1# RUN: %PYTHON %s 2>&1 | FileCheck %s
2# REQUIRES: llvm_has_native_target
3import gc, sys
4from mlir.ir import *
5from mlir.passmanager import *
6from mlir.execution_engine import *
7from mlir.runtime import *
8
9
10# Log everything to stderr and flush so that we have a unified stream to match
11# errors/info emitted by MLIR to stderr.
12def log(*args):
13  print(*args, file=sys.stderr)
14  sys.stderr.flush()
15
16
17def run(f):
18  log("\nTEST:", f.__name__)
19  f()
20  gc.collect()
21  assert Context._get_live_count() == 0
22
23
24# Verify capsule interop.
25# CHECK-LABEL: TEST: testCapsule
26def testCapsule():
27  with Context():
28    module = Module.parse(r"""
29llvm.func @none() {
30  llvm.return
31}
32    """)
33    execution_engine = ExecutionEngine(module)
34    execution_engine_capsule = execution_engine._CAPIPtr
35    # CHECK: mlir.execution_engine.ExecutionEngine._CAPIPtr
36    log(repr(execution_engine_capsule))
37    execution_engine._testing_release()
38    execution_engine1 = ExecutionEngine._CAPICreate(execution_engine_capsule)
39    # CHECK: _mlirExecutionEngine.ExecutionEngine
40    log(repr(execution_engine1))
41
42
43run(testCapsule)
44
45
46# Test invalid ExecutionEngine creation
47# CHECK-LABEL: TEST: testInvalidModule
48def testInvalidModule():
49  with Context():
50    # Builtin function
51    module = Module.parse(r"""
52    func @foo() { return }
53    """)
54    # CHECK: Got RuntimeError:  Failure while creating the ExecutionEngine.
55    try:
56      execution_engine = ExecutionEngine(module)
57    except RuntimeError as e:
58      log("Got RuntimeError: ", e)
59
60
61run(testInvalidModule)
62
63
64def lowerToLLVM(module):
65  import mlir.conversions
66  pm = PassManager.parse(
67      "convert-memref-to-llvm,convert-std-to-llvm,reconcile-unrealized-casts")
68  pm.run(module)
69  return module
70
71
72# Test simple ExecutionEngine execution
73# CHECK-LABEL: TEST: testInvokeVoid
74def testInvokeVoid():
75  with Context():
76    module = Module.parse(r"""
77func @void() attributes { llvm.emit_c_interface } {
78  return
79}
80    """)
81    execution_engine = ExecutionEngine(lowerToLLVM(module))
82    # Nothing to check other than no exception thrown here.
83    execution_engine.invoke("void")
84
85
86run(testInvokeVoid)
87
88
89# Test argument passing and result with a simple float addition.
90# CHECK-LABEL: TEST: testInvokeFloatAdd
91def testInvokeFloatAdd():
92  with Context():
93    module = Module.parse(r"""
94func @add(%arg0: f32, %arg1: f32) -> f32 attributes { llvm.emit_c_interface } {
95  %add = arith.addf %arg0, %arg1 : f32
96  return %add : f32
97}
98    """)
99    execution_engine = ExecutionEngine(lowerToLLVM(module))
100    # Prepare arguments: two input floats and one result.
101    # Arguments must be passed as pointers.
102    c_float_p = ctypes.c_float * 1
103    arg0 = c_float_p(42.)
104    arg1 = c_float_p(2.)
105    res = c_float_p(-1.)
106    execution_engine.invoke("add", arg0, arg1, res)
107    # CHECK: 42.0 + 2.0 = 44.0
108    log("{0} + {1} = {2}".format(arg0[0], arg1[0], res[0]))
109
110
111run(testInvokeFloatAdd)
112
113
114# Test callback
115# CHECK-LABEL: TEST: testBasicCallback
116def testBasicCallback():
117  # Define a callback function that takes a float and an integer and returns a float.
118  @ctypes.CFUNCTYPE(ctypes.c_float, ctypes.c_float, ctypes.c_int)
119  def callback(a, b):
120    return a / 2 + b / 2
121
122  with Context():
123    # The module just forwards to a runtime function known as "some_callback_into_python".
