|
| 1 | +// RUN: %clangxx -fsycl -O2 %s -o %t.out |
| 2 | +#include <CL/sycl.hpp> |
| 3 | +#if (SYCL_EXT_ONEAPI_MATRIX == 2) |
| 4 | +#include <iostream> |
| 5 | + |
| 6 | +using namespace sycl::ext::oneapi::experimental::matrix; |
| 7 | +using bfloat16 = sycl::ext::oneapi::experimental::bfloat16; |
| 8 | + |
| 9 | +static constexpr auto TILE_SZ = 16; |
| 10 | +static constexpr auto TM = TILE_SZ - 1; |
| 11 | +static constexpr auto TN = TILE_SZ - 1; |
| 12 | +static constexpr auto TK = 2 * TILE_SZ - 2; |
| 13 | + |
| 14 | +static constexpr auto SG_SZ = 16; |
| 15 | + |
| 16 | +template <typename T, size_t NUM_ROWS, size_t NUM_COLS> struct big_matrix { |
| 17 | +public: |
| 18 | + T *mat; |
| 19 | + |
| 20 | +public: |
| 21 | + T *get_data() { return mat; } |
| 22 | + void set_data(T *data) { mat = data; } |
| 23 | + big_matrix(T *data) : mat(data) {} |
| 24 | +}; |
| 25 | + |
| 26 | +template <typename T1, typename T2, size_t NUM_ROWS_A, size_t NUM_COLS_A, |
| 27 | + size_t NUM_ROWS_B, size_t NUM_COLS_B, size_t NUM_ROWS_C, |
| 28 | + size_t NUM_COLS_C> |
| 29 | +void matrix_multiply(big_matrix<T1, NUM_ROWS_C, NUM_COLS_C> &C, |
| 30 | + big_matrix<T2, NUM_ROWS_A, NUM_COLS_A> &A, |
| 31 | + big_matrix<T2, NUM_ROWS_B, NUM_COLS_B> &B) { |
| 32 | + size_t M = NUM_ROWS_C; |
| 33 | + size_t N = NUM_COLS_C; |
| 34 | + size_t K = NUM_COLS_A; |
| 35 | + // B => K/4 x N*4, A => M x K, C => M, N |
| 36 | + // stride should be X's cols, e.g., B's stirde = N*4 |
| 37 | + assert(NUM_ROWS_C == NUM_ROWS_A && NUM_COLS_A == NUM_ROWS_B * 2); |
| 38 | + size_t NDRangeM = M / TM; |
| 39 | + size_t NDRangeN = N / TN; |
| 40 | + sycl::buffer<bfloat16, 2> bufA(A.get_data(), sycl::range<2>(M, K)); |
| 41 | + sycl::buffer<bfloat16, 2> bufB(B.get_data(), sycl::range<2>(K, N)); |
| 42 | + sycl::buffer<float, 2> bufC((float *)C.get_data(), sycl::range<2>(M, N)); |
| 43 | + |
| 44 | + sycl::queue q; |
| 45 | + q.submit([&](sycl::handler &cgh) { |
| 46 | + auto accC = bufC.get_access<sycl::access::mode::read_write>(cgh); |
| 47 | + auto accA = bufA.get_access<sycl::access::mode::read_write>(cgh); |
| 48 | + auto accB = bufB.get_access<sycl::access::mode::read_write>(cgh); |
| 49 | + |
| 50 | + cgh.parallel_for<class imatrix>( |
| 51 | + sycl::nd_range<2>({NDRangeM, NDRangeN * SG_SZ}, {1, 1 * SG_SZ}), |
| 52 | + [accA, accB, accC, M, N, K](sycl::nd_item<2> spmd_item) |
| 53 | + |
| 54 | + { |
| 55 | + // The submatrix API has to be accessed by all the workitems in a |
| 56 | + // subgroup these functions will be called once by the subgroup no |
| 57 | + // code divergence between the workitems |
| 58 | + const auto global_idx = spmd_item.get_global_id(0); |
| 59 | + const auto global_idy = spmd_item.get_global_id(1); |
