/third_party/mindspore/mindspore/_extends/graph_kernel/expanders/ |
D | layernorm_grad.py | 25 def _expand(self, graph_builder): argument 34 x = graph_builder.emit('Cast', [x], attrs={'dst_type': 'float32'}) 35 dy = graph_builder.emit('Cast', [dy], attrs={'dst_type': 'float32'}) 36 variance = graph_builder.emit('Cast', [variance], attrs={'dst_type': 'float32'}) 37 mean = graph_builder.emit('Cast', [mean], attrs={'dst_type': 'float32'}) 38 gamma = graph_builder.emit('Cast', [gamma], attrs={'dst_type': 'float32'}) 53 eps = graph_builder.value(x.dtype, epsilon) 54 const_neg_half = graph_builder.value(x.dtype, -0.5) 55 const_neg_two = graph_builder.value(x.dtype, -2.0) 56 const_two = graph_builder.value(x.dtype, 2.0) [all …]
|
D | lamb_apply_optimizer_assign.py | 23 def _expand(self, graph_builder): argument 29 square_grad = graph_builder.emit('Mul', [grad, grad]) 30 mul_3_result = graph_builder.emit('Mul', [square_grad, one_minus_beta_2]) 31 mul_2_result = graph_builder.emit('Mul', [inputv, beta_2]) 32 next_v = graph_builder.emit('Add', [mul_2_result, mul_3_result]) 35 mul_0_result = graph_builder.emit('Mul', [inputm, beta_1]) 36 mul_1_result = graph_builder.emit('Mul', [grad, one_minus_beta_1]) 37 next_m = graph_builder.emit('Add', [mul_0_result, mul_1_result]) 40 const_one = graph_builder.value(beta_2.dtype, 1) 42 beta_1_tensor = graph_builder.emit('BroadcastTo', [beta_1], attrs={'shape': shape}) [all …]
|
D | batchnorm.py | 28 def _expand(self, graph_builder): argument 35 epsilon_v = graph_builder.value(input_scale.dtype, self.attrs['epsilon']) 43 input_x = graph_builder.emit('Cast', [input_x], attrs={'dst_type': input_x_new_type}) 47 res_y, mean_res, variance_res, mean_muls, y_sqrt_rec = self._bn_train(graph_builder) 49 res_y = graph_builder.emit('Cast', [res_y], attrs={'dst_type': input_x_ori_type}) 53 input_mean = graph_builder.emit( 55 input_scale = graph_builder.emit( 57 input_offset = graph_builder.emit( 59 x_sub = graph_builder.emit('Sub', [input_x, input_mean]) 60 x_sub_mul = graph_builder.emit('Mul', [input_scale, x_sub]) [all …]
|
D | gelu_grad.py | 26 def _expand(self, graph_builder): argument 36 const_csvalue = graph_builder.value(input_dy.dtype, self.CSVALUE) 37 … const_csvalue_sqrt_two_div_pi = graph_builder.value(input_dy.dtype, self.CSVALUE_SQRT_TWO_DIV_PI) 38 const_csvalue_tri = graph_builder.value(input_dy.dtype, self.CSVALUE_TRI) 39 const_one = graph_builder.value(input_dy.dtype, 1) 40 const_half = graph_builder.value(input_dy.dtype, 0.5) 43 mul_double = graph_builder.emit('Mul', [input_x, input_x]) 44 mul_double_mul_tri = graph_builder.emit('Mul', [const_csvalue_tri, mul_double]) 45 mul_add_one = graph_builder.emit('Add', [const_one, mul_double_mul_tri]) 46 mul_right = graph_builder.emit('Mul', [const_csvalue_sqrt_two_div_pi, mul_add_one]) [all …]
|
D | lamb_apply_weight_assign.py | 23 def _expand(self, graph_builder): argument 30 data_min = graph_builder.value(dtype, 2**(-126)) 32 data_min = graph_builder.value(dtype, 2*(-24)) 36 const_zero = graph_builder.value(dtype, 0) 37 const_one = graph_builder.value(dtype, 1) 42 g_norm_greater_res = graph_builder.emit('Greater', [g_norm, const_zero]) 43 g_norm_res = graph_builder.emit('Cast', [g_norm_greater_res], attrs={'dst_type': dtype}) 45 g_norm = graph_builder.emit('Add', [g_norm, data_min]) 46 w_norm_g_norm = graph_builder.emit('RealDiv', [w_norm, g_norm]) 48 g_norm_value_1 = graph_builder.emit('Mul', [g_norm_res, w_norm_g_norm]) [all …]
