// Copyright (c) 2012 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
// This webpage shows layout of YV12 and other YUV formats
// http://www.fourcc.org/yuv.php
// The actual conversion is best described here
// http://en.wikipedia.org/wiki/YUV
// An article on optimizing YUV conversion using tables instead of multiplies
// http://lestourtereaux.free.fr/papers/data/yuvrgb.pdf
//
// YV12 is a full plane of Y and a half height, half width chroma planes
// YV16 is a full plane of Y and a full height, half width chroma planes
//
// ARGB pixel format is output, which on little endian is stored as BGRA.
// The alpha is set to 255, allowing the application to use RGBA or RGB32.
#include "media/base/yuv_convert.h"
#include "base/cpu.h"
#include "base/logging.h"
#include "base/memory/scoped_ptr.h"
#include "build/build_config.h"
#include "media/base/simd/convert_rgb_to_yuv.h"
#include "media/base/simd/convert_yuv_to_rgb.h"
#include "media/base/simd/filter_yuv.h"
#if defined(ARCH_CPU_X86_FAMILY)
#if defined(COMPILER_MSVC)
#include <intrin.h>
#else
#include <mmintrin.h>
#endif
#endif
namespace media {
static FilterYUVRowsProc ChooseFilterYUVRowsProc() {
#if defined(ARCH_CPU_X86_FAMILY)
base::CPU cpu;
if (cpu.has_sse2())
return &FilterYUVRows_SSE2;
if (cpu.has_mmx())
return &FilterYUVRows_MMX;
#endif
return &FilterYUVRows_C;
}
static ConvertYUVToRGB32RowProc ChooseConvertYUVToRGB32RowProc() {
#if defined(ARCH_CPU_X86_FAMILY)
base::CPU cpu;
if (cpu.has_sse())
return &ConvertYUVToRGB32Row_SSE;
if (cpu.has_mmx())
return &ConvertYUVToRGB32Row_MMX;
#endif
return &ConvertYUVToRGB32Row_C;
}
static ScaleYUVToRGB32RowProc ChooseScaleYUVToRGB32RowProc() {
#if defined(ARCH_CPU_X86_64)
// Use 64-bits version if possible.
return &ScaleYUVToRGB32Row_SSE2_X64;
#elif defined(ARCH_CPU_X86_FAMILY)
base::CPU cpu;
// Choose the best one on 32-bits system.
if (cpu.has_sse())
return &ScaleYUVToRGB32Row_SSE;
if (cpu.has_mmx())
return &ScaleYUVToRGB32Row_MMX;
#endif // defined(ARCH_CPU_X86_64)
return &ScaleYUVToRGB32Row_C;
}
static ScaleYUVToRGB32RowProc ChooseLinearScaleYUVToRGB32RowProc() {
#if defined(ARCH_CPU_X86_64)
// Use 64-bits version if possible.
return &LinearScaleYUVToRGB32Row_MMX_X64;
#elif defined(ARCH_CPU_X86_FAMILY)
base::CPU cpu;
// 32-bits system.
if (cpu.has_sse())
return &LinearScaleYUVToRGB32Row_SSE;
if (cpu.has_mmx())
return &LinearScaleYUVToRGB32Row_MMX;
#endif // defined(ARCH_CPU_X86_64)
return &LinearScaleYUVToRGB32Row_C;
}
// Empty SIMD registers state after using them.
void EmptyRegisterState() {
#if defined(ARCH_CPU_X86_FAMILY)
static bool checked = false;
static bool has_mmx = false;
if (!checked) {
base::CPU cpu;
has_mmx = cpu.has_mmx();
checked = true;
}
if (has_mmx)
_mm_empty();
#endif
}
// 16.16 fixed point arithmetic
const int kFractionBits = 16;
const int kFractionMax = 1 << kFractionBits;
const int kFractionMask = ((1 << kFractionBits) - 1);
// Scale a frame of YUV to 32 bit ARGB.
