| 1 |
// Copyright (c) the JPEG XL Project Authors. All rights reserved. |
| 2 |
// |
| 3 |
// Use of this source code is governed by a BSD-style |
| 4 |
// license that can be found in the LICENSE file. |
| 5 |
|
| 6 |
#include "lib/jxl/render_pipeline/stage_xyb.h" |
| 7 |
|
| 8 |
#include "lib/jxl/base/common.h" |
| 9 |
#include "lib/jxl/base/sanitizers.h" |
| 10 |
|
| 11 |
#undef HWY_TARGET_INCLUDE |
| 12 |
#define HWY_TARGET_INCLUDE "lib/jxl/render_pipeline/stage_xyb.cc" |
| 13 |
#include <hwy/foreach_target.h> |
| 14 |
#include <hwy/highway.h> |
| 15 |
|
| 16 |
#include "lib/jxl/cms/opsin_params.h" |
| 17 |
#include "lib/jxl/common.h" // JXL_HIGH_PRECISION |
| 18 |
#include "lib/jxl/dec_xyb-inl.h" |
| 19 |
|
| 20 |
HWY_BEFORE_NAMESPACE(); |
| 21 |
namespace jxl { |
| 22 |
namespace HWY_NAMESPACE { |
| 23 |
|
| 24 |
class XYBStage : public RenderPipelineStage { |
| 25 |
public: |
| 26 |
explicit XYBStage(const OutputEncodingInfo& output_encoding_info) |
| 27 |
: RenderPipelineStage(RenderPipelineStage::Settings()), |
| 28 |
opsin_params_(output_encoding_info.opsin_params), |
| 29 |
output_is_xyb_(output_encoding_info.color_encoding.GetColorSpace() == |
| 30 |
ColorSpace::kXYB) {} |
| 31 |
|
| 32 |
Status ProcessRow(const RowInfo& input_rows, const RowInfo& output_rows, |
| 33 |
size_t xextra, size_t xsize, size_t xpos, size_t ypos, |
| 34 |
size_t thread_id) const final { |
| 35 |
const HWY_FULL(float) d; |
| 36 |
JXL_ENSURE(xextra == 0); |
| 37 |
const size_t xsize_v = RoundUpTo(xsize, Lanes(d)); |
| 38 |
float* JXL_RESTRICT row0 = GetInputRow(input_rows, 0, 0); |
| 39 |
float* JXL_RESTRICT row1 = GetInputRow(input_rows, 1, 0); |
| 40 |
float* JXL_RESTRICT row2 = GetInputRow(input_rows, 2, 0); |
| 41 |
// All calculations are lane-wise, still some might require |
| 42 |
// value-dependent behaviour (e.g. NearestInt). Temporary unpoison last |
| 43 |
// vector tail. |
| 44 |
msan::UnpoisonMemory(row0 + xsize, sizeof(float) * (xsize_v - xsize)); |
| 45 |
msan::UnpoisonMemory(row1 + xsize, sizeof(float) * (xsize_v - xsize)); |
| 46 |
msan::UnpoisonMemory(row2 + xsize, sizeof(float) * (xsize_v - xsize)); |
| 47 |
// TODO(eustas): when using frame origin, addresses might be unaligned; |
| 48 |
// making them aligned will void performance penalty. |
| 49 |
if (output_is_xyb_) { |
| 50 |
const auto scale_x = Set(d, jxl::cms::kScaledXYBScale[0]); |
| 51 |
const auto scale_y = Set(d, jxl::cms::kScaledXYBScale[1]); |
| 52 |
const auto scale_bmy = Set(d, jxl::cms::kScaledXYBScale[2]); |
| 53 |
const auto offset_x = Set(d, jxl::cms::kScaledXYBOffset[0]); |
| 54 |
const auto offset_y = Set(d, jxl::cms::kScaledXYBOffset[1]); |
| 55 |
const auto offset_bmy = Set(d, jxl::cms::kScaledXYBOffset[2]); |
| 56 |
for (ssize_t x = -xextra; x < static_cast<ssize_t>(xsize + xextra); |
| 57 |
x += Lanes(d)) { |
| 58 |
const auto in_x = LoadU(d, row0 + x); |
| 59 |
const auto in_y = LoadU(d, row1 + x); |
| 60 |
const auto in_b = LoadU(d, row2 + x); |
| 61 |
auto out_x = Mul(Add(in_x, offset_x), scale_x); |
| 62 |
auto out_y = Mul(Add(in_y, offset_y), scale_y); |
| 63 |
auto out_b = Mul(Add(Sub(in_b, in_y), offset_bmy), scale_bmy); |
| 64 |
StoreU(out_x, d, row0 + x); |
| 65 |
StoreU(out_y, d, row1 + x); |
| 66 |
StoreU(out_b, d, row2 + x); |
| 67 |
} |
| 68 |
} else { |
| 69 |
for (ssize_t x = -xextra; x < static_cast<ssize_t>(xsize + xextra); |
| 70 |
x += Lanes(d)) { |
| 71 |
const auto in_opsin_x = LoadU(d, row0 + x); |
| 72 |
const auto in_opsin_y = LoadU(d, row1 + x); |
| 73 |
const auto in_opsin_b = LoadU(d, row2 + x); |
| 74 |
auto r = Undefined(d); |
| 75 |
auto g = Undefined(d); |
| 76 |
auto b = Undefined(d); |
| 77 |
XybToRgb(d, in_opsin_x, in_opsin_y, in_opsin_b, opsin_params_, &r, &g, |
| 78 |
&b); |
| 79 |
