Support sparse tensor output.

[bixia]

Summary

Add convertFromMLIRSparseTensor to the supporting C shared library to convert
SparseTensorStorage to COO-flavor format.

Add Python routine sparse_tensor_to_coo_tensor to convert sparse tensor storage
pointer to numpy values for COO-flavor format tensor.

Add a Python test for sparse tensor output.

[bixia]

Fix build.

[bixia]

Exclude the tools directory from tests.

[bixia]

Misc fixes.

[aartbik]

mlir/lib/ExecutionEngine/SparseTensorUtils.cpp

1075

Make it a little more clear that these are the *output* parameters (so we expect pointers that will be populated by this method)

1088

also add TODO on generalizing type, just like L1051

1096

does allocation leak?

why not just use a std::vector<uint64_t> for this

std::vector<uint64_t> perm(rank);
std::iota(perm.begin(), perm.end(), 0);

mlir/test/Integration/Dialect/SparseTensor/python/test_elementwise_add_sparse_output.py

23

Make this a TODO since we will have to work on allowing this

117

actually, in Python land we don't need to CHECK output, we can actually just verify that the values are as expected, see the other py test in this dir

mlir/test/Integration/Dialect/SparseTensor/python/tools/np_to_sparse_tensor.py

29

is there a way to do this only once?

[aartbik]

mlir/lib/ExecutionEngine/SparseTensorUtils.cpp

1113

how about using a "base" for this to avoid the * as was done above
(not trusting compiler optimization ;-)

[bixia]

Address review comment.

[bixia]

mlir/lib/ExecutionEngine/SparseTensorUtils.cpp

1096

Good catch, add delete.
The function toCOO only accept c style array.

[bixia]

Fixed test.

[aartbik]

mlir/lib/ExecutionEngine/SparseTensorUtils.cpp

1096

Yes, but you can pass perm.data() for that. I have a slight preference for that since that is the idiom already used at line 1057 and further. This way we have a single way of doing the permutation setup.

1113

did you see this one?

[bixia]

Replace the use of an allocated array with std::vector.

[bixia]

mlir/test/Integration/Dialect/SparseTensor/python/tools/np_to_sparse_tensor.py

29

We probably need to make c_lib a global, and then user set it to None to trigger this initialization.
Shall we do this?

[bixia]

Put the code to look up the c shared library in a function and decorate the function with functools.lru_cache.

[bixia]

mlir/lib/ExecutionEngine/SparseTensorUtils.cpp

1113

Use "base".

mlir/test/Integration/Dialect/SparseTensor/python/tools/np_to_sparse_tensor.py

29

Style guide recommend to avoid global variable. I outlined the look up of c_lib to a routine and added decorator lru_cache.

[bixia]

Fixed python function docstrings.

[bixia]

Fixed docstrings.

[bixia]

Fixed a typo.

[aartbik]

mlir/test/Integration/Dialect/SparseTensor/python/test_elementwise_add_sparse_output.py

117

We will have to check the return value, i.e. something like

if np.allclose(....):

  pass
else:
  quit(f'FAILURE')

[bixia]

Use the result of np.allclose.

[aartbik]

One last suggestion on testing the other output variable as well. but other than that, solid work. Ship it!

mlir/test/Integration/Dialect/SparseTensor/python/test_elementwise_add_sparse_output.py

117

perhaps also test the other output parameters, either with a similar test or by printing and CHECKing the value. E.g. the rank is of course also reflected in the indices, but still nice to add an explicit check in this early phase of developing the tests

mlir/lib/ExecutionEngine/SparseTensorUtils.cpp

for (uint64_t i = 0, base = 0; i < nse; i++) {
     tensor->add(idx, values[i]);
     base += rank;
   }
   // Return sparse tensor storage format as opaque pointer.
   return SparseTensorStorage<uint64_t, uint64_t, double>::newSparseTensor(
       rank, shape, perm.data(), sparse.data(), tensor);
 }
 
