Least memory usage and best performance - rust

I need to store a billion "appearances" and I am looking for the most efficient way to store these with respect to both memory usage and performance. What are, for example, the differences in those respects for a1, a2, a3 in:
struct Appearance<'a> {
identity: &'a u64,
role: &'a str
}
struct AnotherAppearance<'a>((&'a u64, &'a str));
fn main() {
let thing = 42;
let hair_color = "hair color";
let a1 = Appearance {identity: &thing, role: &hair_color};
let a2 = AnotherAppearance((&thing, &hair_color));
let a3 = (&thing, &hair_color);
}
Are there better ways to work with such a structure? Also, is there a way to get detailed information about a1, a2, a3 so that I could see how they are represented in memory for myself?

First, as Ijedrz noted, all your proposed alternatives have the same size. In fact, from the compiler's point of view, they're all identical.
If you're after smaller memory sizes, you might be better off using something like:
struct Appearance {
identity: u32,
role: InternedString,
}
First of all, u32 has 4 billion distinct values, so you definitely don't need a u64 for a billion records. Aside from that, a &u64 is going to be the same size as a u64 on a 64-bit machine, so there's not much point in using it. That a u32 is half the size is a bonus.
Beyond that, &str seems incredibly wasteful. That's going to take two pointers for data which, I assume, is unlikely to change much. If there are many more Appearances than roles, your best bet is to intern the strings and reduce the field to a pointer (or even better: a u32 ID which indirects through another table). There is no interned string in the standard library, but they're not that hard to implement, assuming you can't find one somewhere. Such a structure (assuming InternedString is a u32 ID) would be 8 bytes versus your 24 bytes.
If performance is what you're after, that depends on how you use the structures. That said, &u64 is slower than u64, so changing that will probably help. As for the string, it depends on how you use it. If you mostly do comparisons, an interned string will be faster because you can compare those with a single comparison; comparing regular strings can be much slower since you have to actually look at the contents.

All three variants seem to have the same size:
use std::mem::size_of;
println!("a1: {}", size_of::<Appearance>()); // a1: 24
println!("a2: {}", size_of::<AnotherAppearance>()); // a2: 24
println!("a3: {}", size_of::<(&u64, &str)>()); // a3: 24
So I would just use the one that is most descriptive, i.e. Appearance.

Related

Computing u32 hash with FxHasher fast

I've been recently experimenting with different hash functions in Rust. Started off with the fasthash crate, where many algorithms are implemented; e.g., murmur3 is then called as
let hval = murmur3::hash32_with_seed(&tmp_read_buff, 123 as u32);
This works very fast (e.g., few seconds for 100000000 short inputs). I also stumbled upon FxHash, the algorithm used a lot internally in Firefox (at least initially?). I rolled my version of hashing a byte array with this algorithm as follows
use rustc_hash::FxHasher;
use std::hash::Hasher;
fn hash_with_fx(read_buff: &[u8]) -> u64 {
let mut hasher = FxHasher::default();
for el in read_buff {
hasher.write_u8(*el);
}
return hasher.finish();
}
This works, however, it's about 5x slower. I'd really like to know if I'm missing something apparent here/how could I achieve similar or better speeds to fasthash's e.g., murmur3. My intuition is that with FxHash, the core operation is very simple,
self.hash = self.hash.rotate_left(5).bitxor(i).wrapping_mul(K);
hence it should be one of the fastest.
The FxHasher documentation mentions that:
the speed of the hash function itself is much higher because it works on up to 8 bytes at a time.
But your algorithm completely removes this possibility because you are processing each byte individually. You can make it a lot faster by hashing in chunks of 8 bytes.
fn hash_with_fx(read_buff: &[u8]) -> u64 {
let mut hasher = FxHasher::default();
let mut chunks = read_buff.chunks_exact(8);
for bytes in &mut chunks {
// unwrap is ok because `chunks_exact` provides the guarantee that
// the `bytes` slice is exactly the requested length
let int = u64::from_be_bytes(bytes.try_into().unwrap());
hasher.write_u64(int);
}
for byte in chunks.remainder() {
hasher.write_u8(*byte);
}
hasher.finish()
}
For very small inputs (especially for numbers like 7 bytes) it may introduce a small extra overhead compared with your original code. But, for larger inputs, it ought to be significantly faster.
Bonus material
It should be possible to remove a few extra instructions in the loop by using unwrap_unchecked instead of unwrap. It's completely sound to do so in this case, but it may not be worth introducing unsafe code into your codebase. I would measure the difference before deciding to include unsafe code.

