[Matirx] add support for joint_matrix_copy

Author: yubingex007-a11y

sycl/include/sycl/ext/oneapi/matrix/matrix-unified.hpp

@@ -323,6 +323,34 @@ joint_matrix_load(Group sg,
 #endif // defined(__SYCL_DEVICE_ONLY__)
 }
 
+template <typename Group, typename T1, typename T2, size_t Rows, size_t Cols,
+          use Use1, use Use2, layout Layout>
+void joint_matrix_copy(
+    Group sg, joint_matrix<Group, T1, Use1, Rows, Cols, Layout> &src,
+    joint_matrix<Group, T2, Use2, Rows, Cols, Layout> &dest) {
+#if defined(__SYCL_DEVICE_ONLY__)
+#if defined(__NVPTX__)
+  // cuda code
+#else
+  using storage_element_type =
+      typename oneapi::detail::jm_type_interpretation_helper_trait<
+          T2>::storage_element_type;
+  auto wi_data_c = sycl::ext::intel::experimental::matrix::get_wi_data(sg, src);
+  auto wi_data_d =
+      sycl::ext::intel::experimental::matrix::get_wi_data(sg, dest);
+  for (int i = 0; i < wi_data_c.length(); i++) {
+    wi_data_d[i] = static_cast<storage_element_type>(wi_data_c[i]);
+  }
+#endif // defined(__NVPTX__)
+#else
+  std::ignore = sg;
+  std::ignore = dest;
+  std::ignore = src;
+  throw runtime_error("joint matrix is not supported on host device.",
+                      PI_ERROR_INVALID_DEVICE);
+#endif // defined(__SYCL_DEVICE_ONLY__)
+}
+
 template <typename Group, typename T, size_t NumRows, size_t NumCols,
           access::address_space Space, access::decorated IsDecorated>
 inline __SYCL_ALWAYS_INLINE void joint_matrix_store(

sycl/test-e2e/Matrix/joint_matrix_copy.cpp

@@ -0,0 +1,21 @@
+//==-------- joint_matrix_ss_int8.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: %{build} -o %t.out -DSYCL_EXT_ONEAPI_MATRIX_VERSION=4
+// RUN: %{run} %t.out
+
+#include <iostream>
+#include <sycl/sycl.hpp>
+
+using namespace sycl;
+using namespace sycl::ext::oneapi::experimental::matrix;
+
+#define SG_SZ 16
+
+#include "joint_matrix_copy_impl.hpp"

sycl/test-e2e/Matrix/joint_matrix_copy_impl.hpp

@@ -0,0 +1,146 @@
+#define TM 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 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;
+  // B => K/4 x N*4, A => M x K, C => M, N
+  // stride should be X's cols, e.g., B's stirde = N*4
+  assert(NUM_ROWS_C == NUM_ROWS_A && NUM_COLS_A == NUM_ROWS_B * 4);
+  size_t NDRangeM = M / TM;
+  size_t NDRangeN = N / TN;
+  buffer<int8_t, 2> bufA(A.get_data(), range<2>(M, K));
+  buffer<int8_t, 2> bufB(B.get_data(), range<2>(K, N));
+  buffer<int32_t, 2> bufC(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<sub_group, int8_t, use::a, TM, TK, layout::row_major>
+               sub_a;
+           // For B, we assume B has been already VNNIed.
+           joint_matrix<sub_group, int8_t, use::b, TK, TN,
+                        ext::intel::experimental::matrix::layout::packed>
+               sub_b;
+           joint_matrix<sub_group, int32_t, use::accumulator, TM, TN> sub_c;
+           joint_matrix<sub_group, float, use::a, TM, TK, layout::row_major> sub_d;
+
+           joint_matrix_fill(sg, sub_c, 0);
+           for (int k = 0; k < K / TK; k += 1) {
+             joint_matrix_load(
+                 sg, sub_a,
+                 accA.template get_multi_ptr<access::decorated::no>() +
+                     (sg_startx * TM) * K + k * TK,
+                 K);
+             joint_matrix_copy(sg, sub_a, sub_d);
+             joint_matrix_copy(sg, sub_d, sub_a);
+             joint_matrix_load(
+                 sg, sub_b,
+                 accB.template get_multi_ptr<access::decorated::no>() +
+                     (k * TK / 4) * (N * 4) + sg_starty / SG_SZ * TN * 4,
+                 N * 4);
+             sub_c = joint_matrix_mad(sg, sub_a, sub_b, sub_c);
+           }
+           joint_matrix_store(
+               sg, sub_c,
+               accC.template get_multi_ptr<access::decorated::no>() +
+                   (sg_startx * TM) * N + sg_starty / SG_SZ * TN,
+               N, 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;
+int8_t A[MATRIX_M][MATRIX_K];
+int8_t B[MATRIX_K / 4][MATRIX_N * 4];
+int32_t C[MATRIX_M][MATRIX_N];
+int32_t D[MATRIX_M][MATRIX_N];
+
+void matrix_multiply_ref(int32_t *A_mem, int32_t *B_mem, int32_t *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++) {
+        char *va = (char *)(A_mem + m * K + k);
+        char *vb = (char *)(B_mem + k * N + n);
+        int acc = *(C_mem + m * N + n);
+        for (int i = 0; i < 4; i++) {
+          acc += (va[i] * vb[i]);
+        }
+        *(C_mem + m * N + n) = acc;
+      }
+    }
+}
+
+int main() {
+  for (int i = 0; i < MATRIX_M; i++) {
+    for (int j = 0; j < MATRIX_K; j++) {
+      A[i][j] = i + 2 * j;
+    }
+  }
+  for (int i = 0; i < MATRIX_K / 4; i++) {
+    for (int j = 0; j < MATRIX_N * 4; j++) {
+      B[i][j] = i + j;
+    }
+  }
+  for (int i = 0; i < MATRIX_M; i++) {
+    for (int j = 0; j < MATRIX_N; j++) {
+      C[i][j] = 0;
+      D[i][j] = 0;
+    }
+  }
+
+  big_matrix<int32_t, MATRIX_M, MATRIX_N> MC((int32_t *)&C);
+  big_matrix<int32_t, MATRIX_M, MATRIX_N> MD((int32_t *)&D);
+  big_matrix<int8_t, MATRIX_M, MATRIX_K> MA((int8_t *)&A);
+  big_matrix<int8_t, MATRIX_K / 4, MATRIX_N * 4> MB((int8_t *)&B);
+  matrix_multiply(MC, MA, MB);
+  matrix_multiply_ref((int32_t *)A, (int32_t *)B, (int32_t *)D, MATRIX_M,
+                      MATRIX_N, MATRIX_K / 4);
+
+  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;
+    }
+  }
+  std::cout << (res ? "passed" : "failed") << std::endl;
+  return !res;
+}