124    module = Module.parse(r"""
125func @add(%arg0: f32, %arg1: i32) -> f32 attributes { llvm.emit_c_interface } {
126  %resf = call @some_callback_into_python(%arg0, %arg1) : (f32, i32) -> (f32)
127  return %resf : f32
128}
129func private @some_callback_into_python(f32, i32) -> f32 attributes { llvm.emit_c_interface }
130    """)
131    execution_engine = ExecutionEngine(lowerToLLVM(module))
132    execution_engine.register_runtime("some_callback_into_python", callback)
133
134    # Prepare arguments: two input floats and one result.
135    # Arguments must be passed as pointers.
136    c_float_p = ctypes.c_float * 1
137    c_int_p = ctypes.c_int * 1
138    arg0 = c_float_p(42.)
139    arg1 = c_int_p(2)
140    res = c_float_p(-1.)
141    execution_engine.invoke("add", arg0, arg1, res)
142    # CHECK: 42.0 + 2 = 44.0
143    log("{0} + {1} = {2}".format(arg0[0], arg1[0], res[0] * 2))
144
145
146run(testBasicCallback)
147
148
149# Test callback with an unranked memref
150# CHECK-LABEL: TEST: testUnrankedMemRefCallback
151def testUnrankedMemRefCallback():
152  # Define a callback function that takes an unranked memref, converts it to a numpy array and prints it.
153  @ctypes.CFUNCTYPE(None, ctypes.POINTER(UnrankedMemRefDescriptor))
154  def callback(a):
155    arr = unranked_memref_to_numpy(a, np.float32)
156    log("Inside callback: ")
157    log(arr)
158
159  with Context():
160    # The module just forwards to a runtime function known as "some_callback_into_python".
161    module = Module.parse(r"""
162func @callback_memref(%arg0: memref<*xf32>) attributes { llvm.emit_c_interface } {
163  call @some_callback_into_python(%arg0) : (memref<*xf32>) -> ()
164  return
165}
166func private @some_callback_into_python(memref<*xf32>) -> () attributes { llvm.emit_c_interface }
167""")
168    execution_engine = ExecutionEngine(lowerToLLVM(module))
169    execution_engine.register_runtime("some_callback_into_python", callback)
170    inp_arr = np.array([[1.0, 2.0], [3.0, 4.0]], np.float32)
171    # CHECK: Inside callback:
172    # CHECK{LITERAL}: [[1. 2.]
173    # CHECK{LITERAL}:  [3. 4.]]
174    execution_engine.invoke(
175        "callback_memref",
176        ctypes.pointer(ctypes.pointer(get_unranked_memref_descriptor(inp_arr))),
177    )
178    inp_arr_1 = np.array([5, 6, 7], dtype=np.float32)
179    strided_arr = np.lib.stride_tricks.as_strided(
180        inp_arr_1, strides=(4, 0), shape=(3, 4))
181    # CHECK: Inside callback:
182    # CHECK{LITERAL}: [[5. 5. 5. 5.]
183    # CHECK{LITERAL}:  [6. 6. 6. 6.]
184    # CHECK{LITERAL}:  [7. 7. 7. 7.]]
185    execution_engine.invoke(
186        "callback_memref",
187        ctypes.pointer(
188            ctypes.pointer(get_unranked_memref_descriptor(strided_arr))),
189    )
190
191
192run(testUnrankedMemRefCallback)
193
194
195# Test callback with a ranked memref.
196# CHECK-LABEL: TEST: testRankedMemRefCallback
197def testRankedMemRefCallback():
198  # Define a callback function that takes a ranked memref, converts it to a numpy array and prints it.
199  @ctypes.CFUNCTYPE(
200      None,
201      ctypes.POINTER(
202          make_nd_memref_descriptor(2,
203                                    np.ctypeslib.as_ctypes_type(np.float32))),
204  )
205  def callback(a):
206    arr = ranked_memref_to_numpy(a)
207    log("Inside Callback: ")
208    log(arr)
209
210  with Context():
211    # The module just forwards to a runtime function known as "some_callback_into_python".
