1 /* 2 * Copyright 2013 Google Inc. 3 * 4 * Use of this source code is governed by a BSD-style license that can be 5 * found in the LICENSE file. 6 */ 7 8 #include <emmintrin.h> 9 #include "SkBitmap.h" 10 #include "SkBitmapFilter_opts_SSE2.h" 11 #include "SkBitmapProcState.h" 12 #include "SkColor.h" 13 #include "SkColorPriv.h" 14 #include "SkConvolver.h" 15 #include "SkShader.h" 16 #include "SkUnPreMultiply.h" 17 18 #if 0 19 static inline void print128i(__m128i value) { 20 int *v = (int*) &value; 21 printf("% .11d % .11d % .11d % .11d\n", v[0], v[1], v[2], v[3]); 22 } 23 24 static inline void print128i_16(__m128i value) { 25 short *v = (short*) &value; 26 printf("% .5d % .5d % .5d % .5d % .5d % .5d % .5d % .5d\n", v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]); 27 } 28 29 static inline void print128i_8(__m128i value) { 30 unsigned char *v = (unsigned char*) &value; 31 printf("%.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u\n", 32 v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], 33 v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15] 34 ); 35 } 36 37 static inline void print128f(__m128 value) { 38 float *f = (float*) &value; 39 printf("%3.4f %3.4f %3.4f %3.4f\n", f[0], f[1], f[2], f[3]); 40 } 41 #endif 42 43 // because the border is handled specially, this is guaranteed to have all 16 pixels 44 // available to it without running off the bitmap's edge. 45 46 void highQualityFilter_SSE2(const SkBitmapProcState& s, int x, int y, 47 SkPMColor* SK_RESTRICT colors, int count) { 48 49 const int maxX = s.fBitmap->width(); 50 const int maxY = s.fBitmap->height(); 51 SkAutoTMalloc<SkScalar> xWeights(maxX); 52 53 while (count-- > 0) { 54 SkPoint srcPt; 55 s.fInvProc(s.fInvMatrix, x + 0.5f, y + 0.5f, &srcPt); 56 srcPt.fX -= SK_ScalarHalf; 57 srcPt.fY -= SK_ScalarHalf; 58 59 __m128 weight = _mm_setzero_ps(); 60 __m128 accum = _mm_setzero_ps(); 61 62 int y0 = SkClampMax(SkScalarCeilToInt(srcPt.fY-s.getBitmapFilter()->width()), maxY); 63 int y1 = SkClampMax(SkScalarFloorToInt(srcPt.fY+s.getBitmapFilter()->width()+1), maxY); 64 int x0 = SkClampMax(SkScalarCeilToInt(srcPt.fX-s.getBitmapFilter()->width()), maxX); 65 int x1 = SkClampMax(SkScalarFloorToInt(srcPt.fX+s.getBitmapFilter()->width())+1, maxX); 66 67 for (int srcX = x0; srcX < x1 ; srcX++) { 68 // Looking these up once instead of each loop is a ~15% speedup. 69 xWeights[srcX - x0] = s.getBitmapFilter()->lookupScalar((srcPt.fX - srcX)); 70 } 71 72 for (int srcY = y0; srcY < y1; srcY++) { 73 SkScalar yWeight = s.getBitmapFilter()->lookupScalar((srcPt.fY - srcY)); 74 75 for (int srcX = x0; srcX < x1 ; srcX++) { 76 SkScalar xWeight = xWeights[srcX - x0]; 77 78 SkScalar combined_weight = SkScalarMul(xWeight, yWeight); 79 80 SkPMColor color = *s.fBitmap->getAddr32(srcX, srcY); 81 82 __m128i c = _mm_cvtsi32_si128(color); 83 c = _mm_unpacklo_epi8(c, _mm_setzero_si128()); 84 c = _mm_unpacklo_epi16(c, _mm_setzero_si128()); 85 __m128 cfloat = _mm_cvtepi32_ps(c); 86 87 __m128 weightVector = _mm_set1_ps(combined_weight); 88 accum = _mm_add_ps(accum, _mm_mul_ps(cfloat, weightVector)); 89 weight = _mm_add_ps( weight, weightVector ); 90 } 91 } 92 93 accum = _mm_div_ps(accum, weight); 94 accum = _mm_add_ps(accum, _mm_set1_ps(0.5f)); 95 __m128i accumInt = _mm_cvttps_epi32(accum); 96 accumInt = _mm_packs_epi32(accumInt, _mm_setzero_si128()); 97 accumInt = _mm_packus_epi16(accumInt, _mm_setzero_si128()); 98 SkPMColor c = _mm_cvtsi128_si32(accumInt); 99 100 int a = SkClampMax(SkGetPackedA32(c), 255); 101 int r = SkClampMax(SkGetPackedR32(c), a); 102 int g = SkClampMax(SkGetPackedG32(c), a); 103 int b = SkClampMax(SkGetPackedB32(c), a); 104 105 *colors++ = SkPackARGB32(a, r, g, b); 106 107 x++; 108 } 109 } 110 111 // Convolves horizontally along a single row. The row data is given in 112 // |src_data| and continues for the num_values() of the filter. 113 void convolveHorizontally_SSE2(const unsigned char* src_data, 114 const SkConvolutionFilter1D& filter, 115 unsigned char* out_row, 116 bool /*has_alpha*/) { 117 int num_values = filter.numValues(); 118 119 int filter_offset, filter_length; 120 __m128i zero = _mm_setzero_si128(); 121 __m128i mask[4]; 122 // |mask| will be used to decimate all extra filter coefficients that are 123 // loaded by SIMD when |filter_length| is not divisible by 4. 124 // mask[0] is not used in following algorithm. 