michael@0: // Copyright (c) 2010 The Chromium Authors. All rights reserved.
michael@0: // Use of this source code is governed by a BSD-style license that can be
michael@0: // found in the LICENSE file.
michael@0: 
michael@0: // This webpage shows layout of YV12 and other YUV formats
michael@0: // http://www.fourcc.org/yuv.php
michael@0: // The actual conversion is best described here
michael@0: // http://en.wikipedia.org/wiki/YUV
michael@0: // An article on optimizing YUV conversion using tables instead of multiplies
michael@0: // http://lestourtereaux.free.fr/papers/data/yuvrgb.pdf
michael@0: //
michael@0: // YV12 is a full plane of Y and a half height, half width chroma planes
michael@0: // YV16 is a full plane of Y and a full height, half width chroma planes
michael@0: // YV24 is a full plane of Y and a full height, full width chroma planes
michael@0: //
michael@0: // ARGB pixel format is output, which on little endian is stored as BGRA.
michael@0: // The alpha is set to 255, allowing the application to use RGBA or RGB32.
michael@0: 
michael@0: #include "yuv_convert.h"
michael@0: 
michael@0: // Header for low level row functions.
michael@0: #include "yuv_row.h"
michael@0: #include "mozilla/SSE.h"
michael@0: 
michael@0: namespace mozilla {
michael@0: 
michael@0: namespace gfx {
michael@0:  
michael@0: // 16.16 fixed point arithmetic
michael@0: const int kFractionBits = 16;
michael@0: const int kFractionMax = 1 << kFractionBits;
michael@0: const int kFractionMask = ((1 << kFractionBits) - 1);
michael@0: 
michael@0: NS_GFX_(YUVType) TypeFromSize(int ywidth, 
michael@0:                               int yheight, 
michael@0:                               int cbcrwidth, 
michael@0:                               int cbcrheight)
michael@0: {
michael@0:   if (ywidth == cbcrwidth && yheight == cbcrheight) {
michael@0:     return YV24;
michael@0:   }
michael@0:   else if (ywidth / 2 == cbcrwidth && yheight == cbcrheight) {
michael@0:     return YV16;
michael@0:   }
michael@0:   else {
michael@0:     return YV12;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // Convert a frame of YUV to 32 bit ARGB.
michael@0: NS_GFX_(void) ConvertYCbCrToRGB32(const uint8* y_buf,
michael@0:                                   const uint8* u_buf,
michael@0:                                   const uint8* v_buf,
michael@0:                                   uint8* rgb_buf,
michael@0:                                   int pic_x,
michael@0:                                   int pic_y,
michael@0:                                   int pic_width,
michael@0:                                   int pic_height,
michael@0:                                   int y_pitch,
michael@0:                                   int uv_pitch,
michael@0:                                   int rgb_pitch,
michael@0:                                   YUVType yuv_type) {
michael@0:   unsigned int y_shift = yuv_type == YV12 ? 1 : 0;
michael@0:   unsigned int x_shift = yuv_type == YV24 ? 0 : 1;
michael@0:   // Test for SSE because the optimized code uses movntq, which is not part of MMX.
michael@0:   bool has_sse = supports_mmx() && supports_sse();
michael@0:   // There is no optimized YV24 SSE routine so we check for this and
michael@0:   // fall back to the C code.
michael@0:   has_sse &= yuv_type != YV24;
michael@0:   bool odd_pic_x = yuv_type != YV24 && pic_x % 2 != 0;
michael@0:   int x_width = odd_pic_x ? pic_width - 1 : pic_width;
michael@0: 
michael@0:   for (int y = pic_y; y < pic_height + pic_y; ++y) {
michael@0:     uint8* rgb_row = rgb_buf + (y - pic_y) * rgb_pitch;
michael@0:     const uint8* y_ptr = y_buf + y * y_pitch + pic_x;
michael@0:     const uint8* u_ptr = u_buf + (y >> y_shift) * uv_pitch + (pic_x >> x_shift);
michael@0:     const uint8* v_ptr = v_buf + (y >> y_shift) * uv_pitch + (pic_x >> x_shift);
michael@0: 
michael@0:     if (odd_pic_x) {
michael@0:       // Handle the single odd pixel manually and use the
michael@0:       // fast routines for the remaining.
