Optimal implementation of iterative Kahan summation - c++

Intro
Kahan summation / compensated summation is technique that addresses compilers´ inability to respect the associative property of numbers. Truncation errors results in (a+b)+c not being exactly equal to a+(b+c) and thus accumulate an undesired relative error on longer series of sums, which is a common obstacle in scientific computing.
Task
I desire the optimal implementation of Kahan summation. I suspect that the best performance may be achieved with handcrafted assembly code.
Attempts
The code below calculates the sum of 1000 random numbers in range [0,1] with three approaches.
Standard summation: Naive implementation which accumulates a root mean square relative error that grows as O(sqrt(N))
Kahan summation [g++]: Compensated summation using the c/c++ function "csum". Explanation in comments. Note that some compilers may have default flags that invalidate this implementation (see output below).
Kahan summation [asm]: Compensated summation implemented as "csumasm" using the same algorithm as "csum". Cryptic explanation in comments.
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
extern "C" void csumasm(double&, double, double&);
__asm__(
"csumasm:\n"
"movsd (%rcx), %xmm0\n" //xmm0 = a
"subsd (%r8), %xmm1\n" //xmm1 - r8 (c) | y = b-c
"movapd %xmm0, %xmm2\n"
"addsd %xmm1, %xmm2\n" //xmm2 + xmm1 (y) | b = a+y
"movapd %xmm2, %xmm3\n"
"subsd %xmm0, %xmm3\n" //xmm3 - xmm0 (a) | b - a
"movapd %xmm3, %xmm0\n"
"subsd %xmm1, %xmm0\n" //xmm0 - xmm1 (y) | - y
"movsd %xmm0, (%r8)\n" //xmm0 to c
"movsd %xmm2, (%rcx)\n" //b to a
"ret\n"
);
void csum(double &a,double b,double &c) { //this function adds a and b, and passes c as a compensation term
double y = b-c; //y is the correction of b argument
b = a+y; //add corrected b argument to a argument. The output of the current summation
c = (b-a)-y; //find new error to be passed as a compensation term
a = b;
}
double fun(double fMin, double fMax){
double f = (double)rand()/RAND_MAX;
return fMin + f*(fMax - fMin); //returns random value
}
int main(int argc, char** argv) {
int N = 1000;
srand(0); //use 0 seed for each method
double sum1 = 0;
for (int n = 0; n < N; ++n)
sum1 += fun(0,1);
srand(0);
double sum2 = 0;
double c = 0; //compensation term
for (int n = 0; n < N; ++n)
csum(sum2,fun(0,1),c);
srand(0);
double sum3 = 0;
c = 0;
for (int n = 0; n < N; ++n)
csumasm(sum3,fun(0,1),c);
printf("Standard summation:\n %.16e (error: %.16e)\n\n",sum1,sum1-sum3);
printf("Kahan compensated summation [g++]:\n %.16e (error: %.16e)\n\n",sum2,sum2-sum3);
printf("Kahan compensated summation [asm]:\n %.16e\n",sum3);
return 0;
}
The output with -O3 is:
Standard summation:
5.1991955320902093e+002 (error: -3.4106051316484809e-013)
Kahan compensated summation [g++]:
5.1991955320902127e+002 (error: 0.0000000000000000e+000)
Kahan compensated summation [asm]:
5.1991955320902127e+002
The output with -O3 -ffast-math
Standard summation:
5.1991955320902093e+002 (error: -3.4106051316484809e-013)
Kahan compensated summation [g++]:
5.1991955320902093e+002 (error: -3.4106051316484809e-013)
Kahan compensated summation [asm]:
5.1991955320902127e+002
It is clear that -ffast-math destroys the Kahan summation arithmetic, which is unfortunate because my program requires the use of -ffast-math.
Question
Is it possible to construct a better/faster asm x64 code for Kahan's compensated summation? Perhaps there is a clever way to skip some of the movapd instructions?
If no better asm codes are possible, is there a c++ way to implement Kahan summation that can be used with -ffast-math without devolving to the naive summation? Perhaps a c++ implementation is generally more flexible for the compiler to optimize.
Ideas or suggestions are appreciated.
Further information
The contents of "fun" cannot be inlined, but the "csum" function could be.
The sum must be calculated as an iterative process (the corrected term must be applied on every single addition). This is because the intended summation function takes an input that depends on the previous sum.
The intended summation function is called indefinitely and several hundred million times per second, which motives the pursuit of a high performance low-level implementation.
Higher precision arithmetic such as long double, float128 or arbitrary precision libraries are not to be considered as higher precision solutions due to performance reasons.
Edit: Inlined csum (doesn't make much sense without the full code, but just for reference)
subsd xmm0, QWORD PTR [rsp+32]
movapd xmm1, xmm3
addsd xmm3, xmm0
movsd QWORD PTR [rsp+16], xmm3
subsd xmm3, xmm1
movapd xmm1, xmm3
subsd xmm1, xmm0
movsd QWORD PTR [rsp+32], xmm1

You can put functions that need to not use -ffast-math (like a csum loop) in a separate file that gets compiled without -ffast-math.
Possibly you could also use __attribute__((optimize("no-fast-math"))), but https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html says that optimization-level pragmas and attributes aren't "suitable in production code", unfortunately.
update: apparently part of the question was based on a misunderstanding that -O3 wasn't safe, or something? It is; ISO C++ specifies FP math rules that are like GCC's -fno-fast-math. Compiling everything with just -O3 apparently makes the OP's code run quickly and safely. See the bottom of this answer for workarounds like OpenMP to get some of the benefit of fast-math for some parts of your code without actually enabling -ffast-math.
ICC defaults to fast-path so you have to specifically enable FP=strict for it to be safe with -O3, but gcc/clang default to fully strict FP regardless of other optimization settings. (except -Ofast = -O3 -ffast-math)
You should be able to vectorize Kahan summation by keeping a vector (or four) of totals and an equal number of vectors of compensations. You can do that with intrinsics (as long as you don't enable fast-math for that file).
e.g. use SSE2 __m128d for 2 packed additions per instruction. Or AVX __m256d. On modern x86, addpd / subpd have the same performance as addsd and subsd (1 uop, 3 to 5 cycle latency depending on microarchitecture: https://agner.org/optimize/).
