convolution.c

su 的源代码库

		x0 = x[ifx];
		x1 = x[ifx+1];
		x2 = x[ifx+2];
		x3 = x[ifx+3];
		x4 = x[ifx+4];
		x5 = x[ifx+5];
		x6 = x[ifx+6];
		x7 = x[ifx+7];
		x8 = x[ifx+8];
		x9 = x[ifx+9];
		x10 = x[ifx+10];
		x11 = x[ifx+11];
		x12 = x[ifx+12];
		x13 = x[ifx+13];
		x14 = x[ifx+14];
		x15 = x[ifx+15];
		x16 = x[ifx+16];
		x17 = x[ifx+17];
		x18 = x[ifx+18];
		x19 = x[ifx+19];
		x20 = x[ifx+20];
		x21 = x[ifx+21];
		x22 = x[ifx+22];
		x23 = x[ifx+23];
		for (i=ilow; i<=ihigh-1; i+=2) {
			ya = y[i+1-ifx];  z1 = x0*ya;
			yb = y[i-ifx];  z0 = x0*yb;  z1 += x1*yb;
			ya = y[i-ifx-1];  z0 += x1*ya;  z1 += x2*ya;
			yb = y[i-ifx-2];  z0 += x2*yb;  z1 += x3*yb;
			ya = y[i-ifx-3];  z0 += x3*ya;  z1 += x4*ya;
			yb = y[i-ifx-4];  z0 += x4*yb;  z1 += x5*yb;
			ya = y[i-ifx-5];  z0 += x5*ya;  z1 += x6*ya;
			yb = y[i-ifx-6];  z0 += x6*yb;  z1 += x7*yb;
			ya = y[i-ifx-7];  z0 += x7*ya;  z1 += x8*ya;
			yb = y[i-ifx-8];  z0 += x8*yb;  z1 += x9*yb;
			ya = y[i-ifx-9];  z0 += x9*ya;  z1 += x10*ya;
			yb = y[i-ifx-10];  z0 += x10*yb;  z1 += x11*yb;
			ya = y[i-ifx-11];  z0 += x11*ya;  z1 += x12*ya;
			yb = y[i-ifx-12];  z0 += x12*yb;  z1 += x13*yb;
			ya = y[i-ifx-13];  z0 += x13*ya;  z1 += x14*ya;
			yb = y[i-ifx-14];  z0 += x14*yb;  z1 += x15*yb;
			ya = y[i-ifx-15];  z0 += x15*ya;  z1 += x16*ya;
			yb = y[i-ifx-16];  z0 += x16*yb;  z1 += x17*yb;
			ya = y[i-ifx-17];  z0 += x17*ya;  z1 += x18*ya;
			yb = y[i-ifx-18];  z0 += x18*yb;  z1 += x19*yb;
			ya = y[i-ifx-19];  z0 += x19*ya;  z1 += x20*ya;
			yb = y[i-ifx-20];  z0 += x20*yb;  z1 += x21*yb;
			ya = y[i-ifx-21];  z0 += x21*ya;  z1 += x22*ya;
			yb = y[i-ifx-22];  z0 += x22*yb;  z1 += x23*yb;
			ya = y[i-ifx-23];  z0 += x23*ya;
			z[i+1] = z1;
			z[i] = z0;
		}
	} else if (lx==25) {
		x0 = x[ifx];
		x1 = x[ifx+1];
		x2 = x[ifx+2];
		x3 = x[ifx+3];
		x4 = x[ifx+4];
		x5 = x[ifx+5];
		x6 = x[ifx+6];
		x7 = x[ifx+7];
		x8 = x[ifx+8];
		x9 = x[ifx+9];
		x10 = x[ifx+10];
		x11 = x[ifx+11];
		x12 = x[ifx+12];
		x13 = x[ifx+13];
		x14 = x[ifx+14];
		x15 = x[ifx+15];
		x16 = x[ifx+16];
		x17 = x[ifx+17];
		x18 = x[ifx+18];
		x19 = x[ifx+19];
		x20 = x[ifx+20];
		x21 = x[ifx+21];
		x22 = x[ifx+22];
		x23 = x[ifx+23];
		x24 = x[ifx+24];
		for (i=ilow; i<=ihigh-1; i+=2) {
