vp9_dct_sse2.c revision ba164dffc5a6795bce97fae02b51ccf3330e15e4
1/* 2 * Copyright (c) 2012 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include <emmintrin.h> // SSE2 12#include "vp9/common/vp9_idct.h" // for cospi constants 13 14void vp9_short_fdct4x4_sse2(int16_t *input, int16_t *output, int pitch) { 15 // The 2D transform is done with two passes which are actually pretty 16 // similar. In the first one, we transform the columns and transpose 17 // the results. In the second one, we transform the rows. To achieve that, 18 // as the first pass results are transposed, we tranpose the columns (that 19 // is the transposed rows) and transpose the results (so that it goes back 20 // in normal/row positions). 21 const int stride = pitch >> 1; 22 int pass; 23 // Constants 24 // When we use them, in one case, they are all the same. In all others 25 // it's a pair of them that we need to repeat four times. This is done 26 // by constructing the 32 bit constant corresponding to that pair. 27 const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); 28 const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); 29 const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); 30 const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); 31 const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); 32 const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); 33 const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0); 34 const __m128i kOne = _mm_set1_epi16(1); 35 __m128i in0, in1, in2, in3; 36 // Load inputs. 37 { 38 in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); 39 in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); 40 in2 = _mm_loadl_epi64((const __m128i *)(input + 2 * stride)); 41 in3 = _mm_loadl_epi64((const __m128i *)(input + 3 * stride)); 42 // x = x << 4 43 in0 = _mm_slli_epi16(in0, 4); 44 in1 = _mm_slli_epi16(in1, 4); 45 in2 = _mm_slli_epi16(in2, 4); 46 in3 = _mm_slli_epi16(in3, 4); 47 // if (i == 0 && input[0]) input[0] += 1; 48 { 49 // The mask will only contain wether the first value is zero, all 50 // other comparison will fail as something shifted by 4 (above << 4) 51 // can never be equal to one. To increment in the non-zero case, we 52 // add the mask and one for the first element: 53 // - if zero, mask = -1, v = v - 1 + 1 = v 54 // - if non-zero, mask = 0, v = v + 0 + 1 = v + 1 55 __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a); 56 in0 = _mm_add_epi16(in0, mask); 57 in0 = _mm_add_epi16(in0, k__nonzero_bias_b); 58 } 59 } 60 // Do the two transform/transpose passes 61 for (pass = 0; pass < 2; ++pass) { 62 // Transform 1/2: Add/substract 63 const __m128i r0 = _mm_add_epi16(in0, in3); 64 const __m128i r1 = _mm_add_epi16(in1, in2); 65 const __m128i r2 = _mm_sub_epi16(in1, in2); 66 const __m128i r3 = _mm_sub_epi16(in0, in3); 67 // Transform 1/2: Interleave to do the multiply by constants which gets us 68 // into 32 bits. 69 const __m128i t0 = _mm_unpacklo_epi16(r0, r1); 70 const __m128i t2 = _mm_unpacklo_epi16(r2, r3); 71 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); 72 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); 73 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); 74 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); 75 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 76 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 77 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); 78 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); 79 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 80 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 81 const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); 82 const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); 83 // Combine and transpose 84 const __m128i res0 = _mm_packs_epi32(w0, w2); 85 const __m128i res1 = _mm_packs_epi32(w4, w6); 86 // 00 01 02 03 20 21 22 23 87 // 10 11 12 13 30 31 32 33 88 const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); 89 const __m128i tr0_1 = _mm_unpackhi_epi16(res0, res1); 90 // 00 10 01 11 02 12 03 13 91 // 20 30 21 31 22 32 23 33 92 in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); 93 in2 = _mm_unpackhi_epi32(tr0_0, tr0_1); 94 // 00 10 20 30 01 11 21 31 in0 contains 0 followed by 1 95 // 02 12 22 32 03 13 23 33 in2 contains 2 followed by 3 96 if (0 == pass) { 97 // Extract values in the high part for second pass as transform code 98 // only uses the first four values. 99 in1 = _mm_unpackhi_epi64(in0, in0); 100 in3 = _mm_unpackhi_epi64(in2, in2); 101 } else { 102 // Post-condition output and store it (v + 1) >> 2, taking advantage 103 // of the fact 1/3 are stored just after 0/2. 104 __m128i out01 = _mm_add_epi16(in0, kOne); 105 __m128i out23 = _mm_add_epi16(in2, kOne); 106 out01 = _mm_srai_epi16(out01, 2); 107 out23 = _mm_srai_epi16(out23, 2); 108 _mm_storeu_si128((__m128i *)(output + 0 * 4), out01); 109 _mm_storeu_si128((__m128i *)(output + 2 * 4), out23); 110 } 111 } 112} 113 114void vp9_short_fdct8x4_sse2(int16_t *input, int16_t *output, int pitch) { 115 vp9_short_fdct4x4_sse2(input, output, pitch); 116 vp9_short_fdct4x4_sse2(input + 4, output + 16, pitch); 117} 118 119void vp9_short_fdct8x8_sse2(int16_t *input, int16_t *output, int pitch) { 120 const int stride = pitch >> 1; 121 int pass; 122 // Constants 123 // When we use them, in one case, they are all the same. In all others 124 // it's a pair of them that we need to repeat four times. This is done 125 // by constructing the 32 bit constant corresponding to that pair. 126 const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); 127 const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); 128 const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); 129 const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); 130 const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); 131 const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); 