store neon vector register to memory

This seems like a stupid question, but I can't for the life work out how to do it.
I have a buffer like;
let result_buff: &[u8]
and I have some code like
let anded_value: uint8x16_t = unsafe { vandq_u8(sa1, sa2) };
I just want to copy this result from the vector register to result_buff[x..16], probably via an intrinsic?
basically it's 16 bytes long and I need to work out the syntax to copy it to a particular position in the [u8] / Vec<u8>.
Just use the intrinsic vst1q_u8:
vst1q_u8(result_buff.as_mut_ptr(), anded_value);
You might be making it way more complicated than it actually is ;)
pub unsafe fn convert_to_array(d_in: uint8x16_t) -> [u8; 16]{
transmute(d_in)
}
https://godbolt.org/z/6M94bh4Pf

Performance difference between bitpacking bytes into a u32 vs storing them in a vec<u8>?

Intro:
I'm curious about the performance difference (both cpu and memory usage) of storing small numbers as bitpacked unsigned integers versus vectors of bytes
Example
I'll use the example of storing RGBA values. They're 4 Bytes so it is very tempting to store them as a u32.
However, it would be more readable to store them as a vector of type u8.
As a more detailed example, say I want to store and retrieve the color rgba(255,0,0,255)
This is how I would go about doing the two methods:
// Bitpacked:
let i: u32 = 4278190335;
//binary is 11111111 00000000 00000000 11111111
//In reality I would most likely do something more similar to:
let i: u32 = 255 << 24 + 255; //i think this syntax is right
// Vector:
let v: Vec<u8> = [255,0,0,255];
Then the two red values could be queried with
i >> 24
//or
&v[0]
//both expressions evaluate to 255 (i think. I'm really new to rust <3 )
Question 1
As far as I know, the values of v must be stored on the heap and so there are the performance costs that are associated with that. Are these costs significant enough to make bit packing worth it?
Question 2
Then there's the two expressions i >> 24 and &v[0]. I don't know how fast rust is at bit shifting versus getting values off the heap. I'd test it but I won't have access to a machine with rust installed for a while. Are there any immediate insights someone could give on the drawbacks of these two operations?
Question 3
Finally, is the difference in memory usage as simple as just storing 32 bits on the stack for the u32 versus storing 64 bits on the stack for the pointer v as well as 32 bits on the heap for the values of v?
Sorry if this question is a bit confusing
Using a Vec will be more expensive; as you mentioned, it will need to perform heap allocations, and access will be bounds-checked as well.
That said, if you use an array [u8; 4] instead, the performance compared with a bitpacked u32 representation should be almost identical.
In fact, consider the following simple example:
pub fn get_red_bitpacked(i: u32) -> u8 {
(i >> 24) as u8
}
pub fn get_red_array(v: [u8; 4]) -> u8 {
v[3]
}
pub fn test_bits(colour: u8) -> u8 {
let colour = colour as u32;
let i = (colour << 24) + colour;
get_red_bitpacked(i)
}
pub fn test_arr(colour: u8) -> u8 {
let v = [colour, 0, 0, colour];
get_red_array(v)
}
I took a look on Compiler Explorer, and the compiler decided that get_red_bitpacked and get_red_array were completely identical: so much so it didn't even bother generating code for the former. The two "test" functions obviously optimised to the exact same assembly as well.
example::get_red_array:
mov eax, edi
shr eax, 24
ret
example::test_bits:
mov eax, edi
ret
example::test_arr:
mov eax, edi
ret
Obviously this example was seen through by the compiler: for a proper comparison you should benchmark with actual code. That said, I feel fairly safe in saying that with Rust the performance of u32 versus [u8; 4] for these kinds of operations should be identical in general.
tl;dr use a struct:
struct Color {
r: u8,
g: u8,
b: u8,
a: u8,
}
Maybe use repr(packed) as well.
It gives you the best of all worlds and you can give the channels their name.
Are these costs significant enough to make bit packing worth it?
Heap allocation has a huge cost.
Are there any immediate insights someone could give on the drawbacks of these two operations?
Both are noise compared to allocating memory.

