BiasAdd

将偏置向量 input_bias 沿通道维广播，加到输入张量 input_x 上。

\[\mathrm{output} = \mathrm{input\_x} + \mathrm{input\_bias}\]

NCHW：\(\mathrm{output}(n,c,h,w) = \mathrm{input\_x}(n,c,h,w) + \mathrm{input\_bias}(c)\)
NHWC：\(\mathrm{output}(n,h,w,c) = \mathrm{input\_x}(n,h,w,c) + \mathrm{input\_bias}(c)\)

输入：

input_x - 输入张量地址
input_bias - 偏置地址，长度等于通道数 C；c64 时为交错复数，长度 2C 个 float
dims - 维度数组；NCHW 为 [N, C, H, W]，NHWC 为 [N, H, W, C]
shape_size - 维数（测试常用 4）
data_format - 布局：0 为 NCHW，1 为 NHWC（整型，不是字符串）
length - input_x 元素个数（复数时为复数个数）
core_mask - 核掩码（仅共享存储版本）

输出：

output - 输出地址，形状与 input_x 相同

支持平台：

FT78NE MT7004

备注

FT78NE 支持 int8、int16、int32、fp32、fp64、cplx64、cplx128
MT7004 支持 int16、int32、fp16、fp32、cplx64
data_format 为 int：0 = NCHW，非 0（测试用 1）= NHWC

共享存储版本:

void i8_bias_add_s(int8_t *input_x, int8_t *input_bias, int8_t *output, int *dims, int shape_size, int data_format, int length, int core_mask)

void i16_bias_add_s(int16_t *input_x, int16_t *input_bias, int16_t *output, int *dims, int shape_size, int data_format, int length, int core_mask)

void i32_bias_add_s(int32_t *input_x, int32_t *input_bias, int32_t *output, int *dims, int shape_size, int data_format, int length, int core_mask)

void hp_bias_add_s(float16 *input_x, float16 *input_bias, float16 *output, int *dims, int shape_size, int data_format, int length, int core_mask)

void fp_bias_add_s(float *input_x, float *input_bias, float *output, int *dims, int shape_size, int data_format, int length, int core_mask)

void dp_bias_add_s(double *input_x, double *input_bias, double *output, int *dims, int shape_size, int data_format, int length, int core_mask)

void c64_bias_add_s(float *input_x, float *input_bias, float *output, int *dims, int shape_size, int data_format, int length, int core_mask)

void c128_bias_add_s(double *input_x, double *input_bias, double *output, int *dims, int shape_size, int data_format, int length, int core_mask)

C调用示例：

// MT7004 示例（共享存储多核，DDR 地址，NHWC）
void TestBiasAddSMCFp32(int core_mask) {
    int core_id = get_core_id();
    int logic_core_id = GetLogicCoreId(core_mask, core_id);
    int core_num = GetCoreNum(core_mask);
    float *input_x = (float *)0x81000000;
    float *input_bias = (float *)0x82000000;
    float *output = (float *)0x83000000;
    int *dims = (int *)0x85000000;
    if (logic_core_id == 0) {
        dims[0] = 1; dims[1] = 4; dims[2] = 4; dims[3] = 16; /* N,H,W,C */
    }
    sys_bar(0, core_num);
    fp_bias_add_s(input_x, input_bias, output, dims, 4, 1 /* NHWC */, 256, core_mask);
}

void main() {
    int core_mask = 0b1111;
    TestBiasAddSMCFp32(core_mask);
}

私有存储版本:

void i8_bias_add_p(int8_t *input_x, int8_t *input_bias, int8_t *output, int *dims, int shape_size, int data_format, int length)

void i16_bias_add_p(int16_t *input_x, int16_t *input_bias, int16_t *output, int *dims, int shape_size, int data_format, int length)

void i32_bias_add_p(int32_t *input_x, int32_t *input_bias, int32_t *output, int *dims, int shape_size, int data_format, int length)

void hp_bias_add_p(float16 *input_x, float16 *input_bias, float16 *output, int *dims, int shape_size, int data_format, int length)

void fp_bias_add_p(float *input_x, float *input_bias, float *output, int *dims, int shape_size, int data_format, int length)

void dp_bias_add_p(double *input_x, double *input_bias, double *output, int *dims, int shape_size, int data_format, int length)

void c64_bias_add_p(float *input_x, float *input_bias, float *output, int *dims, int shape_size, int data_format, int length)

void c128_bias_add_p(double *input_x, double *input_bias, double *output, int *dims, int shape_size, int data_format, int length)

C调用示例：

// MT7004 示例（私有存储单核，AM 地址，NHWC）
void TestBiasAddAMFp32(void) {
    float *input_x = (float *)0x10010000;
    float *input_bias = (float *)0x10020000;
    float *output = (float *)0x10030000;
    int *dims = (int *)0x10050000;
    dims[0] = 1; dims[1] = 4; dims[2] = 4; dims[3] = 16; /* N,H,W,C */
    fp_bias_add_p(input_x, input_bias, output, dims, 4, 1 /* NHWC */, 256);
}

void main() {
    TestBiasAddAMFp32();
}