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() - 1; 50 const int maxY = s.fBitmap->height() - 1; 51 52 while (count-- > 0) { 53 SkPoint srcPt; 54 s.fInvProc(s.fInvMatrix, SkIntToScalar(x), 55 SkIntToScalar(y), &srcPt); 56 srcPt.fX -= SK_ScalarHalf; 57 srcPt.fY -= SK_ScalarHalf; 58 59 int sx = SkScalarFloorToInt(srcPt.fX); 60 int sy = SkScalarFloorToInt(srcPt.fY); 61 62 __m128 weight = _mm_setzero_ps(); 63 __m128 accum = _mm_setzero_ps(); 64 65 int y0 = SkTMax(0, int(ceil(sy-s.getBitmapFilter()->width() + 0.5f))); 66 int y1 = SkTMin(maxY, int(floor(sy+s.getBitmapFilter()->width() + 0.5f))); 67 int x0 = SkTMax(0, int(ceil(sx-s.getBitmapFilter()->width() + 0.5f))); 68 int x1 = SkTMin(maxX, int(floor(sx+s.getBitmapFilter()->width() + 0.5f))); 69 70 for (int src_y = y0; src_y <= y1; src_y++) { 71 float yweight = SkScalarToFloat(s.getBitmapFilter()->lookupScalar(srcPt.fY - src_y)); 72 73 for (int src_x = x0; src_x <= x1 ; src_x++) { 74 float xweight = SkScalarToFloat(s.getBitmapFilter()->lookupScalar(srcPt.fX - src_x)); 75 76 float combined_weight = xweight * yweight; 77 78 SkPMColor color = *s.fBitmap->getAddr32(src_x, src_y); 79 80 __m128i c = _mm_cvtsi32_si128( color ); 81 c = _mm_unpacklo_epi8(c, _mm_setzero_si128()); 82 c = _mm_unpacklo_epi16(c, _mm_setzero_si128()); 83 84 __m128 cfloat = _mm_cvtepi32_ps( c ); 85 86 __m128 weightVector = _mm_set1_ps(combined_weight); 87 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 96 __m128i accumInt = _mm_cvtps_epi32( accum ); 97 98 int localResult[4]; 99 _mm_storeu_si128((__m128i *) (localResult), accumInt); 100 int a = SkClampMax(localResult[0], 255); 101 int r = SkClampMax(localResult[1], a); 102 int g = SkClampMax(localResult[2], a); 103 int b = SkClampMax(localResult[3], a); 104 105 *colors++ = SkPackARGB32(a, r, g, b); 106 107 x++; 108 } 109 } 110 111 void highQualityFilter_ScaleOnly_SSE2(const SkBitmapProcState &s, int x, int y, 112 SkPMColor *SK_RESTRICT colors, int count) { 113 const int maxX = s.fBitmap->width() - 1; 114 const int maxY = s.fBitmap->height() - 1; 115 116 SkPoint srcPt; 117 s.fInvProc(s.fInvMatrix, SkIntToScalar(x), 118 SkIntToScalar(y), &srcPt); 119 srcPt.fY -= SK_ScalarHalf; 120 int sy = SkScalarFloorToInt(srcPt.fY); 121 122 int y0 = SkTMax(0, int(ceil(sy-s.getBitmapFilter()->width() + 0.5f))); 123 int y1 = SkTMin(maxY, int(floor(sy+s.getBitmapFilter()->width() + 0.5f))); 124 125 while (count-- > 0) { 126 srcPt.fX -= SK_ScalarHalf; 127 srcPt.fY -= SK_ScalarHalf; 128 129 int sx = SkScalarFloorToInt(srcPt.fX); 130 131 float weight = 0; 132 __m128 accum = _mm_setzero_ps(); 133 134 int x0 = SkTMax(0, int(ceil(sx-s.getBitmapFilter()->width() + 0.5f))); 135 int x1 = SkTMin(maxX, int(floor(sx+s.getBitmapFilter()->width() + 0.5f))); 136 137 for (int src_y = y0; src_y <= y1; src_y++) { 138 float yweight = SkScalarToFloat(s.getBitmapFilter()->lookupScalar(srcPt.fY - src_y)); 139 140 for (int src_x = x0; src_x <= x1 ; src_x++) { 141 float xweight = SkScalarToFloat(s.getBitmapFilter()->lookupScalar(srcPt.fX - src_x)); 142 143 float combined_weight = xweight * yweight; 144 145 SkPMColor color = *s.fBitmap->getAddr32(src_x, src_y); 146 147 __m128 c = _mm_set_ps((float)SkGetPackedB32(color), 148 (float)SkGetPackedG32(color), 149 (float)SkGetPackedR32(color), 150 (float)SkGetPackedA32(color)); 151 152 __m128 weightVector = _mm_set1_ps(combined_weight); 153 154 accum = _mm_add_ps(accum, _mm_mul_ps(c, weightVector)); 155 weight += combined_weight; 156 } 157 } 158 159 __m128 totalWeightVector = _mm_set1_ps(weight); 160 accum = _mm_div_ps(accum, totalWeightVector); 161 accum = _mm_add_ps(accum, _mm_set1_ps(0.5f)); 162 163 float localResult[4]; 164 _mm_storeu_ps(localResult, accum); 165 int a = SkClampMax(int(localResult[0]), 255); 166 int r = SkClampMax(int(localResult[1]), a); 167 int g = SkClampMax(int(localResult[2]), a); 168 int b = SkClampMax(int(localResult[3]), a); 169 170 *colors++ = SkPackARGB32(a, r, g, b); 171 172 x++; 173 174 s.fInvProc(s.fInvMatrix, SkIntToScalar(x), 175 SkIntToScalar(y), &srcPt); 176 } 177 } 178 179 // Convolves horizontally along a single row. The row data is given in 180 // |src_data| and continues for the num_values() of the filter. 