enc_sse2.c revision 8b720228d581a84fd173b6dcb2fa295b59db489a
1// Copyright 2011 Google Inc. All Rights Reserved. 2// 3// Use of this source code is governed by a BSD-style license 4// that can be found in the COPYING file in the root of the source 5// tree. An additional intellectual property rights grant can be found 6// in the file PATENTS. All contributing project authors may 7// be found in the AUTHORS file in the root of the source tree. 8// ----------------------------------------------------------------------------- 9// 10// SSE2 version of speed-critical encoding functions. 11// 12// Author: Christian Duvivier (cduvivier@google.com) 13 14#include "./dsp.h" 15 16#if defined(WEBP_USE_SSE2) 17#include <stdlib.h> // for abs() 18#include <emmintrin.h> 19 20#include "../enc/vp8enci.h" 21 22//------------------------------------------------------------------------------ 23// Quite useful macro for debugging. Left here for convenience. 24 25#if 0 26#include <stdio.h> 27static void PrintReg(const __m128i r, const char* const name, int size) { 28 int n; 29 union { 30 __m128i r; 31 uint8_t i8[16]; 32 uint16_t i16[8]; 33 uint32_t i32[4]; 34 uint64_t i64[2]; 35 } tmp; 36 tmp.r = r; 37 printf("%s\t: ", name); 38 if (size == 8) { 39 for (n = 0; n < 16; ++n) printf("%.2x ", tmp.i8[n]); 40 } else if (size == 16) { 41 for (n = 0; n < 8; ++n) printf("%.4x ", tmp.i16[n]); 42 } else if (size == 32) { 43 for (n = 0; n < 4; ++n) printf("%.8x ", tmp.i32[n]); 44 } else { 45 for (n = 0; n < 2; ++n) printf("%.16lx ", tmp.i64[n]); 46 } 47 printf("\n"); 48} 49#endif 50 51//------------------------------------------------------------------------------ 52// Compute susceptibility based on DCT-coeff histograms: 53// the higher, the "easier" the macroblock is to compress. 54 55static void CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred, 56 int start_block, int end_block, 57 VP8Histogram* const histo) { 58 const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); 59 int j; 60 for (j = start_block; j < end_block; ++j) { 61 int16_t out[16]; 62 int k; 63 64 VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); 65 66 // Convert coefficients to bin (within out[]). 67 { 68 // Load. 69 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); 70 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); 71 // sign(out) = out >> 15 (0x0000 if positive, 0xffff if negative) 72 const __m128i sign0 = _mm_srai_epi16(out0, 15); 73 const __m128i sign1 = _mm_srai_epi16(out1, 15); 74 // abs(out) = (out ^ sign) - sign 75 const __m128i xor0 = _mm_xor_si128(out0, sign0); 76 const __m128i xor1 = _mm_xor_si128(out1, sign1); 77 const __m128i abs0 = _mm_sub_epi16(xor0, sign0); 78 const __m128i abs1 = _mm_sub_epi16(xor1, sign1); 79 // v = abs(out) >> 3 80 const __m128i v0 = _mm_srai_epi16(abs0, 3); 81 const __m128i v1 = _mm_srai_epi16(abs1, 3); 82 // bin = min(v, MAX_COEFF_THRESH) 83 const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); 84 const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); 85 // Store. 86 _mm_storeu_si128((__m128i*)&out[0], bin0); 87 _mm_storeu_si128((__m128i*)&out[8], bin1); 88 } 89 90 // Convert coefficients to bin. 91 for (k = 0; k < 16; ++k) { 92 histo->distribution[out[k]]++; 93 } 94 } 95} 96 97//------------------------------------------------------------------------------ 98// Transforms (Paragraph 14.4) 99 100// Does one or two inverse transforms. 101static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst, 102 int do_two) { 103 // This implementation makes use of 16-bit fixed point versions of two 104 // multiply constants: 105 // K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16 106 // K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16 107 // 108 // To be able to use signed 16-bit integers, we use the following trick to 109 // have constants within range: 110 // - Associated constants are obtained by subtracting the 16-bit fixed point 111 // version of one: 112 // k = K - (1 << 16) => K = k + (1 << 16) 113 // K1 = 85267 => k1 = 20091 114 // K2 = 35468 => k2 = -30068 115 // - The multiplication of a variable by a constant become the sum of the 116 // variable and the multiplication of that variable by the associated 117 // constant: 118 // (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x 119 const __m128i k1 = _mm_set1_epi16(20091); 120 const __m128i k2 = _mm_set1_epi16(-30068); 121 __m128i T0, T1, T2, T3; 122 123 // Load and concatenate the transform coefficients (we'll do two inverse 124 // transforms in parallel). In the case of only one inverse transform, the 125 // second half of the vectors will just contain random value we'll never 126 // use nor store. 