1 // Copyright 2011 Google Inc. All Rights Reserved. 2 // 3 // This code is licensed under the same terms as WebM: 4 // Software License Agreement: http://www.webmproject.org/license/software/ 5 // Additional IP Rights Grant: http://www.webmproject.org/license/additional/ 6 // ----------------------------------------------------------------------------- 7 // 8 // SSE2 version of speed-critical encoding functions. 9 // 10 // Author: Christian Duvivier (cduvivier (at) google.com) 11 12 #include "./dsp.h" 13 14 #if defined(__cplusplus) || defined(c_plusplus) 15 extern "C" { 16 #endif 17 18 #if defined(WEBP_USE_SSE2) 19 #include <stdlib.h> // for abs() 20 #include <emmintrin.h> 21 22 #include "../enc/vp8enci.h" 23 24 //------------------------------------------------------------------------------ 25 // Quite useful macro for debugging. Left here for convenience. 26 27 #if 0 28 #include <stdio.h> 29 static void PrintReg(const __m128i r, const char* const name, int size) { 30 int n; 31 union { 32 __m128i r; 33 uint8_t i8[16]; 34 uint16_t i16[8]; 35 uint32_t i32[4]; 36 uint64_t i64[2]; 37 } tmp; 38 tmp.r = r; 39 printf("%s\t: ", name); 40 if (size == 8) { 41 for (n = 0; n < 16; ++n) printf("%.2x ", tmp.i8[n]); 42 } else if (size == 16) { 43 for (n = 0; n < 8; ++n) printf("%.4x ", tmp.i16[n]); 44 } else if (size == 32) { 45 for (n = 0; n < 4; ++n) printf("%.8x ", tmp.i32[n]); 46 } else { 47 for (n = 0; n < 2; ++n) printf("%.16lx ", tmp.i64[n]); 48 } 49 printf("\n"); 50 } 51 #endif 52 53 //------------------------------------------------------------------------------ 54 // Compute susceptibility based on DCT-coeff histograms: 55 // the higher, the "easier" the macroblock is to compress. 56 57 static void CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred, 58 int start_block, int end_block, 59 VP8Histogram* const histo) { 60 const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); 61 int j; 62 for (j = start_block; j < end_block; ++j) { 63 int16_t out[16]; 64 int k; 65 66 VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); 67 68 // Convert coefficients to bin (within out[]). 69 { 70 // Load. 71 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); 72 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); 73 // sign(out) = out >> 15 (0x0000 if positive, 0xffff if negative) 74 const __m128i sign0 = _mm_srai_epi16(out0, 15); 75 const __m128i sign1 = _mm_srai_epi16(out1, 15); 76 // abs(out) = (out ^ sign) - sign 77 const __m128i xor0 = _mm_xor_si128(out0, sign0); 78 const __m128i xor1 = _mm_xor_si128(out1, sign1); 79 const __m128i abs0 = _mm_sub_epi16(xor0, sign0); 80 const __m128i abs1 = _mm_sub_epi16(xor1, sign1); 81 // v = abs(out) >> 3 82 const __m128i v0 = _mm_srai_epi16(abs0, 3); 83 const __m128i v1 = _mm_srai_epi16(abs1, 3); 84 // bin = min(v, MAX_COEFF_THRESH) 85 const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); 86 const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); 87 // Store. 88 _mm_storeu_si128((__m128i*)&out[0], bin0); 89 _mm_storeu_si128((__m128i*)&out[8], bin1); 90 } 91 92 // Convert coefficients to bin. 93 for (k = 0; k < 16; ++k) { 94 histo->distribution[out[k]]++; 95 } 96 } 97 } 98 99 //------------------------------------------------------------------------------ 100 // Transforms (Paragraph 14.4) 101 102 // Does one or two inverse transforms. 103 static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst, 104 int do_two) { 105 // This implementation makes use of 16-bit fixed point versions of two 106 // multiply constants: 107 // K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16 108 // K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16 109 // 110 // To be able to use signed 16-bit integers, we use the following trick to 111 // have constants within range: 112 // - Associated constants are obtained by subtracting the 16-bit fixed point 113 // version of one: 114 // k = K - (1 << 16) => K = k + (1 << 16) 115 // K1 = 85267 => k1 = 20091 116 // K2 = 35468 => k2 = -30068 117 // - The multiplication of a variable by a constant become the sum of the 118 // variable and the multiplication of that variable by the associated 119 // constant: 120 // (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x 121 const __m128i k1 = _mm_set1_epi16(20091); 122 const __m128i k2 = _mm_set1_epi16(-30068); 123 __m128i T0, T1, T2, T3; 124 125 // Load and concatenate the transform coefficients (we'll do two inverse 126 // transforms in parallel). In the case of only one inverse transform, the 127 // second half of the vectors will just contain random value we'll never 128 // use nor store. 