/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/ |
H A D | jit_generator.hpp | 113 static const Xbyak::Reg64 abi_param1(Xbyak::Operand::RCX), 117 static const Xbyak::Reg64 abi_param1(Xbyak::Operand::RDI), 313 void uni_vpxor(const Xbyak::Xmm &x1, const Xbyak::Xmm &x2, in uni_vpxor() 324 void uni_vpxor(const Xbyak::Ymm &x1, const Xbyak::Ymm &x2, in uni_vpxor() 333 void uni_vpxor(const Xbyak::Zmm &x1, const Xbyak::Zmm &x2, in uni_vpxor() 518 const Xbyak::Operand &op, Xbyak::uint8 imm) { in uni_vshufps() 571 const Xbyak::Operand &op2, const Xbyak::Xmm &buf) { in uni_vdivps() 578 const Xbyak::Operand &op2, const Xbyak::Ymm &buf) { in uni_vdivps() 619 void uni_vpsubd(const Xbyak::Xmm &x1, const Xbyak::Xmm &x2, in uni_vpsubd() 624 void uni_vpsubd(const Xbyak::Ymm &x1, const Xbyak::Ymm &x2, in uni_vpsubd() [all …]
|
H A D | jit_avx512_core_amx_1x1_conv_kernel.hpp | 72 const Xbyak::Reg64 &inp_ptr = r15; 73 const Xbyak::Reg64 &wei_ptr = r14; 74 const Xbyak::Reg64 &out_ptr = r13; 75 const Xbyak::Reg64 &wsp_ptr = r12; 77 const Xbyak::Reg64 ®_bias = r11; 97 const Xbyak::Zmm &zmm_bias = zmm31; 99 const Xbyak::Zmm &zmm_zero = zmm30; 102 const Xbyak::Zmm &zmm_zp = zmm29; 134 return Xbyak::Zmm(idx); in zmm_out() 136 Xbyak::Zmm zmm_mask( [all …]
|
H A D | jit_avx512_core_bf16_dw_conv_kernel.hpp | 72 Xbyak::Zmm zmm_ker_reg = Xbyak::Zmm(0); 73 Xbyak::Zmm zmm_src_reg = Xbyak::Zmm(1); 74 Xbyak::Zmm zmm_prev_dst = Xbyak::Zmm(31); 77 Xbyak::Zmm bf16_emu_reserv_1 = Xbyak::Zmm(26); 78 Xbyak::Zmm bf16_emu_reserv_2 = Xbyak::Zmm(27); 79 Xbyak::Zmm bf16_emu_reserv_3 = Xbyak::Zmm(28); 81 Xbyak::Zmm bf16_emu_reserv_5 = Xbyak::Zmm(29); 149 Xbyak::Zmm zmm_ker_reg = Xbyak::Zmm(0); 150 Xbyak::Zmm zmm_dst_reg = Xbyak::Zmm(1); 232 Xbyak::Zmm zmm_bias_reg = Xbyak::Zmm(0); [all …]
|
H A D | jit_avx512_core_x8s8s32x_conv_kernel.hpp | 97 const Xbyak::Opmask ktail_mask = Xbyak::Opmask(2); 98 const Xbyak::Opmask kblend_mask = Xbyak::Opmask(3); 99 const Xbyak::Opmask postops_mask = Xbyak::Opmask(4); 125 const Xbyak::Zmm zmm_wei = Xbyak::Zmm(31); 126 Xbyak::Zmm zmm_tmp; 127 Xbyak::Zmm zmm_src; 128 Xbyak::Zmm zmm_shifted_zero; 129 Xbyak::Zmm zmm_permute; 147 return Xbyak::Zmm(idx); in zmm_out() 161 return Xbyak::Zmm(idx); in zmm_inp() [all …]
|
H A D | jit_avx512_core_x8s8s32x_1x1_conv_kernel.hpp | 48 const Xbyak::Reg64 reg_last_load = r8; 49 const Xbyak::Reg64 reg_bcast_data = r8; 50 const Xbyak::Reg64 reg_ptr_scales = r8; 53 const Xbyak::Reg64 reg_load_data = r10; 58 const Xbyak::Reg64 reg_scratch = r13; 61 const Xbyak::Reg64 imm_addr64 = r15; 75 const Xbyak::Opmask k_load_dim_mask = Xbyak::Opmask(2); 76 const Xbyak::Opmask k_load_dim_tail_mask = Xbyak::Opmask(3); 77 const Xbyak::Opmask postops_mask = Xbyak::Opmask(4); 78 const Xbyak::Opmask vmask = k7; [all …]
|
/dports/emulators/yuzu/yuzu-0b47f7a46/src/common/x64/ |
