start uop docs (#6291)

* start uop docs

* only need show_labels

* sink comes first

* hotfix: invalid

* touchups

* 2 space indent works

* limit some buffer uops

* better BARRIER doc, Op -> UOp when it makes sense.

* make KernelInfo optional

* more work

relative links don't work

* this can be local in multi reduce+pads

* add UOps.SHAPETRACKER details

* UOps.CONST both types

* nit: local buffer isn't device Buffer, habit

* nit2: dtype -> DType

docs/developer/uop.md

@@ -0,0 +1,11 @@
+::: tinygrad.ops.UOp
+    options:
+        members: false
+        members_order: source
+        show_labels: false
+
+::: tinygrad.ops.UOps
+    options:
+        members: true
+        members_order: source
+        show_labels: false

mkdocs.yml

@@ -23,6 +23,7 @@ nav:
   - Developer:
     - Intro: developer/developer.md
     - Function (autodiff): developer/function.md
+    - UOp: developer/uop.md
     - Runtime:
       - developer/runtime.md
       - HCQ: developer/hcq.md

tinygrad/ops.py

@@ -40,20 +40,165 @@ def identity_element(op:BinaryOps, dt:DType): return dtypes.as_const({BinaryOps.
 
 # the order of these UOps controls the order of the toposort
 class UOps(Enum):
-  # ops that aren't rendered
+  # uops that aren't rendered
-  SINK = auto(); EXT = auto(); EXPAND = auto(); CONTRACT = auto(); SHAPETRACKER = auto(); SWIZZLE = auto()  # noqa: E702
+  SINK = auto()
-  DEFINE_GLOBAL = auto(); DEFINE_VAR = auto(); DEFINE_LOCAL = auto(); DEFINE_ACC = auto() # noqa: E702
+  """
-  CONST = auto(); SPECIAL = auto() # noqa: E702
+  Holds `UOps.STORE`. SINK defines the AST for a Kernel.
-  NOOP = auto(); GEP = auto() # noqa: E702
+
+  - **`dtype`**: `None`
+  - **`src`**: `Tuple[UOp, ...]`, Only local STOREs are allowed.
+  - **`arg`**: `Optional[KernelInfo]`
+
+  NOTE: `ScheduleItem` ASTs do not have the `KernelInfo` arg, `Kernel` inserts this to the SINK later.
+  """
+  EXT = auto()
+  """
+  Holds a single MetaOp. EXT UOps do not need a Kernel.
+
+  - **`dtype`**: Output DType
+  - **`src`**: `Tuple[]`
+  - **`arg`**: (`MetaOps.CUSTOM | MetaOps.COPY | MetaOps.EMPTY | MetaOps.VIEW`, LazyBuffer arg)
+  """
+  EXPAND = auto()
+  CONTRACT = auto()
+  SHAPETRACKER = auto()
+  """
+  Defines the ShapeTracker for a buffer UOp `UOps.LOAD`, `UOps.STORE` or `UOps.CONST`.
+
+  - **`dtype`**: `None`
+  - **`src`**: `Tuple[]`
+  - **`arg`**: `ShapeTracker`
+  """
+  SWIZZLE = auto()
+  DEFINE_GLOBAL = auto()
+  DEFINE_VAR = auto()
+  DEFINE_LOCAL = auto()
+  DEFINE_ACC = auto()
+  CONST = auto()
+  """
+  Defines a single scalar constant value.
+
+  - **`dtype`**: The scalar DType of the value.
+
+  - **`src`**:
+    The scheduler creates a CONST with a single SHAPETRACKER UOp src: `Tuple[UOp]`.
+
+    The Lowerer replaces the SHAPETRACKER with an empty src.
+    It uses the ShapeTracker valid to create a `WHERE` UOp mask with sources: `(The actual CONST UOp, CONST 0, 0.0 or False)`
+
+  - **`arg`**: The value.
+  """
+  SPECIAL = auto()
+  NOOP = auto()
+  GEP = auto()
   # math ops
-  CAST = auto(); BITCAST = auto(); VECTORIZE = auto() # noqa: E702
+  CAST = auto()
-  ALU = auto(); REDUCE = auto(); REDUCE_AXIS = auto(); WMMA = auto() # noqa: E702
+  """
+  - **`dtype`**: The casted scalar DType
+  - **`src`**: `Tuple[UOp]`
+  - **`arg`**: `None`
+  """
+  BITCAST = auto()
+  """
+  - **`dtype`**: The bitcasted scalar DType