So you're effectively doing 8 compensated summations in parallel, each sum getting every 8th input element.
Generating random numbers on the fly with your fun() is significantly slower than reading them from memory. If your normal use-case has data in memory, you should be benchmarking that. Otherwise I guess scalar is interesting.
If you're going to use inline asm, it would be much better to actually use it inline so you can get multiple inputs and multiple outputs in XMM registers with Extended asm, not stored/reloaded through memory.
Defining a stand-alone function that actually takes args by reference looks pretty performance-defeating. (Especially when it doesn't even return either of them as a return value to avoid one of the store/reload chains). Even just making a function call introduces a lot of overhead by clobbering many registers. (Not as bad in Windows x64 as in x86-64 System V where all the XMM regs are call-clobbered, and more of the integer regs.)
Also your stand-alone function is specific to the Windows x64 calling convention so it's less portable than inline asm inside a function would be.
And BTW, clang managed to implement csum(double&, double, double&): with only two movapd instructions, instead of the 3 in your asm (which I assume you copied from GCC's asm output). https://godbolt.org/z/lw6tug. If you can assume AVX is available, you can avoid any.
And BTW, movaps is 1 byte smaller and should be used instead. No CPUs have had separate data domains / forwarding networks for double vs. float, just vec-FP vs. vec-int (vs. GP integer)
But really by far your bet is to get GCC to compile a file or function without -ffast-math. https://gcc.gnu.org/wiki/DontUseInlineAsm. That lets the compiler avoid the movaps instructions when AVX is available, besides letting it optimize better when unrolling.
If you're willing to accept the overhead of a function-call for every element, you might as well let the compiler generate that asm by putting csum in a separate file. (Hopefully link-time optimization respects -fno-fast-math for one file, perhaps by not inlining that function.)
But it would be much better to disable fast-math for the whole function containing the summation loop by putting it in a separate file. You may be stuck choosing where non-inline function-call boundaries need to be, based on compiling some code with fast-math and others without.
Ideally compile all of your code with -O3 -march=native, and profile-guided optimization. Also -flto link-time optimization to enable cross-file inlining.
It's not surprising that -ffast-math breaks Kahan summation: treating FP math as associative is one of the main reasons to use fast-math. If you need other parts of -ffast-math like -fno-math-errno and -fno-trapping-math so math functions can inline better, then enable those manually. Those are basically always safe and a good idea; nobody checks errno after calling sqrt so that requirement to set errno for some inputs is just a terrible misdesign of C that burdens implementations unnecessarily. GCC's -ftrapping-math is on by default even though it's broken (it doesn't always exactly reproduce the number of FP exceptions you'd get if you unmasked any) so it should really be off by default. Turning it off doesn't enable any optimizations that would break NaN propagation, it only tells GCC that the number of exceptions isn't a visible side-effect.
Or maybe try -ffast-math -fno-associative-math for your Kahan summation file, but that's the main one that's needed to auto-vectorize FP loops that involve reductions, and helps in other cases. But still, there are several other valuable optimizations that you'd still get.
Another way to get optimizations that normally require fast-math is #pragma omp simd to enable auto-vectorization with OpenMP even in files compiled without auto-vectorization. You can declare an accumulator variable for a reduction to let gcc reorder operations on it as if they were associative.

Related

Is there anything special about -1 (0xFFFFFFFF) regarding ADC?

In a research project of mine I'm writing C++ code. However, the generated assembly is one of the crucial points of the project. C++ doesn't provide direct access to flag manipulating instructions, in particular, to ADC but this shouldn't be a problem provided the compiler is smart enough to use it. Consider:
constexpr unsigned X = 0;
unsigned f1(unsigned a, unsigned b) {
b += a;
unsigned c = b < a;
return c + b + X;
}
Variable c is a workaround to get my hands on the carry flag and add it to b and X. It looks I got luck and the (g++ -O3, version 9.1) generated code is this:
f1(unsigned int, unsigned int):
add %edi,%esi
mov %esi,%eax
adc $0x0,%eax
retq
For all values of X that I've tested the code is as above (except, of course for the immediate value $0x0 that changes accordingly). I found one exception though: when X == -1 (or 0xFFFFFFFFu or ~0u, ... it really doesn't matter how you spell it) the generated code is:
f1(unsigned int, unsigned int):
xor %eax,%eax
add %edi,%esi
setb %al
lea -0x1(%rsi,%rax,1),%eax
retq
This seems less efficient than the initial code as suggested by indirect measurements (not very scientific though) Am I right? If so, is this a "missing optimization opportunity" kind of bug that is worth reporting?
For what is worth, clang -O3, version 8.8.0, always uses ADC (as I wanted) and icc -O3, version 19.0.1 never does.
I've tried using the intrinsic _addcarry_u32 but it didn't help.
unsigned f2(unsigned a, unsigned b) {
b += a;
unsigned char c = b < a;
_addcarry_u32(c, b, X, &b);
return b;
}
I reckon I might not be using _addcarry_u32 correctly (I couldn't find much info on it). What's the point of using it since it's up to me to provide the carry flag? (Again, introducing c and praying for the compiler to understand the situation.)
I might, actually, be using it correctly. For X == 0 I'm happy:
f2(unsigned int, unsigned int):
add %esi,%edi
mov %edi,%eax
adc $0x0,%eax
retq
For X == -1 I'm unhappy :-(
f2(unsigned int, unsigned int):
add %esi,%edi
mov $0xffffffff,%eax
setb %dl
add $0xff,%dl
adc %edi,%eax
retq
I do get the ADC but this is clearly not the most efficient code. (What's dl doing there? Two instructions to read the carry flag and restore it? Really? I hope I'm very wrong!)
mov + adc $-1, %eax is more efficient than xor-zero + setc + 3-component lea for both latency and uop count on most CPUs, and no worse on any still-relevant CPUs.1
This looks like a gcc missed optimization: it probably sees a special case and latches onto that, shooting itself in the foot and preventing the adc pattern recognition from happening.
I don't know what exactly it saw / was looking for, so yes you should report this as a missed-optimization bug. Or if you want to dig deeper yourself, you could look at the GIMPLE or RTL output after optimization passes and see what happens. If you know anything about GCC's internal representations. Godbolt has a GIMPLE tree-dump window you can add from the same dropdown as "clone compiler".