+/// Converts a sparse tensor to COO-flavored format expressed using C-style
+/// data structures. The expected output parameters are pointers for these
+/// values:
+///
+///   rank:    rank of tensor
+///   nse:     number of specified elements (usually the nonzeros)
+///   shape:   array with dimension size for each rank
+///   values:  a "nse" array with values for all specified elements
+///   indices: a flat "nse x rank" array with indices for all specified elements
+///
+/// The input is a pointer to SparseTensorStorage<P, I, V>, typically returned
+/// from convertToMLIRSparseTensor.
+///
+//  TODO: Currently, values are copied from SparseTensorStorage to
+//  SparseTensorCOO, then to the output. We may want to reduce the number of
+//  copies.
+//
+//  TODO: for now f64 tensors only, no dim ordering, all dimensions compressed
+//
+void convertFromMLIRSparseTensor(void *tensor, uint64_t *p_rank,
+                                 uint64_t *p_nse, uint64_t **p_shape,
+                                 double **p_values, uint64_t **p_indices) {
+  SparseTensorStorage<uint64_t, uint64_t, double> *sparse_tensor =
+      static_cast<SparseTensorStorage<uint64_t, uint64_t, double> *>(tensor);
+  uint64_t rank = sparse_tensor->getRank();
+  std::vector<uint64_t> perm(rank);
+  std::iota(perm.begin(), perm.end(), 0);
+  SparseTensorCOO<double> *coo = sparse_tensor->toCOO(perm.data());
+
+  const std::vector<Element<double>> &elements = coo->getElements();
+  uint64_t nse = elements.size();
+
+  uint64_t *shape = new uint64_t[rank];
+  for (uint64_t i = 0; i < rank; i++)
+    shape[i] = coo->getSizes()[i];
+
+  double *values = new double[nse];
+  uint64_t *indices = new uint64_t[rank * nse];
+
+  for (uint64_t i = 0, base = 0; i < nse; i++) {
+    values[i] = elements[i].value;
+    for (uint64_t j = 0; j < rank; j++)
+      indices[base + j] = elements[i].indices[j];
+    base += rank;
+  }
+
+  delete coo;
+  *p_rank = rank;
+  *p_nse = nse;
+  *p_shape = shape;
+  *p_values = values;
+  *p_indices = indices;
+}
 } // extern "C"
 
 #endif // MLIR_CRUNNERUTILS_DEFINE_FUNCTIONS

mlir/test/Integration/Dialect/SparseTensor/python/test_elementwise_add_sparse_output.py

+# RUN: SUPPORT_LIB=%mlir_runner_utils_dir/libmlir_c_runner_utils%shlibext %PYTHON %s | FileCheck %s
+
+import ctypes
+import numpy as np
+import os
+import sys
+
+import mlir.all_passes_registration
+
+from mlir import ir
+from mlir import runtime as rt
+from mlir import execution_engine
+from mlir import passmanager
+from mlir.dialects import sparse_tensor as st
+from mlir.dialects import builtin
+from mlir.dialects.linalg.opdsl import lang as dsl
+
+_SCRIPT_PATH = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(_SCRIPT_PATH)
+from tools import np_to_sparse_tensor as test_tools
+
+# TODO: Use linalg_structured_op to generate the kernel after making it to
+# handle sparse tensor outputs.
+_KERNEL_STR = """
+#DCSR = #sparse_tensor.encoding<{
+  dimLevelType = [ "compressed", "compressed" ]
+}>
+
+#trait_add_elt = {
+  indexing_maps = [
+    affine_map<(i,j) -> (i,j)>,  // A
+    affine_map<(i,j) -> (i,j)>,  // B
+    affine_map<(i,j) -> (i,j)>   // X (out)
+  ],
+  iterator_types = ["parallel", "parallel"],
+  doc = "X(i,j) = A(i,j) + B(i,j)"
+}
+
+func @sparse_add_elt(
+    %arga: tensor<3x4xf64, #DCSR>, %argb: tensor<3x4xf64, #DCSR>) -> tensor<3x4xf64, #DCSR> {
+  %c3 = arith.constant 3 : index
+  %c4 = arith.constant 4 : index
+  %argx = sparse_tensor.init [%c3, %c4] : tensor<3x4xf64, #DCSR>
+  %0 = linalg.generic #trait_add_elt
+    ins(%arga, %argb: tensor<3x4xf64, #DCSR>, tensor<3x4xf64, #DCSR>)
+    outs(%argx: tensor<3x4xf64, #DCSR>) {
+      ^bb(%a: f64, %b: f64, %x: f64):
+        %1 = arith.addf %a, %b : f64
+        linalg.yield %1 : f64
+  } -> tensor<3x4xf64, #DCSR>
+  return %0 : tensor<3x4xf64, #DCSR>
+}
+
+func @main(%ad: tensor<3x4xf64>, %bd: tensor<3x4xf64>) -> tensor<3x4xf64, #DCSR>
+  attributes { llvm.emit_c_interface } {
+  %a = sparse_tensor.convert %ad : tensor<3x4xf64> to tensor<3x4xf64, #DCSR>
+  %b = sparse_tensor.convert %bd : tensor<3x4xf64> to tensor<3x4xf64, #DCSR>
+  %0 = call @sparse_add_elt(%a, %b) : (tensor<3x4xf64, #DCSR>, tensor<3x4xf64, #DCSR>) -> tensor<3x4xf64, #DCSR>
+  return %0 : tensor<3x4xf64, #DCSR>
+}
+"""
+
+
+class _SparseCompiler:
+  """Sparse compiler passes."""
+
+  def __init__(self):
+    self.pipeline = (
+        f'sparsification,'
+        f'sparse-tensor-conversion,'
+        f'builtin.func(linalg-bufferize,convert-linalg-to-loops,convert-vector-to-scf),'
+        f'convert-scf-to-std,'
+        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'lower-affine,'
+        f'convert-memref-to-llvm,'
+        f'convert-std-to-llvm,'
+        f'reconcile-unrealized-casts')
+
+  def __call__(self, module: ir.Module):
+    passmanager.PassManager.parse(self.pipeline).run(module)
+
+
+def _run_test(support_lib, kernel):
+  """Compiles, runs and checks results."""
+  module = ir.Module.parse(kernel)
+  _SparseCompiler()(module)
+  engine = execution_engine.ExecutionEngine(
+      module, opt_level=0, shared_libs=[support_lib])
+
+  # Set up numpy inputs and buffer for output.
+  a = np.array(
+      [[1.1, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 6.6, 0.0]],
+      np.float64)
+  b = np.array(
+      [[1.1, 0.0, 0.0, 2.8], [0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0]],
+      np.float64)
+
+  mem_a = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(a)))
+  mem_b = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(b)))
+
+  # The sparse tensor output is a pointer to pointer of char.
+  out = ctypes.c_char(0)
+  mem_out = ctypes.pointer(ctypes.pointer(out))
+
+  # Invoke the kernel.
+  engine.invoke('main', mem_a, mem_b, mem_out)
+
+  # Retrieve and check the result.
+  rank, nse, shape, values, indices = test_tools.sparse_tensor_to_coo_tensor(
+      support_lib, mem_out[0], np.float64)
+
+  # CHECK: PASSED
+  if np.allclose(values, [2.2, 2.8, 6.6]) and np.allclose(
+      indices, [[0, 0], [0, 3], [2, 2]]):
+    print('PASSED')
+  else:
+    quit('FAILURE')
+
+
+def test_elementwise_add():
+  # Obtain path to runtime support library.
+  support_lib = os.getenv('SUPPORT_LIB')
+  assert support_lib is not None, 'SUPPORT_LIB is undefined'
+  assert os.path.exists(support_lib), f'{support_lib} does not exist'
+  with ir.Context() as ctx, ir.Location.unknown():
+    _run_test(support_lib, _KERNEL_STR)
+
+
+test_elementwise_add()