181 void convolveHorizontally_SSE2(const unsigned char* src_data, 182 const SkConvolutionFilter1D& filter, 183 unsigned char* out_row, 184 bool /*has_alpha*/) { 185 int num_values = filter.numValues(); 186 187 int filter_offset, filter_length; 188 __m128i zero = _mm_setzero_si128(); 189 __m128i mask[4]; 190 // |mask| will be used to decimate all extra filter coefficients that are 191 // loaded by SIMD when |filter_length| is not divisible by 4. 192 // mask[0] is not used in following algorithm. 193 mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1); 194 mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1); 195 mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1); 196 197 // Output one pixel each iteration, calculating all channels (RGBA) together. 198 for (int out_x = 0; out_x < num_values; out_x++) { 199 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = 200 filter.FilterForValue(out_x, &filter_offset, &filter_length); 201 202 __m128i accum = _mm_setzero_si128(); 203 204 // Compute the first pixel in this row that the filter affects. It will 205 // touch |filter_length| pixels (4 bytes each) after this. 206 const __m128i* row_to_filter = 207 reinterpret_cast<const __m128i*>(&src_data[filter_offset << 2]); 208 209 // We will load and accumulate with four coefficients per iteration. 210 for (int filter_x = 0; filter_x < filter_length >> 2; filter_x++) { 211 212 // Load 4 coefficients => duplicate 1st and 2nd of them for all channels. 213 __m128i coeff, coeff16; 214 // [16] xx xx xx xx c3 c2 c1 c0 215 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 216 // [16] xx xx xx xx c1 c1 c0 c0 217 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 218 // [16] c1 c1 c1 c1 c0 c0 c0 c0 219 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 220 221 // Load four pixels => unpack the first two pixels to 16 bits => 222 // multiply with coefficients => accumulate the convolution result. 223 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 224 __m128i src8 = _mm_loadu_si128(row_to_filter); 225 // [16] a1 b1 g1 r1 a0 b0 g0 r0 226 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 227 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 228 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 229 // [32] a0*c0 b0*c0 g0*c0 r0*c0 230 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 231 accum = _mm_add_epi32(accum, t); 232 // [32] a1*c1 b1*c1 g1*c1 r1*c1 233 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 234 accum = _mm_add_epi32(accum, t); 235 236 // Duplicate 3rd and 4th coefficients for all channels => 237 // unpack the 3rd and 4th pixels to 16 bits => multiply with coefficients 238 // => accumulate the convolution results. 239 // [16] xx xx xx xx c3 c3 c2 c2 240 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 241 // [16] c3 c3 c3 c3 c2 c2 c2 c2 242 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 243 // [16] a3 g3 b3 r3 a2 g2 b2 r2 244 src16 = _mm_unpackhi_epi8(src8, zero); 245 mul_hi = _mm_mulhi_epi16(src16, coeff16); 246 mul_lo = _mm_mullo_epi16(src16, coeff16); 247 // [32] a2*c2 b2*c2 g2*c2 r2*c2 248 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 249 accum = _mm_add_epi32(accum, t); 250 // [32] a3*c3 b3*c3 g3*c3 r3*c3 251 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 252 accum = _mm_add_epi32(accum, t); 253 254 // Advance the pixel and coefficients pointers. 