1 // Copyright 2016 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 // MSA version of encoder dsp functions. 11 // 12 // Author: Prashant Patil (prashant.patil (at) imgtec.com) 13 14 #include "./dsp.h" 15 16 #if defined(WEBP_USE_MSA) 17 18 #include <stdlib.h> 19 #include "./msa_macro.h" 20 #include "../enc/vp8i_enc.h" 21 22 //------------------------------------------------------------------------------ 23 // Transforms 24 25 #define IDCT_1D_W(in0, in1, in2, in3, out0, out1, out2, out3) do { \ 26 v4i32 a1_m, b1_m, c1_m, d1_m; \ 27 const v4i32 cospi8sqrt2minus1 = __msa_fill_w(20091); \ 28 const v4i32 sinpi8sqrt2 = __msa_fill_w(35468); \ 29 v4i32 c_tmp1_m = in1 * sinpi8sqrt2; \ 30 v4i32 c_tmp2_m = in3 * cospi8sqrt2minus1; \ 31 v4i32 d_tmp1_m = in1 * cospi8sqrt2minus1; \ 32 v4i32 d_tmp2_m = in3 * sinpi8sqrt2; \ 33 \ 34 ADDSUB2(in0, in2, a1_m, b1_m); \ 35 SRAI_W2_SW(c_tmp1_m, c_tmp2_m, 16); \ 36 c_tmp2_m = c_tmp2_m + in3; \ 37 c1_m = c_tmp1_m - c_tmp2_m; \ 38 SRAI_W2_SW(d_tmp1_m, d_tmp2_m, 16); \ 39 d_tmp1_m = d_tmp1_m + in1; \ 40 d1_m = d_tmp1_m + d_tmp2_m; \ 41 BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3); \ 42 } while (0) 43 44 static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in, 45 uint8_t* dst) { 46 v8i16 input0, input1; 47 v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3; 48 v4i32 res0, res1, res2, res3; 49 v16i8 dest0, dest1, dest2, dest3; 50 const v16i8 zero = { 0 }; 51 52 LD_SH2(in, 8, input0, input1); 53 UNPCK_SH_SW(input0, in0, in1); 54 UNPCK_SH_SW(input1, in2, in3); 55 IDCT_1D_W(in0, in1, in2, in3, hz0, hz1, hz2, hz3); 56 TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3); 57 IDCT_1D_W(hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3); 58 SRARI_W4_SW(vt0, vt1, vt2, vt3, 3); 59 TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3); 60 LD_SB4(ref, BPS, dest0, dest1, dest2, dest3); 61 ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3, 62 res0, res1, res2, res3); 63 ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3, 64 res0, res1, res2, res3); 65 ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3); 66 CLIP_SW4_0_255(res0, res1, res2, res3); 67 PCKEV_B2_SW(res0, res1, res2, res3, vt0, vt1); 68 res0 = (v4i32)__msa_pckev_b((v16i8)vt0, (v16i8)vt1); 69 ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS); 70 } 71 72 static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst, 73 int do_two) { 74 ITransformOne(ref, in, dst); 75 if (do_two) { 76 ITransformOne(ref + 4, in + 16, dst + 4); 77 } 78 } 79 80 static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) { 81 uint64_t out0, out1, out2, out3; 82 uint32_t in0, in1, in2, in3; 83 v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; 84 v8i16 t0, t1, t2, t3; 85 v16u8 srcl0, srcl1, src0, src1; 86 const v8i16 mask0 = { 0, 4, 8, 12, 1, 5, 9, 13 }; 87 const v8i16 mask1 = { 3, 7, 11, 15, 2, 6, 10, 14 }; 88 const v8i16 mask2 = { 4, 0, 5, 1, 6, 2, 7, 3 }; 89 const v8i16 mask3 = { 0, 4, 1, 5, 2, 6, 3, 7 }; 90 const v8i16 cnst0 = { 2217, -5352, 2217, -5352, 2217, -5352, 2217, -5352 }; 91 const v8i16 cnst1 = { 5352, 2217, 5352, 2217, 5352, 2217, 5352, 2217 }; 92 93 LW4(src, BPS, in0, in1, in2, in3); 94 INSERT_W4_UB(in0, in1, in2, in3, src0); 95 LW4(ref, BPS, in0, in1, in2, in3); 96 INSERT_W4_UB(in0, in1, in2, in3, src1); 97 ILVRL_B2_UB(src0, src1, srcl0, srcl1); 98 HSUB_UB2_SH(srcl0, srcl1, t0, t1); 99 VSHF_H2_SH(t0, t1, t0, t1, mask0, mask1, t2, t3); 100 ADDSUB2(t2, t3, t0, t1); 101 t0 = SRLI_H(t0, 3); 102 VSHF_H2_SH(t0, t0, t1, t1, mask2, mask3, t3, t2); 103 tmp0 = __msa_hadd_s_w(t3, t3); 104 tmp2 = __msa_hsub_s_w(t3, t3); 105 FILL_W2_SW(1812, 937, tmp1, tmp3); 106 DPADD_SH2_SW(t2, t2, cnst0, cnst1, tmp3, tmp1); 107 SRAI_W2_SW(tmp1, tmp3, 9); 108 PCKEV_H2_SH(tmp1, tmp0, tmp3, tmp2, t0, t1); 109 VSHF_H2_SH(t0, t1, t0, t1, mask0, mask1, t2, t3); 110 ADDSUB2(t2, t3, t0, t1); 111 VSHF_H2_SH(t0, t0, t1, t1, mask2, mask3, t3, t2); 112 tmp0 = __msa_hadd_s_w(t3, t3); 113 tmp2 = __msa_hsub_s_w(t3, t3); 114 ADDVI_W2_SW(tmp0, 7, tmp2, 7, tmp0, tmp2); 115 SRAI_W2_SW(tmp0, tmp2, 4); 116 FILL_W2_SW(12000, 51000, tmp1, tmp3); 117 DPADD_SH2_SW(t2, t2, cnst0, cnst1, tmp3, tmp1); 118 SRAI_W2_SW(tmp1, tmp3, 16); 119 UNPCK_R_SH_SW(t1, tmp4); 120 tmp5 = __msa_ceqi_w(tmp4, 0); 121 tmp4 = (v4i32)__msa_nor_v((v16u8)tmp5, (v16u8)tmp5); 122 tmp5 = __msa_fill_w(1); 123 tmp5 = (v4i32)__msa_and_v((v16u8)tmp5, (v16u8)tmp4); 124 tmp1 += tmp5; 125 PCKEV_H2_SH(tmp1, tmp0, tmp3, tmp2, t0, t1); 126 out0 = __msa_copy_s_d((v2i64)t0, 0); 127 out1 = __msa_copy_s_d((v2i64)t0, 1); 128 out2 = __msa_copy_s_d((v2i64)t1, 0); 129 out3 = __msa_copy_s_d((v2i64)t1, 1); 130 SD4(out0, out1, out2, out3, out, 8); 131 } 132 133 static void FTransformWHT(const int16_t* in, int16_t* out) { 134 v8i16 in0 = { 0 }; 135 v8i16 in1 = { 0 }; 136 v8i16 tmp0, tmp1, tmp2, tmp3; 137 v8i16 out0, out1; 138 const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 }; 139 const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 }; 140 const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 }; 141 const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 }; 142 143 in0 = __msa_insert_h(in0, 0, in[ 0]); 144 in0 = __msa_insert_h(in0, 1, in[ 64]); 145 in0 = __msa_insert_h(in0, 2, in[128]); 146 in0 = __msa_insert_h(in0, 3, in[192]); 147 in0 = __msa_insert_h(in0, 4, in[ 16]); 148 in0 = __msa_insert_h(in0, 5, in[ 80]); 149 in0 = __msa_insert_h(in0, 6, in[144]); 150 in0 = __msa_insert_h(in0, 7, in[208]); 151 in1 = __msa_insert_h(in1, 0, in[ 48]); 152 in1 = __msa_insert_h(in1, 1, in[112]); 153 in1 = __msa_insert_h(in1, 2, in[176]); 154 in1 = __msa_insert_h(in1, 3, in[240]); 155 in1 = __msa_insert_h(in1, 4, in[ 32]); 156 in1 = __msa_insert_h(in1, 5, in[ 96]); 157 in1 = __msa_insert_h(in1, 6, in[160]); 158 in1 = __msa_insert_h(in1, 7, in[224]); 159 ADDSUB2(in0, in1, tmp0, tmp1); 160 VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); 161 ADDSUB2(tmp2, tmp3, tmp0, tmp1); 162 VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1); 163 ADDSUB2(in0, in1, tmp0, tmp1); 164 VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); 165 ADDSUB2(tmp2, tmp3, out0, out1); 166 SRAI_H2_SH(out0, out1, 1); 167 ST_SH2(out0, out1, out, 8); 168 } 169 170 static int TTransform(const uint8_t* in, const uint16_t* w) { 171 int sum; 172 uint32_t in0_m, in1_m, in2_m, in3_m; 173 v16i8 src0; 174 v8i16 in0, in1, tmp0, tmp1, tmp2, tmp3; 175 v4i32 dst0, dst1; 176 const v16i8 zero = { 0 }; 177 const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 }; 178 const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 }; 179 const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 }; 180 const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 }; 181 182 LW4(in, BPS, in0_m, in1_m, in2_m, in3_m); 183 INSERT_W4_SB(in0_m, in1_m, in2_m, in3_m, src0); 184 ILVRL_B2_SH(zero, src0, tmp0, tmp1); 185 VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1); 186 ADDSUB2(in0, in1, tmp0, tmp1); 187 VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); 188 ADDSUB2(tmp2, tmp3, tmp0, tmp1); 189 VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1); 190 ADDSUB2(in0, in1, tmp0, tmp1); 191 VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); 192 ADDSUB2(tmp2, tmp3, tmp0, tmp1); 193 tmp0 = __msa_add_a_h(tmp0, (v8i16)zero); 194 tmp1 = __msa_add_a_h(tmp1, (v8i16)zero); 195 LD_SH2(w, 8, tmp2, tmp3); 196 DOTP_SH2_SW(tmp0, tmp1, tmp2, tmp3, dst0, dst1); 197 dst0 = dst0 + dst1; 198 sum = HADD_SW_S32(dst0); 199 return sum; 200 } 201 202 static int Disto4x4(const uint8_t* const a, const uint8_t* const b, 203 const uint16_t* const w) { 204 const int sum1 = TTransform(a, w); 205 const int sum2 = TTransform(b, w); 206 return abs(sum2 - sum1) >> 5; 207 } 208 209 static int Disto16x16(const uint8_t* const a, const uint8_t* const b, 210 const uint16_t* const w) { 211 int D = 0; 212 int x, y; 213 for (y = 0; y < 16 * BPS; y += 4 * BPS) { 214 for (x = 0; x < 16; x += 4) { 215 D += Disto4x4(a + x + y, b + x + y, w); 216 } 217 } 218 return D; 219 } 220 221 //------------------------------------------------------------------------------ 222 // Histogram 223 224 static void CollectHistogram(const uint8_t* ref, const uint8_t* pred, 225 int start_block, int end_block, 226 VP8Histogram* const histo) { 227 int j; 228 int distribution[MAX_COEFF_THRESH + 1] = { 0 }; 229 for (j = start_block; j < end_block; ++j) { 230 int16_t out[16]; 231 VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); 232 { 233 int k; 234 v8i16 coeff0, coeff1; 235 const v8i16 zero = { 0 }; 236 const v8i16 max_coeff_thr = __msa_ldi_h(MAX_COEFF_THRESH); 237 LD_SH2(&out[0], 8, coeff0, coeff1); 238 coeff0 = __msa_add_a_h(coeff0, zero); 239 coeff1 = __msa_add_a_h(coeff1, zero); 240 SRAI_H2_SH(coeff0, coeff1, 3); 241 coeff0 = __msa_min_s_h(coeff0, max_coeff_thr); 242 coeff1 = __msa_min_s_h(coeff1, max_coeff_thr); 243 ST_SH2(coeff0, coeff1, &out[0], 8); 244 for (k = 0; k < 16; ++k) { 245 ++distribution[out[k]]; 246 } 247 } 248 } 249 VP8SetHistogramData(distribution, histo); 250 } 251 252 //------------------------------------------------------------------------------ 253 // Intra predictions 254 255 // luma 4x4 prediction 256 257 #define DST(x, y) dst[(x) + (y) * BPS] 258 #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2) 259 #define AVG2(a, b) (((a) + (b) + 1) >> 1) 260 261 static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) { // vertical 262 const uint64_t val_m = LD(top - 1); 263 const v16u8 A = (v16u8)__msa_insert_d((v2i64)A, 0, val_m); 264 const v16u8 B = SLDI_UB(A, A, 1); 265 const v16u8 C = SLDI_UB(A, A, 2); 266 const v16u8 AC = __msa_ave_u_b(A, C); 267 const v16u8 B2 = __msa_ave_u_b(B, B); 268 const v16u8 R = __msa_aver_u_b(AC, B2); 269 const uint32_t out = __msa_copy_s_w((v4i32)R, 0); 270 SW4(out, out, out, out, dst, BPS); 271 } 272 273 static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) { // horizontal 274 const int X = top[-1]; 275 const int I = top[-2]; 276 const int J = top[-3]; 277 const int K = top[-4]; 278 const int L = top[-5]; 279 WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J)); 280 WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K)); 281 WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L)); 282 WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L)); 283 } 284 285 static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) { 286 uint32_t dc = 4; 287 int i; 288 for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i]; 289 dc >>= 3; 290 dc = dc | (dc << 8) | (dc << 16) | (dc << 24); 291 SW4(dc, dc, dc, dc, dst, BPS); 292 } 293 294 static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) { 295 const uint64_t val_m = LD(top - 5); 296 const v16u8 A1 = (v16u8)__msa_insert_d((v2i64)A1, 0, val_m); 297 const v16u8 A = (v16u8)__msa_insert_b((v16i8)A1, 8, top[3]); 298 const v16u8 B = SLDI_UB(A, A, 1); 299 const v16u8 C = SLDI_UB(A, A, 2); 300 const v16u8 AC = __msa_ave_u_b(A, C); 301 const v16u8 B2 = __msa_ave_u_b(B, B); 302 const v16u8 R0 = __msa_aver_u_b(AC, B2); 303 const v16u8 R1 = SLDI_UB(R0, R0, 1); 304 const v16u8 R2 = SLDI_UB(R1, R1, 1); 305 const v16u8 R3 = SLDI_UB(R2, R2, 1); 306 const uint32_t val0 = __msa_copy_s_w((v4i32)R0, 0); 307 const uint32_t val1 = __msa_copy_s_w((v4i32)R1, 0); 308 const uint32_t val2 = __msa_copy_s_w((v4i32)R2, 0); 309 const uint32_t val3 = __msa_copy_s_w((v4i32)R3, 0); 310 SW4(val3, val2, val1, val0, dst, BPS); 311 } 312 313 static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) { 314 const uint64_t val_m = LD(top); 315 const v16u8 A = (v16u8)__msa_insert_d((v2i64)A, 0, val_m); 316 const v16u8 B = SLDI_UB(A, A, 1); 317 const v16u8 C1 = SLDI_UB(A, A, 2); 318 const v16u8 C = (v16u8)__msa_insert_b((v16i8)C1, 6, top[7]); 319 const v16u8 AC = __msa_ave_u_b(A, C); 320 const v16u8 B2 = __msa_ave_u_b(B, B); 