| OLD | NEW |
|---|
| (Empty) |
| 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 %.3
u %.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 static SK_ALWAYS_INLINE void accum_remainder(const unsigned char* pixels_left, | |
| 44 const SkConvolutionFilter1D::ConvolutionFixed* filter_values, __m128i& a
ccum, int r) { | |
| 45 int remainder[4] = {0}; | |
| 46 for (int i = 0; i < r; i++) { | |
| 47 SkConvolutionFilter1D::ConvolutionFixed coeff = filter_values[i]; | |
| 48 remainder[0] += coeff * pixels_left[i * 4 + 0]; | |
| 49 remainder[1] += coeff * pixels_left[i * 4 + 1]; | |
| 50 remainder[2] += coeff * pixels_left[i * 4 + 2]; | |
| 51 remainder[3] += coeff * pixels_left[i * 4 + 3]; | |
| 52 } | |
| 53 __m128i t = _mm_setr_epi32(remainder[0], remainder[1], remainder[2], remaind
er[3]); | |
| 54 accum = _mm_add_epi32(accum, t); | |
| 55 } | |
| 56 | |
| 57 // Convolves horizontally along a single row. The row data is given in | |
| 58 // |src_data| and continues for the num_values() of the filter. | |
| 59 void convolveHorizontally_SSE2(const unsigned char* src_data, | |
| 60 const SkConvolutionFilter1D& filter, | |
| 61 unsigned char* out_row, | |
| 62 bool /*has_alpha*/) { | |
| 63 int num_values = filter.numValues(); | |
| 64 | |
| 65 int filter_offset, filter_length; | |
| 66 __m128i zero = _mm_setzero_si128(); | |
| 67 | |
| 68 // Output one pixel each iteration, calculating all channels (RGBA) together
. | |
| 69 for (int out_x = 0; out_x < num_values; out_x++) { | |
| 70 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = | |
| 71 filter.FilterForValue(out_x, &filter_offset, &filter_length); | |
| 72 | |
| 73 __m128i accum = _mm_setzero_si128(); | |
| 74 | |
| 75 // Compute the first pixel in this row that the filter affects. It will | |
| 76 // touch |filter_length| pixels (4 bytes each) after this. | |
| 77 const __m128i* row_to_filter = | |
| 78 reinterpret_cast<const __m128i*>(&src_data[filter_offset << 2]); | |
| 79 | |
| 80 // We will load and accumulate with four coefficients per iteration. | |
| 81 for (int filter_x = 0; filter_x < filter_length >> 2; filter_x++) { | |
| 82 | |
| 83 // Load 4 coefficients => duplicate 1st and 2nd of them for all chan
nels. | |
| 84 __m128i coeff, coeff16; | |
| 85 // [16] xx xx xx xx c3 c2 c1 c0 | |
| 86 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_valu
es)); | |
| 87 // [16] xx xx xx xx c1 c1 c0 c0 | |
| 88 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); | |
| 89 // [16] c1 c1 c1 c1 c0 c0 c0 c0 | |
| 90 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); | |
| 91 | |
| 92 // Load four pixels => unpack the first two pixels to 16 bits => | |
| 93 // multiply with coefficients => accumulate the convolution result. | |
| 94 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 | |
| 95 __m128i src8 = _mm_loadu_si128(row_to_filter); | |
| 96 // [16] a1 b1 g1 r1 a0 b0 g0 r0 | |
| 97 __m128i src16 = _mm_unpacklo_epi8(src8, zero); | |
| 98 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); | |
| 99 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); | |
| 100 // [32] a0*c0 b0*c0 g0*c0 r0*c0 | |
| 101 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); | |
| 102 accum = _mm_add_epi32(accum, t); | |
