fix(pad_v2): golden bfloat16 分支 cv 转为 float 避免类型错误-ops-math-AtomGit

Mconversion/batch_to_space_nd/tests/assets/golden.py+109-62

	@@ -11,8 +11,6 @@
11	# ----------------------------------------------------------------------------	11	# ----------------------------------------------------------------------------
12		12
13	import numpy as np	13	import numpy as np
14	-import tensorflow.compat.v1 as tf
15	-tf.disable_v2_behavior()
16		14
17		15
18	__golden__ = {	16	__golden__ = {
	@@ -22,59 +20,96 @@ __golden__ = {
22	}	20	}
23		21
24		22
25	-def _trans_nc1hwc0_to_nhwc(data):	23	+def _ceil_div(x, y):
26	- in_shape = data.shape	24	+ return (x + y - 1) // y
		25	+
		26	+
		27	+def shape_6d_2_5d(tensor):
		28	+ n, di, c1, hi, wi, c0 = tensor.shape
		29	+ tmp_tensor = tensor.reshape(n, di, c1, hi, wi, c0)
		30	+ tmp_tensor = np.transpose(tmp_tensor, axes=(0, 1, 3, 4, 2, 5))
		31	+ new_tensor = tmp_tensor.reshape(n, di, hi, wi, c1 * c0)
		32	+ return new_tensor
		33	+
		34	+
		35	+def shape_5d_2_6d(tensor):
		36	+ import math
		37	+ c0 = 16
		38	+ n, di, hi, wi, c = tensor.shape
		39	+ c1 = math.ceil(c / c0)
		40	+ tmp_tensor = tensor.reshape(n, di, hi, wi, c1, c0)
		41	+ new_tensor = np.transpose(tmp_tensor, axes=(0, 1, 4, 2, 3, 5))
		42	+ return new_tensor
		43	+
		44	+
		45	+def trans_nhwc_to_nc1hwc0(input_data, c0=16):
		46	+ in_shape = np.shape(input_data)
		47	+ axis_n = in_shape[0]
		48	+ axis_h = in_shape[1]
		49	+ axis_w = in_shape[2]
		50	+ axis_c = in_shape[3]
		51	+ axis_c0 = c0
		52	+ axis_c1 = _ceil_div(axis_c, axis_c0)
		53	+ c_pad = 0
		54	+ tmp_input_tensor = np.pad(input_data, ((0, 0), (0, 0), (0, 0), (0, c_pad)), mode="constant", constant_values=(0, 0))
		55	+ tmp_input_tensor = tmp_input_tensor.reshape(axis_n, axis_h, axis_w, axis_c1, axis_c0)
		56	+ output_arr = np.transpose(tmp_input_tensor, axes=(0, 3, 1, 2, 4))
		57	+ return output_arr
		58	+
		59	+
		60	+def trans_nc1hwc0_to_nhwc(input_data):
		61	+ in_shape = np.shape(input_data)
27	axis_n = in_shape[0]	62	axis_n = in_shape[0]
28	axis_c1 = in_shape[1]	63	axis_c1 = in_shape[1]
29	axis_h = in_shape[2]	64	axis_h = in_shape[2]
30	axis_w = in_shape[3]	65	axis_w = in_shape[3]
31	axis_c0 = in_shape[4]	66	axis_c0 = in_shape[4]
32	- tmp_input_tensor = np.transpose(data, axes=(0, 2, 3, 1, 4))	67	+ tmp_input_tensor = np.transpose(input_data, axes=(0, 2, 3, 1, 4))
33	tmp_input_tensor = tmp_input_tensor.reshape(axis_n, axis_h, axis_w, axis_c1 * axis_c0)	68	tmp_input_tensor = tmp_input_tensor.reshape(axis_n, axis_h, axis_w, axis_c1 * axis_c0)
34	return tmp_input_tensor	69	return tmp_input_tensor
35		70
36		71
37	-def _trans_nhwc_to_nc1hwc0(data):	72	+def _high_dim_batch_to_space_nd(input_tensor, block_shape, crops):
38	- in_shape = data.shape	73	+ import tensorflow as tf
39	- axis_n = in_shape[0]
40	- axis_h = in_shape[1]
41	- axis_w = in_shape[2]
42	- axis_c = in_shape[3]
43	- axis_c0 = 16
44	- axis_c1 = (axis_c + axis_c0 - 1) // axis_c0
45	- tmp_input_tensor = data.reshape(axis_n, axis_h, axis_w, axis_c1, axis_c0)
46	- tmp_input_tensor = np.transpose(tmp_input_tensor, axes=(0, 3, 1, 2, 4))
47	- return tmp_input_tensor
48		74
		75	+ M = block_shape.shape[0]
		76	+ input_shape = input_tensor.shape
		77	+ batch_size = input_shape[0]
		78	+ spatial_shape = input_shape[1 : 1 + M]
		79	+ remain_shape = input_shape[1 + M :]
49		80
50	-def _shape_6d_2_5d(data):	81	+ block_prod = tf.reduce_prod(block_shape)
51	- in_shape = data.shape	82	+ new_batch = batch_size // block_prod
52	- axis_d = in_shape[0]
53	- axis_c1 = in_shape[1]
54	- axis_h = in_shape[2]
55	- axis_w = in_shape[3]
56	- axis_n = in_shape[4]
57	- axis_c0 = in_shape[5]
58	- axis_n = axis_n * axis_d
59	- axis_c = axis_c1 * axis_c0
60	- tmp_input_tensor = data.reshape(axis_n, axis_c1, axis_h, axis_w, axis_c0)
61	- tmp_input_tensor = np.transpose(tmp_input_tensor, axes=(0, 2, 3, 1, 4))
