/dports/math/blis/blis-0.8.1/kernels/zen/1/ |
H A D | bli_setv_zen_int.c | 160 _mm256_storeu_pd(x + num_elem_per_reg * 0, alphav); in bli_dsetv_zen_int() 161 _mm256_storeu_pd(x + num_elem_per_reg * 1, alphav); in bli_dsetv_zen_int() 162 _mm256_storeu_pd(x + num_elem_per_reg * 2, alphav); in bli_dsetv_zen_int() 163 _mm256_storeu_pd(x + num_elem_per_reg * 3, alphav); in bli_dsetv_zen_int() 164 _mm256_storeu_pd(x + num_elem_per_reg * 4, alphav); in bli_dsetv_zen_int() 165 _mm256_storeu_pd(x + num_elem_per_reg * 5, alphav); in bli_dsetv_zen_int() 166 _mm256_storeu_pd(x + num_elem_per_reg * 6, alphav); in bli_dsetv_zen_int() 167 _mm256_storeu_pd(x + num_elem_per_reg * 7, alphav); in bli_dsetv_zen_int() 168 _mm256_storeu_pd(x + num_elem_per_reg * 8, alphav); in bli_dsetv_zen_int() 169 _mm256_storeu_pd(x + num_elem_per_reg * 9, alphav); in bli_dsetv_zen_int() [all …]
|
H A D | bli_copyv_zen_int.c | 238 _mm256_storeu_pd(y + num_elem_per_reg * 0, xv[0]); in bli_dcopyv_zen_int() 239 _mm256_storeu_pd(y + num_elem_per_reg * 1, xv[1]); in bli_dcopyv_zen_int() 240 _mm256_storeu_pd(y + num_elem_per_reg * 2, xv[2]); in bli_dcopyv_zen_int() 241 _mm256_storeu_pd(y + num_elem_per_reg * 3, xv[3]); in bli_dcopyv_zen_int() 242 _mm256_storeu_pd(y + num_elem_per_reg * 4, xv[4]); in bli_dcopyv_zen_int() 243 _mm256_storeu_pd(y + num_elem_per_reg * 5, xv[5]); in bli_dcopyv_zen_int() 244 _mm256_storeu_pd(y + num_elem_per_reg * 6, xv[6]); in bli_dcopyv_zen_int() 245 _mm256_storeu_pd(y + num_elem_per_reg * 7, xv[7]); in bli_dcopyv_zen_int() 246 _mm256_storeu_pd(y + num_elem_per_reg * 8, xv[8]); in bli_dcopyv_zen_int() 247 _mm256_storeu_pd(y + num_elem_per_reg * 9, xv[9]); in bli_dcopyv_zen_int() [all …]
|
H A D | bli_swapv_zen_int8.c | 248 _mm256_storeu_pd( (x0 + 0*n_elem_per_reg), yv[0]); in bli_dswapv_zen_int8() 249 _mm256_storeu_pd( (x0 + 1*n_elem_per_reg), yv[1]); in bli_dswapv_zen_int8() 250 _mm256_storeu_pd( (x0 + 2*n_elem_per_reg), yv[2]); in bli_dswapv_zen_int8() 251 _mm256_storeu_pd( (x0 + 3*n_elem_per_reg), yv[3]); in bli_dswapv_zen_int8() 252 _mm256_storeu_pd( (x0 + 4*n_elem_per_reg), yv[4]); in bli_dswapv_zen_int8() 253 _mm256_storeu_pd( (x0 + 5*n_elem_per_reg), yv[5]); in bli_dswapv_zen_int8() 254 _mm256_storeu_pd( (x0 + 6*n_elem_per_reg), yv[6]); in bli_dswapv_zen_int8() 255 _mm256_storeu_pd( (x0 + 7*n_elem_per_reg), yv[7]); in bli_dswapv_zen_int8() 257 _mm256_storeu_pd( (y0 + 0*n_elem_per_reg), xv[0]); in bli_dswapv_zen_int8() 258 _mm256_storeu_pd( (y0 + 1*n_elem_per_reg), xv[1]); in bli_dswapv_zen_int8() [all …]
