/dports/math/openblas/OpenBLAS-0.3.18/kernel/mips/ |
H A D | dgemm_tcopy_8_msa.c | 65 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME() 66 LD_DP4_INC(psrc2, 2, src4, src5, src6, src7); in CNAME() 67 LD_DP4_INC(psrc3, 2, src8, src9, src10, src11); in CNAME() 74 LD_DP4_INC(psrc5, 2, src0, src1, src2, src3); in CNAME() 75 LD_DP4_INC(psrc6, 2, src4, src5, src6, src7); in CNAME() 76 LD_DP4_INC(psrc7, 2, src8, src9, src10, src11); in CNAME() 150 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME() 151 LD_DP4_INC(psrc2, 2, src4, src5, src6, src7); in CNAME() 205 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME() 206 LD_DP4_INC(psrc2, 2, src4, src5, src6, src7); in CNAME() [all …]
|
H A D | zgemm_ncopy_4_msa.c | 52 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME() 53 LD_DP4_INC(psrc2, 2, src4, src5, src6, src7); in CNAME() 54 LD_DP4_INC(psrc3, 2, src8, src9, src10, src11); in CNAME() 55 LD_DP4_INC(psrc4, 2, src12, src13, src14, src15); in CNAME() 95 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME() 96 LD_DP4_INC(psrc2, 2, src4, src5, src6, src7); in CNAME() 126 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME()
|
H A D | zgemm_tcopy_4_msa.c | 59 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME() 60 LD_DP4_INC(psrc2, 2, src4, src5, src6, src7); in CNAME() 61 LD_DP4_INC(psrc3, 2, src8, src9, src10, src11); in CNAME() 62 LD_DP4_INC(psrc4, 2, src12, src13, src14, src15); in CNAME() 106 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME() 107 LD_DP4_INC(psrc2, 2, src4, src5, src6, src7); in CNAME() 141 LD_DP4_INC(psrc1, 2, src0, src1, src2, src3); in CNAME()
|
H A D | zrot_msa.c | 176 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); \ 177 LD_DP4_INC(py, inc_y2, y0, y1, y2, y3); \ 328 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); \ 329 LD_DP4_INC(py, inc_y2, y0, y1, y2, y3); \ 391 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); \ 392 LD_DP4_INC(py, inc_y2, y0, y1, y2, y3); \ 469 LD_DP4_INC(px, inc_x2, x4, x5, x6, x7); \ 470 LD_DP4_INC(py, inc_y2, y4, y5, y6, y7); \ 486 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); \ 487 LD_DP4_INC(py, inc_y2, y0, y1, y2, y3); \ [all …]
|
H A D | drot_msa.c | 204 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 205 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() 365 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 366 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() 437 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 438 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() 515 LD_DP4_INC(px, 2, x4, x5, x6, x7); in CNAME() 516 LD_DP4_INC(py, 2, y4, y5, y6, y7); in CNAME() 532 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 533 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() [all …]
|
H A D | zaxpy_msa.c | 131 LD_DP4_INC(x, 2, x0, x1, x2, x3); in CNAME() 132 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() 240 LD_DP4_INC(x, inc_x2, x0, x1, x2, x3); in CNAME() 241 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() 349 LD_DP4_INC(x, 2, x0, x1, x2, x3); in CNAME() 350 LD_DP4_INC(py, inc_y2, y0, y1, y2, y3); in CNAME() 442 LD_DP4_INC(x, inc_x2, x0, x1, x2, x3); in CNAME() 443 LD_DP4_INC(py, inc_y2, y0, y1, y2, y3); in CNAME()
|
H A D | zswap_msa.c | 152 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 153 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() 215 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); in CNAME() 216 LD_DP4_INC(py, inc_y2, y0, y1, y2, y3); in CNAME()
|
H A D | zdot_msa.c | 97 LD_DP4_INC(x, 2, vx0, vx1, vx2, vx3); in CNAME() 98 LD_DP4_INC(y, 2, vy0, vy1, vy2, vy3); in CNAME() 210 LD_DP4_INC(x, inc_x2, vx0, vx1, vx2, vx3); in CNAME() 211 LD_DP4_INC(y, inc_y2, vy0, vy1, vy2, vy3); in CNAME() 314 LD_DP4_INC(x, inc_x2, vx0, vx1, vx2, vx3); in CNAME() 315 LD_DP4_INC(y, inc_y2, vy0, vy1, vy2, vy3); in CNAME()