129 __m128i in0, in1, in2, in3; 130 { 131 in0 = _mm_loadl_epi64((__m128i*)&in[0]); 132 in1 = _mm_loadl_epi64((__m128i*)&in[4]); 133 in2 = _mm_loadl_epi64((__m128i*)&in[8]); 134 in3 = _mm_loadl_epi64((__m128i*)&in[12]); 135 // a00 a10 a20 a30 x x x x 136 // a01 a11 a21 a31 x x x x 137 // a02 a12 a22 a32 x x x x 138 // a03 a13 a23 a33 x x x x 139 if (do_two) { 140 const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]); 141 const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]); 142 const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]); 143 const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]); 144 in0 = _mm_unpacklo_epi64(in0, inB0); 145 in1 = _mm_unpacklo_epi64(in1, inB1); 146 in2 = _mm_unpacklo_epi64(in2, inB2); 147 in3 = _mm_unpacklo_epi64(in3, inB3); 148 // a00 a10 a20 a30 b00 b10 b20 b30 149 // a01 a11 a21 a31 b01 b11 b21 b31 150 // a02 a12 a22 a32 b02 b12 b22 b32 151 // a03 a13 a23 a33 b03 b13 b23 b33 152 } 153 } 154 155 // Vertical pass and subsequent transpose. 156 { 157 // First pass, c and d calculations are longer because of the "trick" 158 // multiplications. 159 const __m128i a = _mm_add_epi16(in0, in2); 160 const __m128i b = _mm_sub_epi16(in0, in2); 161 // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3 162 const __m128i c1 = _mm_mulhi_epi16(in1, k2); 163 const __m128i c2 = _mm_mulhi_epi16(in3, k1); 164 const __m128i c3 = _mm_sub_epi16(in1, in3); 165 const __m128i c4 = _mm_sub_epi16(c1, c2); 166 const __m128i c = _mm_add_epi16(c3, c4); 167 // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3 168 const __m128i d1 = _mm_mulhi_epi16(in1, k1); 169 const __m128i d2 = _mm_mulhi_epi16(in3, k2); 170 const __m128i d3 = _mm_add_epi16(in1, in3); 171 const __m128i d4 = _mm_add_epi16(d1, d2); 172 const __m128i d = _mm_add_epi16(d3, d4); 173 174 // Second pass. 175 const __m128i tmp0 = _mm_add_epi16(a, d); 176 const __m128i tmp1 = _mm_add_epi16(b, c); 177 const __m128i tmp2 = _mm_sub_epi16(b, c); 178 const __m128i tmp3 = _mm_sub_epi16(a, d); 179 180 // Transpose the two 4x4. 181 // a00 a01 a02 a03 b00 b01 b02 b03 182 // a10 a11 a12 a13 b10 b11 b12 b13 183 // a20 a21 a22 a23 b20 b21 b22 b23 184 // a30 a31 a32 a33 b30 b31 b32 b33 185 const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1); 186 const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3); 187 const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1); 188 const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3); 189 // a00 a10 a01 a11 a02 a12 a03 a13 190 // a20 a30 a21 a31 a22 a32 a23 a33 191 // b00 b10 b01 b11 b02 b12 b03 b13 192 // b20 b30 b21 b31 b22 b32 b23 b33 193 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); 194 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); 195 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); 196 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); 197 // a00 a10 a20 a30 a01 a11 a21 a31 198 // b00 b10 b20 b30 b01 b11 b21 b31 199 // a02 a12 a22 a32 a03 a13 a23 a33 200 // b02 b12 a22 b32 b03 b13 b23 b33 201 T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); 202 T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); 203 T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); 204 T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); 205 // a00 a10 a20 a30 b00 b10 b20 b30 206 // a01 a11 a21 a31 b01 b11 b21 b31 207 // a02 a12 a22 a32 b02 b12 b22 b32 208 // a03 a13 a23 a33 b03 b13 b23 b33 209 } 210 211 // Horizontal pass and subsequent transpose. 212 { 213 // First pass, c and d calculations are longer because of the "trick" 214 // multiplications. 215 const __m128i four = _mm_set1_epi16(4); 216 const __m128i dc = _mm_add_epi16(T0, four); 217 const __m128i a = _mm_add_epi16(dc, T2); 218 const __m128i b = _mm_sub_epi16(dc, T2); 219 // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3 220 const __m128i c1 = _mm_mulhi_epi16(T1, k2); 221 const __m128i c2 = _mm_mulhi_epi16(T3, k1); 222 const __m128i c3 = _mm_sub_epi16(T1, T3); 223 const __m128i c4 = _mm_sub_epi16(c1, c2); 224 const __m128i c = _mm_add_epi16(c3, c4); 225 // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3 226 const __m128i d1 = _mm_mulhi_epi16(T1, k1); 227 const __m128i d2 = _mm_mulhi_epi16(T3, k2); 228 const __m128i d3 = _mm_add_epi16(T1, T3); 229 const __m128i d4 = _mm_add_epi16(d1, d2); 230 const __m128i d = _mm_add_epi16(d3, d4); 231 232 // Second pass. 233 const __m128i tmp0 = _mm_add_epi16(a, d); 234 const __m128i tmp1 = _mm_add_epi16(b, c); 235 const __m128i tmp2 = _mm_sub_epi16(b, c); 236 const __m128i tmp3 = _mm_sub_epi16(a, d); 237 const __m128i shifted0 = _mm_srai_epi16(tmp0, 3); 238 const __m128i shifted1 = _mm_srai_epi16(tmp1, 3); 239 const __m128i shifted2 = _mm_srai_epi16(tmp2, 3); 240 const __m128i shifted3 = _mm_srai_epi16(tmp3, 3); 241 242 // Transpose the two 4x4. 