H A D | xbyak_abi.h | 54 constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax; 55 constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rcx; 56 constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rdx; 57 constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::r8; 58 constexpr inline Xbyak::Reg ABI_PARAM4 = Xbyak::util::r9; 64 Xbyak::util::r8, 105 constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax; 106 constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rdi; 107 constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rsi; 108 constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::rdx; [all …]
|
/dports/emulators/citra/citra-ac98458e0/src/common/x64/ |
H A D | xbyak_abi.h | 54 constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax; 55 constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rcx; 56 constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rdx; 57 constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::r8; 58 constexpr inline Xbyak::Reg ABI_PARAM4 = Xbyak::util::r9; 64 Xbyak::util::r8, 105 constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax; 106 constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rdi; 107 constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rsi; 108 constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::rdx; [all …]
|
/dports/emulators/citra-qt5/citra-ac98458e0/src/common/x64/ |
H A D | xbyak_abi.h | 54 constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax; 55 constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rcx; 56 constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rdx; 57 constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::r8; 58 constexpr inline Xbyak::Reg ABI_PARAM4 = Xbyak::util::r9; 64 Xbyak::util::r8, 105 constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax; 106 constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rdi; 107 constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rsi; 108 constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::rdx; [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/ |
H A D | jit_generator.hpp | 113 static const Xbyak::Reg64 abi_param1(Xbyak::Operand::RCX), 117 static const Xbyak::Reg64 abi_param1(Xbyak::Operand::RDI), 313 void uni_vpxor(const Xbyak::Xmm &x1, const Xbyak::Xmm &x2, in uni_vpxor() 324 void uni_vpxor(const Xbyak::Ymm &x1, const Xbyak::Ymm &x2, in uni_vpxor() 333 void uni_vpxor(const Xbyak::Zmm &x1, const Xbyak::Zmm &x2, in uni_vpxor() 528 const Xbyak::Operand &op, Xbyak::uint8 imm) { in uni_vshufps() 538 const Xbyak::Xmm &x1, const Xbyak::Operand &op, Xbyak::uint8 imm) { in uni_vpshufd() 594 const Xbyak::Operand &op2, const Xbyak::Xmm &buf) { in uni_vdivps() 601 const Xbyak::Operand &op2, const Xbyak::Ymm &buf) { in uni_vdivps() 632 void uni_vphaddd(const Xbyak::Xmm &x, const Xbyak::Xmm &x2, in uni_vphaddd() [all …]
|
H A D | jit_avx512_core_amx_1x1_conv_kernel.hpp | 72 const Xbyak::Reg64 &inp_ptr = r15; 73 const Xbyak::Reg64 &wei_ptr = r14; 74 const Xbyak::Reg64 &out_ptr = r13; 75 const Xbyak::Reg64 &wsp_ptr = r12; 77 const Xbyak::Reg64 ®_bias = r11; 98 const Xbyak::Zmm &zmm_bias = zmm31; 100 const Xbyak::Zmm &zmm_zero = zmm30; 104 const Xbyak::Zmm &zmm_zp = zmm29; 136 return Xbyak::Zmm(idx); in zmm_out() 138 Xbyak::Zmm zmm_mask( [all …]
|
