dsp.h revision 8b720228d581a84fd173b6dcb2fa295b59db489a 
1// Copyright 2011 Google Inc. All Rights Reserved.
2//
3// Use of this source code is governed by a BSD-style license
4// that can be found in the COPYING file in the root of the source
5// tree. An additional intellectual property rights grant can be found
6// in the file PATENTS. All contributing project authors may
7// be found in the AUTHORS file in the root of the source tree.
8// -----------------------------------------------------------------------------
9//
10//   Speed-critical functions.
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#ifndef WEBP_DSP_DSP_H_
15#define WEBP_DSP_DSP_H_
16
17#include "webp/types.h"
18
19#ifdef __cplusplus
20extern "C" {
21#endif
22
23//------------------------------------------------------------------------------
24// CPU detection
25
26#if defined(_MSC_VER) && _MSC_VER > 1310 && \
27    (defined(_M_X64) || defined(_M_IX86))
28#define WEBP_MSC_SSE2  // Visual C++ SSE2 targets
29#endif
30
31#if defined(__SSE2__) || defined(WEBP_MSC_SSE2)
32#define WEBP_USE_SSE2
33#endif
34
35#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__)
36#define WEBP_ANDROID_NEON  // Android targets that might support NEON
37#endif
38
39#if defined(__ARM_NEON__)
40#define WEBP_USE_NEON
41#endif
42
43typedef enum {
44  kSSE2,
45  kSSE3,
46  kNEON
47} CPUFeature;
48// returns true if the CPU supports the feature.
49typedef int (*VP8CPUInfo)(CPUFeature feature);
50extern VP8CPUInfo VP8GetCPUInfo;
51
52//------------------------------------------------------------------------------
53// Encoding
54
55// Transforms
56// VP8Idct: Does one of two inverse transforms. If do_two is set, the transforms
57//          will be done for (ref, in, dst) and (ref + 4, in + 16, dst + 4).
58typedef void (*VP8Idct)(const uint8_t* ref, const int16_t* in, uint8_t* dst,
59                        int do_two);
60typedef void (*VP8Fdct)(const uint8_t* src, const uint8_t* ref, int16_t* out);
61typedef void (*VP8WHT)(const int16_t* in, int16_t* out);
62extern VP8Idct VP8ITransform;
63extern VP8Fdct VP8FTransform;
64extern VP8WHT VP8ITransformWHT;
65extern VP8WHT VP8FTransformWHT;
66// Predictions
67// *dst is the destination block. *top and *left can be NULL.
68typedef void (*VP8IntraPreds)(uint8_t *dst, const uint8_t* left,
69                              const uint8_t* top);
70typedef void (*VP8Intra4Preds)(uint8_t *dst, const uint8_t* top);
71extern VP8Intra4Preds VP8EncPredLuma4;
72extern VP8IntraPreds VP8EncPredLuma16;
73extern VP8IntraPreds VP8EncPredChroma8;
74
75typedef int (*VP8Metric)(const uint8_t* pix, const uint8_t* ref);
76extern VP8Metric VP8SSE16x16, VP8SSE16x8, VP8SSE8x8, VP8SSE4x4;
77typedef int (*VP8WMetric)(const uint8_t* pix, const uint8_t* ref,
78                          const uint16_t* const weights);
79extern VP8WMetric VP8TDisto4x4, VP8TDisto16x16;
80
81typedef void (*VP8BlockCopy)(const uint8_t* src, uint8_t* dst);
82extern VP8BlockCopy VP8Copy4x4;
83// Quantization
84struct VP8Matrix;   // forward declaration
85typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16],
86                                int n, const struct VP8Matrix* const mtx);
87extern VP8QuantizeBlock VP8EncQuantizeBlock;
88
89// specific to 2nd transform:
90typedef int (*VP8QuantizeBlockWHT)(int16_t in[16], int16_t out[16],
91                                   const struct VP8Matrix* const mtx);
92extern VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
93
94// Collect histogram for susceptibility calculation and accumulate in histo[].
95struct VP8Histogram;
96typedef void (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred,
97                          int start_block, int end_block,
98                          struct VP8Histogram* const histo);
99extern const int VP8DspScan[16 + 4 + 4];
100extern VP8CHisto VP8CollectHistogram;
101
102void VP8EncDspInit(void);   // must be called before using any of the above
103
104//------------------------------------------------------------------------------
105// Decoding
106
107typedef void (*VP8DecIdct)(const int16_t* coeffs, uint8_t* dst);
108// when doing two transforms, coeffs is actually int16_t[2][16].
109typedef void (*VP8DecIdct2)(const int16_t* coeffs, uint8_t* dst, int do_two);
110extern VP8DecIdct2 VP8Transform;
111extern VP8DecIdct VP8TransformAC3;
112extern VP8DecIdct VP8TransformUV;
113extern VP8DecIdct VP8TransformDC;
114extern VP8DecIdct VP8TransformDCUV;
115extern VP8WHT VP8TransformWHT;
116
117// *dst is the destination block, with stride BPS. Boundary samples are
118// assumed accessible when needed.
