[SYCL][matrix] add a new test for irregular slicing on packed matrix that calculates sum of rows of the matrix (intel#899)

Author: dkhaldi

SYCL/Matrix/element_wise_irreg_sum_rows.cpp

@@ -0,0 +1,131 @@
+//==-------- element_wise_irreg_sum_rows.cpp  - DPC++ joint_matrix----- ----==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// REQUIRES: matrix
+
+// RUN: %clangxx -fsycl %s -o %t.out
+// RUN: %CPU_RUN_PLACEHOLDER %t.out
+// RUN: %GPU_RUN_PLACEHOLDER %t.out
+
+// this code calculates the sum of rows into a global array of number of rows
+// elements. First, partial reduction is computed inside each SG, then atomic
+// add is used to reduce between SG leaders
+
+#include <CL/sycl.hpp>
+#include <iostream>
+
+using namespace sycl;
+using namespace sycl::ext::oneapi::experimental::matrix;
+
+#define SG_SZ 8
+
+#define TN SG_SZ
+#define TK 32
+
+template <typename T, size_t NUM_ROWS, size_t NUM_COLS> struct big_matrix {
+public:
+  T *mat;
+
+public:
+  T *get_data() { return mat; }
+  void set_data(T *data) { mat = data; }
+  big_matrix(T *data) : mat(data) {}
+};
+
+template <typename T, size_t M, size_t N>
+void sum_rows_ref(
+    accessor<T, 2, access::mode::read, access::target::host_buffer> B,
+    accessor<int, 1, access::mode::read, access::target::host_buffer>
+        sum_rows) {
+  int sum_rows_ref[M] = {0};
+  for (size_t i = 0; i < M; i++) {
+    for (size_t j = 0; j < N; j++) {
+      sum_rows_ref[i] += B[i][j];
+    }
+    auto diff = sum_rows[i] - sum_rows_ref[i];
+    assert(std::fabs(static_cast<int>(diff)) <=
+           std::numeric_limits<int>::epsilon());
+  }
+}
+
+template <typename T, size_t M, size_t N>
+void matrix_sum_rows(queue q, big_matrix<T, M, N> &B, nd_range<2> &r) {
+  buffer<int8_t, 2> bufB(B.get_data(), range<2>(M, N));
+  // size of vector is known because SG size of set by the user in this case
+  int sum_rows[M] = {0};
+  buffer<int> sum_rows_v(sum_rows, M); // there are total of tK/4 * 2, 16 rows
+  q.submit([&](handler &cgh) {
+     auto accB = bufB.get_access<access::mode::read_write>(cgh);
+
+     auto v = sum_rows_v.get_access<access::mode::atomic>(cgh);
+
+     cgh.parallel_for<class add_matrix>(
+         r, [=](nd_item<2> spmd_item) [[intel::reqd_sub_group_size(SG_SZ)]] {
+           const auto global_idx = spmd_item.get_global_id(0);
+           const auto global_idy = spmd_item.get_global_id(1);
+           const auto sg_startx = global_idx - spmd_item.get_local_id(0);
+           const auto sg_starty = global_idy - spmd_item.get_local_id(1);
+
+           ext::oneapi::sub_group sg = spmd_item.get_sub_group();
+
+           joint_matrix<T, TK, TN, matrix_layout::packed_b> sub_b(sg);
+
+           joint_matrix_load(sg, sub_b,
+                             accB.get_pointer() + (global_idx * (TK / 4) * N) +
+                                 sg_starty / SG_SZ * TN * 4,
+                             N, matrix_layout::packed_b);
+           // calculate sum of rows in sum_rows_v[8], there are 8 rows in sub_b
+           // (tK/4)
+           int32_t sum_local_rows[M] = {0}; // 8 local rows, M total
+           // sub_b has 32x8 elements, 32 elements per WI, 4 per WI per row
+           auto data = sub_b.get_wi_data();
+
+           // each WI calculates local sum of rows
+           for (int row = 0; row < TK / 4; row++) { // there are 8 rows
+             for (int i = 0; i < data.length() / (TK / 4); i++) { // 4 per row
+               // i*SG_SIZE index is found based on the round robin
+               // distribution we are using in the implementation
+               sum_local_rows[row + global_idx * (TK / 4)] += data[i + row * 4]
+             }
+             sum_local_rows[row + global_idx * (TK / 4)] = reduce_over_group(
+                 sg, sum_local_rows[row + global_idx * (TK / 4)],
+                 sycl::plus<>());
+
+             // only Groups leader perform the global reduction
+             if (global_idy % 8 == 0) {
+               atomic_fetch_add(v[row + global_idx * (TK / 4)],
+                                sum_local_rows[row + global_idx * (TK / 4)]);
+             }
+           }
+         }); // parallel for
+   }).wait();
+  sum_rows_ref<T, M, N>(bufB.get_access<access::mode::read>(),
+                        sum_rows_v.get_access<access::mode::read>());
+}
+
+static constexpr size_t MATRIX_K = TK / 4 * 2;
+static constexpr size_t MATRIX_N = TN * 4 * 2;
+int8_t B[MATRIX_K][MATRIX_N];
+
+int main() {
+  big_matrix<int8_t, MATRIX_K, MATRIX_N> MB((int8_t *)&B);
+
+  size_t NDRangeK = MATRIX_K / (TK / 4);
+  size_t NDRangeN = (MATRIX_N / 4) / TN;
+  queue q;
+  nd_range<2> r({NDRangeK, NDRangeN * SG_SZ}, {1, 1 * SG_SZ});
+
+  for (int i = 0; i < MATRIX_K; i++) {
+    for (int j = 0; j < MATRIX_N; j++) {
+      B[i][j] = i;
+    }
+  }
+
+  matrix_sum_rows<int8_t, MATRIX_K, MATRIX_N>(q, MB, r);
+
+  return 0;
+}