xref: /aosp_15_r20/external/ComputeLibrary/cl_kernels/nhwc/direct_convolution3d.clembed (revision c217d954acce2dbc11938adb493fc0abd69584f3)

R"(


#ifndef ARM_COMPUTE_HELPER_H
#define ARM_COMPUTE_HELPER_H




#define STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    VSTORE(N0)                                                 \
    (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

#define STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

#define STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

#define STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

#define STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

#define STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

#define STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

#define STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

#define STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                 \
    (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

#define STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
    VSTORE(N0)                                                  \
    (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

#define STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                  \
    (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

#define STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                  \
    (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

#define STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                  \
    (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

#define STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                  \
    (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

#define STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                  \
    (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

#define STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                  \
    (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));



#define CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##0), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

#define CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##1), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

#define CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##2), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

#define CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##3), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

#define CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##4), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

#define CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##5), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

#define CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##6), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

#define CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##7), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

#define CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                         \
    (CONVERT_SAT((BASENAME##8), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

#define CONVERT_STORE_ROW_10(N0, DATA, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    VSTORE(N0)                                                     \
    (CONVERT_SAT((BASENAME##9), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

#define CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                          \
    (CONVERT_SAT((BASENAME##A), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

#define CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                          \
    (CONVERT_SAT((BASENAME##B), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

#define CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                          \
    (CONVERT_SAT((BASENAME##C), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

#define CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                          \
    (CONVERT_SAT((BASENAME##D), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

#define CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                          \
    (CONVERT_SAT((BASENAME##E), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

#define CONVERT_STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE(N0)                                                          \
    (CONVERT_SAT((BASENAME##F), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));




#define STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)



#define CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)



#define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

#define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

#define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

#define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

#define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

#define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

#define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

#define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

#define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
    (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

#define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
    (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

#define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
    (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

#define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
    (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

#define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
    (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

#define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
    (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

#define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
    (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

#define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
    STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
    (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));



#define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)

#define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
    if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y))                                                                                                            \
    {                                                                                                                                                     \
        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                                           \
    }                                                                                                                                                     \
    else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X))                                                                                                        \
    {                                                                                                                                                     \
        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
    }                                                                                                                                                     \
    else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X))                                                                                                        \
    {                                                                                                                                                     \
        STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
    }                                                                                                                                                     \
    else                                                                                                                                                  \
    {                                                                                                                                                     \
        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                               \
    }

#define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) \
    if(!(PARTIAL_COND_X))                                                                                         \
    {                                                                                                             \
        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
    }                                                                                                             \
    else                                                                                                          \
    {                                                                                                             \
        STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
    }

#define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \
    if(!(PARTIAL_COND_Y))                                                                                         \
    {                                                                                                             \
        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
    }                                                                                                             \
    else                                                                                                          \
    {                                                                                                             \
        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
    }


#if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)


#if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0

#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
    STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)

#elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0

#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
    STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y)

#elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0

#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
    STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X)

#else

#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
    STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X)

#endif

#endif


#if defined(PARTIAL_STORE_M0)

#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
    ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0))))
#else
#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
    ((uint)(y * M0))
#endif



#define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond) \
    STORE_BLOCK_PARTIAL_IN_X(1, vec_size, data_type, basename, ptr, 0, 0, leftover, cond)


#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#endif

#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#pragma OPENCL EXTENSION cl_arm_integer_dot_product_int8 : enable
#endif

#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#pragma OPENCL EXTENSION cl_arm_integer_dot_product_accumulate_int8 : enable
#endif

#if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)
#pragma OPENCL EXTENSION cl_arm_printf : enable
#endif

#define GPU_ARCH_MIDGARD 0x100
#define GPU_ARCH_BIFROST 0x200
#define GPU_ARCH_VALHALL 0x300


#define CONCAT(a, b) a##b


#define EXPAND(x) x


#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)


#define REV1(x) ((x))
#define REV2(x) ((x).s10)
#define REV3(x) ((x).s210)
#define REV4(x) ((x).s3210)
#define REV8(x) ((x).s76543210)
#define REV16(x) ((x).sFEDCBA9876543210)



#define REVERSE_STR(x, s) REV##s((x))
#define REVERSE(x, s) REVERSE_STR(x, s)



#define ROT1_0(x) ((x))
#define ROT1_1(x) ((x))

#define ROT2_0(x) ((x))
#define ROT2_1(x) ((x).s10)
#define ROT2_2(x) ((x))

#define ROT3_0(x) ((x))
#define ROT3_1(x) ((x).s201)
#define ROT3_2(x) ((x).s120)
#define ROT3_3(x) ((x))

#define ROT4_0(x) ((x))
#define ROT4_1(x) ((x).s3012)
#define ROT4_2(x) ((x).s2301)
#define ROT4_3(x) ((x).s1230)
#define ROT4_4(x) ((x))

#define ROT8_0(x) ((x))
#define ROT8_1(x) ((x).s70123456)
#define ROT8_2(x) ((x).s67012345)
#define ROT8_3(x) ((x).s56701234)
#define ROT8_4(x) ((x).s45670123)
#define ROT8_5(x) ((x).s34567012)
#define ROT8_6(x) ((x).s23456701)
#define ROT8_7(x) ((x).s12345670)
#define ROT8_8(x) ((x))

#define ROT16_0(x) ((x))
#define ROT16_1(x) ((x).sF0123456789ABCDE)
#define ROT16_2(x) ((x).sEF0123456789ABCD)
#define ROT16_3(x) ((x).sDEF0123456789ABC)
#define ROT16_4(x) ((x).sCDEF0123456789AB)
#define ROT16_5(x) ((x).sBCDEF0123456789A)
#define ROT16_6(x) ((x).sABCDEF0123456789)
#define ROT16_7(x) ((x).s9ABCDEF012345678)
#define ROT16_8(x) ((x).s89ABCDEF01234567)
#define ROT16_9(x) ((x).s789ABCDEF0123456)
#define ROT16_10(x) ((x).s6789ABCDEF012345)
#define ROT16_11(x) ((x).s56789ABCDEF01234)
#define ROT16_12(x) ((x).s456789ABCDEF0123)
#define ROT16_13(x) ((x).s3456789ABCDEF012)
#define ROT16_14(x) ((x).s23456789ABCDEF01)
#define ROT16_15(x) ((x).s123456789ABCDEF0)
#define ROT16_16(x) ((x))



#define ROTATE_STR(x, s, n) ROT##s##_##n(x)
#define ROTATE(x, s, n) ROTATE_STR(x, s, n)



#define V_OFFS1(dt) (dt##1)(0)
#define V_OFFS2(dt) (dt##2)(0, 1)
#define V_OFFS3(dt) (dt##3)(0, 1, 2)
#define V_OFFS4(dt) (dt##4)(0, 1, 2, 3)
#define V_OFFS8(dt) (dt##8)(0, 1, 2, 3, 4, 5, 6, 7)
#define V_OFFS16(dt) (dt##16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15)



#define VEC_OFFS_STR(dt, s) V_OFFS##s(dt)
#define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s)


#define VLOAD_STR(size) vload##size
#define VLOAD(size) VLOAD_STR(size)


#define VLOAD_PARTIAL_STR(size, load_size) vload_partial_##size##_##load_size
#define VLOAD_PARTIAL(size, load_size) VLOAD_PARTIAL_STR(size, load_size)

#define NO_LOAD(data, offs, ptr) \
    {                            \
    }


#define vload_partial_1_0 NO_LOAD
#define vload_partial_1_1 vload1
#define vload_partial_1_2 NO_LOAD
#define vload_partial_1_3 NO_LOAD
#define vload_partial_1_4 NO_LOAD
#define vload_partial_1_5 NO_LOAD
#define vload_partial_1_6 NO_LOAD
#define vload_partial_1_7 NO_LOAD
#define vload_partial_1_8 NO_LOAD
#define vload_partial_1_9 NO_LOAD
#define vload_partial_1_10 NO_LOAD
#define vload_partial_1_11 NO_LOAD
#define vload_partial_1_12 NO_LOAD
#define vload_partial_1_13 NO_LOAD
#define vload_partial_1_14 NO_LOAD
#define vload_partial_1_15 NO_LOAD
#define vload_partial_1_16 NO_LOAD

#define vload_partial_2_0 NO_LOAD
#define vload_partial_2_1 vload_partial_1
#define vload_partial_2_2 vload_partial_2
#define vload_partial_2_3 NO_LOAD
#define vload_partial_2_4 NO_LOAD
#define vload_partial_2_5 NO_LOAD
#define vload_partial_2_6 NO_LOAD
#define vload_partial_2_7 NO_LOAD
#define vload_partial_2_8 NO_LOAD
#define vload_partial_2_9 NO_LOAD
#define vload_partial_2_10 NO_LOAD
#define vload_partial_2_11 NO_LOAD
#define vload_partial_2_12 NO_LOAD
#define vload_partial_2_13 NO_LOAD
#define vload_partial_2_14 NO_LOAD
#define vload_partial_2_15 NO_LOAD
#define vload_partial_2_16 NO_LOAD

#define vload_partial_3_0 NO_LOAD
#define vload_partial_3_1 vload_partial_1
#define vload_partial_3_2 vload_partial_2
#define vload_partial_3_3 vload_partial_3
#define vload_partial_3_4 NO_LOAD
#define vload_partial_3_5 NO_LOAD
#define vload_partial_3_6 NO_LOAD
#define vload_partial_3_7 NO_LOAD
#define vload_partial_3_8 NO_LOAD
#define vload_partial_3_9 NO_LOAD
#define vload_partial_3_10 NO_LOAD
#define vload_partial_3_11 NO_LOAD
#define vload_partial_3_12 NO_LOAD
#define vload_partial_3_13 NO_LOAD
#define vload_partial_3_14 NO_LOAD
#define vload_partial_3_15 NO_LOAD
#define vload_partial_3_16 NO_LOAD