void ScaleYUVToRGB32(const uint8* y_buf,
const uint8* u_buf,
const uint8* v_buf,
uint8* rgb_buf,
int source_width,
int source_height,
int width,
int height,
int y_pitch,
int uv_pitch,
int rgb_pitch,
YUVType yuv_type,
Rotate view_rotate,
ScaleFilter filter) {
static FilterYUVRowsProc filter_proc = NULL;
static ConvertYUVToRGB32RowProc convert_proc = NULL;
static ScaleYUVToRGB32RowProc scale_proc = NULL;
static ScaleYUVToRGB32RowProc linear_scale_proc = NULL;
if (!filter_proc)
filter_proc = ChooseFilterYUVRowsProc();
if (!convert_proc)
convert_proc = ChooseConvertYUVToRGB32RowProc();
if (!scale_proc)
scale_proc = ChooseScaleYUVToRGB32RowProc();
if (!linear_scale_proc)
linear_scale_proc = ChooseLinearScaleYUVToRGB32RowProc();
// Handle zero sized sources and destinations.
if ((yuv_type == YV12 && (source_width < 2 || source_height < 2)) ||
(yuv_type == YV16 && (source_width < 2 || source_height < 1)) ||
width == 0 || height == 0)
return;
// 4096 allows 3 buffers to fit in 12k.
// Helps performance on CPU with 16K L1 cache.
// Large enough for 3830x2160 and 30" displays which are 2560x1600.
const int kFilterBufferSize = 4096;
// Disable filtering if the screen is too big (to avoid buffer overflows).
// This should never happen to regular users: they don't have monitors
// wider than 4096 pixels.
// TODO(fbarchard): Allow rotated videos to filter.
if (source_width > kFilterBufferSize || view_rotate)
filter = FILTER_NONE;
unsigned int y_shift = yuv_type;
// Diagram showing origin and direction of source sampling.
// ->0 4<-
// 7 3
//
// 6 5
// ->1 2<-
// Rotations that start at right side of image.
if ((view_rotate == ROTATE_180) ||
(view_rotate == ROTATE_270) ||
(view_rotate == MIRROR_ROTATE_0) ||
(view_rotate == MIRROR_ROTATE_90)) {
y_buf += source_width - 1;
u_buf += source_width / 2 - 1;
v_buf += source_width / 2 - 1;
source_width = -source_width;
}
// Rotations that start at bottom of image.
if ((view_rotate == ROTATE_90) ||
(view_rotate == ROTATE_180) ||
(view_rotate == MIRROR_ROTATE_90) ||
(view_rotate == MIRROR_ROTATE_180)) {
y_buf += (source_height - 1) * y_pitch;
u_buf += ((source_height >> y_shift) - 1) * uv_pitch;
v_buf += ((source_height >> y_shift) - 1) * uv_pitch;
source_height = -source_height;
}
int source_dx = source_width * kFractionMax / width;
if ((view_rotate == ROTATE_90) ||
(view_rotate == ROTATE_270)) {
int tmp = height;
height = width;
width = tmp;
tmp = source_height;
source_height = source_width;
source_width = tmp;
int source_dy = source_height * kFractionMax / height;
source_dx = ((source_dy >> kFractionBits) * y_pitch) << kFractionBits;
if (view_rotate == ROTATE_90) {
y_pitch = -1;
uv_pitch = -1;
source_height = -source_height;
} else {
y_pitch = 1;
uv_pitch = 1;
}
}
// Need padding because FilterRows() will write 1 to 16 extra pixels
// after the end for SSE2 version.
uint8 yuvbuf[16 + kFilterBufferSize * 3 + 16];
uint8* ybuf =
reinterpret_cast<uint8*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15);
uint8* ubuf = ybuf + kFilterBufferSize;
uint8* vbuf = ubuf + kFilterBufferSize;
// TODO(fbarchard): Fixed point math is off by 1 on negatives.
// We take a y-coordinate in [0,1] space in the source image space, and
// transform to a y-coordinate in [0,1] space in the destination image space.
// Note that the coordinate endpoints lie on pixel boundaries, not on pixel
// centers: e.g. a two-pixel-high image will have pixel centers at 0.25 and
// 0.75. The formula is as follows (in fixed-point arithmetic):
// y_dst = dst_height * ((y_src + 0.5) / src_height)
// dst_pixel = clamp([0, dst_height - 1], floor(y_dst - 0.5))
// Implement this here as an accumulator + delta, to avoid expensive math
// in the loop.
int source_y_subpixel_accum =
((kFractionMax / 2) * source_height) / height - (kFractionMax / 2);
int source_y_subpixel_delta = ((1 << kFractionBits) * source_height) / height;
// TODO(fbarchard): Split this into separate function for better efficiency.
for (int y = 0; y < height; ++y) {
uint8* dest_pixel = rgb_buf + y * rgb_pitch;
int source_y_subpixel = source_y_subpixel_accum;
source_y_subpixel_accum += source_y_subpixel_delta;
if (source_y_subpixel < 0)
source_y_subpixel = 0;
else if (source_y_subpixel > ((source_height - 1) << kFractionBits))
source_y_subpixel = (source_height - 1) << kFractionBits;
const uint8* y_ptr = NULL;
const uint8* u_ptr = NULL;
const uint8* v_ptr = NULL;
// Apply vertical filtering if necessary.
