Home
last modified time | relevance | path

Searched refs:v17 (Results 1 – 25 of 123) sorted by relevance

12345

/third_party/openssl/crypto/modes/asm/arm64/
Dghashv8-armx.S10 ld1 {v17.2d},[x1] //load input H
13 ext v3.16b,v17.16b,v17.16b,#8
15 dup v17.4s,v17.s[1]
18 sshr v17.4s,v17.4s,#31 //broadcast carry bit
22 and v16.16b,v16.16b,v17.16b
34 ext v17.16b,v0.16b,v2.16b,#8 //Karatsuba post-processing
36 eor v1.16b,v1.16b,v17.16b
49 ext v17.16b,v22.16b,v22.16b,#8 //Karatsuba pre-processing
50 eor v17.16b,v17.16b,v22.16b
51 ext v21.16b,v16.16b,v17.16b,#8 //pack Karatsuba pre-processed
[all …]
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/arm64/
DIndirectGemmInt16to32_8x4.S52 ld1 {v17.8h}, [x2], #16
57 smlal v24.4s, v17.4h, v0.h[2]
58 smlal v25.4s, v17.4h, v1.h[2]
59 smlal2 v24.4s, v17.8h, v0.h[3]
60 smlal2 v25.4s, v17.8h, v1.h[3]
73 smlal v26.4s, v17.4h, v2.h[2]
74 smlal v27.4s, v17.4h, v3.h[2]
75 smlal2 v26.4s, v17.8h, v2.h[3]
76 smlal2 v27.4s, v17.8h, v3.h[3]
103 smlal v28.4s, v17.4h, v4.h[2]
[all …]
DConvDw3x3Int8Stride2.S119 ld1 {v17.8b}, [x17], x4
123 ssubl v17.8h, v17.8b, v28.8b
154 smlal v26.4s, v4.4h, v17.4h
155 smlal2 v27.4s, v4.8h, v17.8h
183 ld1r {v17.4s}, [x11]
190 sqrdmulh v24.4s, v24.4s, v17.4s
191 sqrdmulh v25.4s, v25.4s, v17.4s
192 sqrdmulh v26.4s, v26.4s, v17.4s
193 sqrdmulh v27.4s, v27.4s, v17.4s
198 sqrdmulh v24.4s, v24.4s, v17.4s
[all …]
DConvDwInt8Row.S51 ld1 {v16.4s, v17.4s}, [x0], #32
55 smlal2 v17.4s, v20.8h, v2.8h
62 st1 {v16.4s, v17.4s}, [x10], #32
71 ld1 {v16.4s, v17.4s}, [x0], #32
75 smlal2 v17.4s, v20.8h, v2.8h
83 st1 {v16.4s, v17.4s}, [x10], #32
97 ld1 {v16.4s, v17.4s}, [x0], #32
101 smlal2 v17.4s, v20.8h, v2.8h
102 st1 {v16.4s, v17.4s}, [x10], #32
DMatmulInt8Opt.S82 dup v17.4s, wzr
121 sadalp v17.4s, v9.8h
158 addp v16.4s, v16.4s, v17.4s
168 addp v17.4s, v20.4s, v22.4s
176 add v17.4s, v17.4s, v15.4s
195 sub v17.4s, v17.4s, v21.4s
213 sqshl v17.4s, v17.4s, v13.4s
218 sqrdmulh v17.4s, v17.4s, v12.4s
226 and v21.16b, v11.16b, v17.16b
228 sqadd v17.4s, v17.4s, v21.4s
[all …]
DMatmulInt8.S86 dup v17.4s, wzr
133 sadalp v17.4s, v9.8h
171 addp v16.4s, v16.4s, v17.4s
181 addp v17.4s, v20.4s, v22.4s
188 add v17.4s, v17.4s, v15.4s
223 sub v17.4s, v17.4s, v21.4s
229 sqshl v17.4s, v17.4s, v13.4s
235 sqrdmulh v17.4s, v17.4s, v12.4s
244 and v21.16b, v11.16b, v17.16b
246 sqadd v17.4s, v17.4s, v21.4s
[all …]
DWinogradTransRight.S78 fmul v17.4s, v21.4s, v0.s[1]
82 fmla v17.4s, v21.4s, v0.s[3]
84 fadd v17.4s, v16.4s, v17.4s
85 st1 {v17.4s}, [x2], #16
114 fmul v17.4s, v21.4s, v0.s[1]
118 fadd v17.4s, v16.4s, v17.4s
119 st1 {v17.4s}, [x2], #16