62	- tmp_input_tensor = tmp_input_tensor.reshape(axis_n, axis_h, axis_w, axis_c)
63	- return tmp_input_tensor
64		83
		84	+ new_shape = tf.concat(
		85	+ [block_shape, [new_batch], spatial_shape, remain_shape], axis=0
		86	+ )
		87	+ reshaped = tf.reshape(input_tensor, new_shape)
65		88
66	-def _shape_5d_2_6d(data):	89	+ perm = [M]
67	- in_shape = data.shape	90	+ for i in range(M):
68	- axis_n = in_shape[0]	91	+ perm.append(M + 1 + i)
69	- axis_h = in_shape[1]	92	+ perm.append(i)
70	- axis_w = in_shape[2]	93	+ remain_start = M + 1 + M
71	- axis_c = in_shape[3]	94	+ perm.extend(range(remain_start, reshaped.shape.rank))
72	- axis_c0 = 16	95	+ transposed = tf.transpose(reshaped, perm)
73	- axis_c1 = (axis_c + axis_c0 - 1) // axis_c0	96	+
74	- axis_d = 1	97	+ expanded_spatial = spatial_shape * block_shape
75	- tmp_input_tensor = data.reshape(axis_d, axis_n, axis_h, axis_w, axis_c1, axis_c0)	98	+ new_shape2 = tf.concat([[new_batch], expanded_spatial, remain_shape], axis=0)
76	- tmp_input_tensor = np.transpose(tmp_input_tensor, axes=(0, 4, 2, 3, 1, 5))	99	+ merged = tf.reshape(transposed, new_shape2)
77	- return tmp_input_tensor	100	+
		101	+ begin = [0]
		102	+ size = [new_batch]
		103	+ for i in range(M):
		104	+ begin.append(crops[i, 0])
		105	+ size.append(expanded_spatial[i] - crops[i, 0] - crops[i, 1])
		106	+ R = input_tensor.shape.rank - 1 - M
		107	+ for i in range(R):
		108	+ begin.append(0)
		109	+ size.append(remain_shape[i])
		110	+
		111	+ output = tf.slice(merged, begin, size)
		112	+ return output
78		113
79		114
80	def batch_to_space_nd_golden(x, block_shape, crops, **kwargs):	115	def batch_to_space_nd_golden(x, block_shape, crops, **kwargs):
	@@ -86,24 +121,36 @@ def batch_to_space_nd_golden(x, block_shape, crops, **kwargs):
86	input_formats, output_formats, input_ori_formats, output_ori_formats,	121	input_formats, output_formats, input_ori_formats, output_ori_formats,
87	input_dtypes, output_dtypes.	122	input_dtypes, output_dtypes.
88	'''	123	'''
89	- input_format = kwargs.get('input_formats', ['ND'])[0]	124	+ import tensorflow.compat.v1 as tf
90	-	125	+ tf.disable_v2_behavior()
91	- if input_format == 'NC1HWC0':	126	+
92	- x = _trans_nc1hwc0_to_nhwc(x)	127	+ data_x = x
93	- elif input_format == "NDC1HWC0":	128	+ block_shape_arr = np.array(block_shape).astype(np.int64)
94	- x = _shape_6d_2_5d(x)	129	+ crops_arr = np.array(crops).astype(np.int64)
95	-	130	+
96	- tensor_x = tf.placeholder(x.dtype, shape=x.shape)	131	+ input_formats = kwargs.get('input_formats', ())
97	- block_shape_tensor = tf.constant(block_shape.tolist())	132	+ fmt = input_formats[0] if input_formats and len(input_formats) > 0 else 'ND'
98	- crops_tensor = tf.constant(crops.tolist())	133	+ if isinstance(fmt, (list, tuple)):
99	- out = tf.batch_to_space_nd(tensor_x, block_shape_tensor, crops_tensor)	134	+ fmt = fmt[0] if fmt else 'ND'
100	-	135	+
		136	+ if fmt == 'NC1HWC0':
		137	+ data_x = trans_nc1hwc0_to_nhwc(data_x)
		138	+ elif fmt == "NDC1HWC0":
		139	+ data_x = shape_6d_2_5d(data_x)
		140	+
		141	+ tensor_x = tf.placeholder(data_x.dtype, shape=data_x.shape)
		142	+ tf_block_shape = tf.constant(block_shape_arr)
		143	+ tf_crops = tf.constant(crops_arr)
		144	+ if block_shape_arr.shape[0] > 4:
		145	+ out = _high_dim_batch_to_space_nd(tensor_x, tf_block_shape, tf_crops)
		146	+ else:
		147	+ out = tf.batch_to_space_nd(tensor_x, tf_block_shape, tf_crops)
		148	+
101	with tf.Session() as sess:	149	with tf.Session() as sess:
102	- res = sess.run(out, feed_dict={tensor_x: x})	150	+ res = sess.run(out, feed_dict={tensor_x: data_x})
103	-	151	+
104	- if input_format == "NC1HWC0":	152	+ if fmt == "NC1HWC0":
105	- res = _trans_nhwc_to_nc1hwc0(res)	153	+ res = trans_nhwc_to_nc1hwc0(res)
106	- elif input_format == "NDC1HWC0":	154	+ elif fmt == "NDC1HWC0":
107	- res = _shape_5d_2_6d(res)	155	+ res = shape_5d_2_6d(res)
108	-	156	+ return res
109	- return res

