/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/aarch64/ |
H A D | jit_sve_512_x8s8s32x_conv_kernel.cpp | 947 auto regw_tmp = reg_oi; in generate() 981 eor(reg_oi, reg_oi, reg_oi); in generate() 998 adds(reg_oi, reg_oi, 1); in generate() 1009 adds(reg_oi, reg_oi, 1); in generate() 1011 cmp(reg_oi, reg_tmp0_imm); in generate() 1060 mov_imm(reg_oi, n_oi_first_ow_block); in generate() 1066 subs(reg_oi, reg_oi, 1); in generate() 1082 mov_imm(reg_oi, n_oi_last_ow_block); in generate() 1089 mov_imm(reg_oi, n_oi_next_last_ow_block); in generate() 1097 cmp(reg_oi, 0); in generate() [all …]
|
H A D | jit_sve_512_conv_kernel.cpp | 654 add_imm(reg_oi, reg_oi, 1, reg_tmp_imm); // increment in generate() 666 add_imm(reg_oi, reg_oi, 1, reg_tmp_imm); //inc(reg_oi); in generate() 729 sub(reg_oi, reg_oi, 1); // decrement in generate() 730 cmp(reg_oi, 0); in generate() 755 cmp(reg_oi, 0); in generate() 762 sub(reg_oi, reg_oi, 1); // dec(reg_oi); in generate() 763 cmp(reg_oi, 0); in generate() 1978 sub(reg_oi, reg_oi, 1); in generate() 3339 subs(reg_oi, reg_oi, 1); in bias_kernel_2d() 3368 subs(reg_oi, reg_oi, reg_tmp_imm); in bias_kernel_3d() [all …]
|
H A D | jit_uni_dw_conv_kernel_f32.cpp | 280 mov(reg_oi, 0); in ow_loop() 296 add(reg_oi, reg_oi, 1); in ow_loop() 306 add(reg_oi, reg_oi, 1); in ow_loop() 307 cmp(reg_oi, n_oi); in ow_loop()
|
H A D | jit_sve_512_conv_kernel.hpp | 96 reg64_t reg_oi = x11; member 272 reg64_t reg_oi = x11; member 467 reg64_t reg_oi = x12; member
|
H A D | jit_sve_512_x8s8s32x_conv_kernel.hpp | 99 const XReg reg_oi = x3; member
|
H A D | jit_uni_dw_conv_kernel_f32.hpp | 66 reg64_t reg_oi = x11; member
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/aarch64/ |
H A D | jit_sve_512_x8s8s32x_conv_kernel.cpp | 947 auto regw_tmp = reg_oi; in generate() 981 eor(reg_oi, reg_oi, reg_oi); in generate() 998 adds(reg_oi, reg_oi, 1); in generate() 1009 adds(reg_oi, reg_oi, 1); in generate() 1011 cmp(reg_oi, reg_tmp0_imm); in generate() 1060 mov_imm(reg_oi, n_oi_first_ow_block); in generate() 1066 subs(reg_oi, reg_oi, 1); in generate() 1082 mov_imm(reg_oi, n_oi_last_ow_block); in generate() 1089 mov_imm(reg_oi, n_oi_next_last_ow_block); in generate() 1097 cmp(reg_oi, 0); in generate() [all …]
|
H A D | jit_sve_512_conv_kernel.cpp | 654 add_imm(reg_oi, reg_oi, 1, reg_tmp_imm); // increment in generate() 666 add_imm(reg_oi, reg_oi, 1, reg_tmp_imm); //inc(reg_oi); in generate() 729 sub(reg_oi, reg_oi, 1); // decrement in generate() 730 cmp(reg_oi, 0); in generate() 755 cmp(reg_oi, 0); in generate() 762 sub(reg_oi, reg_oi, 1); // dec(reg_oi); in generate() 763 cmp(reg_oi, 0); in generate() 1978 sub(reg_oi, reg_oi, 1); in generate() 3339 subs(reg_oi, reg_oi, 1); in bias_kernel_2d() 3368 subs(reg_oi, reg_oi, reg_tmp_imm); in bias_kernel_3d() [all …]
|
H A D | jit_uni_dw_conv_kernel_f32.cpp | 280 mov(reg_oi, 0); in ow_loop() 296 add(reg_oi, reg_oi, 1); in ow_loop() 306 add(reg_oi, reg_oi, 1); in ow_loop() 307 cmp(reg_oi, n_oi); in ow_loop()
|
H A D | jit_sve_512_conv_kernel.hpp | 96 reg64_t reg_oi = x11; member 272 reg64_t reg_oi = x11; member 467 reg64_t reg_oi = x12; member
|
H A D | jit_sve_512_x8s8s32x_conv_kernel.hpp | 99 const XReg reg_oi = x3; member
|
H A D | jit_uni_dw_conv_kernel_f32.hpp | 66 reg64_t reg_oi = x11; member
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/ |
H A D | jit_avx512_common_conv_kernel.hpp | 80 reg64_t reg_oi = rbx; member 266 reg64_t reg_oi = rbx; member 441 reg64_t reg_oi = r12; member
