compressed_dc.cc (11019B)
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/compressed_dc.h" 7 8 #include <stdint.h> 9 #include <stdlib.h> 10 #include <string.h> 11 12 #include <algorithm> 13 #include <vector> 14 15 #undef HWY_TARGET_INCLUDE 16 #define HWY_TARGET_INCLUDE "lib/jxl/compressed_dc.cc" 17 #include <hwy/aligned_allocator.h> 18 #include <hwy/foreach_target.h> 19 #include <hwy/highway.h> 20 21 #include "lib/jxl/base/compiler_specific.h" 22 #include "lib/jxl/base/data_parallel.h" 23 #include "lib/jxl/base/status.h" 24 #include "lib/jxl/image.h" 25 HWY_BEFORE_NAMESPACE(); 26 namespace jxl { 27 namespace HWY_NAMESPACE { 28 29 using D = HWY_FULL(float); 30 using DScalar = HWY_CAPPED(float, 1); 31 32 // These templates are not found via ADL. 33 using hwy::HWY_NAMESPACE::Abs; 34 using hwy::HWY_NAMESPACE::Add; 35 using hwy::HWY_NAMESPACE::Div; 36 using hwy::HWY_NAMESPACE::Max; 37 using hwy::HWY_NAMESPACE::Mul; 38 using hwy::HWY_NAMESPACE::MulAdd; 39 using hwy::HWY_NAMESPACE::Rebind; 40 using hwy::HWY_NAMESPACE::Sub; 41 using hwy::HWY_NAMESPACE::Vec; 42 using hwy::HWY_NAMESPACE::ZeroIfNegative; 43 44 // TODO(veluca): optimize constants. 45 const float w1 = 0.20345139757231578f; 46 const float w2 = 0.0334829185968739f; 47 const float w0 = 1.0f - 4.0f * (w1 + w2); 48 49 template <class V> 50 V MaxWorkaround(V a, V b) { 51 #if (HWY_TARGET == HWY_AVX3) && HWY_COMPILER_CLANG <= 800 52 // Prevents "Do not know how to split the result of this operator" error 53 return IfThenElse(a > b, a, b); 54 #else 55 return Max(a, b); 56 #endif 57 } 58 59 template <typename D> 60 JXL_INLINE void ComputePixelChannel(const D d, const float dc_factor, 61 const float* JXL_RESTRICT row_top, 62 const float* JXL_RESTRICT row, 63 const float* JXL_RESTRICT row_bottom, 64 Vec<D>* JXL_RESTRICT mc, 65 Vec<D>* JXL_RESTRICT sm, 66 Vec<D>* JXL_RESTRICT gap, size_t x) { 67 const auto tl = LoadU(d, row_top + x - 1); 68 const auto tc = Load(d, row_top + x); 69 const auto tr = LoadU(d, row_top + x + 1); 70 71 const auto ml = LoadU(d, row + x - 1); 72 *mc = Load(d, row + x); 73 const auto mr = LoadU(d, row + x + 1); 74 75 const auto bl = LoadU(d, row_bottom + x - 1); 76 const auto bc = Load(d, row_bottom + x); 77 const auto br = LoadU(d, row_bottom + x + 1); 78 79 const auto w_center = Set(d, w0); 80 const auto w_side = Set(d, w1); 81 const auto w_corner = Set(d, w2); 82 83 const auto corner = Add(Add(tl, tr), Add(bl, br)); 84 const auto side = Add(Add(ml, mr), Add(tc, bc)); 85 *sm = MulAdd(corner, w_corner, MulAdd(side, w_side, Mul(*mc, w_center))); 86 87 const auto dc_quant = Set(d, dc_factor); 88 *gap = MaxWorkaround(*gap, Abs(Div(Sub(*mc, *sm), dc_quant))); 89 } 90 91 template <typename D> 92 JXL_INLINE void ComputePixel( 93 const float* JXL_RESTRICT dc_factors, 94 const float* JXL_RESTRICT* JXL_RESTRICT rows_top, 95 const float* JXL_RESTRICT* JXL_RESTRICT rows, 96 const float* JXL_RESTRICT* JXL_RESTRICT rows_bottom, 97 float* JXL_RESTRICT* JXL_RESTRICT out_rows, size_t x) { 98 const D d; 99 auto mc_x = Undefined(d); 100 auto mc_y = Undefined(d); 101 auto mc_b = Undefined(d); 102 auto sm_x = Undefined(d); 103 auto sm_y = Undefined(d); 104 auto sm_b = Undefined(d); 105 auto gap = Set(d, 0.5f); 106 ComputePixelChannel(d, dc_factors[0], rows_top[0], rows[0], rows_bottom[0], 107 &mc_x, &sm_x, &gap, x); 108 ComputePixelChannel(d, dc_factors[1], rows_top[1], rows[1], rows_bottom[1], 109 &mc_y, &sm_y, &gap, x); 110 ComputePixelChannel(d, dc_factors[2], rows_top[2], rows[2], rows_bottom[2], 