// TODO(fbarchard): Remove memcpy when not necessary.
if (filter & media::FILTER_BILINEAR_V) {
int source_y = source_y_subpixel >> kFractionBits;
y_ptr = y_buf + source_y * y_pitch;
u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
// Vertical scaler uses 16.8 fixed point.
int source_y_fraction =
(source_y_subpixel & kFractionMask) >> 8;
if (source_y_fraction != 0) {
filter_proc(ybuf, y_ptr, y_ptr + y_pitch, source_width,
source_y_fraction);
} else {
memcpy(ybuf, y_ptr, source_width);
}
y_ptr = ybuf;
ybuf[source_width] = ybuf[source_width-1];
int uv_source_width = (source_width + 1) / 2;
int source_uv_fraction;
// For formats with half-height UV planes, each even-numbered pixel row
// should not interpolate, since the next row to interpolate from should
// be a duplicate of the current row.
if (y_shift && (source_y & 0x1) == 0)
source_uv_fraction = 0;
else
source_uv_fraction = source_y_fraction;
if (source_uv_fraction != 0) {
filter_proc(ubuf, u_ptr, u_ptr + uv_pitch, uv_source_width,
source_uv_fraction);
filter_proc(vbuf, v_ptr, v_ptr + uv_pitch, uv_source_width,
source_uv_fraction);
} else {
memcpy(ubuf, u_ptr, uv_source_width);
memcpy(vbuf, v_ptr, uv_source_width);
}
u_ptr = ubuf;
v_ptr = vbuf;
ubuf[uv_source_width] = ubuf[uv_source_width - 1];
vbuf[uv_source_width] = vbuf[uv_source_width - 1];
} else {
// Offset by 1/2 pixel for center sampling.
int source_y = (source_y_subpixel + (kFractionMax / 2)) >> kFractionBits;
y_ptr = y_buf + source_y * y_pitch;
u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
}
if (source_dx == kFractionMax) { // Not scaled
convert_proc(y_ptr, u_ptr, v_ptr, dest_pixel, width);
} else {
if (filter & FILTER_BILINEAR_H) {
linear_scale_proc(y_ptr, u_ptr, v_ptr, dest_pixel, width, source_dx);
} else {
scale_proc(y_ptr, u_ptr, v_ptr, dest_pixel, width, source_dx);
}
}
}
EmptyRegisterState();
}
// Scale a frame of YV12 to 32 bit ARGB for a specific rectangle.
void ScaleYUVToRGB32WithRect(const uint8* y_buf,
const uint8* u_buf,
const uint8* v_buf,
uint8* rgb_buf,
int source_width,
int source_height,
int dest_width,
int dest_height,
int dest_rect_left,
int dest_rect_top,
int dest_rect_right,
int dest_rect_bottom,
int y_pitch,
int uv_pitch,
int rgb_pitch) {
static FilterYUVRowsProc filter_proc = NULL;
if (!filter_proc)
filter_proc = ChooseFilterYUVRowsProc();
// This routine doesn't currently support up-scaling.
CHECK_LE(dest_width, source_width);
CHECK_LE(dest_height, source_height);
// Sanity-check the destination rectangle.
DCHECK(dest_rect_left >= 0 && dest_rect_right <= dest_width);
DCHECK(dest_rect_top >= 0 && dest_rect_bottom <= dest_height);
DCHECK(dest_rect_right > dest_rect_left);
DCHECK(dest_rect_bottom > dest_rect_top);
// Fixed-point value of vertical and horizontal scale down factor.
// Values are in the format 16.16.
int y_step = kFractionMax * source_height / dest_height;
int x_step = kFractionMax * source_width / dest_width;
// Determine the coordinates of the rectangle in 16.16 coords.
// NB: Our origin is the *center* of the top/left pixel, NOT its top/left.