			ya = y[i+1-ifx];  z1 = x0*ya;
			yb = y[i-ifx];  z0 = x0*yb;  z1 += x1*yb;
			ya = y[i-ifx-1];  z0 += x1*ya;  z1 += x2*ya;
			yb = y[i-ifx-2];  z0 += x2*yb;  z1 += x3*yb;
			ya = y[i-ifx-3];  z0 += x3*ya;  z1 += x4*ya;
			yb = y[i-ifx-4];  z0 += x4*yb;  z1 += x5*yb;
			ya = y[i-ifx-5];  z0 += x5*ya;  z1 += x6*ya;
			yb = y[i-ifx-6];  z0 += x6*yb;  z1 += x7*yb;
			ya = y[i-ifx-7];  z0 += x7*ya;  z1 += x8*ya;
			yb = y[i-ifx-8];  z0 += x8*yb;  z1 += x9*yb;
			ya = y[i-ifx-9];  z0 += x9*ya;  z1 += x10*ya;
			yb = y[i-ifx-10];  z0 += x10*yb;  z1 += x11*yb;
			ya = y[i-ifx-11];  z0 += x11*ya;  z1 += x12*ya;
			yb = y[i-ifx-12];  z0 += x12*yb;  z1 += x13*yb;
			ya = y[i-ifx-13];  z0 += x13*ya;  z1 += x14*ya;
			yb = y[i-ifx-14];  z0 += x14*yb;  z1 += x15*yb;
			ya = y[i-ifx-15];  z0 += x15*ya;  z1 += x16*ya;
			yb = y[i-ifx-16];  z0 += x16*yb;  z1 += x17*yb;
			ya = y[i-ifx-17];  z0 += x17*ya;  z1 += x18*ya;
			yb = y[i-ifx-18];  z0 += x18*yb;  z1 += x19*yb;
			ya = y[i-ifx-19];  z0 += x19*ya;  z1 += x20*ya;
			yb = y[i-ifx-20];  z0 += x20*yb;  z1 += x21*yb;
			ya = y[i-ifx-21];  z0 += x21*ya;  z1 += x22*ya;
			yb = y[i-ifx-22];  z0 += x22*yb;  z1 += x23*yb;
			ya = y[i-ifx-23];  z0 += x23*ya;  z1 += x24*ya;
			yb = y[i-ifx-24];  z0 += x24*yb;
			z[i+1] = z1;
			z[i] = z0;
		}
	} else if (lx==26) {
		x0 = x[ifx];
		x1 = x[ifx+1];
		x2 = x[ifx+2];
		x3 = x[ifx+3];
		x4 = x[ifx+4];
		x5 = x[ifx+5];
		x6 = x[ifx+6];
		x7 = x[ifx+7];
		x8 = x[ifx+8];
		x9 = x[ifx+9];
		x10 = x[ifx+10];
		x11 = x[ifx+11];
		x12 = x[ifx+12];
		x13 = x[ifx+13];
		x14 = x[ifx+14];
		x15 = x[ifx+15];
		x16 = x[ifx+16];
		x17 = x[ifx+17];
		x18 = x[ifx+18];
		x19 = x[ifx+19];
		x20 = x[ifx+20];
		x21 = x[ifx+21];
		x22 = x[ifx+22];
		x23 = x[ifx+23];
		x24 = x[ifx+24];
		x25 = x[ifx+25];
		for (i=ilow; i<=ihigh-1; i+=2) {
			ya = y[i+1-ifx];  z1 = x0*ya;
			yb = y[i-ifx];  z0 = x0*yb;  z1 += x1*yb;
			ya = y[i-ifx-1];  z0 += x1*ya;  z1 += x2*ya;
			yb = y[i-ifx-2];  z0 += x2*yb;  z1 += x3*yb;
			ya = y[i-ifx-3];  z0 += x3*ya;  z1 += x4*ya;
			yb = y[i-ifx-4];  z0 += x4*yb;  z1 += x5*yb;
			ya = y[i-ifx-5];  z0 += x5*ya;  z1 += x6*ya;
			yb = y[i-ifx-6];  z0 += x6*yb;  z1 += x7*yb;
			ya = y[i-ifx-7];  z0 += x7*ya;  z1 += x8*ya;
