dsp.h revision a2415724fb3466168b2af5b08bd94ba732c0e753
15821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Copyright 2011 Google Inc. All Rights Reserved. 25821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 35821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This code is licensed under the same terms as WebM: 45821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Software License Agreement: http://www.webmproject.org/license/software/ 55821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ 65821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// ----------------------------------------------------------------------------- 75821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 8f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles)// Speed-critical functions. 97d4cd473f85ac64c3747c96c277f9e506a0d2246Torne (Richard Coles)// 102a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)// Author: Skal (pascal.massimino@gmail.com) 117d4cd473f85ac64c3747c96c277f9e506a0d2246Torne (Richard Coles) 125821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#ifndef WEBP_DSP_DSP_H_ 135821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#define WEBP_DSP_DSP_H_ 145821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 152a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)#include "webp/types.h" 165821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 175821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#if defined(__cplusplus) || defined(c_plusplus) 185821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern "C" { 195821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#endif 205821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 215821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//------------------------------------------------------------------------------ 225821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// CPU detection 235821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 245821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) 255821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#define WEBP_MSC_SSE2 // Visual C++ SSE2 targets 265821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#endif 27f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles) 28f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles)#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) 29f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles)#define WEBP_USE_SSE2 30f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles)#endif 31f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles) 325821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__) 335821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//#define WEBP_ANDROID_NEON // Android targets that might support NEON 345821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//#endif 35f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles) 36f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles)#if defined(__ARM_NEON__) || defined(WEBP_ANDROID_NEON) 37f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles)#define WEBP_USE_NEON 38f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles)#endif 39f2477e01787aa58f445919b809d89e252beef54fTorne (Richard Coles) 405821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef enum { 415821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) kSSE2, 425821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) kSSE3, 435821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) kNEON 445821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)} CPUFeature; 455821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// returns true if the CPU supports the feature. 465821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef int (*VP8CPUInfo)(CPUFeature feature); 475821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8CPUInfo VP8GetCPUInfo; 485821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 495821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//------------------------------------------------------------------------------ 505821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Encoding 515821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 525821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Transforms 532a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)// VP8Idct: Does one of two inverse transforms. If do_two is set, the transforms 54116680a4aac90f2aa7413d9095a592090648e557Ben Murdoch// will be done for (ref, in, dst) and (ref + 4, in + 16, dst + 4). 552a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)typedef void (*VP8Idct)(const uint8_t* ref, const int16_t* in, uint8_t* dst, 565821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) int do_two); 575821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef void (*VP8Fdct)(const uint8_t* src, const uint8_t* ref, int16_t* out); 585821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef void (*VP8WHT)(const int16_t* in, int16_t* out); 595821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8Idct VP8ITransform; 602a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)extern VP8Fdct VP8FTransform; 615d1f7b1de12d16ceb2c938c56701a3e8bfa558f7Torne (Richard Coles)extern VP8WHT VP8ITransformWHT; 625821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8WHT VP8FTransformWHT; 635821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Predictions 645821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// *dst is the destination block. *top and *left can be NULL. 652a99a7e74a7f215066514fe81d2bfa6639d9edddTorne (Richard Coles)typedef void (*VP8IntraPreds)(uint8_t *dst, const uint8_t* left, 665821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const uint8_t* top); 675821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef void (*VP8Intra4Preds)(uint8_t *dst, const uint8_t* top); 685821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8Intra4Preds VP8EncPredLuma4; 695821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8IntraPreds VP8EncPredLuma16; 705821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8IntraPreds VP8EncPredChroma8; 715821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 725821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef int (*VP8Metric)(const