triton-windows 3.3.1.post19__cp312-cp312-win_amd64.whl → 3.5.0.post21__cp312-cp312-win_amd64.whl

triton/runtime/driver.py

@@ -1,60 +1,38 @@
-from ..backends import backends
-from ..backends import DriverBase
+from __future__ import annotations
 
+from ..backends import backends, DriverBase
 
-def _create_driver():
-    actives = [x.driver for x in backends.values() if x.driver.is_active()]
-    if len(actives) != 1:
-        raise RuntimeError(f"{len(actives)} active drivers ({actives}). There should only be one.")
-    return actives[0]()
 
+def _create_driver() -> DriverBase:
+    active_drivers = [x.driver for x in backends.values() if x.driver.is_active()]
+    if len(active_drivers) != 1:
+        raise RuntimeError(f"{len(active_drivers)} active drivers ({active_drivers}). There should only be one.")
+    return active_drivers[0]()
 
-class LazyProxy:
 
-    def __init__(self, init_fn):
-        self._init_fn = init_fn
-        self._obj = None
-
-    def _initialize_obj(self):
-        if self._obj is None:
-            self._obj = self._init_fn()
-
-    def __getattr__(self, name):
-        self._initialize_obj()
-        return getattr(self._obj, name)
-
-    def __setattr__(self, name, value):
-        if name in ["_init_fn", "_obj"]:
-            super().__setattr__(name, value)
-        else:
-            self._initialize_obj()
-            setattr(self._obj, name, value)
-
-    def __delattr__(self, name):
-        self._initialize_obj()
-        delattr(self._obj, name)
-
-    def __repr__(self):
-        if self._obj is None:
-            return f"<{self.__class__.__name__} for {self._init_fn} not yet initialized>"
-        return repr(self._obj)
-
-    def __str__(self):
-        self._initialize_obj()
-        return str(self._obj)
+class DriverConfig:
 
+    def __init__(self) -> None:
+        self._default: DriverBase | None = None
+        self._active: DriverBase | None = None
 
-class DriverConfig:
+    @property
+    def default(self) -> DriverBase:
+        if self._default is None:
+            self._default = _create_driver()
+        return self._default
 
-    def __init__(self):
-        self.default = LazyProxy(_create_driver)
-        self.active = self.default
+    @property
+    def active(self) -> DriverBase:
+        if self._active is None:
+            self._active = self.default
+        return self._active
 
-    def set_active(self, driver: DriverBase):
-        self.active = driver
+    def set_active(self, driver: DriverBase) -> None:
+        self._active = driver
 
-    def reset_active(self):
-        self.active = self.default
+    def reset_active(self) -> None:
+        self._active = self.default
 
 
 driver = DriverConfig()

triton/runtime/interpreter.py

@@ -1,32 +1,36 @@
+from __future__ import annotations
 import ast
 import textwrap
 import inspect
-from typing import Tuple, List
+from typing import Tuple, List, Dict, Callable
 
 import math
 import numpy as np
 
 import triton
 import triton.language as tl
+import dataclasses
 from dataclasses import dataclass
+
+from triton.language.semantic import TritonSemantic
+from triton.tools.tensor_descriptor import TensorDescriptor
 from .errors import InterpreterError
 from functools import partial
 from .._C.libtriton import interpreter as _interpreter
 from .._C.libtriton import ir as _ir
 
 
+@dataclass
 class TensorHandle:
-
-    def __init__(self, data, dtype):
-        '''
-            data: numpy array
-            dtype: triton type, either pointer_type or scalar_type.
-            we don't store block_type here because the shape information is already available in the data field
-            attr: a dictionary of attributes
-        '''
-        self.data = data
-        self.dtype = dtype
-        self.attr = {}
+    '''
+        data: numpy array
+        dtype: triton type, either pointer_type or scalar_type.
+        we don't store block_type here because the shape information is already available in the data field
+        attr: a dictionary of attributes
+    '''
+    data: np.array
+    dtype: tl.dtype
+    attr: Dict = dataclasses.field(default_factory=dict)
 
