[Hybrid Script] Unify the symbol tables to one; support `tvm.containe…

…r.Array` (apache#2366)

docs/langref/hybrid_script.rst

@@ -52,20 +52,23 @@ The current parse interface looks like:
    parser = tvm.hybrid.parse(outer_product, [a, b]) # return the parser of this function
 
 
-If we pass these tvm tensors to this function, it returns a op node:
+If we pass these tvm data structures, like ``Tensor``, ``Var``, ``Expr.*Imm``,
+or ``tvm.container.Array``, to this function, it returns a op node:
 
 .. code-block:: python
 
    a = tvm.placeholder((100, ), name='a')
    b = tvm.placeholder((99, ), name='b')
    c = outer_product(a, b, c) # return the output tensor(s) of the operator
 
-**Under construction, we are still deciding what kind of node should be returned.**
+You can use any methods that can be applied on a TVM ``OpNode``, like create_schedule, although
+so far, the functionality of schedule is as limited as ``ExternOpNode``. At least, it can be built
+to LLVM module.
 
 Tuning
 ~~~~~~
 
-**Under construction, not truly supported yet.**
+**Under construction, not supported yet.**
 
 Follow up the example above, you can use some tvm like interfaces to tune the code: 
 
@@ -86,6 +89,21 @@ Here we use ``range`` aka ``serial``, ``unroll``, ``parallel``, and ``vectorize`
 these **4** keywords to annotate the corresponding types of for loops.
 The the usage is roughly the same as Python standard ``range``.
 
+Besides all the loop types supported in Halide, ``const_range`` is supported for some specific conditions.
+Sometimes, ``tvm.container.Array`` is desired to pass as an argument, but in TVM-HalideIR, there is no
+such support that converts ``tvm.container.Array`` to an ``Expr``. Thus, a limited feature is supported.
+Users can access containers by either constants or constants loops annotated.
+
+.. code-block:: python
+
+   @tvm.hybrid.script
+   def foo(a, b): # b is a tvm.container.Array
+       c = output_tensor(a.shape, a.dtype)
+       for i in const_range(len(a)): # because you have b access, i should be explicitly annotated as const_range
+           c[i] = a[i] + b[i]
+       return c
+
+
 Variables
 ~~~~~~~~~
 
@@ -111,14 +129,14 @@ It regards the first store of a variable as its declaration.
      	  s += a[i, j] # do something with sum
        b[i] = sum # you can still use sum in this level
    a[0] = s # you CANNOT use s here, even though it is allowed in conventional Python
-   b = (1, 2) # this has NOT been supported yet!
 
 
 Attributes
 ~~~~~~~~~~
 
-So far, ONLY tensors' ``shape`` attribute is supported! The ``shape`` atrribute is essentailly a
-tuple, so you MUST access it as an array. Also, currently, only constant-indexed access is supported.
+So far, ONLY tensors' ``shape`` and ``dtype`` attribute are supported!
+The ``shape`` atrribute is essentailly a tuple, so you MUST access it as an array.
+Currently, only constant-indexed access is supported.
 
 .. code-block:: python
 
@@ -133,8 +151,11 @@ Conditional Statement and Expression
 
 .. code-block:: python
 
-   if condition:
-        # do something
+   if condition1 and condition2 and condition3:
+       # do something
+   else:
+       # do something else
+   # Select
    a = b if condition else c
 
 However, NO ``True`` and ``False`` keyword supported yet.
@@ -153,7 +174,9 @@ Array Allocation
 **Under construction, this function will be supported later!**
 
 Use a function call ``allocation(shape, type, share/local)`` to declare an array buffer.
-The basic usage is roughly the same as a normal array.
+The basic usage is roughly the same as a normal ``numpy.array``, and you should access
+high-dim array in ``a[i, j, k]`` fashion instead of ``a[i][j][k]``,
+even for ``tvm.container.Array`` for compilation.
 
 
 Thread Bind
@@ -170,5 +193,5 @@ You can also do loop-thread bind by writing code like this:
 
 Keywords
 ~~~~~~~~
-- For keywords: ``serial``, ``range``, ``unroll``, ``parallel``, ``vectorize``, ``bind``
+- For keywords: ``serial``, ``range``, ``unroll``, ``parallel``, ``vectorize``, ``bind``, ``const_expr``
 - Math keywords: ``log``, ``exp``, ``sigmoid``, ``tanh``, ``power``, ``popcount``

python/tvm/hybrid/calls.py

@@ -12,15 +12,17 @@
 #pylint: disable=redefined-builtin
 
 LOOP_INTRIN = {
-    'range'    : For.Serial,
-    'unroll'   : For.Unrolled,
-    'parallel' : For.Parallel,
-    'vectorize': For.Vectorized,
+    'range'       : For.Serial,
+    'unroll'      : For.Unrolled,
+    'parallel'    : For.Parallel,
+    'vectorize'   : For.Vectorized,
+    'const_range' : (For.Unrolled, ),
 }
 
