1 /* 2 * Copyright (c) 2015 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 #include <stdlib.h> 11 12 #include "./vpx_dsp_rtcd.h" 13 #include "vpx_dsp/mips/macros_msa.h" 14 15 uint32_t vpx_avg_8x8_msa(const uint8_t *src, int32_t src_stride) { 16 uint32_t sum_out; 17 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; 18 v8u16 sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7; 19 v4u32 sum = { 0 }; 20 21 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); 22 HADD_UB4_UH(src0, src1, src2, src3, sum0, sum1, sum2, sum3); 23 HADD_UB4_UH(src4, src5, src6, src7, sum4, sum5, sum6, sum7); 24 ADD4(sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum0, sum2, sum4, sum6); 25 ADD2(sum0, sum2, sum4, sum6, sum0, sum4); 26 sum0 += sum4; 27 28 sum = __msa_hadd_u_w(sum0, sum0); 29 sum0 = (v8u16)__msa_pckev_h((v8i16)sum, (v8i16)sum); 30 sum = __msa_hadd_u_w(sum0, sum0); 31 sum = (v4u32)__msa_srari_w((v4i32)sum, 6); 32 sum_out = __msa_copy_u_w((v4i32)sum, 0); 33 34 return sum_out; 35 } 36 37 uint32_t vpx_avg_4x4_msa(const uint8_t *src, int32_t src_stride) { 38 uint32_t sum_out; 39 uint32_t src0, src1, src2, src3; 40 v16u8 vec = { 0 }; 41 v8u16 sum0; 42 v4u32 sum1; 43 v2u64 sum2; 44 45 LW4(src, src_stride, src0, src1, src2, src3); 46 INSERT_W4_UB(src0, src1, src2, src3, vec); 47 48 sum0 = __msa_hadd_u_h(vec, vec); 49 sum1 = __msa_hadd_u_w(sum0, sum0); 50 sum0 = (v8u16)__msa_pckev_h((v8i16)sum1, (v8i16)sum1); 51 sum1 = __msa_hadd_u_w(sum0, sum0); 52 sum2 = __msa_hadd_u_d(sum1, sum1); 53 sum1 = (v4u32)__msa_srari_w((v4i32)sum2, 4); 54 sum_out = __msa_copy_u_w((v4i32)sum1, 0); 55 56 return sum_out; 57 } 58 59 void vpx_hadamard_8x8_msa(const int16_t *src, int src_stride, int16_t *dst) { 60 v8i16 src0, src1, src2, src3, src4, src5, src6, src7; 61 v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 62 63 LD_SH8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); 64 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 65 tmp6, tmp7, tmp5, tmp3, tmp1); 66 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 67 src5, src7, src6, src3, src2); 68 BUTTERFLY_8(src0, src1, src2, src3, src7, src6, src5, src4, tmp0, tmp7, tmp3, 69 tmp4, tmp5, tmp1, tmp6, tmp2); 70 TRANSPOSE8x8_SH_SH(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, src0, src1, 71 src2, src3, src4, src5, src6, src7); 72 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 73 tmp6, tmp7, tmp5, tmp3, tmp1); 74 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 75 src5, src7, src6, src3, src2); 76 BUTTERFLY_8(src0, src1, src2, src3, src7, src6, src5, src4, tmp0, tmp7, tmp3, 77 tmp4, tmp5, tmp1, tmp6, tmp2); 78 TRANSPOSE8x8_SH_SH(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, src0, src1, 79 src2, src3, src4, src5, src6, src7); 80 ST_SH8(src0, src1, src2, src3, src4, src5, src6, src7, dst, 8); 81 } 82 83 void vpx_hadamard_16x16_msa(const int16_t *src, int src_stride, int16_t *dst) { 84 v8i16 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; 85 v8i16 src11, src12, src13, src14, src15, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; 86 v8i16 tmp6, tmp7, tmp8, tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15; 87 v8i16 res0, res1, res2, res3, res4, res5, res6, res7; 88 89 LD_SH2(src, 8, src0, src8); 90 src += src_stride; 91 LD_SH2(src, 8, src1, src9); 92 src += src_stride; 93 LD_SH2(src, 8, src2, src10); 94 src += src_stride; 95 LD_SH2(src, 8, src3, src11); 96 src += src_stride; 97 LD_SH2(src, 8, src4, src12); 98 src += src_stride; 99 