#define vload_partial_4_0 NO_LOAD
#define vload_partial_4_1 vload_partial_1
#define vload_partial_4_2 vload_partial_2
#define vload_partial_4_3 vload_partial_3
#define vload_partial_4_4 vload_partial_4
#define vload_partial_4_5 NO_LOAD
#define vload_partial_4_6 NO_LOAD
#define vload_partial_4_7 NO_LOAD
#define vload_partial_4_8 NO_LOAD
#define vload_partial_4_9 NO_LOAD
#define vload_partial_4_10 NO_LOAD
#define vload_partial_4_11 NO_LOAD
#define vload_partial_4_12 NO_LOAD
#define vload_partial_4_13 NO_LOAD
#define vload_partial_4_14 NO_LOAD
#define vload_partial_4_15 NO_LOAD
#define vload_partial_4_16 NO_LOAD

#define vload_partial_8_0 NO_LOAD
#define vload_partial_8_1 vload_partial_1
#define vload_partial_8_2 vload_partial_2
#define vload_partial_8_3 vload_partial_3
#define vload_partial_8_4 vload_partial_4
#define vload_partial_8_5 vload_partial_5
#define vload_partial_8_6 vload_partial_6
#define vload_partial_8_7 vload_partial_7
#define vload_partial_8_8 vload_partial_8
#define vload_partial_8_9 NO_LOAD
#define vload_partial_8_10 NO_LOAD
#define vload_partial_8_11 NO_LOAD
#define vload_partial_8_12 NO_LOAD
#define vload_partial_8_13 NO_LOAD
#define vload_partial_8_14 NO_LOAD
#define vload_partial_8_15 NO_LOAD
#define vload_partial_8_16 NO_LOAD

#define vload_partial_16_0 NO_LOAD
#define vload_partial_16_1 vload_partial_1
#define vload_partial_16_2 vload_partial_2
#define vload_partial_16_3 vload_partial_3
#define vload_partial_16_4 vload_partial_4
#define vload_partial_16_5 vload_partial_5
#define vload_partial_16_6 vload_partial_6
#define vload_partial_16_7 vload_partial_7
#define vload_partial_16_8 vload_partial_8
#define vload_partial_16_9 vload_partial_9
#define vload_partial_16_10 vload_partial_10
#define vload_partial_16_11 vload_partial_11
#define vload_partial_16_12 vload_partial_12
#define vload_partial_16_13 vload_partial_13
#define vload_partial_16_14 vload_partial_14
#define vload_partial_16_15 vload_partial_15
#define vload_partial_16_16 vload_partial_16


#define vload_partial_1(DATA, OFFSET, PTR) \
    DATA.s0 = vload1(OFFSET, PTR);

#define vload_partial_2(DATA, OFFSET, PTR) \
    DATA.s01 = vload2(OFFSET, PTR);

#define vload_partial_3(DATA, OFFSET, PTR) \
    DATA.s012 = vload3(OFFSET, PTR);

#define vload_partial_4(DATA, OFFSET, PTR) \
    DATA.s0123 = vload4(OFFSET, PTR);

#define vload_partial_5(DATA, OFFSET, PTR)    \
    vload_partial_4(DATA.s0123, OFFSET, PTR); \
    DATA.s4 = vload1(OFFSET, PTR + 4);

#define vload_partial_6(DATA, OFFSET, PTR)    \
    vload_partial_4(DATA.s0123, OFFSET, PTR); \
    vload_partial_2(DATA.s45, OFFSET, PTR + 4);

#define vload_partial_7(DATA, OFFSET, PTR)    \
    vload_partial_4(DATA.s0123, OFFSET, PTR); \
    vload_partial_3(DATA.s456, OFFSET, PTR + 4);

#define vload_partial_8(DATA, OFFSET, PTR) \
    DATA.s01234567 = vload8(OFFSET, PTR);

#define vload_partial_9(DATA, OFFSET, PTR)        \
    vload_partial_8(DATA.s01234567, OFFSET, PTR); \
    DATA.s8 = vload1(OFFSET, PTR + 8);

#define vload_partial_10(DATA, OFFSET, PTR)       \
    vload_partial_8(DATA.s01234567, OFFSET, PTR); \
    vload_partial_2(DATA.s89, OFFSET, PTR + 8);

#define vload_partial_11(DATA, OFFSET, PTR)       \
    vload_partial_8(DATA.s01234567, OFFSET, PTR); \
    vload_partial_3(DATA.s89A, OFFSET, PTR + 8);

#define vload_partial_12(DATA, OFFSET, PTR)       \
    vload_partial_8(DATA.s01234567, OFFSET, PTR); \
    vload_partial_4(DATA.s89AB, OFFSET, PTR + 8);

#define vload_partial_13(DATA, OFFSET, PTR)       \
    vload_partial_8(DATA.s01234567, OFFSET, PTR); \
    vload_partial_5(DATA.s89ABCDEF, OFFSET, PTR + 8);

#define vload_partial_14(DATA, OFFSET, PTR)       \
    vload_partial_8(DATA.s01234567, OFFSET, PTR); \
    vload_partial_6(DATA.s89ABCDEF, OFFSET, PTR + 8);

#define vload_partial_15(DATA, OFFSET, PTR)       \
    vload_partial_8(DATA.s01234567, OFFSET, PTR); \
    vload_partial_7(DATA.s89ABCDEF, OFFSET, PTR + 8);

#define vload_partial_16(DATA, OFFSET, PTR) \
    DATA = vload16(OFFSET, PTR);



#define PIXEL_UNIT4 1
#define PIXEL_UNIT8 2
#define PIXEL_UNIT16 4


#define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size) PIXEL_UNIT##vec_size
#define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(vec_size) CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size)


#define read_image2d_floatx1(img, x_coord, y_coord) (float4)(read_imagef(img, (int2)(x_coord, y_coord)));
#define read_image2d_floatx2(img, x_coord, y_coord) (float8)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)));
#define read_image2d_floatx4(img, x_coord, y_coord) (float16)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)), read_imagef(img, (int2)(x_coord + 2, y_coord)), read_imagef(img, (int2)(x_coord + 3, y_coord)));

#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
#define read_image2d_halfx1(img, x_coord, y_coord) (half4)(read_imageh(img, (int2)(x_coord, y_coord)));
#define read_image2d_halfx2(img, x_coord, y_coord) (half8)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)));
#define read_image2d_halfx4(img, x_coord, y_coord) (half16)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)), read_imageh(img, (int2)(x_coord + 2, y_coord)), read_imageh(img, (int2)(x_coord + 3, y_coord)));
#endif

#define write_image2d_floatx1(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values));
#define write_image2d_floatx2(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values.s0123), write_imagef(img, (int2)(x_coord + 1, y_coord), values.s4567));
#define write_image2d_floatx4(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values.s0123), write_imagef(img, (int2)(x_coord + 1, y_coord), values.s4567), write_imagef(img, (int2)(x_coord + 2, y_coord), values.s89AB), write_imagef(img, (int2)(x_coord + 3, y_coord), values.sCDEF));

#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
#define write_image2d_halfx1(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values));
#define write_image2d_halfx2(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values.s0123), write_imageh(img, (int2)(x_coord + 1, y_coord), values.s4567));
#define write_image2d_halfx4(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values.s0123), write_imageh(img, (int2)(x_coord + 1, y_coord), values.s4567), write_imageh(img, (int2)(x_coord + 2, y_coord), values.s89AB), write_imageh(img, (int2)(x_coord + 3, y_coord), values.sCDEF));
#endif


#define READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord) read_image2d_##data_type##x##n0(img, x_coord, y_coord)
#define READ_IMAGE2D(data_type, n0, img, x_coord, y_coord) READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord)


#define WRITE_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord, values) write_image2d_##data_type##x##n0(img, x_coord, y_coord, values)
#define WRITE_IMAGE2D(data_type, n0, img, x_coord, y_coord, values) WRITE_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord, values)

#define VSTORE_STR(size) vstore##size
#define VSTORE(size) VSTORE_STR(size)

#define float1 float
#define half1 half
#define char1 char
#define uchar1 uchar
#define short1 short
#define ushort1 ushort
#define int1 int
#define uint1 uint
#define long1 long
#define ulong1 ulong
#define double1 double

#define vload1(OFFSET, PTR) *(OFFSET + PTR)
#define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA


#define VSTORE_PARTIAL_STR(size, store_size) vstore_partial_##size##_##store_size
#define VSTORE_PARTIAL(size, store_size) VSTORE_PARTIAL_STR(size, store_size)

#define NO_STORE(data, offs, ptr) \
    {                             \
    }


#define vstore_partial_1_0 NO_STORE
#define vstore_partial_1_1 vstore1
#define vstore_partial_1_2 NO_STORE
#define vstore_partial_1_3 NO_STORE
#define vstore_partial_1_4 NO_STORE
#define vstore_partial_1_5 NO_STORE
#define vstore_partial_1_6 NO_STORE
#define vstore_partial_1_7 NO_STORE
#define vstore_partial_1_8 NO_STORE
#define vstore_partial_1_9 NO_STORE
#define vstore_partial_1_10 NO_STORE
#define vstore_partial_1_11 NO_STORE
#define vstore_partial_1_12 NO_STORE
#define vstore_partial_1_13 NO_STORE
#define vstore_partial_1_14 NO_STORE
#define vstore_partial_1_15 NO_STORE
#define vstore_partial_1_16 NO_STORE

#define vstore_partial_2_0 NO_STORE
#define vstore_partial_2_1 vstore_partial_1
#define vstore_partial_2_2 vstore_partial_2
#define vstore_partial_2_3 NO_STORE
#define vstore_partial_2_4 NO_STORE
#define vstore_partial_2_5 NO_STORE
#define vstore_partial_2_6 NO_STORE
#define vstore_partial_2_7 NO_STORE
#define vstore_partial_2_8 NO_STORE
#define vstore_partial_2_9 NO_STORE
#define vstore_partial_2_10 NO_STORE
#define vstore_partial_2_11 NO_STORE
#define vstore_partial_2_12 NO_STORE
#define vstore_partial_2_13 NO_STORE
#define vstore_partial_2_14 NO_STORE
#define vstore_partial_2_15 NO_STORE
#define vstore_partial_2_16 NO_STORE

