SplitWithOverlap

沿指定轴（axis）将输入张量切分为多个输出张量。与标准 Split 不同，该算子允许通过 start_indices 和 end_indices 自定义每个输出块的起始和结束位置，从而支持输出块之间的重叠（Overlap）。

\[\text{对于第 } j \text{ 个输出张量，其在 axis 轴上的第 } k \text{ 个元素对应：}\]
\[Output[j]_{(\dots, k, \dots)} = Input_{(\dots, start\_indices[j] + k, \dots)} \quad \text{其中 } 0 \le k < (end\_indices[j] - start\_indices[j])\]
输入：

  • input - 输入张量数据地址。

  • params - 参数数组。

    • axis - 进行切分的轴索引。

    • input_shape - 输入张量的形状数组。

    • input_ndim - 输入张量的维度。

    • num_split - 输出张量的数量。

    • start_indices - 每个输出张量在切分轴上的起始索引数组。

    • end_indices - 每个输出张量在切分轴上的结束索引数组。

  • core_mask(int, 可选) - 核掩码（仅适用于共享存储版本）。

输出：

  • outputs - 各个输出张量中填充了切分后的数据。

支持平台：

FT78NE MT7004

备注

  • FT78NE 支持 int8, int16, int32, fp32, fp64, cplx64, cplx128

  • MT7004 支持 fp16, fp32, int16, int32, cplx64

  • 算子支持不连续切分或有重叠的切分。

  • 每个输出张量在非切分轴上的维度与输入张量保持一致。

共享存储版本:

void i8_split_with_overlap_s(int8_t *input, int8_t *outputs[], long long *params, int core_mask)
void i16_split_with_overlap_s(int16_t *input, int16_t *outputs[], long long *params, int core_mask)
void i32_split_with_overlap_s(int32_t *input, int32_t *outputs[], long long *params, int core_mask)
void hp_split_with_overlap_s(half *input, half *outputs[], long long *params, int core_mask)
void fp_split_with_overlap_s(float *input, float *outputs[], long long *params, int core_mask)
void dp_split_with_overlap_s(double *input, double *outputs[], long long *params, int core_mask)
void c64_split_with_overlap_s(float *input, float *outputs[], long long *params, int core_mask)
void c128_split_with_overlap_s(double *input, double *outputs[], long long *params, int core_mask)

C调用示例：

 1//FT78NE示例（共享存储）
 2#include "78NE/utils.h"
 3
 4int main() {
 5    int core_id = get_core_id();
 6    int logic_core_id = GetLogicCoreId(core_mask, core_id);
 7    int core_num = GetCoreNum(core_mask);
 8
 9    // --- 计算尺寸 ---
10    int input_size = input_shape[0] * input_shape[1] * input_shape[2];
11    int output_sizes[num_split];
12    int total_output_size = 0;
13    int i;
14    for (i = 0; i < num_split; ++i) {
15        output_sizes[i] = input_shape[0] * (end_indices[i] - start_indices[i]) * input_shape[2];
16        total_output_size += output_sizes[i];
17    }
18
19    // --- 内存布局 (DDR空间 - 共享) ---
20    float* input = (float*)0x81000000;
21    float* output_base = (float*)0x82000000;
22    float* check_base = (float*)0x83000000;
23
24    // --- 构建指针数组 (栈空间 - 独享) ---
25    float* outputs[num_split];
26    float* check_outputs[num_split];
27    int current_offset = 0;
28    for (i = 0; i < num_split; ++i) {
29        outputs[i] = output_base + current_offset;
30        check_outputs[i] = check_base + current_offset;
31        current_offset += output_sizes[i];
32    }
33
34    // --- 数据初始化 (DDR空间 - 共享) ---
35    // 根据《多核编程规定》，共享空间变量的并发修改只能发生在单独分支
36    if (logic_core_id == 0) {
37        printf("Test_SplitWithOverlap_multicore_DDR_fp32\n");
38        for (i = 0; i < input_size; ++i) {
39            input[i] = rand() / (float)RAND_MAX * 100 - 50;
40        }
41        // 清零输出和校验区，避免旧数据干扰
42        memset(output_base, 0, total_output_size * sizeof(float));
43        memset(check_base, 0, total_output_size * sizeof(float));
44    }
45
46    long long params[10];
47    params[0] = (long long) axis;
48    params[1] = (long long) input_shape;
49    params[2] = (long long) input_ndim;
50    params[3] = (long long) num_split;
51    params[4] = (long long) start_indices;
52    params[5] = (long long) end_indices;
53
54    sys_bar(0, core_num);
55    fp_split_with_overlap_s(input, outputs, params, core_mask);
56    sys_bar(0, core_num);
57    return 0;
58}

私有存储版本:

void i8_split_with_overlap_p(int8_t *input, int8_t *outputs[], long long *params)
void i16_split_with_overlap_p(int16_t *input, int16_t *outputs[], long long *params)
void i32_split_with_overlap_p(int32_t *input, int32_t *outputs[], long long *params)
void hp_split_with_overlap_p(half *input, half *outputs[], long long *params)
void fp_split_with_overlap_p(float *input, float *outputs[], long long *params)
void dp_split_with_overlap_p(double *input, double *outputs[], long long *params)
void c64_split_with_overlap_p(float *input, float *outputs[], long long *params)
void c128_split_with_overlap_p(double *input, double *outputs[], long long *params)

C调用示例：

 1//MT7004 示例（私有存储）
 2#include <stdio.h>
 3
 4int main() {
 5    int input_size = input_shape[0] * input_shape[1] * input_shape[2];
 6    int output_sizes[num_split];
 7    int total_output_size = 0;
 8    int i;
 9    for (i = 0; i < num_split; ++i) {
10        output_sizes[i] = input_shape[0] * (end_indices[i] - start_indices[i]) * input_shape[2];
11        total_output_size += output_sizes[i];
12    }
13
14    int32_t* input = (int32_t*)0x10010000;
15    int32_t* output_base = (int32_t*)0x10020000;
16    int32_t* check_base = (int32_t*)0x10030000;
17
18    int32_t* outputs[num_split];
19    int32_t* check_outputs[num_split];
20    int current_offset = 0;
21    for (i = 0; i < num_split; ++i) {
22        outputs[i] = output_base + current_offset;
23        check_outputs[i] = check_base + current_offset;
24        current_offset += output_sizes[i];
25    }
26
27    for (i = 0; i < input_size; ++i) {
28        input[i] = (int32_t)i + 0.1f;
29    }
30    memset(output_base, 0, total_output_size * sizeof(int32_t));
31    memset(check_base, 0, total_output_size * sizeof(int32_t));
32
33    long long params[10];
34    params[0] = (long long)axis;
35    params[1] = (long long)input_shape;
36    params[2] = input_ndim;
37    params[3] = num_split;
38    params[4] = (long long)start_indices;
39    params[5] = (long long)end_indices;
40
41    i32_split_with_overlap_p(input, outputs, params);
42    return 0;
43}