127 __m128i in0, in1, in2, in3; 128 { 129 in0 = _mm_loadl_epi64((__m128i*)&in[0]); 130 in1 = _mm_loadl_epi64((__m128i*)&in[4]); 131 in2 = _mm_loadl_epi64((__m128i*)&in[8]); 132 in3 = _mm_loadl_epi64((__m128i*)&in[12]); 133 // a00 a10 a20 a30 x x x x 134 // a01 a11 a21 a31 x x x x 135 // a02 a12 a22 a32 x x x x 136 // a03 a13 a23 a33 x x x x 137 if (do_two) { 138 const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]); 139 const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]); 140 const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]); 141 const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]); 142 in0 = _mm_unpacklo_epi64(in0, inB0); 143 in1 = _mm_unpacklo_epi64(in1, inB1); 144 in2 = _mm_unpacklo_epi64(in2, inB2); 145 in3 = _mm_unpacklo_epi64(in3, inB3); 146 // a00 a10 a20 a30 b00 b10 b20 b30 147 // a01 a11 a21 a31 b01 b11 b21 b31 148 // a02 a12 a22 a32 b02 b12 b22 b32 149 // a03 a13 a23 a33 b03 b13 b23 b33 150 } 151 } 152 153 // Vertical pass and subsequent transpose. 154 { 155 // First pass, c and d calculations are longer because of the "trick" 156 // multiplications. 157 const __m128i a = _mm_add_epi16(in0, in2); 158 const __m128i b = _mm_sub_epi16(in0, in2); 159 // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3 160 const __m128i c1 = _mm_mulhi_epi16(in1, k2); 161 const __m128i c2 = _mm_mulhi_epi16(in3, k1); 162 const __m128i c3 = _mm_sub_epi16(in1, in3); 163 const __m128i c4 = _mm_sub_epi16(c1, c2); 164 const __m128i c = _mm_add_epi16(c3, c4); 165 // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3 166 const __m128i d1 = _mm_mulhi_epi16(in1, k1); 167 const __m128i d2 = _mm_mulhi_epi16(in3, k2); 168 const __m128i d3 = _mm_add_epi16(in1, in3); 169 const __m128i d4 = _mm_add_epi16(d1, d2); 170 const __m128i d = _mm_add_epi16(d3, d4); 171 172 // Second pass. 173 const __m128i tmp0 = _mm_add_epi16(a, d); 174 const __m128i tmp1 = _mm_add_epi16(b, c); 175 const __m128i tmp2 = _mm_sub_epi16(b, c); 176 const __m128i tmp3 = _mm_sub_epi16(a, d); 177 178 // Transpose the two 4x4. 179 // a00 a01 a02 a03 b00 b01 b02 b03 180 // a10 a11 a12 a13 b10 b11 b12 b13 181 // a20 a21 a22 a23 b20 b21 b22 b23 182 // a30 a31 a32 a33 b30 b31 b32 b33 183 const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1); 184 const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3); 185 const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1); 186 const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3); 187 // a00 a10 a01 a11 a02 a12 a03 a13 188 // a20 a30 a21 a31 a22 a32 a23 a33 189 // b00 b10 b01 b11 b02 b12 b03 b13 190 // b20 b30 b21 b31 b22 b32 b23 b33 191 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); 192 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); 193 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); 194 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); 195 // a00 a10 a20 a30 a01 a11 a21 a31 196 // b00 b10 b20 b30 b01 b11 b21 b31 197 // a02 a12 a22 a32 a03 a13 a23 a33 198 // b02 b12 a22 b32 b03 b13 b23 b33 199 T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); 200 T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); 201 T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); 202 T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); 203 // a00 a10 a20 a30 b00 b10 b20 b30 204 // a01 a11 a21 a31 b01 b11 b21 b31 205 // a02 a12 a22 a32 b02 b12 b22 b32 206 // a03 a13 a23 a33 b03 b13 b23 b33 207 } 208 209 // Horizontal pass and subsequent transpose. 210 { 211 // First pass, c and d calculations are longer because of the "trick" 212 // multiplications. 213 const __m128i four = _mm_set1_epi16(4); 214 const __m128i dc = _mm_add_epi16(T0, four); 215 const __m128i a = _mm_add_epi16(dc, T2); 216 const __m128i b = _mm_sub_epi16(dc, T2); 217 // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3 218 const __m128i c1 = _mm_mulhi_epi16(T1, k2); 219 const __m128i c2 = _mm_mulhi_epi16(T3, k1); 220 const __m128i c3 = _mm_sub_epi16(T1, T3); 221 const __m128i c4 = _mm_sub_epi16(c1, c2); 222 const __m128i c = _mm_add_epi16(c3, c4); 223 // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3 224 const __m128i d1 = _mm_mulhi_epi16(T1, k1); 225 const __m128i d2 = _mm_mulhi_epi16(T3, k2); 226 const __m128i d3 = _mm_add_epi16(T1, T3); 227 const __m128i d4 = _mm_add_epi16(d1, d2); 228 const __m128i d = _mm_add_epi16(d3, d4); 229 230 // Second pass. 231 const __m128i tmp0 = _mm_add_epi16(a, d); 232 const __m128i tmp1 = _mm_add_epi16(b, c); 233 const __m128i tmp2 = _mm_sub_epi16(b, c); 234 const __m128i tmp3 = _mm_sub_epi16(a, d); 235 const __m128i shifted0 = _mm_srai_epi16(tmp0, 3); 236 const __m128i shifted1 = _mm_srai_epi16(tmp1, 3); 237 const __m128i shifted2 = _mm_srai_epi16(tmp2, 3); 238 const __m128i shifted3 = _mm_srai_epi16(tmp3, 3); 239 240 // Transpose the two 4x4. 