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 import func 16from mlir.dialects.linalg.opdsl import lang as dsl 17 18_SCRIPT_PATH = os.path.dirname(os.path.abspath(__file__)) 19sys.path.append(_SCRIPT_PATH) 20from tools import sparse_compiler 21 22@dsl.linalg_structured_op 23def sddmm_dsl( 24 A=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.K), 25 B=dsl.TensorDef(dsl.T, dsl.S.K, dsl.S.N), 26 S=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.N), 27 C=dsl.TensorDef(dsl.T, dsl.S.M, dsl.S.N, output=True)): 28 C[dsl.D.m, 29 dsl.D.n] += S[dsl.D.m, dsl.D.n] * A[dsl.D.m, dsl.D.k] * B[dsl.D.k, dsl.D.n] 30 31 32def build_SDDMM(attr: st.EncodingAttr): 33 """Build SDDMM kernel. 34 35 This method generates a linalg op with for matrix multiplication using 36 just the Python API. Effectively, a generic linalg op is constructed 37 that computes C(i,j) += S(i,j) SUM_k A(i,k) B(k,j) for sparse S. 38 """ 39 module = ir.Module.create() 40 f64 = ir.F64Type.get() 41 a = ir.RankedTensorType.get([8, 8], f64) 42 b = ir.RankedTensorType.get([8, 8], f64) 43 c = ir.RankedTensorType.get([8, 8], f64) 44 s = ir.RankedTensorType.get([8, 8], f64, attr) 45 arguments = [a, b, s, c] 46 with ir.InsertionPoint(module.body): 47 48 @func.FuncOp.from_py_func(*arguments) 49 def sddmm(*args): 50 return sddmm_dsl(args[0], args[1], args[2], outs=[args[3]]) 51 52 return module 53 54 55def boilerplate(attr: st.EncodingAttr): 56 """Returns boilerplate code for main driver.""" 57 return f""" 58func @main(%a: tensor<8x8xf64>, 59 %b: tensor<8x8xf64>, 60 %c: tensor<8x8xf64>) -> tensor<8x8xf64> attributes {{ llvm.emit_c_interface }} {{ 61 %t = arith.constant sparse<[[0,0], [0,2], [4,1]], [1.0, 2.0, 3.0]> : tensor<8x8xf64> 62 %s = sparse_tensor.convert %t : tensor<8x8xf64> to tensor<8x8xf64, {attr}> 63 %0 = call @sddmm(%a, %b, %s, %c) : (tensor<8x8xf64>, 64 tensor<8x8xf64>, 65 tensor<8x8xf64, {attr}>, 66 tensor<8x8xf64>) -> tensor<8x8xf64> 67 return %0 : tensor<8x8xf64> 68}} 69""" 70 71 72def build_compile_and_run_SDDMMM(attr: st.EncodingAttr, opt: str, 73 support_lib: str, compiler): 74 # Build. 75 module = build_SDDMM(attr) 76 func = str(module.operation.regions[0].blocks[0].operations[0].operation) 77 module = ir.Module.parse(func + boilerplate(attr)) 78 79 # Compile. 80 compiler(module) 81 engine = execution_engine.ExecutionEngine( 82 module, opt_level=0, shared_libs=[support_lib]) 83 84 # Set up numpy input and buffer for output. 85 a = np.array([[1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1, 8.1], 86 [1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2, 8.2], 87 [1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3, 8.3], 88 [1.4, 2.4, 3.4, 4.4, 5.4, 6.4, 7.4, 8.4], 89 [1.5, 2.5, 3.5, 4.5, 5.5, 6.5, 7.5, 8.5], 90 [1.6, 2.6, 3.6, 4.6, 5.6, 6.6, 7.6, 8.6], 91 [1.7, 2.7, 3.7, 4.7, 5.7, 6.7, 7.7, 8.7], 92 [1.8, 2.8, 3.8, 4.8, 5.8, 6.8, 7.8, 8.8]], np.float64) 93 b = np.ones((8, 8), np.float64) 94 c = np.zeros((8, 8), np.float64) 95 96 mem_a = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(a))) 97 mem_b = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(b))) 98 mem_c = ctypes.pointer(ctypes.pointer(rt.get_ranked_memref_descriptor(c))) 99 100 # Allocate a MemRefDescriptor to receive the output tensor. 101 # The buffer itself is allocated inside the MLIR code generation. 102 ref_out = rt.make_nd_memref_descriptor(2, ctypes.c_double)() 103 mem_out = ctypes.pointer(ctypes.pointer(ref_out)) 104 105 # Invoke the kernel and get numpy output. 106 # Built-in bufferization uses in-out buffers. 107 # TODO: replace with inplace comprehensive bufferization. 108 engine.invoke('main', mem_out, mem_a, mem_b, mem_c) 109 110 # Sanity check on computed result. Only a few elements 111 # are sampled from the full dense matrix multiplication. 112 full_matmul = np.matmul(a, b) 113 expected = np.zeros((8, 8), np.float64) 114 expected[0, 0] = 1.0 * full_matmul[0, 0] 115 expected[0, 2] = 2.0 * full_matmul[0, 2] 116 expected[4, 1] = 3.0 * full_matmul[4, 1] 117 c = rt.ranked_memref_to_numpy(mem_out[0]) 118 if np.allclose(c, expected): 119 pass 120 else: 121 quit(f'FAILURE') 122 123 124def main(): 125 support_lib = os.getenv('SUPPORT_LIB') 126 assert support_lib is not None, 'SUPPORT_LIB is undefined' 127 if not os.path.exists(support_lib): 128 raise FileNotFoundError(errno.ENOENT, os.strerror(errno.ENOENT), 129 support_lib) 130 131 # CHECK-LABEL: TEST: testSDDMMM 132 print('\nTEST: testSDDMMM') 133 with ir.Context() as ctx, ir.Location.unknown(): 134 count = 0 135 # Loop over various ways to compile and annotate the SDDMM kernel with 136 # a *single* sparse tensor. Note that we deliberate do not exhaustively 137 # search the full state space to reduce runtime of the test. It is 138 # straightforward to adapt the code below to explore more combinations. 139 levels = [[st.DimLevelType.dense, st.DimLevelType.dense], 140 [st.DimLevelType.dense, st.DimLevelType.compressed], 141 [st.DimLevelType.compressed, st.DimLevelType.dense], 142 [st.DimLevelType.compressed, st.DimLevelType.compressed]] 143 orderings = [ 144 ir.AffineMap.get_permutation([0, 1]), 145 ir.AffineMap.get_permutation([1, 0]) 146 ] 147 for level in levels: 148 for ordering in orderings: 149 for pwidth in [32]: 150 for iwidth in [32]: 151 for par in [0]: 152 for vec in [0, 1]: 153 for e in [True]: 154 vl = 1 if vec == 0 else 16 155 attr = st.EncodingAttr.get(level, ordering, pwidth, iwidth) 156 opt = (f'parallelization-strategy={par} ' 157 f'vectorization-strategy={vec} ' 158 f'vl={vl} enable-simd-index32={e}') 159 compiler = sparse_compiler.SparseCompiler(options=opt) 160 build_compile_and_run_SDDMMM(attr, opt, support_lib, compiler) 161 count = count + 1 162 # CHECK: Passed 16 tests 163 print('Passed ', count, 'tests') 164 165 166if __name__ == '__main__': 167 main() 168