blob: 8e78eabc1b355b5cddb191eb46b2d919b6506554 [file] [log] [blame]
Pierre Ossman82c7f312009-03-09 13:21:27 +00001/*
2 * simd/jsimd.h
3 *
4 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 *
6 * Based on the x86 SIMD extension for IJG JPEG library,
7 * Copyright (C) 1999-2006, MIYASAKA Masaru.
8 *
9 */
10
11/* Bitmask for supported acceleration methods */
12
13#define JSIMD_NONE 0x00
Pierre Ossman3e0e2de2009-03-09 13:25:30 +000014#define JSIMD_MMX 0x01
Pierre Ossman2c2e54b2009-03-09 13:28:10 +000015#define JSIMD_3DNOW 0x02
Pierre Ossman82c7f312009-03-09 13:21:27 +000016
17/* Short forms of external names for systems with brain-damaged linkers. */
18
19#ifdef NEED_SHORT_EXTERNAL_NAMES
Pierre Ossman3e0e2de2009-03-09 13:25:30 +000020#define jpeg_simd_cpu_support jSiCpuSupport
21#define jsimd_rgb_ycc_convert_mmx jSRGBYCCM
22#define jsimd_ycc_rgb_convert_mmx jSYCCRGBM
23#define jsimd_h2v2_downsample_mmx jSDnH2V2M
24#define jsimd_h2v1_downsample_mmx jSDnH2V1M
25#define jsimd_h2v2_upsample_mmx jSUpH2V2M
26#define jsimd_h2v1_upsample_mmx jSUpH2V1M
27#define jsimd_h2v2_fancy_upsample_mmx jSFUpH2V2M
28#define jsimd_h2v1_fancy_upsample_mmx jSFUpH2V1M
29#define jsimd_h2v2_merged_upsample_mmx jSMUpH2V2M
30#define jsimd_h2v1_merged_upsample_mmx jSMUpH2V1M
31#define jsimd_convsamp_mmx jSConvM
Pierre Ossman2c2e54b2009-03-09 13:28:10 +000032#define jsimd_convsamp_float_3dnow jSConvF3D
Pierre Ossman3e0e2de2009-03-09 13:25:30 +000033#define jsimd_fdct_islow_mmx jSFDMIS
34#define jsimd_fdct_ifast_mmx jSFDMIF
Pierre Ossman2c2e54b2009-03-09 13:28:10 +000035#define jsimd_fdct_float_3dnow jSFD3DF
Pierre Ossman3e0e2de2009-03-09 13:25:30 +000036#define jsimd_quantize_mmx jSQuantM
Pierre Ossman2c2e54b2009-03-09 13:28:10 +000037#define jsimd_quantize_float_3dnow jSQuantF3D
Pierre Ossman3e0e2de2009-03-09 13:25:30 +000038#define jsimd_idct_2x2_mmx jSIDM22
39#define jsimd_idct_4x4_mmx jSIDM44
40#define jsimd_idct_islow_mmx jSIDMIS
41#define jsimd_idct_ifast_mmx jSIDMIF
Pierre Ossman2c2e54b2009-03-09 13:28:10 +000042#define jsimd_idct_float_3dnow jSID3DF
Pierre Ossman82c7f312009-03-09 13:21:27 +000043#endif /* NEED_SHORT_EXTERNAL_NAMES */
44
45/* SIMD Ext: retrieve SIMD/CPU information */
46EXTERN(unsigned int) jpeg_simd_cpu_support JPP((void));
47
Pierre Ossman3e0e2de2009-03-09 13:25:30 +000048/* SIMD Color Space Conversion */
49EXTERN(void) jsimd_rgb_ycc_convert_mmx
50 JPP((JDIMENSION img_width,
51 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
52 JDIMENSION output_row, int num_rows));
53EXTERN(void) jsimd_ycc_rgb_convert_mmx
54 JPP((JDIMENSION out_width,
55 JSAMPIMAGE input_buf, JDIMENSION input_row,
56 JSAMPARRAY output_buf, int num_rows));
57
58/* SIMD Downsample */
59EXTERN(void) jsimd_h2v2_downsample_mmx
60 JPP((JDIMENSION image_width, int max_v_samp_factor,
61 JDIMENSION v_samp_factor, JDIMENSION width_blocks,
62 JSAMPARRAY input_data, JSAMPARRAY output_data));
63EXTERN(void) jsimd_h2v1_downsample_mmx
64 JPP((JDIMENSION image_width, int max_v_samp_factor,
