I have a simple asm code which loads 12 quad registers of NEON, and have paralleled pairwise add instruction along with the load instruction ( to exploit the dual issue capability). I have verified the code here:
http://pulsar.webshaker.net/ccc/sample-d3a7fe78
As one can see, the code is taking around 13 cycles. But when I load the code on the board, the load instructions seems to take more than one cycle per load, I verified and found out that the VPADAL is taking 1 cycle as stated, but VLD1 is taking more than one cycle. Why is that?
I have taken care of the following:
The address is 16 byte aligned.
Have provided the alignment hint in the instruction vld1.64 {d0, d1} [r0,:128]!
Tried preload instruction pld [r0, #192], at places but that seems to add to the cycles instead of actually reducing the latency.
Can someone tell me what am I doing wrong, why this latency?
Other Details:
With reference to cortex-a8
arm-2009q1 cross compiler tool chain
coding in assembly
Your code is executing much slower than expected because as it's currently written, it's causing the perfect storm of pipeline stalls. On any modern CPU with a pipelined architecture, instructions can execute in one cycle under ideal conditions. The ideal conditions are that the instruction is not waiting for memory and doesn't have any register dependencies. The way you've written the code, you're not allowing for the delay in reading from memory and making the next instruction dependent on the results of the read. This is causing the worst possible performance. Also, I'm not sure why you're accumulating the pairwise adds into multiple registers. Try something like this:
veor.u16 q12,q12,q12 # clear accumulated sum
top_of_loop:
vld1.u16 {q0,q1},[r0,:128]!
vld1.u16 {q2,q3},[r0,:128]!
vpadal.u16 q12,q0
vpadal.u16 q12,q1
vpadal.u16 q12,q2
vpadal.u16 q12,q3
vld1.u16 {q0,q1},[r0,:128]!
vld1.u16 {q2,q3},[r0,:128]!
vpadal.u16 q12,q0
vpadal.u16 q12,q1
vpadal.u16 q12,q2
vpadal.u16 q12,q3
subs r1,r1,#8
bne top_of_loop
Experiment with different numbers of load instructions before executing the adds. The point is that you need to allow time for the read to occur before you can use the target register.
Note: Using Q4-Q7 is risky because they're non-volatile registers. On Android you will get random garbage appearing in these (especially Q4).
Related
For fun, I'm writing a bignum library in Rust. My goal (as with most bignum libraries) is to make it as efficient as I can. I'd like it to be efficient even on unusual architectures.
It seems intuitive to me that a CPU will perform arithmetic faster on integers with the native number of bits for the architecture (i.e., u64 for 64-bit machines, u16 for 16-bit machines, etc.) As such, since I want to create a library that is efficient on all architectures, I need to take the target architecture's native integer size into account. The obvious way to do this would be to use the cfg attribute target_pointer_width. For instance, to define the smallest type which will always be able to hold more than the maximum native int size:
#[cfg(target_pointer_width = "16")]
type LargeInt = u32;
#[cfg(target_pointer_width = "32")]
type LargeInt = u64;
#[cfg(target_pointer_width = "64")]
type LargeInt = u128;
However, while looking into this, I came across this comment. It gives an example of an architecture where the native int size is different from the pointer width. Thus, my solution will not work for all architectures. Another potential solution would be to write a build script which codegens a small module which defines LargeInt based on the size of a usize (which we can acquire like so: std::mem::size_of::<usize>().) However, this has the same problem as above, since usize is based on the pointer width as well. A final obvious solution is to simply keep a map of native int sizes for each architecture. However, this solution is inelegant and doesn't scale well, so I'd like to avoid it.
So, my questions: is there a way to find the target's native int size, preferably before compilation, in order to reduce runtime overhead? Is this effort even worth it? That is, is there likely to be a significant difference between using the native int size as opposed to the pointer width?