mlir/test/Integration/Dialect/SparseTensor/python/tools/lit.local.cfg

+# Files in this directory are tools, not tests.
+config.unsupported = True

mlir/test/Integration/Dialect/SparseTensor/python/tools/np_to_sparse_tensor.py

+#  Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+#  See https://llvm.org/LICENSE.txt for license information.
+#  SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#  This file contains functions to process sparse tensor outputs.
+
+import ctypes
+import functools
+import numpy as np
+
+
[e-mail]_cache()
+def _get_c_shared_lib(lib_name: str):
+  """Loads and returns the requested C shared library.
+
+  Args:
+    lib_name: A string representing the C shared library.
+
+  Returns:
+    The C shared library.
+
+  Raises:
+    OSError: If there is any problem in loading the shared library.
+    ValueError:  If the shared library doesn't contain the needed routine.
+  """
+  # This raises OSError exception if there is any problem in loading the shared
+  # library.
+  c_lib = ctypes.CDLL(lib_name)
+
+  try:
+    c_lib.convertFromMLIRSparseTensor.restype = ctypes.c_void_p
+  except Exception as e:
+    raise ValueError('Missing function convertFromMLIRSparseTensor from '
+                     f'the C shared library: {e} ') from e
+
+  return c_lib
+
+
+def sparse_tensor_to_coo_tensor(support_lib, sparse, dtype):
+  """Converts a sparse tensor to COO-flavored format.
+
+  Args:
+     support_lib: A string for the supporting C shared library.
+     sparse: A ctypes.pointer to the sparse tensor descriptor.
+     dtype: The numpy data type for the tensor elements.
+
+  Returns:
+    A tuple that contains the following values:
+    rank: An integer for the rank of the tensor.
+    nse: An interger for the number of non-zero values in the tensor.
+    shape: A 1D numpy array of integers, for the shape of the tensor.
+    values: A 1D numpy array, for the non-zero values in the tensor.
+    indices: A 2D numpy array of integers, representing the indices for the
+      non-zero values in the tensor.
+
+  Raises:
+    OSError: If there is any problem in loading the shared library.
+    ValueError:  If the shared library doesn't contain the needed routine.
+  """
+  c_lib = _get_c_shared_lib(support_lib)
+
+  rank = ctypes.c_ulonglong(0)
+  nse = ctypes.c_ulonglong(0)
+  shape = ctypes.POINTER(ctypes.c_ulonglong)()
+  values = ctypes.POINTER(np.ctypeslib.as_ctypes_type(dtype))()
+  indices = ctypes.POINTER(ctypes.c_ulonglong)()
+  c_lib.convertFromMLIRSparseTensor(sparse, ctypes.byref(rank),
+                                    ctypes.byref(nse), ctypes.byref(shape),
+                                    ctypes.byref(values), ctypes.byref(indices))
+  # Convert the returned values to the corresponding numpy types.
+  shape = np.ctypeslib.as_array(shape, shape=[rank.value])
+  values = np.ctypeslib.as_array(values, shape=[nse.value])
+  indices = np.ctypeslib.as_array(indices, shape=[nse.value, rank.value])
+  return rank, nse, shape, values, indices