26 #define SUB(x, y) (x - y) 27 #define ADD(x, y) (x + y) 28 #define MAX(x, y) max((x), (y)) 29 #define MIN(x, y) min((x), (y)) 30 #define SQUARED_DIFF(x, y) (x - y) * (x - y) 31 #define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE)0), SELECT_VEC_DATA_TYPE(float, VEC_SIZE_OUT)))) 32 #define DIV(x, y) (x / y) 34 #define CONVERT_RTE(x, type) (convert_##type##_rte((x))) 35 #define CONVERT_DOWN(x, type) CONVERT_RTE(x, type) 37 #define OP_FUN_NAME_STR(op) elementwise_operation_##op##_quantized 38 #define OP_FUN_NAME(op) OP_FUN_NAME_STR(op) 40 #if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE) 42 #define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE_OUT) 43 #define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE_OUT) 44 #define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT) 90 #if !defined(IN_PLACE) 93 #endif // !defined(IN_PLACE) 98 #else // VEC_SIZE_IN1 == 1 99 uint in1_x_offs = max((
int)(get_global_id(0) * VEC_SIZE_IN1 - (VEC_SIZE_IN1 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN1), 0);
100 #endif // VEC_SIZE_IN1 == 1 101 #if VEC_SIZE_IN2 == 1 103 #else // VEC_SIZE_IN2 == 1 104 uint in2_x_offs = max((
int)(get_global_id(0) * VEC_SIZE_IN2 - (VEC_SIZE_IN2 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN2), 0);
105 #endif // VEC_SIZE_IN2 == 1 106 #if !defined(IN_PLACE) 107 uint out_x_offs = max((
int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
108 #endif // !defined(IN_PLACE) 111 __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs *
sizeof(DATA_TYPE) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z;
112 __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs *
sizeof(DATA_TYPE) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z;
114 #if !defined(IN_PLACE) 115 out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs *
sizeof(DATA_TYPE) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z;
116 #else // !defined(IN_PLACE) 117 #if defined(SRC1_IN_PLACE) 119 #else //defined(SRC1_IN_PLACE) 121 #endif //defined(SRC1_IN_PLACE) 122 #endif // !defined(IN_PLACE) 124 VEC_INT in_a =
CONVERT((
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(
VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE *)in1_addr)), VEC_INT);
125 VEC_INT in_b =
CONVERT((
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(
VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE *)in2_addr)), VEC_INT);
127 in_a =
SUB(in_a, (VEC_INT)((
int)OFFSET_IN1));
128 in_b =
SUB(in_b, (VEC_INT)((
int)OFFSET_IN2));
136 STORE_VECTOR_SELECT(res, DATA_TYPE, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
#define CONVERT_SAT(x, type)
#define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond)
Store a vector that can only be partial in x.
#define CONVERT_DOWN(x, type)
#define TENSOR3D_DECLARATION(name)
#define VEC_DATA_TYPE(type, size)