     def __bool__(self):
         return bool(self.data.all())
@@ -73,17 +77,19 @@ class BlockPointerHandle:
 class TensorDescHandle:
 
     def __init__(self, base: TensorHandle, shape: List[TensorHandle], strides: List[TensorHandle],
-                 block_shape: List[int]):
+                 block_shape: List[int], padding):
         self.base = base
         self.ndim = len(shape)
         self.shape = shape
         self.strides = strides
         self.block_shape = block_shape
+        self.padding = padding
 
     def validate(self):
         assert self.base.data.item() % 16 == 0, "base must be 16-byte aligned"
         assert len(self.strides) == self.ndim
         assert len(self.block_shape) == self.ndim
+        assert self.ndim >= 1, "descriptor cannot be 0 dimensional"
 
         for stride in self.strides[:-1]:
             assert stride.data.item() % 16 == 0, "stride must be 16-byte aligned"
@@ -103,6 +109,7 @@ class TensorDescHandle:
             off = (offsets[dim].data + np.arange(self.block_shape[dim])).reshape(bcast_dims)
             ptrs = ptrs + (itemsize * off * self.strides[dim].data).astype(np.uint64)
             masks = masks & (0 <= off) & (off < self.shape[dim].data)
+        assert ptrs.dtype == np.uint64
         ptrs = TensorHandle(ptrs, self.base.dtype.scalar)
         return ptrs, masks
 
@@ -114,7 +121,7 @@ class InterpreterOptions:
     sanitize_overflow: bool = True
     arch: str = None
     supported_fp8_dtypes: Tuple[str] = ("fp8e5", "fp8e5b16", "fp8e4nv", "fp8e4b8", "fp8e4b15")
-    deprecated_fp8_dtypes: Tuple[str] = ()
+    deprecated_fp8_dot_operand_dtypes: Tuple[str] = ()
     default_dot_input_precision: str = "tf32"
     allowed_dot_input_precisions: Tuple[str] = ("tf32", "tf32x3", "ieee")
     max_num_imprecise_acc_default: int = 0
@@ -248,8 +255,8 @@ np_umulhi_u64 = np.vectorize(_umulhi_64, otypes=[np.uint64])
 class ExtraFunctions:
 
     @staticmethod
-    def _convert_custom_types(input, dst_ty, fp_downcast_rounding, _builder):
-        return tl.tensor(_builder.create_fp_to_fp(input.handle, dst_ty, fp_downcast_rounding), dst_ty)
+    def _convert_custom_types(input, dst_ty, fp_downcast_rounding, _semantic):
+        return tl.tensor(_semantic.builder.create_fp_to_fp(input.handle, dst_ty, fp_downcast_rounding), dst_ty)
 
 
 class InterpreterBuilder:
@@ -306,6 +313,9 @@ class InterpreterBuilder:
     def get_double_ty(self):
         return tl.float64
 
+    def get_int1_ty(self):
+        return tl.int1
+
     def get_int8_ty(self):
         return tl.int8
 
@@ -587,11 +597,18 @@ class InterpreterBuilder:
             b_data = _convert_float(b_data, b.dtype, tl.float16, None).view(np.float16)
         return TensorHandle(np.matmul(a_data, b_data, dtype=d.data.dtype) + d.data, d.dtype.scalar)
 
-    def create_make_range(self, start, stop):
+    def create_make_range(self, ret_ty, start, stop):
         return TensorHandle(np.arange(start, stop, dtype=np.int32), tl.int32)
 