241 // a00 a01 a02 a03 b00 b01 b02 b03 242 // a10 a11 a12 a13 b10 b11 b12 b13 243 // a20 a21 a22 a23 b20 b21 b22 b23 244 // a30 a31 a32 a33 b30 b31 b32 b33 245 const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1); 246 const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3); 247 const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1); 248 const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3); 249 // a00 a10 a01 a11 a02 a12 a03 a13 250 // a20 a30 a21 a31 a22 a32 a23 a33 251 // b00 b10 b01 b11 b02 b12 b03 b13 252 // b20 b30 b21 b31 b22 b32 b23 b33 253 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); 254 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); 255 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); 256 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); 257 // a00 a10 a20 a30 a01 a11 a21 a31 258 // b00 b10 b20 b30 b01 b11 b21 b31 259 // a02 a12 a22 a32 a03 a13 a23 a33 260 // b02 b12 a22 b32 b03 b13 b23 b33 261 T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); 262 T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); 263 T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); 264 T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); 265 // a00 a10 a20 a30 b00 b10 b20 b30 266 // a01 a11 a21 a31 b01 b11 b21 b31 267 // a02 a12 a22 a32 b02 b12 b22 b32 268 // a03 a13 a23 a33 b03 b13 b23 b33 269 } 270 271 // Add inverse transform to 'ref' and store. 272 { 273 const __m128i zero = _mm_setzero_si128(); 274 // Load the reference(s). 275 __m128i ref0, ref1, ref2, ref3; 276 if (do_two) { 277 // Load eight bytes/pixels per line. 278 ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]); 279 ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]); 280 ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]); 281 ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]); 282 } else { 283 // Load four bytes/pixels per line. 284 ref0 = _mm_cvtsi32_si128(*(int*)&ref[0 * BPS]); 285 ref1 = _mm_cvtsi32_si128(*(int*)&ref[1 * BPS]); 286 ref2 = _mm_cvtsi32_si128(*(int*)&ref[2 * BPS]); 287 ref3 = _mm_cvtsi32_si128(*(int*)&ref[3 * BPS]); 288 } 289 // Convert to 16b. 290 ref0 = _mm_unpacklo_epi8(ref0, zero); 291 ref1 = _mm_unpacklo_epi8(ref1, zero); 292 ref2 = _mm_unpacklo_epi8(ref2, zero); 293 ref3 = _mm_unpacklo_epi8(ref3, zero); 294 // Add the inverse transform(s). 295 ref0 = _mm_add_epi16(ref0, T0); 296 ref1 = _mm_add_epi16(ref1, T1); 297 ref2 = _mm_add_epi16(ref2, T2); 298 ref3 = _mm_add_epi16(ref3, T3); 299 // Unsigned saturate to 8b. 300 ref0 = _mm_packus_epi16(ref0, ref0); 301 ref1 = _mm_packus_epi16(ref1, ref1); 302 ref2 = _mm_packus_epi16(ref2, ref2); 303 ref3 = _mm_packus_epi16(ref3, ref3); 304 // Store the results. 305 if (do_two) { 306 // Store eight bytes/pixels per line. 307 _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0); 308 _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1); 309 _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2); 310 _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3); 311 } else { 312 // Store four bytes/pixels per line. 313 *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(ref0); 314 *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(ref1); 315 *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(ref2); 316 *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(ref3); 317 } 318 } 319} 320 321static void FTransformSSE2(const uint8_t* src, const uint8_t* ref, 322 int16_t* out) { 323 const __m128i zero = _mm_setzero_si128(); 324 const __m128i seven = _mm_set1_epi16(7); 325 const __m128i k937 = _mm_set1_epi32(937); 326 const __m128i k1812 = _mm_set1_epi32(1812); 327 const __m128i k51000 = _mm_set1_epi32(51000); 328 const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16)); 329 const __m128i k5352_2217 = _mm_set_epi16(5352, 2217, 5352, 2217, 330 5352, 2217, 5352, 2217); 331 const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352, 332 2217, -5352, 2217, -5352); 333 const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8); 334 const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8); 335 const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352, 336 2217, 5352, 2217, 5352); 337 const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217, 338 -5352, 2217, -5352, 2217); 339 __m128i v01, v32; 340 341 342 // Difference between src and ref and initial transpose. 343 { 344 // Load src and convert to 16b. 345 const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]); 346 const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]); 347 const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]); 348 const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]); 349 const __m128i src_0 = _mm_unpacklo_epi8(src0, zero); 350 const __m128i src_1 = _mm_unpacklo_epi8(src1, zero); 351 const __m128i src_2 = _mm_unpacklo_epi8(src2, zero); 352 const __m128i src_3 = _mm_unpacklo_epi8(src3, zero); 353 // Load ref and convert to 16b. 354 const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]); 355 const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]); 356 const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]); 357 const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]); 358 const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero); 359 const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero); 360 