| /external/libavc/common/armv8/ |
| D | ih264_inter_pred_chroma_av8.s | 155 umlal v20.8h, v3.8b, v29.8b
157 umlal v20.8h, v5.8b, v30.8b
159 umlal v20.8h, v8.8b, v31.8b
163 umlal v22.8h, v4.8b, v29.8b
165 umlal v22.8h, v6.8b, v30.8b
167 umlal v22.8h, v9.8b, v31.8b
171 umlal v24.8h, v8.8b, v29.8b
173 umlal v24.8h, v10.8b, v30.8b
175 umlal v24.8h, v13.8b, v31.8b
179 umlal v16.8h, v9.8b, v29.8b
[all …]
|
| D | ih264_inter_pred_luma_horz_qpel_vert_qpel_av8.s | 154 umlal v24.8h, v4.8b, v30.8b
155 umlal v24.8h, v6.8b, v30.8b
165 umlal v28.8h, v20.8b, v30.8b
166 umlal v28.8h, v21.8b, v30.8b
171 umlal v24.8h, v5.8b, v30.8b
172 umlal v24.8h, v7.8b, v30.8b
186 umlal v24.8h, v20.8b, v30.8b
187 umlal v24.8h, v21.8b, v30.8b
192 umlal v16.8h, v6.8b, v30.8b
193 umlal v16.8h, v8.8b, v30.8b
[all …]
|
| D | ih264_inter_pred_filters_luma_horz_av8.s | 139 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 (column1,row0)
141 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 (column2,row0)
143 umlal v14.8h, v28.8b, v1.8b //// a0 + a5 + 20a2 (column1,row1)
145 umlal v16.8h, v27.8b, v1.8b //// a0 + a5 + 20a2 (column2,row1)
147 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column1,row0)
149 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column2,row0)
151 umlal v14.8h, v28.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column1,row1)
153 umlal v16.8h, v27.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column2,row1)
192 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 (column1,row2)
194 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 (column2,row2)
[all …]
|
| D | ih264_inter_pred_luma_horz_qpel_av8.s | 147 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 (column1,row0)
149 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 (column2,row0)
151 umlal v14.8h, v28.8b, v1.8b //// a0 + a5 + 20a2 (column1,row1)
153 umlal v16.8h, v27.8b, v1.8b //// a0 + a5 + 20a2 (column2,row1)
155 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column1,row0)
157 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column2,row0)
159 umlal v14.8h, v28.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column1,row1)
161 umlal v16.8h, v27.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column2,row1)
208 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 (column1,row2)
210 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 (column2,row2)
[all …]
|
| /external/libhevc/common/arm64/ |
| D | ihevc_intra_pred_luma_planar.s | 184 umlal v27.8h, v5.8b, v0.8b //(1)(row+1) * src[nt-1]
187 umlal v27.8h, v17.8b, v1.8b //(1)(col+1) * src[3nt+1]
190 umlal v27.8h, v6.8b, v3.8b //(1)(nt-1-row) * src[2nt+1+col]
193 umlal v27.8h, v19.8b, v20.8b //(1)(nt-1-col) * src[2nt-1-row]
201 umlal v30.8h, v5.8b, v0.8b //(2)
204 umlal v30.8h, v17.8b, v1.8b //(2)
206 umlal v30.8h, v6.8b, v3.8b //(2)
207 umlal v30.8h, v19.8b, v21.8b //(2)
215 umlal v28.8h, v5.8b, v0.8b //(3)
218 umlal v28.8h, v17.8b, v1.8b //(3)
[all …]
|
| D | ihevc_inter_pred_filters_luma_vert_w16out.s | 127 umlal v19.8h, v3.8b, v25.8b //mul_res1 = vmlal_u8(mul_res1, src_tmp4, coeffabs_3)//