Mconversion/pad_v2/tests/assets/golden.py+1-1

	@@ -130,7 +130,7 @@ def pad_v2_golden(x, paddings, constant_values, **kwargs):
130	cv = constant_values_arr[0]	130	cv = constant_values_arr[0]
131	elif x.dtype.name == "bfloat16":	131	elif x.dtype.name == "bfloat16":
132	x_tensor = _numpy_to_torch_tensor(x).to(torch.float32)	132	x_tensor = _numpy_to_torch_tensor(x).to(torch.float32)
133	- cv = constant_values_arr[0]	133	+ cv = float(constant_values_arr[0])
134	elif x.dtype.name in fp8dtypes:	134	elif x.dtype.name in fp8dtypes:
135	x_tensor = _numpy_to_torch_tensor(x.view(np.int8))	135	x_tensor = _numpy_to_torch_tensor(x.view(np.int8))
136	constant_values_arr = constant_values_arr.view(np.int8)	136	constant_values_arr = constant_values_arr.view(np.int8)

Mconversion/transpose/tests/assets/golden.py+1-2

	@@ -15,8 +15,7 @@ import numpy as np
15		15
16	__golden__ = {	16	__golden__ = {
17	"kernel": {	17	"kernel": {
18	- "transpose": "transpose_golden",	18	+ "transpose": "transpose_golden"
19	- "transpose_d": "transpose_golden"
20	}	19	}
21	}	20	}
22		21

Mmath/sinc/tests/assets/golden.py+54-10

	@@ -11,7 +11,6 @@
11	# ----------------------------------------------------------------------------	11	# ----------------------------------------------------------------------------
12		12
13	import numpy as np	13	import numpy as np
14	-import torch
15		14
16		15
17	__golden__ = {	16	__golden__ = {
	@@ -21,13 +20,58 @@ __golden__ = {
21	}	20	}
22		21
23		22
24	-def sinc_golden(x, **kwargs):	23	+def _numpy_bfloat16():
25	- ori_dtype = kwargs.get("input_dtypes", ["float32"])[0]	24	+ try:
26	- x_dtype = x.dtype	25	+ from ml_dtypes import bfloat16
27	-	26	+ except ModuleNotFoundError:
28	- if "bfloat16" in str(ori_dtype).lower() or "float16" in str(ori_dtype).lower():	27	+ try:
29	- x_tensor = torch.from_numpy(x.astype(np.float32))	28	+ import tensorflow
30	- output = torch.sinc(x_tensor)	29	+ bfloat16 = tensorflow.bfloat16.as_numpy_dtype
31	- return output.numpy().astype(x_dtype, copy=False)	30	+ except ModuleNotFoundError:
		31	+ raise RuntimeError("ml-dtypes or tensorflow is needed to support bfloat16 dtype!!! "
		32	+ "Please install with `pip3 install ml-dtypes` or `pip3 install tensorflow`")
		33	+ return bfloat16
		34	+
		35	+
		36	+def _numpy_to_torch_tensor(np_array):
		37	+ import torch
		38	+ if np_array is None:
		39	+ return None
		40	+ np_dtype = np_array.dtype.name
		41	+ if "bfloat16" in np_dtype:
		42	+ np_int16 = np_array.view(dtype=np.int16)
		43	+ t_int16 = torch.from_numpy(np_int16)
		44	+ return t_int16.view(torch.bfloat16)
32	else:	45	else:
33	- return np.sinc(x)	46	+ return torch.from_numpy(np_array)
		47	+
		48	+
		49	+def _torch_to_numpy_tensor(torch_tensor):
		50	+ import torch
		51	+ if torch_tensor is None:
		52	+ return None
		53	+ if not isinstance(torch_tensor, torch.Tensor):
		54	+ raise RuntimeError(f"Only support torch.Tensor. But got {type(torch_tensor)}")
		55	+ torch_dtype = torch_tensor.dtype
		56	+ if torch_dtype == torch.bfloat16:
		57	+ t_int16 = torch_tensor.view(torch.int16)
		58	+ np_int16 = t_int16.numpy()
		59	+ return np_int16.view(dtype=_numpy_bfloat16())
		60	+ else:
		61	+ return torch_tensor.numpy()
		62	+
		63	+
		64	+def sinc_golden(x, **kwargs):
		65	+ '''
		66	+ Kernel golden for sinc.
		67	+ All the parameters follow @sinc_def.cpp without outputs.
		68	+ All the input Tensors are numpy.ndarray.
		69	+ kwargs may contain: short_soc_version, input_ori_shapes, output_ori_shapes,
		70	+ input_formats, output_formats, input_ori_formats, output_ori_formats,
		71	+ input_dtypes, output_dtypes.
		72	+ '''
		73	+ import torch
		74	+
		75	+ input_x = _numpy_to_torch_tensor(x)
		76	+ output = torch.sinc(input_x)
		77	+ return _torch_to_numpy_tensor(output)