132 const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); 133 const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); 134 const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); 135 // Load input 136 __m128i in0 = _mm_loadu_si128((const __m128i *)(input + 0 * stride)); 137 __m128i in1 = _mm_loadu_si128((const __m128i *)(input + 1 * stride)); 138 __m128i in2 = _mm_loadu_si128((const __m128i *)(input + 2 * stride)); 139 __m128i in3 = _mm_loadu_si128((const __m128i *)(input + 3 * stride)); 140 __m128i in4 = _mm_loadu_si128((const __m128i *)(input + 4 * stride)); 141 __m128i in5 = _mm_loadu_si128((const __m128i *)(input + 5 * stride)); 142 __m128i in6 = _mm_loadu_si128((const __m128i *)(input + 6 * stride)); 143 __m128i in7 = _mm_loadu_si128((const __m128i *)(input + 7 * stride)); 144 // Pre-condition input (shift by two) 145 in0 = _mm_slli_epi16(in0, 2); 146 in1 = _mm_slli_epi16(in1, 2); 147 in2 = _mm_slli_epi16(in2, 2); 148 in3 = _mm_slli_epi16(in3, 2); 149 in4 = _mm_slli_epi16(in4, 2); 150 in5 = _mm_slli_epi16(in5, 2); 151 in6 = _mm_slli_epi16(in6, 2); 152 in7 = _mm_slli_epi16(in7, 2); 153 154 // We do two passes, first the columns, then the rows. The results of the 155 // first pass are transposed so that the same column code can be reused. The 156 // results of the second pass are also transposed so that the rows (processed 157 // as columns) are put back in row positions. 158 for (pass = 0; pass < 2; pass++) { 159 // To store results of each pass before the transpose. 160 __m128i res0, res1, res2, res3, res4, res5, res6, res7; 161 // Add/substract 162 const __m128i q0 = _mm_add_epi16(in0, in7); 163 const __m128i q1 = _mm_add_epi16(in1, in6); 164 const __m128i q2 = _mm_add_epi16(in2, in5); 165 const __m128i q3 = _mm_add_epi16(in3, in4); 166 const __m128i q4 = _mm_sub_epi16(in3, in4); 167 const __m128i q5 = _mm_sub_epi16(in2, in5); 168 const __m128i q6 = _mm_sub_epi16(in1, in6); 169 const __m128i q7 = _mm_sub_epi16(in0, in7); 170 // Work on first four results 171 { 172 // Add/substract 173 const __m128i r0 = _mm_add_epi16(q0, q3); 174 const __m128i r1 = _mm_add_epi16(q1, q2); 175 const __m128i r2 = _mm_sub_epi16(q1, q2); 176 const __m128i r3 = _mm_sub_epi16(q0, q3); 177 // Interleave to do the multiply by constants which gets us into 32bits 178 const __m128i t0 = _mm_unpacklo_epi16(r0, r1); 179 const __m128i t1 = _mm_unpackhi_epi16(r0, r1); 180 const __m128i t2 = _mm_unpacklo_epi16(r2, r3); 181 const __m128i t3 = _mm_unpackhi_epi16(r2, r3); 182 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); 183 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); 184 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); 185 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); 186 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); 187 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); 188 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); 189 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); 190 // dct_const_round_shift 191 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 192 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 193 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 194 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 195 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); 196 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); 197 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); 198 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); 199 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 200 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 201 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 202 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 203 const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); 204 const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); 205 const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); 206 const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); 207 // Combine 208 res0 = _mm_packs_epi32(w0, w1); 209 res4 = _mm_packs_epi32(w2, w3); 210 res2 = _mm_packs_epi32(w4, w5); 211 res6 = _mm_packs_epi32(w6, w7); 212 } 213 // Work on next four results 214 { 215 // Interleave to do the multiply by constants which gets us into 32bits 216 const __m128i d0 = _mm_unpacklo_epi16(q6, q5); 217 const __m128i d1 = _mm_unpackhi_epi16(q6, q5); 218 const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); 219 const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); 220 const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); 221 const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); 222 // dct_const_round_shift 223 const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); 224 const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); 225 const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); 226 const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); 227 const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); 228 const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); 229 const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); 230 const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); 231 // Combine 232 const __m128i r0 = _mm_packs_epi32(s0, s1); 233 const __m128i r1 = _mm_packs_epi32(s2, s3); 234 // Add/substract 235 const __m128i x0 = _mm_add_epi16(q4, r0); 236 const __m128i x1 = _mm_sub_epi16(q4, r0); 237 const __m128i x2 = _mm_sub_epi16(q7, r1); 238 const __m128i x3 = _mm_add_epi16(q7, r1); 239 // Interleave to do the multiply by constants which gets us into 32bits 240 const __m128i t0 = _mm_unpacklo_epi16(x0, x3); 241 const __m128i t1 = _mm_unpackhi_epi16(x0, x3); 242 const __m128i t2 = _mm_unpacklo_epi16(x1, x2); 243 const __m128i t3 = _mm_unpackhi_epi16(x1, x2); 244 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); 245 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); 246 