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 // Convolves horizontally along a single row. The row data is given in 44 // |src_data| and continues for the num_values() of the filter. 45 void convolveHorizontally_SSE2(const unsigned char* src_data, 46 const SkConvolutionFilter1D& filter, 47 unsigned char* out_row, 48 bool /*has_alpha*/) { 49 int num_values = filter.numValues(); 50 51 int filter_offset, filter_length; 52 __m128i zero = _mm_setzero_si128(); 53 __m128i mask[4]; 54 // |mask| will be used to decimate all extra filter coefficients that are 55 // loaded by SIMD when |filter_length| is not divisible by 4. 56 // mask[0] is not used in following algorithm. 57 mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1); 58 mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1); 59 mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1); 60 61 // Output one pixel each iteration, calculating all channels (RGBA) together. 62 for (int out_x = 0; out_x < num_values; out_x++) { 63 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = 64 filter.FilterForValue(out_x, &filter_offset, &filter_length); 65 66 __m128i accum = _mm_setzero_si128(); 67 68 // Compute the first pixel in this row that the filter affects. It will 69 // touch |filter_length| pixels (4 bytes each) after this. 70 const __m128i* row_to_filter = 71 reinterpret_cast<const __m128i*>(&src_data[filter_offset << 2]); 72 73 // We will load and accumulate with four coefficients per iteration. 74 for (int filter_x = 0; filter_x < filter_length >> 2; filter_x++) { 75 76 // Load 4 coefficients => duplicate 1st and 2nd of them for all channels. 77 __m128i coeff, coeff16; 78 // [16] xx xx xx xx c3 c2 c1 c0 79 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 80 // [16] xx xx xx xx c1 c1 c0 c0 81 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 82 // [16] c1 c1 c1 c1 c0 c0 c0 c0 83 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 84 85 // Load four pixels => unpack the first two pixels to 16 bits => 86 // multiply with coefficients => accumulate the convolution result. 87 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 88 __m128i src8 = _mm_loadu_si128(row_to_filter); 89 // [16] a1 b1 g1 r1 a0 b0 g0 r0 90 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 91 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 92 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 93 // [32] a0*c0 b0*c0 g0*c0 r0*c0 94 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 95 accum = _mm_add_epi32(accum, t); 96 // [32] a1*c1 b1*c1 g1*c1 r1*c1 97 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 98 accum = _mm_add_epi32(accum, t); 99 100 // Duplicate 3rd and 4th coefficients for all channels => 101 // unpack the 3rd and 4th pixels to 16 bits => multiply with coefficients 102 // => accumulate the convolution results. 103 // [16] xx xx xx xx c3 c3 c2 c2 104 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 105 // [16] c3 c3 c3 c3 c2 c2 c2 c2 106 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 107 // [16] a3 g3 b3 r3 a2 g2 b2 r2 108 src16 = _mm_unpackhi_epi8(src8, zero); 109 mul_hi = _mm_mulhi_epi16(src16, coeff16); 110 mul_lo = _mm_mullo_epi16(src16, coeff16); 111 // [32] a2*c2 b2*c2 g2*c2 r2*c2 112 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 113 accum = _mm_add_epi32(accum, t); 114 // [32] a3*c3 b3*c3 g3*c3 r3*c3 115 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 116 accum = _mm_add_epi32(accum, t); 117 118 // Advance the pixel and coefficients pointers. 119 row_to_filter += 1; 120 filter_values += 4; 121 } 122 123 // When |filter_length| is not divisible by 4, we need to decimate some of 124 // the filter coefficient that was loaded incorrectly to zero; Other than 125 // that the algorithm is same with above, exceot that the 4th pixel will be 126 // always absent. 127 int r = filter_length&3; 128 if (r) { 129 // Note: filter_values must be padded to align_up(filter_offset, 8). 130 __m128i coeff, coeff16; 131 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 132 // Mask out extra filter taps. 133 coeff = _mm_and_si128(coeff, mask[r]); 134 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 135 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 136 137 // Note: line buffer must be padded to align_up(filter_offset, 16). 