[RFE] passing data in is too slow via slice assign on memory buffer

[muayyad-alsadi]

Motivation

I made a simple WASM image filter, I tested it on the browser, compare it to pure js version and tried wasmer to run the filter.
There was a huge bottleneck when moving the large data to wasm memory. several seconds instead of milliseconds.
more over assigning multidimensional data was even slower (30x) slower.

Proposed solution

• implement faster assign to a memory slice
• Make slice assign on np.ndarray(A, dtype=np.uint8) and array.array('B') as fast as bytearray
• if multidimensional make unwinding/flattening outside python, move it to the c implementation

class Uint8Array(object)
 |  
 |  __setitem__(self, key, value, /)
 |      Set self[key] to value.

this function seems to walk into the slice items one by one

u8buf.__setitem__(slice(start_offset,end_offset,None), value)

Alternatives

Iterating over the WxHxC multidimensional array in python is a huge wast of time
As the only reason for moving to WASM is to do the hot/active loop.

Additional context

import numpy as np
import array
from wasmer import Store, Module, Instance, Memory, MemoryType
from PIL import Image

memory = Memory(Store(), MemoryType(minimum=200))
u8buf = memory.uint8_view(100)

a3d = np.zeros((1920, 1080,4), dtype=np.uint8)
flat = a3d.flatten()
flata = array.array('B', flat)
img = Image.fromarray(a3d, 'RGBA')
pil_a3d = img.getdata()

offset = 200
size = 1920*1080*4

u8buf[offset:offset+size]=bytearray(flat) # <------ `bytearray` type is fast but not fast enough
u8buf[offset:offset+size]=flat # <-------- `np.ndarray`  is slow
u8buf[offset:offset+size]=flata # <-------- `array.array` is slow
u8buf[offset:offset+size]=a3d # <-------- multidimensional `np.ndarray` is impossible
u8buf[offset:offset+size]=pil_a3d # <-------- `PIL img.getdata()` is impossible

NOTE: I'm using Fedora 36, Python 3.10.9, python wasmer 1.1.0, wasmer-compiler-singlepass 1.1.0

[muayyad-alsadi]

this is also slow

d3=img.getdata()
u8buf[offset:offset+size]=bytearray([ d1 for d2 in d3 for d1 in d2 ])

[muayyad-alsadi]

this is also slow

u8buf[offset:offset+size] = bytearray(itertools.chain(*img.getdata()))

note things I marked as slow are at least 30x times slower.