StoreU(r, d, row0 + x); |
| 80 |
StoreU(g, d, row1 + x); |
| 81 |
StoreU(b, d, row2 + x); |
| 82 |
} |
| 83 |
} |
| 84 |
msan::PoisonMemory(row0 + xsize, sizeof(float) * (xsize_v - xsize)); |
| 85 |
msan::PoisonMemory(row1 + xsize, sizeof(float) * (xsize_v - xsize)); |
| 86 |
msan::PoisonMemory(row2 + xsize, sizeof(float) * (xsize_v - xsize)); |
| 87 |
return true; |
| 88 |
} |
| 89 |
|
| 90 |
RenderPipelineChannelMode GetChannelMode(size_t c) const final { |
| 91 |
return c < 3 ? RenderPipelineChannelMode::kInPlace |
| 92 |
: RenderPipelineChannelMode::kIgnored; |
| 93 |
} |
| 94 |
|
| 95 |
const char* GetName() const override { return "XYB"; } |
| 96 |
|
| 97 |
private: |
| 98 |
const OpsinParams opsin_params_; |
| 99 |
const bool output_is_xyb_; |
| 100 |
}; |
| 101 |
|
| 102 |
std::unique_ptr<RenderPipelineStage> GetXYBStage( |
| 103 |
const OutputEncodingInfo& output_encoding_info) { |
| 104 |
return jxl::make_unique<XYBStage>(output_encoding_info); |
| 105 |
} |
| 106 |
|
| 107 |
// NOLINTNEXTLINE(google-readability-namespace-comments) |
| 108 |
} // namespace HWY_NAMESPACE |
| 109 |
} // namespace jxl |
| 110 |
HWY_AFTER_NAMESPACE(); |
| 111 |
|
| 112 |
#if HWY_ONCE |
| 113 |
namespace jxl { |
| 114 |
|
| 115 |
HWY_EXPORT(GetXYBStage); |
| 116 |
|
| 117 |
std::unique_ptr<RenderPipelineStage> GetXYBStage( |
| 118 |
const OutputEncodingInfo& output_encoding_info) { |
| 119 |
return HWY_DYNAMIC_DISPATCH(GetXYBStage)(output_encoding_info); |
| 120 |
} |
| 121 |
|
| 122 |
#if !JXL_HIGH_PRECISION |
| 123 |
namespace { |
| 124 |
class FastXYBStage : public RenderPipelineStage { |
| 125 |
public: |
| 126 |
FastXYBStage(uint8_t* rgb, size_t stride, size_t width, size_t height, |
| 127 |
bool rgba, bool has_alpha, size_t alpha_c) |
| 128 |
: RenderPipelineStage(RenderPipelineStage::Settings()), |
| 129 |
rgb_(rgb), |
| 130 |
stride_(stride), |
| 131 |
width_(width), |
| 132 |
height_(height), |
| 133 |
rgba_(rgba), |
| 134 |
has_alpha_(has_alpha), |
| 135 |
alpha_c_(alpha_c) {} |
| 136 |
|
| 137 |
Status ProcessRow(const RowInfo& input_rows, const RowInfo& output_rows, |
| 138 |
size_t xextra, size_t xsize, size_t xpos, size_t ypos, |
| 139 |
size_t thread_id) const final { |
| 140 |
if (ypos >= height_) return true; |
| 141 |
JXL_ENSURE(xextra == 0); |
| 142 |
const float* xyba[4] = { |
| 143 |
GetInputRow(input_rows, 0, 0), GetInputRow(input_rows, 1, 0), |
| 144 |
GetInputRow(input_rows, 2, 0), |
| 145 |
has_alpha_ ? GetInputRow(input_rows, alpha_c_, 0) : nullptr}; |
| 146 |
uint8_t* out_buf = rgb_ + stride_ * ypos + (rgba_ ? 4 : 3) * xpos; |
| 147 |
return FastXYBTosRGB8(xyba, out_buf, rgba_, |
| 148 |
xsize + xpos <= width_ ? xsize : width_ - xpos); |
| 149 |
} |
| 150 |
|
| 151 |
RenderPipelineChannelMode GetChannelMode(size_t c) const final { |
| 152 |
return c < 3 || (has_alpha_ && c == alpha_c_) |
| 153 |
? RenderPipelineChannelMode::kInput |
| 154 |
: RenderPipelineChannelMode::kIgnored; |
| 155 |
} |
| 156 |
|
| 157 |
const char* GetName() const override { return "FastXYB"; } |
| 158 |
|
| 159 |
private: |
| 160 |
uint8_t* rgb_; |
| 161 |
size_t stride_; |
| 162 |
size_t width_; |
| 163 |
size_t height_; |
| 164 |
bool rgba_; |
| 165 |
bool has_alpha_; |
| 166 |
size_t alpha_c_; |
| 167 |
std::vector<float> opaque_alpha_; |
| 168 |
}; |
| 169 |
|
| 170 |
} // namespace |
| 171 |
|
| 172 |
std::unique_ptr<RenderPipelineStage> GetFastXYBTosRGB8Stage( |
| 173 |
uint8_t* rgb, size_t stride, size_t width, size_t height, bool rgba, |
| 174 |
bool has_alpha, size_t alpha_c) { |
| 175 |
if (!HasFastXYBTosRGB8()) return nullptr; |
| 176 |
return make_unique<FastXYBStage>(rgb, stride, width, height, rgba, has_alpha, |
| 177 |
alpha_c); |
| 178 |
} |
| 179 |
#endif |
| 180 |
|
| 181 |
} // namespace jxl |
| 182 |
#endif |