138 // We resolve this by use C-version for the last horizontal line. 139 __m128i src8 = _mm_loadu_si128(row_to_filter); 140 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 141 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 142 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 143 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 144 accum = _mm_add_epi32(accum, t); 145 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 146 accum = _mm_add_epi32(accum, t); 147 148 src16 = _mm_unpackhi_epi8(src8, zero); 149 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 150 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); 151 mul_hi = _mm_mulhi_epi16(src16, coeff16); 152 mul_lo = _mm_mullo_epi16(src16, coeff16); 153 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 154 accum = _mm_add_epi32(accum, t); 155 } 156 157 // Shift right for fixed point implementation. 158 accum = _mm_srai_epi32(accum, SkConvolutionFilter1D::kShiftBits); 159 160 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). 161 accum = _mm_packs_epi32(accum, zero); 162 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). 163 accum = _mm_packus_epi16(accum, zero); 164 165 // Store the pixel value of 32 bits. 166 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum); 167 out_row += 4; 168 } 169 } 170 171 // Convolves horizontally along four rows. The row data is given in 172 // |src_data| and continues for the num_values() of the filter. 173 // The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please 174 // refer to that function for detailed comments. 175 void convolve4RowsHorizontally_SSE2(const unsigned char* src_data[4], 176 const SkConvolutionFilter1D& filter, 177 unsigned char* out_row[4]) { 178 int num_values = filter.numValues(); 179 180 int filter_offset, filter_length; 181 __m128i zero = _mm_setzero_si128(); 182 __m128i mask[4]; 183 // |mask| will be used to decimate all extra filter coefficients that are 184 // loaded by SIMD when |filter_length| is not divisible by 4. 185 // mask[0] is not used in following algorithm. 186 mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1); 187 mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1); 188 mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1); 189 190 // Output one pixel each iteration, calculating all channels (RGBA) together. 191 for (int out_x = 0; out_x < num_values; out_x++) { 192 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = 193 filter.FilterForValue(out_x, &filter_offset, &filter_length); 194 195 // four pixels in a column per iteration. 196 __m128i accum0 = _mm_setzero_si128(); 197 __m128i accum1 = _mm_setzero_si128(); 198 __m128i accum2 = _mm_setzero_si128(); 199 __m128i accum3 = _mm_setzero_si128(); 200 int start = (filter_offset<<2); 201 // We will load and accumulate with four coefficients per iteration. 202 for (int filter_x = 0; filter_x < (filter_length >> 2); filter_x++) { 203 __m128i coeff, coeff16lo, coeff16hi; 204 // [16] xx xx xx xx c3 c2 c1 c0 205 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 206 // [16] xx xx xx xx c1 c1 c0 c0 207 coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 208 // [16] c1 c1 c1 c1 c0 c0 c0 c0 209 coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo); 210 // [16] xx xx xx xx c3 c3 c2 c2 211 coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 212 // [16] c3 c3 c3 c3 c2 c2 c2 c2 213 coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi); 214 215 __m128i src8, src16, mul_hi, mul_lo, t; 216 217 #define ITERATION(src, accum) \ 218 src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \ 219 src16 = _mm_unpacklo_epi8(src8, zero); \ 220 mul_hi = _mm_mulhi_epi16(src16, coeff16lo); \ 221 mul_lo = _mm_mullo_epi16(src16, coeff16lo); \ 222 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ 223 accum = _mm_add_epi32(accum, t); \ 224 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ 225 accum = _mm_add_epi32(accum, t); \ 226 src16 = _mm_unpackhi_epi8(src8, zero); \ 227 mul_hi = _mm_mulhi_epi16(src16, coeff16hi); \ 228 mul_lo = _mm_mullo_epi16(src16, coeff16hi); \ 229 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ 230 accum = _mm_add_epi32(accum, t); \ 231 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ 232 accum = _mm_add_epi32(accum, t) 233 234 ITERATION(src_data[0] + start, accum0); 235 ITERATION(src_data[1] + start, accum1); 236 ITERATION(src_data[2] + start, accum2); 237 ITERATION(src_data[3] + start, accum3); 238 239 start += 16; 240 filter_values += 4; 241 } 242 243 int r = filter_length & 3; 244 if (r) { 245 // Note: filter_values must be padded to align_up(filter_offset, 8); 246 __m128i coeff; 247 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values)); 248 // Mask out extra filter taps. 249 coeff = _mm_and_si128(coeff, mask[r]); 250 251 __m128i coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); 252 /* c1 c1 c1 c1 c0 c0 c0 c0 */ 253 coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo); 254 __m128i coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); 255 coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi); 256 257 __m128i src8, src16, mul_hi, mul_lo, t; 258 259 ITERATION(src_data[0] + start, accum0); 260 ITERATION(src_data[1] + start, accum1); 261 ITERATION(src_data[2] + start, accum2); 262 ITERATION(src_data[3] + start, accum3); 263 } 264 265 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 266 accum0 = _mm_packs_epi32(accum0, zero); 267 accum0 = _mm_packus_epi16(accum0, zero); 268 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 269 accum1 = _mm_packs_epi32(accum1, zero); 270 accum1 = _mm_packus_epi16(accum1, zero); 271 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 272 accum2 = _mm_packs_epi32(accum2, zero); 273 accum2 = _mm_packus_epi16(accum2, zero); 274 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); 275 accum3 = _mm_packs_epi32(accum3, zero); 276 accum3 = _mm_packus_epi16(accum3, zero); 277 278 *(reinterpret_cast<int*>(out_row[0])) = _mm_cvtsi128_si32(accum0); 279 *(reinterpret_cast<int*>(out_row[1])) = _mm_cvtsi128_si32(accum1); 280 *(reinterpret_cast<int*>(out_row[2])) = _mm_cvtsi128_si32(accum2); 281 *(reinterpret_cast<int*>(out_row[3])) = _mm_cvtsi128_si32(accum3); 282 283 out_row[0] += 4; 284 out_row[1] += 4; 285 out_row[2] += 4; 286 out_row[3] += 4; 287 } 288 } 289 290 // Does vertical convolution to produce one output row. The filter values and 291 // length are given in the first two parameters. These are applied to each 292 // of the rows pointed to in the |source_data_rows| array, with each row 293 // being |pixel_width| wide. 294 // 295 // The output must have room for |pixel_width * 4| bytes. 296 template<bool has_alpha> 297 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filter_values, 298 int filter_length, 299 unsigned char* const* source_data_rows, 300 int pixel_width, 301 unsigned char* out_row) { 302 int width = pixel_width & ~3; 303 304 __m128i zero = _mm_setzero_si128(); 305 __m128i accum0, accum1, accum2, accum3, coeff16; 306 const __m128i* src; 307 // Output four pixels per iteration (16 bytes). 308 for (int out_x = 0; out_x < width; out_x += 4) { 309 310 // Accumulated result for each pixel. 32 bits per RGBA channel. 311 accum0 = _mm_setzero_si128(); 312 accum1 = _mm_setzero_si128(); 313 accum2 = _mm_setzero_si128(); 314 accum3 = _mm_setzero_si128(); 315 316 // Convolve with one filter coefficient per iteration. 317 for (int filter_y = 0; filter_y < filter_length; filter_y++) { 318 319 // Duplicate the filter coefficient 8 times. 320 // [16] cj cj cj cj cj cj cj cj 321 coeff16 = _mm_set1_epi16(filter_values[filter_y]); 322 323 // Load four pixels (16 bytes) together. 324 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 325 src = reinterpret_cast<const __m128i*>( 326 &source_data_rows[filter_y][out_x << 2]); 327 __m128i src8 = _mm_loadu_si128(src); 328 329 // Unpack 1st and 2nd pixels from 8 bits to 16 bits for each channels => 330 // multiply with current coefficient => accumulate the result. 331 // [16] a1 b1 g1 r1 a0 b0 g0 r0 332 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 333 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 334 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 335 // [32] a0 b0 g0 r0 336 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 337 accum0 = _mm_add_epi32(accum0, t); 338 // [32] a1 b1 g1 r1 339 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 340 accum1 = _mm_add_epi32(accum1, t); 341 342 // Unpack 3rd and 4th pixels from 8 bits to 16 bits for each channels => 343 // multiply with current coefficient => accumulate the result. 344 // [16] a3 b3 g3 r3 a2 b2 g2 r2 345 src16 = _mm_unpackhi_epi8(src8, zero); 346 mul_hi = _mm_mulhi_epi16(src16, coeff16); 347 mul_lo = _mm_mullo_epi16(src16, coeff16); 348 // [32] a2 b2 g2 r2 349 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 350 accum2 = _mm_add_epi32(accum2, t); 351 // [32] a3 b3 g3 r3 352 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 353 accum3 = _mm_add_epi32(accum3, t); 354 } 355 356 // Shift right for fixed point implementation. 357 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 358 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 359 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 360 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); 361 362 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). 363 // [16] a1 b1 g1 r1 a0 b0 g0 r0 364 accum0 = _mm_packs_epi32(accum0, accum1); 365 // [16] a3 b3 g3 r3 a2 b2 g2 r2 366 accum2 = _mm_packs_epi32(accum2, accum3); 367 368 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). 369 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 370 accum0 = _mm_packus_epi16(accum0, accum2); 371 372 if (has_alpha) { 373 // Compute the max(ri, gi, bi) for each pixel. 