It's generally hard (or impossible) to get compilers to emit optimal code for BigNum stuff, that's why https://gmplib.org/ has its low level primitive functions (mpn_... docs) hand-written in assembly for various target architectures with tuning for different micro-architecture, e.g. https://gmplib.org/repo/gmp/file/tip/mpn/x86_64/core2/mul_basecase.asm for the general case of multi-limb * multi-limb numbers. And https://gmplib.org/repo/gmp/file/tip/mpn/x86_64/coreisbr/aors_n.asm for mpn_add_n and mpn_sub_n (Add OR Sub = aors), tuned for SandyBridge-family which doesn't have partial-flag stalls so it can loop with dec/jnz.
Understanding what kind of asm is optimal may be helpful when writing code in a higher level language. Although in practice you can't even get close to that so it sometimes makes sense to use a different technique, like only using values up to 2^30 in 32-bit integers (like CPython does internally, getting the carry-out via a right shift, see the section about Python in this). In Rust you do have access to add_overflow to get the carry-out, but using it is still hard.
For practical use, writing Rust bindings for GMP is probably your best bet, unless that already exists.
Using the largest chunks possible is very good; on all current CPUs, add reg64, reg64 has the same throughput and latency as add reg32, reg32 or reg8. So you get twice as much work done per unit. And carry propagation through 64 bits of result in 1 cycle of latency.
(There are alternate ways to store BigInteger data that can make SIMD useful; #Mysticial explains in Can long integer routines benefit from SSE?. e.g. 30 value bits per 32-bit int, allowing you to defer normalization until after a few addition steps. But every use of such numbers has to be aware of these issues so it's not an easy drop-in replacement.)
In Rust, you probably want to just use u64 regardless of the target, unless you really care about small-number (single-limb) performance on 32-bit targets. Let the compiler build u64 operations for you out of add / adc (add with carry).
The only thing that might need to be ISA-specific is if u128 is not available on some targets. You want to use 64 * 64 => 128-bit full multiply as your building block for multiplication; if the compiler can do that for you with u128 then that's great, especially if it inlines efficiently.
See also discussion in comments under the question.
One stumbling block for getting compilers to emit efficient BigInt addition loops (even inside the body of one unrolled loop) is writing an add that takes a carry input and produces a carry output. Note that x += 0xff..ff + carry=1 needs to produce a carry out even though 0xff..ff + 1 wraps to zero. So in C or Rust, x += y + carry has to check for carry out in both the y+carry and the x+= parts.
It's really hard (probably impossible) to convince compiler back-ends like LLVM to emit a chain of adc instructions. An add/adc is doable when you don't need the carry-out from adc. Or probably if the compiler is doing it for you for u128.overflowing_add
Often compilers will turn the carry flag into a 0 / 1 in a register instead of using adc. You can hopefully avoid that for at least pairs of u64 in addition by combining the input u64 values to u128 for u128.overflowing_add. That will hopefully not cost any asm instructions because a u128 already has to be stored across two separate 64-bit registers, just like two separate u64 values.
So combining up to u128 could just be a local optimization for a function that adds arrays of u64 elements, to get the compiler to suck less.
In my library ibig what I do is:
Select architecture-specific size based on target_arch.
If I don't have a value for an architecture, select 16, 32 or 64 based on target_pointer_width.
If target_pointer_width is not one of these values, use 64.
A C++ standard library implements std::copy with the following code (ignoring all sorts of wrappers and concept checks etc) with the simple loop:
for (; __first != __last; ++__result, ++__first)
*__result = *__first;
Now, suppose I want a general-purpose std::copy-like function for warps (not blocks; not grids) to use for collaboratively copying data from one place to another. Let's even assume for simplicity that the function takes pointers rather than an arbitrary iterator.
Of course, writing general-purpose code in CUDA is often a useless pursuit - since we might be sacrificing a lot of the benefit of using a GPU in the first place in favor of generality - so I'll allow myself some boolean/enum template parameters to possibly select between frequently-occurring cases, avoiding runtime checks. So the signature might be, say:
template <typename T, bool SomeOption, my_enum_t AnotherOption>
T* copy(
T* __restrict__ destination,
const T* __restrict__ source,
size_t length
);
but for each of these cases I'm aiming for optimal performance (or optimal expected performance given that we don't know what other warps are doing).
