| 1 | // Copyright (C) 2016 The Qt Company Ltd. |
| 2 | // SPDX-License-Identifier: LicenseRef-Qt-Commercial OR LGPL-3.0-only OR GPL-2.0-only OR GPL-3.0-only |
| 3 | |
| 4 | #include "qimagescale_p.h" |
| 5 | #include "qimage.h" |
| 6 | #include <private/qdrawhelper_x86_p.h> |
| 7 | #include <private/qsimd_p.h> |
| 8 | |
| 9 | #if QT_CONFIG(thread) && !defined(Q_OS_WASM) |
| 10 | #include <qsemaphore.h> |
| 11 | #include <qthreadpool.h> |
| 12 | #include <private/qthreadpool_p.h> |
| 13 | #endif |
| 14 | |
| 15 | #if defined(QT_COMPILER_SUPPORTS_SSE4_1) |
| 16 | |
| 17 | QT_BEGIN_NAMESPACE |
| 18 | |
| 19 | using namespace QImageScale; |
| 20 | |
| 21 | template<typename T> |
| 22 | static inline void multithread_pixels_function(QImageScaleInfo *isi, int dh, const T &scaleSection) |
| 23 | { |
| 24 | #if QT_CONFIG(thread) && !defined(Q_OS_WASM) |
| 25 | int segments = (qsizetype(isi->sh) * isi->sw) / (1<<16); |
| 26 | segments = std::min(a: segments, b: dh); |
| 27 | QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance(); |
| 28 | if (segments > 1 && threadPool && !threadPool->contains(thread: QThread::currentThread())) { |
| 29 | QSemaphore semaphore; |
| 30 | int y = 0; |
| 31 | for (int i = 0; i < segments; ++i) { |
| 32 | int yn = (dh - y) / (segments - i); |
| 33 | threadPool->start([&, y, yn]() { |
| 34 | scaleSection(y, y + yn); |
| 35 | semaphore.release(n: 1); |
| 36 | }); |
| 37 | y += yn; |
| 38 | } |
| 39 | semaphore.acquire(n: segments); |
| 40 | return; |
| 41 | } |
| 42 | #endif |
| 43 | scaleSection(0, dh); |
| 44 | } |
| 45 | |
| 46 | inline static __m128i Q_DECL_VECTORCALL |
| 47 | qt_qimageScaleAARGBA_helper(const unsigned int *pix, int xyap, int Cxy, int step, const __m128i vxyap, const __m128i vCxy) |
| 48 | { |
| 49 | __m128i vpix = _mm_cvtepu8_epi32(V: _mm_cvtsi32_si128(a: *pix)); |
| 50 | __m128i vx = _mm_mullo_epi32(V1: vpix, V2: vxyap); |
| 51 | int i; |
| 52 | for (i = (1 << 14) - xyap; i > Cxy; i -= Cxy) { |
| 53 | pix += step; |
| 54 | vpix = _mm_cvtepu8_epi32(V: _mm_cvtsi32_si128(a: *pix)); |
| 55 | vx = _mm_add_epi32(a: vx, b: _mm_mullo_epi32(V1: vpix, V2: vCxy)); |
| 56 | } |
| 57 | pix += step; |
| 58 | vpix = _mm_cvtepu8_epi32(V: _mm_cvtsi32_si128(a: *pix)); |
| 59 | vx = _mm_add_epi32(a: vx, b: _mm_mullo_epi32(V1: vpix, V2: _mm_set1_epi32(i: i))); |
| 60 | return vx; |
| 61 | } |
| 62 | |
| 63 | template<bool RGB> |
| 64 | void qt_qimageScaleAARGBA_up_x_down_y_sse4(QImageScaleInfo *isi, unsigned int *dest, |
| 65 | int dw, int dh, int dow, int sow) |
| 66 | { |
| 67 | const unsigned int **ypoints = isi->ypoints; |
| 68 | const int *xpoints = isi->xpoints; |
| 69 | const int *xapoints = isi->xapoints; |
| 70 | const int *yapoints = isi->yapoints; |
| 71 | |
| 72 | const __m128i v256 = _mm_set1_epi32(i: 256); |
| 73 | |
| 74 | /* go through every scanline in the output buffer */ |
| 75 | auto scaleSection = [&] (int yStart, int yEnd) { |
| 76 | for (int y = yStart; y < yEnd; ++y) { |
| 77 | const int Cy = yapoints[y] >> 16; |
| 78 | const int yap = yapoints[y] & 0xffff; |
| 79 | const __m128i vCy = _mm_set1_epi32(i: Cy); |
| 80 | const __m128i vyap = _mm_set1_epi32(i: yap); |
| 81 | |