255 row_to_filter += 1; 256 filter_values += 4; 257 } 258 259 // When |filter_length| is not divisible by 4, we need to decimate some of 260 // the filter coefficient that was loaded incorrectly to zero; Other than 261 // that the algorithm is same with above, exceot that the 4th pixel will be 262 // always absent. 263 int r = filter_length&3; 264 if (r) { 265 // Note: filter_values must be padded to align_up(filter_offset, 8). 266 __m128i coeff, coeff16; 267 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 268 // Mask out extra filter taps. 269 coeff = _mm_and_si128(coeff, mask[r]); 270 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 271 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 272 273 // Note: line buffer must be padded to align_up(filter_offset, 16). 274 // We resolve this by use C-version for the last horizontal line. 275 __m128i src8 = _mm_loadu_si128(row_to_filter); 276 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 277 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 278 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 279 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 280 accum = _mm_add_epi32(accum, t); 281 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 282 accum = _mm_add_epi32(accum, t); 283 284 src16 = _mm_unpackhi_epi8(src8, zero); 285 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 286 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 287 mul_hi = _mm_mulhi_epi16(src16, coeff16); 288 mul_lo = _mm_mullo_epi16(src16, coeff16); 289 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 290 accum = _mm_add_epi32(accum, t); 291 } 292 293 // Shift right for fixed point implementation. 294 accum = _mm_srai_epi32(accum, SkConvolutionFilter1D::kShiftBits); 295 296 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). 297 accum = _mm_packs_epi32(accum, zero); 298 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). 299 accum = _mm_packus_epi16(accum, zero); 300 301 // Store the pixel value of 32 bits. 302 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum); 303 out_row += 4; 304 } 305 } 306 307 // Convolves horizontally along four rows. The row data is given in 308 // |src_data| and continues for the num_values() of the filter. 309 // The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please 310 // refer to that function for detailed comments. 311 void convolve4RowsHorizontally_SSE2(const unsigned char* src_data[4], 312 const SkConvolutionFilter1D& filter, 313 unsigned char* out_row[4]) { 314 int num_values = filter.numValues(); 315 316 int filter_offset, filter_length; 317 __m128i zero = _mm_setzero_si128(); 318 __m128i mask[4]; 319 // |mask| will be used to decimate all extra filter coefficients that are 320 // loaded by SIMD when |filter_length| is not divisible by 4. 321 // mask[0] is not used in following algorithm. 322 mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1); 323 mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1); 324 mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1); 325 326 // Output one pixel each iteration, calculating all channels (RGBA) together. 327 for (int out_x = 0; out_x < num_values; out_x++) { 328 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = 329 filter.FilterForValue(out_x, &filter_offset, &filter_length); 330 331 // four pixels in a column per iteration. 332 __m128i accum0 = _mm_setzero_si128(); 333 __m128i accum1 = _mm_setzero_si128(); 334 __m128i accum2 = _mm_setzero_si128(); 335 __m128i accum3 = _mm_setzero_si128(); 336 int start = (filter_offset<<2); 337 // We will load and accumulate with four coefficients per iteration. 