yading@10: ;***************************************************************************** yading@10: ;* MMX/SSE2/AVX-optimized H.264 deblocking code yading@10: ;***************************************************************************** yading@10: ;* Copyright (C) 2005-2011 x264 project yading@10: ;* yading@10: ;* Authors: Loren Merritt yading@10: ;* Jason Garrett-Glaser yading@10: ;* Oskar Arvidsson yading@10: ;* yading@10: ;* This file is part of FFmpeg. yading@10: ;* yading@10: ;* FFmpeg is free software; you can redistribute it and/or yading@10: ;* modify it under the terms of the GNU Lesser General Public yading@10: ;* License as published by the Free Software Foundation; either yading@10: ;* version 2.1 of the License, or (at your option) any later version. yading@10: ;* yading@10: ;* FFmpeg is distributed in the hope that it will be useful, yading@10: ;* but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: ;* Lesser General Public License for more details. yading@10: ;* yading@10: ;* You should have received a copy of the GNU Lesser General Public yading@10: ;* License along with FFmpeg; if not, write to the Free Software yading@10: ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: ;****************************************************************************** yading@10: yading@10: %include "libavutil/x86/x86util.asm" yading@10: yading@10: SECTION_RODATA yading@10: yading@10: pb_A1: times 16 db 0xA1 yading@10: pb_3_1: times 4 db 3, 1 yading@10: yading@10: SECTION .text yading@10: yading@10: cextern pb_0 yading@10: cextern pb_1 yading@10: cextern pb_3 yading@10: yading@10: ; expands to [base],...,[base+7*stride] yading@10: %define PASS8ROWS(base, base3, stride, stride3) \ yading@10: [base], [base+stride], [base+stride*2], [base3], \ yading@10: [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4] yading@10: yading@10: %define PASS8ROWS(base, base3, stride, stride3, offset) \ yading@10: PASS8ROWS(base+offset, base3+offset, stride, stride3) yading@10: yading@10: ; in: 8 rows of 4 bytes in %4..%11 yading@10: ; out: 4 rows of 8 bytes in m0..m3 yading@10: %macro TRANSPOSE4x8_LOAD 11 yading@10: movh m0, %4 yading@10: movh m2, %5 yading@10: movh m1, %6 yading@10: movh m3, %7 yading@10: punpckl%1 m0, m2 yading@10: punpckl%1 m1, m3 yading@10: mova m2, m0 yading@10: punpckl%2 m0, m1 yading@10: punpckh%2 m2, m1 yading@10: yading@10: movh m4, %8 yading@10: movh m6, %9 yading@10: movh m5, %10 yading@10: movh m7, %11 yading@10: punpckl%1 m4, m6 yading@10: punpckl%1 m5, m7 yading@10: mova m6, m4 yading@10: punpckl%2 m4, m5 yading@10: punpckh%2 m6, m5 yading@10: yading@10: punpckh%3 m1, m0, m4 yading@10: punpckh%3 m3, m2, m6 yading@10: punpckl%3 m0, m4 yading@10: punpckl%3 m2, m6 yading@10: %endmacro yading@10: yading@10: ; in: 4 rows of 8 bytes in m0..m3 yading@10: ; out: 8 rows of 4 bytes in %1..%8 yading@10: %macro TRANSPOSE8x4B_STORE 8 yading@10: punpckhdq m4, m0, m0 yading@10: punpckhdq m5, m1, m1 yading@10: punpckhdq m6, m2, m2 yading@10: yading@10: punpcklbw m0, m1 yading@10: punpcklbw m2, m3 yading@10: punpcklwd m1, m0, m2 yading@10: punpckhwd m0, m2 yading@10: movh %1, m1 yading@10: punpckhdq m1, m1 yading@10: movh %2, m1 yading@10: movh %3, m0 yading@10: punpckhdq m0, m0 yading@10: movh %4, m0 yading@10: yading@10: punpckhdq m3, m3 yading@10: punpcklbw m4, m5 yading@10: punpcklbw m6, m3 yading@10: punpcklwd m5, m4, m6 yading@10: punpckhwd m4, m6 yading@10: movh %5, m5 yading@10: punpckhdq m5, m5 yading@10: movh %6, m5 yading@10: movh %7, m4 yading@10: punpckhdq m4, m4 yading@10: movh %8, m4 yading@10: %endmacro yading@10: yading@10: %macro TRANSPOSE4x8B_LOAD 8 yading@10: TRANSPOSE4x8_LOAD bw, wd, dq, %1, %2, %3, %4, %5, %6, %7, %8 yading@10: %endmacro yading@10: yading@10: %macro SBUTTERFLY3 4 yading@10: punpckh%1 %4, %2, %3 yading@10: punpckl%1 %2, %3 yading@10: %endmacro yading@10: yading@10: ; in: 8 rows of 8 (only the middle 6 pels are used) in %1..