| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/arm64/ |
| D | ConvDwFp32Indirect5x5.S | 50 ld1 {v29.4s}, [x4], #16
67 ld1 {v0.4s}, [x6], #16
68 ld1 {v1.4s}, [x7], #16
69 ld1 {v2.4s}, [x8], #16
70 ld1 {v3.4s}, [x9], #16
71 ld1 {v4.4s}, [x10], #16
73 ld1 {v18.4s}, [x5], #16
74 ld1 {v19.4s}, [x5], #16
75 ld1 {v20.4s}, [x5], #16
76 ld1 {v21.4s}, [x5], #16
[all …]
|
| D | ConvDwFp32Indirect3x3.S | 49 ld1 {v0.4s}, [x12], #16
50 ld1 {v1.4s}, [x13], #16
51 ld1 {v2.4s}, [x14], #16
53 ld1 {v17.4s}, [x9], #16
54 ld1 {v18.4s}, [x9], #16
55 ld1 {v19.4s}, [x9], #16
57 ld1 {v29.4s}, [x10], #16
62 ld1 {v3.4s}, [x15], #16
63 ld1 {v20.4s}, [x9], #16
65 ld1 {v4.4s}, [x16], #16
[all …]
|
| D | ConvDw3x3Fp32Vertical.S | 45 ld1 {v23.4s}, [x3], #16
49 ld1 {v0.4s}, [x9], x5
51 ld1 {v4.4s}, [x10], x13
52 ld1 {v1.4s}, [x9], x5
54 ld1 {v5.4s}, [x10], x13
55 ld1 {v2.4s}, [x11], x5
56 ld1 {v6.4s}, [x12], x13
57 ld1 {v3.4s}, [x11], x5
58 ld1 {v7.4s}, [x12], x13
59 ld1 {v16.4s}, [x9], x5
[all …]
|
| D | ConvDw3x3Fp32Horizontal.S | 45 ld1 {v23.4s}, [x3], #16
49 ld1 {v0.4s}, [x9], x5
51 ld1 {v4.4s}, [x10], x13
53 ld1 {v1.4s}, [x9], x5
54 ld1 {v5.4s}, [x10], x13
56 ld1 {v2.4s}, [x11], x5
58 ld1 {v6.4s}, [x12], x13
59 ld1 {v3.4s}, [x11], x5
60 ld1 {v7.4s}, [x12], x13
61 ld1 {v16.4s}, [x15], x5
[all …]
|
| D | ConvDw3x3Fp32Stride2.S | 58 ld1 {v0.4s}, [x2], x16
59 ld1 {v1.4s}, [x2], x16
60 ld1 {v2.4s}, [x2], x16
61 ld1 {v3.4s}, [x2], x16
62 ld1 {v4.4s}, [x2], x16
63 ld1 {v5.4s}, [x2], x16
64 ld1 {v6.4s}, [x2], x16
65 ld1 {v7.4s}, [x2], x16
66 ld1 {v8.4s}, [x2], x16
71 ld1 {v9.4s}, [x11], x15
[all …]
|
| D | ConvDw3x3Fp32Stride1.S | 58 ld1 {v0.4s}, [x2], x16
59 ld1 {v1.4s}, [x2], x16
60 ld1 {v2.4s}, [x2], x16
61 ld1 {v3.4s}, [x2], x16
62 ld1 {v4.4s}, [x2], x16
63 ld1 {v5.4s}, [x2], x16
64 ld1 {v6.4s}, [x2], x16
65 ld1 {v7.4s}, [x2], x16
66 ld1 {v8.4s}, [x2], x16
71 ld1 {v9.4s}, [x11], x15
[all …]
|
| D | MatVecMulFp32.S | 63 ld1 {v0.4s, v1.4s}, [x15], #32
64 ld1 {v2.4s, v3.4s}, [x7], #32
65 ld1 {v4.4s, v5.4s}, [x10], #32
72 ld1 {v6.4s, v7.4s}, [x11], #32
75 ld1 {v8.4s, v9.4s}, [x12], #32
78 ld1 {v2.4s, v3.4s}, [x7], #32
81 ld1 {v0.4s, v1.4s}, [x15], #32
