Numba中float128浮点数声明、全局设置及类型标注问题咨询

Let's tackle your questions one by one to clear up any confusion about using float128 in Numba:

1. Does nb.jit('f8()') specify a function with no arguments returning float128? #

No — the type code f8 corresponds to 64-bit floating-point numbers (aka float64 or double-precision), not 128-bit. For float128, you need to use either the explicit type name float128 or the type code f16 (since 128 bits = 16 bytes) in your signature. So the correct signature for a no-argument function returning float128 would be:

@nb.jit('float128()')
def func_nb():
    # Your code here
    return nb.float128(1.23)

Or alternatively:

@nb.jit('f16()')
def func_nb():
    return nb.float128(1.23)

Just note that float128 support is platform-dependent (most reliable on x86_64 systems; ARM platforms may have limited or no support).

2. How to make all floating-point numbers in Numba use float128? #

Numba doesn’t have a global switch to force all floating-point variables to float128, but you can enforce this behavior with these practical approaches:

• Explicit function signatures: Define all parameters and return values as float128 in your JIT decorator. For example, a function taking two float128 inputs and returning a float128:

  @nb.jit('float128(float128, float128)')
  def add(a, b):
      return a + b
  

• Initialize variables with float128: Inside your function, explicitly create float128 values using nb.float128() or numpy.float128(). This ensures any derived variables inherit the 128-bit type:

  @nb.jit(nopython=True)
  def high_precision_calc():
      x = nb.float128(0.1)
      y = nb.float128(0.2)
      return x + y  # Result is float128
  

• Use float128 arrays: If working with numpy arrays, pass arrays with dtype=np.float128 — Numba will preserve this type throughout computations:

  import numpy as np
  
  arr = np.array([1.0, 2.0], dtype=np.float128)
  
  @nb.jit(nopython=True)
  def process_arr(arr):
      return arr * nb.float128(3.0)  # Output array remains float128
  

3. How to declare float128 variables in Numba? #

You have a few straightforward ways to declare float128 variables:

• Explicit type conversion: Convert literals or other values to float128 using nb.float128():

  @nb.jit(nopython=True)
  def declare_vars():
      pi = nb.float128(3.14159265358979323846)
      return pi
  

• Type annotations: Use Python type hints with Numba's float128 type (works in nopython mode):

  @nb.jit(nopython=True)
  def annotated_vars():
      x: nb.float128 = 0.1
      y: nb.float128 = x * 2
      return y
  

• Numpy float128 values: Create numpy scalars with np.float128() and assign them to variables — Numba recognizes this type:

  import numpy as np
  
  @nb.jit(nopython=True)
  def numpy_float128():
      val = np.float128(1e-10)
      return val
  

Just a quick reminder: Always test for float128 support on your target platform, as some architectures (like ARM) may not fully implement 128-bit floating-point operations.

内容的提问来源于stack exchange，提问作者kilojoules