%8 yading@10: ; out: 6 rows of 8 in [%9+0*16] .. [%9+5*16] yading@10: %macro TRANSPOSE6x8_MEM 9 yading@10: RESET_MM_PERMUTATION yading@10: movq m0, %1 yading@10: movq m1, %2 yading@10: movq m2, %3 yading@10: movq m3, %4 yading@10: movq m4, %5 yading@10: movq m5, %6 yading@10: movq m6, %7 yading@10: SBUTTERFLY bw, 0, 1, 7 yading@10: SBUTTERFLY bw, 2, 3, 7 yading@10: SBUTTERFLY bw, 4, 5, 7 yading@10: movq [%9+0x10], m3 yading@10: SBUTTERFLY3 bw, m6, %8, m7 yading@10: SBUTTERFLY wd, 0, 2, 3 yading@10: SBUTTERFLY wd, 4, 6, 3 yading@10: punpckhdq m0, m4 yading@10: movq [%9+0x00], m0 yading@10: SBUTTERFLY3 wd, m1, [%9+0x10], m3 yading@10: SBUTTERFLY wd, 5, 7, 0 yading@10: SBUTTERFLY dq, 1, 5, 0 yading@10: SBUTTERFLY dq, 2, 6, 0 yading@10: punpckldq m3, m7 yading@10: movq [%9+0x10], m2 yading@10: movq [%9+0x20], m6 yading@10: movq [%9+0x30], m1 yading@10: movq [%9+0x40], m5 yading@10: movq [%9+0x50], m3 yading@10: RESET_MM_PERMUTATION yading@10: %endmacro yading@10: yading@10: ; in: 8 rows of 8 in %1..%8 yading@10: ; out: 8 rows of 8 in %9..%16 yading@10: %macro TRANSPOSE8x8_MEM 16 yading@10: RESET_MM_PERMUTATION yading@10: movq m0, %1 yading@10: movq m1, %2 yading@10: movq m2, %3 yading@10: movq m3, %4 yading@10: movq m4, %5 yading@10: movq m5, %6 yading@10: movq m6, %7 yading@10: SBUTTERFLY bw, 0, 1, 7 yading@10: SBUTTERFLY bw, 2, 3, 7 yading@10: SBUTTERFLY bw, 4, 5, 7 yading@10: SBUTTERFLY3 bw, m6, %8, m7 yading@10: movq %9, m5 yading@10: SBUTTERFLY wd, 0, 2, 5 yading@10: SBUTTERFLY wd, 4, 6, 5 yading@10: SBUTTERFLY wd, 1, 3, 5 yading@10: movq %11, m6 yading@10: movq m6, %9 yading@10: SBUTTERFLY wd, 6, 7, 5 yading@10: SBUTTERFLY dq, 0, 4, 5 yading@10: SBUTTERFLY dq, 1, 6, 5 yading@10: movq %9, m0 yading@10: movq %10, m4 yading@10: movq %13, m1 yading@10: movq %14, m6 yading@10: SBUTTERFLY3 dq, m2, %11, m0 yading@10: SBUTTERFLY dq, 3, 7, 4 yading@10: movq %11, m2 yading@10: movq %12, m0 yading@10: movq %15, m3 yading@10: movq %16, m7 yading@10: RESET_MM_PERMUTATION yading@10: %endmacro yading@10: yading@10: ; out: %4 = |%1-%2|>%3 yading@10: ; clobbers: %5 yading@10: %macro DIFF_GT 5 yading@10: %if avx_enabled == 0 yading@10: mova %5, %2 yading@10: mova %4, %1 yading@10: psubusb %5, %1 yading@10: psubusb %4, %2 yading@10: %else yading@10: psubusb %5, %2, %1 yading@10: psubusb %4, %1, %2 yading@10: %endif yading@10: por %4, %5 yading@10: psubusb %4, %3 yading@10: %endmacro yading@10: yading@10: ; out: %4 = |%1-%2|>%3 yading@10: ; clobbers: %5 yading@10: %macro DIFF_GT2 5 yading@10: %if ARCH_X86_64 yading@10: psubusb %5, %2, %1 yading@10: psubusb %4, %1, %2 yading@10: %else yading@10: mova %5, %2 yading@10: mova %4, %1 yading@10: psubusb %5, %1 yading@10: psubusb %4, %2 yading@10: %endif yading@10: psubusb %5, %3 yading@10: psubusb %4, %3 yading@10: pcmpeqb %4, %5 yading@10: %endmacro yading@10: yading@10: ; in: m0=p1 m1=p0 m2=q0 m3=q1 %1=alpha-1 %2=beta-1 yading@10: ; out: m5=beta-1, m7=mask, %3=alpha-1 yading@10: ; clobbers: m4,m6 yading@10: %macro LOAD_MASK 2-3 yading@10: movd m4, %1 yading@10: movd m5, %2 yading@10: SPLATW m4, m4 yading@10: SPLATW m5, m5 yading@10: packuswb m4, m4 ; 16x alpha-1 yading@10: packuswb m5, m5 ; 16x beta-1 yading@10: %if %0>2 yading@10: mova %3, m4 yading@10: %endif yading@10: DIFF_GT m1, m2, m4, m7, m6 ; |p0-q0| > alpha-1 yading@10: DIFF_GT m0, m1, m5, m4, m6 ; |p1-p0| > beta-1 yading@10: por m7, m4 yading@10: DIFF_GT m3, m2, m5, m4, m6 ; |q1-q0| > beta-1 yading@10: por m7, m4 yading@10: pxor m6, m6 yading@10: pcmpeqb m7, m6 yading@10: %endmacro yading@10: yading@10: ; in: m0=p1 m1=p0 m2=q0 m3=q1 m7=(tc&mask) yading@10: ; out: m1=p0' m2=q0' yading@10: ; clobbers: m0,3-6 yading@10: %macro DEBLOCK_P0_Q0 0 yading@10: pcmpeqb m4, m4 yading@10: pxor m5, m1, m2 ; p0^q0 yading@10: pxor m3, m4 yading@10: pand m5, [pb_1] ; (p0^q0)&1 yading@10: pavgb m3, m0 ; (p1 - q1 + 256)>>1 yading@10: pxor m4, m1 yading@10: pavgb m3, [pb_3] ; (((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2 yading@10: pavgb m4, m2 ; (q0 - p0 + 256)>>1 yading@10: pavgb m3, m5 yading@10: mova m6, [pb_A1] yading@10: paddusb m3, m4 ; d+128+33 yading@10: psubusb m6, m3 yading@10: psubusb m3, [pb_A1] yading@10: pminub m6, m7 yading@10: pminub m3, m7 yading@10: psubusb m1, m6 yading@10: psubusb m2, m3 yading@10: paddusb m1, m3 yading@10: paddusb m2, m6 yading@10: %endmacro yading@10: yading@10: ; in: m1=p0 m2=q0 yading@10: ; %1=p1 %2=q2 %3=[q2] %4=[q1] %5=tc0 %6=tmp yading@10: ; out: [q1] = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 ) yading@10: ; clobbers: q2, tmp, tc0 yading@10: %macro LUMA_Q1 6 yading@10: pavgb %6, m1, m2 yading@10: pavgb %2, %6 ; avg(p2,avg(p0,q0)) yading@10: pxor %6, %3 yading@10: pand %6, [pb_1] ; (p2^avg(p0,q0))&1 yading@10: psubusb %2, %6 ; (p2+((p0+q0+1)>>1))>>1 yading@10: psubusb %6, %1, %5 yading@10: paddusb %5, %1 yading@10: pmaxub %2, %6 yading@10: pminub %2, %5 yading@10: mova %4, %2 yading@10: %endmacro yading@10: yading@10: %if ARCH_X86_64 yading@10: ;----------------------------------------------------------------------------- yading@10: ; void deblock_v_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro DEBLOCK_LUMA 0 yading@10: cglobal deblock_v_luma_8, 5,5,10 yading@10: movd m8, [r4] ; tc0 yading@10: lea r4, [r1*3] yading@10: dec r2d ; alpha-1 yading@10: neg r4 yading@10: dec r3d ; beta-1 yading@10: add r4, r0 ; pix-3*stride yading@10: yading@10: mova m0, [r4+r1] ; p1 yading@10: mova m1, [r4+2*r1] ; p0 yading@10: mova m2, [r0] ; q0 yading@10: mova m3, [r0+r1] ; q1 yading@10: LOAD_MASK r2d, r3d yading@10: yading@10: punpcklbw m8, m8 yading@10: punpcklbw m8, m8 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0] yading@10: pcmpeqb m9, m9 yading@10: pcmpeqb m9, m8 yading@10: pandn m9, m7 yading@10: pand m8, m9 yading@10: yading@10: movdqa m3, [r4] ; p2 yading@10: DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1 yading@10: pand m6, m9 yading@10: psubb m7, m8, m6 yading@10: pand m6, m8 yading@10: LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4 yading@10: yading@10: movdqa m4, [r0+2*r1] ; q2 yading@10: DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1 yading@10: pand m6, m9 yading@10: pand m8, m6 yading@10: psubb m7, m6 yading@10: mova m3, [r0+r1] yading@10: LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m8, m6 yading@10: yading@10: DEBLOCK_P0_Q0 yading@10: mova [r4+2*r1], m1 yading@10: mova [r0], m2 yading@10: RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) yading@10: ;----------------------------------------------------------------------------- yading@10: INIT_MMX cpuname yading@10: cglobal deblock_h_luma_8, 5,9 yading@10: movsxd r7, r1d yading@10: lea r8, [r7+r7*2] yading@10: lea r6, [r0-4] yading@10: lea r5, [r0-4+r8] yading@10: %if WIN64 yading@10: sub rsp, 0x98 yading@10: %define pix_tmp rsp+0x30 yading@10: %else yading@10: sub rsp, 0x68 yading@10: %define pix_tmp rsp yading@10: %endif yading@10: yading@10: ; transpose 6x16 -> tmp space yading@10: TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r7, r8), pix_tmp yading@10: lea r6, [r6+r7*8] yading@10: lea r5, [r5+r7*8] yading@10: TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r7, r8), pix_tmp+8 yading@10: yading@10: ; vertical filter yading@10: ; alpha, beta, tc0 are still in r2d, r3d, r4 yading@10: ; don't backup r6, r5, r7, r8 because deblock_v_luma_sse2 doesn't use them yading@10: lea r0, [pix_tmp+0x30] yading@10: mov r1d, 0x10 yading@10: %if WIN64 yading@10: mov [rsp+0x20], r4 yading@10: %endif yading@10: call deblock_v_luma_8 yading@10: yading@10: ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter) yading@10: add r6, 2 yading@10: add r5, 2 yading@10: movq m0, [pix_tmp+0x18] yading@10: movq m1, [pix_tmp+0x28] yading@10: movq m2, [pix_tmp+0x38] yading@10: movq m3, [pix_tmp+0x48] yading@10: TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8) yading@10: yading@10: shl r7, 3 yading@10: sub r6, r7 yading@10: sub r5, r7 yading@10: shr r7, 3 yading@10: movq m0, [pix_tmp+0x10] yading@10: movq m1, [pix_tmp+0x20] yading@10: movq m2, [pix_tmp+0x30] yading@10: movq m3, [pix_tmp+0x40] yading@10: TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8) yading@10: yading@10: %if WIN64 yading@10: add rsp, 0x98 yading@10: %else yading@10: add rsp, 0x68 yading@10: %endif yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: DEBLOCK_LUMA yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: DEBLOCK_LUMA yading@10: %endif yading@10: yading@10: %else yading@10: yading@10: %macro DEBLOCK_LUMA 2 yading@10: ;----------------------------------------------------------------------------- yading@10: ; void deblock_v8_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) yading@10: ;----------------------------------------------------------------------------- yading@10: cglobal deblock_%1_luma_8, 5,5,8,2*%2 yading@10: lea r4, [r1*3] yading@10: dec r2 ; alpha-1 yading@10: neg r4 yading@10: dec r3 ; beta-1 yading@10: add r4, r0 ; pix-3*stride yading@10: yading@10: mova m0, [r4+r1] ; p1 yading@10: mova