vector_fpu.inl

hl2 source code. Do not use it illegal.

inline void hk_VecFPU::fpu_add_multiple_row(hk_real *target_adress,hk_real *source_adress,hk_real factor,int size,hk_bool adress_aligned) {
    if(adress_aligned==HK_FALSE) {
	//we have to calculate the block size and shift adresses to lower aligned adresses
	long result_adress = long(source_adress) & hk_VecFPU_MEM_MASK_FLOAT;
	target_adress = (hk_real *)( long (target_adress) & hk_VecFPU_MEM_MASK_FLOAT);
	size += (long(source_adress)-result_adress)>>hk_VecFPU_MEMSHIFT_FLOAT;
	source_adress=(hk_real *)result_adress;
    }

#if defined(IVP_USE_PS2_VU0)

    asm __volatile__("
        mfc1    $8,%3
        qmtc2    $8,vf5		# vf5.x   factor
    1:
	lqc2    vf4,0x0(%1)     # vf4 *source
	vmulx.xyzw vf6,vf4,vf5  # vf6 *source * factor
	lqc2    vf4,0x0(%0)
	addi    %0, 0x10
	vadd.xyzw vf6,vf4,vf6   # vf6 = *dest + factor * *source
	addi    %2,-4
	addi    %1, 0x10
	sqc2    vf6,-0x10(%0)
	bgtz    %2,1b
	nop
	"
	: /* no output */
	: "r" (target_adress), "r" (source_adress) , "r" (size), "f" (factor)
	: "$8" , "memory");
#else
#if 0
#       if defined(IVP_WILLAMETTE)
	    ;
        __m128d factor128=_mm_set1_pd(factor);

        for(int i=size;i>0;i-=hk_VecFPU_SIZE_FLOAT) {
	        __m128d source_d=_mm_load_pd(source_adress);
	        __m128d prod_d=_mm_mul_pd(factor128,source_d);
	        __m128d target_d=_mm_load_pd(target_adress);
	        target_d=_mm_add_pd(prod_d,target_d);
	        _mm_store_pd(target_adress,target_d);

	        target_adress+=hk_VecFPU_SIZE_FLOAT;
	        source_adress+=hk_VecFPU_SIZE_FLOAT;
		}
#endif
#endif
	

        for(int i=size;i>0;i-=hk_VecFPU_SIZE_FLOAT) {
#       if hk_VecFPU_SIZE_FLOAT == 2
	    hk_real a = source_adress[0] * factor;
	    hk_real b = source_adress[1] * factor;
	    a += target_adress[0];
	    b += target_adress[1];
	    target_adress[0] = a;
	    target_adress[1] = b;
#   elif hk_VecFPU_SIZE_FLOAT == 4
	    hk_real a = source_adress[0] * factor;
	    hk_real b = source_adress[1] * factor;
	    hk_real c = source_adress[2] * factor;
	    hk_real d = source_adress[3] * factor;
	    a += target_adress[0];
	    b += target_adress[1];
	    c += target_adress[2];
	    d += target_adress[3];
	    target_adress[0] = a;
	    target_adress[1] = b;
	    target_adress[2] = c;
	    target_adress[3] = d;
#   else 
	shit
#   endif
	    target_adress+=hk_VecFPU_SIZE_FLOAT;
	    source_adress+=hk_VecFPU_SIZE_FLOAT;
		
	}
#endif
}

inline hk_real hk_VecFPU::fpu_large_dot_product(hk_real *base_a, hk_real *base_b, int size, hk_bool adress_aligned) {
    if(adress_aligned==HK_FALSE) {
	    //we have to calculate the block size and shift adresses to lower aligned adresses
	    long result_adress = long(base_a) & hk_VecFPU_MEM_MASK_FLOAT;
	    base_b = (hk_real *)( long (base_b) & hk_VecFPU_MEM_MASK_FLOAT);
	    size += (long(base_a)-result_adress)>>hk_VecFPU_MEMSHIFT_FLOAT;  // because start changed
	    base_a=(hk_real *)result_adress;
    }
#   if defined(IVP_WILLAMETTE)
    IVP_IF_WILLAMETTE_OPT(IVP_Environment_Manager::get_environment_manager()->ivp_willamette_optimization) {