|
H A D | bli_scalv_zen_int10.c | 343 _mm256_storeu_pd( (x0 + 0*n_elem_per_reg), zv[0] ); in bli_dscalv_zen_int10() 344 _mm256_storeu_pd( (x0 + 1*n_elem_per_reg), zv[1] ); in bli_dscalv_zen_int10() 345 _mm256_storeu_pd( (x0 + 2*n_elem_per_reg), zv[2] ); in bli_dscalv_zen_int10() 346 _mm256_storeu_pd( (x0 + 3*n_elem_per_reg), zv[3] ); in bli_dscalv_zen_int10() 347 _mm256_storeu_pd( (x0 + 4*n_elem_per_reg), zv[4] ); in bli_dscalv_zen_int10() 348 _mm256_storeu_pd( (x0 + 5*n_elem_per_reg), zv[5] ); in bli_dscalv_zen_int10() 349 _mm256_storeu_pd( (x0 + 6*n_elem_per_reg), zv[6] ); in bli_dscalv_zen_int10() 350 _mm256_storeu_pd( (x0 + 7*n_elem_per_reg), zv[7] ); in bli_dscalv_zen_int10() 351 _mm256_storeu_pd( (x0 + 8*n_elem_per_reg), zv[8] ); in bli_dscalv_zen_int10() 352 _mm256_storeu_pd( (x0 + 9*n_elem_per_reg), zv[9] ); in bli_dscalv_zen_int10() [all …]
|
H A D | bli_axpyv_zen_int10.c | 334 _mm256_storeu_pd( (y0 + 0*n_elem_per_reg), zv[0] ); in bli_daxpyv_zen_int10() 335 _mm256_storeu_pd( (y0 + 1*n_elem_per_reg), zv[1] ); in bli_daxpyv_zen_int10() 336 _mm256_storeu_pd( (y0 + 2*n_elem_per_reg), zv[2] ); in bli_daxpyv_zen_int10() 337 _mm256_storeu_pd( (y0 + 3*n_elem_per_reg), zv[3] ); in bli_daxpyv_zen_int10() 338 _mm256_storeu_pd( (y0 + 4*n_elem_per_reg), zv[4] ); in bli_daxpyv_zen_int10() 339 _mm256_storeu_pd( (y0 + 5*n_elem_per_reg), zv[5] ); in bli_daxpyv_zen_int10() 340 _mm256_storeu_pd( (y0 + 6*n_elem_per_reg), zv[6] ); in bli_daxpyv_zen_int10() 341 _mm256_storeu_pd( (y0 + 7*n_elem_per_reg), zv[7] ); in bli_daxpyv_zen_int10() 342 _mm256_storeu_pd( (y0 + 8*n_elem_per_reg), zv[8] ); in bli_daxpyv_zen_int10() 343 _mm256_storeu_pd( (y0 + 9*n_elem_per_reg), zv[9] ); in bli_daxpyv_zen_int10() [all …]
|
/dports/math/blasfeo/blasfeo-0.1.2/kernel/avx2/ |
H A D | kernel_dgetrf_pivot_lib.c | 221 _mm256_storeu_pd( &pA[0+lda*0], a_0 ); in kernel_dgetrf_pivot_12_lib() 226 _mm256_storeu_pd( &pA[0+lda*1], c_0 ); in kernel_dgetrf_pivot_12_lib() 230 _mm256_storeu_pd( &pA[0+lda*2], a_i ); in kernel_dgetrf_pivot_12_lib() 234 _mm256_storeu_pd( &pA[0+lda*3], a_i ); in kernel_dgetrf_pivot_12_lib() 238 _mm256_storeu_pd( &pA[0+lda*4], a_i ); in kernel_dgetrf_pivot_12_lib() 242 _mm256_storeu_pd( &pA[0+lda*5], a_i ); in kernel_dgetrf_pivot_12_lib() 246 _mm256_storeu_pd( &pA[0+lda*6], a_i ); in kernel_dgetrf_pivot_12_lib() 250 _mm256_storeu_pd( &pA[0+lda*7], a_i ); in kernel_dgetrf_pivot_12_lib() 254 _mm256_storeu_pd( &pA[0+lda*8], a_i ); in kernel_dgetrf_pivot_12_lib() 258 _mm256_storeu_pd( &pA[0+lda*9], a_i ); in kernel_dgetrf_pivot_12_lib() [all …]