#define vstore_partial_3_0 NO_STORE
#define vstore_partial_3_1 vstore_partial_1
#define vstore_partial_3_2 vstore_partial_2
#define vstore_partial_3_3 vstore_partial_3
#define vstore_partial_3_4 NO_STORE
#define vstore_partial_3_5 NO_STORE
#define vstore_partial_3_6 NO_STORE
#define vstore_partial_3_7 NO_STORE
#define vstore_partial_3_8 NO_STORE
#define vstore_partial_3_9 NO_STORE
#define vstore_partial_3_10 NO_STORE
#define vstore_partial_3_11 NO_STORE
#define vstore_partial_3_12 NO_STORE
#define vstore_partial_3_13 NO_STORE
#define vstore_partial_3_14 NO_STORE
#define vstore_partial_3_15 NO_STORE
#define vstore_partial_3_16 NO_STORE

#define vstore_partial_4_0 NO_STORE
#define vstore_partial_4_1 vstore_partial_1
#define vstore_partial_4_2 vstore_partial_2
#define vstore_partial_4_3 vstore_partial_3
#define vstore_partial_4_4 vstore_partial_4
#define vstore_partial_4_5 NO_STORE
#define vstore_partial_4_6 NO_STORE
#define vstore_partial_4_7 NO_STORE
#define vstore_partial_4_8 NO_STORE
#define vstore_partial_4_9 NO_STORE
#define vstore_partial_4_10 NO_STORE
#define vstore_partial_4_11 NO_STORE
#define vstore_partial_4_12 NO_STORE
#define vstore_partial_4_13 NO_STORE
#define vstore_partial_4_14 NO_STORE
#define vstore_partial_4_15 NO_STORE
#define vstore_partial_4_16 NO_STORE

#define vstore_partial_8_0 NO_STORE
#define vstore_partial_8_1 vstore_partial_1
#define vstore_partial_8_2 vstore_partial_2
#define vstore_partial_8_3 vstore_partial_3
#define vstore_partial_8_4 vstore_partial_4
#define vstore_partial_8_5 vstore_partial_5
#define vstore_partial_8_6 vstore_partial_6
#define vstore_partial_8_7 vstore_partial_7
#define vstore_partial_8_8 vstore_partial_8
#define vstore_partial_8_9 NO_STORE
#define vstore_partial_8_10 NO_STORE
#define vstore_partial_8_11 NO_STORE
#define vstore_partial_8_12 NO_STORE
#define vstore_partial_8_13 NO_STORE
#define vstore_partial_8_14 NO_STORE
#define vstore_partial_8_15 NO_STORE
#define vstore_partial_8_16 NO_STORE

#define vstore_partial_16_0 NO_STORE
#define vstore_partial_16_1 vstore_partial_1
#define vstore_partial_16_2 vstore_partial_2
#define vstore_partial_16_3 vstore_partial_3
#define vstore_partial_16_4 vstore_partial_4
#define vstore_partial_16_5 vstore_partial_5
#define vstore_partial_16_6 vstore_partial_6
#define vstore_partial_16_7 vstore_partial_7
#define vstore_partial_16_8 vstore_partial_8
#define vstore_partial_16_9 vstore_partial_9
#define vstore_partial_16_10 vstore_partial_10
#define vstore_partial_16_11 vstore_partial_11
#define vstore_partial_16_12 vstore_partial_12
#define vstore_partial_16_13 vstore_partial_13
#define vstore_partial_16_14 vstore_partial_14
#define vstore_partial_16_15 vstore_partial_15
#define vstore_partial_16_16 vstore_partial_16


#define vstore_partial_1(DATA, OFFSET, PTR) \
    vstore1(DATA.s0, OFFSET, PTR);

#define vstore_partial_2(DATA, OFFSET, PTR) \
    vstore2(DATA.s01, OFFSET, PTR);

#define vstore_partial_3(DATA, OFFSET, PTR) \
    vstore3(DATA.s012, OFFSET, PTR);

#define vstore_partial_4(DATA, OFFSET, PTR) \
    vstore4(DATA.s0123, OFFSET, PTR);

#define vstore_partial_5(DATA, OFFSET, PTR)    \
    vstore_partial_4(DATA.s0123, OFFSET, PTR); \
    vstore1(DATA.s4, OFFSET, PTR + 4);

#define vstore_partial_6(DATA, OFFSET, PTR)    \
    vstore_partial_4(DATA.s0123, OFFSET, PTR); \
    vstore_partial_2(DATA.s45, OFFSET, PTR + 4);

#define vstore_partial_7(DATA, OFFSET, PTR)    \
    vstore_partial_4(DATA.s0123, OFFSET, PTR); \
    vstore_partial_3(DATA.s456, OFFSET, PTR + 4);

#define vstore_partial_8(DATA, OFFSET, PTR) \
    vstore8(DATA.s01234567, OFFSET, PTR);

#define vstore_partial_9(DATA, OFFSET, PTR)        \
    vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
    vstore1(DATA.s8, OFFSET, PTR + 8);

#define vstore_partial_10(DATA, OFFSET, PTR)       \
    vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
    vstore_partial_2(DATA.s89, OFFSET, PTR + 8);

#define vstore_partial_11(DATA, OFFSET, PTR)       \
    vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
    vstore_partial_3(DATA.s89a, OFFSET, PTR + 8);

#define vstore_partial_12(DATA, OFFSET, PTR)       \
    vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
    vstore_partial_4(DATA.s89ab, OFFSET, PTR + 8);

#define vstore_partial_13(DATA, OFFSET, PTR)       \
    vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
    vstore_partial_5(DATA.s89abcdef, OFFSET, PTR + 8);

#define vstore_partial_14(DATA, OFFSET, PTR)       \
    vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
    vstore_partial_6(DATA.s89abcdef, OFFSET, PTR + 8);

#define vstore_partial_15(DATA, OFFSET, PTR)       \
    vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
    vstore_partial_7(DATA.s89abcdef, OFFSET, PTR + 8);

#define vstore_partial_16(DATA, OFFSET, PTR) \
    vstore16(DATA, OFFSET, PTR);





#define convert_float_sat convert_float
#define convert_float1_sat convert_float
#define convert_float2_sat convert_float2
#define convert_float3_sat convert_float3
#define convert_float4_sat convert_float4
#define convert_float8_sat convert_float8
#define convert_float16_sat convert_float16
#define convert_half_sat convert_float
#define convert_half1_sat convert_half
#define convert_half2_sat convert_half2
#define convert_half3_sat convert_half3
#define convert_half4_sat convert_half4
#define convert_half8_sat convert_half8
#define convert_half16_sat convert_half16

#define convert_float1 convert_float
#define convert_half1 convert_half
#define convert_char1 convert_char
#define convert_uchar1 convert_uchar
#define convert_short1 convert_short
#define convert_ushort1 convert_ushort
#define convert_int1 convert_int
#define convert_uint1 convert_uint
#define convert_long1 convert_long
#define convert_ulong1 convert_ulong
#define convert_double1 convert_double

#define convert_char1_sat convert_char_sat
#define convert_uchar1_sat convert_uchar_sat
#define convert_uchar2_sat convert_uchar2_sat
#define convert_uchar3_sat convert_uchar3_sat
#define convert_uchar4_sat convert_uchar4_sat
#define convert_uchar8_sat convert_uchar8_sat
#define convert_uchar16_sat convert_uchar16_sat
#define convert_short1_sat convert_short_sat
#define convert_ushort1_sat convert_ushort_sat
#define convert_int1_sat convert_int_sat
#define convert_uint1_sat convert_uint_sat
#define convert_long1_sat convert_long_sat
#define convert_ulong1_sat convert_ulong_sat
#define convert_double1_sat convert_double_sat

#define VEC_DATA_TYPE_STR(type, size) type##size
#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)

#define CONVERT_STR(x, type) (convert_##type((x)))
#define CONVERT(x, type) CONVERT_STR(x, type)

#define CONVERT_SAT_STR(x, type) (convert_##type##_sat((x)))
#define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)

#define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x)))
#define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round)

#define select_vec_dt_uchar(size) uchar##size
#define select_vec_dt_char(size) char##size
#define select_vec_dt_ushort(size) ushort##size
#define select_vec_dt_short(size) short##size
#define select_vec_dt_half(size) short##size
#define select_vec_dt_uint(size) uint##size
#define select_vec_dt_int(size) int##size
#define select_vec_dt_float(size) int##size
#define select_vec_dt_ulong(size) ulong##size
#define select_vec_dt_long(size) long##size

#define SELECT_VEC_DATA_TYPE_STR(type, size) select_vec_dt_##type(size)
#define SELECT_VEC_DATA_TYPE(type, size) SELECT_VEC_DATA_TYPE_STR(type, size)
#define SELECT_DATA_TYPE(type) SELECT_VEC_DATA_TYPE_STR(type, 1)

#define signed_int_vec_dt_uchar(size) char##size
#define signed_int_vec_dt_char(size) char##size
#define signed_int_vec_dt_ushort(size) short##size
#define signed_int_vec_dt_short(size) short##size
#define signed_int_vec_dt_half(size) short##size
#define signed_int_vec_dt_uint(size) int##size
#define signed_int_vec_dt_int(size) int##size
#define signed_int_vec_dt_float(size) int##size
#define signed_int_vec_dt_ulong(size) long##size
#define signed_int_vec_dt_long(size) long##size

#define SIGNED_INT_VEC_DATA_TYPE_STR(type, size) signed_int_vec_dt_##type(size)
#define SIGNED_INT_VEC_DATA_TYPE(type, size) SIGNED_INT_VEC_DATA_TYPE_STR(type, size)
#define SIGNED_INT_DATA_TYPE(type) SIGNED_INT_VEC_DATA_TYPE_STR(type, 1)

