'''Copyright (c) 2017-2018 Mozilla Copyright (c) 2022 Amazon Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: - Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. - Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. ''' import numpy as np from .c_writer import CWriter def print_vector(writer, vector, name, dtype='float', reshape_8x4=False, static=True, debug_float=False): if isinstance(writer, CWriter): f = writer.source binary_blob = writer.enable_binary_blob else: f = writer binary_blob = False dtype_suffix = { 'float' : 'float', 'opus_int8' : 'int8', 'opus_uint16' : 'uint16', 'opus_int16' : 'int16', 'int' : 'int', 'qweight': 'qweight' } if binary_blob: f.write( f''' #ifndef USE_WEIGHTS_FILE ''' ) writer.weight_arrays.append(name) if reshape_8x4: vector = vector.reshape((vector.shape[0]//4, 4, vector.shape[1]//8, 8)) vector = vector.transpose((2, 0, 3, 1)) v = np.reshape(vector, (-1)) if debug_float: f.write('#ifndef DISABLE_DEBUG_FLOAT\n') if binary_blob: f.write( f''' #define WEIGHTS_{name}_DEFINED #define WEIGHTS_{name}_TYPE WEIGHT_TYPE_{dtype_suffix[dtype]} ''' ) if static: f.write('static ') f.write(f'const {dtype} {name}[{len(v)}] = {{\n ') for i in range(0, len(v)): f.write(f'{v[i]}') if (i!=len(v)-1): f.write(',') else: break if (i%8==7): f.write("\n ") else: f.write(" ") f.write('\n};\n\n') if debug_float: f.write('#endif /*DISABLE_DEBUG_FLOAT*/\n') if binary_blob: f.write( f''' #endif /* USE_WEIGHTS_FILE */ ''' ) return vector def extract_diagonal(A): """ input shape is (N, k*N) """ N, M = A.shape B = A.copy() assert M % N == 0 k = M // N diags = [] for l in range(k): diag = np.diag(B[:, l * N : (l+1) * N]).copy() B[:, l * N : (l+1) * N] -= np.diag(diag) diags.append(diag) diag = np.concatenate(diags) return diag, B def quantize_weight(weight, scale): scale = scale + 1e-30 Aq = np.round(weight / scale).astype('int') if Aq.max() > 127 or Aq.min() <= -128: raise ValueError("value out of bounds in quantize_weight") Aq = np.clip(np.round(weight / scale).astype('int'), -128, 127) return Aq def print_sparse_weight(writer, A, name, scale=1/128, have_diag=True, quantize=False): N = A.shape[0] M = A.shape[1] W = np.zeros((0,), dtype='int') W0 = np.zeros((0,)) if have_diag: diag, A = extract_diagonal(A) print_vector(writer, diag, name + '_diag') if quantize: Aq = quantize_weight(A, scale) else: Aq = A # extract blocks idx = np.zeros((0,), dtype='int') for i in range(M//8): pos = idx.shape[0] idx = np.append(idx, -1) nb_nonzero = 0 for j in range(N//4): block = A[j*4:(j+1)*4, i*8:(i+1)*8] qblock = Aq[j*4:(j+1)*4, i*8:(i+1)*8] if np.sum(np.abs(block)) > 1e-10: nb_nonzero = nb_nonzero + 1 idx = np.append(idx, j*4) vblock = qblock.transpose((1,0)).reshape((-1,)) W0 = np.concatenate([W0, block.reshape((-1,))]) W = np.concatenate([W, vblock]) idx[pos] = nb_nonzero if quantize: print_vector(writer, W, name + '_int8', reshape_8x4=False, dtype='opus_int8') print_vector(writer, W0, name + '_float', reshape_8x4=False, dtype='float', debug_float=quantize) print_vector(writer, idx, name + '_idx', reshape_8x4=False, dtype='int') return Aq def compute_scaling(weight): """ computes optimal scaling vector for weight of shape (features_in, features_out) """ n_in, n_out = weight.shape assert n_in % 4 == 0 and n_out % 8 == 0 weight_max_abs = np.max(np.abs(weight), axis=0) weight_max_sum = np.max(np.abs(weight[: n_in : 2] + weight[1 : n_in : 2]), axis=0) scale_max = weight_max_abs / 127 scale_sum = weight_max_sum / 129 scale = np.maximum(scale_max, scale_sum) return scale def qn(string): if string == "NULL": return string else: return '"' + string + '"' def print_linear_layer(writer : CWriter, name : str, weight : np.ndarray, bias : np.ndarray, scale : np.ndarray = None, sparse : bool = False, diagonal : bool = False, quantize : bool = True): """ prints linear layer Parameters: ----------- name : str layer name weight: np.ndarray ... scale: np.ndarray or None If None auto scaling will be applied. Otherwise, output channels will be multiplied by scale (the usual broadcasting rules apply). """ if len(weight.shape) != 2: raise ValueError('expecting 2-dim weight array in print_linear_layer') bias_name = "NULL" if bias is None else name + "_bias" subias_name = name + "_subias" if quantize else "NULL" scale_name = name + "_scale" if quantize else "NULL" idx_name = name + "_weights_idx" if sparse else "NULL" float_weight_name = name + "_weights_float" int_weight_name = name + "_weights_int8" if quantize else "NULL" diag_name = name + "_weights_diag" if sparse and diagonal else "NULL" nb_inputs, nb_outputs = weight.shape if scale is None and quantize: scale = compute_scaling(weight) if sparse: weight_q = print_sparse_weight(writer, weight, name + "_weights", scale=scale, have_diag=diagonal, quantize=quantize) else: if quantize: weight_q = quantize_weight(weight, scale) print_vector(writer, weight_q, name + "_weights_int8", dtype='opus_int8', reshape_8x4=True) print_vector(writer, weight, name + "_weights_float", dtype='float', reshape_8x4=False, debug_float=quantize) if quantize: subias = (np.zeros(nb_outputs) if bias is None else bias) - np.sum(weight_q * scale, axis=0) print_vector(writer, subias, name + "_subias") final_scale = scale / 127 * np.ones(nb_outputs) print_vector(writer, final_scale, name + "_scale") if bias is not None: print_vector(writer, bias, name + "_bias") init_call = f'linear_init(&model->{name}, arrays, {qn(bias_name)}, {qn(subias_name)}, {qn(int_weight_name)},' \ + f'{qn(float_weight_name)}, {qn(idx_name)}, {qn(diag_name)}, {qn(scale_name)}, {nb_inputs}, {nb_outputs})' writer.layer_dict[name] = ('LinearLayer', init_call) def print_dense_layer(writer : CWriter, name : str, weight : np.ndarray, bias : np.ndarray, scale=1/128, format : str = 'torch', sparse=False, diagonal=False, quantize=False): if format == 'torch': weight = weight.transpose() print_linear_layer(writer, name, weight, bias, scale=scale, sparse=sparse, diagonal=diagonal, quantize=quantize) writer.header.write(f"\n#define {name.upper()}_OUT_SIZE {weight.shape[1]}\n") def print_conv1d_layer(writer : CWriter, name : str, weight : np.ndarray, bias : np.ndarray, scale=1/128, format : str = 'torch', quantize=False, sparse=False): if format == "torch": # convert to channels last weight = np.transpose(weight, (2, 1, 0)) lin_weight = np.reshape(weight, (-1, weight.shape[-1])) print_linear_layer(writer, name, lin_weight, bias, scale=scale, sparse=sparse, diagonal=False, quantize=quantize) writer.header.write(f"\n#define {name.upper()}_OUT_SIZE {weight.shape[2]}\n") writer.header.write(f"\n#define {name.upper()}_IN_SIZE {weight.shape[1]}\n") writer.header.write(f"\n#define {name.upper()}_STATE_SIZE ({weight.shape[1]} * ({weight.shape[0] - 1}))\n") writer.header.write(f"\n#define {name.upper()}_DELAY {(weight.shape[0] - 1) // 2}\n") # CAVE: delay is not a property of the conv layer return weight.shape[0] * weight.shape[1] def print_conv2d_layer(writer : CWriter, name : str, weight : np.ndarray, bias : np.ndarray, scale : float=1/128, quantize : bool=False): if quantize: print("[print_conv2d_layer] warning: quantize argument ignored") bias_name = name + "_bias" float_weight_name = name + "_weight_float" print_vector(writer, weight, float_weight_name) print_vector(writer, bias, bias_name) # init function out_channels, in_channels, ksize1, ksize2 = weight.shape init_call = f'conv2d_init(&model->{name}, arrays, "{bias_name}", "{float_weight_name}", {in_channels}, {out_channels}, {ksize1}, {ksize2})' writer.layer_dict[name] = ('Conv2dLayer', init_call) def print_gru_layer(writer : CWriter, name : str, weight : np.ndarray, recurrent_weight : np.ndarray, bias : np.ndarray, recurrent_bias : np.ndarray, format : str = 'torch', quantize : bool = False, input_sparse : bool = False, recurrent_sparse : bool = False, scale=1/128, recurrent_scale=1/128 ): if format == "torch": # change gate ordering from rzn to zrn N = weight.shape[0] // 3 for x in [weight, recurrent_weight, bias, recurrent_bias]: if x is None: continue tmp = x[0:N].copy() x[0:N] = x[N:2*N] x[N:2*N] = tmp weight = weight.transpose() recurrent_weight = recurrent_weight.transpose() else: N = weight.shape[1] // 3 print_linear_layer(writer, name + "_input", weight, bias, scale=scale, sparse=input_sparse, quantize=quantize) print_linear_layer(writer, name + "_recurrent", recurrent_weight, recurrent_bias, scale=recurrent_scale, sparse=recurrent_sparse, diagonal=recurrent_sparse, quantize=quantize) # wrapping it up writer.header.write(f"\n#define {name.upper()}_OUT_SIZE {N}\n") writer.header.write(f"\n#define {name.upper()}_STATE_SIZE {N}\n") return N def print_tconv1d_layer(writer : CWriter, name : str, weight : np.ndarray, bias : np.ndarray, stride: int, scale=1/128, quantize=False, sparse=False): in_channels, out_channels, kernel_size = weight.shape linear_weight = weight.transpose(2, 1, 0).reshape(kernel_size * out_channels, in_channels).transpose(1, 0) linear_bias = np.repeat(bias[np.newaxis, :], kernel_size, 0).flatten() print_linear_layer(writer, name, linear_weight, linear_bias, scale=scale, quantize=quantize, sparse=sparse) writer.header.write(f"\n#define {name.upper()}_KERNEL_SIZE {kernel_size}\n") writer.header.write(f"\n#define {name.upper()}_STRIDE {stride}\n") writer.header.write(f"\n#define {name.upper()}_IN_CHANNELS {in_channels}\n") writer.header.write(f"\n#define {name.upper()}_OUT_CHANNELS {out_channels}\n")