The fact that clang compiles it with adc proves that it's legal, i.e. that the asm you want does match the C++ source, and you didn't miss some special case that's stopping the compiler from doing that optimization. (Assuming clang is bug-free, which is the case here.)
That problem can certainly happen if you're not careful, e.g. trying to write a general-case adc function that takes carry in and provides carry-out from the 3-input addition is hard in C, because either of the two additions can carry so you can't just use the sum < a+b idiom after adding the carry to one of the inputs. I'm not sure it's possible to get gcc or clang to emit add/adc/adc where the middle adc has to take carry-in and produce carry-out.
e.g. 0xff...ff + 1 wraps around to 0, so sum = a+b+carry_in / carry_out = sum < a can't optimize to an adc because it needs to ignore carry in the special case where a = -1 and carry_in = 1.
So another guess is that maybe gcc considered doing the + X earlier, and shot itself in the foot because of that special case. That doesn't make a lot of sense, though.
What's the point of using it since it's up to me to provide the carry flag?
You're using _addcarry_u32 correctly.
The point of its existence is to let you express an add with carry in as well as carry out, which is hard in pure C. GCC and clang don't optimize it well, often not just keeping the carry result in CF.
If you only want carry-out, you can provide a 0 as the carry in and it will optimize to add instead of adc, but still give you the carry-out as a C variable.
e.g. to add two 128-bit integers in 32-bit chunks, you can do this
// bad on x86-64 because it doesn't optimize the same as 2x _addcary_u64
// even though __restrict guarantees non-overlap.
void adc_128bit(unsigned *__restrict dst, const unsigned *__restrict src)
{
unsigned char carry;
carry = _addcarry_u32(0, dst[0], src[0], &dst[0]);
carry = _addcarry_u32(carry, dst[1], src[1], &dst[1]);
carry = _addcarry_u32(carry, dst[2], src[2], &dst[2]);
carry = _addcarry_u32(carry, dst[3], src[3], &dst[3]);
}
(On Godbolt with GCC/clang/ICC)
That's very inefficient vs. unsigned __int128 where compilers would just use 64-bit add/adc, but does get clang and ICC to emit a chain of add/adc/adc/adc. GCC makes a mess, using setcc to store CF to an integer for some of the steps, then add dl, -1 to put it back into CF for an adc.
GCC unfortunately sucks at extended-precision / biginteger written in pure C. Clang sometimes does slightly better, but most compilers are bad at it. This is why the lowest-level gmplib functions are hand-written in asm for most architectures.
Footnote 1: or for uop count: equal on Intel Haswell and earlier where adc is 2 uops, except with a zero immediate where Sandybridge-family's decoders special case that as 1 uop.
But the 3-component LEA with a base + index + disp makes it a 3-cycle latency instruction on Intel CPUs, so it's definitely worse.
On Intel Broadwell and later, adc is a 1-uop instruction even with a non-zero immediate, taking advantage of support for 3-input uops introduced with Haswell for FMA.
So equal total uop count but worse latency means that adc would still be a better choice.
https://agner.org/optimize/

How to let GCC compiler turn variable-division into mul(if faster)

int a, b;
scanf("%d %d", &a, &b);
printf("%d\n", (unsigned int)a/(unsigned char)b);
When compiling, I got
...
::00401C1E:: C70424 24304000 MOV DWORD PTR [ESP],403024 %d %d
::00401C25:: E8 36FFFFFF CALL 00401B60 scanf
::00401C2A:: 0FB64C24 1C MOVZX ECX,BYTE PTR [ESP+1C]
::00401C2F:: 8B4424 18 MOV EAX,[ESP+18]
::00401C33:: 31D2 XOR EDX,EDX
::00401C35:: F7F1 DIV ECX
::00401C37:: 894424 04 MOV [ESP+4],EAX
::00401C3B:: C70424 2A304000 MOV DWORD PTR [ESP],40302A %d\x0A
::00401C42:: E8 21FFFFFF CALL 00401B68 printf
Will it be faster if the DIV turn into MUL and use an array to store the mulvalue? If so, how to let the compiler do the optimization?
int main() {
uint a, s=0, i, t;
scanf("%d", &a);
diviuint aa = a;
t = clock();
for (i=0; i<1000000000; i++)
s += i/a;
printf("Result:%10u\n", s);
printf("Time:%12u\n", clock()-t);
return 0;
}
where diviuint(a) make a memory of 1/a and use multiple instead
Using s+=i/aa makes the speed 2 times of s+=i/a
You are correct that finding the multiplicative inverse may be worth it if integer division inside a loop is unavoidable. gcc and clang won't do this for you with run-time constants, though; only compile-time constants. It's too expensive (in code-size) for the compiler to do without being sure it's needed, and the perf gains aren't as big with non compile-time constants. (I'm not confident a speedup will always be possible, depending on how good integer division is on the target microarchitecture.)
Using a multiplicative inverse
If you can't transform things to pull the divide out of the loop, and it runs many iterations, and a significant increase in code-size is with the performance gain (e.g. you aren't bottlenecked on cache misses that hide the div latency), then you might get a speedup from doing for run-time constants what the compiler does for compile-time constants.
Note that different constants need different shifts of the high half of the full-multiply, and some constants need more different shifts than others. (Another way of saying that some of the shift-counts are zero for some constants). So non-compile-time-constant divide-by-multiplying code needs all the shifts, and the shift counts have to be variable-count. (On x86, this is more expensive than immediate-count shifts).
libdivide has an implementation of the necessary math. You can use it to do SIMD-vectorized division, or for scalar, I think. This will definitely provide a big speedup over unpacking to scalar and doing integer division there. I haven't used it myself.
(Intel SSE/AVX doesn't do integer-division in hardware, but provides a variety of multiplies, and fairly efficient variable-count shift instructions. For 16bit elements, there's an instruction that produces only the high half of the multiply. For 32bit elements, there's a widening multiply, so you'd need a shuffle with that.)
Anyway, you could use libdivide to vectorize that add loop, with a horizontal sum at the end.
