filters/resample.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // Avisynth v2.5. Copyright 2002 Ben Rudiak-Gould et al. | ||
| 2 | // http://avisynth.nl | ||
| 3 | |||
| 4 | // This program is free software; you can redistribute it and/or modify | ||
| 5 | // it under the terms of the GNU General Public License as published by | ||
| 6 | // the Free Software Foundation; either version 2 of the License, or | ||
| 7 | // (at your option) any later version. | ||
| 8 | // | ||
| 9 | // This program is distributed in the hope that it will be useful, | ||
| 10 | // but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 11 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 12 | // GNU General Public License for more details. | ||
| 13 | // | ||
| 14 | // You should have received a copy of the GNU General Public License | ||
| 15 | // along with this program; if not, write to the Free Software | ||
| 16 | // Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA, or visit | ||
| 17 | // http://www.gnu.org/copyleft/gpl.html . | ||
| 18 | // | ||
| 19 | // Linking Avisynth statically or dynamically with other modules is making a | ||
| 20 | // combined work based on Avisynth. Thus, the terms and conditions of the GNU | ||
| 21 | // General Public License cover the whole combination. | ||
| 22 | // | ||
| 23 | // As a special exception, the copyright holders of Avisynth give you | ||
| 24 | // permission to link Avisynth with independent modules that communicate with | ||
| 25 | // Avisynth solely through the interfaces defined in avisynth.h, regardless of the license | ||
| 26 | // terms of these independent modules, and to copy and distribute the | ||
| 27 | // resulting combined work under terms of your choice, provided that | ||
| 28 | // every copy of the combined work is accompanied by a complete copy of | ||
| 29 | // the source code of Avisynth (the version of Avisynth used to produce the | ||
| 30 | // combined work), being distributed under the terms of the GNU General | ||
| 31 | // Public License plus this exception. An independent module is a module | ||
| 32 | // which is not derived from or based on Avisynth, such as 3rd-party filters, | ||
| 33 | // import and export plugins, or graphical user interfaces. | ||
| 34 | |||
| 35 | #include "resample.h" | ||
| 36 | #ifdef INTEL_INTRINSICS | ||
| 37 | #include "intel/resample_sse.h" | ||
| 38 | #include "intel/resample_avx2.h" | ||
| 39 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 40 | #include "intel/resample_avx512.h" | ||
| 41 | #endif | ||
| 42 | #include "intel/turn_sse.h" | ||
| 43 | #include "intel/turn_avx2.h" | ||
| 44 | #endif | ||
| 45 | #ifdef NEON_INTRINSICS | ||
| 46 | #include "aarch64/turn_neon.h" | ||
| 47 | #endif | ||
| 48 | |||
| 49 | #include <avs/config.h> | ||
| 50 | |||
| 51 | #include "transform.h" | ||
| 52 | #include "turn.h" | ||
| 53 | #include <avs/alignment.h> | ||
| 54 | #include <avs/minmax.h> | ||
| 55 | #include "../convert/convert_planar.h" | ||
| 56 | #include "../convert/convert_helper.h" | ||
| 57 | |||
| 58 | #include <type_traits> | ||
| 59 | #include <algorithm> | ||
| 60 | |||
| 61 | #include "../core/avs_simd_c.h" | ||
| 62 | #include <cassert> | ||
| 63 | |||
| 64 | // Prepares resampling coefficients for end conditions and/or SIMD processing by: | ||
| 65 | // 1. Sets a "real-life" size for the filter, which at small dimensions can be less than the original | ||
| 66 | // 2. Aligning filter_size to 8 or 16 boundary for SIMD efficiency | ||
| 67 | // 3. Right-aligning coefficients within padded arrays to ensure valid access at boundaries | ||
| 68 | // | ||
| 69 | // Before: After right-alignment (filter_size=4, kernel_size=2): | ||
| 70 | // | ||
| 71 | // offset->| offset-2 ->| | ||
| 72 | // [x][y][ ][ ] [0][0][x][y] | ||
| 73 | // ^ ^ ^ ^ ^ ^ | ||
| 74 | // | | Off-boundary | | | ||
| 75 | // Values used Values used | ||
| 76 | // | ||
| 77 | // This ensures SIMD instructions can safely load full vectors even at image boundaries | ||
| 78 | // while maintaining correct coefficient positioning and proper zero padding. | ||
| 79 | |||
| 80 | |||
| 81 | 23997 | static void checkAndSetOverread(int end_pos, SafeLimit& safelimit, int start_pos, int i, int source_size) { | |
| 82 |
2/2✓ Branch 2 → 3 taken 3975 times.
✓ Branch 2 → 5 taken 20022 times.
|
23997 | if (end_pos >= source_size) { |
| 83 |
2/2✓ Branch 3 → 4 taken 982 times.
✓ Branch 3 → 5 taken 2993 times.
|
3975 | if (!safelimit.overread_possible) { |
| 84 | 982 | safelimit.overread_possible = true; | |
| 85 | 982 | safelimit.source_overread_offset = start_pos; | |
| 86 | 982 | safelimit.source_overread_beyond_targetx = i; | |
| 87 | } | ||
| 88 | } | ||
| 89 | 23997 | } | |
| 90 | |||
| 91 | |||
| 92 | 120 | void resize_prepare_coeffs(ResamplingProgram* p, IScriptEnvironment* env, int filter_size_alignment) { | |
| 93 | 120 | p->filter_size_alignment = filter_size_alignment; | |
| 94 | 120 | p->safelimit_filter_size_aligned.overread_possible = false; | |
| 95 | 120 | p->safelimit_4_pixels.overread_possible = false; | |
| 96 | 120 | p->safelimit_8_pixels.overread_possible = false; | |
| 97 | 120 | p->safelimit_16_pixels.overread_possible = false; | |
| 98 | 120 | p->safelimit_32_pixels.overread_possible = false; | |
| 99 | 120 | p->safelimit_8_pixels_each8th_target.overread_possible = false; | |
| 100 | 120 | p->safelimit_16_pixels_each16th_target.overread_possible = false; | |
| 101 | 120 | p->safelimit_64_pixels_each32th_target.overread_possible = false; // avx512 uint16_t 32 target pixels, handling 64 source pixels in permutex-based resizers | |
| 102 | 120 | p->safelimit_128_pixels_each64th_target.overread_possible = false; // avx512 uint8_t 64 target pixels, handling 128 source pixels in permutex-based resizers | |
| 103 | // FIXME: found out how to make it general safelimit_SOURCEREADPIXELS_pixels_each_TARGETPIXELSATATIME. Not here, in each frame proecssing for sure. | ||
| 104 | |||
| 105 | // note: filter_size_real was the max(kernel_sizes[]) | ||
| 106 | 120 | int filter_size_aligned = AlignNumber(p->filter_size_real, p->filter_size_alignment); | |
| 107 | // FIXME: really this needs to be dynamic based on SIMD used in resizer | ||
| 108 | |||
| 109 | 120 | int target_size_aligned = AlignNumber(p->target_size, ALIGN_RESIZER_TARGET_SIZE); | |
| 110 | |||
| 111 | // align target_size to X units to allow safe, up to X pixels/cycle in H resizers. | ||
| 112 | // also, this is the coeff table Y-size. | ||
| 113 | // e.g. ALIGN_RESIZER_TARGET_SIZE = 64 allows to access coefficient table elements at | ||
| 114 | // current_coeff + filter_size * 63, if we step current_coeff by 64 * filter_size | ||
| 115 | 120 | p->target_size_alignment = ALIGN_RESIZER_TARGET_SIZE; | |
| 116 | |||
| 117 | // Common variables for both float and integer paths | ||
| 118 | 120 | void* new_coeff = nullptr; | |
| 119 | 120 | void* src_coeff = nullptr; | |
| 120 | 120 | size_t element_size = 0; | |
| 121 | |||
| 122 | // allocate for a larger target_size area and nullify the coeffs. | ||
| 123 | // Even between target_size and target_size_aligned. | ||
| 124 |
2/2✓ Branch 4 → 5 taken 12 times.
✓ Branch 4 → 19 taken 108 times.
|
120 | if (p->bits_per_pixel == 32) { |
| 125 | 12 | element_size = sizeof(float); | |
| 126 | 12 | src_coeff = p->pixel_coefficient_float; | |
| 127 | 12 | new_coeff = env->Allocate(element_size * target_size_aligned * filter_size_aligned, 64, AVS_NORMAL_ALLOC); | |
| 128 |
1/2✗ Branch 6 → 7 not taken.
✓ Branch 6 → 9 taken 12 times.
|
12 | if (!new_coeff) { |
| 129 | ✗ | env->Free(new_coeff); | |
| 130 | ✗ | env->ThrowError("Could not reserve memory in a resampler."); | |
| 131 | } | ||
| 132 | 24 | std::fill_n((float*)new_coeff, target_size_aligned * filter_size_aligned, 0.0f); | |
| 133 | } | ||
| 134 | else { | ||
| 135 | 108 | element_size = sizeof(short); | |
| 136 | 108 | src_coeff = p->pixel_coefficient; | |
| 137 | 108 | new_coeff = env->Allocate(element_size * target_size_aligned * filter_size_aligned, 64, AVS_NORMAL_ALLOC); | |
| 138 |
1/2✗ Branch 20 → 21 not taken.
✓ Branch 20 → 23 taken 108 times.
|
108 | if (!new_coeff) { |
| 139 | ✗ | env->Free(new_coeff); | |
| 140 | ✗ | env->ThrowError("Could not reserve memory in a resampler."); | |
| 141 | } | ||
| 142 | 108 | memset(new_coeff, 0, element_size * target_size_aligned * filter_size_aligned); | |
| 143 | } | ||
| 144 | |||
| 145 | 120 | const int last_line = p->source_size - 1; | |
| 146 | |||
| 147 | // Process coefficients - common code for both types | ||
| 148 |
2/2✓ Branch 60 → 25 taken 4421 times.
✓ Branch 60 → 61 taken 120 times.
|
4541 | for (int i = 0; i < p->target_size; i++) { |
| 149 | 4421 | const int kernel_size = p->kernel_sizes[i]; | |
| 150 | 4421 | const int offset = p->pixel_offset[i]; | |
| 151 | 4421 | const int last_coeff_index = offset + p->filter_size_real - 1; | |
| 152 |
2/2✓ Branch 27 → 28 taken 236 times.
✓ Branch 27 → 29 taken 4185 times.
|
4421 | const int shift_needed = last_coeff_index > last_line ? p->filter_size_real - kernel_size : 0; |
| 153 | |||
| 154 | // In order to be able to read 'filter_size_real' number of coefficients safely at the | ||
| 155 | // image boundaries, we right-align the actual coefficients within the allocated filter | ||
| 156 | // size. This will require adjusting (shifting) the pixel offsets as well, and increasing | ||
| 157 | // the smaller kernel sizes, to reflect the new effective size: filter_size_real. | ||
| 158 | |||
| 159 | // Copy coefficients with appropriate shift | ||
| 160 |
2/2✓ Branch 30 → 31 taken 329 times.
✓ Branch 30 → 34 taken 4092 times.
|
4421 | if (p->bits_per_pixel == 32) { |
| 161 | 329 | float* dst = (float*)new_coeff + i * filter_size_aligned; | |
| 162 | 329 | float* src = (float*)src_coeff + i * p->filter_size; | |
| 163 |
2/2✓ Branch 33 → 32 taken 818 times.
✓ Branch 33 → 37 taken 329 times.
|
1147 | for (int j = 0; j < kernel_size; j++) { |
| 164 | 818 | dst[j + shift_needed] = src[j]; | |
| 165 | } | ||
| 166 | } | ||
| 167 | else { | ||
| 168 | 4092 | short* dst = (short*)new_coeff + i * filter_size_aligned; | |
| 169 | 4092 | short* src = (short*)src_coeff + i * p->filter_size; | |
| 170 |
2/2✓ Branch 36 → 35 taken 45080 times.
✓ Branch 36 → 37 taken 4092 times.
|
49172 | for (int j = 0; j < kernel_size; j++) { |
| 171 | 45080 | dst[j + shift_needed] = src[j]; | |
| 172 | } | ||
| 173 | } | ||
| 174 | |||
| 175 | // Update offsets and kernel sizes | ||
| 176 | 4421 | p->pixel_offset[i] -= shift_needed; | |
| 177 | 4421 | p->kernel_sizes[i] += shift_needed; | |
| 178 | |||
| 179 | // left side, already right padded with zero coeffs, we can | ||
| 180 | // change to actual width to the common one | ||
| 181 |
2/2✓ Branch 40 → 41 taken 233 times.
✓ Branch 40 → 43 taken 4188 times.
|
4421 | if(p->kernel_sizes[i] < p->filter_size_real) |
| 182 | 233 | p->kernel_sizes[i] = p->filter_size_real; | |
| 183 | |||
| 184 | // In a horizontal resizer, when reading filter_size_alignment pixels, | ||
| 185 | // we must protect against source scanline overread. | ||
| 186 | // Using this not in only 32-bit float resizers is new in 3.7.4. | ||
| 187 | 4421 | const int start_pos = p->pixel_offset[i]; | |
| 188 | 4421 | const int end_pos = start_pos + p->filter_size_real - 1; | |
| 189 | 4421 | if (end_pos >= p->source_size) { | |
| 190 | // This issue has already been fixed, so it cannot occur. | ||
| 191 | } | ||
| 192 | |||
| 193 | // Check for SIMD optimization limits and record first danger positions. | ||
| 194 | // If reading N pixels starting from `start_pos` would reach past the end | ||
| 195 | // of the source (>= source_size), register that first occurrence for | ||
| 196 | // the corresponding SafeLimit entry so resizers can avoid unsafe wide loads. | ||
| 197 | |||
| 198 | 4421 | checkAndSetOverread(start_pos + filter_size_aligned - 1, p->safelimit_filter_size_aligned, start_pos, i, p->source_size); | |
| 199 | 4421 | checkAndSetOverread(start_pos + 4 - 1, p->safelimit_4_pixels, start_pos, i, p->source_size); | |
| 200 | 4421 | checkAndSetOverread(start_pos + 8 - 1, p->safelimit_8_pixels, start_pos, i, p->source_size); | |
| 201 | 4421 | checkAndSetOverread(start_pos + 16 - 1, p->safelimit_16_pixels, start_pos, i, p->source_size); | |
| 202 | 4421 | checkAndSetOverread(start_pos + 32 - 1, p->safelimit_32_pixels, start_pos, i, p->source_size); | |
| 203 | // for permutex-based AVX2 ks4 float H resizers, where we read 8 pixels at a time exactly from | ||
| 204 | // start_pos of each Nth pixel output block | ||
| 205 |
2/2✓ Branch 50 → 51 taken 591 times.
✓ Branch 50 → 53 taken 3830 times.
|
4421 | if (i % 8 == 0) { |
| 206 | 591 | checkAndSetOverread(start_pos + 8 - 1, p->safelimit_8_pixels_each8th_target, start_pos, i, p->source_size); | |
| 207 | 591 | checkAndSetOverread(start_pos + 16 - 1, p->safelimit_16_pixels_each8th_target, start_pos, i, p->source_size); | |
| 208 | } | ||
| 209 |
2/2✓ Branch 53 → 54 taken 341 times.
✓ Branch 53 → 55 taken 4080 times.
|
4421 | if (i % 16 == 0) |
| 210 | 341 | checkAndSetOverread(start_pos + 16 - 1, p->safelimit_16_pixels_each16th_target, start_pos, i, p->source_size); | |
| 211 |
2/2✓ Branch 55 → 56 taken 213 times.
✓ Branch 55 → 57 taken 4208 times.
|
4421 | if (i % 32 == 0) // avx512 uint16_t 32 target pixels, handling 64 source pixels |
| 212 | 213 | checkAndSetOverread(start_pos + 64 - 1, p->safelimit_64_pixels_each32th_target, start_pos, i, p->source_size); | |
| 213 |
2/2✓ Branch 57 → 58 taken 156 times.
✓ Branch 57 → 59 taken 4265 times.
|
4421 | if (i % 64 == 0) // avx512 uint8_t 64 target pixels, handling 128 source pixels |
| 214 | 156 | checkAndSetOverread(start_pos + 128 - 1, p->safelimit_128_pixels_each64th_target, start_pos, i, p->source_size); | |
| 215 | |||
| 216 | } | ||
| 217 | |||
| 218 | // from now on, kernel_sizes[] has no role, each is filter_size_real | ||
| 219 | 120 | p->kernel_sizes.clear(); | |
| 220 | |||
| 221 | // Fill the extra offset after target_size with fake values. | ||
| 222 | // Our aim is to have a safe, up to 8-32 pixels/cycle simd loop for V and specific H resizers. | ||
| 223 | // Their coeffs will be 0, so they don't count if such coeffs | ||
| 224 | // are multiplied with invalid, though existing pixels. | ||
| 225 |
2/2✓ Branch 62 → 63 taken 99 times.
✓ Branch 62 → 69 taken 21 times.
|
120 | if (p->target_size < target_size_aligned) { |
| 226 | 99 | p->pixel_offset.resize(target_size_aligned); | |
| 227 | 99 | int last_offset = p->pixel_offset[p->target_size - 1]; | |
| 228 |
2/2✓ Branch 68 → 66 taken 5563 times.
✓ Branch 68 → 69 taken 99 times.
|
5662 | for (int i = p->target_size; i < target_size_aligned; ++i) { |
| 229 | 5563 | p->pixel_offset[i] = last_offset; // repeat last valid offset, helps permutex-based H resizers to stay within valid distances | |
| 230 | } | ||
| 231 | } | ||
| 232 | |||
| 233 | // Free old coefficients and assign new ones | ||
| 234 |
2/2✓ Branch 69 → 70 taken 12 times.
✓ Branch 69 → 72 taken 108 times.
|
120 | if (p->bits_per_pixel == 32) { |
| 235 | 12 | env->Free(p->pixel_coefficient_float); | |
| 236 | 12 | p->pixel_coefficient_float = (float*)new_coeff; | |
| 237 | } | ||
| 238 | else { | ||
| 239 | 108 | env->Free(p->pixel_coefficient); | |
| 240 | 108 | p->pixel_coefficient = (short*)new_coeff; | |
| 241 | } | ||
| 242 | |||
| 243 | 120 | p->filter_size = filter_size_aligned; | |
| 244 | // by now coeffs[old_filter_size][target_size] was copied and padded into coeffs[new_filter_size][target_size] | ||
| 245 | 120 | } | |
| 246 | |||
| 247 | /*************************************** | ||
| 248 | ***** Vertical Resizer Assembly ******* | ||
| 249 | ***************************************/ | ||
| 250 | |||
| 251 | template<typename pixel_t> | ||
| 252 | ✗ | static void resize_v_planar_pointresize(BYTE* dst, const BYTE* src, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int target_height, int bits_per_pixel) | |
| 253 | { | ||
| 254 | AVS_UNUSED(bits_per_pixel); | ||
| 255 | |||
| 256 | ✗ | pixel_t* src0 = (pixel_t*)src; | |
| 257 | ✗ | pixel_t* dst0 = (pixel_t*)dst; | |
| 258 | ✗ | src_pitch = src_pitch / sizeof(pixel_t); | |
| 259 | ✗ | dst_pitch = dst_pitch / sizeof(pixel_t); | |
| 260 | |||
| 261 | ✗ | for (int y = 0; y < target_height; y++) { | |
| 262 | ✗ | int offset = program->pixel_offset[y]; | |
| 263 | ✗ | const pixel_t* src_ptr = src0 + src_pitch * offset; | |
| 264 | |||
| 265 | ✗ | memcpy(dst0, src_ptr, width * sizeof(pixel_t)); | |
| 266 | |||
| 267 | ✗ | dst0 += dst_pitch; | |
| 268 | } | ||
| 269 | ✗ | } | |
| 270 | |||
| 271 | // This C implementation isn't optimized for auto-vectorization, | ||
| 272 | // But on x86 MSVC SSE2 settings for 8 bit pixel types it performs better than our | ||
| 273 | // vectorizer-friendly version. | ||
| 274 | // Other compilers benefit from vectorizing by a huge margin for every case. | ||
| 275 | // The vector code which replaced this is already 2x faster for 10-16 bits processing | ||
| 276 | // and achieves a 6x speedup for 32-bit operations. LLVM has even an additional 1.5x-4x speedup | ||
| 277 | // compared to MSVC. | ||
| 278 | // Kept for reference, supports 8, 10-16 and 32 bit pixel types. | ||
| 279 | template<typename pixel_t, bool lessthan16bit> | ||
| 280 | static void resize_v_c_planar(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int target_height, int bits_per_pixel) | ||
| 281 | { | ||
| 282 | |||
| 283 | int filter_size = program->filter_size; | ||
| 284 | |||
| 285 | typedef typename std::conditional < std::is_floating_point<pixel_t>::value, float, short>::type coeff_t; | ||
| 286 | coeff_t* current_coeff; | ||
| 287 | |||
| 288 | if (!std::is_floating_point<pixel_t>::value) | ||
| 289 | current_coeff = (coeff_t*)program->pixel_coefficient; | ||
| 290 | else | ||
| 291 | current_coeff = (coeff_t*)program->pixel_coefficient_float; | ||
| 292 | |||
| 293 | pixel_t* src = (pixel_t*)src8; | ||
| 294 | pixel_t* dst = (pixel_t*)dst8; | ||
| 295 | src_pitch = src_pitch / sizeof(pixel_t); | ||
| 296 | dst_pitch = dst_pitch / sizeof(pixel_t); | ||
| 297 | |||
| 298 | pixel_t limit = 0; | ||
| 299 | if (!std::is_floating_point<pixel_t>::value) { // floats are unscaled and uncapped | ||
| 300 | if constexpr (sizeof(pixel_t) == 1) limit = 255; | ||
| 301 | else if constexpr (sizeof(pixel_t) == 2) limit = pixel_t((1 << bits_per_pixel) - 1); | ||
| 302 | } | ||
| 303 | |||
| 304 | // for 16 bits only | ||
| 305 | const short shifttosigned_short = -32768; | ||
| 306 | const int shiftfromsigned_int = 32768 << FPScale16bits; | ||
| 307 | |||
| 308 | for (int y = 0; y < target_height; y++) { | ||
| 309 | int offset = program->pixel_offset[y]; | ||
| 310 | const int kernel_size = program->kernel_sizes[y]; | ||
| 311 | const pixel_t* src_ptr = src + src_pitch * offset; | ||
| 312 | |||
| 313 | // perhaps helps vectorizing decision | ||
| 314 | const int ksmod4 = kernel_size / 4 * 4; | ||
| 315 | |||
| 316 | for (int x = 0; x < width; x++) { | ||
| 317 | if constexpr (std::is_floating_point<pixel_t>::value) { | ||
| 318 | const float* src2_ptr = src_ptr + x; | ||
| 319 | |||
| 320 | float result = 0; | ||
| 321 | for (int i = 0; i < ksmod4; i += 4) { | ||
| 322 | result += *(src2_ptr + 0 * src_pitch) * current_coeff[i + 0]; | ||
| 323 | result += *(src2_ptr + 1 * src_pitch) * current_coeff[i + 1]; | ||
| 324 | result += *(src2_ptr + 2 * src_pitch) * current_coeff[i + 2]; | ||
| 325 | result += *(src2_ptr + 3 * src_pitch) * current_coeff[i + 3]; | ||
| 326 | src2_ptr += 4 * src_pitch; | ||
| 327 | } | ||
| 328 | for (int i = ksmod4; i < kernel_size; i++) { | ||
| 329 | result += *src2_ptr * current_coeff[i]; | ||
| 330 | src2_ptr += src_pitch; | ||
| 331 | } | ||
| 332 | dst[x] = result; | ||
| 333 | } | ||
| 334 | else if constexpr (sizeof(pixel_t) == 2) { | ||
| 335 | // theoretically, no need for int64 accumulator, | ||
| 336 | // sum of coeffs is 1.0 that is (1 << FPScale16bits) in integer arithmetic | ||
| 337 | const uint16_t* src2_ptr = src_ptr + x; | ||
| 338 | int result = 1 << (FPScale16bits - 1); // rounder; | ||
| 339 | for (int i = 0; i < ksmod4; i += 4) { | ||
| 340 | int val; | ||
| 341 | val = *(src2_ptr + 0 * src_pitch); | ||
| 342 | if constexpr (!lessthan16bit) | ||
| 343 | val = val + shifttosigned_short; | ||
| 344 | result += val * current_coeff[i + 0]; | ||
| 345 | |||
| 346 | val = *(src2_ptr + 1 * src_pitch); | ||
| 347 | if constexpr (!lessthan16bit) | ||
| 348 | val = val + shifttosigned_short; | ||
| 349 | result += val * current_coeff[i + 1]; | ||
| 350 | |||
| 351 | val = *(src2_ptr + 2 * src_pitch); | ||
| 352 | if constexpr (!lessthan16bit) | ||
| 353 | val = val + shifttosigned_short; | ||
| 354 | result += val * current_coeff[i + 2]; | ||
| 355 | |||
| 356 | val = *(src2_ptr + 3 * src_pitch); | ||
| 357 | if constexpr (!lessthan16bit) | ||
| 358 | val = val + shifttosigned_short; | ||
| 359 | result += val * current_coeff[i + 3]; | ||
| 360 | |||
| 361 | src2_ptr += 4 * src_pitch; | ||
| 362 | } | ||
| 363 | for (int i = ksmod4; i < kernel_size; i++) { | ||
| 364 | int val = *src2_ptr; | ||
| 365 | if constexpr (!lessthan16bit) | ||
| 366 | val = val + shifttosigned_short; | ||
| 367 | result += val * current_coeff[i]; | ||
| 368 | src2_ptr += src_pitch; | ||
| 369 | } | ||
| 370 | if constexpr (!lessthan16bit) | ||
| 371 | result = result + shiftfromsigned_int; | ||
| 372 | result = result >> FPScale16bits; | ||
| 373 | result = result > limit ? limit : result < 0 ? 0 : result; // clamp 10..16 bits | ||
| 374 | dst[x] = (uint16_t)result; | ||
| 375 | } | ||
| 376 | else if constexpr (sizeof(pixel_t) == 1) { | ||
| 377 | const uint8_t* src2_ptr = src_ptr + x; | ||
| 378 | int result = 1 << (FPScale8bits - 1); // rounder; | ||
| 379 | for (int i = 0; i < ksmod4; i += 4) { | ||
| 380 | short val; | ||
| 381 | val = *(src2_ptr + 0 * src_pitch); | ||
| 382 | result += val * current_coeff[i + 0]; | ||
| 383 | |||
| 384 | val = *(src2_ptr + 1 * src_pitch); | ||
| 385 | result += val * current_coeff[i + 1]; | ||
| 386 | |||
| 387 | val = *(src2_ptr + 2 * src_pitch); | ||
| 388 | result += val * current_coeff[i + 2]; | ||
| 389 | |||
| 390 | val = *(src2_ptr + 3 * src_pitch); | ||
| 391 | result += val * current_coeff[i + 3]; | ||
| 392 | |||
| 393 | src2_ptr += 4 * src_pitch; | ||
| 394 | } | ||
| 395 | for (int i = ksmod4; i < kernel_size; i++) { | ||
| 396 | short val = *src2_ptr; | ||
| 397 | result += val * current_coeff[i]; | ||
| 398 | src2_ptr += src_pitch; | ||
| 399 | } | ||
| 400 | result = result >> FPScale8bits; | ||
| 401 | result = result > limit ? limit : result < 0 ? 0 : result; // clamp 8 bits | ||
| 402 | dst[x] = (uint8_t)result; | ||
| 403 | } | ||
| 404 | } | ||
| 405 | |||
| 406 | dst += dst_pitch; | ||
| 407 | current_coeff += filter_size; | ||
| 408 | } | ||
| 409 | } | ||
| 410 | |||
| 411 | |||
| 412 | /* | ||
| 413 | |||
| 414 | Benchmarks. | ||
| 415 | |||
| 416 | SetMaxCPU("none") | ||
| 417 | #SetMaxCPU("SSSE3") | ||
| 418 | #SetMaxCPU("AVX2") | ||
| 419 | ColorbarsHD(200,200, pixel_type="YUV444P8") # P8, P10, P16, PS | ||
| 420 | |||
| 421 | wmul=2 # >1: horizontal resizer test | ||
| 422 | hmul=1 # >1: vertical resizer test | ||
| 423 | |||
| 424 | avs=LanczosResize(width*wmul, height*hmul, taps=16) | ||
| 425 | avsr=z_ConvertFormat(width=width*wmul, height=height*hmul, resample_filter="lanczos", filter_param_a=16) | ||
| 426 | fmtconv=fmtc_resample(w=width*wmul, h=height*hmul, kernel="lanczos", taps=16) | ||
| 427 | avs | ||
| 428 | |||
| 429 | Figures in general has no place in source code but it is useful to have a clue | ||
| 430 | and see the huge differences between compilers and code variants. | ||
| 431 | Also, only mazochists want to test integer optimization with MSVC (x86). Wasted time. | ||
| 432 | Use llvm on Arm, clangcl/llvm on Windows x86. | ||
| 433 | |||
| 434 | Horizontals Hor+Vert Verticals | ||
| 435 | [8] [10-14] [16] [32] [8] [10] [8] [10-14] [16] [32] | ||
