190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;
2f71323e297a928af368937089d3ed71239786f86Andreas Huber;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;
4f71323e297a928af368937089d3ed71239786f86Andreas Huber;  Use of this source code is governed by a BSD-style license
5f71323e297a928af368937089d3ed71239786f86Andreas Huber;  that can be found in the LICENSE file in the root of the source
6f71323e297a928af368937089d3ed71239786f86Andreas Huber;  tree. An additional intellectual property rights grant can be found
7f71323e297a928af368937089d3ed71239786f86Andreas Huber;  in the file PATENTS.  All contributing project authors may
8f71323e297a928af368937089d3ed71239786f86Andreas Huber;  be found in the AUTHORS file in the root of the source tree.
990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;
1090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
1190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
1290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber%include "vpx_ports/x86_abi_support.asm"
1390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
1490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;void vp8_short_inv_walsh4x4_sse2(short *input, short *output)
1590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberglobal sym(vp8_short_inv_walsh4x4_sse2)
1690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Hubersym(vp8_short_inv_walsh4x4_sse2):
1790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rbp
1890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov         rbp, rsp
1990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    SHADOW_ARGS_TO_STACK 2
20f71323e297a928af368937089d3ed71239786f86Andreas Huber    SAVE_XMM
2190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rsi
2290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rdi
2390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; end prolog
2490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
2590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rsi, arg(0)
2690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rdi, arg(1)
2790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rax, 3
2890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
2990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm0, [rsi + 0]       ;ip[4] ip[0]
3090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm1, [rsi + 16]      ;ip[12] ip[8]
3190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    shl     rax, 16
3390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    or      rax, 3            ;00030003h
3490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pshufd    xmm2, xmm1, 4eh       ;ip[8] ip[12]
3690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm3, xmm0          ;ip[4] ip[0]
3790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   xmm0, xmm2          ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1
3990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   xmm3, xmm2          ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1
4090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
4190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm4, xmm0
4290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklqdq  xmm0, xmm3          ;d1 a1
4390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhqdq  xmm4, xmm3          ;c1 b1
4490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movd    xmm7, eax
4590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
4690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm1, xmm4          ;c1 b1
4790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   xmm4, xmm0          ;dl+cl a1+b1 aka op[4] op[0]
4890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   xmm0, xmm1          ;d1-c1 a1-b1 aka op[12] op[8]
4990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
5090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;;;temp output
5190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;;  movdqu  [rdi + 0], xmm4
5290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;;  movdqu  [rdi + 16], xmm3
5390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
5490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
5590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 13 12 11 10 03 02 01 00
5690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ;
5790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 33 32 31 30 23 22 21 20
5890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ;
5990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm3, xmm4          ; 13 12 11 10 03 02 01 00
6090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd xmm4, xmm0          ; 23 03 22 02 21 01 20 00
6190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd xmm3, xmm0          ; 33 13 32 12 31 11 30 10
6290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm1, xmm4          ; 23 03 22 02 21 01 20 00
6390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd xmm4, xmm3          ; 31 21 11 01 30 20 10 00
6490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd xmm1, xmm3          ; 33 23 13 03 32 22 12 02
6590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
6690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pshufd    xmm2, xmm1, 4eh       ;ip[8] ip[12]
6790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm3, xmm4          ;ip[4] ip[0]
6890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
6990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pshufd    xmm7, xmm7, 0       ;03 03 03 03 03 03 03 03
7090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
7190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   xmm4, xmm2          ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1
7290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   xmm3, xmm2          ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1
7390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
7490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm5, xmm4
7590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklqdq  xmm4, xmm3          ;d1 a1
7690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhqdq  xmm5, xmm3          ;c1 b1
7790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
7890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm1, xmm5          ;c1 b1
7990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   xmm5, xmm4          ;dl+cl a1+b1 aka op[4] op[0]
8090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   xmm4, xmm1          ;d1-c1 a1-b1 aka op[12] op[8]
8190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
8290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 13 12 11 10 03 02 01 00
8390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ;
8490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 33 32 31 30 23 22 21 20
8590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ;
8690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm0, xmm5          ; 13 12 11 10 03 02 01 00
8790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd xmm5, xmm4          ; 23 03 22 02 21 01 20 00
8890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd xmm0, xmm4          ; 33 13 32 12 31 11 30 10
8990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa    xmm1, xmm5          ; 23 03 22 02 21 01 20 00
9090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd xmm5, xmm0          ; 31 21 11 01 30 20 10 00
9190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd xmm1, xmm0          ; 33 23 13 03 32 22 12 02
9290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
9390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   xmm5, xmm7
9490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   xmm1, xmm7
9590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
9690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psraw   xmm5, 3
9790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psraw   xmm1, 3
9890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
9990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa  [rdi + 0], xmm5
10090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movdqa  [rdi + 16], xmm1
10190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
10290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; begin epilog
10390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop rdi
10490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop rsi
105f71323e297a928af368937089d3ed71239786f86Andreas Huber    RESTORE_XMM
10690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    UNSHADOW_ARGS
10790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop         rbp
10890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ret
10990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
11090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas HuberSECTION_RODATA
11190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberalign 16
11290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberx_s1sqr2:
11390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    times 4 dw 0x8A8C
11490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberalign 16
11590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberx_c1sqr2less1:
11690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    times 4 dw 0x4E7B
11790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberalign 16
11890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberfours:
11990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    times 4 dw 0x0004
120