129 umlal v19.8h, v4.8b, v26.8b //mul_res1 = vmlal_u8(mul_res1, src_tmp1, coeffabs_4)//
133 umlal v19.8h, v6.8b, v28.8b //mul_res1 = vmlal_u8(mul_res1, src_tmp3, coeffabs_6)//
151 umlal v20.8h, v4.8b, v25.8b //mul_res2 = vmlal_u8(mul_res2, src_tmp1, coeffabs_3)//
154 umlal v20.8h, v5.8b, v26.8b //mul_res2 = vmlal_u8(mul_res2, src_tmp2, coeffabs_4)//
159 umlal v20.8h, v7.8b, v28.8b //mul_res2 = vmlal_u8(mul_res2, src_tmp4, coeffabs_6)//
171 umlal v21.8h, v5.8b, v25.8b
172 umlal v21.8h, v6.8b, v26.8b
174 umlal v21.8h, v16.8b, v28.8b
186 umlal v30.8h, v6.8b, v25.8b
[all …]
|
| D | ihevc_intra_pred_chroma_planar.s | 190 umlal v12.8h, v6.8b, v10.8b //(nt-1-row) * src[2nt+1+col]
192 umlal v12.8h, v17.8b, v1.8b //(col+1) * src[3nt+1]
194 umlal v12.8h, v30.8b, v4.8b //(nt-1-col) * src[2nt-1-row]
201 umlal v28.8h, v6.8b, v11.8b
205 umlal v28.8h, v31.8b, v4.8b
207 umlal v28.8h, v25.8b, v1.8b
212 umlal v26.8h, v19.8b, v10.8b //(nt-1-row) * src[2nt+1+col]
214 umlal v26.8h, v17.8b, v1.8b //(col+1) * src[3nt+1]
216 umlal v26.8h, v30.8b, v3.8b //(nt-1-col) * src[2nt-1-row]
225 umlal v24.8h, v19.8b, v11.8b
[all …]
|
| D | ihevc_inter_pred_filters_luma_vert.s | 170 umlal v19.8h, v3.8b, v25.8b //mul_res1 = vmlal_u8(mul_res1, src_tmp4, coeffabs_3)//
172 umlal v19.8h, v4.8b, v26.8b //mul_res1 = vmlal_u8(mul_res1, src_tmp1, coeffabs_4)//
176 umlal v19.8h, v6.8b, v28.8b //mul_res1 = vmlal_u8(mul_res1, src_tmp3, coeffabs_6)//
193 umlal v20.8h, v4.8b, v25.8b //mul_res2 = vmlal_u8(mul_res2, src_tmp1, coeffabs_3)//
196 umlal v20.8h, v5.8b, v26.8b //mul_res2 = vmlal_u8(mul_res2, src_tmp2, coeffabs_4)//
202 umlal v20.8h, v7.8b, v28.8b //mul_res2 = vmlal_u8(mul_res2, src_tmp4, coeffabs_6)//
217 umlal v21.8h, v5.8b, v25.8b
218 umlal v21.8h, v6.8b, v26.8b
220 umlal v21.8h, v16.8b, v28.8b
232 umlal v30.8h, v6.8b, v25.8b
[all …]
|
| D | ihevc_inter_pred_luma_horz_w16out.s | 279 umlal v8.8h, v3.8b, v27.8b
280 umlal v8.8h, v4.8b, v28.8b
282 umlal v8.8h, v6.8b, v30.8b
345 umlal v8.8h, v3.8b, v27.8b
347 umlal v8.8h, v4.8b, v28.8b
351 umlal v8.8h, v6.8b, v30.8b
417 umlal v8.8h, v3.8b, v27.8b //mul_res = vmull_u8(src[0_3], coeffabs_3)//
423 umlal v8.8h, v4.8b, v28.8b //mul_res = vmlal_u8(src[0_4], coeffabs_4)//
427 umlal v8.8h, v6.8b, v30.8b //mul_res = vmlal_u8(src[0_6], coeffabs_6)//
436 umlal v10.8h, v16.8b, v28.8b //mul_res = vmlal_u8(src[0_4], coeffabs_4)//
[all …]
|
| D | ihevc_inter_pred_filters_luma_horz.s | 231 umlal v8.8h, v3.8b, v27.8b //mul_res = vmull_u8(src[0_3], coeffabs_3)//
237 umlal v8.8h, v4.8b, v28.8b //mul_res = vmlal_u8(src[0_4], coeffabs_4)//
241 umlal v8.8h, v6.8b, v30.8b //mul_res = vmlal_u8(src[0_6], coeffabs_6)//
250 umlal v10.8h, v16.8b, v28.8b //mul_res = vmlal_u8(src[0_4], coeffabs_4)//
254 umlal v10.8h, v18.8b, v30.8b //mul_res = vmlal_u8(src[0_6], coeffabs_6)//
258 umlal v10.8h, v13.8b, v25.8b //mul_res = vmlal_u8(src[0_1], coeffabs_1)//
322 umlal v8.8h, v6.8b, v27.8b //mul_res = vmull_u8(src[0_3], coeffabs_3)//
329 umlal v8.8h, v12.8b, v28.8b //mul_res = vmlal_u8(src[0_4], coeffabs_4)//
331 umlal v8.8h, v16.8b, v30.8b //mul_res = vmlal_u8(src[0_6], coeffabs_6)//