321 const v16u8 R0 = __msa_aver_u_b(AC, B2); 322 const v16u8 R1 = SLDI_UB(R0, R0, 1); 323 const v16u8 R2 = SLDI_UB(R1, R1, 1); 324 const v16u8 R3 = SLDI_UB(R2, R2, 1); 325 const uint32_t val0 = __msa_copy_s_w((v4i32)R0, 0); 326 const uint32_t val1 = __msa_copy_s_w((v4i32)R1, 0); 327 const uint32_t val2 = __msa_copy_s_w((v4i32)R2, 0); 328 const uint32_t val3 = __msa_copy_s_w((v4i32)R3, 0); 329 SW4(val0, val1, val2, val3, dst, BPS); 330 } 331 332 static WEBP_INLINE void VR4(uint8_t* dst, const uint8_t* top) { 333 const int X = top[-1]; 334 const int I = top[-2]; 335 const int J = top[-3]; 336 const int K = top[-4]; 337 const int A = top[0]; 338 const int B = top[1]; 339 const int C = top[2]; 340 const int D = top[3]; 341 DST(0, 0) = DST(1, 2) = AVG2(X, A); 342 DST(1, 0) = DST(2, 2) = AVG2(A, B); 343 DST(2, 0) = DST(3, 2) = AVG2(B, C); 344 DST(3, 0) = AVG2(C, D); 345 DST(0, 3) = AVG3(K, J, I); 346 DST(0, 2) = AVG3(J, I, X); 347 DST(0, 1) = DST(1, 3) = AVG3(I, X, A); 348 DST(1, 1) = DST(2, 3) = AVG3(X, A, B); 349 DST(2, 1) = DST(3, 3) = AVG3(A, B, C); 350 DST(3, 1) = AVG3(B, C, D); 351 } 352 353 static WEBP_INLINE void VL4(uint8_t* dst, const uint8_t* top) { 354 const int A = top[0]; 355 const int B = top[1]; 356 const int C = top[2]; 357 const int D = top[3]; 358 const int E = top[4]; 359 const int F = top[5]; 360 const int G = top[6]; 361 const int H = top[7]; 362 DST(0, 0) = AVG2(A, B); 363 DST(1, 0) = DST(0, 2) = AVG2(B, C); 364 DST(2, 0) = DST(1, 2) = AVG2(C, D); 365 DST(3, 0) = DST(2, 2) = AVG2(D, E); 366 DST(0, 1) = AVG3(A, B, C); 367 DST(1, 1) = DST(0, 3) = AVG3(B, C, D); 368 DST(2, 1) = DST(1, 3) = AVG3(C, D, E); 369 DST(3, 1) = DST(2, 3) = AVG3(D, E, F); 370 DST(3, 2) = AVG3(E, F, G); 371 DST(3, 3) = AVG3(F, G, H); 372 } 373 374 static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) { 375 const int I = top[-2]; 376 const int J = top[-3]; 377 const int K = top[-4]; 378 const int L = top[-5]; 379 DST(0, 0) = AVG2(I, J); 380 DST(2, 0) = DST(0, 1) = AVG2(J, K); 381 DST(2, 1) = DST(0, 2) = AVG2(K, L); 382 DST(1, 0) = AVG3(I, J, K); 383 DST(3, 0) = DST(1, 1) = AVG3(J, K, L); 384 DST(3, 1) = DST(1, 2) = AVG3(K, L, L); 385 DST(3, 2) = DST(2, 2) = 386 DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L; 387 } 388 389 static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) { 390 const int X = top[-1]; 391 const int I = top[-2]; 392 const int J = top[-3]; 393 const int K = top[-4]; 394 const int L = top[-5]; 395 const int A = top[0]; 396 const int B = top[1]; 397 const int C = top[2]; 398 DST(0, 0) = DST(2, 1) = AVG2(I, X); 399 DST(0, 1) = DST(2, 2) = AVG2(J, I); 400 DST(0, 2) = DST(2, 3) = AVG2(K, J); 401 DST(0, 3) = AVG2(L, K); 402 DST(3, 0) = AVG3(A, B, C); 403 DST(2, 0) = AVG3(X, A, B); 404 DST(1, 0) = DST(3, 1) = AVG3(I, X, A); 405 DST(1, 1) = DST(3, 2) = AVG3(J, I, X); 406 DST(1, 2) = DST(3, 3) = AVG3(K, J, I); 407 DST(1, 3) = AVG3(L, K, J); 408 } 409 410 static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) { 411 const v16i8 zero = { 0 }; 412 const v8i16 TL = (v8i16)__msa_fill_h(top[-1]); 413 const v8i16 L0 = (v8i16)__msa_fill_h(top[-2]); 414 const v8i16 L1 = (v8i16)__msa_fill_h(top[-3]); 415 const v8i16 L2 = (v8i16)__msa_fill_h(top[-4]); 416 const v8i16 L3 = (v8i16)__msa_fill_h(top[-5]); 417 const v16u8 T1 = LD_UB(top); 418 const v8i16 T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1); 419 const v8i16 d = T - TL; 420 v8i16 r0, r1, r2, r3; 421 ADD4(d, L0, d, L1, d, L2, d, L3, r0, r1, r2, r3); 422 CLIP_SH4_0_255(r0, r1, r2, r3); 423 PCKEV_ST4x4_UB(r0, r1, r2, r3, dst, BPS); 424 } 425 426 #undef DST 427 #undef AVG3 428 #undef AVG2 429 430 static void Intra4Preds(uint8_t* dst, const uint8_t* top) { 431 DC4(I4DC4 + dst, top); 432 TM4(I4TM4 + dst, top); 433 VE4(I4VE4 + dst, top); 434 HE4(I4HE4 + dst, top); 435 RD4(I4RD4 + dst, top); 436 VR4(I4VR4 + dst, top); 437 LD4(I4LD4 + dst, top); 438 VL4(I4VL4 + dst, top); 439 HD4(I4HD4 + dst, top); 440 HU4(I4HU4 + dst, top); 441 } 442 443 // luma 16x16 prediction 444 445 #define STORE16x16(out, dst) do { \ 446 ST_UB8(out, out, out, out, out, out, out, out, dst + 0 * BPS, BPS); \ 447 ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS); \ 448 } while (0) 449 450 static WEBP_INLINE void VerticalPred16x16(uint8_t* dst, const uint8_t* top) { 451 if (top != NULL) { 452 const v16u8 out = LD_UB(top); 453 STORE16x16(out, dst); 454 } else { 455 const v16u8 out = (v16u8)__msa_fill_b(0x7f); 456 STORE16x16(out, dst); 457 } 458 } 459 460 static WEBP_INLINE void HorizontalPred16x16(uint8_t* dst, 461 const uint8_t* left) { 462 if (left != NULL) { 463 int j; 464 for (j = 0; j < 16; j += 4) { 465 const v16u8 L0 = (v16u8)__msa_fill_b(left[0]); 466 const v16u8 L1 = (v16u8)__msa_fill_b(left[1]); 467 const v16u8 L2 = (v16u8)__msa_fill_b(left[2]); 468 const v16u8 L3 = (v16u8)__msa_fill_b(left[3]); 469 ST_UB4(L0, L1, L2, L3, dst, BPS); 470 dst += 4 * BPS; 471 left += 4; 472 } 473 } else { 474 const v16u8 out = (v16u8)__msa_fill_b(0x81); 475 STORE16x16(out, dst); 476 } 477 } 478 479 static WEBP_INLINE void TrueMotion16x16(uint8_t* dst, const uint8_t* left, 480 const uint8_t* top) { 481 if (left != NULL) { 482 if (top != NULL) { 483 int j; 484 v8i16 d1, d2; 485 const v16i8 zero = { 0 }; 486 const v8i16 TL = (v8i16)__msa_fill_h(left[-1]); 487 const v16u8 T = LD_UB(top); 488 ILVRL_B2_SH(zero, T, d1, d2); 489 SUB2(d1, TL, d2, TL, d1, d2); 490 for (j = 0; j < 16; j += 4) { 491 v16i8 t0, t1, t2, t3; 492 v8i16 r0, r1, r2, r3, r4, r5, r6, r7; 493 const v8i16 L0 = (v8i16)__msa_fill_h(left[j + 0]); 494 const v8i16 L1 = (v8i16)__msa_fill_h(left[j + 1]); 495 const v8i16 L2 = (v8i16)__msa_fill_h(left[j + 2]); 496 const v8i16 L3 = (v8i16)__msa_fill_h(left[j + 3]); 497 ADD4(d1, L0, d1, L1, d1, L2, d1, L3, r0, r1, r2, r3); 498 ADD4(d2, L0, d2, L1, d2, L2, d2, L3, r4, r5, r6, r7); 499 CLIP_SH4_0_255(r0, r1, r2, r3); 500 CLIP_SH4_0_255(r4, r5, r6, r7); 501 PCKEV_B4_SB(r4, r0, r5, r1, r6, r2, r7, r3, t0, t1, t2, t3); 502 ST_SB4(t0, t1, t2, t3, dst, BPS); 503 dst += 4 * BPS; 504 } 505 } else { 506 HorizontalPred16x16(dst, left); 507 } 508 } else { 509 if (top != NULL) { 510 VerticalPred16x16(dst, top); 511 } else { 512 const v16u8 out = (v16u8)__msa_fill_b(0x81); 513 STORE16x16(out, dst); 514 } 515 } 516 } 517 518 static WEBP_INLINE void DCMode16x16(uint8_t* dst, const uint8_t* left, 519 const uint8_t* top) { 520 int DC; 521 v16u8 out; 522 if (top != NULL && left != NULL) { 523 const v16u8 rtop = LD_UB(top); 524 const v8u16 dctop = __msa_hadd_u_h(rtop, rtop); 525 const v16u8 rleft = LD_UB(left); 526 const v8u16 dcleft = __msa_hadd_u_h(rleft, rleft); 527 const v8u16 dctemp = dctop + dcleft; 528 DC = HADD_UH_U32(dctemp); 529 DC = (DC + 16) >> 5; 530 } else if (left != NULL) { // left but no top 531 const v16u8 rleft = LD_UB(left); 532 const v8u16 dcleft = __msa_hadd_u_h(rleft, rleft); 533 DC = HADD_UH_U32(dcleft); 534 DC = (DC + DC + 16) >> 5; 535 } else if (top != NULL) { // top but no left 536 const v16u8 rtop = LD_UB(top); 537 const v8u16 dctop = __msa_hadd_u_h(rtop, rtop); 538 DC = HADD_UH_U32(dctop); 539 DC = (DC + DC + 16) >> 5; 540 } else { // no top, no left, nothing. 541 DC = 0x80; 542 } 543 out = (v16u8)__msa_fill_b(DC); 544 STORE16x16(out, dst); 545 } 546 547 static void Intra16Preds(uint8_t* dst, 548 const uint8_t* left, const uint8_t* top) { 549 DCMode16x16(I16DC16 + dst, left, top); 550 VerticalPred16x16(I16VE16 + dst, top); 551 HorizontalPred16x16(I16HE16 + dst, left); 552 TrueMotion16x16(I16TM16 + dst, left, top); 553 } 554 555 // Chroma 8x8 prediction 556 557 #define CALC_DC8(in, out) do { \ 558 const v8u16 temp0 = __msa_hadd_u_h(in, in); \ 559 const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0); \ 560 const v2i64 temp2 = (v2i64)__msa_hadd_u_d(temp1, temp1); \ 561 const v2i64 temp3 = __msa_splati_d(temp2, 1); \ 562 const v2i64 temp4 = temp3 + temp2; \ 