nvmexplorer_src.nvmFI.data_transforms package

Submodules

nvmexplorer_src.nvmFI.data_transforms.bitmask_utils module

nvmexplorer_src.nvmFI.data_transforms.bitmask_utils.encoded_capacity(wmb, data, num_bits)

Returns the total capacity in bits of the bitmask + data

Parameters

  • wmb – weight mask bits – bit vectors to indicate non-zero values in original weight matrix M

  • data – Data corresponding to the non-zero elements of matrix M

  • num_bits – number of bits per data value to use

nvmexplorer_src.nvmFI.data_transforms.bitmask_utils.from_bitmask(wmb, data)

Translate bitmask encoded data back to original matrix M

Parameters

  • wmb – weight mask bits – bit vectors to indicate non-zero values in original weight matrix M

  • data – Data corresponding to the non-zero elements of matrix M

nvmexplorer_src.nvmFI.data_transforms.bitmask_utils.to_bitmask(M)

Returns the bitmask sparse format of matrix ‘M’

Parameters

M – original weight matrix M

nvmexplorer_src.nvmFI.data_transforms.data_transform_utils module

nvmexplorer_src.nvmFI.data_transforms.data_transform_utils.convert_f_mat(v_mlc, conf, int_bits, frac_bits, exp_bias, q_type)

Convert MLC-packed per-storage-cell values back to floating point values

Parameters

  • v_mlc – vector of per-storage-cell values (possible MLC encoding)

  • conf – mapping from number of bits in input value to number of levels per NVM storage cell (e.g., [2, 2] for a 2 bit input into SLCs)

  • int_bits – number of integer bits for data format (if applicable)

  • frac_bits – number of fractional bits for data format (if applicable)

  • exp_bias – exponent bias for data format (if applicable)

  • q_type – data format choice (e.g., signed, unsigned, adaptive floating point)

nvmexplorer_src.nvmFI.data_transforms.data_transform_utils.convert_mlc_mat(num_float, rep_conf, int_bits, frac_bits, exp_bias, q_type)

Format an entire input matrix into per-memory-cell array under MLC config for bit-level fault injection

Parameters

  • num_float – input value (floating point)

  • rep_conf – mapping from number of bits in input value to number of levels per NVM storage cell (e.g., [2, 2] for a 2 bit input into SLCs)

  • int_bits – number of integer bits for data format (if applicable)

  • frac_bits – number of fractional bits for data format (if applicable)

  • exp_bias – exponent bias for data format (if applicable)

  • q_type – data format choice (e.g., signed, unsigned, adaptive floating point)

nvmexplorer_src.nvmFI.data_transforms.data_transform_utils.get_afloat_bias(num_float, n_exp)

Extract bias term for AdaptivFloat data format https://arxiv.org/abs/1909.13271

Parameters

  • num_float – input value as float

  • n_exp – number of exponential bits for adaptivFloar data format

nvmexplorer_src.nvmFI.data_transforms.data_transform_utils.get_binary_array_mat(orig_flt, rep_conf, int_bits, frac_bits, exp_bias, q_type)

Format an input float value into binary array for bit-level fault injection

Parameters

  • orig_flt – input value (floating point)

  • rep_conf – mapping from number of bits in input value to number of levels per NVM storage cell (e.g., [2, 2] for a 2 bit input into SLCs)

  • int_bits – number of integer bits for data format (if applicable)

  • frac_bits – number of fractional bits for data format (if applicable)

  • exp_bias – exponent bias for data format (if applicable)

  • q_type – data format choice (e.g., signed, unsigned, adaptive floating point)

nvmexplorer_src.nvmFI.data_transforms.data_transform_utils.get_q_afloat(num_float, n_bits, n_exp, bias)

Conversion to AdaptivFloat data format https://arxiv.org/abs/1909.13271

Parameters

  • num_float – input value as float

  • n_bits – total number of bits per value (mantissa bits = n_bits - n_exp)

  • n_exp – number of exponential bits for adaptivFloar data format

  • bias – input bias term

nvmexplorer_src.nvmFI.data_transforms.graph_utils module

nvmexplorer_src.nvmFI.data_transforms.graph_utils.gen_adj_mat(G)

Generate an adjacency matrix (size NxN for N node graph) for an input snapPY graph

Parameters

G – snapPY input graph

nvmexplorer_src.nvmFI.data_transforms.graph_utils.gen_graph(M)

Generate a snapPY graph based on an adjacency matrix

Parameters

M – adjacency matrix representing stored graph

nvmexplorer_src.nvmFI.data_transforms.graph_utils.pack_bits(M, bpc=2)

For MLC storage, pack adjacency matrix according to MLC settings

Parameters

  • M – adjacency matrix representing stored graph

  • bpc – bits-per-cell for MLC storage format

nvmexplorer_src.nvmFI.data_transforms.graph_utils.unpack_bits(M_pack, bpc=2)

For MLC storage, unpack adjacency matrix according to MLC settings

Parameters

  • M – adjacency matrix representing stored graph, packed according to MLC storage settings

  • bpc – bits-per-cell for MLC storage format

Module contents