triton-windows 3.3.1.post19__cp312-cp312-win_amd64.whl → 3.4.0.post20__cp312-cp312-win_amd64.whl

triton/experimental/gluon/language/_core.py

@@ -0,0 +1,312 @@
+from __future__ import annotations
+from typing import TypeVar, List, TYPE_CHECKING, Tuple
+from functools import wraps
+
+if TYPE_CHECKING:
+    from triton._C.libtriton.gluon_ir import GluonOpBuilder
+    from ._semantic import GluonSemantic
+
+from ._layouts import SharedLayout, DistributedLayout
+from triton._C.libtriton import ir
+import triton.language.core as tl_core
+from triton.language.core import (
+    constexpr,
+    base_value,
+    base_type,
+    dtype,
+    block_type,  # TODO: block type with layout info
+    pointer_type,
+    void,
+    int1,
+    int8,
+    int16,
+    int32,
+    int64,
+    uint8,
+    uint16,
+    uint32,
+    uint64,
+    float8e5,
+    float8e5b16,
+    float8e4nv,
+    float8e4b8,
+    float8e4b15,
+    float16,
+    bfloat16,
+    float32,
+    float64,
+    _unwrap_if_constexpr,
+    _unwrap_shape,
+    tensor,
+    tuple,
+    tuple_type,
+)
+
+_IMPORT_FROM_TRITON: List[str] = [
+    "expand_dims",
+    "join",
+    "load",
+    "maximum",
+    "minimum",
+    "permute",
+    "program_id",
+    "reduce",
+    "reshape",
+    "split",
+    "static_assert",
+    "static_print",
+    "store",
+    "to_tensor",
+    "where",
+    "inline_asm_elementwise",
+]
+
+__all__ = [
+    "constexpr",
+    "base_value",
+    "base_type",
+    "dtype",
+    "block_type",
+    "pointer_type",
+    "tuple_type",
+    "void",
+    "int1",
+    "int8",
+    "int16",
+    "int32",
+    "int64",
+    "uint8",
+    "uint16",
+    "uint32",
+    "uint64",
+    "float8e5",
+    "float8e5b16",
+    "float8e4nv",
+    "float8e4b8",
+    "float8e4b8",
+    "float8e4b15",
+    "float16",
+    "bfloat16",
+    "float32",
+    "float64",
+    "_unwrap_if_constexpr",
+    "tensor",
+    "tuple",
+    "tuple_type",
+    "thread_barrier",
+    "arange",
+    "full",
+    "convert_layout",
+    "allocate_shared_memory",
+    "shared_memory_descriptor",
+    "warp_specialize",
+    *_IMPORT_FROM_TRITON,
+]
+
+T = TypeVar("T")
+
+# TODO: split these
+GLUON_BUILTIN = "__triton_builtin__"
+
+
+class distributed_type(block_type):
+
+    def __init__(self, element_ty: dtype, shape: List[int], layout):
+        super().__init__(element_ty, shape)
+        self.layout = layout
+        self.name = f"<{self.shape}, {self.element_ty}, {self.layout}>"
+        assert isinstance(layout, DistributedLayout)
+
+    def to_ir(self, builder: ir.builder) -> ir.type:
+        elem_ty = self.element_ty.to_ir(builder)
+        layout = self.layout._to_ir(builder)
+        return builder.get_distributed_ty(elem_ty, self.shape, layout)
+
+    def mangle(self) -> str:
+        elt = self.scalar.mangle()
+        shape = "_".join(map(str, self.shape))
+        layout = self.layout.mangle()
+        return f"{elt}S{shape}SL{layout}L"
+
+    def with_element_ty(self, scalar_ty: dtype) -> block_type:
+        return distributed_type(scalar_ty, self.shape, self.layout)
+
+
+def builtin(fn: T) -> T:
+    """Mark a function as a builtin."""
+    assert callable(fn)
+
+    @wraps(fn)
+    def wrapper(*args, **kwargs):
+        if "_semantic" not in kwargs or kwargs["_semantic"] is None:
+            raise ValueError("Did you forget to add @triton.gluon.jit ? "
+                             "(`_semantic` argument must be provided outside of JIT functions.)")