563 const v16i8 temp5 = (v16i8)__msa_srari_d(temp4, 4); \ 564 const v2i64 temp6 = (v2i64)__msa_splati_b(temp5, 0); \ 565 out = __msa_copy_s_d(temp6, 0); \ 566 } while (0) 567 568 #define STORE8x8(out, dst) do { \ 569 SD4(out, out, out, out, dst + 0 * BPS, BPS); \ 570 SD4(out, out, out, out, dst + 4 * BPS, BPS); \ 571 } while (0) 572 573 static WEBP_INLINE void VerticalPred8x8(uint8_t* dst, const uint8_t* top) { 574 if (top != NULL) { 575 const uint64_t out = LD(top); 576 STORE8x8(out, dst); 577 } else { 578 const uint64_t out = 0x7f7f7f7f7f7f7f7fULL; 579 STORE8x8(out, dst); 580 } 581 } 582 583 static WEBP_INLINE void HorizontalPred8x8(uint8_t* dst, const uint8_t* left) { 584 if (left != NULL) { 585 int j; 586 for (j = 0; j < 8; j += 4) { 587 const v16u8 L0 = (v16u8)__msa_fill_b(left[0]); 588 const v16u8 L1 = (v16u8)__msa_fill_b(left[1]); 589 const v16u8 L2 = (v16u8)__msa_fill_b(left[2]); 590 const v16u8 L3 = (v16u8)__msa_fill_b(left[3]); 591 const uint64_t out0 = __msa_copy_s_d((v2i64)L0, 0); 592 const uint64_t out1 = __msa_copy_s_d((v2i64)L1, 0); 593 const uint64_t out2 = __msa_copy_s_d((v2i64)L2, 0); 594 const uint64_t out3 = __msa_copy_s_d((v2i64)L3, 0); 595 SD4(out0, out1, out2, out3, dst, BPS); 596 dst += 4 * BPS; 597 left += 4; 598 } 599 } else { 600 const uint64_t out = 0x8181818181818181ULL; 601 STORE8x8(out, dst); 602 } 603 } 604 605 static WEBP_INLINE void TrueMotion8x8(uint8_t* dst, const uint8_t* left, 606 const uint8_t* top) { 607 if (left != NULL) { 608 if (top != NULL) { 609 int j; 610 const v8i16 TL = (v8i16)__msa_fill_h(left[-1]); 611 const v16u8 T1 = LD_UB(top); 612 const v16i8 zero = { 0 }; 613 const v8i16 T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1); 614 const v8i16 d = T - TL; 615 for (j = 0; j < 8; j += 4) { 616 uint64_t out0, out1, out2, out3; 617 v16i8 t0, t1; 618 v8i16 r0 = (v8i16)__msa_fill_h(left[j + 0]); 619 v8i16 r1 = (v8i16)__msa_fill_h(left[j + 1]); 620 v8i16 r2 = (v8i16)__msa_fill_h(left[j + 2]); 621 v8i16 r3 = (v8i16)__msa_fill_h(left[j + 3]); 622 ADD4(d, r0, d, r1, d, r2, d, r3, r0, r1, r2, r3); 623 CLIP_SH4_0_255(r0, r1, r2, r3); 624 PCKEV_B2_SB(r1, r0, r3, r2, t0, t1); 625 out0 = __msa_copy_s_d((v2i64)t0, 0); 626 out1 = __msa_copy_s_d((v2i64)t0, 1); 627 out2 = __msa_copy_s_d((v2i64)t1, 0); 628 out3 = __msa_copy_s_d((v2i64)t1, 1); 629 SD4(out0, out1, out2, out3, dst, BPS); 630 dst += 4 * BPS; 631 } 632 } else { 633 HorizontalPred8x8(dst, left); 634 } 635 } else { 636 if (top != NULL) { 637 VerticalPred8x8(dst, top); 638 } else { 639 const uint64_t out = 0x8181818181818181ULL; 640 STORE8x8(out, dst); 641 } 642 } 643 } 644 645 static WEBP_INLINE void DCMode8x8(uint8_t* dst, const uint8_t* left, 646 const uint8_t* top) { 647 uint64_t out; 648 v16u8 src; 649 if (top != NULL && left != NULL) { 650 const uint64_t left_m = LD(left); 651 const uint64_t top_m = LD(top); 652 INSERT_D2_UB(left_m, top_m, src); 653 CALC_DC8(src, out); 654 } else if (left != NULL) { // left but no top 655 const uint64_t left_m = LD(left); 656 INSERT_D2_UB(left_m, left_m, src); 657 CALC_DC8(src, out); 658 } else if (top != NULL) { // top but no left 659 const uint64_t top_m = LD(top); 660 INSERT_D2_UB(top_m, top_m, src); 661 CALC_DC8(src, out); 662 } else { // no top, no left, nothing. 663 src = (v16u8)__msa_fill_b(0x80); 664 out = __msa_copy_s_d((v2i64)src, 0); 665 } 666 STORE8x8(out, dst); 667 } 668 669 static void IntraChromaPreds(uint8_t* dst, const uint8_t* left, 670 const uint8_t* top) { 671 // U block 672 DCMode8x8(C8DC8 + dst, left, top); 673 VerticalPred8x8(C8VE8 + dst, top); 674 HorizontalPred8x8(C8HE8 + dst, left); 675 TrueMotion8x8(C8TM8 + dst, left, top); 676 // V block 677 dst += 8; 678 if (top != NULL) top += 8; 679 if (left != NULL) left += 16; 680 DCMode8x8(C8DC8 + dst, left, top); 681 VerticalPred8x8(C8VE8 + dst, top); 682 HorizontalPred8x8(C8HE8 + dst, left); 683 TrueMotion8x8(C8TM8 + dst, left, top); 684 } 685 686 //------------------------------------------------------------------------------ 687 // Metric 688 689 #define PACK_DOTP_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do { \ 690 v16u8 tmp0, tmp1; \ 691 v8i16 tmp2, tmp3; \ 692 ILVRL_B2_UB(in0, in1, tmp0, tmp1); \ 693 HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ 694 DOTP_SH2_SW(tmp2, tmp3, tmp2, tmp3, out0, out1); \ 695 ILVRL_B2_UB(in2, in3, tmp0, tmp1); \ 696 HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ 697 DOTP_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3); \ 698 } while (0) 699 700 #define PACK_DPADD_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do { \ 701 v16u8 tmp0, tmp1; \ 702 v8i16 tmp2, tmp3; \ 703 ILVRL_B2_UB(in0, in1, tmp0, tmp1); \ 704 HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ 705 DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out0, out1); \ 706 ILVRL_B2_UB(in2, in3, tmp0, tmp1); \ 707 HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ 708 DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3); \ 709 } while (0) 710 711 static int SSE16x16(const uint8_t* a, const uint8_t* b) { 712 uint32_t sum; 713 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; 714 v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; 715 v4i32 out0, out1, out2, out3; 716 717 LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); 718 LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 719 PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3); 720 PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3); 721 PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3); 722 PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3); 723 a += 8 * BPS; 724 b += 8 * BPS; 725 LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); 726 LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 727 PACK_DPADD_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3); 728 PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3); 729 PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3); 730 PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3); 731 out0 += out1; 732 out2 += out3; 733 out0 += out2; 734 sum = HADD_SW_S32(out0); 735 return sum; 736 } 737 738 static int SSE16x8(const uint8_t* a, const uint8_t* b) { 739 uint32_t sum; 740 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; 741 v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; 742 v4i32 out0, out1, out2, out3; 743 744 LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); 745 LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 746 PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3); 747 PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3); 748 PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3); 749 PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3); 750 out0 += out1; 751 out2 += out3; 752 out0 += out2; 753 sum = HADD_SW_S32(out0); 754 return sum; 755 } 756 757 static int SSE8x8(const uint8_t* a, const uint8_t* b) { 758 uint32_t sum; 759 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; 760 v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; 761 v16u8 t0, t1, t2, t3; 762 v4i32 out0, out1, out2, out3; 763 764 LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); 765 LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); 766 ILVR_B4_UB(src0, src1, src2, src3, ref0, ref1, ref2, ref3, t0, t1, t2, t3); 767 PACK_DOTP_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3); 768 ILVR_B4_UB(src4, src5, src6, src7, ref4, ref5, ref6, ref7, t0, t1, t2, t3); 769 PACK_DPADD_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3); 770 out0 += out1; 771 out2 += out3; 772 out0 += out2; 773 sum = HADD_SW_S32(out0); 774 return sum; 775 } 776 777 static int SSE4x4(const uint8_t* a, const uint8_t* b) { 778 uint32_t sum = 0; 779 uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3; 780 v16u8 src, ref, tmp0, tmp1; 781 v8i16 diff0, diff1; 782 v4i32 out0, out1; 783 784 