125 mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1); 126 mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1); 127 mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1); 128 129 // Output one pixel each iteration, calculating all channels (RGBA) together. 130 for (int out_x = 0; out_x < num_values; out_x++) { 131 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = 132 filter.FilterForValue(out_x, &filter_offset, &filter_length); 133 134 __m128i accum = _mm_setzero_si128(); 135 136 // Compute the first pixel in this row that the filter affects. It will 137 // touch |filter_length| pixels (4 bytes each) after this. 138 const __m128i* row_to_filter = 139 reinterpret_cast<const __m128i*>(&src_data[filter_offset << 2]); 140 141 // We will load and accumulate with four coefficients per iteration. 142 for (int filter_x = 0; filter_x < filter_length >> 2; filter_x++) { 143 144 // Load 4 coefficients => duplicate 1st and 2nd of them for all channels. 145 __m128i coeff, coeff16; 146 // [16] xx xx xx xx c3 c2 c1 c0 147 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 148 // [16] xx xx xx xx c1 c1 c0 c0 149 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 150 // [16] c1 c1 c1 c1 c0 c0 c0 c0 151 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 152 153 // Load four pixels => unpack the first two pixels to 16 bits => 154 // multiply with coefficients => accumulate the convolution result. 155 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 156 __m128i src8 = _mm_loadu_si128(row_to_filter); 157 // [16] a1 b1 g1 r1 a0 b0 g0 r0 158 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 159 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 160 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 161 // [32] a0*c0 b0*c0 g0*c0 r0*c0 162 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 163 accum = _mm_add_epi32(accum, t); 164 // [32] a1*c1 b1*c1 g1*c1 r1*c1 165 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 166 accum = _mm_add_epi32(accum, t); 167 168 // Duplicate 3rd and 4th coefficients for all channels => 169 // unpack the 3rd and 4th pixels to 16 bits => multiply with coefficients 170 // => accumulate the convolution results. 171 // [16] xx xx xx xx c3 c3 c2 c2 172 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 173 // [16] c3 c3 c3 c3 c2 c2 c2 c2 174 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 175 // [16] a3 g3 b3 r3 a2 g2 b2 r2 176 src16 = _mm_unpackhi_epi8(src8, zero); 177 mul_hi = _mm_mulhi_epi16(src16, coeff16); 178 mul_lo = _mm_mullo_epi16(src16, coeff16); 179 // [32] a2*c2 b2*c2 g2*c2 r2*c2 180 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 181 accum = _mm_add_epi32(accum, t); 182 // [32] a3*c3 b3*c3 g3*c3 r3*c3 183 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 184 accum = _mm_add_epi32(accum, t); 185 186 // Advance the pixel and coefficients pointers. 187 row_to_filter += 1; 188 filter_values += 4; 189 } 190 191 // When |filter_length| is not divisible by 4, we need to decimate some of 192 // the filter coefficient that was loaded incorrectly to zero; Other than 193 // that the algorithm is same with above, exceot that the 4th pixel will be 194 // always absent. 195 int r = filter_length&3; 196 if (r) { 197 // Note: filter_values must be padded to align_up(filter_offset, 8). 198 __m128i coeff, coeff16; 199 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 200 // Mask out extra filter taps. 201 coeff = _mm_and_si128(coeff, mask[r]); 202 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 203 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 204 205 // Note: line buffer must be padded to align_up(filter_offset, 16). 206 // We resolve this by use C-version for the last horizontal line. 207 __m128i src8 = _mm_loadu_si128(row_to_filter); 208 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 209 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 210 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 211 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 212 accum = _mm_add_epi32(accum, t); 213 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 214 accum = _mm_add_epi32(accum, t); 215 216 src16 = _mm_unpackhi_epi8(src8, zero); 217 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 218 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 219 mul_hi = _mm_mulhi_epi16(src16, coeff16); 220 mul_lo = _mm_mullo_epi16(src16, coeff16); 221 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 222 accum = _mm_add_epi32(accum, t); 223 } 224 225 // Shift right for fixed point implementation. 226 accum = _mm_srai_epi32(accum, SkConvolutionFilter1D::kShiftBits); 227 228 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). 