Amath/truncate_div/tests/assets/golden.py+183-0

@@ -0,0 +1,183 @@
	1	+#!/usr/bin/env python3
	2	+# -- coding: UTF-8 --
	3	+# ----------------------------------------------------------------------------
	4	+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
	5	+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
	6	+# CANN Open Software License Agreement Version 2.0 (the "License").
	7	+# Please refer to the License for details. You may not use this file except in compliance with the License.
	8	+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
	9	+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
	10	+# See LICENSE in the root of the software repository for the full text of the License.
	11	+# ----------------------------------------------------------------------------
	12	+
	13	+import numpy as np
	14	+
	15	+
	16	+__golden__ = {
	17	+ "kernel": {
	18	+ "truncate_div": "truncate_div_golden"
	19	+ }
	20	+}
	21	+
	22	+
	23	+def _broadcast_to_maxshape(shapes):
	24	+ """
	25	+ produce broadcast shape
	26	+ for example:
	27	+ input: shape is [[2, 3], [3, 2, 1], [3, 1, 3]]
	28	+ output: [1, 2, 3], [3, 2, 1], [3, 1, 3], [3, 2, 3]
	29	+ """
	30	+ def _max(_shape):
	31	+ no_one_shape = [s for s in _shape if s != 1]
	32	+ if len(no_one_shape) == 0:
	33	+ max_value = 1
	34	+ else:
	35	+ max_value = no_one_shape[0]
	36	+ return max_value
	37	+ max_dim_length = max(len(list(shape)) for shape in shapes)
	38	+ input_shapes = []
	39	+ for shape in shapes:
	40	+ input_shapes.append([1 for _ in range(max_dim_length - len(shape))] + list(shape))
	41	+ input_shapes = list(map(list, zip(*input_shapes)))
	42	+ max_shape = [_max(shape) for shape in input_shapes]
	43	+ input_shapes = list(map(list, zip(*input_shapes)))
	44	+ return (*input_shapes, max_shape)
	45	+
	46	+
	47	+def _numpy_bfloat16():
	48	+ try:
	49	+ from ml_dtypes import bfloat16
	50	+ except ModuleNotFoundError:
	51	+ try:
	52	+ import tensorflow
	53	+ bfloat16 = tensorflow.bfloat16.as_numpy_dtype
	54	+ except ModuleNotFoundError:
	55	+ raise RuntimeError("ml-dtypes or tensorflow is needed to support bfloat16 dtype!!! "
	56	+ "Please install with `pip3 install ml-dtypes` or `pip3 install tensorflow`")
	57	+ return bfloat16
	58	+
	59	+
	60	+def _numpy_to_torch_tensor(np_array):
	61	+ import torch
	62	+ if np_array is None:
	63	+ return None
	64	+ np_dtype = np_array.dtype.name
	65	+ if "bfloat16" in np_dtype:
	66	+ np_int16 = np_array.view(dtype=np.int16)
	67	+ t_int16 = torch.from_numpy(np_int16)
	68	+ return t_int16.view(torch.bfloat16)
	69	+ else:
	70	+ return torch.from_numpy(np_array)
	71	+
	72	+
	73	+def _torch_to_numpy_tensor(torch_tensor):
	74	+ import torch
	75	+ if torch_tensor is None:
	76	+ return None
	77	+ if not isinstance(torch_tensor, torch.Tensor):
	78	+ raise RuntimeError(f"Only support torch.Tensor. But got {type(torch_tensor)}")
	79	+ torch_dtype = torch_tensor.dtype
	80	+ if torch_dtype == torch.bfloat16:
	81	+ t_int16 = torch_tensor.view(torch.int16)
	82	+ np_int16 = t_int16.numpy()
	83	+ return np_int16.view(dtype=_numpy_bfloat16())
	84	+ else:
	85	+ return torch_tensor.numpy()
	86	+
	87	+
	88	+def truncate_div_golden(x1, x2, **kwargs):
	89	+ '''
	90	+ Kernel golden for truncate_div.
	91	+ All the parameters follow @truncate_div_def.cpp without outputs.
	92	+ All the input Tensors are numpy.ndarray.
	93	+ kwargs may contain: short_soc_version, input_ori_shapes, output_ori_shapes,
	94	+ input_formats, output_formats, input_ori_formats, output_ori_formats,
