- cglobal resample_linear_%1, 1, 7, 5, ctx, min_filter_length_x4, filter2, \
- frac, index, dst, filter_bank
- ; push temp variables to stack
- %define ctx_stackq r0mp
- %define src_stackq r2mp
- %define update_context_stackd r4m
- mov dstq, r1mp
- mov r3, r3mp
- lea r3, [dstq+r3*%2]
- PUSH dword [ctxq+ResampleContext.dst_incr_div]
- PUSH r3
- mov r3, dword [ctxq+ResampleContext.filter_alloc]
- PUSH dword [ctxq+ResampleContext.dst_incr_mod]
- PUSH r3
- shl r3, %3
- PUSH r3
- mov r3, dword [ctxq+ResampleContext.src_incr]
- PUSH dword [ctxq+ResampleContext.phase_count] ; unneeded replacement of phase_mask
- PUSH r3d
- %ifidn %1, int16
- movd m4, [pd_0x4000]
- %else ; float/double
- cvtsi2s%4 xm0, r3d
- movs%4 xm4, [%5]
- divs%4 xm4, xm0
- %endif
- mov min_filter_length_x4d, [ctxq+ResampleContext.filter_length]
- mov indexd, [ctxq+ResampleContext.index]
- shl min_filter_length_x4d, %3
- mov fracd, [ctxq+ResampleContext.frac]
- neg min_filter_length_x4q
- mov filter_bankq, [ctxq+ResampleContext.filter_bank]
- sub r2mp, min_filter_length_x4q
- sub filter_bankq, min_filter_length_x4q
- PUSH min_filter_length_x4q
- PUSH filter_bankq
- PUSH dword [ctxq+ResampleContext.phase_count]
- DEFINE_ARGS filter1, min_filter_count_x4, filter2, frac, index, dst, src
- %define phase_count_stackd dword [rsp+0x0]
- %define filter_bankq dword [rsp+0x4]
- %define min_filter_length_x4q dword [rsp+0x8]
- %define src_incrd dword [rsp+0xc]
- %define phase_mask_stackd dword [rsp+0x10]
- %define filter_alloc_x4q dword [rsp+0x14]
- %define filter_allocd dword [rsp+0x18]
- %define dst_incr_modd dword [rsp+0x1c]
- %define dst_endq dword [rsp+0x20]
- %define dst_incr_divd dword [rsp+0x24]
- mov srcq, r2mp
- %endif
- .loop:
- mov filter1d, filter_allocd
- imul filter1d, indexd
- %if ARCH_X86_64
- mov min_filter_count_x4q, min_filter_len_x4q
- lea filter1q, [filter_bankq+filter1q*%2]
- lea filter2q, [filter1q+filter_allocq*%2]
- %else ; x86-32
- mov min_filter_count_x4q, filter_bankq
- lea filter1q, [min_filter_count_x4q+filter1q*%2]
- mov min_filter_count_x4q, min_filter_length_x4q
- mov filter2q, filter1q
- add filter2q, filter_alloc_x4q
- %endif
- %ifidn %1, int16
- mova m0, m4
- mova m2, m4
- %else ; float/double
- xorps m0, m0, m0
- xorps m2, m2, m2
- %endif
- align 16
- .inner_loop:
- movu m1, [srcq+min_filter_count_x4q*1]
- %ifidn %1, int16
- %if cpuflag(xop)
- vpmadcswd m2, m1, [filter2q+min_filter_count_x4q*1], m2
- vpmadcswd m0, m1, [filter1q+min_filter_count_x4q*1], m0
- %else
- pmaddwd m3, m1, [filter2q+min_filter_count_x4q*1]
- pmaddwd m1, [filter1q+min_filter_count_x4q*1]
- paddd m2, m3
- paddd m0, m1
- %endif ; cpuflag
- %else ; float/double
- %if cpuflag(fma4) || cpuflag(fma3)
- fmaddp%4 m2, m1, [filter2q+min_filter_count_x4q*1], m2
- fmaddp%4 m0, m1, [filter1q+min_filter_count_x4q*1], m0
- %else
- mulp%4 m3, m1, [filter2q+min_filter_count_x4q*1]
- mulp%4 m1, m1, [filter1q+min_filter_count_x4q*1]
- addp%4 m2, m2, m3
- addp%4 m0, m0, m1
- %endif ; cpuflag
- %endif
- add min_filter_count_x4q, mmsize
- js .inner_loop
- %ifidn %1, int16
- %if mmsize == 16
- %if cpuflag(xop)
- vphadddq m2, m2
- vphadddq m0, m0
- %endif
- pshufd m3, m2, q0032
- pshufd m1, m0, q0032
- paddd m2, m3
- paddd m0, m1
- %endif
- %if notcpuflag(xop)
- PSHUFLW m3, m2, q0032
- PSHUFLW m1, m0, q0032
- paddd m2, m3
- paddd m0, m1
- %endif
- psubd m2, m0
- ; This is probably a really bad idea on atom and other machines with a
- ; long transfer latency between GPRs and XMMs (atom). However, it does
- ; make the clip a lot simpler...