LD_SH2(src, 8, src5, src13); 100 src += src_stride; 101 LD_SH2(src, 8, src6, src14); 102 src += src_stride; 103 LD_SH2(src, 8, src7, src15); 104 src += src_stride; 105 106 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 107 tmp6, tmp7, tmp5, tmp3, tmp1); 108 BUTTERFLY_8(src8, src10, src12, src14, src15, src13, src11, src9, tmp8, tmp10, 109 tmp12, tmp14, tmp15, tmp13, tmp11, tmp9); 110 111 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 112 src5, src7, src6, src3, src2); 113 BUTTERFLY_8(src0, src1, src2, src3, src7, src6, src5, src4, tmp0, tmp7, tmp3, 114 tmp4, tmp5, tmp1, tmp6, tmp2); 115 TRANSPOSE8x8_SH_SH(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, src0, src1, 116 src2, src3, src4, src5, src6, src7); 117 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 118 tmp6, tmp7, tmp5, tmp3, tmp1); 119 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 120 src5, src7, src6, src3, src2); 121 BUTTERFLY_8(src0, src1, src2, src3, src7, src6, src5, src4, tmp0, tmp7, tmp3, 122 tmp4, tmp5, tmp1, tmp6, tmp2); 123 TRANSPOSE8x8_SH_SH(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, src0, src1, 124 src2, src11, src4, src5, src6, src7); 125 ST_SH8(src0, src1, src2, src11, src4, src5, src6, src7, dst, 8); 126 127 BUTTERFLY_8(tmp8, tmp9, tmp12, tmp13, tmp15, tmp14, tmp11, tmp10, src8, src9, 128 src12, src13, src15, src14, src11, src10); 129 BUTTERFLY_8(src8, src9, src10, src11, src15, src14, src13, src12, tmp8, tmp15, 130 tmp11, tmp12, tmp13, tmp9, tmp14, tmp10); 131 TRANSPOSE8x8_SH_SH(tmp8, tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, src8, 132 src9, src10, src11, src12, src13, src14, src15); 133 BUTTERFLY_8(src8, src10, src12, src14, src15, src13, src11, src9, tmp8, tmp10, 134 tmp12, tmp14, tmp15, tmp13, tmp11, tmp9); 135 BUTTERFLY_8(tmp8, tmp9, tmp12, tmp13, tmp15, tmp14, tmp11, tmp10, src8, src9, 136 src12, src13, src15, src14, src11, src10); 137 BUTTERFLY_8(src8, src9, src10, src11, src15, src14, src13, src12, tmp8, tmp15, 138 tmp11, tmp12, tmp13, tmp9, tmp14, tmp10); 139 TRANSPOSE8x8_SH_SH(tmp8, tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, res0, 140 res1, res2, res3, res4, res5, res6, res7); 141 142 LD_SH2(src, 8, src0, src8); 143 src += src_stride; 144 LD_SH2(src, 8, src1, src9); 145 src += src_stride; 146 LD_SH2(src, 8, src2, src10); 147 src += src_stride; 148 LD_SH2(src, 8, src3, src11); 149 src += src_stride; 150 151 ST_SH8(res0, res1, res2, res3, res4, res5, res6, res7, dst + 64, 8); 152 153 LD_SH2(src, 8, src4, src12); 154 src += src_stride; 155 LD_SH2(src, 8, src5, src13); 156 src += src_stride; 157 LD_SH2(src, 8, src6, src14); 158 src += src_stride; 159 LD_SH2(src, 8, src7, src15); 160 src += src_stride; 161 162 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 163 tmp6, tmp7, tmp5, tmp3, tmp1); 164 BUTTERFLY_8(src8, src10, src12, src14, src15, src13, src11, src9, tmp8, tmp10, 165 tmp12, tmp14, tmp15, tmp13, tmp11, tmp9); 166 167 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 168 src5, src7, src6, src3, src2); 169 BUTTERFLY_8(src0, src1, src2, src3, src7, src6, src5, src4, tmp0, tmp7, tmp3, 170 tmp4, tmp5, tmp1, tmp6, tmp2); 171 TRANSPOSE8x8_SH_SH(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, src0, src1, 172 src2, src3, src4, src5, src6, src7); 173 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 174 tmp6, tmp7, tmp5, tmp3, tmp1); 175 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 176 src5, src7, src6, src3, src2); 177 