Dependency between Int8 and VulkanMemoryModel capabilities

[149segolte]

Followup from #225, I have been able to use u8 in shaders successfully. But on Apple M2 GPU it does not work due to it not supporting VulkanMemoryModel. Testing this with a GLSL shader, I am able to run the code on macos with just enabling shader_int8 and storage_buffer8_bit_access device features. But when using rust-gpu with capabilities Int8 and StorageBuffer8BitAccess, the code doesn't compile saying VulkanMemoryModel needs to be enabled.

Is this dependency on VulkanMemoryModel something that SPIRV enforces or is there a way for rust-gpu to disable it somehow? I am able to run GLSL with 8 bit integers just fine without VulkanMemoryModel, so I guess it is not a requirement in rust-gpu?

GLSL code:

#version 460
#extension GL_EXT_shader_8bit_storage : require

layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;

layout(set = 0, binding = 0) buffer Data {
    uint8_t data[];
} buf;

void main() {
    uint idx = gl_GlobalInvocationID.x;
    buf.data[idx] *= uint8_t(2);
}

rust-gpu code:

#![no_std]

use glam::UVec3;
use spirv_std::spirv;

#[spirv(compute(threads(64, 1, 1)))]
pub fn compute_shader(
    #[spirv(global_invocation_id)] global_invocation_id: UVec3,
    #[spirv(storage_buffer, descriptor_set = 0, binding = 0)] buf: &mut [u8],
) {
    let idx = (global_invocation_id.x + global_invocation_id.y + global_invocation_id.z) as usize;
    buf[idx] *= 2;
}

Both shaders were tested using vulkano-rs/vulkano, which offers safe bindings on top of ash. I get the following error without enabling VulkanMemoryModel:

Error: a validation error occurred

Caused by:
    create_info.code: uses the SPIR-V capability `VulkanMemoryModel` -- requires one of: device feature `vulkan_memory_model` (Vulkan VUIDs: VUID-VkShaderModuleCreateInfo-pCode-08742)

Enabling it I get:

Error: a validation error occurred

Caused by:
    create_info.enabled_features: contains `vulkan_memory_model`, but this feature is not supported by the physical device

[LegNeato]

Hmm, this appears to be set depending on the target, so when you specify a vulkan target it (naturally) targets the same memory model. There is a way to do the glsl memory model, but you have to specify an opengl target:

rust-gpu/crates/rustc_codegen_spirv/src/target.rs

Line 26 in 698f10a

| TargetEnv::OpenGL_4_5 => MemoryModel::GLSL450,

Here are the opengl targets:

rust-gpu/docs/src/platform-support.md

Line 48 in 698f10a

### OpenGL Targets

Are you using a device that supports vulkan natively or going through the voltenmk translation layer?

I'm again afraid I haven't touched this part of the project but happy to help figure out what is going on.

[LegNeato]

Ah, I see you are on Apple M2 GPU so using MoletnVK. Weird, I would expect it to support the memory model but it does not:

https://github.com/KhronosGroup/MoltenVK/blob/83510e0f3835c3c43651dda087305abc42572e17/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm#L238

[149segolte]

I see. The translation layer does not implement certain things. It does seem weird to provide a target and then not implement a memory model for it.

Anyways, for anyone else looking to use these features, for macos I set:

shader-target = "spirv-unknown-opengl4.5"
capability = ["Int8", "UniformAndStorageBuffer8BitAccess"]
extension = ["SPV_KHR_8bit_storage", "SPV_KHR_storage_buffer_storage_class"]

For a nvidia GPU, I use:

shader-target = "spirv-unknown-vulkan1.2"
capability = ["Int8", "UniformAndStorageBuffer8BitAccess"]

[Firestar99]

Afaik the VulkanMemoryModel formally describes how atomics work on GPUs, as before they've been severely underspec'ed. And we're just following the VulkanMemoryModel with our atomics, and implicitly always require it. But it actually doesn't add any new capabilities to the shaders themselves, so it missing doesn't really change anything.