|
D | batchnorm_grad.py | 28 def _expand(self, graph_builder): argument 46 input_x = graph_builder.emit('Cast', [input_x], attrs={'dst_type': 'float32'}) 48 input_dy = graph_builder.emit('Cast', [input_dy], attrs={'dst_type': 'float32'}) 50 num_rec_v = graph_builder.value(input_scale.dtype, num_rec) 51 …dbeta = graph_builder.emit('ReduceSum', [input_dy], attrs={'reduce_axis': reduce_axis, 'keep_dims'… 57 epsilon_v = graph_builder.value(input_scale.dtype, self.attrs['epsilon']) 58 var_add = graph_builder.emit('Add', [input_save_inv_variance, epsilon_v]) 59 sqrt_var_eps = graph_builder.emit('Sqrt', [var_add]) 61 scalar_one_v = graph_builder.value(input_scale.dtype, scalar_one) 62 inv_variance = graph_builder.emit('RealDiv', [scalar_one_v, sqrt_var_eps]) [all …]
|
D | fused_adam_weight_decay.py | 23 def _expand(self, graph_builder): argument 27 beta_1_mul_m = graph_builder.emit('Mul', [beta_1, m]) 28 one_sub_beta_1_mul_grad = graph_builder.emit('Mul', [one_sub_beta_1, gradient]) 29 next_m = graph_builder.emit('Add', [beta_1_mul_m, one_sub_beta_1_mul_grad]) 30 beta_2_mul_v = graph_builder.emit('Mul', [beta_2, v]) 31 grad_square = graph_builder.emit('Mul', [gradient, gradient]) 32 one_sub_beta_2_mul_grad_square = graph_builder.emit('Mul', [one_sub_beta_2, grad_square]) 33 next_v = graph_builder.emit('Add', [beta_2_mul_v, one_sub_beta_2_mul_grad_square]) 34 sqrt_next_v = graph_builder.emit('Sqrt', [next_v]) 35 sqrt_next_v_add_eps = graph_builder.emit('Add', [sqrt_next_v, eps]) [all …]
|
D | layernorm.py | 27 def _expand(self, graph_builder): argument 35 input_x = graph_builder.emit('Cast', [input_x], attrs={'dst_type': 'float32'}) 36 input_gamma = graph_builder.emit('Cast', [input_gamma], attrs={'dst_type': 'float32'}) 37 input_beta = graph_builder.emit('Cast', [input_beta], attrs={'dst_type': 'float32'}) 62 mean_cof_v = graph_builder.value(input_x.dtype, mean_cof) 65 mean_red = graph_builder.emit('ReduceSum', [input_x], 67 mean = graph_builder.emit('Mul', [mean_red, mean_cof_v]) 69 mean = graph_builder.emit('Reshape', [mean], attrs={'shape': ori_reduced_shape_x}) 72 variance_sub = graph_builder.emit('Sub', [input_x, mean]) 73 variance_mul = graph_builder.emit('Mul', [variance_sub, variance_sub]) [all …]
|
D | fused_adam.py | 23 def _expand(self, graph_builder): argument 26 beta_1_mul_m = graph_builder.emit('Mul', [beta_1, m]) 27 one_sub_beta_1_mul_grad = graph_builder.emit('Mul', [one_sub_beta_1, gradient]) 28 next_m = graph_builder.emit('Add', [beta_1_mul_m, one_sub_beta_1_mul_grad]) 29 beta_2_mul_v = graph_builder.emit('Mul', [beta_2, v]) 30 grad_square = graph_builder.emit('Mul', [gradient, gradient]) 31 one_sub_beta_2_mul_grad_square = graph_builder.emit('Mul', [one_sub_beta_2, grad_square]) 32 next_v = graph_builder.emit('Add', [beta_2_mul_v, one_sub_beta_2_mul_grad_square]) 33 sqrt_next_v = graph_builder.emit('Sqrt', [next_v]) 34 sqrt_next_v_add_eps = graph_builder.emit('Add', [sqrt_next_v, eps]) [all …]
|