| 436 | 177 170 101 208 262 486 417 217 C-RaspberryPi5 gcc 12.2 (code variant) | ||
| 437 | 137 469 88 208 C-RaspberryPi5 gcc 12.2 (code variant) | ||
| 438 | 227 218 147 178 C-RaspberryPi5 gcc 12.2 no vector attrib ??! Quicker than vector attrib version gcc not recommended | ||
| 439 | 416 611 579 404 128 186 362 609 556 624 C-RaspberryPi5 llvm 14 vector attrib | ||
| 440 | 90 185 94 403 C-RaspberryPi5 llvm 14 vector attrib + integer madd@H | ||
| 441 | 180 182 162 67 213 212 206 639 C-RaspberryPi5 llvm 14 no vector attrib | ||
| 442 | 942 985 982 993 976 1114 982 1920 C-ClangCl in 3.7.6 VS2026 SSE2 (MSVC tuned simd_c and even more tuned vector-friendly code) | ||
| 443 | 1270 1238 1186 1550 1555 1560 3670 C-ClangCl in VS2022 AVX2 | ||
| 444 | 1051 1126 1102 C-Intel ICX 2025 SSE2 | ||
| 445 | 1513 2355 1560 1128 1969 C-Intel ICX 2025 SSE4.2 smart madd! | ||
| 446 | 1938 2413 1775 1061 442 453 1136 1126 1037 3511 C-Intel ICX 2025 AVX2 | ||
| 447 | 212 188 187 264 73 64 223 195 198 268 C-MSVC SSE2 3.7.3 | ||
| 448 | 417 463 360 449 424 384 352 C-MSVC SSE2 3.7.4 (some unrolling vs. 3.7.3) | ||
| 449 | 215 215 97 928 79 79 220 744 96 1951 C-MSVC SSE2 (zero optim seen in asm on 8-16, did not tolerate vector-friendly code) | ||
| 450 | 595 638 634 976 552 382 95 1940 C-MSVC SSE2 3.7.6 (MSVC tuned simd_c and vector-friendly code) | ||
| 451 | 591 616 620 700 670 518 1950 C-MSVC SSE2 3.7.6 (MSVC tuned simd_c and even more tuned vector-friendly code) | ||
| 452 | 853 913 924 1107 880 860 713 3777 C-MSVC AVX2 3.7.6 (MSVC tuned simd_c and vector-friendly code) | ||
| 453 | 201 206 99 C-MSVC AVX2 3.7.4 (zero optim seen in asm also on 8-16, not even using SSE2 xmm registers) | ||
| 454 | 597 631 651 C-Intel SSE4.2 3.7.4 code | ||
| 455 | 1183 1193 889 C-Intel AVX2 3.7.4 code | ||
| 456 | 5600 2140 SIMD-avsresize (AVX2 or AVX512?) | ||
| 457 | 2260 840 SIMD-fmtconv (16 bit output for 8 bits) | ||
| 458 | 4578 2614 2560 2250 1490 4220 4534 3887 3570 SIMD-MSVC AVX2 3.7.3 (horizontal was memory-boundary unsafe) | ||
| 459 | 3631 3505 3221 2344 1291 1354 3804 4466 3855 2260 SIMD-MSVC AVX2 3.7.4 Float vertical regression - no time to finish | ||
| 460 | 3720 3478 3130 2385 1566 1480 5014 5288 5077 3810 SIMD-MSVC AVX2 + incrementing offsets in V, 20-25% gain in integer verticals | ||
| 461 | 4730 4612 4233 2487 1390 1471 3792 4476 4380 3942 SIMD-ClangCl AVX2, verticals behind MSVC by surprise | ||
| 462 | 2373 2181 1893 1306 868 723 2660 2137 2670 1886 SIMD-MSVC SSSE3 + incrementing offsets in V, 20-25% gain in integer verticals | ||
| 463 | 2294 2979 2595 1460 859 976 2623 2865 2625 1962 SIMD-Intel ICX 2025 SSSE3 | ||
| 464 | 4395 4616 4160 2570 1664 1720 5110 5870 5085 2999 SIMD-Intel ICX 2025 AVX2 Surprisingly slow at vertical float FIXME, slower than C :) | ||
| 465 | 4755 4453 4255 3540 1447 1537 3560 4592 3920 7837 SIMD-ClangCL 3.7.6 | ||
| 466 | |||
| 467 | * float has different optimization: 8 pixels 2 coeffs | ||
| 468 | ** on aarch64 the float benchmarks included a ConvertBits(8) at the end. | ||
| 469 | |||
| 470 | Non-x86 platforms are the focus of this comparison, as they lack SIMD optimized paths. | ||
| 471 | x86 platforms have their own SIMD optimized paths and do not use C. | ||
| 472 | |||
| 473 | Compilers can give totally different results even after some reordering of the code, | ||
| 474 | See gcc aarch64: the 8 bit case dropped to 2/3 speed while 16 bits increased by a factor of 250% | ||
| 475 | depending on where I put a line. | ||
| 476 | |||
| 477 | Considerations for aarch64 (Armv8a, includes neon, e.g. RPi5 with gcc 12.2, llvm 14.0.6) | ||
| 478 | - gcc is not recommended, their vectorizer is either broken or not yet ready. | ||
| 479 | Using vector attribute gave slower code. | ||
| 480 | - llvm gave consistent fast results. | ||
| 481 | - Vertical resampler: aarch64 + GCC, (8-16 bits) the 4-pixel 4-coefficient version performs best. | ||
| 482 | |||
| 483 | On x86 platforms MSVC is unusable regarding integer vectorization, the more the code helps to | ||
| 484 | recognize vectorization patterns the slower assembly it produces, e.g. horizontal resizer AVX2 16 bit: | ||
| 485 | MSVC 99 fps (634 after tuning), MSVC+clangcl: 1186 fps, Intel+LLVM: 1775 fps. Joke. We use | ||
| 486 | only for internal testing and reference, not for actual builds, everithing is SIMD optimized in Avisynth for x86. | ||
| 487 | |||
| 488 | MSVC on x86 is only capable of vectorizing 32 bit float code, and even then it is not the fastest. | ||
| 489 | There is no reason to not use clangcl for x86 Visual Studio builds. It comes officially with VS2022 for free. | ||
| 490 | 2026 update: simd_c was tuned for better MSVC auto-vectorization. | ||
| 491 | |||
| 492 | */ | ||
| 493 | |||
| 494 | // This is a vectorizer-friendly version of the above function. | ||
| 495 | template<typename pixel_t, bool lessthan16bit> | ||
| 496 | ✗ | void resize_v_c_planar_uint8_16_t_auto_vectorized(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int target_height, int bits_per_pixel) { | |
| 497 | |||
| 498 | ✗ | const short* AVS_RESTRICT current_coeff = program->pixel_coefficient; | |
| 499 | |||
| 500 | ✗ | auto src = reinterpret_cast<const pixel_t*>(src8); | |
| 501 | ✗ | auto dst = reinterpret_cast<pixel_t * AVS_RESTRICT>(dst8); | |
| 502 | ✗ | src_pitch = src_pitch / sizeof(pixel_t); | |
| 503 | ✗ | dst_pitch = dst_pitch / sizeof(pixel_t); | |
| 504 | |||
| 505 | ✗ | int limit = 0; | |
| 506 | ✗ | if constexpr (sizeof(pixel_t) == 1) limit = 255; | |
| 507 | ✗ | else if constexpr (sizeof(pixel_t) == 2) limit = pixel_t((1 << bits_per_pixel) - 1); | |
| 508 | |||
| 509 | // for 16 bits only | ||
| 510 | ✗ | [[maybe_unused]] Int16x8 shifttosigned_short(-32768); | |
| 511 | ✗ | [[maybe_unused]] const Int32x4 shiftfromsigned_int(32768 << FPScale16bits); | |
| 512 | |||
| 513 | ✗ | const int filter_size = program->filter_size; | |
| 514 | ✗ | constexpr int fixpoint_scaler_bits = sizeof(pixel_t) == 2 ? FPScale16bits : FPScale8bits; | |
| 515 | ✗ | const int rounder = 1 << (fixpoint_scaler_bits - 1); | |
| 516 | |||
| 517 | // 8-16 SIMD padding of coeffs is not usable here, work with single coeffs | ||
| 518 | |||
| 519 | ✗ | const int kernel_size = program->filter_size_real; | |
| 520 | |||
| 521 | ✗ | const int ksmod4 = kernel_size / 4 * 4; | |
| 522 | |||
| 523 | ✗ | for (int y = 0; y < target_height; y++) { | |
| 524 | ✗ | int offset = program->pixel_offset[y]; | |
| 525 | ✗ | const pixel_t* src_ptr = src + offset * src_pitch; | |
| 526 | |||
| 527 | // 4 pixels at a time | ||
| 528 | ✗ | for (int x = 0; x < width; x += 4) { | |
| 529 | ✗ | Int32x4 result(rounder); // master accumulator and the initial first coeff part | |
| 530 | ✗ | Int32x4 result_2(0); | |
| 531 | ✗ | Int32x4 result_3(0); | |
| 532 | ✗ | Int32x4 result_4(0); | |
| 533 | ✗ | const pixel_t* AVS_RESTRICT src2_ptr = src_ptr + x; // __restrict here | |
| 534 | ✗ | int i = 0; | |
| 535 | // Process coefficients in pairs or quads for better instruction parallelism. | ||
| 536 | // Depending on platform and compiler, results may vary. | ||
| 537 | // Since on Intel we have SIMD optimized paths, we decide on the | ||
| 538 | // results of a not-yet-optimized aarch64 platform: takeout is 4 pix 4 coeff | ||
| 539 | ✗ | for (; i < ksmod4; i += 4) { | |
| 540 | |||
| 541 | ✗ | Int32x4 src_1, src_2, src_3, src_4; | |
| 542 | |||
| 543 | ✗ | const int coeff_1 = current_coeff[i]; | |
| 544 | ✗ | const int coeff_2 = current_coeff[i + 1]; | |
| 545 | ✗ | const int coeff_3 = current_coeff[i + 2]; | |
| 546 | ✗ | const int coeff_4 = current_coeff[i + 3]; | |
| 547 | |||
| 548 | if constexpr (sizeof(pixel_t) == 1) { | ||
| 549 | // uint8_t | ||
| 550 | auto src8_1 = Uint8x4::from_ptr(src2_ptr); | ||
| 551 | ✗ | auto src8_2 = Uint8x4::from_ptr(src2_ptr + src_pitch); | |
| 552 | ✗ | auto src8_3 = Uint8x4::from_ptr(src2_ptr + 2 * src_pitch); | |
| 553 | ✗ | auto src8_4 = Uint8x4::from_ptr(src2_ptr + 3 * src_pitch); | |
| 554 | |||
| 555 | ✗ | src_1 = Int32x4::convert_from(src8_1); | |
| 556 | ✗ | src_2 = Int32x4::convert_from(src8_2); | |
| 557 | ✗ | src_3 = Int32x4::convert_from(src8_3); | |
| 558 | ✗ | src_4 = Int32x4::convert_from(src8_4); | |
| 559 | } | ||
| 560 | else { | ||
| 561 | // uint16_t | ||
| 562 | // only lo 64 bit (4x16) is used | ||
| 563 | auto src16_1 = Int16x8::from_ptr_lo(reinterpret_cast<const short* AVS_RESTRICT>(src2_ptr)); | ||
| 564 | ✗ | auto src16_2 = Int16x8::from_ptr_lo(reinterpret_cast<const short* AVS_RESTRICT>(src2_ptr + src_pitch)); | |
| 565 | ✗ | auto src16_3 = Int16x8::from_ptr_lo(reinterpret_cast<const short* AVS_RESTRICT>(src2_ptr + 2 * src_pitch)); | |
| 566 | ✗ | auto src16_4 = Int16x8::from_ptr_lo(reinterpret_cast<const short* AVS_RESTRICT>(src2_ptr + 3 * src_pitch)); | |
| 567 | |||
| 568 | // Turn unsigned to signed 16 bit, will be adjusted back before scaling back and storing. | ||
| 569 | // Since there's a little hope that short*short pattern is recognized by the compiler, though | ||
| 570 | // this is different from the horizontal case where hadd can be used. | ||
| 571 | if constexpr (!lessthan16bit) { | ||
| 572 | src16_1 += shifttosigned_short; | ||
| 573 | src16_2 += shifttosigned_short; | ||
| 574 | src16_3 += shifttosigned_short; | ||
| 575 | src16_4 += shifttosigned_short; | ||
| 576 | } | ||
| 577 | // widen short->int | ||
| 578 | ✗ | src_1 = Int32x4::convert_from_lo(src16_1); | |
| 579 | ✗ | src_2 = Int32x4::convert_from_lo(src16_2); | |
| 580 | ✗ | src_3 = Int32x4::convert_from_lo(src16_3); | |
| 581 | ✗ | src_4 = Int32x4::convert_from_lo(src16_4); | |
| 582 | } | ||
| 583 | |||
| 584 | ✗ | result += src_1 * coeff_1; | |
| 585 | ✗ | result_2 += src_2 * coeff_2; | |
| 586 | ✗ | result_3 += src_3 * coeff_3; | |
| 587 | ✗ | result_4 += src_4 * coeff_4; | |
| 588 | |||
| 589 | ✗ | src2_ptr += 4 * src_pitch; | |
| 590 | } | ||
| 591 | |||
| 592 | result += result_2; | ||
| 593 | result_3 += result_4; | ||
| 594 | result += result_3; | ||
| 595 | |||
| 596 | // rest zero or one | ||
| 597 | ✗ | for (; i < kernel_size; ++i) { | |
| 598 | ✗ | Int32x4 src; | |
| 599 | ✗ | const int a_coeff = current_coeff[i]; | |
| 600 | if constexpr (sizeof(pixel_t) == 1) { | ||
| 601 | // uint8_t | ||
| 602 | src.load_from_any_intptr(src2_ptr); | ||
| 603 | } | ||
| 604 | else { | ||
| 605 | // uint16_t | ||
| 606 | auto src16 = Int16x8::from_ptr_lo(reinterpret_cast<const short* AVS_RESTRICT>(src2_ptr)); | ||
| 607 | if constexpr (!lessthan16bit) { | ||
| 608 | src16 += shifttosigned_short; | ||
| 609 | } | ||
| 610 | ✗ | src = Int32x4::convert_from_lo(src16); // widen short->int | |
| 611 | } | ||
| 612 | |||
| 613 | ✗ | result += src * a_coeff; | |
| 614 | |||
| 615 | ✗ | src2_ptr += src_pitch; | |
| 616 | } | ||
| 617 | |||
| 618 | // back to unsigned 16 bit for exact 16 bit source | ||
| 619 | if constexpr (sizeof(pixel_t) == 2 && !lessthan16bit) { | ||
| 620 | result += shiftfromsigned_int; | ||
| 621 | } | ||
| 622 | |||
| 623 | result >>= fixpoint_scaler_bits; | ||
| 624 | |||
| 625 | if constexpr (sizeof(pixel_t) == 1) { | ||
| 626 | ✗ | Uint8x4 result8; | |
| 627 | convert_and_saturate_int32x4_to_uint8x4(result, result8); | ||
| 628 | ✗ | result8.store(dst + x); | |
| 629 | } | ||
| 630 | else { | ||
| 631 | ✗ | Uint16x4 result16; | |
| 632 | if constexpr (lessthan16bit) { | ||
| 633 | ✗ | convert_and_saturate_int32x4_to_uint16x4_limit(result, result16, limit); | |
| 634 | } | ||
| 635 | else { | ||
| 636 | convert_and_saturate_int32x4_to_uint16x4(result, result16); | ||
| 637 | } | ||
| 638 | ✗ | result16.store(dst + x); | |
| 639 | } | ||
| 640 | } | ||
| 641 | |||
| 642 | ✗ | dst += dst_pitch; | |
| 643 | ✗ | current_coeff += filter_size; | |
| 644 | } | ||
| 645 | ✗ | } | |
| 646 | |||
| 647 | |||
| 648 | ✗ | static void resize_v_c_planar_float_auto_vectorized(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int target_height, int bits_per_pixel) { | |
| 649 | |||
| 650 | ✗ | const float* AVS_RESTRICT current_coeff = program->pixel_coefficient_float; | |
| 651 | |||
| 652 | ✗ | auto src = reinterpret_cast<const float*>(src8); | |
| 653 | ✗ | auto dst = reinterpret_cast<float* AVS_RESTRICT>(dst8); | |
| 654 | ✗ | src_pitch = src_pitch / sizeof(float); | |
| 655 | ✗ | dst_pitch = dst_pitch / sizeof(float); | |
| 656 | |||
| 657 | ✗ | const int filter_size = program->filter_size; | |
| 658 | |||
| 659 | // 8-16 SIMD padding of coeffs is not usable here, work with single coeffs | ||
| 660 | ✗ | const int kernel_size = program->filter_size_real; | |
| 661 | ✗ | const int ksmod2 = kernel_size / 2 * 2; // Ensure kernel_size is processed in multiples of 2 | |
| 662 | |||
| 663 | ✗ | for (int y = 0; y < target_height; y++) { | |
| 664 | ✗ | int offset = program->pixel_offset[y]; | |
| 665 | ✗ | const float* src_ptr = src + offset * src_pitch; | |
| 666 | |||
| 667 | ✗ | for (int x = 0; x < width; x += 8) { | |
| 668 | // This function written in vectorizer-friendly C unexpectedly outperformed | ||
| 669 | // our hand-written SSE2 version. | ||
| 670 | // The reason: Float8 class is probably using two 128-bit XMM registers | ||
| 671 | // when compiled with SSE2 enabled by MSVC, processing 8 floats per iteration. | ||
| 672 | // Our original SSE2 function only processed 4 pixels at once, explaining the | ||
| 673 | // performance difference. We've since updated the SSE2 version to also process | ||
| 674 | // 8 pixels simultaneously using two XMM registers. | ||
| 675 | // Processing two coefficients further increases the throughput. | ||
| 676 | ✗ | Float8 result(0.f); | |
| 677 | ✗ | Float8 result_2(0.f); | |
| 678 | ✗ | const float* AVS_RESTRICT src2_ptr = src_ptr + x; | |
| 679 | |||
| 680 | ✗ | for (int i = 0; i < ksmod2; i += 2) { | |
| 681 | ✗ | const float coeff_1 = current_coeff[i]; | |
| 682 | ✗ | const float coeff_2 = current_coeff[i + 1]; | |
| 683 | ✗ | Float8 src_1 = Float8::from_ptr(src2_ptr); | |
| 684 | ✗ | Float8 src_2 = Float8::from_ptr(src2_ptr + src_pitch); | |
| 685 | ✗ | result += src_1 * coeff_1; | |
| 686 | ✗ | result_2 += src_2 * coeff_2; | |
| 687 | |||
| 688 | ✗ | src2_ptr += 2 * src_pitch; | |
| 689 | } | ||
| 690 | |||
| 691 | result += result_2; | ||
| 692 | |||
| 693 | // Process remaining coefficients if kernel_size is odd | ||
| 694 | ✗ | if (ksmod2 < kernel_size) { | |
| 695 | ✗ | const float a_coeff = current_coeff[ksmod2]; | |
| 696 | ✗ | Float8 src = Float8::from_ptr(src2_ptr); | |
| 697 | ✗ | result += src * a_coeff; | |
| 698 | } | ||
| 699 | |||
| 700 | // no rounding, no clamp | ||
| 701 | ✗ | result.store(dst + x); | |
| 702 | } | ||
| 703 | |||
| 704 | ✗ | dst += dst_pitch; | |
| 705 | ✗ | current_coeff += filter_size; | |
| 706 | } | ||
| 707 | ✗ | } | |
| 708 | |||
| 709 | |||
| 710 | /*************************************** | ||
| 711 | ********* Horizontal Resizer** ******** | ||
| 712 | ***************************************/ | ||
| 713 | |||
| 714 | // Only <float> is used, which got some SIMD-C optimizations. | ||
| 715 | // On x86 MSVC the 8-16 bit versions may perform better due to their extremely poor integer | ||
| 716 | // vectorization capabilities. | ||
| 717 | // But gcc and llvm (clangcl) compilers perform very good, their code from the other, | ||
| 718 | // vector-fiendly versions are excellent, as expected in 2025. | ||
| 719 | // Luckily, on x86 there are handcrafted SIMD versions for all 8, 10-16 and 32 bit pixel types. | ||
| 720 | |||
| 721 | template<typename pixel_t, bool lessthan16bit> | ||
| 722 | ✗ | static void resize_h_c_planar(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) { | |
| 723 | ✗ | int filter_size = program->filter_size; | |
| 724 | |||
| 725 | typedef typename std::conditional < std::is_floating_point<pixel_t>::value, float, short>::type coeff_t; | ||
| 726 | const coeff_t* AVS_RESTRICT current_coeff; | ||
| 727 | |||
| 728 | ✗ | pixel_t limit = 0; | |
| 729 | if (!std::is_floating_point<pixel_t>::value) { // floats are unscaled and uncapped | ||
| 730 | if constexpr (sizeof(pixel_t) == 1) limit = 255; | ||
| 731 | else if constexpr (sizeof(pixel_t) == 2) limit = pixel_t((1 << bits_per_pixel) - 1); | ||
| 732 | } | ||
| 733 | |||
| 734 | ✗ | src_pitch = src_pitch / sizeof(pixel_t); | |
| 735 | ✗ | dst_pitch = dst_pitch / sizeof(pixel_t); | |
| 736 | |||
| 737 | ✗ | pixel_t* src = (pixel_t*)src8; | |
| 738 | ✗ | pixel_t* dst = (pixel_t*)dst8; | |
| 739 | |||
| 740 | // for 16 bits only | ||
| 741 | ✗ | const short shifttosigned_short = -32768; | |
| 742 | ✗ | const int shiftfromsigned_int = 32768 << FPScale16bits; | |
| 743 | |||
| 744 | // perhaps helps vectorizing decision | ||
| 745 | ✗ | const int kernel_size = program->filter_size_real; | |
| 746 | ✗ | const int ksmod4 = kernel_size / 4 * 4; | |
| 747 | ✗ | const int ksmod8 = kernel_size / 8 * 8; | |
| 748 | |||
| 749 | // external loop y is much faster | ||
| 750 | ✗ | for (int y = 0; y < height; y++) { | |
| 751 | if (!std::is_floating_point<pixel_t>::value) | ||
| 752 | current_coeff = (const coeff_t* AVS_RESTRICT)program->pixel_coefficient; | ||
| 753 | else | ||
| 754 | ✗ | current_coeff = (const coeff_t* AVS_RESTRICT)program->pixel_coefficient_float; | |
| 755 | |||
| 756 | ✗ | pixel_t* AVS_RESTRICT dst2_ptr = dst + y * dst_pitch; | |
| 757 | ✗ | const pixel_t* src_ptr = src + y * src_pitch; | |
| 758 | |||
| 759 | ✗ | for (int x = 0; x < width; x++) { | |
| 760 | ✗ | int begin = program->pixel_offset[x]; | |
| 761 | ✗ | const pixel_t* AVS_RESTRICT src2_ptr = src_ptr + begin; | |
| 762 | |||
| 763 | if constexpr (std::is_floating_point<pixel_t>::value) { | ||
| 764 | ✗ | Float4 result(0.f); | |
| 765 | ✗ | for (int i = 0; i < ksmod4; i += 4) { | |
| 766 | ✗ | Float4 src = Float4::from_ptr(src2_ptr + i); | |
| 767 | ✗ | Float4 coeff = Float4::from_ptr(current_coeff + i); | |
| 768 | ✗ | result += src * coeff; | |
| 769 | } | ||
| 770 | ✗ | float result_single = result.horiz_add_float(); | |
| 771 | ✗ | for (int i = ksmod4; i < kernel_size; i++) { | |
| 772 | ✗ | result_single += src2_ptr[i] * current_coeff[i]; | |
| 773 | } | ||
| 774 | ✗ | dst2_ptr[x] = result_single; | |
| 775 | } | ||
| 776 | else if constexpr (sizeof(pixel_t) == 2) { | ||
| 777 | // theoretically, no need for int64 accumulator, | ||
| 778 | // sum of coeffs is 1.0 that is (1 << FPScale16bits) in integer arithmetic | ||
| 779 | int result = 1 << (FPScale16bits - 1); // rounder; | ||
| 780 | for (int i = 0; i < ksmod4; i += 4) { | ||
| 781 | int val; | ||
| 782 | val = src2_ptr[i+0]; | ||
| 783 | if constexpr (!lessthan16bit) | ||
| 784 | val = val + shifttosigned_short; | ||
| 785 | result += val * current_coeff[i+0]; | ||
| 786 | |||
| 787 | val = src2_ptr[i+1]; | ||
| 788 | if constexpr (!lessthan16bit) | ||
| 789 | val = val + shifttosigned_short; | ||
| 790 | result += val * current_coeff[i+1]; | ||
| 791 | |||
| 792 | val = src2_ptr[i + 2]; | ||
| 793 | if constexpr (!lessthan16bit) | ||
| 794 | val = val + shifttosigned_short; | ||
| 795 | result += val * current_coeff[i + 2]; | ||
| 796 | |||
| 797 | val = src2_ptr[i + 3]; | ||
| 798 | if constexpr (!lessthan16bit) | ||
| 799 | val = val + shifttosigned_short; | ||
| 800 | result += val * current_coeff[i + 3]; | ||
| 801 | } | ||
| 802 | for (int i = ksmod4; i < kernel_size; i++) { | ||
| 803 | int val = src2_ptr[i]; | ||
| 804 | if constexpr (!lessthan16bit) | ||
| 805 | val = val + shifttosigned_short; | ||
| 806 | result += val * current_coeff[i]; | ||
| 807 | } | ||
| 808 | if constexpr (!lessthan16bit) | ||
| 809 | result = result + shiftfromsigned_int; | ||
| 810 | result = result >> FPScale16bits; | ||
| 811 | result = result > limit ? limit : result < 0 ? 0 : result; // clamp 10..16 bits | ||
| 812 | dst2_ptr[x] = (uint16_t)result; | ||
| 813 | } | ||
| 814 | else if constexpr (sizeof(pixel_t) == 1) { | ||
| 815 | int result = 1 << (FPScale8bits - 1); // rounder; | ||
| 816 | for (int i = 0; i < ksmod8; i += 8) { | ||
| 817 | short val; | ||
| 818 | val = src2_ptr[i + 0]; | ||
| 819 | result += val * current_coeff[i + 0]; | ||
| 820 | val = src2_ptr[i + 1]; | ||
| 821 | result += val * current_coeff[i + 1]; | ||
| 822 | val = src2_ptr[i + 2]; | ||
| 823 | result += val * current_coeff[i + 2]; | ||
| 824 | val = src2_ptr[i + 3]; | ||
| 825 | result += val * current_coeff[i + 3]; | ||
| 826 | val = src2_ptr[i + 4]; | ||
| 827 | result += val * current_coeff[i + 4]; | ||
| 828 | val = src2_ptr[i + 5]; | ||
| 829 | result += val * current_coeff[i + 5]; | ||
| 830 | val = src2_ptr[i + 6]; | ||
| 831 | result += val * current_coeff[i + 6]; | ||
| 832 | val = src2_ptr[i + 7]; | ||
| 833 | result += val * current_coeff[i + 7]; | ||
| 834 | } | ||
| 835 | for (int i = ksmod8; i < ksmod4; i += 4) { | ||
| 836 | short val; | ||
| 837 | val = src2_ptr[i + 0]; | ||
| 838 | result += val * current_coeff[i + 0]; | ||
| 839 | val = src2_ptr[i + 1]; | ||
| 840 | result += val * current_coeff[i + 1]; | ||
| 841 | val = src2_ptr[i + 2]; | ||
| 842 | result += val * current_coeff[i + 2]; | ||
| 843 | val = src2_ptr[i + 3]; | ||
| 844 | result += val * current_coeff[i + 3]; | ||
| 845 | val = src2_ptr[i + 4]; | ||
| 846 | } | ||
| 847 | for (int i = ksmod4; i < kernel_size; i++) { | ||
| 848 | short val = src2_ptr[i]; | ||
| 849 | result += val * current_coeff[i]; | ||
| 850 | } | ||
| 851 | result = result >> FPScale8bits; | ||
| 852 | result = result > limit ? limit : result < 0 ? 0 : result; // clamp 8 bits | ||
| 853 | dst2_ptr[x] = (uint8_t)result; | ||
| 854 | } | ||
| 855 | ✗ | current_coeff += filter_size; | |
| 856 | } | ||
| 857 | } | ||
| 858 | ✗ | } | |
| 859 | |||
| 860 | // Vectorizer-friendly 8-16 bit horizontal resampler | ||
| 861 | |||
| 862 | // 8 pixel instead of real-SIMD 16 to ease register pressure | ||
| 863 | // and the make the compiler task easier | ||
| 864 | template<typename pixel_t, bool lessthan16bit> | ||
| 865 | AVS_FORCEINLINE static void process_two_8pixels_h_uint8_16_core(const pixel_t* AVS_RESTRICT src_ptr, const short* AVS_RESTRICT current_coeff, Int32x4& AVS_RESTRICT result, const Int16x8& shifttosigned_8xshort) { | ||
| 866 | |||
| 867 | ✗ | Int16x8 data; | |
| 868 | // using signed int16 (short) until the coeff multiplication | ||
| 869 | if constexpr (sizeof(pixel_t) == 1) { | ||
| 870 | // pixel_t is uint8_t | ||
| 871 | ✗ | Uint8x8 src = Uint8x8::from_ptr(src_ptr); | |
| 872 | ✗ | data = Int16x8::convert_from(src); // 8 pixels uint8_t->int16 | |
| 873 | } | ||
| 874 | else { | ||
| 875 | // pixel_t is uint16_t, at exact 16 bit bit depth unsigned -> signed 16 bit conversion needed | ||
| 876 | // this mimics the behavior of the SIMD version, which processes signed 16 bit input | ||
| 877 | // data.load(reinterpret_cast<const short * AVS_RESTRICT>(src_ptr)); | ||
| 878 | ✗ | Uint16x8 src = Uint16x8::from_ptr(src_ptr); | |
| 879 | ✗ | data = Int16x8::convert_from(src); // 8 pixels uint8_t->int16 | |
| 880 | |||
| 881 | if constexpr (!lessthan16bit) { | ||
| 882 | ✗ | data += shifttosigned_8xshort; // 16 bit addition is invariant regarding overflow | |
| 883 | } | ||
| 884 | } | ||
| 885 | |||
| 886 | ✗ | Int16x8 coeff = Int16x8::from_ptr(current_coeff); // 8 coeffs | |
| 887 | |||
| 888 | // no need real intel-like madd, but intel compiler can use it (experienced) | ||
| 889 | #if defined(__INTEL_LLVM_COMPILER) | ||
| 890 | // Intel proc + LLVM: only this compiler combo is compiling into madd,which is much faster. | ||
| 891 | Int32x4 madd_result = simul_madd_epi16(data, coeff); | ||
| 892 | #else | ||
| 893 | Int32x4 madd_result = mul16x16_reduce_to_Int32x4(data, coeff); | ||
| 894 | #endif | ||
| 895 | ✗ | result += madd_result; | |
| 896 | // later, the four 32 bit results will be further reduced and added together (hadd) | ||
| 897 | ✗ | } | |
| 898 | |||
| 899 | |||
| 900 | template<typename pixel_t, bool lessthan16bit> | ||
| 901 | AVS_FORCEINLINE static void process_two_4pixels_h_uint8_16_core(const pixel_t* AVS_RESTRICT src_ptr, const short* AVS_RESTRICT current_coeff, Int32x4& AVS_RESTRICT result, const Int32x4& shifttosigned_4xint) { | ||