+        return fn(*args, **kwargs)
+
+    setattr(wrapper, GLUON_BUILTIN, True)
+
+    return wrapper
+
+
+class shared_memory_descriptor_type(base_type):
+
+    def __init__(self, element_ty, shape, layout, alloc_shape):
+        self.element_ty = element_ty
+        self.shape = shape
+        self.layout = layout
+        self.alloc_shape = alloc_shape
+        assert isinstance(layout, SharedLayout)
+
+    def to_ir(self, builder: GluonOpBuilder) -> None:
+        return builder.get_shared_mem_desc_ty(
+            self.element_ty.to_ir(builder),
+            self.shape,
+            self.layout._to_ir(builder),
+            self.alloc_shape,
+        )
+
+    def _unflatten_ir(self, handles: List[ir.Value], cursor: int) -> Tuple[shared_memory_descriptor, int]:
+        value = shared_memory_descriptor(handles[cursor], self.element_ty, self.shape, self.layout, self.alloc_shape)
+        return value, cursor + 1
+
+    def _flatten_ir_types(self, builder: GluonOpBuilder, out: List[ir.type]) -> None:
+        out.append(self.to_ir(builder))
+
+    def __str__(self) -> str:
+        return f"shared_memory_descriptor<{self.element_ty}, {self.shape}, {self.layout}, {self.alloc_shape}>"
+
+    def __eq__(self, other) -> bool:
+        return (type(self) is type(other) and self.shape == other.shape and self.layout == other.layout
+                and self.alloc_shape == other.alloc_shape)
+
+    def __neq__(self, other) -> bool:
+        return not (self == other)
+
+    def mangle(self) -> str:
+        shape_str = "_".join([str(s) for s in self.shape])
+        return f"MD{self.element_ty.mangle()}S{shape_str}SL{self.layout.mangle()}LAS{self.alloc_shape}ASMD"
+
+
+class shared_memory_descriptor(base_value):
+
+    def __init__(self, handle, element_ty, shape, layout, alloc_shape):
+        self.handle = handle
+        self.type = shared_memory_descriptor_type(element_ty, shape, layout, alloc_shape)
+
+    def _flatten_ir(self, handles: List[ir.value]) -> None:
+        handles.append(self.handle)
+
+    @property
+    def dtype(self):
+        return self.type.element_ty
+
+    @property
+    def shape(self):
+        return self.type.shape
+
+    @property
+    def rank(self):
+        return len(self.shape)
+
+    @property
+    def layout(self):
+        return self.type.layout
+
+    def __str__(self) -> str:
+        return str(self.type)
+
+    @builtin
+    def load(self, layout, _semantic: GluonSemantic) -> tensor:
+        layout = _unwrap_if_constexpr(layout)
+        return _semantic.shared_load(self, layout)
+
+    @builtin
+    def store(self, value, _semantic: GluonSemantic) -> None:
+        return _semantic.shared_store(self, value)
+
+    @builtin
+    def slice(self, start, length, dim=0, _semantic: GluonSemantic = None) -> shared_memory_descriptor:
+        start = _unwrap_if_constexpr(start)
+        length = _unwrap_if_constexpr(length)
+        dim = _unwrap_if_constexpr(dim)
+        return _semantic.memdesc_slice(self, start, length, dim)
+
+    @builtin
+    def index(self, index, _semantic: GluonSemantic = None) -> shared_memory_descriptor:
+        index = _unwrap_if_constexpr(index)
+        return _semantic.memdesc_index(self, index)
+
+    @builtin
+    def permute(self, order, _semantic: GluonSemantic) -> shared_memory_descriptor:
+        order = [_unwrap_if_constexpr(o) for o in order]
+        return _semantic.memdesc_trans(self, order)
+
+    @builtin
+    def reshape(self, shape, layout, _semantic: GluonSemantic) -> shared_memory_descriptor:
+        shape = [_unwrap_if_constexpr(s) for s in shape]
+        layout = _unwrap_if_constexpr(layout)
+
+        return _semantic.memdesc_reshape(self, shape, layout)
+
+    @builtin
+    def _reinterpret(self, dtype, shape, layout, _semantic: GluonSemantic = None) -> shared_memory_descriptor:
+        dtype = _unwrap_if_constexpr(dtype)