|
/dports/science/madness/madness-ebb3fd7/src/madness/tensor/mtxmq_graveyard/new_mtxmq/bests/ |
H A D | avx2_rr.cc | 64 _mm256_storeu_pd(c+(i+0)*effj+0, _c_0_0); in mtxmq() 65 _mm256_storeu_pd(c+(i+0)*effj+4, _c_0_4); in mtxmq() 66 _mm256_storeu_pd(c+(i+0)*effj+8, _c_0_8); in mtxmq() 69 _mm256_storeu_pd(c+(i+1)*effj+0, _c_1_0); in mtxmq() 70 _mm256_storeu_pd(c+(i+1)*effj+4, _c_1_4); in mtxmq() 71 _mm256_storeu_pd(c+(i+1)*effj+8, _c_1_8); in mtxmq() 96 _mm256_storeu_pd(c+(i+0)*effj+0, _c_0_0); in mtxmq() 97 _mm256_storeu_pd(c+(i+0)*effj+4, _c_0_4); in mtxmq() 98 _mm256_storeu_pd(c+(i+0)*effj+8, _c_0_8); in mtxmq() 136 _mm256_storeu_pd(c+(i+0)*effj+0, _c_0_0); in mtxmq() [all …]
|
H A D | avx_rr.cc | 63 _mm256_storeu_pd(c+(i+0)*dimj+0, _c_0_0); in mtxmq() 64 _mm256_storeu_pd(c+(i+0)*dimj+4, _c_0_4); in mtxmq() 65 _mm256_storeu_pd(c+(i+0)*dimj+8, _c_0_8); in mtxmq() 68 _mm256_storeu_pd(c+(i+1)*dimj+0, _c_1_0); in mtxmq() 69 _mm256_storeu_pd(c+(i+1)*dimj+4, _c_1_4); in mtxmq() 70 _mm256_storeu_pd(c+(i+1)*dimj+8, _c_1_8); in mtxmq() 95 _mm256_storeu_pd(c+(i+0)*dimj+0, _c_0_0); in mtxmq() 96 _mm256_storeu_pd(c+(i+0)*dimj+4, _c_0_4); in mtxmq() 97 _mm256_storeu_pd(c+(i+0)*dimj+8, _c_0_8); in mtxmq() 135 _mm256_storeu_pd(c+(i+0)*dimj+0, _c_0_0); in mtxmq() [all …]
|
H A D | avx2_rc.cc | 62 _mm256_storeu_pd(c+(i+0)*effj*2+0, _c_0_0); in mtxmq() 63 _mm256_storeu_pd(c+(i+0)*effj*2+4, _c_0_4); in mtxmq() 64 _mm256_storeu_pd(c+(i+0)*effj*2+8, _c_0_8); in mtxmq() 66 _mm256_storeu_pd(c+(i+1)*effj*2+0, _c_1_0); in mtxmq() 67 _mm256_storeu_pd(c+(i+1)*effj*2+4, _c_1_4); in mtxmq() 68 _mm256_storeu_pd(c+(i+1)*effj*2+8, _c_1_8); in mtxmq() 70 _mm256_storeu_pd(c+(i+2)*effj*2+0, _c_2_0); in mtxmq() 71 _mm256_storeu_pd(c+(i+2)*effj*2+4, _c_2_4); in mtxmq() 72 _mm256_storeu_pd(c+(i+2)*effj*2+8, _c_2_8); in mtxmq() 93 _mm256_storeu_pd(c+(i+0)*effj*2+0, _c_0_0); in mtxmq() [all …]
|
H A D | avx2_cr.cc | 64 _mm256_storeu_pd(c+(i+0)*effj*2+0, _c_0_0); in mtxmq() 65 _mm256_storeu_pd(c+(i+0)*effj*2+4, _c_0_4); in mtxmq() 66 _mm256_storeu_pd(c+(i+0)*effj*2+8, _c_0_8); in mtxmq() 68 _mm256_storeu_pd(c+(i+1)*effj*2+0, _c_1_0); in mtxmq() 69 _mm256_storeu_pd(c+(i+1)*effj*2+4, _c_1_4); in mtxmq() 70 _mm256_storeu_pd(c+(i+1)*effj*2+8, _c_1_8); in mtxmq() 72 _mm256_storeu_pd(c+(i+2)*effj*2+0, _c_2_0); in mtxmq() 73 _mm256_storeu_pd(c+(i+2)*effj*2+4, _c_2_4); in mtxmq() 74 _mm256_storeu_pd(c+(i+2)*effj*2+8, _c_2_8); in mtxmq() 95 _mm256_storeu_pd(c+(i+0)*effj*2+0, _c_0_0); in mtxmq() [all …]