82 ld1 {v4.4s, v5.4s}, [x10], #32
90 ld1 {v6.4s, v7.4s}, [x11], #32
93 ld1 {v8.4s, v9.4s}, [x12], #32
[all …]
|
| D | ConvDw3x3Fp32Corner.S | 45 ld1 {v23.4s}, [x3], #16
49 ld1 {v0.4s}, [x9], x5
51 ld1 {v4.4s}, [x10], x13 // weight
53 ld1 {v1.4s}, [x9], x5
54 ld1 {v5.4s}, [x10], x13
55 ld1 {v2.4s}, [x11], x5
56 ld1 {v6.4s}, [x12], x13
57 ld1 {v3.4s}, [x11], x5
58 ld1 {v7.4s}, [x12], x13
69 ld1 {v0.4s}, [x9], x5
[all …]
|
| D | ConvDw3x3Int8Vertical.S | 56 ld1 {v27.4s}, [x9], #16
57 ld1 {v28.4s}, [x10], #16
58 ld1 {v29.4s}, [x11], #16
66 ld1 {v23.4s}, [x3], #16
67 ld1 {v24.4s}, [x3], #16
71 ld1 {v0.8b}, [x12], x5
74 ld1 {v4.8h}, [x13], x21 // weight
76 ld1 {v1.8b}, [x12], x5
78 ld1 {v5.8h}, [x13], x21
79 ld1 {v2.8b}, [x19], x5
[all …]
|
| D | ConvDw3x3Int8Horizontal.S | 57 ld1 {v27.4s}, [x9], #16
58 ld1 {v28.4s}, [x10], #16
59 ld1 {v29.4s}, [x11], #16
71 ld1 {v23.4s}, [x3], #16
72 ld1 {v24.4s}, [x3], #16
76 ld1 {v0.8b}, [x12], x5
79 ld1 {v4.8h}, [x13], x23 // weight
81 ld1 {v1.8b}, [x12], x5
83 ld1 {v5.8h}, [x13], x23
85 ld1 {v2.8b}, [x19], x5
[all …]
|
| D | ConvDw3x3Int8Corner.S | 56 ld1 {v27.4s}, [x9], #16
57 ld1 {v28.4s}, [x10], #16
58 ld1 {v29.4s}, [x11], #16
66 ld1 {v23.4s}, [x3], #16
67 ld1 {v24.4s}, [x3], #16
71 ld1 {v0.8b}, [x12], x5
74 ld1 {v4.8h}, [x13], x21 // weight
76 ld1 {v1.8b}, [x12], x5
78 ld1 {v5.8h}, [x13], x21
79 ld1 {v2.8b}, [x19], x5
[all …]
|
| D | WinogradTransRight.S | 62 ld1 {v0.s}[0], [x13], x10
63 ld1 {v0.s}[1], [x13], x10
64 ld1 {v0.s}[2], [x13], x10
65 ld1 {v0.s}[3], [x13], x10
74 ld1 {v16.4s}, [x2]
75 ld1 {v20.4s}, [x17], #16
77 ld1 {v21.4s}, [x14], #16
79 ld1 {v20.4s}, [x16], #16
81 ld1 {v21.4s}, [x19], #16
100 ld1 {v0.s}[0], [x13], x10
[all …]
|
| D | WinogradTransLeft.S | 64 ld1 {v0.s}[0], [x17], x10
65 ld1 {v0.s}[1], [x17], x10
66 ld1 {v0.s}[2], [x17], x10
67 ld1 {v0.s}[3], [x17], x10
75 ld1 {v16.4s}, [x2]
76 ld1 {v20.4s}, [x14], #16
78 ld1 {v21.4s}, [x20], #16
80 ld1 {v20.4s}, [x16], #16
82 ld1 {v21.4s}, [x19], #16
100 ld1 {v0.s}[0], [x17], x10
[all …]
|
| /third_party/openh264/codec/decoder/core/arm64/ |
| D | intra_pred_aarch64_neon.S | 51 ld1 {v0.b}[4], [x3], x1