-    def create_histogram(self, data, bins):
-        return TensorHandle(np.histogram(data.data, bins=bins, range=(0, bins))[0], tl.int32)
+    def create_histogram(self, data, bins, mask):
+        if mask is None:
+            mask = TensorHandle(np.ones_like(data.data, dtype=bool), tl.int1)
+        # force all masked elements to zero
+        data = np.where(mask.data, data.data, np.zeros_like(data.data))
+        histogram = np.histogram(data, bins=bins, range=(0, bins))[0]
+        # remove overcounted elements
+        histogram[0] -= np.logical_not(mask.data).sum()
+        return TensorHandle(histogram, tl.int32)
 
     def create_gather(self, src, indices, axis):
         return TensorHandle(np.take_along_axis(src.data, indices.data, axis=axis), src.dtype.scalar)
@@ -641,12 +658,16 @@ class InterpreterBuilder:
         # Triton only supports splitting the original tensor into two along the last axis
         return (TensorHandle(val.data[..., 0], val.dtype.scalar), TensorHandle(val.data[..., 1], val.dtype.scalar))
 
-    def create_splat(self, arg, shape):
+    def create_splat(self, ret_ty, arg):
+        shape = ret_ty.shape
         if isinstance(arg.dtype, tl.block_type):
             return TensorHandle(np.full(shape, arg.data[0], dtype=_get_np_dtype(arg.dtype)), arg.dtype.scalar)
         else:  # scalar
             return TensorHandle(np.full(shape, arg.data, dtype=_get_np_dtype(arg.dtype)), arg.dtype.scalar)
 
+    def create_unsplat(self, arg):
+        return TensorHandle(np.full((1, ), arg.data[0], dtype=_get_np_dtype(arg.dtype)), arg.dtype.scalar)
+
     def create_atomic_cas(self, ptr, cmp, val, sem, scope):
         if sem not in self.ir_sem_to_interpreter_sem:
             raise ValueError(f"unsupported semantic {sem}")
@@ -709,14 +730,9 @@ class InterpreterBuilder:
             ret.offsets[i].data += offsets[i].data
         return ret
 
-    def create_make_tensor_descriptor(
-        self,
-        base: TensorHandle,
-        shape: List[TensorHandle],
-        strides: List[TensorHandle],
-        tensor_shape: List[int],
-    ):
-        desc = TensorDescHandle(base, shape, strides, tensor_shape)
+    def create_make_tensor_descriptor(self, base: TensorHandle, shape: List[TensorHandle], strides: List[TensorHandle],
+                                      tensor_shape: List[int], is_signed: bool, padding: str = "zero"):
+        desc = TensorDescHandle(base, shape, strides, tensor_shape, padding)
         desc.validate()
         return desc
 
@@ -724,7 +740,16 @@ class InterpreterBuilder:
                                eviction_policy):
         assert isinstance(desc, TensorDescHandle)
         ptrs, mask = desc.materialize_pointers(indices)
-        return self.create_masked_load(ptrs, mask, other=None, cache_modifier=cache_modifier,
+        dtype_tt = ptrs.get_element_ty()
+        dtype_np = _get_np_dtype(dtype_tt)
+        padding = desc.padding
+        if padding == _ir.PADDING_OPTION.PAD_ZERO:
+            other = TensorHandle(np.zeros_like(ptrs.data, dtype=dtype_np), dtype_tt)
+        elif padding == _ir.PADDING_OPTION.PAD_NAN:
+            other = TensorHandle(np.full_like(ptrs.data, float('nan'), dtype=dtype_np), dtype_tt)
+        else:
+            raise ValueError(f"unsupported padding {padding}")
+        return self.create_masked_load(ptrs, mask, other, cache_modifier=cache_modifier,
                                        eviction_policy=eviction_policy, is_volatile=False)
 
     def create_descriptor_store(self, desc: TensorDescHandle, value: TensorHandle, indices: List[TensorHandle]):
@@ -753,15 +778,18 @@ class InterpreterBuilder:
         np_type = _get_np_dtype(type)
         if "int" in np_type.name:
             return TensorHandle(np.full(1, -1, dtype=np_type), type.scalar)
+        elif np_type == np.bool_:
+            return TensorHandle(np.full(1, True, dtype=np_type), type.scalar)
         else:
             raise TypeError(f"unsupported type {type}")
 
 
 def _patch_attr(obj, name, member, builder):
+    semantic = TritonSemantic(builder)
     new_member = lambda *args, member=member, **kwargs: (member(*args, **
                                                                 {k: v
                                                                  for k, v in kwargs.items()
-                                                                 if k != "_builder"}, _builder=builder))
+                                                                 if k != "_semantic"}, _semantic=semantic))
     setattr(obj, name, new_member)
 
