| /dports/math/blasfeo/blasfeo-0.1.2/kernel/armv8a/ |
| H A D | kernel_dgemm_8x4_lib.S | 101 fmla v0.2d, v16.2d, v24.2d[0]
102 fmla v1.2d, v17.2d, v24.2d[0]
104 fmla v8.2d, v20.2d, v24.2d[0]
105 fmla v9.2d, v21.2d, v24.2d[0]
107 fmla v2.2d, v16.2d, v24.2d[1]
108 fmla v3.2d, v17.2d, v24.2d[1]
113 fmla v4.2d, v16.2d, v25.2d[0]
114 fmla v5.2d, v17.2d, v25.2d[0]
119 fmla v6.2d, v16.2d, v25.2d[1]
120 fmla v7.2d, v17.2d, v25.2d[1]
[all …]
|
| H A D | kernel_dgemm_8x4_lib4.S | 152 fmla v0.2d, v16.2d, v24.2d[0]
153 fmla v1.2d, v17.2d, v24.2d[0]
155 fmla v8.2d, v20.2d, v24.2d[0]
156 fmla v9.2d, v21.2d, v24.2d[0]
158 fmla v2.2d, v16.2d, v24.2d[1]
159 fmla v3.2d, v17.2d, v24.2d[1]
164 fmla v4.2d, v16.2d, v25.2d[0]
165 fmla v5.2d, v17.2d, v25.2d[0]
170 fmla v6.2d, v16.2d, v25.2d[1]
171 fmla v7.2d, v17.2d, v25.2d[1]
[all …]
|
| H A D | kernel_sgemm_8x8_lib4.S | 153 fmla v0.4s, v24.4s, v28.4s[0]
154 fmla v1.4s, v24.4s, v28.4s[1]
156 fmla v2.4s, v24.4s, v28.4s[2]
157 fmla v3.4s, v24.4s, v28.4s[3]
159 fmla v4.4s, v20.4s, v28.4s[0]
160 fmla v5.4s, v20.4s, v28.4s[1]
162 fmla v6.4s, v20.4s, v28.4s[2]
163 fmla v7.4s, v20.4s, v28.4s[3]
165 fmla v8.4s, v24.4s, v16.4s[0]
166 fmla v9.4s, v24.4s, v16.4s[1]
[all …]
|
| H A D | kernel_dgemm_4x4_lib.S | 117 fmla v0.2d, v16.2d, v24.2d[0]
118 fmla v1.2d, v17.2d, v24.2d[0]
121 fmla v2.2d, v16.2d, v24.2d[1]
122 fmla v3.2d, v17.2d, v24.2d[1]
125 fmla v4.2d, v16.2d, v25.2d[0]
126 fmla v5.2d, v17.2d, v25.2d[0]
129 fmla v6.2d, v16.2d, v25.2d[1]
130 fmla v7.2d, v17.2d, v25.2d[1]
134 fmla v8.2d, v18.2d, v26.2d[0]
135 fmla v9.2d, v19.2d, v26.2d[0]
[all …]
|
| H A D | kernel_dgemm_12x4_lib.S | 93 fmla v0.2d, v24.2d, v30.2d[0]
94 fmla v1.2d, v25.2d, v30.2d[0]
95 fmla v8.2d, v26.2d, v30.2d[0]
96 fmla v9.2d, v27.2d, v30.2d[0]
99 fmla v2.2d, v24.2d, v30.2d[1]
100 fmla v3.2d, v25.2d, v30.2d[1]
105 fmla v4.2d, v24.2d, v31.2d[0]
106 fmla v5.2d, v25.2d, v31.2d[0]
111 fmla v6.2d, v24.2d, v31.2d[1]
112 fmla v7.2d, v25.2d, v31.2d[1]
[all …]
|
| H A D | kernel_dgemm_4x4_lib4.S | 155 fmla v0.2d, v16.2d, v24.2d[0]
156 fmla v1.2d, v17.2d, v24.2d[0]
158 fmla v2.2d, v16.2d, v24.2d[1]
159 fmla v3.2d, v17.2d, v24.2d[1]
161 fmla v4.2d, v16.2d, v25.2d[0]
