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 11#include <assert.h> 12#include "./vpx_dsp_rtcd.h" 13#include "vpx_dsp/mips/vpx_convolve_msa.h" 14 15const uint8_t mc_filt_mask_arr[16 * 3] = { 16 /* 8 width cases */ 17 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 18 /* 4 width cases */ 19 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20, 20 /* 4 width cases */ 21 8, 9, 9, 10, 10, 11, 11, 12, 24, 25, 25, 26, 26, 27, 27, 28 22}; 23 24static void common_hv_8ht_8vt_4w_msa(const uint8_t *src, int32_t src_stride, 25 uint8_t *dst, int32_t dst_stride, 26 int8_t *filter_horiz, int8_t *filter_vert, 27 int32_t height) { 28 uint32_t loop_cnt; 29 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; 30 v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3; 31 v16u8 mask0, mask1, mask2, mask3, out; 32 v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6; 33 v8i16 hz_out7, hz_out8, hz_out9, tmp0, tmp1, out0, out1, out2, out3, out4; 34 v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3; 35 36 mask0 = LD_UB(&mc_filt_mask_arr[16]); 37 src -= (3 + 3 * src_stride); 38 39 /* rearranging filter */ 40 filt = LD_SH(filter_horiz); 41 SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3); 42 43 mask1 = mask0 + 2; 44 mask2 = mask0 + 4; 45 mask3 = mask0 + 6; 46 47 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6); 48 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6); 49 src += (7 * src_stride); 50 51 hz_out0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0, 52 filt_hz1, filt_hz2, filt_hz3); 53 hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0, 54 filt_hz1, filt_hz2, filt_hz3); 55 hz_out4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0, 56 filt_hz1, filt_hz2, filt_hz3); 57 hz_out5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0, 58 filt_hz1, filt_hz2, filt_hz3); 59 SLDI_B2_SH(hz_out2, hz_out4, hz_out0, hz_out2, hz_out1, hz_out3, 8); 60 61 filt = LD_SH(filter_vert); 62 SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3); 63 64 ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1); 65 out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4); 66 67 for (loop_cnt = (height >> 2); loop_cnt--;) { 68 LD_SB4(src, src_stride, src7, src8, src9, src10); 69 XORI_B4_128_SB(src7, src8, src9, src10); 70 src += (4 * src_stride); 71 72 hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3, filt_hz0, 73 filt_hz1, filt_hz2, filt_hz3); 74 hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8); 75 out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6); 76 tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1, 77 filt_vt2, filt_vt3); 78 79 hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3, filt_hz0, 80 filt_hz1, filt_hz2, filt_hz3); 81 hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8); 82 out4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8); 83 tmp1 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out4, filt_vt0, filt_vt1, 84 filt_vt2, filt_vt3); 85 SRARI_H2_SH(tmp0, tmp1, FILTER_BITS); 86 SAT_SH2_SH(tmp0, tmp1, 7); 87 out = PCKEV_XORI128_UB(tmp0, tmp1); 88 ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride); 89 dst += (4 * dst_stride); 90 91 hz_out5 = hz_out9; 92 out0 = out2; 93 out1 = out3; 94 out2 = out4; 95 } 96} 97 98static void common_hv_8ht_8vt_8w_msa(const uint8_t *src, int32_t src_stride, 99 uint8_t *dst, int32_t dst_stride, 100 int8_t *filter_horiz, int8_t *filter_vert, 101 int32_t height) { 102 uint32_t loop_cnt; 103 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; 104 v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3; 105 v16u8 mask0, mask1, mask2, mask3, vec0, vec1; 106 v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3; 107 v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6; 108 v8i16 hz_out7, hz_out8, hz_out9, hz_out10, tmp0, tmp1, tmp2, tmp3; 109 v8i16 out0, out1, out2, out3, out4, out5, out6, out7, out8, out9; 110 111 mask0 = LD_UB(&mc_filt_mask_arr[0]); 112 src -= (3 + 3 * src_stride); 113 114 /* rearranging filter */ 115 filt = LD_SH(filter_horiz); 116 SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3); 117 118 mask1 = mask0 + 2; 119 mask2 = mask0 + 4; 120 mask3 = mask0 + 6; 121 122 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6); 123 src += (7 * src_stride); 124 125 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6); 126 hz_out0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0, 127 filt_hz1, filt_hz2, filt_hz3); 128 hz_out1 = HORIZ_8TAP_FILT(src1, src1, mask0, mask1, mask2, mask3, filt_hz0, 129 filt_hz1, filt_hz2, filt_hz3); 130 hz_out2 = HORIZ_8TAP_FILT(src2, src2, mask0, mask1, mask2, mask3, filt_hz0, 131 filt_hz1, filt_hz2, filt_hz3); 132 hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0, 133 filt_hz1, filt_hz2, filt_hz3); 134 hz_out4 = HORIZ_8TAP_FILT(src4, src4, mask0, mask1, mask2, mask3, filt_hz0, 135 filt_hz1, filt_hz2, filt_hz3); 136 hz_out5 = HORIZ_8TAP_FILT(src5, src5, mask0, mask1, mask2, mask3, filt_hz0, 137 filt_hz1, filt_hz2, filt_hz3); 138 hz_out6 = HORIZ_8TAP_FILT(src6, src6, mask0, mask1, mask2, mask3, filt_hz0, 139 filt_hz1, filt_hz2, filt_hz3); 140 141 filt = LD_SH(filter_vert); 142 SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3); 143 144 ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1); 145 ILVEV_B2_SH(hz_out4, hz_out5, hz_out1, hz_out2, out2, out4); 146 ILVEV_B2_SH(hz_out3, hz_out4, hz_out5, hz_out6, out5, out6); 147 148 for (loop_cnt = (height >> 2); loop_cnt--;) { 149 LD_SB4(src, src_stride, src7, src8, src9, src10); 150 src += (4 * src_stride); 151 152 XORI_B4_128_SB(src7, src8, src9, src10); 153 154 hz_out7 = HORIZ_8TAP_FILT(src7, src7, mask0, mask1, mask2, mask3, filt_hz0, 155 filt_hz1, filt_hz2, filt_hz3); 156 out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6); 157 tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1, 158 filt_vt2, filt_vt3); 159 160 hz_out8 = HORIZ_8TAP_FILT(src8, src8, mask0, mask1, mask2, mask3, filt_hz0, 161 filt_hz1, filt_hz2, filt_hz3); 162 out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7); 163 tmp1 = FILT_8TAP_DPADD_S_H(out4, out5, out6, out7, filt_vt0, filt_vt1, 164 filt_vt2, filt_vt3); 165 166 hz_out9 = HORIZ_8TAP_FILT(src9, src9, mask0, mask1, mask2, mask3, filt_hz0, 167 filt_hz1, filt_hz2, filt_hz3); 168 out8 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8); 169 tmp2 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out8, filt_vt0, filt_vt1, 170 filt_vt2, filt_vt3); 171 172 hz_out10 = HORIZ_8TAP_FILT(src10, src10, mask0, mask1, mask2, mask3, 173 filt_hz0, filt_hz1, filt_hz2, filt_hz3); 174 out9 = (v8i16)__msa_ilvev_b((v16i8)hz_out10, (v16i8)hz_out9); 175 tmp3 = FILT_8TAP_DPADD_S_H(out5, out6, out7, out9, filt_vt0, filt_vt1, 176 filt_vt2, filt_vt3); 177 SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); 178 SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7); 179 vec0 = PCKEV_XORI128_UB(tmp0, tmp1); 180 vec1 = PCKEV_XORI128_UB(tmp2, tmp3); 181 ST8x4_UB(vec0, vec1, dst, dst_stride); 182 dst += (4 * dst_stride); 183 184 hz_out6 = hz_out10; 185 out0 = out2; 186 out1 = out3; 187 out2 = out8; 188 out4 = out6; 189 out5 = out7; 190 out6 = out9; 191 } 192} 193 194static void common_hv_8ht_8vt_16w_msa(const uint8_t *src, int32_t src_stride, 195 uint8_t *dst, int32_t dst_stride, 196 int8_t *filter_horiz, int8_t *filter_vert, 197 int32_t height) { 198 int32_t multiple8_cnt; 199 for (multiple8_cnt = 2; multiple8_cnt--;) { 200 common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz, 201 filter_vert, height); 202 src += 8; 203 dst += 8; 204 } 205} 206 207static void common_hv_8ht_8vt_32w_msa(const uint8_t *src, int32_t src_stride, 208 uint8_t *dst, int32_t dst_stride, 209 int8_t *filter_horiz, int8_t *filter_vert, 210 int32_t height) { 211 int32_t multiple8_cnt; 212 for (multiple8_cnt = 4; multiple8_cnt--;) { 213 common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz, 214 filter_vert, height); 215 src += 8; 216 dst += 8; 217 } 218} 219 220static void common_hv_8ht_8vt_64w_msa(const uint8_t *src, int32_t src_stride, 221 uint8_t *dst, int32_t dst_stride, 222 int8_t *filter_horiz, int8_t *filter_vert, 223 int32_t height) { 224 int32_t multiple8_cnt; 225 for (multiple8_cnt = 8; multiple8_cnt--;) { 226 common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz, 227 filter_vert, height); 228 src += 8; 229 dst += 8; 230 } 231} 232 233static void common_hv_2ht_2vt_4x4_msa(const uint8_t *src, int32_t src_stride, 234 uint8_t *dst, int32_t dst_stride, 235 int8_t *filter_horiz, 236 int8_t *filter_vert) { 237 v16i8 src0, src1, src2, src3, src4, mask; 238 v16u8 filt_vt, filt_hz, vec0, vec1, res0, res1; 239 v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, filt, tmp0, tmp1; 240 241 mask = LD_SB(&mc_filt_mask_arr[16]); 242 243 /* rearranging filter */ 244 filt = LD_UH(filter_horiz); 245 filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0); 246 247 filt = LD_UH(filter_vert); 248 filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0); 249 250 LD_SB5(src, src_stride, src0, src1, src2, src3, src4); 251 hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); 252 hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); 253 hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); 254 hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8); 255 hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2); 256 257 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); 258 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); 259 SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); 260 PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1); 261 ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride); 262} 263 264static void common_hv_2ht_2vt_4x8_msa(const uint8_t *src, int32_t src_stride, 265 uint8_t *dst, int32_t dst_stride, 266 int8_t *filter_horiz, 267 int8_t *filter_vert) { 268 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask; 269 v16i8 res0, res1, res2, res3; 270 v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3; 271 v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6; 272 v8u16 hz_out7, hz_out8, vec4, vec5, vec6, vec7, filt; 273 274 mask = LD_SB(&mc_filt_mask_arr[16]); 275 276 /* rearranging filter */ 277 filt = LD_UH(filter_horiz); 278 filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0); 279 280 filt = LD_UH(filter_vert); 281 filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0); 282 283 LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); 284 src += (8 * src_stride); 285 src8 = LD_SB(src); 286 287 hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); 288 hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); 289 hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, FILTER_BITS); 290 hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, FILTER_BITS); 291 hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, FILTER_BITS); 292 SLDI_B3_UH(hz_out2, hz_out4, hz_out6, hz_out0, hz_out2, hz_out4, hz_out1, 293 hz_out3, hz_out5, 8); 294 hz_out7 = (v8u16)__msa_pckod_d((v2i64)hz_out8, (v2i64)hz_out6); 295 296 