+  - **`src`**: `Tuple[UOp]`
+  - **`arg`**: `None`
+  """
+  VECTORIZE = auto()
+  """
+  - **`dtype`**: The upcasted vector DType
+  - **`src`**: `Tuple[UOp, ...]`
+  - **`arg`**: `None`
+
+  NOTE: Length of sources must match `dtype.count`
+  """
+  ALU = auto()
+  """
+  - **`dtype`**: Output DType
+  - **`src`**: `Tuple[UOp] | Tuple[UOp, UOp] | Tuple[UOp, UOp, UOp]`
+  - **`arg`**: `UnaryOps | BinaryOps | TernaryOps`
+  """
+  REDUCE = auto()
+  REDUCE_AXIS = auto()
+  """
+  - **`dtype`**: Output DType
+  - **`src`**: Input to reduce `Tuple[UOp]`
+  - **`arg`**: `(BinaryOps.ADD | BinaryOps.MUL | BinaryOps.MAX, Tuple[int, ...])`
+  """
+  WMMA = auto()
   # memory/assignment ops
-  LOAD = auto(); STORE = auto(); PHI = auto() # noqa: E702
+  LOAD = auto()
+  """
+  - **`dtype`**: Output DType
+  - **`src`**:
+
+    The scheduler and Kernel create LOADs with a SHAPETRACKER uop in src.
+
+    - Normal LOAD: `Tuple[UOp, UOp]`
+      - Buffer UOp `UOps.DEFINE_GLOBAL`.
+      - SHAPETRACKER UOp.
+
+    - Local LOAD: `Tuple[UOp, UOp, UOp]`
+      - Buffer UOp `UOps.DEFINE_LOCAL`.
+      - SHAPETRACKER UOp.
+      - Local UOps.STORE to the same local buffer. We will barrier this later.
+
+    The Lowerer replaces the SHAPETRACKER with an indexing uop and gates the LOAD if needed.
+
+    - Normal LOAD: `Tuple[UOp, UOp]`
+      - Buffer UOp `UOps.DEFINE_GLOBAL`.
+      - Indexing UOp, can only return `dtypes.int32`.
+    - Gated LOAD: `Tuple[UOp, UOp, UOp, UOp]`
+      - Buffer UOp `UOps.DEFINE_GLOBAL`.
+      - Indexing UOp, can only return `dtypes.int32`.
+      - Gate UOp, can only return `dtypes.bool`.
+      - Value if gate is `False`, can only be a `UOps.CONST` with arg 0, 0.0 or `False`.
+    - Barriered LOAD: `Tuple[UOp, UOp, UOp, UOp]`
+      - Buffer UOp `UOps.DEFINE_LOCAL`.
+      - Indexing UOp, can only return `dtypes.int32`.
+      - Gate UOp, can only return `dtypes.bool`.
+      - Barrier UOp `UOps.BARRIER`.
+  - **`arg`**: `None`
+  """
+  STORE = auto()
+  """
+  - **`dtype`**: `None`
+  - **`src`**:
+
+    Similar to LOAD, the scheduler and Kernel create STOREs with a SHAPETRACKER uop in src:
+
+    - Buffer UOp `UOps.DEFINE_GLOBAL` or `UOps.DEFINE_LOCAL`.
+    - SHAPETRACKER UOp.
+    - Value to store.
+
+    The Lowerer replaces the SHAPETRACKER with an indexing uop and gates the STORE if needed.
+
+    - Normal STORE: `Tuple[UOp, UOp, UOp]`
+      - Buffer UOp `UOps.DEFINE_GLOBAL` or `UOps.DEFINE_LOCAL`.
+      - Indexing Op, can only return `dtypes.int32`.
+      - Value to store.
+    - Gated STORE: `Tuple[UOp, UOp, UOp, UOp]`
+      - Buffer UOp `UOps.DEFINE_GLOBAL` or `UOps.DEFINE_LOCAL`.
+      - Indexing UOp, can only return `dtypes.int32`.
+      - Value to store.
+      - Gate UOp, can only return `dtypes.bool`.
+  - **`arg`**: `None`
+  """
+  PHI = auto()
   # control flow ops
-  BARRIER = auto(); IF = auto(); RANGE = auto() # noqa: E702
+  BARRIER = auto()
-  # these two are not graph nodes
+  """
-  ENDRANGE = auto(); ENDIF = auto() # noqa: E702
+  Inserts a warp sync between local stores and local loads.
+
+  - **`dtype`**: `None`
+  - **`src`**: `Tuple[UOp, ...]`, Only local STOREs are allowed.
+  - **`arg`**: `None`
+  """
+  IF = auto()
+  RANGE = auto()
+  # ops that are not graph nodes
+  ENDRANGE = auto()
+  ENDIF = auto()
 
 BUFFER_UOPS = {UOps.LOAD, UOps.STORE, UOps.CONST}