michael@0: ; michael@0: ; Copyright (c) 2010 The WebM project authors. All Rights Reserved. michael@0: ; michael@0: ; Use of this source code is governed by a BSD-style license michael@0: ; that can be found in the LICENSE file in the root of the source michael@0: ; tree. An additional intellectual property rights grant can be found michael@0: ; in the file PATENTS. All contributing project authors may michael@0: ; be found in the AUTHORS file in the root of the source tree. michael@0: ; michael@0: michael@0: michael@0: %include "vpx_ports/x86_abi_support.asm" michael@0: michael@0: ;void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch) michael@0: global sym(vp8_short_walsh4x4_sse2) PRIVATE michael@0: sym(vp8_short_walsh4x4_sse2): michael@0: push rbp michael@0: mov rbp, rsp michael@0: SHADOW_ARGS_TO_STACK 3 michael@0: SAVE_XMM 7 michael@0: GET_GOT rbx michael@0: push rsi michael@0: push rdi michael@0: ; end prolog michael@0: michael@0: mov rsi, arg(0) ; input michael@0: mov rdi, arg(1) ; output michael@0: movsxd rdx, dword ptr arg(2) ; pitch michael@0: michael@0: ; first for loop michael@0: movq xmm0, MMWORD PTR [rsi] ; load input michael@0: movq xmm1, MMWORD PTR [rsi + rdx] michael@0: lea rsi, [rsi + rdx*2] michael@0: movq xmm2, MMWORD PTR [rsi] michael@0: movq xmm3, MMWORD PTR [rsi + rdx] michael@0: michael@0: punpcklwd xmm0, xmm1 michael@0: punpcklwd xmm2, xmm3 michael@0: michael@0: movdqa xmm1, xmm0 michael@0: punpckldq xmm0, xmm2 ; ip[1] ip[0] michael@0: punpckhdq xmm1, xmm2 ; ip[3] ip[2] michael@0: michael@0: movdqa xmm2, xmm0 michael@0: paddw xmm0, xmm1 michael@0: psubw xmm2, xmm1 michael@0: michael@0: psllw xmm0, 2 ; d1 a1 michael@0: psllw xmm2, 2 ; c1 b1 michael@0: michael@0: movdqa xmm1, xmm0 michael@0: punpcklqdq xmm0, xmm2 ; b1 a1 michael@0: punpckhqdq xmm1, xmm2 ; c1 d1 michael@0: michael@0: pxor xmm6, xmm6 michael@0: movq xmm6, xmm0 michael@0: pxor xmm7, xmm7 michael@0: pcmpeqw xmm7, xmm6 michael@0: paddw xmm7, [GLOBAL(c1)] michael@0: michael@0: movdqa xmm2, xmm0 michael@0: paddw xmm0, xmm1 ; b1+c1 a1+d1 michael@0: psubw xmm2, xmm1 ; b1-c1 a1-d1 michael@0: paddw xmm0, xmm7 ; b1+c1 a1+d1+(a1!=0) michael@0: michael@0: ; second for loop michael@0: ; input: 13 9 5 1 12 8 4 0 (xmm0) michael@0: ; 14 10 6 2 15 11 7 3 (xmm2) michael@0: ; after shuffle: michael@0: ; 13 5 9 1 12 4 8 0 (xmm0) michael@0: ; 14 6 10 2 15 7 11 3 (xmm1) michael@0: pshuflw xmm3, xmm0, 0xd8 michael@0: pshufhw xmm0, xmm3, 0xd8 michael@0: pshuflw xmm3, xmm2, 0xd8 michael@0: pshufhw xmm1, xmm3, 0xd8 michael@0: michael@0: movdqa xmm2, xmm0 michael@0: pmaddwd xmm0, [GLOBAL(c1)] ; d11 a11 d10 a10 michael@0: pmaddwd xmm2, [GLOBAL(cn1)] ; c11 b11 c10 b10 michael@0: movdqa xmm3, xmm1 michael@0: pmaddwd xmm1, [GLOBAL(c1)] ; d12 