+        shape = [_unwrap_if_constexpr(s) for s in shape]
+        layout = _unwrap_if_constexpr(layout)
+
+        return _semantic.memdesc_reinterpret(self, dtype, shape, layout)
+
+    @builtin
+    def _keep_alive(self, _semantic: GluonSemantic = None) -> None:
+        return _semantic.shared_dealloc(self)
+
+
+for name in _IMPORT_FROM_TRITON:
+    fn = getattr(tl_core, name)
+    globals()[name] = builtin(fn)
+
+
+@builtin
+def arange(start, end, layout, _semantic=None):
+    start = _unwrap_if_constexpr(start)
+    end = _unwrap_if_constexpr(end)
+    layout = _unwrap_if_constexpr(layout)
+    return _semantic.arange(start, end, layout)
+
+
+@builtin
+def convert_layout(value, layout, _semantic=None):
+    layout = _unwrap_if_constexpr(layout)
+    return _semantic.convert_layout(value, layout)
+
+
+@builtin
+def full(shape, value, dtype, layout, _semantic=None):
+    shape = _unwrap_shape(shape)
+    value = _unwrap_if_constexpr(value)
+    dtype = _unwrap_if_constexpr(dtype)
+    layout = _unwrap_if_constexpr(layout)
+    return _semantic.full(shape, value, dtype, layout)
+
+
+@builtin
+def allocate_shared_memory(element_ty, shape, layout, value=None, _semantic=None):
+    element_ty = _unwrap_if_constexpr(element_ty)
+    shape = _unwrap_if_constexpr(shape)
+    shape = [_unwrap_if_constexpr(s) for s in shape]
+    layout = _unwrap_if_constexpr(layout)
+    return _semantic.allocate_shared(element_ty, shape, layout, value)
+
+
+@builtin
+def warp_specialize(args, default_partition, worker_partitions, worker_num_warps, worker_num_regs,  #
+                    _semantic=None, _generator=None):
+    worker_num_warps = [_unwrap_if_constexpr(w) for w in worker_num_warps]
+    worker_num_regs = [_unwrap_if_constexpr(r) for r in worker_num_regs]
+    return _semantic.warp_specialize(args, default_partition, worker_partitions, worker_num_warps,  #
+                                     worker_num_regs, _generator)
+
+
+@builtin
+def thread_barrier(_semantic=None):
+    return _semantic.debug_barrier()

triton/experimental/gluon/language/_layouts.py

@@ -0,0 +1,230 @@
+from dataclasses import dataclass
+from typing import List, Optional
+from triton.language.core import _unwrap_if_constexpr, _unwrap_shape
+
+__all__ = [
+    "BlockedLayout",
+    "SliceLayout",
+    "DistributedLinearLayout",
+    "NVMMASharedLayout",
+    "SwizzledSharedLayout",
+]
+
+
+def _realize_cta_layout(rank, ctas_per_cga, cta_split_num, cta_order):
+    ctas_per_cga = ctas_per_cga or [1] * rank
+    cta_split_num = cta_split_num or [1] * rank
+    cta_order = cta_order or list(reversed(range(rank)))
+    return ctas_per_cga, cta_split_num, cta_order
+
+
+class DistributedLayout:
+    pass
+
+
+@dataclass(frozen=True)
+class BlockedLayout(DistributedLayout):
+    size_per_thread: List[int]
+    threads_per_warp: List[int]
+    warps_per_cta: List[int]
+    order: List[int]
+    ctas_per_cga: Optional[List[int]] = None
+    cta_split_num: Optional[List[int]] = None
+    cta_order: Optional[List[int]] = None
+
+    def __post_init__(self):
+        super().__setattr__("size_per_thread", _unwrap_if_constexpr(self.size_per_thread))
+        super().__setattr__("threads_per_warp", _unwrap_if_constexpr(self.threads_per_warp))
+        super().__setattr__("warps_per_cta", _unwrap_if_constexpr(self.warps_per_cta))
+        super().__setattr__("order", _unwrap_if_constexpr(self.order))
+        super().__setattr__("ctas_per_cga", _unwrap_if_constexpr(self.ctas_per_cga))
+        super().__setattr__("cta_split_num", _unwrap_if_constexpr(self.cta_split_num))
+        super().__setattr__("cta_order", _unwrap_if_constexpr(self.cta_order))
+
+        rank = len(self.size_per_thread)
+        assert len(self.threads_per_warp) == rank
+        assert len(self.warps_per_cta) == rank
+        assert len(self.order) == rank
+        assert self.ctas_per_cga is None or len(self.ctas_per_cga) == rank
