diff options
Diffstat (limited to 'libswresample/x86')
-rw-r--r-- | libswresample/x86/Makefile | 9 | ||||
-rw-r--r-- | libswresample/x86/audio_convert.asm | 739 | ||||
-rw-r--r-- | libswresample/x86/audio_convert_init.c | 181 | ||||
-rw-r--r-- | libswresample/x86/rematrix.asm | 250 | ||||
-rw-r--r-- | libswresample/x86/rematrix_init.c | 90 | ||||
-rw-r--r-- | libswresample/x86/resample.asm | 605 | ||||
-rw-r--r-- | libswresample/x86/resample_init.c | 90 | ||||
-rw-r--r-- | libswresample/x86/w64xmmtest.c | 29 |
8 files changed, 1993 insertions, 0 deletions
diff --git a/libswresample/x86/Makefile b/libswresample/x86/Makefile new file mode 100644 index 0000000000..be44df56aa --- /dev/null +++ b/libswresample/x86/Makefile @@ -0,0 +1,9 @@ +YASM-OBJS += x86/audio_convert.o\ + x86/rematrix.o\ + x86/resample.o\ + +OBJS += x86/audio_convert_init.o\ + x86/rematrix_init.o\ + x86/resample_init.o\ + +OBJS-$(CONFIG_XMM_CLOBBER_TEST) += x86/w64xmmtest.o diff --git a/libswresample/x86/audio_convert.asm b/libswresample/x86/audio_convert.asm new file mode 100644 index 0000000000..d441636d3c --- /dev/null +++ b/libswresample/x86/audio_convert.asm @@ -0,0 +1,739 @@ +;****************************************************************************** +;* Copyright (c) 2012 Michael Niedermayer +;* +;* 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" + +SECTION_RODATA 32 +flt2pm31: times 8 dd 4.6566129e-10 +flt2p31 : times 8 dd 2147483648.0 +flt2p15 : times 8 dd 32768.0 + +word_unpack_shuf : db 0, 1, 4, 5, 8, 9,12,13, 2, 3, 6, 7,10,11,14,15 + +SECTION .text + + +;to, from, a/u, log2_outsize, log_intsize, const +%macro PACK_2CH 5-7 +cglobal pack_2ch_%2_to_%1_%3, 3, 4, 6, dst, src, len, src2 + mov src2q , [srcq+gprsize] + mov srcq , [srcq] + mov dstq , [dstq] +%ifidn %3, a + test dstq, mmsize-1 + jne pack_2ch_%2_to_%1_u_int %+ SUFFIX + test srcq, mmsize-1 + jne pack_2ch_%2_to_%1_u_int %+ SUFFIX + test src2q, mmsize-1 + jne pack_2ch_%2_to_%1_u_int %+ SUFFIX +%else +pack_2ch_%2_to_%1_u_int %+ SUFFIX: +%endif + lea srcq , [srcq + (1<<%5)*lenq] + lea src2q, [src2q + (1<<%5)*lenq] + lea dstq , [dstq + (2<<%4)*lenq] + neg lenq + %7 m0,m1,m2,m3,m4,m5 +.next: +%if %4 >= %5 + mov%3 m0, [ srcq +(1<<%5)*lenq] + mova m1, m0 + mov%3 m2, [ src2q+(1<<%5)*lenq] +%if %5 == 1 + punpcklwd m0, m2 + punpckhwd m1, m2 +%else + punpckldq m0, m2 + punpckhdq m1, m2 +%endif + %6 m0,m1,m2,m3,m4,m5 +%else + mov%3 m0, [ srcq +(1<<%5)*lenq] + mov%3 m1, [mmsize + srcq +(1<<%5)*lenq] + mov%3 m2, [ src2q+(1<<%5)*lenq] + mov%3 m3, [mmsize + src2q+(1<<%5)*lenq] + %6 m0,m1,m2,m3,m4,m5 + mova m2, m0 + punpcklwd m0, m1 + punpckhwd m2, m1 + SWAP 1,2 +%endif + mov%3 [ dstq+(2<<%4)*lenq], m0 + mov%3 [ mmsize + dstq+(2<<%4)*lenq], m1 +%if %4 > %5 + mov%3 [2*mmsize + dstq+(2<<%4)*lenq], m2 + mov%3 [3*mmsize + dstq+(2<<%4)*lenq], m3 + add lenq, 4*mmsize/(2<<%4) +%else + add lenq, 2*mmsize/(2<<%4) +%endif + jl .next + REP_RET +%endmacro + +%macro UNPACK_2CH 5-7 +cglobal unpack_2ch_%2_to_%1_%3, 3, 4, 7, dst, src, len, dst2 + mov dst2q , [dstq+gprsize] + mov srcq , [srcq] + mov dstq , [dstq] +%ifidn %3, a + test dstq, mmsize-1 + jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX + test srcq, mmsize-1 + jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX + test dst2q, mmsize-1 + jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX +%else +unpack_2ch_%2_to_%1_u_int %+ SUFFIX: +%endif + lea srcq , [srcq + (2<<%5)*lenq] + lea dstq , [dstq + (1<<%4)*lenq] + lea dst2q, [dst2q + (1<<%4)*lenq] + neg lenq + %7 m0,m1,m2,m3,m4,m5 + mova m6, [word_unpack_shuf] +.next: + mov%3 m0, [ srcq +(2<<%5)*lenq] + mov%3 m2, [ mmsize + srcq +(2<<%5)*lenq] +%if %5 == 1 +%ifidn SUFFIX, _ssse3 + pshufb m0, m6 + mova m1, m0 + pshufb m2, m6 + punpcklqdq m0,m2 + punpckhqdq m1,m2 +%else + mova m1, m0 + punpcklwd m0,m2 + punpckhwd m1,m2 + + mova m2, m0 + punpcklwd m0,m1 + punpckhwd m2,m1 + + mova m1, m0 + punpcklwd m0,m2 + punpckhwd m1,m2 +%endif +%else + mova m1, m0 + shufps m0, m2, 10001000b + shufps m1, m2, 11011101b +%endif +%if %4 < %5 + mov%3 m2, [2*mmsize + srcq +(2<<%5)*lenq] + mova m3, m2 + mov%3 m4, [3*mmsize + srcq +(2<<%5)*lenq] + shufps m2, m4, 10001000b + shufps m3, m4, 11011101b + SWAP 1,2 +%endif + %6 m0,m1,m2,m3,m4,m5 + mov%3 [ dstq+(1<<%4)*lenq], m0 +%if %4 > %5 + mov%3 [ dst2q+(1<<%4)*lenq], m2 + mov%3 [ mmsize + dstq+(1<<%4)*lenq], m1 + mov%3 [ mmsize + dst2q+(1<<%4)*lenq], m3 + add lenq, 2*mmsize/(1<<%4) +%else + mov%3 [ dst2q+(1<<%4)*lenq], m1 + add lenq, mmsize/(1<<%4) +%endif + jl .next + REP_RET +%endmacro + +%macro CONV 5-7 +cglobal %2_to_%1_%3, 3, 3, 6, dst, src, len + mov srcq , [srcq] + mov dstq , [dstq] +%ifidn %3, a + test dstq, mmsize-1 + jne %2_to_%1_u_int %+ SUFFIX + test srcq, mmsize-1 + jne %2_to_%1_u_int %+ SUFFIX +%else +%2_to_%1_u_int %+ SUFFIX: +%endif + lea srcq , [srcq + (1<<%5)*lenq] + lea dstq , [dstq + (1<<%4)*lenq] + neg lenq + %7 m0,m1,m2,m3,m4,m5 +.next: + mov%3 m0, [ srcq +(1<<%5)*lenq] + mov%3 m1, [ mmsize + srcq +(1<<%5)*lenq] +%if %4 < %5 + mov%3 m2, [2*mmsize + srcq +(1<<%5)*lenq] + mov%3 m3, [3*mmsize + srcq +(1<<%5)*lenq] +%endif + %6 m0,m1,m2,m3,m4,m5 + mov%3 [ dstq+(1<<%4)*lenq], m0 + mov%3 [ mmsize + dstq+(1<<%4)*lenq], m1 +%if %4 > %5 + mov%3 [2*mmsize + dstq+(1<<%4)*lenq], m2 + mov%3 [3*mmsize + dstq+(1<<%4)*lenq], m3 + add lenq, 4*mmsize/(1<<%4) +%else + add lenq, 2*mmsize/(1<<%4) +%endif + jl .next +%if mmsize == 8 + emms + RET +%else + REP_RET +%endif +%endmacro + +%macro PACK_6CH 8 +cglobal pack_6ch_%2_to_%1_%3, 2, 8, %6, dst, src, src1, src2, src3, src4, src5, len +%if ARCH_X86_64 + mov lend, r2d +%else + %define lend dword r2m +%endif + mov src1q, [srcq+1*gprsize] + mov src2q, [srcq+2*gprsize] + mov src3q, [srcq+3*gprsize] + mov src4q, [srcq+4*gprsize] + mov src5q, [srcq+5*gprsize] + mov srcq, [srcq] + mov dstq, [dstq] +%ifidn %3, a + test dstq, mmsize-1 + jne pack_6ch_%2_to_%1_u_int %+ SUFFIX + test srcq, mmsize-1 + jne pack_6ch_%2_to_%1_u_int %+ SUFFIX + test src1q, mmsize-1 + jne pack_6ch_%2_to_%1_u_int %+ SUFFIX + test src2q, mmsize-1 + jne pack_6ch_%2_to_%1_u_int %+ SUFFIX + test src3q, mmsize-1 + jne pack_6ch_%2_to_%1_u_int %+ SUFFIX + test src4q, mmsize-1 + jne pack_6ch_%2_to_%1_u_int %+ SUFFIX + test src5q, mmsize-1 + jne pack_6ch_%2_to_%1_u_int %+ SUFFIX +%else +pack_6ch_%2_to_%1_u_int %+ SUFFIX: +%endif + sub src1q, srcq + sub src2q, srcq + sub src3q, srcq + sub src4q, srcq + sub src5q, srcq + %8 x,x,x,x,m7,x +.loop: + mov%3 m0, [srcq ] + mov%3 m1, [srcq+src1q] + mov%3 m2, [srcq+src2q] + mov%3 m3, [srcq+src3q] + mov%3 m4, [srcq+src4q] + mov%3 m5, [srcq+src5q] +%if cpuflag(sse) + SBUTTERFLYPS 0, 1, 6 + SBUTTERFLYPS 2, 3, 6 + SBUTTERFLYPS 4, 5, 6 + +%if cpuflag(avx) + blendps m6, m4, m0, 1100b +%else + movaps m6, m4 + shufps m4, m0, q3210 + SWAP 4,6 +%endif + movlhps m0, m2 + movhlps m4, m2 +%if cpuflag(avx) + blendps m2, m5, m1, 1100b +%else + movaps m2, m5 + shufps m5, m1, q3210 + SWAP 2,5 +%endif + movlhps m1, m3 + movhlps m5, m3 + + %7 m0,m6,x,x,m7,m3 + %7 m4,m1,x,x,m7,m3 + %7 m2,m5,x,x,m7,m3 + + mov %+ %3 %+ ps [dstq ], m0 + mov %+ %3 %+ ps [dstq+16], m6 + mov %+ %3 %+ ps [dstq+32], m4 + mov %+ %3 %+ ps [dstq+48], m1 + mov %+ %3 %+ ps [dstq+64], m2 + mov %+ %3 %+ ps [dstq+80], m5 +%else ; mmx + SBUTTERFLY dq, 0, 1, 6 + SBUTTERFLY dq, 2, 3, 6 + SBUTTERFLY dq, 4, 5, 6 + + movq [dstq ], m0 + movq [dstq+ 8], m2 + movq [dstq+16], m4 + movq [dstq+24], m1 + movq [dstq+32], m3 + movq [dstq+40], m5 +%endif + add srcq, mmsize + add dstq, mmsize*6 + sub lend, mmsize/4 + jg .loop +%if mmsize == 8 + emms + RET +%else + REP_RET +%endif +%endmacro + +%macro UNPACK_6CH 8 +cglobal unpack_6ch_%2_to_%1_%3, 2, 8, %6, dst, src, dst1, dst2, dst3, dst4, dst5, len +%if ARCH_X86_64 + mov lend, r2d +%else + %define lend dword r2m +%endif + mov dst1q, [dstq+1*gprsize] + mov dst2q, [dstq+2*gprsize] + mov dst3q, [dstq+3*gprsize] + mov dst4q, [dstq+4*gprsize] + mov dst5q, [dstq+5*gprsize] + mov dstq, [dstq] + mov srcq, [srcq] +%ifidn %3, a + test dstq, mmsize-1 + jne unpack_6ch_%2_to_%1_u_int %+ SUFFIX + test srcq, mmsize-1 + jne unpack_6ch_%2_to_%1_u_int %+ SUFFIX + test dst1q, mmsize-1 + jne unpack_6ch_%2_to_%1_u_int %+ SUFFIX + test dst2q, mmsize-1 + jne unpack_6ch_%2_to_%1_u_int %+ SUFFIX + test dst3q, mmsize-1 + jne unpack_6ch_%2_to_%1_u_int %+ SUFFIX + test dst4q, mmsize-1 + jne unpack_6ch_%2_to_%1_u_int %+ SUFFIX + test dst5q, mmsize-1 + jne unpack_6ch_%2_to_%1_u_int %+ SUFFIX +%else +unpack_6ch_%2_to_%1_u_int %+ SUFFIX: +%endif + sub dst1q, dstq + sub dst2q, dstq + sub dst3q, dstq + sub dst4q, dstq + sub dst5q, dstq + %8 x,x,x,x,m7,x +.loop: + mov%3 m0, [srcq ] + mov%3 m1, [srcq+16] + mov%3 m2, [srcq+32] + mov%3 m3, [srcq+48] + mov%3 m4, [srcq+64] + mov%3 m5, [srcq+80] + + SBUTTERFLYPS 0, 3, 6 + SBUTTERFLYPS 1, 4, 6 + SBUTTERFLYPS 2, 5, 6 + SBUTTERFLYPS 0, 4, 6 + SBUTTERFLYPS 3, 2, 6 + SBUTTERFLYPS 1, 5, 6 + SWAP 1, 4 + SWAP 2, 3 + + %7 m0,m1,x,x,m7,m6 + %7 m2,m3,x,x,m7,m6 + %7 m4,m5,x,x,m7,m6 + + mov %+ %3 %+ ps [dstq ], m0 + mov %+ %3 %+ ps [dstq+dst1q], m1 + mov %+ %3 %+ ps [dstq+dst2q], m2 + mov %+ %3 %+ ps [dstq+dst3q], m3 + mov %+ %3 %+ ps [dstq+dst4q], m4 + mov %+ %3 %+ ps [dstq+dst5q], m5 + + add srcq, mmsize*6 + add dstq, mmsize + sub lend, mmsize/4 + jg .loop + REP_RET +%endmacro + +%define PACK_8CH_GPRS (10 * ARCH_X86_64) + ((6 + HAVE_ALIGNED_STACK) * ARCH_X86_32) + +%macro PACK_8CH 8 +cglobal pack_8ch_%2_to_%1_%3, 2, PACK_8CH_GPRS, %6, ARCH_X86_32*48, dst, src, len, src1, src2, src3, src4, src5, src6, src7 + mov dstq, [dstq] +%if ARCH_X86_32 + DEFINE_ARGS dst, src, src2, src3, src4, src5, src6 + %define lend dword r2m + %define src1q r0q + %define src1m dword [rsp+32] +%if HAVE_ALIGNED_STACK == 0 + DEFINE_ARGS dst, src, src2, src3, src5, src6 + %define src4q r0q + %define src4m dword [rsp+36] +%endif + %define src7q r0q + %define src7m dword [rsp+40] + mov dstm, dstq +%endif + mov src7q, [srcq+7*gprsize] + mov src6q, [srcq+6*gprsize] +%if ARCH_X86_32 + mov src7m, src7q +%endif + mov src5q, [srcq+5*gprsize] + mov src4q, [srcq+4*gprsize] + mov src3q, [srcq+3*gprsize] +%if ARCH_X86_32 && HAVE_ALIGNED_STACK == 0 + mov src4m, src4q +%endif + mov src2q, [srcq+2*gprsize] + mov src1q, [srcq+1*gprsize] + mov srcq, [srcq] +%ifidn %3, a +%if ARCH_X86_32 + test dstmp, mmsize-1 +%else + test dstq, mmsize-1 +%endif + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX + test srcq, mmsize-1 + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX + test src1q, mmsize-1 + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX + test src2q, mmsize-1 + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX + test src3q, mmsize-1 + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX +%if ARCH_X86_32 && HAVE_ALIGNED_STACK == 0 + test src4m, mmsize-1 +%else + test src4q, mmsize-1 +%endif + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX + test src5q, mmsize-1 + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX + test src6q, mmsize-1 + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX +%if ARCH_X86_32 + test src7m, mmsize-1 +%else + test src7q, mmsize-1 +%endif + jne pack_8ch_%2_to_%1_u_int %+ SUFFIX +%else +pack_8ch_%2_to_%1_u_int %+ SUFFIX: +%endif + sub src1q, srcq + sub src2q, srcq + sub src3q, srcq +%if ARCH_X86_64 || HAVE_ALIGNED_STACK + sub src4q, srcq +%else + sub src4m, srcq +%endif + sub src5q, srcq + sub src6q, srcq +%if ARCH_X86_64 + sub src7q, srcq +%else + mov src1m, src1q + sub src7m, srcq +%endif + +%if ARCH_X86_64 + %8 x,x,x,x,m9,x +%elifidn %1, int32 + %define m9 [flt2p31] +%else + %define m9 [flt2pm31] +%endif + +.loop: + mov%3 m0, [srcq ] + mov%3 m1, [srcq+src1q] + mov%3 m2, [srcq+src2q] +%if ARCH_X86_32 && HAVE_ALIGNED_STACK == 0 + mov src4q, src4m +%endif + mov%3 m3, [srcq+src3q] + mov%3 m4, [srcq+src4q] + mov%3 m5, [srcq+src5q] +%if ARCH_X86_32 + mov src7q, src7m +%endif + mov%3 m6, [srcq+src6q] + mov%3 m7, [srcq+src7q] + +%if ARCH_X86_64 + TRANSPOSE8x4D 0, 1, 2, 3, 4, 5, 6, 7, 8 + + %7 m0,m1,x,x,m9,m8 + %7 m2,m3,x,x,m9,m8 + %7 m4,m5,x,x,m9,m8 + %7 m6,m7,x,x,m9,m8 + + mov%3 [dstq], m0 +%else + mov dstq, dstm + + TRANSPOSE8x4D 0, 1, 2, 3, 4, 5, 6, 7, [rsp], [rsp+16], 1 + + %7 m0,m1,x,x,m9,m2 + mova m2, [rsp] + mov%3 [dstq], m0 + %7 m2,m3,x,x,m9,m0 + %7 m4,m5,x,x,m9,m0 + %7 m6,m7,x,x,m9,m0 + +%endif + + mov%3 [dstq+16], m1 + mov%3 [dstq+32], m2 + mov%3 [dstq+48], m3 + mov%3 [dstq+64], m4 + mov%3 [dstq+80], m5 + mov%3 [dstq+96], m6 + mov%3 [dstq+112], m7 + + add srcq, mmsize + add dstq, mmsize*8 +%if ARCH_X86_32 + mov dstm, dstq + mov src1q, src1m +%endif + sub lend, mmsize/4 + jg .loop + REP_RET +%endmacro + +%macro INT16_TO_INT32_N 6 + pxor m2, m2 + pxor m3, m3 + punpcklwd m2, m1 + punpckhwd m3, m1 + SWAP 4,0 + pxor m0, m0 + pxor m1, m1 + punpcklwd m0, m4 + punpckhwd m1, m4 +%endmacro + +%macro INT32_TO_INT16_N 6 + psrad m0, 16 + psrad m1, 16 + psrad m2, 16 + psrad m3, 16 + packssdw m0, m1 + packssdw m2, m3 + SWAP 1,2 +%endmacro + +%macro INT32_TO_FLOAT_INIT 6 + mova %5, [flt2pm31] +%endmacro +%macro INT32_TO_FLOAT_N 6 + cvtdq2ps %1, %1 + cvtdq2ps %2, %2 + mulps %1, %1, %5 + mulps %2, %2, %5 +%endmacro + +%macro FLOAT_TO_INT32_INIT 6 + mova %5, [flt2p31] +%endmacro +%macro FLOAT_TO_INT32_N 6 + mulps %1, %5 + mulps %2, %5 + cvtps2dq %6, %1 + cmpps %1, %1, %5, 5 + paddd %1, %6 + cvtps2dq %6, %2 + cmpps %2, %2, %5, 5 + paddd %2, %6 +%endmacro + +%macro INT16_TO_FLOAT_INIT 6 + mova m5, [flt2pm31] +%endmacro +%macro INT16_TO_FLOAT_N 6 + INT16_TO_INT32_N %1,%2,%3,%4,%5,%6 + cvtdq2ps m0, m0 + cvtdq2ps m1, m1 + cvtdq2ps m2, m2 + cvtdq2ps m3, m3 + mulps m0, m0, m5 + mulps m1, m1, m5 + mulps m2, m2, m5 + mulps m3, m3, m5 +%endmacro + +%macro FLOAT_TO_INT16_INIT 6 + mova m5, [flt2p15] +%endmacro +%macro FLOAT_TO_INT16_N 6 + mulps m0, m5 + mulps m1, m5 + mulps m2, m5 + mulps m3, m5 + cvtps2dq m0, m0 + cvtps2dq m1, m1 + packssdw m0, m1 + cvtps2dq m1, m2 + cvtps2dq m3, m3 + packssdw m1, m3 +%endmacro + +%macro NOP_N 0-6 +%endmacro + +INIT_MMX mmx +CONV int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N +CONV int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N +CONV int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N +CONV int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N + +PACK_6CH float, float, u, 2, 2, 0, NOP_N, NOP_N +PACK_6CH float, float, a, 2, 2, 0, NOP_N, NOP_N + +INIT_XMM sse +PACK_6CH float, float, u, 2, 2, 7, NOP_N, NOP_N +PACK_6CH float, float, a, 2, 2, 7, NOP_N, NOP_N + +UNPACK_6CH float, float, u, 2, 2, 7, NOP_N, NOP_N +UNPACK_6CH float, float, a, 2, 2, 7, NOP_N, NOP_N + +INIT_XMM sse2 +CONV int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N +CONV int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N +CONV int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N +CONV int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N + +PACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N +PACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N +PACK_2CH int32, int32, u, 2, 2, NOP_N, NOP_N +PACK_2CH int32, int32, a, 2, 2, NOP_N, NOP_N +PACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N +PACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N +PACK_2CH int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N +PACK_2CH int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N + +UNPACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N +UNPACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N +UNPACK_2CH int32, int32, u, 2, 2, NOP_N, NOP_N +UNPACK_2CH int32, int32, a, 2, 2, NOP_N, NOP_N +UNPACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N +UNPACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N +UNPACK_2CH int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N +UNPACK_2CH int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N + +CONV float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +CONV float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +CONV int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +CONV int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +CONV float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT +CONV float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT +CONV int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT +CONV int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT + +PACK_2CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_2CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_2CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +PACK_2CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +PACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT +PACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT +PACK_2CH int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT +PACK_2CH int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT + +UNPACK_2CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +UNPACK_2CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +UNPACK_2CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +UNPACK_2CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +UNPACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT +UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT +UNPACK_2CH int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT +UNPACK_2CH int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT + +PACK_6CH float, int32, u, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_6CH float, int32, a, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_6CH int32, float, u, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +PACK_6CH int32, float, a, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT + +UNPACK_6CH float, int32, u, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +UNPACK_6CH float, int32, a, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +UNPACK_6CH int32, float, u, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +UNPACK_6CH int32, float, a, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT + +PACK_8CH float, float, u, 2, 2, 9, NOP_N, NOP_N +PACK_8CH float, float, a, 2, 2, 9, NOP_N, NOP_N + +PACK_8CH float, int32, u, 2, 2, 10, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_8CH float, int32, a, 2, 2, 10, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_8CH int32, float, u, 2, 2, 10, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +PACK_8CH int32, float, a, 2, 2, 10, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT + +INIT_XMM ssse3 +UNPACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N +UNPACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N +UNPACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N +UNPACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N +UNPACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT +UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT + +%if HAVE_AVX_EXTERNAL +INIT_XMM avx +PACK_6CH float, float, u, 2, 2, 8, NOP_N, NOP_N +PACK_6CH float, float, a, 2, 2, 8, NOP_N, NOP_N + +UNPACK_6CH float, float, u, 2, 2, 8, NOP_N, NOP_N +UNPACK_6CH float, float, a, 2, 2, 8, NOP_N, NOP_N + +PACK_6CH float, int32, u, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_6CH float, int32, a, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_6CH int32, float, u, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +PACK_6CH int32, float, a, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT + +UNPACK_6CH float, int32, u, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +UNPACK_6CH float, int32, a, 2, 2, 8, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +UNPACK_6CH int32, float, u, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +UNPACK_6CH int32, float, a, 2, 2, 8, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT + +PACK_8CH float, float, u, 2, 2, 9, NOP_N, NOP_N +PACK_8CH float, float, a, 2, 2, 9, NOP_N, NOP_N + +PACK_8CH float, int32, u, 2, 2, 10, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_8CH float, int32, a, 2, 2, 10, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +PACK_8CH int32, float, u, 2, 2, 10, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +PACK_8CH int32, float, a, 2, 2, 10, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT + +INIT_YMM avx +CONV float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +CONV float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT +%endif + +%if HAVE_AVX2_EXTERNAL +INIT_YMM avx2 +CONV int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +CONV int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT +%endif diff --git a/libswresample/x86/audio_convert_init.c b/libswresample/x86/audio_convert_init.c new file mode 100644 index 0000000000..bb89cf604b --- /dev/null +++ b/libswresample/x86/audio_convert_init.c @@ -0,0 +1,181 @@ +/* + * Copyright (C) 2012 Michael