|
| 1 | +#define TM 8 |
| 2 | +#define TN SG_SZ |
| 3 | +#define TK 16 |
| 4 | + |
| 5 | +static constexpr size_t MATRIX_M = TM * 2; |
| 6 | +static constexpr size_t MATRIX_N = TN * 2; |
| 7 | +static constexpr size_t MATRIX_K = TK * 2; |
| 8 | + |
| 9 | +#define BF16_EPSILON 0.00781250 |
| 10 | + |
| 11 | +template <typename T, size_t NUM_ROWS, size_t NUM_COLS> struct big_matrix { |
| 12 | +private: |
| 13 | + T *mat; |
| 14 | + |
| 15 | +public: |
| 16 | + T *get_data() { return mat; } |
| 17 | + void set_data(T *data) { mat = data; } |
| 18 | + big_matrix(T *data) : mat(data) {} |
| 19 | +}; |
| 20 | + |
| 21 | +// clang-format off |
| 22 | +/* |
| 23 | +Here's how the data is distributed |
| 24 | +W0 --> 0 1 2 3 4 5 6 7 |
| 25 | +wi [0,0] -> i=0, [0, 0] wi [0,1] --> i=0, [0, 1] wi [0,15] --> i=0, [0, 15] |
| 26 | + i=1, [1, 0] i=1, [1, 1] i=1, [1, 15] |
| 27 | + i=2, [2, 0] i=2, [2, 1] ... |
| 28 | + ... .... |
| 29 | + i=7, [7, 0] i=7, [7, 1] |
| 30 | +*/ |
| 31 | +// clang-format on |
| 32 | +std::tuple<uint32_t, uint32_t> get_coord_ref(int i, int wi_number) { |
| 33 | + return std::make_tuple(i, wi_number); |
| 34 | +} |
| 35 | + |
| 36 | +float sum_rows[MATRIX_M] = {0}; |
| 37 | + |
| 38 | +template <typename T1, typename T2, size_t M, size_t N, size_t K> |
| 39 | +void matrix_multiply(big_matrix<T1, M, N> &C, big_matrix<T2, M, K> &A, |
| 40 | + big_matrix<T2, K / 2, N * 2> &B) { |
| 41 | + size_t NDRangeM = M / TM; |
| 42 | + size_t NDRangeN = N / TN; |
| 43 | + buffer<bfloat16, 2> bufA(A.get_data(), range<2>(M, K)); |
| 44 | + buffer<bfloat16, 2> bufB(B.get_data(), range<2>(K, N)); |
| 45 | + buffer<float, 2> bufC((float *)C.get_data(), range<2>(M, N)); |
| 46 | + |
| 47 | + buffer<float> sum_rows_v(sum_rows, M); // there are total of M rows |
| 48 | + |
| 49 | + queue q; |
| 50 | + q.submit([&](handler &cgh) { |
| 51 | + auto accC = bufC. get_access<access::mode::read_write>(cgh); |
| 52 | + auto accA = bufA. get_access<access::mode::read_write>(cgh); |
| 53 | + auto accB = bufB. get_access<access::mode::read_write>(cgh); |
| 54 | + |
| 55 | + auto v = sum_rows_v. get_access<access::mode::read_write>(cgh); |
| 56 | + auto os = sycl::stream(100000, 6144, cgh); |
| 57 | + |
| 58 | + cgh.parallel_for<class imatrix>( |
| 59 | + nd_range<2>({NDRangeM, NDRangeN * SG_SZ}, {1, 1 * SG_SZ}), |
| 60 | + [=](nd_item<2> spmd_item) [[intel::reqd_sub_group_size(SG_SZ)]] |
| 61 | + |
| 62 | + { |
| 63 | + // The submatrix API has to be accessed by all the workitems in a |
| 64 | + // subgroup these functions will be called once by the subgroup no |
| 65 | + // code divergence between the workitems |
| 66 | + const auto global_idx = spmd_item.get_global_id(0); |
| 67 | + const auto global_idy = spmd_item.get_global_id(1); |
