yading@11: ;****************************************************************************** yading@11: ;* Copyright (c) 2012 Michael Niedermayer yading@11: ;* yading@11: ;* This file is part of FFmpeg. yading@11: ;* yading@11: ;* FFmpeg is free software; you can redistribute it and/or yading@11: ;* modify it under the terms of the GNU Lesser General Public yading@11: ;* License as published by the Free Software Foundation; either yading@11: ;* version 2.1 of the License, or (at your option) any later version. yading@11: ;* yading@11: ;* FFmpeg is distributed in the hope that it will be useful, yading@11: ;* but WITHOUT ANY WARRANTY; without even the implied warranty of yading@11: ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@11: ;* Lesser General Public License for more details. yading@11: ;* yading@11: ;* You should have received a copy of the GNU Lesser General Public yading@11: ;* License along with FFmpeg; if not, write to the Free Software yading@11: ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@11: ;****************************************************************************** yading@11: yading@11: %include "libavutil/x86/x86util.asm" yading@11: yading@11: SECTION_RODATA yading@11: align 32 yading@11: flt2pm31: times 8 dd 4.6566129e-10 yading@11: flt2p31 : times 8 dd 2147483648.0 yading@11: flt2p15 : times 8 dd 32768.0 yading@11: yading@11: word_unpack_shuf : db 0, 1, 4, 5, 8, 9,12,13, 2, 3, 6, 7,10,11,14,15 yading@11: yading@11: SECTION .text yading@11: yading@11: yading@11: ;to, from, a/u, log2_outsize, log_intsize, const yading@11: %macro PACK_2CH 5-7 yading@11: cglobal pack_2ch_%2_to_%1_%3, 3, 4, 6, dst, src, len, src2 yading@11: mov src2q , [srcq+gprsize] yading@11: mov srcq , [srcq] yading@11: mov dstq , [dstq] yading@11: %ifidn %3, a yading@11: test dstq, mmsize-1 yading@11: jne pack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: test srcq, mmsize-1 yading@11: jne pack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: test src2q, mmsize-1 yading@11: jne pack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: %else yading@11: pack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: %endif yading@11: lea srcq , [srcq + (1<<%5)*lenq] yading@11: lea src2q, [src2q + (1<<%5)*lenq] yading@11: lea dstq , [dstq + (2<<%4)*lenq] yading@11: neg lenq yading@11: %7 m0,m1,m2,m3,m4,m5 yading@11: .next: yading@11: %if %4 >= %5 yading@11: mov%3 m0, [ srcq +(1<<%5)*lenq] yading@11: mova m1, m0 yading@11: mov%3 m2, [ src2q+(1<<%5)*lenq] yading@11: %if %5 == 1 yading@11: punpcklwd m0, m2 yading@11: punpckhwd m1, m2 yading@11: %else yading@11: punpckldq m0, m2 yading@11: punpckhdq m1, m2 yading@11: %endif yading@11: %6 m0,m1,m2,m3,m4,m5 yading@11: %else yading@11: mov%3 m0, [ srcq +(1<<%5)*lenq] yading@11: mov%3 m1, [mmsize + srcq +(1<<%5)*lenq] yading@11: mov%3 m2, [ src2q+(1<<%5)*lenq] yading@11: mov%3 m3, [mmsize + src2q+(1<<%5)*lenq] yading@11: %6 m0,m1,m2,m3,m4,m5 yading@11: mova m2, m0 yading@11: punpcklwd m0, m1 yading@11: punpckhwd m2, m1 yading@11: SWAP 1,2 yading@11: %endif yading@11: mov%3 [ dstq+(2<<%4)*lenq], m0 yading@11: mov%3 [ mmsize + dstq+(2<<%4)*lenq], m1 yading@11: %if %4 > %5 yading@11: mov%3 [2*mmsize + dstq+(2<<%4)*lenq], m2 yading@11: mov%3 [3*mmsize + dstq+(2<<%4)*lenq], m3 yading@11: add lenq, 4*mmsize/(2<<%4) yading@11: %else yading@11: add lenq, 2*mmsize/(2<<%4) yading@11: %endif yading@11: jl .next yading@11: REP_RET yading@11: %endmacro yading@11: yading@11: %macro UNPACK_2CH 5-7 yading@11: cglobal unpack_2ch_%2_to_%1_%3, 3, 4, 7, dst, src, len, dst2 yading@11: mov dst2q , [dstq+gprsize] yading@11: mov srcq , [srcq] yading@11: mov dstq , [dstq] yading@11: %ifidn %3, a yading@11: test dstq, mmsize-1 yading@11: jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: test srcq, mmsize-1 yading@11: jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: test dst2q, mmsize-1 yading@11: jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: %else yading@11: unpack_2ch_%2_to_%1_u_int %+ SUFFIX yading@11: %endif yading@11: lea srcq , [srcq + (2<<%5)*lenq] yading@11: lea dstq , [dstq + (1<<%4)*lenq] yading@11: lea dst2q, [dst2q + (1<<%4)*lenq] yading@11: neg lenq yading@11: %7 m0,m1,m2,m3,m4,m5 yading@11: mova m6, [word_unpack_shuf] yading@11: .next: yading@11: mov%3 m0, [ srcq +(2<<%5)*lenq] yading@11: mov%3 m2, [ mmsize + srcq +(2<<%5)*lenq] yading@11: %if %5 == 1 yading@11: %ifidn SUFFIX, _ssse3 yading@11: pshufb m0, m6 yading@11: mova m1, m0 yading@11: pshufb m2, m6 yading@11: punpcklqdq m0,m2 yading@11: punpckhqdq m1,m2 yading@11: %else yading@11: mova m1, m0 yading@11: punpcklwd m0,m2 yading@11: punpckhwd m1,m2 yading@11: yading@11: mova m2, m0 yading@11: punpcklwd m0,m1 yading@11: punpckhwd m2,m1 yading@11: yading@11: mova m1, m0 yading@11: punpcklwd m0,m2 yading@11: punpckhwd m1,m2 yading@11: %endif yading@11: %else yading@11: mova m1, m0 yading@11: shufps m0, m2, 10001000b yading@11: shufps m1, m2, 11011101b yading@11: %endif yading@11: %if %4 < %5 yading@11: mov%3 m2, [2*mmsize + srcq +(2<<%5)*lenq] yading@11: mova m3, m2 yading@11: mov%3 m4, [3*mmsize + srcq +(2<<%5)*lenq] yading@11: shufps m2, m4, 10001000b yading@11: shufps m3, m4, 11011101b yading@11: SWAP 1,2 yading@11: %endif yading@11: %6 m0,m1,m2,m3,m4,m5 yading@11: mov%3 [ dstq+(1<<%4)*lenq], m0 yading@11: %if %4 > %5 yading@11: mov%3 [ dst2q+(1<<%4)*lenq], m2 yading@11: mov%3 [ mmsize + dstq+(1<<%4)*lenq], m1 yading@11: mov%3 [ mmsize + dst2q+(1<<%4)*lenq], m3 yading@11: add lenq, 2*mmsize/(1<<%4) yading@11: %else yading@11: mov%3 [ dst2q+(1<<%4)*lenq], m1 yading@11: add lenq, mmsize/(1<<%4) yading@11: %endif yading@11: jl .next yading@11: REP_RET yading@11: %endmacro yading@11: yading@11: %macro CONV 5-7 yading@11: cglobal %2_to_%1_%3, 3, 3, 6, dst, src, len yading@11: mov srcq , [srcq] yading@11: mov dstq , [dstq] yading@11: %ifidn %3, a yading@11: test dstq, mmsize-1 yading@11: jne %2_to_%1_u_int %+ SUFFIX yading@11: test srcq, mmsize-1 yading@11: jne %2_to_%1_u_int %+ SUFFIX yading@11: %else yading@11: %2_to_%1_u_int %+ SUFFIX yading@11: %endif yading@11: lea srcq , [srcq + (1<<%5)*lenq] yading@11: lea dstq , [dstq + (1<<%4)*lenq] yading@11: neg lenq yading@11: %7 m0,m1,m2,m3,m4,m5 yading@11: .next: yading@11: mov%3 m0, [ srcq +(1<<%5)*lenq] yading@11: mov%3 m1, [ mmsize + srcq +(1<<%5)*lenq] yading@11: %if %4 < %5 yading@11: mov%3 m2, [2*mmsize + srcq +(1<<%5)*lenq] yading@11: mov%3 m3, [3*mmsize + srcq +(1<<%5)*lenq] yading@11: %endif yading@11: %6 m0,m1,m2,m3,m4,m5 yading@11: mov%3 [ dstq+(1<<%4)*lenq], m0 yading@11: mov%3 [ mmsize + dstq+(1<<%4)*lenq], m1 yading@11: %if %4 > %5 yading@11: mov%3 [2*mmsize + dstq+(1<<%4)*lenq], m2 yading@11: mov%3 [3*mmsize + dstq+(1<<%4)*lenq], m3 yading@11: add lenq, 4*mmsize/(1<<%4) yading@11: %else yading@11: add lenq, 2*mmsize/(1<<%4) yading@11: %endif yading@11: jl .next yading@11: REP_RET yading@11: %endmacro yading@11: yading@11: %macro PACK_6CH 5-7 yading@11: cglobal pack_6ch_%2_to_%1_%3, 2,8,7, dst, src, src1, src2, src3, src4, src5, len yading@11: %if ARCH_X86_64 yading@11: mov lend, r2d yading@11: %else yading@11: %define lend dword r2m yading@11: %endif yading@11: mov src1q, [srcq+1*gprsize] yading@11: mov src2q, [srcq+2*gprsize] yading@11: mov src3q, [srcq+3*gprsize] yading@11: mov src4q, [srcq+4*gprsize] yading@11: mov src5q, [srcq+5*gprsize] yading@11: mov srcq, [srcq] yading@11: mov dstq, [dstq] yading@11: %ifidn %3, a yading@11: test dstq, mmsize-1 yading@11: jne pack_6ch_%2_to_%1_u_int %+ SUFFIX yading@11: test srcq, mmsize-1 yading@11: jne pack_6ch_%2_to_%1_u_int %+ SUFFIX yading@11: test src2q, mmsize-1 yading@11: jne pack_6ch_%2_to_%1_u_int %+ SUFFIX yading@11: test src3q, mmsize-1 yading@11: jne pack_6ch_%2_to_%1_u_int %+ SUFFIX yading@11: test src4q, mmsize-1 yading@11: jne pack_6ch_%2_to_%1_u_int %+ SUFFIX yading@11: test src5q, mmsize-1 yading@11: jne pack_6ch_%2_to_%1_u_int %+ SUFFIX yading@11: %else yading@11: pack_6ch_%2_to_%1_u_int %+ SUFFIX yading@11: %endif yading@11: sub src1q, srcq yading@11: sub src2q, srcq yading@11: sub src3q, srcq yading@11: sub src4q, srcq yading@11: sub src5q, srcq yading@11: .loop: yading@11: mov%3 m0, [srcq ] yading@11: mov%3 m1, [srcq+src1q] yading@11: mov%3 m2, [srcq+src2q] yading@11: mov%3 m3, [srcq+src3q] yading@11: mov%3 m4, [srcq+src4q] yading@11: mov%3 m5, [srcq+src5q] yading@11: %7 x,x,x,x,m7,x yading@11: %if cpuflag(sse4) yading@11: SBUTTERFLYPS 0, 1, 6 yading@11: SBUTTERFLYPS 2, 3, 6 yading@11: SBUTTERFLYPS 4, 5, 6 yading@11: yading@11: blendps m6, m4, m0, 1100b yading@11: movlhps m0, m2 yading@11: movhlps m4, m2 yading@11: blendps m2, m5, m1, 1100b yading@11: movlhps m1, m3 yading@11: movhlps m5, m3 yading@11: yading@11: %6 m0,m6,x,x,m7,m3 yading@11: %6 m4,m1,x,x,m7,m3 yading@11: %6 m2,m5,x,x,m7,m3 yading@11: yading@11: mov %+ %3 %+ ps [dstq ], m0 yading@11: mov %+ %3 %+ ps [dstq+16], m6 yading@11: mov %+ %3 %+ ps [dstq+32], m4 yading@11: mov %+ %3 %+ ps [dstq+48], m1 yading@11: mov %+ %3 %+ ps [dstq+64], m2 yading@11: mov %+ %3 %+ ps [dstq+80], m5 yading@11: %else ; mmx yading@11: SBUTTERFLY dq, 0, 1, 6 yading@11: SBUTTERFLY dq, 2, 3, 6 yading@11: SBUTTERFLY dq, 4, 5, 6 yading@11: yading@11: movq [dstq ], m0 yading@11: movq [dstq+ 8], m2 yading@11: movq [dstq+16], m4 yading@11: movq [dstq+24], m1 yading@11: movq [dstq+32], m3 yading@11: movq [dstq+40], m5 yading@11: %endif yading@11: add srcq, mmsize yading@11: add dstq, mmsize*6 yading@11: sub lend, mmsize/4 yading@11: jg .loop yading@11: %if mmsize == 8 yading@11: emms yading@11: RET yading@11: %else yading@11: REP_RET yading@11: %endif yading@11: %endmacro yading@11: yading@11: %macro INT16_TO_INT32_N 6 yading@11: pxor m2, m2 yading@11: pxor m3, m3 yading@11: punpcklwd m2, m1 yading@11: punpckhwd m3, m1 yading@11: SWAP 4,0 yading@11: pxor m0, m0 yading@11: pxor m1, m1 yading@11: punpcklwd m0, m4 yading@11: punpckhwd m1, m4 yading@11: %endmacro yading@11: yading@11: %macro INT32_TO_INT16_N 6 yading@11: psrad m0, 16 yading@11: psrad m1, 16 yading@11: psrad m2, 16 yading@11: psrad m3, 16 yading@11: packssdw m0, m1 yading@11: packssdw m2, m3 yading@11: SWAP 1,2 yading@11: %endmacro yading@11: yading@11: %macro INT32_TO_FLOAT_INIT 6 yading@11: mova %5, [flt2pm31] yading@11: %endmacro yading@11: %macro INT32_TO_FLOAT_N 6 yading@11: cvtdq2ps %1, %1 yading@11: cvtdq2ps %2, %2 yading@11: mulps %1, %1, %5 yading@11: mulps %2, %2, %5 yading@11: %endmacro yading@11: yading@11: %macro FLOAT_TO_INT32_INIT 6 yading@11: mova %5, [flt2p31] yading@11: %endmacro yading@11: %macro FLOAT_TO_INT32_N 6 yading@11: mulps %1, %5 yading@11: mulps %2, %5 yading@11: cvtps2dq %6, %1 yading@11: cmpnltps %1, %5 yading@11: paddd %1, %6 yading@11: cvtps2dq %6, %2 yading@11: cmpnltps %2, %5 yading@11: paddd %2, %6 yading@11: %endmacro yading@11: yading@11: %macro INT16_TO_FLOAT_INIT 6 yading@11: mova m5, [flt2pm31] yading@11: %endmacro yading@11: %macro INT16_TO_FLOAT_N 6 yading@11: INT16_TO_INT32_N %1,%2,%3,%4,%5,%6 yading@11: cvtdq2ps m0, m0 yading@11: cvtdq2ps m1, m1 yading@11: cvtdq2ps m2, m2 yading@11: cvtdq2ps m3, m3 yading@11: mulps m0, m0, m5 yading@11: mulps m1, m1, m5 yading@11: mulps m2, m2, m5 yading@11: mulps m3, m3, m5 yading@11: %endmacro yading@11: yading@11: %macro FLOAT_TO_INT16_INIT 6 yading@11: mova m5, [flt2p15] yading@11: %endmacro yading@11: %macro FLOAT_TO_INT16_N 6 yading@11: mulps m0, m5 yading@11: mulps m1, m5 yading@11: mulps m2, m5 yading@11: mulps m3, m5 yading@11: cvtps2dq m0, m0 yading@11: cvtps2dq m1, m1 yading@11: packssdw m0, m1 yading@11: cvtps2dq m1, m2 yading@11: cvtps2dq m3, m3 yading@11: packssdw m1, m3 yading@11: %endmacro yading@11: yading@11: %macro NOP_N 0-6 yading@11: %endmacro yading@11: yading@11: INIT_MMX mmx yading@11: CONV int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: CONV int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: CONV int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: CONV int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: yading@11: PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N yading@11: PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N yading@11: yading@11: INIT_XMM sse2 yading@11: CONV int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: CONV int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: CONV int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: CONV