Wrong output when running dsps_mulc_s16_ functions (DSP-108) about esp-dsp

[peter-marcisovsky]

Hello, the problem here is data format. In the esp-dsp repository, all the _s16_ functions use Q15 fixed point data format. In your code, you are feeding the function with incorrect inputs, which causes an output overflow, and this is the reason why you are getting -1 output for your input.

Peter

[guilhAbreu]

@peter-marcisovsky Wouldn't be the case to implement the shift parameter as it is implemented for dsps_add_s16?

[guilhAbreu]

@peter-marcisovsky Thank you for your time, and I'm sorry to continue to bother you. I took a look at both ANSI and ae32 implementations, and simply just removing the shift instruction solved the problem. It would be very helpful if the shift parameter could be implemented.

For those who are facing the same issue. I'm talking about lines 41 and 24 of the ae32 and ANSI implementation, respectively. I'm appending here the custom implementations that solve the problem for me.

#include "dsps_mulc_platform.h"
#if (dsps_mulc_s16_ae32_enabled == 1)

    .text
    .align  4
    .global dsps_mulc_s16_ae32_custom
    .type   dsps_mulc_s16_ae32_custom,@function

dsps_mulc_s16_ae32_custom: 
// input    - a2
// output   - a3
// len      - a4
// C        - a5
// step_in  - a6
// step_out - a7
// shift    - stack (a8)

    entry	a1, 16
    
    l32i.n	a8, a1,  16     // Load shift to the a8 register
    ssr     a8              // sar = a8

    srli    a4, a4, 1   // a4 = a4>>1
    slli 	a6, a6, 2  	// a6 - step_in<<3, because we load two inputs per loop
    slli 	a7, a7, 1  	// a7 - step_out<<2

    addi    a6, a6, -4;
    addi    a2, a2, -4;

	ldinc m0, a2

    loopnez a4, loop_end_mulc_f32_ae32
        add.n       a2, a2, a6     // input+=step_input;
        mul.DA.LL   m0, a5 
        rsr a8, acchi
        rsr a9, acclo
        src a8, a8, a9  // Here result in a8
    	s16i	a8, a3, 0   // store result to the putput        
        // rsr a9, acclo
    	// s16i	a9, a3, 0   // store result to the putput        
        add.n   a3, a3, a7     // output+=step_out;
        mul.DA.HL   m0, a5         

        rsr a8, acchi
        rsr a9, acclo
	    ldinc       m0,   a2               // load next data
        src a10, a8, a9  // Here result in a8
    	s16i	a10, a3, 0   // store result to the putput
        // // rsr a9, acclo
    	// // s16i	a9, a3, 0   // store result to the putput        
        add.n   a3, a3, a7  // output+=step_out;
loop_end_mulc_f32_ae32:
    movi.n	a2, 0 // return status ESP_OK
    retw.n
    
#endif // dsps_mulc_s16_ae32_enabled

esp_err_t dsps_mulc_s16_ansi_custom(const int16_t *input, int16_t *output, int len, int16_t C, int step_in, int step_out, int shift)
{
    if (NULL == input) return ESP_ERR_DSP_PARAM_OUTOFRANGE;
    if (NULL == output) return ESP_ERR_DSP_PARAM_OUTOFRANGE;

    for (int i = 0 ; i < len ; i++) {
        int32_t acc = (int32_t)input[i * step_in] * (int32_t)C;
        output[i * step_out] = (int16_t)(acc>>shift);
    }
    return ESP_OK;
}