Home
last modified time | relevance | path

Searched refs:v29 (Results 1 – 25 of 74) sorted by relevance

123

/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/arm64/
DConvDwFp32Indirect5x5.S50 ld1 {v29.4s}, [x4], #16
85 fmla v29.4s, v0.4s, v18.4s
88 fmla v29.4s, v1.4s, v19.4s
91 fmla v29.4s, v2.4s, v20.4s
94 fmla v29.4s, v3.4s, v21.4s
97 fmla v29.4s, v4.4s, v22.4s
101 fmla v29.4s, v5.4s, v23.4s
104 fmla v29.4s, v6.4s, v24.4s
107 fmla v29.4s, v7.4s, v25.4s
110 fmla v29.4s, v16.4s, v26.4s
[all …]
DConvDwFp32Indirect3x3.S57 ld1 {v29.4s}, [x10], #16
61 fmla v29.4s, v0.4s, v17.4s
64 fmla v29.4s, v1.4s, v18.4s
67 fmla v29.4s, v2.4s, v19.4s
70 fmla v29.4s, v3.4s, v20.4s
73 fmla v29.4s, v4.4s, v21.4s
76 fmla v29.4s, v5.4s, v22.4s
79 fmla v29.4s, v6.4s, v23.4s
82 fmla v29.4s, v7.4s, v24.4s
85 fmla v29.4s, v16.4s, v25.4s
[all …]
DConvFp32Center.S100 ld1 {v29.4s}, [x16], x15
103 zip1 v20.4s, v28.4s, v29.4s
104 zip2 v21.4s, v28.4s, v29.4s
110 trn2 v29.2d, v20.2d, v22.2d
117 fmla v0.4s, v29.4s, v16.s[1]
118 fmla v1.4s, v29.4s, v17.s[1]
127 fmla v2.4s, v29.4s, v18.s[1]
128 fmla v3.4s, v29.4s, v19.s[1]
137 fmla v4.4s, v29.4s, v20.s[1]
138 fmla v5.4s, v29.4s, v21.s[1]
[all …]
DConvDw3x3Line.S77 fsub v29.4s, v27.4s, v26.4s
78 fadd v29.4s, v29.4s, v25.4s
83 fadd v29.4s, v29.4s, v31.4s
91 fmin v29.4s, v29.4s, v30.4s
95 fmax v29.4s, v29.4s, v27.4s
113 st1 {v29.2s}, [x15]
119 st1 {v29.2s}, [x15]
121 st1 {v29.s}[2], [x18]
125 st1 {v29.4s}, [x15]
DIndirectGemmInt16to32_8x4.S28 mov v29.16b, v28.16b
100 smlal v29.4s, v16.4h, v5.h[0]
102 smlal2 v29.4s, v16.8h, v5.h[1]
104 smlal v29.4s, v17.4h, v5.h[2]
106 smlal2 v29.4s, v17.8h, v5.h[3]
119 smlal v29.4s, v18.4h, v5.h[4]
121 smlal2 v29.4s, v18.8h, v5.h[5]
123 smlal v29.4s, v19.4h, v5.h[6]
125 smlal2 v29.4s, v19.8h, v5.h[7]
186 smlal v29.4s, v16.4h, v5.h[0]
[all …]
DConvDw3x3Int8Corner.S53 ld1r {v29.4s}, [x11]
58 ld1 {v29.4s}, [x11], #16
129 sqrshl v23.4s, v23.4s, v29.4s
130 sqrshl v24.4s, v24.4s, v29.4s
137 sqrshl v23.4s, v23.4s, v29.4s
138 ld1 {v29.4s}, [x11], #16
143 sqrshl v24.4s, v24.4s, v29.4s
144 ld1 {v29.4s}, [x11], #16
189 sqrshl v23.4s, v23.4s, v29.4s
190 sqrshl v24.4s, v24.4s, v29.4s
[all …]
DConvDw3x3Int8Vertical.S53 ld1r {v29.4s}, [x11]
58 ld1 {v29.4s}, [x11], #16
149 sqrshl v23.4s, v23.4s, v29.4s
150 sqrshl v24.4s, v24.4s, v29.4s
157 sqrshl v23.4s, v23.4s, v29.4s
158 ld1 {v29.4s}, [x11], #16
163 sqrshl v24.4s, v24.4s, v29.4s
164 ld1 {v29.4s}, [x11], #16
213 sqrshl v23.4s, v23.4s, v29.4s
214 sqrshl v24.4s, v24.4s, v29.4s
[all …]
DConvDw3x3Int8Horizontal.S54 ld1r {v29.4s}, [x11]
59 ld1 {v29.4s}, [x11], #16
158 sqrshl v23.4s, v23.4s, v29.4s
159 sqrshl v24.4s, v24.4s, v29.4s
166 sqrshl v23.4s, v23.4s, v29.4s
167 ld1 {v29.4s}, [x11], #16