			yb = y[i-ifx-8];  z0 += x8*yb;  z1 += x9*yb;
			ya = y[i-ifx-9];  z0 += x9*ya;  z1 += x10*ya;
			yb = y[i-ifx-10];  z0 += x10*yb;  z1 += x11*yb;
			ya = y[i-ifx-11];  z0 += x11*ya;  z1 += x12*ya;
			yb = y[i-ifx-12];  z0 += x12*yb;  z1 += x13*yb;
			ya = y[i-ifx-13];  z0 += x13*ya;  z1 += x14*ya;
			yb = y[i-ifx-14];  z0 += x14*yb;  z1 += x15*yb;
			ya = y[i-ifx-15];  z0 += x15*ya;  z1 += x16*ya;
			yb = y[i-ifx-16];  z0 += x16*yb;  z1 += x17*yb;
			ya = y[i-ifx-17];  z0 += x17*ya;  z1 += x18*ya;
			yb = y[i-ifx-18];  z0 += x18*yb;  z1 += x19*yb;
			ya = y[i-ifx-19];  z0 += x19*ya;  z1 += x20*ya;
			yb = y[i-ifx-20];  z0 += x20*yb;  z1 += x21*yb;
			ya = y[i-ifx-21];  z0 += x21*ya;  z1 += x22*ya;
			yb = y[i-ifx-22];  z0 += x22*yb;  z1 += x23*yb;
			ya = y[i-ifx-23];  z0 += x23*ya;  z1 += x24*ya;
			yb = y[i-ifx-24];  z0 += x24*yb;  z1 += x25*yb;
			ya = y[i-ifx-25];  z0 += x25*ya;
			z[i+1] = z1;
			z[i] = z0;
		}
	} else if (lx==27) {
		x0 = x[ifx];
		x1 = x[ifx+1];
		x2 = x[ifx+2];
		x3 = x[ifx+3];
		x4 = x[ifx+4];
		x5 = x[ifx+5];
		x6 = x[ifx+6];
		x7 = x[ifx+7];
		x8 = x[ifx+8];
		x9 = x[ifx+9];
		x10 = x[ifx+10];
		x11 = x[ifx+11];
		x12 = x[ifx+12];
		x13 = x[ifx+13];
		x14 = x[ifx+14];
		x15 = x[ifx+15];
		x16 = x[ifx+16];
		x17 = x[ifx+17];
		x18 = x[ifx+18];
		x19 = x[ifx+19];
		x20 = x[ifx+20];
		x21 = x[ifx+21];
		x22 = x[ifx+22];
		x23 = x[ifx+23];
		x24 = x[ifx+24];
		x25 = x[ifx+25];
		x26 = x[ifx+26];
		for (i=ilow; i<=ihigh-1; i+=2) {
			ya = y[i+1-ifx];  z1 = x0*ya;
			yb = y[i-ifx];  z0 = x0*yb;  z1 += x1*yb;
			ya = y[i-ifx-1];  z0 += x1*ya;  z1 += x2*ya;
			yb = y[i-ifx-2];  z0 += x2*yb;  z1 += x3*yb;
			ya = y[i-ifx-3];  z0 += x3*ya;  z1 += x4*ya;
			yb = y[i-ifx-4];  z0 += x4*yb;  z1 += x5*yb;
			ya = y[i-ifx-5];  z0 += x5*ya;  z1 += x6*ya;
			yb = y[i-ifx-6];  z0 += x6*yb;  z1 += x7*yb;
			ya = y[i-ifx-7];  z0 += x7*ya;  z1 += x8*ya;
			yb = y[i-ifx-8];  z0 += x8*yb;  z1 += x9*yb;
			ya = y[i-ifx-9];  z0 += x9*ya;  z1 += x10*ya;
			yb = y[i-ifx-10];  z0 += x10*yb;  z1 += x11*yb;
			ya = y[i-ifx-11];  z0 += x11*ya;  z1 += x12*ya;
			yb = y[i-ifx-12];  z0 += x12*yb;  z1 += x13*yb;
			ya = y[i-ifx-13];  z0 += x13*ya;  z1 += x14*ya;
			yb = y[i-ifx-14];  z0 += x14*yb;  z1 += x15*yb;