212    module = Module.parse(r"""
213func @callback_memref(%arg0: memref<2x2xf32>) attributes { llvm.emit_c_interface } {
214  call @some_callback_into_python(%arg0) : (memref<2x2xf32>) -> ()
215  return
216}
217func private @some_callback_into_python(memref<2x2xf32>) -> () attributes { llvm.emit_c_interface }
218""")
219    execution_engine = ExecutionEngine(lowerToLLVM(module))
220    execution_engine.register_runtime("some_callback_into_python", callback)
221    inp_arr = np.array([[1.0, 5.0], [6.0, 7.0]], np.float32)
222    # CHECK: Inside Callback:
223    # CHECK{LITERAL}: [[1. 5.]
224    # CHECK{LITERAL}:  [6. 7.]]
225    execution_engine.invoke(
226        "callback_memref",
227        ctypes.pointer(ctypes.pointer(get_ranked_memref_descriptor(inp_arr))))
228
229
230run(testRankedMemRefCallback)
231
232
233#  Test addition of two memrefs.
234# CHECK-LABEL: TEST: testMemrefAdd
235def testMemrefAdd():
236  with Context():
237    module = Module.parse("""
238      module  {
239      func @main(%arg0: memref<1xf32>, %arg1: memref<f32>, %arg2: memref<1xf32>) attributes { llvm.emit_c_interface } {
240        %0 = arith.constant 0 : index
241        %1 = memref.load %arg0[%0] : memref<1xf32>
242        %2 = memref.load %arg1[] : memref<f32>
243        %3 = arith.addf %1, %2 : f32
244        memref.store %3, %arg2[%0] : memref<1xf32>
245        return
246      }
247     } """)
248    arg1 = np.array([32.5]).astype(np.float32)
249    arg2 = np.array(6).astype(np.float32)
250    res = np.array([0]).astype(np.float32)
251
252    arg1_memref_ptr = ctypes.pointer(
253        ctypes.pointer(get_ranked_memref_descriptor(arg1)))
254    arg2_memref_ptr = ctypes.pointer(
255        ctypes.pointer(get_ranked_memref_descriptor(arg2)))
256    res_memref_ptr = ctypes.pointer(
257        ctypes.pointer(get_ranked_memref_descriptor(res)))
258
259    execution_engine = ExecutionEngine(lowerToLLVM(module))
260    execution_engine.invoke("main", arg1_memref_ptr, arg2_memref_ptr,
261                            res_memref_ptr)
262    # CHECK: [32.5] + 6.0 = [38.5]
263    log("{0} + {1} = {2}".format(arg1, arg2, res))
264
265
266run(testMemrefAdd)
267
268
269#  Test addition of two 2d_memref
270# CHECK-LABEL: TEST: testDynamicMemrefAdd2D
271def testDynamicMemrefAdd2D():
272  with Context():
273    module = Module.parse("""
274      module  {
275        func @memref_add_2d(%arg0: memref<2x2xf32>, %arg1: memref<?x?xf32>, %arg2: memref<2x2xf32>) attributes {llvm.emit_c_interface} {
276          %c0 = arith.constant 0 : index
277          %c2 = arith.constant 2 : index
278          %c1 = arith.constant 1 : index
279          cf.br ^bb1(%c0 : index)
280        ^bb1(%0: index):  // 2 preds: ^bb0, ^bb5
281          %1 = arith.cmpi slt, %0, %c2 : index
282          cf.cond_br %1, ^bb2, ^bb6
283        ^bb2:  // pred: ^bb1
284          %c0_0 = arith.constant 0 : index
285          %c2_1 = arith.constant 2 : index
286          %c1_2 = arith.constant 1 : index
287          cf.br ^bb3(%c0_0 : index)
288        ^bb3(%2: index):  // 2 preds: ^bb2, ^bb4
289          %3 = arith.cmpi slt, %2, %c2_1 : index
290          cf.cond_br %3, ^bb4, ^bb5
291        ^bb4:  // pred: ^bb3
292          %4 = memref.load %arg0[%0, %2] : memref<2x2xf32>
293          %5 = memref.load %arg1[%0, %2] : memref<?x?xf32>
294          %6 = arith.addf %4, %5 : f32
295          memref.store %6, %arg2[%0, %2] : memref<2x2xf32>
296          %7 = arith.addi %2, %c1_2 : index
