This discussion has been locked.

You can no longer post new replies to this discussion. If you have a question you can start a new discussion

CMSIS confusion

Note: This was originally posted on 27th October 2012 at http://forums.arm.com

Hello all
I have few questions about CMSIS. I have reviewed some libs with basic functions and there is no loops for rest of the samples (Cortex M3/4), for example (Abs function q31):

 /*loop Unrolling */
  blkCnt = blockSize >> 2u;

  /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.   
   ** a second loop below computes the remaining 1 to 3 samples. */
  while(blkCnt > 0u)
  {
    /* C = |A| */
    /* Calculate absolute of input (if -1 then saturated to 0x7fffffff) and then store the results in the destination buffer. */
    in = *pSrc++;
    *pDst++ = (in > 0) ? in : ((in == 0x80000000) ? 0x7fffffff : -in);
    in = *pSrc++;
    *pDst++ = (in > 0) ? in : ((in == 0x80000000) ? 0x7fffffff : -in);
    in = *pSrc++;
    *pDst++ = (in > 0) ? in : ((in == 0x80000000) ? 0x7fffffff : -in);
    in = *pSrc++;
    *pDst++ = (in > 0) ? in : ((in == 0x80000000) ? 0x7fffffff : -in);

    /* Decrement the loop counter */
    blkCnt--;
  }

  /* If the blockSize is not a multiple of 4, compute any remaining output samples here.   
   ** No loop unrolling is used. */
  blkCnt = blockSize % 0x4u;

I think there should be a loop for remaining samples. Can you explain why there is 4 operations in one loop iteration? Is it related with time optimization?

Best Regards

Parents

Dr. Paul Beckmann over 12 years ago

Note: This was originally posted on 27th October 2012 at http://forums.arm.com

Hi Surix,
The DSP functions in the CMSIS library rely on loop unrolling for optimization. The Q31 absolute value function is unrolled by a factor of 4. Unrolling helps to reduce run time by: (1) allowing memory accesses to be grouped together, (2) reducing the associated loop overhead (it takes 3 cycles to do a loop on the M4). The first loop in the function arm_abs_q31.c does processing on multiples of 4 samples. If you look further down in the code you'll see code that does any remaining samples:

while(blkCnt > 0u)
{
    /* C = |A| */
    /* Calculate absolute value of the input (if -1 then saturated to 0x7fffffff) and then store the results in the destination buffer. */
    in = *pSrc++;
    *pDst++ = (in > 0) ? in : ((in == 0x80000000) ? 0x7fffffff : -in);

    /* Decrement the loop counter */
    blkCnt--;
}

Structured this way, the function can handle any length vector.

-Paul
Cancel
Vote up 0 Vote down

Cancel

Reply

Dr. Paul Beckmann over 12 years ago

Note: This was originally posted on 27th October 2012 at http://forums.arm.com

Hi Surix,
The DSP functions in the CMSIS library rely on loop unrolling for optimization. The Q31 absolute value function is unrolled by a factor of 4. Unrolling helps to reduce run time by: (1) allowing memory accesses to be grouped together, (2) reducing the associated loop overhead (it takes 3 cycles to do a loop on the M4). The first loop in the function arm_abs_q31.c does processing on multiples of 4 samples. If you look further down in the code you'll see code that does any remaining samples:

while(blkCnt > 0u)
{
    /* C = |A| */
    /* Calculate absolute value of the input (if -1 then saturated to 0x7fffffff) and then store the results in the destination buffer. */
    in = *pSrc++;
    *pDst++ = (in > 0) ? in : ((in == 0x80000000) ? 0x7fffffff : -in);

    /* Decrement the loop counter */
    blkCnt--;
}

Structured this way, the function can handle any length vector.

-Paul
Cancel
Vote up 0 Vote down

Cancel

Children

No data