任务是使用 p 线程并行化矩阵乘法并使用英特尔 ISPC 编译器进行矢量化

Question

In.ispc file using pthread generates errors as follows: (1) t.ispc:2:13: Error: Illegal to return a "varying" or vector type from exported function "matrix_mult_pl" export void * matrix_mult_pl( void *arg )

(2) t.ispc:2:36: Error: Varying pointer type parameter "arg" is illegal in an exported function. export void * matrix_mult_pl( void *arg )

(3) t.ispc:6:11: Error: syntax error, unexpected 'int'. tid = *(int *)(arg); // get the thread ID assigned sequentially. ^^^

and many more errors. Coder is attached below. Kindly look into the issue of using pthreads in ISPC.

threads.c file

export void * matrix_mult_pl( void *arg )
{
  int rows, cols, j, tid, portion_size, row_start, row_end;

  tid = *(int *)(arg); // get the thread ID assigned sequentially.
  portion_size = size / num_threads;
  row_start = tid * portion_size;
  row_end = (tid+1) * portion_size;

  for (rows = row_start; rows < row_end; ++rows) { // hold row index of 'matrix1'
    for (j = 0; j < size; ++j) { // hold column index of 'matrix2'
     // hold value of a cell
      /* one pass to sum the multiplications of corresponding cells
     in the row vector and column vector. */
      for(cols=0; cols<size; cols++) { 
        result_pl[ rows ][ cols ] += matrix1[ rows ][ j ] * matrix2[ j ][ cols ];
      }
    }
  }
}

threads.ispc file

export void * matrix_mult_pl( void *arg ) { int rows, cols, j, tid, portion_size, row_start, row_end; tid = *(int *)(arg); // get the thread ID assigned sequentially. portion_size = size / num_threads; row_start = tid * portion_size; row_end = (tid+1) * portion_size; for (rows = row_start; rows < row_end; ++rows) { // hold row index of 'matrix1' for (j = 0; j < size; ++j) { // hold column index of 'matrix2' // hold value of a cell /* one pass to sum the multiplications of corresponding cells in the row vector and column vector. */ for(cols=0; cols<size; cols++) { result_pl[ rows ][ cols ] += matrix1[ rows ][ j ] * matrix2[ j ][ cols ]; } } } }

Why is ISPC file not vectorizing the execution with parallelization by pthreads?

Answer 1

The problem you're having is because ISPC defaults to varying type. int x = 0 is the same as varying int x = 0 . This applies to pointer types as well void *arg as void varying * uniform arg and you can't have varying types in exported functions. When porting and first getting started with ISPC it's best to be explicit with uniform and varying keywords.