			ya = y[i-ifx-15];  z0 += x15*ya;  z1 += x16*ya;
			yb = y[i-ifx-16];  z0 += x16*yb;  z1 += x17*yb;
			ya = y[i-ifx-17];  z0 += x17*ya;  z1 += x18*ya;
			yb = y[i-ifx-18];  z0 += x18*yb;  z1 += x19*yb;
			ya = y[i-ifx-19];  z0 += x19*ya;  z1 += x20*ya;
			yb = y[i-ifx-20];  z0 += x20*yb;  z1 += x21*yb;
			ya = y[i-ifx-21];  z0 += x21*ya;  z1 += x22*ya;
			yb = y[i-ifx-22];  z0 += x22*yb;  z1 += x23*yb;
			ya = y[i-ifx-23];  z0 += x23*ya;  z1 += x24*ya;
			yb = y[i-ifx-24];  z0 += x24*yb;  z1 += x25*yb;
			ya = y[i-ifx-25];  z0 += x25*ya;  z1 += x26*ya;
			yb = y[i-ifx-26];  z0 += x26*yb;
			z[i+1] = z1;
			z[i] = z0;
		}
	} else if (lx==28) {
		x0 = x[ifx];
		x1 = x[ifx+1];
		x2 = x[ifx+2];
		x3 = x[ifx+3];
		x4 = x[ifx+4];
		x5 = x[ifx+5];
		x6 = x[ifx+6];
		x7 = x[ifx+7];
		x8 = x[ifx+8];
		x9 = x[ifx+9];
		x10 = x[ifx+10];
		x11 = x[ifx+11];
		x12 = x[ifx+12];
		x13 = x[ifx+13];
		x14 = x[ifx+14];
		x15 = x[ifx+15];
		x16 = x[ifx+16];
		x17 = x[ifx+17];
		x18 = x[ifx+18];
		x19 = x[ifx+19];
		x20 = x[ifx+20];
		x21 = x[ifx+21];
		x22 = x[ifx+22];
		x23 = x[ifx+23];
		x24 = x[ifx+24];
		x25 = x[ifx+25];
		x26 = x[ifx+26];
		x27 = x[ifx+27];
		for (i=ilow; i<=ihigh-1; i+=2) {
			ya = y[i+1-ifx];  z1 = x0*ya;
			yb = y[i-ifx];  z0 = x0*yb;  z1 += x1*yb;
			ya = y[i-ifx-1];  z0 += x1*ya;  z1 += x2*ya;
			yb = y[i-ifx-2];  z0 += x2*yb;  z1 += x3*yb;
			ya = y[i-ifx-3];  z0 += x3*ya;  z1 += x4*ya;
			yb = y[i-ifx-4];  z0 += x4*yb;  z1 += x5*yb;
			ya = y[i-ifx-5];  z0 += x5*ya;  z1 += x6*ya;
			yb = y[i-ifx-6];  z0 += x6*yb;  z1 += x7*yb;
			ya = y[i-ifx-7];  z0 += x7*ya;  z1 += x8*ya;
			yb = y[i-ifx-8];  z0 += x8*yb;  z1 += x9*yb;
			ya = y[i-ifx-9];  z0 += x9*ya;  z1 += x10*ya;
			yb = y[i-ifx-10];  z0 += x10*yb;  z1 += x11*yb;
			ya = y[i-ifx-11];  z0 += x11*ya;  z1 += x12*ya;
			yb = y[i-ifx-12];  z0 += x12*yb;  z1 += x13*yb;
			ya = y[i-ifx-13];  z0 += x13*ya;  z1 += x14*ya;
			yb = y[i-ifx-14];  z0 += x14*yb;  z1 += x15*yb;
			ya = y[i-ifx-15];  z0 += x15*ya;  z1 += x16*ya;
			yb = y[i-ifx-16];  z0 += x16*yb;  z1 += x17*yb;
			ya = y[i-ifx-17];  z0 += x17*ya;  z1 += x18*ya;
			yb = y[i-ifx-18];  z0 += x18*yb;  z1 += x19*yb;
			ya = y[i-ifx-19];  z0 += x19*ya;  z1 += x20*ya;
