[mlir][sparse][python] make imports more selective

[bixia]

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

4–5

I don't see any use of this import?

7

This is the same as
from mlir import ir

I couldn't find any thing in the style guide that says which one is preferable. But there are examples for the style I mentioned in the style guide and I couldn't find an example for the style used here.

8–9

Similar to the previous comment, these can be

from mlir.dialects import sparse_tensor as st
from mlir import runtime as rt

10

This comment is for the import grouping. I couldn't find anything from the google python guide, but I found something from https://www.python.org/dev/peps/pep-0008/#imports

Imports should be grouped in the following order:

Standard library imports.
Related third party imports.
Local application/library specific imports.
You should put a blank line between each group of imports.

24

from mlir.dialects.linalg.opdsl import lang
or
from mlir.dialects.linalg.opdsl import lang as dsl
?

[aartbik]

comments

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

4–5

This is a kitchen sink for registering all passes as side effect (useful until every pass has its own proper import).

7

Changed. I was also not sure which one is generally preferred.

[aartbik]

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

24

I though about that, but then

def matmul_dsl(

  A=TensorDef(T, S.M, S.K),
  B=TensorDef(T, S.K, S.N),
  C=TensorDef(T, S.M, S.N, output=True)):
C[D.m, D.n] += A[D.m, D.k] * B[D.k, D.n]

loses a lot of its charm, since T, S, etc. need prefix
All other examples I could find use .* for this particular one.

WDYT?

[aartbik]

explicitly name all linalg.opdsl.lang symbols

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

24

How about this for mlir.dialects.linalg.opdsl.lang? Just name everything we import, so that the actual DSL def below stays compact.

[bixia]

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

16–20

I understand that you want to import the individual types to simplify the use syntax, but python style guide says this:

Use import statements for packages and modules only, not for individual classes or functions. Imports from the typing module, typing_extensions module, and the six.moves module are exempt from this rule.

[aartbik]

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

16–20

There seems to be various style guides floating around, and they all say slightly different things.
Okay, here is one last attempt, now with just namespace imports and full prefixes....

[aartbik]

names only in imports

[bixia]

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

24

I like the concise syntax as well. I am fine with what you have right now, or adding a comment to this import and use the concise syntax.

from mlir.dialects.linalg.opdsl.lang import *

[aartbik]

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

24

Yeah, I am on the fence too. But let's to the right thing here.
We can always later introduce a style guide rule for our dsl imports ;-)

mlir/test/python/dialects/sparse_tensor/test_SpMM.py

 # RUN: SUPPORT_LIB=%mlir_runner_utils_dir/libmlir_c_runner_utils%shlibext %PYTHON %s | FileCheck %s
 
-import os
 import ctypes
-import mlir.all_passes_registration
+import os
+
 import numpy as np
 
-from mlir.dialects import builtin
-from mlir.dialects.linalg.opdsl.lang import *
-from mlir.dialects.sparse_tensor import *
-from mlir.execution_engine import *
-from mlir.ir import *
-from mlir.passmanager import *
-from mlir.runtime import *
+import mlir.all_passes_registration
+
+from mlir import ir
+from mlir import runtime as rt
+
+from mlir.dialects import sparse_tensor as st
+
+from mlir.dialects.builtin import FuncOp
+from mlir.dialects.linalg.opdsl.lang import D
+from mlir.dialects.linalg.opdsl.lang import S
+from mlir.dialects.linalg.opdsl.lang import T
+from mlir.dialects.linalg.opdsl.lang import linalg_structured_op
+from mlir.dialects.linalg.opdsl.lang import TensorDef
+from mlir.execution_engine import ExecutionEngine
+from mlir.passmanager import PassManager
 
 
 def run(f):
   print('\nTEST:', f.__name__)
   f()
   return f
 
 
 @linalg_structured_op
 def matmul_dsl(
     A=TensorDef(T, S.M, S.K),
     B=TensorDef(T, S.K, S.N),
     C=TensorDef(T, S.M, S.N, output=True)):
   C[D.m, D.n] += A[D.m, D.k] * B[D.k, D.n]
 
 
-def build_SpMM(attr: EncodingAttr):
+def build_SpMM(attr: st.EncodingAttr):
   """Build SpMM kernel.
 
   This method generates a linalg op with for matrix multiplication using
   just the Python API. Effectively, a generic linalg op is constructed
   that computes C(i,j) += A(i,k) * B(k,j) for annotated matrix A.
   """
-  module = Module.create()
+  module = ir.Module.create()
   f64 = ir.F64Type.get()
-  a = RankedTensorType.get([3, 4], f64, attr)
-  b = RankedTensorType.get([4, 2], f64)
-  c = RankedTensorType.get([3, 2], f64)
+  a = ir.RankedTensorType.get([3, 4], f64, attr)
+  b = ir.RankedTensorType.get([4, 2], f64)
+  c = ir.RankedTensorType.get([3, 2], f64)
   arguments = [a, b, c]
-  with InsertionPoint(module.body):
+  with ir.InsertionPoint(module.body):
 
-    @builtin.FuncOp.from_py_func(*arguments)
+    @FuncOp.from_py_func(*arguments)
     def spMxM(*args):
       return matmul_dsl(args[0], args[1], outs=[args[2]])
 
   return module
 
 
-def boilerplate(attr: EncodingAttr):
+def boilerplate(attr: st.EncodingAttr):
   """Returns boilerplate main method.
 