| 103 // [32] a1*c1 b1*c1 g1*c1 r1*c1 | |
| 104 t = _mm_unpackhi_epi16(mul_lo, mul_hi); | |
| 105 accum = _mm_add_epi32(accum, t); | |
| 106 | |
| 107 // Duplicate 3rd and 4th coefficients for all channels => | |
| 108 // unpack the 3rd and 4th pixels to 16 bits => multiply with coeffic
ients | |
| 109 // => accumulate the convolution results. | |
| 110 // [16] xx xx xx xx c3 c3 c2 c2 | |
| 111 coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); | |
| 112 // [16] c3 c3 c3 c3 c2 c2 c2 c2 | |
| 113 coeff16 = _mm_unpacklo_epi16(coeff16, coeff16); | |
| 114 // [16] a3 g3 b3 r3 a2 g2 b2 r2 | |
| 115 src16 = _mm_unpackhi_epi8(src8, zero); | |
| 116 mul_hi = _mm_mulhi_epi16(src16, coeff16); | |
| 117 mul_lo = _mm_mullo_epi16(src16, coeff16); | |
| 118 // [32] a2*c2 b2*c2 g2*c2 r2*c2 | |
| 119 t = _mm_unpacklo_epi16(mul_lo, mul_hi); | |
| 120 accum = _mm_add_epi32(accum, t); | |
| 121 // [32] a3*c3 b3*c3 g3*c3 r3*c3 | |
| 122 t = _mm_unpackhi_epi16(mul_lo, mul_hi); | |
| 123 accum = _mm_add_epi32(accum, t); | |
| 124 | |
| 125 // Advance the pixel and coefficients pointers. | |
| 126 row_to_filter += 1; | |
| 127 filter_values += 4; | |
| 128 } | |
| 129 | |
| 130 // When |filter_length| is not divisible by 4, we accumulate the last 1
- 3 | |
| 131 // coefficients one at a time. | |
| 132 int r = filter_length & 3; | |
| 133 if (r) { | |
| 134 int remainder_offset = (filter_offset + filter_length - r) * 4; | |
| 135 accum_remainder(src_data + remainder_offset, filter_values, accum, r
); | |
| 136 } | |
| 137 | |
| 138 // Shift right for fixed point implementation. | |
| 139 accum = _mm_srai_epi32(accum, SkConvolutionFilter1D::kShiftBits); | |
| 140 | |
| 141 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). | |
| 142 accum = _mm_packs_epi32(accum, zero); | |
| 143 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). | |
| 144 accum = _mm_packus_epi16(accum, zero); | |
| 145 | |
| 146 // Store the pixel value of 32 bits. | |
| 147 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum); | |
| 148 out_row += 4; | |
| 149 } | |
| 150 } | |
| 151 | |
| 152 // Convolves horizontally along four rows. The row data is given in | |
| 153 // |src_data| and continues for the num_values() of the filter. | |
| 154 // The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please | |
| 155 // refer to that function for detailed comments. | |
| 156 void convolve4RowsHorizontally_SSE2(const unsigned char* src_data[4], | |
| 157 const SkConvolutionFilter1D& filter, | |
| 158 unsigned char* out_row[4], | |
| 159 size_t outRowBytes) { | |
| 160 SkDEBUGCODE(const unsigned char* out_row_0_start = out_row[0];) | |
| 161 | |
| 162 int num_values = filter.numValues(); | |
| 163 | |
| 164 int filter_offset, filter_length; | |
| 165 __m128i zero = _mm_setzero_si128(); | |
| 166 | |
| 167 // Output one pixel each iteration, calculating all channels (RGBA) together
. | |
| 168 for (int out_x = 0; out_x < num_values; out_x++) { | |
| 169 const SkConvolutionFilter1D::ConvolutionFixed* filter_values = | |
| 170 filter.FilterForValue(out_x, &filter_offset, &filter_length); | |
| 171 | |
| 172 // four pixels in a column per iteration. | |
| 173 __m128i accum0 = _mm_setzero_si128(); | |