Which factors should I take into consideration when writing such a function? Or in other words: Which cases should I distinguish between in implementing this function?
Notes:
This should target Compute Capabilities 3.0 or better (i.e. Kepler or newer micro-architectures)
I don't want to make a Runtime API memcpy() call. At least, I don't think I do.
Factors I believe should be taken into consideration:
Coalescing memory writes - ensuring that consecutive lanes in a warp write to consecutive memory locations (no gaps).
Type size vs Memory transaction size I - if sizeof(T) is sizeof(T) is 1 or 2, and we have have each lane write a single element, the entire warp would write less than 128B, wasting some of the memory transaction. Instead, we should have each thread place 2 or 4 input elements in a register, and write that
Type size vs Memory transaction size II - For type sizes such that lcm(4, sizeof(T)) > 4, it's not quite clear what to do. How well does the compiler/the GPU handle writes when each lane writes more than 4 bytes? I wonder.
Slack due to the reading of multiple elements at a time - If each thread wishes to read 2 or 4 elements for each write, and write 4-byte integers - we might have 1 or 2 elements at the beginning and the end of the input which must be handled separately.
Slack due to input address mis-alignment - The input is read in 32B transactions (under reasonable assumptions); we thus have to handle the first elements up to the multiple of 32B, and the last elements (after the last such multiple,) differently.
Slack due to output address mis-alignment - The output is written in transactions of upto 128B (or is it just 32B?); we thus have to handle the first elements up to the multiple of this number, and the last elements (after the last such multiple,) differently.
Whether or not T is trivially-copy-constructible. But let's assume that it is.
But it could be that I'm missing some considerations, or that some of the above are redundant.
Factors I've been wondering about:
The block size (i.e. how many other warps are there)
The compute capability (given that it's at least 3)
Whether the source/target is in shared memory / constant memory
Choice of caching mode
I am reading sensor output as square wave(0-5 volt) via oscilloscope. Now I want to measure frequency of one period with Beaglebone. So I should measure the time between two rising edges. However, I don't have any experience with working Beaglebone. Can you give some advices or sample codes about measuring time between rising edges?
How deterministic do you need this to be? If you can tolerate some inaccuracy, you can probably do it on the main Linux OS; if you want to be fancy pants, this seems like a potential use case for the BBB's PRU's (which I unfortunately haven't used so take this with substantial amounts of salt). I would expect you'd be able to write PRU code that just sits with an infinite outerloop and then inside that loop, start looping until it sees the pin shows 0, then starts looping until the pin shows 1 (this is the first rising edge), then starts counting until either the pin shows 0 again (this would then be the falling edge) or another loop to the next rising edge... either way, you could take the counter value and you should be able to directly convert that into time (the PRU is states as having fixed frequency for each instruction, and is a 200Mhz (50ns/instruction). Assuming your loop is something like
#starting with pin low
inner loop 1:
registerX = loadPin
increment counter
jump if zero registerX to inner loop 1
# pin is now high
inner loop 2:
registerX = loadPin
increment counter
jump if one registerX to inner loop 2
# pin is now low again
That should take 3 instructions per counter increment, so you can get the time as 3 * counter * 50 ns.
As suggested by Foon in his answer, the PRUs are a good fit for this task (although depending on your requirements it may be fine to use the ARM processor and standard GPIO). Please note that (as far as I know) both the regular GPIOs and the PRU inputs are based on 3.3V logic, and connecting a 5V signal might fry your board! You will need an additional component or circuit to convert from 5V to 3.3V.
I've written a basic example that measures timing between rising edges on the header pin P8.15 for my own purpose of measuring an engine's rpm. If you decide to use it, you should check the timing results against a known reference. It's about right but I haven't checked it carefully at all. It is implemented using PRU assembly and uses the pypruss python module to simplify interfacing.
Since many of the Project Euler problems require you to do a divisibility check for quite a number of times, I've been trying to figure out the fastest way to perform this task in ZX81 BASIC.
