Can't get unnormalized texture coordinates to work on iOS, but MacOS is fine

[whodges]

I have a basic 256x256 test gradient test texture where (0, 0) is 0, (1, 0) is 1, etc. with each texel incrementing by one until (255, 255) is 65535. I'm not doing anything special when I make it:

VkImageCreateInfo imageInfo{};
imageInfo.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
imageInfo.imageType = VK_IMAGE_TYPE_2D;
imageInfo.extent.width = 256;
imageInfo.extent.height = 256;
imageInfo.extent.depth = 1;
imageInfo.mipLevels = 1;
imageInfo.arrayLayers = 1;
imageInfo.format = VK_FORMAT_R8G8B8A8_UINT;
imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
imageInfo.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
imageInfo.samples = VK_SAMPLE_COUNT_1_BIT;
imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;

VkImageViewCreateInfo createInfo{};
createInfo.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
createInfo.image = image;
createInfo.viewType = VK_IMAGE_VIEW_TYPE_2D;
createInfo.format = VK_FORMAT_R8G8B8A8_UINT;
createInfo.components.r = VK_COMPONENT_SWIZZLE_IDENTITY;
createInfo.components.g = VK_COMPONENT_SWIZZLE_IDENTITY;
createInfo.components.b = VK_COMPONENT_SWIZZLE_IDENTITY;
createInfo.components.a = VK_COMPONENT_SWIZZLE_IDENTITY;
createInfo.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
createInfo.subresourceRange.baseMipLevel = 0;
createInfo.subresourceRange.levelCount = 1;
createInfo.subresourceRange.baseArrayLayer = 0;
createInfo.subresourceRange.layerCount = 1;

VkSamplerCreateInfo samplerInfo{};
samplerInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
samplerInfo.magFilter = VK_FILTER_NEAREST;
samplerInfo.minFilter = VK_FILTER_NEAREST;
samplerInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
samplerInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
samplerInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
samplerInfo.anisotropyEnable = VK_FALSE;
samplerInfo.maxAnisotropy = 1;
samplerInfo.unnormalizedCoordinates = VK_TRUE;
samplerInfo.compareEnable = VK_FALSE;
samplerInfo.compareOp = VK_COMPARE_OP_ALWAYS;
samplerInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;

And I have a straightforward shader:

#version 450
#extension GL_EXT_debug_printf : enable
#extension GL_EXT_scalar_block_layout : enable
#extension GL_ARB_separate_shader_objects : enable

precision highp int;

layout(binding = 0) uniform usampler2D texSampler;

layout(location = 0) in vec2 fragTexCoord;

layout(location = 0) out vec4 outColor;

void main()
{
    uvec2 texSize = textureSize(texSampler, 0);
    uvec2 texCoord = uvec2(texSize.x / 2, texSize.y / 2);
    uvec4 texValue = textureLod(texSampler, texCoord, 0);

    outColor = vec4(texValue.r / 255.0, texValue.g / 255.0, texValue.b / 255.0, 1.0);
}

The shader is just sampling the middle of the texture for this example. If you sample any of the first row, things are fine. (0, 0), (1, 0), (2, 0), etc. are all good. But as soon as you try to sample any row other than y = 0 on iOS, the resulting sampled value is (255, 255, 255, 255) (i.e., the edge value), rather than (128, 128, 128, 128). I've tried both my iPhone 14 Pro and my old iPad 5th gen. Note however that this does work on MacOS.

Am I doing something wrong or missing something completely obvious? I've been poking at it for 3 days now. I checked out the MoltenVk code that translate a Vulkan texture sampler request to a Metal one; honestly it all looks fine. Maybe this is an iOS specific thing?

MTLSamplerDescriptor* MVKSampler::newMTLSamplerDescriptor(const VkSamplerCreateInfo* pCreateInfo) {

	MTLSamplerDescriptor* mtlSampDesc = [MTLSamplerDescriptor new];		// retained
	mtlSampDesc.sAddressMode = getMTLSamplerAddressMode(pCreateInfo->addressModeU);
	mtlSampDesc.tAddressMode = getMTLSamplerAddressMode(pCreateInfo->addressModeV);
    if (!pCreateInfo->unnormalizedCoordinates) {
        mtlSampDesc.rAddressMode = getMTLSamplerAddressMode(pCreateInfo->addressModeW);
    }
#if MVK_MACOS_OR_IOS
	mtlSampDesc.borderColorMVK = mvkMTLSamplerBorderColorFromVkBorderColor(pCreateInfo->borderColor);
#endif

	mtlSampDesc.minFilter = mvkMTLSamplerMinMagFilterFromVkFilter(pCreateInfo->minFilter);
	mtlSampDesc.magFilter = mvkMTLSamplerMinMagFilterFromVkFilter(pCreateInfo->magFilter);
    mtlSampDesc.mipFilter = (pCreateInfo->unnormalizedCoordinates
                             ? MTLSamplerMipFilterNotMipmapped
                             : mvkMTLSamplerMipFilterFromVkSamplerMipmapMode(pCreateInfo->mipmapMode));
#if MVK_USE_METAL_PRIVATE_API
	if (getMVKConfig().useMetalPrivateAPI) {
		mtlSampDesc.lodBiasMVK = pCreateInfo->mipLodBias;
	}
#endif
	mtlSampDesc.lodMinClamp = pCreateInfo->minLod;
	mtlSampDesc.lodMaxClamp = pCreateInfo->maxLod;
	mtlSampDesc.maxAnisotropy = (pCreateInfo->anisotropyEnable
								 ? mvkClamp(pCreateInfo->maxAnisotropy, 1.0f, getDeviceProperties().limits.maxSamplerAnisotropy)
								 : 1);
	mtlSampDesc.normalizedCoordinates = !pCreateInfo->unnormalizedCoordinates;
	mtlSampDesc.supportArgumentBuffers = isUsingMetalArgumentBuffers();

	// If compareEnable is true, but dynamic samplers with depth compare are not available
	// on this device, this sampler must only be used as an immutable sampler, and will
	// be automatically hardcoded into the shader MSL. An error will be triggered if this
	// sampler is used to update or push a descriptor.
	if (pCreateInfo->compareEnable && !_requiresConstExprSampler) {
		mtlSampDesc.compareFunctionMVK = mvkMTLCompareFunctionFromVkCompareOp(pCreateInfo->compareOp);
	}

	return mtlSampDesc;
}

Anyway any thoughts are much appreciated. Maybe I'll just be forced to use normalized coordinates; won't be the end of the world.

[whodges]

Ok, so, turns out if I create the texture with the VK_FORMAT_R8G8B8A8_UNORM format rather than VK_FORMAT_R8G8B8A8_UINT, and use a 'sampler' rather than 'usampler', it works - I can use unnormalized coordinates and get the expected texture value. Not sure why it doesn't work with VK_FORMAT_R8G8B8A8_UINT, but I can work with VK_FORMAT_R8G8B8A8_UNORM.