Fastest way to convert a float to int in C++

Question

What is the fastest and most efficient way to convert a float to an integer in c++ (rounding toward zero)? is it

long ftoint(float x) { unsigned int e = (0x7F + 31) - ((* (unsigned int*) &x & 0x7F800000) >> 23); unsigned int m = 0x80000000 | (* (unsigned int*) &x << 8); return int((m >> e) & -(e < 32)); }

?

Beside not compiling, the ftoint routine also has undefined behavior. — Eljay
– Eljay, Commented Jul 5, 2021 at 14:55
@463035818_is_not_a_number clang encodes your int y = x; as a single instruction: vcvttss2si eax, xmm0. That seems pretty fast. — Adrian Mole
– Adrian Mole, Commented Jul 5, 2021 at 15:02
maybe you got misled by floor from cmath being actually useful. The reason is that it returns a float. If it would return an integer, I have no clue what would it be good for in c++, you dont need floor (or some other fancy stuff) to convert a float to an int — 463035818_is_not_an_ai
– 463035818_is_not_an_ai, Commented Jul 5, 2021 at 15:04
@Random why not just cast the float/double directly to an int/long or have it implicitly converted? That'll be the fastest. — mediocrevegetable1
– mediocrevegetable1, Commented Jul 5, 2021 at 15:07

463035818_is_not_an_ai · Accepted Answer · 2021-07-05 15:24:55Z

Lets compare the following two:

long ftoint(float x) { unsigned int e = (0x7F + 31) - ((* (unsigned int*) &x & 0x7F800000) >> 23); unsigned int m = 0x80000000 | (* (unsigned int*) &x << 8); return int((m >> e) & -(e < 32)); } long ftointfast(float x){ return x; }

Clang with -O3 produces:

ftoint(float): # @ftoint(float) movd eax, xmm0 mov ecx, eax shr ecx, 23 movzx edx, cl mov ecx, 158 sub ecx, edx shl eax, 8 or eax, -2147483648 shr eax, cl xor edx, edx cmp ecx, 32 cmovb edx, eax movsxd rax, edx ret ftointfast(float): # @ftointfast(float) cvttss2si rax, xmm0 ret

I am not fluent in assembly, but I am certain that you cannot get it faster than a single instruction.

std::floor(arg) computes the largest integer value not greater than arg. It returns a floating point value. If you do not need a floating point value but only the integer then you do not need std::floor. You also do not need to compare your solution to std::floor because it does something you don't need. And of course you can just write (assuming x actually fits in the range of long):

long y = x;

or to be explicit

long y = static_cast<long>(x);

it's much more tricky than that and a single instruction can be slower than a bunch of instruction. And it depend on the target cpu and the allowed instruction set. This topic is a school case of "don't try to outsmart your compiler" :)

Stack Exchange Broke The Law · Accepted Answer · 2021-07-05 15:09:16Z

1

Probably with

long ftoint(float x) { return (long)x; }

(or static_cast if you hate C-style casts)

In fact, you don't need a function. You can just write the cast.

answered Jul 5, 2021 at 15:09

Stack Exchange Broke The Law

59.2k7 gold badges57 silver badges103 bronze badges

3 Comments

Ben Voigt Over a year ago

This has "round-toward-zero" behavior, "floor" has round-down behavior. Not the same.

Stack Exchange Broke The Law Over a year ago

However, the question doesn't say it needs to round down.

Random Over a year ago

i went with static_cast<int>(var_b);

Collectives™ on Stack Overflow

Fastest way to convert a float to int in C++

2 Answers 2

1 Comment

3 Comments

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

1 Comment

3 Comments

Related