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); 297 ILVEV_B2_UB(hz_out4, hz_out5, hz_out6, hz_out7, vec2, vec3); 298 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt, vec4, 299 vec5, vec6, vec7); 300 SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS); 301 PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1, res2, 302 res3); 303 ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride); 304 dst += (4 * dst_stride); 305 ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride); 306} 307 308static void common_hv_2ht_2vt_4w_msa(const uint8_t *src, int32_t src_stride, 309 uint8_t *dst, int32_t dst_stride, 310 int8_t *filter_horiz, int8_t *filter_vert, 311 int32_t height) { 312 if (4 == height) { 313 common_hv_2ht_2vt_4x4_msa(src, src_stride, dst, dst_stride, filter_horiz, 314 filter_vert); 315 } else if (8 == height) { 316 common_hv_2ht_2vt_4x8_msa(src, src_stride, dst, dst_stride, filter_horiz, 317 filter_vert); 318 } 319} 320 321static void common_hv_2ht_2vt_8x4_msa(const uint8_t *src, int32_t src_stride, 322 uint8_t *dst, int32_t dst_stride, 323 int8_t *filter_horiz, 324 int8_t *filter_vert) { 325 v16i8 src0, src1, src2, src3, src4, mask, out0, out1; 326 v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3; 327 v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3; 328 v8i16 filt; 329 330 mask = LD_SB(&mc_filt_mask_arr[0]); 331 332 /* rearranging filter */ 333 filt = LD_SH(filter_horiz); 334 filt_hz = (v16u8)__msa_splati_h(filt, 0); 335 336 filt = LD_SH(filter_vert); 337 filt_vt = (v16u8)__msa_splati_h(filt, 0); 338 339 LD_SB5(src, src_stride, src0, src1, src2, src3, src4); 340 341 hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); 342 hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); 343 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); 344 tmp0 = __msa_dotp_u_h(vec0, filt_vt); 345 346 hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); 347 vec1 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); 348 tmp1 = __msa_dotp_u_h(vec1, filt_vt); 349 350 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); 351 vec2 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); 352 tmp2 = __msa_dotp_u_h(vec2, filt_vt); 353 354 hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); 355 vec3 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); 356 tmp3 = __msa_dotp_u_h(vec3, filt_vt); 357 358 SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); 359 PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1); 360 ST8x4_UB(out0, out1, dst, dst_stride); 361} 362 363static void common_hv_2ht_2vt_8x8mult_msa(const uint8_t *src, 364 int32_t src_stride, uint8_t *dst, 365 int32_t dst_stride, 366 int8_t *filter_horiz, 367 int8_t *filter_vert, int32_t height) { 368 uint32_t loop_cnt; 369 v16i8 src0, src1, src2, src3, src4, mask, out0, out1; 370 v16u8 filt_hz, filt_vt, vec0; 371 v8u16 hz_out0, hz_out1, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8; 372 v8i16 filt; 373 374 mask = LD_SB(&mc_filt_mask_arr[0]); 375 376 /* rearranging filter */ 377 filt = LD_SH(filter_horiz); 378 filt_hz = (v16u8)__msa_splati_h(filt, 0); 379 380 filt = LD_SH(filter_vert); 381 filt_vt = (v16u8)__msa_splati_h(filt, 0); 382 383 src0 = LD_SB(src); 384 src += src_stride; 385 386 hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); 387 388 for (loop_cnt = (height >> 3); loop_cnt--;) { 389 LD_SB4(src, src_stride, src1, src2, src3, src4); 390 src += (4 * src_stride); 391 392 hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); 393 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); 394 tmp1 = __msa_dotp_u_h(vec0, filt_vt); 395 396 hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); 397 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); 398 tmp2 = __msa_dotp_u_h(vec0, filt_vt); 399 400 SRARI_H2_UH(tmp1, tmp2, FILTER_BITS); 401 402 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); 403 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); 404 tmp3 = __msa_dotp_u_h(vec0, filt_vt); 405 406 hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); 407 LD_SB4(src, src_stride, src1, src2, src3, src4); 408 src += (4 * src_stride); 409 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); 410 tmp4 = __msa_dotp_u_h(vec0, filt_vt); 411 412 SRARI_H2_UH(tmp3, tmp4, FILTER_BITS); 413 PCKEV_B2_SB(tmp2, tmp1, tmp4, tmp3, out0, out1); 414 ST8x4_UB(out0, out1, dst, dst_stride); 415 dst += (4 * dst_stride); 416 417 hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); 418 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); 419 tmp5 = __msa_dotp_u_h(vec0, filt_vt); 420 421 hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); 422 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); 423 tmp6 = __msa_dotp_u_h(vec0, filt_vt); 424 425 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); 426 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); 427 tmp7 = __msa_dotp_u_h(vec0, filt_vt); 428 429 hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); 430 vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); 431 tmp8 = __msa_dotp_u_h(vec0, filt_vt); 432 433 SRARI_H4_UH(tmp5, tmp6, tmp7, tmp8, FILTER_BITS); 434 PCKEV_B2_SB(tmp6, tmp5, tmp8, tmp7, out0, out1); 435 ST8x4_UB(out0, out1, dst, dst_stride); 436 dst += (4 * dst_stride); 437 } 438} 439 440static void common_hv_2ht_2vt_8w_msa(const uint8_t *src, int32_t src_stride, 441 uint8_t *dst, int32_t dst_stride, 442 int8_t *filter_horiz, int8_t *filter_vert, 443 int32_t height) { 444 if (4 == height) { 445 common_hv_2ht_2vt_8x4_msa(src, src_stride, dst, dst_stride, filter_horiz, 