[SYCL][Matrix] add a new test for slicing feature where matrix size does not multiply sg size (#756)

Signed-off-by: Dounia Khaldi <[email protected]>

Author: dkhaldi

SYCL/Matrix/elemwise_irreg_size_ops_bf16.cpp

@@ -0,0 +1,190 @@
+//==-------- elemwise_irreg_size_ops_bf16.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
+// This test is for element wise operations when matrix size does not multiply
+// SG size. This corner case only applies to AMX. Also, it tests bf16 type.
+// only run this on AMX
+// RUN: %CPU_RUN_PLACEHOLDER %t.out
+
+#include <CL/sycl.hpp>
+#include <iostream>
+
+using namespace sycl;
+using namespace sycl::ext::oneapi::experimental::matrix;
+
+#define SG_SZ 16
+
+// 10x12 is not multiply the sg size, slicing implementation will have to insert
+// padding
+#define TM 10
+#define TN 12
+#define TK 16
+
+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 T1, typename T2, size_t NUM_ROWS_A, size_t NUM_COLS_A,
+          size_t NUM_ROWS_B, size_t NUM_COLS_B, size_t NUM_ROWS_C,
+          size_t NUM_COLS_C>
+void matrix_multiply(big_matrix<T1, NUM_ROWS_C, NUM_COLS_C> &C,
+                     big_matrix<T2, NUM_ROWS_A, NUM_COLS_A> &A,
+                     big_matrix<T2, NUM_ROWS_B, NUM_COLS_B> &B) {
+  size_t M = NUM_ROWS_C;
+  size_t N = NUM_COLS_C;
+  size_t K = NUM_COLS_A;
+
+  assert(NUM_ROWS_C == NUM_ROWS_A && NUM_COLS_A == NUM_ROWS_B * 2);
+  size_t NDRangeM = M / TM;
+  size_t NDRangeN = N / TN;
+  buffer<unsigned short, 2> bufA(A.get_data(), range<2>(M, K));
+  buffer<unsigned short, 2> bufB(B.get_data(), range<2>(K / 2, N * 2));
+  buffer<float, 2> bufC((float *)C.get_data(), range<2>(M, N));
+
+  queue q;
+  q.submit([&](handler &cgh) {
+     auto accC = bufC.get_access<access::mode::read_write>(cgh);
+     auto accA = bufA.get_access<access::mode::read_write>(cgh);
+     auto accB = bufB.get_access<access::mode::read_write>(cgh);
+
+     cgh.parallel_for<class imatrix>(
+         nd_range<2>({NDRangeM, NDRangeN * SG_SZ}, {1, 1 * SG_SZ}),
+         [accA, accB, accC, M, N, K](nd_item<2> spmd_item)
+             [[intel::reqd_sub_group_size(SG_SZ)]]
+
+         {
+           // The submatrix API has to be accessed by all the workitems in a
+           // subgroup these functions will be called once by the subgroup no
+           // code divergence between the workitems
+           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);
+
+           sub_group sg = spmd_item.get_sub_group();
+           joint_matrix<unsigned short, TM, TK> sub_a(sg);
+           // For B, since current implementation does not support non-packed
+           // layout, users need to specify the packed_b layout.
+           // By default, the layout is row_major
+           joint_matrix<unsigned short, TK, TN, matrix_layout::packed_b> sub_b(
+               sg);
+           joint_matrix<float, TM, TN> sub_c(sg);
+           joint_matrix_load(sg, sub_c,
+                             accC.get_pointer() + (sg_startx * TM) * N +
+                                 sg_starty / SG_SZ * TN,
+                             N, matrix_layout::row_major);
+           for (int k = 0; k < K; k += TK) {
+             joint_matrix_load(sg, sub_a,
+                               accA.get_pointer() + (sg_startx * TM) * K + k, K,