- movd eax, m2
- add indexd, dst_incr_divd
- imul fracd
- idiv src_incrd
- movd m1, eax
- add fracd, dst_incr_modd
- paddd m0, m1
- psrad m0, 15
- packssdw m0, m0
- movd [dstq], m0
- ; note that for imul/idiv, I need to move filter to edx/eax for each:
- ; - 32bit: eax=r0[filter1], edx=r2[filter2]
- ; - win64: eax=r6[filter1], edx=r1[todo]
- ; - unix64: eax=r6[filter1], edx=r2[todo]
- %else ; float/double
- ; val += (v2 - val) * (FELEML) frac / c->src_incr;
- %if mmsize == 32
- vextractf128 xm1, m0, 0x1
- vextractf128 xm3, m2, 0x1
- addp%4 xm0, xm1
- addp%4 xm2, xm3
- %endif
- cvtsi2s%4 xm1, fracd
- subp%4 xm2, xm0
- mulp%4 xm1, xm4
- shufp%4 xm1, xm1, q0000
- %if cpuflag(fma4) || cpuflag(fma3)
- fmaddp%4 xm0, xm2, xm1, xm0
- %else
- mulp%4 xm2, xm1
- addp%4 xm0, xm2
- %endif ; cpuflag
- ; horizontal sum & store
- movhlps xm1, xm0
- %ifidn %1, float
- addps xm0, xm1
- shufps xm1, xm0, xm0, q0001
- %endif
- add fracd, dst_incr_modd
- addp%4 xm0, xm1
- add indexd, dst_incr_divd
- movs%4 [dstq], xm0
- %endif
- cmp fracd, src_incrd
- jl .skip
- sub fracd, src_incrd
- inc indexd
- %if UNIX64
- DEFINE_ARGS filter_alloc, dst, filter2, phase_count, index, frac, index_incr, \
- dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
- dst_incr_div, src_incr, src, dst_end, filter_bank
- %elif WIN64
- DEFINE_ARGS phase_count, filter2, src, filter_alloc, index, frac, index_incr, \
- dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
- dst_incr_div, src_incr, dst, dst_end, filter_bank
- %else ; x86-32
- DEFINE_ARGS filter1, phase_count, index_incr, frac, index, dst, src
- %endif
- .skip:
- %if ARCH_X86_32
- mov phase_countd, phase_count_stackd
- %endif
- add dstq, %2
- cmp indexd, phase_countd
- jb .index_skip
- .index_while:
- sub indexd, phase_countd
- lea srcq, [srcq+%2]
- cmp indexd, phase_countd
- jnb .index_while
- .index_skip:
- cmp dstq, dst_endq
- jne .loop
- %if UNIX64
- DEFINE_ARGS ctx, dst, filter2, phase_count, index, frac, index_incr, \
- dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
- dst_incr_div, src_incr, src, dst_end, filter_bank
- %elif WIN64
- DEFINE_ARGS ctx, filter2, src, phase_count, index, frac, index_incr, \
- dst_incr_mod, min_filter_count_x4, min_filter_len_x4, \
- dst_incr_div, src_incr, dst, dst_end, filter_bank
- %else ; x86-32
- DEFINE_ARGS filter1, ctx, update_context, frac, index, dst, src
- %endif
- cmp dword update_context_stackd, 0
- jz .skip_store
- ; strictly speaking, the function should always return the consumed
- ; number of bytes; however, we only use the value if update_context
- ; is true, so let's just leave it uninitialized otherwise
- mov ctxq, ctx_stackq
- movifnidn rax, srcq
- mov [ctxq+ResampleContext.frac ], fracd
- sub rax, src_stackq
- mov [ctxq+ResampleContext.index], indexd
- shr rax, %3
- .skip_store:
- %if ARCH_X86_32
- ADD rsp, 0x28
- %endif
- RET
- %endmacro
- INIT_XMM sse
- RESAMPLE_FNS float, 4, 2, s, pf_1
- %if HAVE_AVX_EXTERNAL
- INIT_YMM avx
- RESAMPLE_FNS float, 4, 2, s, pf_1
- %endif
- %if HAVE_FMA3_EXTERNAL
- INIT_YMM fma3
- RESAMPLE_FNS float, 4, 2, s, pf_1
- %endif
- %if HAVE_FMA4_EXTERNAL
- INIT_XMM fma4
- RESAMPLE_FNS float, 4, 2, s, pf_1
- %endif
- %if ARCH_X86_32
- INIT_MMX mmxext
- RESAMPLE_FNS int16, 2, 1
- %endif
- INIT_XMM sse2
- RESAMPLE_FNS int16, 2, 1
- %if HAVE_XOP_EXTERNAL
- INIT_XMM xop
- RESAMPLE_FNS int16, 2, 1
- %endif
- INIT_XMM sse2
- RESAMPLE_FNS double, 8, 3, d, pdbl_1
- %if HAVE_AVX_EXTERNAL
- INIT_YMM avx
- RESAMPLE_FNS double, 8, 3, d, pdbl_1
- %endif
- %if HAVE_FMA3_EXTERNAL
- INIT_YMM fma3
- RESAMPLE_FNS double, 8, 3, d, pdbl_1
- %endif
|