LW4(a, BPS, src0, src1, src2, src3); 785 LW4(b, BPS, ref0, ref1, ref2, ref3); 786 INSERT_W4_UB(src0, src1, src2, src3, src); 787 INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); 788 ILVRL_B2_UB(src, ref, tmp0, tmp1); 789 HSUB_UB2_SH(tmp0, tmp1, diff0, diff1); 790 DOTP_SH2_SW(diff0, diff1, diff0, diff1, out0, out1); 791 out0 += out1; 792 sum = HADD_SW_S32(out0); 793 return sum; 794 } 795 796 //------------------------------------------------------------------------------ 797 // Quantization 798 799 static int QuantizeBlock(int16_t in[16], int16_t out[16], 800 const VP8Matrix* const mtx) { 801 int sum; 802 v8i16 in0, in1, sh0, sh1, out0, out1; 803 v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, sign0, sign1; 804 v4i32 s0, s1, s2, s3, b0, b1, b2, b3, t0, t1, t2, t3; 805 const v8i16 zero = { 0 }; 806 const v8i16 zigzag0 = { 0, 1, 4, 8, 5, 2, 3, 6 }; 807 const v8i16 zigzag1 = { 9, 12, 13, 10, 7, 11, 14, 15 }; 808 const v8i16 maxlevel = __msa_fill_h(MAX_LEVEL); 809 810 LD_SH2(&in[0], 8, in0, in1); 811 LD_SH2(&mtx->sharpen_[0], 8, sh0, sh1); 812 tmp4 = __msa_add_a_h(in0, zero); 813 tmp5 = __msa_add_a_h(in1, zero); 814 ILVRL_H2_SH(sh0, tmp4, tmp0, tmp1); 815 ILVRL_H2_SH(sh1, tmp5, tmp2, tmp3); 816 HADD_SH4_SW(tmp0, tmp1, tmp2, tmp3, s0, s1, s2, s3); 817 sign0 = (in0 < zero); 818 sign1 = (in1 < zero); // sign 819 LD_SH2(&mtx->iq_[0], 8, tmp0, tmp1); // iq 820 ILVRL_H2_SW(zero, tmp0, t0, t1); 821 ILVRL_H2_SW(zero, tmp1, t2, t3); 822 LD_SW4(&mtx->bias_[0], 4, b0, b1, b2, b3); // bias 823 MUL4(t0, s0, t1, s1, t2, s2, t3, s3, t0, t1, t2, t3); 824 ADD4(b0, t0, b1, t1, b2, t2, b3, t3, b0, b1, b2, b3); 825 SRAI_W4_SW(b0, b1, b2, b3, 17); 826 PCKEV_H2_SH(b1, b0, b3, b2, tmp2, tmp3); 827 tmp0 = (tmp2 > maxlevel); 828 tmp1 = (tmp3 > maxlevel); 829 tmp2 = (v8i16)__msa_bmnz_v((v16u8)tmp2, (v16u8)maxlevel, (v16u8)tmp0); 830 tmp3 = (v8i16)__msa_bmnz_v((v16u8)tmp3, (v16u8)maxlevel, (v16u8)tmp1); 831 SUB2(0, tmp2, 0, tmp3, tmp0, tmp1); 832 tmp2 = (v8i16)__msa_bmnz_v((v16u8)tmp2, (v16u8)tmp0, (v16u8)sign0); 833 tmp3 = (v8i16)__msa_bmnz_v((v16u8)tmp3, (v16u8)tmp1, (v16u8)sign1); 834 LD_SW4(&mtx->zthresh_[0], 4, t0, t1, t2, t3); // zthresh 835 t0 = (s0 > t0); 836 t1 = (s1 > t1); 837 t2 = (s2 > t2); 838 t3 = (s3 > t3); 839 PCKEV_H2_SH(t1, t0, t3, t2, tmp0, tmp1); 840 tmp4 = (v8i16)__msa_bmnz_v((v16u8)zero, (v16u8)tmp2, (v16u8)tmp0); 841 tmp5 = (v8i16)__msa_bmnz_v((v16u8)zero, (v16u8)tmp3, (v16u8)tmp1); 842 LD_SH2(&mtx->q_[0], 8, tmp0, tmp1); 843 MUL2(tmp4, tmp0, tmp5, tmp1, in0, in1); 844 VSHF_H2_SH(tmp4, tmp5, tmp4, tmp5, zigzag0, zigzag1, out0, out1); 845 ST_SH2(in0, in1, &in[0], 8); 846 ST_SH2(out0, out1, &out[0], 8); 847 out0 = __msa_add_a_h(out0, out1); 848 sum = HADD_SH_S32(out0); 849 return (sum > 0); 850 } 851 852 static int Quantize2Blocks(int16_t in[32], int16_t out[32], 853 const VP8Matrix* const mtx) { 854 int nz; 855 nz = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; 856 nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; 857 return nz; 858 } 859 860 //------------------------------------------------------------------------------ 861 // Entry point 862 863 extern void VP8EncDspInitMSA(void); 864 865 WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitMSA(void) { 866 VP8ITransform = ITransform; 867 VP8FTransform = FTransform; 868 VP8FTransformWHT = FTransformWHT; 869 870 VP8TDisto4x4 = Disto4x4; 871 VP8TDisto16x16 = Disto16x16; 872 VP8CollectHistogram = CollectHistogram; 873 874 VP8EncPredLuma4 = Intra4Preds; 875 VP8EncPredLuma16 = Intra16Preds; 876 VP8EncPredChroma8 = IntraChromaPreds; 877 878 VP8SSE16x16 = SSE16x16; 879 VP8SSE16x8 = SSE16x8; 880 VP8SSE8x8 = SSE8x8; 881 VP8SSE4x4 = SSE4x4; 882 883 VP8EncQuantizeBlock = QuantizeBlock; 884 VP8EncQuantize2Blocks = Quantize2Blocks; 885 VP8EncQuantizeBlockWHT = QuantizeBlock; 886 } 887 888 #else // !WEBP_USE_MSA 889 890 WEBP_DSP_INIT_STUB(VP8EncDspInitMSA) 891 892 #endif // WEBP_USE_MSA 893