| 902 | ✗ | Int32x4 data; | |
| 903 | |||
| 904 | // this one is using full int32 internally | ||
| 905 | |||
| 906 | if constexpr (sizeof(pixel_t) == 1) { | ||
| 907 | data.load_from_any_intptr(src_ptr); // 4 pixels uint8_t->int32 | ||
| 908 | } | ||
| 909 | else { | ||
| 910 | // pixel_t is uint16_t, at exact 16 bit bit depth an unsigned -> signed 16 bit conversion needed | ||
| 911 | data.load_from_any_intptr(src_ptr); // 4 pixels uint16_t->int32 | ||
| 912 | |||
| 913 | if constexpr (!lessthan16bit) { | ||
| 914 | ✗ | data += shifttosigned_4xint; | |
| 915 | } | ||
| 916 | } | ||
| 917 | |||
| 918 | // 4x signed 16 bit coeffs to int32 | ||
| 919 | ✗ | Int32x4 coeff_int; | |
| 920 | coeff_int.load_from_any_intptr(current_coeff); | ||
| 921 | |||
| 922 | ✗ | result += data * coeff_int; | |
| 923 | // later, the four 32 bit results will be added together (hadd) | ||
| 924 | ✗ | } | |
| 925 | |||
| 926 | template<bool safe_aligned_mode, typename pixel_t, bool lessthan16bit> | ||
| 927 | AVS_FORCEINLINE static void process_two_pixels_h_uint8_16(const pixel_t* AVS_RESTRICT src_ptr, const int begin1, const int begin2, const short* AVS_RESTRICT current_coeff, const int filter_size, | ||
| 928 | Int32x4& _result1, Int32x4& _result2, const int kernel_size, | ||
| 929 | const Int16x8& shifttosigned) | ||
| 930 | { | ||
| 931 | // Reference parameters (_result1, _result2) have known addresses, so MSVC | ||
| 932 | // treats them as aliased memory and emits scalar extract+writeback on every | ||
| 933 | // loop iteration (movd/vpextrd to memory after each vpaddd). | ||
| 934 | // By copying to local variables, MSVC gains full ownership and keeps the | ||
| 935 | // accumulators in XMM registers for the entire loop, writing back only once | ||
| 936 | // at the end. This alone gave a ~25% speedup (106 -> 183 fps on MSVC). | ||
| 937 | // this problem with Clang/GCC was not observed. | ||
| 938 | ✗ | Int32x4 result1 = _result1; | |
| 939 | ✗ | Int32x4 result2 = _result2; | |
| 940 | |||
| 941 | int ksmod8; | ||
| 942 | // 16 is too much for C optimizer | ||
| 943 | if constexpr (safe_aligned_mode) | ||
| 944 | ✗ | ksmod8 = filter_size / 8 * 8; | |
| 945 | else | ||
| 946 | ✗ | ksmod8 = kernel_size / 8 * 8; // danger zone, scanline overread possible. Use exact unaligned kernel_size | |
| 947 | ✗ | const pixel_t* src_ptr1 = src_ptr + begin1; | |
| 948 | ✗ | const pixel_t* src_ptr2 = src_ptr + begin2; | |
| 949 | ✗ | int i = 0; | |
| 950 | |||
| 951 | // Process 16 elements at a time | ||
| 952 | ✗ | for (; i < ksmod8; i += 8) { | |
| 953 | ✗ | process_two_8pixels_h_uint8_16_core<pixel_t, lessthan16bit>(src_ptr1 + i, current_coeff + i, result1, shifttosigned); | |
| 954 | ✗ | process_two_8pixels_h_uint8_16_core<pixel_t, lessthan16bit>(src_ptr2 + i, current_coeff + filter_size + i , result2, shifttosigned); | |
| 955 | } | ||
| 956 | |||
| 957 | if constexpr (!safe_aligned_mode) { | ||
| 958 | // working with the original, unaligned kernel_size | ||
| 959 | ✗ | if (i == kernel_size) { | |
| 960 | ✗ | _result1 = result1; | |
| 961 | ✗ | _result2 = result2; | |
| 962 | ✗ | return; | |
| 963 | } | ||
| 964 | |||
| 965 | ✗ | const int ksmod4 = kernel_size / 4 * 4; | |
| 966 | // Process 4 elements if needed | ||
| 967 | ✗ | if (i < ksmod4) { | |
| 968 | ✗ | Int32x4 shifttosigned_4; | |
| 969 | ✗ | shifttosigned_4.convert_from_lo(shifttosigned); // copy lower half of 8xshort, 4xshort to 4xint | |
| 970 | ✗ | process_two_4pixels_h_uint8_16_core<pixel_t, lessthan16bit>(src_ptr1 + i, current_coeff + i, result1, shifttosigned_4); | |
| 971 | ✗ | process_two_4pixels_h_uint8_16_core<pixel_t, lessthan16bit>(src_ptr2 + i, current_coeff + filter_size + i, result2, shifttosigned_4); | |
| 972 | ✗ | i += 4; | |
| 973 | ✗ | if (i == kernel_size) { | |
| 974 | ✗ | _result1 = result1; | |
| 975 | ✗ | _result2 = result2; | |
| 976 | ✗ | return; | |
| 977 | } | ||
| 978 | } | ||
| 979 | |||
| 980 | // Process remaining 1-3 elements with scalar operations | ||
| 981 | ✗ | if (i < kernel_size) { | |
| 982 | ✗ | Int32x4 scalar_sum1(0); // like an __m128i | |
| 983 | ✗ | Int32x4 scalar_sum2(0); | |
| 984 | |||
| 985 | ✗ | int index = 0; | |
| 986 | ✗ | for (; i < kernel_size; i++, index++) { | |
| 987 | |||
| 988 | if constexpr (sizeof(pixel_t) == 1) { | ||
| 989 | ✗ | scalar_sum1.set(index, scalar_sum1[index] + src_ptr1[i] * current_coeff[i]); | |
| 990 | ✗ | scalar_sum2.set(index, scalar_sum2[index] + src_ptr2[i] * current_coeff[filter_size + i]); | |
| 991 | } | ||
| 992 | else { | ||
| 993 | // pixel_t is uint16_t | ||
| 994 | ✗ | short val = src_ptr1[i]; | |
| 995 | if constexpr (!lessthan16bit) | ||
| 996 | ✗ | val = val + shifttosigned[0]; // still short | |
| 997 | ✗ | scalar_sum1.set(index, scalar_sum1[index] + val * current_coeff[i]); | |
| 998 | |||
| 999 | ✗ | val = src_ptr2[i]; | |
| 1000 | if constexpr (!lessthan16bit) | ||
| 1001 | ✗ | val = val + shifttosigned[0]; // still short | |
| 1002 | ✗ | scalar_sum2.set(index, scalar_sum2[index] + val * current_coeff[filter_size + i]); | |
| 1003 | } | ||
| 1004 | } | ||
| 1005 | |||
| 1006 | // update result vectors | ||
| 1007 | result1 += scalar_sum1; | ||
| 1008 | result2 += scalar_sum2; | ||
| 1009 | |||
| 1010 | } | ||
| 1011 | } | ||
| 1012 | |||
| 1013 | ✗ | _result1 = result1; | |
| 1014 | ✗ | _result2 = result2; | |
| 1015 | ✗ | } | |
| 1016 | |||
| 1017 | |||
| 1018 | // NO Forceinline! Helps MSVC, by starting a new stack frame and have enough registers again. | ||
| 1019 | template<bool is_safe, typename pixel_t, bool lessthan16bit> | ||
| 1020 | ✗ | static void process_eight_pixels_h_uint8_16(const pixel_t * AVS_RESTRICT src, int x, const short* current_coeff_base, int filter_size, | |
| 1021 | const Int32x4& rounder128, const Int16x8& shifttosigned, const uint16_t clamp_limit, | ||
| 1022 | pixel_t* AVS_RESTRICT dst, | ||
| 1023 | ResamplingProgram* program) | ||
| 1024 | { | ||
| 1025 | ✗ | assert(program->filter_size_alignment >= 16); // code assumes this | |
| 1026 | |||
| 1027 | ✗ | const short* AVS_RESTRICT current_coeff = current_coeff_base + x * filter_size; | |
| 1028 | ✗ | const int unaligned_kernel_size = program->filter_size_real; | |
| 1029 | |||
| 1030 | // Unrolled processing of all 8 pixels | ||
| 1031 | |||
| 1032 | // 0 & 1 | ||
| 1033 | ✗ | Int32x4 result0 = rounder128; | |
| 1034 | ✗ | Int32x4 result1 = rounder128; | |
| 1035 | ✗ | int begin0 = program->pixel_offset[x + 0]; | |
| 1036 | ✗ | int begin1 = program->pixel_offset[x + 1]; | |
| 1037 | process_two_pixels_h_uint8_16<is_safe, pixel_t, lessthan16bit>(src, begin0, begin1, current_coeff, filter_size, result0, result1, unaligned_kernel_size, shifttosigned); | ||
| 1038 | ✗ | current_coeff += 2 * filter_size; | |
| 1039 | |||
| 1040 | // 2 & 3 | ||
| 1041 | ✗ | Int32x4 result2 = rounder128; | |
| 1042 | ✗ | Int32x4 result3 = rounder128; | |
| 1043 | ✗ | begin0 = program->pixel_offset[x + 2]; | |
| 1044 | ✗ | begin1 = program->pixel_offset[x + 3]; | |
| 1045 | process_two_pixels_h_uint8_16<is_safe, pixel_t, lessthan16bit>(src, begin0, begin1, current_coeff, filter_size, result2, result3, unaligned_kernel_size, shifttosigned); | ||
| 1046 | ✗ | current_coeff += 2 * filter_size; | |
| 1047 | |||
| 1048 | ✗ | Int32x4 sumQuad1234; | |
| 1049 | ✗ | sumQuad1234 = make_from_horiz_sums(result0, result1, result2, result3); | |
| 1050 | /* provided and optimized from simd_c.h | ||
| 1051 | sumQuad1234.set(0, result1.horiz_add_int32()); | ||
| 1052 | sumQuad1234.set(1, result1.horiz_add_int32()); | ||
| 1053 | sumQuad1234.set(2, result2.horiz_add_int32()); | ||
| 1054 | sumQuad1234.set(3, result3.horiz_add_int32()); | ||
| 1055 | */ | ||
| 1056 | |||
| 1057 | // 4 & 5 | ||
| 1058 | ✗ | result0 = rounder128; | |
| 1059 | ✗ | result1 = rounder128; | |
| 1060 | ✗ | begin0 = program->pixel_offset[x + 4]; | |
| 1061 | ✗ | begin1 = program->pixel_offset[x + 5]; | |
| 1062 | process_two_pixels_h_uint8_16<is_safe, pixel_t, lessthan16bit>(src, begin0, begin1, current_coeff, filter_size, result0, result1, unaligned_kernel_size, shifttosigned); | ||
| 1063 | ✗ | current_coeff += 2 * filter_size; | |
| 1064 | |||
| 1065 | // 6 & 7 | ||
| 1066 | ✗ | result2 = rounder128; | |
| 1067 | ✗ | result3 = rounder128; | |
| 1068 | ✗ | begin0 = program->pixel_offset[x + 6]; | |
| 1069 | ✗ | begin1 = program->pixel_offset[x + 7]; | |
| 1070 | process_two_pixels_h_uint8_16<is_safe, pixel_t, lessthan16bit>(src, begin0, begin1, current_coeff, filter_size, result2, result3, unaligned_kernel_size, shifttosigned); | ||
| 1071 | // current_coeff += 2 * filter_size; // not needed anymore | ||
| 1072 | |||
| 1073 | ✗ | Int32x4 sumQuad5678; | |
| 1074 | ✗ | sumQuad5678 = make_from_horiz_sums(result0, result1, result2, result3); | |
| 1075 | |||
| 1076 | // correct if signed, scale back, store | ||
| 1077 | if constexpr (sizeof(pixel_t) == 2 && !lessthan16bit) { | ||
| 1078 | ✗ | const Int32x4 shiftfromsigned(32768 << FPScale16bits); | |
| 1079 | sumQuad1234 += shiftfromsigned; | ||
| 1080 | sumQuad5678 += shiftfromsigned; | ||
| 1081 | } | ||
| 1082 | |||
| 1083 | ✗ | constexpr int current_fp_scale_bits = (sizeof(pixel_t) == 1) ? FPScale8bits : FPScale16bits; | |
| 1084 | // scale back, store | ||
| 1085 | sumQuad1234 >>= current_fp_scale_bits; | ||
| 1086 | sumQuad5678 >>= current_fp_scale_bits; | ||
| 1087 | |||
| 1088 | if constexpr (sizeof(pixel_t) == 1) { | ||
| 1089 | ✗ | Uint8x8 result_2x4x_uint8; | |
| 1090 | convert_and_saturate_int32x4x2_to_uint8x8(sumQuad1234, sumQuad5678, result_2x4x_uint8); | ||
| 1091 | ✗ | result_2x4x_uint8.store(&dst[x]); | |
| 1092 | } | ||
| 1093 | else { | ||
| 1094 | // uint16_t 10-16 bit | ||
| 1095 | ✗ | Uint16x8 result_2x4x_uint16_128; | |
| 1096 | if constexpr (lessthan16bit) { | ||
| 1097 | ✗ | convert_and_saturate_int32x4x2_to_uint16x8_limit(sumQuad1234, sumQuad5678, result_2x4x_uint16_128, clamp_limit); | |
| 1098 | } | ||
| 1099 | else { | ||
| 1100 | convert_and_saturate_int32x4x2_to_uint16x8(sumQuad1234, sumQuad5678, result_2x4x_uint16_128); | ||
| 1101 | } | ||
| 1102 | ✗ | result_2x4x_uint16_128.store(&dst[x]); | |
| 1103 | |||
| 1104 | } | ||
| 1105 | ✗ | } | |
| 1106 | |||
| 1107 | //-------- uint8/16_t Horizontal | ||
| 1108 | // 4 pixels at a time. | ||
| 1109 | template<typename pixel_t, bool lessthan16bit> | ||
| 1110 | ✗ | void resizer_h_c_generic_uint8_16_vectorized(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel) { | |
| 1111 | |||
| 1112 | ✗ | const int filter_size = program->filter_size; | |
| 1113 | ✗ | const int current_fp_scale_bits = (sizeof(pixel_t) == 1) ? FPScale8bits : FPScale16bits; | |
| 1114 | ✗ | const Int32x4 rounder128 = { 1 << (current_fp_scale_bits - 1), 0, 0, 0 }; | |
| 1115 | |||
| 1116 | ✗ | const Int16x8 shifttosigned_or_zero128(sizeof(pixel_t) == 1 ? 0 : -32768); | |
| 1117 | |||
| 1118 | ✗ | const uint16_t clamp_limit = (1 << bits_per_pixel) - 1; | |
| 1119 | |||
| 1120 | ✗ | const pixel_t* AVS_RESTRICT src = reinterpret_cast<const pixel_t*>(src8); | |
| 1121 | ✗ | pixel_t* AVS_RESTRICT dst = reinterpret_cast<pixel_t*>(dst8); | |
| 1122 | ✗ | dst_pitch /= sizeof(pixel_t); | |
| 1123 | ✗ | src_pitch /= sizeof(pixel_t); | |
| 1124 | |||
| 1125 | ✗ | const int w_safe_mod8 = (program->safelimit_filter_size_aligned.overread_possible ? program->safelimit_filter_size_aligned.source_overread_beyond_targetx : width) / 8 * 8; | |
| 1126 | |||
| 1127 | ✗ | for (int y = 0; y < height; y++) { | |
| 1128 | ✗ | const short* current_coeff_base = program->pixel_coefficient; | |
| 1129 | |||
| 1130 | // Process safe aligned pixels | ||
| 1131 | ✗ | for (int x = 0; x < w_safe_mod8; x += 8) { | |
| 1132 | ✗ | process_eight_pixels_h_uint8_16<true, pixel_t, lessthan16bit>(src, x, current_coeff_base, filter_size, rounder128, shifttosigned_or_zero128, clamp_limit, dst, program); | |
| 1133 | } | ||
| 1134 | |||
| 1135 | // Process up to the actual kernel size instead of the aligned filter_size to prevent overreading beyond the last source pixel. | ||
| 1136 | // We assume extra offset entries were added to the p->pixel_offset array (aligned to 8 during initialization). | ||
| 1137 | // This may store 1-7 false pixels but it still remain in alignment-safe area. | ||
| 1138 | ✗ | for (int x = w_safe_mod8; x < width; x += 8) { | |
| 1139 | ✗ | process_eight_pixels_h_uint8_16<false, pixel_t, lessthan16bit>(src, x, current_coeff_base, filter_size, rounder128, shifttosigned_or_zero128, clamp_limit, dst, program); | |
| 1140 | } | ||
| 1141 | |||
| 1142 | ✗ | dst += dst_pitch; | |
| 1143 | ✗ | src += src_pitch; | |
| 1144 | } | ||
| 1145 | ✗ | } | |
| 1146 | |||
| 1147 | // 16 bit Horizontal | ||
| 1148 | |||
| 1149 | template void resizer_h_c_generic_uint8_16_vectorized<uint8_t, true>(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel); | ||
| 1150 | template void resizer_h_c_generic_uint8_16_vectorized<uint16_t, false>(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel); | ||
| 1151 | template void resizer_h_c_generic_uint8_16_vectorized<uint16_t, true>(BYTE* dst8, const BYTE* src8, int dst_pitch, int src_pitch, ResamplingProgram* program, int width, int height, int bits_per_pixel); | ||
| 1152 | |||
| 1153 | /******************************************************************** | ||
| 1154 | ***** Declare index of new filters for Avisynth's filter engine ***** | ||
| 1155 | ********************************************************************/ | ||
| 1156 | |||
| 1157 | extern const AVSFunction Resample_filters[] = { | ||
| 1158 | { "PointResize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_PointResize }, | ||
| 1159 | { "BilinearResize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_BilinearResize }, | ||
| 1160 | { "BicubicResize", BUILTIN_FUNC_PREFIX, "cii[b]f[c]f[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_BicubicResize }, | ||
| 1161 | { "LanczosResize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[taps]i[force]i[keep_center]b[placement]s", FilteredResize::Create_LanczosResize}, | ||
| 1162 | { "Lanczos4Resize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_Lanczos4Resize}, | ||
| 1163 | { "BlackmanResize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[taps]i[force]i[keep_center]b[placement]s", FilteredResize::Create_BlackmanResize}, | ||
| 1164 | { "Spline16Resize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_Spline16Resize}, | ||
| 1165 | { "Spline36Resize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_Spline36Resize}, | ||
| 1166 | { "Spline64Resize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_Spline64Resize}, | ||
| 1167 | { "GaussResize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[p]f[b]f[s]f[force]i[keep_center]b[placement]s", FilteredResize::Create_GaussianResize}, | ||
| 1168 | { "SincResize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[taps]i[force]i[keep_center]b[placement]s", FilteredResize::Create_SincResize}, | ||
| 1169 | { "SinPowerResize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[p]f[force]i[keep_center]b[placement]s", FilteredResize::Create_SinPowerResize}, | ||
| 1170 | { "SincLin2Resize", BUILTIN_FUNC_PREFIX, "cii[src_left]f[src_top]f[src_width]f[src_height]f[taps]i[force]i[keep_center]b[placement]s", FilteredResize::Create_SincLin2Resize}, | ||
| 1171 | { "UserDefined2Resize", BUILTIN_FUNC_PREFIX, "cii[b]f[c]f[s]f[src_left]f[src_top]f[src_width]f[src_height]f[force]i[keep_center]b[placement]s", FilteredResize::Create_UserDefined2Resize}, | ||
| 1172 | /** | ||
| 1173 | * Resize(PClip clip, dst_width, dst_height [src_left, src_top, src_width, int src_height,] ) | ||
| 1174 | * | ||
| 1175 | * src_left et al. = when these optional arguments are given, the filter acts just like | ||
| 1176 | * a Crop was performed with those parameters before resizing, only faster | ||
| 1177 | **/ | ||
| 1178 | |||
| 1179 | { 0 } | ||
| 1180 | }; | ||
| 1181 | |||
| 1182 | // Borrowed from fmtconv | ||
| 1183 | // ChromaPlacement.cpp | ||
| 1184 | // Author : Laurent de Soras, 2015 | ||
| 1185 | |||
| 1186 | // Fixes the vertical chroma placement when the picture is interlaced. | ||
| 1187 | // ofs = ordinate to skip between TFF and BFF, relative to the chroma grid. A | ||
| 1188 | // single line of full-res picture is 0.25. | ||
| 1189 | 12 | static inline void ChromaPlacement_fix_itl(double& cp_v, bool interlaced_flag, bool top_flag, double ofs = 0.5) | |
| 1190 | { | ||
| 1191 |
1/2✗ Branch 2 → 3 not taken.
✓ Branch 2 → 4 taken 12 times.
|
12 | assert(cp_v >= 0); |
| 1192 | |||
| 1193 |
1/2✗ Branch 4 → 5 not taken.
✓ Branch 4 → 7 taken 12 times.
|
12 | if (interlaced_flag) |
| 1194 | { | ||
| 1195 | ✗ | cp_v *= 0.5; | |
| 1196 | ✗ | if (!top_flag) | |
| 1197 | { | ||
| 1198 | ✗ | cp_v += ofs; | |
| 1199 | } | ||
| 1200 | } | ||
| 1201 | 12 | } | |
| 1202 | /* | ||
| 1203 | ss_h and ss_v are log2(subsampling) | ||
| 1204 | rgb_flag actually means that chroma subsampling doesn't apply. | ||
| 1205 | |||
| 1206 | http://www.mir.com/DMG/chroma.html | ||
| 1207 | |||
| 1208 | cp_* is the position of the sampling point relative to the frame | ||
| 1209 | top/left border, in the plane coordinates. For reference, the border | ||
| 1210 | of the frame is at 0.5 units of luma from the first luma sampling point. | ||
| 1211 | I. e., the luma sampling point is at the pixel's center. | ||
| 1212 | */ | ||
| 1213 | |||
| 1214 | // PF added BOTTOM, BOTTOM_LEFT, TOP | ||
| 1215 | // Pass ChromaLocation_e::AVS_CHROMA_UNUSED for defaults | ||
| 1216 | // plane index 0:Y, 1:U, 2:V | ||
| 1217 | // cplace is a ChromaLocation_e constant | ||
| 1218 | 10 | static void ChromaPlacement_compute_cplace(double& cp_h, double& cp_v, int cplace, int plane_index, int ss_h, int ss_v, bool rgb_flag, bool interlaced_flag, bool top_flag) | |
| 1219 | { | ||
| 1220 |
3/4✓ Branch 2 → 3 taken 4 times.
✓ Branch 2 → 5 taken 6 times.
✗ Branch 3 → 4 not taken.
✓ Branch 3 → 5 taken 4 times.
|
10 | assert(cplace >= 0 || cplace == ChromaLocation_e::AVS_CHROMA_UNUSED); |
| 1221 |
1/2✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 10 times.
|
10 | assert(cplace < ChromaLocation_e::AVS_CHROMA_DV); |
| 1222 |
1/2✗ Branch 7 → 8 not taken.
✓ Branch 7 → 9 taken 10 times.
|
10 | assert(ss_h >= 0); |
| 1223 |
1/2✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 10 times.
|
10 | assert(ss_v >= 0); |
| 1224 |
1/2✗ Branch 11 → 12 not taken.
✓ Branch 11 → 13 taken 10 times.
|
10 | assert(plane_index >= 0); |
| 1225 | |||
| 1226 | // Generic case for luma, non-subsampled chroma and center (MPEG-1) chroma. | ||
| 1227 | 10 | cp_h = 0.5; | |
| 1228 | 10 | cp_v = 0.5; | |
| 1229 | 10 | ChromaPlacement_fix_itl(cp_v, interlaced_flag, top_flag); | |
| 1230 | |||
| 1231 | // Subsampled chroma | ||
| 1232 |
2/4✓ Branch 14 → 15 taken 10 times.
✗ Branch 14 → 37 not taken.
✓ Branch 15 → 16 taken 10 times.
✗ Branch 15 → 37 not taken.
|
10 | if (!rgb_flag && plane_index > 0) |
| 1233 | { | ||
| 1234 |
2/2✓ Branch 16 → 17 taken 6 times.
✓ Branch 16 → 22 taken 4 times.
|
10 | if (ss_h > 0) // horizontal subsampling 420 411 |
| 1235 | { | ||
| 1236 |
2/2✓ Branch 17 → 18 taken 4 times.
✓ Branch 17 → 21 taken 2 times.
|
6 | if (cplace == ChromaLocation_e::AVS_CHROMA_LEFT // mpeg2 |
| 1237 |
1/2✓ Branch 18 → 19 taken 4 times.
✗ Branch 18 → 21 not taken.
|
4 | || cplace == ChromaLocation_e::AVS_CHROMA_DV |
| 1238 |
2/2✓ Branch 19 → 20 taken 3 times.
✓ Branch 19 → 21 taken 1 time.
|
4 | || cplace == ChromaLocation_e::AVS_CHROMA_TOP_LEFT |
| 1239 |
1/2✗ Branch 20 → 21 not taken.
✓ Branch 20 → 22 taken 3 times.
|
3 | || cplace == ChromaLocation_e::AVS_CHROMA_BOTTOM_LEFT |
| 1240 | ) | ||
| 1241 | { | ||
| 1242 | 3 | cp_h = 0.5 / (1 << ss_h); | |
| 1243 | } | ||
| 1244 | } | ||
| 1245 | |||
| 1246 |
2/2✓ Branch 22 → 23 taken 4 times.
✓ Branch 22 → 37 taken 6 times.
|
10 | if (ss_v == 1) // vertical subsampling 420, 422 |
| 1247 | { | ||
| 1248 |
2/2✓ Branch 23 → 24 taken 1 time.
✓ Branch 23 → 25 taken 3 times.
|
4 | if (cplace == ChromaLocation_e::AVS_CHROMA_LEFT) |
| 1249 | { | ||
| 1250 | 1 | cp_v = 0.5; | |
| 1251 | 1 | ChromaPlacement_fix_itl(cp_v, interlaced_flag, top_flag); | |
| 1252 | } | ||
| 1253 |
1/2✓ Branch 25 → 26 taken 3 times.
✗ Branch 25 → 28 not taken.
|
3 | else if (cplace == ChromaLocation_e::AVS_CHROMA_DV |
| 1254 |
2/2✓ Branch 26 → 27 taken 2 times.
✓ Branch 26 → 28 taken 1 time.
|
3 | || cplace == ChromaLocation_e::AVS_CHROMA_TOP_LEFT |
| 1255 |
1/2✗ Branch 27 → 28 not taken.
✓ Branch 27 → 33 taken 2 times.
|
2 | || cplace == ChromaLocation_e::AVS_CHROMA_TOP |
| 1256 | ) | ||
| 1257 | { | ||
| 1258 | 1 | cp_v = 0.25; | |
| 1259 | 1 | ChromaPlacement_fix_itl(cp_v, interlaced_flag, top_flag, 0.25); | |
| 1260 | |||
| 1261 |
1/4✗ Branch 29 → 30 not taken.
✓ Branch 29 → 32 taken 1 time.
✗ Branch 30 → 31 not taken.
✗ Branch 30 → 32 not taken.
|
1 | if (cplace == ChromaLocation_e::AVS_CHROMA_DV && plane_index == 2) // V |
| 1262 | { | ||
| 1263 | ✗ | cp_v += 0.5; | |
| 1264 | } | ||
| 1265 | } | ||
| 1266 |
1/2✓ Branch 33 → 34 taken 2 times.
✗ Branch 33 → 35 not taken.
|
2 | else if (cplace == ChromaLocation_e::AVS_CHROMA_BOTTOM_LEFT |
| 1267 |
1/2✗ Branch 34 → 35 not taken.
✓ Branch 34 → 36 taken 2 times.
|
2 | || cplace == ChromaLocation_e::AVS_CHROMA_BOTTOM |
| 1268 | ) | ||
| 1269 | { | ||
| 1270 | ✗ | cp_v = 0.75; | |
| 1271 | ✗ | ChromaPlacement_fix_itl(cp_v, interlaced_flag, top_flag, 0.25); | |
| 1272 | } | ||
| 1273 | } // ss_v == 1 | ||
| 1274 | } | ||
| 1275 | 10 | } | |
| 1276 | |||
| 1277 | |||
| 1278 | // returns the requested horizontal or vertical pixel center position | ||
| 1279 | 42 | static void GetCenterShiftForResizers(double& center_pos_luma, double& center_pos_chroma, bool preserve_center, int chroma_placement, VideoInfo &vi, bool for_horizontal) { | |
| 1280 | 42 | double center_pos_h_luma = 0.0; | |
| 1281 | 42 | double center_pos_v_luma = 0.0; | |
| 1282 | // if not needed, these won't be used | ||
| 1283 | 42 | double center_pos_h_chroma = 0.0; | |
| 1284 | 42 | double center_pos_v_chroma = 0.0; | |
| 1285 | |||
| 1286 | // chroma, only if applicable | ||
| 1287 |
8/8✓ Branch 3 → 4 taken 38 times.
✓ Branch 3 → 9 taken 4 times.
✓ Branch 5 → 6 taken 14 times.
✓ Branch 5 → 9 taken 24 times.
✓ Branch 7 → 8 taken 12 times.
✓ Branch 7 → 9 taken 2 times.
✓ Branch 10 → 11 taken 12 times.
✓ Branch 10 → 17 taken 30 times.
|
42 | if (vi.IsPlanar() && vi.NumComponents() > 1 && !vi.IsRGB()) { |
| 1288 | 12 | double cp_s_h = 0; | |
| 1289 | 12 | double cp_s_v = 0; | |
| 1290 | |||
| 1291 |
2/2✓ Branch 11 → 12 taken 10 times.
✓ Branch 11 → 16 taken 2 times.
|
12 | if (preserve_center) { |
| 1292 | // same for source and destination | ||
| 1293 | 10 | int plane_index = 1; // U | |
| 1294 |
1/2✓ Branch 12 → 13 taken 10 times.
✗ Branch 12 → 24 not taken.
|
10 | int src_ss_h = vi.GetPlaneWidthSubsampling(PLANAR_U); |
| 1295 |
1/2✓ Branch 13 → 14 taken 10 times.
✗ Branch 13 → 24 not taken.
|
10 | int src_ss_v = vi.GetPlaneHeightSubsampling(PLANAR_U); |
| 1296 | |||
| 1297 | // int chromaplace = ChromaLocation_e::AVS_CHROMA_CENTER; // MPEG1 | ||
| 1298 | |||
| 1299 | 10 | ChromaPlacement_compute_cplace( | |
| 1300 | cp_s_h, cp_s_v, chroma_placement, plane_index, src_ss_h, src_ss_v, | ||
| 1301 |
1/2✓ Branch 14 → 15 taken 10 times.
✗ Branch 14 → 24 not taken.
|
10 | vi.IsRGB(), |
| 1302 | false, // interlacing flag, we don't handle it here | ||
| 1303 | false // top_flag, we don't handle it here | ||
| 1304 | ); | ||
| 1305 | } | ||
| 1306 | |||
| 1307 | 12 | center_pos_h_chroma = cp_s_h; | |
| 1308 | 12 | center_pos_v_chroma = cp_s_v; | |
| 1309 | } | ||
| 1310 | |||
| 1311 | // luma/rgb planes | ||
| 1312 |
2/2✓ Branch 17 → 18 taken 40 times.
✓ Branch 17 → 19 taken 2 times.
|
42 | if (preserve_center) { |
| 1313 | 40 | center_pos_h_luma = 0.5; | |
| 1314 | 40 | center_pos_v_luma = 0.5; | |
| 1315 | } | ||
| 1316 | else { | ||
| 1317 | 2 | center_pos_h_luma = 0.0; | |
| 1318 | 2 | center_pos_v_luma = 0.0; | |
| 1319 | } | ||
| 1320 | |||
| 1321 | // fill return ref values | ||
| 1322 |
2/2✓ Branch 20 → 21 taken 21 times.