	95	+ input_dtypes, output_dtypes.
	96	+ '''
	97	+ import torch
	98	+
	99	+ type_int = [torch.int8, torch.int32, torch.int64]
	100	+ type_uint = [torch.uint8]
	101	+ type_float = [torch.float, torch.float16, torch.bfloat16]
	102	+
	103	+ _, _, res_shape = _broadcast_to_maxshape([x1.shape, x2.shape])
	104	+ x2_broadcast = np.broadcast_to(x2, res_shape)
	105	+ zero_x2_broadcast_idx = np.where(x2_broadcast == 0)
	106	+
	107	+ output_dtype = kwargs.get("output_dtypes", [None])[0]
	108	+ if output_dtype is None:
	109	+ output_dtype = str(x1.dtype)
	110	+ need_zero_handling = output_dtype in ["int32", "int8", "uint8", "int64"]
	111	+
	112	+ zero_idx = np.where(x2 == 0)
	113	+ if len(zero_idx[0]) > 0 and need_zero_handling:
	114	+ x2[zero_idx] = 1
	115	+
	116	+ x1_t = _numpy_to_torch_tensor(x1)
	117	+ x2_t = _numpy_to_torch_tensor(x2)
	118	+
	119	+ res = 0
	120	+
	121	+ dtype = torch.promote_types(x1_t.dtype, x2_t.dtype)
	122	+ if dtype in (torch.int32, torch.int64):
	123	+ info = torch.iinfo(dtype)
	124	+ min_val, max_val = info.min, info.max
	125	+ dangerous_mask = (x1_t == min_val) & (x2_t == -1)
	126	+ safe_x2 = torch.where(dangerous_mask, torch.ones_like(x2_t), x2_t)
	127	+ res = torch.div(x1_t, safe_x2, rounding_mode="trunc")
	128	+ elif dtype == torch.int16:
	129	+ dangerous_mask = (x2_t == 0)
	130	+ safe_x1 = torch.where(dangerous_mask, torch.full_like(x1_t, -1), x1_t)
	131	+ safe_x2 = torch.where(dangerous_mask, torch.full_like(x2_t, -1), x2_t)
	132	+ res = torch.div(safe_x1, safe_x2, rounding_mode="trunc")
	133	+ else:
	134	+ res = torch.div(x1_t, x2_t, rounding_mode='trunc')
	135	+
	136	+ if len(zero_idx[0]) > 0 and need_zero_handling:
	137	+ x2[zero_idx] = 0
	138	+ if res.dtype in type_int:
	139	+ res[zero_x2_broadcast_idx] = -1
	140	+ if res.dtype in type_uint:
	141	+ res[zero_x2_broadcast_idx] = 255
	142	+
	143	+ if dtype in (torch.int16, torch.int32):
	144	+ info = torch.iinfo(dtype)
	145	+ min_val, max_val = info.min, info.max
	146	+
	147	+ mask = (x1_t == max_val) & (x2_t == min_val)
	148	+ res = torch.where(mask, torch.tensor(0, dtype=dtype), res)
	149	+
	150	+ mask = (x1_t == max_val) & (x2_t == -1)
	151	+ res = torch.where(mask, torch.tensor(-max_val, dtype=dtype), res)
	152	+
	153	+ mask = (x1_t == min_val) & (x2_t == -1)
	154	+ res = torch.where(mask, torch.tensor(min_val, dtype=dtype), res)
	155	+
	156	+ mask = (x2_t == 0)
	157	+ res = torch.where(mask, torch.tensor(-1, dtype=dtype), res)
	158	+
	159	+ if dtype == torch.int64:
	160	+ info = torch.iinfo(dtype)
	161	+ min_val, max_val = info.min, info.max
	162	+
	163	+ mask = (x1_t == max_val) & (x2_t == 0)
	164	+ res = torch.where(mask, torch.tensor(-1, dtype=dtype), res)
	165	+
	166	+ mask = (x1_t == max_val) & (x2_t == -1)
	167	+ res = torch.where(mask, torch.tensor(-max_val, dtype=dtype), res)
	168	+
	169	+ mask = (x1_t == min_val) & (x2_t == 0)
	170	+ res = torch.where(mask, torch.tensor(-1, dtype=dtype), res)
	171	+
	172	+ mask = (x1_t == min_val) & (x2_t == -1)
	173	+ res = torch.where(mask, torch.tensor(min_val, dtype=dtype), res)
	174	+
	175	+ mask = (x1_t >= 0) & (x1_t != max_val) & (x2_t == 0)
	176	+ res = torch.where(mask, torch.tensor(4294967295, dtype=dtype), res)
	177	+
	178	+ mask = (x1_t == -1) & (x2_t == 0)
	179	+ res = torch.where(mask, torch.tensor(-1, dtype=dtype), res)
	180	+
	181	+ res_np = _torch_to_numpy_tensor(res)
	182	+
	183	+ return res_np.astype(output_dtype, copy=False)