	__m128d sum128 =_mm_set1_pd(0.0f);

	int i;
	for(i=size;i>=hk_VecFPU_SIZE_FLOAT; i-= hk_VecFPU_SIZE_FLOAT) {
	    __m128d mult1=_mm_load_pd(base_a);
	    __m128d mult2=_mm_load_pd(base_b);
	    __m128d prod =_mm_mul_pd(mult1,mult2);
	    sum128 =_mm_add_pd(prod,sum128);
	    base_a += hk_VecFPU_SIZE_FLOAT;
	    base_b += hk_VecFPU_SIZE_FLOAT;
	}

	__m128d dummy,low,high;
	low=sum128;
	dummy=sum128;
	//_mm_shuffle_pd(sum128,sum128,1); //swap high and low
	sum128=_mm_unpackhi_pd(sum128,dummy);
	high=sum128;
	__m128d sum64=_mm_add_sd(low,high);
	__m128d res=sum64;
	//_mm_shuffle_pd(sum64,sum64,1); 
	hk_real result_sum; 
	_mm_store_sd(&result_sum,res);
	
	for(;i>=0;i--) {
	    result_sum += base_a[i] * base_b[i];
	}

	return result_sum;
    } else {
	hk_real sum=0.0f;
	for(int i=size-1;i>=0;i--) {
	    sum += base_a[i] * base_b[i];
	}
	return sum;
    } 
#else
    hk_real sum=0.0f;
    for(int i=size-1;i>=0;i--) {
	sum += base_a[i] * base_b[i];
    }
    return sum;
#endif
}

inline void hk_VecFPU::fpu_multiply_row(hk_real *target_adress,hk_real factor,int size,hk_bool adress_aligned) {
    if(adress_aligned==HK_FALSE) {
	//we have to calculate the block size and shift adresses to lower aligned adresses
	long adress,result_adress;
	adress=(long)target_adress;
	result_adress=adress & hk_VecFPU_MEM_MASK_FLOAT;
	size+=(adress-result_adress)>>hk_VecFPU_MEMSHIFT_FLOAT;
	target_adress=(hk_real *)result_adress;
    }
#if 0
#ifdef IVP_WILLAMETTE
        __m128d factor128=_mm_set1_pd(factor);
        int i;
        for(i=size;i>0;i-=hk_VecFPU_SIZE_FLOAT) {
	    __m128d target_d=_mm_load_pd(target_adress);
	    target_d=_mm_mul_pd(factor128,target_d);
	    _mm_store_pd(target_adress,target_d);
	
	    target_adress+=hk_VecFPU_SIZE_FLOAT;
		}    
#endif
#endif
        int i;
        for(i=size;i>0;i-=hk_VecFPU_SIZE_FLOAT) {
#       if hk_VecFPU_SIZE_FLOAT == 2
	    hk_real a = target_adress[0] * factor;
	    hk_real b = target_adress[1] * factor;
	    target_adress[0] = a;
	    target_adress[1] = b;
#       elif hk_VecFPU_SIZE_FLOAT == 4
	    hk_real a = target_adress[0] * factor;
	    hk_real b = target_adress[1] * factor;
	    hk_real c = target_adress[2] * factor;
	    hk_real d = target_adress[3] * factor;
	    target_adress[0] = a;
	    target_adress[1] = b;
	    target_adress[2] = c;
	    target_adress[3] = d;
#       else
	    shit
#       endif
		target_adress+=hk_VecFPU_SIZE_FLOAT;
	
	}
}
// #+# sparc says rui, optimize for non vector units ( hk_VecFPU_SIZE_FLOAT = 4 )
inline void hk_VecFPU::fpu_exchange_rows(hk_real *target_adress1,hk_real *target_adress2,int size,hk_bool adress_aligned) {
    if(adress_aligned==HK_FALSE) {
	//we have to calculate the block size and shift adresses to lower aligned adresses
	long adress,result_adress;
	adress=(long)target_adress1;
	result_adress=adress & hk_VecFPU_MEM_MASK_FLOAT;
	size+=(adress-result_adress)>>hk_VecFPU_MEMSHIFT_FLOAT;
	target_adress1=(hk_real *)result_adress;
	adress=(long)target_adress2;
	adress=adress & hk_VecFPU_MEM_MASK_FLOAT;
	target_adress2=(hk_real *)adress;
    }

