|
| 1 | +// XPU test case (maybe also Habana eventually). |
| 2 | +// |
| 3 | +// This was modified from |
| 4 | +// <URL:https://github.com/intel/llvm-test-suite/blob/intel/SYCL/Matrix/joint_matrix_bfloat16.cpp>. |
| 5 | +// See below for copyright notice. |
| 6 | +// |
| 7 | +// On eX3 XPU node eg n022 (note the 2025 version will not work with this code): |
| 8 | +// |
| 9 | +// $ module load intel/oneapi/2023.1/tbb |
| 10 | +// $ module load intel/oneapi/2023.1/compiler-rt |
| 11 | +// $ module load intel/oneapi/2023.1/compiler |
| 12 | +// $ icpx -fsycl -DSYCL_EXT_ONEAPI_MATRIX_VERSION=4 sycl-mmul.cpp -o sycl-mmul |
| 13 | +// $ ./sycl-mmul & |
| 14 | +// |
| 15 | +// It runs for about 45s on that node. |
| 16 | +// |
| 17 | +// While it's running, run `xpu-smi stats -d 0` or (from Sonar) `xpu-shell -state` or other similar |
| 18 | +// commands to verify that the compute engine is busy, and run both to verify that they are in |
| 19 | +// agreement. |
| 20 | +// |
| 21 | +// To schedule the load on specific devices, use an environment variable, here GPU 1: |
| 22 | +// |
| 23 | +// $ ONEAPI_DEVICE_SELECTOR='*:1' ./sycl-mmul |
| 24 | +// |
| 25 | +// There is a rich device selection language. See |
| 26 | +// <URL:https://github.com/intel/llvm/blob/sycl/sycl/doc/EnvironmentVariables.md#oneapi_device_selector> |
| 27 | +// for further documentation. |
| 28 | + |
| 29 | +//==-------- joint_matrix_bfloat16.cpp - DPC++ joint_matrix----------- ----==// |
| 30 | +// |
| 31 | +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 32 | +// See https://llvm.org/LICENSE.txt for license information. |
| 33 | +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 34 | +// |
| 35 | +//===----------------------------------------------------------------------===// |
| 36 | +// REQUIRES: matrix |
| 37 | + |
| 38 | +// RUN: %clangxx -fsycl %s -o %t.out -DSYCL_EXT_ONEAPI_MATRIX_VERSION=4 |
| 39 | +// RUN: %CPU_RUN_PLACEHOLDER %t.out |
| 40 | +// RUN: %GPU_RUN_PLACEHOLDER %t.out |
| 41 | + |
| 42 | +#include <iostream> |
| 43 | +#include <ctime> |
| 44 | +#include <sycl/sycl.hpp> |
| 45 | + |
| 46 | +using namespace sycl; |
| 47 | +using namespace sycl::ext::oneapi::experimental::matrix; |
| 48 | +using bfloat16 = sycl::ext::oneapi::bfloat16; |
| 49 | + |
| 50 | +#define SG_SZ 16 |
| 51 | + |
| 52 | +#define TM 8 |
| 53 | +#define TN SG_SZ |
| 54 | +#define TK 16 |
| 55 | + |
| 56 | +#define BF16_EPSILON 0.00781250 |
| 57 | + |
| 58 | +template <typename T, size_t NUM_ROWS, size_t NUM_COLS> struct big_matrix { |
| 59 | +private: |
| 60 | + T *mat; |
| 61 | + |
| 62 | +public: |
| 63 | + T *get_data() { return mat; } |
| 64 | + void set_data(T *data) { mat = data; } |
| 65 | + big_matrix(T *data) : mat(data) {} |
| 66 | +}; |
| 67 | + |
| 68 | +template <typename T1, typename T2, size_t M, size_t N, size_t K> |
| 69 | +void matrix_multiply(big_matrix<T1, M, N> &C, big_matrix<T2, M, K> &A, |
| 70 | + big_matrix<T2, K / 2, N * 2> &B) { |
| 71 | + size_t NDRangeM = M / TM; |
| 72 | + size_t NDRangeN = N / TN; |
| 73 | + buffer<bfloat16, 2> bufA(A.get_data(), range<2>(M, K)); |
| 74 | + buffer<bfloat16, 2> bufB(B.get_data(), range<2>(K, N)); |
| 75 | + buffer<float, 2> bufC((float *)C.get_data(), range<2>(M, N)); |
| 76 | + |
| 77 | + queue q; |
| 78 | + q.submit([&](handler &cgh) { |
| 79 | + auto accC = bufC. get_access<access::mode::read_write>(cgh); |
| 80 | + auto accA = bufA. get_access<access::mode::read_write>(cgh); |
| 81 | + auto accB = bufB. get_access<access::mode::read_write>(cgh); |
| 82 | + |
| 83 | + cgh.parallel_for<class imatrix>( |
