OLD | NEW |
1 ; | 1 ; |
2 ; jquanti.asm - sample data conversion and quantization (SSE2) | 2 ; jquanti.asm - sample data conversion and quantization (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 17 matching lines...) Expand all Loading... |
28 ; GLOBAL(void) | 28 ; GLOBAL(void) |
29 ; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col, | 29 ; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col, |
30 ; DCTELEM *workspace); | 30 ; DCTELEM *workspace); |
31 ; | 31 ; |
32 | 32 |
33 %define sample_data ebp+8 ; JSAMPARRAY sample_data | 33 %define sample_data ebp+8 ; JSAMPARRAY sample_data |
34 %define start_col ebp+12 ; JDIMENSION start_col | 34 %define start_col ebp+12 ; JDIMENSION start_col |
35 %define workspace ebp+16 ; DCTELEM *workspace | 35 %define workspace ebp+16 ; DCTELEM *workspace |
36 | 36 |
37 align 16 | 37 align 16 |
38 global EXTN(jsimd_convsamp_sse2) | 38 global EXTN(jsimd_convsamp_sse2) PRIVATE |
39 | 39 |
40 EXTN(jsimd_convsamp_sse2): | 40 EXTN(jsimd_convsamp_sse2): |
41 push ebp | 41 push ebp |
42 mov ebp,esp | 42 mov ebp,esp |
43 push ebx | 43 push ebx |
44 ; push ecx ; need not be preserved | 44 ; push ecx ; need not be preserved |
45 ; push edx ; need not be preserved | 45 ; push edx ; need not be preserved |
46 push esi | 46 push esi |
47 push edi | 47 push edi |
48 | 48 |
(...skipping 61 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
110 | 110 |
111 %define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM) | 111 %define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM) |
112 %define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM) | 112 %define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM) |
113 %define SCALE(m,n,b) XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM) | 113 %define SCALE(m,n,b) XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM) |
114 | 114 |
115 %define coef_block ebp+8 ; JCOEFPTR coef_block | 115 %define coef_block ebp+8 ; JCOEFPTR coef_block |
116 %define divisors ebp+12 ; DCTELEM *divisors | 116 %define divisors ebp+12 ; DCTELEM *divisors |
117 %define workspace ebp+16 ; DCTELEM *workspace | 117 %define workspace ebp+16 ; DCTELEM *workspace |
118 | 118 |
119 align 16 | 119 align 16 |
120 global EXTN(jsimd_quantize_sse2) | 120 global EXTN(jsimd_quantize_sse2) PRIVATE |
121 | 121 |
122 EXTN(jsimd_quantize_sse2): | 122 EXTN(jsimd_quantize_sse2): |
123 push ebp | 123 push ebp |
124 mov ebp,esp | 124 mov ebp,esp |
125 ; push ebx ; unused | 125 ; push ebx ; unused |
126 ; push ecx ; unused | 126 ; push ecx ; unused |
127 ; push edx ; need not be preserved | 127 ; push edx ; need not be preserved |
128 push esi | 128 push esi |
129 push edi | 129 push edi |
130 | 130 |
(...skipping 60 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
191 pop esi | 191 pop esi |
192 ; pop edx ; need not be preserved | 192 ; pop edx ; need not be preserved |
193 ; pop ecx ; unused | 193 ; pop ecx ; unused |
194 ; pop ebx ; unused | 194 ; pop ebx ; unused |
195 pop ebp | 195 pop ebp |
196 ret | 196 ret |
197 | 197 |
198 ; For some reason, the OS X linker does not honor the request to align the | 198 ; For some reason, the OS X linker does not honor the request to align the |
199 ; segment unless we do this. | 199 ; segment unless we do this. |
200 align 16 | 200 align 16 |
OLD | NEW |