297          cf.br ^bb3(%7 : index)
298        ^bb5:  // pred: ^bb3
299          %8 = arith.addi %0, %c1 : index
300          cf.br ^bb1(%8 : index)
301        ^bb6:  // pred: ^bb1
302          return
303        }
304      }
305        """)
306    arg1 = np.random.randn(2, 2).astype(np.float32)
307    arg2 = np.random.randn(2, 2).astype(np.float32)
308    res = np.random.randn(2, 2).astype(np.float32)
309
310    arg1_memref_ptr = ctypes.pointer(
311        ctypes.pointer(get_ranked_memref_descriptor(arg1)))
312    arg2_memref_ptr = ctypes.pointer(
313        ctypes.pointer(get_ranked_memref_descriptor(arg2)))
314    res_memref_ptr = ctypes.pointer(
315        ctypes.pointer(get_ranked_memref_descriptor(res)))
316
317    execution_engine = ExecutionEngine(lowerToLLVM(module))
318    execution_engine.invoke("memref_add_2d", arg1_memref_ptr, arg2_memref_ptr,
319                            res_memref_ptr)
320    # CHECK: True
321    log(np.allclose(arg1 + arg2, res))
322
323
324run(testDynamicMemrefAdd2D)
325
326
327#  Test loading of shared libraries.
328# CHECK-LABEL: TEST: testSharedLibLoad
329def testSharedLibLoad():
330  with Context():
331    module = Module.parse("""
332      module  {
333      func @main(%arg0: memref<1xf32>) attributes { llvm.emit_c_interface } {
334        %c0 = arith.constant 0 : index
335        %cst42 = arith.constant 42.0 : f32
336        memref.store %cst42, %arg0[%c0] : memref<1xf32>
337        %u_memref = memref.cast %arg0 : memref<1xf32> to memref<*xf32>
338        call @print_memref_f32(%u_memref) : (memref<*xf32>) -> ()
339        return
340      }
341      func private @print_memref_f32(memref<*xf32>) attributes { llvm.emit_c_interface }
342     } """)
343    arg0 = np.array([0.0]).astype(np.float32)
344
345    arg0_memref_ptr = ctypes.pointer(
346        ctypes.pointer(get_ranked_memref_descriptor(arg0)))
347
348    execution_engine = ExecutionEngine(
349        lowerToLLVM(module),
350        opt_level=3,
351        shared_libs=[
352            "../../../../lib/libmlir_runner_utils.so",
353            "../../../../lib/libmlir_c_runner_utils.so"
354        ])
355    execution_engine.invoke("main", arg0_memref_ptr)
356    # CHECK: Unranked Memref
357    # CHECK-NEXT: [42]
358
359
360run(testSharedLibLoad)
361
362
363#  Test that nano time clock is available.
364# CHECK-LABEL: TEST: testNanoTime
365def testNanoTime():
366  with Context():
367    module = Module.parse("""
368      module {
369      func @main() attributes { llvm.emit_c_interface } {
370        %now = call @nano_time() : () -> i64
371        %memref = memref.alloca() : memref<1xi64>
372        %c0 = arith.constant 0 : index
373        memref.store %now, %memref[%c0] : memref<1xi64>
374        %u_memref = memref.cast %memref : memref<1xi64> to memref<*xi64>
375        call @print_memref_i64(%u_memref) : (memref<*xi64>) -> ()
376        return
377      }
378      func private @nano_time() -> i64 attributes { llvm.emit_c_interface }
379      func private @print_memref_i64(memref<*xi64>) attributes { llvm.emit_c_interface }
380    }""")
381
382    execution_engine = ExecutionEngine(
383        lowerToLLVM(module),
384        opt_level=3,
385        shared_libs=[
386            "../../../../lib/libmlir_runner_utils.so",
387            "../../../../lib/libmlir_c_runner_utils.so"
388        ])
389    execution_engine.invoke("main")
390    # CHECK: Unranked Memref
391    # CHECK: [{{.*}}]
392
393
394run(testNanoTime)
395