| 60 | + const auto sg_startx = global_idx - spmd_item.get_local_id(0); |
| 61 | + const auto sg_starty = global_idy - spmd_item.get_local_id(1); |
| 62 | + |
| 63 | + sycl::ext::oneapi::sub_group sg = spmd_item.get_sub_group(); |
| 64 | + joint_matrix<bfloat16, TM, TK> sub_a(sg); |
| 65 | + // For B, since current implementation does not support non-packed |
| 66 | + // layout, users need to specify the updated VNNI sizes along with |
| 67 | + // the packed_b layout. By default, the layout is row_major and size |
| 68 | + // is (TK, TN). |
| 69 | + joint_matrix<bfloat16, TK, TN, matrix_layout::packed_b> sub_b(sg); |
| 70 | + joint_matrix<float, TM, TN> sub_c(sg); |
| 71 | + |
| 72 | + // AMX: 8 register tiles : 1k byte size, SMmaxxSKmax =16x64 |
| 73 | + // strideX = X's cols, so strideC = N, strideA = K, strideB = N*4 |
| 74 | + joint_matrix_load(sg, sub_c, |
| 75 | + accC.get_pointer() + (sg_startx * TM) * N + |
| 76 | + sg_starty / SG_SZ * TN, |
| 77 | + N, matrix_layout::row_major); |
| 78 | + for (int k = 0; k < K / TK; k += 1) { // |
| 79 | + joint_matrix_load( |
| 80 | + sg, sub_a, accA.get_pointer() + (sg_startx * TM) * K + k * TK, |
| 81 | + K, matrix_layout::row_major); |
| 82 | + // Assuming B data is already in VNNI format. |
| 83 | + joint_matrix_load(sg, sub_b, |
| 84 | + accB.get_pointer() + (k * TK / 2) * (N * 2) + |
| 85 | + sg_starty / SG_SZ * TN * 2, |
| 86 | + N * 2, matrix_layout::packed_b); |
| 87 | + sub_c = joint_matrix_mad(sg, sub_a, sub_b, sub_c); |
| 88 | + } |
| 89 | + joint_matrix_store(sg, sub_c, |
| 90 | + accC.get_pointer() + (sg_startx * TM) * N + |
| 91 | + sg_starty / SG_SZ * TN, |
| 92 | + N, matrix_layout::row_major); |
| 93 | + }); // parallel for |
| 94 | + }).wait(); |
| 95 | +} |
| 96 | + |
| 97 | +static constexpr size_t MATRIX_M = TM * 2; |
| 98 | +static constexpr size_t MATRIX_N = TN * 2; |
| 99 | +static constexpr size_t MATRIX_K = TK * 2; |
| 100 | +bfloat16 A[MATRIX_M][MATRIX_K]; |
| 101 | +bfloat16 B[MATRIX_K / 2][MATRIX_N * 2]; |
| 102 | +unsigned short Aref[MATRIX_M][MATRIX_K]; |
| 103 | +unsigned short Bref[MATRIX_K / 2][MATRIX_N * 2]; |
| 104 | +float C[MATRIX_M][MATRIX_N]; |
| 105 | +float D[MATRIX_M][MATRIX_N]; |
| 106 | + |
| 107 | +float make_fp32(short x) { |
| 108 | + unsigned int y = x; |
| 109 | + y = y << 16; |
| 110 | + float *res = reinterpret_cast<float *>(&y); |
| 111 | + return *res; |
| 112 | +} |
| 113 | + |
| 114 | +unsigned short make_bf16(float x) { |
| 115 | + int *res = reinterpret_cast<int *>(&x); |
| 116 | + *res = *res >> 16; |
| 117 | + return (unsigned short)*res; |
| 118 | +} |
| 119 | + |
| 120 | +void matrix_multiply_ref(int *A_mem, int *B_mem, int *C_mem, int M, int N, |
| 121 | + int K) { |