162 fmla v5.2d, v17.2d, v25.2d[0]
164 fmla v6.2d, v16.2d, v25.2d[1]
165 fmla v7.2d, v17.2d, v25.2d[1]
169 fmla v8.2d, v18.2d, v26.2d[0]
170 fmla v9.2d, v19.2d, v26.2d[0]
[all …]
|
| H A D | kernel_dgemm_12x4_lib4.S | 198 fmla v0.2d, v24.2d, v30.2d[0]
200 fmla v2.2d, v24.2d, v30.2d[1]
202 fmla v1.2d, v25.2d, v30.2d[0]
207 fmla v3.2d, v25.2d, v30.2d[1]
209 fmla v8.2d, v26.2d, v30.2d[0]
216 fmla v4.2d, v24.2d, v31.2d[0]
218 fmla v6.2d, v24.2d, v31.2d[1]
220 fmla v5.2d, v25.2d, v31.2d[0]
225 fmla v7.2d, v25.2d, v31.2d[1]
234 fmla v9.2d, v27.2d, v30.2d[0]
[all …]
|
| H A D | kernel_sgemm_16x4_lib4.S | 156 fmla v0.4s, v16.4s, v28.4s[0]
157 fmla v4.4s, v18.4s, v28.4s[0]
159 fmla v8.4s, v20.4s, v28.4s[0]
163 fmla v1.4s, v16.4s, v28.4s[1]
164 fmla v5.4s, v18.4s, v28.4s[1]
166 fmla v9.4s, v20.4s, v28.4s[1]
170 fmla v2.4s, v16.4s, v28.4s[2]
171 fmla v6.4s, v18.4s, v28.4s[2]
177 fmla v3.4s, v16.4s, v28.4s[3]
178 fmla v7.4s, v18.4s, v28.4s[3]
[all …]
|
| /dports/graphics/vapoursynth-waifu2x-ncnn-vulkan/vapoursynth-waifu2x-ncnn-vulkan-r4/deps/ncnn/docs/developer-guide/ |
| H A D | arm-a53-a55-dual-issue.md | 29 fmla
31 fmla
33 fmla
36 fmla
38 fmla
40 fmla
43 fmla
44 fmla
45 fmla
48 fmla
[all …]
|
| /dports/graphics/waifu2x-ncnn-vulkan/waifu2x-ncnn-vulkan-20210521/src/ncnn/docs/developer-guide/ |
| H A D | arm-a53-a55-dual-issue.md | 29 fmla
31 fmla
33 fmla
36 fmla
38 fmla
40 fmla
43 fmla
44 fmla
45 fmla
48 fmla
[all …]
|
| /dports/benchmarks/vkpeak/vkpeak-20210430/ncnn/docs/developer-guide/ |
| H A D | arm-a53-a55-dual-issue.md | 29 fmla
31 fmla
33 fmla
36 fmla
38 fmla
40 fmla
43 fmla
44 fmla
45 fmla
48 fmla
[all …]
|
| /dports/misc/ncnn/ncnn-20211208/docs/developer-guide/ |
| H A D | arm-a53-a55-dual-issue.md | 29 fmla
31 fmla
33 fmla
36 fmla
38 fmla
40 fmla
43 fmla
44 fmla
45 fmla
48 fmla
[all …]
|
| /dports/graphics/realsr-ncnn-vulkan/realsr-ncnn-vulkan-20210210/src/ncnn/docs/developer-guide/ |
| H A D | arm-a53-a55-dual-issue.md | 29 fmla
31 fmla
33 fmla
36 fmla
38 fmla
40 fmla
43 fmla
44 fmla
45 fmla
48 fmla
[all …]
|
| /dports/misc/mnn/MNN-1.2.0/source/backend/cpu/arm/arm64/ |
| H A D | MNNPackedMatMul.S | 200 fmla v8.4s, v0.4s, v5.s[1]