Other ways to get the div out of the loop
for (i=0; i<1000000000; i++)
s += i/a;
In your example, you might get better results from using a uint128_t s accumulator and dividing by a outside the loop. A 64bit add/adc pair is pretty cheap. (It wouldn't give identical results, though, because integer division truncates instead of rounding to nearest.)
I think you can account for that by looping with i += a; tmp++, and doing s += tmp*a, to combine all the adds from iterations where i/a is the same. So s += 1 * a accounts for all the iterations from i = [a .. a*2-1]. Obviously that was just a trivial example, and looping more efficiently is usually not actually possible. It's off-topic for this question, but worth saying anyway: Look for big optimizations by re-structuring code or taking advantage of some math before trying to speed up doing the exact same thing faster. Speaking of math, you can use the sum(0..n) = n * (n+1) / 2 formula here, because we can factor a out of a*1 + a*2 + a*3 ... a*max. I may have an off-by-one here, but I'm confident a closed-form simple constant time calculation will give the same answer as the loop for any a:
uint32_t n = 1000000000 / a;
uint32_t s = a * n*(n+1)/2 + 1000000000 % a;
If you just needed i/a in a loop, it might be worth it to do something like:
// another optimization for an unlikely case
for (uint32_t i=0, remainder=0, i_over_a=0 ; i < n ; i++) {
// use i_over_a
++remainder;
if (remainder == a) { // if you don't need the remainder in the loop, it could save an insn or two to count down from a to 0 instead of up from 0 to a, e.g. on x86. But then you need a clever variable name other than remainder.
remainder = 0;
++i_over_a;
}
}
Again, this is unlikely: it only works if you're dividing the loop counter by a constant. However, it should work well. Either a is large so branch mispredicts will be infrequent, or a is (hopefully) small enough for a good branch predictor to recognize the repeating pattern of a-1 branches one way, then 1 branch the other way. The worst-case a value might be 33 or 65 or something, depending on microarchitecture. Branchless asm is probably possible but not worth it. e.g. handle ++i_over_a with an add-with-carry and a conditional move for zeroing. (e.g. x86 pseudo-code cmp a-1, remainder / cmovc remainder, 0 / adc i_over_a, 0. The b (below) condition is just CF==1, same as the c (carry) condition. The branchless asm would be simplified by decrementing from a to 0. (don't need a zeroed reg for cmov, and could have a in a reg instead of a-1))
Replacing DIV with MUL may make sense (but doesn't have to in all cases) when one of the values is known at compile time. When both are user inputs, you don't know what's the range, so all usual tricks will not work.
Basically you need to handle both a and b between INT_MAX and INT_MIN. There's no space left for scaling them up/down. Even if you wanted to extend them to larger types, it would probably take longer time just to invert b and check that the result will be consistent.
The only way to KNOW if div or mul is faster is by testing both in a benchmark [obviously, if you use your above code, you'd mostly measure the time of read/write of the inputs and results, not the actual divide instruction, so you need something where you can isolate the divide instruction(s) from the input and output].
My guess would be that on slightly older processors, mul is a bit faster, on modern processors, div will be as fast as, if not faster than, a lookup of 256 int values.
If you have ONE target system, then it's plausible to test this. If you have several different systems you want to run on, you will have to ensure the "improved code" is faster on at least some of them - and not significantly slower on the rest.
Note also that you would introduce a dependency, which may in itself slow down the sequence of operations - modern CPU's are pretty good at "hiding" latency as long as there are other instructions to execute [so you should use this in an "as realistic scenario as possible"].
There is a wrong assumption in the question. The multiplicative inverse of an integer greater than 1 is a fraction less than one. These don't exist in the world of integers. A lookup table doesn't work because you can't lookup what doesn't exist. Even if you "scale" the dividend the results will not be correct in the sense of being the same as an integer division. Take this example:
printf("%x %x\n", 0x10/0x9, 0x30/0x9);
// prints: 1 5
Assuming a multiplicative inverse existed, both terms are divided by the same divisor (9) so must have the same lookup table value (multiplicative inverse). Any fixed lookup value corresponding to the divisor (9) multiplied by an integer will be precisely 3 times greater in the second term relative to the first term. As you can see from the example, the result of an actual integer division is a 5, not a 3.
You can approximate things by using a scaled lookup table. For instance a lookup table that is the multiplicative inverse when the result is divided by 2^16. You would then multiply by the lookup table value and shift the result 16 bits to the right. Time consuming and requires a 1024 byte lookup table. Even so, this would not produce the same results as an integer divide. A compiler optimization is not going to produce "approximate" results of an integer division.

g++ optimizing arihmetic expression with trivial terms (zero-product terms)

I writing mathematical code with a complicated arithmetic expression that is being used with varying types of inputs, but for the purpose of hyper-optimization AND readability, I want to use it as a macro rather than a function. However, I'm having trouble getting g++ (4.9.2) to optimize in a way I would think is obvious and need help.
The arithmetic operation of interest is in only one location, but the rub is that there are numerous permutations to the input variables. That is, the equation variables can be either a singleton or an array. For the purpose of conserving memory and CPU cache, I want to code for each variation (i.e. array and singleton input). Another element of the problem is that some terms of the equation drop out under certain conditions variables might not just be a singleton, but known a priori to be exactly 0 for certain instances, thus causing terms of the equation to be trivial.
Consider the following simplified, illustrative example
float a[128],b[128],c[128],d[128],e[128],r[128];
float a0, b0, c0, d0, e0;
...
#define ITERATE for (int n=1; n<N; n++)
#define THE_EQ(A,B,C,D,E,F) r[n] = A*(B-C) + D*E + F
...
switch(condition) {
case 1: ITERATE THE_EQ(a[n],b[n],c[n],d[n],e[n],0); break;
case 2: ITERATE THE_EQ( a0 ,b[n],c[n],d[n],e[n],0); break;
case 3: ITERATE THE_EQ( 0 ,b[n],c[n],d[n],e[n],0); break;
case 4: ITERATE THE_EQ(a[n],b[n],c[n],d[n],e[n],0); break;
case 5: ITERATE THE_EQ( a0 ,b[n],c[n], d0 ,e[n],0); break;
case 6: ITERATE THE_EQ( 0 ,b[n],c[n], 0 ,e[n],1); break;
....