DWinogradTransLeft.S79 fmul v17.4s, v21.4s, v0.s[1]
83 fmla v17.4s, v21.4s, v0.s[3]
84 fadd v17.4s, v16.4s, v17.4s
85 st1 {v17.4s}, [x2], #16
112 fmul v17.4s, v21.4s, v0.s[1]
115 fadd v17.4s, v16.4s, v17.4s
116 st1 {v17.4s}, [x2], #16
/third_party/ffmpeg/libavcodec/aarch64/
Dh264cmc_neon.S71 umull v17.8H, v6.8B, v0.8B
73 umlal v17.8H, v7.8B, v1.8B
74 umlal v17.8H, v4.8B, v2.8B
75 umlal v17.8H, v5.8B, v3.8B
79 rshrn v17.8B, v17.8H, #6
82 add v17.8H, v17.8H, v22.8H
84 shrn v17.8B, v17.8H, #6
90 urhadd v17.8B, v17.8B, v21.8B
93 st1 {v17.8B}, [x0], x2
109 umull v17.8H, v6.8B, v0.8B
[all …]
Dsimple_idct_neon.S79 smull\i v17.4S, \y2\l, z1
88 smlal\i v17.4S, \y4\l, z3
110 smlal2 v17.4S, \y2\().8H, z5
120 smlal2 v17.4S, \y4\().8H, z7
129 \pass: add \y3\().4S, v19.4S, v17.4S
138 sub v19.4S, v19.4S, v17.4S
147 trn2 v17.8H, \y1\().8H, \y2\().8H
151 trn1 \y2\().4S, v17.4S, v19.4S
153 trn2 \y4\().4S, v17.4S, v19.4S
184 smlal\i v17.4S, v29\l, z5
[all …]
Dmdct_neon.S39 ld2 {v16.2s,v17.2s}, [x7], x12 // d16=x,n1 d17=x,n0
41 rev64 v17.2s, v17.2s
43 fmul v6.2s, v17.2s, v2.2s
49 fmul v5.2s, v17.2s, v3.2s
57 ld2 {v16.2s,v17.2s}, [x7], x12
59 rev64 v17.2s, v17.2s
61 fmul v6.2s, v17.2s, v2.2s
91 ld2 {v16.2s,v17.2s},[x1], x7 // d16=c1,c0 d18=s1,s0
94 fmul v7.2s, v0.2s, v17.2s
96 fmul v4.2s, v1.2s, v17.2s
[all …]
Dvp9itxfm_neon.S185 sub v17.4h, \c2\().4h, \c3\().4h
186 sub v16.4h, \c0\().4h, v17.4h
190 add \c3\().4h, v17.4h, \c2\().4h
198 sub v17.4h, \c0\().4h, \c2\().4h
201 smull v19.4s, v17.4h, v0.h[0]
217 smull v17.4s, \c0\().4h, v0.h[6]
218 smlsl v17.4s, \c2\().4h, v0.h[4]
220 smlsl v17.4s, \c3\().4h, v0.h[5]
225 add v21.4s, v17.4s, v19.4s
227 add v16.4s, v16.4s, v17.4s
[all …]
Dh264dsp_neon.S55 uabd v17.16B, v20.16B, v16.16B // abs(p2 - p0)
61 cmhi v17.16B, v22.16B, v17.16B // < beta
64 and v17.16B, v17.16B, v21.16B
68 sub v21.16B, v24.16B, v17.16B
92 bsl v17.16B, v23.16B, v18.16B
127 st1 {v17.16B}, [x0], x1
161 transpose_4x16B v17, v16, v0, v19, v21, v23, v25, v27
165 st1 {v17.S}[0], [x0], x1
169 st1 {v17.S}[1], [x0], x1
173 st1 {v17.S}[2], [x0], x1
[all …]
Dfft_neon.S45 ext v17.8b, v3.8b, v2.8b, #4
48 fsub v7.2s, v16.2s, v17.2s // r3-r2,i2-i3
63 ld1 {v16.2s,v17.2s,v18.2s,v19.2s}, [x0]
66 fadd v4.2s, v16.2s, v17.2s // r4+r5,i4+i5
68 fsub v17.2s, v16.2s, v17.2s // r4-r5,i4-i5
73 fmul v26.2s, v17.2s, v28.2s // -a2r*w,a2i*w
79 fmul v24.2s, v17.2s, v28.s[1] // a2r*w,a2i*w
99 fsub v17.2s, v1.2s, v5.2s
104 st1 {v16.2s,v17.2s,v18.2s,v19.2s}, [x0]
113 ld1 {v16.2s,v17.2s,v18.2s,v19.2s}, [x0], #32
[all …]
Dvp9itxfm_16bpp_neon.S234 sub v17.4s, \c2\().4s, \c3\().4s
