| OLD | NEW |
| 1 ; | 1 ; |
| 2 ; jquantf.asm - sample data conversion and quantization (SSE & SSE2) | 2 ; jquantf.asm - sample data conversion and quantization (SSE & 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_float_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col, | 29 ; jsimd_convsamp_float_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col, |
| 30 ; FAST_FLOAT *workspace); | 30 ; FAST_FLOAT *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 ; FAST_FLOAT *workspace | 35 %define workspace ebp+16 ; FAST_FLOAT *workspace |
| 36 | 36 |
| 37 align 16 | 37 align 16 |
| 38 global EXTN(jsimd_convsamp_float_sse2) | 38 global EXTN(jsimd_convsamp_float_sse2) PRIVATE |
| 39 | 39 |
| 40 EXTN(jsimd_convsamp_float_sse2): | 40 EXTN(jsimd_convsamp_float_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 59 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 108 ; GLOBAL(void) | 108 ; GLOBAL(void) |
| 109 ; jsimd_quantize_float_sse2 (JCOEFPTR coef_block, FAST_FLOAT *divisors, | 109 ; jsimd_quantize_float_sse2 (JCOEFPTR coef_block, FAST_FLOAT *divisors, |
| 110 ; FAST_FLOAT *workspace); | 110 ; FAST_FLOAT *workspace); |
| 111 ; | 111 ; |
| 112 | 112 |
| 113 %define coef_block ebp+8 ; JCOEFPTR coef_block | 113 %define coef_block ebp+8 ; JCOEFPTR coef_block |
| 114 %define divisors ebp+12 ; FAST_FLOAT *divisors | 114 %define divisors ebp+12 ; FAST_FLOAT *divisors |
| 115 %define workspace ebp+16 ; FAST_FLOAT *workspace | 115 %define workspace ebp+16 ; FAST_FLOAT *workspace |
| 116 | 116 |
| 117 align 16 | 117 align 16 |
| 118 global EXTN(jsimd_quantize_float_sse2) | 118 global EXTN(jsimd_quantize_float_sse2) PRIVATE |
| 119 | 119 |
| 120 EXTN(jsimd_quantize_float_sse2): | 120 EXTN(jsimd_quantize_float_sse2): |
| 121 push ebp | 121 push ebp |
| 122 mov ebp,esp | 122 mov ebp,esp |
| 123 ; push ebx ; unused | 123 ; push ebx ; unused |
| 124 ; push ecx ; unused | 124 ; push ecx ; unused |
| 125 ; push edx ; need not be preserved | 125 ; push edx ; need not be preserved |
| 126 push esi | 126 push esi |
| 127 push edi | 127 push edi |
| 128 | 128 |
| (...skipping 33 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 162 pop esi | 162 pop esi |
| 163 ; pop edx ; need not be preserved | 163 ; pop edx ; need not be preserved |
| 164 ; pop ecx ; unused | 164 ; pop ecx ; unused |
| 165 ; pop ebx ; unused | 165 ; pop ebx ; unused |
| 166 pop ebp | 166 pop ebp |
| 167 ret | 167 ret |
| 168 | 168 |
| 169 ; For some reason, the OS X linker does not honor the request to align the | 169 ; For some reason, the OS X linker does not honor the request to align the |
| 170 ; segment unless we do this. | 170 ; segment unless we do this. |
| 171 align 16 | 171 align 16 |
| OLD | NEW |