matrix-int8-test-SG-16.cpp

// RUN: %clangxx -fsycl -O2 %s -o %t.out
#include <CL/sycl.hpp>
#if (SYCL_EXT_ONEAPI_MATRIX == 2)
#include <iostream>

using namespace sycl;
using namespace sycl::ext::oneapi::experimental::matrix;

#define TILE_SZ 16
#define TM (TILE_SZ-5)
#define TN (TILE_SZ-6)
#define TK (4 * TILE_SZ-8)

#define SG_SZ 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;
  // 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);

           ext::oneapi::sub_group sg = spmd_item.get_sub_group();
           joint_matrix<int8_t, TM, TK> sub_a(sg);
           // For B, since current implementation does not support non-packed layout,
           // users need to specify the updated VNNI sizes along with the packed_b layout.
           // By default, the layout is row_major and size is (TK, TN).
           joint_matrix<int8_t, TK, TN, matrix_layout::packed_b> sub_b(sg);
           joint_matrix<int32_t, TM, TN> sub_c(sg);

           // AMX: 8 register tiles : 1k byte size, SMmaxxSKmax =16x64
           // strideX = X's cols, so strideC = N, strideA = K, strideB = N*4
           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 / TK; k += 1) {
             joint_matrix_load(
                 sg, sub_a, accA.get_pointer() + (sg_startx * TM) * K + k * TK,
                 K, matrix_layout::row_major);
             // Assuming B data is already in VNNI format.
             joint_matrix_load(sg, sub_b,
                               accB.get_pointer() + (k * TK / 4) * (N * 4) +
                                   sg_starty / SG_SZ * TN * 4,
                               N * 4, matrix_layout::packed_b);
             sub_c = joint_matrix_mad(sg, sub_a, sub_b, sub_c);
           }
           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;
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] = 1;
      D[i][j] = 1;
    }
  }

  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;
    }
  }
  if (res)
    std::cout << "passed\n";
  else
    std::cout << "failed\n";
  for (int i = 0; i < MATRIX_M; i++) {
    for (int j = 0; j < MATRIX_N; j++)
      std::cout << C[i][j] << ", ";
    std::cout << "\n";
  }
  std::cout << std::endl;
  for (int i = 0; i < MATRIX_M; i++) {
    for (int j = 0; j < MATRIX_N; j++)
      std::cout << D[i][j] << ", ";
    std::cout << "\n";
  }
}
#endif // (SYCL_EXT_ONEAPI_MATRIX == 2)