So far I've compared (N/D) to INT(N/D) to check, whether N is dividable by D or not.
I have been thinking about doing the test in Z80 machine code, I haven't yet figured out how to use the variables in the BASIC in the machine code.
How can it be achieved?
You can do this very fast in machine code by subtracting repeatedly. Basically you have a procedure like:
set accumulator to N
subtract D
if carry flag is set then it is not divisible
if zero flag is set then it is divisible
otherwise repeat subtraction until one of the above occurs
The 8 bit version would be something like:
DIVISIBLE_TEST:
LD B,10
LD A,100
DIVISIBLE_TEST_LOOP:
SUB B
JR C, $END_DIVISIBLE_TEST
JR Z, $END_DIVISIBLE_TEST
JR $DIVISIBLE_TEST_LOOP
END_DIVISIBLE_TEST:
LD B,A
LD C,0
RET
Now, you can call from basic using USR. What USR returns is whatever's in the BC register pair, so you would probably want to do something like:
REM poke the memory addresses with the operands to load the registers
POKE X+1, D
POKE X+3, N
LET r = USR X
IF r = 0 THEN GOTO isdivisible
IF r <> 0 THEN GOTO isnotdivisible
This is an introduction I wrote to Z80 which should help you figure this out. This will explain the flags if you're not familiar with them.
There's a load more links to good Z80 stuff from the main site although it is Spectrum rather than ZX81 focused.
A 16 bit version would be quite similar but using register pair operations. If you need to go beyond 16 bits it would get a bit more convoluted.
How you load this is up to you - but the traditional method is using DATA statements and POKEs. You may prefer to have an assembler figure out the machine code for you though!
Your existing solution may be good enough. Only replace it with something faster if you find it to be a bottleneck in profiling.
(Said with a straight face, of course.)
And anyway, on the ZX81 you can just switch to FAST mode.
Don't know if RANDOMIZE USR is available in ZX81 but I think it can be used to call routines in assembly. To pass arguments you might need to use POKE to set some fixed memory locations before executing RANDOMIZE USR.
I remember to find a list of routines implemented in the ROM to support the ZX Basic. I'm sure there are a few to perform floating operation.
An alternative to floating point is to use fixed point math. It's a lot faster in these kind of situations where there is no math coprocessor.
You also might find more information in Sinclair User issues. They published some articles related to programming in the ZX Spectrum
You should place the values in some pre-known memory locations, first. Then use the same locations from within Z80 assembler. There is no parameter passing between the two.
This is based on what I (still) remember of ZX Spectrum 48. Good luck, but you might consider upgrading your hw. ;/
The problem with Z80 machine code is that it has no floating point ops (and no integer divide or multiply, for that matter). Implementing your own FP library in Z80 assembler is not trivial. Of course, you can use the built-in BASIC routines, but then you may as well just stick with BASIC.
I'm not talking about algorithmic stuff (eg use quicksort instead of bubblesort), and I'm not talking about simple things like loop unrolling.
I'm talking about the hardcore stuff. Like Tiny Teensy ELF, The Story of Mel; practically everything in the demoscene, and so on.
I once wrote a brute force RC5 key search that processed two keys at a time, the first key used the integer pipeline, the second key used the SSE pipelines and the two were interleaved at the instruction level. This was then coupled with a supervisor program that ran an instance of the code on each core in the system. In total, the code ran about 25 times faster than a naive C version.
In one (here unnamed) video game engine I worked with, they had rewritten the model-export tool (the thing that turns a Maya mesh into something the game loads) so that instead of just emitting data, it would actually emit the exact stream of microinstructions that would be necessary to render that particular model. It used a genetic algorithm to find the one that would run in the minimum number of cycles. That is to say, the data format for a given model was actually a perfectly-optimized subroutine for rendering just that model. So, drawing a mesh to the screen meant loading it into memory and branching into it.
(This wasn't for a PC, but for a console that had a vector unit separate and parallel to the CPU.)