m1, [r4+2*r1] ; p0 yading@10: mova m2, [r0] ; q0 yading@10: mova m3, [r0+r1] ; q1 yading@10: LOAD_MASK r2, r3 yading@10: yading@10: mov r3, r4mp yading@10: pcmpeqb m3, m3 yading@10: movd m4, [r3] ; tc0 yading@10: punpcklbw m4, m4 yading@10: punpcklbw m4, m4 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0] yading@10: mova [esp+%2], m4 ; tc yading@10: pcmpgtb m4, m3 yading@10: mova m3, [r4] ; p2 yading@10: pand m4, m7 yading@10: mova [esp], m4 ; mask yading@10: yading@10: DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1 yading@10: pand m6, m4 yading@10: pand m4, [esp+%2] ; tc yading@10: psubb m7, m4, m6 yading@10: pand m6, m4 yading@10: LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4 yading@10: yading@10: mova m4, [r0+2*r1] ; q2 yading@10: DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1 yading@10: pand m6, [esp] ; mask yading@10: mova m5, [esp+%2] ; tc yading@10: psubb m7, m6 yading@10: pand m5, m6 yading@10: mova m3, [r0+r1] yading@10: LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m5, m6 yading@10: yading@10: DEBLOCK_P0_Q0 yading@10: mova [r4+2*r1], m1 yading@10: mova [r0], m2 yading@10: RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) yading@10: ;----------------------------------------------------------------------------- yading@10: INIT_MMX cpuname yading@10: cglobal deblock_h_luma_8, 0,5,8,0x60+HAVE_ALIGNED_STACK*12 yading@10: mov r0, r0mp yading@10: mov r3, r1m yading@10: lea r4, [r3*3] yading@10: sub r0, 4 yading@10: lea r1, [r0+r4] yading@10: %define pix_tmp esp+12*HAVE_ALIGNED_STACK yading@10: yading@10: ; transpose 6x16 -> tmp space yading@10: TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp yading@10: lea r0, [r0+r3*8] yading@10: lea r1, [r1+r3*8] yading@10: TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp+8 yading@10: yading@10: ; vertical filter yading@10: lea r0, [pix_tmp+0x30] yading@10: PUSH dword r4m yading@10: PUSH dword r3m yading@10: PUSH dword r2m yading@10: PUSH dword 16 yading@10: PUSH dword r0 yading@10: call deblock_%1_luma_8 yading@10: %ifidn %1, v8 yading@10: add dword [esp ], 8 ; pix_tmp+0x38 yading@10: add dword [esp+16], 2 ; tc0+2 yading@10: call deblock_%1_luma_8 yading@10: %endif yading@10: ADD esp, 20 yading@10: yading@10: ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter) yading@10: mov r0, r0mp yading@10: sub r0, 2 yading@10: yading@10: movq m0, [pix_tmp+0x10] yading@10: movq m1, [pix_tmp+0x20] yading@10: lea r1, [r0+r4] yading@10: movq m2, [pix_tmp+0x30] yading@10: movq m3, [pix_tmp+0x40] yading@10: TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4) yading@10: yading@10: lea r0, [r0+r3*8] yading@10: lea r1, [r1+r3*8] yading@10: movq m0, [pix_tmp+0x18] yading@10: movq m1, [pix_tmp+0x28] yading@10: movq m2, [pix_tmp+0x38] yading@10: movq m3, [pix_tmp+0x48] yading@10: TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4) yading@10: yading@10: RET yading@10: %endmacro ; DEBLOCK_LUMA yading@10: yading@10: INIT_MMX mmxext yading@10: DEBLOCK_LUMA v8, 8 yading@10: INIT_XMM sse2 yading@10: DEBLOCK_LUMA v, 16 yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: DEBLOCK_LUMA v, 16 yading@10: %endif yading@10: yading@10: %endif ; ARCH yading@10: yading@10: yading@10: yading@10: %macro LUMA_INTRA_P012 4 ; p0..p3 in memory yading@10: %if ARCH_X86_64 yading@10: pavgb t0, p2, p1 yading@10: pavgb t1, p0, q0 yading@10: %else yading@10: mova t0, p2 yading@10: mova t1, p0 yading@10: pavgb t0, p1 yading@10: pavgb t1, q0 yading@10: %endif yading@10: pavgb t0, t1 ; ((p2+p1+1)/2 + (p0+q0+1)/2 + 1)/2 yading@10: mova t5, t1 yading@10: %if ARCH_X86_64 yading@10: paddb t2, p2, p1 yading@10: paddb t3, p0, q0 yading@10: %else yading@10: mova t2, p2 yading@10: mova t3, p0 yading@10: paddb t2, p1 yading@10: paddb t3, q0 yading@10: %endif yading@10: paddb t2, t3 yading@10: mova t3, t2 yading@10: mova t4, t2 yading@10: psrlw t2, 1 yading@10: pavgb t2, mpb_0 yading@10: pxor t2, t0 yading@10: pand t2, mpb_1 yading@10: psubb t0, t2 ; p1' = (p2+p1+p0+q0+2)/4; yading@10: yading@10: %if ARCH_X86_64 yading@10: pavgb t1, p2, q1 yading@10: psubb t2, p2, q1 yading@10: %else yading@10: mova t1, p2 yading@10: mova t2, p2 yading@10: pavgb t1, q1 yading@10: psubb t2, q1 yading@10: %endif yading@10: paddb t3, t3 yading@10: psubb t3, t2 ; p2+2*p1+2*p0+2*q0+q1 yading@10: pand t2, mpb_1 yading@10: psubb t1, t2 yading@10: pavgb t1, p1 yading@10: pavgb t1, t5 ; (((p2+q1)/2 + p1+1)/2 + (p0+q0+1)/2 + 1)/2 yading@10: psrlw t3, 2 yading@10: pavgb t3, mpb_0 yading@10: pxor t3, t1 yading@10: pand t3, mpb_1 yading@10: psubb t1, t3 ; p0'a = (p2+2*p1+2*p0+2*q0+q1+4)/8 yading@10: yading@10: pxor t3, p0, q1 yading@10: pavgb t2, p0, q1 yading@10: pand t3, mpb_1 yading@10: psubb t2, t3 yading@10: pavgb t2, p1 ; p0'b = (2*p1+p0+q0+2)/4 yading@10: yading@10: pxor t1, t2 yading@10: pxor t2, p0 yading@10: pand t1, mask1p yading@10: pand t2, mask0 yading@10: pxor t1, t2 yading@10: pxor t1, p0 yading@10: mova %1, t1 ; store p0 yading@10: yading@10: mova t1, %4 ; p3 yading@10: paddb t2, t1, p2 yading@10: pavgb t1, p2 yading@10: pavgb t1, t0 ; (p3+p2+1)/2 + (p2+p1+p0+q0+2)/4 yading@10: paddb t2, t2 yading@10: paddb t2, t4 ; 2*p3+3*p2+p1+p0+q0 yading@10: psrlw t2, 2 yading@10: pavgb t2, mpb_0 yading@10: pxor t2, t1 yading@10: pand t2, mpb_1 yading@10: psubb t1, t2 ; p2' = (2*p3+3*p2+p1+p0+q0+4)/8 yading@10: yading@10: pxor t0, p1 yading@10: pxor t1, p2 yading@10: pand t0, mask1p yading@10: pand t1, mask1p yading@10: pxor t0, p1 yading@10: pxor t1, p2 yading@10: mova %2, t0 ; store p1 yading@10: mova %3, t1 ; store p2 yading@10: %endmacro yading@10: yading@10: %macro LUMA_INTRA_SWAP_PQ 0 yading@10: %define q1 m0 yading@10: %define q0 m1 yading@10: %define p0 m2 yading@10: %define p1 m3 yading@10: %define p2 q2 yading@10: %define mask1p mask1q yading@10: %endmacro yading@10: yading@10: %macro DEBLOCK_LUMA_INTRA 1 yading@10: %define p1 m0 yading@10: %define p0 m1 yading@10: %define q0 m2 yading@10: %define q1 m3 yading@10: %define t0 m4 yading@10: %define t1 m5 yading@10: %define t2 m6 yading@10: %define t3 m7 yading@10: %if ARCH_X86_64 yading@10: %define p2 m8 yading@10: %define q2 m9 yading@10: %define t4 m10 yading@10: %define t5 m11 yading@10: %define mask0 m12 yading@10: %define mask1p m13 yading@10: %if WIN64 yading@10: %define mask1q [rsp] yading@10: %else yading@10: %define mask1q [rsp-24] yading@10: %endif yading@10: %define mpb_0 m14 yading@10: %define mpb_1 m15 yading@10: %else yading@10: %define spill(x) [esp+16*x] yading@10: %define p2 [r4+r1] yading@10: %define q2 [r0+2*r1] yading@10: %define t4 spill(0) yading@10: %define t5 spill(1) yading@10: %define mask0 spill(2) yading@10: %define mask1p spill(3) yading@10: %define mask1q spill(4) yading@10: %define mpb_0 [pb_0] yading@10: %define mpb_1 [pb_1] yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void deblock_v_luma_intra( uint8_t *pix, int stride, int alpha, int beta ) yading@10: ;----------------------------------------------------------------------------- yading@10: %if WIN64 yading@10: cglobal deblock_%1_luma_intra_8, 4,6,16,0x10 yading@10: %else yading@10: cglobal deblock_%1_luma_intra_8, 4,6,16,ARCH_X86_64*0x50-0x50 yading@10: %endif yading@10: lea r4, [r1*4] yading@10: lea r5, [r1*3] ; 3*stride yading@10: dec r2d ; alpha-1 yading@10: jl .end yading@10: neg r4 yading@10: dec r3d ; beta-1 yading@10: jl .end yading@10: add r4, r0 ; pix-4*stride yading@10: mova p1, [r4+2*r1] yading@10: mova p0, [r4+r5] yading@10: mova q0, [r0] yading@10: mova q1, [r0+r1] yading@10: %if ARCH_X86_64 yading@10: pxor mpb_0, mpb_0 yading@10: mova mpb_1, [pb_1] yading@10: LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0 yading@10: SWAP 7, 12 ; m12=mask0 yading@10: pavgb t5, mpb_0 yading@10: pavgb t5, mpb_1 ; alpha/4+1 yading@10: movdqa p2, [r4+r1] yading@10: movdqa q2, [r0+2*r1] yading@10: DIFF_GT2 p0, q0, t5, t0, t3 ; t0 = |p0-q0| > alpha/4+1 yading@10: DIFF_GT2 p0, p2, m5, t2, t5 ; mask1 = |p2-p0| > beta-1 yading@10: DIFF_GT2 q0, q2, m5, t4, t5 ; t4 = |q2-q0| > beta-1 yading@10: pand t0, mask0 yading@10: pand t4, t0 yading@10: pand t2, t0 yading@10: mova mask1q, t4 yading@10: mova mask1p, t2 yading@10: %else yading@10: LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0 yading@10: mova m4, t5 yading@10: mova mask0, m7 yading@10: pavgb m4, [pb_0] yading@10: pavgb m4, [pb_1] ; alpha/4+1 yading@10: DIFF_GT2 p0, q0, m4, m6, m7 ; m6 = |p0-q0| > alpha/4+1 yading@10: pand m6, mask0 yading@10: DIFF_GT2 p0, p2, m5, m4, m7 ; m4 = |p2-p0| > beta-1 yading@10: pand m4, m6 yading@10: mova mask1p, m4 yading@10: DIFF_GT2 q0, q2, m5, m4, m7 ; m4 = |q2-q0| > beta-1 yading@10: pand m4, m6 