/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/ |
H A D | jit_avx2_conv_kernel_f32.cpp | 89 size_t inp_off = get_input_offset( in oh_step_unroll_kw() 150 size_t inp_off = get_input_offset( in oh_step_nopad() 399 get_input_offset(0, filter_h_to_input(1)) in width_blk_step() 1455 sub(aux_reg_input, get_input_offset(0, jcp.iw * jcp.ih)); in od_step_comeback_pointers() 1468 sub(reg_input, get_input_offset(0, jcp.iw)); in oh_step_comeback_pointers() 1637 get_input_offset(0, jcp.iw) - get_input_offset(ic_block, 0)); in compute_oh_step_unroll_ow() 1674 get_input_offset(0, jcp.iw) - get_input_offset(ic_tail, 0)); in compute_oh_step_unroll_ow() 1686 add(aux_reg_input, get_input_offset(0, jcp.ih * jcp.iw)); in compute_oh_step_unroll_ow() 1779 get_input_offset(0, jcp.iw) - get_input_offset(ic_block, 0)); in compute_oh_step_common() 1787 add(aux_reg_input, get_input_offset(0, jcp.ih * jcp.iw)); in compute_oh_step_common() [all …]
|
H A D | jit_sse41_conv_kernel_f32.cpp | 80 size_t inp_off = get_input_offset( in oh_step_unroll_kw() 115 size_t inp_off = get_input_offset( in oh_step_nopad() 133 add(aux_reg_input, get_input_offset(0, filter_w_to_input(1))); in oh_step_nopad() 259 sub(aux_reg_input, get_input_offset(0, filter_w_to_input(kw))); in width_blk_step() 260 add(aux_reg_input, get_input_offset(0, filter_h_to_input(1))); in width_blk_step() 264 add(aux_reg_input, get_input_offset(0, filter_h_to_input(1))); in width_blk_step() 324 add(reg_input, get_input_offset(0, filter_w_to_input(0, ur_w, l_pad))); in solve_common() 335 add(reg_input, get_input_offset(0, filter_w_to_input(0, ur_w))); in solve_common() 345 add(reg_input, get_input_offset(0, filter_w_to_input(0, ur_w))); in solve_common()
|
H A D | jit_sse41_conv_kernel_f32.hpp | 82 inline dim_t get_input_offset(int i_ic, int i_iw) { in get_input_offset() function
|
H A D | jit_avx2_conv_kernel_f32.hpp | 97 inline dim_t get_input_offset(int i_ic, int i_iw) { in get_input_offset() function 290 inline dim_t get_input_offset(int i_ic, int i_iw) { in get_input_offset() function
|
H A D | jit_avx512_common_conv_kernel.hpp | 141 inline size_t get_input_offset(int ki, int ic, int oi, int pad_l) { in get_input_offset() function
|
H A D | jit_uni_dw_conv_kernel_f32.cpp | 149 auto get_input_offset = [=](int ii, int ci, int rep) { in apply_filter_unrolled() local 197 const int inp_off = get_input_offset(ii, ch, i); in apply_filter_unrolled() 222 const int inp_off = get_input_offset(ii, ch, i); in apply_filter_unrolled()
|
H A D | jit_avx512_common_conv_kernel.cpp | 826 = get_input_offset(ki, ic, j, pad_l); in compute_loop_fma() 975 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core() 996 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core()
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/ |
H A D | jit_avx2_conv_kernel_f32.cpp | 89 size_t inp_off = get_input_offset( in oh_step_unroll_kw() 150 size_t inp_off = get_input_offset( in oh_step_nopad() 388 get_input_offset(0, filter_h_to_input(1)) in width_blk_step() 1445 sub(aux_reg_input, get_input_offset(0, jcp.iw * jcp.ih)); in od_step_comeback_pointers() 1458 sub(reg_input, get_input_offset(0, jcp.iw)); in oh_step_comeback_pointers() 1627 get_input_offset(0, jcp.iw) - get_input_offset(ic_block, 0)); in compute_oh_step_unroll_ow() 1664 get_input_offset(0, jcp.iw) - get_input_offset(ic_tail, 0)); in compute_oh_step_unroll_ow() 1676 add(aux_reg_input, get_input_offset(0, jcp.ih * jcp.iw)); in compute_oh_step_unroll_ow() 1769 get_input_offset(0, jcp.iw) - get_input_offset(ic_block, 0)); in compute_oh_step_common() 1777 add(aux_reg_input, get_input_offset(0, jcp.ih * jcp.iw)); in compute_oh_step_common() [all …]
|
H A D | jit_sse41_conv_kernel_f32.cpp | 80 size_t inp_off = get_input_offset( in oh_step_unroll_kw() 115 size_t inp_off = get_input_offset( in oh_step_nopad() 133 add(aux_reg_input, get_input_offset(0, filter_w_to_input(1))); in oh_step_nopad() 252 sub(aux_reg_input, get_input_offset(0, filter_w_to_input(kw))); in width_blk_step() 253 add(aux_reg_input, get_input_offset(0, filter_h_to_input(1))); in width_blk_step() 257 add(aux_reg_input, get_input_offset(0, filter_h_to_input(1))); in width_blk_step() 317 add(reg_input, get_input_offset(0, filter_w_to_input(0, ur_w, l_pad))); in solve_common() 328 add(reg_input, get_input_offset(0, filter_w_to_input(0, ur_w))); in solve_common() 338 add(reg_input, get_input_offset(0, filter_w_to_input(0, ur_w))); in solve_common()
|
H A D | jit_sse41_conv_kernel_f32.hpp | 83 inline dim_t get_input_offset(int i_ic, int i_iw) { in get_input_offset() function
|
H A D | jit_avx2_conv_kernel_f32.hpp | 100 inline dim_t get_input_offset(int i_ic, int i_iw) { in get_input_offset() function 293 inline dim_t get_input_offset(int i_ic, int i_iw) { in get_input_offset() function
|
H A D | jit_avx512_common_conv_kernel.hpp | 145 inline size_t get_input_offset(int ki, int ic, int oi, int pad_l) { in get_input_offset() function
|
H A D | jit_uni_dw_conv_kernel_f32.cpp | 149 auto get_input_offset = [=](int ii, int ci, int rep) { in apply_filter_unrolled() local 197 const int inp_off = get_input_offset(ii, ch, i); in apply_filter_unrolled() 222 const int inp_off = get_input_offset(ii, ch, i); in apply_filter_unrolled()
|
H A D | jit_avx512_common_conv_kernel.cpp | 813 = get_input_offset(ki, ic, j, pad_l); in compute_loop_fma() 962 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core() 983 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core()
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/aarch64/ |
H A D | jit_sve_512_conv_kernel.hpp | 185 inline size_t get_input_offset(int ki, int ic, int oi, int pad_l) { in get_input_offset() function
|
H A D | jit_sve_512_conv_kernel.cpp | 457 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core() 473 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core()
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/aarch64/ |
H A D | jit_sve_512_conv_kernel.hpp | 185 inline size_t get_input_offset(int ki, int ic, int oi, int pad_l) { in get_input_offset() function
|
H A D | jit_sve_512_conv_kernel.cpp | 457 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core() 473 = get_input_offset(ki, ic, jj, pad_l); in compute_loop_fma_core()
|