229 accum = _mm_packs_epi32(accum, zero); 230 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). 231 accum = _mm_packus_epi16(accum, zero); 232 233 // Store the pixel value of 32 bits. 234 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum); 235 out_row += 4; 236 } 237 } 238 239 // Convolves horizontally along four rows. The row data is given in 240 // |src_data| and continues for the num_values() of the filter. 241 // The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please 242 // refer to that function for detailed comments. 243 void convolve4RowsHorizontally_SSE2(const unsigned char* src_data[4], 244 const SkConvolutionFilter1D& filter, 245 unsigned char* out_row[4]) { 246 int num_values = filter.numValues(); 247 248 int filter_offset, filter_length; 249 __m128i zero = _mm_setzero_si128(); 250 __m128i mask[4]; 251 // |mask| will be used to decimate all extra filter coefficients that are 252 // loaded by SIMD when |filter_length| is not divisible by 4. 253 // mask[0] is not used in following algorithm. 254 mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1); 255 mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1); 256 mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1); 257 258 // Output one pixel each iteration, calculating all channels (RGBA) together. 259 for (int out_x = 0; out_x < num_values; out_x++) { 260 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = 261 filter.FilterForValue(out_x, &filter_offset, &filter_length); 262 263 // four pixels in a column per iteration. 264 __m128i accum0 = _mm_setzero_si128(); 265 __m128i accum1 = _mm_setzero_si128(); 266 __m128i accum2 = _mm_setzero_si128(); 267 __m128i accum3 = _mm_setzero_si128(); 268 int start = (filter_offset<<2); 269 // We will load and accumulate with four coefficients per iteration. 270 for (int filter_x = 0; filter_x < (filter_length >> 2); filter_x++) { 271 __m128i coeff, coeff16lo, coeff16hi; 272 // [16] xx xx xx xx c3 c2 c1 c0 273 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 274 // [16] xx xx xx xx c1 c1 c0 c0 275 coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 276 // [16] c1 c1 c1 c1 c0 c0 c0 c0 277 coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo); 278 // [16] xx xx xx xx c3 c3 c2 c2 279 coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 280 // [16] c3 c3 c3 c3 c2 c2 c2 c2 281 coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi); 282 283 __m128i src8, src16, mul_hi, mul_lo, t; 284 285 #define ITERATION(src, accum) \ 286 src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \ 287 src16 = _mm_unpacklo_epi8(src8, zero); \ 288 mul_hi = _mm_mulhi_epi16(src16, coeff16lo); \ 289 mul_lo = _mm_mullo_epi16(src16, coeff16lo); \ 290 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ 291 accum = _mm_add_epi32(accum, t); \ 292 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ 293 accum = _mm_add_epi32(accum, t); \ 294 src16 = _mm_unpackhi_epi8(src8, zero); \ 295 mul_hi = _mm_mulhi_epi16(src16, coeff16hi); \ 296 mul_lo = _mm_mullo_epi16(src16, coeff16hi); \ 297 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ 298 accum = _mm_add_epi32(accum, t); \ 299 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ 300 accum = _mm_add_epi32(accum, t) 301 302 ITERATION(src_data[0] + start, accum0); 303 ITERATION(src_data[1] + start, accum1); 304 ITERATION(src_data[2] + start, accum2); 305 ITERATION(src_data[3] + start, accum3); 306 307 start += 16; 308 filter_values += 4; 309 } 310 311 int r = filter_length & 3; 312 if (r) { 313 // Note: filter_values must be padded to align_up(filter_offset, 8); 314 __m128i coeff; 315 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 316 // Mask out extra filter taps. 317 coeff = _mm_and_si128(coeff, mask[r]); 318 319 __m128i coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 320 /* c1 c1 c1 c1 c0 c0 c0 c0 */ 321 coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo); 322 __m128i coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 323 coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi); 324 325 __m128i src8, src16, mul_hi, mul_lo, t; 326 327 ITERATION(src_data[0] + start, accum0); 328 ITERATION(src_data[1] + start, accum1); 329 ITERATION(src_data[2] + start, accum2); 330 ITERATION(src_data[3] + start, accum3); 331 } 332 333 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 334 accum0 = _mm_packs_epi32(accum0, zero); 335 accum0 = _mm_packus_epi16(accum0, zero); 336 