243 // a00 a01 a02 a03 b00 b01 b02 b03 244 // a10 a11 a12 a13 b10 b11 b12 b13 245 // a20 a21 a22 a23 b20 b21 b22 b23 246 // a30 a31 a32 a33 b30 b31 b32 b33 247 const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1); 248 const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3); 249 const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1); 250 const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3); 251 // a00 a10 a01 a11 a02 a12 a03 a13 252 // a20 a30 a21 a31 a22 a32 a23 a33 253 // b00 b10 b01 b11 b02 b12 b03 b13 254 // b20 b30 b21 b31 b22 b32 b23 b33 255 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); 256 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); 257 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); 258 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); 259 // a00 a10 a20 a30 a01 a11 a21 a31 260 // b00 b10 b20 b30 b01 b11 b21 b31 261 // a02 a12 a22 a32 a03 a13 a23 a33 262 // b02 b12 a22 b32 b03 b13 b23 b33 263 T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); 264 T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); 265 T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); 266 T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); 267 // a00 a10 a20 a30 b00 b10 b20 b30 268 // a01 a11 a21 a31 b01 b11 b21 b31 269 // a02 a12 a22 a32 b02 b12 b22 b32 270 // a03 a13 a23 a33 b03 b13 b23 b33 271 } 272 273 // Add inverse transform to 'ref' and store. 274 { 275 const __m128i zero = _mm_setzero_si128(); 276 // Load the reference(s). 277 __m128i ref0, ref1, ref2, ref3; 278 if (do_two) { 279 // Load eight bytes/pixels per line. 280 ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]); 281 ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]); 282 ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]); 283 ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]); 284 } else { 285 // Load four bytes/pixels per line. 286 ref0 = _mm_cvtsi32_si128(*(int*)&ref[0 * BPS]); 287 ref1 = _mm_cvtsi32_si128(*(int*)&ref[1 * BPS]); 288 ref2 = _mm_cvtsi32_si128(*(int*)&ref[2 * BPS]); 289 ref3 = _mm_cvtsi32_si128(*(int*)&ref[3 * BPS]); 290 } 291 // Convert to 16b. 292 ref0 = _mm_unpacklo_epi8(ref0, zero); 293 ref1 = _mm_unpacklo_epi8(ref1, zero); 294 ref2 = _mm_unpacklo_epi8(ref2, zero); 295 ref3 = _mm_unpacklo_epi8(ref3, zero); 296 // Add the inverse transform(s). 297 ref0 = _mm_add_epi16(ref0, T0); 298 ref1 = _mm_add_epi16(ref1, T1); 299 ref2 = _mm_add_epi16(ref2, T2); 300 ref3 = _mm_add_epi16(ref3, T3); 301 // Unsigned saturate to 8b. 302 ref0 = _mm_packus_epi16(ref0, ref0); 303 ref1 = _mm_packus_epi16(ref1, ref1); 304 ref2 = _mm_packus_epi16(ref2, ref2); 305 ref3 = _mm_packus_epi16(ref3, ref3); 306 // Store the results. 307 if (do_two) { 308 // Store eight bytes/pixels per line. 309 _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0); 310 _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1); 311 _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2); 312 _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3); 313 } else { 314 // Store four bytes/pixels per line. 315 *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(ref0); 316 *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(ref1); 317 *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(ref2); 318 *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(ref3); 319 } 320 } 321 } 322 323 static void FTransformSSE2(const uint8_t* src, const uint8_t* ref, 324 int16_t* out) { 325 const __m128i zero = _mm_setzero_si128(); 326 const __m128i seven = _mm_set1_epi16(7); 327 const __m128i k937 = _mm_set1_epi32(937); 328 const __m128i k1812 = _mm_set1_epi32(1812); 329 const __m128i k51000 = _mm_set1_epi32(51000); 330 const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16)); 331 const __m128i k5352_2217 = _mm_set_epi16(5352, 2217, 5352, 2217, 332 5352, 2217, 5352, 2217); 333 const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352, 334 2217, -5352, 2217, -5352); 335 const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8); 336 const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8); 337 const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352, 338 2217, 5352, 2217, 5352); 339 const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217, 340 -5352, 2217, -5352, 2217); 341 __m128i v01, v32; 342 343 344 // Difference between src and ref and initial transpose. 