amf_delayntapmoutoffchip_ds.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_DelayNTapMOutOffChip_DS.h"

void AMF_DelayNTapMOutOffChip_DS_Render(AMF_DelayNTapMOutOffChip_DS * restrict instance,float * restrict * buffers,int tickSize);

#define EPSILON .0001

#if 1
#pragma optimize_for_speed
SEG_MOD_FAST_CODE  void AMF_DelayNTapMOutOffChip_DS_Render(AMF_DelayNTapMOutOffChip_DS * restrict instance,float * restrict * buffers,int tickSize) {
    int delaySize = instance->delaySize*2; 					// length in samples of delay buffer * 2
    float *delayBuffer = &instance->delayBuffer[0];		 	// delay buffer
    int writeIndex = instance->delayWritePtr-delayBuffer;	// delay index
    int *delays = &instance->delays[0];
    int numTaps = instance->numTaps;
    int i, j;
    float * restrict in;
    float * restrict out;
	float pm *bufPtr;
	float pm *readBufPtr;
    int outputCount = instance->b.outputPinCount;
    float pm *tmpbuf = ((float pm *)((int)buffers[1 + outputCount]));               // scratch buffer
    float *amps = &instance->amps[0];                    // pm
    float ampL, ampR;
    int outNum;
    int *outWritten = (int *)buffers[2 + outputCount];             // Boolean, scratch buffer
	
 	in = buffers[0];
		
	// write loop
/*   	// weird kludge to get the compiler to actually set the B0 register...
    // update delayBuffer with the input values
	bufPtr = (float pm *)((int)delayBuffer);
	bufPtr = (float pm *)__builtin_circptr(bufPtr, writeIndex, (float pm *)((int)delayBuffer), delaySize);		
#pragma loop_count(8,256,8)
#pragma no_alias
    for (i=0; i<2*tickSize; i++) {
		*bufPtr = in[i];
		bufPtr = (float pm *)__builtin_circptr(bufPtr, 1, (float pm *)((int)delayBuffer), delaySize);		
    } */
    
/*asm("	R0 = dm(I0,M6);\
	LCNTR=%4, do (pc,1) until lce;\
		R0 = dm(I0,M6), pm(I12,M14)=R0;"\
	:"+I12"(instance->delayWritePtr) \
	:"B12"(delayBuffer),"L12"(delaySize),"#I0"(in),"D"(2*tickSize) \
	:"R0","I0");*/
		
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"(delaySize),"#I0"(in),"d"(2*tickSize), "z"(writeIndex) \
	:"R0","I0");	
		
    // clear outWritten flag for each output channel
#pragma loop_count(1,256,1)
    for (outNum=0; outNum<outputCount; outNum++) {
        outWritten[outNum] = 0;
    }

#pragma loop_count(1,256,1)
    for (j=0; j<numTaps; j++) {
        // bring the data with length of tickSize on that tap on chip
/*		readBufPtr = (float pm *)((int)delayBuffer);
		readBufPtr = (float pm *)__builtin_circptr(readBufPtr, writeIndex-delays[j]*2, (float pm *)((int)delayBuffer), delaySize);
    		
#pragma loop_count(8,256,8)
#pragma no_alias		
        for (i=0; i<2*tickSize; i++) {
			tmpbuf[i] = *readBufPtr;
			readBufPtr = (float pm *)__builtin_circptr(readBufPtr, 1, (float pm *)((int)delayBuffer), delaySize);		
    	} */

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"(delaySize),"#I0"(tmpbuf),"d"(2*tickSize),"z"(writeIndex-delays[j]*2) \
	:"R0","I0","I11");
          
#pragma loop_count(1,256,1)
        for (outNum=0; outNum<outputCount; outNum++) {
            ampL = amps[2*(j*outputCount+outNum)];
            ampR = amps[2*(j*outputCount+outNum)+1];
            if (ampL > EPSILON || ampL < -EPSILON || ampR > EPSILON || ampR < -EPSILON) {
                out = buffers[1+outNum];
                if (!outWritten[outNum]) {
                    outWritten[outNum] = 1;
#pragma loop_count(8,128,8)
#pragma vector_for
#pragma no_alias
                    for (i=0; i<2*tickSize; i+=2) {
     					// Should be able to use SIMD here                       
     					out[i] = tmpbuf[i] * ampL;
                        out[i+1] = tmpbuf[i+1] * ampR;
                    }
                }
                else 
#pragma loop_count(8,128,8)
#pragma vector_for
#pragma no_alias
                    for (i=0; i<2*tickSize; i+=2) {
     					// Should be able to use SIMD here                       
                        out[i] += tmpbuf[i] * ampL;
                        out[i+1] += tmpbuf[i+1] * ampR;
                    }
            }
        }
    }

    // clear output if that channel was not written
#pragma loop_count(1,256,1)
    for (outNum=0; outNum<outputCount; outNum++) {
        if (!outWritten[outNum]) {
            out = buffers[1+outNum];
            for (i=0; i<2*tickSize; i++) {
     			// Should be able to use SIMD here                       
                out[i] = 0.0;
            }
        }
    }
	
//    instance->delayWritePtr = bufPtr;
}
#endif

SEG_MOD_SLOW_CONST const AMF_ModuleClass AMFClassDelayNTapMOutOffChip_DS = {
    
    /* Flags */
    AMFModuleClassFlag_VARIABLE_PIN_COUNT,
     
    (AMF_RenderFunction)AMF_DelayNTapMOutOffChip_DS_Render,  // render function 

    /* Default bypass */
	(void *)0,
};