From b2d5cfabcf1d373b7abbe67daeeb09846c703ae4 Mon Sep 17 00:00:00 2001
From: Zhou Wei <1183042833@qq.com>
Date: Fri, 21 Apr 2023 21:16:28 +0800
Subject: [PATCH] fix BF16 dtype bug of to_tensor, due to numpy not support
 BF16 (#53151) (#53152)

---
 python/paddle/fluid/data_feeder.py            | 26 ++++++++++++++++++-
 .../fluid/tests/unittests/test_var_base.py    | 20 ++++++++++++++
 python/paddle/tensor/creation.py              |  7 ++++-
 3 files changed, 51 insertions(+), 2 deletions(-)

diff --git a/python/paddle/fluid/data_feeder.py b/python/paddle/fluid/data_feeder.py
index b4643b3afb5..77abddce03d 100644
--- a/python/paddle/fluid/data_feeder.py
+++ b/python/paddle/fluid/data_feeder.py
@@ -17,6 +17,7 @@ import numpy as np
 import os
 import multiprocessing
 import warnings
+import struct
 
 from .framework import (
     Variable,
@@ -45,6 +46,27 @@ _PADDLE_DTYPE_2_NUMPY_DTYPE = {
 }
 
 
+def copy_bits_from_float_to_uint16(f):
+    return struct.unpack('<I', struct.pack('<f', f))[0] >> 16
+
+
+def convert_float_to_uint16(data, data_format="NCHW"):
+    if data.size == 0:
+        return data.view(np.uint16)
+
+    if data_format == "NHWC":
+        data = np.transpose(data, [0, 3, 1, 2])
+
+    new_data = []
+    for x in np.nditer(data):
+        new_data.append(np.uint16(copy_bits_from_float_to_uint16(x)))
+    new_data = np.reshape(new_data, data.shape).view(np.uint16)
+
+    if data_format == "NHWC":
+        new_data = np.transpose(new_output, [0, 2, 3, 1])
+    return new_data
+
+
 def convert_dtype(dtype):
     if isinstance(dtype, core.VarDesc.VarType):
         if dtype in _PADDLE_DTYPE_2_NUMPY_DTYPE:
@@ -86,7 +108,9 @@ def convert_dtype(dtype):
             # type, so it will not be handled by the previous branch. We need
             # to convert it to str here.
             return str(dtype)
-        # NOTE(zhangbo): Now numpy does not support bfloat, and paddle use uint16 to represent bfloat16, and there binaries are consistent.
+        # NOTE(zhangbo): Now numpy does not support bfloat, so use numpy.uint16 to represent paddle.bfloat16, there binaries are consistent.
+        # If cast ndarray to uint16 and trans to tensor, should not ndarray.astype('uint16') directly
+        # should use function 'convert_float_to_uint16' above, otherwise bits is wrong
         if dtype in ['bfloat16']:
             return 'uint16'
 
diff --git a/python/paddle/fluid/tests/unittests/test_var_base.py b/python/paddle/fluid/tests/unittests/test_var_base.py
index 24920eb375c..d31ccaf86dd 100644
--- a/python/paddle/fluid/tests/unittests/test_var_base.py
+++ b/python/paddle/fluid/tests/unittests/test_var_base.py
@@ -246,6 +246,26 @@ class TestVarBase(unittest.TestCase):
                 np.testing.assert_array_equal(x.numpy(), numpy_array)
                 self.assertEqual(x.type, core.VarDesc.VarType.LOD_TENSOR)
 
+                # test dtype bfloat16
+                x = paddle.to_tensor(-1e6, dtype=paddle.bfloat16)
+                self.assertEqual(x.dtype, core.VarDesc.VarType.BF16)
+                self.assertTrue(x == -999424.0)
+
+                x = paddle.to_tensor([-1e6, -1e6, -1e6], dtype='bfloat16')
+                self.assertEqual(x.dtype, core.VarDesc.VarType.BF16)
+                self.assertTrue(x[0] == -999424.0)
+                self.assertTrue(x[1] == -999424.0)
+                self.assertTrue(x[2] == -999424.0)
+
+                x = paddle.to_tensor(
+                    -1e6, dtype=paddle.bfloat16, stop_gradient=False
+                )
+                self.assertEqual(x.dtype, core.VarDesc.VarType.BF16)
+                self.assertTrue(x == -999424.0)
+                y = x * x
+                y.backward()
+                self.assertTrue(x.grad == -999424.0 * 2)
+
                 with self.assertRaises(ValueError):
                     paddle.randn([3, 2, 2]).item()
                 with self.assertRaises(ValueError):
diff --git a/python/paddle/tensor/creation.py b/python/paddle/tensor/creation.py
index 43e6b9c6d86..8e92dfa3a5e 100644
--- a/python/paddle/tensor/creation.py
+++ b/python/paddle/tensor/creation.py
@@ -28,6 +28,7 @@ from ..fluid.data_feeder import (
     check_type,
     check_variable_and_dtype,
     convert_dtype,
+    convert_float_to_uint16,
 )
 from ..fluid.framework import (
     Variable,
@@ -613,7 +614,11 @@ def _to_tensor_non_static(data, dtype=None, place=None, stop_gradient=True):
                 data = data.astype(default_type)
 
     if dtype and convert_dtype(dtype) != data.dtype:
-        data = data.astype(convert_dtype(dtype))
+        if convert_dtype(dtype) in ['uint16']:
+            # should not ndarray.astype('uint16') directly, data bits is wrong
+            data = convert_float_to_uint16(data.astype('float32'))
+        else:
+            data = data.astype(convert_dtype(dtype))
 
     if _in_eager_without_dygraph_check() and isinstance(data, np.ndarray):
         return core.eager.Tensor(
-- 
GitLab