119typedef void (*VP8PredFunc)(uint8_t* dst);
120extern const VP8PredFunc VP8PredLuma16[/* NUM_B_DC_MODES */];
121extern const VP8PredFunc VP8PredChroma8[/* NUM_B_DC_MODES */];
122extern const VP8PredFunc VP8PredLuma4[/* NUM_BMODES */];
123
124// simple filter (only for luma)
125typedef void (*VP8SimpleFilterFunc)(uint8_t* p, int stride, int thresh);
126extern VP8SimpleFilterFunc VP8SimpleVFilter16;
127extern VP8SimpleFilterFunc VP8SimpleHFilter16;
128extern VP8SimpleFilterFunc VP8SimpleVFilter16i;  // filter 3 inner edges
129extern VP8SimpleFilterFunc VP8SimpleHFilter16i;
130
131// regular filter (on both macroblock edges and inner edges)
132typedef void (*VP8LumaFilterFunc)(uint8_t* luma, int stride,
133                                  int thresh, int ithresh, int hev_t);
134typedef void (*VP8ChromaFilterFunc)(uint8_t* u, uint8_t* v, int stride,
135                                    int thresh, int ithresh, int hev_t);
136// on outer edge
137extern VP8LumaFilterFunc VP8VFilter16;
138extern VP8LumaFilterFunc VP8HFilter16;
139extern VP8ChromaFilterFunc VP8VFilter8;
140extern VP8ChromaFilterFunc VP8HFilter8;
141
142// on inner edge
143extern VP8LumaFilterFunc VP8VFilter16i;   // filtering 3 inner edges altogether
144extern VP8LumaFilterFunc VP8HFilter16i;
145extern VP8ChromaFilterFunc VP8VFilter8i;  // filtering u and v altogether
146extern VP8ChromaFilterFunc VP8HFilter8i;
147
148// must be called before anything using the above
149void VP8DspInit(void);
150
151//------------------------------------------------------------------------------
152// WebP I/O
153
154#define FANCY_UPSAMPLING   // undefined to remove fancy upsampling support
155
156// Convert a pair of y/u/v lines together to the output rgb/a colorspace.
157// bottom_y can be NULL if only one line of output is needed (at top/bottom).
158typedef void (*WebPUpsampleLinePairFunc)(
159    const uint8_t* top_y, const uint8_t* bottom_y,
160    const uint8_t* top_u, const uint8_t* top_v,
161    const uint8_t* cur_u, const uint8_t* cur_v,
162    uint8_t* top_dst, uint8_t* bottom_dst, int len);
163
164#ifdef FANCY_UPSAMPLING
165
166// Fancy upsampling functions to convert YUV to RGB(A) modes
167extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
168
169// Initializes SSE2 version of the fancy upsamplers.
170void WebPInitUpsamplersSSE2(void);
171
172// NEON version
173void WebPInitUpsamplersNEON(void);
174
175#endif    // FANCY_UPSAMPLING
176
177// Point-sampling methods.
178typedef void (*WebPSampleLinePairFunc)(
179    const uint8_t* top_y, const uint8_t* bottom_y,
180    const uint8_t* u, const uint8_t* v,
181    uint8_t* top_dst, uint8_t* bottom_dst, int len);
182
183extern const WebPSampleLinePairFunc WebPSamplers[/* MODE_LAST */];
184
185// General function for converting two lines of ARGB or RGBA.
186// 'alpha_is_last' should be true if 0xff000000 is stored in memory as
187// as 0x00, 0x00, 0x00, 0xff (little endian).
188WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last);
189
190// YUV444->RGB converters
191typedef void (*WebPYUV444Converter)(const uint8_t* y,
192                                    const uint8_t* u, const uint8_t* v,
193                                    uint8_t* dst, int len);
194
195extern const WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
196
197// Main function to be called
198void WebPInitUpsamplers(void);
199
200//------------------------------------------------------------------------------
201// Pre-multiply planes with alpha values
202
203// Apply alpha pre-multiply on an rgba, bgra or argb plane of size w * h.
204// alpha_first should be 0 for argb, 1 for rgba or bgra (where alpha is last).
205extern void (*WebPApplyAlphaMultiply)(
206    uint8_t* rgba, int alpha_first, int w, int h, int stride);
207
208// Same, buf specifically for RGBA4444 format
209extern void (*WebPApplyAlphaMultiply4444)(
210    uint8_t* rgba4444, int w, int h, int stride);
211
212// To be called first before using the above.
213void WebPInitPremultiply(void);
214
215void WebPInitPremultiplySSE2(void);   // should not be called directly.
216void WebPInitPremultiplyNEON(void);
217
218//------------------------------------------------------------------------------
219
220#ifdef __cplusplus
221}    // extern "C"
222#endif
223
224#endif  /* WEBP_DSP_DSP_H_ */
225