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); 247 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); 248 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); 249 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); 250 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); 251 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); 252 // dct_const_round_shift 253 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 254 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 255 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 256 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 257 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); 258 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); 259 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); 260 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); 261 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 262 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 263 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 264 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 265 const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); 266 const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); 267 const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); 268 const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); 269 // Combine 270 res1 = _mm_packs_epi32(w0, w1); 271 res7 = _mm_packs_epi32(w2, w3); 272 res5 = _mm_packs_epi32(w4, w5); 273 res3 = _mm_packs_epi32(w6, w7); 274 } 275 // Transpose the 8x8. 276 { 277 // 00 01 02 03 04 05 06 07 278 // 10 11 12 13 14 15 16 17 279 // 20 21 22 23 24 25 26 27 280 // 30 31 32 33 34 35 36 37 281 // 40 41 42 43 44 45 46 47 282 // 50 51 52 53 54 55 56 57 283 // 60 61 62 63 64 65 66 67 284 // 70 71 72 73 74 75 76 77 285 const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); 286 const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); 287 const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); 288 const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); 289 const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); 290 const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); 291 const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); 292 const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); 293 // 00 10 01 11 02 12 03 13 294 // 20 30 21 31 22 32 23 33 295 // 04 14 05 15 06 16 07 17 296 // 24 34 25 35 26 36 27 37 297 // 40 50 41 51 42 52 43 53 298 // 60 70 61 71 62 72 63 73 299 // 54 54 55 55 56 56 57 57 300 // 64 74 65 75 66 76 67 77 301 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); 302 const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); 303 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); 304 const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); 305 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); 306 const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); 307 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); 308 const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); 309 // 00 10 20 30 01 11 21 31 310 // 40 50 60 70 41 51 61 71 311 // 02 12 22 32 03 13 23 33 312 // 42 52 62 72 43 53 63 73 313 // 04 14 24 34 05 15 21 36 314 // 44 54 64 74 45 55 61 76 315 // 06 16 26 36 07 17 27 37 316 // 46 56 66 76 47 57 67 77 317 in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); 318 in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); 319 in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); 320 in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); 321 in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); 322 in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); 323 in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); 324 in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); 325 // 00 10 20 30 40 50 60 70 326 // 01 11 21 31 41 51 61 71 327 // 02 12 22 32 42 52 62 72 328 // 03 13 23 33 43 53 63 73 329 // 04 14 24 34 44 54 64 74 330 // 05 15 25 35 45 55 65 75 331 // 06 16 26 36 46 56 66 76 332 // 07 17 27 37 47 57 67 77 333 } 334 } 335 // Post-condition output and store it 336 { 337 // Post-condition (division by two) 338 // division of two 16 bits signed numbers using shifts 339 // n / 2 = (n - (n >> 15)) >> 1 340 const __m128i sign_in0 = _mm_srai_epi16(in0, 15); 341 const __m128i sign_in1 = _mm_srai_epi16(in1, 15); 342 const __m128i sign_in2 = _mm_srai_epi16(in2, 15); 343 const __m128i sign_in3 = _mm_srai_epi16(in3, 15); 344 const __m128i sign_in4 = _mm_srai_epi16(in4, 15); 345 const __m128i sign_in5 = _mm_srai_epi16(in5, 15); 346 const __m128i sign_in6 = _mm_srai_epi16(in6, 15); 347 const __m128i sign_in7 = _mm_srai_epi16(in7, 15); 348 in0 = _mm_sub_epi16(in0, sign_in0); 349 in1 = _mm_sub_epi16(in1, sign_in1); 350 in2 = _mm_sub_epi16(in2, sign_in2); 351 in3 = _mm_sub_epi16(in3, sign_in3); 352 in4 = _mm_sub_epi16(in4, sign_in4); 353 in5 = _mm_sub_epi16(in5, sign_in5); 354 in6 = _mm_sub_epi16(in6, sign_in6); 355 in7 = _mm_sub_epi16(in7, sign_in7); 356 in0 = _mm_srai_epi16(in0, 1); 357 in1 = _mm_srai_epi16(in1, 1); 358 in2 = _mm_srai_epi16(in2, 1); 359 in3 = _mm_srai_epi16(in3, 1); 360 in4 = _mm_srai_epi16(in4, 1); 361 in5 = _mm_srai_epi16(in5, 1); 362 in6 = _mm_srai_epi16(in6, 1); 363 in7 = _mm_srai_epi16(in7, 1); 364 // store results 365 _mm_storeu_si128((__m128i *)(output + 0 * 8), in0); 366 _mm_storeu_si128((__m128i *)(output + 1 * 8), in1); 367 _mm_storeu_si128((__m128i *)(output + 2 * 8), in2); 368 _mm_storeu_si128((__m128i *)(output + 3 * 8), in3); 369 _mm_storeu_si128((__m128i *)(output + 4 * 8), in4); 370 _mm_storeu_si128((__m128i *)(output + 5 * 8), in5); 371 _mm_storeu_si128((__m128i *)(output + 6 * 8), in6); 372 _mm_storeu_si128((__m128i *)(output + 7 * 8), in7); 373 } 374} 375 376void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) { 377 // The 2D transform is done with two passes which are actually pretty 378 // similar. In the first one, we transform the columns and transpose 379 // the results. In the second one, we transform the rows. To achieve that, 380 // as the first pass results are transposed, we tranpose the columns (that 381 // is the transposed rows) and transpose the results (so that it goes back 382 // in normal/row positions). 