In the early days of DOS when we used floppy discs for all data transport there were viruses as well. One common way for viruses to infect different computers was to copy a virus bootloader into the bootsector of an inserted floppydisc. When the user inserted the floppydisc into another computer and rebooted without remembering to remove the floppy, the virus was run and infected the harddrive bootsector, thus permanently infecting the host PC. A particulary annoying virus I was infected by was called "Form", to battle this I wrote a custom floppy bootsector that had the following features:
Validate the bootsector of the host harddrive and make sure it was not infected.
Validate the floppy bootsector and
make sure that it was not infected.
Code to remove the virus from the
harddrive if it was infected.
Code to duplicate the antivirus
bootsector to another floppy if a
special key was pressed.
Code to boot the harddrive if all was
well, and no infections was found.
This was done in the program space of a bootsector, about 440 bytes :)
The biggest problem for my mates was the very cryptic messages displayed because I needed all the space for code. It was like "FFVD RM?", which meant "FindForm Virus Detected, Remove?"
I was quite happy with that piece of code. The optimization was program size, not speed. Two quite different optimizations in assembly.
My favorite is the floating point inverse square root via integer operations. This is a cool little hack on how floating point values are stored and can execute faster (even doing a 1/result is faster than the stock-standard square root function) or produce more accurate results than the standard methods.
In c/c++ the code is: (sourced from Wikipedia)
float InvSqrt (float x)
{
float xhalf = 0.5f*x;
int i = *(int*)&x;
i = 0x5f3759df - (i>>1); // Now this is what you call a real magic number
x = *(float*)&i;
x = x*(1.5f - xhalf*x*x);
return x;
}
A Very Biological Optimisation
Quick background: Triplets of DNA nucleotides (A, C, G and T) encode amino acids, which are joined into proteins, which are what make up most of most living things.
Ordinarily, each different protein requires a separate sequence of DNA triplets (its "gene") to encode its amino acids -- so e.g. 3 proteins of lengths 30, 40, and 50 would require 90 + 120 + 150 = 360 nucleotides in total. However, in viruses, space is at a premium -- so some viruses overlap the DNA sequences for different genes, using the fact that there are 6 possible "reading frames" to use for DNA-to-protein translation (namely starting from a position that is divisible by 3; from a position that divides 3 with remainder 1; or from a position that divides 3 with remainder 2; and the same again, but reading the sequence in reverse.)
For comparison: Try writing an x86 assembly language program where the 300-byte function doFoo() begins at offset 0x1000... and another 200-byte function doBar() starts at offset 0x1001! (I propose a name for this competition: Are you smarter than Hepatitis B?)
That's hardcore space optimisation!
UPDATE: Links to further info:
Reading Frames on Wikipedia suggests Hepatitis B and "Barley Yellow Dwarf" virus (a plant virus) both overlap reading frames.
Hepatitis B genome info on Wikipedia. Seems that different reading-frame subunits produce different variations of a surface protein.
Or you could google for "overlapping reading frames"
Seems this can even happen in mammals! Extensively overlapping reading frames in a second mammalian gene is a 2001 scientific paper by Marilyn Kozak that talks about a "second" gene in rat with "extensive overlapping reading frames". (This is quite surprising as mammals have a genome structure that provides ample room for separate genes for separate proteins.) Haven't read beyond the abstract myself.
I wrote a tile-based game engine for the Apple IIgs in 65816 assembly language a few years ago. This was a fairly slow machine and programming "on the metal" is a virtual requirement for coaxing out acceptable performance.
In order to quickly update the graphics screen one has to map the stack to the screen in order to use some special instructions that allow one to update 4 screen pixels in only 5 machine cycles. This is nothing particularly fantastic and is described in detail in IIgs Tech Note #70. The hard-core bit was how I had to organize the code to make it flexible enough to be a general-purpose library while still maintaining maximum speed.
I decomposed the graphics screen into scan lines and created a 246 byte code buffer to insert the specialized 65816 opcodes. The 246 bytes are needed because each scan line of the graphics screen is 80 words wide and 1 additional word is required on each end for smooth scrolling. The Push Effective Address (PEA) instruction takes up 3 bytes, so 3 * (80 + 1 + 1) = 246 bytes.