✓ Branch 20 → 22 taken 21 times.
|
42 | if (for_horizontal) { |
| 1323 | 21 | center_pos_luma = center_pos_h_luma; | |
| 1324 | 21 | center_pos_chroma = center_pos_h_chroma; | |
| 1325 | } | ||
| 1326 | else { | ||
| 1327 | // vertical | ||
| 1328 | 21 | center_pos_luma = center_pos_v_luma; | |
| 1329 | 21 | center_pos_chroma = center_pos_v_chroma; | |
| 1330 | } | ||
| 1331 | |||
| 1332 | 42 | } | |
| 1333 | |||
| 1334 | 22 | FilteredResizeH::FilteredResizeH(PClip _child, double subrange_left, double subrange_width, | |
| 1335 | 22 | int target_width, ResamplingFunction* func, bool preserve_center, int chroma_placement, IScriptEnvironment* env) | |
| 1336 | : GenericVideoFilter(_child), | ||
| 1337 | 22 | resampling_program_luma(nullptr), resampling_program_chroma(nullptr), | |
| 1338 | 22 | resampler_h_luma(nullptr), resampler_h_chroma(nullptr), | |
| 1339 | 22 | resampler_h_luma_mt(nullptr), resampler_h_chroma_mt(nullptr), | |
| 1340 | 22 | use_alternative_h_avx512_mt_fallback(false), | |
| 1341 | 22 | resampler_luma(nullptr), resampler_chroma(nullptr), | |
| 1342 |
2/4✓ Branch 2 → 3 taken 22 times.
✗ Branch 2 → 105 not taken.
✓ Branch 3 → 4 taken 22 times.
✗ Branch 3 → 103 not taken.
|
22 | num_threads(0) |
| 1343 | |||
| 1344 | { | ||
| 1345 | 22 | src_width = vi.width; | |
| 1346 | 22 | src_height = vi.height; | |
| 1347 | 22 | dst_width = target_width; | |
| 1348 | 22 | dst_height = vi.height; | |
| 1349 | |||
| 1350 |
1/2✓ Branch 5 → 6 taken 22 times.
✗ Branch 5 → 106 not taken.
|
22 | pixelsize = vi.ComponentSize(); // AVS16 |
| 1351 |
1/2✓ Branch 6 → 7 taken 22 times.
✗ Branch 6 → 106 not taken.
|
22 | bits_per_pixel = vi.BitsPerComponent(); |
| 1352 |
1/2✓ Branch 7 → 8 taken 22 times.
✗ Branch 7 → 106 not taken.
|
22 | grey = vi.IsY(); |
| 1353 | |||
| 1354 |
5/8✓ Branch 8 → 9 taken 22 times.
✗ Branch 8 → 106 not taken.
✓ Branch 9 → 10 taken 21 times.
✓ Branch 9 → 12 taken 1 time.
✓ Branch 10 → 11 taken 21 times.
✗ Branch 10 → 106 not taken.
✗ Branch 11 → 12 not taken.
✓ Branch 11 → 13 taken 21 times.
|
22 | bool isRGBPfamily = vi.IsPlanarRGB() || vi.IsPlanarRGBA(); |
| 1355 | |||
| 1356 |
1/2✗ Branch 14 → 15 not taken.
✓ Branch 14 → 16 taken 22 times.
|
22 | if (target_width <= 0) { |
| 1357 | ✗ | env->ThrowError("Resize: Width must be greater than 0."); | |
| 1358 | } | ||
| 1359 | |||
| 1360 |
9/10✓ Branch 16 → 17 taken 22 times.
✗ Branch 16 → 106 not taken.
✓ Branch 17 → 18 taken 20 times.
✓ Branch 17 → 21 taken 2 times.
✓ Branch 18 → 19 taken 8 times.
✓ Branch 18 → 21 taken 12 times.
✓ Branch 19 → 20 taken 7 times.
✓ Branch 19 → 21 taken 1 time.
✓ Branch 22 → 23 taken 7 times.
✓ Branch 22 → 26 taken 15 times.
|
22 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { |
| 1361 |
1/2✓ Branch 23 → 24 taken 7 times.
✗ Branch 23 → 106 not taken.
|
7 | const int mask = (1 << vi.GetPlaneWidthSubsampling(PLANAR_U)) - 1; |
| 1362 | |||
| 1363 |
2/2✓ Branch 24 → 25 taken 1 time.
✓ Branch 24 → 26 taken 6 times.
|
7 | if (target_width & mask) |
| 1364 |
1/2✗ Branch 25 → 26 not taken.
✓ Branch 25 → 106 taken 1 time.
|
1 | env->ThrowError("Resize: Planar destination height must be a multiple of %d.", mask + 1); |
| 1365 | } | ||
| 1366 | |||
| 1367 | double center_pos_h_luma; | ||
| 1368 | double center_pos_h_chroma; | ||
| 1369 |
1/2✓ Branch 26 → 27 taken 21 times.
✗ Branch 26 → 106 not taken.
|
21 | GetCenterShiftForResizers(center_pos_h_luma, center_pos_h_chroma, preserve_center, chroma_placement, vi, true /* for horizontal */); |
| 1370 | // 3.7.4- parameter, old Avisynth behavior: 0.5, 0.5 | ||
| 1371 | |||
| 1372 | // Main resampling program | ||
| 1373 |
1/2✓ Branch 27 → 28 taken 21 times.
✗ Branch 27 → 106 not taken.
|
21 | resampling_program_luma = func->GetResamplingProgram(vi.width, subrange_left, subrange_width, target_width, bits_per_pixel, |
| 1374 | center_pos_h_luma, center_pos_h_luma, // for resizing it's the same for source and dest | ||
| 1375 | env); | ||
| 1376 |
9/10✓ Branch 28 → 29 taken 21 times.
✗ Branch 28 → 106 not taken.
✓ Branch 29 → 30 taken 19 times.
✓ Branch 29 → 33 taken 2 times.
✓ Branch 30 → 31 taken 7 times.
✓ Branch 30 → 33 taken 12 times.
✓ Branch 31 → 32 taken 6 times.
✓ Branch 31 → 33 taken 1 time.
✓ Branch 34 → 35 taken 6 times.
✓ Branch 34 → 38 taken 15 times.
|
21 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { |
| 1377 |
1/2✓ Branch 35 → 36 taken 6 times.
✗ Branch 35 → 106 not taken.
|
6 | const int shift = vi.GetPlaneWidthSubsampling(PLANAR_U); |
| 1378 | 6 | const int div = 1 << shift; | |
| 1379 | |||
| 1380 | |||
| 1381 | 6 | resampling_program_chroma = func->GetResamplingProgram( | |
| 1382 |
1/2✓ Branch 36 → 37 taken 6 times.
✗ Branch 36 → 106 not taken.
|
6 | vi.width >> shift, |
| 1383 | subrange_left / div, | ||
| 1384 | subrange_width / div, | ||
| 1385 | target_width >> shift, | ||
| 1386 | bits_per_pixel, | ||
| 1387 | center_pos_h_chroma, center_pos_h_chroma, // horizontal | ||
| 1388 | env); | ||
| 1389 | } | ||
| 1390 | |||
| 1391 | // when not fast_resize, then we use vertical resizers between turnleft/turnright | ||
| 1392 | #ifdef INTEL_INTRINSICS | ||
| 1393 |
1/2✓ Branch 38 → 39 taken 21 times.
✗ Branch 38 → 106 not taken.
|
21 | int cpu = env->GetCPUFlags(); |
| 1394 | 21 | bool has_sse2 = (cpu & CPUF_SSE2) != 0; | |
| 1395 | 21 | bool has_avx2 = (cpu & CPUF_AVX2) != 0; | |
| 1396 | #elif defined(NEON_INTRINSICS) | ||
| 1397 | int cpu = env->GetCPUFlags(); | ||
| 1398 | bool has_neon = (cpu & CPUF_ARM_NEON) != 0; | ||
| 1399 | #else | ||
| 1400 | int cpu = 0; | ||
| 1401 | #endif | ||
| 1402 | |||
| 1403 |
1/2✓ Branch 39 → 40 taken 21 times.
✗ Branch 39 → 106 not taken.
|
21 | fast_resize = vi.IsPlanar(); |
| 1404 | // PF 2025: H is not slower than V in C implementation. | ||
| 1405 | // Still, H resizers are incompatible with packed RGB formats | ||
| 1406 | |||
| 1407 |
2/2✓ Branch 40 → 41 taken 2 times.
✓ Branch 40 → 95 taken 19 times.
|
21 | if (!fast_resize) { |
| 1408 | |||
| 1409 | // nonfast-resize: using V resizer for horizontal resizing between a turnleft/right | ||
| 1410 | // For packed RGB formats this is the only way | ||
| 1411 | |||
| 1412 |
1/2✓ Branch 41 → 42 taken 2 times.
✗ Branch 41 → 106 not taken.
|
2 | resampler_luma = FilteredResizeV::GetResampler(cpu, pixelsize, bits_per_pixel, resampling_program_luma, env); |
| 1413 | |||
| 1414 |
3/10✓ Branch 42 → 43 taken 2 times.
✗ Branch 42 → 106 not taken.
✗ Branch 43 → 44 not taken.
✓ Branch 43 → 47 taken 2 times.
✗ Branch 44 → 45 not taken.
✗ Branch 44 → 47 not taken.
✗ Branch 45 → 46 not taken.
✗ Branch 45 → 47 not taken.
✗ Branch 48 → 49 not taken.
✓ Branch 48 → 51 taken 2 times.
|
2 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { |
| 1415 | ✗ | resampler_chroma = FilteredResizeV::GetResampler(cpu, pixelsize, bits_per_pixel, resampling_program_chroma, env); | |
| 1416 | } | ||
| 1417 | |||
| 1418 | // Temporary buffer size for turns | ||
| 1419 |
1/2✓ Branch 51 → 52 taken 2 times.
✗ Branch 51 → 106 not taken.
|
2 | temp_1_pitch = AlignNumber(vi.BytesFromPixels(src_height), FRAME_ALIGN); |
| 1420 |
1/2✓ Branch 53 → 54 taken 2 times.
✗ Branch 53 → 106 not taken.
|
2 | temp_2_pitch = AlignNumber(vi.BytesFromPixels(dst_height), FRAME_ALIGN); |
| 1421 | |||
| 1422 | // Initialize Turn function | ||
| 1423 | // see turn.cpp | ||
| 1424 |
3/4✓ Branch 55 → 56 taken 2 times.
✗ Branch 55 → 106 not taken.
✓ Branch 56 → 57 taken 1 time.
✓ Branch 56 → 58 taken 1 time.
|
2 | if (vi.IsRGB24()) { |
| 1425 | #ifdef INTEL_INTRINSICS | ||
| 1426 | // no intel intentionally | ||
| 1427 | #endif | ||
| 1428 | 1 | turn_left = turn_left_rgb24; | |
| 1429 | 1 | turn_right = turn_right_rgb24; | |
| 1430 | } | ||
| 1431 |
2/4✓ Branch 58 → 59 taken 1 time.
✗ Branch 58 → 106 not taken.
✗ Branch 59 → 60 not taken.
✓ Branch 59 → 65 taken 1 time.
|
1 | else if (vi.IsRGB32()) { |
| 1432 | #ifdef INTEL_INTRINSICS | ||
| 1433 | ✗ | if (has_avx2) { | |
| 1434 | ✗ | turn_left = turn_left_rgb32_avx2; | |
| 1435 | ✗ | turn_right = turn_right_rgb32_avx2; | |
| 1436 | } | ||
| 1437 | ✗ | else if (has_sse2) { | |
| 1438 | ✗ | turn_left = turn_left_rgb32_sse2; | |
| 1439 | ✗ | turn_right = turn_right_rgb32_sse2; | |
| 1440 | } | ||
| 1441 | else | ||
| 1442 | #elif NEON_INTRINSICS | ||
| 1443 | if (has_neon) { | ||
| 1444 | turn_left = turn_left_rgb32_neon; | ||
| 1445 | turn_right = turn_right_rgb32_neon; | ||
| 1446 | } | ||
| 1447 | else | ||
| 1448 | #endif | ||
| 1449 | { | ||
| 1450 | ✗ | turn_left = turn_left_rgb32_c; | |
| 1451 | ✗ | turn_right = turn_right_rgb32_c; | |
| 1452 | } | ||
| 1453 | } | ||
| 1454 |
2/4✓ Branch 65 → 66 taken 1 time.
✗ Branch 65 → 106 not taken.
✗ Branch 66 → 67 not taken.
✓ Branch 66 → 68 taken 1 time.
|
1 | else if (vi.IsRGB48()) { |
| 1455 | #ifdef INTEL_INTRINSICS | ||
| 1456 | // no intel intentionally | ||
| 1457 | #endif | ||
| 1458 | ✗ | turn_left = turn_left_rgb48_c; | |
| 1459 | ✗ | turn_right = turn_right_rgb48_c; | |
| 1460 | } | ||
| 1461 |
2/4✓ Branch 68 → 69 taken 1 time.
✗ Branch 68 → 106 not taken.
✓ Branch 69 → 70 taken 1 time.
✗ Branch 69 → 75 not taken.
|
1 | else if (vi.IsRGB64()) { |
| 1462 | #ifdef INTEL_INTRINSICS | ||
| 1463 |
1/2✓ Branch 70 → 71 taken 1 time.
✗ Branch 70 → 72 not taken.
|
1 | if (has_avx2) { |
| 1464 | 1 | turn_left = turn_left_rgb64_avx2; | |
| 1465 | 1 | turn_right = turn_right_rgb64_avx2; | |
| 1466 | } | ||
| 1467 | ✗ | else if (has_sse2) { | |
| 1468 | ✗ | turn_left = turn_left_rgb64_sse2; | |
| 1469 | ✗ | turn_right = turn_right_rgb64_sse2; | |
| 1470 | } | ||
| 1471 | else | ||
| 1472 | #elif defined(NEON_INTRINSICS) | ||
| 1473 | if (has_neon) { | ||
| 1474 | turn_left = turn_left_rgb64_neon; | ||
| 1475 | turn_right = turn_right_rgb64_neon; | ||
| 1476 | } | ||
| 1477 | else | ||
| 1478 | #endif | ||
| 1479 | { | ||
| 1480 | ✗ | turn_left = turn_left_rgb64_c; | |
| 1481 | ✗ | turn_right = turn_right_rgb64_c; | |
| 1482 | } | ||
| 1483 | } | ||
| 1484 | else { | ||
| 1485 | ✗ | switch (vi.ComponentSize()) {// AVS16 | |
| 1486 | ✗ | case 1: // 8 bit | |
| 1487 | #ifdef INTEL_INTRINSICS | ||
| 1488 | ✗ | if (has_avx2) { | |
| 1489 | ✗ | turn_left = turn_left_plane_8_avx2; | |
| 1490 | ✗ | turn_right = turn_right_plane_8_avx2; | |
| 1491 | } | ||
| 1492 | ✗ | else if (has_sse2) { | |
| 1493 | ✗ | turn_left = turn_left_plane_8_sse2; | |
| 1494 | ✗ | turn_right = turn_right_plane_8_sse2; | |
| 1495 | } | ||
| 1496 | else | ||
| 1497 | #elif defined(NEON_INTRINSICS) | ||
| 1498 | if (has_neon) { | ||
| 1499 | turn_left = turn_left_plane_8_neon; | ||
| 1500 | turn_right = turn_right_plane_8_neon; | ||
| 1501 | } | ||
| 1502 | else | ||
| 1503 | #endif | ||
| 1504 | { | ||
| 1505 | ✗ | turn_left = turn_left_plane_8_c; | |
| 1506 | ✗ | turn_right = turn_right_plane_8_c; | |
| 1507 | } | ||
| 1508 | ✗ | break; | |
| 1509 | ✗ | case 2: // 16 bit | |
| 1510 | #ifdef INTEL_INTRINSICS | ||
| 1511 | ✗ | if (has_avx2) { | |
| 1512 | ✗ | turn_left = turn_left_plane_16_avx2; | |
| 1513 | ✗ | turn_right = turn_right_plane_16_avx2; | |
| 1514 | } | ||
| 1515 | ✗ | else if (has_sse2) { | |
| 1516 | ✗ | turn_left = turn_left_plane_16_sse2; | |
| 1517 | ✗ | turn_right = turn_right_plane_16_sse2; | |
| 1518 | } | ||
| 1519 | else | ||
| 1520 | #elif defined(NEON_INTRINSICS) | ||
| 1521 | if (has_neon) { | ||
| 1522 | turn_left = turn_left_plane_16_neon; | ||
| 1523 | turn_right = turn_right_plane_16_neon; | ||
| 1524 | } | ||
| 1525 | else | ||
| 1526 | #endif | ||
| 1527 | { | ||
| 1528 | ✗ | turn_left = turn_left_plane_16_c; | |
| 1529 | ✗ | turn_right = turn_right_plane_16_c; | |
| 1530 | } | ||
| 1531 | ✗ | break; | |
| 1532 | ✗ | default: // 32 bit | |
| 1533 | #ifdef INTEL_INTRINSICS | ||
| 1534 | ✗ | if (has_avx2) { | |
| 1535 | ✗ | turn_left = turn_left_plane_32_avx2; | |
| 1536 | ✗ | turn_right = turn_right_plane_32_avx2; | |
| 1537 | } | ||
| 1538 | ✗ | else if (has_sse2) { | |
| 1539 | ✗ | turn_left = turn_left_plane_32_sse2; | |
| 1540 | ✗ | turn_right = turn_right_plane_32_sse2; | |
| 1541 | } | ||
| 1542 | else | ||
| 1543 | #endif | ||
| 1544 | { | ||
| 1545 | ✗ | turn_left = turn_left_plane_32_c; | |
| 1546 | ✗ | turn_right = turn_right_plane_32_c; | |
| 1547 | } | ||
| 1548 | } | ||
| 1549 | } | ||
| 1550 | } | ||
| 1551 | else { | ||
| 1552 | // planar format (or Y) | ||
| 1553 |
1/2✓ Branch 95 → 96 taken 19 times.
✗ Branch 95 → 106 not taken.
|
19 | resampler_h_luma = GetResampler(cpu, pixelsize, bits_per_pixel, resampling_program_luma, /*out*/resampler_h_luma_mt, env); |
| 1554 | |||
| 1555 |
4/4✓ Branch 96 → 97 taken 7 times.
✓ Branch 96 → 100 taken 12 times.
✓ Branch 97 → 98 taken 6 times.
✓ Branch 97 → 100 taken 1 time.
|
19 | if (!grey && !isRGBPfamily) { |
| 1556 |
1/2✓ Branch 98 → 99 taken 6 times.
✗ Branch 98 → 106 not taken.
|
6 | resampler_h_chroma = GetResampler(cpu, pixelsize, bits_per_pixel, resampling_program_chroma, /*out*/resampler_h_chroma_mt, env); |
| 1557 | } | ||
| 1558 | |||
| 1559 |
1/2✓ Branch 100 → 101 taken 19 times.
✗ Branch 100 → 102 not taken.
|
19 | if (!use_alternative_h_avx512_mt_fallback) { |
| 1560 | 19 | resampler_h_luma_mt = nullptr; | |
| 1561 | 19 | resampler_h_chroma_mt = nullptr; | |
| 1562 | } | ||
| 1563 | } | ||
| 1564 | // Change target video info size | ||
| 1565 | 21 | vi.width = target_width; | |
| 1566 | 22 | } | |
| 1567 | |||
| 1568 | 13 | PVideoFrame __stdcall FilteredResizeH::GetFrame(int n, IScriptEnvironment* env) | |
| 1569 | { | ||
| 1570 |
1/2✓ Branch 3 → 4 taken 13 times.
✗ Branch 3 → 198 not taken.
|
13 | PVideoFrame src = child->GetFrame(n, env); |
| 1571 |
1/2✓ Branch 4 → 5 taken 13 times.
✗ Branch 4 → 196 not taken.
|
13 | PVideoFrame dst = env->NewVideoFrameP(vi, &src); |
| 1572 | |||
| 1573 |
5/8✓ Branch 5 → 6 taken 13 times.
✗ Branch 5 → 194 not taken.
✓ Branch 6 → 7 taken 12 times.
✓ Branch 6 → 9 taken 1 time.
✓ Branch 7 → 8 taken 12 times.
✗ Branch 7 → 194 not taken.
✗ Branch 8 → 9 not taken.
✓ Branch 8 → 10 taken 12 times.
|
13 | bool isRGBPfamily = vi.IsPlanarRGB() || vi.IsPlanarRGBA(); |
| 1574 | |||
| 1575 |
2/2✓ Branch 11 → 12 taken 2 times.
✓ Branch 11 → 118 taken 11 times.
|
13 | if (!fast_resize) { |
| 1576 | // e.g. not aligned, not mod4 | ||
| 1577 | // temp_1_pitch and temp_2_pitch is pixelsize-aware | ||
| 1578 |
1/2✓ Branch 12 → 13 taken 2 times.
✗ Branch 12 → 194 not taken.
|
2 | BYTE* temp_1 = static_cast<BYTE*>(env->Allocate(temp_1_pitch * src_width, FRAME_ALIGN, AVS_POOLED_ALLOC)); |
| 1579 |
1/2✓ Branch 13 → 14 taken 2 times.
✗ Branch 13 → 194 not taken.
|
2 | BYTE* temp_2 = static_cast<BYTE*>(env->Allocate(temp_2_pitch * dst_width, FRAME_ALIGN, AVS_POOLED_ALLOC)); |
| 1580 |
2/4✓ Branch 14 → 15 taken 2 times.
✗ Branch 14 → 16 not taken.
✗ Branch 15 → 16 not taken.
✓ Branch 15 → 19 taken 2 times.
|
2 | if (!temp_1 || !temp_2) { |
| 1581 | ✗ | env->Free(temp_1); | |
| 1582 | ✗ | env->Free(temp_2); | |
| 1583 | ✗ | env->ThrowError("Could not reserve memory in a resampler."); | |
| 1584 | } | ||
| 1585 | |||
| 1586 |
4/8✓ Branch 19 → 20 taken 2 times.
✗ Branch 19 → 194 not taken.
✓ Branch 20 → 21 taken 2 times.
✗ Branch 20 → 22 not taken.
✗ Branch 21 → 22 not taken.
✓ Branch 21 → 23 taken 2 times.
✗ Branch 24 → 25 not taken.
✓ Branch 24 → 102 taken 2 times.
|
2 | if (!vi.IsRGB() || isRGBPfamily) { |
| 1587 | // Y/G Plane | ||
| 1588 | ✗ | turn_right(src->GetReadPtr(), temp_1, src_width * pixelsize, src_height, src->GetPitch(), temp_1_pitch); // * pixelsize: turn_right needs GetPlaneWidth full size | |
| 1589 | ✗ | resampler_luma(temp_2, temp_1, temp_2_pitch, temp_1_pitch, resampling_program_luma, src_height, dst_width, bits_per_pixel); | |
| 1590 | ✗ | turn_left(temp_2, dst->GetWritePtr(), dst_height * pixelsize, dst_width, temp_2_pitch, dst->GetPitch()); | |
| 1591 | |||
| 1592 | ✗ | if (isRGBPfamily) | |
| 1593 | { | ||
| 1594 | ✗ | turn_right(src->GetReadPtr(PLANAR_B), temp_1, src_width * pixelsize, src_height, src->GetPitch(PLANAR_B), temp_1_pitch); // * pixelsize: turn_right needs GetPlaneWidth full size | |
| 1595 | ✗ | resampler_luma(temp_2, temp_1, temp_2_pitch, temp_1_pitch, resampling_program_luma, src_height, dst_width, bits_per_pixel); | |
| 1596 | ✗ | turn_left(temp_2, dst->GetWritePtr(PLANAR_B), dst_height * pixelsize, dst_width, temp_2_pitch, dst->GetPitch(PLANAR_B)); | |
| 1597 | |||
| 1598 | ✗ | turn_right(src->GetReadPtr(PLANAR_R), temp_1, src_width * pixelsize, src_height, src->GetPitch(PLANAR_R), temp_1_pitch); // * pixelsize: turn_right needs GetPlaneWidth full size | |
| 1599 | ✗ | resampler_luma(temp_2, temp_1, temp_2_pitch, temp_1_pitch, resampling_program_luma, src_height, dst_width, bits_per_pixel); | |
| 1600 | ✗ | turn_left(temp_2, dst->GetWritePtr(PLANAR_R), dst_height * pixelsize, dst_width, temp_2_pitch, dst->GetPitch(PLANAR_R)); | |
| 1601 | } | ||
| 1602 | ✗ | else if (!grey) { | |
| 1603 | ✗ | const int shift = vi.GetPlaneWidthSubsampling(PLANAR_U); | |
| 1604 | ✗ | const int shift_h = vi.GetPlaneHeightSubsampling(PLANAR_U); | |
| 1605 | |||
| 1606 | ✗ | const int src_chroma_width = src_width >> shift; | |
| 1607 | ✗ | const int dst_chroma_width = dst_width >> shift; | |
| 1608 | ✗ | const int src_chroma_height = src_height >> shift_h; | |
| 1609 | ✗ | const int dst_chroma_height = dst_height >> shift_h; | |
| 1610 | |||
| 1611 | // turn_xxx: width * pixelsize: needs GetPlaneWidth-like full size | ||
| 1612 | // U Plane | ||
| 1613 | ✗ | turn_right(src->GetReadPtr(PLANAR_U), temp_1, src_chroma_width * pixelsize, src_chroma_height, src->GetPitch(PLANAR_U), temp_1_pitch); | |
| 1614 | ✗ | resampler_luma(temp_2, temp_1, temp_2_pitch, temp_1_pitch, resampling_program_chroma, src_chroma_height, dst_chroma_width, bits_per_pixel); | |
| 1615 | ✗ | turn_left(temp_2, dst->GetWritePtr(PLANAR_U), dst_chroma_height * pixelsize, dst_chroma_width, temp_2_pitch, dst->GetPitch(PLANAR_U)); | |
| 1616 | |||
| 1617 | // V Plane | ||
| 1618 | ✗ | turn_right(src->GetReadPtr(PLANAR_V), temp_1, src_chroma_width * pixelsize, src_chroma_height, src->GetPitch(PLANAR_V), temp_1_pitch); | |
| 1619 | ✗ | resampler_luma(temp_2, temp_1, temp_2_pitch, temp_1_pitch, resampling_program_chroma, src_chroma_height, dst_chroma_width, bits_per_pixel); | |
| 1620 | ✗ | turn_left(temp_2, dst->GetWritePtr(PLANAR_V), dst_chroma_height * pixelsize, dst_chroma_width, temp_2_pitch, dst->GetPitch(PLANAR_V)); | |
| 1621 | } | ||
| 1622 | ✗ | if (vi.IsYUVA() || vi.IsPlanarRGBA()) | |
| 1623 | { | ||
| 1624 | ✗ | turn_right(src->GetReadPtr(PLANAR_A), temp_1, src_width * pixelsize, src_height, src->GetPitch(PLANAR_A), temp_1_pitch); // * pixelsize: turn_right needs GetPlaneWidth full size | |
| 1625 | ✗ | resampler_luma(temp_2, temp_1, temp_2_pitch, temp_1_pitch, resampling_program_luma, src_height, dst_width, bits_per_pixel); | |
| 1626 | ✗ | turn_left(temp_2, dst->GetWritePtr(PLANAR_A), dst_height * pixelsize, dst_width, temp_2_pitch, dst->GetPitch(PLANAR_A)); | |
| 1627 | } | ||
| 1628 | |||
| 1629 | } | ||
| 1630 | else { | ||
| 1631 | // packed RGB | ||
| 1632 | // First left, then right. Reason: packed RGB bottom to top. Right+left shifts RGB24/RGB32 image to the opposite horizontal direction | ||
| 1633 |
4/8✓ Branch 103 → 104 taken 2 times.
✗ Branch 103 → 194 not taken.
✓ Branch 104 → 105 taken 2 times.
✗ Branch 104 → 194 not taken.
✓ Branch 106 → 107 taken 2 times.
✗ Branch 106 → 194 not taken.
✓ Branch 107 → 108 taken 2 times.
✗ Branch 107 → 194 not taken.
|
2 | turn_left(src->GetReadPtr(), temp_1, vi.BytesFromPixels(src_width), src_height, src->GetPitch(), temp_1_pitch); |
| 1634 |
2/4✓ Branch 108 → 109 taken 2 times.
✗ Branch 108 → 194 not taken.
✓ Branch 109 → 110 taken 2 times.
✗ Branch 109 → 194 not taken.
|
2 | resampler_luma(temp_2, temp_1, temp_2_pitch, temp_1_pitch, resampling_program_luma, vi.BytesFromPixels(src_height) / pixelsize, dst_width, bits_per_pixel); |
| 1635 |
4/8✓ Branch 111 → 112 taken 2 times.
✗ Branch 111 → 194 not taken.
✓ Branch 112 → 113 taken 2 times.
✗ Branch 112 → 194 not taken.
✓ Branch 114 → 115 taken 2 times.
✗ Branch 114 → 194 not taken.
✓ Branch 115 → 116 taken 2 times.
✗ Branch 115 → 194 not taken.
|
2 | turn_right(temp_2, dst->GetWritePtr(), vi.BytesFromPixels(dst_height), dst_width, temp_2_pitch, dst->GetPitch()); |
| 1636 | } | ||
| 1637 | |||
| 1638 |
1/2✓ Branch 116 → 117 taken 2 times.
✗ Branch 116 → 194 not taken.
|
2 | env->Free(temp_1); |
| 1639 |
1/2✓ Branch 117 → 191 taken 2 times.
✗ Branch 117 → 194 not taken.
|
2 | env->Free(temp_2); |
| 1640 | } | ||
| 1641 | else { | ||
| 1642 | // depending on MT or not, select proper resizer if alternative is available | ||
| 1643 |
1/4✗ Branch 118 → 119 not taken.
✓ Branch 118 → 121 taken 11 times.
✗ Branch 119 → 120 not taken.
✗ Branch 119 → 121 not taken.
|
11 | ResamplerH current_resampler_h_luma = (num_threads > 1 && resampler_h_luma_mt != nullptr) ? resampler_h_luma_mt : resampler_h_luma; |
| 1644 |
1/4✗ Branch 122 → 123 not taken.
✓ Branch 122 → 125 taken 11 times.
✗ Branch 123 → 124 not taken.
✗ Branch 123 → 125 not taken.
|
11 | ResamplerH current_resampler_h_chroma = (num_threads > 1 && resampler_h_chroma_mt != nullptr) ? resampler_h_chroma_mt : resampler_h_chroma; |
| 1645 | |||
| 1646 | // Y Plane | ||
| 1647 |
5/10✓ Branch 127 → 128 taken 11 times.
✗ Branch 127 → 194 not taken.
✓ Branch 129 → 130 taken 11 times.
✗ Branch 129 → 194 not taken.
✓ Branch 131 → 132 taken 11 times.
✗ Branch 131 → 194 not taken.
✓ Branch 133 → 134 taken 11 times.
✗ Branch 133 → 194 not taken.
✓ Branch 134 → 135 taken 11 times.
✗ Branch 134 → 194 not taken.
|
11 | current_resampler_h_luma(dst->GetWritePtr(), src->GetReadPtr(), dst->GetPitch(), src->GetPitch(), resampling_program_luma, dst_width, dst_height, bits_per_pixel); |
| 1648 | |||
| 1649 |
2/2✓ Branch 135 → 136 taken 1 time.
✓ Branch 135 → 154 taken 10 times.
|
11 | if (isRGBPfamily) { |
| 1650 |
5/10✓ Branch 137 → 138 taken 1 time.
✗ Branch 137 → 194 not taken.
✓ Branch 139 → 140 taken 1 time.