65 JDIMENSION v_samp_factor, JDIMENSION width_blocks,
66 JSAMPARRAY input_data, JSAMPARRAY output_data));
67
68/* SIMD Upsample */
69EXTERN(void) jsimd_h2v2_upsample_mmx
70 JPP((int max_v_samp_factor, JDIMENSION output_width,
71 JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
72EXTERN(void) jsimd_h2v1_upsample_mmx
73 JPP((int max_v_samp_factor, JDIMENSION output_width,
74 JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
75
76EXTERN(void) jsimd_h2v2_fancy_upsample_mmx
77 JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
78 JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
79EXTERN(void) jsimd_h2v1_fancy_upsample_mmx
80 JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
81 JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
82
83EXTERN(void) jsimd_h2v2_merged_upsample_mmx
84 JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
85 JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
86EXTERN(void) jsimd_h2v1_merged_upsample_mmx
87 JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
88 JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
89
90/* SIMD Sample Conversion */
91EXTERN(void) jsimd_convsamp_mmx JPP((JSAMPARRAY sample_data,
92 JDIMENSION start_col,
93 DCTELEM * workspace));
94
Pierre Ossman2c2e54b2009-03-09 13:28:10 +000095EXTERN(void) jsimd_convsamp_float_3dnow JPP((JSAMPARRAY sample_data,
96 JDIMENSION start_col,
97 FAST_FLOAT * workspace));
98
Pierre Ossman3e0e2de2009-03-09 13:25:30 +000099/* SIMD Forward DCT */
100EXTERN(void) jsimd_fdct_islow_mmx JPP((DCTELEM * data));
101EXTERN(void) jsimd_fdct_ifast_mmx JPP((DCTELEM * data));
102
Pierre Ossman2c2e54b2009-03-09 13:28:10 +0000103EXTERN(void) jsimd_fdct_float_3dnow JPP((FAST_FLOAT * data));
104
Pierre Ossman3e0e2de2009-03-09 13:25:30 +0000105/* SIMD Quantization */
106EXTERN(void) jsimd_quantize_mmx JPP((JCOEFPTR coef_block,
107 DCTELEM * divisors,
108 DCTELEM * workspace));
109
Pierre Ossman2c2e54b2009-03-09 13:28:10 +0000110EXTERN(void) jsimd_quantize_float_3dnow JPP((JCOEFPTR coef_block,
111 FAST_FLOAT * divisors,
112 FAST_FLOAT * workspace));
113
Pierre Ossman3e0e2de2009-03-09 13:25:30 +0000114/* SIMD Reduced Inverse DCT */
115EXTERN(void) jsimd_idct_2x2_mmx JPP((void * dct_table,
116 JCOEFPTR coef_block,
117 JSAMPARRAY output_buf,
118 JDIMENSION output_col));
119EXTERN(void) jsimd_idct_4x4_mmx JPP((void * dct_table,
120 JCOEFPTR coef_block,
121 JSAMPARRAY output_buf,
122 JDIMENSION output_col));
123
124/* SIMD Inverse DCT */
125EXTERN(void) jsimd_idct_islow_mmx JPP((void * dct_table,
126 JCOEFPTR coef_block,
127 JSAMPARRAY output_buf,
128 JDIMENSION output_col));
129EXTERN(void) jsimd_idct_ifast_mmx JPP((void * dct_table,
130 JCOEFPTR coef_block,
131 JSAMPARRAY output_buf,
132 JDIMENSION output_col));
133
Pierre Ossman2c2e54b2009-03-09 13:28:10 +0000134EXTERN(void) jsimd_idct_float_3dnow JPP((void * dct_table,
135 JCOEFPTR coef_block,
136 JSAMPARRAY output_buf,
137 JDIMENSION output_col));
138