const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero); 361 const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero); 362 // Compute difference. -> 00 01 02 03 00 00 00 00 363 const __m128i diff0 = _mm_sub_epi16(src_0, ref_0); 364 const __m128i diff1 = _mm_sub_epi16(src_1, ref_1); 365 const __m128i diff2 = _mm_sub_epi16(src_2, ref_2); 366 const __m128i diff3 = _mm_sub_epi16(src_3, ref_3); 367 368 369 // Unpack and shuffle 370 // 00 01 02 03 0 0 0 0 371 // 10 11 12 13 0 0 0 0 372 // 20 21 22 23 0 0 0 0 373 // 30 31 32 33 0 0 0 0 374 const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1); 375 const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3); 376 // 00 01 10 11 02 03 12 13 377 // 20 21 30 31 22 23 32 33 378 const __m128i shuf01_p = 379 _mm_shufflehi_epi16(shuf01, _MM_SHUFFLE(2, 3, 0, 1)); 380 const __m128i shuf23_p = 381 _mm_shufflehi_epi16(shuf23, _MM_SHUFFLE(2, 3, 0, 1)); 382 // 00 01 10 11 03 02 13 12 383 // 20 21 30 31 23 22 33 32 384 const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p); 385 const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p); 386 // 00 01 10 11 20 21 30 31 387 // 03 02 13 12 23 22 33 32 388 const __m128i a01 = _mm_add_epi16(s01, s32); 389 const __m128i a32 = _mm_sub_epi16(s01, s32); 390 // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ] 391 // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ] 392 393 const __m128i tmp0 = _mm_madd_epi16(a01, k88p); // [ (a0 + a1) << 3, ... ] 394 const __m128i tmp2 = _mm_madd_epi16(a01, k88m); // [ (a0 - a1) << 3, ... ] 395 const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p); 396 const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m); 397 const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812); 398 const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937); 399 const __m128i tmp1 = _mm_srai_epi32(tmp1_2, 9); 400 const __m128i tmp3 = _mm_srai_epi32(tmp3_2, 9); 401 const __m128i s03 = _mm_packs_epi32(tmp0, tmp2); 402 const __m128i s12 = _mm_packs_epi32(tmp1, tmp3); 403 const __m128i s_lo = _mm_unpacklo_epi16(s03, s12); // 0 1 0 1 0 1... 404 const __m128i s_hi = _mm_unpackhi_epi16(s03, s12); // 2 3 2 3 2 3 405 const __m128i v23 = _mm_unpackhi_epi32(s_lo, s_hi); 406 v01 = _mm_unpacklo_epi32(s_lo, s_hi); 407 v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2)); // 3 2 3 2 3 2.. 408 } 409 410 // Second pass 411 { 412 // Same operations are done on the (0,3) and (1,2) pairs. 413 // a0 = v0 + v3 414 // a1 = v1 + v2 415 // a3 = v0 - v3 416 // a2 = v1 - v2 417 const __m128i a01 = _mm_add_epi16(v01, v32); 418 const __m128i a32 = _mm_sub_epi16(v01, v32); 419 const __m128i a11 = _mm_unpackhi_epi64(a01, a01); 420 const __m128i a22 = _mm_unpackhi_epi64(a32, a32); 421 const __m128i a01_plus_7 = _mm_add_epi16(a01, seven); 422 423 // d0 = (a0 + a1 + 7) >> 4; 424 // d2 = (a0 - a1 + 7) >> 4; 425 const __m128i c0 = _mm_add_epi16(a01_plus_7, a11); 426 const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11); 427 const __m128i d0 = _mm_srai_epi16(c0, 4); 428 const __m128i d2 = _mm_srai_epi16(c2, 4); 429 430 // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16) 431 // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16) 432 const __m128i b23 = _mm_unpacklo_epi16(a22, a32); 433 const __m128i c1 = _mm_madd_epi16(b23, k5352_2217); 434 const __m128i c3 = _mm_madd_epi16(b23, k2217_5352); 435 const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one); 436 const __m128i d3 = _mm_add_epi32(c3, k51000); 437 const __m128i e1 = _mm_srai_epi32(d1, 16); 438 const __m128i e3 = _mm_srai_epi32(d3, 16); 439 const __m128i f1 = _mm_packs_epi32(e1, e1); 440 const __m128i f3 = _mm_packs_epi32(e3, e3); 441 // f1 = f1 + (a3 != 0); 442 // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the 443 // desired (0, 1), we add one earlier through k12000_plus_one. 444 // -> f1 = f1 + 1 - (a3 == 0) 445 const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero)); 446 447 _mm_storel_epi64((__m128i*)&out[ 0], d0); 448 _mm_storel_epi64((__m128i*)&out[ 4], g1); 449 _mm_storel_epi64((__m128i*)&out[ 8], d2); 450 _mm_storel_epi64((__m128i*)&out[12], f3); 451 } 452} 453 454static void FTransformWHTSSE2(const int16_t* in, int16_t* out) { 455 int32_t tmp[16]; 456 int i; 457 for (i = 0; i < 4; ++i, in += 64) { 458 const int a0 = (in[0 * 16] + in[2 * 16]); 459 const int a1 = (in[1 * 16] + in[3 * 16]); 460 const int a2 = (in[1 * 16] - in[3 * 16]); 461 const int a3 = (in[0 * 16] - in[2 * 16]); 462 tmp[0 + i * 4] = a0 + a1; 463 tmp[1 + i * 4] = a3 + a2; 464 tmp[2 + i * 4] = a3 - a2; 465 tmp[3 + i * 4] = a0 - a1; 466 } 467 { 468 const __m128i src0 = _mm_loadu_si128((__m128i*)&tmp[0]); 469 const __m128i src1 = _mm_loadu_si128((__m128i*)&tmp[4]); 470 const __m128i src2 = _mm_loadu_si128((__m128i*)&tmp[8]); 471 const __m128i src3 = _mm_loadu_si128((__m128i*)&tmp[12]); 472 const __m128i a0 = _mm_add_epi32(src0, src2); 473 const __m128i a1 = _mm_add_epi32(src1, src3); 474 const __m128i a2 = _mm_sub_epi32(src1, src3); 475 const __m128i a3 = _mm_sub_epi32(src0, src2); 476 const __m128i b0 = _mm_srai_epi32(_mm_add_epi32(a0, a1), 1); 477 const __m128i b1 = _mm_srai_epi32(_mm_add_epi32(a3, a2), 1); 478 const __m128i b2 = _mm_srai_epi32(_mm_sub_epi32(a3, a2), 1); 479 const __m128i b3 = _mm_srai_epi32(_mm_sub_epi32(a0, a1), 1); 480 const __m128i out0 = _mm_packs_epi32(b0, b1); 481 const __m128i out1 = _mm_packs_epi32(b2, b3); 482 _mm_storeu_si128((__m128i*)&out[0], out0); 483 _mm_storeu_si128((__m128i*)&out[8], out1); 484 } 485} 486 487//------------------------------------------------------------------------------ 488// Metric 489 490static int SSE_Nx4SSE2(const uint8_t* a, const uint8_t* b, 491 int num_quads, int do_16) { 492 const __m128i zero = _mm_setzero_si128(); 493 __m128i sum1 = zero; 494 __m128i sum2 = zero; 495 496 while (num_quads-- > 0) { 497 // Note: for the !do_16 case, we read 16 pixels instead of 8 but that's ok, 498 // thanks to buffer over-allocation to that effect. 499 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]); 500 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]); 501 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]); 502 const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]); 503 const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]); 504 const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]); 505 const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]); 506 const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]); 507 508 // compute clip0(a-b) and clip0(b-a) 509 const __m128i a0p = _mm_subs_epu8(a0, b0); 510 const __m128i a0m = _mm_subs_epu8(b0, a0); 511 const __m128i a1p = _mm_subs_epu8(a1, b1); 512 const __m128i a1m = _mm_subs_epu8(b1, a1); 513 const __m128i a2p = _mm_subs_epu8(a2, b2); 514 const __m128i a2m = _mm_subs_epu8(b2, a2); 515 const __m128i a3p = _mm_subs_epu8(a3, b3); 516 const __m128i a3m = _mm_subs_epu8(b3, a3); 517 518 // compute |a-b| with 8b arithmetic as clip0(a-b) | clip0(b-a) 519 const __m128i diff0 = _mm_or_si128(a0p, a0m); 520 const __m128i diff1 = _mm_or_si128(a1p, a1m); 521 const __m128i diff2 = _mm_or_si128(a2p, a2m); 522 const __m128i diff3 = _mm_or_si128(a3p, a3m); 523 524 // unpack (only four operations, instead of eight) 525 const __m128i low0 = _mm_unpacklo_epi8(diff0, zero); 526 const __m128i low1 = _mm_unpacklo_epi8(diff1, zero); 527 const __m128i low2 = _mm_unpacklo_epi8(diff2, zero); 528 const __m128i low3 = _mm_unpacklo_epi8(diff3, zero); 529 530 // multiply with self 531 const __m128i low_madd0 = _mm_madd_epi16(low0, low0); 532 const __m128i low_madd1 = _mm_madd_epi16(low1, low1); 533 const __m128i low_madd2 = _mm_madd_epi16(low2, low2); 534 const __m128i low_madd3 = _mm_madd_epi16(low3, low3); 535 536 // collect in a cascading way 537 const __m128i low_sum0 = _mm_add_epi32(low_madd0, low_madd1); 538 const __m128i low_sum1 = _mm_add_epi32(low_madd2, low_madd3); 539 sum1 = _mm_add_epi32(sum1, low_sum0); 540 sum2 = _mm_add_epi32(sum2, low_sum1); 541 542 if (do_16) { // if necessary, process the higher 8 bytes similarly 543 const __m128i hi0 = _mm_unpackhi_epi8(diff0, zero); 544 const __m128i hi1 = _mm_unpackhi_epi8(diff1, zero); 545 const __m128i hi2 = _mm_unpackhi_epi8(diff2, zero); 546 const __m128i hi3 = _mm_unpackhi_epi8(diff3, zero); 547 548 const __m128i hi_madd0 = _mm_madd_epi16(hi0, hi0); 549 const __m128i hi_madd1 = _mm_madd_epi16(hi1, hi1); 550 const __m128i hi_madd2 = _mm_madd_epi16(hi2, hi2); 551 const __m128i hi_madd3 = _mm_madd_epi16(hi3, hi3); 552 const __m128i hi_sum0 = _mm_add_epi32(hi_madd0, hi_madd1); 553 const __m128i hi_sum1 = _mm_add_epi32(hi_madd2, hi_madd3); 554 sum1 = _mm_add_epi32(sum1, hi_sum0); 555 sum2 = _mm_add_epi32(sum2, hi_sum1); 556 } 557 a += 4 * BPS; 558 b += 4 * BPS; 559 } 560 { 561 int32_t tmp[4]; 562 const __m128i sum = _mm_add_epi32(sum1, sum2); 563 _mm_storeu_si128((__m128i*)tmp, sum); 564 return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); 565 } 566} 567 568static int SSE16x16SSE2(const uint8_t* a, const uint8_t* b) { 569 return SSE_Nx4SSE2(a, b, 4, 1); 570} 571 572static int SSE16x8SSE2(const uint8_t* a, const uint8_t* b) { 573 return SSE_Nx4SSE2(a, b, 2, 1); 574} 575 576static int SSE8x8SSE2(const uint8_t* a, const uint8_t* b) { 577 return SSE_Nx4SSE2(a, b, 2, 0); 578} 579 580static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) { 581 const __m128i zero = _mm_setzero_si128(); 582 583 // Load values. Note that we read 8 pixels instead of 4, 584 // but the a/b buffers are over-allocated to that effect. 