201 fmla v9.4s, v0.4s, v5.s[1]
202 fmla v10.4s, v0.4s, v5.s[1]
203 fmla v11.4s, v0.4s, v5.s[1]
205 fmla v12.4s, v0.4s, v5.s[1]
206 fmla v13.4s, v0.4s, v5.s[1]
207 fmla v14.4s, v0.4s, v5.s[1]
208 fmla v15.4s, v0.4s, v5.s[1]
210 fmla v16.4s, v0.4s, v5.s[1]
211 fmla v17.4s, v0.4s, v5.s[1]
[all …]
|
| H A D | MNNPackedMatMul_BF16.S | 228 fmla v8.4s, v0.4s, v5.s[1]
229 fmla v9.4s, v0.4s, v5.s[1]
230 fmla v10.4s, v0.4s, v5.s[1]
231 fmla v11.4s, v0.4s, v5.s[1]
233 fmla v12.4s, v0.4s, v5.s[1]
234 fmla v13.4s, v0.4s, v5.s[1]
235 fmla v14.4s, v0.4s, v5.s[1]
236 fmla v15.4s, v0.4s, v5.s[1]
238 fmla v16.4s, v0.4s, v5.s[1]
239 fmla v17.4s, v0.4s, v5.s[1]
[all …]
|
| /dports/math/openblas/OpenBLAS-0.3.18/kernel/arm64/ |
| H A D | dgemm_kernel_4x8.S | 186 fmla v16.2d, v0.2d, v8.d[0]
187 fmla v17.2d, v1.2d, v8.d[0]
188 fmla v18.2d, v0.2d, v8.d[1]
189 fmla v19.2d, v1.2d, v8.d[1]
191 fmla v20.2d, v0.2d, v9.d[0]
192 fmla v21.2d, v1.2d, v9.d[0]
193 fmla v22.2d, v0.2d, v9.d[1]
194 fmla v23.2d, v1.2d, v9.d[1]
277 fmla v16.2d, v0.2d, v8.d[0]
278 fmla v17.2d, v1.2d, v8.d[0]
[all …]
|
| H A D | sgemm_kernel_8x8_cortexa53.S | 618 fmla s16, s0, v4.s[0]
619 fmla s18, s0, v4.s[1]
620 fmla s20, s0, v4.s[2]
621 fmla s22, s0, v4.s[3]
623 fmla s24, s0, v5.s[0]
624 fmla s26, s0, v5.s[1]
625 fmla s28, s0, v5.s[2]
626 fmla s30, s0, v5.s[3]
633 fmla s0, s16, alphaV0
639 fmla s2, s18, alphaV2
[all …]
|
| H A D | sgemm_kernel_8x8.S | 604 fmla s16, s0, v4.s[0]
605 fmla s18, s0, v4.s[1]
606 fmla s20, s0, v4.s[2]
607 fmla s22, s0, v4.s[3]
608 fmla s24, s0, v5.s[0]
609 fmla s26, s0, v5.s[1]
610 fmla s28, s0, v5.s[2]
611 fmla s30, s0, v5.s[3]
618 fmla s0, s16, alphaV0
624 fmla s2, s18, alphaV2
[all …]
|
| H A D | dgemm_kernel_8x4_thunderx2t99.S | 194 fmla v16.2d, v0.2d, v8.d[0]
195 fmla v20.2d, v0.2d, v8.d[1]
196 fmla v24.2d, v0.2d, v9.d[0]
197 fmla v28.2d, v0.2d, v9.d[1]
201 fmla v17.2d, v1.2d, v8.d[0]
202 fmla v25.2d, v1.2d, v9.d[0]
203 fmla v21.2d, v1.2d, v8.d[1]
204 fmla v29.2d, v1.2d, v9.d[1]
208 fmla v18.2d, v2.2d, v8.d[0]
209 fmla v22.2d, v2.2d, v8.d[1]
[all …]
|
| H A D | sgemm_kernel_16x4.S | 197 fmla v16.4s, v0.4s, v8.s[0]
198 fmla v17.4s, v1.4s, v8.s[0]
202 fmla v18.4s, v2.4s, v8.s[0]