345 { 346 // Load src and convert to 16b. 347 const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]); 348 const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]); 349 const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]); 350 const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]); 351 const __m128i src_0 = _mm_unpacklo_epi8(src0, zero); 352 const __m128i src_1 = _mm_unpacklo_epi8(src1, zero); 353 const __m128i src_2 = _mm_unpacklo_epi8(src2, zero); 354 const __m128i src_3 = _mm_unpacklo_epi8(src3, zero); 355 // Load ref and convert to 16b. 356 const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]); 357 const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]); 358 const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]); 359 const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]); 360 const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero); 361 const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero); 362 const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero); 363 const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero); 364 // Compute difference. -> 00 01 02 03 00 00 00 00 365 const __m128i diff0 = _mm_sub_epi16(src_0, ref_0); 366 const __m128i diff1 = _mm_sub_epi16(src_1, ref_1); 367 const __m128i diff2 = _mm_sub_epi16(src_2, ref_2); 368 const __m128i diff3 = _mm_sub_epi16(src_3, ref_3); 369 370 371 // Unpack and shuffle 372 // 00 01 02 03 0 0 0 0 373 // 10 11 12 13 0 0 0 0 374 // 20 21 22 23 0 0 0 0 375 // 30 31 32 33 0 0 0 0 376 const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1); 377 const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3); 378 // 00 01 10 11 02 03 12 13 379 // 20 21 30 31 22 23 32 33 380 const __m128i shuf01_p = 381 _mm_shufflehi_epi16(shuf01, _MM_SHUFFLE(2, 3, 0, 1)); 382 const __m128i shuf23_p = 383 _mm_shufflehi_epi16(shuf23, _MM_SHUFFLE(2, 3, 0, 1)); 384 // 00 01 10 11 03 02 13 12 385 // 20 21 30 31 23 22 33 32 386 const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p); 387 const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p); 388 // 00 01 10 11 20 21 30 31 389 // 03 02 13 12 23 22 33 32 390 const __m128i a01 = _mm_add_epi16(s01, s32); 391 const __m128i a32 = _mm_sub_epi16(s01, s32); 392 // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ] 393 // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ] 394 395 const __m128i tmp0 = _mm_madd_epi16(a01, k88p); // [ (a0 + a1) << 3, ... ] 396 const __m128i tmp2 = _mm_madd_epi16(a01, k88m); // [ (a0 - a1) << 3, ... ] 397 const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p); 398 const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m); 399 const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812); 400 const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937); 401 const __m128i tmp1 = _mm_srai_epi32(tmp1_2, 9); 402 const __m128i tmp3 = _mm_srai_epi32(tmp3_2, 9); 403 const __m128i s03 = _mm_packs_epi32(tmp0, tmp2); 404 const __m128i s12 = _mm_packs_epi32(tmp1, tmp3); 405 const __m128i s_lo = _mm_unpacklo_epi16(s03, s12); // 0 1 0 1 0 1... 406 const __m128i s_hi = _mm_unpackhi_epi16(s03, s12); // 2 3 2 3 2 3 407 const __m128i v23 = _mm_unpackhi_epi32(s_lo, s_hi); 408 v01 = _mm_unpacklo_epi32(s_lo, s_hi); 409 v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2)); // 3 2 3 2 3 2.. 410 } 411 412 // Second pass 413 { 414 // Same operations are done on the (0,3) and (1,2) pairs. 415 // a0 = v0 + v3 416 // a1 = v1 + v2 417 // a3 = v0 - v3 418 // a2 = v1 - v2 419 const __m128i a01 = _mm_add_epi16(v01, v32); 420 const __m128i a32 = _mm_sub_epi16(v01, v32); 421 const __m128i a11 = _mm_unpackhi_epi64(a01, a01); 422 const __m128i a22 = _mm_unpackhi_epi64(a32, a32); 423 const __m128i a01_plus_7 = _mm_add_epi16(a01, seven); 424 425 // d0 = (a0 + a1 + 7) >> 4; 426 // d2 = (a0 - a1 + 7) >> 4; 427 const __m128i c0 = _mm_add_epi16(a01_plus_7, a11); 428 const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11); 429 const __m128i d0 = _mm_srai_epi16(c0, 4); 430 const __m128i d2 = _mm_srai_epi16(c2, 4); 431 432 // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16) 433 // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16) 434 const __m128i b23 = _mm_unpacklo_epi16(a22, a32); 435 const __m128i c1 = _mm_madd_epi16(b23, k5352_2217); 436 const __m128i c3 = _mm_madd_epi16(b23, k2217_5352); 437 const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one); 438 const __m128i d3 = _mm_add_epi32(c3, k51000); 439 const __m128i e1 = _mm_srai_epi32(d1, 16); 440 const __m128i e3 = _mm_srai_epi32(d3, 16); 441 const __m128i f1 = _mm_packs_epi32(e1, e1); 442 const __m128i f3 = _mm_packs_epi32(e3, e3); 443 // f1 = f1 + (a3 != 0); 444 // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the 445 // desired (0, 1), we add one earlier through k12000_plus_one. 446 // -> f1 = f1 + 1 - (a3 == 0) 447 const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero)); 448 449 _mm_storel_epi64((__m128i*)&out[ 0], d0); 450 _mm_storel_epi64((__m128i*)&out[ 4], g1); 451 _mm_storel_epi64((__m128i*)&out[ 8], d2); 452 _mm_storel_epi64((__m128i*)&out[12], f3); 453 } 454 } 455 456 //------------------------------------------------------------------------------ 457 // Metric 458 459 static int SSE_Nx4SSE2(const uint8_t* a, const uint8_t* b, 460 int num_quads, int do_16) { 461 const __m128i zero = _mm_setzero_si128(); 462 __m128i sum1 = zero; 463 __m128i sum2 = zero; 464 465 while (num_quads-- > 0) { 466 // Note: for the !do_16 case, we read 16 pixels instead of 8 but that's ok, 467 // thanks to buffer over-allocation to that effect. 