+        assert self.cta_split_num is None or len(self.cta_split_num) == rank
+        assert self.cta_order is None or len(self.cta_order) == rank
+
+    def _to_ir(self, builder):
+        rank = len(self.size_per_thread)
+        ctas_per_cga, cta_split_num, cta_order = _realize_cta_layout(rank, self.ctas_per_cga, self.cta_split_num,
+                                                                     self.cta_order)
+        return builder.get_blocked_layout(
+            self.size_per_thread,
+            self.threads_per_warp,
+            self.warps_per_cta,
+            self.order,
+            ctas_per_cga,
+            cta_split_num,
+            cta_order,
+        )
+
+    def mangle(self) -> str:
+
+        def stringify(x):
+            if x is None:
+                return ""
+            return "_".join(map(str, x))
+
+        size_per_thread = stringify(self.size_per_thread)
+        threads_per_warp = stringify(self.threads_per_warp)
+        warps_per_cta = stringify(self.warps_per_cta)
+        order = stringify(self.order)
+        ctas_per_cga = stringify(self.ctas_per_cga)
+        cta_split_num = stringify(self.cta_split_num)
+        cta_order = stringify(self.cta_order)
+        return f"B{size_per_thread}B{threads_per_warp}B{warps_per_cta}B{order}B{ctas_per_cga}B{cta_split_num}B{cta_order}B"
+
+
+@dataclass(frozen=True)
+class SliceLayout(DistributedLayout):
+    dim: int
+    parent: DistributedLayout
+
+    def __post_init__(self):
+        super().__setattr__("dim", _unwrap_if_constexpr(self.dim))
+        super().__setattr__("parent", _unwrap_if_constexpr(self.parent))
+
+    def _to_ir(self, builder):
+        return builder.get_slice_layout(
+            self.dim,
+            self.parent._to_ir(builder),
+        )
+
+    def mangle(self) -> str:
+        return f"SL{self.dim}_{self.parent.mangle()}SL"
+
+
+@dataclass(frozen=True)
+class DistributedLinearLayout(DistributedLayout):
+    reg_bases: List[List[int]]
+    lane_bases: List[List[int]]
+    warp_bases: List[List[int]]
+    block_bases: List[List[int]]
+    shape: List[int]
+
+    def __post_init__(self):
+        super().__setattr__("reg_bases", _unwrap_shape(self.reg_bases))
+        super().__setattr__("lane_bases", _unwrap_shape(self.lane_bases))
+        super().__setattr__("warp_bases", _unwrap_shape(self.warp_bases))
+        super().__setattr__("block_bases", _unwrap_shape(self.block_bases))
+        super().__setattr__("shape", _unwrap_shape(self.shape))
+
+        rank = len(self.shape)
+
+        for basis in self.reg_bases:
+            assert len(basis) == rank
+        for basis in self.lane_bases:
+            assert len(basis) == rank
+        for basis in self.warp_bases:
+            assert len(basis) == rank
+        for basis in self.block_bases:
+            assert len(basis) == rank
+
+    def _to_ir(self, builder):
+        return builder.get_distributed_linear_layout(self.reg_bases, self.lane_bases, self.warp_bases, self.block_bases,
+                                                     self.shape)
+
+    def mangle(self):
+        return f"DLL{self.reg_bases}_{self.lane_bases}_{self.warp_bases}_{self.block_bases}_{self.shape}DLL"
+
+
+class SharedLayout:
+    pass
+
+
+@dataclass(frozen=True)
+class NVMMASharedLayout(SharedLayout):
+    swizzle_byte_width: int
+    element_bitwidth: int
+    rank: int
+    transposed: bool = False
+    fp4_padded: bool = False
+    ctas_per_cga: Optional[List[int]] = None
+    cta_split_num: Optional[List[int]] = None
+    cta_order: Optional[List[int]] = None
+
+    def __post_init__(self):
+        super().__setattr__("swizzle_byte_width", _unwrap_if_constexpr(self.swizzle_byte_width))
+        super().__setattr__("element_bitwidth", _unwrap_if_constexpr(self.element_bitwidth))
+        super().__setattr__("rank", _unwrap_if_constexpr(self.rank))