203 fmla v19.4s, v3.4s, v8.s[0]
205 fmla v20.4s, v0.4s, v9.s[0]
206 fmla v21.4s, v1.4s, v9.s[0]
210 fmla v22.4s, v2.4s, v9.s[0]
211 fmla v23.4s, v3.4s, v9.s[0]
298 fmla v16.4s, v0.4s, v8.s[0]
299 fmla v20.4s, v0.4s, v9.s[0]
[all …]
|
| H A D | dgemm_kernel_4x4.S | 247 fmla v16.2d, v0.2d, v8.d[0]
253 fmla v18.2d, v2.2d, v8.d[0]
259 fmla v20.2d, v0.2d, v9.d[0]
264 fmla v22.2d, v2.2d, v9.d[0]
270 fmla v21.2d, v1.2d, v9.d[0]
276 fmla v23.2d, v3.2d, v9.d[0]
279 fmla v17.2d, v1.2d, v8.d[0]
285 fmla v19.2d, v3.2d, v8.d[0]
318 fmla v16.2d, v0.2d, v8.d[0]
320 fmla v20.2d, v0.2d, v9.d[0]
[all …]
|
| H A D | dgemm_kernel_8x4.S | 196 fmla v16.2d, v0.2d, v8.d[0]
197 fmla v20.2d, v0.2d, v9.d[0]
206 fmla v17.2d, v1.2d, v8.d[0]
211 fmla v21.2d, v1.2d, v9.d[0]
216 fmla v18.2d, v2.2d, v8.d[0]
217 fmla v22.2d, v2.2d, v9.d[0]
223 fmla v19.2d, v3.2d, v8.d[0]
224 fmla v23.2d, v3.2d, v9.d[0]
296 fmla v16.2d, v0.2d, v8.d[0]
297 fmla v20.2d, v0.2d, v9.d[0]
[all …]
|
| H A D | sgemm_kernel_4x4.S | 292 fmla v16.4s, v0.4s, v8.s[0]
293 fmla v17.4s, v2.4s, v8.s[0]
298 fmla v18.4s, v4.4s, v8.s[0]
299 fmla v19.4s, v6.4s, v8.s[0]
304 fmla v20.4s, v0.4s, v8.s[1]
305 fmla v21.4s, v2.4s, v8.s[1]
310 fmla v22.4s, v4.4s, v8.s[1]
311 fmla v23.4s, v6.4s, v8.s[1]
316 fmla v24.4s, v0.4s, v8.s[2]
317 fmla v25.4s, v2.4s, v8.s[2]
[all …]
|
| /dports/lang/eisl/eisl-2.35/library/ |
| H A D | formula.lsp | 71 (infip fmla))
73 (defun infip (fmla)
74 (if (atom fmla)
75 fmla
79 (if (or (< (weight (op fmla)) 5) (> (weight (op fmla)) 7))
80 (inf2 (cdr fmla) optr (cons (infip (car fmla)) opln))
81 (inf3 (cddr fmla) optr (cons (list (op fmla) (infip (arg1 fmla))) opln))))
87 (inf1 (cdr fmla) (cons (car fmla) optr) opln))
90 fmla
100 (inf1 (cdr fmla) (cons (car fmla) optr) opln))
[all …]
|
| /dports/misc/mnn/MNN-1.2.0/source/backend/arm82/asm/arm64/ |
| H A D | MNNPackedMatMulFP16.S | 197 fmla v8.8h, v0.8h, v5.h[1]
198 fmla v9.8h, v0.8h, v5.h[1]
199 fmla v10.8h, v0.8h, v5.h[1]
200 fmla v11.8h, v0.8h, v5.h[1]
202 fmla v12.8h, v0.8h, v5.h[1]
203 fmla v13.8h, v0.8h, v5.h[1]
204 fmla v14.8h, v0.8h, v5.h[1]
205 fmla v15.8h, v0.8h, v5.h[1]
207 fmla v16.8h, v0.8h, v5.h[1]
340 fmla v8.8h, v0.8h, v5.h[1]
[all …]
|