383 const int stride = pitch >> 1; 384 int pass; 385 // We need an intermediate buffer between passes. 386 int16_t intermediate[256]; 387 int16_t *in = input; 388 int16_t *out = intermediate; 389 // Constants 390 // When we use them, in one case, they are all the same. In all others 391 // it's a pair of them that we need to repeat four times. This is done 392 // by constructing the 32 bit constant corresponding to that pair. 393 const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); 394 const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); 395 const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); 396 const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64); 397 const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); 398 const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); 399 const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); 400 const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); 401 const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); 402 const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64); 403 const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64); 404 const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); 405 const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); 406 const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64); 407 const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64); 408 const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); 409 const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); 410 const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); 411 const __m128i kOne = _mm_set1_epi16(1); 412 // Do the two transform/transpose passes 413 for (pass = 0; pass < 2; ++pass) { 414 // We process eight columns (transposed rows in second pass) at a time. 415 int column_start; 416 for (column_start = 0; column_start < 16; column_start += 8) { 417 __m128i in00, in01, in02, in03, in04, in05, in06, in07; 418 __m128i in08, in09, in10, in11, in12, in13, in14, in15; 419 __m128i input0, input1, input2, input3, input4, input5, input6, input7; 420 __m128i step1_0, step1_1, step1_2, step1_3; 421 __m128i step1_4, step1_5, step1_6, step1_7; 422 __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6; 423 __m128i step3_0, step3_1, step3_2, step3_3; 424 __m128i step3_4, step3_5, step3_6, step3_7; 425 __m128i res00, res01, res02, res03, res04, res05, res06, res07; 426 __m128i res08, res09, res10, res11, res12, res13, res14, res15; 427 // Load and pre-condition input. 428 if (0 == pass) { 429 in00 = _mm_loadu_si128((const __m128i *)(in + 0 * stride)); 430 in01 = _mm_loadu_si128((const __m128i *)(in + 1 * stride)); 431 in02 = _mm_loadu_si128((const __m128i *)(in + 2 * stride)); 432 in03 = _mm_loadu_si128((const __m128i *)(in + 3 * stride)); 433 in04 = _mm_loadu_si128((const __m128i *)(in + 4 * stride)); 434 in05 = _mm_loadu_si128((const __m128i *)(in + 5 * stride)); 435 in06 = _mm_loadu_si128((const __m128i *)(in + 6 * stride)); 436 in07 = _mm_loadu_si128((const __m128i *)(in + 7 * stride)); 437 in08 = _mm_loadu_si128((const __m128i *)(in + 8 * stride)); 438 in09 = _mm_loadu_si128((const __m128i *)(in + 9 * stride)); 439 in10 = _mm_loadu_si128((const __m128i *)(in + 10 * stride)); 440 in11 = _mm_loadu_si128((const __m128i *)(in + 11 * stride)); 441 in12 = _mm_loadu_si128((const __m128i *)(in + 12 * stride)); 442 in13 = _mm_loadu_si128((const __m128i *)(in + 13 * stride)); 443 in14 = _mm_loadu_si128((const __m128i *)(in + 14 * stride)); 444 in15 = _mm_loadu_si128((const __m128i *)(in + 15 * stride)); 445 // x = x << 2 446 in00 = _mm_slli_epi16(in00, 2); 447 in01 = _mm_slli_epi16(in01, 2); 448 in02 = _mm_slli_epi16(in02, 2); 449 in03 = _mm_slli_epi16(in03, 2); 450 in04 = _mm_slli_epi16(in04, 2); 451 in05 = _mm_slli_epi16(in05, 2); 452 in06 = _mm_slli_epi16(in06, 2); 453 in07 = _mm_slli_epi16(in07, 2); 454 in08 = _mm_slli_epi16(in08, 2); 455 in09 = _mm_slli_epi16(in09, 2); 456 in10 = _mm_slli_epi16(in10, 2); 457 in11 = _mm_slli_epi16(in11, 2); 458 in12 = _mm_slli_epi16(in12, 2); 459 in13 = _mm_slli_epi16(in13, 2); 460 in14 = _mm_slli_epi16(in14, 2); 461 in15 = _mm_slli_epi16(in15, 2); 462 } else { 463 in00 = _mm_loadu_si128((const __m128i *)(in + 0 * 16)); 464 in01 = _mm_loadu_si128((const __m128i *)(in + 1 * 16)); 465 in02 = _mm_loadu_si128((const __m128i *)(in + 2 * 16)); 466 in03 = _mm_loadu_si128((const __m128i *)(in + 3 * 16)); 467 in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 16)); 468 in05 = _mm_loadu_si128((const __m128i *)(in + 5 * 16)); 469 in06 = _mm_loadu_si128((const __m128i *)(in + 6 * 16)); 470 in07 = _mm_loadu_si128((const __m128i *)(in + 7 * 16)); 471 in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 16)); 472 in09 = _mm_loadu_si128((const __m128i *)(in + 9 * 16)); 473 in10 = _mm_loadu_si128((const __m128i *)(in + 10 * 16)); 474 in11 = _mm_loadu_si128((const __m128i *)(in + 11 * 16)); 475 in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 16)); 476 in13 = _mm_loadu_si128((const __m128i *)(in + 13 * 16)); 477 in14 = _mm_loadu_si128((const __m128i *)(in + 14 * 16)); 478 in15 = _mm_loadu_si128((const __m128i *)(in + 15 * 16)); 479 // x = (x + 1) >> 2 480 in00 = _mm_add_epi16(in00, kOne); 481 in01 = _mm_add_epi16(in01, kOne); 482 in02 = _mm_add_epi16(in02, kOne); 483 in03 = _mm_add_epi16(in03, kOne); 484 in04 = _mm_add_epi16(in04, kOne); 485 in05 = _mm_add_epi16(in05, kOne); 486 in06 = _mm_add_epi16(in06, kOne); 487 in07 = _mm_add_epi16(in07, kOne); 488 in08 = _mm_add_epi16(in08, kOne); 489 in09 = _mm_add_epi16(in09, kOne); 490 in10 = _mm_add_epi16(in10, kOne); 491 in11 = _mm_add_epi16(in11, kOne); 492 in12 = _mm_add_epi16(in12, kOne); 493 in13 = _mm_add_epi16(in13, kOne); 494 in14 = _mm_add_epi16(in14, kOne); 495 in15 = _mm_add_epi16(in15, kOne); 496 in00 = _mm_srai_epi16(in00, 2); 497 in01 = _mm_srai_epi16(in01, 2); 498 in02 = _mm_srai_epi16(in02, 2); 499 in03 = _mm_srai_epi16(in03, 2); 500 in04 = _mm_srai_epi16(in04, 2); 501 in05 = _mm_srai_epi16(in05, 2); 502 in06 = _mm_srai_epi16(in06, 2); 503 in07 = _mm_srai_epi16(in07, 2); 504 in08 = _mm_srai_epi16(in08, 2); 505 in09 = _mm_srai_epi16(in09, 2); 506 in10 = _mm_srai_epi16(in10, 2); 507 in11 = _mm_srai_epi16(in11, 2); 508 in12 = _mm_srai_epi16(in12, 2); 509 in13 = _mm_srai_epi16(in13, 2); 510 in14 = _mm_srai_epi16(in14, 2); 511 in15 = _mm_srai_epi16(in15, 2); 512 } 513 in += 8; 514 // Calculate input for the first 8 results. 