| 174 __m128i accum1 = _mm_setzero_si128(); | |
| 175 __m128i accum2 = _mm_setzero_si128(); | |
| 176 __m128i accum3 = _mm_setzero_si128(); | |
| 177 int start = (filter_offset<<2); | |
| 178 // We will load and accumulate with four coefficients per iteration. | |
| 179 for (int filter_x = 0; filter_x < (filter_length >> 2); filter_x++) { | |
| 180 __m128i coeff, coeff16lo, coeff16hi; | |
| 181 // [16] xx xx xx xx c3 c2 c1 c0 | |
| 182 coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_valu
es)); | |
| 183 // [16] xx xx xx xx c1 c1 c0 c0 | |
| 184 coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0)); | |
| 185 // [16] c1 c1 c1 c1 c0 c0 c0 c0 | |
| 186 coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo); | |
| 187 // [16] xx xx xx xx c3 c3 c2 c2 | |
| 188 coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2)); | |
| 189 // [16] c3 c3 c3 c3 c2 c2 c2 c2 | |
| 190 coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi); | |
| 191 | |
| 192 __m128i src8, src16, mul_hi, mul_lo, t; | |
| 193 | |
| 194 #define ITERATION(src, accum) \ | |
| 195 src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \ | |
| 196 src16 = _mm_unpacklo_epi8(src8, zero); \ | |
| 197 mul_hi = _mm_mulhi_epi16(src16, coeff16lo); \ | |
| 198 mul_lo = _mm_mullo_epi16(src16, coeff16lo); \ | |
| 199 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ | |
| 200 accum = _mm_add_epi32(accum, t); \ | |
| 201 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ | |
| 202 accum = _mm_add_epi32(accum, t); \ | |
| 203 src16 = _mm_unpackhi_epi8(src8, zero); \ | |
| 204 mul_hi = _mm_mulhi_epi16(src16, coeff16hi); \ | |
| 205 mul_lo = _mm_mullo_epi16(src16, coeff16hi); \ | |
| 206 t = _mm_unpacklo_epi16(mul_lo, mul_hi); \ | |
| 207 accum = _mm_add_epi32(accum, t); \ | |
| 208 t = _mm_unpackhi_epi16(mul_lo, mul_hi); \ | |
| 209 accum = _mm_add_epi32(accum, t) | |
| 210 | |
| 211 ITERATION(src_data[0] + start, accum0); | |
| 212 ITERATION(src_data[1] + start, accum1); | |
| 213 ITERATION(src_data[2] + start, accum2); | |
| 214 ITERATION(src_data[3] + start, accum3); | |
| 215 | |
| 216 start += 16; | |
| 217 filter_values += 4; | |
| 218 } | |
| 219 | |
| 220 int r = filter_length & 3; | |
| 221 if (r) { | |
| 222 int remainder_offset = (filter_offset + filter_length - r) * 4; | |
| 223 accum_remainder(src_data[0] + remainder_offset, filter_values, accum
0, r); | |
| 224 accum_remainder(src_data[1] + remainder_offset, filter_values, accum
1, r); | |
| 225 accum_remainder(src_data[2] + remainder_offset, filter_values, accum
2, r); | |
| 226 accum_remainder(src_data[3] + remainder_offset, filter_values, accum
3, r); | |
| 227 } | |
| 228 | |
| 229 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); | |
| 230 accum0 = _mm_packs_epi32(accum0, zero); | |
| 231 accum0 = _mm_packus_epi16(accum0, zero); | |
| 232 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); | |
| 233 accum1 = _mm_packs_epi32(accum1, zero); | |
| 234 accum1 = _mm_packus_epi16(accum1, zero); | |
| 235 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); | |
| 236 accum2 = _mm_packs_epi32(accum2, zero); | |
| 237 accum2 = _mm_packus_epi16(accum2, zero); | |
| 238 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); | |