#define sum_reduce_1(x) (x)
#define sum_reduce_2(x) ((x).s0) + ((x).s1)
#define sum_reduce_3(x) sum_reduce_2((x).s01) + ((x).s2)
#define sum_reduce_4(x) sum_reduce_2((x).s01) + sum_reduce_2((x).s23)
#define sum_reduce_8(x) sum_reduce_4((x).s0123) + sum_reduce_4((x).s4567)
#define sum_reduce_16(x) sum_reduce_8((x).s01234567) + sum_reduce_8((x).s89ABCDEF)

#define SUM_REDUCE_STR(x, size) sum_reduce_##size(x)
#define SUM_REDUCE(x, size) SUM_REDUCE_STR(x, size)

#define prod_reduce_1(x) (x)
#define prod_reduce_2(x) ((x).s0) * ((x).s1)
#define prod_reduce_3(x) prod_reduce_2((x).s01) * ((x).s2)
#define prod_reduce_4(x) prod_reduce_2((x).s01) * prod_reduce_2((x).s23)
#define prod_reduce_8(x) prod_reduce_4((x).s0123) * prod_reduce_4((x).s4567)
#define prod_reduce_16(x) prod_reduce_8((x).s01234567) * prod_reduce_8((x).s89ABCDEF)

#define PROD_REDUCE_STR(x, size) prod_reduce_##size(x)
#define PROD_REDUCE(x, size) PROD_REDUCE_STR(x, size)

#define max_reduce_1(x) (x)
#define max_reduce_2(x) max(((x).s0), ((x).s1))
#define max_reduce_3(x) max(max_reduce_2((x).s01), ((x).s2))
#define max_reduce_4(x) max(max_reduce_2((x).s01), max_reduce_2((x).s23))
#define max_reduce_8(x) max(max_reduce_4((x).s0123), max_reduce_4((x).s4567))
#define max_reduce_16(x) max(max_reduce_8((x).s01234567), max_reduce_8((x).s89ABCDEF))

#define MAX_REDUCE_STR(x, size) max_reduce_##size(x)
#define MAX_REDUCE(x, size) MAX_REDUCE_STR(x, size)

#define VECTOR_DECLARATION(name)     \
    __global uchar *name##_ptr,      \
    uint        name##_stride_x, \
    uint        name##_step_x,   \
    uint        name##_offset_first_element_in_bytes

#define IMAGE_DECLARATION(name)      \
    __global uchar *name##_ptr,      \
    uint        name##_stride_x, \
    uint        name##_step_x,   \
    uint        name##_stride_y, \
    uint        name##_step_y,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR3D_DECLARATION(name)   \
    __global uchar *name##_ptr,      \
    uint        name##_stride_x, \
    uint        name##_step_x,   \
    uint        name##_stride_y, \
    uint        name##_step_y,   \
    uint        name##_stride_z, \
    uint        name##_step_z,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR4D_DECLARATION(name)   \
    __global uchar *name##_ptr,      \
    uint        name##_stride_x, \
    uint        name##_step_x,   \
    uint        name##_stride_y, \
    uint        name##_step_y,   \
    uint        name##_stride_z, \
    uint        name##_step_z,   \
    uint        name##_stride_w, \
    uint        name##_step_w,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR5D_DECLARATION(name)   \
    __global uchar *name##_ptr,      \
    uint        name##_stride_x, \
    uint        name##_step_x,   \
    uint        name##_stride_y, \
    uint        name##_step_y,   \
    uint        name##_stride_z, \
    uint        name##_step_z,   \
    uint        name##_stride_w, \
    uint        name##_step_w,   \
    uint        name##_stride_v, \
    uint        name##_step_v,   \
    uint        name##_offset_first_element_in_bytes

#define CONVERT_TO_VECTOR_STRUCT(name) \
    update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)

#define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \
    update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0)

#define CONVERT_TO_IMAGE_STRUCT(name) \
    update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)

#define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \
    update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0)

#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
    update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \
    update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, name##_step_z)

#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
    update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

#define CONVERT_TO_TENSOR3D_STRUCT(name)                                                                                                           \
    update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                 name##_stride_z, name##_step_z)

#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \
    update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0)

#define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size)                                                                                                 \
    update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                 name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size)

#define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \
    update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size)

#define CONVERT_TO_TENSOR3D_STRUCT_NO_UPDATE_PTR(name)                                                                                       \
    tensor3D_ptr_no_update(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                           name##_stride_z, name##_step_z)


typedef struct Vector
{
    __global uchar *ptr;
    int             offset_first_element_in_bytes;
    int             stride_x;
} Vector;


typedef struct Image
{
    __global uchar *ptr;
    int             offset_first_element_in_bytes;
    int             stride_x;
    int             stride_y;
} Image;


typedef struct Tensor3D
{
    __global uchar *ptr;
    int             offset_first_element_in_bytes;
    int             stride_x;
    int             stride_y;
    int             stride_z;
} Tensor3D;


typedef struct Tensor4D
{
    __global uchar *ptr;
    int             offset_first_element_in_bytes;
    int             stride_x;
    int             stride_y;
    int             stride_z;
    int             stride_w;
} Tensor4D;


inline Vector update_vector_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x)
{
    Vector vector =
    {
        .ptr                           = ptr,
        .offset_first_element_in_bytes = offset_first_element_in_bytes,
        .stride_x                      = stride_x,
    };
    vector.ptr += vector.offset_first_element_in_bytes + get_global_id(0) * step_x;
    return vector;
}


inline Image update_image_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
{
    Image img =
    {
        .ptr                           = ptr,
        .offset_first_element_in_bytes = offset_first_element_in_bytes,
        .stride_x                      = stride_x,
        .stride_y                      = stride_y
    };
    img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y;
    return img;
}


inline Image update_image_from_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
{
    Image img =
    {
        .ptr                           = ptr,
        .offset_first_element_in_bytes = offset_first_element_in_bytes,
        .stride_x                      = stride_x,
        .stride_y                      = stride_y
    };
    img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
    return img;
}


inline Tensor3D update_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
{
    Tensor3D tensor =
    {
        .ptr                           = ptr,
        .offset_first_element_in_bytes = offset_first_element_in_bytes,
        .stride_x                      = stride_x,
        .stride_y                      = stride_y,
        .stride_z                      = stride_z
    };
    tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
    return tensor;
}


inline Tensor3D tensor3D_ptr_no_update(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
{
    Tensor3D tensor =
    {
        .ptr                           = ptr,
        .offset_first_element_in_bytes = offset_first_element_in_bytes,
        .stride_x                      = stride_x,
        .stride_y                      = stride_y,
        .stride_z                      = stride_z
    };
    return tensor;
}

inline Tensor4D update_tensor4D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z, uint stride_w,
                                             uint step_w,
                                             uint mod_size)
{
    Tensor4D tensor =
    {
        .ptr                           = ptr,
        .offset_first_element_in_bytes = offset_first_element_in_bytes,
        .stride_x                      = stride_x,
        .stride_y                      = stride_y,
        .stride_z                      = stride_z,
        .stride_w                      = stride_w
    };

    tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + (get_global_id(2) % mod_size) * step_z + (get_global_id(2) / mod_size) * step_w;
    return tensor;
}


inline __global const uchar *vector_offset(const Vector *vec, int x)
{
    return vec->ptr + x * vec->stride_x;
}


inline __global uchar *offset(const Image *img, int x, int y)
{
    return img->ptr + x * img->stride_x + y * img->stride_y;
}


inline __global const uchar *tensor3D_offset(const Tensor3D *tensor, int x, int y, int z)
{
    return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z;
}


inline __global const uchar *tensor4D_offset(const Tensor4D *tensor, int x, int y, int z, int w)
{
    return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + w * tensor->stride_w;
}


inline __global const uchar *tensor3D_index2ptr(const Tensor3D *tensor, uint width, uint height, uint depth, uint index)
{
    uint num_elements = width * height;

    const uint z = index / num_elements;

    index %= num_elements;

    const uint y = index / width;

    index %= width;

    const uint x = index;

    return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + tensor->offset_first_element_in_bytes;
}

#endif

#ifndef SRC_CORE_CL_CL_KERNELS_TILE_HELPERS
#define SRC_CORE_CL_CL_KERNELS_TILE_HELPERS




#define TILE_VECTOR_SIZE1 1
#define TILE_VECTOR_SIZE2 2
#define TILE_VECTOR_SIZE3 3
#define TILE_VECTOR_SIZE4 4
#define TILE_VECTOR_SIZE5 8
#define TILE_VECTOR_SIZE6 8
#define TILE_VECTOR_SIZE7 8
#define TILE_VECTOR_SIZE8 8
#define TILE_VECTOR_SIZE9 16
#define TILE_VECTOR_SIZE10 16
#define TILE_VECTOR_SIZE11 16
#define TILE_VECTOR_SIZE12 16
#define TILE_VECTOR_SIZE13 16
#define TILE_VECTOR_SIZE14 16
#define TILE_VECTOR_SIZE15 16
#define TILE_VECTOR_SIZE16 16

#define TILE_VECTOR_TYPE1(DATA_TYPE) DATA_TYPE##1
#define TILE_VECTOR_TYPE2(DATA_TYPE) DATA_TYPE##2
#define TILE_VECTOR_TYPE3(DATA_TYPE) DATA_TYPE##3
#define TILE_VECTOR_TYPE4(DATA_TYPE) DATA_TYPE##4
#define TILE_VECTOR_TYPE5(DATA_TYPE) DATA_TYPE##8
#define TILE_VECTOR_TYPE6(DATA_TYPE) DATA_TYPE##8
#define TILE_VECTOR_TYPE7(DATA_TYPE) DATA_TYPE##8
#define TILE_VECTOR_TYPE8(DATA_TYPE) DATA_TYPE##8
#define TILE_VECTOR_TYPE9(DATA_TYPE) DATA_TYPE##16
#define TILE_VECTOR_TYPE10(DATA_TYPE) DATA_TYPE##16
#define TILE_VECTOR_TYPE11(DATA_TYPE) DATA_TYPE##16
#define TILE_VECTOR_TYPE12(DATA_TYPE) DATA_TYPE##16
#define TILE_VECTOR_TYPE13(DATA_TYPE) DATA_TYPE##16
#define TILE_VECTOR_TYPE14(DATA_TYPE) DATA_TYPE##16
#define TILE_VECTOR_TYPE15(DATA_TYPE) DATA_TYPE##16
#define TILE_VECTOR_TYPE16(DATA_TYPE) DATA_TYPE##16


