amf_delaylatticeoffchip.c

ADI SHARC DSP 音频算法标准模块库

// Copyright(c) 2005 Analog Devices, Inc. All Rights Reserved. ADI Confidential.

#include "AMF_DelayLatticeOffChip.h"

void AMF_DelayLatticeOffChip_Render(AMF_DelayLatticeOffChip * restrict instance,float * restrict * buffers, int tickSize);

#if 1

#pragma optimize_for_speed
SEG_MOD_FAST_CODE  void AMF_DelayLatticeOffChip_Render(AMF_DelayLatticeOffChip * restrict instance,float * restrict * buffers,int tickSize) {
    int delay = instance->delay;							// delay in samples
    float *delayBuffer = &instance->delayBuffer[0];			// delay buffer
    int writeIndex = instance->delayWritePtr-delayBuffer;	// delay index
    int i;
    float tmp;
    float *in = buffers[0];
    float *receive = buffers[1]; 
    float *out = buffers[2];
    float *send = buffers[3];
	float *readBufPtr;
	float *writePtr;
	float coef = instance->coef;
    float pm *tmpbuf = ((float pm *)((int)buffers[4]));               // scratch buffer

    // bring the data from delayBuffer on chip
/*    readBufPtr = (float pm *)((int)delayBuffer);
    readBufPtr = (float pm *)__builtin_circptr(readBufPtr, writeIndex, (float pm *)((int)delayBuffer), delay);
    		
#pragma loop_count(8,128,8)
#pragma no_alias		
    for (i=0; i<tickSize; i++) {
	    tmpbuf[i] = *readBufPtr;
		readBufPtr = (float pm *)__builtin_circptr(readBufPtr, 1, (float pm *)((int)delayBuffer), delay);		
    } */

asm("modify(i11,%5);\
		R0=pm(I11,M14);\
	LCNTR=%4, do (pc,1) until lce;\
		dm(I0,M6)=R0, R0=pm(I11,M14);"\
	: \
	:"#I11"(delayBuffer), "B11"(delayBuffer),"L11"(delay),"#I0"(tmpbuf),"d"(tickSize),"z"(writeIndex) \
	:"R0","I0","I11");
          
#pragma loop_count(8,128,8)
	for (i=0; i<tickSize; i++) {
		//tmp = *writePtr+coef*in[i];
        tmp = tmpbuf[i] + coef*in[i];
		send[i] = in[i] - coef*tmp; // make sure we read in[] before writing out[], to allow aliasing
 		out[i] = tmp;
    }    

	// weird kludge to get the compiler to actually set the B0, B1 registers...
/*	writePtr = delayBuffer;
	writePtr = __builtin_circptr(writePtr, writeIndex, delayBuffer, delay);		

#pragma loop_count(8,128,8)
#pragma no_alias
	for (i=0; i<tickSize; i++) {
 		*writePtr = receive[i]; // the processed send signal is delayed
		writePtr = __builtin_circptr(writePtr, 1, delayBuffer, delay);		
    }*/

asm("modify(I12,%6);\
        R0 = dm(I0,M6);\
	LCNTR=%5, do (pc,1) until lce;\
		R0 = dm(I0,M6), pm(I12,M14)=R0;\
	%0=I12;"\
	:"=d"(instance->delayWritePtr) \
	:"#I12"(delayBuffer), "B12"(delayBuffer),"L12"(delay),"#I0"(receive),"d"(tickSize), "z"(writeIndex) \
	:"R0","I0");

//   	instance->delayWritePtr = writePtr;
}
#endif


SEG_MOD_SLOW_CONST const AMF_ModuleClass AMFClassDelayLatticeOffChip = {

	/** Flags. This is not used, and the member should probably be removed. */
	0,

	/** Reference to render function. */
    (AMF_RenderFunction)AMF_DelayLatticeOffChip_Render,  // render function 

    /* Default bypass */
	(void *)0,
    
	/* Input descriptor - 2 inputs, and they are mono. */
	2, 0,

	/* Output descriptor - 2 outputs, and they are mono. */
    2, 0,
};