468 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]); 469 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]); 470 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]); 471 const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]); 472 const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]); 473 const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]); 474 const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]); 475 const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]); 476 477 // compute clip0(a-b) and clip0(b-a) 478 const __m128i a0p = _mm_subs_epu8(a0, b0); 479 const __m128i a0m = _mm_subs_epu8(b0, a0); 480 const __m128i a1p = _mm_subs_epu8(a1, b1); 481 const __m128i a1m = _mm_subs_epu8(b1, a1); 482 const __m128i a2p = _mm_subs_epu8(a2, b2); 483 const __m128i a2m = _mm_subs_epu8(b2, a2); 484 const __m128i a3p = _mm_subs_epu8(a3, b3); 485 const __m128i a3m = _mm_subs_epu8(b3, a3); 486 487 // compute |a-b| with 8b arithmetic as clip0(a-b) | clip0(b-a) 488 const __m128i diff0 = _mm_or_si128(a0p, a0m); 489 const __m128i diff1 = _mm_or_si128(a1p, a1m); 490 const __m128i diff2 = _mm_or_si128(a2p, a2m); 491 const __m128i diff3 = _mm_or_si128(a3p, a3m); 492 493 // unpack (only four operations, instead of eight) 494 const __m128i low0 = _mm_unpacklo_epi8(diff0, zero); 495 const __m128i low1 = _mm_unpacklo_epi8(diff1, zero); 496 const __m128i low2 = _mm_unpacklo_epi8(diff2, zero); 497 const __m128i low3 = _mm_unpacklo_epi8(diff3, zero); 498 499 // multiply with self 500 const __m128i low_madd0 = _mm_madd_epi16(low0, low0); 501 const __m128i low_madd1 = _mm_madd_epi16(low1, low1); 502 const __m128i low_madd2 = _mm_madd_epi16(low2, low2); 503 const __m128i low_madd3 = _mm_madd_epi16(low3, low3); 504 505 // collect in a cascading way 506 const __m128i low_sum0 = _mm_add_epi32(low_madd0, low_madd1); 507 const __m128i low_sum1 = _mm_add_epi32(low_madd2, low_madd3); 508 sum1 = _mm_add_epi32(sum1, low_sum0); 509 sum2 = _mm_add_epi32(sum2, low_sum1); 510 511 if (do_16) { // if necessary, process the higher 8 bytes similarly 512 const __m128i hi0 = _mm_unpackhi_epi8(diff0, zero); 513 const __m128i hi1 = _mm_unpackhi_epi8(diff1, zero); 514 const __m128i hi2 = _mm_unpackhi_epi8(diff2, zero); 515 const __m128i hi3 = _mm_unpackhi_epi8(diff3, zero); 516 517 const __m128i hi_madd0 = _mm_madd_epi16(hi0, hi0); 518 const __m128i hi_madd1 = _mm_madd_epi16(hi1, hi1); 519 const __m128i hi_madd2 = _mm_madd_epi16(hi2, hi2); 520 const __m128i hi_madd3 = _mm_madd_epi16(hi3, hi3); 521 const __m128i hi_sum0 = _mm_add_epi32(hi_madd0, hi_madd1); 522 const __m128i hi_sum1 = _mm_add_epi32(hi_madd2, hi_madd3); 523 sum1 = _mm_add_epi32(sum1, hi_sum0); 524 sum2 = _mm_add_epi32(sum2, hi_sum1); 525 } 526 a += 4 * BPS; 527 b += 4 * BPS; 528 } 529 { 530 int32_t tmp[4]; 531 const __m128i sum = _mm_add_epi32(sum1, sum2); 532 _mm_storeu_si128((__m128i*)tmp, sum); 533 return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); 534 } 535 } 536 537 static int SSE16x16SSE2(const uint8_t* a, const uint8_t* b) { 538 return SSE_Nx4SSE2(a, b, 4, 1); 539 } 540 541 static int SSE16x8SSE2(const uint8_t* a, const uint8_t* b) { 542 return SSE_Nx4SSE2(a, b, 2, 1); 543 } 544 545 static int SSE8x8SSE2(const uint8_t* a, const uint8_t* b) { 546 return SSE_Nx4SSE2(a, b, 2, 0); 547 } 548 549 static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) { 550 const __m128i zero = _mm_setzero_si128(); 551 552 // Load values. Note that we read 8 pixels instead of 4, 553 // but the a/b buffers are over-allocated to that effect. 554 const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]); 555 const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]); 556 const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]); 557 const __m128i a3 = _mm_loadl_epi64((__m128i*)&a[BPS * 3]); 558 const __m128i b0 = _mm_loadl_epi64((__m128i*)&b[BPS * 0]); 559 const __m128i b1 = _mm_loadl_epi64((__m128i*)&b[BPS * 1]); 560 const __m128i b2 = _mm_loadl_epi64((__m128i*)&b[BPS * 2]); 561 const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]); 562 563 // Combine pair of lines and convert to 16b. 564 const __m128i a01 = _mm_unpacklo_epi32(a0, a1); 565 const __m128i a23 = _mm_unpacklo_epi32(a2, a3); 566 const __m128i b01 = _mm_unpacklo_epi32(b0, b1); 567 const __m128i b23 = _mm_unpacklo_epi32(b2, b3); 568 const __m128i a01s = _mm_unpacklo_epi8(a01, zero); 569 const __m128i a23s = _mm_unpacklo_epi8(a23, zero); 570 const __m128i b01s = _mm_unpacklo_epi8(b01, zero); 571 const __m128i b23s = _mm_unpacklo_epi8(b23, zero); 572 573 // Compute differences; (a-b)^2 = (abs(a-b))^2 = (sat8(a-b) + sat8(b-a))^2 574 // TODO(cduvivier): Dissassemble and figure out why this is fastest. We don't 575 // need absolute values, there is no need to do calculation 576 // in 8bit as we are already in 16bit, ... Yet this is what 577 // benchmarks the fastest! 