			yb = y[i-ifx-20];  z0 += x20*yb;  z1 += x21*yb;
			ya = y[i-ifx-21];  z0 += x21*ya;  z1 += x22*ya;
			yb = y[i-ifx-22];  z0 += x22*yb;  z1 += x23*yb;
			ya = y[i-ifx-23];  z0 += x23*ya;  z1 += x24*ya;
			yb = y[i-ifx-24];  z0 += x24*yb;  z1 += x25*yb;
			ya = y[i-ifx-25];  z0 += x25*ya;  z1 += x26*ya;
			yb = y[i-ifx-26];  z0 += x26*yb;  z1 += x27*yb;
			ya = y[i-ifx-27];  z0 += x27*ya;
			z[i+1] = z1;
			z[i] = z0;
		}
	} else if (lx==29) {
		x0 = x[ifx];
		x1 = x[ifx+1];
		x2 = x[ifx+2];
		x3 = x[ifx+3];
		x4 = x[ifx+4];
		x5 = x[ifx+5];
		x6 = x[ifx+6];
		x7 = x[ifx+7];
		x8 = x[ifx+8];
		x9 = x[ifx+9];
		x10 = x[ifx+10];
		x11 = x[ifx+11];
		x12 = x[ifx+12];
		x13 = x[ifx+13];
		x14 = x[ifx+14];
		x15 = x[ifx+15];
		x16 = x[ifx+16];
		x17 = x[ifx+17];
		x18 = x[ifx+18];
		x19 = x[ifx+19];
		x20 = x[ifx+20];
		x21 = x[ifx+21];
		x22 = x[ifx+22];
		x23 = x[ifx+23];
		x24 = x[ifx+24];
		x25 = x[ifx+25];
		x26 = x[ifx+26];
		x27 = x[ifx+27];
		x28 = x[ifx+28];
		for (i=ilow; i<=ihigh-1; i+=2) {
			ya = y[i+1-ifx];  z1 = x0*ya;
			yb = y[i-ifx];  z0 = x0*yb;  z1 += x1*yb;
			ya = y[i-ifx-1];  z0 += x1*ya;  z1 += x2*ya;
			yb = y[i-ifx-2];  z0 += x2*yb;  z1 += x3*yb;
			ya = y[i-ifx-3];  z0 += x3*ya;  z1 += x4*ya;
			yb = y[i-ifx-4];  z0 += x4*yb;  z1 += x5*yb;
			ya = y[i-ifx-5];  z0 += x5*ya;  z1 += x6*ya;
			yb = y[i-ifx-6];  z0 += x6*yb;  z1 += x7*yb;
			ya = y[i-ifx-7];  z0 += x7*ya;  z1 += x8*ya;
			yb = y[i-ifx-8];  z0 += x8*yb;  z1 += x9*yb;
			ya = y[i-ifx-9];  z0 += x9*ya;  z1 += x10*ya;
			yb = y[i-ifx-10];  z0 += x10*yb;  z1 += x11*yb;
			ya = y[i-ifx-11];  z0 += x11*ya;  z1 += x12*ya;
			yb = y[i-ifx-12];  z0 += x12*yb;  z1 += x13*yb;
			ya = y[i-ifx-13];  z0 += x13*ya;  z1 += x14*ya;
			yb = y[i-ifx-14];  z0 += x14*yb;  z1 += x15*yb;
			ya = y[i-ifx-15];  z0 += x15*ya;  z1 += x16*ya;
			yb = y[i-ifx-16];  z0 += x16*yb;  z1 += x17*yb;
			ya = y[i-ifx-17];  z0 += x17*ya;  z1 += x18*ya;
			yb = y[i-ifx-18];  z0 += x18*yb;  z1 += x19*yb;
			ya = y[i-ifx-19];  z0 += x19*ya;  z1 += x20*ya;
			yb = y[i-ifx-20];  z0 += x20*yb;  z1 += x21*yb;
			ya = y[i-ifx-21];  z0 += x21*ya;  z1 += x22*ya;
			yb = y[i-ifx-22];  z0 += x22*yb;  z1 += x23*yb;
			ya = y[i-ifx-23];  z0 += x23*ya;  z1 += x24*ya;
			yb = y[i-ifx-24];  z0 += x24*yb;  z1 += x25*yb;