} // switch
The real code has a more complex equation/macro and has many more permutations of input variable, but the above code is my test of g++ compiler/optimizer. What I'm finding is that g++ is not optimizing as I had hoped. That is, g++ is not dropping, when appropriate, the trivial arithmetic operations.
For example, in case 3, I expect the final microcode free of any b[n] - c[n] related operation. And for case 6, I expect elements in array r[n] to be simply set to 1, with no arithmetic operations at all. Unfortunately, neither is the case. :(
The optimizations that does occur are the AVX vectorization and using VXORPS to generate the zero value. But with the optimally generated 0, g++ still goes thru with the trival 0*(b[n]-c[n]) and 0*e[n] operations.
Witness:
; case 6: ITERATE THE_EQ(0,b[n],c[n],0,e[n],1); break;
0x400778 vxorps xmm1,xmm1,xmm1 ; generate 0
...
0x4007a0 vmovaps ymm0,YMMWORD PTR [rsi+rax*1]
0x4007a5 vmulps ymm2,ymm1,YMMWORD PTR [rdx+rax*1] ; 0*e[n]
0x4007aa vsubps ymm0,ymm0,YMMWORD PTR [rcx+rax*1] ; b[n]-c[n]
0x4007af vmulps ymm0,ymm0,ymm1 ; (b[n]-c[n])*0
0x4007b3 vaddps ymm0,ymm0,ymm2 ; 0*(b[n]-c[n])+0*e[n]
0x4007b7 vaddps ymm0,ymm0,ymm3 ; ....+1
0x4007bb vmovaps YMMWORD PTR [rdi+rax*1],ymm0
0x4007c0 add rax,0x20
0x4007c4 cmp rax,0x200
0x4007ca jne 0x4007a0 <main()+128>
Obviously, this is NOT optimal. Is there a technique (via PRAGMA or compile flag) that will encourage g++ drop those trivial arithmetic operations? Or is there another method, other than hardcoding 60+ different expressions, to get to my desired level of optimization.
Command is
g++-4.9 -Wall -fexceptions -O3 -g -march=corei7-avx main.cpp
with g++ 4.9.2
/* Maybe fold x * 0 to 0. The expressions aren't the same
when x is NaN, since x * 0 is also NaN. Nor are they the
same in modes with signed zeros, since multiplying a
negative value by 0 gives -0, not +0. */
The optimization in gcc is guarded by a few conditions. -ffast-math is the easiest way to convince gcc that you do not care about those cases.

Fast multiplication/division by 2 for floats and doubles (C/C++)

In the software I'm writing, I'm doing millions of multiplication or division by 2 (or powers of 2) of my values. I would really like these values to be int so that I could access the bitshift operators
int a = 1;
int b = a<<24
However, I cannot, and I have to stick with doubles.
My question is : as there is a standard representation of doubles (sign, exponent, mantissa), is there a way to play with the exponent to get fast multiplications/divisions by a power of 2?
I can even assume that the number of bits is going to be fixed (the software will work on machines that will always have 64 bits long doubles)
P.S : And yes, the algorithm mostly does these operations only. This is the bottleneck (it's already multithreaded).
Edit : Or am I completely mistaken and clever compilers already optimize things for me?
Temporary results (with Qt to measure time, overkill, but I don't care):
#include <QtCore/QCoreApplication>
#include <QtCore/QElapsedTimer>
#include <QtCore/QDebug>
#include <iostream>
#include <math.h>
using namespace std;
int main(int argc, char *argv[])
{
QCoreApplication a(argc, argv);
while(true)
{
QElapsedTimer timer;
timer.start();
int n=100000000;
volatile double d=12.4;
volatile double D;
for(unsigned int i=0; i<n; ++i)
{
//D = d*32; // 200 ms
//D = d*(1<<5); // 200 ms
D = ldexp (d,5); // 6000 ms
}
qDebug() << "The operation took" << timer.elapsed() << "milliseconds";
}
return a.exec();
}
Runs suggest that D = d*(1<<5); and D = d*32; run in the same time (200 ms) whereas D = ldexp (d,5); is much slower (6000 ms). I know that this is a micro benchmark, and that suddenly, my RAM has exploded because Chrome has suddenly asked to compute Pi in my back every single time I run ldexp(), so this benchmark is worth nothing. But I'll keep it nevertheless.
On the other had, I'm having trouble doing reinterpret_cast<uint64_t *> because there's a const violation (seems the volatile keyword interferes)
This is one of those highly-application specific things. It may help in some cases and not in others. (In the vast majority of cases, a straight-forward multiplication is still best.)
The "intuitive" way of doing this is just to extract the bits into a 64-bit integer and add the shift value directly into the exponent. (this will work as long as you don't hit NAN or INF)
So something like this:
union{
uint64 i;
double f;
};
f = 123.;
i += 0x0010000000000000ull;
// Check for zero. And if it matters, denormals as well.
Note that this code is not C-compliant in any way, and is shown just to illustrate the idea. Any attempt to implement this should be done directly in assembly or SSE intrinsics.
However, in most cases the overhead of moving the data from the FP unit to the integer unit (and back) will cost much more than just doing a multiplication outright. This is especially the case for pre-SSE era where the value needs to be stored from the x87 FPU into memory and then read back into the integer registers.
In the SSE era, the Integer SSE and FP SSE use the same ISA registers (though they still have separate register files). According the Agner Fog, there's a 1 to 2 cycle penalty for moving data between the Integer SSE and FP SSE execution units. So the cost is much better than the x87 era, but it's still there.
All-in-all, it will depend on what else you have on your pipeline. But in most cases, multiplying will still be faster. I've run into this exact same problem before so I'm speaking from first-hand experience.
Now with 256-bit AVX instructions that only support FP instructions, there's even less of an incentive to play tricks like this.
How about ldexp?
Any half-decent compiler will generate optimal code on your platform.
But as #Clinton points out, simply writing it in the "obvious" way should do just as well. Multiplying and dividing by powers of two is child's play for a modern compiler.
Directly munging the floating point representation, besides being non-portable, will almost certainly be no faster (and might well be slower).
And of course, you should not waste time even thinking about this question unless your profiling tool tells you to. But the kind of people who listen to this advice will never need it, and the ones who need it will never listen.
