|
| 1 | +//===---joint_matrix_16bit_impl.hpp - DPC++ joint_matrix----------------===// |
| 2 | +// |
| 3 | +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | +// See https://llvm.org/LICENSE.txt for license information. |
| 5 | +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | +// |
| 7 | +//===----------------------------------------------------------------------===// |
| 8 | + |
| 9 | +template <typename Tab, typename TAcc, typename TResult, size_t TM, size_t TN, |
| 10 | + size_t TK, layout B_layout> |
| 11 | +class imatrix; |
| 12 | + |
| 13 | +template <typename Tab, typename TAcc, typename TResult, size_t M, size_t N, |
| 14 | + size_t K, size_t TM, size_t TN, size_t TK, layout B_layout, size_t VF> |
| 15 | +void matrix_multiply(big_matrix<TResult, M, N> &D, big_matrix<TAcc, M, N> &C, |
| 16 | + big_matrix<Tab, M, K> &A, |
| 17 | + big_matrix<Tab, K / VF, N * VF> &B) { |
| 18 | + size_t NDRangeM = M / TM; |
| 19 | + size_t NDRangeN = N / TN; |
| 20 | + buffer<Tab, 2> bufA(A.get_data(), range<2>(M, K)); |
| 21 | + buffer<Tab, 2> bufB(B.get_data(), range<2>(K, N)); |
| 22 | + buffer<TAcc, 2> bufC((TAcc *)C.get_data(), range<2>(M, N)); |
| 23 | + buffer<TResult, 2> bufD((TResult *)D.get_data(), range<2>(M, N)); |
| 24 | + queue q; |
| 25 | + size_t sg_size = |
| 26 | + get_sg_size<imatrix<Tab, TAcc, TResult, TM, TN, TK, B_layout>>(q); |
| 27 | + |
| 28 | + q.submit([&](handler &cgh) { |
| 29 | + accessor accA{bufA, cgh}; |
| 30 | + accessor accB{bufB, cgh}; |
| 31 | + accessor accC{bufC, cgh}; |
| 32 | + accessor accD{bufD, cgh}; |
| 33 | + |
| 34 | + cgh.parallel_for<imatrix<Tab, TAcc, TResult, TM, TN, TK, B_layout>>( |
| 35 | + nd_range<2>({NDRangeM, NDRangeN * sg_size}, {1, 1 * sg_size}), |
| 36 | + [=](nd_item<2> spmd_item) |
| 37 | +#ifdef SG_SZ |
| 38 | + [[sycl::reqd_sub_group_size(SG_SZ)]] |
| 39 | +#endif |
| 40 | + { |
| 41 | + // The submatrix API has to be accessed by all the workitems in a |
| 42 | + // subgroup these functions will be called once by the subgroup no |
| 43 | + // code divergence between the workitems |
| 44 | + const auto global_idx = spmd_item.get_global_id(0); |
| 45 | + const auto global_idy = spmd_item.get_global_id(1); |
| 46 | + const auto sg_startx = global_idx - spmd_item.get_local_id(0); |
| 47 | + const auto sg_starty = global_idy - spmd_item.get_local_id(1); |
| 48 | + |
| 49 | + sub_group sg = spmd_item.get_sub_group(); |
| 50 | + joint_matrix<sub_group, Tab, use::a, TM, TK, layout::row_major> |
| 51 | + sub_a; |
| 52 | + joint_matrix<sub_group, Tab, use::b, TK, TN, B_layout> sub_b; |
| 53 | + joint_matrix<sub_group, TAcc, use::accumulator, TM, TN> sub_c; |
| 54 | + joint_matrix<sub_group, TResult, use::accumulator, TM, TN> sub_d; |
| 55 | + |
| 56 | + joint_matrix_load( |
| 57 | + sg, sub_c, |
| 58 | + accC. template get_multi_ptr<access::decorated::no>() + |
| 59 | + (sg_startx * TM) * N + sg_starty / sg_size * TN, |
| 60 | + N, layout::row_major); |
| 61 | + |
| 62 | + for (int k = 0; k < K / TK; k += 1) { |
| 63 | + joint_matrix_load( |
| 64 | + sg, sub_a, |
| 65 | + accA. template get_multi_ptr<access::decorated::no>() + |
| 66 | + (sg_startx * TM) * K + k * TK, |
| 67 | + K); |
| 68 | + joint_matrix_load( |
| 69 | + sg, sub_b, |
| 70 | + accB. template get_multi_ptr<access::decorated::no>() + |