578 const __m128i d0 = _mm_subs_epu8(a01s, b01s); 579 const __m128i d1 = _mm_subs_epu8(b01s, a01s); 580 const __m128i d2 = _mm_subs_epu8(a23s, b23s); 581 const __m128i d3 = _mm_subs_epu8(b23s, a23s); 582 583 // Square and add them all together. 584 const __m128i madd0 = _mm_madd_epi16(d0, d0); 585 const __m128i madd1 = _mm_madd_epi16(d1, d1); 586 const __m128i madd2 = _mm_madd_epi16(d2, d2); 587 const __m128i madd3 = _mm_madd_epi16(d3, d3); 588 const __m128i sum0 = _mm_add_epi32(madd0, madd1); 589 const __m128i sum1 = _mm_add_epi32(madd2, madd3); 590 const __m128i sum2 = _mm_add_epi32(sum0, sum1); 591 592 int32_t tmp[4]; 593 _mm_storeu_si128((__m128i*)tmp, sum2); 594 return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); 595 } 596 597 //------------------------------------------------------------------------------ 598 // Texture distortion 599 // 600 // We try to match the spectral content (weighted) between source and 601 // reconstructed samples. 602 603 // Hadamard transform 604 // Returns the difference between the weighted sum of the absolute value of 605 // transformed coefficients. 606 static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB, 607 const uint16_t* const w) { 608 int32_t sum[4]; 609 __m128i tmp_0, tmp_1, tmp_2, tmp_3; 610 const __m128i zero = _mm_setzero_si128(); 611 612 // Load, combine and tranpose inputs. 613 { 614 const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]); 615 const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]); 616 const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]); 617 const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]); 618 const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]); 619 const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]); 620 const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]); 621 const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]); 622 623 // Combine inA and inB (we'll do two transforms in parallel). 624 const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0); 625 const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1); 626 const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2); 627 const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3); 628 // a00 b00 a01 b01 a02 b03 a03 b03 0 0 0 0 0 0 0 0 629 // a10 b10 a11 b11 a12 b12 a13 b13 0 0 0 0 0 0 0 0 630 // a20 b20 a21 b21 a22 b22 a23 b23 0 0 0 0 0 0 0 0 631 // a30 b30 a31 b31 a32 b32 a33 b33 0 0 0 0 0 0 0 0 632 633 // Transpose the two 4x4, discarding the filling zeroes. 634 const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2); 635 const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3); 636 // a00 a20 b00 b20 a01 a21 b01 b21 a02 a22 b02 b22 a03 a23 b03 b23 637 // a10 a30 b10 b30 a11 a31 b11 b31 a12 a32 b12 b32 a13 a33 b13 b33 638 const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1); 639 const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1); 640 // a00 a10 a20 a30 b00 b10 b20 b30 a01 a11 a21 a31 b01 b11 b21 b31 641 // a02 a12 a22 a32 b02 b12 b22 b32 a03 a13 a23 a33 b03 b13 b23 b33 642 643 // Convert to 16b. 644 tmp_0 = _mm_unpacklo_epi8(transpose1_0, zero); 645 tmp_1 = _mm_unpackhi_epi8(transpose1_0, zero); 646 tmp_2 = _mm_unpacklo_epi8(transpose1_1, zero); 647 tmp_3 = _mm_unpackhi_epi8(transpose1_1, zero); 648 // a00 a10 a20 a30 b00 b10 b20 b30 649 // a01 a11 a21 a31 b01 b11 b21 b31 650 // a02 a12 a22 a32 b02 b12 b22 b32 651 // a03 a13 a23 a33 b03 b13 b23 b33 652 } 653 654 // Horizontal pass and subsequent transpose. 655 { 656 // Calculate a and b (two 4x4 at once). 657 const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); 658 const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); 659 const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); 660 const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); 661 const __m128i b0 = _mm_add_epi16(a0, a1); 662 const __m128i b1 = _mm_add_epi16(a3, a2); 663 const __m128i b2 = _mm_sub_epi16(a3, a2); 664 const __m128i b3 = _mm_sub_epi16(a0, a1); 665 // a00 a01 a02 a03 b00 b01 b02 b03 666 // a10 a11 a12 a13 b10 b11 b12 b13 667 // a20 a21 a22 a23 b20 b21 b22 b23 668 // a30 a31 a32 a33 b30 b31 b32 b33 669 670 // Transpose the two 4x4. 