H A D | jit_avx512_core_bf16_dw_conv_kernel.hpp | 74 Xbyak::Zmm zmm_ker_reg = Xbyak::Zmm(0); 75 Xbyak::Zmm zmm_src_reg = Xbyak::Zmm(1); 76 Xbyak::Zmm zmm_prev_dst = Xbyak::Zmm(31); 79 Xbyak::Zmm bf16_emu_reserv_1 = Xbyak::Zmm(26); 80 Xbyak::Zmm bf16_emu_reserv_2 = Xbyak::Zmm(27); 81 Xbyak::Zmm bf16_emu_reserv_3 = Xbyak::Zmm(28); 83 Xbyak::Zmm bf16_emu_reserv_5 = Xbyak::Zmm(29); 152 Xbyak::Zmm zmm_ker_reg = Xbyak::Zmm(0); 153 Xbyak::Zmm zmm_dst_reg = Xbyak::Zmm(1); 235 Xbyak::Zmm zmm_bias_reg = Xbyak::Zmm(0); [all …]
|
H A D | jit_avx512_core_x8s8s32x_conv_kernel.hpp | 101 const Xbyak::Opmask ktail_mask = Xbyak::Opmask(2); 102 const Xbyak::Opmask kblend_mask = Xbyak::Opmask(3); 103 const Xbyak::Opmask postops_mask = Xbyak::Opmask(4); 130 const Xbyak::Zmm zmm_wei = Xbyak::Zmm(31); 131 Xbyak::Zmm zmm_tmp; 132 Xbyak::Zmm zmm_src; 133 Xbyak::Zmm zmm_shifted_zero; 134 Xbyak::Zmm zmm_permute; 152 return Xbyak::Zmm(idx); in zmm_out() 166 return Xbyak::Zmm(idx); in zmm_inp() [all …]
|
H A D | jit_avx512_core_x8s8s32x_1x1_conv_kernel.hpp | 48 const Xbyak::Reg64 reg_last_load = r8; 49 const Xbyak::Reg64 reg_bcast_data = r8; 50 const Xbyak::Reg64 reg_ptr_scales = r8; 53 const Xbyak::Reg64 reg_load_data = r10; 58 const Xbyak::Reg64 reg_scratch = r13; 61 const Xbyak::Reg64 imm_addr64 = r15; 75 const Xbyak::Opmask k_load_dim_mask = Xbyak::Opmask(2); 76 const Xbyak::Opmask k_load_dim_tail_mask = Xbyak::Opmask(3); 77 const Xbyak::Opmask postops_mask = Xbyak::Opmask(4); 78 const Xbyak::Opmask vmask = k7; [all …]
|
/dports/emulators/yuzu/yuzu-0b47f7a46/externals/dynarmic/tests/ |
H A D | cpu_info.cpp | 14 Xbyak::util::Cpu cpu_info; 16 std::make_pair(Xbyak::util::Cpu::tMMX, "MMX"), 17 std::make_pair(Xbyak::util::Cpu::tMMX2, "MMX2"), 18 std::make_pair(Xbyak::util::Cpu::tCMOV, "CMOV"), 19 std::make_pair(Xbyak::util::Cpu::tSSE, "SSE"), 20 std::make_pair(Xbyak::util::Cpu::tSSE2, "SSE2"), 21 std::make_pair(Xbyak::util::Cpu::tSSE3, "SSE3"), 30 std::make_pair(Xbyak::util::Cpu::tAVX, "AVX"), 31 std::make_pair(Xbyak::util::Cpu::tFMA, "FMA"), 32 std::make_pair(Xbyak::util::Cpu::t3DN, "3DN"), [all …]
|
/dports/emulators/citra-qt5/citra-ac98458e0/externals/dynarmic/tests/ |
H A D | cpu_info.cpp | 9 Xbyak::util::Cpu cpu_info; 11 std::make_pair(Xbyak::util::Cpu::tMMX, "MMX"), 12 std::make_pair(Xbyak::util::Cpu::tMMX2, "MMX2"), 13 std::make_pair(Xbyak::util::Cpu::tCMOV, "CMOV"), 14 std::make_pair(Xbyak::util::Cpu::tSSE, "SSE"), 15 std::make_pair(Xbyak::util::Cpu::tSSE2, "SSE2"), 16 std::make_pair(Xbyak::util::Cpu::tSSE3, "SSE3"), 25 std::make_pair(Xbyak::util::Cpu::tAVX, "AVX"), 26 std::make_pair(Xbyak::util::Cpu::tFMA, "FMA"), 27 std::make_pair(Xbyak::util::Cpu::t3DN, "3DN"), [all …]
|
/dports/emulators/citra/citra-ac98458e0/externals/dynarmic/tests/ |
H A D | cpu_info.cpp | 9 Xbyak::util::Cpu cpu_info; 11 std::make_pair(Xbyak::util::Cpu::tMMX, "MMX"), 12 std::make_pair(Xbyak::util::Cpu::tMMX2, "MMX2"), 13 std::make_pair(Xbyak::util::Cpu::tCMOV, "CMOV"), 14 std::make_pair(Xbyak::util::Cpu::tSSE, "SSE"), 15 std::make_pair(Xbyak::util::Cpu::tSSE2, "SSE2"), 16 std::make_pair(Xbyak::util::Cpu::tSSE3, "SSE3"), 25 std::make_pair(Xbyak::util::Cpu::tAVX, "AVX"), 26 std::make_pair(Xbyak::util::Cpu::tFMA, "FMA"), 27 std::make_pair(Xbyak::util::Cpu::t3DN, "3DN"), [all …]