int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: yading@11: PACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N yading@11: PACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N yading@11: PACK_2CH int32, int32, u, 2, 2, NOP_N, NOP_N yading@11: PACK_2CH int32, int32, a, 2, 2, NOP_N, NOP_N yading@11: PACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: PACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: PACK_2CH int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: PACK_2CH int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: yading@11: UNPACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N yading@11: UNPACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N yading@11: UNPACK_2CH int32, int32, u, 2, 2, NOP_N, NOP_N yading@11: UNPACK_2CH int32, int32, a, 2, 2, NOP_N, NOP_N yading@11: UNPACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: UNPACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: UNPACK_2CH int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: UNPACK_2CH int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N yading@11: yading@11: CONV float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: CONV float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: CONV int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: CONV int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: CONV float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: CONV float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: CONV int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT yading@11: CONV int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT yading@11: yading@11: PACK_2CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: PACK_2CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: PACK_2CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: PACK_2CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: PACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: PACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: PACK_2CH int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT yading@11: PACK_2CH int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT yading@11: yading@11: UNPACK_2CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: UNPACK_2CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: UNPACK_2CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: UNPACK_2CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: UNPACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: UNPACK_2CH int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT yading@11: UNPACK_2CH int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT yading@11: yading@11: yading@11: INIT_XMM ssse3 yading@11: UNPACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N yading@11: UNPACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N yading@11: UNPACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: UNPACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N yading@11: UNPACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT yading@11: yading@11: INIT_XMM sse4 yading@11: PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N yading@11: PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N yading@11: yading@11: PACK_6CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: PACK_6CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: PACK_6CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: PACK_6CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: yading@11: %if HAVE_AVX_EXTERNAL yading@11: INIT_XMM avx yading@11: PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N yading@11: PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N yading@11: yading@11: PACK_6CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: PACK_6CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: PACK_6CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: PACK_6CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT yading@11: yading@11: INIT_YMM avx yading@11: CONV float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: CONV float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT yading@11: %endif