1# RUN: SUPPORT_LIB=%mlir_runner_utils_dir/libmlir_c_runner_utils%shlibext \
2# RUN:   %PYTHON %s | FileCheck %s
3
4import ctypes
5import numpy as np
6import os
7import sys
8
9from mlir import ir
10from mlir import runtime as rt
11from mlir import execution_engine
12
13from mlir.dialects import sparse_tensor as st
14from mlir.dialects import builtin
15from mlir.dialects.linalg.opdsl import lang as dsl
16
17_SCRIPT_PATH = os.path.dirname(os.path.abspath(__file__))
18sys.path.append(_SCRIPT_PATH)
19from tools import sparse_compiler
20
21@dsl.linalg_structured_op
22def matmul_dsl(
23    A=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.K),
24    B=dsl.TensorDef(dsl.T, dsl.S.K, dsl.S.N),
25    C=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.N, output=True)):
26  C[dsl.D.m, dsl.D.n] += A[dsl.D.m, dsl.D.k] * B[dsl.D.k, dsl.D.n]
27
28
29def build_SpMM(attr: st.EncodingAttr):
30  """Build SpMM kernel.
31
32  This method generates a linalg op with for matrix multiplication using
33  just the Python API. Effectively, a generic linalg op is constructed
34  that computes C(i,j) += A(i,k) * B(k,j) for annotated matrix A.
35  """
36  module = ir.Module.create()
37  f64 = ir.F64Type.get()
38  a = ir.RankedTensorType.get([3, 4], f64, attr)
39  b = ir.RankedTensorType.get([4, 2], f64)
40  c = ir.RankedTensorType.get([3, 2], f64)
41  arguments = [a, b, c]
42  with ir.InsertionPoint(module.body):
43
44    @builtin.FuncOp.from_py_func(*arguments)
45    def spMxM(*args):
46      return matmul_dsl(args[0], args[1], outs=[args[2]])
47
48  return module
49
50
51def boilerplate(attr: st.EncodingAttr):
52  """Returns boilerplate main method.
53
54  This method sets up a boilerplate main method that takes three tensors
55  (a, b, c), converts the first tensor a into s sparse tensor, and then
56  calls the sparse kernel for matrix multiplication. For convenience,
57  this part is purely done as string input.
58  """
59  return f"""
60func @main(%ad: tensor<3x4xf64>, %b: tensor<4x2xf64>, %c: tensor<3x2xf64>) -> tensor<3x2xf64>
61  attributes {{ llvm.emit_c_interface }} {{
62  %a = sparse_tensor.convert %ad : tensor<3x4xf64> to tensor<3x4xf64, {attr}>
63  %0 = call @spMxM(%a, %b, %c) : (tensor<3x4xf64, {attr}>,
64                                  tensor<4x2xf64>,
65                                  tensor<3x2xf64>) -> tensor<3x2xf64>
66  return %0 : tensor<3x2xf64>
67}}
68"""
69
70
71def build_compile_and_run_SpMM(attr: st.EncodingAttr, support_lib: str,
72                               compiler):
73  # Build.
74  module = build_SpMM(attr)
75  func = str(module.operation.regions[0].blocks[0].operations[0].operation)
76  module = ir.Module.parse(func + boilerplate(attr))
77
78  # Compile.
79  compiler(module)
80  engine = execution_engine.ExecutionEngine(
81      module, opt_level=0, shared_libs=[support_lib])
82
83  # Set up numpy input and buffer for output.
84  a = np.array(
85      [[1.1, 0.0, 0.0, 1.4], [0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 3.3, 0.0]],
86      np.float64)
87  b = np.array([[1.0, 2.0], [4.0, 3.0], [5.0, 6.0], [8.0, 7.0]], np.float64)
88  c = np.zeros((3, 2), np.float64)
89
90  mem_a = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(a)))
91  mem_b = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(b)))
92  mem_c = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(c)))
93  # Allocate a MemRefDescriptor to receive the output tensor.
94  # The buffer itself is allocated inside the MLIR code generation.
95  ref_out = rt.make_nd_memref_descriptor(2, ctypes.c_double)()
96  mem_out = ctypes.pointer(ctypes.pointer(ref_out))
97
98  # Invoke the kernel and get numpy output.
99  # Built-in bufferization uses in-out buffers.
100  # TODO: replace with inplace comprehensive bufferization.
101  engine.invoke('main', mem_out, mem_a, mem_b, mem_c)
102
103  # Sanity check on computed result.
104  expected = np.matmul(a, b);
105  c = rt.ranked_memref_to_numpy(mem_out[0])
106  if np.allclose(c, expected):
107    pass
108  else:
109    quit(f'FAILURE')
110
111
112def main():
113  support_lib = os.getenv('SUPPORT_LIB')
114  assert support_lib is not None, 'SUPPORT_LIB is undefined'
115  if not os.path.exists(support_lib):
116    raise FileNotFoundError(errno.ENOENT, os.strerror(errno.ENOENT), support_lib)
117
118  # CHECK-LABEL: TEST: testSpMM
119  print('\nTEST: testSpMM')
120  with ir.Context() as ctx, ir.Location.unknown():
121    count = 0
122    # Loop over various ways to compile and annotate the SpMM kernel with
123    # a *single* sparse tensor. Note that we deliberate do not exhaustively
124    # search the full state space to reduce runtime of the test. It is
125    # straightforward to adapt the code below to explore more combinations.
126    par = 0
127    vec = 0
128    vl = 1
129    e = False
130    opt = (f'parallelization-strategy={par} '
131           f'vectorization-strategy={vec} '
132           f'vl={vl} enable-simd-index32={e}')
133    levels = [[st.DimLevelType.dense, st.DimLevelType.dense],
134              [st.DimLevelType.dense, st.DimLevelType.compressed],
135              [st.DimLevelType.compressed, st.DimLevelType.dense],
136              [st.DimLevelType.compressed, st.DimLevelType.compressed]]
137    orderings = [
138        ir.AffineMap.get_permutation([0, 1]),
139        ir.AffineMap.get_permutation([1, 0])
140    ]
141    bitwidths = [0]
142    for level in levels:
143      for ordering in orderings:
144        for pwidth in bitwidths:
145          for iwidth in bitwidths:
146            attr = st.EncodingAttr.get(level, ordering, pwidth, iwidth)
147            compiler = sparse_compiler.SparseCompiler(options=opt)
148            build_compile_and_run_SpMM(attr, support_lib, compiler)
149            count = count + 1
150    # CHECK: Passed 8 tests
151    print('Passed ', count, 'tests')
152
153
154if __name__ == '__main__':
155  main()
156