a12 d13 a13 michael@0: pmaddwd xmm3, [GLOBAL(cn1)] ; c12 b12 c13 b13 michael@0: michael@0: pshufd xmm4, xmm0, 0xd8 ; d11 d10 a11 a10 michael@0: pshufd xmm5, xmm2, 0xd8 ; c11 c10 b11 b10 michael@0: pshufd xmm6, xmm1, 0x72 ; d13 d12 a13 a12 michael@0: pshufd xmm7, xmm3, 0x72 ; c13 c12 b13 b12 michael@0: michael@0: movdqa xmm0, xmm4 michael@0: punpcklqdq xmm0, xmm5 ; b11 b10 a11 a10 michael@0: punpckhqdq xmm4, xmm5 ; c11 c10 d11 d10 michael@0: movdqa xmm1, xmm6 michael@0: punpcklqdq xmm1, xmm7 ; b13 b12 a13 a12 michael@0: punpckhqdq xmm6, xmm7 ; c13 c12 d13 d12 michael@0: michael@0: movdqa xmm2, xmm0 michael@0: paddd xmm0, xmm4 ; b21 b20 a21 a20 michael@0: psubd xmm2, xmm4 ; c21 c20 d21 d20 michael@0: movdqa xmm3, xmm1 michael@0: paddd xmm1, xmm6 ; b23 b22 a23 a22 michael@0: psubd xmm3, xmm6 ; c23 c22 d23 d22 michael@0: michael@0: pxor xmm4, xmm4 michael@0: movdqa xmm5, xmm4 michael@0: pcmpgtd xmm4, xmm0 michael@0: pcmpgtd xmm5, xmm2 michael@0: pand xmm4, [GLOBAL(cd1)] michael@0: pand xmm5, [GLOBAL(cd1)] michael@0: michael@0: pxor xmm6, xmm6 michael@0: movdqa xmm7, xmm6 michael@0: pcmpgtd xmm6, xmm1 michael@0: pcmpgtd xmm7, xmm3 michael@0: pand xmm6, [GLOBAL(cd1)] michael@0: pand xmm7, [GLOBAL(cd1)] michael@0: michael@0: paddd xmm0, xmm4 michael@0: paddd xmm2, xmm5 michael@0: paddd xmm0, [GLOBAL(cd3)] michael@0: paddd xmm2, [GLOBAL(cd3)] michael@0: paddd xmm1, xmm6 michael@0: paddd xmm3, xmm7 michael@0: paddd xmm1, [GLOBAL(cd3)] michael@0: paddd xmm3, [GLOBAL(cd3)] michael@0: michael@0: psrad xmm0, 3 michael@0: psrad xmm1, 3 michael@0: psrad xmm2, 3 michael@0: psrad xmm3, 3 michael@0: movdqa xmm4, xmm0 michael@0: punpcklqdq xmm0, xmm1 ; a23 a22 a21 a20 michael@0: punpckhqdq xmm4, xmm1 ; b23 b22 b21 b20 michael@0: movdqa xmm5, xmm2 michael@0: punpckhqdq xmm2, xmm3 ; c23 c22 c21 c20 michael@0: punpcklqdq xmm5, xmm3 ; d23 d22 d21 d20 michael@0: michael@0: packssdw xmm0, xmm4 ; b23 b22 b21 b20 a23 a22 a21 a20 michael@0: packssdw xmm2, xmm5 ; d23 d22 d21 d20 c23 c22 c21 c20 michael@0: michael@0: movdqa XMMWORD PTR [rdi], xmm0 michael@0: movdqa XMMWORD PTR [rdi + 16], xmm2 michael@0: michael@0: ; begin epilog michael@0: pop rdi michael@0: pop rsi michael@0: RESTORE_GOT michael@0: RESTORE_XMM michael@0: UNSHADOW_ARGS michael@0: pop rbp michael@0: ret michael@0: michael@0: SECTION_RODATA michael@0: align 16 michael@0: c1: michael@0: dw 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001 michael@0: align 16 michael@0: cn1: michael@0: dw 0x0001, 0xffff, 0x0001, 0xffff, 0x0001, 0xffff, 0x0001, 0xffff michael@0: align 16 michael@0: cd1: michael@0: dd 0x00000001, 0x00000001, 0x00000001, 0x00000001 michael@0: align 16 michael@0: cd3: michael@0: dd 0x00000003, 0x00000003, 0x00000003, 0x00000003