Amath/truncate_div/tests/st/arch35/ttk_kernel_truncate_div_st.csv+6-0

@@ -0,0 +1,6 @@
	1	+testcase_name,network_name,op_name,input_dtypes,input_ori_shapes,output_ori_shapes,input_ori_formats,output_ori_formats,attributes,input_shapes,output_dtypes,output_shapes,input_formats,output_formats,input_data_ranges,precision_tolerances,absolute_precision,output_inplace_indexes,output_shape_unknown_indexes,is_enabled,remark,soc_series,priority,dump_file_prefix,manual_input_binaries,manual_golden_binaries
	2	+truncate_div_custom_15,UNKNOWN,truncate_div,"('float16', 'float16')","((13, 17, 21), ())","((13, 17, 21),)","('ND', 'ND')","('ND',)",{},"((13, 17, 21), ())","('float16',)","((13, 17, 21),)","('ND', 'ND')","('ND',)","((-65504, 65504), (-1, 1))",,1e-08,(),(),True,,,0,,(),()
	3	+truncate_div_custom_2,UNKNOWN,truncate_div,"('float32', 'float32')","((), (2048, 5, 2))","((2048, 5, 2),)","('ND', 'ND')","('ND',)",{},"((), (2048, 5, 2))","('float32',)","((2048, 5, 2),)","('ND', 'ND')","('ND',)","((-10000, 10000), (-10000, 10000))",,1e-08,(),(),True,,,0,,(),()
	4	+truncate_div_custom_51,UNKNOWN,truncate_div,"('int32', 'float32')","((3, 5, 7, 99), (3, 1, 1, 99))","((3, 5, 7, 99),)","('ND', 'ND')","('ND',)",{},"((3, 5, 7, 99), (3, 1, 1, 99))","('float32',)","((3, 5, 7, 99),)","('ND', 'ND')","('ND',)","((-2147483648, 2147483647), (-10, 10))",,1e-08,(),(),True,,,0,,(),()
	5	+truncate_div_custom_48,UNKNOWN,truncate_div,"('int64', 'int64')","((17772, 1, 2, 1, 2, 1, 2, 1), (2, 2, 2, 2, 2, 2, 2))","((17772, 2, 2, 2, 2, 2, 2, 2),)","('ND', 'ND')","('ND',)",{},"((17772, 1, 2, 1, 2, 1, 2, 1), (2, 2, 2, 2, 2, 2, 2))","('int64',)","((17772, 2, 2, 2, 2, 2, 2, 2),)","('ND', 'ND')","('ND',)","((-9223372036854775808, 9223372036854775807), (-9223372036854775808, 9223372036854775807))",,1e-08,(),(),True,,,0,,(),()
	6	+truncate_div_custom_50,UNKNOWN,truncate_div,"('float16', 'float32')","((3, 5, 7, 99), (3, 1, 1, 99))","((3, 5, 7, 99),)","('ND', 'ND')","('ND',)",{},"((3, 5, 7, 99), (3, 1, 1, 99))","('float32',)","((3, 5, 7, 99),)","('ND', 'ND')","('ND',)","((-3.4028235e+38, 3.4028235e+38), (-1, 1))",,1e-08,(),(),True,,,0,,(),()