michael@0:       FastConvertYUVToRGB32Row_C(y_ptr++,
michael@0:                                  u_ptr++,
michael@0:                                  v_ptr++,
michael@0:                                  rgb_row,
michael@0:                                  1,
michael@0:                                  x_shift);
michael@0:       rgb_row += 4;
michael@0:     }
michael@0: 
michael@0:     if (has_sse) {
michael@0:       FastConvertYUVToRGB32Row(y_ptr,
michael@0:                                u_ptr,
michael@0:                                v_ptr,
michael@0:                                rgb_row,
michael@0:                                x_width);
michael@0:     }
michael@0:     else {
michael@0:       FastConvertYUVToRGB32Row_C(y_ptr,
michael@0:                                  u_ptr,
michael@0:                                  v_ptr,
michael@0:                                  rgb_row,
michael@0:                                  x_width,
michael@0:                                  x_shift);
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   // MMX used for FastConvertYUVToRGB32Row requires emms instruction.
michael@0:   if (has_sse)
michael@0:     EMMS();
michael@0: }
michael@0: 
michael@0: // C version does 8 at a time to mimic MMX code
michael@0: static void FilterRows_C(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
michael@0:                          int source_width, int source_y_fraction) {
michael@0:   int y1_fraction = source_y_fraction;
michael@0:   int y0_fraction = 256 - y1_fraction;
michael@0:   uint8* end = ybuf + source_width;
michael@0:   do {
michael@0:     ybuf[0] = (y0_ptr[0] * y0_fraction + y1_ptr[0] * y1_fraction) >> 8;
michael@0:     ybuf[1] = (y0_ptr[1] * y0_fraction + y1_ptr[1] * y1_fraction) >> 8;
michael@0:     ybuf[2] = (y0_ptr[2] * y0_fraction + y1_ptr[2] * y1_fraction) >> 8;
michael@0:     ybuf[3] = (y0_ptr[3] * y0_fraction + y1_ptr[3] * y1_fraction) >> 8;
michael@0:     ybuf[4] = (y0_ptr[4] * y0_fraction + y1_ptr[4] * y1_fraction) >> 8;
michael@0:     ybuf[5] = (y0_ptr[5] * y0_fraction + y1_ptr[5] * y1_fraction) >> 8;
michael@0:     ybuf[6] = (y0_ptr[6] * y0_fraction + y1_ptr[6] * y1_fraction) >> 8;
michael@0:     ybuf[7] = (y0_ptr[7] * y0_fraction + y1_ptr[7] * y1_fraction) >> 8;
michael@0:     y0_ptr += 8;
michael@0:     y1_ptr += 8;
michael@0:     ybuf += 8;
michael@0:   } while (ybuf < end);
michael@0: }
michael@0: 
michael@0: #ifdef MOZILLA_MAY_SUPPORT_MMX
michael@0: void FilterRows_MMX(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
michael@0:                     int source_width, int source_y_fraction);
michael@0: #endif
michael@0: 
michael@0: #ifdef MOZILLA_MAY_SUPPORT_SSE2
michael@0: void FilterRows_SSE2(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
michael@0:                      int source_width, int source_y_fraction);
michael@0: #endif
michael@0: 
michael@0: static inline void FilterRows(uint8* ybuf, const uint8* y0_ptr,
michael@0:                               const uint8* y1_ptr, int source_width,
michael@0:                               int source_y_fraction) {
michael@0: #ifdef MOZILLA_MAY_SUPPORT_SSE2
michael@0:   if (mozilla::supports_sse2()) {
michael@0:     FilterRows_SSE2(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
michael@0:     return;
michael@0:   }
michael@0: #endif
michael@0: 
michael@0: #ifdef MOZILLA_MAY_SUPPORT_MMX
michael@0:   if (mozilla::supports_mmx()) {
michael@0:     FilterRows_MMX(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
michael@0:     return;
michael@0:   }
michael@0: #endif
michael@0: 
michael@0:   FilterRows_C(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
michael@0: }
michael@0: 
michael@0: 
michael@0: // Scale a frame of YUV to 32 bit ARGB.
michael@0: NS_GFX_(void) ScaleYCbCrToRGB32(const uint8* y_buf,
michael@0:                                 const uint8* u_buf,
michael@0:                                 const uint8* v_buf,
michael@0:                                 uint8* rgb_buf,
michael@0:                                 int source_width,
michael@0:                                 int source_height,
michael@0:                                 int width,
michael@0:                                 int height,
michael@0:                                 int y_pitch,
michael@0:                                 int uv_pitch,
michael@0:                                 int rgb_pitch,
michael@0:                                 YUVType yuv_type,
michael@0:                                 Rotate view_rotate,
michael@0:                                 ScaleFilter filter) {
michael@0:   bool has_mmx = supports_mmx();
michael@0: 
michael@0:   // 4096 allows 3 buffers to fit in 12k.
michael@0:   // Helps performance on CPU with 16K L1 cache.