235 sub v16.4s, \c0\().4s, v17.4s
239 add \c3\().4s, v17.4s, \c2\().4s
251 sub v17.4s, \c0\().4s, \c2\().4s
254 mul v24.4s, v17.4s, v0.s[0]
272 sub v17.4s, \c0\().4s, \c2\().4s
277 smull v24.2d, v17.2s, v0.s[0]
278 smull2 v25.2d, v17.4s, v0.s[0]
318 smull2 v17.2d, \c0\().4s, v1.s[0]
320 smlal2 v17.2d, \c2\().4s, v1.s[1]
[all …]
Dvp9mc_16bpp_neon.S38 ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x0], x1
43 urhadd v17.8h, v17.8h, v21.8h
47 st1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x5], x1
60 ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x0], x1
65 urhadd v17.8h, v17.8h, v21.8h
69 st1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x5], x1
178 ld1 {v16.8h, v17.8h, v18.8h}, [x7], #48
181 ld1 {v16.8h, v17.8h}, [x7]
193 smull v26.4s, v17.4h, v0.h[0]
195 smull2 v27.4s, v17.8h, v0.h[0]
[all …]
/third_party/json/include/nlohmann/detail/
Dmacro_scope.hpp219 …15, v16, v17) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE17(func, v2, v3, v4, v5, v6, v7, v… argument
220 …, v16, v17, v18) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE18(func, v2, v3, v4, v5, v6, v7… argument
221 …16, v17, v18, v19) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE19(func, v2, v3, v4, v5, v6, … argument
222 …, v17, v18, v19, v20) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE20(func, v2, v3, v4, v5, v… argument
223v17, v18, v19, v20, v21) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE21(func, v2, v3, v4, v5… argument
224v17, v18, v19, v20, v21, v22) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE22(func, v2, v3, v… argument
225v17, v18, v19, v20, v21, v22, v23) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE23(func, v2, … argument
226v17, v18, v19, v20, v21, v22, v23, v24) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE24(func,… argument
227v17, v18, v19, v20, v21, v22, v23, v24, v25) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE25(… argument
228v17, v18, v19, v20, v21, v22, v23, v24, v25, v26) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PAS… argument
[all …]
/third_party/ffmpeg/libavutil/aarch64/
Dfloat_dsp_neon.S33 fmul v17.4S, v1.4S, v5.4S
36 st1 {v16.4S, v17.4S}, [x0], #32
45 ld1 {v16.4S, v17.4S}, [x0], #32
50 fmla v17.4S, v5.4S, v0.S[0]
53 st1 {v16.4S, v17.4S}, [x0], #32
116 fmul v17.4S, v0.4S, v2.4S // s0 * wi
119 rev64 v17.4S, v17.4S
122 ext v17.16B, v17.16B, v17.16B, #8 // (s0 * wi)_rev
124 fmla v17.4S, v1.4S, v3.4S // (s0 * wi)_rev + s1 * wj
128 st1 {v17.4S}, [x5], x7
[all …]
/third_party/openssl/crypto/aes/asm/arm64/
Daesv8-armx.S300 ld1 {v16.4s,v17.4s},[x3] // load key schedule...
