OLD | NEW |
1 ; | 1 ; |
2 ; jdmrgss2.asm - merged upsampling/color conversion (SSE2) | 2 ; jdmrgss2.asm - merged upsampling/color conversion (SSE2) |
3 ; | 3 ; |
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB | 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
5 ; | 5 ; |
6 ; Based on | 6 ; Based on |
7 ; x86 SIMD extension for IJG JPEG library | 7 ; x86 SIMD extension for IJG JPEG library |
8 ; Copyright (C) 1999-2006, MIYASAKA Masaru. | 8 ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc | 9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
10 ; | 10 ; |
(...skipping 24 matching lines...) Expand all Loading... |
35 %define input_buf(b) (b)+12 ; JSAMPIMAGE input_buf | 35 %define input_buf(b) (b)+12 ; JSAMPIMAGE input_buf |
36 %define in_row_group_ctr(b) (b)+16 ; JDIMENSION in_row_group_ctr | 36 %define in_row_group_ctr(b) (b)+16 ; JDIMENSION in_row_group_ctr |
37 %define output_buf(b) (b)+20 ; JSAMPARRAY output_buf | 37 %define output_buf(b) (b)+20 ; JSAMPARRAY output_buf |
38 | 38 |
39 %define original_ebp ebp+0 | 39 %define original_ebp ebp+0 |
40 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] | 40 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] |
41 %define WK_NUM 3 | 41 %define WK_NUM 3 |
42 %define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr | 42 %define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr |
43 | 43 |
44 align 16 | 44 align 16 |
45 » global» EXTN(jsimd_h2v1_merged_upsample_sse2) | 45 » global» EXTN(jsimd_h2v1_merged_upsample_sse2) PRIVATE |
46 | 46 |
47 EXTN(jsimd_h2v1_merged_upsample_sse2): | 47 EXTN(jsimd_h2v1_merged_upsample_sse2): |
48 push ebp | 48 push ebp |
49 mov eax,esp ; eax = original ebp | 49 mov eax,esp ; eax = original ebp |
50 sub esp, byte 4 | 50 sub esp, byte 4 |
51 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits | 51 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits |
52 mov [esp],eax | 52 mov [esp],eax |
53 mov ebp,esp ; ebp = aligned ebp | 53 mov ebp,esp ; ebp = aligned ebp |
54 lea esp, [wk(0)] | 54 lea esp, [wk(0)] |
55 pushpic eax ; make a room for GOT address | 55 pushpic eax ; make a room for GOT address |
(...skipping 446 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
502 ; JDIMENSION in_row_group_ctr, | 502 ; JDIMENSION in_row_group_ctr, |
503 ; JSAMPARRAY output_buf); | 503 ; JSAMPARRAY output_buf); |
504 ; | 504 ; |
505 | 505 |
506 %define output_width(b) (b)+8 ; JDIMENSION output_width | 506 %define output_width(b) (b)+8 ; JDIMENSION output_width |
507 %define input_buf(b) (b)+12 ; JSAMPIMAGE input_buf | 507 %define input_buf(b) (b)+12 ; JSAMPIMAGE input_buf |
508 %define in_row_group_ctr(b) (b)+16 ; JDIMENSION in_row_group_ctr | 508 %define in_row_group_ctr(b) (b)+16 ; JDIMENSION in_row_group_ctr |
509 %define output_buf(b) (b)+20 ; JSAMPARRAY output_buf | 509 %define output_buf(b) (b)+20 ; JSAMPARRAY output_buf |
510 | 510 |
511 align 16 | 511 align 16 |
512 » global» EXTN(jsimd_h2v2_merged_upsample_sse2) | 512 » global» EXTN(jsimd_h2v2_merged_upsample_sse2) PRIVATE |
513 | 513 |
514 EXTN(jsimd_h2v2_merged_upsample_sse2): | 514 EXTN(jsimd_h2v2_merged_upsample_sse2): |
515 push ebp | 515 push ebp |
516 mov ebp,esp | 516 mov ebp,esp |
517 push ebx | 517 push ebx |
518 ; push ecx ; need not be preserved | 518 ; push ecx ; need not be preserved |
519 ; push edx ; need not be preserved | 519 ; push edx ; need not be preserved |
520 push esi | 520 push esi |
521 push edi | 521 push edi |
522 | 522 |
(...skipping 32 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
555 pop esi | 555 pop esi |
556 ; pop edx ; need not be preserved | 556 ; pop edx ; need not be preserved |
557 ; pop ecx ; need not be preserved | 557 ; pop ecx ; need not be preserved |
558 pop ebx | 558 pop ebx |
559 pop ebp | 559 pop ebp |
560 ret | 560 ret |
561 | 561 |
562 ; For some reason, the OS X linker does not honor the request to align the | 562 ; For some reason, the OS X linker does not honor the request to align the |
563 ; segment unless we do this. | 563 ; segment unless we do this. |
564 align 16 | 564 align 16 |
OLD | NEW |