#define TILE(DATA_TYPE, H, W, BASENAME) TILE_STR(DATA_TYPE, H, W, BASENAME)
#define TILE_STR(DATA_TYPE, H, W, BASENAME) \
    union {                                 \
        DATA_TYPE                      s[TILE_VECTOR_SIZE##W];                  \
        TILE_VECTOR_TYPE##W(DATA_TYPE) v;                     \
    } BASENAME[H]

#define TENSOR4D_IMAGE(name)          \
    __read_only image2d_t name##_img, \
    __global uchar *name##_ptr,       \
    uint            name##_stride_x,  \
    uint            name##_step_x,    \
    uint            name##_stride_y,  \
    uint            name##_step_y,    \
    uint            name##_stride_z,  \
    uint            name##_step_z,    \
    uint            name##_stride_w,  \
    uint            name##_step_w,    \
    uint            name##_offset_first_element_in_bytes

#define TENSOR4D_BUFFER(name)    \
    __global uchar *name##_ptr,  \
    uint        name##_stride_x, \
    uint        name##_step_x,   \
    uint        name##_stride_y, \
    uint        name##_step_y,   \
    uint        name##_stride_z, \
    uint        name##_step_z,   \
    uint        name##_stride_w, \
    uint        name##_step_w,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR4D_STR(name, type) TENSOR4D_##type(name)
#define TENSOR4D(name, type) TENSOR4D_STR(name, type)

#define TENSOR4D_T_IMAGE(name)          \
    __read_only image2d_t name##_img, \
    __global uchar *name##_ptr,       \
    uint        name##_stride_y, \
    uint        name##_stride_z, \
    uint        name##_stride_w, \
    uint        name##_c,   \
    uint        name##_w,   \
    uint        name##_h,   \
    uint        name##_n,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR4D_T_BUFFER(name)    \
    __global uchar *name##_ptr,  \
    uint        name##_stride_y, \
    uint        name##_stride_z, \
    uint        name##_stride_w, \
    uint        name##_c,   \
    uint        name##_w,   \
    uint        name##_h,   \
    uint        name##_n,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR4D_T_STR(name, type) TENSOR4D_T_##type(name)


#define TENSOR4D_T(name, type) TENSOR4D_T_STR(name, type)

#define TENSOR4D_RO_T_IMAGE(name)          \
    __read_only image2d_t name##_img, \
    TENSOR4D_T_BUFFER(name)

#define TENSOR4D_RO_T_BUFFER(name) TENSOR4D_T_BUFFER(name)

#define TENSOR4D_RO_T_STR(name, type) TENSOR4D_RO_T_##type(name)


#define TENSOR4D_RO_T(name, type) TENSOR4D_RO_T_STR(name, type)

#define TENSOR4D_WO_T_IMAGE(name)          \
    __write_only image2d_t name##_img, \
    TENSOR4D_T_BUFFER(name)

#define TENSOR4D_WO_T_BUFFER(name) TENSOR4D_T_BUFFER(name)

#define TENSOR4D_WO_T_STR(name, type) TENSOR4D_WO_T_##type(name)


#define TENSOR4D_WO_T(name, type) TENSOR4D_WO_T_STR(name, type)

#define TENSOR3D_T_IMAGE(name)          \
    __read_only image2d_t name##_img, \
    __global uchar *name##_ptr,       \
    uint        name##_stride_y, \
    uint        name##_stride_z, \
    uint        name##_w,   \
    uint        name##_h,   \
    uint        name##_n,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR3D_T_BUFFER(name)    \
    __global uchar *name##_ptr,  \
    uint        name##_stride_y, \
    uint        name##_stride_z, \
    uint        name##_w,   \
    uint        name##_h,   \
    uint        name##_n,   \
    uint        name##_offset_first_element_in_bytes

#define TENSOR3D_T_STR(name, type) TENSOR3D_T_##type(name)
#define TENSOR3D_T(name, type) TENSOR3D_T_STR(name, type)

#if !defined(UNROLL_WITH_PRAGMA)
#define UNROLL_INCR(idx, step, macro) idx += (step); (macro)

#define LOOP_UNROLLING_1(idx, step, macro) (macro)
#define LOOP_UNROLLING_2(idx, step, macro) LOOP_UNROLLING_1(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_3(idx, step, macro) LOOP_UNROLLING_2(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_4(idx, step, macro) LOOP_UNROLLING_3(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_5(idx, step, macro) LOOP_UNROLLING_4(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_6(idx, step, macro) LOOP_UNROLLING_5(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_7(idx, step, macro) LOOP_UNROLLING_6(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_8(idx, step, macro) LOOP_UNROLLING_7(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_9(idx, step, macro) LOOP_UNROLLING_8(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_10(idx, step, macro) LOOP_UNROLLING_9(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_11(idx, step, macro) LOOP_UNROLLING_10(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_12(idx, step, macro) LOOP_UNROLLING_11(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_13(idx, step, macro) LOOP_UNROLLING_12(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_14(idx, step, macro) LOOP_UNROLLING_13(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_15(idx, step, macro) LOOP_UNROLLING_14(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_16(idx, step, macro) LOOP_UNROLLING_15(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_17(idx, step, macro) LOOP_UNROLLING_16(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_18(idx, step, macro) LOOP_UNROLLING_17(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_19(idx, step, macro) LOOP_UNROLLING_18(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_20(idx, step, macro) LOOP_UNROLLING_19(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_21(idx, step, macro) LOOP_UNROLLING_20(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_22(idx, step, macro) LOOP_UNROLLING_21(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_23(idx, step, macro) LOOP_UNROLLING_22(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_24(idx, step, macro) LOOP_UNROLLING_23(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_25(idx, step, macro) LOOP_UNROLLING_24(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_26(idx, step, macro) LOOP_UNROLLING_25(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_27(idx, step, macro) LOOP_UNROLLING_26(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_28(idx, step, macro) LOOP_UNROLLING_27(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_29(idx, step, macro) LOOP_UNROLLING_28(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_30(idx, step, macro) LOOP_UNROLLING_29(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_31(idx, step, macro) LOOP_UNROLLING_30(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_32(idx, step, macro) LOOP_UNROLLING_31(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_33(idx, step, macro) LOOP_UNROLLING_32(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_34(idx, step, macro) LOOP_UNROLLING_33(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_35(idx, step, macro) LOOP_UNROLLING_34(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_36(idx, step, macro) LOOP_UNROLLING_35(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_37(idx, step, macro) LOOP_UNROLLING_36(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_38(idx, step, macro) LOOP_UNROLLING_37(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_39(idx, step, macro) LOOP_UNROLLING_38(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_40(idx, step, macro) LOOP_UNROLLING_39(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_41(idx, step, macro) LOOP_UNROLLING_40(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_42(idx, step, macro) LOOP_UNROLLING_41(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_43(idx, step, macro) LOOP_UNROLLING_42(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_44(idx, step, macro) LOOP_UNROLLING_43(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_45(idx, step, macro) LOOP_UNROLLING_44(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_46(idx, step, macro) LOOP_UNROLLING_45(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_47(idx, step, macro) LOOP_UNROLLING_46(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_48(idx, step, macro) LOOP_UNROLLING_47(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_49(idx, step, macro) LOOP_UNROLLING_48(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_50(idx, step, macro) LOOP_UNROLLING_49(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_51(idx, step, macro) LOOP_UNROLLING_50(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_52(idx, step, macro) LOOP_UNROLLING_51(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_53(idx, step, macro) LOOP_UNROLLING_52(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_54(idx, step, macro) LOOP_UNROLLING_53(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_55(idx, step, macro) LOOP_UNROLLING_54(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_56(idx, step, macro) LOOP_UNROLLING_55(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_57(idx, step, macro) LOOP_UNROLLING_56(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_58(idx, step, macro) LOOP_UNROLLING_57(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_59(idx, step, macro) LOOP_UNROLLING_58(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_60(idx, step, macro) LOOP_UNROLLING_59(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_61(idx, step, macro) LOOP_UNROLLING_60(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_62(idx, step, macro) LOOP_UNROLLING_61(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_63(idx, step, macro) LOOP_UNROLLING_62(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_64(idx, step, macro) LOOP_UNROLLING_63(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_65(idx, step, macro) LOOP_UNROLLING_64(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_66(idx, step, macro) LOOP_UNROLLING_65(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_67(idx, step, macro) LOOP_UNROLLING_66(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_68(idx, step, macro) LOOP_UNROLLING_67(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_69(idx, step, macro) LOOP_UNROLLING_68(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_70(idx, step, macro) LOOP_UNROLLING_69(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_71(idx, step, macro) LOOP_UNROLLING_70(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_72(idx, step, macro) LOOP_UNROLLING_71(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_73(idx, step, macro) LOOP_UNROLLING_72(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_74(idx, step, macro) LOOP_UNROLLING_73(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_75(idx, step, macro) LOOP_UNROLLING_74(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_76(idx, step, macro) LOOP_UNROLLING_75(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_77(idx, step, macro) LOOP_UNROLLING_76(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_78(idx, step, macro) LOOP_UNROLLING_77(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_79(idx, step, macro) LOOP_UNROLLING_78(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_80(idx, step, macro) LOOP_UNROLLING_79(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_81(idx, step, macro) LOOP_UNROLLING_80(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_82(idx, step, macro) LOOP_UNROLLING_81(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_83(idx, step, macro) LOOP_UNROLLING_82(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_84(idx, step, macro) LOOP_UNROLLING_83(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_85(idx, step, macro) LOOP_UNROLLING_84(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_86(idx, step, macro) LOOP_UNROLLING_85(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_87(idx, step, macro) LOOP_UNROLLING_86(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_88(idx, step, macro) LOOP_UNROLLING_87(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_89(idx, step, macro) LOOP_UNROLLING_88(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_90(idx, step, macro) LOOP_UNROLLING_89(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_91(idx, step, macro) LOOP_UNROLLING_90(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_92(idx, step, macro) LOOP_UNROLLING_91(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_93(idx, step, macro) LOOP_UNROLLING_92(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_94(idx, step, macro) LOOP_UNROLLING_93(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_95(idx, step, macro) LOOP_UNROLLING_94(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_96(idx, step, macro) LOOP_UNROLLING_95(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_97(idx, step, macro) LOOP_UNROLLING_96(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_98(idx, step, macro) LOOP_UNROLLING_97(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_99(idx, step, macro) LOOP_UNROLLING_98(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_100(idx, step, macro) LOOP_UNROLLING_99(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_101(idx, step, macro) LOOP_UNROLLING_100(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_102(idx, step, macro) LOOP_UNROLLING_101(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_103(idx, step, macro) LOOP_UNROLLING_102(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_104(idx, step, macro) LOOP_UNROLLING_103(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_105(idx, step, macro) LOOP_UNROLLING_104(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_106(idx, step, macro) LOOP_UNROLLING_105(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_107(idx, step, macro) LOOP_UNROLLING_106(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_108(idx, step, macro) LOOP_UNROLLING_107(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_109(idx, step, macro) LOOP_UNROLLING_108(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_110(idx, step, macro) LOOP_UNROLLING_109(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_111(idx, step, macro) LOOP_UNROLLING_110(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_112(idx, step, macro) LOOP_UNROLLING_111(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_113(idx, step, macro) LOOP_UNROLLING_112(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_114(idx, step, macro) LOOP_UNROLLING_113(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_115(idx, step, macro) LOOP_UNROLLING_114(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_116(idx, step, macro) LOOP_UNROLLING_115(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_117(idx, step, macro) LOOP_UNROLLING_116(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_118(idx, step, macro) LOOP_UNROLLING_117(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_119(idx, step, macro) LOOP_UNROLLING_118(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_120(idx, step, macro) LOOP_UNROLLING_119(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_121(idx, step, macro) LOOP_UNROLLING_120(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_122(idx, step, macro) LOOP_UNROLLING_121(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_123(idx, step, macro) LOOP_UNROLLING_122(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_124(idx, step, macro) LOOP_UNROLLING_123(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_125(idx, step, macro) LOOP_UNROLLING_124(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_126(idx, step, macro) LOOP_UNROLLING_125(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_127(idx, step, macro) LOOP_UNROLLING_126(idx, step, macro); UNROLL_INCR(idx, step, macro)
#define LOOP_UNROLLING_128(idx, step, macro) LOOP_UNROLLING_127(idx, step, macro); UNROLL_INCR(idx, step, macro)