// If we're down-scaling by more than a factor of two, we start with a 50%
// fraction to avoid degenerating to point-sampling - we should really just
// fix the fraction at 50% for all pixels in that case.
int source_left = dest_rect_left * x_step;
int source_right = (dest_rect_right - 1) * x_step;
if (x_step < kFractionMax * 2) {
source_left += ((x_step - kFractionMax) / 2);
source_right += ((x_step - kFractionMax) / 2);
} else {
source_left += kFractionMax / 2;
source_right += kFractionMax / 2;
}
int source_top = dest_rect_top * y_step;
if (y_step < kFractionMax * 2) {
source_top += ((y_step - kFractionMax) / 2);
} else {
source_top += kFractionMax / 2;
}
// Determine the parts of the Y, U and V buffers to interpolate.
int source_y_left = source_left >> kFractionBits;
int source_y_right = std::min(
(source_right >> kFractionBits) + 2,
source_width + 1);
int source_uv_left = source_y_left / 2;
int source_uv_right = std::min(
(source_right >> (kFractionBits + 1)) + 2,
(source_width + 1) / 2);
int source_y_width = source_y_right - source_y_left;
int source_uv_width = source_uv_right - source_uv_left;
// Determine number of pixels in each output row.
int dest_rect_width = dest_rect_right - dest_rect_left;
// Intermediate buffer for vertical interpolation.
// 4096 bytes allows 3 buffers to fit in 12k, which fits in a 16K L1 cache,
// and is bigger than most users will generally need.
// The buffer is 16-byte aligned and padded with 16 extra bytes; some of the
// FilterYUVRowProcs have alignment requirements, and the SSE version can
// write up to 16 bytes past the end of the buffer.
const int kFilterBufferSize = 4096;
if (source_width > kFilterBufferSize)
filter_proc = NULL;
uint8 yuv_temp[16 + kFilterBufferSize * 3 + 16];
uint8* y_temp =
reinterpret_cast<uint8*>(
reinterpret_cast<uintptr_t>(yuv_temp + 15) & ~15);
uint8* u_temp = y_temp + kFilterBufferSize;
uint8* v_temp = u_temp + kFilterBufferSize;
// Move to the top-left pixel of output.
rgb_buf += dest_rect_top * rgb_pitch;
rgb_buf += dest_rect_left * 4;
// For each destination row perform interpolation and color space
// conversion to produce the output.
for (int row = dest_rect_top; row < dest_rect_bottom; ++row) {
// Round the fixed-point y position to get the current row.
int source_row = source_top >> kFractionBits;
int source_uv_row = source_row / 2;
DCHECK(source_row < source_height);
// Locate the first row for each plane for interpolation.
const uint8* y0_ptr = y_buf + y_pitch * source_row + source_y_left;
const uint8* u0_ptr = u_buf + uv_pitch * source_uv_row + source_uv_left;
const uint8* v0_ptr = v_buf + uv_pitch * source_uv_row + source_uv_left;
const uint8* y1_ptr = NULL;
const uint8* u1_ptr = NULL;
const uint8* v1_ptr = NULL;
// Locate the second row for interpolation, being careful not to overrun.
if (source_row + 1 >= source_height) {
y1_ptr = y0_ptr;
} else {
y1_ptr = y0_ptr + y_pitch;
}
if (source_uv_row + 1 >= (source_height + 1) / 2) {
u1_ptr = u0_ptr;
v1_ptr = v0_ptr;
} else {
u1_ptr = u0_ptr + uv_pitch;
v1_ptr = v0_ptr + uv_pitch;
}
if (filter_proc) {
// Vertical scaler uses 16.8 fixed point.
int fraction = (source_top & kFractionMask) >> 8;
filter_proc(y_temp + source_y_left, y0_ptr, y1_ptr,
source_y_width, fraction);
filter_proc(u_temp + source_uv_left, u0_ptr, u1_ptr,
source_uv_width, fraction);
filter_proc(v_temp + source_uv_left, v0_ptr, v1_ptr,
source_uv_width, fraction);
// Perform horizontal interpolation and color space conversion.
// TODO(hclam): Use the MMX version after more testing.
LinearScaleYUVToRGB32RowWithRange_C(
y_temp, u_temp, v_temp, rgb_buf,
dest_rect_width, source_left, x_step);
} else {
// If the frame is too large then we linear scale a single row.
