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_1_mmx(short *input, short *output)
1590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberglobal sym(vp8_short_inv_walsh4x4_1_mmx)
1690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Hubersym(vp8_short_inv_walsh4x4_1_mmx):
1790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rbp
1890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov         rbp, rsp
1990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    SHADOW_ARGS_TO_STACK 2
2090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rsi
2190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rdi
2290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; end prolog
2390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
2490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rsi, arg(0)
2590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rax, 3
2690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
2790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rdi, arg(1)
2890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    add     rax, [rsi]          ;input[0] + 3
2990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movd    mm0, eax
3190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd mm0, mm0          ;x x val val
3390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckldq mm0, mm0          ;val val val val
3590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psraw   mm0, 3            ;(input[0] + 3) >> 3
3790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 0], mm0
3990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 8], mm0
4090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 16], mm0
4190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 24], mm0
4290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
4390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; begin epilog
4490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop rdi
4590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop rsi
4690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    UNSHADOW_ARGS
4790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop         rbp
4890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ret
4990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
5090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;void vp8_short_inv_walsh4x4_mmx(short *input, short *output)
5190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberglobal sym(vp8_short_inv_walsh4x4_mmx)
5290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Hubersym(vp8_short_inv_walsh4x4_mmx):
5390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rbp
5490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov         rbp, rsp
5590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    SHADOW_ARGS_TO_STACK 2
5690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rsi
5790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push        rdi
5890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; end prolog
5990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
6090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rax, 3
6190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rsi, arg(0)
6290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    mov     rdi, arg(1)
6390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    shl     rax, 16
6490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
6590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm0, [rsi + 0]        ;ip[0]
6690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm1, [rsi + 8]        ;ip[4]
6790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    or      rax, 3            ;00030003h
6890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
6990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm2, [rsi + 16]       ;ip[8]
7090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm3, [rsi + 24]       ;ip[12]
7190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
72538f6170b788de7408b06efc6613dc98579aa6a6Andreas Huber    movq    mm7, rax
7390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm4, mm0
7490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
7590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd mm7, mm7          ;0003000300030003h
7690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm5, mm1
7790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
7890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm4, mm3          ;ip[0] + ip[12] aka al
7990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm5, mm2          ;ip[4] + ip[8] aka bl
8090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm6, mm4          ;temp al
8290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm4, mm5          ;al + bl
8490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm6, mm5          ;al - bl
8590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm0, mm3          ;ip[0] - ip[12] aka d1
8790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm1, mm2          ;ip[4] - ip[8] aka c1
8890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm5, mm0          ;temp dl
9090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
9190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm0, mm1          ;dl + cl
9290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm5, mm1          ;dl - cl
9390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
9490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 03 02 01 00
9590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 13 12 11 10
9690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 23 22 21 20
9790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; 33 32 31 30
9890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
9990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm3, mm4          ; 03 02 01 00
10090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd mm4, mm0          ; 11 01 10 00
10190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd mm3, mm0          ; 13 03 12 02
10290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
10390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm1, mm6          ; 23 22 21 20
10490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd mm6, mm5          ; 31 21 30 20
10590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd mm1, mm5          ; 33 23 32 22
10690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
10790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm0, mm4          ; 11 01 10 00
10890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm2, mm3          ; 13 03 12 02
10990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
11090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckldq mm0, mm6          ; 30 20 10 00 aka ip[0]
11190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhdq mm4, mm6          ; 31 21 11 01 aka ip[4]
11290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
11390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckldq mm2, mm1          ; 32 22 12 02 aka ip[8]
11490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhdq mm3, mm1          ; 33 23 13 03 aka ip[12]
11590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;~~~~~~~~~~~~~~~~~~~~~
11690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm1, mm0
11790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm5, mm4
11890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
11990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm1, mm3          ;ip[0] + ip[12] aka al
12090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm5, mm2          ;ip[4] + ip[8] aka bl
12190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
12290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm6, mm1          ;temp al
12390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
12490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm1, mm5          ;al + bl
12590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm6, mm5          ;al - bl
12690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
12790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm0, mm3          ;ip[0] - ip[12] aka d1
12890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm4, mm2          ;ip[4] - ip[8] aka c1
12990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
13090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm5, mm0          ;temp dl
13190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
13290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm0, mm4          ;dl + cl
13390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psubw   mm5, mm4          ;dl - cl
13490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;~~~~~~~~~~~~~~~~~~~~~
13590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm3, mm1          ; 03 02 01 00
13690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd mm1, mm0          ; 11 01 10 00
13790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd mm3, mm0          ; 13 03 12 02
13890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
13990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm4, mm6          ; 23 22 21 20
14090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpcklwd mm6, mm5          ; 31 21 30 20
14190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhwd mm4, mm5          ; 33 23 32 22
14290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
14390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm0, mm1          ; 11 01 10 00
14490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq    mm2, mm3          ; 13 03 12 02
14590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
14690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckldq mm0, mm6          ; 30 20 10 00 aka ip[0]
14790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhdq mm1, mm6          ; 31 21 11 01 aka ip[4]
14890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
14990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckldq mm2, mm4          ; 32 22 12 02 aka ip[8]
15090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    punpckhdq mm3, mm4          ; 33 23 13 03 aka ip[12]
15190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
15290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm0, mm7
15390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm1, mm7
15490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm2, mm7
15590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    paddw   mm3, mm7
15690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
15790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psraw   mm0, 3
15890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psraw   mm1, 3
15990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psraw   mm2, 3
16090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    psraw   mm3, 3
16190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
16290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 0], mm0
16390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 8], mm1
16490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 16], mm2
16590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    movq  [rdi + 24], mm3
16690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
16790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ; begin epilog
16890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop rdi
16990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop rsi
17090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    UNSHADOW_ARGS
17190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop         rbp
17290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ret
17390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
174