BUTTERFLY_8(src0, src1, src2, src3, src7, src6, src5, src4, tmp0, tmp7, tmp3, 178 tmp4, tmp5, tmp1, tmp6, tmp2); 179 TRANSPOSE8x8_SH_SH(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, src0, src1, 180 src2, src3, src4, src5, src6, src7); 181 ST_SH8(src0, src1, src2, src3, src4, src5, src6, src7, dst + 2 * 64, 8); 182 183 BUTTERFLY_8(tmp8, tmp9, tmp12, tmp13, tmp15, tmp14, tmp11, tmp10, src8, src9, 184 src12, src13, src15, src14, src11, src10); 185 BUTTERFLY_8(src8, src9, src10, src11, src15, src14, src13, src12, tmp8, tmp15, 186 tmp11, tmp12, tmp13, tmp9, tmp14, tmp10); 187 TRANSPOSE8x8_SH_SH(tmp8, tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, src8, 188 src9, src10, src11, src12, src13, src14, src15); 189 BUTTERFLY_8(src8, src10, src12, src14, src15, src13, src11, src9, tmp8, tmp10, 190 tmp12, tmp14, tmp15, tmp13, tmp11, tmp9); 191 BUTTERFLY_8(tmp8, tmp9, tmp12, tmp13, tmp15, tmp14, tmp11, tmp10, src8, src9, 192 src12, src13, src15, src14, src11, src10); 193 BUTTERFLY_8(src8, src9, src10, src11, src15, src14, src13, src12, tmp8, tmp15, 194 tmp11, tmp12, tmp13, tmp9, tmp14, tmp10); 195 TRANSPOSE8x8_SH_SH(tmp8, tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, res0, 196 res1, res2, res3, res4, res5, res6, res7); 197 ST_SH8(res0, res1, res2, res3, res4, res5, res6, res7, dst + 3 * 64, 8); 198 199 LD_SH4(dst, 64, src0, src1, src2, src3); 200 LD_SH4(dst + 8, 64, src4, src5, src6, src7); 201 202 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 203 tmp6, tmp7, tmp5, tmp3, tmp1); 204 SRA_4V(tmp0, tmp1, tmp2, tmp3, 1); 205 SRA_4V(tmp4, tmp5, tmp6, tmp7, 1); 206 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 207 src5, src7, src6, src3, src2); 208 209 ST_SH4(src0, src1, src2, src3, dst, 64); 210 ST_SH4(src4, src5, src6, src7, dst + 8, 64); 211 dst += 16; 212 213 LD_SH4(dst, 64, src0, src1, src2, src3); 214 LD_SH4(dst + 8, 64, src4, src5, src6, src7); 215 216 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 217 tmp6, tmp7, tmp5, tmp3, tmp1); 218 SRA_4V(tmp0, tmp1, tmp2, tmp3, 1); 219 SRA_4V(tmp4, tmp5, tmp6, tmp7, 1); 220 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 221 src5, src7, src6, src3, src2); 222 223 ST_SH4(src0, src1, src2, src3, dst, 64); 224 ST_SH4(src4, src5, src6, src7, dst + 8, 64); 225 dst += 16; 226 227 LD_SH4(dst, 64, src0, src1, src2, src3); 228 LD_SH4(dst + 8, 64, src4, src5, src6, src7); 229 230 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 231 tmp6, tmp7, tmp5, tmp3, tmp1); 232 SRA_4V(tmp0, tmp1, tmp2, tmp3, 1); 233 SRA_4V(tmp4, tmp5, tmp6, tmp7, 1); 234 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 235 src5, src7, src6, src3, src2); 236 237 ST_SH4(src0, src1, src2, src3, dst, 64); 238 ST_SH4(src4, src5, src6, src7, dst + 8, 64); 239 dst += 16; 240 241 LD_SH4(dst, 64, src0, src1, src2, src3); 242 LD_SH4(dst + 8, 64, src4, src5, src6, src7); 243 244 BUTTERFLY_8(src0, src2, src4, src6, src7, src5, src3, src1, tmp0, tmp2, tmp4, 245 tmp6, tmp7, tmp5, tmp3, tmp1); 246 SRA_4V(tmp0, tmp1, tmp2, tmp3, 1); 247 SRA_4V(tmp4, tmp5, tmp6, tmp7, 1); 248 BUTTERFLY_8(tmp0, tmp1, tmp4, tmp5, tmp7, tmp6, tmp3, tmp2, src0, src1, src4, 249 src5, src7, src6, src3, src2); 250 251 ST_SH4(src0, src1, src2, src3, dst, 64); 252 ST_SH4(src4, src5, src6, src7, dst + 8, 64); 253 } 254 255 int vpx_satd_msa(const int16_t *data, int length) { 256 int i, satd; 257 v8i16 src0, src1, src2, src3, src4, src5, src6, src7; 258 v8i16 src8, src9, src10, src11, src12, src13, src14, src15; 259 v8i16 zero = { 0 }; 260 v8u16 