package spu.pb

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: RuntimeConfig

• bool disable_matmul_pack = 1

  disable the ciphertext packing for matmul

• bool enable_mul_lsb_error = 2

  allow least significant bits error for point-wise mul

• CheetahOtKind ot_kind = 3

  Setup for cheetah ot

Used in: CheetahConfig

• YACL_Ferret = 0

• YACL_Softspoken = 1

• EMP_Ferret = 2

Used in: TTPBeaverConfig

• string certificate = 1

  Certificate in PEM format, supported both file path and raw string

• string private_key = 2

  Private key in PEM format, supported both file path and raw string based on prefix

• string ca_file_path = 3

  The trusted CA file to verify the peer's certificate If empty, use the system default CA files

• int32 verify_depth = 4

  Maximum depth of the certificate chain for verification If 0, turn off the verification

• SourceIRType ir_type = 1

  Input IR type

• bytes ir_txt = 2

  IR

• repeated Visibility input_visibility = 3

  Input visibilities

• bool enable_pretty_print = 1

  Pretty print

• string pretty_print_dump_dir = 2

• XLAPrettyPrintKind xla_pp_kind = 3

• bool disable_sqrt_plus_epsilon_rewrite = 10

  Disable sqrt(x) + eps to sqrt(x+eps) rewrite

• bool disable_div_sqrt_rewrite = 11

  Disable x/sqrt(y) to x*rsqrt(y) rewrite

• bool disable_reduce_truncation_optimization = 12

  Disable reduce truncation optimization

• bool disable_maxpooling_optimization = 13

  Disable maxpooling optimization

• bool disallow_mix_types_opts = 14

  Disallow mix type operations

• bool disable_select_optimization = 15

  Disable SelectOp optimization

• bool enable_optimize_denominator_with_broadcast = 16

  Enable optimize x/bcast(y) -> x * bcast(1/y)

• bool disable_deallocation_insertion = 17

  Disable deallocation insertion pass

• bool disable_partial_sort_optimization = 28

  Disable sort->topk rewrite when only partial sort is required

The SPU datatype

Used in: ValueMetaProto

• DT_INVALID = 0

• DT_I1 = 1

  1bit integer (bool).

• DT_I8 = 2

  int8

• DT_U8 = 3

  uint8

• DT_I16 = 4

  int16

• DT_U16 = 5

  uint16

• DT_I32 = 6

  int32

• DT_U32 = 7

  uint32

• DT_I64 = 8

  int64

• DT_U64 = 9

  uint64

• DT_F16 = 10

  half

• DT_F32 = 11

  float

• DT_F64 = 12

  double

The executable format accepted by SPU runtime. - Inputs should be prepared before running executable. - Output is maintained after execution, and can be fetched by output name. ```python rt = spu.Runtime(...) # create an spu runtime. rt.set_var('x', ...) # set variable to the runtime. exe = spu.ExecutableProto( # prepare the executable. name = 'balabala', input_names = ['x'], output_names = ['y'], code = ...) rt.run(exe) # run the executable. y = rt.get_var('y') # get the executable from spu runtime. ```

• string name = 1

  The name of the executable.

• repeated string input_names = 2

  The input names.

• repeated string output_names = 4

  The output names.

• bytes code = 6

  The bytecode of the program, with format IR_MLIR_SPU.

A security parameter type. The secure evaluation is based on some algebraic structure (ring or field),

Used in: RuntimeConfig

• FT_INVALID = 0

• FM32 = 1

  Ring 2^32

• FM64 = 2

  Ring 2^64

• FM128 = 3

  Ring 2^128

The protocol kind.

Used in: RuntimeConfig

• PROT_INVALID = 0

  Invalid protocol.

• REF2K = 1

  The reference implementation in `ring^2k`, note: this 'protocol' only behave-like a fixed point secure protocol without any security guarantee. Hence, it should only be selected for debugging purposes.

• SEMI2K = 2

  A semi-honest multi-party protocol. This protocol requires a trusted third party to generate the offline correlated randoms. Currently, SecretFlow by default ships this protocol with a trusted first party. Hence, it should only be used for debugging purposes.

• ABY3 = 3

  A honest majority 3PC-protocol. SecretFlow provides the semi-honest implementation without Yao.

• CHEETAH = 4

  The famous [Cheetah](https://eprint.iacr.org/2022/207) protocol, a very fast 2PC protocol.

• SECURENN = 5

  A semi-honest 3PC-protocol for Neural Network, P2 as the helper, (https://eprint.iacr.org/2018/442)

Plaintext type SPU runtime does not process with plaintext directly, plaintext type is mainly used for IO purposes, when converting a plaintext buffer to an SPU buffer, we have to let spu know which type the plaintext buffer is.