52 ld1 {v0.b}[5], [x3], x1
53 ld1 {v0.b}[6], [x3], x1
54 ld1 {v0.b}[7], [x3]
79 ld1 {v0.8b}, [x2]
99 ld1 {v0.8b}, [x2]
120 ld1 {v0.8b}, [x2]
138 ld1 {v0.8b}, [x2]
158 ld1 {v0.s}[1], [x2]
160 ld1 {v0.b}[3], [x2], x1
[all …]
|
| /third_party/openh264/codec/encoder/core/arm64/ |
| D | intra_pred_aarch64_neon.S | 51 ld1 {v0.b}[4], [x4], x2
52 ld1 {v0.b}[5], [x4], x2
53 ld1 {v0.b}[6], [x4], x2
54 ld1 {v0.b}[7], [x4]
79 ld1 {v0.8b}, [x3]
99 ld1 {v0.8b}, [x3]
120 ld1 {v0.8b}, [x3]
138 ld1 {v0.8b}, [x3]
158 ld1 {v0.s}[1], [x3]
160 ld1 {v0.b}[3], [x3], x2
[all …]
|
| D | pixel_aarch64_neon.S | 50 ld1 {v0.8b}, [x0], x1
51 ld1 {v1.8b}, [x0], x1
52 ld1 {v2.8b}, [x0], x1
53 ld1 {v3.8b}, [x0], x1
54 ld1 {v4.8b}, [x0], x1
55 ld1 {v5.8b}, [x0], x1
56 ld1 {v6.8b}, [x0], x1
57 ld1 {v7.8b}, [x0], x1
61 ld1 {v0.16b}, [x0], x1
62 ld1 {v1.16b}, [x0], x1
[all …]
|
| /third_party/ffmpeg/libswresample/aarch64/ |
| D | audio_convert_neon.S | 27 ld1 {v0.4s}, [x1], #16
29 ld1 {v1.4s}, [x1], #16
36 ld1 {v2.4s}, [x1], #16
39 ld1 {v3.4s}, [x1], #16
44 ld1 {v0.4s}, [x1], #16
46 ld1 {v1.4s}, [x1], #16
52 2: ld1 {v2.4s}, [x1], #16
55 ld1 {v3.4s}, [x1], #16
72 ld1 {v0.4s}, [x4], #16
74 ld1 {v1.4s}, [x4], #16
[all …]
|
| /third_party/ffmpeg/libavresample/aarch64/ |
| D | audio_convert_neon.S | 27 ld1 {v0.4s}, [x1], #16
29 ld1 {v1.4s}, [x1], #16
36 ld1 {v2.4s}, [x1], #16
39 ld1 {v3.4s}, [x1], #16
44 ld1 {v0.4s}, [x1], #16
46 ld1 {v1.4s}, [x1], #16
52 2: ld1 {v2.4s}, [x1], #16
55 ld1 {v3.4s}, [x1], #16
72 ld1 {v0.4s}, [x4], #16
74 ld1 {v1.4s}, [x4], #16
[all …]
|
| /third_party/ffmpeg/libavutil/aarch64/ |
| D | float_dsp_neon.S | 28 ld1 {v0.4S, v1.4S}, [x1], #32
29 ld1 {v2.4S, v3.4S}, [x1], #32
30 ld1 {v4.4S, v5.4S}, [x2], #32
31 ld1 {v6.4S, v7.4S}, [x2], #32
45 ld1 {v16.4S, v17.4S}, [x0], #32
46 ld1 {v18.4S, v19.4S}, [x0], x3
47 ld1 {v4.4S, v5.4S}, [x1], #32
48 ld1 {v6.4S, v7.4S}, [x1], #32
64 ld1 {v0.4S, v1.4S}, [x1], #32
67 ld1 {v2.4S, v3.4S}, [x1], #32
[all …]
|
| /third_party/ffmpeg/libavcodec/aarch64/ |
| D | hpeldsp_neon.S | 29 1: ld1 {v0.16B}, [x1], x2
30 ld1 {v1.16B}, [x1], x2
31 ld1 {v2.16B}, [x1], x2
32 ld1 {v3.16B}, [x1], x2
34 ld1 {v4.16B}, [x12], x2