michael@0:   // Large enough for 3830x2160 and 30" displays which are 2560x1600.
michael@0:   const int kFilterBufferSize = 4096;
michael@0:   // Disable filtering if the screen is too big (to avoid buffer overflows).
michael@0:   // This should never happen to regular users: they don't have monitors
michael@0:   // wider than 4096 pixels.
michael@0:   // TODO(fbarchard): Allow rotated videos to filter.
michael@0:   if (source_width > kFilterBufferSize || view_rotate)
michael@0:     filter = FILTER_NONE;
michael@0: 
michael@0:   unsigned int y_shift = yuv_type == YV12 ? 1 : 0;
michael@0:   // Diagram showing origin and direction of source sampling.
michael@0:   // ->0   4<-
michael@0:   // 7       3
michael@0:   //
michael@0:   // 6       5
michael@0:   // ->1   2<-
michael@0:   // Rotations that start at right side of image.
michael@0:   if ((view_rotate == ROTATE_180) ||
michael@0:       (view_rotate == ROTATE_270) ||
michael@0:       (view_rotate == MIRROR_ROTATE_0) ||
michael@0:       (view_rotate == MIRROR_ROTATE_90)) {
michael@0:     y_buf += source_width - 1;
michael@0:     u_buf += source_width / 2 - 1;
michael@0:     v_buf += source_width / 2 - 1;
michael@0:     source_width = -source_width;
michael@0:   }
michael@0:   // Rotations that start at bottom of image.
michael@0:   if ((view_rotate == ROTATE_90) ||
michael@0:       (view_rotate == ROTATE_180) ||
michael@0:       (view_rotate == MIRROR_ROTATE_90) ||
michael@0:       (view_rotate == MIRROR_ROTATE_180)) {
michael@0:     y_buf += (source_height - 1) * y_pitch;
michael@0:     u_buf += ((source_height >> y_shift) - 1) * uv_pitch;
michael@0:     v_buf += ((source_height >> y_shift) - 1) * uv_pitch;
michael@0:     source_height = -source_height;
michael@0:   }
michael@0: 
michael@0:   // Handle zero sized destination.
michael@0:   if (width == 0 || height == 0)
michael@0:     return;
michael@0:   int source_dx = source_width * kFractionMax / width;
michael@0:   int source_dy = source_height * kFractionMax / height;
michael@0:   int source_dx_uv = source_dx;
michael@0: 
michael@0:   if ((view_rotate == ROTATE_90) ||
michael@0:       (view_rotate == ROTATE_270)) {
michael@0:     int tmp = height;
michael@0:     height = width;
michael@0:     width = tmp;
michael@0:     tmp = source_height;
michael@0:     source_height = source_width;
michael@0:     source_width = tmp;
michael@0:     int original_dx = source_dx;
michael@0:     int original_dy = source_dy;
michael@0:     source_dx = ((original_dy >> kFractionBits) * y_pitch) << kFractionBits;
michael@0:     source_dx_uv = ((original_dy >> kFractionBits) * uv_pitch) << kFractionBits;
michael@0:     source_dy = original_dx;
michael@0:     if (view_rotate == ROTATE_90) {
michael@0:       y_pitch = -1;
michael@0:       uv_pitch = -1;
michael@0:       source_height = -source_height;
michael@0:     } else {
michael@0:       y_pitch = 1;
michael@0:       uv_pitch = 1;
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   // Need padding because FilterRows() will write 1 to 16 extra pixels
michael@0:   // after the end for SSE2 version.
michael@0:   uint8 yuvbuf[16 + kFilterBufferSize * 3 + 16];
michael@0:   uint8* ybuf =
michael@0:       reinterpret_cast<uint8*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15);
michael@0:   uint8* ubuf = ybuf + kFilterBufferSize;
michael@0:   uint8* vbuf = ubuf + kFilterBufferSize;
michael@0:   // TODO(fbarchard): Fixed point math is off by 1 on negatives.
michael@0:   int yscale_fixed = (source_height << kFractionBits) / height;
michael@0: 
michael@0:   // TODO(fbarchard): Split this into separate function for better efficiency.
michael@0:   for (int y = 0; y < height; ++y) {
michael@0:     uint8* dest_pixel = rgb_buf + y * rgb_pitch;
michael@0:     int source_y_subpixel = (y * yscale_fixed);
michael@0:     if (yscale_fixed >= (kFractionMax * 2)) {
michael@0:       source_y_subpixel += kFractionMax / 2;  // For 1/2 or less, center filter.