#define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
    {                                                          \
        type idx = start;                                      \
        LOOP_UNROLLING_##num(idx, step, macro);                \
    }
#else
#define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
    {                                                          \
        _Pragma("unroll")                                      \
        for(type idx = start; idx < (num * step); idx += step) \
        {                                                      \
            (macro);                                           \
        }                                                      \
    }
#endif
#define LOOP_UNROLLING(type, idx, start, step, num, macro) LOOP_UNROLLING_STR(type, idx, start, step, num, macro)


#define GET_SPATIAL_IDX(IDX, N0, PARTIAL_N0) (max((int)(get_global_id(IDX) * N0 - (N0 - PARTIAL_N0) % N0), 0))


#define DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c)
#define DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT##K0##_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c)
#define DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        c += (C_DATA_TYPE)(a) * (C_DATA_TYPE)(b);     \
    })
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_khr_integer_dot_product)
#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)));
#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0));
#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((a), (b));
#elif defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)), (c));
#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0), (c));
#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((a), (b), (c));
#elif defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)));
#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0));
#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((a), (b));
#else
#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c)   \
    ({                                                  \
        c += (C_DATA_TYPE)(a).s0 * (C_DATA_TYPE)(b).s0; \
        c += (C_DATA_TYPE)(a).s1 * (C_DATA_TYPE)(b).s1; \
    })
#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c)   \
    ({                                                  \
        DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c);  \
        c += (C_DATA_TYPE)(a).s2 * (C_DATA_TYPE)(b).s2; \
    })
#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, x, y, val)   \
    ({                                                    \
        val += (C_DATA_TYPE)(x).s0 * (C_DATA_TYPE)(y).s0; \
        val += (C_DATA_TYPE)(x).s1 * (C_DATA_TYPE)(y).s1; \
        val += (C_DATA_TYPE)(x).s2 * (C_DATA_TYPE)(y).s2; \
        val += (C_DATA_TYPE)(x).s3 * (C_DATA_TYPE)(y).s3; \
    })
#endif
#define DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c);     \
        DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s4), ((b).s4), c);     \
    })
#define DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c);     \
        DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s45), ((b).s45), c);     \
    })
#define DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c);     \
        DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s456), ((b).s456), c);     \
    })
#define DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c);     \
        DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c);     \
    })
#define DOT_PRODUCT9_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c);     \
        DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s8), ((b).s8), c);     \
    })
#define DOT_PRODUCT10_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c);     \
        DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89), ((b).s89), c);     \
    })
#define DOT_PRODUCT11_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c);     \
        DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89A), ((b).s89A), c);     \
    })
#define DOT_PRODUCT12_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c);     \
        DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89AB), ((b).s89AB), c);     \
    })
#define DOT_PRODUCT13_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c);     \
        DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABC), ((b).s89ABC), c);     \
    })
#define DOT_PRODUCT14_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c);     \
        DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCD), ((b).s89ABCD), c);     \
    })
#define DOT_PRODUCT15_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c);     \
        DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCDE), ((b).s89ABCDE), c);     \
    })
#define DOT_PRODUCT16_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
    ({                                                 \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c);      \
        DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c);      \
    })


#define REDUCE_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c)
#define REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, (TILE_VECTOR_TYPE##K0(B_DATA_TYPE))1, c)


#define V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y)
#define V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y)
#define V_LOAD_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) \
    VLOAD(WIDTH)                                                \
    (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * (STRIDE_Y)))
#define V_LOAD_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) READ_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y))


#define V_STORE(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) V_STORE_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES)
#define V_STORE_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) V_STORE_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES)
#define V_STORE_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) \
    VSTORE(WIDTH)                                                \
    (VALUES, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * (STRIDE_Y)))
#define V_STORE_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) WRITE_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y), VALUES)


#define T_LOAD(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, Y, YI_MULTIPLIER, STRIDE_Y, dst)                      \
    ({                                                                                                                 \
        LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                          \
        {                                                                                                              \
            dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, ((Y) + _i * (int)(YI_MULTIPLIER)), STRIDE_Y); \
        })                                                                                                             \
    })


#define T_LOAD_INDIRECT(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, STRIDE_Y, indirect_y, dst)    \
    ({                                                                                                  \
        LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                           \
        {                                                                                               \
            dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, (indirect_y[_i].v), STRIDE_Y); \
        })                                                                                              \
    })


#define T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, dst, indirect_y)                                                      \
    ({                                                                                                                                                                                             \
        if(WIDTH1_CONDITION)                                                                                                                                                                       \
        {                                                                                                                                                                                          \
            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
            {                                                                                                                                                                                      \
                VLOAD_PARTIAL(WIDTH0, WIDTH1)                                                         \
                (dst[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y));               \
            })                                                                                                                                                                                     \
        }                                                                                                                                                                                          \
        else                                                                                                                                                                                       \
        {                                                                                                                                                                                          \
            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
            {                                                                                                                                                                                      \
                dst[HEIGHT - 1 - _i].v = V_LOAD(DATA_TYPE, WIDTH0, TENSOR_TYPE, TENSOR, X, (indirect_y[HEIGHT - 1 - _i].v), STRIDE_Y); \
            })                                                                                                                                                                                     \
        }                                                                                                                                                                                          \
    })

#define T_LOAD_NHWC(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, dst)   \
    ({                                                                                                                                                \
        LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT,                                                                                                   \
        {                                                                                                                                             \
            LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH,                                                                                                \
            {                                                                                                                                         \
                int _src_y = (X) + _xk + ((Y) + _yk) * (TENSOR_WIDTH);                                                                                \
                _src_y    += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT);                                                                        \
                int _src_valid_y = (((X) + _xk) >= 0 && ((X) + _xk) < (int)(TENSOR_WIDTH) && ((Y) + _yk) >= 0 && ((Y) + _yk) < (int)(TENSOR_HEIGHT)); \
                if(_src_valid_y != 0)                                                                                                                 \
                {                                                                                                                                     \
                    dst[_xk + _yk * (TILE_WIDTH)].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y);                     \
                }                                                                                                                                     \
            })                                                                                                                                        \
        })                                                                                                                                            \
    })


#define T_LOAD_NHWC_WITH_DILATION(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, DILATION_X, DILATION_Y, BOUNDARY_CHECK, dst)         \
    ({ \
        LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \
        { \
            LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \
            { \
                int _src_y = (X) + _xk * (DILATION_X); \
                int _src_z = ((Y) + _yk * (DILATION_Y)); \
                int _src_w    = (B); \
                bool _src_valid_y = (((X) + _xk * (DILATION_X)) >= 0) && (((X) + _xk * (DILATION_X)) < (int)(TENSOR_WIDTH)) && (((Y) + _yk * (DILATION_Y)) >= 0) && (((Y) + _yk * (DILATION_Y)) < (int)(TENSOR_HEIGHT)); \
                if(!(BOUNDARY_CHECK)) \
                { \
                    dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS)                                                \
                    (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \
                } \
                else \
                { \
                    if(_src_valid_y) \
                    { \
                        dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS)                                                \
                    (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \
                    }                                                                                                                                                                                                 \
                } \
            })                                                                                                                                                                                                             \
        })                                                                                                                                                                                                             \
    })


#define T_LOAD_NHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, xi, yi, dst)                \
    ({                                                                                                                                                                \
        LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA,                                                                                                                      \
        {                                                                                                                                                             \
            int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH);                                                                                          \
            _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT);                                                                                               \
            int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT)); \
            if(_src_valid_y != 0)                                                                                                                                     \
            {                                                                                                                                                         \
                dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y);                                                               \
            }                                                                                                                                                         \
        })                                                                                                                                                            \
    })