172 sqrshl v24.4s, v24.4s, v29.4s
173 ld1 {v29.4s}, [x11], #16
222 sqrshl v23.4s, v23.4s, v29.4s
223 sqrshl v24.4s, v24.4s, v29.4s
[all …]
DConvDwInt8PostAlign4.S38 dup v29.4s, w3
86 add v0.4s, v0.4s, v29.4s
87 add v1.4s, v1.4s, v29.4s
88 add v2.4s, v2.4s, v29.4s
89 add v3.4s, v3.4s, v29.4s
146 add v0.4s, v0.4s, v29.4s
147 add v1.4s, v1.4s, v29.4s
178 add v0.4s, v0.4s, v29.4s
DPostFuncInt8C4Neon64.S67 dup v29.4s, w10
117 add v0.4s, v0.4s, v29.4s
118 add v1.4s, v1.4s, v29.4s
119 add v2.4s, v2.4s, v29.4s
120 add v3.4s, v3.4s, v29.4s
158 add v0.4s, v0.4s, v29.4s
197 add v0.4s, v0.4s, v29.4s
220 add v0.4s, v0.4s, v29.4s
243 add v0.4s, v0.4s, v29.4s
/third_party/openh264/codec/encoder/core/arm64/
Dintra_pred_sad_3_opt_aarch64_neon.S229 uabdl v29.8h, v0.8b, v3.8b
230 uabal v29.8h, v4.8b, v5.8b //top
244 uabal v29.8h, v0.8b, v3.8b
245 uabal v29.8h, v4.8b, v5.8b //top
261 uabal v29.8h, v0.8b, v3.8b
262 uabal v29.8h, v4.8b, v5.8b //top
271 saddlv s29, v29.8h
302 uabdl v29.8h, v0.8b, v3.8b
303 uabal2 v29.8h, v0.16b,v3.16b //top
315 uabal v29.8h, v0.8b, v3.8b
[all …]
Dpixel_aarch64_neon.S43 saddlv s29, v29.8h
46 st4 {v28.s, v29.s, v30.s, v31.s}[0], [x4]
94 uab\arg0\()l v29.8h, v0.8b, v18.8b
95 uabal v29.8h, v1.8b, v19.8b
96 uabal v29.8h, v2.8b, v20.8b
97 uabal v29.8h, v3.8b, v21.8b
98 uabal v29.8h, v4.8b, v22.8b
99 uabal v29.8h, v5.8b, v23.8b
100 uabal v29.8h, v6.8b, v24.8b
101 uabal v29.8h, v7.8b, v25.8b
[all …]
/third_party/openh264/codec/processing/src/arm64/
Dvaa_calc_aarch64_neon.S161 add v29.2s, v20.2s, v21.2s
163 add v29.2s, v29.2s, v30.2s
164 add v28.2s, v28.2s, v29.2s
180 uaddlp v29.4s, v30.8h
182 uadalp v29.4s, v30.8h // p_sqdiff
199 uadalp v29.4s, v30.8h
201 uadalp v29.4s, v30.8h // p_sqdiff
220 uadalp v29.4s, v30.8h
222 uadalp v29.4s, v30.8h // p_sqdiff
239 uadalp v29.4s, v30.8h
[all …]
/third_party/json/include/nlohmann/detail/
Dmacro_scope.hpp231v29) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE29(func, v2, v3, v4, v5, v6, v7, v8, v9, v1… argument
232v29, v30) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE30(func, v2, v3, v4, v5, v6, v7, v8, v… argument
233v29, v30, v31) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE31(func, v2, v3, v4, v5, v6, v7, … argument
234v29, v30, v31, v32) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE32(func, v2, v3, v4, v5, v6,… argument
235v29, v30, v31, v32, v33) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE33(func, v2, v3, v4, v5… argument
236v29, v30, v31, v32, v33, v34) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE34(func, v2, v3, v… argument
237v29, v30, v31, v32, v33, v34, v35) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE35(func, v2, … argument