D | gelu.py | 24 def _expand(self, graph_builder): argument 32 mul_0 = graph_builder.emit('Mul', [input_x, input_x]) 33 pow_0 = graph_builder.emit('Mul', [mul_0, input_x]) 34 const_csvalue = graph_builder.value(pow_0.dtype, self.CSVALUE) 35 mul_1 = graph_builder.emit('Mul', [pow_0, const_csvalue]) 36 tanh_res = graph_builder.emit('Add', [input_x, mul_1]) 37 … const_csvalue_sqrt_two_div_pi = graph_builder.value(tanh_res.dtype, self.CSVALUE_SQRT_TWO_DIV_PI) 38 y = graph_builder.emit('Mul', [tanh_res, const_csvalue_sqrt_two_div_pi]) 41 tanh_y = graph_builder.emit('Tanh', [y]) 42 const_one = graph_builder.value(tanh_y.dtype, 1) [all …]
|
D | softmax_cross_entropy_with_logits.py | 24 def _expand(self, graph_builder): argument 29 … max_x = graph_builder.emit('ReduceMax', [logits], attrs={'reduce_axis': axis, 'keep_dims': True}) 30 data_sub = graph_builder.emit('Sub', [logits, max_x]) 31 data_exp = graph_builder.emit('Exp', [data_sub]) 32 …data_expsum = graph_builder.emit('ReduceSum', [data_exp], attrs={'reduce_axis': axis, 'keep_dims':… 33 data_softmax = graph_builder.emit('RealDiv', [data_exp, data_expsum]) 34 const_eps = graph_builder.value(logits.dtype, 0.000001) 35 data_softmax_safety = graph_builder.emit("Maximum", [data_softmax, const_eps]) 36 softmax_log = graph_builder.emit('Log', [data_softmax_safety]) 37 label_mul_log = graph_builder.emit('Mul', [label, softmax_log]) [all …]
|
D | sigmoid_cross_entropy_with_logits.py | 23 def _expand(self, graph_builder): argument 30 const_one = graph_builder.value(logits.dtype, 1.0) 31 const_zero = graph_builder.value(logits.dtype, 0.0) 32 max_logits = graph_builder.emit('Maximum', [logits, const_zero]) 33 logits_mul_labels = graph_builder.emit('Mul', [logits, labels]) 34 abs_logits = graph_builder.emit('Abs', [logits]) 35 neg_abs_logits = graph_builder.emit('Neg', [abs_logits]) 36 exp_neg_abs_logits = graph_builder.emit('Exp', [neg_abs_logits]) 37 one_add_exp_neg_abs_logits = graph_builder.emit('Add', [const_one, exp_neg_abs_logits]) 38 log_one_add_exp_neg_abs_logits = graph_builder.emit('Log', [one_add_exp_neg_abs_logits]) [all …]
|
D | softmax.py | 27 def _expand(self, graph_builder): argument 47 input_x_f16 = graph_builder.emit('Cast', [input_x], attrs={'dst_type': 'float16'}) 48 …max_x_f16 = graph_builder.emit('ReduceMax', [input_x_f16], attrs={'reduce_axis': axis, 'keep_dims'… 49 max_x = graph_builder.emit('Cast', [max_x_f16], attrs={'dst_type': ori_dtype}) 51 … max_x = graph_builder.emit('ReduceMax', [input_x], attrs={'reduce_axis': axis, 'keep_dims': True}) 54 max_x = graph_builder.emit('Cast', [max_x], attrs={'dst_type': "float32"}) 55 input_x = graph_builder.emit('Cast', [input_x], attrs={'dst_type': "float32"}) 58 max_x = graph_builder.emit('Reshape', [max_x], attrs={'shape': ori_reduced_shape}) 59 data_sub = graph_builder.emit('Sub', [input_x, max_x]) 60 data_exp = graph_builder.emit('Exp', [data_sub]) [all …]
|
D | erfc.py | 22 def _expand(self, graph_builder): argument 26 const_one = graph_builder.value("float32", 1) 27 input_x = graph_builder.emit('Cast', [input_x], attrs={'dst_type': "float32"}) 28 erf_result = graph_builder.emit('Erf', [input_x]) 29 result = graph_builder.emit('Sub', [const_one, erf_result]) 30 result = graph_builder.emit('Cast', [result], attrs={'dst_type': "float16"}) 32 const_one = graph_builder.value(input_x.dtype, 1) 33 erf_result = graph_builder.emit('Erf', [input_x]) 34 result = graph_builder.emit('Sub', [const_one, erf_result])