|
H A D | avx_cr.cc | 65 _mm256_storeu_pd(xc+(i+0)*dimj*2+0, _c_0_0); in mtxmq() 66 _mm256_storeu_pd(xc+(i+0)*dimj*2+4, _c_0_4); in mtxmq() 67 _mm256_storeu_pd(xc+(i+0)*dimj*2+8, _c_0_8); in mtxmq() 68 _mm256_storeu_pd(xc+(i+1)*dimj*2+0, _c_1_0); in mtxmq() 69 _mm256_storeu_pd(xc+(i+1)*dimj*2+4, _c_1_4); in mtxmq() 70 _mm256_storeu_pd(xc+(i+1)*dimj*2+8, _c_1_8); in mtxmq() 71 _mm256_storeu_pd(xc+(i+2)*dimj*2+0, _c_2_0); in mtxmq() 72 _mm256_storeu_pd(xc+(i+2)*dimj*2+4, _c_2_4); in mtxmq() 73 _mm256_storeu_pd(xc+(i+2)*dimj*2+8, _c_2_8); in mtxmq() 74 _mm256_storeu_pd(xc+(i+3)*dimj*2+0, _c_3_0); in mtxmq() [all …]
|
H A D | avx2_cc.cc | 79 _mm256_storeu_pd(xc+(i+0)*effj*2+0, _c_0_0); in mtxmq() 80 _mm256_storeu_pd(xc+(i+0)*effj*2+4, _c_0_4); in mtxmq() 81 _mm256_storeu_pd(xc+(i+0)*effj*2+8, _c_0_8); in mtxmq() 84 _mm256_storeu_pd(xc+(i+1)*effj*2+0, _c_1_0); in mtxmq() 85 _mm256_storeu_pd(xc+(i+1)*effj*2+4, _c_1_4); in mtxmq() 86 _mm256_storeu_pd(xc+(i+1)*effj*2+8, _c_1_8); in mtxmq() 139 _mm256_storeu_pd(xc+(i+0)*effj*2+0, _c_0_0); in mtxmq() 140 _mm256_storeu_pd(xc+(i+0)*effj*2+4, _c_0_4); in mtxmq() 141 _mm256_storeu_pd(xc+(i+0)*effj*2+8, _c_0_8); in mtxmq() 144 _mm256_storeu_pd(xc+(i+1)*effj*2+0, _c_1_0); in mtxmq() [all …]
|
H A D | avx_rc.cc | 57 _mm256_storeu_pd(xc+(i+0)*effj*2+0, _c_0_0); in mtxmq() 58 _mm256_storeu_pd(xc+(i+0)*effj*2+4, _c_0_4); in mtxmq() 59 _mm256_storeu_pd(xc+(i+0)*effj*2+8, _c_0_8); in mtxmq() 60 _mm256_storeu_pd(xc+(i+1)*effj*2+0, _c_1_0); in mtxmq() 61 _mm256_storeu_pd(xc+(i+1)*effj*2+4, _c_1_4); in mtxmq() 62 _mm256_storeu_pd(xc+(i+1)*effj*2+8, _c_1_8); in mtxmq() 63 _mm256_storeu_pd(xc+(i+2)*effj*2+0, _c_2_0); in mtxmq() 64 _mm256_storeu_pd(xc+(i+2)*effj*2+4, _c_2_4); in mtxmq() 65 _mm256_storeu_pd(xc+(i+2)*effj*2+8, _c_2_8); in mtxmq() 96 _mm256_storeu_pd(xc+(i+0)*effj*2+0, _c_0_0); in mtxmq() [all …]
|
H A D | avx_cc.cc | 70 _mm256_storeu_pd(xc+(i+0)*dimj*2+0, _c_0_0); in mtxmq() 71 _mm256_storeu_pd(xc+(i+0)*dimj*2+4, _c_0_4); in mtxmq() 72 _mm256_storeu_pd(xc+(i+0)*dimj*2+8, _c_0_8); in mtxmq() 74 _mm256_storeu_pd(xc+(i+1)*dimj*2+0, _c_1_0); in mtxmq() 75 _mm256_storeu_pd(xc+(i+1)*dimj*2+4, _c_1_4); in mtxmq() 76 _mm256_storeu_pd(xc+(i+1)*dimj*2+8, _c_1_8); in mtxmq() 120 _mm256_storeu_pd(xc+(i+0)*dimj*2+0, _c_0_0); in mtxmq() 121 _mm256_storeu_pd(xc+(i+0)*dimj*2+4, _c_0_4); in mtxmq() 122 _mm256_storeu_pd(xc+(i+0)*dimj*2+8, _c_0_8); in mtxmq() 124 _mm256_storeu_pd(xc+(i+1)*dimj*2+0, _c_1_0); in mtxmq() [all …]
|