tmp0_h, tmp1_h, tmp2_h, tmp3_h, tmp4_h, tmp5_h, tmp6_h, tmp7_h; 261 v4u32 tmp0_w = { 0 }; 262 263 if (16 == length) { 264 LD_SH2(data, 8, src0, src1); 265 tmp0_h = (v8u16)__msa_asub_s_h(src0, zero); 266 tmp1_h = (v8u16)__msa_asub_s_h(src1, zero); 267 tmp0_w = __msa_hadd_u_w(tmp0_h, tmp0_h); 268 tmp0_w += __msa_hadd_u_w(tmp1_h, tmp1_h); 269 satd = HADD_UW_U32(tmp0_w); 270 } else if (64 == length) { 271 LD_SH8(data, 8, src0, src1, src2, src3, src4, src5, src6, src7); 272 273 tmp0_h = (v8u16)__msa_asub_s_h(src0, zero); 274 tmp1_h = (v8u16)__msa_asub_s_h(src1, zero); 275 tmp2_h = (v8u16)__msa_asub_s_h(src2, zero); 276 tmp3_h = (v8u16)__msa_asub_s_h(src3, zero); 277 tmp4_h = (v8u16)__msa_asub_s_h(src4, zero); 278 tmp5_h = (v8u16)__msa_asub_s_h(src5, zero); 279 tmp6_h = (v8u16)__msa_asub_s_h(src6, zero); 280 tmp7_h = (v8u16)__msa_asub_s_h(src7, zero); 281 282 tmp0_w = __msa_hadd_u_w(tmp0_h, tmp0_h); 283 tmp0_w += __msa_hadd_u_w(tmp1_h, tmp1_h); 284 tmp0_w += __msa_hadd_u_w(tmp2_h, tmp2_h); 285 tmp0_w += __msa_hadd_u_w(tmp3_h, tmp3_h); 286 tmp0_w += __msa_hadd_u_w(tmp4_h, tmp4_h); 287 tmp0_w += __msa_hadd_u_w(tmp5_h, tmp5_h); 288 tmp0_w += __msa_hadd_u_w(tmp6_h, tmp6_h); 289 tmp0_w += __msa_hadd_u_w(tmp7_h, tmp7_h); 290 291 satd = HADD_UW_U32(tmp0_w); 292 } else if (256 == length) { 293 for (i = 0; i < 2; ++i) { 294 LD_SH8(data, 8, src0, src1, src2, src3, src4, src5, src6, src7); 295 data += 8 * 8; 296 LD_SH8(data, 8, src8, src9, src10, src11, src12, src13, src14, src15); 297 data += 8 * 8; 298 299 tmp0_h = (v8u16)__msa_asub_s_h(src0, zero); 300 tmp1_h = (v8u16)__msa_asub_s_h(src1, zero); 301 tmp2_h = (v8u16)__msa_asub_s_h(src2, zero); 302 tmp3_h = (v8u16)__msa_asub_s_h(src3, zero); 303 tmp4_h = (v8u16)__msa_asub_s_h(src4, zero); 304 tmp5_h = (v8u16)__msa_asub_s_h(src5, zero); 305 tmp6_h = (v8u16)__msa_asub_s_h(src6, zero); 306 tmp7_h = (v8u16)__msa_asub_s_h(src7, zero); 307 308 tmp0_w += __msa_hadd_u_w(tmp0_h, tmp0_h); 309 tmp0_w += __msa_hadd_u_w(tmp1_h, tmp1_h); 310 tmp0_w += __msa_hadd_u_w(tmp2_h, tmp2_h); 311 tmp0_w += __msa_hadd_u_w(tmp3_h, tmp3_h); 312 tmp0_w += __msa_hadd_u_w(tmp4_h, tmp4_h); 313 tmp0_w += __msa_hadd_u_w(tmp5_h, tmp5_h); 314 tmp0_w += __msa_hadd_u_w(tmp6_h, tmp6_h); 315 tmp0_w += __msa_hadd_u_w(tmp7_h, tmp7_h); 316 317 tmp0_h = (v8u16)__msa_asub_s_h(src8, zero); 318 tmp1_h = (v8u16)__msa_asub_s_h(src9, zero); 319 tmp2_h = (v8u16)__msa_asub_s_h(src10, zero); 320 tmp3_h = (v8u16)__msa_asub_s_h(src11, zero); 321 tmp4_h = (v8u16)__msa_asub_s_h(src12, zero); 322 tmp5_h = (v8u16)__msa_asub_s_h(src13, zero); 323 tmp6_h = (v8u16)__msa_asub_s_h(src14, zero); 324 tmp7_h = (v8u16)__msa_asub_s_h(src15, zero); 325 326 tmp0_w += __msa_hadd_u_w(tmp0_h, tmp0_h); 327 tmp0_w += __msa_hadd_u_w(tmp1_h, tmp1_h); 328 tmp0_w += __msa_hadd_u_w(tmp2_h, tmp2_h); 329 tmp0_w += __msa_hadd_u_w(tmp3_h, tmp3_h); 330 tmp0_w += __msa_hadd_u_w(tmp4_h, tmp4_h); 331 tmp0_w += __msa_hadd_u_w(tmp5_h, tmp5_h); 332 tmp0_w += __msa_hadd_u_w(tmp6_h, tmp6_h); 333 tmp0_w += __msa_hadd_u_w(tmp7_h, tmp7_h); 334 } 335 336 satd = HADD_UW_U32(tmp0_w); 337 } else if (1024 == length) { 338 for (i = 0; i < 8; ++i) { 339 LD_SH8(data, 8, src0, src1, src2, src3, src4, src5, src6, src7); 340 data += 8 * 8; 341 LD_SH8(data, 8, src8, src9, src10, src11, src12, src13, src14, src15); 342 data += 8 * 8; 343 344 tmp0_h = (v8u16)__msa_asub_s_h(src0, zero); 345 tmp1_h = (v8u16)__msa_asub_s_h(src1, zero); 346 tmp2_h = (v8u16)__msa_asub_s_h(src2, zero); 347 tmp3_h = (v8u16)__msa_asub_s_h(src3, zero); 348 tmp4_h = (v8u16)__msa_asub_s_h(src4, zero); 349 tmp5_h = (v8u16)__msa_asub_s_h(src5, zero); 350 tmp6_h = (v8u16)__msa_asub_s_h(src6, zero); 351 tmp7_h = (v8u16)__msa_asub_s_h(src7, zero); 352 353 tmp0_w += __msa_hadd_u_w(tmp0_h, tmp0_h); 354 tmp0_w += __msa_hadd_u_w(tmp1_h, tmp1_h); 355 tmp0_w += __msa_hadd_u_w(tmp2_h, tmp2_h); 356 tmp0_w += __msa_hadd_u_w(tmp3_h, tmp3_h); 357 tmp0_w += __msa_hadd_u_w(tmp4_h, tmp4_h); 358 tmp0_w += __msa_hadd_u_w(tmp5_h, tmp5_h); 359 tmp0_w += __msa_hadd_u_w(tmp6_h, tmp6_h); 360 tmp0_w += __msa_hadd_u_w(tmp7_h, tmp7_h); 361 362 tmp0_h = (v8u16)__msa_asub_s_h(src8, zero); 363 tmp1_h = (v8u16)__msa_asub_s_h(src9, zero); 364 tmp2_h = (v8u16)__msa_asub_s_h(src10, zero); 365 tmp3_h = (v8u16)__msa_asub_s_h(src11, zero); 366 tmp4_h = (v8u16)__msa_asub_s_h(src12, zero); 367 tmp5_h = (v8u16)__msa_asub_s_h(src13, zero); 368 tmp6_h = (v8u16)__msa_asub_s_h(src14, zero); 369 tmp7_h = (v8u16)__msa_asub_s_h(src15, zero); 370 371 tmp0_w += __msa_hadd_u_w(tmp0_h, tmp0_h); 372 tmp0_w += __msa_hadd_u_w(tmp1_h, tmp1_h); 373 tmp0_w += __msa_hadd_u_w(tmp2_h, tmp2_h); 374 tmp0_w += __msa_hadd_u_w(tmp3_h, tmp3_h); 375 tmp0_w += __msa_hadd_u_w(tmp4_h, tmp4_h); 376 tmp0_w += __msa_hadd_u_w(tmp5_h, tmp5_h); 377 tmp0_w += __msa_hadd_u_w(tmp6_h, tmp6_h); 378 tmp0_w += __msa_hadd_u_w(tmp7_h, tmp7_h); 379 } 380 381 satd = HADD_UW_U32(tmp0_w); 382 } else { 383 satd = 0; 384 385 for (i = 0; i < length; ++i) { 386 satd += abs(data[i]); 387 } 388 } 389 390 return satd; 391 } 392 393 void vpx_int_pro_row_msa(int16_t hbuf[16], const uint8_t *ref, 394 const int ref_stride, const int height) { 395 int i; 396 v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; 397 v8i16 hbuf_r = { 0 }; 398 v8i16 hbuf_l = { 0 }; 399 v8i16 ref0_r, ref0_l, ref1_r, ref1_l, ref2_r, ref2_l, ref3_r, ref3_l; 400 v8i16 ref4_r, ref4_l, ref5_r, ref5_l, ref6_r, ref6_l, ref7_r, ref7_l; 401 402 if (16 == height) { 403 for (i = 2; i--;) { 404 LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 405 ref += 8 * ref_stride; 406 UNPCK_UB_SH(ref0, ref0_r, ref0_l); 407 UNPCK_UB_SH(ref1, ref1_r, ref1_l); 408 UNPCK_UB_SH(ref2, ref2_r, ref2_l); 409 UNPCK_UB_SH(ref3, ref3_r, ref3_l); 410 UNPCK_UB_SH(ref4, ref4_r, ref4_l); 411 UNPCK_UB_SH(ref5, ref5_r, ref5_l); 412 UNPCK_UB_SH(ref6, ref6_r, ref6_l); 413 UNPCK_UB_SH(ref7, ref7_r, ref7_l); 414 ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l, 415 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 416 ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l, 417 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 418 ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l, 419 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 420 ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l, 421 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 422 } 423 424 SRA_2V(hbuf_r, hbuf_l, 3); 425 ST_SH2(hbuf_r, hbuf_l, hbuf, 8); 426 } else if (32 == height) { 427 for (i = 2; i--;) { 428 LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 429 ref += 8 * ref_stride; 430 UNPCK_UB_SH(ref0, ref0_r, ref0_l); 431 UNPCK_UB_SH(ref1, ref1_r, ref1_l); 432 UNPCK_UB_SH(ref2, ref2_r, ref2_l); 433 UNPCK_UB_SH(ref3, ref3_r, ref3_l); 434 UNPCK_UB_SH(ref4, ref4_r, ref4_l); 435 UNPCK_UB_SH(ref5, ref5_r, ref5_l); 436 UNPCK_UB_SH(ref6, ref6_r, ref6_l); 437 UNPCK_UB_SH(ref7, ref7_r, ref7_l); 438 ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l, 439 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 