+
 def _range(annotation, args):
     """Handling TVM loop types"""
-    n = len(args)
+    n = args.__len__()
     if n == 1:
         low, ext = _api.const(0, dtype='int32'), args[0]
     else:
@@ -33,13 +35,13 @@ def _range(annotation, args):
     return iter_var, low, ext, for_type
 
 
-range = unroll = vectorize = parallel = _range #pylint: disable=invalid-name
+range = unroll = vectorize = parallel = const_range = _range #pylint: disable=invalid-name
 
 
 def bind(func_id, args):
     """Handling TVM thread binding"""
     _internal_assert(func_id == "bind", "This function cannot be directly invoked!")
-    _internal_assert(len(args) == 2, "A loop bind should only have 2 arguments!")
+    _internal_assert(args.__len__() == 2, "A loop bind should only have 2 arguments!")
     _internal_assert(isinstance(args[0], str), \
                      "A loop bind's first argument should be a string!")
     iter_var = _api.thread_axis(args[0])
@@ -56,7 +58,7 @@ def _math_intrin(func_id, args):
 
 
 def _min_max(func_id, args):
-    _internal_assert(len(args) == 2, "Max/Min function should have 2 elements")
+    _internal_assert(args.__len__() == 2, "Max/Min function should have 2 elements")
     return getattr(_make, func_id.title())(args[0], args[1])
 
 
@@ -66,7 +68,7 @@ def _min_max(func_id, args):
 def _allocate_tensor(func_id, args):
     """Handling TVM tensor allocation.
     You may refer hybrid.intrin.allocate for more details."""
-    n = len(args)
+    n = args.__len__()
     _internal_assert(isinstance(_api.convert(args[0]), Array), \
                      "allocate's first argument should be a tuple of shape!")
     shape = args[0]
@@ -89,4 +91,16 @@ def _allocate_tensor(func_id, args):
         scope = 'global' if func_id != 'output_tensor' else 'output'
     return (shape, dtype, scope)
 
+
 output_tensor = allocate = _allocate_tensor #pylint: disable=invalid-name
+
+
+def len(func_id, args):
+    """Iterpret the len function"""
+    _internal_assert(args.__len__() == 1, "Only 1 argument is expected!")
+    _internal_assert(func_id == "len", "This function cannot be directly invoked!")
+    try:
+        return _api.convert(args[0].__len__())
+    except: #pylint: disable=bare-except
+        _internal_assert(args[0].shape.__len__() == 1, "Only one-dimension array can get len")
+        return _api.convert(args[0].shape[0])

python/tvm/hybrid/intrin.py

@@ -2,32 +2,19 @@
 
 import numpy
 
-class _range(object):
-    """Base class of the loop ranges in hybrid script"""
-    def __init__(self, a, b=None):
-        if b is None:
-            self.low = 0
-            self.ext = a
-        else:
-            self.low = a
-            self.ext = b
+
+class bind(object): #pylint: disable=invalid-name
+    """GPU bind software emulataion runtime."""
+    def __init__(self, _, ext):
+        self.ext = ext
 
     def __iter__(self):
         i = 0
         while i < self.ext:
-            yield i + self.low
+            yield i
             i += 1
 
 
-class bind(_range): #pylint: disable=invalid-name
-    def __init__(self, tag, ext):
-        super(bind, self).__init__(ext)
-        self.tag = tag
-
-
-unroll = vectorize = parallel = _range #pylint: disable=invalid-name
-
-
 def allocate(shape, dtype='float32', scope='global'): #pylint: disable=unused-argument
     """Allocate a buffer with given shape
 
@@ -47,7 +34,6 @@ def allocate(shape, dtype='float32', scope='global'): #pylint: disable=unused-ar
     """
     return numpy.zeros(shape).astype(dtype)
 
-output_tensor = allocate #pylint: disable=invalid-name
 
 def popcount(x):
     """
@@ -87,17 +73,19 @@ def sigmoid(x):
 
 
 HYBRID_GLOBALS = {
-    'unroll'       : unroll,
-    'vectorize'    : vectorize,
-    'parallel'     : parallel,
-    'allocate'     : allocate,
-    'output_tensor': output_tensor,
+    'len'          : len,
+    'unroll'       : range,
+    'vectorize'    : range,
+    'parallel'     : range,
+    'const_range'  : range,
     'bind'         : bind,
+    'allocate'     : allocate,
+    'output_tensor': allocate,
     'sqrt'         : numpy.sqrt,
     'log'          : numpy.log,
     'tanh'         : numpy.tanh,
     'power'        : numpy.power,
     'exp'          : numpy.exp,
     'sigmoid'      : sigmoid,
-    'popcount'     : popcount
+    'popcount'     : popcount,
 }