446 filter_vert); 447 } else { 448 common_hv_2ht_2vt_8x8mult_msa(src, src_stride, dst, dst_stride, 449 filter_horiz, filter_vert, height); 450 } 451} 452 453static void common_hv_2ht_2vt_16w_msa(const uint8_t *src, int32_t src_stride, 454 uint8_t *dst, int32_t dst_stride, 455 int8_t *filter_horiz, int8_t *filter_vert, 456 int32_t height) { 457 uint32_t loop_cnt; 458 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask; 459 v16u8 filt_hz, filt_vt, vec0, vec1; 460 v8u16 tmp1, tmp2, hz_out0, hz_out1, hz_out2, hz_out3; 461 v8i16 filt; 462 463 mask = LD_SB(&mc_filt_mask_arr[0]); 464 465 /* rearranging filter */ 466 filt = LD_SH(filter_horiz); 467 filt_hz = (v16u8)__msa_splati_h(filt, 0); 468 469 filt = LD_SH(filter_vert); 470 filt_vt = (v16u8)__msa_splati_h(filt, 0); 471 472 LD_SB2(src, 8, src0, src1); 473 src += src_stride; 474 475 hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); 476 hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); 477 478 for (loop_cnt = (height >> 2); loop_cnt--;) { 479 LD_SB4(src, src_stride, src0, src2, src4, src6); 480 LD_SB4(src + 8, src_stride, src1, src3, src5, src7); 481 src += (4 * src_stride); 482 483 hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); 484 hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); 485 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); 486 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2); 487 SRARI_H2_UH(tmp1, tmp2, FILTER_BITS); 488 PCKEV_ST_SB(tmp1, tmp2, dst); 489 dst += dst_stride; 490 491 hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); 492 hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); 493 ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); 494 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2); 495 SRARI_H2_UH(tmp1, tmp2, FILTER_BITS); 496 PCKEV_ST_SB(tmp1, tmp2, dst); 497 dst += dst_stride; 498 499 hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); 500 hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS); 501 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); 502 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2); 503 SRARI_H2_UH(tmp1, tmp2, FILTER_BITS); 504 PCKEV_ST_SB(tmp1, tmp2, dst); 505 dst += dst_stride; 506 507 hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS); 508 hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS); 509 ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); 510 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2); 511 SRARI_H2_UH(tmp1, tmp2, FILTER_BITS); 512 PCKEV_ST_SB(tmp1, tmp2, dst); 513 dst += dst_stride; 514 } 515} 516 517static void common_hv_2ht_2vt_32w_msa(const uint8_t *src, int32_t src_stride, 518 uint8_t *dst, int32_t dst_stride, 519 int8_t *filter_horiz, int8_t *filter_vert, 520 int32_t height) { 521 int32_t multiple8_cnt; 522 for (multiple8_cnt = 2; multiple8_cnt--;) { 523 common_hv_2ht_2vt_16w_msa(src, src_stride, dst, dst_stride, filter_horiz, 524 filter_vert, height); 525 src += 16; 526 dst += 16; 527 } 528} 529 530static void common_hv_2ht_2vt_64w_msa(const uint8_t *src, int32_t src_stride, 531 uint8_t *dst, int32_t dst_stride, 532 int8_t *filter_horiz, int8_t *filter_vert, 533 int32_t height) { 534 int32_t multiple8_cnt; 535 for (multiple8_cnt = 4; multiple8_cnt--;) { 536 common_hv_2ht_2vt_16w_msa(src, src_stride, dst, dst_stride, filter_horiz, 537 filter_vert, height); 538 src += 16; 539 dst += 16; 540 } 541} 542 543void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, 544 ptrdiff_t dst_stride, const InterpKernel *filter, 545 int x0_q4, int32_t x_step_q4, int y0_q4, 546 int32_t y_step_q4, int32_t w, int32_t h) { 547 const int16_t *const filter_x = filter[x0_q4]; 548 const int16_t *const filter_y = filter[y0_q4]; 549 int8_t cnt, filt_hor[8], filt_ver[8]; 550 551 assert(x_step_q4 == 16); 552 assert(y_step_q4 == 16); 553 assert(((const int32_t *)filter_x)[1] != 0x800000); 554 assert(((const int32_t *)filter_y)[1] != 0x800000); 555 556 for (cnt = 0; cnt < 8; ++cnt) { 557 filt_hor[cnt] = filter_x[cnt]; 558 filt_ver[cnt] = filter_y[cnt]; 559 } 560 561 if (((const int32_t *)filter_x)[0] == 0 && 562 ((const int32_t *)filter_y)[0] == 0) { 563 switch (w) { 564 case 4: 565 common_hv_2ht_2vt_4w_msa(src, (int32_t)src_stride, dst, 566 (int32_t)dst_stride, &filt_hor[3], 567 &filt_ver[3], (int32_t)h); 568 break; 569 case 8: 570 common_hv_2ht_2vt_8w_msa(src, (int32_t)src_stride, dst, 571 (int32_t)dst_stride, &filt_hor[3], 572 &filt_ver[3], (int32_t)h); 573 break; 574 case 16: 575 common_hv_2ht_2vt_16w_msa(src, (int32_t)src_stride, dst, 576 (int32_t)dst_stride, &filt_hor[3], 577 &filt_ver[3], (int32_t)h); 578 break; 579 case 32: 580 common_hv_2ht_2vt_32w_msa(src, (int32_t)src_stride, dst, 581 (int32_t)dst_stride, &filt_hor[3], 582 &filt_ver[3], (int32_t)h); 583 break; 584 case 64: 585 common_hv_2ht_2vt_64w_msa(src, (int32_t)src_stride, dst, 586 (int32_t)dst_stride, &filt_hor[3], 587 &filt_ver[3], (int32_t)h); 588 break; 589 default: 590 vpx_convolve8_c(src, src_stride, dst, dst_stride, filter, x0_q4, 591 x_step_q4, y0_q4, y_step_q4, w, h); 592 break; 593 } 594 } else if (((const int32_t *)filter_x)[0] == 0 || 595 ((const int32_t *)filter_y)[0] == 0) { 596 vpx_convolve8_c(src, src_stride, dst, dst_stride, filter, x0_q4, x_step_q4, 597 y0_q4, y_step_q4, w, h); 598 } else { 599 switch (w) { 600 case 4: 601 common_hv_8ht_8vt_4w_msa(src, (int32_t)src_stride, dst, 602 (int32_t)dst_stride, filt_hor, filt_ver, 603 (int32_t)h); 604 break; 605 case 8: 606 common_hv_8ht_8vt_8w_msa(src, (int32_t)src_stride, dst, 