From 126974b45b3021b24c3500fd3e8d35965b4700cb Mon Sep 17 00:00:00 2001 From: skal Date: Thu, 14 Mar 2013 20:28:29 +0100 Subject: [PATCH] add LUT-free reference code for YUV->RGB conversion. Reported to eventually be 4% on ARM (see https://code.google.com/p/webp/issues/detail?id=134 for details) We might activate it selectively later... Output values is not bitwise the same as the LUT-based version, but difference is only +/-1 at max. Change-Id: I1cc790ff4459885ed2ae2e72f31c5f3740095f07 --- src/dsp/yuv.c | 8 +++- src/dsp/yuv.h | 119 ++++++++++++++++++++++++++++++++++++++++++++------ 2 files changed, 112 insertions(+), 15 deletions(-) diff --git a/src/dsp/yuv.c b/src/dsp/yuv.c index 38895281..f8988bae 100644 --- a/src/dsp/yuv.c +++ b/src/dsp/yuv.c @@ -15,7 +15,7 @@ extern "C" { #endif -enum { YUV_HALF = 1 << (YUV_FIX - 1) }; +#ifdef WEBP_YUV_USE_TABLE int16_t VP8kVToR[256], VP8kUToB[256]; int32_t VP8kVToG[256], VP8kUToG[256]; @@ -62,6 +62,12 @@ void VP8YUVInit(void) { done = 1; } +#else + +void VP8YUVInit(void) {} + +#endif // WEBP_YUV_USE_TABLE + #if defined(__cplusplus) || defined(c_plusplus) } // extern "C" #endif diff --git a/src/dsp/yuv.h b/src/dsp/yuv.h index add167ea..126404b6 100644 --- a/src/dsp/yuv.h +++ b/src/dsp/yuv.h @@ -19,8 +19,9 @@ // G = 1.164 * (Y-16) - 0.813 * (V-128) - 0.391 * (U-128) // B = 1.164 * (Y-16) + 2.018 * (U-128) // where Y is in the [16,235] range, and U/V in the [16,240] range. -// But the common term 1.164 * (Y-16) can be handled as an offset in the -// VP8kClip[] table. So the formulae should be read as: +// In the table-lookup version (WEBP_YUV_USE_TABLE), the common factor +// "1.164 * (Y-16)" can be handled as an offset in the VP8kClip[] table. +// So in this case the formulae should be read as: // R = 1.164 * [Y + 1.371 * (V-128) ] - 18.624 // G = 1.164 * [Y - 0.698 * (V-128) - 0.336 * (U-128)] - 18.624 // B = 1.164 * [Y + 1.733 * (U-128)] - 18.624 @@ -33,6 +34,9 @@ #include "../dec/decode_vp8.h" +// Define the following to use the LUT-based code: +#define WEBP_YUV_USE_TABLE + #if defined(WEBP_EXPERIMENTAL_FEATURES) // Do NOT activate this feature for real compression. This is only experimental! // This flag is for comparison purpose against JPEG's "YUVj" natural colorspace. @@ -51,9 +55,14 @@ extern "C" { #endif enum { YUV_FIX = 16, // fixed-point precision + YUV_HALF = 1 << (YUV_FIX - 1), + YUV_MASK = (256 << YUV_FIX) - 1, YUV_RANGE_MIN = -227, // min value of r/g/b output YUV_RANGE_MAX = 256 + 226 // max value of r/g/b output }; + +#ifdef WEBP_YUV_USE_TABLE + extern int16_t VP8kVToR[256], VP8kUToB[256]; extern int32_t VP8kVToG[256], VP8kUToG[256]; extern uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN]; @@ -69,6 +78,16 @@ static WEBP_INLINE void VP8YuvToRgb(uint8_t y, uint8_t u, uint8_t v, rgb[2] = VP8kClip[y + b_off - YUV_RANGE_MIN]; } +static WEBP_INLINE void VP8YuvToBgr(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const bgr) { + const int r_off = VP8kVToR[v]; + const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX; + const int b_off = VP8kUToB[u]; + bgr[0] = VP8kClip[y + b_off - YUV_RANGE_MIN]; + bgr[1] = VP8kClip[y + g_off - YUV_RANGE_MIN]; + bgr[2] = VP8kClip[y + r_off - YUV_RANGE_MIN]; +} + static WEBP_INLINE void VP8YuvToRgb565(uint8_t y, uint8_t u, uint8_t v, uint8_t* const rgb) { const int r_off = VP8kVToR[v]; @@ -87,12 +106,6 @@ static WEBP_INLINE void VP8YuvToRgb565(uint8_t y, uint8_t u, uint8_t v, #endif } -static