238v29, v30, v31, v32, v33, v34, v35, v36) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE36(func,… argument
239v29, v30, v31, v32, v33, v34, v35, v36, v37) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE37(… argument
240v29, v30, v31, v32, v33, v34, v35, v36, v37, v38) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PAS… argument
[all …]
/third_party/ffmpeg/libavcodec/aarch64/
Dh264idct_neon.S220 sub v16.8H, v29.8H, v27.8H
223 add v19.8H, v29.8H, v27.8H
226 add va.8H, va.8H, v29.8H
230 sshr v29.8H, v29.8H, #1
234 add va.8H, va.8H, v29.8H
238 sshr v29.8H, va.8H, #2
242 add v17.8H, v17.8H, v29.8H
251 sub v29.8H, v28.8H, v17.8H
259 sub v29.8H, v28.8H, v17.8H
277 ld1 {v28.8H, v29.8H}, [x1]
[all …]
Dvp9itxfm_neon.S334 butterfly_8h v28, v29, v17, v21 // v28 = t4, v29 = t5a
338 dmbutterfly0 v31, v29, v31, v29, v2, v3, v4, v5, v6, v7 // v31 = t6, v29 = t5
342 butterfly_8h v18, v21, v27, v29 // q13 = out[2], q10 = out[5]
348 … dmbutterfly_l v28, v29, v30, v31, v21, v18, v1.h[3], v1.h[2] // v28,v29 = t3a, v30,v31 = t2a
355 dbutterfly_n v30, v31, v28, v29, v16, v18, v6, v7, v26, v27 // v30 = t3, v31 = t7
361 … dmbutterfly0 v19, v20, v6, v7, v24, v26, v27, v28, v29, v30 // v19 = -out[3], v20 = out[4]
364 … dmbutterfly_l v26, v27, v28, v29, v5, v3, v0.h[2], v0.h[3] // v26,v27 = t5a, v28,v29 = t4a
367 dbutterfly_n v17, v30, v28, v29, v2, v3, v6, v7, v24, v25 // v17 = -out[1], v30 = t6
528 butterfly_8h v25, v29, v29, v17 // v25 = t15a, v29 = t12a
531 dmbutterfly0 v28, v27, v29, v28, v21, v29, v16, v17, v30, v31 // v28 = t12, v27 = t11
[all …]
Dmpegaudiodsp_neon.S59 movi v29.2d, #0
61 trn1 v31.2d, v29.2d, v30.2d
62 trn2 v30.2d, v30.2d, v29.2d
63 trn1 v16.2d, v16.2d, v29.2d
105 ext v28.16b, v29.16b, v28.16b, #8
191 ext v28.16b, v28.16b, v29.16b, #8
193 ext v28.16b, v29.16b, v28.16b, #8
Dvp9itxfm_16bpp_neon.S582 ld1 {v28.4s,v29.4s,v30.4s,v31.4s}, [x2], #64
591 idct8 v17, v19, v21, v23, v25, v27, v29, v31, v2, v3, v4, v5, v6, v7
594 \txfm1\()8 v17, v19, v21, v23, v25, v27, v29, v31, v4, v5, v6, v7, v8, v9
598 …transpose_8x8s v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v4…
602 idct8 v17, v19, v21, v23, v25, v27, v29, v31, v2, v3, v4, v5, v6, v7
605 \txfm2\()8 v17, v19, v21, v23, v25, v27, v29, v31, v4, v5, v6, v7, v8, v9
644 srshr v29.4s, v29.4s, #5
661 uaddw2 v29.4s, v29.4s, v6.8h
675 sqxtun2 v6.8h, v29.4s
769 butterfly_4s v25, v29, v29, v17 // v25 = t15a, v29 = t12a
[all …]
Dsimple_idct_neon.S170 mov x5, v29.D[\i - 1]
184 smlal\i v17.4S, v29\l, z5
185 smlsl\i v18.4S, v29\l, z1
186 smlal\i v5.4S, v29\l, z7
187 smlal\i v6.4S, v29\l, z3
229 idct_row4_neon v28, v29, v30, v31, 2
267 idct_row4_neon v28, v29, v30, v31, 2
293 zip1 v29.2D, v4.2D, v18.2D
314 uaddw v26.8H, v29.8H, v22.8B