How to convert Vec<Rgb<u8>> to Vec<u8>

Using the Piston image crate, I can write an image by feeding it a Vec<u8>, but my actual data is Vec<Rgb<u8>> (because that is a lot easier to deal with, and I want to grow it dynamically).
How can I convert Vec<Rgb<u8>> to Vec<u8>? Rgb<u8> is really [u8; 3]. Does this have to be an unsafe conversion?
The answer depends on whether you are fine with copying the data. If copying is not an issue for you, you can do something like this:
let img: Vec<Rgb<u8>> = ...;
let buf: Vec<u8> = img.iter().flat_map(|rgb| rgb.data.iter()).cloned().collect();
If you want to perform the conversion without copying, though, we first need to make sure that your source and destination types actually have the same memory layout. Rust makes very few guarantees about the memory layout of structs. It currently does not even guarantee that a struct with a single member has the same memory layout as the member itself.
In this particular case, the Rust memory layout is not relevant though, since Rgb is defined as
#[repr(C)]
pub struct Rgb<T: Primitive> {
pub data: [T; 3],
}
The #[repr(C)] attribute specifies that the memory layout of the struct should be the same as an equivalent C struct. The C memory layout is not fully specified in the C standard, but according to the unsafe code guidelines, there are some rules that hold for "most" platforms:
Field order is preserved.
The first field begins at offset 0.
Assuming the struct is not packed, each field's offset is aligned to the ABI-mandated alignment for that field's type, possibly creating unused padding bits.
The total size of the struct is rounded up to its overall alignment.
As pointed out in the comments, the C standard theoretically allows additional padding at the end of the struct. However, the Piston image library itself makes the assumption that a slice of channel data has the same memory layout as the Rgb struct, so if you are on a platform where this assumption does not hold, all bets are off anyway (and I couldnt' find any evidence that such a platform exists).
Rust does guarantee that arrays, slices and vectors are densely packed, and that structs and arrays have an alignment equal to the maximum alignment of their elements. Together with the assumption that the layout of Rgb is as specified by the rules I quotes above, this guarantees that Rgb<u8> is indeed laid out as three consecutive bytes in memory, and that Vec<Rgb<u8>> is indeed a consecutive, densely packed buffer of RGB values, so our conversion is safe. We still need to use unsafe code to write it:
let p = img.as_mut_ptr();
let len = img.len() * mem::size_of::<Rgb<u8>>();
let cap = img.capacity() * mem::size_of::<Rgb<u8>>();
mem::forget(img);
let buf: Vec<u8> = unsafe { Vec::from_raw_parts(p as *mut u8, len, cap) };
If you want to protect against the case that there is additional padding at the end of Rgb, you can check whether size_of::<Rgb<u8>>() is indeed 3. If it is, you can use the unsafe non-copying version, otherwise you have to use the first version above.
You choose the Vec<Rgb<u8>> storage format because it's easier to deal with and you want it to grow dynamically. But as you noticed, there's no guarantee of compatibility of its storage with a Vec<u8>, and no safe conversion.
Why not take the problem the other way and build a convenient facade for a Vec<u8> ?
type Rgb = [u8; 3];
#[derive(Debug)]
struct Img(Vec<u8>);
impl Img {
fn new() -> Img {
Img(Vec::new())
}
fn push(&mut self, rgb: &Rgb) {
self.0.push(rgb[0]);
self.0.push(rgb[1]);
self.0.push(rgb[2]);
}
// other convenient methods
}
fn main() {
let mut img = Img::new();
let rgb : Rgb = [1, 2, 3];
img.push(&rgb);
img.push(&rgb);
println!("{:?}", img);
}

What happens if I call Vec::from_raw_parts with a smaller capacity than the pointer actually has?