671 const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1); 672 const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3); 673 const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1); 674 const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3); 675 // a00 a10 a01 a11 a02 a12 a03 a13 676 // a20 a30 a21 a31 a22 a32 a23 a33 677 // b00 b10 b01 b11 b02 b12 b03 b13 678 // b20 b30 b21 b31 b22 b32 b23 b33 679 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); 680 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); 681 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); 682 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); 683 // a00 a10 a20 a30 a01 a11 a21 a31 684 // b00 b10 b20 b30 b01 b11 b21 b31 685 // a02 a12 a22 a32 a03 a13 a23 a33 686 // b02 b12 a22 b32 b03 b13 b23 b33 687 tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); 688 tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); 689 tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); 690 tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); 691 // a00 a10 a20 a30 b00 b10 b20 b30 692 // a01 a11 a21 a31 b01 b11 b21 b31 693 // a02 a12 a22 a32 b02 b12 b22 b32 694 // a03 a13 a23 a33 b03 b13 b23 b33 695 } 696 697 // Vertical pass and difference of weighted sums. 698 { 699 // Load all inputs. 700 // TODO(cduvivier): Make variable declarations and allocations aligned so 701 // we can use _mm_load_si128 instead of _mm_loadu_si128. 702 const __m128i w_0 = _mm_loadu_si128((__m128i*)&w[0]); 703 const __m128i w_8 = _mm_loadu_si128((__m128i*)&w[8]); 704 705 // Calculate a and b (two 4x4 at once). 706 const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); 707 const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); 708 const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); 709 const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); 710 const __m128i b0 = _mm_add_epi16(a0, a1); 711 const __m128i b1 = _mm_add_epi16(a3, a2); 712 const __m128i b2 = _mm_sub_epi16(a3, a2); 713 const __m128i b3 = _mm_sub_epi16(a0, a1); 714 715 // Separate the transforms of inA and inB. 716 __m128i A_b0 = _mm_unpacklo_epi64(b0, b1); 717 __m128i A_b2 = _mm_unpacklo_epi64(b2, b3); 718 __m128i B_b0 = _mm_unpackhi_epi64(b0, b1); 719 __m128i B_b2 = _mm_unpackhi_epi64(b2, b3); 720 721 { 722 // sign(b) = b >> 15 (0x0000 if positive, 0xffff if negative) 723 const __m128i sign_A_b0 = _mm_srai_epi16(A_b0, 15); 724 const __m128i sign_A_b2 = _mm_srai_epi16(A_b2, 15); 725 const __m128i sign_B_b0 = _mm_srai_epi16(B_b0, 15); 726 const __m128i sign_B_b2 = _mm_srai_epi16(B_b2, 15); 727 728 // b = abs(b) = (b ^ sign) - sign 729 A_b0 = _mm_xor_si128(A_b0, sign_A_b0); 730 A_b2 = _mm_xor_si128(A_b2, sign_A_b2); 731 B_b0 = _mm_xor_si128(B_b0, sign_B_b0); 732 B_b2 = _mm_xor_si128(B_b2, sign_B_b2); 733 A_b0 = _mm_sub_epi16(A_b0, sign_A_b0); 734 A_b2 = _mm_sub_epi16(A_b2, sign_A_b2); 735 B_b0 = _mm_sub_epi16(B_b0, sign_B_b0); 736 B_b2 = _mm_sub_epi16(B_b2, sign_B_b2); 737 } 738 739 // weighted sums 740 A_b0 = _mm_madd_epi16(A_b0, w_0); 741 A_b2 = _mm_madd_epi16(A_b2, w_8); 742 B_b0 = _mm_madd_epi16(B_b0, w_0); 743 B_b2 = _mm_madd_epi16(B_b2, w_8); 744 A_b0 = _mm_add_epi32(A_b0, A_b2); 745 B_b0 = _mm_add_epi32(B_b0, B_b2); 746 747 // difference of weighted sums 748 A_b0 = _mm_sub_epi32(A_b0, B_b0); 749 _mm_storeu_si128((__m128i*)&sum[0], A_b0); 750 } 751 return sum[0] + sum[1] + sum[2] + sum[3]; 752 } 753 754 static int Disto4x4SSE2(const uint8_t* const a, const uint8_t* const b, 755 const uint16_t* const w) { 756 const int diff_sum = TTransformSSE2(a, b, w); 757 return abs(diff_sum) >> 5; 758 } 759 760 static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b, 761 const uint16_t* const w) { 762 int D = 0; 763 int x, y; 764 for (y = 0; y < 16 * BPS; y += 4 * BPS) { 765 for (x = 0; x < 16; x += 4) { 766 D += Disto4x4SSE2(a + x + y, b + x + y, w); 767 } 768 } 769 return D; 770 } 771 772 //------------------------------------------------------------------------------ 773 // Quantization 774 // 775 776 // Simple quantization 777 static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16], 778 int n, const VP8Matrix* const mtx) { 779 const __m128i max_coeff_2047 = _mm_set1_epi16(2047); 780 const __m128i zero = _mm_setzero_si128(); 781 __m128i coeff0, coeff8; 782 __m128i out0, out8; 783 __m128i packed_out; 784 785 // Load all inputs. 786 // TODO(cduvivier): Make variable declarations and allocations aligned so that 787 // we can use _mm_load_si128 instead of _mm_loadu_si128. 788 __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]); 789 __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]); 790 const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]); 791 const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]); 792 const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]); 793 const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]); 794 const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]); 795 const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]); 796 const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]); 797 const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]); 798 const __m128i zthresh0 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[0]); 