| 122 | + // tiling |
| 123 | + for (int m = 0; m < M; m++) |
| 124 | + for (int n = 0; n < N; n++) { |
| 125 | + for (int k = 0; k < K; k++) { |
| 126 | + short *va = (short *)(A_mem + m * K + k); |
| 127 | + short *vb = (short *)(B_mem + k * N + n); |
| 128 | + float acc = *((float *)(C_mem + m * N + n)); |
| 129 | + // FIXME: Should we do reduce-add in another version? |
| 130 | + for (int i = 0; i < 2; i++) { |
| 131 | + acc += (make_fp32(va[i]) * make_fp32(vb[i])); |
| 132 | + } |
| 133 | + *((float *)(C_mem + m * N + n)) = acc; |
| 134 | + } |
| 135 | + } |
| 136 | +} |
| 137 | + |
| 138 | +int main() { |
| 139 | + for (int i = 0; i < MATRIX_M; i++) { |
| 140 | + for (int j = 0; j < MATRIX_K; j++) { |
| 141 | + // Ee create bfloat16 from unsigned short since float-to-bfloat's |
| 142 | + // conversion is not allowed. |
| 143 | + A[i][j] = bfloat16::from_bits(make_bf16(1.0f * (i + j))); |
| 144 | + Aref[i][j] = make_bf16(1.0f * (i + j)); |
| 145 | + } |
| 146 | + } |
| 147 | + for (int i = 0; i < MATRIX_K / 2; i++) { |
| 148 | + for (int j = 0; j < MATRIX_N * 2; j++) { |
| 149 | + B[i][j] = bfloat16::from_bits((make_bf16(2.0f * i + 3.0f * j))); |
| 150 | + Bref[i][j] = make_bf16(2.0f * i + 3.0f * j); |
| 151 | + } |
| 152 | + } |
| 153 | + for (int i = 0; i < MATRIX_M; i++) { |
| 154 | + for (int j = 0; j < MATRIX_N; j++) { |
| 155 | + C[i][j] = 1.0; |
| 156 | + D[i][j] = 1.0; |
| 157 | + } |
| 158 | + } |
| 159 | + |
| 160 | + big_matrix<float, MATRIX_M, MATRIX_N> MC((float *)&C); |
| 161 | + big_matrix<float, MATRIX_M, MATRIX_N> MD((float *)&D); |
| 162 | + big_matrix<bfloat16, MATRIX_M, MATRIX_K> MA((bfloat16 *)&A); |
| 163 | + big_matrix<bfloat16, MATRIX_K / 2, MATRIX_N * 2> MB((bfloat16 *)&B); |
| 164 | + matrix_multiply(MC, MA, MB); |
| 165 | + matrix_multiply_ref((int32_t *)Aref, (int32_t *)Bref, (int32_t *)D, MATRIX_M, |
| 166 | + MATRIX_N, MATRIX_K / 2); |
| 167 | + |
| 168 | + bool res = true; |
| 169 | + for (int i = 0; i < MATRIX_M; i++) { |
| 170 | + for (int j = 0; j < MATRIX_N; j++) { |
| 171 | + if (C[i][j] != D[i][j]) |
| 172 | + res = false; |
| 173 | + } |
| 174 | + } |
| 175 | + if (res) |
| 176 | + std::cout << "passed\n"; |
| 177 | + else |
| 178 | + std::cout << "failed\n"; |
| 179 | + for (int i = 0; i < MATRIX_M; i++) { |
| 180 | + for (int j = 0; j < MATRIX_N; j++) |
| 181 | + std::cout << C[i][j] << ", "; |
| 182 | + std::cout << "\n"; |
| 183 | + } |
| 184 | + std::cout << std::endl; |
| 185 | + for (int i = 0; i < MATRIX_M; i++) { |
| 186 | + for (int j = 0; j < MATRIX_N; j++) |
| 187 | + std::cout << D[i][j] << ", "; |
| 188 | + std::cout << "\n"; |
| 189 | + } |
| 190 | +} |
| 191 | +#endif // (SYCL_EXT_ONEAPI_MATRIX == 2) |
0 commit comments