使用NodeJS N-API计算浮点值时出现问题

Question

Hello everyone,

Coming from the web dev world. I am currently trying to do some C code which is converting an RGB value to an XYZ value that can be used by NodeJS through N-API. The issue that I have is regarding to float calculation . Below is the explanation of my issue:

Based on this C code below, this code is trying to convert an RGB value to an XYZ value.

char * colorType = getStringValue(env, funcParams[2], SPACELen);
// m is either the value srgb | adobeRgb
Matrix m = getEnumFromStr(colorType);
Rgb * rgb = getRGBFromJSObj(env, funcParams[0]);
xyz = generateXyzFromRgb(rgb, m);

And I am using this JS snippet to call my library

const rgb = {
  r: 255,
  g: 255,
  b: 255
};

const xyz = lib.getXyzFromRgb(rgb, "srgb", 10000);
expect(xyz).to.be.deep.equal({
  x: 0.9504,
  y: 1,
  z: 1.0888
});

If everything should be all right the output should be like the one below

{
 x: 0.9504,
 y: 1,
 z: 1.0888 
}

However the output that I have is this one

{
 x: 0.9502,
 y: 0.9997,
 z: 1.0886 
}

As you can see the output is totally wrong. However this wrong output only happened on my local machine (OSX) and only when I am trying to do the conversion by using the JS snippet.

Indeed, when I am trying to run the conversion with this piece of code below directly through Xcode the output is correct

// RGB and & m variable is outputing the same value as the conversion done by the C code above
xyz = generateXyzFromRgb(rgb, m);

Moreover when I'm trying to call the JS code through travis which also running OSX and on Ubuntu through Docker the JS code also output the right value .

Can it be more related to the hardware or the way I am compiling my libraries or else ?

Thank you in advance.

Answer 1

In javascript v8 engine. Actually only exist smi and double. readFloatBE: when float pass to v8, the float will be cast to double. it will promote to the double when you go out and rounding into a float.

First, if you really want to get value like C, you need to manually specify the rounded digits and re-instantiate the Number object with the Number.prototype.toPrecision function rounded up:

For your reference:

var v = 5.2
var buffer = new Buffer(5)
buffer.writeFloatBE(v)
var g = buffer.readFloatBE()

console.log(v)
console.log(g)
console.log(v==g)
console.log(Number(g.toPrecision(5)))

Answer 2

Indeed, as post by Pointy and user10683038 JavaScript numbers are 64 bits.

As I coded this feature the numeric type used was float (was not thinking at making a Node module at first glance). Later on I decided to do a conversion from float to double using a small util method like below

double v = *(double *) arg;
status = napi_create_double(env, v, &value);

Strangely it appear that I might have randomly loosing decimal precision while doing this conversion while it should have not (32bit to 64bit) or I might have missing out something.

Later on I refactor the same code but this time by using double instead and it works as it should. Many thanks to all of you.

Note : As I used this util code above method accross the NodeJS binding codebase I am surprised that I haven't got a decimal precision error earlier.