• PT_INVALID = 0

• PT_I8 = 1

  int8_t

• PT_U8 = 2

  uint8_t

• PT_I16 = 3

  int16_t

• PT_U16 = 4

  uint16_t

• PT_I32 = 5

  int32_t

• PT_U32 = 6

  uint32_t

• PT_I64 = 7

  int64_t

• PT_U64 = 8

  uint64_t

• PT_I128 = 9

  int128_t

• PT_U128 = 10

  uint128_t

• PT_I1 = 11

  bool

• PT_F16 = 30

  half

• PT_F32 = 31

  float

• PT_F64 = 32

  double

• PT_CF32 = 50

  complex float

• PT_CF64 = 51

  complex double

The SPU runtime configuration.

• ProtocolKind protocol = 1

  The protocol kind.

• FieldType field = 2

  The field type.

• int64 fxp_fraction_bits = 3

  Number of fraction bits of fixed-point number. 0(default) indicates implementation defined.

• int32 max_concurrency = 4

  Max number of cores

• bool enable_action_trace = 10

  When enabled, runtime prints verbose info of the call stack, debug purpose only.

• bool enable_type_checker = 11

  When enabled, runtime checks runtime type infos against the compile-time ones, exceptions are raised if mismatches happen. Note: Runtime outputs prefer runtime type infos even when flag is on.

• bool enable_pphlo_trace = 12

  When enabled, runtime prints executed pphlo list, debug purpose only.

• bool enable_runtime_snapshot = 13

  When enabled, runtime dumps executed executables in the dump_dir, debug purpose only.

• string snapshot_dump_dir = 14

• bool enable_pphlo_profile = 15

  When enabled, runtime records detailed pphlo timing data, debug purpose only. WARNING: the `send bytes` information is only accurate when `experimental_enable_inter_op_par` and `experimental_enable_intra_op_par` options are disabled.

• bool enable_hal_profile = 16

  When enabled, runtime records detailed hal timing data, debug purpose only. WARNING: the `send bytes` information is only accurate when `experimental_enable_inter_op_par` and `experimental_enable_intra_op_par` options are disabled.

• uint64 public_random_seed = 19

  The public random variable generated by the runtime, the concrete prg function is implementation defined. Note: this seed only applies to `public variable` only, it has nothing to do with security.

• uint64 share_max_chunk_size = 20

  max chunk size for Value::toProto default: 128 * 1024 * 1024

• RuntimeConfig.SortMethod sort_method = 21

  SPU supports multiple sorting algorithms. -for 2pc, only sorting network is supported. -for 2.5pc or 3pc, all these algorithms are supported. -for stable sort, only radix sort is supported.

• int64 quick_sort_threshold = 22

  threshold for quick sort, when the size of the array is less than this value, use merge sort instead

• int64 fxp_div_goldschmidt_iters = 50

  The iterations use in f_div with Goldschmidt method. 0(default) indicates implementation defined.

• RuntimeConfig.ExpMode fxp_exp_mode = 51

  The exponent approximation method.

• int64 fxp_exp_iters = 52

  Number of iterations of `exp` approximation, 0(default) indicates impl defined.

• RuntimeConfig.LogMode fxp_log_mode = 53

  The logarithm approximation method.

• int64 fxp_log_iters = 54

  Number of iterations of `log` approximation, 0(default) indicates impl-defined.

• int64 fxp_log_orders = 55

  Number of orders of `log` approximation, 0(default) indicates impl defined.

• RuntimeConfig.SigmoidMode sigmoid_mode = 56

  The sigmoid function approximation model.

• bool enable_lower_accuracy_rsqrt = 57

  Enable a simpler rsqrt approximation

• int64 sine_cosine_iters = 58

  Sine/Cosine approximation iterations

• RuntimeConfig.BeaverType beaver_type = 70

  beaver config, works for semi2k and spdz2k for now.

• optional TTPBeaverConfig ttp_beaver_config = 71

  TrustedThirdParty configs.

• optional CheetahConfig cheetah_2pc_config = 72

  Cheetah 2PC configs.

• bool trunc_allow_msb_error = 73

  For protocol like SecureML, the most significant bit may have error with low probability, which lead to huge calculation error.

• bool experimental_disable_mmul_split = 100

  Experimental: DO NOT USE

• bool experimental_enable_inter_op_par = 101

  Inter op parallel, aka, DAG level parallel.

• bool experimental_enable_intra_op_par = 102

  Intra op parallel, aka, hal/mpc level parallel.

• bool experimental_disable_vectorization = 103

  Disable kernel level vectorization.

• uint64 experimental_inter_op_concurrency = 104

  Inter op concurrency.

