Bitwise shift of buffer in CUDA [duplicate] - c++

Question

What is the best way to implement a bitwise memmove? The method should take an additional destination and source bit-offset and the count should be in bits too.
I saw that ARM provides a non-standard _membitmove, which does exactly what I need, but I couldn't find its source.
Bind's bitset includes isc_bitstring_copy, but it's not efficient
I'm aware that the C standard library doesn't provide such a method, but I also couldn't find any third-party code providing a similar method.

Answer 1

Assuming "best" means "easiest", you can copy bits one by one. Conceptually, an address of a bit is an object (struct) that has a pointer to a byte in memory and an index of a bit in the byte.
struct pointer_to_bit
{
uint8_t* p;
int b;
};
void membitmovebl(
void *dest,
const void *src,
int dest_offset,
int src_offset,
size_t nbits)
{
// Create pointers to bits
struct pointer_to_bit d = {dest, dest_offset};
struct pointer_to_bit s = {src, src_offset};
// Bring the bit offsets to range (0...7)
d.p += d.b / 8; // replace division by right-shift if bit offset can be negative
d.b %= 8; // replace "%=8" by "&=7" if bit offset can be negative
s.p += s.b / 8;
s.b %= 8;
// Determine whether it's OK to loop forward
if (d.p < s.p || d.p == s.p && d.b <= s.b)
{
// Copy bits one by one
for (size_t i = 0; i < nbits; i++)
{
// Read 1 bit
int bit = (*s.p >> s.b) & 1;
// Write 1 bit
*d.p &= ~(1 << d.b);
*d.p |= bit << d.b;
// Advance pointers
if (++s.b == 8)
{
s.b = 0;
++s.p;
}
if (++d.b == 8)
{
d.b = 0;
++d.p;
}
}
}
else
{
// Copy stuff backwards - essentially the same code but ++ replaced by --
}
}
If you want to write a version optimized for speed, you will have to do copying by bytes (or, better, words), unroll loops, and handle a number of special cases (memmove does that; you will have to do more because your function is more complicated).
P.S. Oh, seeing that you call isc_bitstring_copy inefficient, you probably want the speed optimization. You can use the following idea:
Start copying bits individually until the destination is byte-aligned (d.b == 0). Then, it is easy to copy 8 bits at once, doing some bit twiddling. Do this until there are less than 8 bits left to copy; then continue copying bits one by one.
// Copy 8 bits from s to d and advance pointers
*d.p = *s.p++ >> s.b;
*d.p++ |= *s.p << (8 - s.b);
P.P.S Oh, and seeing your comment on what you are going to use the code for, you don't really need to implement all the versions (byte/halfword/word, big/little-endian); you only want the easiest one - the one working with words (uint32_t).

Answer 2

Here is a partial implementation (not tested). There are obvious efficiency and usability improvements.
Copy n bytes from src to dest (not overlapping src), and shift bits at dest rightwards by bit bits, 0 <= bit <= 7. This assumes that the least significant bits are at the right of the bytes
void memcpy_with_bitshift(unsigned char *dest, unsigned char *src, size_t n, int bit)
{
int i;
memcpy(dest, src, n);
for (i = 0; i < n; i++) {
dest[i] >> bit;
}
for (i = 0; i < n; i++) {
dest[i+1] |= (src[i] << (8 - bit));
}
}
Some improvements to be made:
Don't overwrite first bit bits at beginning of dest.
Merge loops
Have a way to copy a number of bits not divisible by 8
Fix for >8 bits in a char