✗ Branch 139 → 194 not taken.
✓ Branch 141 → 142 taken 1 time.
✗ Branch 141 → 194 not taken.
✓ Branch 143 → 144 taken 1 time.
✗ Branch 143 → 194 not taken.
✓ Branch 144 → 145 taken 1 time.
✗ Branch 144 → 194 not taken.
|
1 | current_resampler_h_luma(dst->GetWritePtr(PLANAR_B), src->GetReadPtr(PLANAR_B), dst->GetPitch(PLANAR_B), src->GetPitch(PLANAR_B), resampling_program_luma, dst_width, dst_height, bits_per_pixel); |
| 1651 |
5/10✓ Branch 146 → 147 taken 1 time.
✗ Branch 146 → 194 not taken.
✓ Branch 148 → 149 taken 1 time.
✗ Branch 148 → 194 not taken.
✓ Branch 150 → 151 taken 1 time.
✗ Branch 150 → 194 not taken.
✓ Branch 152 → 153 taken 1 time.
✗ Branch 152 → 194 not taken.
✓ Branch 153 → 175 taken 1 time.
✗ Branch 153 → 194 not taken.
|
1 | current_resampler_h_luma(dst->GetWritePtr(PLANAR_R), src->GetReadPtr(PLANAR_R), dst->GetPitch(PLANAR_R), src->GetPitch(PLANAR_R), resampling_program_luma, dst_width, dst_height, bits_per_pixel); |
| 1652 | } | ||
| 1653 |
2/2✓ Branch 154 → 155 taken 6 times.
✓ Branch 154 → 175 taken 4 times.
|
10 | else if (!grey) { |
| 1654 |
1/2✓ Branch 155 → 156 taken 6 times.
✗ Branch 155 → 194 not taken.
|
6 | const int dst_chroma_width = dst_width >> vi.GetPlaneWidthSubsampling(PLANAR_U); |
| 1655 |
1/2✓ Branch 156 → 157 taken 6 times.
✗ Branch 156 → 194 not taken.
|
6 | const int dst_chroma_height = dst_height >> vi.GetPlaneHeightSubsampling(PLANAR_U); |
| 1656 | |||
| 1657 | // U Plane | ||
| 1658 |
5/10✓ Branch 158 → 159 taken 6 times.
✗ Branch 158 → 194 not taken.
✓ Branch 160 → 161 taken 6 times.
✗ Branch 160 → 194 not taken.
✓ Branch 162 → 163 taken 6 times.
✗ Branch 162 → 194 not taken.
✓ Branch 164 → 165 taken 6 times.
✗ Branch 164 → 194 not taken.
✓ Branch 165 → 166 taken 6 times.
✗ Branch 165 → 194 not taken.
|
6 | current_resampler_h_chroma(dst->GetWritePtr(PLANAR_U), src->GetReadPtr(PLANAR_U), dst->GetPitch(PLANAR_U), src->GetPitch(PLANAR_U), resampling_program_chroma, dst_chroma_width, dst_chroma_height, bits_per_pixel); |
| 1659 | |||
| 1660 | // V Plane | ||
| 1661 |
5/10✓ Branch 167 → 168 taken 6 times.
✗ Branch 167 → 194 not taken.
✓ Branch 169 → 170 taken 6 times.
✗ Branch 169 → 194 not taken.
✓ Branch 171 → 172 taken 6 times.
✗ Branch 171 → 194 not taken.
✓ Branch 173 → 174 taken 6 times.
✗ Branch 173 → 194 not taken.
✓ Branch 174 → 175 taken 6 times.
✗ Branch 174 → 194 not taken.
|
6 | current_resampler_h_chroma(dst->GetWritePtr(PLANAR_V), src->GetReadPtr(PLANAR_V), dst->GetPitch(PLANAR_V), src->GetPitch(PLANAR_V), resampling_program_chroma, dst_chroma_width, dst_chroma_height, bits_per_pixel); |
| 1662 | } | ||
| 1663 |
7/10✓ Branch 175 → 176 taken 11 times.
✗ Branch 175 → 194 not taken.
✓ Branch 176 → 177 taken 10 times.
✓ Branch 176 → 179 taken 1 time.
✓ Branch 177 → 178 taken 10 times.
✗ Branch 177 → 194 not taken.
✗ Branch 178 → 179 not taken.
✓ Branch 178 → 180 taken 10 times.
✓ Branch 181 → 182 taken 1 time.
✓ Branch 181 → 191 taken 10 times.
|
11 | if (vi.IsYUVA() || vi.IsPlanarRGBA()) |
| 1664 | { | ||
| 1665 |
5/10✓ Branch 183 → 184 taken 1 time.
✗ Branch 183 → 194 not taken.
✓ Branch 185 → 186 taken 1 time.
✗ Branch 185 → 194 not taken.
✓ Branch 187 → 188 taken 1 time.
✗ Branch 187 → 194 not taken.
✓ Branch 189 → 190 taken 1 time.
✗ Branch 189 → 194 not taken.
✓ Branch 190 → 191 taken 1 time.
✗ Branch 190 → 194 not taken.
|
1 | current_resampler_h_luma(dst->GetWritePtr(PLANAR_A), src->GetReadPtr(PLANAR_A), dst->GetPitch(PLANAR_A), src->GetPitch(PLANAR_A), resampling_program_luma, dst_width, dst_height, bits_per_pixel); |
| 1666 | } | ||
| 1667 | |||
| 1668 | } | ||
| 1669 | |||
| 1670 | 13 | return dst; | |
| 1671 | 13 | } | |
| 1672 | |||
| 1673 | 25 | ResamplerH FilteredResizeH::GetResampler(int CPU, int pixelsize, int bits_per_pixel, ResamplingProgram* program, ResamplerH &out_resampler_h_alternative_for_mt, IScriptEnvironment* env) | |
| 1674 | { | ||
| 1675 | 25 | out_resampler_h_alternative_for_mt = nullptr; | |
| 1676 | 25 | int simd_coeff_count_padding = 8; // even for _ks16_float this is enough, it works differently inside | |
| 1677 | |||
| 1678 | // Both 8-bit and 16-bit SSSE3 and AVX2 horizontal resizers benefit from processing 16 pixels per cycle. | ||
| 1679 | // Floats also use 32 bytes, but since 32/sizeof(float) = 8, processing 16 pixels is unnecessary. | ||
| 1680 | // Even in C, the code is optimized to be vector-friendly. | ||
| 1681 |
3/4✓ Branch 2 → 3 taken 1 time.
✓ Branch 2 → 4 taken 24 times.
✓ Branch 3 → 4 taken 1 time.
✗ Branch 3 → 5 not taken.
|
25 | if (pixelsize == 1 || pixelsize == 2) |
| 1682 | 25 | simd_coeff_count_padding = 16; // FIXME: make it 8 | |
| 1683 | |||
| 1684 | // Not only does it prepare and pad for SIMD/vector code, but it also corrects, reorders, and equalizes coefficients | ||
| 1685 | // at the right and bottom ends, since we may have variable kernel sizes due to boundary conditions. | ||
| 1686 | 25 | resize_prepare_coeffs(program, env, simd_coeff_count_padding); | |
| 1687 | |||
| 1688 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 1689 | 25 | const bool has_AVX512_base = (CPU & CPUF_AVX512_BASE) == CPUF_AVX512_BASE; // group flag! | |
| 1690 | 25 | const bool has_AVX512_fast = (CPU & CPUF_AVX512_FAST) == CPUF_AVX512_FAST; // group flag! | |
| 1691 | #endif | ||
| 1692 | |||
| 1693 |
2/2✓ Branch 6 → 7 taken 24 times.
✓ Branch 6 → 47 taken 1 time.
|
25 | if (pixelsize == 1) |
| 1694 | { | ||
| 1695 | #ifdef INTEL_INTRINSICS | ||
| 1696 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 1697 |
1/2✗ Branch 7 → 8 not taken.
✓ Branch 7 → 42 taken 24 times.
|
24 | if (has_AVX512_base) { |
| 1698 | #ifdef PF_GENERIC_UINT_TEST | ||
| 1699 | return resizer_h_avx512_generic_uint8_t; // PF debug | ||
| 1700 | #endif | ||
| 1701 | // feature flag, grouping many avx512 features | ||
| 1702 | // in case of optimized avx512_permutex_vstripe resizer found, set alternative resizer for MT use | ||
| 1703 | ✗ | out_resampler_h_alternative_for_mt = resizer_h_avx2_generic_uint8_t; // AVX2 should present if AVX512 present | |
| 1704 | // | ||
| 1705 | // AVX-512 uint8 horizontal resizer selection (pretransposed-coeffs variants, normal builds): | ||
| 1706 | // | ||
| 1707 | // Function | Pixels/iter | Source window | Use case | ||
| 1708 | // ------------------------------------------|-------------|---------------|------------------------------------------ | ||
| 1709 | // mpz_ks4_pretransposed_coeffs | 64 | 128 bytes | filter_size <= 4, upscale / mild downscale | ||
| 1710 | // mpz_ks8_pretransposed_coeffs | 64 | 128 bytes | filter_size <= 8, upscale / mild downscale (1st choice) | ||
| 1711 | // 2s32_ks8_pretransposed_coeffs | 2x32 | 2x128 bytes | filter_size <= 8, heavier downscale (~0.5x), two independent source strips | ||
| 1712 | // mpz_ks16_pretransposed_coeffs | 32 | 128 bytes | filter_size <= 16, upscale / mild downscale | ||
| 1713 | // 2s32_ks64_pretransposed_coeffs | 2x32 | 2x128 bytes | filter_size up to ~64, variable inner loop over taps | ||
| 1714 | // | ||
| 1715 | // "mpz" = maskz-permutex: single 128-byte source window per 64 (or 32) target pixels. | ||
| 1716 | // "2s32" = two strips of 32: two independent 128-byte windows to cover wider source spans. | ||
| 1717 | // VNNI path: uses native vpermi2b (VBMI) + dpwssd. BASE path: simulates vpermi2b via | ||
| 1718 | // precomputed word-index/shift-amount pairs fed to vpermw+vpsrlvw (avoids costly VBMI instruction). | ||
| 1719 | // | ||
| 1720 | ✗ | if (program->filter_size_real <= 4) { | |
| 1721 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(64/*iSamplesInTheGroup*/, 128/*permutex_index_diff_limit*/, 4/*kernel_size*/)) { | |
| 1722 | /* | ||
| 1723 | (Rocket Lake i7 - 11700, Expr-vertical-stripes + BicubicResize(width*2,height)) | ||
| 1724 | Contenders | ||
| 1725 | AVX512 Fast | ||
| 1726 | resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_vnni 2760fps (VNNI is not even really used in clangcl) | ||
| 1727 | resize_h_planar_uint8_avx512_permutex_vstripe_ks4_vbmi 2548fps | ||
| 1728 | AVX512 base; No VNNI, No VBMI, both simulating the 8 bit VBMI shuffle | ||
| 1729 | resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_base 1478fps (register pressure - twice as much _mm512_permutex2var_epi8 simulation than non-mpz) | ||
| 1730 | resize_h_planar_uint8_avx512_permutex_vstripe_ks4_base 2493fps (despite the _mm512_permutex2var_epi8 simulation, this is only 2-3% slower than vbmi version | ||
| 1731 | resizer_h_avx2_generic_uint8_t 752fps | ||
| 1732 | |||
| 1733 | Winners: (Fast) resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_vnni (Base) resize_h_planar_uint8_avx512_permutex_vstripe_ks4_base | ||
| 1734 | */ | ||
| 1735 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 64/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 4/*fixed_kernel_size*/); | |
| 1736 | ✗ | if (has_AVX512_fast) | |
| 1737 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_vnni; | |
| 1738 | else | ||
| 1739 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks4_pretransposed_coeffs_base; | |
| 1740 | } | ||
| 1741 | } | ||
| 1742 | ✗ | if (program->filter_size_real <= 8) { | |
| 1743 | /* | ||
| 1744 | resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8 | ||
| 1745 | - support more downsampling ratios, like | ||
| 1746 | Bicubic/BilinearResize(width/2) and even SinPowResize(width/2) for downsampling of UHD 4k to FHD is working. | ||
| 1747 | - Expected to support scaling ratios from about a bit below 0.5 to infinity (with filter support <=2). | ||
| 1748 | |||
| 1749 | resize_h_planar_uint8_avx512_permutex_vstripe_ks8 | ||
| 1750 | - faster with scale ratios from about 1.0 to infinity (with filter support <=4). | ||
| 1751 | |||
| 1752 | These two functions selected in order from faster to slower. | ||
| 1753 | */ | ||
| 1754 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(64/*iSamplesInTheGroup*/, 128/*permutex_index_diff_limit*/, 8/*kernel_size*/)) { // first try faster ks8 | |
| 1755 | /* | ||
| 1756 | Contenders | ||
| 1757 | AVX512 Fast | ||
| 1758 | resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_vnni 3420fps | ||
| 1759 | resize_h_planar_uint8_avx512_permutex_vstripe_ks8_vbmi 3240fps | ||
| 1760 | AVX512 base | ||
| 1761 | resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_base 1720fps (register pressure) | ||
| 1762 | resize_h_planar_uint8_avx512_permutex_vstripe_ks8_base 2940fps | ||
| 1763 | |||
| 1764 | Winners: (Fast) resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_vnni (Base) resize_h_planar_uint8_avx512_permutex_vstripe_ks8_base | ||
| 1765 | */ | ||
| 1766 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 64/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 8/*fixed_kernel_size*/); | |
| 1767 | ✗ | if (has_AVX512_fast) | |
| 1768 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_vnni; | |
| 1769 | else | ||
| 1770 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks8_pretransposed_coeffs_base; | |
| 1771 | } | ||
| 1772 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(32/*iSamplesInTheGroup*/, 128/*permutex_index_diff_limit*/, 8/*kernel_size*/)) { // slower ks8 but more downsample ratio for /2 | |
| 1773 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 32/*iSamplesInTheGroup*/, 2/*iGroupsCount*/, 8/*fixed_kernel_size*/); | |
| 1774 | ✗ | if (has_AVX512_fast) | |
| 1775 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_vnni; | |
| 1776 | else | ||
| 1777 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_2s32_ks8_pretransposed_coeffs_base; | |
| 1778 | } | ||
| 1779 | } | ||
| 1780 | ✗ | if (program->filter_size_real <= 16) { | |
| 1781 | // yes: LanczosResize(int(width*0.9 + 0.5), height, taps=4) kernel size 9 (K) | ||
| 1782 | // yes: LanczosResize(int(width*1.1 + 0.5), height, taps=5) kernel size 10 (L) | ||
| 1783 | // yes: LanczosResize(int(width*1.1 + 0.5), height, taps=6) kernel size 12 (M) | ||
| 1784 | // yes: LanczosResize(int(width*0.5 + 0.5), height, taps=3) kernel size 12 (N) (in float only 2s8_ks16 covered this resampling ratio) | ||
| 1785 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(32/*iSamplesInTheGroup*/, 128/*permutex_index_diff_limit*/, 16/*kernel_size*/)) { | |
| 1786 | /* | ||
| 1787 | Contenders: | ||
| 1788 | AVX512 fast | ||
| 1789 | resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_vnni 3440fps | ||
| 1790 | resize_h_planar_uint8_avx512_permutex_vstripe_ks16_vbmi 3037fps | ||
| 1791 | AVX512 base | ||
| 1792 | resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_base 1686fps (register pressure) | ||
| 1793 | resize_h_planar_uint8_avx512_permutex_vstripe_ks16_base 2909fps | ||
| 1794 | |||
| 1795 | Winners: (Fast) resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_vnni (Base) resize_h_planar_uint8_avx512_permutex_vstripe_ks16_base | ||
| 1796 | */ | ||
| 1797 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 32/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 16/*fixed_kernel_size*/); | |
| 1798 | ✗ | if (has_AVX512_fast) | |
| 1799 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_vnni; | |
| 1800 | else | ||
| 1801 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_ks16_pretransposed_coeffs_base; | |
| 1802 | } | ||
| 1803 | } | ||
| 1804 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(32/*iSamplesInTheGroup*/, 128/*permutex_index_diff_limit*/, program->filter_size_real/*kernel_size*/)) | |
| 1805 | { | ||
| 1806 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 32/*iSamplesInTheGroup*/, 2/*iGroupsCount*/, 0/*fixed_kernel_size: variable-loop kernel*/); | |
| 1807 | ✗ | if (has_AVX512_fast) | |
| 1808 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_vnni; | |
| 1809 | else | ||
| 1810 | ✗ | return resize_h_planar_uint8_avx512_permutex_vstripe_mpz_2s32_ks64_pretransposed_coeffs_base; | |
| 1811 | } | ||
| 1812 | ✗ | out_resampler_h_alternative_for_mt = nullptr; // not needed | |
| 1813 | } | ||
| 1814 | #endif | ||
| 1815 |
1/2✓ Branch 42 → 43 taken 24 times.
✗ Branch 42 → 44 not taken.
|
24 | if (CPU & CPUF_AVX2) { |
| 1816 | 24 | return resizer_h_avx2_generic_uint8_t; | |
| 1817 | } | ||
| 1818 | ✗ | if (CPU & CPUF_SSSE3) { | |
| 1819 | ✗ | return resizer_h_ssse3_generic_uint8_16<uint8_t, true>; | |
| 1820 | } | ||
| 1821 | #endif | ||
| 1822 | ✗ | return resizer_h_c_generic_uint8_16_vectorized<uint8_t, true>; | |
| 1823 | //return resize_h_c_planar<uint8_t, 1>; | ||
| 1824 | } | ||
| 1825 |
1/2✓ Branch 47 → 48 taken 1 time.
✗ Branch 47 → 117 not taken.
|
1 | else if (pixelsize == 2) { |
| 1826 | #ifdef INTEL_INTRINSICS | ||
| 1827 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 1828 |
1/2✗ Branch 48 → 49 not taken.
✓ Branch 48 → 106 taken 1 time.
|
1 | if (has_AVX512_base) { |
| 1829 | #ifdef PF_GENERIC_UINT_TEST | ||
| 1830 | if (bits_per_pixel < 16) | ||
| 1831 | return resizer_h_avx512_generic_uint16_t<true>; // PF debug | ||
| 1832 | else | ||
| 1833 | return resizer_h_avx512_generic_uint16_t<false>; // PF debug | ||
| 1834 | #endif | ||
| 1835 | ✗ | if (bits_per_pixel < 16) | |
| 1836 | ✗ | out_resampler_h_alternative_for_mt = resizer_h_avx2_generic_uint16_t<true>; // AVX2 should present if AVX512 present | |
| 1837 | else | ||
| 1838 | ✗ | out_resampler_h_alternative_for_mt = resizer_h_avx2_generic_uint16_t<false>; | |
| 1839 | |||
| 1840 | // feature flag, grouping many avx512 features | ||
| 1841 | ✗ | if (program->filter_size_real <= 4) { | |
| 1842 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(32/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 4/*kernel_size*/)) | |
| 1843 | { | ||
| 1844 | /* | ||
| 1845 | Contenders : | ||
| 1846 | AVX512 Fast | ||
| 1847 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks4_vnni 2578fps | ||
| 1848 | resize_h_planar_uint16_avx512_permutex_vstripe_ks4 2310fps (only base) | ||
| 1849 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4 2643fps | ||
| 1850 | AVX512 Base | ||
| 1851 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks4_base 2556fps | ||
| 1852 | resize_h_planar_uint16_avx512_permutex_vstripe_ks4 2310fps (only base) | ||
| 1853 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4 2685fps | ||
| 1854 | Fazit: The MP versions' difference is only two VNNI instructions between BASE/FAST, in benchmarks zero visible speed benefit is seen. | ||
| 1855 | Winners: (Both mp) (Fast) resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks4_vnni (Base) resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks4_base | ||
| 1856 | */ | ||
| 1857 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 64/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 4/*fixed_kernel_size*/); | |
| 1858 | ✗ | if (bits_per_pixel < 16) { | |
| 1859 | ✗ | if (has_AVX512_fast) | |
| 1860 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_vnni<true>; | |
| 1861 | else | ||
| 1862 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_base<true>; | |
| 1863 | } | ||
| 1864 | else { | ||
| 1865 | ✗ | if (has_AVX512_fast) | |
| 1866 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_vnni<false>; | |
| 1867 | else | ||
| 1868 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks4_pretransposed_coeffs_base<false>; | |
| 1869 | } | ||
| 1870 | } | ||
| 1871 | } | ||
| 1872 | ✗ | if (program->filter_size_real <= 8) { | |
| 1873 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(32/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 8/*kernel_size*/)) { | |
| 1874 | /* | ||
| 1875 | Contenders: | ||
| 1876 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks8_vnni | ||
| 1877 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks8_base | ||
| 1878 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_vnni (20260118) Case Q | ||
| 1879 | resize_h_planar_uint16_avx512_permutex_vstripe_ks8 (base avx512 only, no special intructions) | ||
| 1880 | Fazit: The MP versions' difference is only two VNNI instructions between BASE/FAST, in benchmarks 1-2% visible speed benefit is seen. | ||
| 1881 | Winners: (Both mp) (Fast) resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks8_vnni (Base) resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks8_base | ||
| 1882 | */ | ||
| 1883 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 64/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 8/*fixed_kernel_size*/); | |
| 1884 | ✗ | if (bits_per_pixel < 16) { | |
| 1885 | ✗ | if (has_AVX512_fast) | |
| 1886 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_vnni<true>; | |
| 1887 | else | ||
| 1888 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_base<true>; | |
| 1889 | } | ||
| 1890 | else { | ||
| 1891 | ✗ | if (has_AVX512_fast) | |
| 1892 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_vnni<false>; | |
| 1893 | else | ||
| 1894 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_2s32_ks8_pretransposed_coeffs_base<false>; | |
| 1895 | } | ||
| 1896 | } // check(32/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 8/*kernel_size*/) | ||
| 1897 | |||
| 1898 | // | ||
| 1899 | // Analysis C_VNNI C_BASE E_VNNI E_BASE R_VNNI R_BASE | ||
| 1900 | // resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8 3588 3604 3576 3601 3426 3611 [fps] | ||
| 1901 | // resize_h_planar_uint16_avx512_permutex_vstripe_2s16_ks8 3383 3269 3378 3287 3293 3270 | ||
| 1902 | // Case C LanczosResize(int(width*0.5 + 0.5), height, taps=1) kernel size 4, downsampling | ||
| 1903 | // Case E LanczosResize(int(width*0.5 + 0.5), height, taps=2) kernel size 8 | ||
| 1904 | // Case R: LanczosResize(int(width*0.5 + 0.5), height, taps=2) kernel size 8 | ||
| 1905 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(16/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 8/*kernel_size*/)) { | |
| 1906 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 64/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 8/*fixed_kernel_size*/); | |
| 1907 | ✗ | if (bits_per_pixel < 16) { | |
| 1908 | ✗ | if (has_AVX512_fast) | |
| 1909 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_vnni<true>; | |
| 1910 | else | ||
| 1911 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_base<true>; | |
| 1912 | } | ||
| 1913 | else { | ||
| 1914 | // bits_per_pixel == 16 | ||
| 1915 | ✗ | if (has_AVX512_fast) | |
| 1916 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_vnni<false>; | |
| 1917 | else | ||
| 1918 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks8_pretransposed_coeffs_base<false>; | |
| 1919 | } | ||
| 1920 | } // check(16/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 8/*kernel_size*/) | ||
| 1921 | |||
| 1922 | } // if (program->filter_size_real <= 8) | ||
| 1923 | |||
| 1924 | ✗ | if (program->filter_size_real <= 16) { | |
| 1925 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(32/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, 16/*kernel_size*/)) | |
| 1926 | { | ||
| 1927 | // yes: LanczosResize(int(width*0.9 + 0.5), height, taps=4) kernel size 9 (K) | ||
| 1928 | // yes: LanczosResize(int(width*1.1 + 0.5), height, taps=5) kernel size 10 (L) | ||
| 1929 | // yes: LanczosResize(int(width*1.1 + 0.5), height, taps=6) kernel size 12 (M) | ||
| 1930 | // no: LanczosResize(int(width*0.5 + 0.5), height, taps=3) kernel size 12 (N) (in float only 2s8_ks16 covered this resampling ratio) | ||
| 1931 | /* | ||
| 1932 | Contenders (none): | ||
| 1933 | AVX512 Fast | ||
| 1934 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_vnni 1851 1853 2189 | ||
| 1935 | AVX512 Base | ||
| 1936 | resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_base 1889 1869 2203 (? within measurement error, but quicker than VNNI??) | ||
| 1937 | resizer_h_avx2_generic_uint16_t (fallback) 1156 1085 1292 | ||
| 1938 | Winners: (Both mp) (Fast) resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks8_vnni (Base) resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks8_base | ||
| 1939 | */ | ||
| 1940 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 32/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 16/*fixed_kernel_size*/); | |
| 1941 | ✗ | if (bits_per_pixel < 16) { | |
| 1942 | ✗ | if (has_AVX512_fast) | |
| 1943 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_vnni<true>; | |
| 1944 | else | ||
| 1945 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_base<true>; | |
| 1946 | } | ||
| 1947 | else { | ||
| 1948 | ✗ | if (has_AVX512_fast) | |
| 1949 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_vnni<false>; | |
| 1950 | else | ||
| 1951 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_ks16_pretransposed_coeffs_base<false>; | |
| 1952 | } | ||
| 1953 | } | ||
| 1954 | } | ||
| 1955 | // "ks48" catch-all: handles any filter_size_real that fits within the | ||
| 1956 | // _mm512_permutex2var_epi16 index range of 64 uint16 elements (two ZMMs). | ||
| 1957 | // The check condition is: pixel_offset[x+15] - pixel_offset[x] + filter_size_real - 1 < 64. | ||
| 1958 | // At 1:1 scale the 16-pixel group spans 15 source positions, leaving room for | ||
| 1959 | // filter_size_real up to 48 (max even value: 15 + 48 - 1 = 62 < 64; 49 rounds up | ||
| 1960 | // to 50 → 15 + 49 = 64, fails). For downscaling the offset span grows and the | ||
| 1961 | // usable kernel shrinks — the check enforces this automatically per x-group. | ||
| 1962 | // The function itself has no hard 48 limit; it loops over filter_size_real directly. | ||
| 1963 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(16/*iSamplesInTheGroup*/, 64/*permutex_index_diff_limit*/, program->filter_size_real/*kernel_size*/)) | |
| 1964 | { | ||
| 1965 | ✗ | resize_prepare_coeffs_AVX512_H(program, env, 64/*iSamplesInTheGroup*/, 1/*iGroupsCount*/, 0/*fixed_kernel_size: variable-loop kernel*/); | |
| 1966 | ✗ | if (bits_per_pixel < 16) { | |
| 1967 | ✗ | if (has_AVX512_fast) | |
| 1968 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_vnni<true>; | |
| 1969 | else | ||
| 1970 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_base<true>; | |
| 1971 | } | ||
| 1972 | else { | ||
| 1973 | ✗ | if (has_AVX512_fast) | |
| 1974 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_vnni<false>; | |
| 1975 | else | ||
| 1976 | ✗ | return resize_h_planar_uint16_avx512_permutex_vstripe_mp_4s16_ks48_pretransposed_coeffs_base<false>; | |
| 1977 | } | ||
| 1978 | } | ||
| 1979 | ✗ | out_resampler_h_alternative_for_mt = nullptr; // not needed | |
| 1980 | } // has_AVX512_base | ||
| 1981 | #endif | ||
| 1982 |
1/2✓ Branch 106 → 107 taken 1 time.
✗ Branch 106 → 110 not taken.
|
1 | if (CPU & CPUF_AVX2) { |
| 1983 |
1/2✗ Branch 107 → 108 not taken.