| 239 accum3 = _mm_packs_epi32(accum3, zero); | |
| 240 accum3 = _mm_packus_epi16(accum3, zero); | |
| 241 | |
| 242 // We seem to be running off the edge here (chromium:491660). | |
| 243 SkASSERT(((size_t)out_row[0] - (size_t)out_row_0_start) < outRowBytes); | |
| 244 | |
| 245 *(reinterpret_cast<int*>(out_row[0])) = _mm_cvtsi128_si32(accum0); | |
| 246 *(reinterpret_cast<int*>(out_row[1])) = _mm_cvtsi128_si32(accum1); | |
| 247 *(reinterpret_cast<int*>(out_row[2])) = _mm_cvtsi128_si32(accum2); | |
| 248 *(reinterpret_cast<int*>(out_row[3])) = _mm_cvtsi128_si32(accum3); | |
| 249 | |
| 250 out_row[0] += 4; | |
| 251 out_row[1] += 4; | |
| 252 out_row[2] += 4; | |
| 253 out_row[3] += 4; | |
| 254 } | |
| 255 } | |
| 256 | |
| 257 // Does vertical convolution to produce one output row. The filter values and | |
| 258 // length are given in the first two parameters. These are applied to each | |
| 259 // of the rows pointed to in the |source_data_rows| array, with each row | |
| 260 // being |pixel_width| wide. | |
| 261 // | |
| 262 // The output must have room for |pixel_width * 4| bytes. | |
| 263 template<bool has_alpha> | |
| 264 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filt
er_values, | |
| 265 int filter_length, | |
| 266 unsigned char* const* source_data_rows, | |
| 267 int pixel_width, | |
| 268 unsigned char* out_row) { | |
| 269 int width = pixel_width & ~3; | |
| 270 | |
| 271 __m128i zero = _mm_setzero_si128(); | |
| 272 __m128i accum0, accum1, accum2, accum3, coeff16; | |
| 273 const __m128i* src; | |
| 274 // Output four pixels per iteration (16 bytes). | |
| 275 for (int out_x = 0; out_x < width; out_x += 4) { | |
| 276 | |
| 277 // Accumulated result for each pixel. 32 bits per RGBA channel. | |
| 278 accum0 = _mm_setzero_si128(); | |
| 279 accum1 = _mm_setzero_si128(); | |
| 280 accum2 = _mm_setzero_si128(); | |
| 281 accum3 = _mm_setzero_si128(); | |
| 282 | |
| 283 // Convolve with one filter coefficient per iteration. | |
| 284 for (int filter_y = 0; filter_y < filter_length; filter_y++) { | |
| 285 | |
| 286 // Duplicate the filter coefficient 8 times. | |
| 287 // [16] cj cj cj cj cj cj cj cj | |
| 288 coeff16 = _mm_set1_epi16(filter_values[filter_y]); | |
| 289 | |
| 290 // Load four pixels (16 bytes) together. | |
| 291 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 | |
| 292 src = reinterpret_cast<const __m128i*>( | |
| 293 &source_data_rows[filter_y][out_x << 2]); | |
| 294 __m128i src8 = _mm_loadu_si128(src); | |
| 295 | |
| 296 // Unpack 1st and 2nd pixels from 8 bits to 16 bits for each channel
s => | |
| 297 // multiply with current coefficient => accumulate the result. | |
| 298 // [16] a1 b1 g1 r1 a0 b0 g0 r0 | |
| 299 __m128i src16 = _mm_unpacklo_epi8(src8, zero); | |
| 300 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); | |
| 301 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); | |
| 302 // [32] a0 b0 g0 r0 | |
| 303 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); | |
| 304 accum0 = _mm_add_epi32(accum0, t); | |
| 305 // [32] a1 b1 g1 r1 | |
| 306 t = _mm_unpackhi_epi16(mul_lo, mul_hi); | |
| 307 accum1 = _mm_add_epi32(accum1, t); | |
| 308 | |
| 309 // Unpack 3rd and 4th pixels from 8 bits to 16 bits for each channel