585 const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]); 586 const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]); 587 const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]); 588 const __m128i a3 = _mm_loadl_epi64((__m128i*)&a[BPS * 3]); 589 const __m128i b0 = _mm_loadl_epi64((__m128i*)&b[BPS * 0]); 590 const __m128i b1 = _mm_loadl_epi64((__m128i*)&b[BPS * 1]); 591 const __m128i b2 = _mm_loadl_epi64((__m128i*)&b[BPS * 2]); 592 const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]); 593 594 // Combine pair of lines and convert to 16b. 595 const __m128i a01 = _mm_unpacklo_epi32(a0, a1); 596 const __m128i a23 = _mm_unpacklo_epi32(a2, a3); 597 const __m128i b01 = _mm_unpacklo_epi32(b0, b1); 598 const __m128i b23 = _mm_unpacklo_epi32(b2, b3); 599 const __m128i a01s = _mm_unpacklo_epi8(a01, zero); 600 const __m128i a23s = _mm_unpacklo_epi8(a23, zero); 601 const __m128i b01s = _mm_unpacklo_epi8(b01, zero); 602 const __m128i b23s = _mm_unpacklo_epi8(b23, zero); 603 604 // Compute differences; (a-b)^2 = (abs(a-b))^2 = (sat8(a-b) + sat8(b-a))^2 605 // TODO(cduvivier): Dissassemble and figure out why this is fastest. We don't 606 // need absolute values, there is no need to do calculation 607 // in 8bit as we are already in 16bit, ... Yet this is what 608 // benchmarks the fastest! 609 const __m128i d0 = _mm_subs_epu8(a01s, b01s); 610 const __m128i d1 = _mm_subs_epu8(b01s, a01s); 611 const __m128i d2 = _mm_subs_epu8(a23s, b23s); 612 const __m128i d3 = _mm_subs_epu8(b23s, a23s); 613 614 // Square and add them all together. 615 const __m128i madd0 = _mm_madd_epi16(d0, d0); 616 const __m128i madd1 = _mm_madd_epi16(d1, d1); 617 const __m128i madd2 = _mm_madd_epi16(d2, d2); 618 const __m128i madd3 = _mm_madd_epi16(d3, d3); 619 const __m128i sum0 = _mm_add_epi32(madd0, madd1); 620 const __m128i sum1 = _mm_add_epi32(madd2, madd3); 621 const __m128i sum2 = _mm_add_epi32(sum0, sum1); 622 623 int32_t tmp[4]; 624 _mm_storeu_si128((__m128i*)tmp, sum2); 625 return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); 626} 627 628//------------------------------------------------------------------------------ 629// Texture distortion 630// 631// We try to match the spectral content (weighted) between source and 632// reconstructed samples. 633 634// Hadamard transform 635// Returns the difference between the weighted sum of the absolute value of 636// transformed coefficients. 637static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB, 638 const uint16_t* const w) { 639 int32_t sum[4]; 640 __m128i tmp_0, tmp_1, tmp_2, tmp_3; 641 const __m128i zero = _mm_setzero_si128(); 642 643 // Load, combine and transpose inputs. 644 { 645 const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]); 646 const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]); 647 const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]); 648 const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]); 649 const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]); 650 const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]); 651 const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]); 652 const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]); 653 654 // Combine inA and inB (we'll do two transforms in parallel). 655 const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0); 656 const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1); 657 const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2); 658 const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3); 659 // a00 b00 a01 b01 a02 b03 a03 b03 0 0 0 0 0 0 0 0 660 // a10 b10 a11 b11 a12 b12 a13 b13 0 0 0 0 0 0 0 0 661 // a20 b20 a21 b21 a22 b22 a23 b23 0 0 0 0 0 0 0 0 662 // a30 b30 a31 b31 a32 b32 a33 b33 0 0 0 0 0 0 0 0 663 664 // Transpose the two 4x4, discarding the filling zeroes. 665 const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2); 666 const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3); 667 // a00 a20 b00 b20 a01 a21 b01 b21 a02 a22 b02 b22 a03 a23 b03 b23 668 // a10 a30 b10 b30 a11 a31 b11 b31 a12 a32 b12 b32 a13 a33 b13 b33 669 const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1); 670 const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1); 671 // a00 a10 a20 a30 b00 b10 b20 b30 a01 a11 a21 a31 b01 b11 b21 b31 672 // a02 a12 a22 a32 b02 b12 b22 b32 a03 a13 a23 a33 b03 b13 b23 b33 673 674 // Convert to 16b. 675 tmp_0 = _mm_unpacklo_epi8(transpose1_0, zero); 676 tmp_1 = _mm_unpackhi_epi8(transpose1_0, zero); 677 tmp_2 = _mm_unpacklo_epi8(transpose1_1, zero); 678 tmp_3 = _mm_unpackhi_epi8(transpose1_1, zero); 679 // a00 a10 a20 a30 b00 b10 b20 b30 680 // a01 a11 a21 a31 b01 b11 b21 b31 681 // a02 a12 a22 a32 b02 b12 b22 b32 682 // a03 a13 a23 a33 b03 b13 b23 b33 683 } 684 685 // Horizontal pass and subsequent transpose. 686 { 687 // Calculate a and b (two 4x4 at once). 688 const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); 689 const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); 690 const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); 691 const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); 692 const __m128i b0 = _mm_add_epi16(a0, a1); 693 const __m128i b1 = _mm_add_epi16(a3, a2); 694 const __m128i b2 = _mm_sub_epi16(a3, a2); 695 const __m128i b3 = _mm_sub_epi16(a0, a1); 696 // a00 a01 a02 a03 b00 b01 b02 b03 697 // a10 a11 a12 a13 b10 b11 b12 b13 698 // a20 a21 a22 a23 b20 b21 b22 b23 699 // a30 a31 a32 a33 b30 b31 b32 b33 700 701 // Transpose the two 4x4. 