#if 0
        int i;
        for(i=size;i>0;i-=hk_VecFPU_SIZE_FLOAT) {
	    __m128d a_d=_mm_load_pd(target_adress1);
	    __m128d b_d=_mm_load_pd(target_adress2);
	    _mm_store_pd(target_adress1,b_d);    
	    _mm_store_pd(target_adress2,a_d);    
	    
	    target_adress1+=hk_VecFPU_SIZE_FLOAT;
	    target_adress2+=hk_VecFPU_SIZE_FLOAT;
	}
#endif
	
        int i;
        for(i=size;i>0;i-=hk_VecFPU_SIZE_FLOAT) {
	        hk_real h;
#       if hk_VecFPU_SIZE_FLOAT == 2
	        h=target_adress1[0];  target_adress1[0]=target_adress2[0];  target_adress2[0]=h;
	        h=target_adress1[1];  target_adress1[1]=target_adress2[1];  target_adress2[1]=h;
#       elif hk_VecFPU_SIZE_FLOAT == 4
	        h=target_adress1[0];  target_adress1[0]=target_adress2[0];  target_adress2[0]=h;
	        h=target_adress1[1];  target_adress1[1]=target_adress2[1];  target_adress2[1]=h;
	        h=target_adress1[2];  target_adress1[2]=target_adress2[2];  target_adress2[2]=h;
	        h=target_adress1[3];  target_adress1[3]=target_adress2[3];  target_adress2[3]=h;
#       else
        shit
#       endif
	    target_adress1+=hk_VecFPU_SIZE_FLOAT;
	    target_adress2+=hk_VecFPU_SIZE_FLOAT;
		}
	
}

inline void hk_VecFPU::fpu_copy_rows(hk_real *target_adress,hk_real *source_adress,int size,hk_bool adress_aligned) {
    if(adress_aligned==HK_FALSE) {
	//we have to calculate the block size and shift adresses to lower aligned adresses
	long adress,result_adress;
	adress=(long)source_adress;
	result_adress=adress & hk_VecFPU_MEM_MASK_FLOAT;
	size+=(adress-result_adress)>>hk_VecFPU_MEMSHIFT_FLOAT;
	source_adress=(hk_real *)result_adress;
	adress=(long)target_adress;
	adress=adress & hk_VecFPU_MEM_MASK_FLOAT;
	target_adress=(hk_real *)adress;
    }
#if 0
        int i;
        for(i=size-1;i>=0;i-=hk_VecFPU_SIZE_FLOAT) {
	        __m128d target_d=_mm_load_pd(source_adress);
	        _mm_store_pd(target_adress,target_d);

	        target_adress+=hk_VecFPU_SIZE_FLOAT;
	        source_adress+=hk_VecFPU_SIZE_FLOAT;
		}
	}
#endif
        int i;
        for(i=size-1;i>=0;i-=hk_VecFPU_SIZE_FLOAT) {
	    int j;
	    for(j=0;j<hk_VecFPU_SIZE_FLOAT;j++) {
	        target_adress[j] = source_adress[j];
		}
	    target_adress+=hk_VecFPU_SIZE_FLOAT;
	    source_adress+=hk_VecFPU_SIZE_FLOAT;
		
	}
}

inline void hk_VecFPU::fpu_set_row_to_zero(hk_real *target_adress,int size,hk_bool adress_aligned) {
    if(adress_aligned==HK_FALSE) {
        long adress,result_adress;
	adress=(long)target_adress;
	result_adress=adress & hk_VecFPU_MEM_MASK_FLOAT;
	size+=(adress-result_adress)>>hk_VecFPU_MEMSHIFT_FLOAT;
	target_adress=(hk_real *)result_adress;
    }
#if 0
        __m128d zero128=_mm_set1_pd(0.0f);
        int i;
        for(i=size-1;i>=0;i-=hk_VecFPU_SIZE_FLOAT) {	
	        _mm_store_pd(target_adress,zero128);

	        target_adress+=hk_VecFPU_SIZE_FLOAT;
		} 
	}
#endif
        int i;
        for(i=size-1;i>=0;i-=hk_VecFPU_SIZE_FLOAT) {
	        int j;
	        for(j=0;j<hk_VecFPU_SIZE_FLOAT;j++) {
	            target_adress[j] = 0.0f;
			}
	        target_adress+=hk_VecFPU_SIZE_FLOAT;
		 
	}
}

inline int hk_VecFPU::calc_aligned_row_len(int unaligned_len,hk_real *dummy_type) { //dummy type is used for overloading
    return (unaligned_len+hk_VecFPU_SIZE_FLOAT-1)&hk_VecFPU_MASK_FLOAT;
}

//----------------------------------------------------------------------------------------------------------------------