334 aese v0.16b,v17.16b
342 ld1 {v17.4s},[x12]
348 aese v0.16b,v17.16b
350 ld1 {v17.4s},[x3]
357 aese v0.16b,v17.16b
370 ld1 {v17.4s},[x7] // re-pre-load rndkey[1]
391 aese v0.16b,v17.16b
442 aesd v0.16b,v17.16b
444 aesd v1.16b,v17.16b
[all …]
/third_party/openh264/codec/decoder/core/arm64/
Dblock_add_aarch64_neon.S73 ROW_TRANSFORM_1_STEP v0, v1, v2, v3, v16, v17, v18, v19, v4, v5
74 TRANSFORM_4BYTES v0, v1, v2, v3, v16, v17, v18, v19
77 trn2 v17.4s, v0.4s, v1.4s //[0 1 2 3]+[4 5 6 7]-->[1 5 3 7]
82 trn1 v1.2d, v17.2d, v19.2d //[1 5 3 7]+[9 13 11 15]-->[1 5 9 13]
83 trn2 v3.2d, v17.2d, v19.2d //[1 5 3 7]+[9 13 11 15]-->[3 7 11 15]
84 COL_TRANSFORM_1_STEP v0, v1, v2, v3, v16, v17, v18, v19
86 TRANSFORM_4BYTES v0, v1, v2, v3, v16, v17, v18, v19
91 ld1 {v17.s}[0],[x0],x1
92 ld1 {v17.s}[1],[x0]
100 uxtl v3.8h,v17.8b
/third_party/openh264/codec/common/arm64/
Ddeblocking_aarch64_neon.S307 dup v17.16b, w3 //beta
325 MASK_MATRIX v1, v2, v3, v4, v16, v17, v18
333 DIFF_LUMA_LT4_P1_Q1 v0, v1, v2, v3, v17, v18, v6, v7, v19, v20
336 DIFF_LUMA_LT4_P1_Q1 v5, v4, v3, v2, v17, v18, v6, v7, v21, v22
366 dup v17.16b, w3 //beta
379 MASK_MATRIX v2, v3, v4, v5, v16, v17, v18
387 cmhi v21.16b, v17.16b, v21.16b //bDetaP2P0
391 cmhi v22.16b, v17.16b, v22.16b //bDetaQ2Q0
399 DIFF_LUMA_EQ4_P2P1P0_1 v0, v1, v2, v3, v4, v5, v23, v19, v17, v16
400 DIFF_LUMA_EQ4_P2P1P0_2 v25, v1, v2, v3, v4, v5, v24, v19, v17, v16
[all …]
Dcopy_mb_aarch64_neon.S130 LOAD16_ALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x2, x3
132 STORE16_ALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x0, x1
138 LOAD16_ALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x2, x3
140 STORE16_ALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x0, x1
152 LOAD16_UNALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x2, x3
154 STORE16_UNALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x0, x1
160 LOAD16_UNALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x2, x3
162 STORE16_UNALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x0, x1
174 LOAD16_UNALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x2, x3
176 STORE16_UNALIGNED_DATA_WITH_STRIDE v16, v17, v18, v19, x0, x1
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/fp16/
DWinogradTransRightFp16.S70 fmul v17.4h, v21.4h, v0.h[1]
74 fmla v17.4h, v21.4h, v0.h[3]
76 fadd v17.4h, v16.4h, v17.4h
77 st1 {v17.4h}, [x2], #8
106 fmul v17.4h, v21.4h, v0.h[1]
110 fadd v17.4h, v16.4h, v17.4h
111 st1 {v17.4h}, [x2], #8
DWinogradTransLeftFp16.S71 fmul v17.4h, v21.4h, v0.h[1]
75 fmla v17.4h, v21.4h, v0.h[3]
76 fadd v17.4h, v16.4h, v17.4h
77 st1 {v17.4h}, [x2], #8
104 fmul v17.4h, v21.4h, v0.h[1]
107 fadd v17.4h, v16.4h, v17.4h
108 st1 {v17.4h}, [x2], #8
DMatmulFp16Opt.S85 dup v17.4s, wzr
108 fmla v17.8h, v8.8h, v0.h[1]
125 fmla v17.8h, v9.8h, v2.h[1]
141 fmla v17.8h, v10.8h, v4.h[1]
157 fmla v17.8h, v11.8h, v6.h[1]
182 fmla v17.8h, v2.8h, v0.h[1]
205 fadd v17.8h, v17.8h, v0.8h
231 fmin v17.8h, v17.8h, v2.8h
250 fmax v17.8h, v17.8h, v2.8h
280 dup v17.4s, wzr
[all …]

12345