702 const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1); 703 const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3); 704 const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1); 705 const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3); 706 // a00 a10 a01 a11 a02 a12 a03 a13 707 // a20 a30 a21 a31 a22 a32 a23 a33 708 // b00 b10 b01 b11 b02 b12 b03 b13 709 // b20 b30 b21 b31 b22 b32 b23 b33 710 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); 711 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); 712 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); 713 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); 714 // a00 a10 a20 a30 a01 a11 a21 a31 715 // b00 b10 b20 b30 b01 b11 b21 b31 716 // a02 a12 a22 a32 a03 a13 a23 a33 717 // b02 b12 a22 b32 b03 b13 b23 b33 718 tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); 719 tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); 720 tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); 721 tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); 722 // a00 a10 a20 a30 b00 b10 b20 b30 723 // a01 a11 a21 a31 b01 b11 b21 b31 724 // a02 a12 a22 a32 b02 b12 b22 b32 725 // a03 a13 a23 a33 b03 b13 b23 b33 726 } 727 728 // Vertical pass and difference of weighted sums. 729 { 730 // Load all inputs. 731 // TODO(cduvivier): Make variable declarations and allocations aligned so 732 // we can use _mm_load_si128 instead of _mm_loadu_si128. 733 const __m128i w_0 = _mm_loadu_si128((__m128i*)&w[0]); 734 const __m128i w_8 = _mm_loadu_si128((__m128i*)&w[8]); 735 736 // Calculate a and b (two 4x4 at once). 737 const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); 738 const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); 739 const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); 740 const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); 741 const __m128i b0 = _mm_add_epi16(a0, a1); 742 const __m128i b1 = _mm_add_epi16(a3, a2); 743 const __m128i b2 = _mm_sub_epi16(a3, a2); 744 const __m128i b3 = _mm_sub_epi16(a0, a1); 745 746 // Separate the transforms of inA and inB. 747 __m128i A_b0 = _mm_unpacklo_epi64(b0, b1); 748 __m128i A_b2 = _mm_unpacklo_epi64(b2, b3); 749 __m128i B_b0 = _mm_unpackhi_epi64(b0, b1); 750 __m128i B_b2 = _mm_unpackhi_epi64(b2, b3); 751 752 { 753 // sign(b) = b >> 15 (0x0000 if positive, 0xffff if negative) 754 const __m128i sign_A_b0 = _mm_srai_epi16(A_b0, 15); 755 const __m128i sign_A_b2 = _mm_srai_epi16(A_b2, 15); 756 const __m128i sign_B_b0 = _mm_srai_epi16(B_b0, 15); 757 const __m128i sign_B_b2 = _mm_srai_epi16(B_b2, 15); 758 759 // b = abs(b) = (b ^ sign) - sign 760 A_b0 = _mm_xor_si128(A_b0, sign_A_b0); 761 A_b2 = _mm_xor_si128(A_b2, sign_A_b2); 762 B_b0 = _mm_xor_si128(B_b0, sign_B_b0); 763 B_b2 = _mm_xor_si128(B_b2, sign_B_b2); 764 A_b0 = _mm_sub_epi16(A_b0, sign_A_b0); 765 A_b2 = _mm_sub_epi16(A_b2, sign_A_b2); 766 B_b0 = _mm_sub_epi16(B_b0, sign_B_b0); 767 B_b2 = _mm_sub_epi16(B_b2, sign_B_b2); 768 } 769 770 // weighted sums 771 A_b0 = _mm_madd_epi16(A_b0, w_0); 772 A_b2 = _mm_madd_epi16(A_b2, w_8); 773 B_b0 = _mm_madd_epi16(B_b0, w_0); 774 B_b2 = _mm_madd_epi16(B_b2, w_8); 775 A_b0 = _mm_add_epi32(A_b0, A_b2); 776 B_b0 = _mm_add_epi32(B_b0, B_b2); 777 778 // difference of weighted sums 779 A_b0 = _mm_sub_epi32(A_b0, B_b0); 780 _mm_storeu_si128((__m128i*)&sum[0], A_b0); 781 } 782 return sum[0] + sum[1] + sum[2] + sum[3]; 783} 784 785static int Disto4x4SSE2(const uint8_t* const a, const uint8_t* const b, 786 const uint16_t* const w) { 787 const int diff_sum = TTransformSSE2(a, b, w); 788 return abs(diff_sum) >> 5; 789} 790 791static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b, 792 const uint16_t* const w) { 793 int D = 0; 794 int x, y; 795 for (y = 0; y < 16 * BPS; y += 4 * BPS) { 796 for (x = 0; x < 16; x += 4) { 797 D += Disto4x4SSE2(a + x + y, b + x + y, w); 798 } 799 } 800 return D; 801} 802 803//------------------------------------------------------------------------------ 804// Quantization 805// 806 807// Simple quantization 808static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16], 809 int n, const VP8Matrix* const mtx) { 810 const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL); 811 const __m128i zero = _mm_setzero_si128(); 812 __m128i coeff0, coeff8; 813 __m128i out0, out8; 814 __m128i packed_out; 815 816 // Load all inputs. 817 // TODO(cduvivier): Make variable declarations and allocations aligned so that 818 // we can use _mm_load_si128 instead of _mm_loadu_si128. 