Niedermayer (michaelni@gmx.at) + * + * This file is part of libswresample + * + * libswresample 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. + * + * libswresample 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 libswresample; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include "libavutil/x86/cpu.h" +#include "libswresample/swresample_internal.h" +#include "libswresample/audioconvert.h" + +#define PROTO(pre, in, out, cap) void ff ## pre ## in## _to_ ##out## _a_ ##cap(uint8_t **dst, const uint8_t **src, int len); +#define PROTO2(pre, out, cap) PROTO(pre, int16, out, cap) PROTO(pre, int32, out, cap) PROTO(pre, float, out, cap) +#define PROTO3(pre, cap) PROTO2(pre, int16, cap) PROTO2(pre, int32, cap) PROTO2(pre, float, cap) +#define PROTO4(pre) PROTO3(pre, mmx) PROTO3(pre, sse) PROTO3(pre, sse2) PROTO3(pre, ssse3) PROTO3(pre, sse4) PROTO3(pre, avx) PROTO3(pre, avx2) +PROTO4(_) +PROTO4(_pack_2ch_) +PROTO4(_pack_6ch_) +PROTO4(_pack_8ch_) +PROTO4(_unpack_2ch_) +PROTO4(_unpack_6ch_) + +av_cold void swri_audio_convert_init_x86(struct AudioConvert *ac, + enum AVSampleFormat out_fmt, + enum AVSampleFormat in_fmt, + int channels){ + int mm_flags = av_get_cpu_flags(); + + ac->simd_f= NULL; + +//FIXME add memcpy case + +#define MULTI_CAPS_FUNC(flag, cap) \ + if (EXTERNAL_##flag(mm_flags)) {\ + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S16 || out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_S16P)\ + ac->simd_f = ff_int16_to_int32_a_ ## cap;\ + if( out_fmt == AV_SAMPLE_FMT_S16 && in_fmt == AV_SAMPLE_FMT_S32 || out_fmt == AV_SAMPLE_FMT_S16P && in_fmt == AV_SAMPLE_FMT_S32P)\ + ac->simd_f = ff_int32_to_int16_a_ ## cap;\ + } + +MULTI_CAPS_FUNC(MMX, mmx) +MULTI_CAPS_FUNC(SSE2, sse2) + + if(EXTERNAL_MMX(mm_flags)) { + if(channels == 6) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_6ch_float_to_float_a_mmx; + } + } + if(EXTERNAL_SSE(mm_flags)) { + if(channels == 6) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_6ch_float_to_float_a_sse; + + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_FLT || out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_S32) + ac->simd_f = ff_unpack_6ch_float_to_float_a_sse; + } + } + if(EXTERNAL_SSE2(mm_flags)) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S32 || out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_int32_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S16 || out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S16P) + ac->simd_f = ff_int16_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_FLT || out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_float_to_int32_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S16 && in_fmt == AV_SAMPLE_FMT_FLT || out_fmt == AV_SAMPLE_FMT_S16P && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_float_to_int16_a_sse2; + + if(channels == 2) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_2ch_int32_to_int32_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S16 && in_fmt == AV_SAMPLE_FMT_S16P) + ac->simd_f = ff_pack_2ch_int16_to_int16_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S16P) + ac->simd_f = ff_pack_2ch_int16_to_int32_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S16 && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_2ch_int32_to_int16_a_sse2; + + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_FLT || out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_S32) + ac->simd_f = ff_unpack_2ch_int32_to_int32_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S16P && in_fmt == AV_SAMPLE_FMT_S16) + ac->simd_f = ff_unpack_2ch_int16_to_int16_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_S16) + ac->simd_f = ff_unpack_2ch_int16_to_int32_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S16P && in_fmt == AV_SAMPLE_FMT_S32) + ac->simd_f = ff_unpack_2ch_int32_to_int16_a_sse2; + + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_2ch_int32_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_pack_2ch_float_to_int32_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S16P) + ac->simd_f = ff_pack_2ch_int16_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S16 && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_pack_2ch_float_to_int16_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S32) + ac->simd_f = ff_unpack_2ch_int32_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_FLT) + ac->simd_f = ff_unpack_2ch_float_to_int32_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S16) + ac->simd_f = ff_unpack_2ch_int16_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S16P && in_fmt == AV_SAMPLE_FMT_FLT) + ac->simd_f = ff_unpack_2ch_float_to_int16_a_sse2; + } + if(channels == 6) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_6ch_int32_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_pack_6ch_float_to_int32_a_sse2; + + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S32) + ac->simd_f = ff_unpack_6ch_int32_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_FLT) + ac->simd_f = ff_unpack_6ch_float_to_int32_a_sse2; + } + if(channels == 8) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_8ch_float_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_8ch_int32_to_float_a_sse2; + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_pack_8ch_float_to_int32_a_sse2; + } + } + if(EXTERNAL_SSSE3(mm_flags)) { + if(channels == 2) { + if( out_fmt == AV_SAMPLE_FMT_S16P && in_fmt == AV_SAMPLE_FMT_S16) + ac->simd_f = ff_unpack_2ch_int16_to_int16_a_ssse3; + if( out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_S16) + ac->simd_f = ff_unpack_2ch_int16_to_int32_a_ssse3; + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S16) + ac->simd_f = ff_unpack_2ch_int16_to_float_a_ssse3; + } + } + if(EXTERNAL_AVX_FAST(mm_flags)) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S32 || out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_int32_to_float_a_avx; + } + if(EXTERNAL_AVX(mm_flags)) { + if(channels == 6) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_6ch_float_to_float_a_avx; + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_6ch_int32_to_float_a_avx; + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_pack_6ch_float_to_int32_a_avx; + + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_FLT || out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_S32) + ac->simd_f = ff_unpack_6ch_float_to_float_a_avx; + if( out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S32) + ac->simd_f = ff_unpack_6ch_int32_to_float_a_avx; + if( out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_FLT) + ac->simd_f = ff_unpack_6ch_float_to_int32_a_avx; + } + if(channels == 8) { + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_8ch_float_to_float_a_avx; + if( out_fmt == AV_SAMPLE_FMT_FLT && in_fmt == AV_SAMPLE_FMT_S32P) + ac->simd_f = ff_pack_8ch_int32_to_float_a_avx; + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_pack_8ch_float_to_int32_a_avx; + } + } + if(EXTERNAL_AVX2_FAST(mm_flags)) { + if( out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_FLT || out_fmt == AV_SAMPLE_FMT_S32P && in_fmt == AV_SAMPLE_FMT_FLTP) + ac->simd_f = ff_float_to_int32_a_avx2; + } +} diff --git a/libswresample/x86/rematrix.asm b/libswresample/x86/rematrix.asm new file mode 100644 index 0000000000..7984b9a729 --- /dev/null +++ b/libswresample/x86/rematrix.asm @@ -0,0 +1,250 @@ +;****************************************************************************** +;* Copyright (c) 2012 Michael Niedermayer +;* +;* 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" + + +SECTION_RODATA 32 +dw1: times 8 dd 1 +w1 : times 16 dw 1 + +SECTION .text + +%macro MIX2_FLT 1 +cglobal mix_2_1_%1_float, 7, 7, 6, out, in1, in2, coeffp, index1, index2, len +%ifidn %1, a + test in1q, mmsize-1 + jne mix_2_1_float_u_int %+ SUFFIX + test in2q, mmsize-1 + jne mix_2_1_float_u_int %+ SUFFIX + test outq, mmsize-1 + jne mix_2_1_float_u_int %+ SUFFIX +%else +mix_2_1_float_u_int %+ SUFFIX: +%endif + VBROADCASTSS m4, [coeffpq + 4*index1q] + VBROADCASTSS m5, [coeffpq + 4*index2q] + shl lend , 2 + add in1q , lenq + add in2q , lenq + add outq , lenq + neg lenq +.next: +%ifidn %1, a + mulps m0, m4, [in1q + lenq ] + mulps m1, m5, [in2q + lenq ] + mulps m2, m4, [in1q + lenq + mmsize] + mulps m3, m5, [in2q + lenq + mmsize] +%else + movu m0, [in1q + lenq ] + movu m1, [in2q + lenq ] + movu m2, [in1q + lenq + mmsize] + movu m3, [in2q + lenq + mmsize] + mulps m0, m0, m4 + mulps m1, m1, m5 + mulps m2, m2, m4 + mulps m3, m3, m5 +%endif + addps m0, m0, m1 + addps m2, m2, m3 + mov%1 [outq + lenq ], m0 + mov%1 [outq + lenq + mmsize], m2 + add lenq, mmsize*2 + jl .next + REP_RET +%endmacro + +%macro MIX1_FLT 1 +cglobal mix_1_1_%1_float, 5, 5, 3, out, in, coeffp, index, len +%ifidn %1, a + test inq, mmsize-1 + jne mix_1_1_float_u_int %+ SUFFIX + test outq, mmsize-1 + jne mix_1_1_float_u_int %+ SUFFIX +%else +mix_1_1_float_u_int %+ SUFFIX: +%endif + VBROADCASTSS m2, [coeffpq + 4*indexq] + shl lenq , 2 + add inq , lenq + add outq , lenq + neg lenq +.next: +%ifidn %1, a + mulps m0, m2, [inq + lenq ] + mulps m1, m2, [inq + lenq + mmsize] +%else + movu m0, [inq + lenq ] + movu m1, [inq + lenq + mmsize] + mulps m0, m0, m2 + mulps m1, m1, m2 +%endif + mov%1 [outq + lenq ], m0 + mov%1 [outq + lenq + mmsize], m1 + add lenq, mmsize*2 + jl .next + REP_RET +%endmacro + +%macro MIX1_INT16 1 +cglobal mix_1_1_%1_int16, 5, 5, 6, out, in, coeffp, index, len +%ifidn %1, a + test inq, mmsize-1 + jne mix_1_1_int16_u_int %+ SUFFIX + test outq, mmsize-1 + jne mix_1_1_int16_u_int %+ SUFFIX +%else +mix_1_1_int16_u_int %+ SUFFIX: +%endif + movd m4, [coeffpq + 4*indexq] + SPLATW m5, m4 + psllq m4, 32 + psrlq m4, 48 + mova m0, [w1] + psllw m0, m4 + psrlw m0, 1 + punpcklwd m5, m0 + add lenq , lenq + add inq , lenq + add outq , lenq + neg lenq +.next: + mov%1 m0, [inq + lenq ] + mov%1 m2, [inq + lenq + mmsize] + mova m1, m0 + mova m3, m2 + punpcklwd m0, [w1] + punpckhwd m1, [w1] + punpcklwd m2, [w1] + punpckhwd m3, [w1] + pmaddwd m0, m5 + pmaddwd m1, m5 + pmaddwd m2, m5 + pmaddwd m3, m5 + psrad m0, m4 + psrad m1, m4 + psrad m2, m4 + psrad m3, m4 + packssdw m0, m1 + packssdw m2, m3 + mov%1 [outq + lenq ], m0 + mov%1 [outq + lenq + mmsize], m2 + add lenq, mmsize*2 + jl .next +%if mmsize == 8 + emms + RET +%else + REP_RET +%endif +%endmacro + +%macro MIX2_INT16 1 +cglobal mix_2_1_%1_int16, 