LinearScaleYUVToRGB32RowWithRange_C(
y0_ptr, u0_ptr, v0_ptr, rgb_buf,
dest_rect_width, source_left, x_step);
}
// Advance vertically in the source and destination image.
source_top += y_step;
rgb_buf += rgb_pitch;
}
EmptyRegisterState();
}
void ConvertRGB32ToYUV(const uint8* rgbframe,
uint8* yplane,
uint8* uplane,
uint8* vplane,
int width,
int height,
int rgbstride,
int ystride,
int uvstride) {
static void (*convert_proc)(const uint8*, uint8*, uint8*, uint8*,
int, int, int, int, int) = NULL;
if (!convert_proc) {
#if defined(ARCH_CPU_ARM_FAMILY) || defined(ARCH_CPU_MIPS_FAMILY)
// For ARM and MIPS processors, always use C version.
// TODO(hclam): Implement a NEON version.
convert_proc = &ConvertRGB32ToYUV_C;
#else
// TODO(hclam): Switch to SSSE3 version when the cyan problem is solved.
// See: crbug.com/100462
base::CPU cpu;
if (cpu.has_sse2())
convert_proc = &ConvertRGB32ToYUV_SSE2;
else
convert_proc = &ConvertRGB32ToYUV_C;
#endif
}
convert_proc(rgbframe, yplane, uplane, vplane, width, height,
rgbstride, ystride, uvstride);
}
void ConvertRGB24ToYUV(const uint8* rgbframe,
uint8* yplane,
uint8* uplane,
uint8* vplane,
int width,
int height,
int rgbstride,
int ystride,
int uvstride) {
#if defined(ARCH_CPU_ARM_FAMILY) || defined(ARCH_CPU_MIPS_FAMILY)
ConvertRGB24ToYUV_C(rgbframe, yplane, uplane, vplane, width, height,
rgbstride, ystride, uvstride);
#else
static void (*convert_proc)(const uint8*, uint8*, uint8*, uint8*,
int, int, int, int, int) = NULL;
if (!convert_proc) {
base::CPU cpu;
if (cpu.has_ssse3())
convert_proc = &ConvertRGB24ToYUV_SSSE3;
else
convert_proc = &ConvertRGB24ToYUV_C;
}
convert_proc(rgbframe, yplane, uplane, vplane, width, height,
rgbstride, ystride, uvstride);
#endif
}
void ConvertYUY2ToYUV(const uint8* src,
uint8* yplane,
uint8* uplane,
uint8* vplane,
int width,
int height) {
for (int i = 0; i < height / 2; ++i) {
for (int j = 0; j < (width / 2); ++j) {
yplane[0] = src[0];
*uplane = src[1];
yplane[1] = src[2];
*vplane = src[3];
src += 4;
yplane += 2;
uplane++;
vplane++;
}
for (int j = 0; j < (width / 2); ++j) {
yplane[0] = src[0];
yplane[1] = src[2];
src += 4;
yplane += 2;
}
}
}
void ConvertNV21ToYUV(const uint8* src,
uint8* yplane,
uint8* uplane,
uint8* vplane,
int width,
int height) {
int y_plane_size = width * height;
memcpy(yplane, src, y_plane_size);
src += y_plane_size;
int u_plane_size = y_plane_size >> 2;
for (int i = 0; i < u_plane_size; ++i) {
*vplane++ = *src++;
*uplane++ = *src++;
}
}
void ConvertYUVToRGB32(const uint8* yplane,
const uint8* uplane,
const uint8* vplane,
uint8* rgbframe,
int width,
int height,
int ystride,
int uvstride,
int rgbstride,
YUVType yuv_type) {
#if defined(ARCH_CPU_ARM_FAMILY) || defined(ARCH_CPU_MIPS_FAMILY)
ConvertYUVToRGB32_C(yplane, uplane, vplane, rgbframe,
width, height, ystride, uvstride, rgbstride, yuv_type);
#else
static ConvertYUVToRGB32Proc convert_proc = NULL;
if (!convert_proc) {
base::CPU cpu;
if (cpu.has_sse())
convert_proc = &ConvertYUVToRGB32_SSE;
else if (cpu.has_mmx())
convert_proc = &ConvertYUVToRGB32_MMX;
else
convert_proc = &ConvertYUVToRGB32_C;
}
convert_proc(yplane, uplane, vplane, rgbframe,
width, height, ystride, uvstride, rgbstride, yuv_type);
#endif
}
} // namespace media