Hi Matthew thanks for the reply. Yes performance is an issue thats why I try to do away with the paddings. Because I am calling memcpy many times.

Hi Matthew thanks for the reply. Yes performance is an issue thats why I try to do away with the paddings. Because I am calling memcpy many times.

Hi

The usuage of Ne10 fir function is as follows:

void ne10_fir_float_c (const ne10_fir_instance_f32_t * S,

ne10_float32_t * pSrc,

ne10_float32_t * pDst,

ne10_uint32_t blockSize).

The NEON version has the same interface.

Before using it, you need to call initialization function:

ne10_result_t ne10_fir_init_float (ne10_fir_instance_f32_t * S,

ne10_uint16_t numTaps,

ne10_float32_t * pCoeffs,

ne10_float32_t * pState,

ne10_uint32_t blockSize)

{

/* Assign filter taps */

S->numTaps = numTaps;

/* Assign coefficient pointer */

S->pCoeffs = pCoeffs;

/* Clear state buffer and the size of state buffer is (blockSize + numTaps - 1) */

memset (pState, 0, (numTaps + (blockSize - 1u)) * sizeof (ne10_float32_t));

/* Assign state pointer */

S->pState = pState;

return NE10_OK;

}

Currently, the input and output have the same length. If you want to acheive your result, padding is one of resolutions. But padding needs too many memcpy. Another resolution is that you need to modify the neon code according to your requirement.

Hi Zhang Yang,

Thanks for the reply. I tried the first method zero padding with memcpy and the outcome was disastrous. I am more inclined to use the second method or use fft and ifft to replace the filter function. If I were to start on the second method, may I know how should I start? Do I need to change the ASM code?

Kelvin