uint8_t* pix, const uint8_t* ref); 73868fa2fe829687343ffae624259930155e16dbd8Torne (Richard Coles)extern VP8Metric VP8SSE16x16, VP8SSE16x8, VP8SSE8x8, VP8SSE4x4; 745821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef int (*VP8WMetric)(const uint8_t* pix, const uint8_t* ref, 755821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const uint16_t* const weights); 765821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8WMetric VP8TDisto4x4, VP8TDisto16x16; 775821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 78868fa2fe829687343ffae624259930155e16dbd8Torne (Richard Coles)typedef void (*VP8BlockCopy)(const uint8_t* src, uint8_t* dst); 795821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8BlockCopy VP8Copy4x4; 805821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Quantization 815821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)struct VP8Matrix; // forward declaration 825821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16], 835821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) int n, const struct VP8Matrix* const mtx); 845821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)extern VP8QuantizeBlock VP8EncQuantizeBlock; 855821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 865821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Collect histogram for susceptibility calculation and accumulate in histo[]. 87struct VP8Histogram; 88typedef void (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred, 89 int start_block, int end_block, 90 struct VP8Histogram* const histo); 91extern const int VP8DspScan[16 + 4 + 4]; 92extern VP8CHisto VP8CollectHistogram; 93 94void VP8EncDspInit(void); // must be called before using any of the above 95 96//------------------------------------------------------------------------------ 97// Decoding 98 99typedef void (*VP8DecIdct)(const int16_t* coeffs, uint8_t* dst); 100// when doing two transforms, coeffs is actually int16_t[2][16]. 101typedef void (*VP8DecIdct2)(const int16_t* coeffs, uint8_t* dst, int do_two); 102extern VP8DecIdct2 VP8Transform; 103extern VP8DecIdct VP8TransformUV; 104extern VP8DecIdct VP8TransformDC; 105extern VP8DecIdct VP8TransformDCUV; 106extern void (*VP8TransformWHT)(const int16_t* in, int16_t* out); 107 108// *dst is the destination block, with stride BPS. Boundary samples are 109// assumed accessible when needed. 110typedef void (*VP8PredFunc)(uint8_t* dst); 111extern const VP8PredFunc VP8PredLuma16[/* NUM_B_DC_MODES */]; 112extern const VP8PredFunc VP8PredChroma8[/* NUM_B_DC_MODES */]; 113extern const VP8PredFunc VP8PredLuma4[/* NUM_BMODES */]; 114 115// simple filter (only for luma) 116typedef void (*VP8SimpleFilterFunc)(uint8_t* p, int stride, int thresh); 117extern VP8SimpleFilterFunc VP8SimpleVFilter16; 118extern VP8SimpleFilterFunc VP8SimpleHFilter16; 119extern VP8SimpleFilterFunc VP8SimpleVFilter16i; // filter 3 inner edges 120extern VP8SimpleFilterFunc VP8SimpleHFilter16i; 121 122// regular filter (on both macroblock edges and inner edges) 123typedef void (*VP8LumaFilterFunc)(uint8_t* luma, int stride, 124 int thresh, int ithresh, int hev_t); 125typedef void (*VP8ChromaFilterFunc)(uint8_t* u, uint8_t* v, int stride, 126 int thresh, int ithresh, int hev_t); 127// on outer edge 128extern VP8LumaFilterFunc VP8VFilter16; 129extern VP8LumaFilterFunc VP8HFilter16; 130extern VP8ChromaFilterFunc VP8VFilter8; 131extern VP8ChromaFilterFunc VP8HFilter8; 132 133// on inner edge 134extern VP8LumaFilterFunc VP8VFilter16i; // filtering 3 inner edges altogether 135extern VP8LumaFilterFunc VP8HFilter16i; 136extern VP8ChromaFilterFunc VP8VFilter8i; // filtering u and v altogether 137extern VP8ChromaFilterFunc VP8HFilter8i; 138 139// must be called before anything using the above 140void VP8DspInit(void); 141 142//------------------------------------------------------------------------------ 143// WebP I/O 144 145#define FANCY_UPSAMPLING // undefined to remove fancy upsampling support 146 147typedef void (*WebPUpsampleLinePairFunc)( 148 const uint8_t* top_y, const uint8_t* bottom_y, 149 const uint8_t* top_u, const uint8_t* top_v, 150 const uint8_t* cur_u, const uint8_t* cur_v, 151 uint8_t* top_dst, uint8_t* bottom_dst, int len); 152 153#ifdef FANCY_UPSAMPLING 154 155// Fancy upsampling functions to convert YUV to RGB(A) modes 156extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; 157 158// Initializes SSE2 version of the fancy upsamplers. 159void WebPInitUpsamplersSSE2(void); 160 161#endif // FANCY_UPSAMPLING 162 163// Point-sampling methods. 164typedef void (*WebPSampleLinePairFunc)( 165 const uint8_t* top_y, const uint8_t* bottom_y, 166 const uint8_t* u, const uint8_t* v, 167 uint8_t* top_dst, uint8_t* bottom_dst, int len); 168 169extern const WebPSampleLinePairFunc WebPSamplers[/* MODE_LAST */]; 170 171// General function for converting two lines of ARGB or RGBA. 172// 'alpha_is_last' should be true if 0xff000000 is stored in memory as 173// as 0x00, 0x00, 0x00, 0xff (little endian). 174WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last); 175 176// YUV444->RGB converters 177typedef void (*WebPYUV444Converter)(const uint8_t* y, 178 const uint8_t* u, const uint8_t* v, 179 uint8_t* dst, int len); 180 181extern const WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */]; 182 183// Main function to be called 184void WebPInitUpsamplers(void); 185 186//------------------------------------------------------------------------------ 187// Pre-multiply planes with alpha values 188 189// Apply alpha pre-multiply on an rgba, bgra or argb plane of size w * h. 190// alpha_first should be 0 for argb, 1 for rgba or bgra (where alpha is last). 191extern void (*WebPApplyAlphaMultiply)( 192 uint8_t* rgba, int alpha_first, int w, int h, int stride); 193 194// Same, buf specifically for RGBA4444 format 195extern void (*WebPApplyAlphaMultiply4444)( 196 uint8_t* rgba4444, int w, int h, int stride); 197 198// To be called first before using the above. 199void WebPInitPremultiply(void); 200 201void WebPInitPremultiplySSE2(void); // should not be called directly. 202 203//------------------------------------------------------------------------------ 204 205#if defined(__cplusplus) || defined(c_plusplus) 206} // extern "C" 207#endif 208 209#endif /* WEBP_DSP_DSP_H_ */ 210