819 __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]); 820 __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]); 821 const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]); 822 const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]); 823 const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]); 824 const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]); 825 const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]); 826 const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]); 827 const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]); 828 const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]); 829 830 // sign(in) = in >> 15 (0x0000 if positive, 0xffff if negative) 831 const __m128i sign0 = _mm_srai_epi16(in0, 15); 832 const __m128i sign8 = _mm_srai_epi16(in8, 15); 833 834 // coeff = abs(in) = (in ^ sign) - sign 835 coeff0 = _mm_xor_si128(in0, sign0); 836 coeff8 = _mm_xor_si128(in8, sign8); 837 coeff0 = _mm_sub_epi16(coeff0, sign0); 838 coeff8 = _mm_sub_epi16(coeff8, sign8); 839 840 // coeff = abs(in) + sharpen 841 coeff0 = _mm_add_epi16(coeff0, sharpen0); 842 coeff8 = _mm_add_epi16(coeff8, sharpen8); 843 844 // out = (coeff * iQ + B) >> QFIX; 845 { 846 // doing calculations with 32b precision (QFIX=17) 847 // out = (coeff * iQ) 848 __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0); 849 __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0); 850 __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8); 851 __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8); 852 __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H); 853 __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H); 854 __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H); 855 __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H); 856 // expand bias from 16b to 32b 857 __m128i bias_00 = _mm_unpacklo_epi16(bias0, zero); 858 __m128i bias_04 = _mm_unpackhi_epi16(bias0, zero); 859 __m128i bias_08 = _mm_unpacklo_epi16(bias8, zero); 860 __m128i bias_12 = _mm_unpackhi_epi16(bias8, zero); 861 // out = (coeff * iQ + B) 862 out_00 = _mm_add_epi32(out_00, bias_00); 863 out_04 = _mm_add_epi32(out_04, bias_04); 864 out_08 = _mm_add_epi32(out_08, bias_08); 865 out_12 = _mm_add_epi32(out_12, bias_12); 866 // out = (coeff * iQ + B) >> QFIX; 867 out_00 = _mm_srai_epi32(out_00, QFIX); 868 out_04 = _mm_srai_epi32(out_04, QFIX); 869 out_08 = _mm_srai_epi32(out_08, QFIX); 870 out_12 = _mm_srai_epi32(out_12, QFIX); 871 872 // pack result as 16b 873 out0 = _mm_packs_epi32(out_00, out_04); 874 out8 = _mm_packs_epi32(out_08, out_12); 875 876 // if (coeff > 2047) coeff = 2047 877 out0 = _mm_min_epi16(out0, max_coeff_2047); 878 out8 = _mm_min_epi16(out8, max_coeff_2047); 879 } 880 881 // get sign back (if (sign[j]) out_n = -out_n) 882 out0 = _mm_xor_si128(out0, sign0); 883 out8 = _mm_xor_si128(out8, sign8); 884 out0 = _mm_sub_epi16(out0, sign0); 885 out8 = _mm_sub_epi16(out8, sign8); 886 887 // in = out * Q 888 in0 = _mm_mullo_epi16(out0, q0); 889 in8 = _mm_mullo_epi16(out8, q8); 890 891 _mm_storeu_si128((__m128i*)&in[0], in0); 892 _mm_storeu_si128((__m128i*)&in[8], in8); 893 894 // zigzag the output before storing it. 895 // 896 // The zigzag pattern can almost be reproduced with a small sequence of 897 // shuffles. After it, we only need to swap the 7th (ending up in third 898 // position instead of twelfth) and 8th values. 899 { 900 __m128i outZ0, outZ8; 901 outZ0 = _mm_shufflehi_epi16(out0, _MM_SHUFFLE(2, 1, 3, 0)); 902 outZ0 = _mm_shuffle_epi32 (outZ0, _MM_SHUFFLE(3, 1, 2, 0)); 903 outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2)); 904 outZ8 = _mm_shufflelo_epi16(out8, _MM_SHUFFLE(3, 0, 2, 1)); 905 outZ8 = _mm_shuffle_epi32 (outZ8, _MM_SHUFFLE(3, 1, 2, 0)); 906 outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0)); 907 _mm_storeu_si128((__m128i*)&out[0], outZ0); 908 _mm_storeu_si128((__m128i*)&out[8], outZ8); 909 packed_out = _mm_packs_epi16(outZ0, outZ8); 910 } 911 { 912 const int16_t outZ_12 = out[12]; 913 const int16_t outZ_3 = out[3]; 914 out[3] = outZ_12; 915 out[12] = outZ_3; 916 } 917 918 // detect if all 'out' values are zeroes or not 919 { 920 int32_t tmp[4]; 921 _mm_storeu_si128((__m128i*)tmp, packed_out); 922 if (n) { 923 tmp[0] &= ~0xff; 924 } 925 return (tmp[3] || tmp[2] || tmp[1] || tmp[0]); 926 } 927} 928 929static int QuantizeBlockWHTSSE2(int16_t in[16], int16_t out[16], 930 const VP8Matrix* const mtx) { 931 return QuantizeBlockSSE2(in, out, 0, mtx); 932} 933 934#endif // WEBP_USE_SSE2 935 936//------------------------------------------------------------------------------ 937// Entry point 938 939extern void VP8EncDspInitSSE2(void); 940 941void VP8EncDspInitSSE2(void) { 942#if defined(WEBP_USE_SSE2) 943 VP8CollectHistogram = CollectHistogramSSE2; 944 VP8EncQuantizeBlock = QuantizeBlockSSE2; 945 VP8EncQuantizeBlockWHT = QuantizeBlockWHTSSE2; 946 VP8ITransform = ITransformSSE2; 947 VP8FTransform = FTransformSSE2; 948 VP8FTransformWHT = FTransformWHTSSE2; 949 VP8SSE16x16 = SSE16x16SSE2; 950 VP8SSE16x8 = SSE16x8SSE2; 951 VP8SSE8x8 = SSE8x8SSE2; 952 VP8SSE4x4 = SSE4x4SSE2; 953 VP8TDisto4x4 = Disto4x4SSE2; 954 VP8TDisto16x16 = Disto16x16SSE2; 955#endif // WEBP_USE_SSE2 956} 957 958