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/gemm/f32/ |
H A D | jit_avx2_kernel_sgemm_kern.cpp | 234 vpxorq(Xbyak::Zmm(i), Xbyak::Zmm(i), Xbyak::Zmm(i)); in generate() 301 loop<Xbyak::Zmm, Xbyak::Zmm, Xbyak::Address, Xbyak::Xmm, in generate() 310 loop<Xbyak::Ymm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 323 loop<Xbyak::Zmm, Xbyak::Ymm, Xbyak::Address, Xbyak::Xmm, in generate() 329 loop<Xbyak::Xmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 344 loop<Xbyak::Zmm, Xbyak::Ymm, Xbyak::Operand, Xbyak::Xmm, in generate() 349 loop<Xbyak::Xmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 361 loop<Xbyak::Zmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 367 loop<Xbyak::Xmm, Xbyak::Xmm, Xbyak::Address, Xbyak::Xmm, in generate() 381 loop<Xbyak::Zmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/gemm/f32/ |
H A D | jit_avx2_kernel_sgemm_kern.cpp | 234 vpxorq(Xbyak::Zmm(i), Xbyak::Zmm(i), Xbyak::Zmm(i)); in generate() 301 loop<Xbyak::Zmm, Xbyak::Zmm, Xbyak::Address, Xbyak::Xmm, in generate() 310 loop<Xbyak::Ymm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 323 loop<Xbyak::Zmm, Xbyak::Ymm, Xbyak::Address, Xbyak::Xmm, in generate() 329 loop<Xbyak::Xmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 344 loop<Xbyak::Zmm, Xbyak::Ymm, Xbyak::Operand, Xbyak::Xmm, in generate() 349 loop<Xbyak::Xmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 361 loop<Xbyak::Zmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() 367 loop<Xbyak::Xmm, Xbyak::Xmm, Xbyak::Address, Xbyak::Xmm, in generate() 381 loop<Xbyak::Zmm, Xbyak::Xmm, Xbyak::Operand, Xbyak::Xmm, in generate() [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/injectors/ |
H A D | jit_uni_binary_injector.hpp | 105 const Xbyak::Reg64 &rhs_addr_reg, 111 const Xbyak::Reg64 &rhs_addr_reg, 118 const Xbyak::Reg64 &rhs_addr_reg, 135 const Xbyak::Opmask &tail_opmask, const Xbyak::Reg64 ®_tail_size, 150 Xbyak::Reg64 rhs_addr_reg; 151 Xbyak::Reg64 rhs_helper_reg; 158 Xbyak::Opmask tail_opmask; 160 Xbyak::Reg64 reg_tail_size; 197 Xbyak::Reg64 param1; 499 Xbyak::Address remove_bcast_bit(const Xbyak::Address &rhs_addr) const; [all …]
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/gemm/amx/ |
H A D | jit_avx512_core_amx_copy_kern.hpp | 42 void load(const Xbyak::Xmm &dst, const Xbyak::Address &src, bool corner); 43 void store(const Xbyak::Address &dst, const Xbyak::Xmm &src); 45 void transpose(int s, const Xbyak::Ymm &dst1, const Xbyak::Ymm &dst2, 46 const Xbyak::Ymm &src1, const Xbyak::Ymm &src2); 48 void amxtrans8(const Xbyak::Ymm &dst1, const Xbyak::Ymm &dst2, 49 const Xbyak::Ymm &src1, const Xbyak::Ymm &src2, 50 const Xbyak::Ymm &src3, const Xbyak::Ymm &src4); 51 void amxtrans16(const Xbyak::Ymm &dst1, const Xbyak::Ymm &dst2, 52 const Xbyak::Ymm &src1, const Xbyak::Ymm &src2); 92 Xbyak::Ymm tmp1_, tmp2_; [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/gemm/amx/ |