#define T_LOAD2D_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) T_LOAD2D_INDIRECT_STR(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst)
#define T_LOAD2D_INDIRECT_STR(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) T_LOAD2D_INDIRECT_##TENSOR_TYPE(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst)
#define T_LOAD2D_INDIRECT_BUFFER(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) \
    ({ \
        LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \
        { \
            if(yi[0].s[_i] >= 0) \
            { \
                dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, yi[0].s[_i], STRIDE_Y); \
            } \
        }) \
    })

#define T_LOAD2D_INDIRECT_IMAGE(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) \
    ({ \
        LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \
        { \
            dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, yi[0].s[_i], STRIDE_Y); \
        }) \
    })


#define T_LOAD_NDHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Z, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, TENSOR_DEPTH, STRIDE_Y, xi, yi, zi, dst) \
    ({                                                                                                                                                                \
        LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA,                                                                                                                      \
        {                                                                                                                                                             \
            int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH) + ((Z) + zi[_i].v) * (TENSOR_WIDTH * TENSOR_HEIGHT);                                      \
            _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT) * (int)(TENSOR_DEPTH);                                                                         \
            int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT)   \
                             && ((Z) + zi[_i].v) >= 0 && ((Z) + zi[_i].v) < (int)(TENSOR_DEPTH));                                                                     \
            if(_src_valid_y != 0)                                                                                                                                     \
            {                                                                                                                                                         \
                dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y);                                                               \
            }                                                                                                                                                         \
        })                                                                                                                                                            \
    })


#define T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, src, indirect_y)                                                      \
    ({                                                                                                                                                                                             \
        if(WIDTH1_CONDITION)                                                                                                                                                                       \
        {                                                                                                                                                                                          \
            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
            {                                                                                                                                                                                      \
                VSTORE_PARTIAL(WIDTH0, WIDTH1)                                                                                                                                                     \
                (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \
            })                                                                                                                                                                                     \
        }                                                                                                                                                                                          \
        else                                                                                                                                                                                       \
        {                                                                                                                                                                                          \
            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
            {                                                                                                                                                                                      \
                VSTORE(WIDTH0)                                                                                                                                                                     \
                (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \
            })                                                                                                                                                                                     \
        }                                                                                                                                                                                          \
    })


#define T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, lhs, rhs, dst)        \
    ({                                                                                               \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                                                           \
        {                                                                                            \
            ACC_DATA_TYPE _tm = 0;                                                                   \
            LOOP_UNROLLING(int, _k0, 0, 1, K0,                                                       \
            {                                                                                        \
                _tm += ((ACC_DATA_TYPE)lhs[_m0].s[_k0] * (ACC_DATA_TYPE)WEI_OFFSET);                 \
            })                                                                                       \
            LOOP_UNROLLING(int, _n0, 0, 1, N0,                                                       \
            {                                                                                        \
                dst[_m0].s[_n0] += _tm;                                                              \
                LOOP_UNROLLING(int, _k0, 0, 1, K0,                                                   \
                {                                                                                    \
                    dst[_m0].s[_n0] += ((ACC_DATA_TYPE)rhs[_n0].s[_k0] * (ACC_DATA_TYPE)SRC_OFFSET); \
                })                                                                                   \
            })                                                                                       \
        })                                                                                          \
    })


#define T_QUANTIZE8(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst)
#define T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_##QUANTIZATION_TYPE(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst)


#define T_QUANTIZE8_PER_TENSOR(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst)                          \
    ({ \
        LOOP_UNROLLING(int, _m0, 0, 1, M0, \
        { \
            LOOP_UNROLLING(int, _n0, 0, 1, N0, \
            { \
                SRC_DATA_TYPE _tmp = 0; \
                SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
                _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \
                SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \
                long a_64 = (long)(_src); \
                long b_64 = (long)(DST_MULTIPLIER); \
                long ab_64 = a_64 * b_64; \
                long mask1 = 1 << 30; \
                long mask2 = 1 - (1 << 30); \
                long is_positive_or_zero = ab_64 >= 0; \
                long nudge = select(mask2, mask1, is_positive_or_zero); \
                SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
                _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
                if(DST_SHIFT >= 0) \
                { \
                    long mask = ((((int)1) << DST_SHIFT) - (long)1); \
                    long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \
                    _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \
                } \
                _tmp += DST_OFFSET; \
                dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE);                                                                            \
            })                                                                                                                                          \
        })                                                                                                                                          \
    })


#define T_QUANTIZE8_PER_CHANNEL(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst)                          \
    ({ \
        LOOP_UNROLLING(int, _m0, 0, 1, M0, \
        { \
            LOOP_UNROLLING(int, _n0, 0, 1, N0, \
            { \
                SRC_DATA_TYPE _tmp = 0; \
                SRC_DATA_TYPE _tmp2 = 0; \
                SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
                SRC_DATA_TYPE _dst_multiplier = dst_multipliers[0].s[_n0]; \
                SRC_DATA_TYPE _dst_shift = dst_shifts[0].s[_n0]; \
                _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-_dst_shift)), ((SRC_DATA_TYPE)_dst_shift < (SRC_DATA_TYPE)0)); \
                SRC_DATA_TYPE overflow = _src == _dst_multiplier && _src == INT_MIN; \
                long a_64 = (long)(_src); \
                long b_64 = (long)(_dst_multiplier); \
                long ab_64 = a_64 * b_64; \
                long mask1 = 1 << 30; \
                long mask2 = 1 - (1 << 30); \
                long is_positive_or_zero = ab_64 >= 0; \
                long nudge = select(mask2, mask1, is_positive_or_zero); \
                SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
                _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
                long mask = ((((int)1) << _dst_shift) - (int)1); \
                long threshold = (mask >> 1) + any(_tmp); \
                _tmp2 = _tmp >> _dst_shift; \
                _tmp2 += select(0, 1, (_tmp & mask) > threshold); \
                _tmp = select(_tmp, _tmp2, _dst_shift >= 0); \
                _tmp += DST_OFFSET; \
                dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE);                                                                            \
            })                                                                                                                                          \
        })                                                                                                                                         \
    })


#define T_QUANTIZE8_ASYMMETRIC(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst)                          \
    ({ \
        LOOP_UNROLLING(int, _m0, 0, 1, M0, \
        { \
            LOOP_UNROLLING(int, _n0, 0, 1, N0, \
            { \
                SRC_DATA_TYPE _tmp = 0; \
                SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
                _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \
                SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \
                long a_64 = (long)(_src); \
                long b_64 = (long)(DST_MULTIPLIER); \
                long ab_64 = a_64 * b_64; \
                long mask1 = 1 << 30; \
                long mask2 = 1 - (1 << 30); \
                long is_positive_or_zero = ab_64 >= 0; \
                long nudge = select(mask2, mask1, is_positive_or_zero); \
                SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
                _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
                if(DST_SHIFT >= 0) \
                { \
                    long mask = ((((int)1) << DST_SHIFT) - (int)1); \
                    long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \
                    _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \
                } \
                _tmp += DST_OFFSET; \
                dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE);                                                                            \
            })                                                                                                                                          \
        })                                                                                                                                          \
    })


#define T_ROWSET_MASK(DATA_TYPE, M0, N0, VALUE_TO_SET, a, mask)                                                                                            \
    ({                                                                                                                                                     \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                                                                                                                 \
        {                                                                                                                                                  \
            LOOP_UNROLLING(int, _n0, 0, 1, N0,                                                                                                             \
            {                                                                                                                                              \
                a[_m0].s[_n0] = select((DATA_TYPE)(a[_m0].s[_n0]), (DATA_TYPE)(VALUE_TO_SET), (SELECT_DATA_TYPE(DATA_TYPE))(mask[_m0].v == (DATA_TYPE)0)); \
            })                                                                                                                                             \
        })                                                                                                                                                 \
    })


#define T_ACTIVATION(DATA_TYPE, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, src, dst)               \
    ({                                                                                         \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                                                     \
        {                                                                                      \
            dst[_m0].v = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, N0, src[_m0].v, A_VAL, B_VAL); \
        })                                                                                     \
    })


#define relu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (max((DATA_TYPE)ZERO_VALUE, x))

#define brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)ZERO_VALUE, x)))

#define lu_brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL))

#define hard_swish_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x * ((min(max((DATA_TYPE)(x + (DATA_TYPE)3.f), (DATA_TYPE)0.f), (DATA_TYPE)6.f)) * (DATA_TYPE)0.166666667f))

#define identity_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x)

#define ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) op##_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x)
#define ACTIVATION_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x)

#define V_ADD(A_VAL, B_VAL) ((A_VAL) + (B_VAL))
#define V_SUB(A_VAL, B_VAL) ((A_VAL) - (B_VAL))
#define V_DIV(A_VAL, B_VAL) ((A_VAL) / (B_VAL))
#define V_MUL(A_VAL, B_VAL) ((A_VAL) * (B_VAL))


#define T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_VALUE, A_VAL, B_VAL, src, dst)               \
    ({ \
        LOOP_UNROLLING(int, _m0, 0, 1, M0, \
        { \
            dst[_m0].v = ACTIVATION_QUANTIZED(ACTIVATION_TYPE, DATA_TYPE, N0, ZERO_VALUE, A_VAL, B_VAL, src[_m0].v); \
        })                                                                                          \
    })


#define T_ADD(DATA_TYPE, M0, N0, lhs, rhs, dst) \
    ({                                                            \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                        \
        {                                                         \
            dst[_m0].v = lhs[_m0].v + rhs[_m0].v; \
        })                                                        \
    })


#define T_ADD_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \
    ({                                                            \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                        \
        {                                                         \
            dst[_m0].v = lhs[_m0].v + (DATA_TYPE)rhs_constant;               \
        })                                                        \
    })

#define T_ELTWISE_BROADCAST_ADD_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
#define T_ELTWISE_BROADCAST_LHS_X_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
#define T_ELTWISE_BROADCAST_RHS_X_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)

#define T_ELTWISE_BROADCAST_LHS_X_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
#define T_ELTWISE_BROADCAST_RHS_X_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)