 
@@ -822,12 +850,10 @@ class ReduceScanOpInterface:
 
     def apply(self, input):
         if not isinstance(input, tuple):
-            input = (input, )
+            return self.apply((input, ))[0]
         self.check_tensor(input)
-        return self.apply_impl(input)
-
-    def apply_impl(self, input):
-        raise NotImplementedError("apply_impl not implemented")
+        ret = self.apply_impl(input)
+        return tuple(ret) if isinstance(ret, (list, tuple)) else (ret, )
 
 
 class ReduceOps(ReduceScanOpInterface):
@@ -887,7 +913,7 @@ class ReduceOps(ReduceScanOpInterface):
                 # Take a scalar
                 data = data.item()
             ret.append(self.to_tensor(data, input[i].dtype))
-        return ret[0] if len(ret) == 1 else tuple(ret)
+        return ret
 
     def min_max(self, input, val_reduce_op, idx_reduce_op=None):
         # If input is a tuple, it must be (val, index), and we only take val
@@ -916,9 +942,9 @@ class ReduceOps(ReduceScanOpInterface):
         elif self.combine_fn == tl.standard._argmax_combine_tie_break_left:
             return self.min_max(input[0], val_reduce_op=np.max, idx_reduce_op=np.argmax)
         elif self.combine_fn == tl.standard._elementwise_max:
-            return self.min_max(input[0], val_reduce_op=np.max, idx_reduce_op=None)
+            return self.min_max(input[0], val_reduce_op=np.nanmax, idx_reduce_op=None)
         elif self.combine_fn == tl.standard._elementwise_min:
-            return self.min_max(input[0], val_reduce_op=np.min, idx_reduce_op=None)
+            return self.min_max(input[0], val_reduce_op=np.nanmin, idx_reduce_op=None)
         elif self.combine_fn == tl.standard._sum_combine:
             return self.sum(input[0])
         else:
@@ -985,7 +1011,7 @@ class ScanOps(ReduceScanOpInterface):
         if self.reverse:
             for arg in ret:
                 arg.handle.data = np.flip(arg.handle.data, axis=self.axis)
-        return len(ret) == 1 and ret[0] or tuple(ret)
+        return ret
 
 
 def _patch_reduce_scan():
@@ -1092,7 +1118,7 @@ def _patch_lang(fn):
             _patch_builtin(lang.math, interpreter_builder)
         _patch_lang_tensor(lang.tensor)
         _patch_lang_core(lang)
-    _patch_builtin(tl.core._experimental_tensor_descriptor_base, interpreter_builder)
+    _patch_builtin(tl.core.tensor_descriptor_base, interpreter_builder)
 
 
 def _tuple_create(arg, contents):
@@ -1107,7 +1133,7 @@ def _tuple_create(arg, contents):
 # TODO: wrap everything in triton tensors
 def _implicit_cvt(arg):
     if isinstance(arg, int):
-        ty = tl.str_to_ty(triton.runtime.jit.mangle_type(arg))
+        ty = tl.str_to_ty(triton.runtime.jit.mangle_type(arg), None)
         dtype = np.int32
         if -2**31 <= arg < 2**31:
             dtype = np.int32
@@ -1122,15 +1148,25 @@ def _implicit_cvt(arg):
         handle = TensorHandle(np.array([arg], dtype=dtype), ty)
         return tl.tensor(handle, ty)
     if hasattr(arg, "data_ptr"):
-        ty = tl.str_to_ty(triton.runtime.jit.mangle_type(arg))
+        ty = tl.str_to_ty(triton.runtime.jit.mangle_type(arg), None)
         handle = TensorHandle(np.array([arg.data_ptr()], dtype=np.uint64), ty)
         return tl.tensor(handle, ty)
     elif isinstance(arg, tuple):
         return _tuple_create(arg, map(_implicit_cvt, arg))
+    elif isinstance(arg, TensorDescriptor):
+        strides = [_implicit_cvt(s) for s in arg.strides]
+        assert arg.strides[-1] == 1
+        strides[-1] = tl.constexpr(1)
+        semantic = TritonSemantic(InterpreterBuilder())
+        return semantic.make_tensor_descriptor(base=_implicit_cvt(arg.base),
+                                               shape=[_implicit_cvt(s) for s in arg.shape], strides=strides,
+                                               block_shape=[tl.constexpr(b)
+                                                            for b in arg.block_shape], padding_option=arg.padding)
     return arg
 
