在FPGA的情况下,除法运算(/)是否昂贵?是否可以通过基本移位操作执行两个Q15格式编号(16位定点编号)的划分?
有人可以提供一些例子来帮助我吗?
提前致谢!
答案 0 :(得分:0)
定点运算只是整数运算,抛出了一些比例.Q15是一个纯粹的小数格式,存储为带符号的16位整数,比例因子为2 15 ,能够表示区间[-1,1]中的值。很明显,除数的大小超过了被除数的大小时,除法中的商数超过可表示的范围时,除法仅在Q15中有意义。
在着手定制Verilog定点分区之前,您需要检查FPGA供应商的库产品,作为定点库,包括管道分区。还有一些可能相关的源项目,例如this one。
当使用整数除法运算符进行定点除法时,我们需要调整除法除去比例因子的事实,即(a * 2 scale )/(b * 2 < sup> scale )=(a / b),而正确的定点结果是(a / b * 2 scale )。这可以通过将被除数乘以2 scale 来轻松解决,如下面的C实现:
int16_t div_q15 (int16_t dividend, int16_t divisor)
{
return (int16_t)(((int32_t)dividend << 15) / (int32_t)divisor);
}
Wikipedia对如何使用加,减和移位操作逐位实现二进制除法进行了合理的研究。这些方法与小学教授的长手分工密切相关。对于FPGA,如this paper所指出的那样,如果经常使用非恢复方法,则使用非恢复方法,例如:
Nikolay Sorokin,“在FPGA上实现高速定点分频器”。 计算机科学与技术杂志技术,Vol。 2006年4月6日第1期,第8-11页。
这是C代码,显示了非恢复方法如何用于16位二进制补码操作数的划分:
/* bit-wise non-restoring two's complement division */
void int16_div (int16_t dividend, int16_t divisor, int16_t *quot, int16_t *rem)
{
const int operand_bits = (int) (sizeof (int16_t) * CHAR_BIT);
uint16_t d = (uint16_t)divisor;
uint16_t nd = 0 - d; /* -divisor */
uint16_t r, q = 0; /* remainder, quotient */
uint32_t dd = (uint32_t)d << operand_bits; /* expanded divisor */
uint32_t pp = dividend; /* partial remainder */
int i;
for (i = operand_bits - 1; i >= 0; i--) {
if ((int32_t)(pp ^ dd) < 0) {
q = (q << 1) + 0; /* record quotient bit -1 (as 0) */
pp = (pp << 1) + dd;
} else {
q = (q << 1) + 1; /* record quotient bit +1 (as 1) */
pp = (pp << 1) - dd;
}
}
/* convert quotient from digit set {-1,1} to plain two's complement */
q = (q << 1) + 1;
/* remainder is upper half of partial remainder */
r = (uint16_t)(pp >> operand_bits);
/* fix up cases where we worked past a partial remainder of zero */
if (r == d) { /* remainder equal to divisor */
q = q + 1;
r = 0;
} else if (r == nd) { /* remainder equal to -divisor */
q = q - 1;
r = 0;
}
/* for truncating division, remainder must have same sign as dividend */
if (r && ((int16_t)(dividend ^ r) < 0)) {
if ((int16_t)q < 0) {
q = q + 1;
r = r - d;
} else {
q = q - 1;
r = r + d;
}
}
*quot = (int16_t)q;
*rem = (int16_t)r;
}
请注意,有多种方法可以处理非恢复分部中出现的各种特殊情况。例如,经常看到检测到零部分余数pp
的代码,并且在这种情况下早期退出超过商位的循环。在这里,我假设FPGA实现将完全展开循环以创建流水线实现,在这种情况下,提前终止没有帮助。相反,最终修正适用于那些受忽略零部分余数影响的商数。
为了从上面创建一个Q15部门,我们只需做一个改变:结合股息的扩大规模。而不是:
uint32_t pp = dividend; /* partial remainder */
我们现在用这个:
uint32_t pp = dividend << 15; /* partial remainder; incorporate Q15 scaling */
生成的C代码(抱歉,我不会提供读取使用的Verilog代码)包括测试框架:
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <limits.h>
#include <math.h>
/* bit-wise non-restoring two's complement division */
void q15_div (int16_t dividend, int16_t divisor, int16_t *quot, int16_t *rem)
{
const int operand_bits = (int) (sizeof (int16_t) * CHAR_BIT);
uint16_t d = (uint16_t)divisor;
uint16_t nd = 0 - d; /* -divisor */
uint16_t r, q = 0; /* remainder, quotient */
uint32_t dd = (uint32_t)d << operand_bits; /* expanded divisor */
uint32_t pp = dividend << 15; /* partial remainder, incorporate Q15 scaling */
int i;
for (i = operand_bits - 1; i >= 0; i--) {
if ((int32_t)(pp ^ dd) < 0) {
q = (q << 1) + 0; /* record quotient bit -1 (as 0) */
pp = (pp << 1) + dd;
} else {
q = (q << 1) + 1; /* record quotient bit +1 (as 1) */
pp = (pp << 1) - dd;
}
}
/* convert quotient from digit set {-1,1} to plain two's complement */
q = (q << 1) + 1;
/* remainder is upper half of partial remainder */
r = (uint16_t)(pp >> operand_bits);
/* fix up cases where we worked past a partial remainder of zero */
if (r == d) { /* remainder equal to divisor */
q = q + 1;
r = 0;
} else if (r == nd) { /* remainder equal to -divisor */
q = q - 1;
r = 0;
}
/* for truncating division, remainder must have same sign as dividend */
if (r && ((int16_t)(dividend ^ r) < 0)) {
if ((int16_t)q < 0) {
q = q + 1;
r = r - d;
} else {
q = q - 1;
r = r + d;
}
}
*quot = (int16_t)q;
*rem = (int16_t)r;
}
int main (void)
{
uint16_t dividend, divisor, ref_q, res_q, res_r;
double quot, fxscale = (1 << 15);
dividend = 0;
do {
printf ("\r%04x", dividend);
divisor = 1;
do {
quot = trunc (fxscale * (int16_t)dividend / (int16_t)divisor);
/* Q15 can only represent numbers in [-1, 1) */
if ((quot >= -1.0) && (quot < 1.0)) {
ref_q = (int16_t)((((int32_t)(int16_t)dividend) << 15) /
((int32_t)(int16_t)divisor));
q15_div ((int16_t)dividend, (int16_t)divisor,
(int16_t *)&res_q, (int16_t *)&res_r);
if (res_q != ref_q) {
printf ("!r dividend=%04x (%f) divisor=%04x (%f) res=%04x (%f) ref=%04x (%f)\n",
dividend, (int16_t)dividend / fxscale,
divisor, (int16_t)divisor / fxscale,
res_q, (int16_t)res_q / fxscale,
ref_q, (int16_t)ref_q / fxscale);
}
}
divisor++;
} while (divisor);
dividend++;
} while (dividend);
return EXIT_SUCCESS;
}