Add unsigned integer (uint) type support to PyTorch operators by updating AT_DISPATCH macros...
This skill helps add support for unsigned integer types (uint16, uint32, uint64) to PyTorch operators by updating their AT_DISPATCH macros.
Use this skill when:
Add unsigned types to existing dispatch:
// Before
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_ALL_TYPES));
// After (method 1: add unsigned types explicitly)
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES));
// After (method 2: use V2 integral types if AT_INTEGRAL_TYPES present)
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_INTEGRAL_TYPES_V2), AT_EXPAND(AT_FLOATING_TYPES));
Unsigned type groups:
AT_BAREBONES_UNSIGNED_TYPES: kUInt16, kUInt32, kUInt64AT_INTEGRAL_TYPES_V2: AT_INTEGRAL_TYPES + AT_BAREBONES_UNSIGNED_TYPESRelationship:
AT_INTEGRAL_TYPES // kByte, kChar, kInt, kLong, kShort
AT_BAREBONES_UNSIGNED_TYPES // kUInt16, kUInt32, kUInt64
AT_INTEGRAL_TYPES_V2 // INTEGRAL_TYPES + BAREBONES_UNSIGNED_TYPES
Check if the file uses AT_DISPATCH_V2:
If using old AT_DISPATCH:
If already using AT_DISPATCH_V2:
Identify what type groups are currently in use:
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
// body
}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBFloat16);
^^^^^^^^^^^^^^^^^^^^^^^^^
Current type coverage
Common patterns:
AT_EXPAND(AT_ALL_TYPES) → includes AT_INTEGRAL_TYPES + AT_FLOATING_TYPESAT_EXPAND(AT_INTEGRAL_TYPES) → signed integers onlyAT_EXPAND(AT_FLOATING_TYPES) → floating point typesTwo approaches:
Method 1: Add AT_BAREBONES_UNSIGNED_TYPES explicitly
AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES) to the type listMethod 2: Substitute AT_INTEGRAL_TYPES with AT_INTEGRAL_TYPES_V2
AT_EXPAND(AT_INTEGRAL_TYPES)Method 1 example:
// Before
AT_DISPATCH_V2(
dtype,
"min_values_cuda",
AT_WRAP([&]() {
kernel_impl<scalar_t>(iter);
}),
AT_EXPAND(AT_ALL_TYPES),
kBFloat16, kHalf, kBool
);
// After (add unsigned types)
AT_DISPATCH_V2(
dtype,
"min_values_cuda",
AT_WRAP([&]() {
kernel_impl<scalar_t>(iter);
}),
AT_EXPAND(AT_ALL_TYPES),
AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES),
kBFloat16, kHalf, kBool
);
Method 2 example:
// Before
AT_DISPATCH_V2(
dtype,
"integral_op",
AT_WRAP([&]() {
kernel<scalar_t>();
}),
AT_EXPAND(AT_INTEGRAL_TYPES)
);
// After (substitute with V2)
AT_DISPATCH_V2(
dtype,
"integral_op",
AT_WRAP([&]() {
kernel<scalar_t>();
}),
AT_EXPAND(AT_INTEGRAL_TYPES_V2)
);
If the dispatch uses AT_EXPAND(AT_ALL_TYPES):
AT_ALL_TYPES = AT_INTEGRAL_TYPES + AT_FLOATING_TYPESAT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES) to the listIf the dispatch separately lists INTEGRAL and FLOATING:
// Before
AT_EXPAND(AT_INTEGRAL_TYPES), AT_EXPAND(AT_FLOATING_TYPES)
// After (Method 2 preferred)
AT_EXPAND(AT_INTEGRAL_TYPES_V2), AT_EXPAND(AT_FLOATING_TYPES)
Check the file for ALL dispatch macros that need uint support:
Check that:
AT_EXPAND()// Before
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBFloat16);
// After
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kHalf, kBFloat16);
// Before
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_INTEGRAL_TYPES), AT_EXPAND(AT_FLOATING_TYPES));
// After
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_INTEGRAL_TYPES_V2), AT_EXPAND(AT_FLOATING_TYPES));
// Before (needs v2 conversion first)
AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, dtype, "op", [&]() {
kernel<scalar_t>();
});
// After v2 conversion
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBFloat16);
// After adding uint support
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kHalf, kBFloat16);
For a file with multiple functions:
void min_values_kernel_cuda(TensorIterator& iter) {
AT_DISPATCH_V2(iter.dtype(), "min_values_cuda", AT_WRAP([&]() {
impl<scalar_t>(iter);
}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf);
// ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
// Added uint support
}
void min_launch_kernel(TensorIterator &iter) {
AT_DISPATCH_V2(iter.input_dtype(), "min_cuda", AT_WRAP([&]() {
gpu_reduce_kernel<scalar_t>(iter);
}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf);
// ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
// Added uint support here too
}
Use this decision tree to determine the approach:
Is the file using AT_DISPATCH_V2?
├─ No → Use at-dispatch-v2 skill first, then continue
└─ Yes
└─ Does it use AT_EXPAND(AT_INTEGRAL_TYPES)?
├─ Yes → Replace with AT_EXPAND(AT_INTEGRAL_TYPES_V2)
└─ No → Add AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES) to type list
If the operator only supports floating point types, don't add uint support:
// Leave as-is - floating point only operator
AT_DISPATCH_V2(dtype, "float_op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_FLOATING_TYPES), kHalf);
Unsigned types work alongside complex types:
AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
kernel<scalar_t>();
}), AT_EXPAND(AT_ALL_TYPES),
AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES),
AT_EXPAND(AT_COMPLEX_TYPES),
kHalf, kBFloat16);
Check if uint types are already present:
AT_INTEGRAL_TYPES_V2 is used → already has uint supportAT_BAREBONES_UNSIGNED_TYPES is already in list → already has uint supportWhen asked to add uint support:
After adding uint support, the operator should accept uint16, uint32, and uint64 tensors. The user is responsible for functional testing.