[update]
OK, so I just tried ldexp with g++ 4.5.2. The cmath header inlines it as a call to __builtin_ldexp, which in turn...
...emits a call to the libm ldexp function. I would have thought this builtin would be trivial to optimize, but I guess the GCC developers never got around to it.
So, multiplying by 1 << p is probably your best bet, as you have discovered.
You can pretty safely assume IEEE 754 formatting, the details of which can get pretty gnarley (esp. when you get into subnormals). In the common cases, however, this should work:
const int DOUBLE_EXP_SHIFT = 52;
const unsigned long long DOUBLE_MANT_MASK = (1ull << DOUBLE_EXP_SHIFT) - 1ull;
const unsigned long long DOUBLE_EXP_MASK = ((1ull << 63) - 1) & ~DOUBLE_MANT_MASK;
void unsafe_shl(double* d, int shift) {
unsigned long long* i = (unsigned long long*)d;
if ((*i & DOUBLE_EXP_MASK) && ((*i & DOUBLE_EXP_MASK) != DOUBLE_EXP_MASK)) {
*i += (unsigned long long)shift << DOUBLE_EXP_SHIFT;
} else if (*i) {
*d *= (1 << shift);
}
}
EDIT: After doing some timing, this method is oddly slower than the double method on my compiler and machine, even stripped to the minimum executed code:
double ds[0x1000];
for (int i = 0; i != 0x1000; i++)
ds[i] = 1.2;
clock_t t = clock();
for (int j = 0; j != 1000000; j++)
for (int i = 0; i != 0x1000; i++)
#if DOUBLE_SHIFT
ds[i] *= 1 << 4;
#else
((unsigned int*)&ds[i])[1] += 4 << 20;
#endif
clock_t e = clock();
printf("%g\n", (float)(e - t) / CLOCKS_PER_SEC);
In the DOUBLE_SHIFT completes in 1.6 seconds, with an inner loop of
movupd xmm0,xmmword ptr [ecx]
lea ecx,[ecx+10h]
mulpd xmm0,xmm1
movupd xmmword ptr [ecx-10h],xmm0
Versus 2.4 seconds otherwise, with an inner loop of:
add dword ptr [ecx],400000h
lea ecx, [ecx+8]
Truly unexpected!
EDIT 2: Mystery solved! One of the changes for VC11 is now it always vectorizes floating point loops, effectively forcing /arch:SSE2, though VC10, even with /arch:SSE2 is still worse with 3.0 seconds with an inner loop of:
movsd xmm1,mmword ptr [esp+eax*8+38h]
mulsd xmm1,xmm0
movsd mmword ptr [esp+eax*8+38h],xmm1
inc eax
VC10 without /arch:SSE2 (even with /arch:SSE) is 5.3 seconds... with 1/100th of the iterations!!, inner loop:
fld qword ptr [esp+eax*8+38h]
inc eax
fmul st,st(1)
fstp qword ptr [esp+eax*8+30h]
I knew the x87 FP stack was aweful, but 500 times worse is kinda ridiculous. You probably won't see these kinds of speedups converting, i.e. matrix ops to SSE or int hacks, since this is the worst case loading into the FP stack, doing one op, and storing from it, but it's a good example for why x87 is not the way to go for anything perf. related.
The fastest way to do this is probably:
x *= (1 << p);
This sort of thing may simply be done by calling an machine instruction to add p to the exponent. Telling the compiler to instead extract the some bits with a mask and doing something manually to it will probably make things slower, not faster.
Remember, C/C++ is not assembly language. Using a bitshift operator does not necessarily compile to a bitshift assembly operation, not does using multiplication necessarily compile to multiplication. There's all sorts of weird and wonderful things going on like what registers are being used and what instructions can be run simultaneously which I'm not smart enough to understand. But your compiler, with many man years of knowledge and experience and lots of computational power, is much better at making these judgements.
p.s. Keep in mind, if your doubles are in an array or some other flat data structure, your compiler might be really smart and use SSE to multiple 2, or even 4 doubles at the same time. However, doing a lot of bit shifting is probably going to confuse your compiler and prevent this optimisation.
Since c++17 you can also use hexadecimal floating literals. That way you can multiply by higher powers of 2. For instance:
d *= 0x1p64;
will multiply d by 2^64. I use it to implement my fast integer arithmetic in a conversion to double.
What other operations does this algorithm require? You might be able to break your floats into int pairs (sign/mantissa and magnitude), do your processing, and reconstitute them at the end.
Multiplying by 2 can be replaced by an addition: x *= 2 is equivalent to x += x.
Division by 2 can be replaced by multiplication by 0.5. Multiplication is usually significantly faster than division.
Although there is little/no practical benefit to treating powers of two specially for float of double types there is a case for this for double-double types. Double-double multiplication and division is complicated in general but is trivial for multiplying and dividing by a power of two.
E.g. for
typedef struct {double hi; double lo;} doubledouble;
doubledouble x;
x.hi*=2, x.lo*=2; //multiply x by 2
x.hi/=2, x.lo/=2; //divide x by 2
In fact I have overloaded << and >> for doubledouble so that it's analogous to integers.
//x is a doubledouble type
x << 2 // multiply x by four;
x >> 3 // divide x by eight.
Depending on what you're multiplying, if you have data that is recurring enough, a look up table might provide better performance, at the expense of memory.

How efficient is an if statement compared to a test that doesn't use an if? (C++)

I need a program to get the smaller of two numbers, and I'm wondering if using a standard "if x is less than y"
int a, b, low;
if (a < b) low = a;
else low = b;
is more or less efficient than this:
int a, b, low;
low = b + ((a - b) & ((a - b) >> 31));
(or the variation of putting int delta = a - b at the top and rerplacing instances of a - b with that).
I'm just wondering which one of these would be more efficient (or if the difference is too miniscule to be relevant), and the efficiency of if-else statements versus alternatives in general.
(Disclaimer: the following deals with very low-level optimizations that are most often not necessary. If you keep reading, you waive your right to complain that computers are fast and there is never any reason to worry about this sort of thing.)
One advantage of eliminating an if statement is that you avoid branch prediction penalties.