440 ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l, 441 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 442 ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l, 443 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 444 ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l, 445 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 446 LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 447 ref += 8 * ref_stride; 448 UNPCK_UB_SH(ref0, ref0_r, ref0_l); 449 UNPCK_UB_SH(ref1, ref1_r, ref1_l); 450 UNPCK_UB_SH(ref2, ref2_r, ref2_l); 451 UNPCK_UB_SH(ref3, ref3_r, ref3_l); 452 UNPCK_UB_SH(ref4, ref4_r, ref4_l); 453 UNPCK_UB_SH(ref5, ref5_r, ref5_l); 454 UNPCK_UB_SH(ref6, ref6_r, ref6_l); 455 UNPCK_UB_SH(ref7, ref7_r, ref7_l); 456 ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l, 457 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 458 ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l, 459 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 460 ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l, 461 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 462 ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l, 463 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 464 } 465 466 SRA_2V(hbuf_r, hbuf_l, 4); 467 ST_SH2(hbuf_r, hbuf_l, hbuf, 8); 468 } else if (64 == height) { 469 for (i = 4; i--;) { 470 LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 471 ref += 8 * ref_stride; 472 UNPCK_UB_SH(ref0, ref0_r, ref0_l); 473 UNPCK_UB_SH(ref1, ref1_r, ref1_l); 474 UNPCK_UB_SH(ref2, ref2_r, ref2_l); 475 UNPCK_UB_SH(ref3, ref3_r, ref3_l); 476 UNPCK_UB_SH(ref4, ref4_r, ref4_l); 477 UNPCK_UB_SH(ref5, ref5_r, ref5_l); 478 UNPCK_UB_SH(ref6, ref6_r, ref6_l); 479 UNPCK_UB_SH(ref7, ref7_r, ref7_l); 480 ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l, 481 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 482 ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l, 483 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 484 ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l, 485 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 486 ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l, 487 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 488 LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 489 ref += 8 * ref_stride; 490 UNPCK_UB_SH(ref0, ref0_r, ref0_l); 491 UNPCK_UB_SH(ref1, ref1_r, ref1_l); 492 UNPCK_UB_SH(ref2, ref2_r, ref2_l); 493 UNPCK_UB_SH(ref3, ref3_r, ref3_l); 494 UNPCK_UB_SH(ref4, ref4_r, ref4_l); 495 UNPCK_UB_SH(ref5, ref5_r, ref5_l); 496 UNPCK_UB_SH(ref6, ref6_r, ref6_l); 497 UNPCK_UB_SH(ref7, ref7_r, ref7_l); 498 ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l, 499 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 500 ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l, 501 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 502 ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l, 503 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 504 ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l, 505 hbuf_r, hbuf_l, hbuf_r, hbuf_l); 506 } 507 508 SRA_2V(hbuf_r, hbuf_l, 5); 509 ST_SH2(hbuf_r, hbuf_l, hbuf, 8); 510 } else { 511 const int norm_factor = height >> 1; 512 int cnt; 513 514 for (cnt = 0; cnt < 16; cnt++) { 515 hbuf[cnt] = 0; 516 } 517 518 for (i = 0; i < height; ++i) { 519 for (cnt = 0; cnt < 16; cnt++) { 520 hbuf[cnt] += ref[cnt]; 521 } 522 523 ref += ref_stride; 524 } 525 526 for (cnt = 0; cnt < 16; cnt++) { 527 