s => | |
| 310 // multiply with current coefficient => accumulate the result. | |
| 311 // [16] a3 b3 g3 r3 a2 b2 g2 r2 | |
| 312 src16 = _mm_unpackhi_epi8(src8, zero); | |
| 313 mul_hi = _mm_mulhi_epi16(src16, coeff16); | |
| 314 mul_lo = _mm_mullo_epi16(src16, coeff16); | |
| 315 // [32] a2 b2 g2 r2 | |
| 316 t = _mm_unpacklo_epi16(mul_lo, mul_hi); | |
| 317 accum2 = _mm_add_epi32(accum2, t); | |
| 318 // [32] a3 b3 g3 r3 | |
| 319 t = _mm_unpackhi_epi16(mul_lo, mul_hi); | |
| 320 accum3 = _mm_add_epi32(accum3, t); | |
| 321 } | |
| 322 | |
| 323 // Shift right for fixed point implementation. | |
| 324 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); | |
| 325 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); | |
| 326 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); | |
| 327 accum3 = _mm_srai_epi32(accum3, SkConvolutionFilter1D::kShiftBits); | |
| 328 | |
| 329 // Packing 32 bits |accum| to 16 bits per channel (signed saturation). | |
| 330 // [16] a1 b1 g1 r1 a0 b0 g0 r0 | |
| 331 accum0 = _mm_packs_epi32(accum0, accum1); | |
| 332 // [16] a3 b3 g3 r3 a2 b2 g2 r2 | |
| 333 accum2 = _mm_packs_epi32(accum2, accum3); | |
| 334 | |
| 335 // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation). | |
| 336 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 | |
| 337 accum0 = _mm_packus_epi16(accum0, accum2); | |
| 338 | |
| 339 if (has_alpha) { | |
| 340 // Compute the max(ri, gi, bi) for each pixel. | |
| 341 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 | |
| 342 __m128i a = _mm_srli_epi32(accum0, 8); | |
| 343 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 | |
| 344 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. | |
| 345 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 | |
| 346 a = _mm_srli_epi32(accum0, 16); | |
| 347 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 | |
| 348 b = _mm_max_epu8(a, b); // Max of r and g and b. | |
| 349 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 | |
| 350 b = _mm_slli_epi32(b, 24); | |
| 351 | |
| 352 // Make sure the value of alpha channel is always larger than maximu
m | |
| 353 // value of color channels. | |
| 354 accum0 = _mm_max_epu8(b, accum0); | |
| 355 } else { | |
| 356 // Set value of alpha channels to 0xFF. | |
| 357 __m128i mask = _mm_set1_epi32(0xff000000); | |
| 358 accum0 = _mm_or_si128(accum0, mask); | |
| 359 } | |
| 360 | |
| 361 // Store the convolution result (16 bytes) and advance the pixel pointer
s. | |
| 362 _mm_storeu_si128(reinterpret_cast<__m128i*>(out_row), accum0); | |
| 363 out_row += 16; | |
| 364 } | |
| 365 | |
| 366 // When the width of the output is not divisible by 4, We need to save one | |
| 367 // pixel (4 bytes) each time. And also the fourth pixel is always absent. | |
| 368 if (pixel_width & 3) { | |
| 369 accum0 = _mm_setzero_si128(); | |
| 370 accum1 = _mm_setzero_si128(); | |
| 371 accum2 = _mm_setzero_si128(); | |
| 372 for (int filter_y = 0; filter_y < filter_length; ++filter_y) { | |
| 373 coeff16 = _mm_set1_epi16(filter_values[filter_y]); | |
| 374 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 | |
| 375 src = reinterpret_cast<const __m128i*>( | |
| 376 &source_data_rows[filter_y][width<<2]); | |