			ya = y[i-ifx-25];  z0 += x25*ya;  z1 += x26*ya;
			yb = y[i-ifx-26];  z0 += x26*yb;  z1 += x27*yb;
			ya = y[i-ifx-27];  z0 += x27*ya;  z1 += x28*ya;
			yb = y[i-ifx-28];  z0 += x28*yb;
			z[i+1] = z1;
			z[i] = z0;
		}
	} else if (lx==30) {
		x0 = x[ifx];
		x1 = x[ifx+1];
		x2 = x[ifx+2];
		x3 = x[ifx+3];
		x4 = x[ifx+4];
		x5 = x[ifx+5];
		x6 = x[ifx+6];
		x7 = x[ifx+7];
		x8 = x[ifx+8];
		x9 = x[ifx+9];
		x10 = x[ifx+10];
		x11 = x[ifx+11];
		x12 = x[ifx+12];
		x13 = x[ifx+13];
		x14 = x[ifx+14];
		x15 = x[ifx+15];
		x16 = x[ifx+16];
		x17 = x[ifx+17];
		x18 = x[ifx+18];
		x19 = x[ifx+19];
		x20 = x[ifx+20];
		x21 = x[ifx+21];
		x22 = x[ifx+22];
		x23 = x[ifx+23];
		x24 = x[ifx+24];
		x25 = x[ifx+25];
		x26 = x[ifx+26];
		x27 = x[ifx+27];
		x28 = x[ifx+28];
		x29 = x[ifx+29];
		for (i=ilow; i<=ihigh-1; i+=2) {
			ya = y[i+1-ifx];  z1 = x0*ya;
			yb = y[i-ifx];  z0 = x0*yb;  z1 += x1*yb;
			ya = y[i-ifx-1];  z0 += x1*ya;  z1 += x2*ya;
			yb = y[i-ifx-2];  z0 += x2*yb;  z1 += x3*yb;
			ya = y[i-ifx-3];  z0 += x3*ya;  z1 += x4*ya;
			yb = y[i-ifx-4];  z0 += x4*yb;  z1 += x5*yb;
			ya = y[i-ifx-5];  z0 += x5*ya;  z1 += x6*ya;
			yb = y[i-ifx-6];  z0 += x6*yb;  z1 += x7*yb;
			ya = y[i-ifx-7];  z0 += x7*ya;  z1 += x8*ya;
			yb = y[i-ifx-8];  z0 += x8*yb;  z1 += x9*yb;
			ya = y[i-ifx-9];  z0 += x9*ya;  z1 += x10*ya;
			yb = y[i-ifx-10];  z0 += x10*yb;  z1 += x11*yb;
			ya = y[i-ifx-11];  z0 += x11*ya;  z1 += x12*ya;
			yb = y[i-ifx-12];  z0 += x12*yb;  z1 += x13*yb;
			ya = y[i-ifx-13];  z0 += x13*ya;  z1 += x14*ya;
			yb = y[i-ifx-14];  z0 += x14*yb;  z1 += x15*yb;
			ya = y[i-ifx-15];  z0 += x15*ya;  z1 += x16*ya;
			yb = y[i-ifx-16];  z0 += x16*yb;  z1 += x17*yb;
			ya = y[i-ifx-17];  z0 += x17*ya;  z1 += x18*ya;
			yb = y[i-ifx-18];  z0 += x18*yb;  z1 += x19*yb;
			ya = y[i-ifx-19];  z0 += x19*ya;  z1 += x20*ya;
			yb = y[i-ifx-20];  z0 += x20*yb;  z1 += x21*yb;
			ya = y[i-ifx-21];  z0 += x21*ya;  z1 += x22*ya;
			yb = y[i-ifx-22];  z0 += x22*yb;  z1 += x23*yb;
			ya = y[i-ifx-23];  z0 += x23*ya;  z1 += x24*ya;
			yb = y[i-ifx-24];  z0 += x24*yb;  z1 += x25*yb;
			ya = y[i-ifx-25];  z0 += x25*ya;  z1 += x26*ya;
			yb = y[i-ifx-26];  z0 += x26*yb;  z1 += x27*yb;
