Public channel for discussing Numba usage. Don't post confidential info here! Consider posting questions to: https://numba.discourse.group/ !
Shouldn't these kernels be overlapping?
import numpy as np
import numba as nb
from numba import cuda
import math
hundred_twenty_eight_floats = np.zeros(128)
hundred_twenty_eight_floats[:] = list(range(128))
@cuda.jit(nb.void(nb.float64[::1], nb.bool_))
def cycle(vals,update_early):
offset = 64 if update_early else 0
for i in range(5000):
vals[cuda.grid(1) + offset] = math.sin(vals[cuda.grid(1) + offset])
stream = cuda.stream()
stream2 = cuda.stream()
for i in range(25):
cycle[2, 32, stream](hundred_twenty_eight_floats, False)
cycle[2, 32, stream2](hundred_twenty_eight_floats, True)
when I actually run it I get:
And if your GPU doesn't have enough warps to execute them at the same time, it wont.
needed to move it to device, note the tiny d-to-h
import numpy as np
import numba as nb
from numba import cuda
import math
hundred_twenty_eight_floats_h = np.zeros(128)
hundred_twenty_eight_floats_h[:] = list(range(128))
hundred_twenty_eight_floats = cuda.to_device(hundred_twenty_eight_floats_h)
@cuda.jit(nb.void(nb.float64[::1], nb.bool_))
def cycle(vals,update_early):
offset = 64 if update_early else 0
for i in range(5000):
vals[cuda.grid(1) + offset] = math.sin(vals[cuda.grid(1) + offset])
stream = cuda.stream()
stream2 = cuda.stream()
for i in range(25):
cycle[2, 32, stream](hundred_twenty_eight_floats, False)
cycle[2, 32, stream2](hundred_twenty_eight_floats, True)
hundred_twenty_eight_floats.to_host()
print(hundred_twenty_eight_floats_h.sum())
No option to skip the check.
e.g.
@jitdecorator(error_model='numpy'). Docs https://numba.readthedocs.io/en/stable/reference/jit-compilation.html?highlight=error_model#numba.jit
that might help?
Hi folks, I have a problem and I am not sure how to solve it. I have a cfunc to which I want to pass a list of pointers (they can point at different types). I want to use carray to get view on this pointers data. Is there a way to cast an int64 to a pointer? I couldn't find anything usefull.
23 replies
doing x[:] = 0 is slow
Hi I am trying to install numba through virtualenv (pip install numba) in a jupyter hub (remote instance) but I get the following error:
error: Command "gcc -pthread -B /sw/spack-rhel6/jupyterhub/jupyterhub/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -Inumba -I/mnt/lustre01/pf/b/b381465/kernels/data/include -I/sw/spack-rhel6/jupyterhub/jupyterhub/include/python3.8 -c numba/_devicearray.cpp -o build/temp.linux-x86_64-3.8/numba/_devicearray.o -std=c++11" failed with exit status 1
Llvmlite is already installed and operative but I can't never get to install numba..
70 replies
Hi :) I was wondering what happens when I call an
njit function (A) with parallel=True and prange inside from a prange loop in another @njit(parallel=True) function (B). Will the prange in the first function (A) be sequential when called from within the second function (B)?
2 replies
Hello, was wondering what this issue means-----
I'm trying to run esda.moran.Moran_Local, which is a function that examines spatial autocorrelation across geographic space.
LoweringError: Failed in nopython mode pipeline (step: nopython mode backend)
Operands must be the same type, got (i64, i32)I'm trying to run esda.moran.Moran_Local, which is a function that examines spatial autocorrelation across geographic space.