|
H A D | ddot_msa.c | 96 LD_DP4_INC(x, 2, vx0, vx1, vx2, vx3); in CNAME() 97 LD_DP4_INC(y, 2, vy0, vy1, vy2, vy3); in CNAME()
|
H A D | daxpy_msa.c | 101 LD_DP4_INC(x, 2, x0, x1, x2, x3); in CNAME() 102 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME() 180 LD_DP4_INC(py, 2, y0, y1, y2, y3); in CNAME()
|
H A D | dgemm_kernel_8x4_msa.c | 104 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_core_msa() 144 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_core_msa() 171 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_core_msa() 210 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_core_msa() 898 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_non_core_msa() 916 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_non_core_msa() 932 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_non_core_msa() 951 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_non_core_msa() 1367 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_non_core_msa() 1380 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); in dgemmkernel_8x4_non_core_msa() [all …]
|
H A D | zasum_msa.c | 137 LD_DP4_INC(x, 2, src0, src1, src2, src3); in CNAME() 234 LD_DP4_INC(x, inc_x, src0, src1, src2, src3); in CNAME()
|
H A D | zcopy_msa.c | 131 LD_DP4_INC(x, 2, x0, x1, x2, x3); in CNAME() 199 LD_DP4_INC(x, inc_x, x0, x1, x2, x3); in CNAME()
|
H A D | dasum_msa.c | 137 LD_DP4_INC(x, 2, src0, src1, src2, src3); in CNAME() 238 LD_DP4_INC(x, inc_x, src0, src1, src2, src3); in CNAME()
|
H A D | zscal_msa.c | 217 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 347 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 439 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME() 552 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); in CNAME() 609 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); in CNAME() 680 LD_DP4_INC(px, inc_x2, x0, x1, x2, x3); in CNAME()
|
H A D | dswap_msa.c | 151 LD_DP4_INC(px, 2, xv0, xv1, xv2, xv3); in CNAME() 152 LD_DP4_INC(py, 2, yv0, yv1, yv2, yv3); in CNAME()
|
H A D | zgemm_kernel_4x4_msa.c | 33 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); \ 34 LD_DP4_INC(pb0, 2, src_b0, src_b1, src_b2, src_b3); \ 91 LD_DP4_INC(pb0, 2, src_b0, src_b1, src_b2, src_b3); \ 127 LD_DP4_INC(pb0, 2, src_b0, src_b1, src_b2, src_b3); \ 148 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); \ 218 LD_DP4_INC(pa0, 2, src_a0, src_a1, src_a2, src_a3); \
|
H A D | dcopy_msa.c | 131 LD_DP4_INC(x, 2, x0, x1, x2, x3); in CNAME()
|
H A D | macros_msa.h | 258 #define LD_DP4_INC(psrc, stride, out0, \ macro 297 LD_DP4_INC(psrc, stride, out0, out1, out2, out3); \ 298 LD_DP4_INC(psrc, stride, out4, out5, out6, out7); \
|
H A D | dtrsm_kernel_RN_8x4_msa.c | 70 LD_DP4_INC(a, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_rn_msa() 97 LD_DP4_INC(a, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_rn_msa() 129 LD_DP4_INC(a, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_rn_msa()
|
H A D | dtrsm_kernel_RT_8x4_msa.c | 70 LD_DP4_INC(pba, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_rt_msa() 97 LD_DP4_INC(pba, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_rt_msa() 129 LD_DP4_INC(pba, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_rt_msa()
|
H A D | dtrsm_kernel_LT_8x4_msa.c | 90 LD_DP4_INC(a, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_lt_msa() 117 LD_DP4_INC(a, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_lt_msa() 149 LD_DP4_INC(a, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_lt_msa()
|
H A D | dtrsm_kernel_LN_8x4_msa.c | 89 LD_DP4_INC(pba, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_ln_msa() 116 LD_DP4_INC(pba, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_ln_msa() 148 LD_DP4_INC(pba, 2, src_a0, src_a1, src_a2, src_a3); in dsolve_8x4_ln_msa()
|
H A D | dscal_msa.c | 215 LD_DP4_INC(px, 2, x0, x1, x2, x3); in CNAME()
|