/dports/math/blasfeo/blasfeo-0.1.2/kernel/avx/ |
H A D | kernel_dgetrf_pivot_lib.c | 217 _mm256_storeu_pd( &pA[0+lda*0], a_0 ); in kernel_dgetrf_pivot_8_lib() 223 _mm256_storeu_pd( &pA[0+lda*1], c_0 ); in kernel_dgetrf_pivot_8_lib() 228 _mm256_storeu_pd( &pA[0+lda*2], a_i ); in kernel_dgetrf_pivot_8_lib() 233 _mm256_storeu_pd( &pA[0+lda*3], a_i ); in kernel_dgetrf_pivot_8_lib() 238 _mm256_storeu_pd( &pA[0+lda*4], a_i ); in kernel_dgetrf_pivot_8_lib() 243 _mm256_storeu_pd( &pA[0+lda*5], a_i ); in kernel_dgetrf_pivot_8_lib() 248 _mm256_storeu_pd( &pA[0+lda*6], a_i ); in kernel_dgetrf_pivot_8_lib() 253 _mm256_storeu_pd( &pA[0+lda*7], a_i ); in kernel_dgetrf_pivot_8_lib() 421 _mm256_storeu_pd( &pA[0+lda*1], a_0 ); in kernel_dgetrf_pivot_8_lib() 427 _mm256_storeu_pd( &pA[0+lda*2], c_0 ); in kernel_dgetrf_pivot_8_lib() [all …]
|
/dports/science/madness/madness-ebb3fd7/src/madness/tensor/mtxmq_graveyard/ |
H A D | mtxmq_avx_rjh.cc | 346 _mm256_storeu_pd(tmp, ci0j0); in mTxmq_core() 352 _mm256_storeu_pd(tmp, ci0j1); in mTxmq_core() 358 _mm256_storeu_pd(tmp, ci0j2); in mTxmq_core() 364 _mm256_storeu_pd(tmp, ci0j3); in mTxmq_core() 370 _mm256_storeu_pd(tmp, ci0j4); in mTxmq_core() 376 _mm256_storeu_pd(tmp, ci0j5); in mTxmq_core() 537 _mm256_storeu_pd(tmp, ci0j0); in mTxmq_core() 543 _mm256_storeu_pd(tmp, ci0j1); in mTxmq_core() 549 _mm256_storeu_pd(tmp, ci0j2); in mTxmq_core() 555 _mm256_storeu_pd(tmp, ci0j3); in mTxmq_core() [all …]
|
H A D | mtxmq.cc | 349 _mm256_storeu_pd(tmp, ci0j0); in mTxmq_core() 355 _mm256_storeu_pd(tmp, ci0j1); in mTxmq_core() 361 _mm256_storeu_pd(tmp, ci0j2); in mTxmq_core() 367 _mm256_storeu_pd(tmp, ci0j3); in mTxmq_core() 373 _mm256_storeu_pd(tmp, ci0j4); in mTxmq_core() 379 _mm256_storeu_pd(tmp, ci0j5); in mTxmq_core() 540 _mm256_storeu_pd(tmp, ci0j0); in mTxmq_core() 546 _mm256_storeu_pd(tmp, ci0j1); in mTxmq_core() 552 _mm256_storeu_pd(tmp, ci0j2); in mTxmq_core() 558 _mm256_storeu_pd(tmp, ci0j3); in mTxmq_core() [all …]
|
/dports/math/openblas/OpenBLAS-0.3.18/kernel/x86_64/ |
H A D | drot_microk_haswell-2.c | 44 _mm256_storeu_pd(&x[i + 0], t0); in drot_kernel() 45 _mm256_storeu_pd(&x[i + 4], t1); in drot_kernel() 46 _mm256_storeu_pd(&x[i + 8], t2); in drot_kernel() 47 _mm256_storeu_pd(&x[i +12], t3); in drot_kernel() 59 _mm256_storeu_pd(&y[i + 0], t0); in drot_kernel() 60 _mm256_storeu_pd(&y[i + 4], t1); in drot_kernel() 61 _mm256_storeu_pd(&y[i + 8], t2); in drot_kernel() 62 _mm256_storeu_pd(&y[i +12], t3); in drot_kernel() 72 _mm256_storeu_pd(&x[i], t0); in drot_kernel() 76 _mm256_storeu_pd(&y[i], t0); in drot_kernel()