338 for (int filter_x = 0; filter_x < (filter_length >> 2); filter_x++) { 339 __m128i coeff, coeff16lo, coeff16hi; 340 // [16] xx xx xx xx c3 c2 c1 c0 341 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 342 // [16] xx xx xx xx c1 c1 c0 c0 343 coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 344 // [16] c1 c1 c1 c1 c0 c0 c0 c0 345 coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo); 346 // [16] xx xx xx xx c3 c3 c2 c2 347 coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 348 // [16] c3 c3 c3 c3 c2 c2 c2 c2 349 coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi); 350 351 __m128i src8, src16, mul_hi, mul_lo, t; 352 353 #define ITERATION(src, accum) \ 354 src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \ 355 src16 = _mm_unpacklo_epi8(src8, zero); \ 356 mul_hi = _mm_mulhi_epi16(src16, coeff16lo); \ 357 mul_lo = _mm_mullo_epi16(src16, coeff16lo); \ 358 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ 359 accum = _mm_add_epi32(accum, t); \ 360 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ 361 accum = _mm_add_epi32(accum, t); \ 362 src16 = _mm_unpackhi_epi8(src8, zero); \ 363 mul_hi = _mm_mulhi_epi16(src16, coeff16hi); \ 364 mul_lo = _mm_mullo_epi16(src16, coeff16hi); \ 365 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ 366 accum = _mm_add_epi32(accum, t); \ 367 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ 368 accum = _mm_add_epi32(accum, t) 369 370 ITERATION(src_data[0] + start, accum0); 371 ITERATION(src_data[1] + start, accum1); 372 ITERATION(src_data[2] + start, accum2); 373 ITERATION(src_data[3] + start, accum3); 374 375 start += 16; 376 filter_values += 4; 377 } 378 379 int r = filter_length & 3; 380 if (r) { 381 // Note: filter_values must be padded to align_up(filter_offset, 8); 382 __m128i coeff; 383 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 384 // Mask out extra filter taps. 385 coeff = _mm_and_si128(coeff, mask[r]); 386 387 __m128i coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 388 /* c1 c1 c1 c1 c0 c0 c0 c0 */ 389 coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo); 390 __m128i coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 391 coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi); 392 393 __m128i src8, src16, mul_hi, mul_lo, t; 394 395 ITERATION(src_data[0] + start, accum0); 396 ITERATION(src_data[1] + start, accum1); 397 ITERATION(src_data[2] + start, accum2); 398 ITERATION(src_data[3] + start, accum3); 399 } 400 401 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 402 accum0 = _mm_packs_epi32(accum0, zero); 403 accum0 = _mm_packus_epi16(accum0, zero); 404 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 405 accum1 = _mm_packs_epi32(accum1, zero); 406 accum1 = _mm_packus_epi16(accum1, zero); 407 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 408 accum2 = _mm_packs_epi32(accum2, zero); 409 accum2 = _mm_packus_epi16(accum2, zero); 410 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); 411 accum3 = _mm_packs_epi32(accum3, zero); 412 accum3 = _mm_packus_epi16(accum3, zero); 413 414 *(reinterpret_cast<int*>(out_row[0])) = _mm_cvtsi128_si32(accum0); 415 *(reinterpret_cast<int*>(out_row[1])) = _mm_cvtsi128_si32(accum1); 416 *(reinterpret_cast<int*>(out_row[2])) = _mm_cvtsi128_si32(accum2); 417 *(reinterpret_cast<int*>(out_row[3])) = _mm_cvtsi128_si32(accum3); 418 419 out_row[0] += 4; 420 out_row[1] += 4; 421 out_row[2] += 4; 422 out_row[3] += 4; 423 } 424 } 425 426 // Does vertical convolution to produce one output row. The filter values and 427 // length are given in the first two parameters. These are applied to each 428 // of the rows pointed to in the |source_data_rows| array, with each row 429 // being |pixel_width| wide. 430 // 431 // The output must have room for |pixel_width * 4| bytes. 