✓ Branch 107 → 109 taken 1 time.
|
1 | if (bits_per_pixel < 16) |
| 1984 | ✗ | return resizer_h_avx2_generic_uint16_t<true>; | |
| 1985 | else | ||
| 1986 | 1 | return resizer_h_avx2_generic_uint16_t<false>; | |
| 1987 | } | ||
| 1988 | ✗ | if (CPU & CPUF_SSSE3) { | |
| 1989 | ✗ | if (bits_per_pixel < 16) | |
| 1990 | ✗ | return resizer_h_ssse3_generic_uint8_16<uint16_t, true>; | |
| 1991 | else | ||
| 1992 | ✗ | return resizer_h_ssse3_generic_uint8_16<uint16_t, false>; | |
| 1993 | } | ||
| 1994 | #endif | ||
| 1995 | ✗ | if (bits_per_pixel == 16) | |
| 1996 | ✗ | return resizer_h_c_generic_uint8_16_vectorized<uint16_t, false>; | |
| 1997 | // return resize_h_c_planar<uint16_t, 0>; | ||
| 1998 | else | ||
| 1999 | ✗ | return resizer_h_c_generic_uint8_16_vectorized<uint16_t, true>; | |
| 2000 | // return resize_h_c_planar<uint16_t, 1>; | ||
| 2001 | } | ||
| 2002 | else { //if (pixelsize == 4) | ||
| 2003 | #ifdef INTEL_INTRINSICS | ||
| 2004 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 2005 | ✗ | if (has_AVX512_base) { | |
| 2006 | // feature flag, grouping many avx512 features | ||
| 2007 | |||
| 2008 | // these perform very poorly in Prefetch, so we provide alternative generic version for MT | ||
| 2009 | |||
| 2010 | ✗ | if (program->filter_size_real <= 4) { | |
| 2011 | // up to 4 coeffs it can be highly optimized with transposes, gather/permutex choice | ||
| 2012 | ✗ | out_resampler_h_alternative_for_mt = resizer_h_avx512_generic_float_pix16_sub4_ks_4_8_16; // jolly joker | |
| 2013 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(16 /*iSamplesInTheGroup*/, 32 /*permutex_index_diff_limit*/, 4 /*kernel_size*/)) { | |
| 2014 | ✗ | return resize_h_planar_float_avx512_permutex_vstripe_ks4; | |
| 2015 | } | ||
| 2016 | ✗ | return resize_h_planar_float_avx512_transpose_vstripe_ks4; | |
| 2017 | } | ||
| 2018 | ✗ | if (program->filter_size_real <= 8) { | |
| 2019 | // up to 8 coeffs it can be highly optimized with transposes, gather/permutex choice | ||
| 2020 | ✗ | out_resampler_h_alternative_for_mt = resizer_h_avx512_generic_float_pix16_sub4_ks_4_8_16; // jolly joker | |
| 2021 | // first check 16 pixels per cycle version, probably resize_h_planar_float_avx512_permutex_vstripe_2s8_ks8 is faster, | ||
| 2022 | // if not possible, then 8 pixels per cycle | ||
| 2023 | ✗ | if (program->resize_h_planar_gather_permutex_vstripe_check(16/*iSamplesInTheGroup*/, 32/*permutex_index_diff_limit*/, 8/*kernel_size*/)) { | |
| 2024 | // 16 pixels per cycle version of permutex was not possible, try 2x8 version | ||
| 2025 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(8/*iSamplesInTheGroup*/, 32/*permutex_index_diff_limit*/, 8/*kernel_size*/)) { | |
| 2026 | ✗ | return resize_h_planar_float_avx512_permutex_vstripe_2s8_ks8; // 2x8 output version: better than transpose and generic | |
| 2027 | } | ||
| 2028 | ✗ | return resize_h_planar_float_avx512_transpose_vstripe_ks8; | |
| 2029 | // Speed ranking fps, just to have a clue, higher is better. | ||
| 2030 | // resize_h_planar_float_avx512_permutex_vstripe_2s8_ks8: 3482 | ||
| 2031 | // resize_h_planar_float_avx512_transpose_vstripe_ks8: 3186 | ||
| 2032 | // generic resizer_h_avx512_generic_float_pix16_sub4_ks_4_8_16: 2772 | ||
| 2033 | } | ||
| 2034 | // Speed ranking fps, just to have a clue, higher is better. | ||
| 2035 | // resize_h_planar_float_avx512_permutex_vstripe_2s8_ks8: 2040 2390 1221 | ||
| 2036 | // resize_h_planar_float_avx512_permutex_vstripe_ks8: 2847 3236 1775 | ||
| 2037 | ✗ | return resize_h_planar_float_avx512_permutex_vstripe_ks8; | |
| 2038 | } | ||
| 2039 | |||
| 2040 | ✗ | if (program->filter_size_real <= 16) { | |
| 2041 | // Dispatcher for float ks16 (filter_size_real 9..16). | ||
| 2042 | // | ||
| 2043 | // check(N, 32, ks) returns TRUE when N consecutive output pixels' source span | ||
| 2044 | // EXCEEDS 32 floats (two ZMMs), making _mm512_permutex2var_ps unusable for that | ||
| 2045 | // group size. FALSE means the group fits and permutex is valid. | ||
| 2046 | // | ||
| 2047 | // Strategy: try the widest feasible group first (best parallelism), fall back | ||
| 2048 | // to progressively narrower groups, then to generic if none fit. | ||
| 2049 | // | ||
| 2050 | // check(16)=false → ks16: all 16-px groups fit in 32 sources (mild downscale / upscale) | ||
| 2051 | // check(16)=true, | ||
| 2052 | // check(8)=false → 2s8: 16-px groups too wide, but 8-px groups fit (moderate downscale) | ||
| 2053 | // check(8)=true, | ||
| 2054 | // check(4)=false → 4s4: even 8-px groups too wide, but 4-px groups fit (heavy downscale) | ||
| 2055 | // check(4)=true → generic: even 4-px groups too wide (very heavy downscale) | ||
| 2056 | // | ||
| 2057 | // Note: 4s4_ks16 is benchmarked SLOWER than generic in its applicable range because | ||
| 2058 | // at such heavy downscale ratios the source taps are nearly consecutive in memory, | ||
| 2059 | // so the generic's sequential-load FMA tree outperforms 4x permutex2var + 3x blend | ||
| 2060 | // (112 port-5 ops vs generic's streaming loads). Kept for possible future use. | ||
| 2061 | ✗ | out_resampler_h_alternative_for_mt = resizer_h_avx512_generic_float_pix16_sub4_ks_4_8_16; // jolly joker | |
| 2062 | ✗ | if (program->resize_h_planar_gather_permutex_vstripe_check(16/*iSamplesInTheGroup*/, 32/*permutex_index_diff_limit*/, 16/*kernel_size*/)) { | |
| 2063 | // 16-px groups are too wide for permutex2var_ps — ks16 cannot be used. | ||
| 2064 | ✗ | if (!program->resize_h_planar_gather_permutex_vstripe_check(8/*iSamplesInTheGroup*/, 32/*permutex_index_diff_limit*/, 16/*kernel_size*/)) { | |
| 2065 | // 8-px groups fit: use 2s8_ks16 (2 independent groups of 8, each in a 32-float window). | ||
| 2066 | // generic resizer_h_avx512_generic_float_pix16_sub4_ks_4_8_16: 2650 fps | ||
| 2067 | // 2s8: 3392 fps | ||
| 2068 | ✗ | resize_prepare_coeffs_AVX512_float_H(program, env); | |
| 2069 | ✗ | return resize_h_planar_float_avx512_permutex_vstripe_2s8_ks16; | |
| 2070 | } | ||
| 2071 | // 8-px groups are also too wide. | ||
| 2072 | #if 0 | ||
| 2073 | // 4s4_ks16 DISABLED: slower than generic in this range (heavy downscale). | ||
| 2074 | // At such ratios source taps are nearly sequential, so generic's streaming | ||
| 2075 | // loads beat 4x permutex2var + 3x blend (112 vs ~48 port-5 ops per y). | ||
| 2076 | // generic: 4081 fps, 4s4_ks16: 3113 fps | ||
| 2077 | if (!program->resize_h_planar_gather_permutex_vstripe_check(4/*iSamplesInTheGroup*/, 32/*permutex_index_diff_limit*/, 16/*kernel_size*/)) { | ||
| 2078 | resize_prepare_coeffs_AVX512_float_H(program, env); | ||
| 2079 | return resize_h_planar_float_avx512_permutex_vstripe_4s4_ks16; | ||
| 2080 | } | ||
| 2081 | #endif | ||
| 2082 | // Even 4-px groups too wide (or 4s4 disabled): fall back to generic. | ||
| 2083 | ✗ | return resizer_h_avx512_generic_float_pix16_sub4_ks_4_8_16; | |
| 2084 | } | ||
| 2085 | // 16-px groups fit: use ks16 (single source group, best case). | ||
| 2086 | // ks16: 5500 fps (reference), generic: 2650 fps | ||
| 2087 | ✗ | resize_prepare_coeffs_AVX512_float_H(program, env); | |
| 2088 | ✗ | return resize_h_planar_float_avx512_permutex_vstripe_ks16; | |
| 2089 | } | ||
| 2090 | |||
| 2091 | ✗ | return resizer_h_avx512_generic_float_pix16_sub4_ks_4_8_16; | |
| 2092 | // other candidates were tested: | ||
| 2093 | // return resizer_h_avx512_generic_float_pix8_sub8_ks16; | ||
| 2094 | // return resizer_h_avx512_generic_float_pix16_sub16_ks8; | ||
| 2095 | // return resizer_h_avx512_generic_float_pix32_sub8_ks8; | ||
| 2096 | // return resizer_h_avx2_generic_float_pix8_sub2; // like AVX2 version | ||
| 2097 | // return resizer_h_avx512_generic_float_pix8_sub2; // like AVX2 version | ||
| 2098 | // return resizer_h_avx512_generic_float_pix8_sub4_ks8; | ||
| 2099 | // return resizer_h_avx512_generic_float_pix16_sub4_ks4; | ||
| 2100 | // return resizer_h_avx512_generic_float_pix16_sub4_ks8; | ||
| 2101 | // return resizer_h_avx2_generic_float; | ||
| 2102 | } | ||
| 2103 | #endif | ||
| 2104 | ✗ | if (CPU & CPUF_AVX2) { | |
| 2105 | // up to 4 coeffs it can be highly optimized with transposes, gather/permutex choice | ||
| 2106 | // These perform very poorly in Prefetch, so we provide alternative generic version for MT | ||
| 2107 | ✗ | out_resampler_h_alternative_for_mt = resize_h_planar_float_avx2_permutex_vstripe_ks4; // jolly joker | |
| 2108 | ✗ | if (program->filter_size_real <= 4) { | |
| 2109 | ✗ | if (program->resize_h_planar_gather_permutex_vstripe_check(8 /*iSamplesInTheGroup*/, 8 /*permutex_index_diff_limit*/, 4 /*kernel_size*/)) { | |
| 2110 | ✗ | switch (program->filter_size_real) { | |
| 2111 | ✗ | case 1: return resize_h_planar_float_avx2_transpose_vstripe_ks4<1>; break; | |
| 2112 | ✗ | case 2: return resize_h_planar_float_avx2_transpose_vstripe_ks4<2>; break; | |
| 2113 | ✗ | case 3: return resize_h_planar_float_avx2_transpose_vstripe_ks4<3>; break; | |
| 2114 | ✗ | case 4: return resize_h_planar_float_avx2_transpose_vstripe_ks4<0>; break; | |
| 2115 | } | ||
| 2116 | } | ||
| 2117 | ✗ | return resize_h_planar_float_avx2_permutex_vstripe_ks4; | |
| 2118 | } | ||
| 2119 | #if 0 | ||
| 2120 | // ks8 DISABLED: slower than generic (1102 fps vs 1196 fps). | ||
| 2121 | // Root cause: avx2_permutex2var_ps (dual-source simulation) costs 3 port-5 ops per call | ||
| 2122 | // (2x permutevar8x32 + blendv). With 8 taps: 24 port-5 ops per y-iteration, making port 5 | ||
| 2123 | // the bottleneck at ~24 cycles. The generic uses gather (ports 2+3, load units) instead, | ||
| 2124 | // not competing on port 5 at all. Single-source ks4 uses 1 port-5 per tap (4 total) and wins; | ||
| 2125 | // dual-source ks8 uses 3x that and loses. Crossover is around tap 5. | ||
| 2126 | // Pre-transposing coefficients is feasible but saves only outside-y-loop cost (amortized | ||
| 2127 | // over height scanlines = negligible) and does not touch the y-loop bottleneck. | ||
| 2128 | if (program->filter_size_real <= 8) { | ||
| 2129 | // check(8,16,8)=false: 8 output pixels' source span fits in 16 floats (two YMMs via avx2_permutex2var_ps) | ||
| 2130 | if (!program->resize_h_planar_gather_permutex_vstripe_check(8, 16, 8)) { | ||
| 2131 | out_resampler_h_alternative_for_mt = resizer_h_avx2_generic_float_pix16_sub4_ks_4_8_16; | ||
| 2132 | return resize_h_planar_float_avx2_permutex_vstripe_ks8; | ||
| 2133 | } | ||
| 2134 | } | ||
| 2135 | #endif | ||
| 2136 | ✗ | return resizer_h_avx2_generic_float_pix16_sub4_ks_4_8_16; // new generic, like avx512 version | |
| 2137 | // return resizer_h_avx2_generic_float; old generic would be named pix8_sub2_ks8 | ||
| 2138 | } | ||
| 2139 | ✗ | if (CPU & CPUF_SSSE3) { | |
| 2140 | // up to 4 coeffs it can be highly optimized with transposes | ||
| 2141 | // These perform very poorly in Prefetch, so we provide alternative generic version for MT | ||
| 2142 | ✗ | if (program->filter_size_real <= 4) | |
| 2143 | ✗ | out_resampler_h_alternative_for_mt = resizer_h_ssse3_generic_float; // jolly joker | |
| 2144 | ✗ | switch (program->filter_size_real) { | |
| 2145 | ✗ | case 1: return resize_h_planar_float_sse_transpose_vstripe_ks4<1>; break; | |
| 2146 | ✗ | case 2: return resize_h_planar_float_sse_transpose_vstripe_ks4<2>; break; | |
| 2147 | ✗ | case 3: return resize_h_planar_float_sse_transpose_vstripe_ks4<3>; break; | |
| 2148 | ✗ | case 4: return resize_h_planar_float_sse_transpose_vstripe_ks4<0>; break; | |
| 2149 | ✗ | default: return resizer_h_ssse3_generic_float; | |
| 2150 | } | ||
| 2151 | } | ||
| 2152 | #endif | ||
| 2153 | ✗ | return resize_h_c_planar<float, 0>; | |
| 2154 | } | ||
| 2155 | } | ||
| 2156 | |||
| 2157 | 33 | FilteredResizeH::~FilteredResizeH(void) | |
| 2158 | { | ||
| 2159 |
2/4✓ Branch 2 → 3 taken 21 times.
✗ Branch 2 → 6 not taken.
✓ Branch 3 → 4 taken 21 times.
✗ Branch 3 → 6 not taken.
|
21 | if (resampling_program_luma) { delete resampling_program_luma; } |
| 2160 |
3/4✓ Branch 6 → 7 taken 6 times.
✓ Branch 6 → 10 taken 15 times.
✓ Branch 7 → 8 taken 6 times.
✗ Branch 7 → 10 not taken.
|
21 | if (resampling_program_chroma) { delete resampling_program_chroma; } |
| 2161 | 33 | } | |
| 2162 | |||
| 2163 | /*************************************** | ||
| 2164 | ***** Filtered Resize - Vertical ****** | ||
| 2165 | ***************************************/ | ||
| 2166 | |||
| 2167 | 22 | FilteredResizeV::FilteredResizeV(PClip _child, double subrange_top, double subrange_height, | |
| 2168 | int target_height, ResamplingFunction* func, | ||
| 2169 | bool preserve_center, int chroma_placement, | ||
| 2170 | 22 | IScriptEnvironment* env) | |
| 2171 | : GenericVideoFilter(_child), | ||
| 2172 |
2/4✓ Branch 2 → 3 taken 22 times.
✗ Branch 2 → 51 not taken.
✓ Branch 3 → 4 taken 22 times.
✗ Branch 3 → 49 not taken.
|
22 | resampling_program_luma(0), resampling_program_chroma(0) |
| 2173 | { | ||
| 2174 |
1/2✗ Branch 5 → 6 not taken.
✓ Branch 5 → 7 taken 22 times.
|
22 | if (target_height <= 0) |
| 2175 | ✗ | env->ThrowError("Resize: Height must be greater than 0."); | |
| 2176 | |||
| 2177 |
1/2✓ Branch 7 → 8 taken 22 times.
✗ Branch 7 → 52 not taken.
|
22 | pixelsize = vi.ComponentSize(); // AVS16 |
| 2178 |
1/2✓ Branch 8 → 9 taken 22 times.
✗ Branch 8 → 52 not taken.
|
22 | bits_per_pixel = vi.BitsPerComponent(); |
| 2179 |
1/2✓ Branch 9 → 10 taken 22 times.
✗ Branch 9 → 52 not taken.
|
22 | grey = vi.IsY(); |
| 2180 |
5/8✓ Branch 10 → 11 taken 22 times.
✗ Branch 10 → 52 not taken.
✓ Branch 11 → 12 taken 21 times.
✓ Branch 11 → 14 taken 1 time.
✓ Branch 12 → 13 taken 21 times.
✗ Branch 12 → 52 not taken.
✗ Branch 13 → 14 not taken.
✓ Branch 13 → 15 taken 21 times.
|
22 | bool isRGBPfamily = vi.IsPlanarRGB() || vi.IsPlanarRGBA(); |
| 2181 | |||
| 2182 |
9/10✓ Branch 16 → 17 taken 22 times.
✗ Branch 16 → 52 not taken.
✓ Branch 17 → 18 taken 20 times.
✓ Branch 17 → 21 taken 2 times.
✓ Branch 18 → 19 taken 8 times.
✓ Branch 18 → 21 taken 12 times.
✓ Branch 19 → 20 taken 7 times.
✓ Branch 19 → 21 taken 1 time.
✓ Branch 22 → 23 taken 7 times.
✓ Branch 22 → 26 taken 15 times.
|
22 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { |
| 2183 |
1/2✓ Branch 23 → 24 taken 7 times.
✗ Branch 23 → 52 not taken.
|
7 | const int mask = (1 << vi.GetPlaneHeightSubsampling(PLANAR_U)) - 1; |
| 2184 | |||
| 2185 |
2/2✓ Branch 24 → 25 taken 1 time.
✓ Branch 24 → 26 taken 6 times.
|
7 | if (target_height & mask) |
| 2186 |
1/2✗ Branch 25 → 26 not taken.
✓ Branch 25 → 52 taken 1 time.
|
1 | env->ThrowError("Resize: Planar destination height must be a multiple of %d.", mask + 1); |
| 2187 | } | ||
| 2188 | |||
| 2189 |
7/8✓ Branch 26 → 27 taken 21 times.
✗ Branch 26 → 52 not taken.
✓ Branch 27 → 28 taken 3 times.
✓ Branch 27 → 30 taken 18 times.
✓ Branch 28 → 29 taken 2 times.
✓ Branch 28 → 30 taken 1 time.
✓ Branch 31 → 32 taken 2 times.
✓ Branch 31 → 33 taken 19 times.
|
21 | if (vi.IsRGB() && !isRGBPfamily) |
| 2190 | 2 | subrange_top = vi.height - subrange_top - subrange_height; // packed RGB upside down | |
| 2191 | |||
| 2192 | #ifdef INTEL_INTRINSICS | ||
| 2193 |
1/2✓ Branch 33 → 34 taken 21 times.
✗ Branch 33 → 52 not taken.
|
21 | int cpu = env->GetCPUFlags(); |
| 2194 | #else | ||
| 2195 | int cpu = 0; | ||
| 2196 | #endif | ||
| 2197 | |||
| 2198 | double center_pos_v_luma; | ||
| 2199 | double center_pos_v_chroma; | ||
| 2200 |
1/2✓ Branch 34 → 35 taken 21 times.
✗ Branch 34 → 52 not taken.
|
21 | GetCenterShiftForResizers(center_pos_v_luma, center_pos_v_chroma, preserve_center, chroma_placement, vi, false /* for vertical */); |
| 2201 | // 3.7.4- parameter, old Avisynth behavior: 0.5, 0.5 | ||
| 2202 | |||
| 2203 | // Create resampling program and pitch table | ||
| 2204 |
1/2✓ Branch 35 → 36 taken 21 times.
✗ Branch 35 → 52 not taken.
|
21 | resampling_program_luma = func->GetResamplingProgram(vi.height, subrange_top, subrange_height, target_height, bits_per_pixel, |
| 2205 | center_pos_v_luma, center_pos_v_luma, // for resizing it's the same for source and dest | ||
| 2206 | env); | ||
| 2207 |
1/2✓ Branch 36 → 37 taken 21 times.
✗ Branch 36 → 52 not taken.
|
21 | resampler_luma = GetResampler(cpu, pixelsize, bits_per_pixel, resampling_program_luma, env); |
| 2208 | |||
| 2209 |
9/10✓ Branch 37 → 38 taken 21 times.
✗ Branch 37 → 52 not taken.
✓ Branch 38 → 39 taken 19 times.
✓ Branch 38 → 42 taken 2 times.
✓ Branch 39 → 40 taken 7 times.
✓ Branch 39 → 42 taken 12 times.
✓ Branch 40 → 41 taken 6 times.
✓ Branch 40 → 42 taken 1 time.
✓ Branch 43 → 44 taken 6 times.
✓ Branch 43 → 48 taken 15 times.
|
21 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { |
| 2210 |
1/2✓ Branch 44 → 45 taken 6 times.
✗ Branch 44 → 52 not taken.
|
6 | const int shift = vi.GetPlaneHeightSubsampling(PLANAR_U); |
| 2211 | 6 | const int div = 1 << shift; | |
| 2212 | |||
| 2213 | 12 | resampling_program_chroma = func->GetResamplingProgram( | |
| 2214 |
1/2✓ Branch 45 → 46 taken 6 times.
✗ Branch 45 → 52 not taken.
|
6 | vi.height >> shift, |
| 2215 | subrange_top / div, | ||
| 2216 | subrange_height / div, | ||
| 2217 | target_height >> shift, | ||
| 2218 | bits_per_pixel, | ||
| 2219 | center_pos_v_chroma, center_pos_v_chroma, // for resizing it's the same for source and dest | ||
| 2220 | env); | ||
| 2221 | |||
| 2222 |
1/2✓ Branch 46 → 47 taken 6 times.
✗ Branch 46 → 52 not taken.
|
6 | resampler_chroma = GetResampler(cpu, pixelsize, bits_per_pixel, resampling_program_chroma, env); |
| 2223 | } | ||
| 2224 | |||
| 2225 | // Change target video info size | ||
| 2226 | 21 | vi.height = target_height; | |
| 2227 | 22 | } | |
| 2228 | |||
| 2229 | 13 | PVideoFrame __stdcall FilteredResizeV::GetFrame(int n, IScriptEnvironment* env) | |
| 2230 | { | ||
| 2231 |
1/2✓ Branch 3 → 4 taken 13 times.
✗ Branch 3 → 93 not taken.
|
13 | PVideoFrame src = child->GetFrame(n, env); |
| 2232 |
1/2✓ Branch 4 → 5 taken 13 times.
✗ Branch 4 → 91 not taken.
|
13 | PVideoFrame dst = env->NewVideoFrameP(vi, &src); |
| 2233 |
1/2✓ Branch 6 → 7 taken 13 times.
✗ Branch 6 → 89 not taken.
|
13 | int src_pitch = src->GetPitch(); |
| 2234 |
1/2✓ Branch 8 → 9 taken 13 times.
✗ Branch 8 → 89 not taken.
|
13 | int dst_pitch = dst->GetPitch(); |
| 2235 |
1/2✓ Branch 10 → 11 taken 13 times.
✗ Branch 10 → 89 not taken.
|
13 | const BYTE* srcp = src->GetReadPtr(); |
| 2236 |
1/2✓ Branch 12 → 13 taken 13 times.
✗ Branch 12 → 89 not taken.
|
13 | BYTE* dstp = dst->GetWritePtr(); |
| 2237 | |||
| 2238 |
5/8✓ Branch 13 → 14 taken 13 times.
✗ Branch 13 → 89 not taken.
✓ Branch 14 → 15 taken 12 times.
✓ Branch 14 → 17 taken 1 time.
✓ Branch 15 → 16 taken 12 times.
✗ Branch 15 → 89 not taken.
✗ Branch 16 → 17 not taken.
✓ Branch 16 → 18 taken 12 times.
|
13 | bool isRGBPfamily = vi.IsPlanarRGB() || vi.IsPlanarRGBA(); |
| 2239 | |||
| 2240 | // Do resizing | ||
| 2241 |
4/6✓ Branch 19 → 20 taken 13 times.
✗ Branch 19 → 89 not taken.
✓ Branch 20 → 21 taken 11 times.
✓ Branch 20 → 22 taken 2 times.
✓ Branch 22 → 23 taken 2 times.
✗ Branch 22 → 89 not taken.
|
13 | int work_width = vi.IsPlanar() ? vi.width : vi.BytesFromPixels(vi.width) / pixelsize; // packed RGB: or vi.width * vi.NumComponent() |
| 2242 |
1/2✓ Branch 24 → 25 taken 13 times.
✗ Branch 24 → 89 not taken.
|
13 | resampler_luma(dstp, srcp, dst_pitch, src_pitch, resampling_program_luma, work_width, vi.height, bits_per_pixel); |
| 2243 |
2/2✓ Branch 25 → 26 taken 1 time.
✓ Branch 25 → 44 taken 12 times.
|
13 | if (isRGBPfamily) |
| 2244 | { | ||
| 2245 |
1/2✓ Branch 27 → 28 taken 1 time.
✗ Branch 27 → 89 not taken.
|
1 | src_pitch = src->GetPitch(PLANAR_B); |
| 2246 |
1/2✓ Branch 29 → 30 taken 1 time.
✗ Branch 29 → 89 not taken.
|
1 | dst_pitch = dst->GetPitch(PLANAR_B); |
| 2247 |
1/2✓ Branch 31 → 32 taken 1 time.
✗ Branch 31 → 89 not taken.
|
1 | srcp = src->GetReadPtr(PLANAR_B); |
| 2248 |
1/2✓ Branch 33 → 34 taken 1 time.
✗ Branch 33 → 89 not taken.
|
1 | dstp = dst->GetWritePtr(PLANAR_B); |
| 2249 | |||
| 2250 |
1/2✓ Branch 34 → 35 taken 1 time.
✗ Branch 34 → 89 not taken.
|
1 | resampler_luma(dstp, srcp, dst_pitch, src_pitch, resampling_program_luma, work_width, vi.height, bits_per_pixel); |
| 2251 | |||
| 2252 |
1/2✓ Branch 36 → 37 taken 1 time.
✗ Branch 36 → 89 not taken.
|
1 | src_pitch = src->GetPitch(PLANAR_R); |
| 2253 |
1/2✓ Branch 38 → 39 taken 1 time.
✗ Branch 38 → 89 not taken.
|
1 | dst_pitch = dst->GetPitch(PLANAR_R); |
| 2254 |
1/2✓ Branch 40 → 41 taken 1 time.
✗ Branch 40 → 89 not taken.
|
1 | srcp = src->GetReadPtr(PLANAR_R); |
| 2255 |
1/2✓ Branch 42 → 43 taken 1 time.
✗ Branch 42 → 89 not taken.
|
1 | dstp = dst->GetWritePtr(PLANAR_R); |
| 2256 | |||
| 2257 |
1/2✓ Branch 43 → 70 taken 1 time.
✗ Branch 43 → 89 not taken.
|
1 | resampler_luma(dstp, srcp, dst_pitch, src_pitch, resampling_program_luma, work_width, vi.height, bits_per_pixel); |
| 2258 | } | ||
| 2259 |
7/8✓ Branch 44 → 45 taken 8 times.
✓ Branch 44 → 48 taken 4 times.
✓ Branch 45 → 46 taken 8 times.
✗ Branch 45 → 89 not taken.
✓ Branch 46 → 47 taken 6 times.
✓ Branch 46 → 48 taken 2 times.
✓ Branch 49 → 50 taken 6 times.
✓ Branch 49 → 70 taken 6 times.
|
12 | else if (!grey && vi.IsPlanar()) { |
| 2260 |
1/2✓ Branch 50 → 51 taken 6 times.
✗ Branch 50 → 89 not taken.
|
6 | int width = vi.width >> vi.GetPlaneWidthSubsampling(PLANAR_U); |
| 2261 |
1/2✓ Branch 51 → 52 taken 6 times.
✗ Branch 51 → 89 not taken.
|
6 | int height = vi.height >> vi.GetPlaneHeightSubsampling(PLANAR_U); |
| 2262 | |||
| 2263 | // Plane U resizing | ||
| 2264 |
1/2✓ Branch 53 → 54 taken 6 times.
✗ Branch 53 → 89 not taken.
|
6 | src_pitch = src->GetPitch(PLANAR_U); |
| 2265 |
1/2✓ Branch 55 → 56 taken 6 times.
✗ Branch 55 → 89 not taken.
|
6 | dst_pitch = dst->GetPitch(PLANAR_U); |
| 2266 |
1/2✓ Branch 57 → 58 taken 6 times.
✗ Branch 57 → 89 not taken.
|
6 | srcp = src->GetReadPtr(PLANAR_U); |
| 2267 |
1/2✓ Branch 59 → 60 taken 6 times.
✗ Branch 59 → 89 not taken.
|
6 | dstp = dst->GetWritePtr(PLANAR_U); |
| 2268 | |||
| 2269 |
1/2✓ Branch 60 → 61 taken 6 times.
✗ Branch 60 → 89 not taken.
|
6 | resampler_chroma(dstp, srcp, dst_pitch, src_pitch, resampling_program_chroma, width, height, bits_per_pixel); |
| 2270 | |||
| 2271 | // Plane V resizing | ||
| 2272 |
1/2✓ Branch 62 → 63 taken 6 times.
✗ Branch 62 → 89 not taken.
|
6 | src_pitch = src->GetPitch(PLANAR_V); |
| 2273 |
1/2✓ Branch 64 → 65 taken 6 times.
✗ Branch 64 → 89 not taken.
|
6 | dst_pitch = dst->GetPitch(PLANAR_V); |
| 2274 |
1/2✓ Branch 66 → 67 taken 6 times.
✗ Branch 66 → 89 not taken.
|
6 | srcp = src->GetReadPtr(PLANAR_V); |
| 2275 |
1/2✓ Branch 68 → 69 taken 6 times.
✗ Branch 68 → 89 not taken.
|
6 | dstp = dst->GetWritePtr(PLANAR_V); |
| 2276 | |||
| 2277 |
1/2✓ Branch 69 → 70 taken 6 times.
✗ Branch 69 → 89 not taken.
|
6 | resampler_chroma(dstp, srcp, dst_pitch, src_pitch, resampling_program_chroma, width, height, bits_per_pixel); |
| 2278 | } | ||
| 2279 | |||
| 2280 |
7/10✓ Branch 70 → 71 taken 13 times.
✗ Branch 70 → 89 not taken.
✓ Branch 71 → 72 taken 12 times.
✓ Branch 71 → 74 taken 1 time.
✓ Branch 72 → 73 taken 12 times.
✗ Branch 72 → 89 not taken.
✗ Branch 73 → 74 not taken.
✓ Branch 73 → 75 taken 12 times.
✓ Branch 76 → 77 taken 1 time.
✓ Branch 76 → 86 taken 12 times.
|
13 | if (vi.IsYUVA() || vi.IsPlanarRGBA()) { |
| 2281 |
1/2✓ Branch 78 → 79 taken 1 time.
✗ Branch 78 → 89 not taken.
|
1 | src_pitch = src->GetPitch(PLANAR_A); |
| 2282 |
1/2✓ Branch 80 → 81 taken 1 time.
✗ Branch 80 → 89 not taken.
|
1 | dst_pitch = dst->GetPitch(PLANAR_A); |
| 2283 |
1/2✓ Branch 82 → 83 taken 1 time.
✗ Branch 82 → 89 not taken.
|
1 | srcp = src->GetReadPtr(PLANAR_A); |
| 2284 |
1/2✓ Branch 84 → 85 taken 1 time.
✗ Branch 84 → 89 not taken.
|
1 | dstp = dst->GetWritePtr(PLANAR_A); |
| 2285 |
1/2✓ Branch 85 → 86 taken 1 time.
✗ Branch 85 → 89 not taken.
|
1 | resampler_luma(dstp, srcp, dst_pitch, src_pitch, resampling_program_luma, work_width, vi.height, bits_per_pixel); |
| 2286 | } | ||
| 2287 | |||
| 2288 | 13 | return dst; | |
| 2289 | 13 | } | |
| 2290 | |||
| 2291 | 29 | ResamplerV FilteredResizeV::GetResampler(int CPU, int pixelsize, int bits_per_pixel, ResamplingProgram* program, IScriptEnvironment* env) | |
| 2292 | { | ||
| 2293 | |||
| 2294 | 29 | resize_prepare_coeffs(program, env, 8); | |
| 2295 | // for SIMD friendliness and more: consolidate the kernel_size vs filter_size at the end. | ||
| 2296 | // See comments at FilteredResizeH::GetResampler | ||
| 2297 | |||
| 2298 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 2299 | 29 | const bool has_AVX512_base = (CPU & CPUF_AVX512_BASE) == CPUF_AVX512_BASE; | |
| 2300 | #endif | ||
| 2301 | |||
| 2302 |
1/2✗ Branch 3 → 4 not taken.
✓ Branch 3 → 8 taken 29 times.
|
29 | if (program->filter_size == 1) { |
| 2303 | // Fast pointresize | ||
| 2304 | ✗ | switch (pixelsize) // AVS16 | |
| 2305 | { | ||
| 2306 | ✗ | case 1: return resize_v_planar_pointresize<uint8_t>; | |
| 2307 | ✗ | case 2: return resize_v_planar_pointresize<uint16_t>; | |
| 2308 | ✗ | default: // case 4: | |
| 2309 | ✗ | return resize_v_planar_pointresize<float>; | |
| 2310 | } | ||
| 2311 | } | ||
| 2312 | else { | ||
| 2313 | // Other resizers | ||
| 2314 |
2/2✓ Branch 8 → 9 taken 26 times.
