sharpyuv: increase precision of gamma<->linear conversion

Change-Id: I261bae3628315bda4ec0dafb8798c7512dd03a36
This commit is contained in:
Maryla 2022-06-01 15:38:44 +02:00
parent 266cbbc511
commit 2d607ee646
8 changed files with 173 additions and 120 deletions

View File

@ -37,6 +37,7 @@ sharpyuv_srcs := \
sharpyuv/sharpyuv.c \
sharpyuv/sharpyuv_csp.c \
sharpyuv/sharpyuv_dsp.c \
sharpyuv/sharpyuv_gamma.c \
sharpyuv/sharpyuv_neon.$(NEON) \
sharpyuv/sharpyuv_sse2.c \

View File

@ -178,6 +178,7 @@ SHARPYUV_OBJS = \
$(DIROBJ)\sharpyuv\sharpyuv.obj \
$(DIROBJ)\sharpyuv\sharpyuv_csp.obj \
$(DIROBJ)\sharpyuv\sharpyuv_dsp.obj \
$(DIROBJ)\sharpyuv\sharpyuv_gamma.obj \
$(DIROBJ)\sharpyuv\sharpyuv_neon.obj \
$(DIROBJ)\sharpyuv\sharpyuv_sse2.obj \

View File

@ -109,6 +109,7 @@ model {
include "sharpyuv.c"
include "sharpyuv_csp.c"
include "sharpyuv_dsp.c"
include "sharpyuv_gamma.c"
include "sharpyuv_neon.c"
include "sharpyuv_sse2.c"
srcDir "src/dec"

View File

@ -129,6 +129,7 @@ SHARPYUV_OBJS = \
sharpyuv/sharpyuv.o \
sharpyuv/sharpyuv_csp.o \
sharpyuv/sharpyuv_dsp.o \
sharpyuv/sharpyuv_gamma.o \
sharpyuv/sharpyuv_neon.o \
sharpyuv/sharpyuv_sse2.o \

View File

@ -22,6 +22,7 @@ libsharpyuv_neon_la_CFLAGS = $(AM_CFLAGS) $(NEON_FLAGS)
libsharpyuv_la_SOURCES =
libsharpyuv_la_SOURCES += sharpyuv_csp.c sharpyuv_csp.h
libsharpyuv_la_SOURCES += sharpyuv_dsp.c sharpyuv_dsp.h
libsharpyuv_la_SOURCES += sharpyuv_gamma.c sharpyuv_gamma.h
libsharpyuv_la_SOURCES += sharpyuv.c sharpyuv.h
libsharpyuv_la_CPPFLAGS = $(AM_CPPFLAGS)