7, 7, 8, out, in1, in2, coeffp, index1, index2, len +%ifidn %1, a + test in1q, mmsize-1 + jne mix_2_1_int16_u_int %+ SUFFIX + test in2q, mmsize-1 + jne mix_2_1_int16_u_int %+ SUFFIX + test outq, mmsize-1 + jne mix_2_1_int16_u_int %+ SUFFIX +%else +mix_2_1_int16_u_int %+ SUFFIX: +%endif + movd m4, [coeffpq + 4*index1q] + movd m6, [coeffpq + 4*index2q] + SPLATW m5, m4 + SPLATW m6, m6 + psllq m4, 32 + psrlq m4, 48 + mova m7, [dw1] + pslld m7, m4 + psrld m7, 1 + punpcklwd m5, m6 + add lend , lend + add in1q , lenq + add in2q , lenq + add outq , lenq + neg lenq +.next: + mov%1 m0, [in1q + lenq ] + mov%1 m2, [in2q + lenq ] + mova m1, m0 + punpcklwd m0, m2 + punpckhwd m1, m2 + + mov%1 m2, [in1q + lenq + mmsize] + mov%1 m6, [in2q + lenq + mmsize] + mova m3, m2 + punpcklwd m2, m6 + punpckhwd m3, m6 + + pmaddwd m0, m5 + pmaddwd m1, m5 + pmaddwd m2, m5 + pmaddwd m3, m5 + paddd m0, m7 + paddd m1, m7 + paddd m2, m7 + paddd m3, m7 + psrad m0, m4 + psrad m1, m4 + psrad m2, m4 + psrad m3, m4 + packssdw m0, m1 + packssdw m2, m3 + mov%1 [outq + lenq ], m0 + mov%1 [outq + lenq + mmsize], m2 + add lenq, mmsize*2 + jl .next +%if mmsize == 8 + emms + RET +%else + REP_RET +%endif +%endmacro + + +INIT_MMX mmx +MIX1_INT16 u +MIX1_INT16 a +MIX2_INT16 u +MIX2_INT16 a + +INIT_XMM sse +MIX2_FLT u +MIX2_FLT a +MIX1_FLT u +MIX1_FLT a + +INIT_XMM sse2 +MIX1_INT16 u +MIX1_INT16 a +MIX2_INT16 u +MIX2_INT16 a + +%if HAVE_AVX_EXTERNAL +INIT_YMM avx +MIX2_FLT u +MIX2_FLT a +MIX1_FLT u +MIX1_FLT a +%endif diff --git a/libswresample/x86/rematrix_init.c b/libswresample/x86/rematrix_init.c new file mode 100644 index 0000000000..5f2c5fe170 --- /dev/null +++ b/libswresample/x86/rematrix_init.c @@ -0,0 +1,90 @@ +/* + * Copyright (C) 2012 Michael Niedermayer (michaelni@gmx.at) + * + * This file is part of libswresample + * + * libswresample 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. + * + * libswresample 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 libswresample; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include "libavutil/x86/cpu.h" +#include "libswresample/swresample_internal.h" + +#define D(type, simd) \ +mix_1_1_func_type ff_mix_1_1_a_## type ## _ ## simd;\ +mix_2_1_func_type ff_mix_2_1_a_## type ## _ ## simd; + +D(float, sse) +D(float, avx) +D(int16, mmx) +D(int16, sse2) + +av_cold int swri_rematrix_init_x86(struct SwrContext *s){ +#if HAVE_YASM + int mm_flags = av_get_cpu_flags(); + int nb_in = av_get_channel_layout_nb_channels(s->in_ch_layout); + int nb_out = av_get_channel_layout_nb_channels(s->out_ch_layout); + int num = nb_in * nb_out; + int i,j; + + s->mix_1_1_simd = NULL; + s->mix_2_1_simd = NULL; + + if (s->midbuf.fmt == AV_SAMPLE_FMT_S16P){ + if(EXTERNAL_MMX(mm_flags)) { + s->mix_1_1_simd = ff_mix_1_1_a_int16_mmx; + s->mix_2_1_simd = ff_mix_2_1_a_int16_mmx; + } + if(EXTERNAL_SSE2(mm_flags)) { + s->mix_1_1_simd = ff_mix_1_1_a_int16_sse2; + s->mix_2_1_simd = ff_mix_2_1_a_int16_sse2; + } + s->native_simd_matrix = av_mallocz_array(num, 2 * sizeof(int16_t)); + s->native_simd_one = av_mallocz(2 * sizeof(int16_t)); + if (!s->native_simd_matrix || !s->native_simd_one) + return AVERROR(ENOMEM); + + for(i=0; i<nb_out; i++){ + int sh = 0; + for(j=0; j<nb_in; j++) + sh = FFMAX(sh, FFABS(((int*)s->native_matrix)[i * nb_in + j])); + sh = FFMAX(av_log2(sh) - 14, 0); + for(j=0; j<nb_in; j++) { + ((int16_t*)s->native_simd_matrix)[2*(i * nb_in + j)+1] = 15 - sh; + ((int16_t*)s->native_simd_matrix)[2*(i * nb_in + j)] = + ((((int*)s->native_matrix)[i * nb_in + j]) + (1<<sh>>1)) >> sh; + } + } + ((int16_t*)s->native_simd_one)[1] = 14; + ((int16_t*)s->native_simd_one)[0] = 16384; + } else if(s->midbuf.fmt == AV_SAMPLE_FMT_FLTP){ + if(EXTERNAL_SSE(mm_flags)) { + s->mix_1_1_simd = ff_mix_1_1_a_float_sse; + s->mix_2_1_simd = ff_mix_2_1_a_float_sse; + } + if(EXTERNAL_AVX_FAST(mm_flags)) { + s->mix_1_1_simd = ff_mix_1_1_a_float_avx; + s->mix_2_1_simd = ff_mix_2_1_a_float_avx; + } + s->native_simd_matrix = av_mallocz_array(num, sizeof(float)); + s->native_simd_one = av_mallocz(sizeof(float)); + if (!s->native_simd_matrix || !s->native_simd_one) + return AVERROR(ENOMEM); + memcpy(s->native_simd_matrix, s->native_matrix, num * sizeof(float)); + memcpy(s->native_simd_one, s->native_one, sizeof(float)); + } +#endif + + return 0; +} diff --git a/libswresample/x86/resample.asm b/libswresample/x86/resample.asm new file mode 100644 index 0000000000..4989aa6991 --- /dev/null +++ b/libswresample/x86/resample.asm @@ -0,0 +1,605 @@ +;****************************************************************************** +;* 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_shift: resd 1 + .phase_mask: 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 + +%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_shift, 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 + mov phase_maskd, [ctxq+ResampleContext.phase_mask] +%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_shift] + mov edi, [ctxq+ResampleContext.filter_alloc] + + DEFINE_ARGS filter_alloc, dst, src, phase_shift, 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_shift] + + DEFINE_ARGS phase_shift, 