374 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 375 __m128i a = _mm_srli_epi32(accum0, 8); 376 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 377 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. 378 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 379 a = _mm_srli_epi32(accum0, 16); 380 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 381 b = _mm_max_epu8(a, b); // Max of r and g and b. 382 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 383 b = _mm_slli_epi32(b, 24); 384 385 // Make sure the value of alpha channel is always larger than maximum 386 // value of color channels. 387 accum0 = _mm_max_epu8(b, accum0); 388 } else { 389 // Set value of alpha channels to 0xFF. 390 __m128i mask = _mm_set1_epi32(0xff000000); 391 accum0 = _mm_or_si128(accum0, mask); 392 } 393 394 // Store the convolution result (16 bytes) and advance the pixel pointers. 395 _mm_storeu_si128(reinterpret_cast<__m128i*>(out_row), accum0); 396 out_row += 16; 397 } 398 399 // When the width of the output is not divisible by 4, We need to save one 400 // pixel (4 bytes) each time. And also the fourth pixel is always absent. 401 if (pixel_width & 3) { 402 accum0 = _mm_setzero_si128(); 403 accum1 = _mm_setzero_si128(); 404 accum2 = _mm_setzero_si128(); 405 for (int filter_y = 0; filter_y < filter_length; ++filter_y) { 406 coeff16 = _mm_set1_epi16(filter_values[filter_y]); 407 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 408 src = reinterpret_cast<const __m128i*>( 409 &source_data_rows[filter_y][width<<2]); 410 __m128i src8 = _mm_loadu_si128(src); 411 // [16] a1 b1 g1 r1 a0 b0 g0 r0 412 __m128i src16 = _mm_unpacklo_epi8(src8, zero); 413 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); 414 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); 415 // [32] a0 b0 g0 r0 416 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); 417 accum0 = _mm_add_epi32(accum0, t); 418 // [32] a1 b1 g1 r1 419 t = _mm_unpackhi_epi16(mul_lo, mul_hi); 420 accum1 = _mm_add_epi32(accum1, t); 421 // [16] a3 b3 g3 r3 a2 b2 g2 r2 422 src16 = _mm_unpackhi_epi8(src8, zero); 423 mul_hi = _mm_mulhi_epi16(src16, coeff16); 424 mul_lo = _mm_mullo_epi16(src16, coeff16); 425 // [32] a2 b2 g2 r2 426 t = _mm_unpacklo_epi16(mul_lo, mul_hi); 427 accum2 = _mm_add_epi32(accum2, t); 428 } 429 430 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); 431 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); 432 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); 433 // [16] a1 b1 g1 r1 a0 b0 g0 r0 434 accum0 = _mm_packs_epi32(accum0, accum1); 435 // [16] a3 b3 g3 r3 a2 b2 g2 r2 436 accum2 = _mm_packs_epi32(accum2, zero); 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 if (has_alpha) { 440 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 441 __m128i a = _mm_srli_epi32(accum0, 8); 442 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 443 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. 444 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 445 a = _mm_srli_epi32(accum0, 16); 446 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 447 b = _mm_max_epu8(a, b); // Max of r and g and b. 448 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 449 b = _mm_slli_epi32(b, 24); 450 accum0 = _mm_max_epu8(b, accum0); 451 } else { 452 __m128i mask = _mm_set1_epi32(0xff000000); 453 accum0 = _mm_or_si128(accum0, mask); 454 } 455 456 for (int out_x = width; out_x < pixel_width; out_x++) { 457 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum0); 458 accum0 = _mm_srli_si128(accum0, 4); 459 out_row += 4; 460 } 461 } 462 } 463 464 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filter_values, 465 int filter_length, 466 unsigned char* const* source_data_rows, 467 int pixel_width, 468 unsigned char* out_row, 469 bool has_alpha) { 470 if (has_alpha) { 471 convolveVertically_SSE2<true>(filter_values, 472 filter_length, 473 source_data_rows, 474 pixel_width, 475 out_row); 476 } else { 477 convolveVertically_SSE2<false>(filter_values, 478 filter_length, 479 source_data_rows, 480 pixel_width, 481 out_row); 482 } 483 } 484 485 void applySIMDPadding_SSE2(SkConvolutionFilter1D *filter) { 486 // Padding |paddingCount| of more dummy coefficients after the coefficients 487 // of last filter to prevent SIMD instructions which load 8 or 16 bytes 488 // together to access invalid memory areas. We are not trying to align the 489 // coefficients right now due to the opaqueness of <vector> implementation. 490 // This has to be done after all |AddFilter| calls. 491 for (int i = 0; i < 8; ++i) { 492 filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0)); 493 } 494 } 495