432 template<bool has_alpha> 433 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filter_values, 434 int filter_length, 435 unsigned char* const* source_data_rows, 436 int pixel_width, 437 unsigned char* out_row) { 438 int width = pixel_width & ~3; 439 440 __m128i zero = _mm_setzero_si128(); 441 __m128i accum0, accum1, accum2, accum3, coeff16; 442 const __m128i* src; 443 // Output four pixels per iteration (16 bytes). 444 for (int out_x = 0; out_x < width; out_x += 4) { 445 446 // Accumulated result for each pixel. 32 bits per RGBA channel. 447 accum0 = _mm_setzero_si128(); 448 accum1 = _mm_setzero_si128(); 449 accum2 = _mm_setzero_si128(); 450 accum3 = _mm_setzero_si128(); 451 452 // Convolve with one filter coefficient per iteration. 453 for (int filter_y = 0; filter_y < filter_length; filter_y++) { 454 455 // Duplicate the filter coefficient 8 times. 456 // [16] cj cj cj cj cj cj cj cj 457 coeff16 = _mm_set1_epi16(filter_values[filter_y]); 458 459 // Load four pixels (16 bytes) together. 460 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 461 src = reinterpret_cast<const __m128i*>( 462 &source_data_rows[filter_y][out_x << 2]); 463 __m128i src8 = _mm_loadu_si128(src); 464 465 // Unpack 1st and 2nd pixels from 8 bits to 16 bits for each channels => 466 // multiply with current coefficient => accumulate the result. 467 // [16] a1 b1 g1 r1 a0 b0 g0 r0 468 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 469 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 470 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 471 // [32] a0 b0 g0 r0 472 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 473 accum0 = _mm_add_epi32(accum0, t); 474 // [32] a1 b1 g1 r1 475 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 476 accum1 = _mm_add_epi32(accum1, t); 477 478 // Unpack 3rd and 4th pixels from 8 bits to 16 bits for each channels => 479 // multiply with current coefficient => accumulate the result. 480 // [16] a3 b3 g3 r3 a2 b2 g2 r2 481 src16 = _mm_unpackhi_epi8(src8, zero); 482 mul_hi = _mm_mulhi_epi16(src16, coeff16); 483 mul_lo = _mm_mullo_epi16(src16, coeff16); 484 // [32] a2 b2 g2 r2 485 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 486 accum2 = _mm_add_epi32(accum2, t); 487 // [32] a3 b3 g3 r3 488 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 489 accum3 = _mm_add_epi32(accum3, t); 490 } 491 492 // Shift right for fixed point implementation. 493 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 494 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 495 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 496 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); 497 498 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). 499 // [16] a1 b1 g1 r1 a0 b0 g0 r0 500 accum0 = _mm_packs_epi32(accum0, accum1); 501 // [16] a3 b3 g3 r3 a2 b2 g2 r2 502 accum2 = _mm_packs_epi32(accum2, accum3); 503 504 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). 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); 507 508 if (has_alpha) { 509 // Compute the max(ri, gi, bi) for each pixel. 510 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 511 __m128i a = _mm_srli_epi32(accum0, 8); 512 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 513 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. 514 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 515 a = _mm_srli_epi32(accum0, 16); 516 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 517 b = _mm_max_epu8(a, b); // Max of r and g and b. 518 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 519 b = _mm_slli_epi32(b, 24); 520 521 // Make sure the value of alpha channel is always larger than maximum 522 // value of color channels. 523 accum0 = _mm_max_epu8(b, accum0); 524 } else { 525 // Set value of alpha channels to 0xFF. 526 __m128i mask = _mm_set1_epi32(0xff000000); 527 accum0 = _mm_or_si128(accum0, mask); 528 } 529 530 // Store the convolution result (16 bytes) and advance the pixel pointers. 