WEBP_INLINE void VP8YuvToArgb(uint8_t y, uint8_t u, uint8_t v, - uint8_t* const argb) { - argb[0] = 0xff; - VP8YuvToRgb(y, u, v, argb + 1); -} - static WEBP_INLINE void VP8YuvToRgba4444(uint8_t y, uint8_t u, uint8_t v, uint8_t* const argb) { const int r_off = VP8kVToR[v]; @@ -110,14 +123,92 @@ static WEBP_INLINE void VP8YuvToRgba4444(uint8_t y, uint8_t u, uint8_t v, #endif } +#else // Table-free version (slower on x86) + +// These constants are 16b fixed-point version of ITU-R BT.601 constants +#define kYScale 76309 // 1.164 = 255 / 219 +#define kVToR 104597 // 1.596 = 255 / 112 * 0.701 +#define kUToG 25674 // 0.391 = 255 / 112 * 0.886 * 0.114 / 0.587 +#define kVToG 53278 // 0.813 = 255 / 112 * 0.701 * 0.299 / 0.587 +#define kUToB 132201 // 2.018 = 255 / 112 * 0.886 +#define kRCst (-kYScale * 16 - kVToR * 128 + YUV_HALF) +#define kGCst (-kYScale * 16 + kUToG * 128 + kVToG * 128 + YUV_HALF) +#define kBCst (-kYScale * 16 - kUToB * 128 + YUV_HALF) + +static WEBP_INLINE uint8_t VP8Clip8(int v) { + return ((v & ~YUV_MASK) == 0) ? (uint8_t)(v >> YUV_FIX) + : (v < 0) ? 0u : 255u; +} + +static WEBP_INLINE uint8_t VP8ClipN(int v, int N) { // clip to N bits + return ((v & ~YUV_MASK) == 0) ? (uint8_t)(v >> (YUV_FIX + (8 - N))) + : (v < 0) ? 0u : (255u >> (8 - N)); +} + +static WEBP_INLINE int VP8YUVToR(int y, int v) { + return kYScale * y + kVToR * v + kRCst; +} + +static WEBP_INLINE int VP8YUVToG(int y, int u, int v) { + return kYScale * y - kUToG * u - kVToG * v + kGCst; +} + +static WEBP_INLINE int VP8YUVToB(int y, int u) { + return kYScale * y + kUToB * u + kBCst; +} + +static WEBP_INLINE void VP8YuvToRgb(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const rgb) { + rgb[0] = VP8Clip8(VP8YUVToR(y, v)); + rgb[1] = VP8Clip8(VP8YUVToG(y, u, v)); + rgb[2] = VP8Clip8(VP8YUVToB(y, u)); +} + static WEBP_INLINE void VP8YuvToBgr(uint8_t y, uint8_t u, uint8_t v, uint8_t* const bgr) { - const int r_off = VP8kVToR[v]; - const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX; - const int b_off = VP8kUToB[u]; - bgr[0] = VP8kClip[y + b_off - YUV_RANGE_MIN]; - bgr[1] = VP8kClip[y + g_off - YUV_RANGE_MIN]; - bgr[2] = VP8kClip[y + r_off - YUV_RANGE_MIN]; + bgr[0] = VP8Clip8(VP8YUVToB(y, u)); + bgr[1] = VP8Clip8(VP8YUVToG(y, u, v)); + bgr[2] = VP8Clip8(VP8YUVToR(y, v)); +} + +static WEBP_INLINE void VP8YuvToRgb565(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const rgb) { + const int r = VP8Clip8(VP8YUVToR(y, u)); + const int g = VP8ClipN(VP8YUVToG(y, u, v), 6); + const int b = VP8ClipN(VP8YUVToB(y, v), 5); + const uint8_t rg = (r & 0xf8) | (g >> 3); + const uint8_t gb = (g << 5) | b; +#ifdef WEBP_SWAP_16BIT_CSP + rgb[0] = gb; + rgb[1] = rg; +#else + rgb[0] = rg; + rgb[1] = gb; +#endif +} + +static WEBP_INLINE void VP8YuvToRgba4444(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const argb) { + const int r = VP8Clip8(VP8YUVToR(y, u)); + const int g = VP8ClipN(VP8YUVToG(y, u, v), 4); + const int b = VP8Clip8(VP8YUVToB(y, v)); + const uint8_t rg = (r & 0xf0) | g; + const uint8_t ba = b | 0x0f; // overwrite the lower 4 bits +#ifdef WEBP_SWAP_16BIT_CSP + argb[0] = ba; + argb[1] = rg; +#else + argb[0] = rg; + argb[1] = ba; +#endif +} + +#endif // WEBP_YUV_USE_TABLE + +static WEBP_INLINE void VP8YuvToArgb(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const argb) { + argb[0] = 0xff; + VP8YuvToRgb(y, u, v, argb + 1); } static WEBP_INLINE void VP8YuvToBgra(uint8_t y, uint8_t u, uint8_t v,