346 umlal v20.8h, v7.8b, v27.8b
[all …]
|
| D | ihevc_inter_pred_chroma_vert_w16out.s | 155 …umlal v6.8h, v4.8b, v2.8b //vmlal_u8(mul_res1, vreinterpret_u8_u32(src_tmp3), coeffa…
160 umlal v4.8h, v16.8b, v2.8b
198 umlal v4.8h, v7.8b, v2.8b
242 umlal v30.8h, v6.8b, v2.8b
252 umlal v28.8h, v7.8b, v2.8b
261 umlal v26.8h, v16.8b, v2.8b
272 umlal v24.8h, v17.8b, v2.8b
290 umlal v30.8h, v6.8b, v2.8b
307 umlal v28.8h, v7.8b, v2.8b
322 umlal v26.8h, v16.8b, v2.8b
[all …]
|
| D | ihevc_intra_pred_chroma_mode_27_to_33.s | 172 umlal v10.8h, v9.8b, v31.8b //(i row)vmull_u8(ref_main_idx_1, dup_const_fract)
184 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
198 umlal v18.8h, v17.8b, v27.8b //(iii)vmull_u8(ref_main_idx_1, dup_const_fract)
216 umlal v22.8h, v21.8b, v25.8b //(iv)vmull_u8(ref_main_idx_1, dup_const_fract)
233 umlal v10.8h, v9.8b, v31.8b //(v)vmull_u8(ref_main_idx_1, dup_const_fract)
249 umlal v14.8h, v13.8b, v29.8b //(vi)vmull_u8(ref_main_idx_1, dup_const_fract)
263 umlal v18.8h, v17.8b, v27.8b //(vii)vmull_u8(ref_main_idx_1, dup_const_fract)
298 umlal v22.8h, v21.8b, v25.8b //(viii)vmull_u8(ref_main_idx_1, dup_const_fract)
314 umlal v10.8h, v9.8b, v31.8b //(i)vmull_u8(ref_main_idx_1, dup_const_fract)
329 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
[all …]
|
| D | ihevc_intra_pred_luma_mode_27_to_33.s | 177 umlal v10.8h, v9.8b, v31.8b //(i row)vmull_u8(ref_main_idx_1, dup_const_fract)
189 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
204 umlal v18.8h, v17.8b, v27.8b //(iii)vmull_u8(ref_main_idx_1, dup_const_fract)
221 umlal v22.8h, v21.8b, v25.8b //(iv)vmull_u8(ref_main_idx_1, dup_const_fract)
238 umlal v10.8h, v9.8b, v31.8b //(v)vmull_u8(ref_main_idx_1, dup_const_fract)
254 umlal v14.8h, v13.8b, v29.8b //(vi)vmull_u8(ref_main_idx_1, dup_const_fract)
268 umlal v18.8h, v17.8b, v27.8b //(vii)vmull_u8(ref_main_idx_1, dup_const_fract)
303 umlal v22.8h, v21.8b, v25.8b //(viii)vmull_u8(ref_main_idx_1, dup_const_fract)
319 umlal v10.8h, v9.8b, v31.8b //(i)vmull_u8(ref_main_idx_1, dup_const_fract)
335 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
[all …]
|
| D | ihevc_inter_pred_chroma_horz_w16out.s | 215 umlal v30.8h, v4.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
233 umlal v28.8h, v5.8b, v26.8b
275 umlal v22.8h, v12.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
294 umlal v20.8h, v13.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
312 umlal v30.8h, v4.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
328 umlal v28.8h, v5.8b, v26.8b
354 umlal v22.8h, v12.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
376 umlal v20.8h, v13.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
393 umlal v30.8h, v4.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
406 umlal v28.8h, v5.8b, v26.8b
[all …]
|
| D | ihevc_inter_pred_chroma_vert.s | 152 …umlal v6.8h, v4.8b, v2.8b //vmlal_u8(mul_res1, vreinterpret_u8_u32(src_tmp3), coeffa…
157 umlal v4.8h, v16.8b, v2.8b
197 umlal v4.8h, v7.8b, v2.8b
241 umlal v30.8h, v6.8b, v2.8b
251 umlal v28.8h, v7.8b, v2.8b
261 umlal v26.8h, v16.8b, v2.8b
273 umlal v24.8h, v17.8b, v2.8b