10 replies
import numpy as np
import numba as nb
from numba import cuda
sixty_four = np.zeros(64)
blocks_of_thirty_two = sixty_four.reshape(-1,32)
nb_sixty_four = cuda.to_device(sixty_four)
nb_sixty_four.reshape(-1, 32)
Hi, I am trying to call a AOT function from C (without any python interpreter). With a simple function like
double square(double) you can easily call it if you define extern double square(double) asm ("cfunc._ZN10mymodule11square$2419Ed");.The case of a function likeunicode_type myrepeat(unicode_type s, int count) seems to be a bit more complex. I am not quite sure how the signature of this function should look like and how I should allocate my unicode_type argument (NRT or Py_BuildValue?). Maybe you guys can give me a hint?
4 replies
@gmarkall @stuartarchibald This might be to little to late given the recent announcement (https://developer.nvidia.com/cuda-python#) but I finally published the ctypes CUDA bindings. I'm not sure if there is a better place to put this but here's the link if anyone wants to take a look. It is still very much a work in progress but most of the work in on the higher level Python API.
https://github.com/uchytilc/PyCu
https://github.com/uchytilc/PyCu
20 replies
Hi! I am busy trying to compose a relatively complicated piece of numba code using generated_jit/overload. However, whilst moving stuff to generated jit I noticed that performance was getting steadily worse. I then coded up the following example which demonstrates a fairly large discrepancy between jitted/generated_jit/overload code. I am not quite sure if this qualifies as a bug and wanted to be sure that I am not doing something obviously wrong. I will place the code in a reply to this message.
11 replies
Hi all, when someone passes in a view of an array, e.g. the columns of a C_CONTIGUOUS 2D array, I assume numba doesn't make a copy?
3 replies
I am having trouble initialising a jitclass with a singular instance of numpy dtype. The following is the code where I am getting the error.
import numpy as np
from numba import jitclass, typeof
counter_dtype = np.dtype([('element', np.int32, 5)])
one_counter = np.zeros(1, dtype=counter_dtype)[0]
spec = [
("counter", typeof(one_counter))
]
@jitclass(spec)
class UpdatingStuff:
def __init__(self, counter):
self.counter = counter
UpdatingStuff(one_counter)
Hi,
I have an issue with caching. On each run of the program, numba creates a new cache file, although nothing has changed. The
Is this the result of a known limitation? What could I do to investigate?
I have an issue with caching. On each run of the program, numba creates a new cache file, although nothing has changed. The
NUMBA_DEBUG_CACHE output has no useful output about why the cache is recreated. This amounted to a lot of files taking some GB of storage when this program was executed a couple thousand times.Is this the result of a known limitation? What could I do to investigate?
32 replies
I wanted to see if I could improve performance of some pandas string functions. Specifically
dfq['c'] = dfq['b'] + '|' + dfq['a]') and c_array = dfq['c'].unique(). I saw that numba now supports str operations. But so far my attempts have failed. It would be useful to combine these 2 operations into a single function.
Here is my attempt at the first
@njit(parallel=True)
def concat_pipe_str(prefix, suffix):
for i in range(len(prefix)):
prefix[i] = prefix[i] + '|' + suffix[i]
return prefix
I am familiar with use the signature as a part of the decorator, but there is no numba type for str
I get this error
TypingError: Failed in nopython mode pipeline (step: nopython frontend)
non-precise type array(pyobject, 1d, C)
During: typing of argument at <ipython-input-2-577501f82f1f> (120)
File "<ipython-input-2-577501f82f1f>", line 120:
def concat_pipe_str(prefix, suffix):
for i in range(len(prefix)):
^
I am projecting a bunch of triangles from a 3D triangular mesh onto a 2D detector with many pixels. I currently handle all triangles in parallel with a cuda kernel. Each triangle is only a little work to project because they are small so their projection overlaps with only a few pixels
However, sometimes my algorithm runs into large triangles, which cover almost the entire detector, causing one cuda thread to loop over all detector pixels sequentially. This is so slow it almost locks the machine.
To solve this I think it might help to change this loop to a new kernel call, but in numba I cannot call a kernel from a kernel. Is there some solution to this in numba? Or do I have to translate the entire thing to C++ and write a wrapper?
9 replies