I have a vector of u8 that I want to interpret as a vector of u32. It is assumed that the bytes are in the right order. I don't want to allocate new memory and copy bytes after casting. I got the following to work:
use std::mem;
fn reinterpret(mut v: Vec<u8>) -> Option<Vec<u32>> {
let v_len = v.len();
v.shrink_to_fit();
if v_len % 4 != 0 {
None
} else {
let v_cap = v.capacity();
let v_ptr = v.as_mut_ptr();
println!("{:?}|{:?}|{:?}", v_len, v_cap, v_ptr);
let v_reinterpret = unsafe { Vec::from_raw_parts(v_ptr as *mut u32, v_len / 4, v_cap / 4) };
println!("{:?}|{:?}|{:?}",
v_reinterpret.len(),
v_reinterpret.capacity(),
v_reinterpret.as_ptr());
println!("{:?}", v_reinterpret);
println!("{:?}", v); // v is still alive, but is same as rebuilt
mem::forget(v);
Some(v_reinterpret)
}
}
fn main() {
let mut v: Vec<u8> = vec![1, 1, 1, 1, 1, 1, 1, 1];
let test = reinterpret(v);
println!("{:?}", test);
}
However, there's an obvious problem here. From the shrink_to_fit documentation:
It will drop down as close as possible to the length but the allocator may still inform the vector that there is space for a few more elements.
Does this mean that my capacity may still not be a multiple of the size of u32 after calling shrink_to_fit? If in from_raw_parts I set capacity to v_len/4 with v.capacity() not an exact multiple of 4, do I leak those 1-3 bytes, or will they go back into the memory pool because of mem::forget on v?
Is there any other problem I am overlooking here?
I think moving v into reinterpret guarantees that it's not accessible from that point on, so there's only one owner from the mem::forget(v) call onwards.
This is an old question, and it looks like it has a working solution in the comments. I've just written up what exactly goes wrong here, and some solutions that one might create/use in today's Rust.
This is undefined behavior
Vec::from_raw_parts is an unsafe function, and thus you must satisfy its invariants, or you invoke undefined behavior.
Quoting from the documentation for Vec::from_raw_parts:
ptr needs to have been previously allocated via String/Vec (at least, it's highly likely to be incorrect if it wasn't).
T needs to have the same size and alignment as what ptr was allocated with. (T having a less strict alignment is not sufficient, the alignment really needs to be equal to satsify the dealloc requirement that memory must be allocated and deallocated with the same layout.)
length needs to be less than or equal to capacity.
capacity needs to be the capacity that the pointer was allocated with.
So, to answer your question, if capacity is not equal to the capacity of the original vec, then you've broken this invariant. This gives you undefined behavior.
Note that the requirement isn't on size_of::<T>() * capacity either, though, which brings us to the next topic.
Is there any other problem I am overlooking here?
Three things.
First, the function as written is disregarding another requirement of from_raw_parts. Specifically, T must have the same size as alignment as the original T. u32 is four times as big as u8, so this again breaks this requirement. Even if capacity*size remains the same, size isn't, and capacity isn't. This function will never be sound as implemented.
Second, even if all of the above was valid, you've also ignored the alignment. u32 must be aligned to 4-byte boundaries, while a Vec<u8> is only guaranteed to be aligned to a 1-byte boundary.
A comment on the OP mentions:
I think on x86_64, misalignment will have performance penalty
It's worth noting that while this may be true of machine language, it is not true for Rust. The rust reference explicitly states "A value of alignment n must only be stored at an address that is a multiple of n." This is a hard requirement.
Why the exact type requirement?
