Incorrect translation of spirv with float64 (dvec2)

[dhardy]

KAS's mandlebrot example does not render with 64-bit shaders (--features shader64) since upgrading to wgpu v0.10 (leaving a black background, which is not the clear colour). It runs correctly with 32-bit shaders (almost identical code aside from the fragment shader).

An API trace is attached: trace.tar.gz

Using wgpu::Features::SPIRV_SHADER_PASSTHROUGH the example works correctly, therefore I deduce the error lies with Naga rather than (other parts of) WGPU. The following patch achieves this:

diff --git a/examples/mandlebrot/mandlebrot.rs b/examples/mandlebrot/mandlebrot.rs
index bf1b202b..03e23cbb 100644
--- a/examples/mandlebrot/mandlebrot.rs
+++ b/examples/mandlebrot/mandlebrot.rs
@@ -45,10 +45,12 @@ impl Shaders {
         let vertex = device.create_shader_module(&include_spirv!("shader.vert.spv"));
         // Note: we don't use wgpu::include_spirv since it forces validation,
         // which cannot currently deal with double precision floats (dvec2).
-        let fragment = device.create_shader_module(&wgpu::ShaderModuleDescriptor {
-            label: Some("fragment shader"),
-            source: wgpu::util::make_spirv(FRAG_SHADER),
-        });
+        let fragment = unsafe {
+            device.create_shader_module_spirv(&wgpu::ShaderModuleDescriptorSpirV {
+                label: Some("fragment shader"),
+                source: wgpu::util::make_spirv_raw(FRAG_SHADER),
+            })
+        };
 
         Shaders { vertex, fragment }
     }
@@ -94,7 +96,9 @@ impl CustomPipeBuilder for PipeBuilder {
     fn device_descriptor() -> wgpu::DeviceDescriptor<'static> {
         wgpu::DeviceDescriptor {
             label: None,
-            features: wgpu::Features::PUSH_CONSTANTS | SHADER_FLOAT64,
+            features: wgpu::Features::PUSH_CONSTANTS
+                | wgpu::Features::SPIRV_SHADER_PASSTHROUGH
+                | SHADER_FLOAT64,
             limits: wgpu::Limits {
                 max_push_constant_size: size_of::<PushConstants>().cast(),
                 ..Default::default()

Tested on: Fedora 34 x86_64 RX570

[LPGhatguy]

I think I've run into this. It is not front-end specific; Naga also has trouble with the GLSL frontend and WGSL backend. It always emits f64 instead of f32. I found this while working on Crevice.

This seems to happen with all f64-based data types, including dvec4 and dmat4.

GLSL input:

#version 450

struct FourF64 {
        double x;
        double y;
        double z;
        double w;
};

layout(std140, set = 0, binding = 0) readonly buffer INPUT {
    FourF64 in_data;
};

layout(std140, set = 0, binding = 1) buffer OUTPUT {
    FourF64 out_data;
};

void main() {
    out_data = in_data;
}

WGSL output:

struct FourF64_ {
    x: f32;
    y: f32;
    z: f32;
    w: f32;
};

[[block]]
struct INPUT {
    in_data: FourF64_;
};

[[block]]
struct OUTPUT {
    out_data: FourF64_;
};

[[group(0), binding(0)]]
var<storage> global: INPUT;
[[group(0), binding(1)]]
var<storage, read_write> global1: OUTPUT;

fn main1() {
    let e4: FourF64_ = global.in_data;
    global1.out_data = e4;
    return;
}

[[stage(compute), workgroup_size(1, 1, 1)]]
fn main() {
    main1();
    return;
}

[teoxoy]

I tried the latest version of the example and noticed it's working. I narrowed it down to naga v0.7.2 which gets it going.

It looks like it has been fixed by #1515.