111 &mc_b, &sm_b, &gap, x); 112 auto factor = MulAdd(Set(d, -4.0f), gap, Set(d, 3.0f)); 113 factor = ZeroIfNegative(factor); 114 115 auto out = MulAdd(Sub(sm_x, mc_x), factor, mc_x); 116 Store(out, d, out_rows[0] + x); 117 out = MulAdd(Sub(sm_y, mc_y), factor, mc_y); 118 Store(out, d, out_rows[1] + x); 119 out = MulAdd(Sub(sm_b, mc_b), factor, mc_b); 120 Store(out, d, out_rows[2] + x); 121 } 122 123 Status AdaptiveDCSmoothing(const float* dc_factors, Image3F* dc, 124 ThreadPool* pool) { 125 const size_t xsize = dc->xsize(); 126 const size_t ysize = dc->ysize(); 127 if (ysize <= 2 || xsize <= 2) return true; 128 129 // TODO(veluca): use tile-based processing? 130 // TODO(veluca): decide if changes to the y channel should be propagated to 131 // the x and b channels through color correlation. 132 JXL_ASSERT(w1 + w2 < 0.25f); 133 134 JXL_ASSIGN_OR_RETURN(Image3F smoothed, Image3F::Create(xsize, ysize)); 135 // Fill in borders that the loop below will not. First and last are unused. 136 for (size_t c = 0; c < 3; c++) { 137 for (size_t y : {static_cast<size_t>(0), ysize - 1}) { 138 memcpy(smoothed.PlaneRow(c, y), dc->PlaneRow(c, y), 139 xsize * sizeof(float)); 140 } 141 } 142 auto process_row = [&](const uint32_t y, size_t /*thread*/) { 143 const float* JXL_RESTRICT rows_top[3]{ 144 dc->ConstPlaneRow(0, y - 1), 145 dc->ConstPlaneRow(1, y - 1), 146 dc->ConstPlaneRow(2, y - 1), 147 }; 148 const float* JXL_RESTRICT rows[3] = { 149 dc->ConstPlaneRow(0, y), 150 dc->ConstPlaneRow(1, y), 151 dc->ConstPlaneRow(2, y), 152 }; 153 const float* JXL_RESTRICT rows_bottom[3] = { 154 dc->ConstPlaneRow(0, y + 1), 155 dc->ConstPlaneRow(1, y + 1), 156 dc->ConstPlaneRow(2, y + 1), 157 }; 158 float* JXL_RESTRICT rows_out[3] = { 159 smoothed.PlaneRow(0, y), 160 smoothed.PlaneRow(1, y), 161 smoothed.PlaneRow(2, y), 162 }; 163 for (size_t x : {static_cast<size_t>(0), xsize - 1}) { 164 for (size_t c = 0; c < 3; c++) { 165 rows_out[c][x] = rows[c][x]; 166 } 167 } 168 169 size_t x = 1; 170 // First pixels 171 const size_t N = Lanes(D()); 172 for (; x < std::min(N, xsize - 1); x++) { 173 ComputePixel<DScalar>(dc_factors, rows_top, rows, rows_bottom, rows_out, 174 x); 175 } 176 // Full vectors. 177 for (; x + N <= xsize - 1; x += N) { 178 ComputePixel<D>(dc_factors, rows_top, rows, rows_bottom, rows_out, x); 179 } 180 // Last pixels. 181 for (; x < xsize - 1; x++) { 182 ComputePixel<DScalar>(dc_factors, rows_top, rows, rows_bottom, rows_out, 183 x); 184 } 185 }; 186 JXL_CHECK(RunOnPool(pool, 1, ysize - 1, ThreadPool::NoInit, process_row, 187 "DCSmoothingRow")); 188 dc->Swap(smoothed); 189 return true; 190 } 191 192 // DC dequantization. 193 void DequantDC(const Rect& r, Image3F* dc, ImageB* quant_dc, const Image& in, 194 const float* dc_factors, float mul, const float* cfl_factors, 195 const YCbCrChromaSubsampling& chroma_subsampling, 196 const BlockCtxMap& bctx) { 197 const HWY_FULL(float) df; 198 const Rebind<pixel_type, HWY_FULL(float)> di; // assumes pixel_type <= float 199 if (chroma_subsampling.Is444()) { 200 const auto fac_x = Set(df, dc_factors[0] * mul); 201 const auto fac_y = Set(df, dc_factors[1] * mul); 202 const auto fac_b = Set(df, dc_factors[2] * mul); 203 const auto cfl_fac_x = Set(df, cfl_factors[0]); 204 const auto cfl_fac_b = Set(df, cfl_factors[2]); 205 for (size_t y = 0; y < r.ysize(); y++) { 206 float* dec_row_x = r.PlaneRow(dc, 0, y); 207 float* dec_row_y = r.PlaneRow(dc, 1, y); 208 float* dec_row_b = r.PlaneRow(dc, 2, y); 209 const int32_t* quant_row_x = in.channel[1].plane.Row(y); 210 const int32_t* quant_row_y = in.channel[0].plane.Row(y); 211 const int32_t* quant_row_b = in.channel[2].plane.Row(y); 212 for (size_t x = 0; x < r.xsize(); x += Lanes(di)) { 213 const auto in_q_x = Load(di, quant_row_x + x); 214 const auto in_q_y = Load(di, quant_row_y + x); 215 const auto in_q_b = Load(di, quant_row_b + x); 216 const auto in_x = Mul(ConvertTo(df, in_q_x), fac_x); 217 const auto in_y = Mul(ConvertTo(df, in_q_y), fac_y); 218 const auto in_b = Mul(ConvertTo(df, in_q_b), fac_b); 219 Store(in_y, df, dec_row_y + x); 220 Store(MulAdd(in_y, cfl_fac_x, in_x), df, dec_row_x + x); 221 Store(MulAdd(in_y, cfl_fac_b, in_b), df, dec_row_b + x); 222 } 223 } 224 } else { 225 for (size_t c : {1, 0, 2}) { 226 Rect rect(r.x0() >> chroma_subsampling.HShift(c), 227 r.y0() >> chroma_subsampling.VShift(c), 228 r.xsize() >> chroma_subsampling.HShift(c), 229 r.ysize() >> chroma_subsampling.VShift(c)); 230 const auto fac = Set(df, dc_factors[c] * mul); 231 const Channel& ch = in.channel[c < 2 ? c ^ 1 : c]; 232 for (size_t y = 0; y < rect.ysize(); y++) { 233 const int32_t* quant_row = ch.plane.Row(y); 234 float* row = rect.PlaneRow(dc, c, y); 235 for (size_t x = 0; x < rect.xsize(); x += Lanes(di)) { 236 const auto in_q = Load(di, quant_row + x); 237 const auto in = Mul(ConvertTo(df, in_q), fac); 238 Store(in, df, row + x); 239 } 240 } 241 } 242 } 243 if (bctx.num_dc_ctxs <= 1) { 244 for (size_t y = 0; y < r.ysize(); y++) { 245 uint8_t* qdc_row = r.Row(quant_dc, y); 246 memset(qdc_row, 0, sizeof(*qdc_row) * r.xsize()); 247 } 248 } else { 249 for (size_t y = 0; y < r.ysize(); y++) { 250 uint8_t* qdc_row_val = r.Row(quant_dc, y); 251 const int32_t* quant_row_x = 252 in.channel[1].plane.Row(y >> chroma_subsampling.VShift(0)); 253 const int32_t* quant_row_y = 254 in.channel[0].plane.Row(y >> chroma_subsampling.VShift(1)); 255 const int32_t* quant_row_b = 256 in.channel[2].plane.Row(y >> chroma_subsampling.VShift(2)); 257 for (size_t x = 0; x < r.xsize(); x++) { 258 int bucket_x = 0; 259 int bucket_y = 0; 260 int bucket_b = 0; 261 for (int t : bctx.dc_thresholds[0]) { 262 if (quant_row_x[x >> chroma_subsampling.HShift(0)] > t) bucket_x++; 263 } 264 for (int t : bctx.dc_thresholds[1]) { 265 if (quant_row_y[x >> chroma_subsampling.HShift(1)] > t) bucket_y++; 266 } 267 for (int t : bctx.dc_thresholds[2]) { 268 if (quant_row_b[x >> chroma_subsampling.HShift(2)] > t) bucket_b++; 269 } 270 int bucket = bucket_x; 271 bucket *= bctx.dc_thresholds[2].size() + 1; 272 bucket += bucket_b; 273 bucket *= bctx.dc_thresholds[1].size() + 1; 274 bucket += bucket_y; 275 qdc_row_val[x] = bucket; 276 } 277 } 278 } 279 } 280 281 // NOLINTNEXTLINE(google-readability-namespace-comments) 282 } // namespace HWY_NAMESPACE 283 } // namespace jxl 284 HWY_AFTER_NAMESPACE(); 285 286 #if HWY_ONCE 287 namespace jxl { 288 289 HWY_EXPORT(DequantDC); 290 HWY_EXPORT(AdaptiveDCSmoothing); 291 Status AdaptiveDCSmoothing(const float* dc_factors, Image3F* dc, 292 ThreadPool* pool) { 293 return HWY_DYNAMIC_DISPATCH(AdaptiveDCSmoothing)(dc_factors, dc, pool); 294 } 295 296 void DequantDC(const Rect& r, Image3F* dc, ImageB* quant_dc, const Image& in, 297 const float* dc_factors, float mul, const float* cfl_factors, 298 const YCbCrChromaSubsampling& chroma_subsampling, 299 const BlockCtxMap& bctx) { 300 HWY_DYNAMIC_DISPATCH(DequantDC) 301 (r, dc, quant_dc, in, dc_factors, mul, cfl_factors, chroma_subsampling, bctx); 302 } 303 304 } // namespace jxl 305 #endif // HWY_ONCE