+        super().__setattr__("transposed", _unwrap_if_constexpr(self.transposed))
+        super().__setattr__("fp4_padded", _unwrap_if_constexpr(self.fp4_padded))
+        super().__setattr__("ctas_per_cga", _unwrap_if_constexpr(self.ctas_per_cga))
+        super().__setattr__("cta_split_num", _unwrap_if_constexpr(self.cta_split_num))
+        super().__setattr__("cta_order", _unwrap_if_constexpr(self.cta_order))
+
+        assert self.element_bitwidth in [8, 16, 32, 64]
+        assert self.swizzle_byte_width in [0, 32, 64, 128]
+        rank = self.rank
+        assert self.ctas_per_cga is None or len(self.ctas_per_cga) == rank
+        assert self.cta_split_num is None or len(self.cta_split_num) == rank
+        assert self.cta_order is None or len(self.cta_order) == rank
+
+    def _to_ir(self, builder):
+        ctas_per_cga, cta_split_num, cta_order = _realize_cta_layout(self.rank, self.ctas_per_cga, self.cta_split_num,
+                                                                     self.cta_order)
+        return builder.get_nvmma_shared_layout(
+            self.swizzle_byte_width,
+            self.element_bitwidth,
+            self.transposed,
+            self.fp4_padded,
+            ctas_per_cga,
+            cta_split_num,
+            cta_order,
+        )
+
+    def mangle(self) -> str:
+        return f"NVMMA_{self.swizzle_byte_width}_{self.element_bitwidth}_{self.transposed}_{self.fp4_padded}_NVMMA"
+
+
+@dataclass(frozen=True, eq=True)
+class SwizzledSharedLayout(SharedLayout):
+    vec: int
+    per_phase: int
+    max_phase: int
+    order: List[int]
+    ctas_per_cga: Optional[List[int]] = None
+    cta_split_num: Optional[List[int]] = None
+    cta_order: Optional[List[int]] = None
+
+    def __post_init__(self):
+        super().__setattr__("vec", _unwrap_if_constexpr(self.vec))
+        super().__setattr__("per_phase", _unwrap_if_constexpr(self.per_phase))
+        super().__setattr__("max_phase", _unwrap_if_constexpr(self.max_phase))
+        super().__setattr__("order", _unwrap_if_constexpr(self.order))
+        super().__setattr__("ctas_per_cga", _unwrap_if_constexpr(self.ctas_per_cga))
+        super().__setattr__("cta_split_num", _unwrap_if_constexpr(self.cta_split_num))
+        super().__setattr__("cta_order", _unwrap_if_constexpr(self.cta_order))
+
+        rank = len(self.order)
+        assert self.ctas_per_cga is None or len(self.ctas_per_cga) == rank
+        assert self.cta_split_num is None or len(self.cta_split_num) == rank
+        assert self.cta_order is None or len(self.cta_order) == rank
+
+    def _to_ir(self, builder):
+        rank = len(self.order)
+        ctas_per_cga, cta_split_num, cta_order = _realize_cta_layout(rank, self.ctas_per_cga, self.cta_split_num,
+                                                                     self.cta_order)
+        return builder.get_swizzled_shared_layout(
+            _unwrap_if_constexpr(self.vec),
+            _unwrap_if_constexpr(self.per_phase),
+            _unwrap_if_constexpr(self.max_phase),
+            self.order,
+            ctas_per_cga,
+            cta_split_num,
+            cta_order,
+        )
+
+    def mangle(self) -> str:
+
+        def stringify(x):
+            if x is None:
+                return ""
+            return "_".join(map(str, x))
+
+        return f"SSS_{self.vec}_{self.per_phase}_{self.max_phase}_{stringify(self.order)}_{stringify(self.ctas_per_cga)}_{stringify(self.cta_split_num)}_{stringify(self.cta_order)}_SSS"

triton/experimental/gluon/language/_math.py

@@ -0,0 +1,12 @@
+# flake8: noqa
+import triton.language.math as tl_math
+from ._core import builtin
+
+__all__ = [
+    "umulhi", "exp", "exp2", "fma", "log", "log2", "cos", "rsqrt", "sin", "sqrt", "sqrt_rn", "abs", "fdiv", "div_rn",
+    "erf", "floor", "ceil"
+]
+
+for name in __all__:
+    fn = getattr(tl_math, name)
+    globals()[name] = builtin(fn)