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_shift, 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_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_shiftd, [ctxq+ResampleContext.phase_shift] + + DEFINE_ARGS src, phase_shift, 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 + addps 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_shift, 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_shift, 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_shift, dst, frac, index, index_incr +%endif + +.skip: + mov index_incrd, indexd + add dstq, %2 + and indexd, phase_maskd + sar index_incrd, phase_shiftb + lea srcq, [srcq+index_incrq*%2] + cmp dstq, dst_endq + jne .loop + +%if ARCH_X86_64 + DEFINE_ARGS ctx, dst, src, phase_shift, 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_shift, 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_shift, 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 + mov phase_maskd, [ctxq+ResampleContext.phase_mask] +%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 phase_mask_stackd, phase_maskd + 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_shift] + mov edi, [ctxq+ResampleContext.filter_alloc] + + DEFINE_ARGS filter_alloc, dst, filter2, phase_shift, 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_shift] + + DEFINE_ARGS phase_shift, 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_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_shift] + + DEFINE_ARGS filter1, min_filter_count_x4, filter2, frac, index, dst, src + +%define phase_shift_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 + addps xm0, xm1 + addps 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_shift, 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_shift, 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_shift, index_incr, frac, index, dst, src +%endif + +.skip: +%if ARCH_X86_32 + mov phase_shiftd, phase_shift_stackd +%endif + mov index_incrd, indexd + add dstq, %2 + and indexd, phase_mask_stackd + sar index_incrd, phase_shiftb + lea srcq, [srcq+index_incrq*%2] + cmp dstq, dst_endq + jne .loop + +%if UNIX64 + DEFINE_ARGS ctx, dst, filter2, phase_shift, 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_shift, 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 diff --git a/libswresample/x86/resample_init.c b/libswresample/x86/resample_init.c new file mode 100644 index 0000000000..9d7d5cf89e --- /dev/null +++ b/libswresample/x86/resample_init.c @@ -0,0 +1,90 @@ +/* + * audio resampling + * Copyright (c) 2004-2012 Michael Niedermayer <michaelni@gmx.at> + * + * 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 + */ + +/** + * @file + * audio resampling + * @author Michael Niedermayer <michaelni@gmx.at> + */ + +#include "libavutil/x86/cpu.h" +#include "libswresample/resample.h" + +#define RESAMPLE_FUNCS(type, opt) \ +int ff_resample_common_##type##_##opt(ResampleContext *c, void *dst, \ + const void *src, int sz, int upd); \ +int ff_resample_linear_##type##_##opt(ResampleContext *c, void *dst, \ + const void *src, int sz, int upd) + +RESAMPLE_FUNCS(int16, mmxext); +RESAMPLE_FUNCS(int16, sse2); +RESAMPLE_FUNCS(int16, xop); +RESAMPLE_FUNCS(float, sse); +RESAMPLE_FUNCS(float, avx); +RESAMPLE_FUNCS(float, fma3); +RESAMPLE_FUNCS(float, fma4); +RESAMPLE_FUNCS(double, sse2); + +av_cold void swri_resample_dsp_x86_init(ResampleContext *c) +{ + int av_unused mm_flags = av_get_cpu_flags(); + + switch(c->format){ + case AV_SAMPLE_FMT_S16P: + if (ARCH_X86_32 && EXTERNAL_MMXEXT(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_int16_mmxext + : ff_resample_common_int16_mmxext; + } + if (EXTERNAL_SSE2(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_int16_sse2 + : ff_resample_common_int16_sse2; + } + if (EXTERNAL_XOP(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_int16_xop + : ff_resample_common_int16_xop; + } + break; + case AV_SAMPLE_FMT_FLTP: + if (EXTERNAL_SSE(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_float_sse + : ff_resample_common_float_sse; + } + if (EXTERNAL_AVX_FAST(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_float_avx + : ff_resample_common_float_avx; + } + if (EXTERNAL_FMA3_FAST(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_float_fma3 + : ff_resample_common_float_fma3; + } + if (EXTERNAL_FMA4(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_float_fma4 + : ff_resample_common_float_fma4; + } + break; + case AV_SAMPLE_FMT_DBLP: + if (EXTERNAL_SSE2(mm_flags)) { + c->dsp.resample = c->linear ? ff_resample_linear_double_sse2 + : ff_resample_common_double_sse2; + } + break; + } +} diff --git a/libswresample/x86/w64xmmtest.c b/libswresample/x86/w64xmmtest.c new file mode 100644 index 0000000000..9cddb4a858 --- /dev/null +++ b/libswresample/x86/w64xmmtest.c @@ -0,0 +1,29 @@ +/* + * check XMM registers for clobbers on Win64 + * Copyright (c) 2013 Martin Storsjo + * + * 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 "libswresample/swresample.h" +#include "libavutil/x86/w64xmmtest.h" + +wrap(swr_convert(struct SwrContext *s, uint8_t **out, int out_count, + const uint8_t **in , int in_count)) +{ + testxmmclobbers(swr_convert, s, out, out_count, in, in_count); +} |