Vivado HLS实现离散余弦变换（DCT）

程序员文章站 2022-07-14 22:26:38

...

1.DCT原理

DCT经常用来对信号进行有损数据压缩，这是由于DCT具有很强的能量集中性，大部分自然信号的能量主要集中在DCT变换后的低频部分，具体的DCT公式这里就不多推导了（笔者也不是很懂）。二维DCT变换是在一维DCT变换的基础上，再进行一次DCT变换，二维DCT变换方法主要由三个步骤组成（信号矩阵8*8）：

（1）对行进行一维DCT得到A。

（2）对A进行行列转置，然后对行进行一维DCT，得到B。

（3）对B进行转置。

2.HLS实现DCT

添加源文件（source）：

一维DCT：

void dct_1d(dct_data_t src[DCT_SIZE], dct_data_t dst[DCT_SIZE])
{
   unsigned int k, n;
   int tmp;
   const dct_data_t dct_coeff_table[DCT_SIZE][DCT_SIZE] = {
#include "dct_coeff_table.txt"
   };

DCT_Outer_Loop:
   for (k = 0; k < DCT_SIZE; k++) {
DCT_Inner_Loop:
      for(n = 0, tmp = 0; n < DCT_SIZE; n++) {
         int coeff = (int)dct_coeff_table[k][n];
         tmp += src[n] * coeff;
      }
      dst[k] = DESCALE(tmp, CONST_BITS);
   }
}

定义dct_coeff_table为系数矩阵（一堆cosin）。
二维DCT：

void dct_2d(dct_data_t in_block[DCT_SIZE][DCT_SIZE],
      dct_data_t out_block[DCT_SIZE][DCT_SIZE])
{
   dct_data_t row_outbuf[DCT_SIZE][DCT_SIZE];
   dct_data_t col_outbuf[DCT_SIZE][DCT_SIZE], col_inbuf[DCT_SIZE][DCT_SIZE];
   unsigned i, j;

   // DCT rows
Row_DCT_Loop:
   for(i = 0; i < DCT_SIZE; i++) {
      dct_1d(in_block[i], row_outbuf[i]);
   }
   // Transpose data in order to re-use 1D DCT code
Xpose_Row_Outer_Loop:
   for (j = 0; j < DCT_SIZE; j++)
Xpose_Row_Inner_Loop:
      for(i = 0; i < DCT_SIZE; i++)
         col_inbuf[j][i] = row_outbuf[i][j];
   // DCT columns
Col_DCT_Loop:
   for (i = 0; i < DCT_SIZE; i++) {
      dct_1d(col_inbuf[i], col_outbuf[i]);
   }
   // Transpose data back into natural order
Xpose_Col_Outer_Loop:
   for (j = 0; j < DCT_SIZE; j++)
Xpose_Col_Inner_Loop:
      for(i = 0; i < DCT_SIZE; i++)
         out_block[j][i] = col_outbuf[i][j];
}

输入缓存buffer：

void read_data(short input[N], short buf[DCT_SIZE][DCT_SIZE])
{
   int r, c;

RD_Loop_Row:
   for (r = 0; r < DCT_SIZE; r++) {
RD_Loop_Col:
      for (c = 0; c < DCT_SIZE; c++)
         buf[r][c] = input[r * DCT_SIZE + c];
   }
}

输出缓存buffer：

void write_data(short buf[DCT_SIZE][DCT_SIZE], short output[N])
{
   int r, c;

WR_Loop_Row:
   for (r = 0; r < DCT_SIZE; r++) {
WR_Loop_Col:
      for (c = 0; c < DCT_SIZE; c++)
         output[r * DCT_SIZE + c] = buf[r][c];
   }
}

dct：

void dct(short input[N], short output[N])
{

   short buf_2d_in[DCT_SIZE][DCT_SIZE];
   short buf_2d_out[DCT_SIZE][DCT_SIZE];

   // Read input data. Fill the internal buffer.
   read_data(input, buf_2d_in);

   dct_2d(buf_2d_in, buf_2d_out);

   // Write out the results.
   write_data(buf_2d_out, output);
}

添加头文件（File—>new File）：

#ifndef __DCT_H__
#define __DCT_H__

#define DW 16
#define N 1024/DW
#define NUM_TRANS 16

typedef short dct_data_t;

#define DCT_SIZE 8    /* defines the input matrix as 8x8 */
#define CONST_BITS  13
#define DESCALE(x,n)  (((x) + (1 << ((n)-1))) >> n)

void dct(short input[N], short output[N]);

#endif // __DCT_H__ not defined

至此，所有的源文件和头文件全部生成。对源文件进行综合（Synthesis）：

时序：

Vivado HLS实现离散余弦变换（DCT）

资源消耗：

Vivado HLS实现离散余弦变换（DCT）

接下来就是写testbench，测试这个RTL模块功能到底正确不正确：

testbench：

int main() {
   short a[N], b[N], b_expected[N];
   int retval = 0, i;
   FILE *fp;

   fp=fopen("in.dat","r");
   for (i=0; i<N; i++){
      int tmp;
      fscanf(fp, "%d", &tmp);
      a[i] = tmp;
   }
   fclose(fp);

   fp=fopen("out.golden.dat","r");
   for (i=0; i<N; i++){
      int tmp;
      fscanf(fp, "%d", &tmp);
      b_expected[i] = tmp;
   }
   fclose(fp);

   dct(a, b);

   for (i = 0; i < N; ++i) {
      if(b[i] != b_expected[i]){
         printf("Incorrect output on sample %d. Expected %d, Received %d \n", i, b_expected[i], b[i]);
         retval = 2;
      }
   }

#if 0 // Optionally write out computed values
   fp=fopen("out.dat","w");
   for (i=0; i<N; i++){
      fprintf(fp, "%d\n", b[i]);
   }
   fclose(fp);
#endif

   if(retval != (2)){
      printf("    *** *** *** *** \n");
      printf("    Results are good \n");
      printf("    *** *** *** *** \n");
   } else {
      printf("    *** *** *** *** \n");
      printf("    BAD!! %d \n", retval);
      printf("    *** *** *** *** \n");
   }
   return retval;
}

添加数据文件：

与创造testbench.c一样并且！放在testbench目录下，才能正常读取数据。

最后！进行RTL/C协同仿真，即Solution—>Run C/RTL Co-simulation，或者主界面综合按钮旁边的小对号。（协同仿真时总也找不到数据文件，第一次果然是紧张，最后对testbench进行了C simulation，仿佛就过了）。