			ya = y[i-ifx-27];  z0 += x27*ya;  z1 += x28*ya;
			yb = y[i-ifx-28];  z0 += x28*yb;  z1 += x29*yb;
			ya = y[i-ifx-29];  z0 += x29*ya;
			z[i+1] = z1;
			z[i] = z0;
		}
	}
	if (ihigh>=ilow && (ihigh-ilow)%2==0) {
		ilow = ihigh;
		jlow = ifx;
		jhigh = ilx;
		for (i=ilow; i<=ihigh; ++i) {
			for (j=jlow,sum=0.0; j<=jhigh; ++j)
				sum += x[j]*y[i-j];
			z[i] = sum;
		}
	}
	
	/* ROLLING OFF:  ily+ifx < i <= ily+ilx */
	ilow = ily+ifx+1;  if (ilow<ifz) ilow = ifz;
	ihigh = ily+ilx;  if (ihigh>ilz) ihigh = ilz;
	jlow = ilow-ily;
	jhigh = ilx;
	for (i=ilow; i<=ihigh; ++i,++jlow) {
		for (j=jlow,sum=0.0; j<=jhigh; ++j)
			sum += x[j]*y[i-j];
		z[i] = sum;
	}
	
	/* OFF RIGHT:  ily+ilx < i */
	ilow = ily+ilx+1;  if (ilow<ifz) ilow = ifz;
	ihigh = ilz;
	for (i=ilow; i<=ihigh; ++i)
		z[i] = 0.0;
}
#endif

#ifdef TEST
static void conv1 (int lx, int ifx, float *x,
	   int ly, int ify, float *y, 
	   int lz, int ifz, float *z)
{
	int ilx=ifx+lx-1,ily=ify+ly-1,ilz=ifz+lz-1,i,j,jlow,jhigh;
	float sum;
	
	x -= ifx;  y -= ify;  z -= ifz;
	for (i=ifz; i<=ilz; ++i) {
		jlow = i-ily;  if (jlow<ifx) jlow = ifx;
		jhigh = i-ify;  if (jhigh>ilx) jhigh = ilx;
		for (j=jlow,sum=0.0; j<=jhigh; ++j)
			sum += x[j]*y[i-j];
		z[i] = sum;
	}
}
#include "cwp.h"
main()
{
	int lx,ly,lz,ifx,ify,ifz,i;
	float *x,*y,*z,*z1;
	
	/* loop over tests */
	while(1) {
	
		/* make parameters */
		lx = 1+100*franuni();
		ly = 1+100*franuni();
		lz = 1+100*franuni();
		ifx = -100+200*franuni();
		ify = -100+200*franuni();
		ifz = -100+200*franuni();
		
		/* allocate space */
		x = alloc1float(lx);
		y = alloc1float(ly);
		z = alloc1float(lz);
		z1 = alloc1float(lz);
			
		/* fill x and y arrays */
		for (i=0; i<lx; i++)
			x[i] = i+1;
		for (i=0; i<ly; i++)
			y[i] = i+1;

		/* convolve and check */
		conv1(lx,ifx,x,ly,ify,y,lz,ifz,z1);
		conv(lx,ifx,x,ly,ify,y,lz,ifz,z);
		for (i=0; i<lz; ++i)
			if (z[i]!=z1[i]) break;
		if (i<lz) {
			printf("%10d %10d %10d %10d %10d %10d\n",
				lx,ifx,ly,ify,lz,ifz);
			printf("z1[%d]=%g != z[%d]=%g\n",
				i+ifz,z1[i],i+ifz,z[i]);
			pp1d(stdout,"simple",lz,ifz,z1);
			pp1d(stdout,"optimized",lz,ifz,z);
			exit(-1);
		}
				
		/* free space */
		free1float(x);
		free1float(y);
		free1float(z);
		free1float(z1);
	}
}
#endif