515 { 516 input0 = _mm_add_epi16(in00, in15); 517 input1 = _mm_add_epi16(in01, in14); 518 input2 = _mm_add_epi16(in02, in13); 519 input3 = _mm_add_epi16(in03, in12); 520 input4 = _mm_add_epi16(in04, in11); 521 input5 = _mm_add_epi16(in05, in10); 522 input6 = _mm_add_epi16(in06, in09); 523 input7 = _mm_add_epi16(in07, in08); 524 } 525 // Calculate input for the next 8 results. 526 { 527 step1_0 = _mm_sub_epi16(in07, in08); 528 step1_1 = _mm_sub_epi16(in06, in09); 529 step1_2 = _mm_sub_epi16(in05, in10); 530 step1_3 = _mm_sub_epi16(in04, in11); 531 step1_4 = _mm_sub_epi16(in03, in12); 532 step1_5 = _mm_sub_epi16(in02, in13); 533 step1_6 = _mm_sub_epi16(in01, in14); 534 step1_7 = _mm_sub_epi16(in00, in15); 535 } 536 // Work on the first eight values; fdct8_1d(input, even_results); 537 { 538 // Add/substract 539 const __m128i q0 = _mm_add_epi16(input0, input7); 540 const __m128i q1 = _mm_add_epi16(input1, input6); 541 const __m128i q2 = _mm_add_epi16(input2, input5); 542 const __m128i q3 = _mm_add_epi16(input3, input4); 543 const __m128i q4 = _mm_sub_epi16(input3, input4); 544 const __m128i q5 = _mm_sub_epi16(input2, input5); 545 const __m128i q6 = _mm_sub_epi16(input1, input6); 546 const __m128i q7 = _mm_sub_epi16(input0, input7); 547 // Work on first four results 548 { 549 // Add/substract 550 const __m128i r0 = _mm_add_epi16(q0, q3); 551 const __m128i r1 = _mm_add_epi16(q1, q2); 552 const __m128i r2 = _mm_sub_epi16(q1, q2); 553 const __m128i r3 = _mm_sub_epi16(q0, q3); 554 // Interleave to do the multiply by constants which gets us 555 // into 32 bits. 556 const __m128i t0 = _mm_unpacklo_epi16(r0, r1); 557 const __m128i t1 = _mm_unpackhi_epi16(r0, r1); 558 const __m128i t2 = _mm_unpacklo_epi16(r2, r3); 559 const __m128i t3 = _mm_unpackhi_epi16(r2, r3); 560 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); 561 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); 562 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); 563 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); 564 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); 565 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); 566 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); 567 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); 568 // dct_const_round_shift 569 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 570 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 571 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 572 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 573 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); 574 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); 575 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); 576 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); 577 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 578 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 579 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 580 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 581 const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); 582 const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); 583 const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); 584 const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); 585 // Combine 586 res00 = _mm_packs_epi32(w0, w1); 587 res08 = _mm_packs_epi32(w2, w3); 588 res04 = _mm_packs_epi32(w4, w5); 589 res12 = _mm_packs_epi32(w6, w7); 590 } 591 // Work on next four results 592 { 593 // Interleave to do the multiply by constants which gets us 594 // into 32 bits. 595 const __m128i d0 = _mm_unpacklo_epi16(q6, q5); 596 const __m128i d1 = _mm_unpackhi_epi16(q6, q5); 597 const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); 598 const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); 599 const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); 600 const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); 601 // dct_const_round_shift 602 const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); 603 const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); 604 const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); 605 const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); 606 const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); 607 const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); 608 const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); 609 const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); 610 // Combine 611 const __m128i r0 = _mm_packs_epi32(s0, s1); 612 const __m128i r1 = _mm_packs_epi32(s2, s3); 613 // Add/substract 614 const __m128i x0 = _mm_add_epi16(q4, r0); 615 const __m128i x1 = _mm_sub_epi16(q4, r0); 616 const __m128i x2 = _mm_sub_epi16(q7, r1); 617 const __m128i x3 = _mm_add_epi16(q7, r1); 618 // Interleave to do the multiply by constants which gets us 619 // into 32 bits. 620 const __m128i t0 = _mm_unpacklo_epi16(x0, x3); 621 const __m128i t1 = _mm_unpackhi_epi16(x0, x3); 622 const __m128i t2 = _mm_unpacklo_epi16(x1, x2); 623 const __m128i t3 = _mm_unpackhi_epi16(x1, x2); 624 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); 625 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); 626 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); 627 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); 628 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); 629 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); 630 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); 631 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); 632 // dct_const_round_shift 633 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 634 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 635 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 636 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 637 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); 638 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); 639 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); 640 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); 641 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 642 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 643 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 644 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 645 const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); 646 const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); 647 const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); 648 const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); 649 // Combine 650 res02 = _mm_packs_epi32(w0, w1); 651 res14 = _mm_packs_epi32(w2, w3); 652 res10 = _mm_packs_epi32(w4, w5); 653 res06 = _mm_packs_epi32(w6, w7); 654 } 655 } 656 // Work on the next eight values; step1 -> odd_results 657 { 658 // step 2 659 { 660 const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); 661 const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); 662 const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); 663 const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); 664 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_m16); 665 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_m16); 666 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_m16); 667 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_m16); 668 // dct_const_round_shift 669 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 670 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 671 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 672 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 673 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 674 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 675 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 676 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 677 // Combine 678 step2_2 = _mm_packs_epi32(w0, w1); 679 step2_3 = _mm_packs_epi32(w2, w3); 680 } 681 { 682 const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); 683 const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); 684 const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); 685 const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); 686 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); 687 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); 688 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_p16); 689 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_p16); 690 // dct_const_round_shift 691 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 692 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 693 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 694 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 695 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 696 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 697 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 698 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 699 // Combine 700 step2_5 = _mm_packs_epi32(w0, w1); 701 step2_4 = _mm_packs_epi32(w2, w3); 702 } 703 // step 3 704 { 705 step3_0 = _mm_add_epi16(step1_0, step2_3); 706 step3_1 = _mm_add_epi16(step1_1, step2_2); 707 step3_2 = _mm_sub_epi16(step1_1, step2_2); 708 step3_3 = _mm_sub_epi16(step1_0, step2_3); 709 step3_4 = _mm_sub_epi16(step1_7, step2_4); 710 step3_5 = _mm_sub_epi16(step1_6, step2_5); 711 step3_6 = _mm_add_epi16(step1_6, step2_5); 712 step3_7 = _mm_add_epi16(step1_7, step2_4); 713 } 714 // step 4 715 { 716 const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); 717 const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); 718 const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); 719 const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); 720 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m08_p24); 721 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m08_p24); 722 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m24_m08); 723 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m24_m08); 724 // dct_const_round_shift 725 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 726 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 727 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 728 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 729 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 730 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 731 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 732 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 733 // Combine 734 step2_1 = _mm_packs_epi32(w0, w1); 735 step2_2 = _mm_packs_epi32(w2, w3); 736 } 737 { 738 const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); 739 const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); 740 const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); 741 const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); 742 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p24_p08); 743 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p24_p08); 