• bool experimental_enable_colocated_optimization = 105

  Enable use of private type

• bool experimental_enable_exp_prime = 106

  enable experimental exp prime method

• uint32 experimental_exp_prime_offset = 107

  The offset parameter for exp prime methods. control the valid range of exp prime method. valid range is: ((47 - offset - 2fxp)/log_2(e), (125 - 2fxp - offset)/log_2(e)) clamp to value would be lower bound: (48 - offset - 2fxp)/log_2(e) higher bound: (124 - 2fxp - offset)/log_2(e) default offset is 13, 0 offset is not supported.

• bool experimental_exp_prime_disable_lower_bound = 108

  whether to apply the clamping lower bound default to enable it

• bool experimental_exp_prime_enable_upper_bound = 109

  whether to apply the clamping upper bound default to disable it

Used in: RuntimeConfig

• TrustedFirstParty = 0

  Assume first party (rank0) as trusted party to generate beaver triple. WARNING: It is NOT SAFE and SHOULD NOT BE used in production.

• TrustedThirdParty = 1

  Generate beaver triple through an additional trusted third party.

• MultiParty = 2

  Generate beaver triple through multi-party.

The exponential approximation method.

Used in: RuntimeConfig

• EXP_DEFAULT = 0

  Implementation defined.

• EXP_PADE = 1

  The pade approximation.

• EXP_TAYLOR = 2

  Taylor series approximation.

• EXP_PRIME = 3

  exp prime only available for some implementations

The logarithm approximation method.

Used in: RuntimeConfig

• LOG_DEFAULT = 0

  Implementation defined.

• LOG_PADE = 1

  The pade approximation.

• LOG_NEWTON = 2

  The newton approximation.

• LOG_MINMAX = 3

  The minmax approximation.

The sigmoid approximation method.

Used in: RuntimeConfig

• SIGMOID_DEFAULT = 0

  Implementation defined.

• SIGMOID_MM1 = 1

  Minmax approximation one order. f(x) = 0.5 + 0.125 * x

• SIGMOID_SEG3 = 2

  Piece-wise simulation. f(x) = 0.5 + 0.125x if -4 <= x <= 4 1 if x > 4 0 if -4 > x

• SIGMOID_REAL = 3

  The real definition, which depends on exp's accuracy. f(x) = 1 / (1 + exp(-x))

Used in: RuntimeConfig

• SORT_DEFAULT = 0

  Implementation defined.

• SORT_RADIX = 1

  The radix sort (stable sort, need efficient shuffle).

• SORT_QUICK = 2

  The quick sort (unstable, need efficient shuffle).

• SORT_NETWORK = 3

  The odd-even sorting network (unstable, most general).

@exclude The shape information.

Used in: ValueMetaProto

• repeated int64 dims = 1

//////////////////////////////////////////////////////////////////////// Compiler relate definition ////////////////////////////////////////////////////////////////////////

Used in: CompilationSource

• XLA = 0

• STABLEHLO = 1

Used in: RuntimeConfig

• string server_host = 1

  TrustedThirdParty beaver server's remote ip:port or load-balance uri.

• int32 adjust_rank = 2

  which rank do adjust rpc call, usually choose the rank closer to the server.

• string asym_crypto_schema = 3

  asym_crypto_schema: support ["SM2"] Will support 25519 in the future, after yacl supported it.

• string server_public_key = 4

  Server's public key in PEM format

• string transport_protocol = 5

  Transport protocol, support ["http", "h2"]

• optional ClientSSLConfig ssl_config = 6

  Configurations related to SSL

The spu Value proto, used for spu value serialization.

• uint64 total_bytes = 1

  chunk info

• uint64 chunk_offset = 2

• bytes content = 3

  chunk bytes

• DataType data_type = 1

  The data type.

• bool is_complex = 2

• Visibility visibility = 3

  The data visibility.

• optional ShapeProto shape = 4

  The shape of the value.

• string storage_type = 5

  The storage type, defined by the underline evaluation engine. i.e. `aby3.AShr<FM64>` means an aby3 arithmetic share in FM64. usually, the application does not care about this attribute.

The visibility type. SPU is a secure evaluation runtime, but not all data are secret, some of them are publicly known to all parties, marking them as public will improve performance significantly.

Used in: CompilationSource, ValueMetaProto

• VIS_INVALID = 0

• VIS_SECRET = 1

  Invisible(unknown) for all or some of the parties.

• VIS_PUBLIC = 2

  Visible(public) for all parties.

• VIS_PRIVATE = 3

  Visible for only one party

Used in: CompilerOptions

• TEXT = 0

• DOT = 1

• HTML = 2