| 68 | + const auto sg_startx = global_idx - spmd_item.get_local_id(0); |
| 69 | + const auto sg_starty = global_idy - spmd_item.get_local_id(1); |
| 70 | + |
| 71 | + sub_group sg = spmd_item.get_sub_group(); |
| 72 | + joint_matrix<sub_group, bfloat16, use::a, TM, TK, layout::row_major> |
| 73 | + sub_a; |
| 74 | + // For B, we assume B has been already VNNIed. |
| 75 | + joint_matrix<sub_group, bfloat16, use::b, TK, TN, |
| 76 | + ext::intel::experimental::matrix::layout::packed> |
| 77 | + sub_b; |
| 78 | + joint_matrix<sub_group, float, use::accumulator, TM, TN> sub_c; |
| 79 | + |
| 80 | + joint_matrix_load(sg, sub_c, |
| 81 | + accC.get_pointer() + (sg_startx * TM) * N + |
| 82 | + sg_starty / SG_SZ * TN, |
| 83 | + N, layout::row_major); |
| 84 | + for (int k = 0; k < K / TK; k += 1) { // |
| 85 | + joint_matrix_load( |
| 86 | + sg, sub_a, accA.get_pointer() + (sg_startx * TM) * K + k * TK, |
| 87 | + K); |
| 88 | + joint_matrix_load(sg, sub_b, |
| 89 | + accB.get_pointer() + (k * TK / 2) * (N * 2) + |
| 90 | + sg_starty / SG_SZ * TN * 2, |
| 91 | + N * 2); |
| 92 | + sub_c = joint_matrix_mad(sg, sub_a, sub_b, sub_c); |
| 93 | + } |
| 94 | + joint_matrix_store(sg, sub_c, |
| 95 | + accC.get_pointer() + (sg_startx * TM) * N + |
| 96 | + sg_starty / SG_SZ * TN, |
| 97 | + N, layout::row_major); |
| 98 | + |
| 99 | + float sum_local_rows[M] = {0}; // 8 local rows, M total |
| 100 | + auto data = |
| 101 | + sycl::ext::intel::experimental::matrix::get_wi_data(sg, sub_c); |
| 102 | + |
| 103 | + // Keep track of rows handled in this WI |
| 104 | + int32_t handled_rows[M] = {-1}; |
| 105 | + size_t |
| 106 | + global_index; // Index into the result array that holds the sums. |
| 107 | + |
| 108 | + for (int i = 0; i < data.length(); ++i) { |
| 109 | + auto dataItem = data[i]; |
| 110 | + auto [row, col] = dataItem.get_coord(); |
| 111 | + // get_coord_ref(i, spmd_item.get_local_id(1)); |
| 112 | + global_index = row + global_idx * TM; |
| 113 | + |
| 114 | + sum_local_rows[global_index] += data[i]; |
| 115 | + |
| 116 | + handled_rows[global_index] = 1; |
| 117 | + } |
| 118 | + |
| 119 | + for (int j = 0; j < M; j++) { |
| 120 | + if (handled_rows[j] == 1) { |
| 121 | + global_index = j; |
| 122 | + sum_local_rows[global_index] = reduce_over_group( |
| 123 | + sg, sum_local_rows[global_index], sycl::plus<>()); |
| 124 | + // only Groups leader perform the global reduction |
| 125 | + if (global_idy % SG_SZ == 0) { |
| 126 | + sycl::atomic_ref<float, sycl::memory_order::relaxed, |
| 127 | + sycl::memory_scope::device> |
| 128 | + aref(v[global_index]); |
| 129 | + aref.fetch_add(sum_local_rows[global_index]); |
| 130 | + } |
| 131 | + } |
| 132 | + } |
| 133 | + }); // parallel for |
| 134 | + }).wait(); |
| 135 | +} |