hbuf[cnt] /= norm_factor; 528 } 529 } 530 } 531 532 int16_t vpx_int_pro_col_msa(const uint8_t *ref, const int width) { 533 int16_t sum; 534 v16u8 ref0, ref1, ref2, ref3; 535 v8u16 ref0_h; 536 537 if (16 == width) { 538 ref0 = LD_UB(ref); 539 ref0_h = __msa_hadd_u_h(ref0, ref0); 540 sum = HADD_UH_U32(ref0_h); 541 } else if (32 == width) { 542 LD_UB2(ref, 16, ref0, ref1); 543 ref0_h = __msa_hadd_u_h(ref0, ref0); 544 ref0_h += __msa_hadd_u_h(ref1, ref1); 545 sum = HADD_UH_U32(ref0_h); 546 } else if (64 == width) { 547 LD_UB4(ref, 16, ref0, ref1, ref2, ref3); 548 ref0_h = __msa_hadd_u_h(ref0, ref0); 549 ref0_h += __msa_hadd_u_h(ref1, ref1); 550 ref0_h += __msa_hadd_u_h(ref2, ref2); 551 ref0_h += __msa_hadd_u_h(ref3, ref3); 552 sum = HADD_UH_U32(ref0_h); 553 } else { 554 int idx; 555 556 sum = 0; 557 for (idx = 0; idx < width; ++idx) { 558 sum += ref[idx]; 559 } 560 } 561 562 return sum; 563 } 564 565 int vpx_vector_var_msa(const int16_t *ref, const int16_t *src, const int bwl) { 566 int sse, mean, var; 567 v8i16 src0, src1, src2, src3, src4, src5, src6, src7, ref0, ref1, ref2; 568 v8i16 ref3, ref4, ref5, ref6, ref7, src_l0_m, src_l1_m, src_l2_m, src_l3_m; 569 v8i16 src_l4_m, src_l5_m, src_l6_m, src_l7_m; 570 v4i32 res_l0_m, res_l1_m, res_l2_m, res_l3_m, res_l4_m, res_l5_m, res_l6_m; 571 v4i32 res_l7_m, mean_v; 572 v2i64 sse_v; 573 574 if (2 == bwl) { 575 LD_SH2(src, 8, src0, src1); 576 LD_SH2(ref, 8, ref0, ref1); 577 578 ILVRL_H2_SH(src0, ref0, src_l0_m, src_l1_m); 579 ILVRL_H2_SH(src1, ref1, src_l2_m, src_l3_m); 580 HSUB_UH2_SW(src_l0_m, src_l1_m, res_l0_m, res_l1_m); 581 HSUB_UH2_SW(src_l2_m, src_l3_m, res_l2_m, res_l3_m); 582 sse_v = __msa_dotp_s_d(res_l0_m, res_l0_m); 583 sse_v = __msa_dpadd_s_d(sse_v, res_l1_m, res_l1_m); 584 DPADD_SD2_SD(res_l2_m, res_l3_m, sse_v, sse_v); 585 mean_v = res_l0_m + res_l1_m; 586 mean_v += res_l2_m + res_l3_m; 587 588 sse_v += __msa_splati_d(sse_v, 1); 589 sse = __msa_copy_s_w((v4i32)sse_v, 0); 590 591 mean = HADD_SW_S32(mean_v); 592 } else if (3 == bwl) { 593 LD_SH4(src, 8, src0, src1, src2, src3); 594 LD_SH4(ref, 8, ref0, ref1, ref2, ref3); 595 596 ILVRL_H2_SH(src0, ref0, src_l0_m, src_l1_m); 597 ILVRL_H2_SH(src1, ref1, src_l2_m, src_l3_m); 598 ILVRL_H2_SH(src2, ref2, src_l4_m, src_l5_m); 599 ILVRL_H2_SH(src3, ref3, src_l6_m, src_l7_m); 600 HSUB_UH2_SW(src_l0_m, src_l1_m, res_l0_m, res_l1_m); 601 HSUB_UH2_SW(src_l2_m, src_l3_m, res_l2_m, res_l3_m); 602 HSUB_UH2_SW(src_l4_m, src_l5_m, res_l4_m, res_l5_m); 603 HSUB_UH2_SW(src_l6_m, src_l7_m, res_l6_m, res_l7_m); 604 sse_v = __msa_dotp_s_d(res_l0_m, res_l0_m); 605 sse_v = __msa_dpadd_s_d(sse_v, res_l1_m, res_l1_m); 606 DPADD_SD2_SD(res_l2_m, res_l3_m, sse_v, sse_v); 607 DPADD_SD2_SD(res_l4_m, res_l5_m, sse_v, sse_v); 608 DPADD_SD2_SD(res_l6_m, res_l7_m, sse_v, sse_v); 609 mean_v = res_l0_m + res_l1_m; 610 mean_v += res_l2_m + res_l3_m; 611 mean_v += res_l4_m + res_l5_m; 612 mean_v += res_l6_m + res_l7_m; 613 614 sse_v += __msa_splati_d(sse_v, 1); 615 sse = __msa_copy_s_w((v4i32)sse_v, 0); 616 617 mean = HADD_SW_S32(mean_v); 618 } else if (4 == bwl) { 619 LD_SH8(src, 8, src0, src1, src2, src3, src4, src5, src6, src7); 620 LD_SH8(ref, 8, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 621 622 ILVRL_H2_SH(src0, ref0, src_l0_m, src_l1_m); 623 ILVRL_H2_SH(src1, ref1, src_l2_m, src_l3_m); 624 ILVRL_H2_SH(src2, ref2, src_l4_m, src_l5_m); 625 ILVRL_H2_SH(src3, ref3, src_l6_m, src_l7_m); 626 HSUB_UH2_SW(src_l0_m, src_l1_m, res_l0_m, res_l1_m); 627 