Vec::from_raw_parts seems like it's pretty strict, and that's for a reason. In Rust, the allocator API operates not only on allocation size, but on a Layout, which is the combination of size, number of things, and alignment of individual elements. In C with memalloc, all the allocator can rely upon is that the size is the same, and some minimum alignment. In Rust, though, it's allowed to rely on the entire Layout, and invoke undefined behavior if not.
So in order to correctly deallocate the memory, Vec needs to know the exact type that it was allocated with. By converting a Vec<u32> into Vec<u8>, it no longer knows this information, and so it can no longer properly deallocate this memory.
Alternative - Transforming slices
Vec::from_raw_parts's strictness comes from the fact that it needs to deallocate the memory. If we create a borrowing slice, &[u32] instead, we no longer need to deal with it! There is no capacity when turning a &[u8] into &[u32], so we should be all good, right?
Well, almost. You still have to deal with alignment. Primitives are generally aligned to their size, so a [u8] is only guaranteed to be aligned to 1-byte boundaries, while [u32] must be aligned to a 4-byte boundary.
If you want to chance it, though, and create a [u32] if possible, there's a function for that - <[T]>::align_to:
pub unsafe fn align_to<U>(&self) -> (&[T], &[U], &[T])
This will trim of any starting and ending misaligned values, and then give you a slice in the middle of your new type. It's unsafe, but the only invariant you need to satisfy is that the elements in the middle slice are valid.
It's sound to reinterpret 4 u8 values as a u32 value, so we're good.
Putting it all together, a sound version of the original function would look like this. This operates on borrowed rather than owned values, but given that reinterpreting an owned Vec is instant-undefined-behavior in any case, I think it's safe to say this is the closest sound function:
use std::mem;
fn reinterpret(v: &[u8]) -> Option<&[u32]> {
let (trimmed_front, u32s, trimmed_back) = unsafe { v.align_to::<u32>() };
if trimmed_front.is_empty() && trimmed_back.is_empty() {
Some(u32s)
} else {
// either alignment % 4 != 0 or len % 4 != 0, so we can't do this op
None
}
}
fn main() {
let mut v: Vec<u8> = vec![1, 1, 1, 1, 1, 1, 1, 1];
let test = reinterpret(&v);
println!("{:?}", test);
}
As a note, this could also be done with std::slice::from_raw_parts rather than align_to. However, that requires manually dealing with the alignment, and all it really gives is more things we need to ensure we're doing right. Well, that and compatibility with older compilers - align_to was introduced in 2018 in Rust 1.30.0, and wouldn't have existed when this question was asked.
Alternative - Copying
If you do need a Vec<u32> for long term data storage, I think the best option is to just allocate new memory. The old memory is allocated for u8s anyways, and wouldn't work.
This can be made fairly simple with some functional programming:
fn reinterpret(v: &[u8]) -> Option<Vec<u32>> {
let v_len = v.len();
if v_len % 4 != 0 {
None
} else {
let result = v
.chunks_exact(4)
.map(|chunk: &[u8]| -> u32 {
let chunk: [u8; 4] = chunk.try_into().unwrap();
let value = u32::from_ne_bytes(chunk);
value
})
.collect();
Some(result)
}
}
First, we use <[T]>::chunks_exact to iterate over chunks of 4 u8s. Next, try_into to convert from &[u8] to [u8; 4]. The &[u8] is guaranteed to be length 4, so this never fails.
We use u32::from_ne_bytes to convert the bytes into a u32 using native endianness. If interacting with a network protocol, or on-disk serialization, then using from_be_bytes or from_le_bytes may be preferable. And finally, we collect to turn our result back into a Vec<u32>.
As a last note, a truly general solution might use both of these techniques. If we change the return type to Cow<'_, [u32]>, we could return aligned, borrowed data if it works, and allocate a new array if it doesn't! Not quite the best of both worlds, but close.

Resources