36 ld1 {v5.16B}, [x12], x2
38 ld1 {v6.16B}, [x12], x2
40 ld1 {v7.16B}, [x12], x2
53 1: ld1 {v0.16B, v1.16B}, [x1], x2
54 ld1 {v2.16B, v3.16B}, [x1], x2
[all …]
|
| D | vp8dsp_neon.S | 30 ld1 {v0.4h - v3.4h}, [x1]
86 ld1 {v0.8b - v3.8b}, [x1]
128 ld1 {v24.s}[0], [x0], x2
130 ld1 {v25.s}[0], [x0], x2
133 ld1 {v26.s}[0], [x0], x2
136 ld1 {v27.s}[0], [x0], x2
178 ld1 {v0.8b}, [x0], x2
180 ld1 {v1.8b}, [x0], x2
182 ld1 {v2.8b}, [x0], x2
184 ld1 {v3.8b}, [x0], x2
[all …]
|
| D | vp9lpf_neon.S | 633 ld1 {v20.8b}, [x9], x1 // p3
634 ld1 {v24.8b}, [x0], x1 // q0
635 ld1 {v21.8b}, [x9], x1 // p2
636 ld1 {v25.8b}, [x0], x1 // q1
637 ld1 {v22.8b}, [x9], x1 // p1
638 ld1 {v26.8b}, [x0], x1 // q2
639 ld1 {v23.8b}, [x9], x1 // p0
640 ld1 {v27.8b}, [x0], x1 // q3
657 ld1 {v20.16b}, [x9], x1 // p3
658 ld1 {v24.16b}, [x0], x1 // q0
[all …]
|
| D | vp9mc_16bpp_neon.S | 33 ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x2], #64
34 ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x0], #64
35 ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x2], x3
38 ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x0], x1
55 ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x2], x3
56 ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x0], x1
57 ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [x2], x3
60 ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [x0], x1
76 ld1 {v2.8h, v3.8h}, [x2], x3
77 ld1 {v0.8h, v1.8h}, [x0]
[all …]
|
| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/assembly/fp16/ |
| D | WinogradTransRightFp16.S | 54 ld1 {v0.h}[0], [x13], x10
55 ld1 {v0.h}[1], [x13], x10
56 ld1 {v0.h}[2], [x13], x10
57 ld1 {v0.h}[3], [x13], x10
66 ld1 {v16.4h}, [x2]
67 ld1 {v20.4h}, [x17], #8
69 ld1 {v21.4h}, [x14], #8
71 ld1 {v20.4h}, [x16], #8
73 ld1 {v21.4h}, [x19], #8
92 ld1 {v0.h}[0], [x13], x10
[all …]
|
| D | WinogradTransLeftFp16.S | 56 ld1 {v0.h}[0], [x17], x10
57 ld1 {v0.h}[1], [x17], x10
58 ld1 {v0.h}[2], [x17], x10
59 ld1 {v0.h}[3], [x17], x10
67 ld1 {v16.4h}, [x2]
68 ld1 {v20.4h}, [x14], #8
70 ld1 {v21.4h}, [x20], #8
72 ld1 {v20.4h}, [x16], #8
74 ld1 {v21.4h}, [x19], #8
92 ld1 {v0.h}[0], [x17], x10
[all …]
|