+                               matrix_layout::row_major);
+             // Assume we alreay in vnni format.
+             joint_matrix_load(sg, sub_b,
+                               accB.get_pointer() + (k) * (N) +
+                                   sg_starty / SG_SZ * TN * 2,
+                               N * 2, matrix_layout::packed_b);
+             sub_c = joint_matrix_mad(sg, sub_a, sub_b, sub_c);
+           }
+           auto wi_slice_c = sub_c.get_wi_data();
+           for (int i = 0; i < wi_slice_c.length(); i++) {
+             wi_slice_c[i] += 5.0;
+           }
+           joint_matrix_store(sg, sub_c,
+                              accC.get_pointer() + (sg_startx * TM) * N +
+                                  sg_starty / SG_SZ * TN,
+                              N, matrix_layout::row_major);
+         }); // parallel for
+   }).wait();
+}
+
+static constexpr size_t MATRIX_M = TM * 2;
+static constexpr size_t MATRIX_N = TN * 2;
+static constexpr size_t MATRIX_K = TK * 2;
+unsigned short A[MATRIX_M][MATRIX_K];
+unsigned short B[MATRIX_K / 2][MATRIX_N * 2];
+float C[MATRIX_M][MATRIX_N];
+float D[MATRIX_M][MATRIX_N];
+
+float make_fp32(short x) {
+  unsigned int y = x;
+  y = y << 16;
+  float *res = reinterpret_cast<float *>(&y);
+  return *res;
+}
+
+unsigned short make_bf16(float x) {
+  int *res = reinterpret_cast<int *>(&x);
+  *res = *res >> 16;
+  return (unsigned short)*res;
+}
+
+void matrix_multiply_ref(int *A_mem, int *B_mem, int *C_mem, int M, int N,
+                         int K) {
+  // tiling
+  for (int m = 0; m < M; m++)
+    for (int n = 0; n < N; n++) {
+      for (int k = 0; k < K; k++) {
+        short *va = (short *)(A_mem + m * K + k);
+        short *vb = (short *)(B_mem + k * N + n);
+        float acc = *((float *)(C_mem + m * N + n));
+        // FIXME: Should we do reduce-add in another version?
+        for (int i = 0; i < 2; i++) {
+          acc += (make_fp32(va[i]) * make_fp32(vb[i]));
+        }
+        *((float *)(C_mem + m * N + n)) = acc;
+      }
+      *((float *)(C_mem + m * N + n)) += 5.0;
+    }
+}
+
+int main() {
+  for (int i = 0; i < MATRIX_M; i++) {
+    for (int j = 0; j < MATRIX_K; j++) {
+      A[i][j] = make_bf16(1.0f * (i + j));
+    }
+  }
+  for (int i = 0; i < MATRIX_K / 2; i++) {
+    for (int j = 0; j < MATRIX_N * 2; j++) {
+      B[i][j] = make_bf16(2.0f * i + 3.0f * j);
+    }
+  }
+  for (int i = 0; i < MATRIX_M; i++) {
+    for (int j = 0; j < MATRIX_N; j++) {
+      C[i][j] = 1.0;
+      D[i][j] = 1.0;
+    }
+  }
+
+  big_matrix<float, MATRIX_M, MATRIX_N> MC((float *)&C);
+  big_matrix<float, MATRIX_M, MATRIX_N> MD((float *)&D);
+  big_matrix<unsigned short, MATRIX_M, MATRIX_K> MA((unsigned short *)&A);
+  big_matrix<unsigned short, MATRIX_K / 2, MATRIX_N * 2> MB(
+      (unsigned short *)&B);
+  matrix_multiply(MC, MA, MB);
+  matrix_multiply_ref((int32_t *)A, (int32_t *)B, (int32_t *)D, MATRIX_M,
+                      MATRIX_N, MATRIX_K / 2);
+
+  bool res = true;
+  for (int i = 0; i < MATRIX_M; i++) {
+    for (int j = 0; j < MATRIX_N; j++) {
+      if (C[i][j] != D[i][j])
+        res = false;
+    }
+  }
+  if (res)
+    std::cout << "passed\n";
+  else
+    std::cout << "failed\n";
+}