yading@10: mova mask1q, m4 yading@10: %endif yading@10: LUMA_INTRA_P012 [r4+r5], [r4+2*r1], [r4+r1], [r4] yading@10: LUMA_INTRA_SWAP_PQ yading@10: LUMA_INTRA_P012 [r0], [r0+r1], [r0+2*r1], [r0+r5] yading@10: .end: yading@10: RET yading@10: yading@10: INIT_MMX cpuname yading@10: %if ARCH_X86_64 yading@10: ;----------------------------------------------------------------------------- yading@10: ; void deblock_h_luma_intra( uint8_t *pix, int stride, int alpha, int beta ) yading@10: ;----------------------------------------------------------------------------- yading@10: cglobal deblock_h_luma_intra_8, 4,9 yading@10: movsxd r7, r1d yading@10: lea r8, [r7*3] yading@10: lea r6, [r0-4] yading@10: lea r5, [r0-4+r8] yading@10: sub rsp, 0x88 yading@10: %define pix_tmp rsp yading@10: yading@10: ; transpose 8x16 -> tmp space yading@10: TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r7, r8), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30) yading@10: lea r6, [r6+r7*8] yading@10: lea r5, [r5+r7*8] yading@10: TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r7, r8), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30) yading@10: yading@10: lea r0, [pix_tmp+0x40] yading@10: mov r1, 0x10 yading@10: call deblock_v_luma_intra_8 yading@10: yading@10: ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8) yading@10: lea r5, [r6+r8] yading@10: TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8) yading@10: shl r7, 3 yading@10: sub r6, r7 yading@10: sub r5, r7 yading@10: shr r7, 3 yading@10: TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8) yading@10: add rsp, 0x88 yading@10: RET yading@10: %else yading@10: cglobal deblock_h_luma_intra_8, 2,4,8,0x80 yading@10: lea r3, [r1*3] yading@10: sub r0, 4 yading@10: lea r2, [r0+r3] yading@10: %define pix_tmp rsp yading@10: yading@10: ; transpose 8x16 -> tmp space yading@10: TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30) yading@10: lea r0, [r0+r1*8] yading@10: lea r2, [r2+r1*8] yading@10: TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30) yading@10: yading@10: lea r0, [pix_tmp+0x40] yading@10: PUSH dword r3m yading@10: PUSH dword r2m yading@10: PUSH dword 16 yading@10: PUSH r0 yading@10: call deblock_%1_luma_intra_8 yading@10: %ifidn %1, v8 yading@10: add dword [rsp], 8 ; pix_tmp+8 yading@10: call deblock_%1_luma_intra_8 yading@10: %endif yading@10: ADD esp, 16 yading@10: yading@10: mov r1, r1m yading@10: mov r0, r0mp yading@10: lea r3, [r1*3] yading@10: sub r0, 4 yading@10: lea r2, [r0+r3] yading@10: ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8) yading@10: TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3) yading@10: lea r0, [r0+r1*8] yading@10: lea r2, [r2+r1*8] yading@10: TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3) yading@10: RET yading@10: %endif ; ARCH_X86_64 yading@10: %endmacro ; DEBLOCK_LUMA_INTRA yading@10: yading@10: INIT_XMM sse2 yading@10: DEBLOCK_LUMA_INTRA v yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: DEBLOCK_LUMA_INTRA v yading@10: %endif yading@10: %if ARCH_X86_64 == 0 yading@10: INIT_MMX mmxext yading@10: DEBLOCK_LUMA_INTRA v8 yading@10: %endif yading@10: yading@10: INIT_MMX mmxext yading@10: yading@10: %macro CHROMA_V_START 0 yading@10: dec r2d ; alpha-1 yading@10: dec r3d ; beta-1 yading@10: mov t5, r0 yading@10: sub t5, r1 yading@10: sub t5, r1 yading@10: %endmacro yading@10: yading@10: %macro CHROMA_H_START 0 yading@10: dec r2d yading@10: dec r3d yading@10: sub r0, 2 yading@10: lea t6, [r1*3] yading@10: mov t5, r0 yading@10: add r0, t6 yading@10: %endmacro yading@10: yading@10: %define t5 r5 yading@10: %define t6 r6 yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void ff_deblock_v_chroma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) yading@10: ;----------------------------------------------------------------------------- yading@10: cglobal deblock_v_chroma_8, 5,6 yading@10: CHROMA_V_START yading@10: movq m0, [t5] yading@10: movq m1, [t5+r1] yading@10: movq m2, [r0] yading@10: movq m3, [r0+r1] yading@10: call ff_chroma_inter_body_mmxext yading@10: movq [t5+r1], m1 yading@10: movq [r0], m2 yading@10: RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void ff_deblock_h_chroma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) yading@10: ;----------------------------------------------------------------------------- yading@10: cglobal deblock_h_chroma_8, 5,7 yading@10: %if UNIX64 yading@10: %define buf0 [rsp-24] yading@10: %define buf1 [rsp-16] yading@10: %elif WIN64 yading@10: sub rsp, 16 yading@10: %define buf0 [rsp] yading@10: %define buf1 [rsp+8] yading@10: %else yading@10: %define buf0 r0m yading@10: %define buf1 r2m yading@10: %endif yading@10: CHROMA_H_START yading@10: TRANSPOSE4x8_LOAD bw, wd, dq, PASS8ROWS(t5, r0, r1, t6) yading@10: movq buf0, m0 yading@10: movq buf1, m3 yading@10: LOAD_MASK r2d, r3d yading@10: movd m6, [r4] ; tc0 yading@10: punpcklbw m6, m6 yading@10: pand m7, m6 yading@10: DEBLOCK_P0_Q0 yading@10: movq m0, buf0 yading@10: movq m3, buf1 yading@10: TRANSPOSE8x4B_STORE PASS8ROWS(t5, r0, r1, t6) yading@10: %if WIN64 yading@10: add rsp, 16 yading@10: %endif yading@10: RET yading@10: yading@10: ALIGN 16 yading@10: ff_chroma_inter_body_mmxext: yading@10: LOAD_MASK r2d, r3d yading@10: movd m6, [r4] ; tc0 yading@10: punpcklbw m6, m6 yading@10: pand m7, m6 yading@10: DEBLOCK_P0_Q0 yading@10: ret yading@10: yading@10: yading@10: yading@10: ; in: %1=p0 %2=p1 %3=q1 yading@10: ; out: p0 = (p0 + q1 + 2*p1 + 2) >> 2 yading@10: %macro CHROMA_INTRA_P0 3 yading@10: movq m4, %1 yading@10: pxor m4, %3 yading@10: pand m4, [pb_1] ; m4 = (p0^q1)&1 yading@10: pavgb %1, %3 yading@10: psubusb %1, m4 yading@10: pavgb %1, %2 ; dst = avg(p1, avg(p0,q1) - ((p0^q1)&1)) yading@10: %endmacro yading@10: yading@10: %define t5 r4 yading@10: %define t6 r5 yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void ff_deblock_v_chroma_intra( uint8_t *pix, int stride, int alpha, int beta ) yading@10: ;----------------------------------------------------------------------------- yading@10: cglobal deblock_v_chroma_intra_8, 4,5 yading@10: CHROMA_V_START yading@10: movq m0, [t5] yading@10: movq m1, [t5+r1] yading@10: movq m2, [r0] yading@10: movq m3, [r0+r1] yading@10: call ff_chroma_intra_body_mmxext yading@10: movq [t5+r1], m1 yading@10: movq [r0], m2 yading@10: RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void ff_deblock_h_chroma_intra( uint8_t *pix, int stride, int alpha, int beta ) yading@10: ;----------------------------------------------------------------------------- yading@10: cglobal deblock_h_chroma_intra_8, 4,6 yading@10: CHROMA_H_START yading@10: TRANSPOSE4x8_LOAD bw, wd, dq, PASS8ROWS(t5, r0, r1, t6) yading@10: call ff_chroma_intra_body_mmxext yading@10: TRANSPOSE8x4B_STORE PASS8ROWS(t5, r0, r1, t6) yading@10: RET yading@10: yading@10: ALIGN 16 yading@10: ff_chroma_intra_body_mmxext: yading@10: LOAD_MASK r2d, r3d yading@10: movq m5, m1 yading@10: movq m6, m2 yading@10: CHROMA_INTRA_P0 m1, m0, m3 yading@10: CHROMA_INTRA_P0 m2, m3, m0 yading@10: psubb m1, m5 yading@10: psubb m2, m6 yading@10: pand m1, m7 yading@10: pand m2, m7 yading@10: paddb m1, m5 yading@10: paddb m2, m6 yading@10: ret yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void h264_loop_filter_strength(int16_t bs[2][4][4], uint8_t nnz[40], yading@10: ; int8_t ref[2][40], int16_t mv[2][40][2], yading@10: ; int bidir, int edges, int step, yading@10: ; int mask_mv0, int mask_mv1, int field); yading@10: ; yading@10: ; bidir is 0 or 1 yading@10: ; edges is 1 or 4 yading@10: ; step is 1 or 2 yading@10: ; mask_mv0 is 0 or 3 yading@10: ; mask_mv1 is 0 or 1 yading@10: ; field is 0 or 1 yading@10: ;----------------------------------------------------------------------------- yading@10: %macro loop_filter_strength_iteration 7 ; edges, step, mask_mv, yading@10: ; dir, d_idx, mask_dir, bidir yading@10: %define edgesd %1 yading@10: %define stepd %2 yading@10: %define mask_mvd %3 yading@10: %define dir %4 yading@10: %define d_idx %5 yading@10: %define mask_dir %6 yading@10: %define bidir %7 yading@10: xor b_idxd, b_idxd ; for (b_idx = 0; b_idx < edges; b_idx += step) yading@10: %%.b_idx_loop: yading@10: %if mask_dir == 0 yading@10: pxor m0, m0 yading@10: %endif yading@10: test b_idxd, dword mask_mvd yading@10: jnz %%.skip_loop_iter ; if (!