|
D | logsoftmax.py | 25 def _expand(self, graph_builder): argument 35 input_x_f16 = graph_builder.emit('Cast', [input_x], attrs={'dst_type': 'float16'}) 36 …max_x_f16 = graph_builder.emit('ReduceMax', [input_x_f16], attrs={'reduce_axis': axis, 'keep_dims'… 37 max_x = graph_builder.emit('Cast', [max_x_f16], attrs={'dst_type': ori_dtype}) 39 … max_x = graph_builder.emit('ReduceMax', [input_x], attrs={'reduce_axis': axis, 'keep_dims': True}) 40 data_sub = graph_builder.emit('Sub', [input_x, max_x]) 41 data_exp = graph_builder.emit('Exp', [data_sub]) 42 …data_expsum = graph_builder.emit('ReduceSum', [data_exp], attrs={'reduce_axis': axis, 'keep_dims':… 43 log_expsum = graph_builder.emit('Log', [data_expsum]) 44 result = graph_builder.emit('Sub', [data_sub, log_expsum])
|
D | sigmoid_cross_entropy_with_logits_grad.py | 23 def _expand(self, graph_builder): argument 28 const_one = graph_builder.value(logits.dtype, 1.0) 29 neg_x = graph_builder.emit('Neg', [logits]) 30 exp_neg_x = graph_builder.emit('Exp', [neg_x]) 31 add_exp = graph_builder.emit('Add', [const_one, exp_neg_x]) 32 sigmoid_res = graph_builder.emit('RealDiv', [const_one, add_exp]) 33 sigmoid_res_sub_label = graph_builder.emit('Sub', [sigmoid_res, label]) 34 res = graph_builder.emit('Mul', [sigmoid_res_sub_label, dout])
|
D | gkdropout.py | 24 def _expand(self, graph_builder): argument 28 r_keep_prob = graph_builder.value(input_x.dtype, 1.0 / keep_prob) 29 keep_prob = graph_builder.value(input_x.dtype, keep_prob) 32 input_mask = graph_builder.emit('Cast', [input_mask], attrs={'dst_type': input_x.dtype}) 33 mask = graph_builder.emit('LessEqual', [input_mask, keep_prob]) # output is bool type 34 mask = graph_builder.emit('Cast', [mask], attrs={'dst_type': input_x.dtype}) 37 result = graph_builder.emit('Mul', [r_keep_prob, input_x]) 38 result = graph_builder.emit('Mul', [result, mask])
|
D | maximum_grad.py | 30 def _expand(self, graph_builder): argument 32 ge_result = graph_builder.emit('GreaterEqual', [input_x, input_y]) 33 ge_result = graph_builder.emit('Cast', [ge_result], attrs={'dst_type': input_x.dtype}) 34 dx = graph_builder.emit('Mul', [ge_result, input_dout]) 35 dy = graph_builder.emit('Sub', [input_dout, dx]) 40 …dx_reduce = graph_builder.emit('ReduceSum', [dx], attrs={'reduce_axis': reduce_axis_x, 'keep_dims'… 42 dx_out = graph_builder.emit('Reshape', [dx_reduce], attrs={'shape': input_x.shape}) 49 …dy_reduce = graph_builder.emit('ReduceSum', [dy], attrs={'reduce_axis': reduce_axis_y, 'keep_dims'… 51 dy_out = graph_builder.emit('Reshape', [dy_reduce], attrs={'shape': input_y.shape})
|
D | minimum_grad.py | 29 def _expand(self, graph_builder): argument 32 le_result = graph_builder.emit('LessEqual', [input_x, input_y]) 33 le_result = graph_builder.emit('Cast', [le_result], attrs={'dst_type': input_x.dtype}) 34 dx = graph_builder.emit('Mul', [le_result, input_dout]) 35 dy = graph_builder.emit('Sub', [input_dout, dx]) 43 …dx_reduce = graph_builder.emit('ReduceSum', [dx], attrs={'reduce_axis': reduce_axis_x, 'keep_dims'… 45 dx_out = graph_builder.emit('Reshape', [dx_reduce], attrs={'shape': input_x.shape}) 52 …dy_reduce = graph_builder.emit('ReduceSum', [dy], attrs={'reduce_axis': reduce_axis_y, 'keep_dims'… 54 dy_out = graph_builder.emit('Reshape', [dy_reduce], attrs={'shape': input_y.shape})