|
H A D | dgemm_tcopy_8_skylakex.c | 136 _mm256_storeu_pd(boffset2 + 0, row1); in CNAME() 137 _mm256_storeu_pd(boffset2 + 4, row2); in CNAME() 138 _mm256_storeu_pd(boffset2 + 8, row3); in CNAME() 139 _mm256_storeu_pd(boffset2 + 12, row4); in CNAME() 140 _mm256_storeu_pd(boffset2 + 16, row5); in CNAME() 141 _mm256_storeu_pd(boffset2 + 20, row6); in CNAME() 142 _mm256_storeu_pd(boffset2 + 24, row7); in CNAME() 143 _mm256_storeu_pd(boffset2 + 28, row8); in CNAME() 262 _mm256_storeu_pd(boffset2 + 0, row1); in CNAME() 263 _mm256_storeu_pd(boffset2 + 4, row2); in CNAME() [all …]
|
/dports/audio/gogglesmm/gogglesmm-1.2.2/cfox/lib/ |
H A D | FXMat4d.cpp | 58 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in FXMat4d() 59 _mm256_storeu_pd(&m[1][0],_mm256_set1_pd(s)); in FXMat4d() 60 _mm256_storeu_pd(&m[2][0],_mm256_set1_pd(s)); in FXMat4d() 61 _mm256_storeu_pd(&m[3][0],_mm256_set1_pd(s)); in FXMat4d() 223 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in operator =() 224 _mm256_storeu_pd(&m[1][0],_mm256_set1_pd(s)); in operator =() 225 _mm256_storeu_pd(&m[2][0],_mm256_set1_pd(s)); in operator =() 226 _mm256_storeu_pd(&m[3][0],_mm256_set1_pd(s)); in operator =() 333 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in set() 334 _mm256_storeu_pd(&m[1][0],_mm256_set1_pd(s)); in set() [all …]
|
H A D | FXMat2d.cpp | 50 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in FXMat2d() 64 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(&s[0][0])); in FXMat2d() 90 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(s)); in FXMat2d() 130 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in operator =() 145 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(&s[0][0])); in operator =() 173 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(s)); in operator =() 188 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in set() 203 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(&s[0][0])); in set() 231 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(s)); in set() 370 _mm256_storeu_pd(&m[0][0],_mm256_set_pd(1.0,0.0,0.0,1.0)); in identity() [all …]
|
/dports/x11-toolkits/fox17/fox-1.7.77/lib/ |