#define T_ELTWISE_BROADCAST_DIV_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)

#define T_ELTWISE_BROADCAST_LHS_X_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
#define T_ELTWISE_BROADCAST_RHS_X_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)


#define T_SCALE_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \
    ({                                                            \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                        \
        {                                                         \
            dst[_m0].v = lhs[_m0].v * (DATA_TYPE)rhs_constant; \
        })                                                        \
    })


#define T_ELTWISE_BROADCAST_X(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \
    ({                                                      \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                  \
        {                                                   \
            dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)));             \
        })                                                  \
    })


#define T_ELTWISE_BROADCAST_LHS_X(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \
    ({                                                      \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                  \
        {                                                   \
            dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)));             \
        })                                                  \
    })

#define T_ELTWISE_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
#define T_ELTWISE_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
#define T_ELTWISE_DIV(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
#define T_ELTWISE_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)


#define T_ELTWISE(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \
    ({                                                      \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                  \
        {                                                   \
            dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)));             \
        })                                                  \
    })


#define T_FLOOR(DST_DATA_TYPE, M0, N0, src, dst) \
    ({                                                      \
        LOOP_UNROLLING(int, _m0, 0, 1, M0,                  \
        {                                                   \
            dst[_m0].v = floor(CONVERT(src[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)));             \
        })                                                  \
    })


#define T_MMUL(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, LHS_LAYOUT, RHS_LAYOUT, lhs, rhs, dst) T_MMUL_##LHS_LAYOUT##_##RHS_LAYOUT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T_char_char_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T_uchar_uchar_uint(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T_uchar_uchar_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
#define T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)                       \
    {                                                                                     \
        LOOP_UNROLLING(int, _m, 0, 1, M0,                                                 \
        {                                                                                 \
            LOOP_UNROLLING(int, _n, 0, 1, N0,                                             \
            {                                                                             \
                LOOP_UNROLLING(int, _k, 0, 1, K0,                                         \
                {                                                                         \
                    dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_m].s[_k]), (DST_DATA_TYPE)(rhs[_n].s[_k]), dst[_m].s[_n]); \
                })                                                                        \
            })                                                                            \
        })                                                                                \
    }

#define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)                            \
    ({ \
        LOOP_UNROLLING(int, _m, 0, 1, M0, \
        { \
            LOOP_UNROLLING(int, _n, 0, 1, N0, \
            { \
                DOT_PRODUCT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, K0, (lhs[_m].v), (rhs[_n].v), dst[_m].s[_n]); \
            })                                                                                             \
        })                                                                                             \
    })

#endif




__kernel void direct_convolution3d_ndhwc(
    TENSOR4D(src, BUFFER),
    TENSOR4D(dst, BUFFER),
    TENSOR4D(wei, BUFFER)
#if defined(HAS_BIAS)
    ,
    VECTOR_DECLARATION(bia)
#endif
)
{
#define _IWEI_WIDTH WEI_WIDTH
#define _IWEI_HEIGHT WEI_HEIGHT
#define _IWEI_DEPTH WEI_DEPTH
#define _ISRC_WIDTH SRC_WIDTH
#define _ISRC_HEIGHT SRC_HEIGHT
#define _ISRC_DEPTH SRC_DEPTH
#define _ISRC_CHANNELS SRC_CHANNELS
#define _IDST_WIDTH DST_WIDTH
#define _IDST_HEIGHT DST_HEIGHT
#define _IDST_DEPTH DST_DEPTH
#define _IDST_CHANNELS DST_CHANNELS
#define _IY_MULTIPLIER (_IWEI_WIDTH * _IWEI_HEIGHT * _IWEI_DEPTH)


#if defined(IS_QUANTIZED)
#define _IOUTPUT_TILE cq
#else
#define _IOUTPUT_TILE c
#endif

    const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0);
    const int mout = GET_SPATIAL_IDX(1, M0, 0);
    const int bout = GET_SPATIAL_IDX(2, 1, 0);

    TILE(int, M0, 1, xi);
    TILE(int, M0, 1, yi);
    TILE(int, M0, 1, zi);


    LOOP_UNROLLING(int, i, 0, 1, M0,
    {
        xi[i].v = ((mout + i) % _IDST_WIDTH) * STRIDE_X;
        yi[i].v = (((mout + i) / _IDST_WIDTH) % _IDST_HEIGHT) * STRIDE_Y;
        zi[i].v = (((mout + i) / (_IDST_WIDTH * _IDST_HEIGHT)) % _IDST_DEPTH) * STRIDE_Z;

        xi[i].v -= PAD_LEFT;
        yi[i].v -= PAD_TOP;
        zi[i].v -= PAD_FRONT;
    })


    TILE(ACC_DATA_TYPE, M0, N0, c);

    LOOP_UNROLLING(int, i, 0, 1, M0,
    {
        c[i].v = (ACC_DATA_TYPE)0;
    })

    for(int i = 0; i < _IY_MULTIPLIER; ++i)
    {
        int ck = 0;
        int xk = i % _IWEI_WIDTH;
        int yk = (i / _IWEI_WIDTH) % _IWEI_HEIGHT;
        int zk = i / (_IWEI_WIDTH * _IWEI_HEIGHT);

        int k = 0;
        for(; k <= (_ISRC_CHANNELS - K0); k += K0)
        {
            TILE(DATA_TYPE, M0, K0, a);
            TILE(DATA_TYPE, N0, K0, b);

            LOOP_UNROLLING(int, i, 0, 1, M0,
            {
                a[i].v = ZERO_VALUE;
            })


            T_LOAD_NDHWC_INDIRECT(DATA_TYPE, M0, K0, BUFFER, src, bout, zk, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, _ISRC_DEPTH, src_stride_y, xi, yi, zi, a);


            const int b_offs = k + (xk * _ISRC_CHANNELS) + (yk * _ISRC_CHANNELS * _IWEI_WIDTH) + (zk * _ISRC_CHANNELS * _IWEI_WIDTH * _IWEI_HEIGHT);
            LOOP_UNROLLING(int, i, 0, 1, N0,
            {
                if((cout + i) < _IDST_CHANNELS)
                {
                    LOOP_UNROLLING(int, j, 0, 1, K0,
                    {
                        b[i].s[j] = *(__global DATA_TYPE *)(wei_ptr + wei_offset_first_element_in_bytes + (cout + i) * sizeof(DATA_TYPE) + j * wei_stride_y + b_offs * wei_stride_y);
                    })
                }
            })


            T_MMUL(DATA_TYPE, DATA_TYPE, ACC_DATA_TYPE, M0, N0, K0, NT, T, a, b, c);



            T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, a, b, c);

            ck += K0;
        }

#if((_ISRC_CHANNELS % K0) != 0)

        for(; k < _ISRC_CHANNELS; ++k)
        {
            TILE(DATA_TYPE, M0, 1, a);
            TILE(DATA_TYPE, N0, 1, b);

            LOOP_UNROLLING(int, i, 0, 1, M0,
            {
                a[i].v = ZERO_VALUE;
            })


            T_LOAD_NDHWC_INDIRECT(DATA_TYPE, M0, 1, BUFFER, src, bout, zk, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, _ISRC_DEPTH, src_stride_y, xi, yi, zi, a);


            const int b_offs = k + (xk * _ISRC_CHANNELS) + (yk * _ISRC_CHANNELS * _IWEI_WIDTH) + (zk * _ISRC_CHANNELS * _IWEI_WIDTH * _IWEI_HEIGHT);
            LOOP_UNROLLING(int, i, 0, 1, N0,
            {
                if((cout + i) < _IDST_CHANNELS)
                {
                    b[i].v = *(__global DATA_TYPE *)(wei_ptr + wei_offset_first_element_in_bytes + (cout + i) * sizeof(DATA_TYPE) + b_offs * wei_stride_y);
                }
            })


            T_MMUL(DATA_TYPE, DATA_TYPE, ACC_DATA_TYPE, M0, N0, 1, NT, T, a, b, c);



            T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, 1, SRC_OFFSET, WEI_OFFSET, a, b, c);

            ++ck;
        }
#endif
    }



    T_ADD_CONSTANT(ACC_DATA_TYPE, M0, N0, c, (_IWEI_WIDTH * _IWEI_HEIGHT * _IWEI_DEPTH * _ISRC_CHANNELS * SRC_OFFSET * WEI_OFFSET), c);

#if defined(HAS_BIAS)
    TILE(BIA_DATA_TYPE, 1, N0, bias0);

    if((cout + N0) <= _IDST_CHANNELS)
    {
        bias0[0].v = VLOAD(N0)(0, (__global BIA_DATA_TYPE *)(bia_ptr + bia_offset_first_element_in_bytes + cout * sizeof(BIA_DATA_TYPE)));
    }
    else
    {
        VLOAD_PARTIAL(N0, PARTIAL_N0)
        (bias0[0].v, 0, (__global BIA_DATA_TYPE *)(bia_ptr + bia_offset_first_element_in_bytes + cout * sizeof(BIA_DATA_TYPE)));
    }


    T_ELTWISE_BROADCAST_ADD_X(ACC_DATA_TYPE, M0, N0, c, bias0, c);

#endif

    TILE(uint, M0, 1, dst_indirect_y);


    LOOP_UNROLLING(int, i, 0, 1, M0,
    {
        dst_indirect_y[i].v = (uint)min(mout + i, (int)(_IDST_WIDTH *_IDST_HEIGHT * _IDST_DEPTH) - 1);
        dst_indirect_y[i].v += bout * (int)(_IDST_WIDTH *_IDST_HEIGHT * _IDST_DEPTH);
    })

#if defined(IS_QUANTIZED)
    TILE(DATA_TYPE, M0, N0, cq);


    T_QUANTIZE8_ASYMMETRIC(ACC_DATA_TYPE, DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, c, cq);
#endif

    bool x_cond = PARTIAL_N0 != 0 && get_global_id(0) == 0;


    T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_N0, BUFFER, dst, cout, dst_stride_y, x_cond, _IOUTPUT_TILE, dst_indirect_y);
})"