| 84 | + nd_range<2>({NDRangeM, NDRangeN * SG_SZ}, {1, 1 * SG_SZ}), |
| 85 | + [=](nd_item<2> spmd_item) [[intel::reqd_sub_group_size(SG_SZ)]] |
| 86 | + |
| 87 | + { |
| 88 | + // The submatrix API has to be accessed by all the workitems in a |
| 89 | + // subgroup these functions will be called once by the subgroup no |
| 90 | + // code divergence between the workitems |
| 91 | + const auto global_idx = spmd_item.get_global_id(0); |
| 92 | + const auto global_idy = spmd_item.get_global_id(1); |
| 93 | + const auto sg_startx = global_idx - spmd_item.get_local_id(0); |
| 94 | + const auto sg_starty = global_idy - spmd_item.get_local_id(1); |
| 95 | + |
| 96 | + sub_group sg = spmd_item.get_sub_group(); |
| 97 | + joint_matrix<sub_group, bfloat16, use::a, TM, TK, layout::row_major> |
| 98 | + sub_a; |
| 99 | + // For B, we assume B has been already VNNIed. |
| 100 | + joint_matrix<sub_group, bfloat16, use::b, TK, TN, |
| 101 | + ext::intel::experimental::matrix::layout::packed> |
| 102 | + sub_b; |
| 103 | + joint_matrix<sub_group, float, use::accumulator, TM, TN> sub_c; |
| 104 | + |
| 105 | + joint_matrix_load(sg, sub_c, |
| 106 | + accC.get_pointer() + (sg_startx * TM) * N + |
| 107 | + sg_starty / SG_SZ * TN, |
| 108 | + N, layout::row_major); |
| 109 | + for (int k = 0; k < K / TK; k += 1) { // |
| 110 | + joint_matrix_load( |
| 111 | + sg, sub_a, accA.get_pointer() + (sg_startx * TM) * K + k * TK, |
| 112 | + K); |
| 113 | + joint_matrix_load(sg, sub_b, |
| 114 | + accB.get_pointer() + (k * TK / 2) * (N * 2) + |
| 115 | + sg_starty / SG_SZ * TN * 2, |
| 116 | + N * 2); |
| 117 | + sub_c = joint_matrix_mad(sg, sub_a, sub_b, sub_c); |
| 118 | + } |
| 119 | + joint_matrix_store(sg, sub_c, |
| 120 | + accC.get_pointer() + (sg_startx * TM) * N + |
| 121 | + sg_starty / SG_SZ * TN, |
| 122 | + N, layout::row_major); |
| 123 | + }); // parallel for |
| 124 | + }).wait(); |
| 125 | +} |
| 126 | + |
| 127 | +static constexpr size_t MATRIX_M = TM * 2000; |
| 128 | +static constexpr size_t MATRIX_N = TN * 2000; |
| 129 | +static constexpr size_t MATRIX_K = TK * 2000; |
| 130 | + |
| 131 | +int main() { |
| 132 | + // bfloat16 A[MATRIX_M][MATRIX_K]; |
| 133 | + // bfloat16 B[MATRIX_K / 2][MATRIX_N * 2]; |
| 134 | + // float C[MATRIX_M][MATRIX_N]; |
| 135 | + |
| 136 | + // Dynamic allocation or the linker will toss its cookies. |
| 137 | + bfloat16 (*A)[MATRIX_K] = (bfloat16 (*)[MATRIX_K])malloc(2*MATRIX_M*MATRIX_K); |
| 138 | + bfloat16 (*B)[MATRIX_N * 2] = (bfloat16(*)[MATRIX_N * 2])malloc(2*(MATRIX_K/2)*(MATRIX_N)*2); |
| 139 | + float (*C)[MATRIX_N] = (float(*)[MATRIX_N])malloc(4*MATRIX_M*MATRIX_N); |
| 140 | + |
| 141 | + for (int i = 0; i < MATRIX_M; i++) { |
| 142 | + for (int j = 0; j < MATRIX_K; j++) { |
| 143 | + A[i][j] = bfloat16(1.0f * (i + j)); |
| 144 | + } |
| 145 | + } |
| 146 | + for (int i = 0; i < MATRIX_K / 2; i++) { |
| 147 | + for (int j = 0; j < MATRIX_N * 2; j++) { |
| 148 | + B[i][j] = bfloat16(2.0f * i + 3.0f * j); |
| 149 | + } |
| 150 | + } |
| 151 | + for (int i = 0; i < MATRIX_M; i++) { |
| 152 | + for (int j = 0; j < MATRIX_N; j++) { |
| 153 | + C[i][j] = 1.0; |
| 154 | + } |
| 155 | + } |
| 156 | + |
| 157 | + big_matrix<float, MATRIX_M, MATRIX_N> MC((float *)C); |
| 158 | + big_matrix<bfloat16, MATRIX_M, MATRIX_K> MA((bfloat16 *)A); |
| 159 | + big_matrix<bfloat16, MATRIX_K / 2, MATRIX_N * 2> MB((bfloat16 *)B); |
| 160 | + |
| 161 | + time_t then = time(NULL); |
| 162 | + matrix_multiply(MC, MA, MB); |
| 163 | + time_t now = time(NULL); |
| 164 | + printf("Running time: %lds\n", now-then); |
| 165 | +} |
0 commit comments