(b_idx & mask_mv)) yading@10: %if bidir == 1 yading@10: movd m2, [refq+b_idxq+d_idx+12] ; { ref0[bn] } yading@10: punpckldq m2, [refq+b_idxq+d_idx+52] ; { ref0[bn], ref1[bn] } yading@10: pshufw m0, [refq+b_idxq+12], 0x44 ; { ref0[b], ref0[b] } yading@10: pshufw m1, [refq+b_idxq+52], 0x44 ; { ref1[b], ref1[b] } yading@10: pshufw m3, m2, 0x4E ; { ref1[bn], ref0[bn] } yading@10: psubb m0, m2 ; { ref0[b] != ref0[bn], yading@10: ; ref0[b] != ref1[bn] } yading@10: psubb m1, m3 ; { ref1[b] != ref1[bn], yading@10: ; ref1[b] != ref0[bn] } yading@10: yading@10: por m0, m1 yading@10: mova m1, [mvq+b_idxq*4+(d_idx+12)*4] yading@10: mova m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize] yading@10: mova m3, m1 yading@10: mova m4, m2 yading@10: psubw m1, [mvq+b_idxq*4+12*4] yading@10: psubw m2, [mvq+b_idxq*4+12*4+mmsize] yading@10: psubw m3, [mvq+b_idxq*4+52*4] yading@10: psubw m4, [mvq+b_idxq*4+52*4+mmsize] yading@10: packsswb m1, m2 yading@10: packsswb m3, m4 yading@10: paddb m1, m6 yading@10: paddb m3, m6 yading@10: psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit yading@10: psubusb m3, m5 yading@10: packsswb m1, m3 yading@10: yading@10: por m0, m1 yading@10: mova m1, [mvq+b_idxq*4+(d_idx+52)*4] yading@10: mova m2, [mvq+b_idxq*4+(d_idx+52)*4+mmsize] yading@10: mova m3, m1 yading@10: mova m4, m2 yading@10: psubw m1, [mvq+b_idxq*4+12*4] yading@10: psubw m2, [mvq+b_idxq*4+12*4+mmsize] yading@10: psubw m3, [mvq+b_idxq*4+52*4] yading@10: psubw m4, [mvq+b_idxq*4+52*4+mmsize] yading@10: packsswb m1, m2 yading@10: packsswb m3, m4 yading@10: paddb m1, m6 yading@10: paddb m3, m6 yading@10: psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit yading@10: psubusb m3, m5 yading@10: packsswb m1, m3 yading@10: yading@10: pshufw m1, m1, 0x4E yading@10: por m0, m1 yading@10: pshufw m1, m0, 0x4E yading@10: pminub m0, m1 yading@10: %else ; bidir == 0 yading@10: movd m0, [refq+b_idxq+12] yading@10: psubb m0, [refq+b_idxq+d_idx+12] ; ref[b] != ref[bn] yading@10: yading@10: mova m1, [mvq+b_idxq*4+12*4] yading@10: mova m2, [mvq+b_idxq*4+12*4+mmsize] yading@10: psubw m1, [mvq+b_idxq*4+(d_idx+12)*4] yading@10: psubw m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize] yading@10: packsswb m1, m2 yading@10: paddb m1, m6 yading@10: psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit yading@10: packsswb m1, m1 yading@10: por m0, m1 yading@10: %endif ; bidir == 1/0 yading@10: yading@10: %%.skip_loop_iter: yading@10: movd m1, [nnzq+b_idxq+12] yading@10: por m1, [nnzq+b_idxq+d_idx+12] ; nnz[b] || nnz[bn] yading@10: yading@10: pminub m1, m7 yading@10: pminub m0, m7 yading@10: psllw m1, 1 yading@10: pxor m2, m2 yading@10: pmaxub m1, m0 yading@10: punpcklbw m1, m2 yading@10: movq [bsq+b_idxq+32*dir], m1 yading@10: yading@10: add b_idxd, dword stepd yading@10: cmp b_idxd, dword edgesd yading@10: jl %%.b_idx_loop yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: cglobal h264_loop_filter_strength, 9, 9, 0, bs, nnz, ref, mv, bidir, edges, \ yading@10: step, mask_mv0, mask_mv1, field yading@10: %define b_idxq bidirq yading@10: %define b_idxd bidird yading@10: cmp dword fieldm, 0 yading@10: mova m7, [pb_1] yading@10: mova m5, [pb_3] yading@10: je .nofield yading@10: mova m5, [pb_3_1] yading@10: .nofield: yading@10: mova m6, m5 yading@10: paddb m5, m5 yading@10: yading@10: shl dword stepd, 3 yading@10: shl dword edgesd, 3 yading@10: %if ARCH_X86_32 yading@10: %define mask_mv0d mask_mv0m yading@10: %define mask_mv1d mask_mv1m yading@10: %endif yading@10: shl dword mask_mv1d, 3 yading@10: shl dword mask_mv0d, 3 yading@10: yading@10: cmp dword bidird, 0 yading@10: jne .bidir yading@10: loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8, 0, 0 yading@10: loop_filter_strength_iteration 32, 8, mask_mv0d, 0, -1, -1, 0 yading@10: yading@10: mova m0, [bsq+mmsize*0] yading@10: mova m1, [bsq+mmsize*1] yading@10: mova m2, [bsq+mmsize*2] yading@10: mova m3, [bsq+mmsize*3] yading@10: TRANSPOSE4x4W 0, 1, 2, 3, 4 yading@10: mova [bsq+mmsize*0], m0 yading@10: mova [bsq+mmsize*1], m1 yading@10: mova [bsq+mmsize*2], m2 yading@10: mova [bsq+mmsize*3], m3 yading@10: RET yading@10: yading@10: .bidir: yading@10: loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8, 0, 1 yading@10: loop_filter_strength_iteration 32, 8, mask_mv0d, 0, -1, -1, 1 yading@10: yading@10: mova m0, [bsq+mmsize*0] yading@10: mova m1, [bsq+mmsize*1] yading@10: mova m2, [bsq+mmsize*2] yading@10: mova m3, [bsq+mmsize*3] yading@10: TRANSPOSE4x4W 0, 1, 2, 3, 4 yading@10: mova [bsq+mmsize*0], m0 yading@10: mova [bsq+mmsize*1], m1 yading@10: mova [bsq+mmsize*2], m2 yading@10: mova [bsq+mmsize*3], m3 yading@10: RET