 
 interpreter_builder = InterpreterBuilder()
+interpreter_semantic = TritonSemantic(interpreter_builder)
 
 
 def _unwrap_tensor(t):
@@ -1162,6 +1198,14 @@ class GridExecutor:
         def _to_cpu(arg):
             if isinstance(arg, tuple):
                 return _tuple_create(arg, map(_to_cpu, arg))
+            elif isinstance(arg, TensorDescriptor):
+                return TensorDescriptor(
+                    _to_cpu(arg.base),
+                    arg.shape,
+                    arg.strides,
+                    arg.block_shape,
+                    arg.padding,
+                )
             elif not hasattr(arg, "data_ptr"):
                 return arg
 
@@ -1195,6 +1239,8 @@ class GridExecutor:
             elif isinstance(arg_dev, tuple):
                 for (arg_dev, arg_hst) in zip(arg_dev, arg_hst):
                     _from_cpu(arg_dev, arg_hst)
+            elif isinstance(arg_dev, TensorDescriptor):
+                _from_cpu(arg_dev.base, arg_hst.base)
 
         for arg_dev, arg_hst in zip(args_dev, args_hst):
             _from_cpu(arg_dev, arg_hst)
@@ -1235,6 +1281,8 @@ class GridExecutor:
                         interpreter_builder.set_grid_idx(x, y, z)
                         self.fn(**args)
         except Exception as e:
+            if triton.knobs.compilation.front_end_debugging:
+                raise
             raise InterpreterError(repr(e)) from e
         # copy arguments back to propagate side-effects
         self._restore_args_dev(args_dev, args_hst, kwargs, kwargs_hst)
@@ -1249,14 +1297,10 @@ class ASTTransformer(ast.NodeTransformer):
         if len(names) > 1:
             raise ValueError("Multiple assignments are not supported")
         # Modify the assignment x = value to
-        # triton.language.semantic.to_tensor(value, interpreter_builder, False)
+        # interpreter_semantic.to_tensor(value, False)
         node.value = ast.Call(
-            func=ast.Attribute(
-                value=ast.Attribute(
-                    value=ast.Attribute(value=ast.Name(id='triton', ctx=ast.Load()), attr='language', ctx=ast.Load()),
-                    attr='semantic', ctx=ast.Load()), attr='to_tensor', ctx=ast.Load()),
-            args=[node.value, ast.Name(id='interpreter_builder', ctx=ast.Load()),
-                  ast.Constant(value=False)], keywords=[])
+            func=ast.Attribute(value=ast.Name(id="interpreter_semantic", ctx=ast.Load()), attr="to_tensor",
+                               ctx=ast.Load()), args=[node.value, ast.Constant(value=False)], keywords=[])
         return node
 
 
@@ -1331,11 +1375,12 @@ class FunctionRewriter:
 
 class InterpretedFunction:
     # Cache all rewritten functions
-    rewritten_fn = {}
+    rewritten_fn: Dict[Callable, Callable] = {}
 
     def __init__(self, fn, **kwargs) -> None:
         self.fn = fn
         self.rewriter = FunctionRewriter(fn, **kwargs)
+        self.kwargs = kwargs
 
         def run(*args, **kwargs):
             grid = kwargs["grid"]