531 _mm_storeu_si128(reinterpret_cast<__m128i*>(out_row), accum0); 532 out_row += 16; 533 } 534 535 // When the width of the output is not divisible by 4, We need to save one 536 // pixel (4 bytes) each time. And also the fourth pixel is always absent. 537 if (pixel_width & 3) { 538 accum0 = _mm_setzero_si128(); 539 accum1 = _mm_setzero_si128(); 540 accum2 = _mm_setzero_si128(); 541 for (int filter_y = 0; filter_y < filter_length; ++filter_y) { 542 coeff16 = _mm_set1_epi16(filter_values[filter_y]); 543 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 544 src = reinterpret_cast<const __m128i*>( 545 &source_data_rows[filter_y][width<<2]); 546 __m128i src8 = _mm_loadu_si128(src); 547 // [16] a1 b1 g1 r1 a0 b0 g0 r0 548 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 549 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 550 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 551 // [32] a0 b0 g0 r0 552 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 553 accum0 = _mm_add_epi32(accum0, t); 554 // [32] a1 b1 g1 r1 555 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 556 accum1 = _mm_add_epi32(accum1, t); 557 // [16] a3 b3 g3 r3 a2 b2 g2 r2 558 src16 = _mm_unpackhi_epi8(src8, zero); 559 mul_hi = _mm_mulhi_epi16(src16, coeff16); 560 mul_lo = _mm_mullo_epi16(src16, coeff16); 561 // [32] a2 b2 g2 r2 562 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 563 accum2 = _mm_add_epi32(accum2, t); 564 } 565 566 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 567 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 568 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 569 // [16] a1 b1 g1 r1 a0 b0 g0 r0 570 accum0 = _mm_packs_epi32(accum0, accum1); 571 // [16] a3 b3 g3 r3 a2 b2 g2 r2 572 accum2 = _mm_packs_epi32(accum2, zero); 573 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 574 accum0 = _mm_packus_epi16(accum0, accum2); 575 if (has_alpha) { 576 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 577 __m128i a = _mm_srli_epi32(accum0, 8); 578 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 579 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. 580 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 581 a = _mm_srli_epi32(accum0, 16); 582 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 583 b = _mm_max_epu8(a, b); // Max of r and g and b. 584 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 585 b = _mm_slli_epi32(b, 24); 586 accum0 = _mm_max_epu8(b, accum0); 587 } else { 588 __m128i mask = _mm_set1_epi32(0xff000000); 589 accum0 = _mm_or_si128(accum0, mask); 590 } 591 592 for (int out_x = width; out_x < pixel_width; out_x++) { 593 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum0); 594 accum0 = _mm_srli_si128(accum0, 4); 595 out_row += 4; 596 } 597 } 598 } 599 600 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filter_values, 601 int filter_length, 602 unsigned char* const* source_data_rows, 603 int pixel_width, 604 unsigned char* out_row, 605 bool has_alpha) { 606 if (has_alpha) { 607 convolveVertically_SSE2<true>(filter_values, 608 filter_length, 609 source_data_rows, 610 pixel_width, 611 out_row); 612 } else { 613 convolveVertically_SSE2<false>(filter_values, 614 filter_length, 615 source_data_rows, 616 pixel_width, 617 out_row); 618 } 619 } 620 621 void applySIMDPadding_SSE2(SkConvolutionFilter1D *filter) { 622 // Padding |paddingCount| of more dummy coefficients after the coefficients 623 // of last filter to prevent SIMD instructions which load 8 or 16 bytes 624 // together to access invalid memory areas. We are not trying to align the 625 // coefficients right now due to the opaqueness of <vector> implementation. 626 // This has to be done after all |AddFilter| calls. 627 for (int i = 0; i < 8; ++i) { 628 filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0)); 629 } 630 } 631