293 umlal v30.8h, v6.8b, v2.8b
310 umlal v28.8h, v7.8b, v2.8b
328 umlal v26.8h, v16.8b, v2.8b
[all …]
|
| D | ihevc_intra_pred_luma_mode_3_to_9.s | 191 umlal v24.8h, v13.8b, v6.8b //mul (row 0)
201 umlal v22.8h, v17.8b, v6.8b //mul (row 1)
212 umlal v20.8h, v15.8b, v6.8b //mul (row 2)
223 umlal v18.8h, v25.8b, v6.8b //mul (row 3)
234 umlal v24.8h, v13.8b, v6.8b //mul (row 4)
245 umlal v22.8h, v17.8b, v6.8b //mul (row 5)
256 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
260 umlal v18.8h, v25.8b, v6.8b //mul (row 7)
319 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
345 umlal v18.8h, v25.8b, v6.8b //mul (row 7)
[all …]
|
| D | ihevc_inter_pred_chroma_horz.s | 200 umlal v30.8h, v4.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
215 umlal v28.8h, v5.8b, v26.8b
265 umlal v22.8h, v12.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
287 umlal v20.8h, v13.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
305 umlal v30.8h, v4.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
328 umlal v28.8h, v5.8b, v26.8b
361 umlal v22.8h, v12.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
379 umlal v20.8h, v13.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
396 umlal v30.8h, v4.8b, v26.8b //mul_res = vmlsl_u8(src[0_0], coeffabs_0)//
412 umlal v28.8h, v5.8b, v26.8b
[all …]
|
| D | ihevc_intra_pred_filters_luma_mode_19_to_25.s | 287 umlal v10.8h, v9.8b, v31.8b //(i row)vmull_u8(ref_main_idx_1, dup_const_fract)
298 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
313 umlal v18.8h, v17.8b, v27.8b //(iii)vmull_u8(ref_main_idx_1, dup_const_fract)
329 umlal v22.8h, v21.8b, v25.8b //(iv)vmull_u8(ref_main_idx_1, dup_const_fract)
345 umlal v10.8h, v9.8b, v31.8b //(v)vmull_u8(ref_main_idx_1, dup_const_fract)
360 umlal v14.8h, v13.8b, v29.8b //(vi)vmull_u8(ref_main_idx_1, dup_const_fract)
374 umlal v18.8h, v17.8b, v27.8b //(vii)vmull_u8(ref_main_idx_1, dup_const_fract)
409 umlal v22.8h, v21.8b, v25.8b //(viii)vmull_u8(ref_main_idx_1, dup_const_fract)
425 umlal v10.8h, v9.8b, v31.8b //(i)vmull_u8(ref_main_idx_1, dup_const_fract)
442 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
[all …]
|
| D | ihevc_intra_pred_filters_chroma_mode_19_to_25.s | 284 umlal v23.8h, v19.8b, v31.8b //(i row)vmull_u8(ref_main_idx_1, dup_const_fract)
295 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
309 umlal v18.8h, v17.8b, v27.8b //(iii)vmull_u8(ref_main_idx_1, dup_const_fract)
326 umlal v22.8h, v21.8b, v25.8b //(iv)vmull_u8(ref_main_idx_1, dup_const_fract)
342 umlal v23.8h, v19.8b, v31.8b //(v)vmull_u8(ref_main_idx_1, dup_const_fract)
357 umlal v14.8h, v13.8b, v29.8b //(vi)vmull_u8(ref_main_idx_1, dup_const_fract)
374 umlal v18.8h, v17.8b, v27.8b //(vii)vmull_u8(ref_main_idx_1, dup_const_fract)
410 umlal v22.8h, v21.8b, v25.8b //(viii)vmull_u8(ref_main_idx_1, dup_const_fract)
426 umlal v23.8h, v19.8b, v31.8b //(i)vmull_u8(ref_main_idx_1, dup_const_fract)
442 umlal v14.8h, v13.8b, v29.8b //(ii)vmull_u8(ref_main_idx_1, dup_const_fract)
[all …]
|
| D | ihevc_intra_pred_filters_luma_mode_11_to_17.s | 311 umlal v24.8h, v13.8b, v6.8b //mul (row 0)