View File

@ -21,6 +21,7 @@
#include "src/webp/types.h"
#include "src/dsp/cpu.h"
#include "sharpyuv/sharpyuv_dsp.h"
#include "sharpyuv/sharpyuv_gamma.h"
//------------------------------------------------------------------------------
// Sharp RGB->YUV conversion
@ -45,100 +46,6 @@ static int GetPrecisionShift(int rgb_bit_depth) {
typedef int16_t fixed_t; // signed type with extra precision for UV
typedef uint16_t fixed_y_t; // unsigned type with extra precision for W
//------------------------------------------------------------------------------
// Code for gamma correction
// Gamma correction compensates loss of resolution during chroma subsampling.
// Size of pre-computed table for converting from gamma to linear.
#define GAMMA_TO_LINEAR_TAB_BITS 10
#define GAMMA_TO_LINEAR_TAB_SIZE (1 << GAMMA_TO_LINEAR_TAB_BITS)
static uint32_t kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE + 2];
// Size of pre-computed table for converting from linear to gamma.
#define LINEAR_TO_GAMMA_TAB_BITS 8
#define LINEAR_TO_GAMMA_TAB_SIZE (1 << LINEAR_TO_GAMMA_TAB_BITS)
static uint32_t kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE + 2];
static const double kGammaF = 1. / 0.45;
#define GAMMA_TO_LINEAR_BITS 14
static volatile int kGammaTablesSOk = 0;
static void InitGammaTablesS(void) {
assert(2 * GAMMA_TO_LINEAR_BITS < 32); // we use uint32_t intermediate values
if (!kGammaTablesSOk) {
int v;
const double a = 0.09929682680944;
const double thresh = 0.018053968510807;
// Precompute gamma to linear table.
{
const double norm = 1. / GAMMA_TO_LINEAR_TAB_SIZE;
const double a_rec = 1. / (1. + a);
const double final_scale = 1 << GAMMA_TO_LINEAR_BITS;
for (v = 0; v <= GAMMA_TO_LINEAR_TAB_SIZE; ++v) {
const double g = norm * v;
double value;
if (g <= thresh * 4.5) {
value = g / 4.5;
} else {
value = pow(a_rec * (g + a), kGammaF);
}
kGammaToLinearTabS[v] = (uint32_t)(value * final_scale + .5);
}
// to prevent small rounding errors to cause read-overflow:
kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE + 1] =
kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE];
}
// Precompute linear to gamma table.
{
const double scale = 1. / LINEAR_TO_GAMMA_TAB_SIZE;
for (v = 0; v <= LINEAR_TO_GAMMA_TAB_SIZE; ++v) {
const double g = scale * v;
double value;
if (g <= thresh) {
value = 4.5 * g;
} else {
value = (1. + a) * pow(g, 1. / kGammaF) - a;
}
kLinearToGammaTabS[v] =
(uint32_t)(GAMMA_TO_LINEAR_TAB_SIZE * value + 0.5);
}
// to prevent small rounding errors to cause read-overflow:
kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE + 1] =
kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE];
}
kGammaTablesSOk = 1;
}
}
static WEBP_INLINE uint32_t FixedPointInterpolation(int v, uint32_t* tab,
int tab_pos_shift,
int tab_value_shift) {
const uint32_t tab_pos = v >> tab_pos_shift;
// fractional part, in 'tab_pos_shift' fixed-point precision
const uint32_t x = v - (tab_pos << tab_pos_shift); // fractional part
// v0 / v1 are in kGammaToLinearBits fixed-point precision (range [0..1])
const uint32_t v0 = tab[tab_pos + 0] << tab_value_shift;
const uint32_t v1 = tab[tab_pos + 1] << tab_value_shift;
// Final interpolation.
const uint32_t v2 = (v1 - v0) * x; // note: v1 >= v0.
const int half = (tab_pos_shift > 0) ? 1 << (tab_pos_shift - 1) : 0;
const uint32_t result = v0 + ((v2 + half) >> tab_pos_shift);
return result;
}
static WEBP_INLINE uint32_t GammaToLinear(int v, int bit_depth) {
const int shift = GAMMA_TO_LINEAR_TAB_BITS - bit_depth;
if (shift > 0) {
return kGammaToLinearTabS[v << shift];
}
return FixedPointInterpolation(v, kGammaToLinearTabS, -shift, 0);
}