Branch prediction penalties are generally only a problem when the branch is not easily predicted. A branch is easily predicted when it is almost always taken/not taken, or it follows a simple pattern. For example, the branch in a loop statement is taken every time except the last one, so it is easily predicted. However, if you have code like
a = random() % 10
if (a < 5)
print "Less"
else
print "Greater"
then this branch is not easily predicted, and will frequently incur the prediction penalty associated with clearing the cache and rolling back instructions that were executed in the wrong part of the branch.
One way to avoid these kinds of penalties is to use the ternary (?:) operator. In simple cases, the compiler will generate conditional move instructions rather than branches.
So
int a, b, low;
if (a < b) low = a;
else low = b;
becomes
int a, b, low;
low = (a < b) ? a : b
and in the second case a branching instruction is not necessary. Additionally, it is much clearer and more readable than your bit-twiddling implementation.
Of course, this is a micro-optimization which is unlikely to have significant impact on your code.
Simple answer: One conditional jump is going to be more efficient than two subtractions, one addition, a bitwise and, and a shift operation combined. I've been sufficiently schooled on this point (see the comments) that I'm no longer even confident enough to say that it's usually more efficient.
Pragmatic answer: Either way, you're not paying nearly as much for the extra CPU cycles as you are for the time it takes a programmer to figure out what that second example is doing. Program for readability first, efficiency second.
Compiling this on gcc 4.3.4, amd64 (core 2 duo), Linux:
int foo1(int a, int b)
{
int low;
if (a < b) low = a;
else low = b;
return low;
}
int foo2(int a, int b)
{
int low;
low = b + ((a - b) & ((a - b) >> 31));
return low;
}
I get:
foo1:
cmpl %edi, %esi
cmovle %esi, %edi
movl %edi, %eax
ret
foo2:
subl %esi, %edi
movl %edi, %eax
sarl $31, %eax
andl %edi, %eax
addl %esi, %eax
ret
...which I'm pretty sure won't count for branch predictions, since the code doesn't jump. Also, the non if-statement version is 2 instructions longer. I think I will continue coding, and let the compiler do it's job.
Like with any low-level optimization, test it on the target CPU/board setup.
On my compiler (gcc 4.5.1 on x86_64), the first example becomes
cmpl %ebx, %eax
cmovle %eax, %esi
The second example becomes
subl %eax, %ebx
movl %ebx, %edx
sarl $31, %edx
andl %ebx, %edx
leal (%rdx,%rax), %esi
Not sure if the first one is faster in all cases, but I would bet it is.
The biggest problem is that your second example won't work on 64-bit machines.
However, even neglecting that, modern compilers are smart enough to consider branchless prediction in every case possible, and compare the estimated speeds. So, you second example will most likely actually be slower
There will be no difference between the if statement and using a ternary operator, as even most dumb compilers are smart enough to recognize this special case.
[Edit] Because I think this is such an interesting topic, I've written a blog post on it.
Either way, the assembly will only be a few instructions and either way it'll take picoseconds for those instructions to execute.
I would profile the application an concentrate your optimization efforts to something more worthwhile.
Also, the time saved by this type of optimization will not be worth the time wasted by anyone trying to maintain it.
For simple statements like this, I find the ternary operator very intuitive:
low = (a < b) ? a : b;
Clear and concise.
For something as simple as this, why not just experiment and try it out?
Generally, you'd profile first, identify this as a hotspot, experiment with a change, and view the result.
I wrote a simple program that compares both techniques passing in random numbers (so that we don't see perfect branch prediction) with Visual C++ 2010. The difference between the approaches on my machine for 100,000,000 iteration? Less than 50ms total, and the if version tended to be faster. Looking at the codegen, the compiler successfully converted the simple if to a cmovl instruction, avoiding a branch altogether.
One thing to be wary of when you get into really bit-fiddly kinds of hacks is how they may interact with compiler optimizations that take place after inlining. For example, the readable procedure
int foo (int a, int b) {
return ((a < b) ? a : b);
}
is likely to be compiled into something very efficient in any case, but in some cases it may be even better. Suppose, for example, that someone writes
int bar = foo (x, x+3);
After inlining, the compiler will recognize that 3 is positive, and may then make use of the fact that signed overflow is undefined to eliminate the test altogether, to get
int bar = x;
It's much less clear how the compiler should optimize your second implementation in this context. This is a rather contrived example, of course, but similar optimizations actually are important in practice. Of course you shouldn't accept bad compiler output when performance is critical, but it's likely wise to see if you can find clear code that produces good output before you resort to code that the next, amazingly improved, version of the compiler won't be able to optimize to death.
One thing I will point out that I haven't noticed mention that an optimization like this can easily be overwhelmed by other issues. For example, if you are running this routine on two large arrays of numbers (or worse yet, pairs of number scattered in memory), the cost of fetching the values on today's CPUs can easily stall the CPU's execution pipelines.
I'm just wondering which one of these
would be more efficient (or if the
difference is to miniscule to be
relevant), and the efficiency of
if-else statements versus alternatives
in general.
Desktop/server CPUs are optimized for pipelining. Second is theoretically faster because CPU doesn't have to branch and can utilize multiple ALUs to evaluate parts of expression in parallel. More non-branching code with intermixed independent operations are best for such CPUs. (But even that is negated now by modern "conditional" CPU instructions which allow to make the first code branch-less too.)
On embedded CPUs branching if often less expensive (relatively to everything else), nor they have many spare ALUs to evaluate operations out-of-order (that's if they support out-of-order execution at all). Less code/data is better - caches are small too. (I have even seen uses of buble-sort in embedded applications: the algorithm uses least of memory/code and fast enough for small amounts of information.)
Important: do not forget about the compiler optimizations. Using many tricks, the compilers sometimes can remove the branching themselves: inlining, constant propagation, refactoring, etc.
But in the end I would say that yes, the difference is minuscule to be relevant. In long term, readable code wins.
The way things go on the CPU front, it is more rewarding to invest time now in making the code multi-threaded and OpenCL capable.
Why low = a; in the if and low = a; in the else? And, why 31? If 31 has anything to do with CPU word size, what if the code is to be run on a CPU of different size?