The graphics screen is rendered by jumping to an address within the 246 byte code buffer that corresponds to the right edge of the screen and patching in a BRanch Always (BRA) instruction into the code at the word immediately following the left-most word. The BRA instruction takes a signed 8-bit offset as its argument, so it just barely has the range to jump out of the code buffer.
Even this isn't too terribly difficult, but the real hard-core optimization comes in here. My graphics engine actually supported two independent background layers and animated tiles by using different 3-byte code sequences depending on the mode:
Background 1 uses a Push Effective Address (PEA) instruction
Background 2 uses a Load Indirect Indexed (LDA ($00),y) instruction followed by a push (PHA)
Animated tiles use a Load Direct Page Indexed (LDA $00,x) instruction followed by a push (PHA)
The critical restriction is that both of the 65816 registers (X and Y) are used to reference data and cannot be modified. Further the direct page register (D) is set based on the origin of the second background and cannot be changed; the data bank register is set to the data bank that holds pixel data for the second background and cannot be changed; the stack pointer (S) is mapped to graphics screen, so there is no possibility of jumping to a subroutine and returning.
Given these restrictions, I had the need to quickly handle cases where a word that is about to be pushed onto the stack is mixed, i.e. half comes from Background 1 and half from Background 2. My solution was to trade memory for speed. Because all of the normal registers were in use, I only had the Program Counter (PC) register to work with. My solution was the following:
Define a code fragment to do the blend in the same 64K program bank as the code buffer
Create a copy of this code for each of the 82 words
There is a 1-1 correspondence, so the return from the code fragment can be a hard-coded address
Done! We have a hard-coded subroutine that does not affect the CPU registers.
Here is the actual code fragments
code_buff: PEA $0000 ; rightmost word (16-bits = 4 pixels)
PEA $0000 ; background 1
PEA $0000 ; background 1
PEA $0000 ; background 1
LDA (72),y ; background 2
PHA
LDA (70),y ; background 2
PHA
JMP word_68 ; mix the data
word_68_rtn: PEA $0000 ; more background 1
...
PEA $0000
BRA *+40 ; patched exit code
...
word_68: LDA (68),y ; load data for background 2
AND #$00FF ; mask
ORA #$AB00 ; blend with data from background 1
PHA
JMP word_68_rtn ; jump back
word_66: LDA (66),y
...
The end result was a near-optimal blitter that has minimal overhead and cranks out more than 15 frames per second at 320x200 on a 2.5 MHz CPU with a 1 MB/s memory bus.
Michael Abrash's "Zen of Assembly Language" had some nifty stuff, though I admit I don't recall specifics off the top of my head.
Actually it seems like everything Abrash wrote had some nifty optimization stuff in it.
The Stalin Scheme compiler is pretty crazy in that aspect.
I once saw a switch statement with a lot of empty cases, a comment at the head of the switch said something along the lines of:
Added case statements that are never hit because the compiler only turns the switch into a jump-table if there are more than N cases
I forget what N was. This was in the source code for Windows that was leaked in 2004.
I've gone to the Intel (or AMD) architecture references to see what instructions there are. movsx - move with sign extension is awesome for moving little signed values into big spaces, for example, in one instruction.
Likewise, if you know you only use 16-bit values, but you can access all of EAX, EBX, ECX, EDX , etc- then you have 8 very fast locations for values - just rotate the registers by 16 bits to access the other values.
The EFF DES cracker, which used custom-built hardware to generate candidate keys (the hardware they made could prove a key isn't the solution, but could not prove a key was the solution) which were then tested with a more conventional code.
The FSG 2.0 packer made by a Polish team, specifically made for packing executables made with assembly. If packing assembly isn't impressive enough (what's supposed to be almost as low as possible) the loader it comes with is 158 bytes and fully functional. If you try packing any assembly made .exe with something like UPX, it will throw a NotCompressableException at you ;)