321 umlal v22.8h, v17.8b, v6.8b //mul (row 1)
332 umlal v20.8h, v15.8b, v6.8b //mul (row 2)
343 umlal v18.8h, v25.8b, v6.8b //mul (row 3)
354 umlal v24.8h, v13.8b, v6.8b //mul (row 4)
365 umlal v22.8h, v17.8b, v6.8b //mul (row 5)
376 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
380 umlal v18.8h, v25.8b, v6.8b //mul (row 7)
440 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
465 umlal v18.8h, v25.8b, v6.8b //mul (row 7)
[all …]
|
| D | ihevc_intra_pred_chroma_mode_3_to_9.s | 190 umlal v24.8h, v13.8b, v6.8b //mul (row 0)
200 umlal v22.8h, v17.8b, v6.8b //mul (row 1)
211 umlal v20.8h, v15.8b, v6.8b //mul (row 2)
222 umlal v18.8h, v23.8b, v6.8b //mul (row 3)
233 umlal v24.8h, v13.8b, v6.8b //mul (row 4)
246 umlal v22.8h, v17.8b, v6.8b //mul (row 5)
257 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
261 umlal v18.8h, v23.8b, v6.8b //mul (row 7)
325 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
353 umlal v18.8h, v23.8b, v6.8b //mul (row 7)
[all …]
|
| D | ihevc_intra_pred_filters_chroma_mode_11_to_17.s | 312 umlal v24.8h, v13.8b, v6.8b //mul (row 0)
322 umlal v22.8h, v17.8b, v6.8b //mul (row 1)
333 umlal v20.8h, v15.8b, v6.8b //mul (row 2)
344 umlal v18.8h, v25.8b, v6.8b //mul (row 3)
355 umlal v24.8h, v13.8b, v6.8b //mul (row 4)
368 umlal v22.8h, v17.8b, v6.8b //mul (row 5)
379 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
383 umlal v18.8h, v25.8b, v6.8b //mul (row 7)
450 umlal v20.8h, v15.8b, v6.8b //mul (row 6)
481 umlal v18.8h, v25.8b, v6.8b //mul (row 7)
[all …]
|
| /external/libavc/encoder/armv8/ |
| D | ih264e_half_pel_av8.s | 121 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 (column1,row0)
123 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 (column2,row0)
125 umlal v12.8h, v29.8b, v1.8b //// a0 + a5 + 20a2 (column3,row0)
127 umlal v14.8h, v28.8b, v1.8b //// a0 + a5 + 20a2 (column1,row1)
130 umlal v16.8h, v27.8b, v1.8b //// a0 + a5 + 20a2 (column2,row1)
132 umlal v18.8h, v26.8b, v1.8b //// a0 + a5 + 20a2 (column3,row1)
134 umlal v8.8h, v31.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column1,row0)
136 umlal v10.8h, v30.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column2,row0)
138 umlal v12.8h, v29.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column3,row0)
140 umlal v14.8h, v28.8b, v1.8b //// a0 + a5 + 20a2 + 20a3 (column1,row1)
[all …]
|
| /external/llvm/test/CodeGen/ARM/ |
| D | longMAC.ll | 9 ;CHECK: umlal
29 ;CHECK: umlal
52 ; + Without "$Rd = $R", this can be satisfied without a mov before the umlal
55 ; function, both after the umlal. With it, *some* move has to happen
56 ; before the umlal.
59 ; CHECK-V7-LE-LABEL: umlal r0, r1, r0, r0
61 ; CHECK-V7-BE-LABEL: umlal r1, r0, r1, r1
65 ; CHECK-LE: umlal [[RDLO]], r1, r0, r0
68 ; CHECK-BE: umlal [[RDLO]], r0, r1, r1
|
| /external/boringssl/linux-arm/crypto/bn/ |
| D | armv4-mont.S | 64 umlal r10,r12,r6,r8 @ np[0]*n0+"t[0]"
72 umlal r10,r11,r5,r2 @ ap[j]*bp[0]
74 umlal r12,r14,r6,r8 @ np[j]*n0
100 umlal r10,r11,r5,r2 @ ap[0]*bp[i]+tp[0]
104 umlal r10,r12,r6,r8 @ np[0]*n0+"tp[0]"
112 umlal r10,r11,r5,r2 @ ap[j]*bp[i]
114 umlal r12,r14,r6,r8 @ np[j]*n0
|