| 377 __m128i src8 = _mm_loadu_si128(src); | |
| 378 // [16] a1 b1 g1 r1 a0 b0 g0 r0 | |
| 379 __m128i src16 = _mm_unpacklo_epi8(src8, zero); | |
| 380 __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16); | |
| 381 __m128i mul_lo = _mm_mullo_epi16(src16, coeff16); | |
| 382 // [32] a0 b0 g0 r0 | |
| 383 __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi); | |
| 384 accum0 = _mm_add_epi32(accum0, t); | |
| 385 // [32] a1 b1 g1 r1 | |
| 386 t = _mm_unpackhi_epi16(mul_lo, mul_hi); | |
| 387 accum1 = _mm_add_epi32(accum1, t); | |
| 388 // [16] a3 b3 g3 r3 a2 b2 g2 r2 | |
| 389 src16 = _mm_unpackhi_epi8(src8, zero); | |
| 390 mul_hi = _mm_mulhi_epi16(src16, coeff16); | |
| 391 mul_lo = _mm_mullo_epi16(src16, coeff16); | |
| 392 // [32] a2 b2 g2 r2 | |
| 393 t = _mm_unpacklo_epi16(mul_lo, mul_hi); | |
| 394 accum2 = _mm_add_epi32(accum2, t); | |
| 395 } | |
| 396 | |
| 397 accum0 = _mm_srai_epi32(accum0, SkConvolutionFilter1D::kShiftBits); | |
| 398 accum1 = _mm_srai_epi32(accum1, SkConvolutionFilter1D::kShiftBits); | |
| 399 accum2 = _mm_srai_epi32(accum2, SkConvolutionFilter1D::kShiftBits); | |
| 400 // [16] a1 b1 g1 r1 a0 b0 g0 r0 | |
| 401 accum0 = _mm_packs_epi32(accum0, accum1); | |
| 402 // [16] a3 b3 g3 r3 a2 b2 g2 r2 | |
| 403 accum2 = _mm_packs_epi32(accum2, zero); | |
| 404 // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 | |
| 405 accum0 = _mm_packus_epi16(accum0, accum2); | |
| 406 if (has_alpha) { | |
| 407 // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0 | |
| 408 __m128i a = _mm_srli_epi32(accum0, 8); | |
| 409 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 | |
| 410 __m128i b = _mm_max_epu8(a, accum0); // Max of r and g. | |
| 411 // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0 | |
| 412 a = _mm_srli_epi32(accum0, 16); | |
| 413 // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0 | |
| 414 b = _mm_max_epu8(a, b); // Max of r and g and b. | |
| 415 // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00 | |
| 416 b = _mm_slli_epi32(b, 24); | |
| 417 accum0 = _mm_max_epu8(b, accum0); | |
| 418 } else { | |
| 419 __m128i mask = _mm_set1_epi32(0xff000000); | |
| 420 accum0 = _mm_or_si128(accum0, mask); | |
| 421 } | |
| 422 | |
| 423 for (int out_x = width; out_x < pixel_width; out_x++) { | |
| 424 *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum0); | |
| 425 accum0 = _mm_srli_si128(accum0, 4); | |
| 426 out_row += 4; | |
| 427 } | |
| 428 } | |
| 429 } | |
| 430 | |
| 431 void convolveVertically_SSE2(const SkConvolutionFilter1D::ConvolutionFixed* filt
er_values, | |
| 432 int filter_length, | |
| 433 unsigned char* const* source_data_rows, | |
| 434 int pixel_width, | |
| 435 unsigned char* out_row, | |
| 436 bool has_alpha) { | |
| 437 if (has_alpha) { | |
| 438 convolveVertically_SSE2<true>(filter_values, | |
| 439 filter_length, | |
| 440 source_data_rows, | |
| 441 pixel_width, | |
| 442 out_row); | |
| 443 } else { | |
| 444 convolveVertically_SSE2<false>(filter_values, | |
| 445 filter_length, | |
| 446 source_data_rows, | |
| 447 pixel_width, | |
| 448 out_row); | |
| 449 } | |
| 450 } | |
| OLD | NEW |
|---|
Chromium Code Reviews
Issue 2500113004:
Port convolve functions to SkOpts (Closed)