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 337 accum1 = _mm_packs_epi32(accum1, zero); 338 accum1 = _mm_packus_epi16(accum1, zero); 339 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 340 accum2 = _mm_packs_epi32(accum2, zero); 341 accum2 = _mm_packus_epi16(accum2, zero); 342 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); 343 accum3 = _mm_packs_epi32(accum3, zero); 344 accum3 = _mm_packus_epi16(accum3, zero); 345 346 *(reinterpret_cast<int*>(out_row[0])) = _mm_cvtsi128_si32(accum0); 347 *(reinterpret_cast<int*>(out_row[1])) = _mm_cvtsi128_si32(accum1); 348 *(reinterpret_cast<int*>(out_row[2])) = _mm_cvtsi128_si32(accum2); 349 *(reinterpret_cast<int*>(out_row[3])) = _mm_cvtsi128_si32(accum3); 350 351 out_row[0] += 4; 352 out_row[1] += 4; 353 out_row[2] += 4; 354 out_row[3] += 4; 355 } 356 } 357 358 // Does vertical convolution to produce one output row. The filter values and 359 // length are given in the first two parameters. These are applied to each 360 // of the rows pointed to in the |source_data_rows| array, with each row 361 // being |pixel_width| wide. 362 // 363 // The output must have room for |pixel_width * 4| bytes. 364 template<bool has_alpha> 365 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filter_values, 366 int filter_length, 367 unsigned char* const* source_data_rows, 368 int pixel_width, 369 unsigned char* out_row) { 370 int width = pixel_width & ~3; 371 372 __m128i zero = _mm_setzero_si128(); 373 __m128i accum0, accum1, accum2, accum3, coeff16; 374 const __m128i* src; 375 // Output four pixels per iteration (16 bytes). 376 for (int out_x = 0; out_x < width; out_x += 4) { 377 378 // Accumulated result for each pixel. 32 bits per RGBA channel. 379 accum0 = _mm_setzero_si128(); 380 accum1 = _mm_setzero_si128(); 381 accum2 = _mm_setzero_si128(); 382 accum3 = _mm_setzero_si128(); 383 384 // Convolve with one filter coefficient per iteration. 385 for (int filter_y = 0; filter_y < filter_length; filter_y++) { 386 387 // Duplicate the filter coefficient 8 times. 388 // [16] cj cj cj cj cj cj cj cj 389 coeff16 = _mm_set1_epi16(filter_values[filter_y]); 390 391 // Load four pixels (16 bytes) together. 392 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 393 src = reinterpret_cast<const __m128i*>( 394 &source_data_rows[filter_y][out_x << 2]); 395 __m128i src8 = _mm_loadu_si128(src); 396 397 // Unpack 1st and 2nd pixels from 8 bits to 16 bits for each channels => 398 // multiply with current coefficient => accumulate the result. 399 // [16] a1 b1 g1 r1 a0 b0 g0 r0 400 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 401 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 402 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 403 // [32] a0 b0 g0 r0 404 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 405 accum0 = _mm_add_epi32(accum0, t); 406 // [32] a1 b1 g1 r1 407 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 408 accum1 = _mm_add_epi32(accum1, t); 409 410 // Unpack 3rd and 4th pixels from 8 bits to 16 bits for each channels => 411 // multiply with current coefficient => accumulate the result. 412 // [16] a3 b3 g3 r3 a2 b2 g2 r2 413 src16 = _mm_unpackhi_epi8(src8, zero); 414 mul_hi = _mm_mulhi_epi16(src16, coeff16); 415 mul_lo = _mm_mullo_epi16(src16, coeff16); 416 // [32] a2 b2 g2 r2 417 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 418 accum2 = _mm_add_epi32(accum2, t); 419 // [32] a3 b3 g3 r3 420 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 421 accum3 = _mm_add_epi32(accum3, t); 422 } 423 424 // Shift right for fixed point implementation. 425 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 426 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 427 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 428 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); 429 430 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). 431 // [16] a1 b1 g1 r1 a0 b0 g0 r0 432 accum0 = _mm_packs_epi32(accum0, accum1); 433 // [16] a3 b3 g3 r3 a2 b2 g2 r2 434 accum2 = _mm_packs_epi32(accum2, accum3); 435 436 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). 437 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 438 accum0 = _mm_packus_epi16(accum0, accum2); 439 440 if (has_alpha) { 441 // Compute the max(ri, gi, bi) for each pixel. 442 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 443 __m128i a = _mm_srli_epi32(accum0, 8); 444 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 445 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. 