| 71 | + (k * TK / VF) * (N * VF) + sg_starty / sg_size * TN * VF, |
| 72 | + N * VF); |
| 73 | + |
| 74 | + joint_matrix_mad(sg, sub_d, sub_a, sub_b, sub_c); |
| 75 | + joint_matrix_copy(sg, sub_d, sub_c); |
| 76 | + } |
| 77 | + |
| 78 | + joint_matrix_store( |
| 79 | + sg, sub_d, |
| 80 | + accD. template get_multi_ptr<access::decorated::no>() + |
| 81 | + (sg_startx * TM) * N + sg_starty / sg_size * TN, |
| 82 | + N, layout::row_major); |
| 83 | + }); // parallel for |
| 84 | + }).wait(); |
| 85 | +} |
| 86 | + |
| 87 | +template <typename Tab, typename TAcc, typename TResult, size_t TM, size_t TN, |
| 88 | + size_t TK, layout B_layout, size_t VF> |
| 89 | +void test() { |
| 90 | + std::cout << "Testing: " << TM << " x " << TN << " x " << TK |
| 91 | + << " [TM x TN x TK]" << std::endl; |
| 92 | + |
| 93 | + static constexpr size_t MATRIX_M = TM * 2; |
| 94 | + static constexpr size_t MATRIX_N = TN * 2; |
| 95 | + static constexpr size_t MATRIX_K = TK * 2; |
| 96 | + Tab A[MATRIX_M][MATRIX_K]; |
| 97 | + Tab B[MATRIX_K / VF][MATRIX_N * VF]; |
| 98 | + TAcc C[MATRIX_M][MATRIX_N]; |
| 99 | + TResult D[MATRIX_M][MATRIX_N]; |
| 100 | + TResult DRef[MATRIX_M][MATRIX_N]; |
| 101 | + |
| 102 | + matrix_rand<Tab>(MATRIX_M, MATRIX_K, (Tab *)A, Tab(1)); |
| 103 | + matrix_rand<Tab>(MATRIX_K / VF, MATRIX_N * VF, (Tab *)B, Tab(1)); |
| 104 | + |
| 105 | + matrix_fill(MATRIX_M, MATRIX_N, (TAcc *)C, TAcc(1)); |
| 106 | + matrix_fill(MATRIX_M, MATRIX_N, (TResult *)D, TResult(1)); |
| 107 | + matrix_fill(MATRIX_M, MATRIX_N, (TResult *)DRef, TResult(1)); |
| 108 | + |
| 109 | + big_matrix<TAcc, MATRIX_M, MATRIX_N> MC((TAcc *)&C); |
| 110 | + big_matrix<TResult, MATRIX_M, MATRIX_N> MD((TResult *)&D); |
| 111 | + big_matrix<Tab, MATRIX_M, MATRIX_K> MA((Tab *)&A); |
| 112 | + big_matrix<Tab, MATRIX_K / VF, MATRIX_N * VF> MB((Tab *)&B); |
| 113 | + |
| 114 | + matrix_multiply<Tab, TAcc, TResult, MATRIX_M, MATRIX_N, MATRIX_K, TM, TN, TK, |
| 115 | + B_layout, VF>(MD, MC, MA, MB); |
| 116 | + matrix_multiply_ref<Tab, Tab, TResult, VF>( |
| 117 | + (Tab *)A, (Tab *)B, (TResult *)DRef, MATRIX_M, MATRIX_N, MATRIX_K / VF); |
| 118 | + assert(matrix_compare(MATRIX_M, MATRIX_N, (TResult *)D, (TResult *)DRef)); |
| 119 | +} |
| 120 | + |
| 121 | +template <typename TLow, typename THigh, size_t TM, size_t TN, size_t TK, |
| 122 | + layout B_layout, size_t VF> |
| 123 | +void test_combo() { |
| 124 | + test<TLow, TLow, THigh, TM, TN, TK, B_layout, VF>(); |
| 125 | + test<TLow, THigh, TLow, TM, TN, TK, B_layout, VF>(); |
| 126 | + test<TLow, TLow, TLow, TM, TN, TK, B_layout, VF>(); |
| 127 | + test<TLow, THigh, THigh, TM, TN, TK, B_layout, VF>(); |
| 128 | +} |
| 129 | + |
| 130 | +template <typename TLow, typename THigh, layout B_layout, size_t VF> |
| 131 | +void test_all() { |
| 132 | + test_combo<TLow, THigh, /*TM*/ 8, /*TN*/ 16, /*TK*/ 16, B_layout, VF>(); |
| 133 | + test_combo<TLow, THigh, /*TM*/ 16, /*TN*/ 16, /*TK*/ 16, B_layout, VF>(); |
| 134 | + test_combo<TLow, THigh, /*TM*/ 1, /*TN*/ 64, /*TK*/ 16, B_layout, VF>(); |
| 135 | + test_combo<TLow, THigh, /*TM*/ 1, /*TN*/ 64, /*TK*/ 32, B_layout, VF>(); |
| 136 | + test_combo<TLow, THigh, /*TM*/ 32, /*TN*/ 64, /*TK*/ 16, B_layout, VF>(); |
| 137 | + test_combo<TLow, THigh, /*TM*/ 32, /*TN*/ 64, /*TK*/ 32, B_layout, VF>(); |
| 138 | +} |
0 commit comments