stb_image_resize2: assert/crash on resizing 2x2 to 1x1 images with wrap

[aras-p]

Describe the bug When resizing very small images (like 2x2 into 1x1) with STBIR_EDGE_WRAP and STBIR_FILTER_DEFAULT, there's a debug assert and a possible later crash.

stb_image_resize2.h version v2.04.

To Reproduce

This code:

#include <stdint.h>
#define STB_IMAGE_RESIZE_IMPLEMENTATION
#include "stb_image_resize2.h"

int main()
{
    int srcWidth = 2, srcHeight = 2, dstWidth = 1, dstHeight = 1;
    uint8_t* src = new uint8_t[srcWidth * srcHeight * 4];
    memset(src, 0xAB, srcWidth * srcHeight * 4);
    uint8_t* dst = new uint8_t[dstWidth * dstHeight * 4];
    stbir_resize(src, srcWidth, srcHeight, 0,
        dst, dstWidth, dstHeight, 0,
        STBIR_RGBA, STBIR_TYPE_UINT8,
        STBIR_EDGE_WRAP, STBIR_FILTER_DEFAULT);
    delete[] src;
    delete[] dst;
}

Triggers STBIR_ASSERT( info->ring_buffer_num_entries <= info->alloc_ring_buffer_num_entries ); at line 7028. info->ring_buffer_num_entries is 6, info->alloc_ring_buffer_num_entries is 5.

Changing wrap mode to clamp, or filter to box does not trigger the assert/crash.

Expected behavior I'd expect no asserts and no crashes.

[jeffrbig2]

I will scope this out, this weekend!

[jeffatrad]

Also, I should probably eventually special case 16x16 and smaller downsamples - that could probably be quite a bit faster without going through all the general machinery. Although these are already fast, I suppose.

[nothings]

I wouldn't bother special-casing them; it's hard to imagine a scenario where they'd be a critical path if they're already reasonably fast.

[jeffrbig2]

OK, fixed in #1603

[nothings]

Fixed by #1605.