|
D | sigmoid.py | 22 def _expand(self, graph_builder): argument 26 const_one = graph_builder.value(input_x.dtype, 1.0) 27 neg_x = graph_builder.emit('Neg', [input_x]) 28 exp_neg_x = graph_builder.emit('Exp', [neg_x]) 29 add_exp = graph_builder.emit('Add', [const_one, exp_neg_x]) 30 res = graph_builder.emit('RealDiv', [const_one, add_exp])
|
/third_party/mindspore/mindspore/_extends/graph_kernel/expanders/complex/ |
D | div.py | 24 def _expand(self, graph_builder): argument 27 x_real = graph_builder.emit('CReal', [input_x]) 28 y_real = graph_builder.emit('CReal', [input_y]) 29 x_imag = graph_builder.emit('CImag', [input_x]) 30 y_imag = graph_builder.emit('CImag', [input_y]) 31 squre_y_real = graph_builder.emit('Mul', [y_real, y_real]) 32 squre_y_imag = graph_builder.emit('Mul', [y_imag, y_imag]) 33 final_denominator = graph_builder.emit('Add', [squre_y_real, squre_y_imag]) 34 x_real_mul_y_real = graph_builder.emit('Mul', [x_real, y_real]) 35 x_imag_mul_y_imag = graph_builder.emit('Mul', [x_imag, y_imag]) [all …]
|
D | mul.py | 24 def _expand(self, graph_builder): argument 27 x_real = graph_builder.emit('CReal', [input_x]) 28 y_real = graph_builder.emit('CReal', [input_y]) 29 x_imag = graph_builder.emit('CImag', [input_x]) 30 y_imag = graph_builder.emit('CImag', [input_y]) 31 x_real_mul_y_real = graph_builder.emit('Mul', [x_real, y_real]) 32 x_imag_mul_y_imag = graph_builder.emit('Mul', [x_imag, y_imag]) 33 x_real_mul_y_imag = graph_builder.emit('Mul', [x_real, y_imag]) 34 x_imag_mul_y_real = graph_builder.emit('Mul', [x_imag, y_real]) 35 result_real = graph_builder.emit('Sub', [x_real_mul_y_real, x_imag_mul_y_imag]) [all …]
|
D | add.py | 24 def _expand(self, graph_builder): argument 26 x_real = graph_builder.emit('CReal', [input_x]) 27 y_real = graph_builder.emit('CReal', [input_y]) 28 x_imag = graph_builder.emit('CImag', [input_x]) 29 y_imag = graph_builder.emit('CImag', [input_y]) 30 result_real = graph_builder.emit('Add', [x_real, y_real]) 31 result_imag = graph_builder.emit('Add', [x_imag, y_imag]) 32 result = graph_builder.emit('Complex', [result_real, result_imag])
|
D | sub.py | 24 def _expand(self, graph_builder): argument 26 x_real = graph_builder.emit('CReal', [input_x]) 27 y_real = graph_builder.emit('CReal', [input_y]) 28 x_imag = graph_builder.emit('CImag', [input_x]) 29 y_imag = graph_builder.emit('CImag', [input_y]) 30 result_real = graph_builder.emit('Sub', [x_real, y_real]) 31 result_imag = graph_builder.emit('Sub', [x_imag, y_imag]) 32 result = graph_builder.emit('Complex', [result_real, result_imag])
|
D | abs.py | 22 def _expand(self, graph_builder): argument 24 x_real = graph_builder.emit('CReal', [input_x]) 25 x_imag = graph_builder.emit('CImag', [input_x]) 26 squre_x_real = graph_builder.emit('Mul', [x_real, x_real]) 27 squre_x_imag = graph_builder.emit('Mul', [x_imag, x_imag]) 28 squre_sum = graph_builder.emit('Add', [squre_x_real, squre_x_imag]) 29 result = graph_builder.emit('Sqrt', [squre_sum])
|