744 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m08_p24); 745 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m08_p24); 746 // dct_const_round_shift 747 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 748 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 749 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 750 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 751 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 752 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 753 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 754 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 755 // Combine 756 step2_6 = _mm_packs_epi32(w0, w1); 757 step2_5 = _mm_packs_epi32(w2, w3); 758 } 759 // step 5 760 { 761 step1_0 = _mm_add_epi16(step3_0, step2_1); 762 step1_1 = _mm_sub_epi16(step3_0, step2_1); 763 step1_2 = _mm_sub_epi16(step3_3, step2_2); 764 step1_3 = _mm_add_epi16(step3_3, step2_2); 765 step1_4 = _mm_add_epi16(step3_4, step2_5); 766 step1_5 = _mm_sub_epi16(step3_4, step2_5); 767 step1_6 = _mm_sub_epi16(step3_7, step2_6); 768 step1_7 = _mm_add_epi16(step3_7, step2_6); 769 } 770 // step 6 771 { 772 const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); 773 const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); 774 const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); 775 const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); 776 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p30_p02); 777 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p30_p02); 778 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p14_p18); 779 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p14_p18); 780 // dct_const_round_shift 781 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 782 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 783 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 784 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 785 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 786 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 787 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 788 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 789 // Combine 790 res01 = _mm_packs_epi32(w0, w1); 791 res09 = _mm_packs_epi32(w2, w3); 792 } 793 { 794 const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); 795 const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); 796 const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); 797 const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); 798 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p22_p10); 799 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p22_p10); 800 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p06_p26); 801 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p06_p26); 802 // dct_const_round_shift 803 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 804 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 805 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 806 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 807 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 808 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 809 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 810 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 811 // Combine 812 res05 = _mm_packs_epi32(w0, w1); 813 res13 = _mm_packs_epi32(w2, w3); 814 } 815 { 816 const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); 817 const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); 818 const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); 819 const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); 820 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m10_p22); 821 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m10_p22); 822 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m26_p06); 823 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m26_p06); 824 // dct_const_round_shift 825 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 826 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 827 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 828 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 829 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 830 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 831 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 832 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 833 // Combine 834 res11 = _mm_packs_epi32(w0, w1); 835 res03 = _mm_packs_epi32(w2, w3); 836 } 837 { 838 const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); 839 const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); 840 const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); 841 const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); 842 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m02_p30); 843 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m02_p30); 844 const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m18_p14); 845 const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m18_p14); 846 // dct_const_round_shift 847 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); 848 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); 849 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); 850 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); 851 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); 852 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); 853 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); 854 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); 855 // Combine 856 res15 = _mm_packs_epi32(w0, w1); 857 res07 = _mm_packs_epi32(w2, w3); 858 } 859 } 860 // Transpose the results, do it as two 8x8 transposes. 