OpenCL floor function 没有返回正确的数字

Question

I am trying to convert a CUDA code snippet in OpenCL.

CUDA code:

# Variable point is CUDA float4.
# X_MIN_RANGE, Y_MIN_RANGE, PILLAR_SIZE and GRID_X_SIZE are all defined as
# constants in a header file.

int idx = floorf((point.x - X_MIN_RANGE)/PILLAR_SIZE);
int idy = floorf((point.y - Y_MIN_RANGE)/PILLAR_SIZE);
unsigned int index = idy * GRID_X_SIZE + idx;

OpenCL code:

# Variable point_coord is OpenCL float4.
# X_MIN_RANGE, Y_MIN_RANGE, PILLAR_SIZE and GRID_X_SIZE are defined as constants using #define.

#define X_MIN_RANGE 0.0
#define Y_MIN_RANGE -39.68
#define PILLAR_SIZE 0.16
#define GRID_X_SIZE 432

/*... Some code here...*/

int idx = floor((point.x - X_MIN_RANGE) / PILLAR_SIZE);
int idy = floor((point.y - Y_MIN_RANGE) / PILLAR_SIZE);
unsigned int index = idy * GRID_X_SIZE + idx;

I have some issues with the results of floor. For instance, for a point of coordinates {12.48, -10.629, -0.223, 0.4}, X_MIN_RANGE = 0.0, Y_MIN_RANGE = -39.68, PILLAR_SIZE = 0.16 and GRID_X_SIZE = 432 I'm expecting to see:

idx = floor((12.48 - 0.0) / 0.16) = floor(78.0) = 78
idy = floor((-10.69 + 39.68) / 0.16) = floor(181.1875) = 181

index = 181 * 432 + 78 = 78270

For the CUDA code, I am getting the correct result, however for the OpenCL code I am getting 78269 for the same coordinates as the floor operation for idx returns 77 instead of 77. I had a look for similar examples and it always happens when the decimal part of the argument in floor is 0, so I assume that the result falls on the wrong side of the integer.

As this part of the code needs to be quite accurate, do you know how can I make sure floor rounds integers correctly? I have alredy tried to rearrange the operations and to use double instead of float, but that didn't help.

Answer 1

(point.x - X_MIN_RANGE)/PILLAR_SIZE is calculated as double , because all defines are double . So the result is a double .

If you then use floorf (the f at the end stands for float ), it will be casted to a float before the function call. If you use floor , it stays a double .

The float value is indeed 78.00000 exact, so floorf returns 78 .

The double is 77.99999713897705078125 and then floor returns 77 .

You can fix that by either using float for the calculation or cast the result before the function call. You could also add a very small number, like 0.0001 before the floor to remove floating point inexactness.