HSUB_UH2_SW(src_l2_m, src_l3_m, res_l2_m, res_l3_m); 628 HSUB_UH2_SW(src_l4_m, src_l5_m, res_l4_m, res_l5_m); 629 HSUB_UH2_SW(src_l6_m, src_l7_m, res_l6_m, res_l7_m); 630 sse_v = __msa_dotp_s_d(res_l0_m, res_l0_m); 631 sse_v = __msa_dpadd_s_d(sse_v, res_l1_m, res_l1_m); 632 DPADD_SD2_SD(res_l2_m, res_l3_m, sse_v, sse_v); 633 DPADD_SD2_SD(res_l4_m, res_l5_m, sse_v, sse_v); 634 DPADD_SD2_SD(res_l6_m, res_l7_m, sse_v, sse_v); 635 mean_v = res_l0_m + res_l1_m; 636 mean_v += res_l2_m + res_l3_m; 637 mean_v += res_l4_m + res_l5_m; 638 mean_v += res_l6_m + res_l7_m; 639 640 ILVRL_H2_SH(src4, ref4, src_l0_m, src_l1_m); 641 ILVRL_H2_SH(src5, ref5, src_l2_m, src_l3_m); 642 ILVRL_H2_SH(src6, ref6, src_l4_m, src_l5_m); 643 ILVRL_H2_SH(src7, ref7, src_l6_m, src_l7_m); 644 HSUB_UH2_SW(src_l0_m, src_l1_m, res_l0_m, res_l1_m); 645 HSUB_UH2_SW(src_l2_m, src_l3_m, res_l2_m, res_l3_m); 646 HSUB_UH2_SW(src_l4_m, src_l5_m, res_l4_m, res_l5_m); 647 HSUB_UH2_SW(src_l6_m, src_l7_m, res_l6_m, res_l7_m); 648 DPADD_SD2_SD(res_l0_m, res_l1_m, sse_v, sse_v); 649 DPADD_SD2_SD(res_l2_m, res_l3_m, sse_v, sse_v); 650 DPADD_SD2_SD(res_l4_m, res_l5_m, sse_v, sse_v); 651 DPADD_SD2_SD(res_l6_m, res_l7_m, sse_v, sse_v); 652 mean_v += res_l0_m + res_l1_m; 653 mean_v += res_l2_m + res_l3_m; 654 mean_v += res_l4_m + res_l5_m; 655 mean_v += res_l6_m + res_l7_m; 656 657 sse_v += __msa_splati_d(sse_v, 1); 658 sse = __msa_copy_s_w((v4i32)sse_v, 0); 659 660 mean = HADD_SW_S32(mean_v); 661 } else { 662 int i; 663 const int width = 4 << bwl; 664 665 sse = 0; 666 mean = 0; 667 668 for (i = 0; i < width; ++i) { 669 const int diff = ref[i] - src[i]; 670 671 mean += diff; 672 sse += diff * diff; 673 } 674 } 675 676 var = sse - ((mean * mean) >> (bwl + 2)); 677 678 return var; 679 } 680 681 void vpx_minmax_8x8_msa(const uint8_t *s, int p, const uint8_t *d, int dp, 682 int *min, int *max) { 683 v16u8 s0, s1, s2, s3, s4, s5, s6, s7, d0, d1, d2, d3, d4, d5, d6, d7; 684 v16u8 diff0, diff1, diff2, diff3, min0, min1, max0, max1; 685 686 LD_UB8(s, p, s0, s1, s2, s3, s4, s5, s6, s7); 687 LD_UB8(d, dp, d0, d1, d2, d3, d4, d5, d6, d7); 688 PCKEV_D4_UB(s1, s0, s3, s2, s5, s4, s7, s6, s0, s1, s2, s3); 689 PCKEV_D4_UB(d1, d0, d3, d2, d5, d4, d7, d6, d0, d1, d2, d3); 690 691 diff0 = __msa_asub_u_b(s0, d0); 692 diff1 = __msa_asub_u_b(s1, d1); 693 diff2 = __msa_asub_u_b(s2, d2); 694 diff3 = __msa_asub_u_b(s3, d3); 695 696 min0 = __msa_min_u_b(diff0, diff1); 697 min1 = __msa_min_u_b(diff2, diff3); 698 min0 = __msa_min_u_b(min0, min1); 699 700 max0 = __msa_max_u_b(diff0, diff1); 701 max1 = __msa_max_u_b(diff2, diff3); 702 max0 = __msa_max_u_b(max0, max1); 703 704 min1 = (v16u8)__msa_sldi_b((v16i8)min1, (v16i8)min0, 8); 705 min0 = __msa_min_u_b(min0, min1); 706 max1 = (v16u8)__msa_sldi_b((v16i8)max1, (v16i8)max0, 8); 707 max0 = __msa_max_u_b(max0, max1); 708 709 min1 = (v16u8)__msa_sldi_b((v16i8)min1, (v16i8)min0, 4); 710 min0 = __msa_min_u_b(min0, min1); 711 max1 = (v16u8)__msa_sldi_b((v16i8)max1, (v16i8)max0, 4); 712 max0 = __msa_max_u_b(max0, max1); 713 714 min1 = (v16u8)__msa_sldi_b((v16i8)min1, (v16i8)min0, 2); 715 min0 = __msa_min_u_b(min0, min1); 716 max1 = (v16u8)__msa_sldi_b((v16i8)max1, (v16i8)max0, 2); 717 max0 = __msa_max_u_b(max0, max1); 718 719 min1 = (v16u8)__msa_sldi_b((v16i8)min1, (v16i8)min0, 1); 720 min0 = __msa_min_u_b(min0, min1); 721 max1 = (v16u8)__msa_sldi_b((v16i8)max1, (v16i8)max0, 1); 722 max0 = __msa_max_u_b(max0, max1); 723 724 *min = min0[0]; 725 *max = max0[0]; 726 } 727