332 idct_row4_neon v28, v29, v30, v31, 2
[all …]
Dhevcdsp_idct_neon.S250 mov v29.16b, v28.16b
254 smlsl\p2 v29.4s, \in2, v0.h[0] //e1
259 add \out1, v29.4s, v31.4s
260 sub \out2, v29.4s, v31.4s
285 smull\p1 v29.4s, \in1\in1t, v0.h[5]
288 sum_sub v29.4s, \in3\in3t, v0.h[7], -, \p1
292 sum_sub v29.4s, \in5\in5t, v0.h[4], -, \p2
296 sum_sub v29.4s, \in7\in7t, v0.h[6], -, \p2
312 add v28.4s, v25.4s, v29.4s
313 sub v25.4s, v25.4s, v29.4s
[all …]
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/opt/
DMatmulDpInt8Opt.S95 dup v29.4s, wzr
120 sdot v29.4s, v2.16b, v0.4b[3]
143 add v29.4s, v29.4s, v1.4s
168 sub v29.4s, v29.4s, v15.4s
205 sub v29.4s, v29.4s, v5.4s
245 sqshl v29.4s, v29.4s, v1.4s
262 sqrdmulh v29.4s, v29.4s, v5.4s
321 and v1.16b, v9.16b, v29.16b
323 sqadd v29.4s, v29.4s, v1.4s
324 srshl v29.4s, v29.4s, v9.4s
[all …]
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/fp16/
DMatmulFp16.S70 dup v29.4s, wzr
94 fmla v29.8h, v8.8h, v1.h[5]
112 fmla v29.8h, v9.8h, v3.h[5]
130 fmla v29.8h, v10.8h, v5.h[5]
147 fmla v29.8h, v11.8h, v7.h[5]
164 fmla v29.8h, v12.8h, v1.h[5]
180 fmla v29.8h, v13.8h, v3.h[5]
196 fmla v29.8h, v14.8h, v5.h[5]
212 fmla v29.8h, v15.8h, v7.h[5]
238 fmla v29.8h, v8.8h, v1.h[5]
[all …]
/third_party/libjpeg-turbo/simd/arm64/
Djsimd_neon.S445 cmeq v29.8h, v7.8h, #0
451 and v12.16b, v29.16b, v30.16b
483 trn1 v29.8h, v4.8h, v5.8h
490 trn1 v2.4s, v28.4s, v29.4s
494 trn2 v4.4s, v28.4s, v29.4s
520 add v29.4s, v27.4s, v19.4s /* tmp11h tmp11 = tmp1 + tmp2; */
583 add v23.4s, v29.4s, v15.4s /* tmp11 + tmp2 */
585 sub v25.4s, v29.4s, v15.4s /* tmp11 - tmp2 */
589 sub v29.4s, v30.4s, v13.4s /* tmp12 - tmp1 */
608 …shrn2 v7.8h, v29.4s, #16 /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS+PA…
[all …]
/third_party/flutter/skia/third_party/externals/libjpeg-turbo/simd/arm64/
Djsimd_neon.S234 cmeq v29.8h, v7.8h, #0
240 and v12.16b, v29.16b, v30.16b
272 trn1 v29.8h, v4.8h, v5.8h
279 trn1 v2.4s, v28.4s, v29.4s
283 trn2 v4.4s, v28.4s, v29.4s
309 add v29.4s, v27.4s, v19.4s /* tmp11h tmp11 = tmp1 + tmp2; */
372 add v23.4s, v29.4s, v15.4s /* tmp11 + tmp2 */
374 sub v25.4s, v29.4s, v15.4s /* tmp11 - tmp2 */
378 sub v29.4s, v30.4s, v13.4s /* tmp12 - tmp1 */
397 …shrn2 v7.8h, v29.4s, #16 /* wsptr[DCTSIZE*5] = (int)DESCALE(tmp12 - tmp1, CONST_BITS+PA…
[all …]
/third_party/openssl/crypto/modes/asm/arm64/
Dghashv8-armx.S291 pmull v29.1q,v20.1d,v25.1d //H·Ii+3
301 eor v29.16b,v29.16b,v16.16b
310 eor v29.16b,v29.16b,v7.16b
337 eor v0.16b,v0.16b,v29.16b
345 pmull v29.1q,v20.1d,v25.1d //H·Ii+3
361 eor v29.16b,v29.16b,v16.16b
374 eor v29.16b,v29.16b,v7.16b
391 eor v0.16b,v0.16b,v29.16b
420 pmull v29.1q,v20.1d,v24.1d //H·Ii+2
438 eor v29.16b,v29.16b,v7.16b
[all …]

123