Lines Matching full:into
119 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs0 into ymm2 in Run()
120 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mul add lhs rhs1 into ymm3 in Run()
121 "vpaddd %%ymm2, %%ymm4, %%ymm4 \n\t" // add muladd lhs + rhs0 into ymm4 in Run()
122 "vpaddd %%ymm3, %%ymm5, %%ymm5 \n\t" // add muladd lhs + rhs1 into ymm5 in Run()
125 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rh3 into ymm2 in Run()
126 "vpshufd $0xff,%%ymm1,%%ymm3 \n\t" // mov rhs 3 element into all ymm3 in Run()
127 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mul add lhs rh4 into ymm3 in Run()
128 "vpaddd %%ymm2, %%ymm6, %%ymm6 \n\t" // add muladd lhs + rhs2 into ymm6 in Run()
129 "vpaddd %%ymm3, %%ymm7, %%ymm7 \n\t" // add muladd lhs + rhs3 into ymm7 in Run()
141 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs0 into ymm2 in Run()
142 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mul add lhs rhs1 into ymm3 in Run()
143 "vpaddd %%ymm2, %%ymm8, %%ymm8 \n\t" // add muladd lhs + rhs0 into ymm8 in Run()
144 "vpaddd %%ymm3, %%ymm9, %%ymm9 \n\t" // add muladd lhs + rhs1 into ymm9 in Run()
148 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs2 into ymm2 in Run()
149 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mul add lhs rhs3 into ymm3 in Run()
150 "vpaddd %%ymm2, %%ymm10, %%ymm10 \n\t" // add muladd lhs + rhs2 into ymm10 in Run()
151 "vpaddd %%ymm3, %%ymm11, %%ymm11 \n\t" // add muladd lhs + rhs3 into ymm11 in Run()
161 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs0 into ymm2 in Run()
162 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mul add lhs rhs1 into ymm3 in Run()
163 "vpaddd %%ymm2, %%ymm12, %%ymm12 \n\t" // add muladd lhs + rhs0 into ymm8 in Run()
164 "vpaddd %%ymm3, %%ymm13, %%ymm13 \n\t" // add muladd lhs + rhs1 into ymm9 in Run()
172 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs2 into ymm2 in Run()
173 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mul add lhs rhs3 into ymm3 in Run()
174 "vpaddd %%ymm2, %%ymm14, %%ymm14 \n\t" // add muladd lhs + rhs2 into ymm10 in Run()
175 "vpaddd %%ymm3, %%ymm15, %%ymm15 \n\t" // add muladd lhs + rhs3 into ymm11 in Run()
186 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // muladd lhs rhs0 into ymm2 in Run()
187 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // muladd lhs rhs1 into ymm3 in Run()
193 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs2 into ymm2 in Run()
194 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mull add lhs rhs3 into ymm3 in Run()
206 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // muladd lhs rhs0 into ymm2 in Run()
207 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // muladd lhs rhs1 into ymm3 in Run()
213 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs2 into ymm2 in Run()
214 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mull add lhs rhs3 into ymm3 in Run()
224 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // muladd lhs rhs0 into ymm2 in Run()
225 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // muladd lhs rhs1 into ymm3 in Run()
232 "vpmaddwd %%ymm0, %%ymm2, %%ymm2 \n\t" // mul add lhs rhs2 into ymm2 in Run()
233 "vpmaddwd %%ymm0, %%ymm3, %%ymm3 \n\t" // mull add lhs rhs3 into ymm3 in Run()
252 "vpmovzxbw (%[rhs_ptr]), %%ymm1 \n\t" // get rhs into ymm1 in Run()
256 "vpmovzxbw (%[lhs_ptr]), %%ymm0 \n\t" // lhs in into ymm0 in Run()
257 "vpshufd $0x00,%%ymm1,%%ymm2 \n\t" // rhs element 0 into ymm2 in Run()
258 "vpshufd $0x55,%%ymm1,%%ymm3 \n\t" // rhs element 1 into ymm3 in Run()
263 "vpshufd $0xaa,%%ymm1,%%ymm2 \n\t" // rhs element 2 into ymm2 in Run()
264 "vpshufd $0xff,%%ymm1,%%ymm3 \n\t" // rhs element 3 into ymm3 in Run()
267 "vpaddd %%ymm2, %%ymm6, %%ymm6 \n\t" // acc element 2 into ymm6 in Run()
268 "vpaddd %%ymm3, %%ymm7, %%ymm7 \n\t" // acc element 3 into ymm7 in Run()
271 "vpmovzxbw 0x10(%[lhs_ptr]), %%ymm0 \n\t" // lhs in into ymm0 in Run()
272 "vpshufd $0x00, %%ymm1, %%ymm2 \n\t" // rhs element 0 into ymm2 in Run()
273 "vpshufd $0x55, %%ymm1, %%ymm3 \n\t" // rhs element 1 into ymm3 in Run()
278 "vpshufd $0xaa,%%ymm1,%%ymm2 \n\t" // rhs element 2 into ymm2 in Run()
279 "vpshufd $0xff,%%ymm1,%%ymm3 \n\t" // rhs element 3 into ymm3 in Run()
282 "vpaddd %%ymm2, %%ymm10, %%ymm10 \n\t" // acc element 2 into ymm10 in Run()
283 "vpaddd %%ymm3, %%ymm11, %%ymm11 \n\t" // acc element 3 into ymm11 in Run()
286 "vpshufd $0x00, %%ymm1, %%ymm2 \n\t" // rhs element 0 into ymm2 in Run()
287 "vpshufd $0x55, %%ymm1, %%ymm3 \n\t" // rhs element 1 into ymm3 in Run()
292 "vpshufd $0xaa,%%ymm1,%%ymm2 \n\t" // rhs element 2 into ymm2 in Run()
293 "vpshufd $0xff,%%ymm1,%%ymm3 \n\t" // rhs element 3 into ymm3 in Run()
296 "vpaddd %%ymm2, %%ymm14, %%ymm14 \n\t" // acc element 2 into ymm14 in Run()
297 "vpaddd %%ymm3, %%ymm15, %%ymm15 \n\t" // acc element 3 into ymm15 in Run()