|
H A D | jit_avx512_common_conv_kernel.cpp | 1175 xor_(reg_oi, reg_oi); in generate() 1182 inc(reg_oi); in generate() 1260 dec(reg_oi); in generate() 1287 cmp(reg_oi, 0); in generate() 1295 dec(reg_oi); in generate() 2724 sub(reg_oi, 1); in generate() 4139 mov(reg_oi, jcp.ow); in bias_kernel_2d() 4150 dec(reg_oi); in bias_kernel_2d() 4180 cmp(reg_oi, 0); in bias_kernel_3d() 4186 imul(reg_oi, reg_tmp); in bias_kernel_3d() [all …]
|
H A D | jit_avx512_core_bf16_conv_kernel.cpp | 669 xor_(reg_oi, reg_oi); in generate() 683 inc(reg_oi); in generate() 693 inc(reg_oi); in generate() 743 mov(reg_oi, n_oi_first_ow_block); in generate() 748 dec(reg_oi); in generate() 763 mov(reg_oi, n_oi_last_ow_block); in generate() 773 cmp(reg_oi, 0); in generate() 779 dec(reg_oi); in generate() 1478 mov(reg_oi, pretail_n_oi); in generate() 1486 mov(reg_oi, body_n_oi); in generate() [all …]
|
H A D | jit_avx512_core_x8s8s32x_conv_kernel.cpp | 956 Reg32 regw_tmp = reg_oi.cvt32(); in generate() 963 const Reg32 regw_tmp = reg_oi.cvt32(); in generate() 1080 mov(reg_oi, n_urw_per_ow_block); in generate() 1098 cmove(reg_oi, reg_scratch); in generate() 1155 dec(reg_oi); in generate() 1181 mov(reg_oi, n_oi_middle_block_loop); in generate() 1188 dec(reg_oi); in generate()
|
H A D | jit_avx512_core_bf16_conv_kernel.hpp | 71 reg64_t reg_oi = rdx; member 291 reg64_t reg_oi = rbx; member 530 reg64_t reg_oi = r12; member
|
H A D | jit_avx512_core_bf16_dw_conv_kernel.cpp | 486 xor_(reg_oi, reg_oi); in loop_ow() 502 inc(reg_oi); in loop_ow() 512 inc(reg_oi); in loop_ow() 513 cmp(reg_oi, n_oi); in loop_ow()
|
H A D | jit_uni_x8s8s32x_conv_kernel.hpp | 77 const Xbyak::Reg64 reg_oi = rbx; member
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/ |
H A D | jit_avx512_common_conv_kernel.hpp | 80 reg64_t reg_oi = rbx; member 270 reg64_t reg_oi = rbx; member 445 reg64_t reg_oi = r12; member
|
H A D | jit_avx512_common_conv_kernel.cpp | 1162 xor_(reg_oi, reg_oi); in generate() 1169 inc(reg_oi); in generate() 1247 dec(reg_oi); in generate() 1274 cmp(reg_oi, 0); in generate() 1282 dec(reg_oi); in generate() 2712 sub(reg_oi, 1); in generate() 4127 mov(reg_oi, jcp.ow); in bias_kernel_2d() 4138 dec(reg_oi); in bias_kernel_2d() 4168 cmp(reg_oi, 0); in bias_kernel_3d() 4174 imul(reg_oi, reg_tmp); in bias_kernel_3d() [all …]
|
H A D | jit_avx512_core_bf16_conv_kernel.cpp | 657 xor_(reg_oi, reg_oi); in generate() 671 inc(reg_oi); in generate() 681 inc(reg_oi); in generate() 731 mov(reg_oi, n_oi_first_ow_block); in generate() 736 dec(reg_oi); in generate() 751 mov(reg_oi, n_oi_last_ow_block); in generate() 761 cmp(reg_oi, 0); in generate() 767 dec(reg_oi); in generate() 1469 mov(reg_oi, pretail_n_oi); in generate() 1477 mov(reg_oi, body_n_oi); in generate() [all …]
|
H A D | jit_avx512_core_x8s8s32x_conv_kernel.cpp | 958 Reg32 regw_tmp = reg_oi.cvt32(); in generate() 965 const Reg32 regw_tmp = reg_oi.cvt32(); in generate() 1082 mov(reg_oi, n_urw_per_ow_block); in generate() 1100 cmove(reg_oi, reg_scratch); in generate() 1157 dec(reg_oi); in generate() 1183 mov(reg_oi, n_oi_middle_block_loop); in generate() 1190 dec(reg_oi); in generate()
|
H A D | jit_avx512_core_bf16_conv_kernel.hpp | 71 reg64_t reg_oi = rdx; member 294 reg64_t reg_oi = rbx; member 533 reg64_t reg_oi = r12; member
|
H A D | jit_avx512_core_bf16_dw_conv_kernel.cpp | 481 xor_(reg_oi, reg_oi); in loop_ow() 497 inc(reg_oi); in loop_ow() 507 inc(reg_oi); in loop_ow() 508 cmp(reg_oi, n_oi); in loop_ow()
|