ScatterNd

依据指定的索引 indices，将更新值 updates 累加到输出张量 output 的对应位置。

\[output[indices_i] = output[indices_i] + updates_i\]
输入：

  • output - 输出张量的起始地址（计算前作为基础值，计算后存储结果）。

  • output_shape - 输出张量的形状数组。

  • output_ndim - 输出张量的维度数。

  • indices - 索引数据地址，其形状通常为 (num_slices, indices_depth)。

  • indices_shape - 索引张量的形状数组。

  • indices_ndim - 索引张量的维度数。

  • updates - 更新数据地址。

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

输出：

  • output - 计算结果地址。

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

  • 该算子在多核实现中由于涉及随机访存写，通常直接在 DDR 空间操作。

  • 张量维度最大支持 8 维。

共享存储版本:

void i8_scatter_nd_s(int8_t *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, int8_t *updates, int core_mask)
void i16_scatter_nd_s(int16_t *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, int16_t *updates, int core_mask)
void i32_scatter_nd_s(int32_t *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, int32_t *updates, int core_mask)
void hp_scatter_nd_s(half *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, half *updates, int core_mask)
void fp_scatter_nd_s(float *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, float *updates, int core_mask)
void dp_scatter_nd_s(double *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, double *updates, int core_mask)
void c64_scatter_nd_s(float *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, float *updates, int core_mask)
void c128_scatter_nd_s(double *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, double *updates, int core_mask)

C调用示例：

 1//FT78NE示例（共享存储）
 2#include <stdio.h>
 3#include "78NE/utils.h"
 4
 5int main() {
 6    float *output = (float *)0xA0000000;   // 基础输出张量在 DDR
 7    float *updates = (float *)0xB0000000;  // 更新值在 DDR
 8    int *indices = (int *)0xC0000000;      // 索引在 DDR
 9    int out_shape[] = {4, 4, 4};
10    int ind_shape[] = {5, 2};
11    int out_ndim = 3;
12    int ind_ndim = 2;
13    int core_mask = 0x0B;
14
15    fp_scatter_nd_s(output, out_shape, out_ndim, indices, ind_shape, ind_ndim, updates, core_mask);
16    return 0;
17}

私有存储版本:

void i8_scatter_nd_p(int8_t *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, int8_t *updates)
void i16_scatter_nd_p(int16_t *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, int16_t *updates)
void i32_scatter_nd_p(int32_t *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, int32_t *updates)
void hp_scatter_nd_p(half *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, half *updates)
void fp_scatter_nd_p(float *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, float *updates)
void dp_scatter_nd_p(double *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, double *updates)
void c64_scatter_nd_p(float *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, float *updates)
void c128_scatter_nd_p(double *output, int *output_shape, int output_ndim, int *indices, int *indices_shape, int indices_ndim, double *updates)

C调用示例：

 1//MT7004 示例
 2#include <stdio.h>
 3
 4int main() {
 5    float *output = (float *)0x10810000;
 6    float *updates = (float *)0x10820000;
 7    int *indices = (int *)0x10830000;
 8    int out_shape[] = {4, 4, 4};
 9    int ind_shape[] = {5, 2};
10    int out_ndim = 3;
11    int ind_ndim = 2;
12
13    fp_scatter_nd_p(output, out_shape, out_ndim, indices, ind_shape, ind_ndim, updates);
14    return 0;
15}