Amath/truncate_mod/tests/assets/golden.py+135-0

@@ -0,0 +1,135 @@
	1	+#!/usr/bin/env python3
	2	+# -- coding: UTF-8 --
	3	+# ----------------------------------------------------------------------------
	4	+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
	5	+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
	6	+# CANN Open Software License Agreement Version 2.0 (the "License").
	7	+# Please refer to the License for details. You may not use this file except in compliance with the License.
	8	+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
	9	+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
	10	+# See LICENSE in the root of the software repository for the full text of the License.
	11	+# ----------------------------------------------------------------------------
	12	+
	13	+import numpy as np
	14	+
	15	+
	16	+__golden__ = {
	17	+ "kernel": {
	18	+ "truncate_mod": "truncate_mod_golden"
	19	+ }
	20	+}
	21	+
	22	+
	23	+def _broadcast_to_maxshape(shapes):
	24	+ """
	25	+ produce broadcast shape
	26	+ for example:
	27	+ input: shape is [[2, 3], [3, 2, 1], [3, 1, 3]]
	28	+ output: [1, 2, 3], [3, 2, 1], [3, 1, 3], [3, 2, 3]
	29	+ """
	30	+ def _max(_shape):
	31	+ no_one_shape = [s for s in _shape if s != 1]
	32	+ if len(no_one_shape) == 0:
	33	+ max_value = 1
	34	+ else:
	35	+ max_value = no_one_shape[0]
	36	+ return max_value
	37	+ max_dim_length = max(len(list(shape)) for shape in shapes)
	38	+ input_shapes = []
	39	+ for shape in shapes:
	40	+ input_shapes.append([1 for _ in range(max_dim_length - len(shape))] + list(shape))
	41	+ input_shapes = list(map(list, zip(*input_shapes)))
	42	+ max_shape = [_max(shape) for shape in input_shapes]
	43	+ input_shapes = list(map(list, zip(*input_shapes)))
	44	+ return (*input_shapes, max_shape)
	45	+
	46	+
	47	+def _numpy_bfloat16():
	48	+ try:
	49	+ from ml_dtypes import bfloat16
	50	+ except ModuleNotFoundError:
	51	+ try:
	52	+ import tensorflow
	53	+ bfloat16 = tensorflow.bfloat16.as_numpy_dtype
	54	+ except ModuleNotFoundError:
	55	+ raise RuntimeError("ml-dtypes or tensorflow is needed to support bfloat16 dtype!!! "
	56	+ "Please install with `pip3 install ml-dtypes` or `pip3 install tensorflow`")
	57	+ return bfloat16
	58	+
	59	+
	60	+def _numpy_to_torch_tensor(np_array):
	61	+ import torch
	62	+ if np_array is None:
	63	+ return None
	64	+ np_dtype = np_array.dtype.name
	65	+ if "bfloat16" in np_dtype:
	66	+ np_int16 = np_array.view(dtype=np.int16)
	67	+ t_int16 = torch.from_numpy(np_int16)
	68	+ return t_int16.view(torch.bfloat16)
	69	+ else:
	70	+ return torch.from_numpy(np_array)
	71	+
	72	+
	73	+def _torch_to_numpy_tensor(torch_tensor):
	74	+ import torch
	75	+ if torch_tensor is None:
	76	+ return None
	77	+ if not isinstance(torch_tensor, torch.Tensor):
	78	+ raise RuntimeError(f"Only support torch.Tensor. But got {type(torch_tensor)}")
	79	+ torch_dtype = torch_tensor.dtype
	80	+ if torch_dtype == torch.bfloat16:
	81	+ t_int16 = torch_tensor.view(torch.int16)
	82	+ np_int16 = t_int16.numpy()
	83	+ return np_int16.view(dtype=_numpy_bfloat16())
	84	+ else:
	85	+ return torch_tensor.numpy()
	86	+
	87	+
	88	+def truncate_mod_golden(x1, x2, **kwargs):
	89	+ '''
	90	+ Kernel golden for truncate_mod.
	91	+ All the parameters follow @truncate_mod_def.cpp without outputs.
	92	+ All the input Tensors are numpy.ndarray.
	93	+ kwargs may contain: short_soc_version, input_ori_shapes, output_ori_shapes,
	94	+ input_formats, output_formats, input_ori_formats, output_ori_formats,
	95	+ input_dtypes, output_dtypes.
	96	+ '''
	97	+ import torch
	98	+
	99	+ output_dtype = kwargs.get("output_dtypes", [None])[0]
	100	+ if output_dtype is None:
	101	+ output_dtype = str(x1.dtype)
	102	+
	103	+ type_int = [torch.int8, torch.int16, torch.int32, torch.int64]
	104	+ type_uint = [torch.uint8, torch.uint16, torch.uint32, torch.uint64]
	105	+ type_float = [torch.float16, torch.bfloat16, torch.float, torch.float64]
	106	+
	107	+ # copy
	108	+ x1 = x1.copy()
	109	+ x2 = x2.copy()
	110	+
	111	+ # 除零保护
	112	+ _, _, res_shape = _broadcast_to_maxshape([x1.shape, x2.shape])
	113	+ X2_broadcast = np.broadcast_to(x2, res_shape)
	114	+ zero_X2_broadcast_idx = np.where(X2_broadcast == 0)
	115	+
	116	+ zero_idx = np.where(x2 == 0)
	117	+ if zero_idx:
	118	+ x2[zero_idx] = 1
	119	+
	120	+ x1 = _numpy_to_torch_tensor(x1)
	121	+ x2 = _numpy_to_torch_tensor(x2)
	122	+ res = torch.fmod(x1, x2)
	123	+
	124	+ # 除零保护
	125	+ if zero_idx:
	126	+ x2[zero_idx] = 0
	127	+ if res.dtype in type_int:
	128	+ res[zero_X2_broadcast_idx] = -1
	129	+ if res.dtype in type_uint:
	130	+ res[zero_X2_broadcast_idx] = 255
	131	+ if res.dtype in type_float:
	132	+ res[zero_X2_broadcast_idx] = torch.nan
	133	+
	134	+ res_np = _torch_to_numpy_tensor(res)
	135	+ return res_np.astype(output_dtype, copy=False)

