OLD | NEW |
1 ; | 1 ; |
2 ; jcclrss2-64.asm - colorspace conversion (64-bit SSE2) | 2 ; jcclrss2-64.asm - colorspace conversion (64-bit SSE2) |
3 ; | 3 ; |
4 ; x86 SIMD extension for IJG JPEG library | 4 ; x86 SIMD extension for IJG JPEG library |
5 ; Copyright (C) 1999-2006, MIYASAKA Masaru. | 5 ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
6 ; Copyright (C) 2009, D. R. Commander. | 6 ; Copyright (C) 2009, D. R. Commander. |
7 ; For conditions of distribution and use, see copyright notice in jsimdext.inc | 7 ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
8 ; | 8 ; |
9 ; This file should be assembled with NASM (Netwide Assembler), | 9 ; This file should be assembled with NASM (Netwide Assembler), |
10 ; can *not* be assembled with Microsoft's MASM or any compatible | 10 ; can *not* be assembled with Microsoft's MASM or any compatible |
(...skipping 21 matching lines...) Expand all Loading... |
32 ; r11 = JSAMPARRAY input_buf | 32 ; r11 = JSAMPARRAY input_buf |
33 ; r12 = JSAMPIMAGE output_buf | 33 ; r12 = JSAMPIMAGE output_buf |
34 ; r13 = JDIMENSION output_row | 34 ; r13 = JDIMENSION output_row |
35 ; r14 = int num_rows | 35 ; r14 = int num_rows |
36 | 36 |
37 %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] | 37 %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] |
38 %define WK_NUM 8 | 38 %define WK_NUM 8 |
39 | 39 |
40 align 16 | 40 align 16 |
41 | 41 |
42 » global» EXTN(jsimd_rgb_ycc_convert_sse2) | 42 » global» EXTN(jsimd_rgb_ycc_convert_sse2) PRIVATE |
43 | 43 |
44 EXTN(jsimd_rgb_ycc_convert_sse2): | 44 EXTN(jsimd_rgb_ycc_convert_sse2): |
45 push rbp | 45 push rbp |
46 mov rax,rsp ; rax = original rbp | 46 mov rax,rsp ; rax = original rbp |
47 sub rsp, byte 4 | 47 sub rsp, byte 4 |
48 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits | 48 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits |
49 mov [rsp],rax | 49 mov [rsp],rax |
50 mov rbp,rsp ; rbp = aligned rbp | 50 mov rbp,rsp ; rbp = aligned rbp |
51 lea rsp, [wk(0)] | 51 lea rsp, [wk(0)] |
52 collect_args | 52 collect_args |
(...skipping 425 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
478 pop rbx | 478 pop rbx |
479 uncollect_args | 479 uncollect_args |
480 mov rsp,rbp ; rsp <- aligned rbp | 480 mov rsp,rbp ; rsp <- aligned rbp |
481 pop rsp ; rsp <- original rbp | 481 pop rsp ; rsp <- original rbp |
482 pop rbp | 482 pop rbp |
483 ret | 483 ret |
484 | 484 |
485 ; For some reason, the OS X linker does not honor the request to align the | 485 ; For some reason, the OS X linker does not honor the request to align the |
486 ; segment unless we do this. | 486 ; segment unless we do this. |
487 align 16 | 487 align 16 |
OLD | NEW |