static WEBP_INLINE uint32_t LinearToGamma(uint32_t value, int bit_depth) {
const uint32_t v = value << LINEAR_TO_GAMMA_TAB_BITS;
return FixedPointInterpolation(v, kLinearToGammaTabS, GAMMA_TO_LINEAR_BITS,
bit_depth - GAMMA_TO_LINEAR_TAB_BITS);
}
//------------------------------------------------------------------------------
static uint8_t clip_8b(fixed_t v) {
@ -161,13 +68,14 @@ static int RGBToGray(int64_t r, int64_t g, int64_t b) {
return (int)(luma >> YUV_FIX);
}
static uint32_t ScaleDown(int a, int b, int c, int d, int rgb_bit_depth) {
static uint32_t ScaleDown(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
int rgb_bit_depth) {
const int bit_depth = rgb_bit_depth + GetPrecisionShift(rgb_bit_depth);
const uint32_t A = GammaToLinear(a, bit_depth);
const uint32_t B = GammaToLinear(b, bit_depth);
const uint32_t C = GammaToLinear(c, bit_depth);
const uint32_t D = GammaToLinear(d, bit_depth);
return LinearToGamma((A + B + C + D + 2) >> 2, bit_depth);
const uint32_t A = SharpYuvGammaToLinear(a, bit_depth);
const uint32_t B = SharpYuvGammaToLinear(b, bit_depth);
const uint32_t C = SharpYuvGammaToLinear(c, bit_depth);
const uint32_t D = SharpYuvGammaToLinear(d, bit_depth);
return SharpYuvLinearToGamma((A + B + C + D + 2) >> 2, bit_depth);
}
static WEBP_INLINE void UpdateW(const fixed_y_t* src, fixed_y_t* dst, int w,
@ -175,11 +83,11 @@ static WEBP_INLINE void UpdateW(const fixed_y_t* src, fixed_y_t* dst, int w,
const int bit_depth = rgb_bit_depth + GetPrecisionShift(rgb_bit_depth);
int i;
for (i = 0; i < w; ++i) {
const uint32_t R = GammaToLinear(src[0 * w + i], bit_depth);
const uint32_t G = GammaToLinear(src[1 * w + i], bit_depth);
const uint32_t B = GammaToLinear(src[2 * w + i], bit_depth);
const uint32_t R = SharpYuvGammaToLinear(src[0 * w + i], bit_depth);
const uint32_t G = SharpYuvGammaToLinear(src[1 * w + i], bit_depth);
const uint32_t B = SharpYuvGammaToLinear(src[2 * w + i], bit_depth);
const uint32_t Y = RGBToGray(R, G, B);
dst[i] = (fixed_y_t)LinearToGamma(Y, bit_depth);
dst[i] = (fixed_y_t)SharpYuvLinearToGamma(Y, bit_depth);
}
}
@ -227,15 +135,6 @@ static WEBP_INLINE int Shift(int v, int shift) {
return (shift >= 0) ? (v << shift) : (v >> -shift);
}
static WEBP_INLINE fixed_y_t ChangePrecision(uint16_t a, int shift) {
if (shift == 0) return a;
if (shift < 0) {
const int rounding = 1 << (-shift - 1);
return (a + rounding) >> -shift;
}
return ((fixed_y_t)a << shift);
}
static void ImportOneRow(const uint8_t* const r_ptr,
const uint8_t* const g_ptr,
const uint8_t* const b_ptr,
@ -252,13 +151,13 @@ static void ImportOneRow(const uint8_t* const r_ptr,
const int off = i * step;
const int shift = GetPrecisionShift(rgb_bit_depth);
if (rgb_bit_depth == 8) {
dst[i + 0 * w] = ChangePrecision(r_ptr[off], shift);
dst[i + 1 * w] = ChangePrecision(g_ptr[off], shift);
dst[i + 2 * w] = ChangePrecision(b_ptr[off], shift);
dst[i + 0 * w] = Shift(r_ptr[off], shift);
dst[i + 1 * w] = Shift(g_ptr[off], shift);
dst[i + 2 * w] = Shift(b_ptr[off], shift);
} else {
dst[i + 0 * w] = ChangePrecision(((uint16_t*)r_ptr)[off], shift);
dst[i + 1 * w] = ChangePrecision(((uint16_t*)g_ptr)[off], shift);
dst[i + 2 * w] = ChangePrecision(((uint16_t*)b_ptr)[off], shift);
dst[i + 0 * w] = Shift(((uint16_t*)r_ptr)[off], shift);
dst[i + 1 * w] = Shift(((uint16_t*)g_ptr)[off], shift);
dst[i + 2 * w] = Shift(((uint16_t*)b_ptr)[off], shift);
}
}
if (pic_width & 1) { // replicate rightmost pixel
@ -527,7 +426,7 @@ void SharpYuvInit(VP8CPUInfo cpu_info_func) {
SharpYuvInitDsp(cpu_info_func);
if (!initialized) {
InitGammaTablesS();
SharpYuvInitGammaTables();
}
sharpyuv_last_cpuinfo_used = cpu_info_func;