Amath/truncate_mod/tests/st/arch35/ttk_kernel_truncate_mod_st.csv+6-0

@@ -0,0 +1,6 @@
	1	+testcase_name,network_name,op_name,input_dtypes,input_ori_shapes,output_ori_shapes,input_ori_formats,output_ori_formats,attributes,input_shapes,output_dtypes,output_shapes,input_formats,output_formats,input_data_ranges,precision_tolerances,absolute_precision,output_inplace_indexes,output_shape_unknown_indexes,is_enabled,remark,soc_series,priority,dump_file_prefix,manual_input_binaries,manual_golden_binaries
	2	+truncate_mod_3,UNKNOWN,truncate_mod,"('bfloat16', 'bfloat16')","((1, 79, 1, 112, 1), (1, 79, 1, 112, 1))","((1, 79, 1, 112, 1),)","('NCDHW', 'NCDHW')","('NCDHW',)",{},"((1, 79, 1, 112, 1), (1, 79, 1, 112, 1))","('bfloat16',)","((1, 79, 1, 112, 1),)","('NCDHW', 'NCDHW')","('NCDHW',)","((10, 1000), (0.0001, 100))","((0.004, 0.004),)",1e-08,(),(),True,,,0,,(),()
	3	+truncate_mod_31,UNKNOWN,truncate_mod,"('float16', 'float16')","((1, 96, 1, 1, 96), (1, 96, 1, 1, 96))","((1, 96, 1, 1, 96),)","('NDHWC', 'NDHWC')","('NDHWC',)",{},"((1, 96, 1, 1, 96), (1, 96, 1, 1, 96))","('float16',)","((1, 96, 1, 1, 96),)","('NDHWC', 'NDHWC')","('NDHWC',)","((-1, 1), (-1, 1))","((0.001, 0.001),)",1e-08,(),(),True,,,0,,(),()
	4	+truncate_mod_14,UNKNOWN,truncate_mod,"('float32', 'float32')","((96, 1), (96, 1))","((96, 1),)","('ND', 'ND')","('ND',)",{},"((1, 6, 16, 16), (1, 6, 16, 16))","('float32',)","((1, 6, 16, 16),)","('FRACTAL_NZ', 'FRACTAL_NZ')","('FRACTAL_NZ',)","((-1, 1), (-1, 1))","((0.0001, 0.0001),)",1e-08,(),(),True,,,0,,(),()
	5	+truncate_mod_6,UNKNOWN,truncate_mod,"('int8', 'int8')","((1, 1, 7, 1), (1, 1, 7, 1))","((1, 1, 7, 1),)","('NCHW', 'NCHW')","('NCHW',)",{},"((7, 1, 16, 32), (7, 1, 16, 32))","('int8',)","((7, 1, 16, 32),)","('FRACTAL_Z', 'FRACTAL_Z')","('FRACTAL_Z',)","((-100, 100), (-100, 100))","((0.001, 0.001),)",1e-08,(),(),True,,,0,,(),()
	6	+truncate_mod_21,UNKNOWN,truncate_mod,"('int8', 'int8')","((1, 1, 1, 320), (1, 1, 1, 320))","((1, 1, 1, 320),)","('NCHW', 'NCHW')","('NCHW',)",{},"((1, 1, 1, 320), (1, 1, 1, 320))","('int8',)","((1, 1, 1, 320),)","('NCHW', 'NCHW')","('NCHW',)","((-100, 100), (-100, 100))","((0.001, 0.001),)",1e-08,(),(),True,,,0,,(),()