607 (int32_t)dst_stride, filt_hor, filt_ver, 608 (int32_t)h); 609 break; 610 case 16: 611 common_hv_8ht_8vt_16w_msa(src, (int32_t)src_stride, dst, 612 (int32_t)dst_stride, filt_hor, filt_ver, 613 (int32_t)h); 614 break; 615 case 32: 616 common_hv_8ht_8vt_32w_msa(src, (int32_t)src_stride, dst, 617 (int32_t)dst_stride, filt_hor, filt_ver, 618 (int32_t)h); 619 break; 620 case 64: 621 common_hv_8ht_8vt_64w_msa(src, (int32_t)src_stride, dst, 622 (int32_t)dst_stride, filt_hor, filt_ver, 623 (int32_t)h); 624 break; 625 default: 626 vpx_convolve8_c(src, src_stride, dst, dst_stride, filter, x0_q4, 627 x_step_q4, y0_q4, y_step_q4, w, h); 628 break; 629 } 630 } 631} 632 633static void filter_horiz_w4_msa(const uint8_t *src_x, ptrdiff_t src_pitch, 634 uint8_t *dst, const int16_t *x_filter) { 635 uint64_t srcd0, srcd1, srcd2, srcd3; 636 uint32_t res; 637 v16u8 src0 = { 0 }, src1 = { 0 }, dst0; 638 v16i8 out0, out1; 639 v16i8 shf1 = { 0, 8, 16, 24, 4, 12, 20, 28, 1, 9, 17, 25, 5, 13, 21, 29 }; 640 v16i8 shf2 = shf1 + 2; 641 v16i8 filt_shf0 = { 0, 1, 0, 1, 0, 1, 0, 1, 8, 9, 8, 9, 8, 9, 8, 9 }; 642 v16i8 filt_shf1 = filt_shf0 + 2; 643 v16i8 filt_shf2 = filt_shf0 + 4; 644 v16i8 filt_shf3 = filt_shf0 + 6; 645 v8i16 filt, src0_h, src1_h, src2_h, src3_h, filt0, filt1, filt2, filt3; 646 647 LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3); 648 INSERT_D2_UB(srcd0, srcd1, src0); 649 INSERT_D2_UB(srcd2, srcd3, src1); 650 VSHF_B2_SB(src0, src1, src0, src1, shf1, shf2, out0, out1); 651 XORI_B2_128_SB(out0, out1); 652 UNPCK_SB_SH(out0, src0_h, src1_h); 653 UNPCK_SB_SH(out1, src2_h, src3_h); 654 655 filt = LD_SH(x_filter); 656 VSHF_B2_SH(filt, filt, filt, filt, filt_shf0, filt_shf1, filt0, filt1); 657 VSHF_B2_SH(filt, filt, filt, filt, filt_shf2, filt_shf3, filt2, filt3); 658 659 src0_h *= filt0; 660 src0_h += src1_h * filt1; 661 src0_h += src2_h * filt2; 662 src0_h += src3_h * filt3; 663 664 src1_h = (v8i16)__msa_sldi_b((v16i8)src0_h, (v16i8)src0_h, 8); 665 666 src0_h = __msa_adds_s_h(src0_h, src1_h); 667 src0_h = __msa_srari_h(src0_h, FILTER_BITS); 668 src0_h = __msa_sat_s_h(src0_h, 7); 669 dst0 = PCKEV_XORI128_UB(src0_h, src0_h); 670 res = __msa_copy_u_w((v4i32)dst0, 0); 671 SW(res, dst); 672} 673 674static void filter_horiz_w8_msa(const uint8_t *src_x, ptrdiff_t src_pitch, 675 uint8_t *dst, const int16_t *x_filter) { 676 uint64_t srcd0, srcd1, srcd2, srcd3; 677 v16u8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 }; 678 v16u8 tmp0, tmp1, tmp2, tmp3, dst0; 679 v16i8 out0, out1, out2, out3; 680 v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 }; 681 v16i8 shf2 = shf1 + 4; 682 v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; 683 v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; 684 685 LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3); 686 INSERT_D2_UB(srcd0, srcd1, src0); 687 INSERT_D2_UB(srcd2, srcd3, src1); 688 LD4(src_x + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); 689 INSERT_D2_UB(srcd0, srcd1, src2); 690 INSERT_D2_UB(srcd2, srcd3, src3); 691 692 filt = LD_SH(x_filter); 693 SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); 694 SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); 695 696 // transpose 697 VSHF_B2_UB(src0, src1, src0, src1, shf1, shf2, tmp0, tmp1); 698 VSHF_B2_UB(src2, src3, src2, src3, shf1, shf2, tmp2, tmp3); 699 ILVRL_W2_SB(tmp2, tmp0, out0, out1); 700 ILVRL_W2_SB(tmp3, tmp1, out2, out3); 701 702 XORI_B4_128_SB(out0, out1, out2, out3); 703 UNPCK_SB_SH(out0, src0_h, src1_h); 704 UNPCK_SB_SH(out1, src2_h, src3_h); 705 UNPCK_SB_SH(out2, src4_h, src5_h); 706 UNPCK_SB_SH(out3, src6_h, src7_h); 707 708 src0_h *= filt0; 709 src4_h *= filt4; 710 src0_h += src1_h * filt1; 711 src4_h += src5_h * filt5; 712 src0_h += src2_h * filt2; 713 src4_h += src6_h * filt6; 714 src0_h += src3_h * filt3; 715 src4_h += src7_h * filt7; 716 717 src0_h = __msa_adds_s_h(src0_h, src4_h); 718 src0_h = __msa_srari_h(src0_h, FILTER_BITS); 719 src0_h = __msa_sat_s_h(src0_h, 7); 720 dst0 = PCKEV_XORI128_UB(src0_h, src0_h); 721 ST8x1_UB(dst0, dst); 722} 723 724static void filter_horiz_w16_msa(const uint8_t *src_x, ptrdiff_t src_pitch, 725 uint8_t *dst, const int16_t *x_filter) { 726 uint64_t srcd0, srcd1, srcd2, srcd3; 727 v16u8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 }; 728 v16u8 src4 = { 0 }, src5 = { 0 }, src6 = { 0 }, src7 = { 0 }; 729 v16u8 tmp0, tmp1, tmp2, tmp3, dst0; 730 v16i8 out0, out1, out2, out3, out4, out5, out6, out7; 731 v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 }; 732 v16i8 shf2 = shf1 + 4; 733 v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; 734 v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; 735 v8i16 dst0_h, dst1_h, dst2_h, dst3_h; 736 737 LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3); 738 INSERT_D2_UB(srcd0, srcd1, src0); 739 INSERT_D2_UB(srcd2, srcd3, src1); 740 LD4(src_x + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); 741 INSERT_D2_UB(srcd0, srcd1, src2); 742 INSERT_D2_UB(srcd2, srcd3, src3); 743 LD4(src_x + 8 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); 744 INSERT_D2_UB(srcd0, srcd1, src4); 745 INSERT_D2_UB(srcd2, srcd3, src5); 746 LD4(src_x + 12 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); 747 INSERT_D2_UB(srcd0, srcd1, src6); 748 INSERT_D2_UB(srcd2, srcd3, src7); 749 750 filt = LD_SH(x_filter); 751 SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); 752 SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); 753 754 // transpose 755 