114
sharpyuv/sharpyuv_gamma.c Normal file
View File

@ -0,0 +1,114 @@
// Copyright 2022 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// Gamma correction utilities.
#include "sharpyuv/sharpyuv_gamma.h"
#include <assert.h>
#include <math.h>
#include <stdint.h>
#include "src/webp/types.h"
// Gamma correction compensates loss of resolution during chroma subsampling.
// Size of pre-computed table for converting from gamma to linear.
#define GAMMA_TO_LINEAR_TAB_BITS 10
#define GAMMA_TO_LINEAR_TAB_SIZE (1 << GAMMA_TO_LINEAR_TAB_BITS)
static uint32_t kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE + 2];
#define LINEAR_TO_GAMMA_TAB_BITS 9
#define LINEAR_TO_GAMMA_TAB_SIZE (1 << LINEAR_TO_GAMMA_TAB_BITS)
static uint32_t kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE + 2];
static const double kGammaF = 1. / 0.45;
#define GAMMA_TO_LINEAR_BITS 16
static volatile int kGammaTablesSOk = 0;
void SharpYuvInitGammaTables(void) {
assert(GAMMA_TO_LINEAR_BITS <= 16);
if (!kGammaTablesSOk) {
int v;
const double a = 0.09929682680944;
const double thresh = 0.018053968510807;
const double final_scale = 1 << GAMMA_TO_LINEAR_BITS;
// Precompute gamma to linear table.
{
const double norm = 1. / GAMMA_TO_LINEAR_TAB_SIZE;
const double a_rec = 1. / (1. + a);
for (v = 0; v <= GAMMA_TO_LINEAR_TAB_SIZE; ++v) {
const double g = norm * v;
double value;
if (g <= thresh * 4.5) {
value = g / 4.5;
} else {
value = pow(a_rec * (g + a), kGammaF);
}
kGammaToLinearTabS[v] = (uint32_t)(value * final_scale + .5);
}
// to prevent small rounding errors to cause read-overflow:
kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE + 1] =
kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE];
}
// Precompute linear to gamma table.
{
const double scale = 1. / LINEAR_TO_GAMMA_TAB_SIZE;
for (v = 0; v <= LINEAR_TO_GAMMA_TAB_SIZE; ++v) {
const double g = scale * v;
double value;
if (g <= thresh) {
value = 4.5 * g;
} else {
value = (1. + a) * pow(g, 1. / kGammaF) - a;
}
kLinearToGammaTabS[v] =
(uint32_t)(final_scale * value + 0.5);
}
// to prevent small rounding errors to cause read-overflow:
kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE + 1] =
kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE];
}
kGammaTablesSOk = 1;
}
}
static WEBP_INLINE int Shift(int v, int shift) {
return (shift >= 0) ? (v << shift) : (v >> -shift);
}
static WEBP_INLINE uint32_t FixedPointInterpolation(int v, uint32_t* tab,
int tab_pos_shift_right,
int tab_value_shift) {
const uint32_t tab_pos = Shift(v, -tab_pos_shift_right);
// fractional part, in 'tab_pos_shift' fixed-point precision
const uint32_t x = v - (tab_pos << tab_pos_shift_right); // fractional part
// v0 / v1 are in kGammaToLinearBits fixed-point precision (range [0..1])
const uint32_t v0 = Shift(tab[tab_pos + 0], tab_value_shift);
const uint32_t v1 = Shift(tab[tab_pos + 1], tab_value_shift);
// Final interpolation.
const uint32_t v2 = (v1 - v0) * x; // note: v1 >= v0.
const int half =
(tab_pos_shift_right > 0) ? 1 << (tab_pos_shift_right - 1) : 0;
const uint32_t result = v0 + ((v2 + half) >> tab_pos_shift_right);
return result;
}
uint32_t SharpYuvGammaToLinear(uint16_t v, int bit_depth) {
const int shift = GAMMA_TO_LINEAR_TAB_BITS - bit_depth;
if (shift > 0) {
return kGammaToLinearTabS[v << shift];
}
return FixedPointInterpolation(v, kGammaToLinearTabS, -shift, 0);
}
uint16_t SharpYuvLinearToGamma(uint32_t value, int bit_depth) {
return FixedPointInterpolation(
value, kLinearToGammaTabS,
(GAMMA_TO_LINEAR_BITS - LINEAR_TO_GAMMA_TAB_BITS),
bit_depth - GAMMA_TO_LINEAR_BITS);
}

35
sharpyuv/sharpyuv_gamma.h Normal file
View File

@ -0,0 +1,35 @@
// Copyright 2022 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// Gamma correction utilities.
#ifndef WEBP_SHARPYUV_SHARPYUV_GAMMA_H_
#define WEBP_SHARPYUV_SHARPYUV_GAMMA_H_
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
// Initializes precomputed tables. Must be called once before calling
// SharpYuvGammaToLinear or SharpYuvLinearToGamma.
void SharpYuvInitGammaTables(void);
// Converts a gamma color value on 'bit_depth' bits to a 16 bit linear value.
uint32_t SharpYuvGammaToLinear(uint16_t v, int bit_depth);
// Converts a 16 bit linear color value to a gamma value on 'bit_depth' bits.
uint16_t SharpYuvLinearToGamma(uint32_t value, int bit_depth);
#ifdef __cplusplus
} // extern "C"
#endif
#endif // WEBP_SHARPYUV_SHARPYUV_GAMMA_H_