warp.atomic_sub#

warp.atomic_sub(arr: Array[Any], i: Int, value: Any) → Any#
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

For composite dtypes (vectors, matrices, quaternions, and transforms), the update is atomic only per component. Another thread can observe arr[i] with only some of its components updated.

On GPU, concurrent floating-point updates to the same element are applied in an unspecified order, so rounding can make the result differ between runs. Enable deterministic mode for reproducible results; see Deterministic Execution. In deterministic mode, compiling a kernel that uses the returned value raises WarpDeterminismError.

This function does not support float16 or bfloat16 on GPUs with compute capability below 7.0, or float64 below 6.0. Kernels compiled for those GPUs skip these updates without raising an error. Use float32 on those GPUs instead.

Currently, the backward pass does not recompute the returned value, so gradients that depend on it are incorrect. If the backward pass needs the returned value, for example as an index, save it to an array in the forward pass and return it from a custom replay function; see Example 2: Custom Replay Function. Even with a custom replay function, autodiff treats the returned value as a constant, so arr receives no gradient from expressions that use it.

Parameters:

  • arr – Array containing the element to update. Its scalar type must be int32, uint32, int64, uint64, float16, bfloat16, float32, or float64.

  • i – Index of the element to update.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i] before this update. For composite dtypes, the result may not equal any value that arr[i] held at a single point in time.

warp.atomic_sub(arr: Array[Any], i: Int, j: Int, value: Any) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j] before this update.

warp.atomic_sub(arr: Array[Any], i: Int, j: Int, k: Int, value: Any) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • k – Third index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j,k] before this update.

warp.atomic_sub(
arr: Array[Any],
i: Int,
j: Int,
k: Int,
l: Int,
value: Any,
) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • k – Third index of the element.

  • l – Fourth index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j,k,l] before this update.

warp.atomic_sub(arr: FabricArray[Any], i: Int, value: Any) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

For composite dtypes (vectors, matrices, quaternions, and transforms), the update is atomic only per component. Another thread can observe arr[i] with only some of its components updated.

On GPU, concurrent floating-point updates to the same element are applied in an unspecified order, so rounding can make the result differ between runs. Enable deterministic mode for reproducible results; see Deterministic Execution. In deterministic mode, compiling a kernel that uses the returned value raises WarpDeterminismError.

This function does not support float16 or bfloat16 on GPUs with compute capability below 7.0, or float64 below 6.0. Kernels compiled for those GPUs skip these updates without raising an error. Use float32 on those GPUs instead.

Currently, the backward pass does not recompute the returned value, so gradients that depend on it are incorrect. If the backward pass needs the returned value, for example as an index, save it to an array in the forward pass and return it from a custom replay function; see Example 2: Custom Replay Function. Even with a custom replay function, autodiff treats the returned value as a constant, so arr receives no gradient from expressions that use it.

Parameters:

  • arr – Array containing the element to update. Its scalar type must be int32, uint32, int64, uint64, float16, bfloat16, float32, or float64.

  • i – Index of the element to update.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i] before this update. For composite dtypes, the result may not equal any value that arr[i] held at a single point in time.

warp.atomic_sub(arr: FabricArray[Any], i: Int, j: Int, value: Any) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j] before this update.

warp.atomic_sub(
arr: FabricArray[Any],
i: Int,
j: Int,
k: Int,
value: Any,
) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • k – Third index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j,k] before this update.

warp.atomic_sub(
arr: FabricArray[Any],
i: Int,
j: Int,
k: Int,
l: Int,
value: Any,
) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • k – Third index of the element.

  • l – Fourth index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j,k,l] before this update.

warp.atomic_sub(arr: IndexedFabricArray[Any], i: Int, value: Any) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

For composite dtypes (vectors, matrices, quaternions, and transforms), the update is atomic only per component. Another thread can observe arr[i] with only some of its components updated.

On GPU, concurrent floating-point updates to the same element are applied in an unspecified order, so rounding can make the result differ between runs. Enable deterministic mode for reproducible results; see Deterministic Execution. In deterministic mode, compiling a kernel that uses the returned value raises WarpDeterminismError.

This function does not support float16 or bfloat16 on GPUs with compute capability below 7.0, or float64 below 6.0. Kernels compiled for those GPUs skip these updates without raising an error. Use float32 on those GPUs instead.

Currently, the backward pass does not recompute the returned value, so gradients that depend on it are incorrect. If the backward pass needs the returned value, for example as an index, save it to an array in the forward pass and return it from a custom replay function; see Example 2: Custom Replay Function. Even with a custom replay function, autodiff treats the returned value as a constant, so arr receives no gradient from expressions that use it.

Parameters:

  • arr – Array containing the element to update. Its scalar type must be int32, uint32, int64, uint64, float16, bfloat16, float32, or float64.

  • i – Index of the element to update.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i] before this update. For composite dtypes, the result may not equal any value that arr[i] held at a single point in time.

warp.atomic_sub(
arr: IndexedFabricArray[Any],
i: Int,
j: Int,
value: Any,
) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j] before this update.

warp.atomic_sub(
arr: IndexedFabricArray[Any],
i: Int,
j: Int,
k: Int,
value: Any,
) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • k – Third index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j,k] before this update.

warp.atomic_sub(
arr: IndexedFabricArray[Any],
i: Int,
j: Int,
k: Int,
l: Int,
value: Any,
) → Any
  • Kernel
  • Differentiable

Atomically subtract a value from an array element in place.

See the wp.atomic_sub(arr, i, value) overload for dtype, atomicity, determinism, and autodiff details.

Parameters:

  • arr – Array containing the element to update.

  • i – First index of the element.

  • j – Second index of the element.

  • k – Third index of the element.

  • l – Fourth index of the element.

  • value – Amount to subtract. value must have the same dtype as arr.

Returns:

The value of arr[i,j,k,l] before this update.