446 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 447 a = _mm_srli_epi32(accum0, 16); 448 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 449 b = _mm_max_epu8(a, b); // Max of r and g and b. 450 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 451 b = _mm_slli_epi32(b, 24); 452 453 // Make sure the value of alpha channel is always larger than maximum 454 // value of color channels. 455 accum0 = _mm_max_epu8(b, accum0); 456 } else { 457 // Set value of alpha channels to 0xFF. 458 __m128i mask = _mm_set1_epi32(0xff000000); 459 accum0 = _mm_or_si128(accum0, mask); 460 } 461 462 // Store the convolution result (16 bytes) and advance the pixel pointers. 463 _mm_storeu_si128(reinterpret_cast<__m128i*>(out_row), accum0); 464 out_row += 16; 465 } 466 467 // When the width of the output is not divisible by 4, We need to save one 468 // pixel (4 bytes) each time. And also the fourth pixel is always absent. 469 if (pixel_width & 3) { 470 accum0 = _mm_setzero_si128(); 471 accum1 = _mm_setzero_si128(); 472 accum2 = _mm_setzero_si128(); 473 for (int filter_y = 0; filter_y < filter_length; ++filter_y) { 474 coeff16 = _mm_set1_epi16(filter_values[filter_y]); 475 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 476 src = reinterpret_cast<const __m128i*>( 477 &source_data_rows[filter_y][width<<2]); 478 __m128i src8 = _mm_loadu_si128(src); 479 // [16] a1 b1 g1 r1 a0 b0 g0 r0 480 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 481 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 482 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 483 // [32] a0 b0 g0 r0 484 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 485 accum0 = _mm_add_epi32(accum0, t); 486 // [32] a1 b1 g1 r1 487 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 488 accum1 = _mm_add_epi32(accum1, t); 489 // [16] a3 b3 g3 r3 a2 b2 g2 r2 490 src16 = _mm_unpackhi_epi8(src8, zero); 491 mul_hi = _mm_mulhi_epi16(src16, coeff16); 492 mul_lo = _mm_mullo_epi16(src16, coeff16); 493 // [32] a2 b2 g2 r2 494 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 495 accum2 = _mm_add_epi32(accum2, t); 496 } 497 498 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 499 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 500 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 501 // [16] a1 b1 g1 r1 a0 b0 g0 r0 502 accum0 = _mm_packs_epi32(accum0, accum1); 503 // [16] a3 b3 g3 r3 a2 b2 g2 r2 504 accum2 = _mm_packs_epi32(accum2, zero); 505 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 506 accum0 = _mm_packus_epi16(accum0, accum2); 507 if (has_alpha) { 508 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 509 __m128i a = _mm_srli_epi32(accum0, 8); 510 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 511 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. 512 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 513 a = _mm_srli_epi32(accum0, 16); 514 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 515 b = _mm_max_epu8(a, b); // Max of r and g and b. 516 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 517 b = _mm_slli_epi32(b, 24); 518 accum0 = _mm_max_epu8(b, accum0); 519 } else { 520 __m128i mask = _mm_set1_epi32(0xff000000); 521 accum0 = _mm_or_si128(accum0, mask); 522 } 523 524 for (int out_x = width; out_x < pixel_width; out_x++) { 525 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum0); 526 accum0 = _mm_srli_si128(accum0, 4); 527 out_row += 4; 528 } 529 } 530 } 531 532 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filter_values, 533 int filter_length, 534 unsigned char* const* source_data_rows, 535 int pixel_width, 536 unsigned char* out_row, 537 bool has_alpha) { 538 if (has_alpha) { 539 convolveVertically_SSE2<true>(filter_values, 540 filter_length, 541 source_data_rows, 542 pixel_width, 543 out_row); 544 } else { 545 convolveVertically_SSE2<false>(filter_values, 546 filter_length, 547 source_data_rows, 548 pixel_width, 549 out_row); 550 } 551 } 552 553 void applySIMDPadding_SSE2(SkConvolutionFilter1D *filter) { 554 // Padding |paddingCount| of more dummy coefficients after the coefficients 555 // of last filter to prevent SIMD instructions which load 8 or 16 bytes 556 // together to access invalid memory areas. We are not trying to align the 557 // coefficients right now due to the opaqueness of <vector> implementation. 558 // This has to be done after all |AddFilter| calls. 559 for (int i = 0; i < 8; ++i) { 560 filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0)); 561 } 562 } 563