| 82 | unsigned int *dptr = dest + (y * dow); |
| 83 | for (int x = 0; x < dw; x++) { |
| 84 | const unsigned int *sptr = ypoints[y] + xpoints[x]; |
| 85 | __m128i vx = qt_qimageScaleAARGBA_helper(pix: sptr, xyap: yap, Cxy: Cy, step: sow, vxyap: vyap, vCxy: vCy); |
| 86 | |
| 87 | const int xap = xapoints[x]; |
| 88 | if (xap > 0) { |
| 89 | const __m128i vxap = _mm_set1_epi32(i: xap); |
| 90 | const __m128i vinvxap = _mm_sub_epi32(a: v256, b: vxap); |
| 91 | __m128i vr = qt_qimageScaleAARGBA_helper(pix: sptr + 1, xyap: yap, Cxy: Cy, step: sow, vxyap: vyap, vCxy: vCy); |
| 92 | |
| 93 | vx = _mm_mullo_epi32(V1: vx, V2: vinvxap); |
| 94 | vr = _mm_mullo_epi32(V1: vr, V2: vxap); |
| 95 | vx = _mm_add_epi32(a: vx, b: vr); |
| 96 | vx = _mm_srli_epi32(a: vx, count: 8); |
| 97 | } |
| 98 | vx = _mm_srli_epi32(a: vx, count: 14); |
| 99 | vx = _mm_packus_epi32(V1: vx, V2: vx); |
| 100 | vx = _mm_packus_epi16(a: vx, b: vx); |
| 101 | *dptr = _mm_cvtsi128_si32(a: vx); |
| 102 | if (RGB) |
| 103 | *dptr |= 0xff000000; |
| 104 | dptr++; |
| 105 | } |
| 106 | } |
| 107 | }; |
| 108 | multithread_pixels_function(isi, dh, scaleSection); |
| 109 | } |
| 110 | |
| 111 | template<bool RGB> |
| 112 | void qt_qimageScaleAARGBA_down_x_up_y_sse4(QImageScaleInfo *isi, unsigned int *dest, |
| 113 | int dw, int dh, int dow, int sow) |
| 114 | { |
| 115 | const unsigned int **ypoints = isi->ypoints; |
| 116 | int *xpoints = isi->xpoints; |
| 117 | int *xapoints = isi->xapoints; |
| 118 | int *yapoints = isi->yapoints; |
| 119 | |
| 120 | const __m128i v256 = _mm_set1_epi32(i: 256); |
| 121 | |
| 122 | /* go through every scanline in the output buffer */ |
| 123 | auto scaleSection = [&] (int yStart, int yEnd) { |
| 124 | for (int y = yStart; y < yEnd; ++y) { |
| 125 | unsigned int *dptr = dest + (y * dow); |
| 126 | for (int x = 0; x < dw; x++) { |
| 127 | int Cx = xapoints[x] >> 16; |
| 128 | int xap = xapoints[x] & 0xffff; |
| 129 | const __m128i vCx = _mm_set1_epi32(i: Cx); |
| 130 | const __m128i vxap = _mm_set1_epi32(i: xap); |
| 131 | |
| 132 | const unsigned int *sptr = ypoints[y] + xpoints[x]; |
| 133 | __m128i vx = qt_qimageScaleAARGBA_helper(pix: sptr, xyap: xap, Cxy: Cx, step: 1, vxyap: vxap, vCxy: vCx); |
| 134 | |
| 135 | int yap = yapoints[y]; |
| 136 | if (yap > 0) { |
| 137 | const __m128i vyap = _mm_set1_epi32(i: yap); |
| 138 | const __m128i vinvyap = _mm_sub_epi32(a: v256, b: vyap); |
| 139 | __m128i vr = qt_qimageScaleAARGBA_helper(pix: sptr + sow, xyap: xap, Cxy: Cx, step: 1, vxyap: vxap, vCxy: vCx); |
| 140 | |
| 141 | vx = _mm_mullo_epi32(V1: vx, V2: vinvyap); |
| 142 | vr = _mm_mullo_epi32(V1: vr, V2: vyap); |
| 143 | vx = _mm_add_epi32(a: vx, b: vr); |
| 144 | vx = _mm_srli_epi32(a: vx, count: 8); |
| 145 | } |
| 146 | vx = _mm_srli_epi32(a: vx, count: 14); |
| 147 | vx = _mm_packus_epi32(V1: vx, V2: vx); |
| 148 | vx = _mm_packus_epi16(a: vx, b: vx); |
| 149 | *dptr = _mm_cvtsi128_si32(a: vx); |
| 150 | if (RGB) |
| 151 | *dptr |= 0xff000000; |
| 152 | dptr++; |
| 153 | } |
| 154 | } |
| 155 | }; |
| 156 | multithread_pixels_function(isi, dh, scaleSection); |
| 157 | } |
| 158 | |
| 159 | template<bool RGB> |
| 160 | void qt_qimageScaleAARGBA_down_xy_sse4(QImageScaleInfo *isi, unsigned int *dest, |