The if..else way looks more readable. I like programs to be as readable to humans as they are to the compilers.
profile results with gcc -o foo -g -p -O0, Solaris 9 v240
%Time Seconds Cumsecs #Calls msec/call Name
36.8 0.21 0.21 8424829 0.0000 foo2
28.1 0.16 0.37 1 160. main
17.5 0.10 0.4716850667 0.0000 _mcount
17.5 0.10 0.57 8424829 0.0000 foo1
0.0 0.00 0.57 4 0. atexit
0.0 0.00 0.57 1 0. _fpsetsticky
0.0 0.00 0.57 1 0. _exithandle
0.0 0.00 0.57 1 0. _profil
0.0 0.00 0.57 1000 0.000 rand
0.0 0.00 0.57 1 0. exit
code:
int
foo1 (int a, int b, int low)
{
if (a < b)
low = a;
else
low = b;
return low;
}
int
foo2 (int a, int b, int low)
{
low = (a < b) ? a : b;
return low;
}
int main()
{
int low=0;
int a=0;
int b=0;
int i=500;
while (i--)
{
for(a=rand(), b=rand(); a; a--)
{
low=foo1(a,b,low);
low=foo2(a,b,low);
}
}
return 0;
}
Based on data, in the above environment, the exact opposite of several beliefs stated here were not found to be true. Note the 'in this environment' If construct was faster than ternary ? : construct
I had written ternary logic simulator not so long ago, and this question was viable to me, as it directly affects my interpretator execution speed; I was required to simulate tons and tons of ternary logic gates as fast as possible.
In a binary-coded-ternary system one trit is packed in two bits. Most significant bit means negative and least significant means positive one. Case "11" should not occur, but it must be handled properly and threated as 0.
Consider inline int bct_decoder( unsigned bctData ) function, which should return our formatted trit as regular integer -1, 0 or 1; As i observed there are 4 approaches: i called them "cond", "mod", "math" and "lut"; Lets investigate them
First is based on jz|jnz and jl|jb conditional jumps, thus cond. Its performance is not good at all, because relies on a branch predictor. And even worse - it varies, because it is unknown if there will be one branch or two a priori. And here is an example:
inline int bct_decoder_cond( unsigned bctData ) {
unsigned lsB = bctData & 1;
unsigned msB = bctData >> 1;
return
( lsB == msB ) ? 0 : // most possible -> make zero fastest branch
( lsB > msB ) ? 1 : -1;
}
This is slowest version, it could involve 2 branches in worst case and this is something where binary logic fails. On my 3770k it prodices around 200MIPS on average on random data. (here and after - each test is average from 1000 tries on randomly filled 2mb dataset)
Next one relies on modulo operator and its speed is somewhere in between first and third, but is definetely faster - 600 MIPS:
inline int bct_decoder_mod( unsigned bctData ) {
return ( int )( ( bctData + 1 ) % 3 ) - 1;
}
Next one is branchless approach, which involves only maths, thus math; it does not assume jump instrunctions at all:
inline int bct_decoder_math( unsigned bctData ) {
return ( int )( bctData & 1 ) - ( int )( bctData >> 1 );
}
This does what is should, and behaves really great. To compare, performance estimate is 1000 MIPS, and it is 5x faster than branched version. Probably branched version is slowed down due to lack of native 2-bit signed int support. But in my application it is quite good version in itself.
If this is not enough then we can go futher, having something special. Next is called lookup table approach:
inline int bct_decoder_lut( unsigned bctData ) {
static const int decoderLUT[] = { 0, 1, -1, 0 };
return decoderLUT[ bctData & 0x3 ];
}
In my case one trit occupied only 2 bits, so lut table was only 2b*4 = 8 bytes, and was worth trying. It fits in cache and works blazing fast at 1400-1600 MIPS, here is where my measurement accuracy is going down. And that is is 1.5x speedup from fast math approach. That's because you just have precalculated result and single AND instruction. Sadly caches are small and (if your index length is greater than several bits) you simply cannot use it.
So i think i answered your question, on what what could branched/branchless code be like. Answer is much better and with detailed samples, real world application and real performance measurements results.
Updated answer taking the current (2018) state of compiler vectorization. Please see danben's answer for the general case where vectorization is not a concern.
TLDR summary: avoiding ifs can help with vectorization.
Because SIMD would be too complex to allow branching on some elements, but not others, any code containing an if statement will fail to be vectorized unless the compiler knows a "superoptimization" technique that can rewrite it into a branchless set of operations. I don't know of any compilers that are doing this as an integrated part of the vectorization pass (Clang does some of this independently, but not specificly toward helping vectorization AFAIK)
Using the OP's provided example:
int a, b, low;
low = b + ((a - b) & ((a - b) >> 31));
Many compilers can vectorize this to be something approximately equivalent to:
__m128i low128i(__m128i a, __m128i b){
__m128i diff, tmp;
diff = _mm_sub_epi32(a,b);
tmp = _mm_srai_epi32(diff, 31);
tmp = _mm_and_si128(tmp,diff);
return _mm_add_epi32(tmp,b);
}
This optimization would require the data to be layed out in a fashion that would allow for it, but it could be extended to __m256i with avx2 or __m512i with avx512 (and even unroll loops further to take advantage of additional registers) or other simd instructions on other architectures. Another plus is that these instructions are all low latency, high-throughput instructions (latencies of ~1 and reciprocal throughputs in the range of 0.33 to 0.5 - so really fast relative to non-vectorized code)
I see no reason why compilers couldn't optimize an if statement to a vectorized conditional move (except that the corresponding x86 operations only work on memory locations and have low throughput and other architectures like arm may lack it entirely) but it could be done by doing something like:
void lowhi128i(__m128i *a, __m128i *b){ // does both low and high
__m128i _a=*a, _b=*b;
__m128i lomask = _mm_cmpgt_epi32(_a,_b),
__m128i himask = _mm_cmpgt_epi32(_b,_a);
_mm_maskmoveu_si128(_b,lomask,a);
_mm_maskmoveu_si128(_a,himask,b);
}
However this would have a much higher latency due to memory reads and writes and lower throughput (higher/worse reciprocal throughput) than the example above.
Unless you're really trying to buckle down on efficiency, I don't think this is something you need to worry about.
My simple thought though is that the if would be quicker because it's comparing one thing, while the other code is doing several operations. But again, I imagine that the difference is minuscule.
If it is for Gnu C++, try this
int min = i <? j;
I have not profiled it but I think it is definitely the one to beat.