861 { 862 // 00 01 02 03 04 05 06 07 863 // 10 11 12 13 14 15 16 17 864 // 20 21 22 23 24 25 26 27 865 // 30 31 32 33 34 35 36 37 866 // 40 41 42 43 44 45 46 47 867 // 50 51 52 53 54 55 56 57 868 // 60 61 62 63 64 65 66 67 869 // 70 71 72 73 74 75 76 77 870 const __m128i tr0_0 = _mm_unpacklo_epi16(res00, res01); 871 const __m128i tr0_1 = _mm_unpacklo_epi16(res02, res03); 872 const __m128i tr0_2 = _mm_unpackhi_epi16(res00, res01); 873 const __m128i tr0_3 = _mm_unpackhi_epi16(res02, res03); 874 const __m128i tr0_4 = _mm_unpacklo_epi16(res04, res05); 875 const __m128i tr0_5 = _mm_unpacklo_epi16(res06, res07); 876 const __m128i tr0_6 = _mm_unpackhi_epi16(res04, res05); 877 const __m128i tr0_7 = _mm_unpackhi_epi16(res06, res07); 878 // 00 10 01 11 02 12 03 13 879 // 20 30 21 31 22 32 23 33 880 // 04 14 05 15 06 16 07 17 881 // 24 34 25 35 26 36 27 37 882 // 40 50 41 51 42 52 43 53 883 // 60 70 61 71 62 72 63 73 884 // 54 54 55 55 56 56 57 57 885 // 64 74 65 75 66 76 67 77 886 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); 887 const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); 888 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); 889 const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); 890 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); 891 const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); 892 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); 893 const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); 894 // 00 10 20 30 01 11 21 31 895 // 40 50 60 70 41 51 61 71 896 // 02 12 22 32 03 13 23 33 897 // 42 52 62 72 43 53 63 73 898 // 04 14 24 34 05 15 21 36 899 // 44 54 64 74 45 55 61 76 900 // 06 16 26 36 07 17 27 37 901 // 46 56 66 76 47 57 67 77 902 const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); 903 const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); 904 const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); 905 const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); 906 const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); 907 const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); 908 const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); 909 const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); 910 // 00 10 20 30 40 50 60 70 911 // 01 11 21 31 41 51 61 71 912 // 02 12 22 32 42 52 62 72 913 // 03 13 23 33 43 53 63 73 914 // 04 14 24 34 44 54 64 74 915 // 05 15 25 35 45 55 65 75 916 // 06 16 26 36 46 56 66 76 917 // 07 17 27 37 47 57 67 77 918 _mm_storeu_si128((__m128i *)(out + 0 * 16), tr2_0); 919 _mm_storeu_si128((__m128i *)(out + 1 * 16), tr2_1); 920 _mm_storeu_si128((__m128i *)(out + 2 * 16), tr2_2); 921 _mm_storeu_si128((__m128i *)(out + 3 * 16), tr2_3); 922 _mm_storeu_si128((__m128i *)(out + 4 * 16), tr2_4); 923 _mm_storeu_si128((__m128i *)(out + 5 * 16), tr2_5); 924 _mm_storeu_si128((__m128i *)(out + 6 * 16), tr2_6); 925 _mm_storeu_si128((__m128i *)(out + 7 * 16), tr2_7); 926 } 927 { 928 // 00 01 02 03 04 05 06 07 929 // 10 11 12 13 14 15 16 17 930 // 20 21 22 23 24 25 26 27 931 // 30 31 32 33 34 35 36 37 932 // 40 41 42 43 44 45 46 47 933 // 50 51 52 53 54 55 56 57 934 // 60 61 62 63 64 65 66 67 935 // 70 71 72 73 74 75 76 77 936 const __m128i tr0_0 = _mm_unpacklo_epi16(res08, res09); 937 const __m128i tr0_1 = _mm_unpacklo_epi16(res10, res11); 938 const __m128i tr0_2 = _mm_unpackhi_epi16(res08, res09); 939 const __m128i tr0_3 = _mm_unpackhi_epi16(res10, res11); 940 const __m128i tr0_4 = _mm_unpacklo_epi16(res12, res13); 941 const __m128i tr0_5 = _mm_unpacklo_epi16(res14, res15); 942 const __m128i tr0_6 = _mm_unpackhi_epi16(res12, res13); 943 const __m128i tr0_7 = _mm_unpackhi_epi16(res14, res15); 944 // 00 10 01 11 02 12 03 13 945 // 20 30 21 31 22 32 23 33 946 // 04 14 05 15 06 16 07 17 947 // 24 34 25 35 26 36 27 37 948 // 40 50 41 51 42 52 43 53 949 // 60 70 61 71 62 72 63 73 950 // 54 54 55 55 56 56 57 57 951 // 64 74 65 75 66 76 67 77 952 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); 953 const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); 954 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); 955 const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); 956 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); 957 const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); 958 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); 959 const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); 960 // 00 10 20 30 01 11 21 31 961 // 40 50 60 70 41 51 61 71 962 // 02 12 22 32 03 13 23 33 963 // 42 52 62 72 43 53 63 73 964 // 04 14 24 34 05 15 21 36 965 // 44 54 64 74 45 55 61 76 966 // 06 16 26 36 07 17 27 37 967 // 46 56 66 76 47 57 67 77 968 const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); 969 const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); 970 const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); 971 const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); 972 const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); 973 const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); 974 const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); 975 const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); 976 // 00 10 20 30 40 50 60 70 977 // 01 11 21 31 41 51 61 71 978 // 02 12 22 32 42 52 62 72 979 // 03 13 23 33 43 53 63 73 980 // 04 14 24 34 44 54 64 74 981 // 05 15 25 35 45 55 65 75 982 // 06 16 26 36 46 56 66 76 983 // 07 17 27 37 47 57 67 77 984 // Store results 985 _mm_storeu_si128((__m128i *)(out + 8 + 0 * 16), tr2_0); 986 _mm_storeu_si128((__m128i *)(out + 8 + 1 * 16), tr2_1); 987 _mm_storeu_si128((__m128i *)(out + 8 + 2 * 16), tr2_2); 988 _mm_storeu_si128((__m128i *)(out + 8 + 3 * 16), tr2_3); 989 _mm_storeu_si128((__m128i *)(out + 8 + 4 * 16), tr2_4); 990 _mm_storeu_si128((__m128i *)(out + 8 + 5 * 16), tr2_5); 991 _mm_storeu_si128((__m128i *)(out + 8 + 6 * 16), tr2_6); 992 _mm_storeu_si128((__m128i *)(out + 8 + 7 * 16), tr2_7); 993 } 994 out += 8*16; 995 } 996 // Setup in/out for next pass. 997 in = intermediate; 998 out = output; 999 } 1000} 1001