Update broken docs/abstractions.py for changed ops, and add to CI (#930)

* fix and add to ci

* still have those

* ocd

* update other doc

.github/workflows/test.yml

@@ -55,6 +55,8 @@ jobs:
         python-version: 3.8
     - name: Install Dependencies
       run: pip install -e '.[testing]' --extra-index-url https://download.pytorch.org/whl/cpu
+    - name: Test Docs
+      run: python docs/abstractions.py
     - name: Run Pytest
       run: python -m pytest -s -v -n=auto test/
 

README.md

@@ -88,7 +88,7 @@ tinygrad already supports numerous accelerators, including:
 - [x] Triton
 - [x] PyTorch
 
-And it is easy to add more! Your accelerator of choice only needs to support a total of 20 (optionally 21) low level ops.
+And it is easy to add more! Your accelerator of choice only needs to support a total of 26 (optionally 27) low level ops.
 More information can be found in the [documentation for adding new accelerators](/docs/adding_new_accelerators.md).
 
 ## Installation

docs/abstractions.py

@@ -98,16 +98,17 @@ class LazyOp:
   src: Tuple[Union[LazyOp, LazyBuffer], ...]   # the sources
   arg: Optional[Any] = None                    # and an optional static argument
 
-# there's currently 20 Ops you have to implement for an accelerator.
-class UnaryOps(Enum):    NOOP = auto(); EXP = auto(); LOG = auto(); NEG = auto(); NOT = auto()
-class BinaryOps(Enum):   ADD = auto();  SUB = auto(); MUL = auto(); DIV = auto(); POW = auto(); CMPEQ = auto(); MAX = auto()
+# there's currently 27 Ops you have to implement for an accelerator.
+class UnaryOps(Enum):    NOOP = auto(); EXP = auto(); LOG = auto(); CAST = auto(); SIN = auto()
+class BinaryOps(Enum):   ADD = auto();  SUB = auto(); MUL = auto(); DIV = auto();  POW = auto(); CMPEQ = auto(); MAX = auto()
 class ReduceOps(Enum):   SUM = auto();  MAX = auto()
 class MovementOps(Enum): RESHAPE = auto(); PERMUTE = auto(); EXPAND = auto(); PAD = auto(); SHRINK = auto(); STRIDE = auto()
-class LoadOps(Enum):     FROMCPU = auto()
+class FusedOps(Enum):    MULACC = auto()
+class LoadOps(Enum):     EMPTY = auto(); RAND = auto(); CONST = auto(); FROM = auto(); CONTIGUOUS = auto(); CUSTOM = auto()
 # NOTE: if you have a CompiledBuffer(DeviceBuffer)
 #       you do not need to implement the MovementOps
 #       as they are handled by the ShapeTracker(in tinygrad/shape/shapetracker.py, code 7/10)
-Op = Union[UnaryOps, BinaryOps, ReduceOps, MovementOps, LoadOps]
+Op = Union[UnaryOps, BinaryOps, ReduceOps, MovementOps, FusedOps, LoadOps]
 
 # most of tinygrad/lazy.py is concerned with fusing Ops into LazyOps ASTs that map to GPUKernels
 # it's beyond the scope of this tutorial, but you can read the file if interested
@@ -130,11 +131,11 @@ assert lazyop.op == BinaryOps.ADD
 assert len(lazyop.src) == 2
 
 # the first source is the 2, it comes from the CPU
-# the source is a LazyBuffer, since FROMCPU cannot be folded into LazyOp ASTs
+# the source is a LazyBuffer, holding the data as an ndarray
 # again, a LazyOp AST is like a GPU kernel. you have to copy the data on the device first
 print(lazyop.src[0].op)
-assert lazyop.src[0].op.op == LoadOps.FROMCPU
-assert lazyop.src[0].op.arg == [2], "the arg of the FROMCPU LazyOP is the [2.]"
+assert lazyop.src[0].op.op == LoadOps.FROM
+assert lazyop.src[0].op.src[0].realized.toCPU()[0] == 2, "the arg of the FROM LazyOP is a LazyBuffer holding [2.]"
 assert result.lazydata.realized is None, "the LazyBuffer is not realized yet"
 
 # now we realize the LazyBuffer

docs/adding_new_accelerators.md

@@ -1,16 +1,17 @@
 # Adding a new accelerator to tinygrad
 
-It's pretty easy to add a new accelerator to tinygrad. All you need to do is implement a total of 20 (optionally 21) low level ops. Then tinygrad takes care of the rest, handling derivatives and syntactic sugar.
+It's pretty easy to add a new accelerator to tinygrad. All you need to do is implement a total of 26 (optionally 27) low level ops. Then tinygrad takes care of the rest, handling derivatives and syntactic sugar.
 
 ## llops
 
-These are the ops that you must implement for your accelerator of choice.
+These are the ops that you must implement for your accelerator of choice. Compiled Accelerators do not need to implement movement_ops, as they are handled b the ShapeTracker.
 ```
 Buffer                                                       # class of memory on this device
 unary_op  (NOOP, EXP, LOG, CAST, SIN)                        # A -> A
 reduce_op (SUM, MAX)                                         # A -> B (smaller size, B has 1 in shape)
 binary_op (ADD, SUB, MUL, DIV, POW, CMPEQ, MAX)              # A + A -> A (all the same size)
 movement_op (EXPAND, RESHAPE, PERMUTE, PAD, SHRINK, STRIDE)  # A -> B (different size)
+load_op   (EMPTY, RAND, CONST, FROM, CONTIGUOUS, CUSTOM)     # -> A   (initialize data on device)
 fused_op [[optional]] (MULACC)                               # A * A -> B
 ```