✓ Branch 8 → 16 taken 3 times.
|
29 | if (pixelsize == 1) |
| 2315 | { | ||
| 2316 | #ifdef INTEL_INTRINSICS | ||
| 2317 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 2318 |
1/2✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 26 times.
|
26 | if (has_AVX512_base) |
| 2319 | ✗ | return resize_v_avx512_planar_uint8_t_w_sr; | |
| 2320 | #endif | ||
| 2321 |
1/2✓ Branch 11 → 12 taken 26 times.
✗ Branch 11 → 13 not taken.
|
26 | if (CPU & CPUF_AVX2) |
| 2322 | 26 | return resize_v_avx2_planar_uint8_t; | |
| 2323 | ✗ | if (CPU & CPUF_SSE2) | |
| 2324 | ✗ | return resize_v_sse2_planar; | |
| 2325 | #ifdef X86_32 | ||
| 2326 | if (CPU & CPUF_MMX) | ||
| 2327 | return resize_v_mmx_planar; | ||
| 2328 | #endif | ||
| 2329 | #endif | ||
| 2330 | // C version | ||
| 2331 | ✗ | return resize_v_c_planar_uint8_16_t_auto_vectorized<uint8_t, true>; | |
| 2332 | } | ||
| 2333 |
1/2✓ Branch 16 → 17 taken 3 times.
✗ Branch 16 → 32 not taken.
|
3 | else if (pixelsize == 2) |
| 2334 | { | ||
| 2335 | #ifdef INTEL_INTRINSICS | ||
| 2336 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 2337 |
1/2✗ Branch 17 → 18 not taken.
✓ Branch 17 → 21 taken 3 times.
|
3 | if (has_AVX512_base) { |
| 2338 | ✗ | if (bits_per_pixel < 16) | |
| 2339 | ✗ | return resize_v_avx512_planar_uint16_t_w_sr<true>; | |
| 2340 | else | ||
| 2341 | ✗ | return resize_v_avx512_planar_uint16_t_w_sr<false>; | |
| 2342 | } | ||
| 2343 | #endif | ||
| 2344 |
1/2✓ Branch 21 → 22 taken 3 times.
✗ Branch 21 → 25 not taken.
|
3 | if (CPU & CPUF_AVX2) { |
| 2345 |
1/2✗ Branch 22 → 23 not taken.
✓ Branch 22 → 24 taken 3 times.
|
3 | if (bits_per_pixel < 16) |
| 2346 | ✗ | return resize_v_avx2_planar_uint16_t<true>; | |
| 2347 | else | ||
| 2348 | 3 | return resize_v_avx2_planar_uint16_t<false>; | |
| 2349 | } | ||
| 2350 | ✗ | if (CPU & CPUF_SSE2) { | |
| 2351 | ✗ | if (bits_per_pixel < 16) | |
| 2352 | ✗ | return resize_v_sse2_planar_uint16_t<true>; | |
| 2353 | else | ||
| 2354 | ✗ | return resize_v_sse2_planar_uint16_t<false>; | |
| 2355 | } | ||
| 2356 | #endif | ||
| 2357 | // C version | ||
| 2358 | ✗ | if(bits_per_pixel == 16) | |
| 2359 | ✗ | return resize_v_c_planar_uint8_16_t_auto_vectorized<uint16_t, false>; | |
| 2360 | else | ||
| 2361 | ✗ | return resize_v_c_planar_uint8_16_t_auto_vectorized<uint16_t, true>; | |
| 2362 | } | ||
| 2363 | else // pixelsize== 4 | ||
| 2364 | { | ||
| 2365 | #ifdef INTEL_INTRINSICS | ||
| 2366 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 2367 | ✗ | if (has_AVX512_base) { | |
| 2368 | // return resize_v_avx512_planar_float; // Old, base version, quicker than avx2 version | ||
| 2369 | // This one is about equal to avx2 version, but only with clang, | ||
| 2370 | // it seems that clang is too good and, probably unrolls the old function version | ||
| 2371 | // out-of-box so much better than MSVC, that it competes with the _w_sr version. | ||
| 2372 | // With MSVC its no-brainer to use avx512 | ||
| 2373 | ✗ | return resize_v_avx512_planar_float_w_sr; | |
| 2374 | } | ||
| 2375 | #endif | ||
| 2376 | ✗ | if (CPU & CPUF_AVX2) { | |
| 2377 | ✗ | return resize_v_avx2_planar_float_w_sr; | |
| 2378 | // a memory-optimized version of resize_v_avx2_planar_float | ||
| 2379 | } | ||
| 2380 | ✗ | if (CPU & CPUF_SSE2) { | |
| 2381 | ✗ | return resize_v_sse2_planar_float; | |
| 2382 | } | ||
| 2383 | #endif | ||
| 2384 | ✗ | return resize_v_c_planar_float_auto_vectorized; | |
| 2385 | } | ||
| 2386 | } | ||
| 2387 | } | ||
| 2388 | |||
| 2389 | 33 | FilteredResizeV::~FilteredResizeV(void) | |
| 2390 | { | ||
| 2391 |
2/4✓ Branch 2 → 3 taken 21 times.
✗ Branch 2 → 6 not taken.
✓ Branch 3 → 4 taken 21 times.
✗ Branch 3 → 6 not taken.
|
21 | if (resampling_program_luma) { delete resampling_program_luma; } |
| 2392 |
3/4✓ Branch 6 → 7 taken 6 times.
✓ Branch 6 → 10 taken 15 times.
✓ Branch 7 → 8 taken 6 times.
✗ Branch 7 → 10 not taken.
|
21 | if (resampling_program_chroma) { delete resampling_program_chroma; } |
| 2393 | 33 | } | |
| 2394 | |||
| 2395 | |||
| 2396 | /********************************************** | ||
| 2397 | ******* Resampling Factory Methods ******* | ||
| 2398 | **********************************************/ | ||
| 2399 | |||
| 2400 | 12 | PClip FilteredResize::CreateResizeH(PClip clip, double subrange_left, double subrange_width, int target_width, bool force, | |
| 2401 | ResamplingFunction* func, bool preserve_center, int chroma_placement, IScriptEnvironment* env) | ||
| 2402 | { | ||
| 2403 |
1/2✓ Branch 3 → 4 taken 12 times.
✗ Branch 3 → 63 not taken.
|
12 | const VideoInfo& vi = clip->GetVideoInfo(); |
| 2404 |
4/8✓ Branch 4 → 5 taken 12 times.
✗ Branch 4 → 10 not taken.
✓ Branch 5 → 6 taken 7 times.
✓ Branch 5 → 10 taken 5 times.
✗ Branch 6 → 7 not taken.
✓ Branch 6 → 10 taken 7 times.
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 10 not taken.
|
12 | if (!force && subrange_left == 0 && subrange_width == target_width && subrange_width == vi.width) { |
| 2405 | ✗ | return clip; | |
| 2406 | } | ||
| 2407 | /* | ||
| 2408 | // intentionally left here: don't use crop at special edge cases to avoid inconsistent results across params/color spaces | ||
| 2409 | if (subrange_left == int(subrange_left) && subrange_width == target_width | ||
| 2410 | && subrange_left >= 0 && subrange_left + subrange_width <= vi.width) { | ||
| 2411 | const int mask = ((vi.IsYUV() || vi.IsYUVA()) && !vi.IsY()) ? (1 << vi.GetPlaneWidthSubsampling(PLANAR_U)) - 1 : 0; | ||
| 2412 | |||
| 2413 | if (((int(subrange_left) | int(subrange_width)) & mask) == 0) | ||
| 2414 | return new Crop(int(subrange_left), 0, int(subrange_width), vi.height, 0, clip, env); | ||
| 2415 | } | ||
| 2416 | */ | ||
| 2417 | // Convert interleaved yuv to planar yuv | ||
| 2418 |
1/2✓ Branch 10 → 11 taken 12 times.
✗ Branch 10 → 63 not taken.
|
12 | PClip result = clip; |
| 2419 |
2/4✓ Branch 11 → 12 taken 12 times.
✗ Branch 11 → 61 not taken.
✗ Branch 12 → 13 not taken.
✓ Branch 12 → 21 taken 12 times.
|
12 | if (vi.IsYUY2()) { |
| 2420 | ✗ | result = new ConvertYUY2ToYV16_or_Y(result, false /*to_y*/, env); | |
| 2421 | } | ||
| 2422 | |||
| 2423 |
5/12✓ Branch 21 → 22 taken 12 times.
✗ Branch 21 → 61 not taken.
✓ Branch 22 → 23 taken 12 times.
✗ Branch 22 → 51 not taken.
✓ Branch 23 → 24 taken 12 times.
✗ Branch 23 → 49 not taken.
✓ Branch 24 → 25 taken 12 times.
✗ Branch 24 → 49 not taken.
✗ Branch 26 → 27 not taken.
✓ Branch 26 → 28 taken 12 times.
✗ Branch 52 → 53 not taken.
✗ Branch 52 → 54 not taken.
|
12 | result = new FilteredResizeH(result, subrange_left, subrange_width, target_width, func, preserve_center, chroma_placement, env); |
| 2424 | |||
| 2425 |
2/4✓ Branch 28 → 29 taken 12 times.
✗ Branch 28 → 61 not taken.
✗ Branch 29 → 30 not taken.
✓ Branch 29 → 38 taken 12 times.
|
12 | if (vi.IsYUY2()) { |
| 2426 | ✗ | result = new ConvertYV16ToYUY2(result, env); | |
| 2427 | } | ||
| 2428 | |||
| 2429 |
1/2✓ Branch 38 → 39 taken 12 times.
✗ Branch 38 → 61 not taken.
|
12 | return result; |
| 2430 | 12 | } | |
| 2431 | |||
| 2432 | |||
| 2433 | 12 | PClip FilteredResize::CreateResizeV(PClip clip, double subrange_top, double subrange_height, int target_height, bool force, | |
| 2434 | ResamplingFunction* func, bool preserve_center, int chroma_placement, IScriptEnvironment* env) | ||
| 2435 | { | ||
| 2436 | 12 | const VideoInfo& vi = clip->GetVideoInfo(); | |
| 2437 |
4/8✓ Branch 4 → 5 taken 12 times.
✗ Branch 4 → 9 not taken.
✓ Branch 5 → 6 taken 7 times.
✓ Branch 5 → 9 taken 5 times.
✗ Branch 6 → 7 not taken.
✓ Branch 6 → 9 taken 7 times.
✗ Branch 7 → 8 not taken.
✗ Branch 7 → 9 not taken.
|
12 | if (!force && subrange_top == 0 && subrange_height == target_height && subrange_height == vi.height) { |
| 2438 | ✗ | return clip; | |
| 2439 | } | ||
| 2440 | /* | ||
| 2441 | // intentionally left here: don't use crop at special edge cases to avoid inconsistent results across params/color spaces | ||
| 2442 | if (subrange_top == int(subrange_top) && subrange_height == target_height | ||
| 2443 | && subrange_top >= 0 && subrange_top + subrange_height <= vi.height) { | ||
| 2444 | const int mask = ((vi.IsYUV() || vi.IsYUVA()) && !vi.IsY()) ? (1 << vi.GetPlaneHeightSubsampling(PLANAR_U)) - 1 : 0; | ||
| 2445 | |||
| 2446 | if (((int(subrange_top) | int(subrange_height)) & mask) == 0) | ||
| 2447 | return new Crop(0, int(subrange_top), vi.width, int(subrange_height), 0, clip, env); | ||
| 2448 | } | ||
| 2449 | */ | ||
| 2450 |
4/10✓ Branch 10 → 11 taken 12 times.
✗ Branch 10 → 20 not taken.
✓ Branch 11 → 12 taken 12 times.
✗ Branch 11 → 18 not taken.
✓ Branch 12 → 13 taken 12 times.
✗ Branch 12 → 18 not taken.
✗ Branch 14 → 15 not taken.
✓ Branch 14 → 16 taken 12 times.
✗ Branch 21 → 22 not taken.
✗ Branch 21 → 23 not taken.
|
12 | return new FilteredResizeV(clip, subrange_top, subrange_height, target_height, func, preserve_center, chroma_placement, env); |
| 2451 | } | ||
| 2452 | |||
| 2453 | |||
| 2454 | 12 | PClip FilteredResize::CreateResize(PClip clip, int target_width, int target_height, const AVSValue* args, int force, | |
| 2455 | ResamplingFunction* f, | ||
| 2456 | bool preserve_center, const char* placement_name, const int forced_chroma_placement, | ||
| 2457 | IScriptEnvironment* env) | ||
| 2458 | { | ||
| 2459 | // args 0-1-2-3: left-top-width-height | ||
| 2460 |
1/2✓ Branch 3 → 4 taken 12 times.
✗ Branch 3 → 91 not taken.
|
12 | VideoInfo vi = clip->GetVideoInfo(); |
| 2461 |
2/4✓ Branch 4 → 5 taken 12 times.
✗ Branch 4 → 91 not taken.
✓ Branch 5 → 6 taken 12 times.
✗ Branch 5 → 91 not taken.
|
12 | const double subrange_left = args[0].AsFloat(0), subrange_top = args[1].AsFloat(0); |
| 2462 | |||
| 2463 |
2/4✓ Branch 6 → 7 taken 12 times.
✗ Branch 6 → 91 not taken.
✓ Branch 7 → 8 taken 12 times.
✗ Branch 7 → 91 not taken.
|
12 | double subrange_width = args[2].AsDblDef(vi.width), subrange_height = args[3].AsDblDef(vi.height); |
| 2464 | // Crop style syntax | ||
| 2465 |
2/2✓ Branch 8 → 9 taken 4 times.
✓ Branch 8 → 10 taken 8 times.
|
12 | if (subrange_width <= 0.0) subrange_width = vi.width - subrange_left + subrange_width; |
| 2466 |
2/2✓ Branch 10 → 11 taken 4 times.
✓ Branch 10 → 12 taken 8 times.
|
12 | if (subrange_height <= 0.0) subrange_height = vi.height - subrange_top + subrange_height; |
| 2467 | |||
| 2468 |
1/2✓ Branch 12 → 13 taken 12 times.
✗ Branch 12 → 91 not taken.
|
12 | PClip result; |
| 2469 | // ensure that the intermediate area is maximal | ||
| 2470 | |||
| 2471 | 12 | const double area_FirstH = subrange_height * target_width; | |
| 2472 | 12 | const double area_FirstV = subrange_width * target_height; | |
| 2473 | |||
| 2474 | // "minimal area" logic is not necessarily faster because H and V resizers are not the same speed. | ||
| 2475 | // so we keep the traditional max area logic, which is for quality | ||
| 2476 | |||
| 2477 | // use forced_chroma_placement >= 0 and placement_name == nullptr together | ||
| 2478 | 12 | int chroma_placement = forced_chroma_placement >= 0 ? forced_chroma_placement : ChromaLocation_e::AVS_CHROMA_UNUSED; | |
| 2479 |
2/2✓ Branch 13 → 14 taken 8 times.
✓ Branch 13 → 29 taken 4 times.
|
12 | if (placement_name) { |
| 2480 | // no format-oriented defaults | ||
| 2481 |
7/14✓ Branch 14 → 15 taken 8 times.
✗ Branch 14 → 89 not taken.
✓ Branch 15 → 16 taken 8 times.
✗ Branch 15 → 20 not taken.
✓ Branch 16 → 17 taken 8 times.
✗ Branch 16 → 89 not taken.
✓ Branch 17 → 18 taken 8 times.
✗ Branch 17 → 20 not taken.
✓ Branch 18 → 19 taken 8 times.
✗ Branch 18 → 89 not taken.
✗ Branch 19 → 20 not taken.
✓ Branch 19 → 21 taken 8 times.
✗ Branch 22 → 23 not taken.
✓ Branch 22 → 29 taken 8 times.
|
8 | if (vi.IsYV411() || vi.Is420() || vi.Is422()) { |
| 2482 | // placement explicite parameter like in ConvertToXXX or Text | ||
| 2483 | // input frame properties, if "auto" | ||
| 2484 | // When called from ConvertToXXX, chroma is not involved. | ||
| 2485 | ✗ | auto frame0 = clip->GetFrame(0, env); | |
| 2486 | ✗ | const AVSMap* props = env->getFramePropsRO(frame0); | |
| 2487 | ✗ | chromaloc_parse_merge_with_props(vi, placement_name, props, /* ref*/chroma_placement, ChromaLocation_e::AVS_CHROMA_UNUSED /*default*/, env); | |
| 2488 | ✗ | } | |
| 2489 | |||
| 2490 | } | ||
| 2491 | |||
| 2492 | // 0 - return unchanged if no resize needed | ||
| 2493 | // 1 - force H | ||
| 2494 | // 2 - force V | ||
| 2495 | // 3 - force H and V | ||
| 2496 |
2/4✓ Branch 29 → 30 taken 12 times.
✗ Branch 29 → 31 not taken.
✗ Branch 30 → 31 not taken.
✓ Branch 30 → 32 taken 12 times.
|
12 | const bool force_H = force == 1 || force == 3; |
| 2497 |
2/4✓ Branch 33 → 34 taken 12 times.
✗ Branch 33 → 35 not taken.
✗ Branch 34 → 35 not taken.
✓ Branch 34 → 36 taken 12 times.
|
12 | const bool force_V = force == 2 || force == 3; |
| 2498 | #ifdef DTL2D | ||
| 2499 | if (force == 3) // not very good manual forcing of special 2pass mode, better to nake selection if both H and V resizs required, currently for test only | ||
| 2500 | result = new FilteredResize_2p(clip, | ||
| 2501 | subrange_left, subrange_width, target_width, | ||
| 2502 | subrange_top, subrange_height, target_height, | ||
| 2503 | f, preserve_center, chroma_placement, env); | ||
| 2504 | else | ||
| 2505 | #endif | ||
| 2506 | { | ||
| 2507 |
1/2✗ Branch 37 → 38 not taken.
✓ Branch 37 → 49 taken 12 times.
|
12 | if (area_FirstH < area_FirstV) |
| 2508 | { | ||
| 2509 | ✗ | result = CreateResizeV(clip, subrange_top, subrange_height, target_height, force_V, f, preserve_center, chroma_placement, env); | |
| 2510 | ✗ | result = CreateResizeH(result, subrange_left, subrange_width, target_width, force_H, f, preserve_center, chroma_placement, env); | |
| 2511 | } | ||
| 2512 | else | ||
| 2513 | { | ||
| 2514 |
3/6✓ Branch 49 → 50 taken 12 times.
✗ Branch 49 → 81 not taken.
✓ Branch 50 → 51 taken 12 times.
✗ Branch 50 → 79 not taken.
✓ Branch 51 → 52 taken 12 times.
✗ Branch 51 → 77 not taken.
|
12 | result = CreateResizeH(clip, subrange_left, subrange_width, target_width, force_H, f, preserve_center, chroma_placement, env); |
| 2515 |
3/6✓ Branch 54 → 55 taken 12 times.
✗ Branch 54 → 87 not taken.
✓ Branch 55 → 56 taken 12 times.
✗ Branch 55 → 85 not taken.
✓ Branch 56 → 57 taken 12 times.
✗ Branch 56 → 83 not taken.
|
12 | result = CreateResizeV(result, subrange_top, subrange_height, target_height, force_V, f, preserve_center, chroma_placement, env); |
| 2516 | } | ||
| 2517 | } | ||
| 2518 | 12 | return result; | |
| 2519 | ✗ | } | |
| 2520 | |||
| 2521 | 4 | AVSValue __cdecl FilteredResize::Create_PointResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2522 | { | ||
| 2523 | 4 | auto f = PointFilter(); | |
| 2524 |
2/4✓ Branch 2 → 3 taken 4 times.
✗ Branch 2 → 29 not taken.
✓ Branch 3 → 4 taken 4 times.
✗ Branch 3 → 29 not taken.
|
4 | const int force = args[7].AsInt(0); |
| 2525 | |||
| 2526 |
2/4✓ Branch 4 → 5 taken 4 times.
✗ Branch 4 → 29 not taken.
✓ Branch 5 → 6 taken 4 times.
✗ Branch 5 → 29 not taken.
|
4 | bool preserve_center = args[8].AsBool(true); // [keep_center] default Avisynth |
| 2527 |
2/4✓ Branch 6 → 7 taken 4 times.
✗ Branch 6 → 29 not taken.
✓ Branch 7 → 8 taken 4 times.
✗ Branch 7 → 29 not taken.
|
4 | const char* placement_name = args[9].AsString("auto"); // [placement]s |
| 2528 | 4 | const int forced_chroma_placement = -1; // no force, used internally | |
| 2529 | |||
| 2530 |
9/18✓ Branch 8 → 9 taken 4 times.
✗ Branch 8 → 28 not taken.
✓ Branch 9 → 10 taken 4 times.
✗ Branch 9 → 28 not taken.
✓ Branch 10 → 11 taken 4 times.
✗ Branch 10 → 28 not taken.
✓ Branch 11 → 12 taken 4 times.
✗ Branch 11 → 28 not taken.
✓ Branch 12 → 13 taken 4 times.
✗ Branch 12 → 28 not taken.
✓ Branch 13 → 14 taken 4 times.
✗ Branch 13 → 27 not taken.
✓ Branch 14 → 15 taken 4 times.
✗ Branch 14 → 27 not taken.
✓ Branch 15 → 16 taken 4 times.
✗ Branch 15 → 25 not taken.
✓ Branch 16 → 17 taken 4 times.
✗ Branch 16 → 23 not taken.
|
8 | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); |
| 2531 | 4 | } | |
| 2532 | |||
| 2533 | |||
| 2534 | ✗ | AVSValue __cdecl FilteredResize::Create_BilinearResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2535 | { | ||
| 2536 | ✗ | auto f = TriangleFilter(); | |
| 2537 | ✗ | const int force = args[7].AsInt(0); | |
| 2538 | |||
| 2539 | ✗ | bool preserve_center = args[8].AsBool(true); // [keep_center] default Avisynth | |
| 2540 | ✗ | const char* placement_name = args[9].AsString("auto"); // [placement]s | |
| 2541 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2542 | |||
| 2543 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2544 | ✗ | } | |
| 2545 | |||
| 2546 | |||
| 2547 | ✗ | AVSValue __cdecl FilteredResize::Create_BicubicResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2548 | { | ||
| 2549 | ✗ | auto f = MitchellNetravaliFilter(args[3].AsDblDef(1. / 3.), args[4].AsDblDef(1. / 3.)); | |
| 2550 | ✗ | const int force = args[9].AsInt(0); | |
| 2551 | |||
| 2552 | ✗ | bool preserve_center = args[10].AsBool(true); // [keep_center] default Avisynth | |
| 2553 | ✗ | const char* placement_name = args[11].AsString("auto"); // [placement]s | |
| 2554 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2555 | |||
| 2556 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[5], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2557 | ✗ | } | |
| 2558 | |||
| 2559 | ✗ | AVSValue __cdecl FilteredResize::Create_LanczosResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2560 | { | ||
| 2561 | ✗ | auto f = LanczosFilter(args[7].AsInt(3)); | |
| 2562 | ✗ | const int force = args[8].AsInt(0); | |
| 2563 | |||
| 2564 | ✗ | bool preserve_center = args[9].AsBool(true); // [keep_center] default Avisynth | |
| 2565 | ✗ | const char* placement_name = args[10].AsString("auto"); // [placement]s | |
| 2566 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2567 | |||
| 2568 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2569 | ✗ | } | |
| 2570 | |||
| 2571 | ✗ | AVSValue __cdecl FilteredResize::Create_Lanczos4Resize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2572 | { | ||
| 2573 | ✗ | auto f = LanczosFilter(4); | |
| 2574 | ✗ | const int force = args[7].AsInt(0); | |
| 2575 | |||
| 2576 | ✗ | bool preserve_center = args[8].AsBool(true); // [keep_center] default Avisynth | |
| 2577 | ✗ | const char* placement_name = args[9].AsString("auto"); // [placement]s | |
| 2578 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2579 | |||
| 2580 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2581 | ✗ | } | |
| 2582 | |||
| 2583 | ✗ | AVSValue __cdecl FilteredResize::Create_BlackmanResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2584 | { | ||
| 2585 | ✗ | auto f = BlackmanFilter(args[7].AsInt(4)); | |
| 2586 | ✗ | const int force = args[8].AsInt(0); | |
| 2587 | |||
| 2588 | ✗ | bool preserve_center = args[8].AsBool(true); // [keep_center] default Avisynth | |
| 2589 | ✗ | const char* placement_name = args[9].AsString("auto"); // [placement]s | |
| 2590 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2591 | |||
| 2592 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2593 | ✗ | } | |
| 2594 | |||
| 2595 | ✗ | AVSValue __cdecl FilteredResize::Create_Spline16Resize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2596 | { | ||
| 2597 | ✗ | auto f = Spline16Filter(); | |
| 2598 | ✗ | const int force = args[7].AsInt(0); | |
| 2599 | |||
| 2600 | ✗ | bool preserve_center = args[8].AsBool(true); // [keep_center] default Avisynth | |
| 2601 | ✗ | const char* placement_name = args[9].AsString("auto"); // [placement]s | |
| 2602 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2603 | |||
| 2604 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2605 | ✗ | } | |
| 2606 | |||
| 2607 | ✗ | AVSValue __cdecl FilteredResize::Create_Spline36Resize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2608 | { | ||
| 2609 | ✗ | auto f = Spline36Filter(); | |
| 2610 | ✗ | const int force = args[7].AsInt(0); | |
| 2611 | |||
| 2612 | ✗ | bool preserve_center = args[8].AsBool(true); // [keep_center] default Avisynth | |
| 2613 | ✗ | const char* placement_name = args[9].AsString("auto"); // [placement]s | |
| 2614 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2615 | |||
| 2616 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2617 | ✗ | } | |
| 2618 | |||
| 2619 | ✗ | AVSValue __cdecl FilteredResize::Create_Spline64Resize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2620 | { | ||
| 2621 | ✗ | auto f = Spline64Filter(); | |
| 2622 | ✗ | const int force = args[7].AsInt(0); | |
| 2623 | |||
| 2624 | ✗ | bool preserve_center = args[8].AsBool(true); // [keep_center] default Avisynth | |
| 2625 | ✗ | const char* placement_name = args[9].AsString("auto"); // [placement]s | |
| 2626 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2627 | |||
| 2628 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2629 | ✗ | } | |
| 2630 | |||
| 2631 | ✗ | AVSValue __cdecl FilteredResize::Create_GaussianResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2632 | { | ||
| 2633 | ✗ | auto f = GaussianFilter(args[7].AsFloat(30.0f), args[8].AsFloat(2.0f), args[9].AsFloat(4.0f)); // defaults at two more places | |
| 2634 | ✗ | const int force = args[10].AsInt(0); | |
| 2635 | |||
| 2636 | ✗ | bool preserve_center = args[11].AsBool(true); // [keep_center] default Avisynth | |
| 2637 | ✗ | const char* placement_name = args[12].AsString("auto"); // [placement]s | |
| 2638 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2639 | |||
| 2640 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2641 | ✗ | } | |
| 2642 | |||
| 2643 | ✗ | AVSValue __cdecl FilteredResize::Create_SincResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2644 | { | ||
| 2645 | ✗ | auto f = SincFilter(args[7].AsInt(4)); | |
| 2646 | ✗ | const int force = args[8].AsInt(0); | |
| 2647 | |||
| 2648 | ✗ | bool preserve_center = args[9].AsBool(true); // [keep_center] default Avisynth | |
| 2649 | ✗ | const char * placement_name = args[10].AsString("auto"); // [placement]s | |
| 2650 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2651 | |||
| 2652 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2653 | ✗ | } | |
| 2654 | |||
| 2655 | // like GaussianFilter(); optional P | ||
| 2656 | ✗ | AVSValue __cdecl FilteredResize::Create_SinPowerResize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2657 | { | ||
| 2658 | ✗ | auto f = SinPowerFilter(args[7].AsFloat(2.5f)); | |
| 2659 | ✗ | const int force = args[8].AsInt(0); | |
| 2660 | |||
| 2661 | ✗ | bool preserve_center = args[9].AsBool(true); // [keep_center] default Avisynth | |
| 2662 | ✗ | const char* placement_name = args[10].AsString("auto"); // [placement]s | |
| 2663 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2664 | |||
| 2665 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2666 | ✗ | } | |
| 2667 | |||
| 2668 | // like SincFilter or LanczosFilter: optional Taps | ||
| 2669 | ✗ | AVSValue __cdecl FilteredResize::Create_SincLin2Resize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2670 | { | ||
| 2671 | ✗ | auto f = SincLin2Filter(args[7].AsInt(15)); | |
| 2672 | ✗ | const int force = args[8].AsInt(0); | |
| 2673 | |||
| 2674 | ✗ | bool preserve_center = args[9].AsBool(true); // [keep_center] default Avisynth | |
| 2675 | ✗ | const char* placement_name = args[10].AsString("auto"); // [placement]s | |
| 2676 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2677 | |||
| 2678 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[3], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2679 | ✗ | } | |
| 2680 | |||
| 2681 | // like bicubic, plus 's'upport: optional B and C and S | ||
| 2682 | ✗ | AVSValue __cdecl FilteredResize::Create_UserDefined2Resize(AVSValue args, void*, IScriptEnvironment* env) | |
| 2683 | { | ||
| 2684 | ✗ | auto f = UserDefined2Filter(args[3].AsFloat(121.0f), args[4].AsFloat(19.0f), args[5].AsFloat(2.3f)); | |
| 2685 | ✗ | const int force = args[10].AsInt(0); | |
| 2686 | |||
| 2687 | ✗ | bool preserve_center = args[11].AsBool(true); // [keep_center] default Avisynth | |
| 2688 | ✗ | const char* placement_name = args[12].AsString("auto"); // [placement]s | |
| 2689 | ✗ | const int forced_chroma_placement = -1; // no force, used internally | |
| 2690 | |||
| 2691 | ✗ | return CreateResize(args[0].AsClip(), args[1].AsInt(), args[2].AsInt(), &args[6], force, &f, preserve_center, placement_name, forced_chroma_placement, env); | |
| 2692 | ✗ | } | |
| 2693 | |||
| 2694 | #ifdef DTL2D | ||
| 2695 | |||
| 2696 | /*************************************** | ||
| 2697 | ***** Filtered Resize - 2p ****** | ||
| 2698 | ***************************************/ | ||
| 2699 | |||
| 2700 | FilteredResize_2p::FilteredResize_2p(PClip _child, | ||
| 2701 | double subrange_left, double subrange_width, int target_width, | ||
| 2702 | double subrange_top, double subrange_height, int target_height, | ||