H A D | jit_avx512_core_amx_copy_kern.hpp | 42 void load(const Xbyak::Xmm &dst, const Xbyak::Address &src, bool corner); 43 void store(const Xbyak::Address &dst, const Xbyak::Xmm &src); 45 void transpose(int s, const Xbyak::Ymm &dst1, const Xbyak::Ymm &dst2, 46 const Xbyak::Ymm &src1, const Xbyak::Ymm &src2); 48 void amxtrans8(const Xbyak::Ymm &dst1, const Xbyak::Ymm &dst2, 49 const Xbyak::Ymm &src1, const Xbyak::Ymm &src2, 50 const Xbyak::Ymm &src3, const Xbyak::Ymm &src4); 51 void amxtrans16(const Xbyak::Ymm &dst1, const Xbyak::Ymm &dst2, 52 const Xbyak::Ymm &src1, const Xbyak::Ymm &src2); 92 Xbyak::Ymm tmp1_, tmp2_; [all …]
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/gemm/s8x8s32/ |
H A D | jit_avx512_core_kernel_gemv_s8x8s32_kern.hpp | 41 void vnni(Xbyak::Zmm acc, Xbyak::Zmm b, Xbyak::Zmm a, vnni_op_t op); 42 void n_loop_body(int nreg_acc, Xbyak::Reg64 A, Xbyak::Reg64 lda, 43 Xbyak::Reg64 X, int use_mask, Xbyak::Opmask mask_n); 45 Xbyak::Zmm, Xbyak::Zmm, Xbyak::Zmm, Xbyak::Zmm, Xbyak::Zmm); 46 void update_c(int, Xbyak::Reg64, int, Xbyak::Opmask); 63 Xbyak::Zmm zmm_tmp = Xbyak::Zmm(0); 64 Xbyak::Xmm xmm_beta = Xbyak::Xmm(1); 66 Xbyak::Zmm zmm_1_s16 = Xbyak::Zmm(2); // avx512_core 67 Xbyak::Zmm zmm_1_u1 = Xbyak::Zmm(2); // s8s8, avx512_core_vnni 68 Xbyak::Zmm zmm_128_u8 = Xbyak::Zmm(3); // s8s8 [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/gemm/s8x8s32/ |
H A D | jit_avx512_core_kernel_gemv_s8x8s32_kern.hpp | 41 void vnni(Xbyak::Zmm acc, Xbyak::Zmm b, Xbyak::Zmm a, vnni_op_t op); 42 void n_loop_body(int nreg_acc, Xbyak::Reg64 A, Xbyak::Reg64 lda, 43 Xbyak::Reg64 X, int use_mask, Xbyak::Opmask mask_n); 45 Xbyak::Zmm, Xbyak::Zmm, Xbyak::Zmm, Xbyak::Zmm, Xbyak::Zmm); 46 void update_c(int, Xbyak::Reg64, int, Xbyak::Opmask); 63 Xbyak::Zmm zmm_tmp = Xbyak::Zmm(0); 64 Xbyak::Xmm xmm_beta = Xbyak::Xmm(1); 66 Xbyak::Zmm zmm_1_s16 = Xbyak::Zmm(2); // avx512_core 67 Xbyak::Zmm zmm_1_u1 = Xbyak::Zmm(2); // s8s8, avx512_core_vnni 68 Xbyak::Zmm zmm_128_u8 = Xbyak::Zmm(3); // s8s8 [all …]
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/utils/ |
H A D | jit_io_helper.hpp | 60 Xbyak::Opmask tail_opmask_ = Xbyak::Opmask(); 62 Xbyak::Reg64 reg_tmp_ = Xbyak::Reg64(); 76 Xbyak::Zmm bf16_emu_reserv_1_ = Xbyak::Zmm(28); 77 Xbyak::Zmm bf16_emu_reserv_2_ = Xbyak::Zmm(29); 78 Xbyak::Zmm bf16_emu_reserv_3_ = Xbyak::Zmm(30); 79 Xbyak::Reg64 reg_tmp_ = Xbyak::util::rax; 80 Xbyak::Zmm bf16_emu_reserv_4_ = Xbyak::Zmm(31); 94 Xbyak::Reg64 reg_tmp_ = Xbyak::Reg64(); 108 Xbyak::Opmask full_opmask_ = Xbyak::Opmask(); 110 Xbyak::Reg64 reg_tmp_ = Xbyak::Reg64(); [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/utils/ |
H A D | jit_io_helper.hpp | 60 Xbyak::Opmask tail_opmask_ = Xbyak::Opmask(); 62 Xbyak::Reg64 reg_tmp_ = Xbyak::Reg64(); 76 Xbyak::Zmm bf16_emu_reserv_1_ = Xbyak::Zmm(28); 77 Xbyak::Zmm bf16_emu_reserv_2_ = Xbyak::Zmm(29); 78 Xbyak::Zmm bf16_emu_reserv_3_ = Xbyak::Zmm(30); 79 Xbyak::Reg64 reg_tmp_ = Xbyak::util::rax; 80 Xbyak::Zmm bf16_emu_reserv_4_ = Xbyak::Zmm(31); 94 Xbyak::Reg64 reg_tmp_ = Xbyak::Reg64(); 108 Xbyak::Opmask full_opmask_ = Xbyak::Opmask(); 110 Xbyak::Reg64 reg_tmp_ = Xbyak::Reg64(); [all …]
|