   This method sets up a boilerplate main method that calls the generated
   sparse kernel. For convenience, this part is purely done as string input.
   """
   return f"""
 func @main(%c: tensor<3x2xf64>) -> tensor<3x2xf64>
   attributes {{ llvm.emit_c_interface }} {{
   %0 = constant dense<[ [ 1.1,  0.0,  0.0,  1.4 ],
                         [ 0.0,  0.0,  0.0,  0.0 ],
                         [ 0.0,  0.0,  3.3,  0.0 ]]> : tensor<3x4xf64>
   %a = sparse_tensor.convert %0 : tensor<3x4xf64> to tensor<3x4xf64, {attr}>
   %b = constant dense<[ [ 1.0,  2.0 ],
                         [ 4.0,  3.0 ],
                         [ 5.0,  6.0 ],
                         [ 8.0,  7.0 ]]> : tensor<4x2xf64>
   %1 = call @spMxM(%a, %b, %c) : (tensor<3x4xf64, {attr}>,
                                   tensor<4x2xf64>,
                                   tensor<3x2xf64>) -> tensor<3x2xf64>
   return %1 : tensor<3x2xf64>
 }}
 """
 
 
-def build_compile_and_run_SpMM(attr: EncodingAttr, support_lib: str, compiler):
+def build_compile_and_run_SpMM(attr: st.EncodingAttr, support_lib: str,
+                               compiler):
   # Build.
   module = build_SpMM(attr)
   func = str(module.operation.regions[0].blocks[0].operations[0].operation)
-  module = Module.parse(func + boilerplate(attr))
+  module = ir.Module.parse(func + boilerplate(attr))
   # Compile.
   compiler(module)
   execution_engine = ExecutionEngine(
       module, opt_level=0, shared_libs=[support_lib])
   # Set up numpy input, invoke the kernel, and get numpy output.
   # Built-in bufferization uses in-out buffers.
   # TODO: replace with inplace comprehensive bufferization.
   Cin = np.zeros((3, 2), np.double)
   Cout = np.zeros((3, 2), np.double)
   Cin_memref_ptr = ctypes.pointer(
-      ctypes.pointer(get_ranked_memref_descriptor(Cin)))
+      ctypes.pointer(rt.get_ranked_memref_descriptor(Cin)))
   Cout_memref_ptr = ctypes.pointer(
-      ctypes.pointer(get_ranked_memref_descriptor(Cout)))
+      ctypes.pointer(rt.get_ranked_memref_descriptor(Cout)))
   execution_engine.invoke('main', Cout_memref_ptr, Cin_memref_ptr)
-  Cresult = ranked_memref_to_numpy(Cout_memref_ptr[0])
+  Cresult = rt.ranked_memref_to_numpy(Cout_memref_ptr[0])
 
   # Sanity check on computed result.
   expected = [[12.3, 12.0], [0.0, 0.0], [16.5, 19.8]]
   if np.allclose(Cresult, expected):
     pass
   else:
     quit(f'FAILURE')
 
pipeline = (
         f'func-bufferize,'
         f'tensor-constant-bufferize,'
         f'builtin.func(tensor-bufferize,std-bufferize,finalizing-bufferize),'
         f'convert-vector-to-llvm{{reassociate-fp-reductions=1 enable-index-optimizations=1}},'
         f'convert-memref-to-llvm,'
         f'convert-std-to-llvm')
     self.pipeline = pipeline
 
-  def __call__(self, module: Module):
+  def __call__(self, module: ir.Module):
     PassManager.parse(self.pipeline).run(module)
 
 
 # CHECK-LABEL: TEST: testSpMM
 # CHECK: Passed 72 tests
 @run
 def testSpMM():
   support_lib = os.getenv('SUPPORT_LIB')
-  with Context() as ctx, Location.unknown():
+  with ir.Context() as ctx, ir.Location.unknown():
     count = 0
     # Fixed compiler optimization strategy.
     # TODO: explore state space here too
     par = 0
     vec = 0
     vl = 1
     e = False
     opt = (f'parallelization-strategy={par} '
            f'vectorization-strategy={vec} '
            f'vl={vl} enable-simd-index32={e}')
     # Exhaustive loop over various ways to annotate a kernel with
     # a *single* sparse tensor. Even this subset already gives
     # quite a large state space!
-    levels = [[DimLevelType.dense, DimLevelType.dense],
-              [DimLevelType.dense, DimLevelType.compressed],
-              [DimLevelType.compressed, DimLevelType.dense],
-              [DimLevelType.compressed, DimLevelType.compressed]]
+    levels = [[st.DimLevelType.dense, st.DimLevelType.dense],
+              [st.DimLevelType.dense, st.DimLevelType.compressed],
+              [st.DimLevelType.compressed, st.DimLevelType.dense],
+              [st.DimLevelType.compressed, st.DimLevelType.compressed]]
     orderings = [
-        AffineMap.get_permutation([0, 1]),
-        AffineMap.get_permutation([1, 0])
+        ir.AffineMap.get_permutation([0, 1]),
+        ir.AffineMap.get_permutation([1, 0])
     ]
     bitwidths = [0, 8, 32]
-    for levels in levels:
+    for level in levels:
       for ordering in orderings:
         for pwidth in bitwidths:
           for iwidth in bitwidths:
-            attr = EncodingAttr.get(levels, ordering, pwidth, iwidth)
+            attr = st.EncodingAttr.get(level, ordering, pwidth, iwidth)
             compiler = SparseCompiler(options=opt)
             build_compile_and_run_SpMM(attr, support_lib, compiler)
             count = count + 1
     print('Passed ', count, 'tests')