| 136 | + |
| 137 | +bfloat16 A[MATRIX_M][MATRIX_K]; |
| 138 | +bfloat16 B[MATRIX_K / 2][MATRIX_N * 2]; |
| 139 | +float C[MATRIX_M][MATRIX_N]; |
| 140 | +float D[MATRIX_M][MATRIX_N]; |
| 141 | + |
| 142 | +float make_fp32(bfloat16 x) { |
| 143 | + unsigned int y = *((int *)&x); |
| 144 | + y = y << 16; |
| 145 | + float *res = reinterpret_cast<float *>(&y); |
| 146 | + return *res; |
| 147 | +} |
| 148 | + |
| 149 | +void matrix_multiply_ref(int *A_mem, int *B_mem, int *C_mem, int M, int N, |
| 150 | + int K) { |
| 151 | + for (int m = 0; m < M; m++) |
| 152 | + for (int n = 0; n < N; n++) { |
| 153 | + for (int k = 0; k < K; k++) { |
| 154 | + // Because B was assumed VNNIed |
| 155 | + bfloat16 *va = (bfloat16 *)(A_mem + m * K + k); |
| 156 | + bfloat16 *vb = (bfloat16 *)(B_mem + k * N + n); |
| 157 | + float acc = *((float *)(C_mem + m * N + n)); |
| 158 | + for (int i = 0; i < 2; i++) { |
| 159 | + acc += (make_fp32(va[i]) * make_fp32(vb[i])); |
| 160 | + } |
| 161 | + *((float *)(C_mem + m * N + n)) = acc; |
| 162 | + } |
| 163 | + } |
| 164 | +} |
| 165 | + |
| 166 | +int main() { |
| 167 | + for (int i = 0; i < MATRIX_M; i++) { |
| 168 | + for (int j = 0; j < MATRIX_K; j++) { |
| 169 | + A[i][j] = bfloat16(1.0f * (i + j)); |
| 170 | + } |
| 171 | + } |
| 172 | + for (int i = 0; i < MATRIX_K / 2; i++) { |
| 173 | + for (int j = 0; j < MATRIX_N * 2; j++) { |
| 174 | + B[i][j] = bfloat16(2.0f * i + 3.0f * j); |
| 175 | + } |
| 176 | + } |
| 177 | + for (int i = 0; i < MATRIX_M; i++) { |
| 178 | + for (int j = 0; j < MATRIX_N; j++) { |
| 179 | + C[i][j] = 1.0; |
| 180 | + D[i][j] = 1.0; |
| 181 | + } |
| 182 | + } |
| 183 | + |
| 184 | + big_matrix<float, MATRIX_M, MATRIX_N> MC((float *)&C); |
| 185 | + big_matrix<float, MATRIX_M, MATRIX_N> MD((float *)&D); |
| 186 | + big_matrix<bfloat16, MATRIX_M, MATRIX_K> MA((bfloat16 *)&A); |
| 187 | + big_matrix<bfloat16, MATRIX_K / 2, MATRIX_N * 2> MB((bfloat16 *)&B); |
| 188 | + matrix_multiply(MC, MA, MB); |
| 189 | + matrix_multiply_ref((int32_t *)A, (int32_t *)B, (int32_t *)D, MATRIX_M, |
| 190 | + MATRIX_N, MATRIX_K / 2); |
| 191 | + |
| 192 | + bool res = true; |
| 193 | + float sum_rows_ref[MATRIX_M] = {0}; |
| 194 | + |
| 195 | + for (int i = 0; i < MATRIX_M; i++) { |
| 196 | + for (int j = 0; j < MATRIX_N; j++) { |
| 197 | + // std::cout << C[i][j] << " "; |
| 198 | + if ((fabs(C[i][j]) - fabs(D[i][j])) > BF16_EPSILON) |
| 199 | + res = false; |
| 200 | + sum_rows_ref[i] += C[i][j]; |
| 201 | + } |
| 202 | + if ((fabs(sum_rows_ref[i]) - fabs(sum_rows[i])) > BF16_EPSILON) |
| 203 | + res = false; |
| 204 | + // std::cout << "\n"; |
| 205 | + } |
| 206 | + std::cout << (res ? "passed" : "failed") << std::endl; |
| 207 | + return !res; |
| 208 | +} |
0 commit comments