VSHF_B2_UB(src0, src1, src0, src1, shf1, shf2, tmp0, tmp1); 756 VSHF_B2_UB(src2, src3, src2, src3, shf1, shf2, tmp2, tmp3); 757 ILVRL_W2_SB(tmp2, tmp0, out0, out1); 758 ILVRL_W2_SB(tmp3, tmp1, out2, out3); 759 XORI_B4_128_SB(out0, out1, out2, out3); 760 761 UNPCK_SB_SH(out0, src0_h, src1_h); 762 UNPCK_SB_SH(out1, src2_h, src3_h); 763 UNPCK_SB_SH(out2, src4_h, src5_h); 764 UNPCK_SB_SH(out3, src6_h, src7_h); 765 766 VSHF_B2_UB(src4, src5, src4, src5, shf1, shf2, tmp0, tmp1); 767 VSHF_B2_UB(src6, src7, src6, src7, shf1, shf2, tmp2, tmp3); 768 ILVRL_W2_SB(tmp2, tmp0, out4, out5); 769 ILVRL_W2_SB(tmp3, tmp1, out6, out7); 770 XORI_B4_128_SB(out4, out5, out6, out7); 771 772 dst0_h = src0_h * filt0; 773 dst1_h = src4_h * filt4; 774 dst0_h += src1_h * filt1; 775 dst1_h += src5_h * filt5; 776 dst0_h += src2_h * filt2; 777 dst1_h += src6_h * filt6; 778 dst0_h += src3_h * filt3; 779 dst1_h += src7_h * filt7; 780 781 UNPCK_SB_SH(out4, src0_h, src1_h); 782 UNPCK_SB_SH(out5, src2_h, src3_h); 783 UNPCK_SB_SH(out6, src4_h, src5_h); 784 UNPCK_SB_SH(out7, src6_h, src7_h); 785 786 dst2_h = src0_h * filt0; 787 dst3_h = src4_h * filt4; 788 dst2_h += src1_h * filt1; 789 dst3_h += src5_h * filt5; 790 dst2_h += src2_h * filt2; 791 dst3_h += src6_h * filt6; 792 dst2_h += src3_h * filt3; 793 dst3_h += src7_h * filt7; 794 795 ADDS_SH2_SH(dst0_h, dst1_h, dst2_h, dst3_h, dst0_h, dst2_h); 796 SRARI_H2_SH(dst0_h, dst2_h, FILTER_BITS); 797 SAT_SH2_SH(dst0_h, dst2_h, 7); 798 dst0 = PCKEV_XORI128_UB(dst0_h, dst2_h); 799 ST_UB(dst0, dst); 800} 801 802static void transpose4x4_to_dst(const uint8_t *src, uint8_t *dst, 803 ptrdiff_t dst_stride) { 804 v16u8 in0; 805 v16i8 out0 = { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 }; 806 807 in0 = LD_UB(src); 808 out0 = __msa_vshf_b(out0, (v16i8)in0, (v16i8)in0); 809 ST4x4_UB(out0, out0, 0, 1, 2, 3, dst, dst_stride); 810} 811 812static void transpose8x8_to_dst(const uint8_t *src, uint8_t *dst, 813 ptrdiff_t dst_stride) { 814 v16u8 in0, in1, in2, in3, out0, out1, out2, out3, tmp0, tmp1, tmp2, tmp3; 815 v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 }; 816 v16i8 shf2 = shf1 + 4; 817 818 LD_UB4(src, 16, in0, in1, in2, in3); 819 VSHF_B2_UB(in0, in1, in0, in1, shf1, shf2, tmp0, tmp1); 820 VSHF_B2_UB(in2, in3, in2, in3, shf1, shf2, tmp2, tmp3); 821 ILVRL_W2_UB(tmp2, tmp0, out0, out1); 822 ILVRL_W2_UB(tmp3, tmp1, out2, out3); 823 ST8x4_UB(out0, out1, dst, dst_stride); 824 ST8x4_UB(out2, out3, dst + 4 * dst_stride, dst_stride); 825} 826 827static void transpose16x16_to_dst(const uint8_t *src, uint8_t *dst, 828 ptrdiff_t dst_stride) { 829 v16u8 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11, in12; 830 v16u8 in13, in14, in15, out0, out1, out2, out3, out4, out5, out6, out7, out8; 831 v16u8 out9, out10, out11, out12, out13, out14, out15; 832 833 LD_UB8(src, 16, in0, in1, in2, in3, in4, in5, in6, in7); 834 LD_UB8(src + 16 * 8, 16, in8, in9, in10, in11, in12, in13, in14, in15); 835 836 TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, 837 in11, in12, in13, in14, in15, out0, out1, out2, out3, 838 out4, out5, out6, out7); 839 ST_UB8(out0, out1, out2, out3, out4, out5, out6, out7, dst, dst_stride); 840 dst += 8 * dst_stride; 841 842 SLDI_B4_0_UB(in0, in1, in2, in3, in0, in1, in2, in3, 8); 843 SLDI_B4_0_UB(in4, in5, in6, in7, in4, in5, in6, in7, 8); 844 SLDI_B4_0_UB(in8, in9, in10, in11, in8, in9, in10, in11, 8); 845 SLDI_B4_0_UB(in12, in13, in14, in15, in12, in13, in14, in15, 8); 846 847 TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, 848 in11, in12, in13, in14, in15, out8, out9, out10, out11, 849 out12, out13, out14, out15); 850 ST_UB8(out8, out9, out10, out11, out12, out13, out14, out15, dst, dst_stride); 851} 852 853static void scaledconvolve_horiz_w4(const uint8_t *src, ptrdiff_t src_stride, 854 uint8_t *dst, ptrdiff_t dst_stride, 855 const InterpKernel *x_filters, int x0_q4, 856 int x_step_q4, int h) { 857 DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]); 858 int y, z, i; 859 src -= SUBPEL_TAPS / 2 - 1; 860 861 for (y = 0; y < h; y += 4) { 862 int x_q4 = x0_q4; 863 for (z = 0; z < 4; ++z) { 864 const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; 865 const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; 866 867 if (x_q4 & SUBPEL_MASK) { 868 filter_horiz_w4_msa(src_x, src_stride, temp + (z * 4), x_filter); 869 } else { 870 for (i = 0; i < 4; ++i) { 871 temp[z * 4 + i] = src_x[i * src_stride + 3]; 872 } 873 } 874 875 x_q4 += x_step_q4; 876 } 877 878 transpose4x4_to_dst(temp, dst, dst_stride); 879 880 src += src_stride * 4; 881 dst += dst_stride * 4; 882 } 883} 884 885static void scaledconvolve_horiz_w8(const uint8_t *src, ptrdiff_t src_stride, 886 uint8_t *dst, ptrdiff_t dst_stride, 887 const InterpKernel *x_filters, int x0_q4, 888 int x_step_q4, int h) { 889 DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]); 890 int y, z, i; 891 src -= SUBPEL_TAPS / 2 - 1; 892 893 // This function processes 8x8 areas. The intermediate height is not always 894 // a multiple of 8, so force it to be a multiple of 8 here. 895 y = h + (8 - (h & 0x7)); 896 897 do { 898 int x_q4 = x0_q4; 899 for (z = 0; z < 8; ++z) { 900 const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; 901 const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; 902 903 if (x_q4 & SUBPEL_MASK) { 904 filter_horiz_w8_msa(src_x, src_stride, temp + (z * 8), x_filter); 905 } else { 906 for (i = 0; i < 8; ++i) { 907 temp[z * 8 + i] = src_x[3 + i * src_stride]; 908 } 909 } 910 911 x_q4 += x_step_q4; 912 } 913 914 