799 const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]); 800 801 // sign(in) = in >> 15 (0x0000 if positive, 0xffff if negative) 802 const __m128i sign0 = _mm_srai_epi16(in0, 15); 803 const __m128i sign8 = _mm_srai_epi16(in8, 15); 804 805 // coeff = abs(in) = (in ^ sign) - sign 806 coeff0 = _mm_xor_si128(in0, sign0); 807 coeff8 = _mm_xor_si128(in8, sign8); 808 coeff0 = _mm_sub_epi16(coeff0, sign0); 809 coeff8 = _mm_sub_epi16(coeff8, sign8); 810 811 // coeff = abs(in) + sharpen 812 coeff0 = _mm_add_epi16(coeff0, sharpen0); 813 coeff8 = _mm_add_epi16(coeff8, sharpen8); 814 815 // if (coeff > 2047) coeff = 2047 816 coeff0 = _mm_min_epi16(coeff0, max_coeff_2047); 817 coeff8 = _mm_min_epi16(coeff8, max_coeff_2047); 818 819 // out = (coeff * iQ + B) >> QFIX; 820 { 821 // doing calculations with 32b precision (QFIX=17) 822 // out = (coeff * iQ) 823 __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0); 824 __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0); 825 __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8); 826 __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8); 827 __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H); 828 __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H); 829 __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H); 830 __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H); 831 // expand bias from 16b to 32b 832 __m128i bias_00 = _mm_unpacklo_epi16(bias0, zero); 833 __m128i bias_04 = _mm_unpackhi_epi16(bias0, zero); 834 __m128i bias_08 = _mm_unpacklo_epi16(bias8, zero); 835 __m128i bias_12 = _mm_unpackhi_epi16(bias8, zero); 836 // out = (coeff * iQ + B) 837 out_00 = _mm_add_epi32(out_00, bias_00); 838 out_04 = _mm_add_epi32(out_04, bias_04); 839 out_08 = _mm_add_epi32(out_08, bias_08); 840 out_12 = _mm_add_epi32(out_12, bias_12); 841 // out = (coeff * iQ + B) >> QFIX; 842 out_00 = _mm_srai_epi32(out_00, QFIX); 843 out_04 = _mm_srai_epi32(out_04, QFIX); 844 out_08 = _mm_srai_epi32(out_08, QFIX); 845 out_12 = _mm_srai_epi32(out_12, QFIX); 846 // pack result as 16b 847 out0 = _mm_packs_epi32(out_00, out_04); 848 out8 = _mm_packs_epi32(out_08, out_12); 849 } 850 851 // get sign back (if (sign[j]) out_n = -out_n) 852 out0 = _mm_xor_si128(out0, sign0); 853 out8 = _mm_xor_si128(out8, sign8); 854 out0 = _mm_sub_epi16(out0, sign0); 855 out8 = _mm_sub_epi16(out8, sign8); 856 857 // in = out * Q 858 in0 = _mm_mullo_epi16(out0, q0); 859 in8 = _mm_mullo_epi16(out8, q8); 860 861 // if (coeff <= mtx->zthresh_) {in=0; out=0;} 862 { 863 __m128i cmp0 = _mm_cmpgt_epi16(coeff0, zthresh0); 864 __m128i cmp8 = _mm_cmpgt_epi16(coeff8, zthresh8); 865 in0 = _mm_and_si128(in0, cmp0); 866 in8 = _mm_and_si128(in8, cmp8); 867 _mm_storeu_si128((__m128i*)&in[0], in0); 868 _mm_storeu_si128((__m128i*)&in[8], in8); 869 out0 = _mm_and_si128(out0, cmp0); 870 out8 = _mm_and_si128(out8, cmp8); 871 } 872 873 // zigzag the output before storing it. 874 // 875 // The zigzag pattern can almost be reproduced with a small sequence of 876 // shuffles. After it, we only need to swap the 7th (ending up in third 877 // position instead of twelfth) and 8th values. 878 { 879 __m128i outZ0, outZ8; 880 outZ0 = _mm_shufflehi_epi16(out0, _MM_SHUFFLE(2, 1, 3, 0)); 881 outZ0 = _mm_shuffle_epi32 (outZ0, _MM_SHUFFLE(3, 1, 2, 0)); 882 outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2)); 883 outZ8 = _mm_shufflelo_epi16(out8, _MM_SHUFFLE(3, 0, 2, 1)); 884 outZ8 = _mm_shuffle_epi32 (outZ8, _MM_SHUFFLE(3, 1, 2, 0)); 885 outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0)); 886 _mm_storeu_si128((__m128i*)&out[0], outZ0); 887 _mm_storeu_si128((__m128i*)&out[8], outZ8); 888 packed_out = _mm_packs_epi16(outZ0, outZ8); 889 } 890 { 891 const int16_t outZ_12 = out[12]; 892 const int16_t outZ_3 = out[3]; 893 out[3] = outZ_12; 894 out[12] = outZ_3; 895 } 896 897 // detect if all 'out' values are zeroes or not 898 { 899 int32_t tmp[4]; 900 _mm_storeu_si128((__m128i*)tmp, packed_out); 901 if (n) { 902 tmp[0] &= ~0xff; 903 } 904 return (tmp[3] || tmp[2] || tmp[1] || tmp[0]); 905 } 906 } 907 908 #endif // WEBP_USE_SSE2 909 910 //------------------------------------------------------------------------------ 911 // Entry point 912 913 extern void VP8EncDspInitSSE2(void); 914 915 void VP8EncDspInitSSE2(void) { 916 #if defined(WEBP_USE_SSE2) 917 VP8CollectHistogram = CollectHistogramSSE2; 918 VP8EncQuantizeBlock = QuantizeBlockSSE2; 919 VP8ITransform = ITransformSSE2; 920 VP8FTransform = FTransformSSE2; 921 VP8SSE16x16 = SSE16x16SSE2; 922 VP8SSE16x8 = SSE16x8SSE2; 923 VP8SSE8x8 = SSE8x8SSE2; 924 VP8SSE4x4 = SSE4x4SSE2; 925 VP8TDisto4x4 = Disto4x4SSE2; 926 VP8TDisto16x16 = Disto16x16SSE2; 927 #endif // WEBP_USE_SSE2 928 } 929 930 #if defined(__cplusplus) || defined(c_plusplus) 931 } // extern "C" 932 #endif 933