michael@0:     }
michael@0:     int source_y = source_y_subpixel >> kFractionBits;
michael@0: 
michael@0:     const uint8* y0_ptr = y_buf + source_y * y_pitch;
michael@0:     const uint8* y1_ptr = y0_ptr + y_pitch;
michael@0: 
michael@0:     const uint8* u0_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
michael@0:     const uint8* u1_ptr = u0_ptr + uv_pitch;
michael@0:     const uint8* v0_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
michael@0:     const uint8* v1_ptr = v0_ptr + uv_pitch;
michael@0: 
michael@0:     // vertical scaler uses 16.8 fixed point
michael@0:     int source_y_fraction = (source_y_subpixel & kFractionMask) >> 8;
michael@0:     int source_uv_fraction =
michael@0:         ((source_y_subpixel >> y_shift) & kFractionMask) >> 8;
michael@0: 
michael@0:     const uint8* y_ptr = y0_ptr;
michael@0:     const uint8* u_ptr = u0_ptr;
michael@0:     const uint8* v_ptr = v0_ptr;
michael@0:     // Apply vertical filtering if necessary.
michael@0:     // TODO(fbarchard): Remove memcpy when not necessary.
michael@0:     if (filter & mozilla::gfx::FILTER_BILINEAR_V) {
michael@0:       if (yscale_fixed != kFractionMax &&
michael@0:           source_y_fraction && ((source_y + 1) < source_height)) {
michael@0:         FilterRows(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
michael@0:       } else {
michael@0:         memcpy(ybuf, y0_ptr, source_width);
michael@0:       }
michael@0:       y_ptr = ybuf;
michael@0:       ybuf[source_width] = ybuf[source_width-1];
michael@0:       int uv_source_width = (source_width + 1) / 2;
michael@0:       if (yscale_fixed != kFractionMax &&
michael@0:           source_uv_fraction &&
michael@0:           (((source_y >> y_shift) + 1) < (source_height >> y_shift))) {
michael@0:         FilterRows(ubuf, u0_ptr, u1_ptr, uv_source_width, source_uv_fraction);
michael@0:         FilterRows(vbuf, v0_ptr, v1_ptr, uv_source_width, source_uv_fraction);
michael@0:       } else {
michael@0:         memcpy(ubuf, u0_ptr, uv_source_width);
michael@0:         memcpy(vbuf, v0_ptr, uv_source_width);
michael@0:       }
michael@0:       u_ptr = ubuf;
michael@0:       v_ptr = vbuf;
michael@0:       ubuf[uv_source_width] = ubuf[uv_source_width - 1];
michael@0:       vbuf[uv_source_width] = vbuf[uv_source_width - 1];
michael@0:     }
michael@0:     if (source_dx == kFractionMax) {  // Not scaled
michael@0:       FastConvertYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
michael@0:                                dest_pixel, width);
michael@0:     } else if (filter & FILTER_BILINEAR_H) {
michael@0:         LinearScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
michael@0:                                  dest_pixel, width, source_dx);
michael@0:     } else {
michael@0: // Specialized scalers and rotation.
michael@0: #if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_MSC_VER) && defined(_M_IX86)
michael@0:       if(mozilla::supports_sse()) {
michael@0:         if (width == (source_width * 2)) {
michael@0:           DoubleYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
michael@0:                                   dest_pixel, width);
michael@0:         } else if ((source_dx & kFractionMask) == 0) {
michael@0:           // Scaling by integer scale factor. ie half.
michael@0:           ConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
michael@0:                                    dest_pixel, width,
michael@0:                                    source_dx >> kFractionBits);
michael@0:         } else if (source_dx_uv == source_dx) {  // Not rotated.
michael@0:           ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
michael@0:                              dest_pixel, width, source_dx);
michael@0:         } else {
michael@0:           RotateConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
michael@0:                                          dest_pixel, width,
michael@0:                                          source_dx >> kFractionBits,
michael@0:                                          source_dx_uv >> kFractionBits);
michael@0:         }
michael@0:       }
michael@0:       else {
michael@0:         ScaleYUVToRGB32Row_C(y_ptr, u_ptr, v_ptr,
michael@0:                              dest_pixel, width, source_dx);
michael@0:       }
michael@0: #else
michael@0:       (void)source_dx_uv;
michael@0:       ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
michael@0:                          dest_pixel, width, source_dx);
michael@0: #endif
michael@0:     }
michael@0:   }
michael@0:   // MMX used for FastConvertYUVToRGB32Row and FilterRows requires emms.
michael@0:   if (has_mmx)
michael@0:     EMMS();
michael@0: }
michael@0: 
michael@0: }  // namespace gfx
michael@0: }  // namespace mozilla