transpose8x8_to_dst(temp, dst, dst_stride); 915 916 src += src_stride * 8; 917 dst += dst_stride * 8; 918 } while (y -= 8); 919} 920 921static void scaledconvolve_horiz_mul16(const uint8_t *src, ptrdiff_t src_stride, 922 uint8_t *dst, ptrdiff_t dst_stride, 923 const InterpKernel *x_filters, int x0_q4, 924 int x_step_q4, int w, int h) { 925 DECLARE_ALIGNED(16, uint8_t, temp[16 * 16]); 926 int x, y, z, i; 927 928 src -= SUBPEL_TAPS / 2 - 1; 929 930 // This function processes 16x16 areas. The intermediate height is not always 931 // a multiple of 16, so force it to be a multiple of 8 here. 932 y = h + (16 - (h & 0xF)); 933 934 do { 935 int x_q4 = x0_q4; 936 for (x = 0; x < w; x += 16) { 937 for (z = 0; z < 16; ++z) { 938 const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; 939 const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; 940 941 if (x_q4 & SUBPEL_MASK) { 942 filter_horiz_w16_msa(src_x, src_stride, temp + (z * 16), x_filter); 943 } else { 944 for (i = 0; i < 16; ++i) { 945 temp[z * 16 + i] = src_x[3 + i * src_stride]; 946 } 947 } 948 949 x_q4 += x_step_q4; 950 } 951 952 transpose16x16_to_dst(temp, dst + x, dst_stride); 953 } 954 955 src += src_stride * 16; 956 dst += dst_stride * 16; 957 } while (y -= 16); 958} 959 960static void filter_vert_w4_msa(const uint8_t *src_y, ptrdiff_t src_pitch, 961 uint8_t *dst, const int16_t *y_filter) { 962 uint32_t srcw0, srcw1, srcw2, srcw3, srcw4, srcw5, srcw6, srcw7; 963 uint32_t res; 964 v16u8 src0 = { 0 }, src1 = { 0 }, dst0; 965 v16i8 out0, out1; 966 v16i8 shf1 = { 0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23 }; 967 v16i8 shf2 = shf1 + 8; 968 v16i8 filt_shf0 = { 0, 1, 0, 1, 0, 1, 0, 1, 8, 9, 8, 9, 8, 9, 8, 9 }; 969 v16i8 filt_shf1 = filt_shf0 + 2; 970 v16i8 filt_shf2 = filt_shf0 + 4; 971 v16i8 filt_shf3 = filt_shf0 + 6; 972 v8i16 filt, src0_h, src1_h, src2_h, src3_h; 973 v8i16 filt0, filt1, filt2, filt3; 974 975 LW4(src_y, src_pitch, srcw0, srcw1, srcw2, srcw3); 976 LW4(src_y + 4 * src_pitch, src_pitch, srcw4, srcw5, srcw6, srcw7); 977 INSERT_W4_UB(srcw0, srcw1, srcw2, srcw3, src0); 978 INSERT_W4_UB(srcw4, srcw5, srcw6, srcw7, src1); 979 VSHF_B2_SB(src0, src1, src0, src1, shf1, shf2, out0, out1); 980 XORI_B2_128_SB(out0, out1); 981 UNPCK_SB_SH(out0, src0_h, src1_h); 982 UNPCK_SB_SH(out1, src2_h, src3_h); 983 984 filt = LD_SH(y_filter); 985 VSHF_B2_SH(filt, filt, filt, filt, filt_shf0, filt_shf1, filt0, filt1); 986 VSHF_B2_SH(filt, filt, filt, filt, filt_shf2, filt_shf3, filt2, filt3); 987 988 src0_h *= filt0; 989 src0_h += src1_h * filt1; 990 src0_h += src2_h * filt2; 991 src0_h += src3_h * filt3; 992 993 src1_h = (v8i16)__msa_sldi_b((v16i8)src0_h, (v16i8)src0_h, 8); 994 995 src0_h = __msa_adds_s_h(src0_h, src1_h); 996 src0_h = __msa_srari_h(src0_h, FILTER_BITS); 997 src0_h = __msa_sat_s_h(src0_h, 7); 998 dst0 = PCKEV_XORI128_UB(src0_h, src0_h); 999 res = __msa_copy_u_w((v4i32)dst0, 0); 1000 SW(res, dst); 1001} 1002 1003static void filter_vert_w8_msa(const uint8_t *src_y, ptrdiff_t src_pitch, 1004 uint8_t *dst, const int16_t *y_filter) { 1005 uint64_t srcd0, srcd1, srcd2, srcd3; 1006 v16u8 dst0; 1007 v16i8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 }; 1008 v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; 1009 v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; 1010 1011 LD4(src_y, src_pitch, srcd0, srcd1, srcd2, srcd3); 1012 INSERT_D2_SB(srcd0, srcd1, src0); 1013 INSERT_D2_SB(srcd2, srcd3, src1); 1014 LD4(src_y + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); 1015 INSERT_D2_SB(srcd0, srcd1, src2); 1016 INSERT_D2_SB(srcd2, srcd3, src3); 1017 1018 filt = LD_SH(y_filter); 1019 SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); 1020 SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); 1021 1022 XORI_B4_128_SB(src0, src1, src2, src3); 1023 UNPCK_SB_SH(src0, src0_h, src1_h); 1024 UNPCK_SB_SH(src1, src2_h, src3_h); 1025 UNPCK_SB_SH(src2, src4_h, src5_h); 1026 UNPCK_SB_SH(src3, src6_h, src7_h); 1027 1028 src0_h *= filt0; 1029 src4_h *= filt4; 1030 src0_h += src1_h * filt1; 1031 src4_h += src5_h * filt5; 1032 src0_h += src2_h * filt2; 1033 src4_h += src6_h * filt6; 1034 src0_h += src3_h * filt3; 1035 src4_h += src7_h * filt7; 1036 1037 src0_h = __msa_adds_s_h(src0_h, src4_h); 1038 src0_h = __msa_srari_h(src0_h, FILTER_BITS); 1039 src0_h = __msa_sat_s_h(src0_h, 7); 1040 dst0 = PCKEV_XORI128_UB(src0_h, src0_h); 1041 ST8x1_UB(dst0, dst); 1042} 1043 1044static void filter_vert_mul_w16_msa(const uint8_t *src_y, ptrdiff_t src_pitch, 1045 uint8_t *dst, const int16_t *y_filter, 1046 int w) { 1047 int x; 1048 v16u8 dst0; 1049 v16i8 src0, src1, src2, src3, src4, src5, src6, src7; 1050 v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; 1051 v8i16 src8_h, src9_h, src10_h, src11_h, src12_h, src13_h, src14_h, src15_h; 1052 v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; 1053 1054 filt = LD_SH(y_filter); 1055 SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); 1056 SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); 1057 1058 for (x = 0; x < w; x += 16) { 1059 LD_SB8(src_y, src_pitch, src0, src1, src2, src3, src4, src5, src6, src7); 1060 src_y += 16; 1061 1062 XORI_B4_128_SB(src0, src1, src2, src3); 1063 XORI_B4_128_SB(src4, src5, src6, src7); 1064 UNPCK_SB_SH(src0, src0_h, src1_h); 1065 UNPCK_SB_SH(src1, src2_h, src3_h); 1066 UNPCK_SB_SH(src2, src4_h, src5_h); 1067 UNPCK_SB_SH(src3, src6_h, src7_h); 1068 UNPCK_SB_SH(src4, src8_h, src9_h); 1069 UNPCK_SB_SH(src5, src10_h, src11_h); 1070 UNPCK_SB_SH(src6, src12_h, src13_h); 