H A D | FXMat4d.cpp | 58 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in FXMat4d() 59 _mm256_storeu_pd(&m[1][0],_mm256_set1_pd(s)); in FXMat4d() 60 _mm256_storeu_pd(&m[2][0],_mm256_set1_pd(s)); in FXMat4d() 61 _mm256_storeu_pd(&m[3][0],_mm256_set1_pd(s)); in FXMat4d() 223 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in operator =() 224 _mm256_storeu_pd(&m[1][0],_mm256_set1_pd(s)); in operator =() 225 _mm256_storeu_pd(&m[2][0],_mm256_set1_pd(s)); in operator =() 226 _mm256_storeu_pd(&m[3][0],_mm256_set1_pd(s)); in operator =() 333 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in set() 334 _mm256_storeu_pd(&m[1][0],_mm256_set1_pd(s)); in set() [all …]
|
H A D | FXMat2d.cpp | 50 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in FXMat2d() 64 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(&s[0][0])); in FXMat2d() 90 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(s)); in FXMat2d() 130 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in operator =() 145 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(&s[0][0])); in operator =() 173 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(s)); in operator =() 188 _mm256_storeu_pd(&m[0][0],_mm256_set1_pd(s)); in set() 203 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(&s[0][0])); in set() 231 _mm256_storeu_pd(&m[0][0],_mm256_loadu_pd(s)); in set() 370 _mm256_storeu_pd(&m[0][0],_mm256_set_pd(1.0,0.0,0.0,1.0)); in identity() [all …]
|
/dports/math/hpipm/hpipm-0.1.1/ipm_core/ |
H A D | d_core_qp_ipm_aux_avx.c | 80 _mm256_storeu_pd( &Gamma[ii+0], y_tmp0 ); in d_compute_Gamma_gamma_qp() 98 _mm256_storeu_pd( &Gamma[ii], y_tmp0 ); in d_compute_Gamma_gamma_qp() 102 _mm256_storeu_pd( &gamma[ii], y_tmp0 ); in d_compute_Gamma_gamma_qp() 155 _mm256_storeu_pd( &gamma[ii], y_tmp0 ); in d_compute_gamma_qp() 195 _mm256_storeu_pd( &dt[ii+0], y_dt0 ); in d_compute_lam_t_qp() 196 _mm256_storeu_pd( &dt[ii+4], y_dt1 ); in d_compute_lam_t_qp() 215 _mm256_storeu_pd( &dt[ii+0], y_dt0 ); in d_compute_lam_t_qp() 408 _mm256_storeu_pd( &v[ii], y_tmp0 ); in d_update_var_qp() 424 _mm256_storeu_pd( &pi[ii], y_tmp0 ); in d_update_var_qp() 453 _mm256_storeu_pd( &lam[ii], y_tmp0 ); in d_update_var_qp() [all …]
|
/dports/audio/spectacle-lv2/spectacle-2.0-1-g27b1525/thirdparty/hiir-1.33/hiir/ |
H A D | PhaseHalfPi4F64Avx.hpp | 162 _mm256_storeu_pd (out_0_ptr, out_0); in process_block() 163 _mm256_storeu_pd (out_1_ptr, out_1); in process_block() 178 _mm256_storeu_pd (out_0_ptr + ofs_0, out_0); in process_block() 179 _mm256_storeu_pd (out_1_ptr + ofs_0, out_1); in process_block() 188 _mm256_storeu_pd (out_0_ptr + ofs_1, out_0); in process_block() 189 _mm256_storeu_pd (out_1_ptr + ofs_1, out_1); in process_block() 203 _mm256_storeu_pd (out_0_ptr + ofs, out_0); in process_block() 204 _mm256_storeu_pd (out_1_ptr + ofs, out_1); in process_block()
|