| 161 | int dw, int dh, int dow, int sow) |
| 162 | { |
| 163 | const unsigned int **ypoints = isi->ypoints; |
| 164 | int *xpoints = isi->xpoints; |
| 165 | int *xapoints = isi->xapoints; |
| 166 | int *yapoints = isi->yapoints; |
| 167 | |
| 168 | auto scaleSection = [&] (int yStart, int yEnd) { |
| 169 | for (int y = yStart; y < yEnd; ++y) { |
| 170 | int Cy = yapoints[y] >> 16; |
| 171 | int yap = yapoints[y] & 0xffff; |
| 172 | const __m128i vCy = _mm_set1_epi32(i: Cy); |
| 173 | const __m128i vyap = _mm_set1_epi32(i: yap); |
| 174 | |
| 175 | unsigned int *dptr = dest + (y * dow); |
| 176 | for (int x = 0; x < dw; x++) { |
| 177 | const int Cx = xapoints[x] >> 16; |
| 178 | const int xap = xapoints[x] & 0xffff; |
| 179 | const __m128i vCx = _mm_set1_epi32(i: Cx); |
| 180 | const __m128i vxap = _mm_set1_epi32(i: xap); |
| 181 | |
| 182 | const unsigned int *sptr = ypoints[y] + xpoints[x]; |
| 183 | __m128i vx = qt_qimageScaleAARGBA_helper(pix: sptr, xyap: xap, Cxy: Cx, step: 1, vxyap: vxap, vCxy: vCx); |
| 184 | __m128i vr = _mm_mullo_epi32(V1: _mm_srli_epi32(a: vx, count: 4), V2: vyap); |
| 185 | |
| 186 | int j; |
| 187 | for (j = (1 << 14) - yap; j > Cy; j -= Cy) { |
| 188 | sptr += sow; |
| 189 | vx = qt_qimageScaleAARGBA_helper(pix: sptr, xyap: xap, Cxy: Cx, step: 1, vxyap: vxap, vCxy: vCx); |
| 190 | vr = _mm_add_epi32(a: vr, b: _mm_mullo_epi32(V1: _mm_srli_epi32(a: vx, count: 4), V2: vCy)); |
| 191 | } |
| 192 | sptr += sow; |
| 193 | vx = qt_qimageScaleAARGBA_helper(pix: sptr, xyap: xap, Cxy: Cx, step: 1, vxyap: vxap, vCxy: vCx); |
| 194 | vr = _mm_add_epi32(a: vr, b: _mm_mullo_epi32(V1: _mm_srli_epi32(a: vx, count: 4), V2: _mm_set1_epi32(i: j))); |
| 195 | |
| 196 | vr = _mm_srli_epi32(a: vr, count: 24); |
| 197 | vr = _mm_packus_epi32(V1: vr, V2: _mm_setzero_si128()); |
| 198 | vr = _mm_packus_epi16(a: vr, b: _mm_setzero_si128()); |
| 199 | *dptr = _mm_cvtsi128_si32(a: vr); |
| 200 | if (RGB) |
| 201 | *dptr |= 0xff000000; |
| 202 | dptr++; |
| 203 | } |
| 204 | } |
| 205 | }; |
| 206 | multithread_pixels_function(isi, dh, scaleSection); |
| 207 | } |
| 208 | |
| 209 | template void qt_qimageScaleAARGBA_up_x_down_y_sse4<false>(QImageScaleInfo *isi, unsigned int *dest, |
| 210 | int dw, int dh, int dow, int sow); |
| 211 | |
| 212 | template void qt_qimageScaleAARGBA_up_x_down_y_sse4<true>(QImageScaleInfo *isi, unsigned int *dest, |
| 213 | int dw, int dh, int dow, int sow); |
| 214 | |
| 215 | template void qt_qimageScaleAARGBA_down_x_up_y_sse4<false>(QImageScaleInfo *isi, unsigned int *dest, |
| 216 | int dw, int dh, int dow, int sow); |
| 217 | |
| 218 | template void qt_qimageScaleAARGBA_down_x_up_y_sse4<true>(QImageScaleInfo *isi, unsigned int *dest, |
| 219 | int dw, int dh, int dow, int sow); |
| 220 | |
| 221 | template void qt_qimageScaleAARGBA_down_xy_sse4<false>(QImageScaleInfo *isi, unsigned int *dest, |
| 222 | int dw, int dh, int dow, int sow); |
| 223 | |
| 224 | template void qt_qimageScaleAARGBA_down_xy_sse4<true>(QImageScaleInfo *isi, unsigned int *dest, |
| 225 | int dw, int dh, int dow, int sow); |
| 226 | |
| 227 | QT_END_NAMESPACE |
| 228 | |
| 229 | #endif |
| 230 | |