1071 UNPCK_SB_SH(src7, src14_h, src15_h); 1072 1073 src0_h *= filt0; 1074 src1_h *= filt0; 1075 src8_h *= filt4; 1076 src9_h *= filt4; 1077 src0_h += src2_h * filt1; 1078 src1_h += src3_h * filt1; 1079 src8_h += src10_h * filt5; 1080 src9_h += src11_h * filt5; 1081 src0_h += src4_h * filt2; 1082 src1_h += src5_h * filt2; 1083 src8_h += src12_h * filt6; 1084 src9_h += src13_h * filt6; 1085 src0_h += src6_h * filt3; 1086 src1_h += src7_h * filt3; 1087 src8_h += src14_h * filt7; 1088 src9_h += src15_h * filt7; 1089 1090 ADDS_SH2_SH(src0_h, src8_h, src1_h, src9_h, src0_h, src1_h); 1091 SRARI_H2_SH(src0_h, src1_h, FILTER_BITS); 1092 SAT_SH2_SH(src0_h, src1_h, 7); 1093 dst0 = PCKEV_XORI128_UB(src0_h, src1_h); 1094 ST_UB(dst0, dst); 1095 dst += 16; 1096 } 1097} 1098 1099static void scaledconvolve_vert_w4(const uint8_t *src, ptrdiff_t src_stride, 1100 uint8_t *dst, ptrdiff_t dst_stride, 1101 const InterpKernel *y_filters, int y0_q4, 1102 int y_step_q4, int h) { 1103 int y; 1104 int y_q4 = y0_q4; 1105 1106 src -= src_stride * (SUBPEL_TAPS / 2 - 1); 1107 1108 for (y = 0; y < h; ++y) { 1109 const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; 1110 const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; 1111 1112 if (y_q4 & SUBPEL_MASK) { 1113 filter_vert_w4_msa(src_y, src_stride, &dst[y * dst_stride], y_filter); 1114 } else { 1115 uint32_t srcd = LW(src_y + 3 * src_stride); 1116 SW(srcd, dst + y * dst_stride); 1117 } 1118 1119 y_q4 += y_step_q4; 1120 } 1121} 1122 1123static void scaledconvolve_vert_w8(const uint8_t *src, ptrdiff_t src_stride, 1124 uint8_t *dst, ptrdiff_t dst_stride, 1125 const InterpKernel *y_filters, int y0_q4, 1126 int y_step_q4, int h) { 1127 int y; 1128 int y_q4 = y0_q4; 1129 1130 src -= src_stride * (SUBPEL_TAPS / 2 - 1); 1131 1132 for (y = 0; y < h; ++y) { 1133 const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; 1134 const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; 1135 1136 if (y_q4 & SUBPEL_MASK) { 1137 filter_vert_w8_msa(src_y, src_stride, &dst[y * dst_stride], y_filter); 1138 } else { 1139 uint64_t srcd = LD(src_y + 3 * src_stride); 1140 SD(srcd, dst + y * dst_stride); 1141 } 1142 1143 y_q4 += y_step_q4; 1144 } 1145} 1146 1147static void scaledconvolve_vert_mul16(const uint8_t *src, ptrdiff_t src_stride, 1148 uint8_t *dst, ptrdiff_t dst_stride, 1149 const InterpKernel *y_filters, int y0_q4, 1150 int y_step_q4, int w, int h) { 1151 int x, y; 1152 int y_q4 = y0_q4; 1153 src -= src_stride * (SUBPEL_TAPS / 2 - 1); 1154 1155 for (y = 0; y < h; ++y) { 1156 const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; 1157 const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; 1158 1159 if (y_q4 & SUBPEL_MASK) { 1160 filter_vert_mul_w16_msa(src_y, src_stride, &dst[y * dst_stride], y_filter, 1161 w); 1162 } else { 1163 for (x = 0; x < w; ++x) { 1164 dst[x + y * dst_stride] = src_y[x + 3 * src_stride]; 1165 } 1166 } 1167 1168 y_q4 += y_step_q4; 1169 } 1170} 1171 1172void vpx_scaled_2d_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, 1173 ptrdiff_t dst_stride, const InterpKernel *filter, 1174 int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, 1175 int w, int h) { 1176 // Note: Fixed size intermediate buffer, temp, places limits on parameters. 1177 // 2d filtering proceeds in 2 steps: 1178 // (1) Interpolate horizontally into an intermediate buffer, temp. 1179 // (2) Interpolate temp vertically to derive the sub-pixel result. 1180 // Deriving the maximum number of rows in the temp buffer (135): 1181 // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative). 1182 // --Largest block size is 64x64 pixels. 1183 // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the 1184 // original frame (in 1/16th pixel units). 1185 // --Must round-up because block may be located at sub-pixel position. 1186 // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails. 1187 // --((64 - 1) * 32 + 15) >> 4 + 8 = 135. 1188 // --Require an additional 8 rows for the horiz_w8 transpose tail. 1189 DECLARE_ALIGNED(16, uint8_t, temp[(135 + 8) * 64]); 1190 const int intermediate_height = 1191 (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS; 1192 1193 assert(w <= 64); 1194 assert(h <= 64); 1195 assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32)); 1196 assert(x_step_q4 <= 64); 1197 1198 if ((0 == x0_q4) && (16 == x_step_q4) && (0 == y0_q4) && (16 == y_step_q4)) { 1199 vpx_convolve_copy_msa(src, src_stride, dst, dst_stride, filter, x0_q4, 1200 x_step_q4, y0_q4, y_step_q4, w, h); 1201 } else { 1202 if (w >= 16) { 1203 scaledconvolve_horiz_mul16(src - src_stride * (SUBPEL_TAPS / 2 - 1), 1204 src_stride, temp, 64, filter, x0_q4, x_step_q4, 1205 w, intermediate_height); 1206 } else if (w == 8) { 1207 scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1), 1208 src_stride, temp, 64, filter, x0_q4, x_step_q4, 1209 intermediate_height); 1210 } else { 1211 scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1), 1212 src_stride, temp, 64, filter, x0_q4, x_step_q4, 1213 intermediate_height); 1214 } 1215 1216 if (w >= 16) { 1217 scaledconvolve_vert_mul16(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, 1218 dst_stride, filter, y0_q4, y_step_q4, w, h); 1219 } else if (w == 8) { 1220 scaledconvolve_vert_w8(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, 1221 dst_stride, filter, y0_q4, y_step_q4, h); 1222 } else { 1223 scaledconvolve_vert_w4(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, 1224 dst_stride, filter, y0_q4, y_step_q4, h); 1225 } 1226 } 1227} 1228