| 2703 | ResamplingFunction* func, bool preserve_center, int chroma_placement, IScriptEnvironment* env) | ||
| 2704 | : GenericVideoFilter(_child), | ||
| 2705 | resampling_program_luma_h(0), resampling_program_chroma_h(0), | ||
| 2706 | resampling_program_luma_v(0), resampling_program_chroma_v(0) | ||
| 2707 | { | ||
| 2708 | if (target_height <= 0) | ||
| 2709 | env->ThrowError("Resize: Height must be greater than 0."); | ||
| 2710 | |||
| 2711 | if (target_width <= 0) | ||
| 2712 | env->ThrowError("Resize: Width must be greater than 0."); | ||
| 2713 | |||
| 2714 | // set class globals | ||
| 2715 | src_width = vi.width; | ||
| 2716 | src_height = vi.height; | ||
| 2717 | dst_width = target_width; | ||
| 2718 | dst_height = target_height; | ||
| 2719 | |||
| 2720 | pixelsize = vi.ComponentSize(); // AVS16 | ||
| 2721 | bits_per_pixel = vi.BitsPerComponent(); | ||
| 2722 | grey = vi.IsY(); | ||
| 2723 | bool isRGBPfamily = vi.IsPlanarRGB() || vi.IsPlanarRGBA(); | ||
| 2724 | |||
| 2725 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { | ||
| 2726 | const int mask = (1 << vi.GetPlaneHeightSubsampling(PLANAR_U)) - 1; | ||
| 2727 | |||
| 2728 | if (target_height & mask) | ||
| 2729 | env->ThrowError("Resize: Planar destination height must be a multiple of %d.", mask + 1); | ||
| 2730 | } | ||
| 2731 | |||
| 2732 | if (vi.IsRGB() && !isRGBPfamily) | ||
| 2733 | subrange_top = vi.height - subrange_top - subrange_height; // packed RGB upside down | ||
| 2734 | |||
| 2735 | #ifdef INTEL_INTRINSICS | ||
| 2736 | int cpu = env->GetCPUFlags(); | ||
| 2737 | #else | ||
| 2738 | int cpu = 0; | ||
| 2739 | #endif | ||
| 2740 | |||
| 2741 | double center_pos_v_luma; | ||
| 2742 | double center_pos_v_chroma; | ||
| 2743 | GetCenterShiftForResizers(center_pos_v_luma, center_pos_v_chroma, preserve_center, chroma_placement, vi, false /* for vertical */); | ||
| 2744 | |||
| 2745 | double center_pos_h_luma; | ||
| 2746 | double center_pos_h_chroma; | ||
| 2747 | GetCenterShiftForResizers(center_pos_h_luma, center_pos_h_chroma, preserve_center, chroma_placement, vi, true /* for horizontal */); | ||
| 2748 | // 3.7.4- parameter, old Avisynth behavior: 0.5, 0.5 | ||
| 2749 | |||
| 2750 | // Create resampling program and pitch table for H | ||
| 2751 | resampling_program_luma_h = func->GetResamplingProgram(vi.width, subrange_left, subrange_width, target_width, bits_per_pixel, | ||
| 2752 | center_pos_h_luma, center_pos_h_luma, // for resizing it's the same for source and dest | ||
| 2753 | env); | ||
| 2754 | resampler_luma_h = GetResamplerH(cpu, pixelsize, bits_per_pixel, resampling_program_luma_h, env); | ||
| 2755 | |||
| 2756 | // Create resampling program and pitch table for V | ||
| 2757 | resampling_program_luma_v = func->GetResamplingProgram(vi.height, subrange_top, subrange_height, target_height, bits_per_pixel, | ||
| 2758 | center_pos_v_luma, center_pos_v_luma, // for resizing it's the same for source and dest | ||
| 2759 | env); | ||
| 2760 | resampler_luma_v = GetResamplerV(cpu, pixelsize, bits_per_pixel, resampling_program_luma_v, env); | ||
| 2761 | |||
| 2762 | |||
| 2763 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { | ||
| 2764 | const int shift = vi.GetPlaneHeightSubsampling(PLANAR_U); | ||
| 2765 | const int div = 1 << shift; | ||
| 2766 | |||
| 2767 | resampling_program_chroma_v = func->GetResamplingProgram( | ||
| 2768 | vi.height >> shift, | ||
| 2769 | subrange_top / div, | ||
| 2770 | subrange_height / div, | ||
| 2771 | target_height >> shift, | ||
| 2772 | bits_per_pixel, | ||
| 2773 | center_pos_v_chroma, center_pos_v_chroma, // for resizing it's the same for source and dest | ||
| 2774 | env); | ||
| 2775 | |||
| 2776 | resampler_chroma_v = GetResamplerV(cpu, pixelsize, bits_per_pixel, resampling_program_chroma_v, env); | ||
| 2777 | } | ||
| 2778 | |||
| 2779 | if (vi.IsPlanar() && !grey && !isRGBPfamily) { | ||
| 2780 | const int shift = vi.GetPlaneWidthSubsampling(PLANAR_U); | ||
| 2781 | const int div = 1 << shift; | ||
| 2782 | |||
| 2783 | resampling_program_chroma_h = func->GetResamplingProgram( | ||
| 2784 | vi.width >> shift, | ||
| 2785 | subrange_left / div, | ||
| 2786 | subrange_width / div, | ||
| 2787 | target_width >> shift, | ||
| 2788 | bits_per_pixel, | ||
| 2789 | center_pos_h_chroma, center_pos_h_chroma, // horizontal | ||
| 2790 | env); | ||
| 2791 | |||
| 2792 | resampler_chroma_h = GetResamplerH(cpu, pixelsize, bits_per_pixel, resampling_program_chroma_h, env); | ||
| 2793 | } | ||
| 2794 | |||
| 2795 | // Change target video info size | ||
| 2796 | vi.height = target_height; | ||
| 2797 | vi.width = target_width; | ||
| 2798 | } | ||
| 2799 | |||
| 2800 | #if 0 // expected worse in performance - left for performance tests | ||
| 2801 | PVideoFrame __stdcall FilteredResize_2p::GetFrame(int n, IScriptEnvironment* env) // use env->Allocate() to get temp buf from other allocated memory - it is NOT returned to the memory pool for the NewVideoFrameP for the downstream filter to write to ? | ||
| 2802 | { | ||
| 2803 | PVideoFrame src = child->GetFrame(n, env); | ||
| 2804 | PVideoFrame dst = env->NewVideoFrameP(vi, &src); | ||
| 2805 | int src_pitch = src->GetPitch(); | ||
| 2806 | int dst_pitch = dst->GetPitch(); | ||
| 2807 | const BYTE* srcp = src->GetReadPtr(); | ||
| 2808 | BYTE* dstp = dst->GetWritePtr(); // for first (largest ?) plane or for single ? | ||
| 2809 | |||
| 2810 | bool isRGBPfamily = vi.IsPlanarRGB() || vi.IsPlanarRGBA(); | ||
| 2811 | |||
| 2812 | BYTE* temp_1 = static_cast<BYTE*>(env->Allocate(dst_pitch * dst_height, FRAME_ALIGN, AVS_POOLED_ALLOC)); | ||
| 2813 | if (!temp_1 ) { | ||
| 2814 | env->Free(temp_1); | ||
| 2815 | env->ThrowError("Could not reserve temp memory in a resampler_2p."); | ||
| 2816 | } | ||
| 2817 | |||
| 2818 | // Do resizing, single plane by plane | ||
| 2819 | resampler_luma_h(temp_1, srcp, dst_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2820 | int work_height = vi.IsPlanar() ? vi.width : vi.BytesFromPixels(vi.width) / pixelsize; // packed RGB: or vi.width * vi.NumComponent() | ||
| 2821 | resampler_luma_v(dstp, temp_1, dst_pitch, dst_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2822 | |||
| 2823 | if (isRGBPfamily) | ||
| 2824 | { | ||
| 2825 | src_pitch = src->GetPitch(PLANAR_B); | ||
| 2826 | dst_pitch = dst->GetPitch(PLANAR_B); | ||
| 2827 | srcp = src->GetReadPtr(PLANAR_B); | ||
| 2828 | dstp = dst->GetWritePtr(PLANAR_B); | ||
| 2829 | |||
| 2830 | resampler_luma_h(temp_1, srcp, dst_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2831 | int work_height = vi.IsPlanar() ? vi.width : vi.BytesFromPixels(vi.width) / pixelsize; // packed RGB: or vi.width * vi.NumComponent() | ||
| 2832 | resampler_luma_v(dstp, temp_1, dst_pitch, dst_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2833 | |||
| 2834 | src_pitch = src->GetPitch(PLANAR_R); | ||
| 2835 | dst_pitch = dst->GetPitch(PLANAR_R); | ||
| 2836 | srcp = src->GetReadPtr(PLANAR_R); | ||
| 2837 | dstp = dst->GetWritePtr(PLANAR_R); | ||
| 2838 | |||
| 2839 | resampler_luma_h(temp_1, srcp, dst_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2840 | resampler_luma_v(dstp, temp_1, dst_pitch, dst_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2841 | |||
| 2842 | } | ||
| 2843 | else if (!grey && vi.IsPlanar()) { | ||
| 2844 | int width = vi.width >> vi.GetPlaneWidthSubsampling(PLANAR_U); | ||
| 2845 | int height = vi.height >> vi.GetPlaneHeightSubsampling(PLANAR_U); | ||
| 2846 | |||
| 2847 | // Plane U resizing | ||
| 2848 | src_pitch = src->GetPitch(PLANAR_U); | ||
| 2849 | dst_pitch = dst->GetPitch(PLANAR_U); | ||
| 2850 | srcp = src->GetReadPtr(PLANAR_U); | ||
| 2851 | dstp = dst->GetWritePtr(PLANAR_U); | ||
| 2852 | |||
| 2853 | resampler_chroma_h(temp_1, srcp, dst_pitch, src_pitch, resampling_program_chroma_h, width, src_height >> vi.GetPlaneHeightSubsampling(PLANAR_U), bits_per_pixel); | ||
| 2854 | resampler_chroma_v(dstp, temp_1, dst_pitch, dst_pitch, resampling_program_chroma_v, width, height, bits_per_pixel); | ||
| 2855 | |||
| 2856 | // Plane V resizing | ||
| 2857 | src_pitch = src->GetPitch(PLANAR_V); | ||
| 2858 | dst_pitch = dst->GetPitch(PLANAR_V); | ||
| 2859 | srcp = src->GetReadPtr(PLANAR_V); | ||
| 2860 | dstp = dst->GetWritePtr(PLANAR_V); | ||
| 2861 | |||
| 2862 | resampler_chroma_h(temp_1, srcp, dst_pitch, src_pitch, resampling_program_chroma_h, width, src_height >> vi.GetPlaneHeightSubsampling(PLANAR_U), bits_per_pixel); | ||
| 2863 | resampler_chroma_v(dstp, temp_1, dst_pitch, dst_pitch, resampling_program_chroma_v, width, height, bits_per_pixel); | ||
| 2864 | |||
| 2865 | } | ||
| 2866 | |||
| 2867 | if (vi.IsYUVA() || vi.IsPlanarRGBA()) { | ||
| 2868 | src_pitch = src->GetPitch(PLANAR_A); | ||
| 2869 | dst_pitch = dst->GetPitch(PLANAR_A); | ||
| 2870 | srcp = src->GetReadPtr(PLANAR_A); | ||
| 2871 | dstp = dst->GetWritePtr(PLANAR_A); | ||
| 2872 | |||
| 2873 | resampler_luma_h(temp_1, srcp, dst_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2874 | int work_height = vi.IsPlanar() ? vi.width : vi.BytesFromPixels(vi.width) / pixelsize; // packed RGB: or vi.width * vi.NumComponent() | ||
| 2875 | resampler_luma_v(dstp, temp_1, dst_pitch, dst_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2876 | } | ||
| 2877 | |||
| 2878 | env->Free(temp_1); | ||
| 2879 | |||
| 2880 | return dst; | ||
| 2881 | } | ||
| 2882 | #endif | ||
| 2883 | |||
| 2884 | PVideoFrame __stdcall FilteredResize_2p::GetFrame(int n, IScriptEnvironment* env) // use NewVideoFrame as temp buf to return it in the vfb pool after exit this filter | ||
| 2885 | { | ||
| 2886 | PVideoFrame src = child->GetFrame(n, env); | ||
| 2887 | PVideoFrame dst = env->NewVideoFrameP(vi, &src); | ||
| 2888 | |||
| 2889 | PVideoFrame tmp = env->NewVideoFrame(vi); // no need frame properties copy, use as temporal buffer only and its refcount will be zeroed at function exit with object auto-release/destructor (PVideoFrame::~PVideoFrame() ) | ||
| 2890 | /* | ||
| 2891 | Here we need to ask ScriptEnvironment to look for output format of downstream filter ? So it is not trans-in-place filter we can request frame buffer larger and left | ||
| 2892 | it unused after exiting this function. Only in this case there is a big probability the env->NewVideoFrameP(vi, &src); for downstream filter call will return this same virtual address buffer | ||
| 2893 | to the downstream filter and it can be (at least partially) overwritten saving from useless downloading from CPU cache. It is new TODO idea for modification of ScriptEnvironment vfb memory management. | ||
| 2894 | After this will be implemented - we can use such method of requesting temp buffer (frame) to use in 2pass resize. | ||
| 2895 | If this is last filter in a chain - simply request lowest possible sized frame. | ||
| 2896 | |||
| 2897 | Update 30.05.2025: The expected transfer of tmp buf address to downstream fiter dst frame sometime happens - but how frequently it happens in real scripts running - need to be discovered. | ||
| 2898 | |||
| 2899 | As env->Allocate/Free buffers are definitely worse (only good if downstream filter will request same temp buf for write) - this temp method expected to be faster (as first expectations). | ||
| 2900 | */ | ||
| 2901 | |||
| 2902 | int src_pitch = src->GetPitch(); | ||
| 2903 | int dst_pitch = dst->GetPitch(); | ||
| 2904 | const BYTE* srcp = src->GetReadPtr(); | ||
| 2905 | BYTE* dstp = dst->GetWritePtr(); // for first (largest ?) plane or for single ? | ||
| 2906 | |||
| 2907 | bool isRGBPfamily = vi.IsPlanarRGB() || vi.IsPlanarRGBA(); | ||
| 2908 | |||
| 2909 | const BYTE* tmp_srcp = tmp->GetReadPtr(); | ||
| 2910 | BYTE* tmp_dstp = tmp->GetWritePtr(); // for first (largest ?) plane or for single ? | ||
| 2911 | int tmp_pitch = tmp->GetPitch(); | ||
| 2912 | |||
| 2913 | // Do resizing, single plane by plane | ||
| 2914 | resampler_luma_h(tmp_dstp, srcp, tmp_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2915 | int work_height = vi.IsPlanar() ? vi.width : vi.BytesFromPixels(vi.width) / pixelsize; // packed RGB: or vi.width * vi.NumComponent() | ||
| 2916 | resampler_luma_v(dstp, tmp_srcp, dst_pitch, tmp_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2917 | |||
| 2918 | |||
| 2919 | if (isRGBPfamily) | ||
| 2920 | { | ||
| 2921 | src_pitch = src->GetPitch(PLANAR_B); | ||
| 2922 | dst_pitch = dst->GetPitch(PLANAR_B); | ||
| 2923 | srcp = src->GetReadPtr(PLANAR_B); | ||
| 2924 | dstp = dst->GetWritePtr(PLANAR_B); | ||
| 2925 | |||
| 2926 | resampler_luma_h(tmp_dstp, srcp, tmp_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2927 | int work_height = vi.IsPlanar() ? vi.width : vi.BytesFromPixels(vi.width) / pixelsize; // packed RGB: or vi.width * vi.NumComponent() | ||
| 2928 | resampler_luma_v(dstp, tmp_srcp, dst_pitch, tmp_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2929 | |||
| 2930 | src_pitch = src->GetPitch(PLANAR_R); | ||
| 2931 | dst_pitch = dst->GetPitch(PLANAR_R); | ||
| 2932 | srcp = src->GetReadPtr(PLANAR_R); | ||
| 2933 | dstp = dst->GetWritePtr(PLANAR_R); | ||
| 2934 | |||
| 2935 | resampler_luma_h(tmp_dstp, srcp, tmp_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2936 | resampler_luma_v(dstp, tmp_srcp, dst_pitch, tmp_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2937 | |||
| 2938 | } | ||
| 2939 | else if (!grey && vi.IsPlanar()) { | ||
| 2940 | int width = vi.width >> vi.GetPlaneWidthSubsampling(PLANAR_U); | ||
| 2941 | int height = vi.height >> vi.GetPlaneHeightSubsampling(PLANAR_U); | ||
| 2942 | |||
| 2943 | // Plane U resizing | ||
| 2944 | src_pitch = src->GetPitch(PLANAR_U); | ||
| 2945 | dst_pitch = dst->GetPitch(PLANAR_U); | ||
| 2946 | srcp = src->GetReadPtr(PLANAR_U); | ||
| 2947 | dstp = dst->GetWritePtr(PLANAR_U); | ||
| 2948 | |||
| 2949 | resampler_chroma_h(tmp_dstp, srcp, tmp_pitch, src_pitch, resampling_program_chroma_h, width, src_height >> vi.GetPlaneHeightSubsampling(PLANAR_U), bits_per_pixel); | ||
| 2950 | resampler_chroma_v(dstp, tmp_dstp, dst_pitch, tmp_pitch, resampling_program_chroma_v, width, height, bits_per_pixel); | ||
| 2951 | |||
| 2952 | // Plane V resizing | ||
| 2953 | src_pitch = src->GetPitch(PLANAR_V); | ||
| 2954 | dst_pitch = dst->GetPitch(PLANAR_V); | ||
| 2955 | srcp = src->GetReadPtr(PLANAR_V); | ||
| 2956 | dstp = dst->GetWritePtr(PLANAR_V); | ||
| 2957 | |||
| 2958 | resampler_chroma_h(tmp_dstp, srcp, tmp_pitch, src_pitch, resampling_program_chroma_h, width, src_height >> vi.GetPlaneHeightSubsampling(PLANAR_V), bits_per_pixel); | ||
| 2959 | resampler_chroma_v(dstp, tmp_dstp, dst_pitch, tmp_pitch, resampling_program_chroma_v, width, height, bits_per_pixel); | ||
| 2960 | |||
| 2961 | } | ||
| 2962 | |||
| 2963 | if (vi.IsYUVA() || vi.IsPlanarRGBA()) { | ||
| 2964 | src_pitch = src->GetPitch(PLANAR_A); | ||
| 2965 | dst_pitch = dst->GetPitch(PLANAR_A); | ||
| 2966 | srcp = src->GetReadPtr(PLANAR_A); | ||
| 2967 | dstp = dst->GetWritePtr(PLANAR_A); | ||
| 2968 | |||
| 2969 | resampler_luma_h(tmp_dstp, srcp, tmp_pitch, src_pitch, resampling_program_luma_h, dst_width, src_height, bits_per_pixel); | ||
| 2970 | int work_height = vi.IsPlanar() ? vi.width : vi.BytesFromPixels(vi.width) / pixelsize; // packed RGB: or vi.width * vi.NumComponent() | ||
| 2971 | resampler_luma_v(dstp, tmp_dstp, dst_pitch, tmp_pitch, resampling_program_luma_v, work_height, vi.height, bits_per_pixel); | ||
| 2972 | } | ||
| 2973 | |||
| 2974 | return dst; | ||
| 2975 | } | ||
| 2976 | |||
| 2977 | |||
| 2978 | ResamplerV FilteredResize_2p::GetResamplerV(int CPU, int pixelsize, int bits_per_pixel, ResamplingProgram* program, IScriptEnvironment* env) // may be somehow call same method from FilteredResizeV class ? | ||
| 2979 | { | ||
| 2980 | |||
| 2981 | resize_prepare_coeffs(program, env, 8); | ||
| 2982 | // for SIMD friendliness and more: consolidate the kernel_size vs filter_size at the end. | ||
| 2983 | // See comments at FilteredResizeH::GetResampler | ||
| 2984 | |||
| 2985 | if (program->filter_size == 1) { | ||
| 2986 | // Fast pointresize | ||
| 2987 | switch (pixelsize) // AVS16 | ||
| 2988 | { | ||
| 2989 | case 1: return resize_v_planar_pointresize<uint8_t>; | ||
| 2990 | case 2: return resize_v_planar_pointresize<uint16_t>; | ||
| 2991 | default: // case 4: | ||
| 2992 | return resize_v_planar_pointresize<float>; | ||
| 2993 | } | ||
| 2994 | } | ||
| 2995 | else { | ||
| 2996 | // Other resizers | ||
| 2997 | if (pixelsize == 1) | ||
| 2998 | { | ||
| 2999 | #ifdef INTEL_INTRINSICS | ||
| 3000 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 3001 | if (CPU & CPUF_AVX512F) | ||
| 3002 | return resize_v_avx512_planar_uint8_t_w_sr; | ||
| 3003 | #endif | ||
| 3004 | if (CPU & CPUF_AVX2) | ||
| 3005 | return resize_v_avx2_planar_uint8_t; | ||
| 3006 | if (CPU & CPUF_SSE2) | ||
| 3007 | return resize_v_sse2_planar; | ||
| 3008 | #ifdef X86_32 | ||
| 3009 | if (CPU & CPUF_MMX) | ||
| 3010 | return resize_v_mmx_planar; | ||
| 3011 | #endif | ||
| 3012 | #endif | ||
| 3013 | // C version | ||
| 3014 | return resize_v_c_planar_uint8_16_t_auto_vectorized<uint8_t, true>; | ||
| 3015 | } | ||
| 3016 | else if (pixelsize == 2) | ||
| 3017 | { | ||
| 3018 | #ifdef INTEL_INTRINSICS | ||
| 3019 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 3020 | if (CPU & CPUF_AVX512F) | ||
| 3021 | if (bits_per_pixel < 16) | ||
| 3022 | return resize_v_avx512_planar_uint16_t_w_sr<true>; | ||
| 3023 | else | ||
| 3024 | return resize_v_avx512_planar_uint16_t_w_sr<false>; | ||
| 3025 | #endif | ||
| 3026 | if (CPU & CPUF_AVX2) { | ||
| 3027 | if (bits_per_pixel < 16) | ||
| 3028 | return resize_v_avx2_planar_uint16_t<true>; | ||
| 3029 | else | ||
| 3030 | return resize_v_avx2_planar_uint16_t<false>; | ||
| 3031 | } | ||
| 3032 | if (CPU & CPUF_SSE2) { | ||
| 3033 | if (bits_per_pixel < 16) | ||
| 3034 | return resize_v_sse2_planar_uint16_t<true>; | ||
| 3035 | else | ||
| 3036 | return resize_v_sse2_planar_uint16_t<false>; | ||
| 3037 | } | ||
| 3038 | #endif | ||
| 3039 | // C version | ||
| 3040 | if (bits_per_pixel == 16) | ||
| 3041 | return resize_v_c_planar_uint8_16_t_auto_vectorized<uint16_t, false>; | ||
| 3042 | else | ||
| 3043 | return resize_v_c_planar_uint8_16_t_auto_vectorized<uint16_t, true>; | ||
| 3044 | } | ||
| 3045 | else // pixelsize== 4 | ||
| 3046 | { | ||
| 3047 | #ifdef INTEL_INTRINSICS | ||
| 3048 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 3049 | if (CPU & CPUF_AVX512F) { | ||
| 3050 | // return resize_v_avx512_planar_float; | ||
| 3051 | return resize_v_avx512_planar_float_w_sr; | ||
| 3052 | } | ||
| 3053 | #endif | ||
| 3054 | if (CPU & CPUF_AVX2) { | ||
| 3055 | return resize_v_avx2_planar_float; | ||
| 3056 | } | ||
| 3057 | if (CPU & CPUF_SSE2) { | ||
| 3058 | return resize_v_sse2_planar_float; | ||
| 3059 | } | ||
| 3060 | #endif | ||
| 3061 | return resize_v_c_planar_float_auto_vectorized; | ||
| 3062 | } | ||
| 3063 | } | ||
| 3064 | } | ||
| 3065 | |||
| 3066 | ResamplerH FilteredResize_2p::GetResamplerH(int CPU, int pixelsize, int bits_per_pixel, ResamplingProgram* program, IScriptEnvironment* env) // may be somehow call same method from FilteredResizeH class ? | ||
| 3067 | { | ||
| 3068 | int simd_coeff_count_padding = 8; | ||
| 3069 | |||
| 3070 | // Both 8-bit and 16-bit SSSE3 and AVX2 horizontal resizers benefit from processing 16 pixels per cycle. | ||
| 3071 | // Floats also use 32 bytes, but since 32/sizeof(float) = 8, processing 16 pixels is unnecessary. | ||
| 3072 | // Even in C, the code is optimized to be vector-friendly. | ||
| 3073 | if (pixelsize == 1 || pixelsize == 2) | ||
| 3074 | simd_coeff_count_padding = 16; | ||
| 3075 | |||
| 3076 | // Not only does it prepare and pad for SIMD/vector code, but it also corrects, reorders, and equalizes coefficients | ||
| 3077 | // at the right and bottom ends, since we may have variable kernel sizes due to boundary conditions. | ||
| 3078 | resize_prepare_coeffs(program, env, simd_coeff_count_padding); | ||
| 3079 | |||
| 3080 | if (pixelsize == 1) | ||
| 3081 | { | ||
| 3082 | #ifdef INTEL_INTRINSICS | ||
| 3083 | if (CPU & CPUF_AVX2) { | ||
| 3084 | return resizer_h_avx2_generic_uint8_t; | ||
| 3085 | } | ||
| 3086 | if (CPU & CPUF_SSSE3) { | ||
| 3087 | return resizer_h_ssse3_generic_uint8_16<uint8_t, true>; | ||
| 3088 | } | ||
| 3089 | #endif | ||
| 3090 | return resizer_h_c_generic_uint8_16_vectorized<uint8_t, true>; | ||
| 3091 | //return resize_h_c_planar<uint8_t, 1>; | ||
| 3092 | } | ||
| 3093 | else if (pixelsize == 2) { | ||
| 3094 | #ifdef INTEL_INTRINSICS | ||
| 3095 | if (CPU & CPUF_AVX2) { | ||
| 3096 | if (bits_per_pixel < 16) | ||
| 3097 | return resizer_h_avx2_generic_uint16_t<true>; | ||
| 3098 | else | ||
| 3099 | return resizer_h_avx2_generic_uint16_t<false>; | ||
| 3100 | } | ||
| 3101 | if (CPU & CPUF_SSSE3) { | ||
| 3102 | if (bits_per_pixel < 16) | ||
| 3103 | return resizer_h_ssse3_generic_uint8_16<uint16_t, true>; | ||
| 3104 | else | ||
| 3105 | return resizer_h_ssse3_generic_uint8_16<uint16_t, false>; | ||
| 3106 | } | ||
| 3107 | #endif | ||
| 3108 | if (bits_per_pixel == 16) | ||
| 3109 | return resizer_h_c_generic_uint8_16_vectorized<uint16_t, false>; | ||
| 3110 | // return resize_h_c_planar<uint16_t, 0>; | ||
| 3111 | else | ||
| 3112 | return resizer_h_c_generic_uint8_16_vectorized<uint16_t, true>; | ||
| 3113 | // return resize_h_c_planar<uint16_t, 1>; | ||
| 3114 | } | ||
| 3115 | else { //if (pixelsize == 4) | ||
| 3116 | #ifdef INTEL_INTRINSICS | ||
| 3117 | #ifdef INTEL_INTRINSICS_AVX512 | ||
| 3118 | if ((CPU & CPUF_AVX512F) && program->filter_size_real <= 4) { | ||
| 3119 | //return resize_h_planar_float_avx2_permutex_vstripe_ks4; | ||
| 3120 | switch (program->filter_size_real) { | ||
| 3121 | /* case 1: return resize_h_planar_float_avx512_transpose_vstripe_ks4<1>; break; | ||
| 3122 | case 2: return resize_h_planar_float_avx512_transpose_vstripe_ks4<2>; break; | ||
| 3123 | case 3: return resize_h_planar_float_avx512_transpose_vstripe_ks4<3>; break; | ||
| 3124 | case 4: return resize_h_planar_float_avx512_transpose_vstripe_ks4<0>; break;*/ | ||
| 3125 | case 1: return resize_h_planar_float_avx512_gather_permutex_vstripe_ks4_2w<1>; break; | ||
| 3126 | case 2: return resize_h_planar_float_avx512_gather_permutex_vstripe_ks4_2w<2>; break; | ||
| 3127 | case 3: return resize_h_planar_float_avx512_gather_permutex_vstripe_ks4_2w<3>; break; | ||
| 3128 | case 4: return resize_h_planar_float_avx512_gather_permutex_vstripe_ks4_2w<0>; break; | ||
| 3129 | } | ||
| 3130 | } | ||
| 3131 | #endif | ||
| 3132 | if (CPU & CPUF_AVX2) { | ||
| 3133 | //return resize_h_planar_float_avx2_permutex_vstripe_ks4; | ||
| 3134 | |||
| 3135 | switch (program->filter_size_real) { | ||
| 3136 | /* case 1: return resize_h_planar_float_avx_transpose_vstripe_ks4<1>; break; | ||
| 3137 | case 2: return resize_h_planar_float_avx_transpose_vstripe_ks4<2>; break; | ||
| 3138 | case 3: return resize_h_planar_float_avx_transpose_vstripe_ks4<3>; break; | ||
| 3139 | case 4: return resize_h_planar_float_avx_transpose_vstripe_ks4<0>; break;*/ | ||
| 3140 | case 1: return resize_h_planar_float_avx2_gather_permutex_vstripe_ks4<1>; break; | ||
| 3141 | case 2: return resize_h_planar_float_avx2_gather_permutex_vstripe_ks4<2>; break; | ||
| 3142 | case 3: return resize_h_planar_float_avx2_gather_permutex_vstripe_ks4<3>; break; | ||
| 3143 | case 4: return resize_h_planar_float_avx2_gather_permutex_vstripe_ks4<0>; break; | ||
| 3144 | default: return resizer_h_avx2_generic_float; | ||
| 3145 | } | ||
| 3146 | |||
| 3147 | } | ||
| 3148 | if (CPU & CPUF_SSSE3) { | ||
| 3149 | // return resizer_h_ssse3_generic_float; | ||
| 3150 | switch (program->filter_size_real) { | ||
| 3151 | case 1: return resize_h_planar_float_sse_transpose_vstripe_ks4<1>; break; | ||
| 3152 | case 2: return resize_h_planar_float_sse_transpose_vstripe_ks4<2>; break; | ||
| 3153 | case 3: return resize_h_planar_float_sse_transpose_vstripe_ks4<3>; break; | ||
| 3154 | case 4: return resize_h_planar_float_sse_transpose_vstripe_ks4<0>; break; | ||
| 3155 | default: return resizer_h_ssse3_generic_float; | ||
| 3156 | } | ||
| 3157 | } | ||
| 3158 | #endif | ||
| 3159 | return resize_h_c_planar<float, 0>; | ||
| 3160 | } | ||
| 3161 | } | ||
| 3162 | |||
| 3163 | |||
| 3164 | FilteredResize_2p::~FilteredResize_2p(void) | ||
| 3165 | { | ||
| 3166 | if (resampling_program_luma_h) { delete resampling_program_luma_h; } | ||
| 3167 | if (resampling_program_chroma_h) { delete resampling_program_chroma_h; } | ||
| 3168 | |||
| 3169 | if (resampling_program_luma_v) { delete resampling_program_luma_v; } | ||
| 3170 | if (resampling_program_chroma_v) { delete resampling_program_chroma_v; } | ||
| 3171 | |||
| 3172 | |||
| 3173 | } | ||
| 3174 | #endif | ||
| 3175 |