1 // RUN: %libomptarget-compilexx-run-and-check-generic 2 3 #include <cassert> 4 #include <iostream> 5 #include <memory> 6 #include <vector> 7 8 class BlockMatrix { 9 private: 10 const int rowsPerBlock; 11 const int colsPerBlock; 12 const long nRows; 13 const long nCols; 14 const int nBlocksPerRow; 15 const int nBlocksPerCol; 16 std::vector<std::vector<std::unique_ptr<float[]>>> Blocks; 17 18 public: 19 BlockMatrix(const int _rowsPerBlock, const int _colsPerBlock, 20 const long _nRows, const long _nCols) 21 : rowsPerBlock(_rowsPerBlock), colsPerBlock(_colsPerBlock), nRows(_nRows), 22 nCols(_nCols), nBlocksPerRow(_nRows / _rowsPerBlock), 23 nBlocksPerCol(_nCols / _colsPerBlock), Blocks(nBlocksPerCol) { 24 for (int i = 0; i < nBlocksPerCol; i++) { 25 for (int j = 0; j < nBlocksPerRow; j++) { 26 Blocks[i].emplace_back(new float[_rowsPerBlock * _colsPerBlock]); 27 } 28 } 29 }; 30 31 // Initialize the BlockMatrix from 2D arrays 32 void Initialize(const std::vector<float> &matrix) { 33 for (int i = 0; i < nBlocksPerCol; i++) 34 for (int j = 0; j < nBlocksPerRow; j++) { 35 float *CurrBlock = GetBlock(i, j); 36 for (int ii = 0; ii < colsPerBlock; ++ii) 37 for (int jj = 0; jj < rowsPerBlock; ++jj) { 38 int curri = i * colsPerBlock + ii; 39 int currj = j * rowsPerBlock + jj; 40 CurrBlock[ii + jj * colsPerBlock] = matrix[curri + currj * nCols]; 41 } 42 } 43 } 44 45 long Compare(const std::vector<float> &matrix) const { 46 long fail = 0; 47 for (int i = 0; i < nBlocksPerCol; i++) 48 for (int j = 0; j < nBlocksPerRow; j++) { 49 float *CurrBlock = GetBlock(i, j); 50 for (int ii = 0; ii < colsPerBlock; ++ii) 51 for (int jj = 0; jj < rowsPerBlock; ++jj) { 52 int curri = i * colsPerBlock + ii; 53 int currj = j * rowsPerBlock + jj; 54 float m_value = matrix[curri + currj * nCols]; 55 float bm_value = CurrBlock[ii + jj * colsPerBlock]; 56 if (bm_value != m_value) { 57 fail++; 58 } 59 } 60 } 61 return fail; 62 } 63 64 float *GetBlock(int i, int j) const { 65 assert(i < nBlocksPerCol && j < nBlocksPerRow && "Accessing outside block"); 66 return Blocks[i][j].get(); 67 } 68 }; 69 70 constexpr const int BS = 256; 71 constexpr const int N = 1024; 72 73 int BlockMatMul_TargetNowait(BlockMatrix &A, BlockMatrix &B, BlockMatrix &C) { 74 #pragma omp parallel 75 #pragma omp master 76 for (int i = 0; i < N / BS; ++i) 77 for (int j = 0; j < N / BS; ++j) { 78 float *BlockC = C.GetBlock(i, j); 79 for (int k = 0; k < N / BS; ++k) { 80 float *BlockA = A.GetBlock(i, k); 81 float *BlockB = B.GetBlock(k, j); 82 // clang-format off 83 #pragma omp target depend(in: BlockA[0], BlockB[0]) depend(inout: BlockC[0]) \ 84 map(to: BlockA[:BS * BS], BlockB[:BS * BS]) \ 85 map(tofrom: BlockC[:BS * BS]) nowait 86 // clang-format on 87 #pragma omp parallel for 88 for (int ii = 0; ii < BS; ii++) 89 for (int jj = 0; jj < BS; jj++) { 90 for (int kk = 0; kk < BS; ++kk) 91 BlockC[ii + jj * BS] += 92 BlockA[ii + kk * BS] * BlockB[kk + jj * BS]; 93 } 94 } 95 } 96 return 0; 97 } 98 99 void Matmul(const std::vector<float> &a, const std::vector<float> &b, 100 std::vector<float> &c) { 101 for (int i = 0; i < N; ++i) { 102 for (int j = 0; j < N; ++j) { 103 float sum = 0.0; 104 for (int k = 0; k < N; ++k) { 105 sum = sum + a[i * N + k] * b[k * N + j]; 106 } 107 c[i * N + j] = sum; 108 } 109 } 110 } 111 112 int main(int argc, char *argv[]) { 113 std::vector<float> a(N * N); 114 std::vector<float> b(N * N); 115 std::vector<float> c(N * N, 0.0); 116 117 for (int i = 0; i < N; ++i) { 118 for (int j = 0; j < N; ++j) { 119 a[i * N + j] = b[i * N + j] = i + j % 100; 120 } 121 } 122 123 auto BlockedA = BlockMatrix(BS, BS, N, N); 124 BlockedA.Initialize(a); 125 BlockedA.Compare(a); 126 auto BlockedB = BlockMatrix(BS, BS, N, N); 127 BlockedB.Initialize(b); 128 BlockedB.Compare(b); 129 130 Matmul(a, b, c); 131 132 auto BlockedC = BlockMatrix(BS, BS, N, N); 133 BlockMatMul_TargetNowait(BlockedA, BlockedB, BlockedC); 134 135 if (BlockedC.Compare(c) > 0) { 136 return 1; 137 } 138 139 std::cout << "PASS\n"; 140 141 return 0; 142 } 143 144 // CHECK: PASS 145